Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procarray.c
4 : * POSTGRES process array code.
5 : *
6 : *
7 : * This module maintains arrays of PGPROC substructures, as well as associated
8 : * arrays in ProcGlobal, for all active backends. Although there are several
9 : * uses for this, the principal one is as a means of determining the set of
10 : * currently running transactions.
11 : *
12 : * Because of various subtle race conditions it is critical that a backend
13 : * hold the correct locks while setting or clearing its xid (in
14 : * ProcGlobal->xids[]/MyProc->xid). See notes in
15 : * src/backend/access/transam/README.
16 : *
17 : * The process arrays now also include structures representing prepared
18 : * transactions. The xid and subxids fields of these are valid, as are the
19 : * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 : * at need by checking for pid == 0.
21 : *
22 : * During hot standby, we also keep a list of XIDs representing transactions
23 : * that are known to be running on the primary (or more precisely, were running
24 : * as of the current point in the WAL stream). This list is kept in the
25 : * KnownAssignedXids array, and is updated by watching the sequence of
26 : * arriving XIDs. This is necessary because if we leave those XIDs out of
27 : * snapshots taken for standby queries, then they will appear to be already
28 : * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 : * array represents standby processes, which by definition are not running
30 : * transactions that have XIDs.
31 : *
32 : * It is perhaps possible for a backend on the primary to terminate without
33 : * writing an abort record for its transaction. While that shouldn't really
34 : * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 : * ourselves by pruning the array when a valid list of running XIDs arrives.
36 : *
37 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/subtrans.h"
51 : #include "access/transam.h"
52 : #include "access/twophase.h"
53 : #include "access/xact.h"
54 : #include "access/xlogutils.h"
55 : #include "catalog/catalog.h"
56 : #include "catalog/pg_authid.h"
57 : #include "miscadmin.h"
58 : #include "pgstat.h"
59 : #include "postmaster/bgworker.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/procsignal.h"
64 : #include "storage/subsystems.h"
65 : #include "utils/acl.h"
66 : #include "utils/builtins.h"
67 : #include "utils/injection_point.h"
68 : #include "utils/lsyscache.h"
69 : #include "utils/rel.h"
70 : #include "utils/snapmgr.h"
71 : #include "utils/wait_event.h"
72 :
73 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
74 :
75 : /* Our shared memory area */
76 : typedef struct ProcArrayStruct
77 : {
78 : int numProcs; /* number of valid procs entries */
79 : int maxProcs; /* allocated size of procs array */
80 :
81 : /*
82 : * Known assigned XIDs handling
83 : */
84 : int maxKnownAssignedXids; /* allocated size of array */
85 : int numKnownAssignedXids; /* current # of valid entries */
86 : int tailKnownAssignedXids; /* index of oldest valid element */
87 : int headKnownAssignedXids; /* index of newest element, + 1 */
88 :
89 : /*
90 : * Highest subxid that has been removed from KnownAssignedXids array to
91 : * prevent overflow; or InvalidTransactionId if none. We track this for
92 : * similar reasons to tracking overflowing cached subxids in PGPROC
93 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
94 : * lock to read it.
95 : */
96 : TransactionId lastOverflowedXid;
97 :
98 : /* oldest xmin of any replication slot */
99 : TransactionId replication_slot_xmin;
100 : /* oldest catalog xmin of any replication slot */
101 : TransactionId replication_slot_catalog_xmin;
102 :
103 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
104 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
105 : } ProcArrayStruct;
106 :
107 : static void ProcArrayShmemRequest(void *arg);
108 : static void ProcArrayShmemInit(void *arg);
109 : static void ProcArrayShmemAttach(void *arg);
110 :
111 : static ProcArrayStruct *procArray;
112 :
113 : const struct ShmemCallbacks ProcArrayShmemCallbacks = {
114 : .request_fn = ProcArrayShmemRequest,
115 : .init_fn = ProcArrayShmemInit,
116 : .attach_fn = ProcArrayShmemAttach,
117 : };
118 :
119 : /*
120 : * State for the GlobalVisTest* family of functions. Those functions can
121 : * e.g. be used to decide if a deleted row can be removed without violating
122 : * MVCC semantics: If the deleted row's xmax is not considered to be running
123 : * by anyone, the row can be removed.
124 : *
125 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
126 : * cutoff XID while building a snapshot (looking at the frequently changing
127 : * xmins scales badly). Instead we compute two boundaries while building the
128 : * snapshot:
129 : *
130 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
131 : * definitely_needed are definitely still visible.
132 : *
133 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
134 : * definitely be removed
135 : *
136 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
137 : * && XID < definitely_needed), the boundaries can be recomputed (using
138 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
139 : * maintaining an accurate value all the time.
140 : *
141 : * As it is not cheap to compute accurate boundaries, we limit the number of
142 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
143 : *
144 : *
145 : * There are three backend lifetime instances of this struct, optimized for
146 : * different types of relations. As e.g. a normal user defined table in one
147 : * database is inaccessible to backends connected to another database, a test
148 : * specific to a relation can be more aggressive than a test for a shared
149 : * relation. Currently we track four different states:
150 : *
151 : * 1) GlobalVisSharedRels, which only considers an XID's
152 : * effects visible-to-everyone if neither snapshots in any database, nor a
153 : * replication slot's xmin, nor a replication slot's catalog_xmin might
154 : * still consider XID as running.
155 : *
156 : * 2) GlobalVisCatalogRels, which only considers an XID's
157 : * effects visible-to-everyone if neither snapshots in the current
158 : * database, nor a replication slot's xmin, nor a replication slot's
159 : * catalog_xmin might still consider XID as running.
160 : *
161 : * I.e. the difference to GlobalVisSharedRels is that
162 : * snapshot in other databases are ignored.
163 : *
164 : * 3) GlobalVisDataRels, which only considers an XID's
165 : * effects visible-to-everyone if neither snapshots in the current
166 : * database, nor a replication slot's xmin consider XID as running.
167 : *
168 : * I.e. the difference to GlobalVisCatalogRels is that
169 : * replication slot's catalog_xmin is not taken into account.
170 : *
171 : * 4) GlobalVisTempRels, which only considers the current session, as temp
172 : * tables are not visible to other sessions.
173 : *
174 : * GlobalVisTestFor(relation) returns the appropriate state
175 : * for the relation.
176 : *
177 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
178 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
179 : * prevent maybe_needed to become old enough after the GetSnapshotData()
180 : * call.
181 : *
182 : * The typedef is in the header.
183 : */
184 : struct GlobalVisState
185 : {
186 : /* XIDs >= are considered running by some backend */
187 : FullTransactionId definitely_needed;
188 :
189 : /* XIDs < are not considered to be running by any backend */
190 : FullTransactionId maybe_needed;
191 : };
192 :
193 : /*
194 : * Result of ComputeXidHorizons().
195 : */
196 : typedef struct ComputeXidHorizonsResult
197 : {
198 : /*
199 : * The value of TransamVariables->latestCompletedXid when
200 : * ComputeXidHorizons() held ProcArrayLock.
201 : */
202 : FullTransactionId latest_completed;
203 :
204 : /*
205 : * The same for procArray->replication_slot_xmin and
206 : * procArray->replication_slot_catalog_xmin.
207 : */
208 : TransactionId slot_xmin;
209 : TransactionId slot_catalog_xmin;
210 :
211 : /*
212 : * Oldest xid that any backend might still consider running. This needs to
213 : * include processes running VACUUM, in contrast to the normal visibility
214 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
215 : * determining visibility, but doesn't care about rows above its xmin to
216 : * be removed.
217 : *
218 : * This likely should only be needed to determine whether pg_subtrans can
219 : * be truncated. It currently includes the effects of replication slots,
220 : * for historical reasons. But that could likely be changed.
221 : */
222 : TransactionId oldest_considered_running;
223 :
224 : /*
225 : * Oldest xid for which deleted tuples need to be retained in shared
226 : * tables.
227 : *
228 : * This includes the effects of replication slots. If that's not desired,
229 : * look at shared_oldest_nonremovable_raw;
230 : */
231 : TransactionId shared_oldest_nonremovable;
232 :
233 : /*
234 : * Oldest xid that may be necessary to retain in shared tables. This is
235 : * the same as shared_oldest_nonremovable, except that is not affected by
236 : * replication slot's catalog_xmin.
237 : *
238 : * This is mainly useful to be able to send the catalog_xmin to upstream
239 : * streaming replication servers via hot_standby_feedback, so they can
240 : * apply the limit only when accessing catalog tables.
241 : */
242 : TransactionId shared_oldest_nonremovable_raw;
243 :
244 : /*
245 : * Oldest xid for which deleted tuples need to be retained in non-shared
246 : * catalog tables.
247 : */
248 : TransactionId catalog_oldest_nonremovable;
249 :
250 : /*
251 : * Oldest xid for which deleted tuples need to be retained in normal user
252 : * defined tables.
253 : */
254 : TransactionId data_oldest_nonremovable;
255 :
256 : /*
257 : * Oldest xid for which deleted tuples need to be retained in this
258 : * session's temporary tables.
259 : */
260 : TransactionId temp_oldest_nonremovable;
261 : } ComputeXidHorizonsResult;
262 :
263 : /*
264 : * Return value for GlobalVisHorizonKindForRel().
265 : */
266 : typedef enum GlobalVisHorizonKind
267 : {
268 : VISHORIZON_SHARED,
269 : VISHORIZON_CATALOG,
270 : VISHORIZON_DATA,
271 : VISHORIZON_TEMP,
272 : } GlobalVisHorizonKind;
273 :
274 : /*
275 : * Reason codes for KnownAssignedXidsCompress().
276 : */
277 : typedef enum KAXCompressReason
278 : {
279 : KAX_NO_SPACE, /* need to free up space at array end */
280 : KAX_PRUNE, /* we just pruned old entries */
281 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
282 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
283 : } KAXCompressReason;
284 :
285 : static PGPROC *allProcs;
286 :
287 : /*
288 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
289 : */
290 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
291 :
292 : /*
293 : * Bookkeeping for tracking emulated transactions in recovery
294 : */
295 :
296 : static TransactionId *KnownAssignedXids;
297 :
298 : static bool *KnownAssignedXidsValid;
299 :
300 : static TransactionId latestObservedXid = InvalidTransactionId;
301 :
302 : /*
303 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
304 : * the highest xid that might still be running that we don't have in
305 : * KnownAssignedXids.
306 : */
307 : static TransactionId standbySnapshotPendingXmin;
308 :
309 : /*
310 : * State for visibility checks on different types of relations. See struct
311 : * GlobalVisState for details. As shared, catalog, normal and temporary
312 : * relations can have different horizons, one such state exists for each.
313 : */
314 : static GlobalVisState GlobalVisSharedRels;
315 : static GlobalVisState GlobalVisCatalogRels;
316 : static GlobalVisState GlobalVisDataRels;
317 : static GlobalVisState GlobalVisTempRels;
318 :
319 : /*
320 : * This backend's RecentXmin at the last time the accurate xmin horizon was
321 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
322 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
323 : */
324 : static TransactionId ComputeXidHorizonsResultLastXmin;
325 :
326 : #ifdef XIDCACHE_DEBUG
327 :
328 : /* counters for XidCache measurement */
329 : static long xc_by_recent_xmin = 0;
330 : static long xc_by_known_xact = 0;
331 : static long xc_by_my_xact = 0;
332 : static long xc_by_latest_xid = 0;
333 : static long xc_by_main_xid = 0;
334 : static long xc_by_child_xid = 0;
335 : static long xc_by_known_assigned = 0;
336 : static long xc_no_overflow = 0;
337 : static long xc_slow_answer = 0;
338 :
339 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
340 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
341 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
342 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
343 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
344 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
345 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
346 : #define xc_no_overflow_inc() (xc_no_overflow++)
347 : #define xc_slow_answer_inc() (xc_slow_answer++)
348 :
349 : static void DisplayXidCache(void);
350 : #else /* !XIDCACHE_DEBUG */
351 :
352 : #define xc_by_recent_xmin_inc() ((void) 0)
353 : #define xc_by_known_xact_inc() ((void) 0)
354 : #define xc_by_my_xact_inc() ((void) 0)
355 : #define xc_by_latest_xid_inc() ((void) 0)
356 : #define xc_by_main_xid_inc() ((void) 0)
357 : #define xc_by_child_xid_inc() ((void) 0)
358 : #define xc_by_known_assigned_inc() ((void) 0)
359 : #define xc_no_overflow_inc() ((void) 0)
360 : #define xc_slow_answer_inc() ((void) 0)
361 : #endif /* XIDCACHE_DEBUG */
362 :
363 : /* Primitives for KnownAssignedXids array handling for standby */
364 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
365 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
366 : bool exclusive_lock);
367 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
368 : static bool KnownAssignedXidExists(TransactionId xid);
369 : static void KnownAssignedXidsRemove(TransactionId xid);
370 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
371 : TransactionId *subxids);
372 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
373 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
374 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
375 : TransactionId *xmin,
376 : TransactionId xmax);
377 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
378 : static void KnownAssignedXidsDisplay(int trace_level);
379 : static void KnownAssignedXidsReset(void);
380 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
381 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
382 : static void MaintainLatestCompletedXid(TransactionId latestXid);
383 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
384 :
385 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
386 : TransactionId xid);
387 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
388 :
389 : /*
390 : * Register the shared PGPROC array during postmaster startup.
391 : */
392 : static void
393 1245 : ProcArrayShmemRequest(void *arg)
394 : {
395 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
396 :
397 : /*
398 : * During Hot Standby processing we have a data structure called
399 : * KnownAssignedXids, created in shared memory. Local data structures are
400 : * also created in various backends during GetSnapshotData(),
401 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
402 : * main structures created in those functions must be identically sized,
403 : * since we may at times copy the whole of the data structures around. We
404 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
405 : *
406 : * Ideally we'd only create this structure if we were actually doing hot
407 : * standby in the current run, but we don't know that yet at the time
408 : * shared memory is being set up.
409 : */
410 : #define TOTAL_MAX_CACHED_SUBXIDS \
411 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
412 :
413 1245 : if (EnableHotStandby)
414 : {
415 1239 : ShmemRequestStruct(.name = "KnownAssignedXids",
416 : .size = mul_size(sizeof(TransactionId), TOTAL_MAX_CACHED_SUBXIDS),
417 : .ptr = (void **) &KnownAssignedXids,
418 : );
419 :
420 1239 : ShmemRequestStruct(.name = "KnownAssignedXidsValid",
421 : .size = mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
422 : .ptr = (void **) &KnownAssignedXidsValid,
423 : );
424 : }
425 :
426 : /* Register the ProcArray shared structure */
427 1245 : ShmemRequestStruct(.name = "Proc Array",
428 : .size = add_size(offsetof(ProcArrayStruct, pgprocnos),
429 : mul_size(sizeof(int), PROCARRAY_MAXPROCS)),
430 : .ptr = (void **) &procArray,
431 : );
432 1245 : }
433 :
434 : /*
435 : * Initialize the shared PGPROC array during postmaster startup.
436 : */
437 : static void
438 1242 : ProcArrayShmemInit(void *arg)
439 : {
440 1242 : procArray->numProcs = 0;
441 1242 : procArray->maxProcs = PROCARRAY_MAXPROCS;
442 1242 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
443 1242 : procArray->numKnownAssignedXids = 0;
444 1242 : procArray->tailKnownAssignedXids = 0;
445 1242 : procArray->headKnownAssignedXids = 0;
446 1242 : procArray->lastOverflowedXid = InvalidTransactionId;
447 1242 : procArray->replication_slot_xmin = InvalidTransactionId;
448 1242 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
449 1242 : TransamVariables->xactCompletionCount = 1;
450 :
451 1242 : allProcs = ProcGlobal->allProcs;
452 1242 : }
453 :
454 : static void
455 0 : ProcArrayShmemAttach(void *arg)
456 : {
457 0 : allProcs = ProcGlobal->allProcs;
458 0 : }
459 :
460 : /*
461 : * Add the specified PGPROC to the shared array.
462 : */
463 : void
464 20542 : ProcArrayAdd(PGPROC *proc)
465 : {
466 20542 : int pgprocno = GetNumberFromPGProc(proc);
467 20542 : ProcArrayStruct *arrayP = procArray;
468 : int index;
469 : int movecount;
470 :
471 : /* See ProcGlobal comment explaining why both locks are held */
472 20542 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
473 20542 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
474 :
475 20542 : if (arrayP->numProcs >= arrayP->maxProcs)
476 : {
477 : /*
478 : * Oops, no room. (This really shouldn't happen, since there is a
479 : * fixed supply of PGPROC structs too, and so we should have failed
480 : * earlier.)
481 : */
482 0 : ereport(FATAL,
483 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
484 : errmsg("sorry, too many clients already")));
485 : }
486 :
487 : /*
488 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
489 : * locality of references much better. This is useful while traversing the
490 : * ProcArray because there is an increased likelihood of finding the next
491 : * PGPROC structure in the cache.
492 : *
493 : * Since the occurrence of adding/removing a proc is much lower than the
494 : * access to the ProcArray itself, the overhead should be marginal
495 : */
496 52366 : for (index = 0; index < arrayP->numProcs; index++)
497 : {
498 46827 : int this_procno = arrayP->pgprocnos[index];
499 :
500 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
501 : Assert(allProcs[this_procno].pgxactoff == index);
502 :
503 : /* If we have found our right position in the array, break */
504 46827 : if (this_procno > pgprocno)
505 15003 : break;
506 : }
507 :
508 20542 : movecount = arrayP->numProcs - index;
509 20542 : memmove(&arrayP->pgprocnos[index + 1],
510 20542 : &arrayP->pgprocnos[index],
511 : movecount * sizeof(*arrayP->pgprocnos));
512 20542 : memmove(&ProcGlobal->xids[index + 1],
513 20542 : &ProcGlobal->xids[index],
514 : movecount * sizeof(*ProcGlobal->xids));
515 20542 : memmove(&ProcGlobal->subxidStates[index + 1],
516 20542 : &ProcGlobal->subxidStates[index],
517 : movecount * sizeof(*ProcGlobal->subxidStates));
518 20542 : memmove(&ProcGlobal->statusFlags[index + 1],
519 20542 : &ProcGlobal->statusFlags[index],
520 : movecount * sizeof(*ProcGlobal->statusFlags));
521 :
522 20542 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
523 20542 : proc->pgxactoff = index;
524 20542 : ProcGlobal->xids[index] = proc->xid;
525 20542 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
526 20542 : ProcGlobal->statusFlags[index] = proc->statusFlags;
527 :
528 20542 : arrayP->numProcs++;
529 :
530 : /* adjust pgxactoff for all following PGPROCs */
531 20542 : index++;
532 54546 : for (; index < arrayP->numProcs; index++)
533 : {
534 34004 : int procno = arrayP->pgprocnos[index];
535 :
536 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
537 : Assert(allProcs[procno].pgxactoff == index - 1);
538 :
539 34004 : allProcs[procno].pgxactoff = index;
540 : }
541 :
542 : /*
543 : * Release in reversed acquisition order, to reduce frequency of having to
544 : * wait for XidGenLock while holding ProcArrayLock.
545 : */
546 20542 : LWLockRelease(XidGenLock);
547 20542 : LWLockRelease(ProcArrayLock);
548 20542 : }
549 :
550 : /*
551 : * Remove the specified PGPROC from the shared array.
552 : *
553 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
554 : * array, and thus causing it to appear as "not running" anymore. In this
555 : * case we must advance latestCompletedXid. (This is essentially the same
556 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
557 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
558 : * twophase.c depends on the latter.)
559 : */
560 : void
561 20514 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
562 : {
563 20514 : ProcArrayStruct *arrayP = procArray;
564 : int myoff;
565 : int movecount;
566 :
567 : #ifdef XIDCACHE_DEBUG
568 : /* dump stats at backend shutdown, but not prepared-xact end */
569 : if (proc->pid != 0)
570 : DisplayXidCache();
571 : #endif
572 :
573 : /* See ProcGlobal comment explaining why both locks are held */
574 20514 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
575 20514 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
576 :
577 20514 : myoff = proc->pgxactoff;
578 :
579 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
580 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
581 :
582 20514 : if (TransactionIdIsValid(latestXid))
583 : {
584 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
585 :
586 : /* Advance global latestCompletedXid while holding the lock */
587 338 : MaintainLatestCompletedXid(latestXid);
588 :
589 : /* Same with xactCompletionCount */
590 338 : TransamVariables->xactCompletionCount++;
591 :
592 338 : ProcGlobal->xids[myoff] = InvalidTransactionId;
593 338 : ProcGlobal->subxidStates[myoff].overflowed = false;
594 338 : ProcGlobal->subxidStates[myoff].count = 0;
595 : }
596 : else
597 : {
598 : /* Shouldn't be trying to remove a live transaction here */
599 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
600 : }
601 :
602 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
603 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
604 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
605 :
606 20514 : ProcGlobal->statusFlags[myoff] = 0;
607 :
608 : /* Keep the PGPROC array sorted. See notes above */
609 20514 : movecount = arrayP->numProcs - myoff - 1;
610 20514 : memmove(&arrayP->pgprocnos[myoff],
611 20514 : &arrayP->pgprocnos[myoff + 1],
612 : movecount * sizeof(*arrayP->pgprocnos));
613 20514 : memmove(&ProcGlobal->xids[myoff],
614 20514 : &ProcGlobal->xids[myoff + 1],
615 : movecount * sizeof(*ProcGlobal->xids));
616 20514 : memmove(&ProcGlobal->subxidStates[myoff],
617 20514 : &ProcGlobal->subxidStates[myoff + 1],
618 : movecount * sizeof(*ProcGlobal->subxidStates));
619 20514 : memmove(&ProcGlobal->statusFlags[myoff],
620 20514 : &ProcGlobal->statusFlags[myoff + 1],
621 : movecount * sizeof(*ProcGlobal->statusFlags));
622 :
623 20514 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
624 20514 : arrayP->numProcs--;
625 :
626 : /*
627 : * Adjust pgxactoff of following procs for removed PGPROC (note that
628 : * numProcs already has been decremented).
629 : */
630 58577 : for (int index = myoff; index < arrayP->numProcs; index++)
631 : {
632 38063 : int procno = arrayP->pgprocnos[index];
633 :
634 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
635 : Assert(allProcs[procno].pgxactoff - 1 == index);
636 :
637 38063 : allProcs[procno].pgxactoff = index;
638 : }
639 :
640 : /*
641 : * Release in reversed acquisition order, to reduce frequency of having to
642 : * wait for XidGenLock while holding ProcArrayLock.
643 : */
644 20514 : LWLockRelease(XidGenLock);
645 20514 : LWLockRelease(ProcArrayLock);
646 20514 : }
647 :
648 :
649 : /*
650 : * ProcArrayEndTransaction -- mark a transaction as no longer running
651 : *
652 : * This is used interchangeably for commit and abort cases. The transaction
653 : * commit/abort must already be reported to WAL and pg_xact.
654 : *
655 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
656 : * latestXid is the latest Xid among the transaction's main XID and
657 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
658 : * the caller to pass latestXid, instead of computing it from the PGPROC's
659 : * contents, because the subxid information in the PGPROC might be
660 : * incomplete.)
661 : */
662 : void
663 650703 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
664 : {
665 650703 : if (TransactionIdIsValid(latestXid))
666 : {
667 : /*
668 : * We must lock ProcArrayLock while clearing our advertised XID, so
669 : * that we do not exit the set of "running" transactions while someone
670 : * else is taking a snapshot. See discussion in
671 : * src/backend/access/transam/README.
672 : */
673 : Assert(TransactionIdIsValid(proc->xid));
674 :
675 : /*
676 : * If we can immediately acquire ProcArrayLock, we clear our own XID
677 : * and release the lock. If not, use group XID clearing to improve
678 : * efficiency.
679 : */
680 168334 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
681 : {
682 168191 : ProcArrayEndTransactionInternal(proc, latestXid);
683 168191 : LWLockRelease(ProcArrayLock);
684 : }
685 : else
686 143 : ProcArrayGroupClearXid(proc, latestXid);
687 : }
688 : else
689 : {
690 : /*
691 : * If we have no XID, we don't need to lock, since we won't affect
692 : * anyone else's calculation of a snapshot. We might change their
693 : * estimate of global xmin, but that's OK.
694 : */
695 : Assert(!TransactionIdIsValid(proc->xid));
696 : Assert(proc->subxidStatus.count == 0);
697 : Assert(!proc->subxidStatus.overflowed);
698 :
699 482369 : proc->vxid.lxid = InvalidLocalTransactionId;
700 482369 : proc->xmin = InvalidTransactionId;
701 :
702 : /* be sure this is cleared in abort */
703 482369 : proc->delayChkptFlags = 0;
704 :
705 : /* must be cleared with xid/xmin: */
706 : /* avoid unnecessarily dirtying shared cachelines */
707 482369 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
708 : {
709 : Assert(!LWLockHeldByMe(ProcArrayLock));
710 125714 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
711 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
712 125714 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
713 125714 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
714 125714 : LWLockRelease(ProcArrayLock);
715 : }
716 : }
717 650703 : }
718 :
719 : /*
720 : * Mark a write transaction as no longer running.
721 : *
722 : * We don't do any locking here; caller must handle that.
723 : */
724 : static inline void
725 168334 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
726 : {
727 168334 : int pgxactoff = proc->pgxactoff;
728 :
729 : /*
730 : * Note: we need exclusive lock here because we're going to change other
731 : * processes' PGPROC entries.
732 : */
733 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
734 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
735 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
736 :
737 168334 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
738 168334 : proc->xid = InvalidTransactionId;
739 168334 : proc->vxid.lxid = InvalidLocalTransactionId;
740 168334 : proc->xmin = InvalidTransactionId;
741 :
742 : /* be sure this is cleared in abort */
743 168334 : proc->delayChkptFlags = 0;
744 :
745 : /* must be cleared with xid/xmin: */
746 : /* avoid unnecessarily dirtying shared cachelines */
747 168334 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
748 : {
749 923 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
750 923 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
751 : }
752 :
753 : /* Clear the subtransaction-XID cache too while holding the lock */
754 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
755 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
756 168334 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
757 : {
758 604 : ProcGlobal->subxidStates[pgxactoff].count = 0;
759 604 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
760 604 : proc->subxidStatus.count = 0;
761 604 : proc->subxidStatus.overflowed = false;
762 : }
763 :
764 : /* Also advance global latestCompletedXid while holding the lock */
765 168334 : MaintainLatestCompletedXid(latestXid);
766 :
767 : /* Same with xactCompletionCount */
768 168334 : TransamVariables->xactCompletionCount++;
769 168334 : }
770 :
771 : /*
772 : * ProcArrayGroupClearXid -- group XID clearing
773 : *
774 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
775 : * commit time, add ourselves to a list of processes that need their XIDs
776 : * cleared. The first process to add itself to the list will acquire
777 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
778 : * on behalf of all group members. This avoids a great deal of contention
779 : * around ProcArrayLock when many processes are trying to commit at once,
780 : * since the lock need not be repeatedly handed off from one committing
781 : * process to the next.
782 : */
783 : static void
784 143 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
785 : {
786 143 : int pgprocno = GetNumberFromPGProc(proc);
787 143 : PROC_HDR *procglobal = ProcGlobal;
788 : uint32 nextidx;
789 : uint32 wakeidx;
790 :
791 : /* We should definitely have an XID to clear. */
792 : Assert(TransactionIdIsValid(proc->xid));
793 :
794 : /* Add ourselves to the list of processes needing a group XID clear. */
795 143 : proc->procArrayGroupMember = true;
796 143 : proc->procArrayGroupMemberXid = latestXid;
797 143 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
798 : while (true)
799 : {
800 143 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
801 :
802 143 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
803 : &nextidx,
804 : (uint32) pgprocno))
805 143 : break;
806 : }
807 :
808 : /*
809 : * If the list was not empty, the leader will clear our XID. It is
810 : * impossible to have followers without a leader because the first process
811 : * that has added itself to the list will always have nextidx as
812 : * INVALID_PROC_NUMBER.
813 : */
814 143 : if (nextidx != INVALID_PROC_NUMBER)
815 : {
816 10 : int extraWaits = 0;
817 :
818 : /* Sleep until the leader clears our XID. */
819 10 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
820 : for (;;)
821 : {
822 : /* acts as a read barrier */
823 10 : PGSemaphoreLock(proc->sem);
824 10 : if (!proc->procArrayGroupMember)
825 10 : break;
826 0 : extraWaits++;
827 : }
828 10 : pgstat_report_wait_end();
829 :
830 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
831 :
832 : /* Fix semaphore count for any absorbed wakeups */
833 10 : while (extraWaits-- > 0)
834 0 : PGSemaphoreUnlock(proc->sem);
835 10 : return;
836 : }
837 :
838 : /* We are the leader. Acquire the lock on behalf of everyone. */
839 133 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
840 :
841 : /*
842 : * Now that we've got the lock, clear the list of processes waiting for
843 : * group XID clearing, saving a pointer to the head of the list. Trying
844 : * to pop elements one at a time could lead to an ABA problem.
845 : */
846 133 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
847 : INVALID_PROC_NUMBER);
848 :
849 : /* Remember head of list so we can perform wakeups after dropping lock. */
850 133 : wakeidx = nextidx;
851 :
852 : /* Walk the list and clear all XIDs. */
853 276 : while (nextidx != INVALID_PROC_NUMBER)
854 : {
855 143 : PGPROC *nextproc = &allProcs[nextidx];
856 :
857 143 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
858 :
859 : /* Move to next proc in list. */
860 143 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
861 : }
862 :
863 : /* We're done with the lock now. */
864 133 : LWLockRelease(ProcArrayLock);
865 :
866 : /*
867 : * Now that we've released the lock, go back and wake everybody up. We
868 : * don't do this under the lock so as to keep lock hold times to a
869 : * minimum. The system calls we need to perform to wake other processes
870 : * up are probably much slower than the simple memory writes we did while
871 : * holding the lock.
872 : */
873 276 : while (wakeidx != INVALID_PROC_NUMBER)
874 : {
875 143 : PGPROC *nextproc = &allProcs[wakeidx];
876 :
877 143 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
878 143 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
879 :
880 : /* ensure all previous writes are visible before follower continues. */
881 143 : pg_write_barrier();
882 :
883 143 : nextproc->procArrayGroupMember = false;
884 :
885 143 : if (nextproc != MyProc)
886 10 : PGSemaphoreUnlock(nextproc->sem);
887 : }
888 : }
889 :
890 : /*
891 : * ProcArrayClearTransaction -- clear the transaction fields
892 : *
893 : * This is used after successfully preparing a 2-phase transaction. We are
894 : * not actually reporting the transaction's XID as no longer running --- it
895 : * will still appear as running because the 2PC's gxact is in the ProcArray
896 : * too. We just have to clear out our own PGPROC.
897 : */
898 : void
899 333 : ProcArrayClearTransaction(PGPROC *proc)
900 : {
901 : int pgxactoff;
902 :
903 : /*
904 : * Currently we need to lock ProcArrayLock exclusively here, as we
905 : * increment xactCompletionCount below. We also need it at least in shared
906 : * mode for pgproc->pgxactoff to stay the same below.
907 : *
908 : * We could however, as this action does not actually change anyone's view
909 : * of the set of running XIDs (our entry is duplicate with the gxact that
910 : * has already been inserted into the ProcArray), lower the lock level to
911 : * shared if we were to make xactCompletionCount an atomic variable. But
912 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
913 : * enough for this not to be the bottleneck. If it ever becomes a
914 : * bottleneck it may also be worth considering to combine this with the
915 : * subsequent ProcArrayRemove()
916 : */
917 333 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
918 :
919 333 : pgxactoff = proc->pgxactoff;
920 :
921 333 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
922 333 : proc->xid = InvalidTransactionId;
923 :
924 333 : proc->vxid.lxid = InvalidLocalTransactionId;
925 333 : proc->xmin = InvalidTransactionId;
926 :
927 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
928 : Assert(!proc->delayChkptFlags);
929 :
930 : /*
931 : * Need to increment completion count even though transaction hasn't
932 : * really committed yet. The reason for that is that GetSnapshotData()
933 : * omits the xid of the current transaction, thus without the increment we
934 : * otherwise could end up reusing the snapshot later. Which would be bad,
935 : * because it might not count the prepared transaction as running.
936 : */
937 333 : TransamVariables->xactCompletionCount++;
938 :
939 : /* Clear the subtransaction-XID cache too */
940 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
941 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
942 333 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
943 : {
944 108 : ProcGlobal->subxidStates[pgxactoff].count = 0;
945 108 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
946 108 : proc->subxidStatus.count = 0;
947 108 : proc->subxidStatus.overflowed = false;
948 : }
949 :
950 333 : LWLockRelease(ProcArrayLock);
951 333 : }
952 :
953 : /*
954 : * Update TransamVariables->latestCompletedXid to point to latestXid if
955 : * currently older.
956 : */
957 : static void
958 169532 : MaintainLatestCompletedXid(TransactionId latestXid)
959 : {
960 169532 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
961 :
962 : Assert(FullTransactionIdIsValid(cur_latest));
963 : Assert(!RecoveryInProgress());
964 : Assert(LWLockHeldByMe(ProcArrayLock));
965 :
966 169532 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
967 : {
968 153364 : TransamVariables->latestCompletedXid =
969 153364 : FullXidRelativeTo(cur_latest, latestXid);
970 : }
971 :
972 : Assert(IsBootstrapProcessingMode() ||
973 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
974 169532 : }
975 :
976 : /*
977 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
978 : */
979 : static void
980 24804 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
981 : {
982 24804 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
983 : FullTransactionId rel;
984 :
985 : Assert(AmStartupProcess() || !IsUnderPostmaster);
986 : Assert(LWLockHeldByMe(ProcArrayLock));
987 :
988 : /*
989 : * Need a FullTransactionId to compare latestXid with. Can't rely on
990 : * latestCompletedXid to be initialized in recovery. But in recovery it's
991 : * safe to access nextXid without a lock for the startup process.
992 : */
993 24804 : rel = TransamVariables->nextXid;
994 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
995 :
996 49485 : if (!FullTransactionIdIsValid(cur_latest) ||
997 24681 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
998 : {
999 18959 : TransamVariables->latestCompletedXid =
1000 18959 : FullXidRelativeTo(rel, latestXid);
1001 : }
1002 :
1003 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1004 24804 : }
1005 :
1006 : /*
1007 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1008 : *
1009 : * Remember up to where the startup process initialized the CLOG and subtrans
1010 : * so we can ensure it's initialized gaplessly up to the point where necessary
1011 : * while in recovery.
1012 : */
1013 : void
1014 123 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1015 : {
1016 : Assert(standbyState == STANDBY_INITIALIZED);
1017 : Assert(TransactionIdIsNormal(initializedUptoXID));
1018 :
1019 : /*
1020 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1021 : * to, so we can extend it from that point onwards in
1022 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1023 : * ProcArrayApplyRecoveryInfo().
1024 : */
1025 123 : latestObservedXid = initializedUptoXID;
1026 123 : TransactionIdRetreat(latestObservedXid);
1027 123 : }
1028 :
1029 : /*
1030 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1031 : *
1032 : * Takes us through 3 states: Initialized, Pending and Ready.
1033 : * Normal case is to go all the way to Ready straight away, though there
1034 : * are atypical cases where we need to take it in steps.
1035 : *
1036 : * Use the data about running transactions on the primary to create the initial
1037 : * state of KnownAssignedXids. We also use these records to regularly prune
1038 : * KnownAssignedXids because we know it is possible that some transactions
1039 : * with FATAL errors fail to write abort records, which could cause eventual
1040 : * overflow.
1041 : *
1042 : * See comments for LogStandbySnapshot().
1043 : */
1044 : void
1045 852 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1046 : {
1047 : TransactionId *xids;
1048 : TransactionId advanceNextXid;
1049 : int nxids;
1050 : int i;
1051 :
1052 : Assert(standbyState >= STANDBY_INITIALIZED);
1053 : Assert(TransactionIdIsValid(running->nextXid));
1054 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1055 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1056 :
1057 : /*
1058 : * Remove stale transactions, if any.
1059 : */
1060 852 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1061 :
1062 : /*
1063 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1064 : * because we will need it up to date for accessing two-phase transactions
1065 : * in StandbyReleaseOldLocks().
1066 : */
1067 852 : advanceNextXid = running->nextXid;
1068 852 : TransactionIdRetreat(advanceNextXid);
1069 852 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1070 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1071 :
1072 : /*
1073 : * Remove stale locks, if any.
1074 : */
1075 852 : StandbyReleaseOldLocks(running->oldestRunningXid);
1076 :
1077 : /*
1078 : * If our snapshot is already valid, nothing else to do...
1079 : */
1080 852 : if (standbyState == STANDBY_SNAPSHOT_READY)
1081 729 : return;
1082 :
1083 : /*
1084 : * If our initial RunningTransactionsData had an overflowed snapshot then
1085 : * we knew we were missing some subxids from our snapshot. If we continue
1086 : * to see overflowed snapshots then we might never be able to start up, so
1087 : * we make another test to see if our snapshot is now valid. We know that
1088 : * the missing subxids are equal to or earlier than nextXid. After we
1089 : * initialise we continue to apply changes during recovery, so once the
1090 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1091 : * know that we no longer have missing information and can mark the
1092 : * snapshot as valid.
1093 : */
1094 123 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1095 : {
1096 : /*
1097 : * If the snapshot isn't overflowed or if its empty we can reset our
1098 : * pending state and use this snapshot instead.
1099 : */
1100 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1101 : {
1102 : /*
1103 : * If we have already collected known assigned xids, we need to
1104 : * throw them away before we apply the recovery snapshot.
1105 : */
1106 0 : KnownAssignedXidsReset();
1107 0 : standbyState = STANDBY_INITIALIZED;
1108 : }
1109 : else
1110 : {
1111 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1112 : running->oldestRunningXid))
1113 : {
1114 0 : standbyState = STANDBY_SNAPSHOT_READY;
1115 0 : elog(DEBUG1,
1116 : "recovery snapshots are now enabled");
1117 : }
1118 : else
1119 0 : elog(DEBUG1,
1120 : "recovery snapshot waiting for non-overflowed snapshot or "
1121 : "until oldest active xid on standby is at least %u (now %u)",
1122 : standbySnapshotPendingXmin,
1123 : running->oldestRunningXid);
1124 0 : return;
1125 : }
1126 : }
1127 :
1128 : Assert(standbyState == STANDBY_INITIALIZED);
1129 :
1130 : /*
1131 : * NB: this can be reached at least twice, so make sure new code can deal
1132 : * with that.
1133 : */
1134 :
1135 : /*
1136 : * Nobody else is running yet, but take locks anyhow
1137 : */
1138 123 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1139 :
1140 : /*
1141 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1142 : * sort them first.
1143 : *
1144 : * Some of the new xids are top-level xids and some are subtransactions.
1145 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1146 : * aren't overflowed then all xids will fit in snapshot and so we don't
1147 : * need subtrans. If we later overflow, an xid assignment record will add
1148 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1149 : * don't have enough information to correctly update subtrans anyway.
1150 : */
1151 :
1152 : /*
1153 : * Allocate a temporary array to avoid modifying the array passed as
1154 : * argument.
1155 : */
1156 123 : xids = palloc_array(TransactionId, running->xcnt + running->subxcnt);
1157 :
1158 : /*
1159 : * Add to the temp array any xids which have not already completed.
1160 : */
1161 123 : nxids = 0;
1162 129 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1163 : {
1164 6 : TransactionId xid = running->xids[i];
1165 :
1166 : /*
1167 : * The running-xacts snapshot can contain xids that were still visible
1168 : * in the procarray when the snapshot was taken, but were already
1169 : * WAL-logged as completed. They're not running anymore, so ignore
1170 : * them.
1171 : */
1172 6 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1173 1 : continue;
1174 :
1175 5 : xids[nxids++] = xid;
1176 : }
1177 :
1178 123 : if (nxids > 0)
1179 : {
1180 5 : if (procArray->numKnownAssignedXids != 0)
1181 : {
1182 0 : LWLockRelease(ProcArrayLock);
1183 0 : elog(ERROR, "KnownAssignedXids is not empty");
1184 : }
1185 :
1186 : /*
1187 : * Sort the array so that we can add them safely into
1188 : * KnownAssignedXids.
1189 : *
1190 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1191 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1192 : * come from RUNNING_XACTS, which means there are only normal XIDs
1193 : * from the same epoch, so this is safe.
1194 : */
1195 5 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1196 :
1197 : /*
1198 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1199 : * snapshot may include duplicated xids because of prepared
1200 : * transactions, so ignore them.
1201 : */
1202 10 : for (i = 0; i < nxids; i++)
1203 : {
1204 5 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1205 : {
1206 0 : elog(DEBUG1,
1207 : "found duplicated transaction %u for KnownAssignedXids insertion",
1208 : xids[i]);
1209 0 : continue;
1210 : }
1211 5 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1212 : }
1213 :
1214 5 : KnownAssignedXidsDisplay(DEBUG3);
1215 : }
1216 :
1217 123 : pfree(xids);
1218 :
1219 : /*
1220 : * latestObservedXid is at least set to the point where SUBTRANS was
1221 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1222 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1223 : * subtrans from thereon, up to nextXid - 1.
1224 : *
1225 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1226 : * because we've just added xids to the known assigned xids machinery that
1227 : * haven't gone through RecordKnownAssignedTransactionId().
1228 : */
1229 : Assert(TransactionIdIsNormal(latestObservedXid));
1230 123 : TransactionIdAdvance(latestObservedXid);
1231 246 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1232 : {
1233 0 : ExtendSUBTRANS(latestObservedXid);
1234 0 : TransactionIdAdvance(latestObservedXid);
1235 : }
1236 123 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1237 :
1238 : /* ----------
1239 : * Now we've got the running xids we need to set the global values that
1240 : * are used to track snapshots as they evolve further.
1241 : *
1242 : * - latestCompletedXid which will be the xmax for snapshots
1243 : * - lastOverflowedXid which shows whether snapshots overflow
1244 : * - nextXid
1245 : *
1246 : * If the snapshot overflowed, then we still initialise with what we know,
1247 : * but the recovery snapshot isn't fully valid yet because we know there
1248 : * are some subxids missing. We don't know the specific subxids that are
1249 : * missing, so conservatively assume the last one is latestObservedXid.
1250 : * ----------
1251 : */
1252 123 : if (running->subxid_status == SUBXIDS_MISSING)
1253 : {
1254 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1255 :
1256 0 : standbySnapshotPendingXmin = latestObservedXid;
1257 0 : procArray->lastOverflowedXid = latestObservedXid;
1258 : }
1259 : else
1260 : {
1261 123 : standbyState = STANDBY_SNAPSHOT_READY;
1262 :
1263 123 : standbySnapshotPendingXmin = InvalidTransactionId;
1264 :
1265 : /*
1266 : * If the 'xids' array didn't include all subtransactions, we have to
1267 : * mark any snapshots taken as overflowed.
1268 : */
1269 123 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1270 26 : procArray->lastOverflowedXid = latestObservedXid;
1271 : else
1272 : {
1273 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1274 97 : procArray->lastOverflowedXid = InvalidTransactionId;
1275 : }
1276 : }
1277 :
1278 : /*
1279 : * If a transaction wrote a commit record in the gap between taking and
1280 : * logging the snapshot then latestCompletedXid may already be higher than
1281 : * the value from the snapshot, so check before we use the incoming value.
1282 : * It also might not yet be set at all.
1283 : */
1284 123 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1285 :
1286 : /*
1287 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1288 : * nobody can see it yet.
1289 : */
1290 :
1291 123 : LWLockRelease(ProcArrayLock);
1292 :
1293 123 : KnownAssignedXidsDisplay(DEBUG3);
1294 123 : if (standbyState == STANDBY_SNAPSHOT_READY)
1295 123 : elog(DEBUG1, "recovery snapshots are now enabled");
1296 : else
1297 0 : elog(DEBUG1,
1298 : "recovery snapshot waiting for non-overflowed snapshot or "
1299 : "until oldest active xid on standby is at least %u (now %u)",
1300 : standbySnapshotPendingXmin,
1301 : running->oldestRunningXid);
1302 : }
1303 :
1304 : /*
1305 : * ProcArrayApplyXidAssignment
1306 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1307 : */
1308 : void
1309 21 : ProcArrayApplyXidAssignment(TransactionId topxid,
1310 : int nsubxids, TransactionId *subxids)
1311 : {
1312 : TransactionId max_xid;
1313 : int i;
1314 :
1315 : Assert(standbyState >= STANDBY_INITIALIZED);
1316 :
1317 21 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1318 :
1319 : /*
1320 : * Mark all the subtransactions as observed.
1321 : *
1322 : * NOTE: This will fail if the subxid contains too many previously
1323 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1324 : * as the code stands, because xid-assignment records should never contain
1325 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1326 : */
1327 21 : RecordKnownAssignedTransactionIds(max_xid);
1328 :
1329 : /*
1330 : * Notice that we update pg_subtrans with the top-level xid, rather than
1331 : * the parent xid. This is a difference between normal processing and
1332 : * recovery, yet is still correct in all cases. The reason is that
1333 : * subtransaction commit is not marked in clog until commit processing, so
1334 : * all aborted subtransactions have already been clearly marked in clog.
1335 : * As a result we are able to refer directly to the top-level
1336 : * transaction's state rather than skipping through all the intermediate
1337 : * states in the subtransaction tree. This should be the first time we
1338 : * have attempted to SubTransSetParent().
1339 : */
1340 1365 : for (i = 0; i < nsubxids; i++)
1341 1344 : SubTransSetParent(subxids[i], topxid);
1342 :
1343 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1344 21 : if (standbyState == STANDBY_INITIALIZED)
1345 0 : return;
1346 :
1347 : /*
1348 : * Uses same locking as transaction commit
1349 : */
1350 21 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1351 :
1352 : /*
1353 : * Remove subxids from known-assigned-xacts.
1354 : */
1355 21 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1356 :
1357 : /*
1358 : * Advance lastOverflowedXid to be at least the last of these subxids.
1359 : */
1360 21 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1361 21 : procArray->lastOverflowedXid = max_xid;
1362 :
1363 21 : LWLockRelease(ProcArrayLock);
1364 : }
1365 :
1366 : /*
1367 : * TransactionIdIsInProgress -- is given transaction running in some backend
1368 : *
1369 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1370 : * there are four possibilities for finding a running transaction:
1371 : *
1372 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1373 : * by looking at ProcGlobal->xids.
1374 : *
1375 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1376 : * We can find this out cheaply too.
1377 : *
1378 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1379 : * if the Xid is running on the primary.
1380 : *
1381 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1382 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1383 : * This is the slowest way, but sadly it has to be done always if the others
1384 : * failed, unless we see that the cached subxact sets are complete (none have
1385 : * overflowed).
1386 : *
1387 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1388 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1389 : * This buys back some concurrency (and we can't retrieve the main Xids from
1390 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1391 : */
1392 : bool
1393 21335216 : TransactionIdIsInProgress(TransactionId xid)
1394 : {
1395 : static TransactionId *xids = NULL;
1396 : static TransactionId *other_xids;
1397 : XidCacheStatus *other_subxidstates;
1398 21335216 : int nxids = 0;
1399 21335216 : ProcArrayStruct *arrayP = procArray;
1400 : TransactionId topxid;
1401 : TransactionId latestCompletedXid;
1402 : int mypgxactoff;
1403 : int numProcs;
1404 : int j;
1405 :
1406 : /*
1407 : * Don't bother checking a transaction older than RecentXmin; it could not
1408 : * possibly still be running. (Note: in particular, this guarantees that
1409 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1410 : * running.)
1411 : */
1412 21335216 : if (TransactionIdPrecedes(xid, RecentXmin))
1413 : {
1414 : xc_by_recent_xmin_inc();
1415 13249882 : return false;
1416 : }
1417 :
1418 : /*
1419 : * We may have just checked the status of this transaction, so if it is
1420 : * already known to be completed, we can fall out without any access to
1421 : * shared memory.
1422 : */
1423 8085334 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1424 : {
1425 : xc_by_known_xact_inc();
1426 1306246 : return false;
1427 : }
1428 :
1429 : /*
1430 : * Also, we can handle our own transaction (and subtransactions) without
1431 : * any access to shared memory.
1432 : */
1433 6779088 : if (TransactionIdIsCurrentTransactionId(xid))
1434 : {
1435 : xc_by_my_xact_inc();
1436 220712 : return true;
1437 : }
1438 :
1439 : /*
1440 : * If first time through, get workspace to remember main XIDs in. We
1441 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1442 : */
1443 6558376 : if (xids == NULL)
1444 : {
1445 : /*
1446 : * In hot standby mode, reserve enough space to hold all xids in the
1447 : * known-assigned list. If we later finish recovery, we no longer need
1448 : * the bigger array, but we don't bother to shrink it.
1449 : */
1450 1726 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1451 :
1452 1726 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1453 1726 : if (xids == NULL)
1454 0 : ereport(ERROR,
1455 : (errcode(ERRCODE_OUT_OF_MEMORY),
1456 : errmsg("out of memory")));
1457 : }
1458 :
1459 6558376 : other_xids = ProcGlobal->xids;
1460 6558376 : other_subxidstates = ProcGlobal->subxidStates;
1461 :
1462 6558376 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1463 :
1464 : /*
1465 : * Now that we have the lock, we can check latestCompletedXid; if the
1466 : * target Xid is after that, it's surely still running.
1467 : */
1468 6558376 : latestCompletedXid =
1469 6558376 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1470 6558376 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1471 : {
1472 6481191 : LWLockRelease(ProcArrayLock);
1473 : xc_by_latest_xid_inc();
1474 6481191 : return true;
1475 : }
1476 :
1477 : /* No shortcuts, gotta grovel through the array */
1478 77185 : mypgxactoff = MyProc->pgxactoff;
1479 77185 : numProcs = arrayP->numProcs;
1480 270765 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1481 : {
1482 : int pgprocno;
1483 : PGPROC *proc;
1484 : TransactionId pxid;
1485 : int pxids;
1486 :
1487 : /* Ignore ourselves --- dealt with it above */
1488 254251 : if (pgxactoff == mypgxactoff)
1489 18534 : continue;
1490 :
1491 : /* Fetch xid just once - see GetNewTransactionId */
1492 235717 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1493 :
1494 235717 : if (!TransactionIdIsValid(pxid))
1495 127348 : continue;
1496 :
1497 : /*
1498 : * Step 1: check the main Xid
1499 : */
1500 108369 : if (TransactionIdEquals(pxid, xid))
1501 : {
1502 60532 : LWLockRelease(ProcArrayLock);
1503 : xc_by_main_xid_inc();
1504 60532 : return true;
1505 : }
1506 :
1507 : /*
1508 : * We can ignore main Xids that are younger than the target Xid, since
1509 : * the target could not possibly be their child.
1510 : */
1511 47837 : if (TransactionIdPrecedes(xid, pxid))
1512 22661 : continue;
1513 :
1514 : /*
1515 : * Step 2: check the cached child-Xids arrays
1516 : */
1517 25176 : pxids = other_subxidstates[pgxactoff].count;
1518 25176 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1519 25176 : pgprocno = arrayP->pgprocnos[pgxactoff];
1520 25176 : proc = &allProcs[pgprocno];
1521 44148 : for (j = pxids - 1; j >= 0; j--)
1522 : {
1523 : /* Fetch xid just once - see GetNewTransactionId */
1524 19111 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1525 :
1526 19111 : if (TransactionIdEquals(cxid, xid))
1527 : {
1528 139 : LWLockRelease(ProcArrayLock);
1529 : xc_by_child_xid_inc();
1530 139 : return true;
1531 : }
1532 : }
1533 :
1534 : /*
1535 : * Save the main Xid for step 4. We only need to remember main Xids
1536 : * that have uncached children. (Note: there is no race condition
1537 : * here because the overflowed flag cannot be cleared, only set, while
1538 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1539 : * worry about.)
1540 : */
1541 25037 : if (other_subxidstates[pgxactoff].overflowed)
1542 227 : xids[nxids++] = pxid;
1543 : }
1544 :
1545 : /*
1546 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1547 : * in the list must be treated as running.
1548 : */
1549 16514 : if (RecoveryInProgress())
1550 : {
1551 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1552 : Assert(nxids == 0);
1553 :
1554 1 : if (KnownAssignedXidExists(xid))
1555 : {
1556 0 : LWLockRelease(ProcArrayLock);
1557 : xc_by_known_assigned_inc();
1558 0 : return true;
1559 : }
1560 :
1561 : /*
1562 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1563 : * too. Fetch all xids from KnownAssignedXids that are lower than
1564 : * xid, since if xid is a subtransaction its parent will always have a
1565 : * lower value. Note we will collect both main and subXIDs here, but
1566 : * there's no help for it.
1567 : */
1568 1 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1569 0 : nxids = KnownAssignedXidsGet(xids, xid);
1570 : }
1571 :
1572 16514 : LWLockRelease(ProcArrayLock);
1573 :
1574 : /*
1575 : * If none of the relevant caches overflowed, we know the Xid is not
1576 : * running without even looking at pg_subtrans.
1577 : */
1578 16514 : if (nxids == 0)
1579 : {
1580 : xc_no_overflow_inc();
1581 16287 : cachedXidIsNotInProgress = xid;
1582 16287 : return false;
1583 : }
1584 :
1585 : /*
1586 : * Step 4: have to check pg_subtrans.
1587 : *
1588 : * At this point, we know it's either a subtransaction of one of the Xids
1589 : * in xids[], or it's not running. If it's an already-failed
1590 : * subtransaction, we want to say "not running" even though its parent may
1591 : * still be running. So first, check pg_xact to see if it's been aborted.
1592 : */
1593 : xc_slow_answer_inc();
1594 :
1595 227 : if (TransactionIdDidAbort(xid))
1596 : {
1597 0 : cachedXidIsNotInProgress = xid;
1598 0 : return false;
1599 : }
1600 :
1601 : /*
1602 : * It isn't aborted, so check whether the transaction tree it belongs to
1603 : * is still running (or, more precisely, whether it was running when we
1604 : * held ProcArrayLock).
1605 : */
1606 227 : topxid = SubTransGetTopmostTransaction(xid);
1607 : Assert(TransactionIdIsValid(topxid));
1608 384 : if (!TransactionIdEquals(topxid, xid) &&
1609 157 : pg_lfind32(topxid, xids, nxids))
1610 157 : return true;
1611 :
1612 70 : cachedXidIsNotInProgress = xid;
1613 70 : return false;
1614 : }
1615 :
1616 :
1617 : /*
1618 : * Determine XID horizons.
1619 : *
1620 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1621 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1622 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1623 : * GlobalVisState).
1624 : *
1625 : * See the definition of ComputeXidHorizonsResult for the various computed
1626 : * horizons.
1627 : *
1628 : * For VACUUM separate horizons (used to decide which deleted tuples must
1629 : * be preserved), for shared and non-shared tables are computed. For shared
1630 : * relations backends in all databases must be considered, but for non-shared
1631 : * relations that's not required, since only backends in my own database could
1632 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1633 : * VACUUMs because (a) they must be working on other tables, and (b) they
1634 : * don't need to do snapshot-based lookups.
1635 : *
1636 : * This also computes a horizon used to truncate pg_subtrans. For that
1637 : * backends in all databases have to be considered, and concurrently running
1638 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1639 : * accesses.
1640 : *
1641 : * Note: we include all currently running xids in the set of considered xids.
1642 : * This ensures that if a just-started xact has not yet set its snapshot,
1643 : * when it does set the snapshot it cannot set xmin less than what we compute.
1644 : * See notes in src/backend/access/transam/README.
1645 : *
1646 : * Note: despite the above, it's possible for the calculated values to move
1647 : * backwards on repeated calls. The calculated values are conservative, so
1648 : * that anything older is definitely not considered as running by anyone
1649 : * anymore, but the exact values calculated depend on a number of things. For
1650 : * example, if there are no transactions running in the current database, the
1651 : * horizon for normal tables will be latestCompletedXid. If a transaction
1652 : * begins after that, its xmin will include in-progress transactions in other
1653 : * databases that started earlier, so another call will return a lower value.
1654 : * Nonetheless it is safe to vacuum a table in the current database with the
1655 : * first result. There are also replication-related effects: a walsender
1656 : * process can set its xmin based on transactions that are no longer running
1657 : * on the primary but are still being replayed on the standby, thus possibly
1658 : * making the values go backwards. In this case there is a possibility that
1659 : * we lose data that the standby would like to have, but unless the standby
1660 : * uses a replication slot to make its xmin persistent there is little we can
1661 : * do about that --- data is only protected if the walsender runs continuously
1662 : * while queries are executed on the standby. (The Hot Standby code deals
1663 : * with such cases by failing standby queries that needed to access
1664 : * already-removed data, so there's no integrity bug.)
1665 : *
1666 : * Note: the approximate horizons (see definition of GlobalVisState) are
1667 : * updated by the computations done here. That's currently required for
1668 : * correctness and a small optimization. Without doing so it's possible that
1669 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1670 : * horizon than later when deciding which tuples can be removed - which the
1671 : * code doesn't expect (breaking HOT).
1672 : */
1673 : static void
1674 242499 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1675 : {
1676 242499 : ProcArrayStruct *arrayP = procArray;
1677 : TransactionId kaxmin;
1678 242499 : bool in_recovery = RecoveryInProgress();
1679 242499 : TransactionId *other_xids = ProcGlobal->xids;
1680 :
1681 : /* inferred after ProcArrayLock is released */
1682 242499 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1683 :
1684 242499 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1685 :
1686 242499 : h->latest_completed = TransamVariables->latestCompletedXid;
1687 :
1688 : /*
1689 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1690 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1691 : * and so protects us against overestimating the result due to future
1692 : * additions.
1693 : */
1694 : {
1695 : TransactionId initial;
1696 :
1697 242499 : initial = XidFromFullTransactionId(h->latest_completed);
1698 : Assert(TransactionIdIsValid(initial));
1699 242499 : TransactionIdAdvance(initial);
1700 :
1701 242499 : h->oldest_considered_running = initial;
1702 242499 : h->shared_oldest_nonremovable = initial;
1703 242499 : h->data_oldest_nonremovable = initial;
1704 :
1705 : /*
1706 : * Only modifications made by this backend affect the horizon for
1707 : * temporary relations. Instead of a check in each iteration of the
1708 : * loop over all PGPROCs it is cheaper to just initialize to the
1709 : * current top-level xid any.
1710 : *
1711 : * Without an assigned xid we could use a horizon as aggressive as
1712 : * GetNewTransactionId(), but we can get away with the much cheaper
1713 : * latestCompletedXid + 1: If this backend has no xid there, by
1714 : * definition, can't be any newer changes in the temp table than
1715 : * latestCompletedXid.
1716 : */
1717 242499 : if (TransactionIdIsValid(MyProc->xid))
1718 49767 : h->temp_oldest_nonremovable = MyProc->xid;
1719 : else
1720 192732 : h->temp_oldest_nonremovable = initial;
1721 : }
1722 :
1723 : /*
1724 : * Fetch slot horizons while ProcArrayLock is held - the
1725 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1726 : * the lock.
1727 : */
1728 242499 : h->slot_xmin = procArray->replication_slot_xmin;
1729 242499 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1730 :
1731 1548023 : for (int index = 0; index < arrayP->numProcs; index++)
1732 : {
1733 1305524 : int pgprocno = arrayP->pgprocnos[index];
1734 1305524 : PGPROC *proc = &allProcs[pgprocno];
1735 1305524 : int8 statusFlags = ProcGlobal->statusFlags[index];
1736 : TransactionId xid;
1737 : TransactionId xmin;
1738 :
1739 : /* Fetch xid just once - see GetNewTransactionId */
1740 1305524 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1741 1305524 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1742 :
1743 : /*
1744 : * Consider both the transaction's Xmin, and its Xid.
1745 : *
1746 : * We must check both because a transaction might have an Xmin but not
1747 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1748 : * some not-yet-set Xmin.
1749 : */
1750 1305524 : xmin = TransactionIdOlder(xmin, xid);
1751 :
1752 : /* if neither is set, this proc doesn't influence the horizon */
1753 1305524 : if (!TransactionIdIsValid(xmin))
1754 614659 : continue;
1755 :
1756 : /*
1757 : * Don't ignore any procs when determining which transactions might be
1758 : * considered running. While slots should ensure logical decoding
1759 : * backends are protected even without this check, it can't hurt to
1760 : * include them here as well..
1761 : */
1762 690865 : h->oldest_considered_running =
1763 690865 : TransactionIdOlder(h->oldest_considered_running, xmin);
1764 :
1765 : /*
1766 : * Skip over backends either vacuuming (which is ok with rows being
1767 : * removed, as long as pg_subtrans is not truncated) or doing logical
1768 : * decoding (which manages xmin separately, check below).
1769 : */
1770 690865 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1771 234986 : continue;
1772 :
1773 : /* shared tables need to take backends in all databases into account */
1774 455879 : h->shared_oldest_nonremovable =
1775 455879 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1776 :
1777 : /*
1778 : * Normally sessions in other databases are ignored for anything but
1779 : * the shared horizon.
1780 : *
1781 : * However, include them when MyDatabaseId is not (yet) set. A
1782 : * backend in the process of starting up must not compute a "too
1783 : * aggressive" horizon, otherwise we could end up using it to prune
1784 : * still-needed data away. If the current backend never connects to a
1785 : * database this is harmless, because data_oldest_nonremovable will
1786 : * never be utilized.
1787 : *
1788 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1789 : * be included. (This flag is used for hot standby feedback, which
1790 : * can't be tied to a specific database.)
1791 : *
1792 : * Also, while in recovery we cannot compute an accurate per-database
1793 : * horizon, as all xids are managed via the KnownAssignedXids
1794 : * machinery.
1795 : */
1796 455879 : if (proc->databaseId == MyDatabaseId ||
1797 22828 : MyDatabaseId == InvalidOid ||
1798 12825 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1799 : in_recovery)
1800 : {
1801 443059 : h->data_oldest_nonremovable =
1802 443059 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1803 : }
1804 : }
1805 :
1806 : /*
1807 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1808 : * after lock is released.
1809 : */
1810 242499 : if (in_recovery)
1811 396 : kaxmin = KnownAssignedXidsGetOldestXmin();
1812 :
1813 : /*
1814 : * No other information from shared state is needed, release the lock
1815 : * immediately. The rest of the computations can be done without a lock.
1816 : */
1817 242499 : LWLockRelease(ProcArrayLock);
1818 :
1819 242499 : if (in_recovery)
1820 : {
1821 396 : h->oldest_considered_running =
1822 396 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1823 396 : h->shared_oldest_nonremovable =
1824 396 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1825 396 : h->data_oldest_nonremovable =
1826 396 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1827 : /* temp relations cannot be accessed in recovery */
1828 : }
1829 :
1830 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1831 : h->shared_oldest_nonremovable));
1832 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1833 : h->data_oldest_nonremovable));
1834 :
1835 : /*
1836 : * Check whether there are replication slots requiring an older xmin.
1837 : */
1838 242499 : h->shared_oldest_nonremovable =
1839 242499 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1840 242499 : h->data_oldest_nonremovable =
1841 242499 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1842 :
1843 : /*
1844 : * The only difference between catalog / data horizons is that the slot's
1845 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1846 : * for logical decoding). Initialize with data horizon, and then back up
1847 : * further if necessary. Have to back up the shared horizon as well, since
1848 : * that also can contain catalogs.
1849 : */
1850 242499 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1851 242499 : h->shared_oldest_nonremovable =
1852 242499 : TransactionIdOlder(h->shared_oldest_nonremovable,
1853 : h->slot_catalog_xmin);
1854 242499 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1855 242499 : h->catalog_oldest_nonremovable =
1856 242499 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1857 : h->slot_catalog_xmin);
1858 :
1859 : /*
1860 : * It's possible that slots backed up the horizons further than
1861 : * oldest_considered_running. Fix.
1862 : */
1863 242499 : h->oldest_considered_running =
1864 242499 : TransactionIdOlder(h->oldest_considered_running,
1865 : h->shared_oldest_nonremovable);
1866 242499 : h->oldest_considered_running =
1867 242499 : TransactionIdOlder(h->oldest_considered_running,
1868 : h->catalog_oldest_nonremovable);
1869 242499 : h->oldest_considered_running =
1870 242499 : TransactionIdOlder(h->oldest_considered_running,
1871 : h->data_oldest_nonremovable);
1872 :
1873 : /*
1874 : * shared horizons have to be at least as old as the oldest visible in
1875 : * current db
1876 : */
1877 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1878 : h->data_oldest_nonremovable));
1879 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1880 : h->catalog_oldest_nonremovable));
1881 :
1882 : /*
1883 : * Horizons need to ensure that pg_subtrans access is still possible for
1884 : * the relevant backends.
1885 : */
1886 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1887 : h->shared_oldest_nonremovable));
1888 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1889 : h->catalog_oldest_nonremovable));
1890 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1891 : h->data_oldest_nonremovable));
1892 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1893 : h->temp_oldest_nonremovable));
1894 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1895 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1896 : h->slot_xmin));
1897 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1898 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1899 : h->slot_catalog_xmin));
1900 :
1901 : /* update approximate horizons with the computed horizons */
1902 242499 : GlobalVisUpdateApply(h);
1903 242499 : }
1904 :
1905 : /*
1906 : * Determine what kind of visibility horizon needs to be used for a
1907 : * relation. If rel is NULL, the most conservative horizon is used.
1908 : */
1909 : static inline GlobalVisHorizonKind
1910 20582504 : GlobalVisHorizonKindForRel(Relation rel)
1911 : {
1912 : /*
1913 : * Other relkinds currently don't contain xids, nor always the necessary
1914 : * logical decoding markers.
1915 : */
1916 : Assert(!rel ||
1917 : rel->rd_rel->relkind == RELKIND_RELATION ||
1918 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1919 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1920 :
1921 20582504 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1922 166440 : return VISHORIZON_SHARED;
1923 20416064 : else if (IsCatalogRelation(rel) ||
1924 15539405 : RelationIsAccessibleInLogicalDecoding(rel))
1925 4876664 : return VISHORIZON_CATALOG;
1926 15539400 : else if (!RELATION_IS_LOCAL(rel))
1927 15415680 : return VISHORIZON_DATA;
1928 : else
1929 123720 : return VISHORIZON_TEMP;
1930 : }
1931 :
1932 : /*
1933 : * Return the oldest XID for which deleted tuples must be preserved in the
1934 : * passed table.
1935 : *
1936 : * If rel is not NULL the horizon may be considerably more recent than
1937 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1938 : * that is correct (but not optimal) for all relations will be returned.
1939 : *
1940 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1941 : * the passed in table.
1942 : */
1943 : TransactionId
1944 154693 : GetOldestNonRemovableTransactionId(Relation rel)
1945 : {
1946 : ComputeXidHorizonsResult horizons;
1947 :
1948 154693 : ComputeXidHorizons(&horizons);
1949 :
1950 154693 : switch (GlobalVisHorizonKindForRel(rel))
1951 : {
1952 22263 : case VISHORIZON_SHARED:
1953 22263 : return horizons.shared_oldest_nonremovable;
1954 95026 : case VISHORIZON_CATALOG:
1955 95026 : return horizons.catalog_oldest_nonremovable;
1956 21427 : case VISHORIZON_DATA:
1957 21427 : return horizons.data_oldest_nonremovable;
1958 15977 : case VISHORIZON_TEMP:
1959 15977 : return horizons.temp_oldest_nonremovable;
1960 : }
1961 :
1962 : /* just to prevent compiler warnings */
1963 0 : return InvalidTransactionId;
1964 : }
1965 :
1966 : /*
1967 : * Return the oldest transaction id any currently running backend might still
1968 : * consider running. This should not be used for visibility / pruning
1969 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1970 : * decisions like up to where pg_subtrans can be truncated.
1971 : */
1972 : TransactionId
1973 1915 : GetOldestTransactionIdConsideredRunning(void)
1974 : {
1975 : ComputeXidHorizonsResult horizons;
1976 :
1977 1915 : ComputeXidHorizons(&horizons);
1978 :
1979 1915 : return horizons.oldest_considered_running;
1980 : }
1981 :
1982 : /*
1983 : * Return the visibility horizons for a hot standby feedback message.
1984 : */
1985 : void
1986 56 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1987 : {
1988 : ComputeXidHorizonsResult horizons;
1989 :
1990 56 : ComputeXidHorizons(&horizons);
1991 :
1992 : /*
1993 : * Don't want to use shared_oldest_nonremovable here, as that contains the
1994 : * effect of replication slot's catalog_xmin. We want to send a separate
1995 : * feedback for the catalog horizon, so the primary can remove data table
1996 : * contents more aggressively.
1997 : */
1998 56 : *xmin = horizons.shared_oldest_nonremovable_raw;
1999 56 : *catalog_xmin = horizons.slot_catalog_xmin;
2000 56 : }
2001 :
2002 : /*
2003 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2004 : *
2005 : * We have to export this for use by snapmgr.c.
2006 : */
2007 : int
2008 38163 : GetMaxSnapshotXidCount(void)
2009 : {
2010 38163 : return procArray->maxProcs;
2011 : }
2012 :
2013 : /*
2014 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2015 : *
2016 : * We have to export this for use by snapmgr.c.
2017 : */
2018 : int
2019 37949 : GetMaxSnapshotSubxidCount(void)
2020 : {
2021 37949 : return TOTAL_MAX_CACHED_SUBXIDS;
2022 : }
2023 :
2024 : /*
2025 : * Helper function for GetSnapshotData() that checks if the bulk of the
2026 : * visibility information in the snapshot is still valid. If so, it updates
2027 : * the fields that need to change and returns true. Otherwise it returns
2028 : * false.
2029 : *
2030 : * This very likely can be evolved to not need ProcArrayLock held (at very
2031 : * least in the case we already hold a snapshot), but that's for another day.
2032 : */
2033 : static bool
2034 2697720 : GetSnapshotDataReuse(Snapshot snapshot)
2035 : {
2036 : uint64 curXactCompletionCount;
2037 :
2038 : Assert(LWLockHeldByMe(ProcArrayLock));
2039 :
2040 2697720 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2041 37932 : return false;
2042 :
2043 2659788 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2044 2659788 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2045 465603 : return false;
2046 :
2047 : /*
2048 : * If the current xactCompletionCount is still the same as it was at the
2049 : * time the snapshot was built, we can be sure that rebuilding the
2050 : * contents of the snapshot the hard way would result in the same snapshot
2051 : * contents:
2052 : *
2053 : * As explained in transam/README, the set of xids considered running by
2054 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2055 : * contents only depend on transactions with xids and xactCompletionCount
2056 : * is incremented whenever a transaction with an xid finishes (while
2057 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2058 : * ensures we would detect if the snapshot would have changed.
2059 : *
2060 : * As the snapshot contents are the same as it was before, it is safe to
2061 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2062 : * visible under the snapshot could already have been removed (that'd
2063 : * require the set of running transactions to change) and it fulfills the
2064 : * requirement that concurrent GetSnapshotData() calls yield the same
2065 : * xmin.
2066 : */
2067 2194185 : if (!TransactionIdIsValid(MyProc->xmin))
2068 816983 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2069 :
2070 2194185 : RecentXmin = snapshot->xmin;
2071 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2072 :
2073 2194185 : snapshot->curcid = GetCurrentCommandId(false);
2074 2194185 : snapshot->active_count = 0;
2075 2194185 : snapshot->regd_count = 0;
2076 2194185 : snapshot->copied = false;
2077 :
2078 2194185 : return true;
2079 : }
2080 :
2081 : /*
2082 : * GetSnapshotData -- returns information about running transactions.
2083 : *
2084 : * The returned snapshot includes xmin (lowest still-running xact ID),
2085 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2086 : * in the range xmin <= xid < xmax. It is used as follows:
2087 : * All xact IDs < xmin are considered finished.
2088 : * All xact IDs >= xmax are considered still running.
2089 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2090 : * it is considered running or not.
2091 : * This ensures that the set of transactions seen as "running" by the
2092 : * current xact will not change after it takes the snapshot.
2093 : *
2094 : * All running top-level XIDs are included in the snapshot, except for lazy
2095 : * VACUUM processes. We also try to include running subtransaction XIDs,
2096 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2097 : * information may not be available. If we find any overflowed subxid arrays,
2098 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2099 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2100 : *
2101 : * We also update the following backend-global variables:
2102 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2103 : * current transaction (this is the same as MyProc->xmin).
2104 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2105 : * older than this are known not running any more.
2106 : *
2107 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2108 : * for the benefit of the GlobalVisTest* family of functions.
2109 : *
2110 : * Note: this function should probably not be called with an argument that's
2111 : * not statically allocated (see xip allocation below).
2112 : */
2113 : Snapshot
2114 2697720 : GetSnapshotData(Snapshot snapshot)
2115 : {
2116 2697720 : ProcArrayStruct *arrayP = procArray;
2117 2697720 : TransactionId *other_xids = ProcGlobal->xids;
2118 : TransactionId xmin;
2119 : TransactionId xmax;
2120 2697720 : int count = 0;
2121 2697720 : int subcount = 0;
2122 2697720 : bool suboverflowed = false;
2123 : FullTransactionId latest_completed;
2124 : TransactionId oldestxid;
2125 : int mypgxactoff;
2126 : TransactionId myxid;
2127 : uint64 curXactCompletionCount;
2128 :
2129 2697720 : TransactionId replication_slot_xmin = InvalidTransactionId;
2130 2697720 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2131 :
2132 : Assert(snapshot != NULL);
2133 :
2134 : /*
2135 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2136 : * be sufficient. But it seems better to do the malloc while not holding
2137 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2138 : * more subxip storage than is probably needed.
2139 : *
2140 : * This does open a possibility for avoiding repeated malloc/free: since
2141 : * maxProcs does not change at runtime, we can simply reuse the previous
2142 : * xip arrays if any. (This relies on the fact that all callers pass
2143 : * static SnapshotData structs.)
2144 : */
2145 2697720 : if (snapshot->xip == NULL)
2146 : {
2147 : /*
2148 : * First call for this snapshot. Snapshot is same size whether or not
2149 : * we are in recovery, see later comments.
2150 : */
2151 37924 : snapshot->xip = (TransactionId *)
2152 37924 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2153 37924 : if (snapshot->xip == NULL)
2154 0 : ereport(ERROR,
2155 : (errcode(ERRCODE_OUT_OF_MEMORY),
2156 : errmsg("out of memory")));
2157 : Assert(snapshot->subxip == NULL);
2158 37924 : snapshot->subxip = (TransactionId *)
2159 37924 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2160 37924 : if (snapshot->subxip == NULL)
2161 : {
2162 : /*
2163 : * Clean up the Snapshot state before throwing the error, so that
2164 : * a retry does not see a partially-initialized snapshot.
2165 : */
2166 0 : free(snapshot->xip);
2167 0 : snapshot->xip = NULL;
2168 0 : ereport(ERROR,
2169 : (errcode(ERRCODE_OUT_OF_MEMORY),
2170 : errmsg("out of memory")));
2171 : }
2172 : }
2173 :
2174 : /*
2175 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2176 : * going to set MyProc->xmin.
2177 : */
2178 2697720 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2179 :
2180 2697720 : if (GetSnapshotDataReuse(snapshot))
2181 : {
2182 2194185 : LWLockRelease(ProcArrayLock);
2183 2194185 : return snapshot;
2184 : }
2185 :
2186 503535 : latest_completed = TransamVariables->latestCompletedXid;
2187 503535 : mypgxactoff = MyProc->pgxactoff;
2188 503535 : myxid = other_xids[mypgxactoff];
2189 : Assert(myxid == MyProc->xid);
2190 :
2191 503535 : oldestxid = TransamVariables->oldestXid;
2192 503535 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2193 :
2194 : /* xmax is always latestCompletedXid + 1 */
2195 503535 : xmax = XidFromFullTransactionId(latest_completed);
2196 503535 : TransactionIdAdvance(xmax);
2197 : Assert(TransactionIdIsNormal(xmax));
2198 :
2199 : /* initialize xmin calculation with xmax */
2200 503535 : xmin = xmax;
2201 :
2202 : /* take own xid into account, saves a check inside the loop */
2203 503535 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2204 34519 : xmin = myxid;
2205 :
2206 503535 : snapshot->takenDuringRecovery = RecoveryInProgress();
2207 :
2208 503535 : if (!snapshot->takenDuringRecovery)
2209 : {
2210 501934 : int numProcs = arrayP->numProcs;
2211 501934 : TransactionId *xip = snapshot->xip;
2212 501934 : int *pgprocnos = arrayP->pgprocnos;
2213 501934 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2214 501934 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2215 :
2216 : /*
2217 : * First collect set of pgxactoff/xids that need to be included in the
2218 : * snapshot.
2219 : */
2220 4447088 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2221 : {
2222 : /* Fetch xid just once - see GetNewTransactionId */
2223 3945154 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2224 : uint8 statusFlags;
2225 :
2226 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2227 :
2228 : /*
2229 : * If the transaction has no XID assigned, we can skip it; it
2230 : * won't have sub-XIDs either.
2231 : */
2232 3945154 : if (likely(xid == InvalidTransactionId))
2233 3181473 : continue;
2234 :
2235 : /*
2236 : * We don't include our own XIDs (if any) in the snapshot. It
2237 : * needs to be included in the xmin computation, but we did so
2238 : * outside the loop.
2239 : */
2240 763681 : if (pgxactoff == mypgxactoff)
2241 68320 : continue;
2242 :
2243 : /*
2244 : * The only way we are able to get here with a non-normal xid is
2245 : * during bootstrap - with this backend using
2246 : * BootstrapTransactionId. But the above test should filter that
2247 : * out.
2248 : */
2249 : Assert(TransactionIdIsNormal(xid));
2250 :
2251 : /*
2252 : * If the XID is >= xmax, we can skip it; such transactions will
2253 : * be treated as running anyway (and any sub-XIDs will also be >=
2254 : * xmax).
2255 : */
2256 695361 : if (!NormalTransactionIdPrecedes(xid, xmax))
2257 205607 : continue;
2258 :
2259 : /*
2260 : * Skip over backends doing logical decoding which manages xmin
2261 : * separately (check below) and ones running LAZY VACUUM.
2262 : */
2263 489754 : statusFlags = allStatusFlags[pgxactoff];
2264 489754 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2265 76 : continue;
2266 :
2267 489678 : if (NormalTransactionIdPrecedes(xid, xmin))
2268 301960 : xmin = xid;
2269 :
2270 : /* Add XID to snapshot. */
2271 489678 : xip[count++] = xid;
2272 :
2273 : /*
2274 : * Save subtransaction XIDs if possible (if we've already
2275 : * overflowed, there's no point). Note that the subxact XIDs must
2276 : * be later than their parent, so no need to check them against
2277 : * xmin. We could filter against xmax, but it seems better not to
2278 : * do that much work while holding the ProcArrayLock.
2279 : *
2280 : * The other backend can add more subxids concurrently, but cannot
2281 : * remove any. Hence it's important to fetch nxids just once.
2282 : * Should be safe to use memcpy, though. (We needn't worry about
2283 : * missing any xids added concurrently, because they must postdate
2284 : * xmax.)
2285 : *
2286 : * Again, our own XIDs are not included in the snapshot.
2287 : */
2288 489678 : if (!suboverflowed)
2289 : {
2290 :
2291 489674 : if (subxidStates[pgxactoff].overflowed)
2292 605 : suboverflowed = true;
2293 : else
2294 : {
2295 489069 : int nsubxids = subxidStates[pgxactoff].count;
2296 :
2297 489069 : if (nsubxids > 0)
2298 : {
2299 8510 : int pgprocno = pgprocnos[pgxactoff];
2300 8510 : PGPROC *proc = &allProcs[pgprocno];
2301 :
2302 8510 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2303 :
2304 8510 : memcpy(snapshot->subxip + subcount,
2305 8510 : proc->subxids.xids,
2306 : nsubxids * sizeof(TransactionId));
2307 8510 : subcount += nsubxids;
2308 : }
2309 : }
2310 : }
2311 : }
2312 : }
2313 : else
2314 : {
2315 : /*
2316 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2317 : *
2318 : * We store all xids directly into subxip[]. Here's why:
2319 : *
2320 : * In recovery we don't know which xids are top-level and which are
2321 : * subxacts, a design choice that greatly simplifies xid processing.
2322 : *
2323 : * It seems like we would want to try to put xids into xip[] only, but
2324 : * that is fairly small. We would either need to make that bigger or
2325 : * to increase the rate at which we WAL-log xid assignment; neither is
2326 : * an appealing choice.
2327 : *
2328 : * We could try to store xids into xip[] first and then into subxip[]
2329 : * if there are too many xids. That only works if the snapshot doesn't
2330 : * overflow because we do not search subxip[] in that case. A simpler
2331 : * way is to just store all xids in the subxip array because this is
2332 : * by far the bigger array. We just leave the xip array empty.
2333 : *
2334 : * Either way we need to change the way XidInMVCCSnapshot() works
2335 : * depending upon when the snapshot was taken, or change normal
2336 : * snapshot processing so it matches.
2337 : *
2338 : * Note: It is possible for recovery to end before we finish taking
2339 : * the snapshot, and for newly assigned transaction ids to be added to
2340 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2341 : * those newly added transaction ids would be filtered away, so we
2342 : * need not be concerned about them.
2343 : */
2344 1601 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2345 : xmax);
2346 :
2347 1601 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2348 8 : suboverflowed = true;
2349 : }
2350 :
2351 :
2352 : /*
2353 : * Fetch into local variable while ProcArrayLock is held - the
2354 : * LWLockRelease below is a barrier, ensuring this happens inside the
2355 : * lock.
2356 : */
2357 503535 : replication_slot_xmin = procArray->replication_slot_xmin;
2358 503535 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2359 :
2360 503535 : if (!TransactionIdIsValid(MyProc->xmin))
2361 274188 : MyProc->xmin = TransactionXmin = xmin;
2362 :
2363 503535 : LWLockRelease(ProcArrayLock);
2364 :
2365 : /* maintain state for GlobalVis* */
2366 : {
2367 : TransactionId def_vis_xid;
2368 : TransactionId def_vis_xid_data;
2369 : FullTransactionId def_vis_fxid;
2370 : FullTransactionId def_vis_fxid_data;
2371 : FullTransactionId oldestfxid;
2372 :
2373 : /*
2374 : * Converting oldestXid is only safe when xid horizon cannot advance,
2375 : * i.e. holding locks. While we don't hold the lock anymore, all the
2376 : * necessary data has been gathered with lock held.
2377 : */
2378 503535 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2379 :
2380 : /* Check whether there's a replication slot requiring an older xmin. */
2381 : def_vis_xid_data =
2382 503535 : TransactionIdOlder(xmin, replication_slot_xmin);
2383 :
2384 : /*
2385 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2386 : * if older than this xid.
2387 : */
2388 503535 : def_vis_xid = def_vis_xid_data;
2389 :
2390 : /*
2391 : * Check whether there's a replication slot requiring an older catalog
2392 : * xmin.
2393 : */
2394 : def_vis_xid =
2395 503535 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2396 :
2397 503535 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2398 503535 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2399 :
2400 : /*
2401 : * Check if we can increase upper bound. As a previous
2402 : * GlobalVisUpdate() might have computed more aggressive values, don't
2403 : * overwrite them if so.
2404 : */
2405 : GlobalVisSharedRels.definitely_needed =
2406 503535 : FullTransactionIdNewer(def_vis_fxid,
2407 : GlobalVisSharedRels.definitely_needed);
2408 : GlobalVisCatalogRels.definitely_needed =
2409 503535 : FullTransactionIdNewer(def_vis_fxid,
2410 : GlobalVisCatalogRels.definitely_needed);
2411 : GlobalVisDataRels.definitely_needed =
2412 503535 : FullTransactionIdNewer(def_vis_fxid_data,
2413 : GlobalVisDataRels.definitely_needed);
2414 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2415 503535 : if (TransactionIdIsNormal(myxid))
2416 : GlobalVisTempRels.definitely_needed =
2417 68206 : FullXidRelativeTo(latest_completed, myxid);
2418 : else
2419 : {
2420 435329 : GlobalVisTempRels.definitely_needed = latest_completed;
2421 435329 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2422 : }
2423 :
2424 : /*
2425 : * Check if we know that we can initialize or increase the lower
2426 : * bound. Currently the only cheap way to do so is to use
2427 : * TransamVariables->oldestXid as input.
2428 : *
2429 : * We should definitely be able to do better. We could e.g. put a
2430 : * global lower bound value into TransamVariables.
2431 : */
2432 : GlobalVisSharedRels.maybe_needed =
2433 503535 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2434 : oldestfxid);
2435 : GlobalVisCatalogRels.maybe_needed =
2436 503535 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2437 : oldestfxid);
2438 : GlobalVisDataRels.maybe_needed =
2439 503535 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2440 : oldestfxid);
2441 : /* accurate value known */
2442 503535 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2443 : }
2444 :
2445 503535 : RecentXmin = xmin;
2446 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2447 :
2448 503535 : snapshot->xmin = xmin;
2449 503535 : snapshot->xmax = xmax;
2450 503535 : snapshot->xcnt = count;
2451 503535 : snapshot->subxcnt = subcount;
2452 503535 : snapshot->suboverflowed = suboverflowed;
2453 503535 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2454 :
2455 503535 : snapshot->curcid = GetCurrentCommandId(false);
2456 :
2457 : /*
2458 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2459 : * not copied in persistent memory.
2460 : */
2461 503535 : snapshot->active_count = 0;
2462 503535 : snapshot->regd_count = 0;
2463 503535 : snapshot->copied = false;
2464 :
2465 503535 : return snapshot;
2466 : }
2467 :
2468 : /*
2469 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2470 : *
2471 : * This is called when installing a snapshot imported from another
2472 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2473 : * check that the source transaction is still running, and we'd better do
2474 : * that atomically with installing the new xmin.
2475 : *
2476 : * Returns true if successful, false if source xact is no longer running.
2477 : */
2478 : bool
2479 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2480 : VirtualTransactionId *sourcevxid)
2481 : {
2482 16 : bool result = false;
2483 16 : ProcArrayStruct *arrayP = procArray;
2484 : int index;
2485 :
2486 : Assert(TransactionIdIsNormal(xmin));
2487 16 : if (!sourcevxid)
2488 0 : return false;
2489 :
2490 : /* Get lock so source xact can't end while we're doing this */
2491 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2492 :
2493 : /*
2494 : * Find the PGPROC entry of the source transaction. (This could use
2495 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2496 : * performance critical.)
2497 : */
2498 16 : for (index = 0; index < arrayP->numProcs; index++)
2499 : {
2500 16 : int pgprocno = arrayP->pgprocnos[index];
2501 16 : PGPROC *proc = &allProcs[pgprocno];
2502 16 : int statusFlags = ProcGlobal->statusFlags[index];
2503 : TransactionId xid;
2504 :
2505 : /* Ignore procs running LAZY VACUUM */
2506 16 : if (statusFlags & PROC_IN_VACUUM)
2507 0 : continue;
2508 :
2509 : /* We are only interested in the specific virtual transaction. */
2510 16 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2511 0 : continue;
2512 16 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2513 0 : continue;
2514 :
2515 : /*
2516 : * We check the transaction's database ID for paranoia's sake: if it's
2517 : * in another DB then its xmin does not cover us. Caller should have
2518 : * detected this already, so we just treat any funny cases as
2519 : * "transaction not found".
2520 : */
2521 16 : if (proc->databaseId != MyDatabaseId)
2522 0 : continue;
2523 :
2524 : /*
2525 : * Likewise, let's just make real sure its xmin does cover us.
2526 : */
2527 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2528 16 : if (!TransactionIdIsNormal(xid) ||
2529 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2530 0 : continue;
2531 :
2532 : /*
2533 : * We're good. Install the new xmin. As in GetSnapshotData, set
2534 : * TransactionXmin too. (Note that because snapmgr.c called
2535 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2536 : * we don't check that.)
2537 : */
2538 16 : MyProc->xmin = TransactionXmin = xmin;
2539 :
2540 16 : result = true;
2541 16 : break;
2542 : }
2543 :
2544 16 : LWLockRelease(ProcArrayLock);
2545 :
2546 16 : return result;
2547 : }
2548 :
2549 : /*
2550 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2551 : *
2552 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2553 : * PGPROC of the transaction from which we imported the snapshot, rather than
2554 : * an XID.
2555 : *
2556 : * Note that this function also copies statusFlags from the source `proc` in
2557 : * order to avoid the case where MyProc's xmin needs to be skipped for
2558 : * computing xid horizon.
2559 : *
2560 : * Returns true if successful, false if source xact is no longer running.
2561 : */
2562 : bool
2563 2228 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2564 : {
2565 2228 : bool result = false;
2566 : TransactionId xid;
2567 :
2568 : Assert(TransactionIdIsNormal(xmin));
2569 : Assert(proc != NULL);
2570 :
2571 : /*
2572 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2573 : */
2574 2228 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2575 :
2576 : /*
2577 : * Be certain that the referenced PGPROC has an advertised xmin which is
2578 : * no later than the one we're installing, so that the system-wide xmin
2579 : * can't go backwards. Also, make sure it's running in the same database,
2580 : * so that the per-database xmin cannot go backwards.
2581 : */
2582 2228 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2583 2228 : if (proc->databaseId == MyDatabaseId &&
2584 2228 : TransactionIdIsNormal(xid) &&
2585 2228 : TransactionIdPrecedesOrEquals(xid, xmin))
2586 : {
2587 : /*
2588 : * Install xmin and propagate the statusFlags that affect how the
2589 : * value is interpreted by vacuum.
2590 : */
2591 2228 : MyProc->xmin = TransactionXmin = xmin;
2592 2228 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2593 2228 : (proc->statusFlags & PROC_XMIN_FLAGS);
2594 2228 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2595 :
2596 2228 : result = true;
2597 : }
2598 :
2599 2228 : LWLockRelease(ProcArrayLock);
2600 :
2601 2228 : return result;
2602 : }
2603 :
2604 : /*
2605 : * GetRunningTransactionData -- returns information about running transactions.
2606 : *
2607 : * Similar to GetSnapshotData but returns more information. We include
2608 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2609 : * prepared transactions.
2610 : *
2611 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2612 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2613 : * array until the caller has WAL-logged this snapshot, and releases the
2614 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2615 : * lock is released.
2616 : *
2617 : * The returned data structure is statically allocated; caller should not
2618 : * modify it, and must not assume it is valid past the next call.
2619 : *
2620 : * This is never executed during recovery so there is no need to look at
2621 : * KnownAssignedXids.
2622 : *
2623 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2624 : * may return entries with duplicated TransactionId values coming from
2625 : * transaction finishing to prepare. Nothing is done about duplicated
2626 : * entries here to not hold on ProcArrayLock more than necessary.
2627 : *
2628 : * We don't worry about updating other counters, we want to keep this as
2629 : * simple as possible and leave GetSnapshotData() as the primary code for
2630 : * that bookkeeping.
2631 : *
2632 : * Note that if any transaction has overflowed its cached subtransactions
2633 : * then there is no real need include any subtransactions.
2634 : */
2635 : RunningTransactions
2636 1535 : GetRunningTransactionData(void)
2637 : {
2638 : /* result workspace */
2639 : static RunningTransactionsData CurrentRunningXactsData;
2640 :
2641 1535 : ProcArrayStruct *arrayP = procArray;
2642 1535 : TransactionId *other_xids = ProcGlobal->xids;
2643 1535 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2644 : TransactionId latestCompletedXid;
2645 : TransactionId oldestRunningXid;
2646 : TransactionId oldestDatabaseRunningXid;
2647 : TransactionId *xids;
2648 : int index;
2649 : int count;
2650 : int subcount;
2651 : bool suboverflowed;
2652 :
2653 : Assert(!RecoveryInProgress());
2654 :
2655 : /*
2656 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2657 : * be sufficient. But it seems better to do the malloc while not holding
2658 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2659 : * more subxip storage than is probably needed.
2660 : *
2661 : * Should only be allocated in bgwriter, since only ever executed during
2662 : * checkpoints.
2663 : */
2664 1535 : if (CurrentRunningXacts->xids == NULL)
2665 : {
2666 : /*
2667 : * First call
2668 : */
2669 621 : CurrentRunningXacts->xids = (TransactionId *)
2670 621 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2671 621 : if (CurrentRunningXacts->xids == NULL)
2672 0 : ereport(ERROR,
2673 : (errcode(ERRCODE_OUT_OF_MEMORY),
2674 : errmsg("out of memory")));
2675 : }
2676 :
2677 1535 : xids = CurrentRunningXacts->xids;
2678 :
2679 1535 : count = subcount = 0;
2680 1535 : suboverflowed = false;
2681 :
2682 : /*
2683 : * Ensure that no xids enter or leave the procarray while we obtain
2684 : * snapshot.
2685 : */
2686 1535 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2687 1535 : LWLockAcquire(XidGenLock, LW_SHARED);
2688 :
2689 1535 : latestCompletedXid =
2690 1535 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2691 1535 : oldestDatabaseRunningXid = oldestRunningXid =
2692 1535 : XidFromFullTransactionId(TransamVariables->nextXid);
2693 :
2694 : /*
2695 : * Spin over procArray collecting all xids
2696 : */
2697 7541 : for (index = 0; index < arrayP->numProcs; index++)
2698 : {
2699 : TransactionId xid;
2700 :
2701 : /* Fetch xid just once - see GetNewTransactionId */
2702 6006 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2703 :
2704 : /*
2705 : * We don't need to store transactions that don't have a TransactionId
2706 : * yet because they will not show as running on a standby server.
2707 : */
2708 6006 : if (!TransactionIdIsValid(xid))
2709 5073 : continue;
2710 :
2711 : /*
2712 : * Be careful not to exclude any xids before calculating the values of
2713 : * oldestRunningXid and suboverflowed, since these are used to clean
2714 : * up transaction information held on standbys.
2715 : */
2716 933 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2717 664 : oldestRunningXid = xid;
2718 :
2719 : /*
2720 : * Also, update the oldest running xid within the current database. As
2721 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2722 : * TransactionId comparison first.
2723 : */
2724 933 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2725 : {
2726 933 : int pgprocno = arrayP->pgprocnos[index];
2727 933 : PGPROC *proc = &allProcs[pgprocno];
2728 :
2729 933 : if (proc->databaseId == MyDatabaseId)
2730 235 : oldestDatabaseRunningXid = xid;
2731 : }
2732 :
2733 933 : if (ProcGlobal->subxidStates[index].overflowed)
2734 1 : suboverflowed = true;
2735 :
2736 : /*
2737 : * If we wished to exclude xids this would be the right place for it.
2738 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2739 : * but they do during truncation at the end when they get the lock and
2740 : * truncate, so it is not much of a problem to include them if they
2741 : * are seen and it is cleaner to include them.
2742 : */
2743 :
2744 933 : xids[count++] = xid;
2745 : }
2746 :
2747 : /*
2748 : * Spin over procArray collecting all subxids, but only if there hasn't
2749 : * been a suboverflow.
2750 : */
2751 1535 : if (!suboverflowed)
2752 : {
2753 1534 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2754 :
2755 7535 : for (index = 0; index < arrayP->numProcs; index++)
2756 : {
2757 6001 : int pgprocno = arrayP->pgprocnos[index];
2758 6001 : PGPROC *proc = &allProcs[pgprocno];
2759 : int nsubxids;
2760 :
2761 : /*
2762 : * Save subtransaction XIDs. Other backends can't add or remove
2763 : * entries while we're holding XidGenLock.
2764 : */
2765 6001 : nsubxids = other_subxidstates[index].count;
2766 6001 : if (nsubxids > 0)
2767 : {
2768 : /* barrier not really required, as XidGenLock is held, but ... */
2769 15 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2770 :
2771 15 : memcpy(&xids[count], proc->subxids.xids,
2772 : nsubxids * sizeof(TransactionId));
2773 15 : count += nsubxids;
2774 15 : subcount += nsubxids;
2775 :
2776 : /*
2777 : * Top-level XID of a transaction is always less than any of
2778 : * its subxids, so we don't need to check if any of the
2779 : * subxids are smaller than oldestRunningXid
2780 : */
2781 : }
2782 : }
2783 : }
2784 :
2785 : /*
2786 : * It's important *not* to include the limits set by slots here because
2787 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2788 : * were to be included here the initial value could never increase because
2789 : * of a circular dependency where slots only increase their limits when
2790 : * running xacts increases oldestRunningXid and running xacts only
2791 : * increases if slots do.
2792 : */
2793 :
2794 1535 : CurrentRunningXacts->xcnt = count - subcount;
2795 1535 : CurrentRunningXacts->subxcnt = subcount;
2796 1535 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2797 1535 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2798 1535 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2799 1535 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2800 1535 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2801 :
2802 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2803 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2804 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2805 :
2806 : /* We don't release the locks here, the caller is responsible for that */
2807 :
2808 1535 : return CurrentRunningXacts;
2809 : }
2810 :
2811 : /*
2812 : * GetOldestActiveTransactionId()
2813 : *
2814 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2815 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2816 : *
2817 : * If allDbs is true, we look at all databases, though there is no need to
2818 : * include WALSender since this has no effect on hot standby conflicts. If
2819 : * allDbs is false, skip processes attached to other databases.
2820 : *
2821 : * This is never executed during recovery so there is no need to look at
2822 : * KnownAssignedXids.
2823 : *
2824 : * We don't worry about updating other counters, we want to keep this as
2825 : * simple as possible and leave GetSnapshotData() as the primary code for
2826 : * that bookkeeping.
2827 : *
2828 : * inCommitOnly indicates getting the oldestActiveXid among the transactions
2829 : * in the commit critical section.
2830 : */
2831 : TransactionId
2832 4479 : GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2833 : {
2834 4479 : ProcArrayStruct *arrayP = procArray;
2835 4479 : TransactionId *other_xids = ProcGlobal->xids;
2836 : TransactionId oldestRunningXid;
2837 : int index;
2838 :
2839 : Assert(!RecoveryInProgress());
2840 :
2841 : /*
2842 : * Read nextXid, as the upper bound of what's still active.
2843 : *
2844 : * Reading a TransactionId is atomic, but we must grab the lock to make
2845 : * sure that all XIDs < nextXid are already present in the proc array (or
2846 : * have already completed), when we spin over it.
2847 : */
2848 4479 : LWLockAcquire(XidGenLock, LW_SHARED);
2849 4479 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2850 4479 : LWLockRelease(XidGenLock);
2851 :
2852 : /*
2853 : * Spin over procArray collecting all xids and subxids.
2854 : */
2855 4479 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2856 25923 : for (index = 0; index < arrayP->numProcs; index++)
2857 : {
2858 : TransactionId xid;
2859 21444 : int pgprocno = arrayP->pgprocnos[index];
2860 21444 : PGPROC *proc = &allProcs[pgprocno];
2861 :
2862 : /* Fetch xid just once - see GetNewTransactionId */
2863 21444 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2864 :
2865 21444 : if (!TransactionIdIsNormal(xid))
2866 17185 : continue;
2867 :
2868 4259 : if (inCommitOnly &&
2869 3382 : (proc->delayChkptFlags & DELAY_CHKPT_IN_COMMIT) == 0)
2870 3300 : continue;
2871 :
2872 959 : if (!allDbs && proc->databaseId != MyDatabaseId)
2873 0 : continue;
2874 :
2875 959 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2876 696 : oldestRunningXid = xid;
2877 :
2878 : /*
2879 : * Top-level XID of a transaction is always less than any of its
2880 : * subxids, so we don't need to check if any of the subxids are
2881 : * smaller than oldestRunningXid
2882 : */
2883 : }
2884 4479 : LWLockRelease(ProcArrayLock);
2885 :
2886 4479 : return oldestRunningXid;
2887 : }
2888 :
2889 : /*
2890 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2891 : *
2892 : * Returns the oldest xid that we can guarantee not to have been affected by
2893 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2894 : * transaction aborted. Note that the value can (and most of the time will) be
2895 : * much more conservative than what really has been affected by vacuum, but we
2896 : * currently don't have better data available.
2897 : *
2898 : * This is useful to initialize the cutoff xid after which a new changeset
2899 : * extraction replication slot can start decoding changes.
2900 : *
2901 : * Must be called with ProcArrayLock held either shared or exclusively,
2902 : * although most callers will want to use exclusive mode since it is expected
2903 : * that the caller will immediately use the xid to peg the xmin horizon.
2904 : */
2905 : TransactionId
2906 746 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2907 : {
2908 746 : ProcArrayStruct *arrayP = procArray;
2909 : TransactionId oldestSafeXid;
2910 : int index;
2911 746 : bool recovery_in_progress = RecoveryInProgress();
2912 :
2913 : Assert(LWLockHeldByMe(ProcArrayLock));
2914 :
2915 : /*
2916 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2917 : * running. If no transaction with xid were running concurrently a new xid
2918 : * could influence the RecentXmin et al.
2919 : *
2920 : * We initialize the computation to nextXid since that's guaranteed to be
2921 : * a safe, albeit pessimal, value.
2922 : */
2923 746 : LWLockAcquire(XidGenLock, LW_SHARED);
2924 746 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2925 :
2926 : /*
2927 : * If there's already a slot pegging the xmin horizon, we can start with
2928 : * that value, it's guaranteed to be safe since it's computed by this
2929 : * routine initially and has been enforced since. We can always use the
2930 : * slot's general xmin horizon, but the catalog horizon is only usable
2931 : * when only catalog data is going to be looked at.
2932 : */
2933 987 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2934 241 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2935 : oldestSafeXid))
2936 10 : oldestSafeXid = procArray->replication_slot_xmin;
2937 :
2938 746 : if (catalogOnly &&
2939 376 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2940 80 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2941 : oldestSafeXid))
2942 31 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2943 :
2944 : /*
2945 : * If we're not in recovery, we walk over the procarray and collect the
2946 : * lowest xid. Since we're called with ProcArrayLock held and have
2947 : * acquired XidGenLock, no entries can vanish concurrently, since
2948 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2949 : * with ProcArrayLock held.
2950 : *
2951 : * In recovery we can't lower the safe value besides what we've computed
2952 : * above, so we'll have to wait a bit longer there. We unfortunately can
2953 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2954 : * machinery can miss values and return an older value than is safe.
2955 : */
2956 746 : if (!recovery_in_progress)
2957 : {
2958 710 : TransactionId *other_xids = ProcGlobal->xids;
2959 :
2960 : /*
2961 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2962 : */
2963 3624 : for (index = 0; index < arrayP->numProcs; index++)
2964 : {
2965 : TransactionId xid;
2966 :
2967 : /* Fetch xid just once - see GetNewTransactionId */
2968 2914 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2969 :
2970 2914 : if (!TransactionIdIsNormal(xid))
2971 2907 : continue;
2972 :
2973 7 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2974 6 : oldestSafeXid = xid;
2975 : }
2976 : }
2977 :
2978 746 : LWLockRelease(XidGenLock);
2979 :
2980 746 : return oldestSafeXid;
2981 : }
2982 :
2983 : /*
2984 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2985 : * delaying checkpoint because they have critical actions in progress.
2986 : *
2987 : * Constructs an array of VXIDs of transactions that are currently in commit
2988 : * critical sections, as shown by having specified delayChkptFlags bits set
2989 : * in their PGPROC.
2990 : *
2991 : * Returns a palloc'd array that should be freed by the caller.
2992 : * *nvxids is the number of valid entries.
2993 : *
2994 : * Note that because backends set or clear delayChkptFlags without holding any
2995 : * lock, the result is somewhat indeterminate, but we don't really care. Even
2996 : * in a multiprocessor with delayed writes to shared memory, it should be
2997 : * certain that setting of delayChkptFlags will propagate to shared memory
2998 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
2999 : * delayChkptFlags if it's already inserted its commit record. Whether it
3000 : * takes a little while for clearing of delayChkptFlags to propagate is
3001 : * unimportant for correctness.
3002 : */
3003 : VirtualTransactionId *
3004 3462 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3005 : {
3006 : VirtualTransactionId *vxids;
3007 3462 : ProcArrayStruct *arrayP = procArray;
3008 3462 : int count = 0;
3009 : int index;
3010 :
3011 : Assert(type != 0);
3012 :
3013 : /* allocate what's certainly enough result space */
3014 3462 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3015 :
3016 3462 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3017 :
3018 11359 : for (index = 0; index < arrayP->numProcs; index++)
3019 : {
3020 7897 : int pgprocno = arrayP->pgprocnos[index];
3021 7897 : PGPROC *proc = &allProcs[pgprocno];
3022 :
3023 7897 : if ((proc->delayChkptFlags & type) != 0)
3024 : {
3025 : VirtualTransactionId vxid;
3026 :
3027 24 : GET_VXID_FROM_PGPROC(vxid, *proc);
3028 24 : if (VirtualTransactionIdIsValid(vxid))
3029 24 : vxids[count++] = vxid;
3030 : }
3031 : }
3032 :
3033 3462 : LWLockRelease(ProcArrayLock);
3034 :
3035 3462 : *nvxids = count;
3036 3462 : return vxids;
3037 : }
3038 :
3039 : /*
3040 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3041 : *
3042 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3043 : * of the specified VXIDs are still in critical sections of code.
3044 : *
3045 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3046 : * those numbers should be small enough for it not to be a problem.
3047 : */
3048 : bool
3049 19 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3050 : {
3051 19 : bool result = false;
3052 19 : ProcArrayStruct *arrayP = procArray;
3053 : int index;
3054 :
3055 : Assert(type != 0);
3056 :
3057 19 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3058 :
3059 212 : for (index = 0; index < arrayP->numProcs; index++)
3060 : {
3061 193 : int pgprocno = arrayP->pgprocnos[index];
3062 193 : PGPROC *proc = &allProcs[pgprocno];
3063 : VirtualTransactionId vxid;
3064 :
3065 193 : GET_VXID_FROM_PGPROC(vxid, *proc);
3066 :
3067 193 : if ((proc->delayChkptFlags & type) != 0 &&
3068 8 : VirtualTransactionIdIsValid(vxid))
3069 : {
3070 : int i;
3071 :
3072 20 : for (i = 0; i < nvxids; i++)
3073 : {
3074 12 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3075 : {
3076 0 : result = true;
3077 0 : break;
3078 : }
3079 : }
3080 8 : if (result)
3081 0 : break;
3082 : }
3083 : }
3084 :
3085 19 : LWLockRelease(ProcArrayLock);
3086 :
3087 19 : return result;
3088 : }
3089 :
3090 : /*
3091 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3092 : *
3093 : * The result may be out of date arbitrarily quickly, so the caller
3094 : * must be careful about how this information is used. NULL is
3095 : * returned if the backend is not active.
3096 : */
3097 : PGPROC *
3098 671 : ProcNumberGetProc(ProcNumber procNumber)
3099 : {
3100 : PGPROC *result;
3101 :
3102 671 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3103 1 : return NULL;
3104 670 : result = GetPGProcByNumber(procNumber);
3105 :
3106 670 : if (result->pid == 0)
3107 3 : return NULL;
3108 :
3109 667 : return result;
3110 : }
3111 :
3112 : /*
3113 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3114 : *
3115 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3116 : * result may be out of date arbitrarily quickly, so the caller must be
3117 : * careful about how this information is used.
3118 : */
3119 : void
3120 8285 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3121 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3122 : {
3123 : PGPROC *proc;
3124 :
3125 8285 : *xid = InvalidTransactionId;
3126 8285 : *xmin = InvalidTransactionId;
3127 8285 : *nsubxid = 0;
3128 8285 : *overflowed = false;
3129 :
3130 8285 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3131 0 : return;
3132 8285 : proc = GetPGProcByNumber(procNumber);
3133 :
3134 : /* Need to lock out additions/removals of backends */
3135 8285 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3136 :
3137 8285 : if (proc->pid != 0)
3138 : {
3139 8285 : *xid = proc->xid;
3140 8285 : *xmin = proc->xmin;
3141 8285 : *nsubxid = proc->subxidStatus.count;
3142 8285 : *overflowed = proc->subxidStatus.overflowed;
3143 : }
3144 :
3145 8285 : LWLockRelease(ProcArrayLock);
3146 : }
3147 :
3148 : /*
3149 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3150 : *
3151 : * Returns NULL if not found. Note that it is up to the caller to be
3152 : * sure that the question remains meaningful for long enough for the
3153 : * answer to be used ...
3154 : */
3155 : PGPROC *
3156 9128 : BackendPidGetProc(int pid)
3157 : {
3158 : PGPROC *result;
3159 :
3160 9128 : if (pid == 0) /* never match dummy PGPROCs */
3161 4 : return NULL;
3162 :
3163 9124 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3164 :
3165 9124 : result = BackendPidGetProcWithLock(pid);
3166 :
3167 9124 : LWLockRelease(ProcArrayLock);
3168 :
3169 9124 : return result;
3170 : }
3171 :
3172 : /*
3173 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3174 : *
3175 : * Same as above, except caller must be holding ProcArrayLock. The found
3176 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3177 : */
3178 : PGPROC *
3179 10819 : BackendPidGetProcWithLock(int pid)
3180 : {
3181 10819 : PGPROC *result = NULL;
3182 10819 : ProcArrayStruct *arrayP = procArray;
3183 : int index;
3184 :
3185 10819 : if (pid == 0) /* never match dummy PGPROCs */
3186 0 : return NULL;
3187 :
3188 42388 : for (index = 0; index < arrayP->numProcs; index++)
3189 : {
3190 38506 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3191 :
3192 38506 : if (proc->pid == pid)
3193 : {
3194 6937 : result = proc;
3195 6937 : break;
3196 : }
3197 : }
3198 :
3199 10819 : return result;
3200 : }
3201 :
3202 : /*
3203 : * BackendXidGetPid -- get a backend's pid given its XID
3204 : *
3205 : * Returns 0 if not found or it's a prepared transaction. Note that
3206 : * it is up to the caller to be sure that the question remains
3207 : * meaningful for long enough for the answer to be used ...
3208 : *
3209 : * Only main transaction Ids are considered. This function is mainly
3210 : * useful for determining what backend owns a lock.
3211 : *
3212 : * Beware that not every xact has an XID assigned. However, as long as you
3213 : * only call this using an XID found on disk, you're safe.
3214 : */
3215 : int
3216 30 : BackendXidGetPid(TransactionId xid)
3217 : {
3218 30 : int result = 0;
3219 30 : ProcArrayStruct *arrayP = procArray;
3220 30 : TransactionId *other_xids = ProcGlobal->xids;
3221 : int index;
3222 :
3223 30 : if (xid == InvalidTransactionId) /* never match invalid xid */
3224 0 : return 0;
3225 :
3226 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3227 :
3228 92 : for (index = 0; index < arrayP->numProcs; index++)
3229 : {
3230 84 : if (other_xids[index] == xid)
3231 : {
3232 22 : int pgprocno = arrayP->pgprocnos[index];
3233 22 : PGPROC *proc = &allProcs[pgprocno];
3234 :
3235 22 : result = proc->pid;
3236 22 : break;
3237 : }
3238 : }
3239 :
3240 30 : LWLockRelease(ProcArrayLock);
3241 :
3242 30 : return result;
3243 : }
3244 :
3245 : /*
3246 : * IsBackendPid -- is a given pid a running backend
3247 : *
3248 : * This is not called by the backend, but is called by external modules.
3249 : */
3250 : bool
3251 2 : IsBackendPid(int pid)
3252 : {
3253 2 : return (BackendPidGetProc(pid) != NULL);
3254 : }
3255 :
3256 :
3257 : /*
3258 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3259 : *
3260 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3261 : *
3262 : * The arguments allow filtering the set of VXIDs returned. Our own process
3263 : * is always skipped. In addition:
3264 : * If limitXmin is not InvalidTransactionId, skip processes with
3265 : * xmin > limitXmin.
3266 : * If excludeXmin0 is true, skip processes with xmin = 0.
3267 : * If allDbs is false, skip processes attached to other databases.
3268 : * If excludeVacuum isn't zero, skip processes for which
3269 : * (statusFlags & excludeVacuum) is not zero.
3270 : *
3271 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3272 : * allow skipping backends whose oldest live snapshot is no older than
3273 : * some snapshot we have. Since we examine the procarray with only shared
3274 : * lock, there are race conditions: a backend could set its xmin just after
3275 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3276 : * other backend could have set its xmin *before* we look. We know however
3277 : * that such a backend must have held shared ProcArrayLock overlapping our
3278 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3279 : * any snapshot the other backend is taking concurrently with our scan cannot
3280 : * consider any transactions as still running that we think are committed
3281 : * (since backends must hold ProcArrayLock exclusive to commit).
3282 : */
3283 : VirtualTransactionId *
3284 529 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3285 : bool allDbs, int excludeVacuum,
3286 : int *nvxids)
3287 : {
3288 : VirtualTransactionId *vxids;
3289 529 : ProcArrayStruct *arrayP = procArray;
3290 529 : int count = 0;
3291 : int index;
3292 :
3293 : /* allocate what's certainly enough result space */
3294 529 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3295 :
3296 529 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3297 :
3298 3270 : for (index = 0; index < arrayP->numProcs; index++)
3299 : {
3300 2741 : int pgprocno = arrayP->pgprocnos[index];
3301 2741 : PGPROC *proc = &allProcs[pgprocno];
3302 2741 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3303 :
3304 2741 : if (proc == MyProc)
3305 529 : continue;
3306 :
3307 2212 : if (excludeVacuum & statusFlags)
3308 14 : continue;
3309 :
3310 2198 : if (allDbs || proc->databaseId == MyDatabaseId)
3311 : {
3312 : /* Fetch xmin just once - might change on us */
3313 1055 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3314 :
3315 1055 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3316 614 : continue;
3317 :
3318 : /*
3319 : * InvalidTransactionId precedes all other XIDs, so a proc that
3320 : * hasn't set xmin yet will not be rejected by this test.
3321 : */
3322 882 : if (!TransactionIdIsValid(limitXmin) ||
3323 441 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3324 : {
3325 : VirtualTransactionId vxid;
3326 :
3327 404 : GET_VXID_FROM_PGPROC(vxid, *proc);
3328 404 : if (VirtualTransactionIdIsValid(vxid))
3329 404 : vxids[count++] = vxid;
3330 : }
3331 : }
3332 : }
3333 :
3334 529 : LWLockRelease(ProcArrayLock);
3335 :
3336 529 : *nvxids = count;
3337 529 : return vxids;
3338 : }
3339 :
3340 : /*
3341 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3342 : *
3343 : * Usage is limited to conflict resolution during recovery on standby servers.
3344 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3345 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3346 : * determine a safe snapshotConflictHorizon value.
3347 : *
3348 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3349 : * so we're not worried if they have a snapshot or not, nor does it really
3350 : * matter what type of lock we hold. Caller must avoid calling here with
3351 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3352 : * during original execution, since that actually indicates that there is
3353 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3354 : * convention for InvalidTransactionId values is the opposite of our own!).
3355 : *
3356 : * All callers that are checking xmins always now supply a valid and useful
3357 : * value for limitXmin. The limitXmin is always lower than the lowest
3358 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3359 : * because we only care about cleanup records that are cleaning up tuple
3360 : * versions from committed transactions. In that case they will only occur
3361 : * at the point where the record is less than the lowest running xid. That
3362 : * allows us to say that if any backend takes a snapshot concurrently with
3363 : * us then the conflict assessment made here would never include the snapshot
3364 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3365 : * concurrent snapshots when limitXmin is valid. We might think about adding
3366 : * Assert(limitXmin < lowest(KnownAssignedXids))
3367 : * but that would not be true in the case of FATAL errors lagging in array,
3368 : * but we already know those are bogus anyway, so we skip that test.
3369 : *
3370 : * If dbOid is valid we skip backends attached to other databases.
3371 : *
3372 : * Be careful to *not* pfree the result from this function. We reuse
3373 : * this array sufficiently often that we use malloc for the result.
3374 : */
3375 : VirtualTransactionId *
3376 16169 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3377 : {
3378 : static VirtualTransactionId *vxids;
3379 16169 : ProcArrayStruct *arrayP = procArray;
3380 16169 : int count = 0;
3381 : int index;
3382 :
3383 : /*
3384 : * If first time through, get workspace to remember main XIDs in. We
3385 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3386 : * result space, remembering room for a terminator.
3387 : */
3388 16169 : if (vxids == NULL)
3389 : {
3390 19 : vxids = (VirtualTransactionId *)
3391 19 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3392 19 : if (vxids == NULL)
3393 0 : ereport(ERROR,
3394 : (errcode(ERRCODE_OUT_OF_MEMORY),
3395 : errmsg("out of memory")));
3396 : }
3397 :
3398 16169 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3399 :
3400 16405 : for (index = 0; index < arrayP->numProcs; index++)
3401 : {
3402 236 : int pgprocno = arrayP->pgprocnos[index];
3403 236 : PGPROC *proc = &allProcs[pgprocno];
3404 :
3405 : /* Exclude prepared transactions */
3406 236 : if (proc->pid == 0)
3407 0 : continue;
3408 :
3409 236 : if (!OidIsValid(dbOid) ||
3410 230 : proc->databaseId == dbOid)
3411 : {
3412 : /* Fetch xmin just once - can't change on us, but good coding */
3413 23 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3414 :
3415 : /*
3416 : * We ignore an invalid pxmin because this means that backend has
3417 : * no snapshot currently. We hold a Share lock to avoid contention
3418 : * with users taking snapshots. That is not a problem because the
3419 : * current xmin is always at least one higher than the latest
3420 : * removed xid, so any new snapshot would never conflict with the
3421 : * test here.
3422 : */
3423 23 : if (!TransactionIdIsValid(limitXmin) ||
3424 5 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3425 : {
3426 : VirtualTransactionId vxid;
3427 :
3428 2 : GET_VXID_FROM_PGPROC(vxid, *proc);
3429 2 : if (VirtualTransactionIdIsValid(vxid))
3430 2 : vxids[count++] = vxid;
3431 : }
3432 : }
3433 : }
3434 :
3435 16169 : LWLockRelease(ProcArrayLock);
3436 :
3437 : /* add the terminator */
3438 16169 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3439 16169 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3440 :
3441 16169 : return vxids;
3442 : }
3443 :
3444 : /*
3445 : * SignalRecoveryConflict -- signal that a process is blocking recovery
3446 : *
3447 : * The 'pid' is redundant with 'proc', but it acts as a cross-check to
3448 : * detect process had exited and the PGPROC entry was reused for a different
3449 : * process.
3450 : *
3451 : * Returns true if the process was signaled, or false if not found.
3452 : */
3453 : bool
3454 5 : SignalRecoveryConflict(PGPROC *proc, pid_t pid, RecoveryConflictReason reason)
3455 : {
3456 5 : bool found = false;
3457 :
3458 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3459 :
3460 : /*
3461 : * Kill the pid if it's still here. If not, that's what we wanted so
3462 : * ignore any errors.
3463 : */
3464 5 : if (proc->pid == pid)
3465 : {
3466 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3467 :
3468 : /* wake up the process */
3469 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, GetNumberFromPGProc(proc));
3470 5 : found = true;
3471 : }
3472 :
3473 5 : LWLockRelease(ProcArrayLock);
3474 :
3475 5 : return found;
3476 : }
3477 :
3478 : /*
3479 : * SignalRecoveryConflictWithVirtualXID -- signal that a VXID is blocking recovery
3480 : *
3481 : * Like SignalRecoveryConflict, but the target is identified by VXID
3482 : */
3483 : bool
3484 5 : SignalRecoveryConflictWithVirtualXID(VirtualTransactionId vxid, RecoveryConflictReason reason)
3485 : {
3486 5 : ProcArrayStruct *arrayP = procArray;
3487 : int index;
3488 5 : pid_t pid = 0;
3489 :
3490 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3491 :
3492 5 : for (index = 0; index < arrayP->numProcs; index++)
3493 : {
3494 5 : int pgprocno = arrayP->pgprocnos[index];
3495 5 : PGPROC *proc = &allProcs[pgprocno];
3496 : VirtualTransactionId procvxid;
3497 :
3498 5 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3499 :
3500 5 : if (procvxid.procNumber == vxid.procNumber &&
3501 5 : procvxid.localTransactionId == vxid.localTransactionId)
3502 : {
3503 5 : pid = proc->pid;
3504 5 : if (pid != 0)
3505 : {
3506 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3507 :
3508 : /*
3509 : * Kill the pid if it's still here. If not, that's what we
3510 : * wanted so ignore any errors.
3511 : */
3512 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, vxid.procNumber);
3513 : }
3514 5 : break;
3515 : }
3516 : }
3517 :
3518 5 : LWLockRelease(ProcArrayLock);
3519 :
3520 5 : return pid != 0;
3521 : }
3522 :
3523 : /*
3524 : * SignalRecoveryConflictWithDatabase -- signal backends using specified database
3525 : *
3526 : * Like SignalRecoveryConflict, but signals all backends using the database.
3527 : */
3528 : void
3529 9 : SignalRecoveryConflictWithDatabase(Oid databaseid, RecoveryConflictReason reason)
3530 : {
3531 9 : ProcArrayStruct *arrayP = procArray;
3532 : int index;
3533 :
3534 : /* tell all backends to die */
3535 9 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3536 :
3537 18 : for (index = 0; index < arrayP->numProcs; index++)
3538 : {
3539 9 : int pgprocno = arrayP->pgprocnos[index];
3540 9 : PGPROC *proc = &allProcs[pgprocno];
3541 :
3542 9 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3543 : {
3544 : VirtualTransactionId procvxid;
3545 : pid_t pid;
3546 :
3547 9 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3548 :
3549 9 : pid = proc->pid;
3550 9 : if (pid != 0)
3551 : {
3552 9 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3553 :
3554 : /*
3555 : * Kill the pid if it's still here. If not, that's what we
3556 : * wanted so ignore any errors.
3557 : */
3558 9 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, procvxid.procNumber);
3559 : }
3560 : }
3561 : }
3562 :
3563 9 : LWLockRelease(ProcArrayLock);
3564 9 : }
3565 :
3566 : /*
3567 : * MinimumActiveBackends --- count backends (other than myself) that are
3568 : * in active transactions. Return true if the count exceeds the
3569 : * minimum threshold passed. This is used as a heuristic to decide if
3570 : * a pre-XLOG-flush delay is worthwhile during commit.
3571 : *
3572 : * Do not count backends that are blocked waiting for locks, since they are
3573 : * not going to get to run until someone else commits.
3574 : */
3575 : bool
3576 0 : MinimumActiveBackends(int min)
3577 : {
3578 0 : ProcArrayStruct *arrayP = procArray;
3579 0 : int count = 0;
3580 : int index;
3581 :
3582 : /* Quick short-circuit if no minimum is specified */
3583 0 : if (min == 0)
3584 0 : return true;
3585 :
3586 : /*
3587 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3588 : * bogus, but since we are only testing fields for zero or nonzero, it
3589 : * should be OK. The result is only used for heuristic purposes anyway...
3590 : */
3591 0 : for (index = 0; index < arrayP->numProcs; index++)
3592 : {
3593 0 : int pgprocno = arrayP->pgprocnos[index];
3594 0 : PGPROC *proc = &allProcs[pgprocno];
3595 :
3596 : /*
3597 : * Since we're not holding a lock, need to be prepared to deal with
3598 : * garbage, as someone could have incremented numProcs but not yet
3599 : * filled the structure.
3600 : *
3601 : * If someone just decremented numProcs, 'proc' could also point to a
3602 : * PGPROC entry that's no longer in the array. It still points to a
3603 : * PGPROC struct, though, because freed PGPROC entries just go to the
3604 : * free list and are recycled. Its contents are nonsense in that case,
3605 : * but that's acceptable for this function.
3606 : */
3607 0 : if (pgprocno == -1)
3608 0 : continue; /* do not count deleted entries */
3609 0 : if (proc == MyProc)
3610 0 : continue; /* do not count myself */
3611 0 : if (proc->xid == InvalidTransactionId)
3612 0 : continue; /* do not count if no XID assigned */
3613 0 : if (proc->pid == 0)
3614 0 : continue; /* do not count prepared xacts */
3615 0 : if (proc->waitLock != NULL)
3616 0 : continue; /* do not count if blocked on a lock */
3617 0 : count++;
3618 0 : if (count >= min)
3619 0 : break;
3620 : }
3621 :
3622 0 : return count >= min;
3623 : }
3624 :
3625 : /*
3626 : * CountDBBackends --- count backends that are using specified database
3627 : */
3628 : int
3629 17 : CountDBBackends(Oid databaseid)
3630 : {
3631 17 : ProcArrayStruct *arrayP = procArray;
3632 17 : int count = 0;
3633 : int index;
3634 :
3635 17 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3636 :
3637 24 : for (index = 0; index < arrayP->numProcs; index++)
3638 : {
3639 7 : int pgprocno = arrayP->pgprocnos[index];
3640 7 : PGPROC *proc = &allProcs[pgprocno];
3641 :
3642 7 : if (proc->pid == 0)
3643 0 : continue; /* do not count prepared xacts */
3644 7 : if (!OidIsValid(databaseid) ||
3645 7 : proc->databaseId == databaseid)
3646 2 : count++;
3647 : }
3648 :
3649 17 : LWLockRelease(ProcArrayLock);
3650 :
3651 17 : return count;
3652 : }
3653 :
3654 : /*
3655 : * CountDBConnections --- counts database backends (only regular backends)
3656 : */
3657 : int
3658 0 : CountDBConnections(Oid databaseid)
3659 : {
3660 0 : ProcArrayStruct *arrayP = procArray;
3661 0 : int count = 0;
3662 : int index;
3663 :
3664 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3665 :
3666 0 : for (index = 0; index < arrayP->numProcs; index++)
3667 : {
3668 0 : int pgprocno = arrayP->pgprocnos[index];
3669 0 : PGPROC *proc = &allProcs[pgprocno];
3670 :
3671 0 : if (proc->pid == 0)
3672 0 : continue; /* do not count prepared xacts */
3673 0 : if (proc->backendType != B_BACKEND)
3674 0 : continue; /* count only regular backend processes */
3675 0 : if (!OidIsValid(databaseid) ||
3676 0 : proc->databaseId == databaseid)
3677 0 : count++;
3678 : }
3679 :
3680 0 : LWLockRelease(ProcArrayLock);
3681 :
3682 0 : return count;
3683 : }
3684 :
3685 : /*
3686 : * CountUserBackends --- count backends that are used by specified user
3687 : * (only regular backends, not any type of background worker)
3688 : */
3689 : int
3690 0 : CountUserBackends(Oid roleid)
3691 : {
3692 0 : ProcArrayStruct *arrayP = procArray;
3693 0 : int count = 0;
3694 : int index;
3695 :
3696 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3697 :
3698 0 : for (index = 0; index < arrayP->numProcs; index++)
3699 : {
3700 0 : int pgprocno = arrayP->pgprocnos[index];
3701 0 : PGPROC *proc = &allProcs[pgprocno];
3702 :
3703 0 : if (proc->pid == 0)
3704 0 : continue; /* do not count prepared xacts */
3705 0 : if (proc->backendType != B_BACKEND)
3706 0 : continue; /* count only regular backend processes */
3707 0 : if (proc->roleId == roleid)
3708 0 : count++;
3709 : }
3710 :
3711 0 : LWLockRelease(ProcArrayLock);
3712 :
3713 0 : return count;
3714 : }
3715 :
3716 : /*
3717 : * CountOtherDBBackends -- check for other backends running in the given DB
3718 : *
3719 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3720 : * for them to exit (or 0.3s for testing purposes). Autovacuum backends are
3721 : * encouraged to exit early by sending them SIGTERM, but normal user backends
3722 : * are just waited for. If background workers connected to this database are
3723 : * marked as interruptible, they are terminated.
3724 : *
3725 : * The current backend is always ignored; it is caller's responsibility to
3726 : * check whether the current backend uses the given DB, if it's important.
3727 : *
3728 : * Returns true if there are (still) other backends in the DB, false if not.
3729 : * Also, *nbackends and *nprepared are set to the number of other backends
3730 : * and prepared transactions in the DB, respectively.
3731 : *
3732 : * This function is used to interlock DROP DATABASE and related commands
3733 : * against there being any active backends in the target DB --- dropping the
3734 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3735 : * detect here the possibility of a newly-started backend that is trying to
3736 : * connect to the doomed database, so additional interlocking is needed during
3737 : * backend startup. The caller should normally hold an exclusive lock on the
3738 : * target DB before calling this, which is one reason we mustn't wait
3739 : * indefinitely.
3740 : */
3741 : bool
3742 496 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3743 : {
3744 496 : ProcArrayStruct *arrayP = procArray;
3745 :
3746 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3747 : int autovac_pids[MAXAUTOVACPIDS];
3748 :
3749 : /*
3750 : * Retry up to 50 times with 100ms between attempts (max 5s total). Can be
3751 : * reduced to 3 attempts (max 0.3s total) to speed up tests.
3752 : */
3753 496 : int ntries = 50;
3754 :
3755 : #ifdef USE_INJECTION_POINTS
3756 496 : if (IS_INJECTION_POINT_ATTACHED("procarray-reduce-count"))
3757 1 : ntries = 3;
3758 : #endif
3759 :
3760 503 : for (int tries = 0; tries < ntries; tries++)
3761 : {
3762 502 : int nautovacs = 0;
3763 502 : bool found = false;
3764 : int index;
3765 :
3766 502 : CHECK_FOR_INTERRUPTS();
3767 :
3768 502 : *nbackends = *nprepared = 0;
3769 :
3770 502 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3771 :
3772 1890 : for (index = 0; index < arrayP->numProcs; index++)
3773 : {
3774 1388 : int pgprocno = arrayP->pgprocnos[index];
3775 1388 : PGPROC *proc = &allProcs[pgprocno];
3776 1388 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3777 :
3778 1388 : if (proc->databaseId != databaseId)
3779 1266 : continue;
3780 122 : if (proc == MyProc)
3781 115 : continue;
3782 :
3783 7 : found = true;
3784 :
3785 7 : if (proc->pid == 0)
3786 0 : (*nprepared)++;
3787 : else
3788 : {
3789 7 : (*nbackends)++;
3790 7 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3791 : nautovacs < MAXAUTOVACPIDS)
3792 0 : autovac_pids[nautovacs++] = proc->pid;
3793 : }
3794 : }
3795 :
3796 502 : LWLockRelease(ProcArrayLock);
3797 :
3798 502 : if (!found)
3799 495 : return false; /* no conflicting backends, so done */
3800 :
3801 : /*
3802 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3803 : * postpone this step until after the loop because we don't want to
3804 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3805 : * block kill() inside the kernel...
3806 : */
3807 7 : for (index = 0; index < nautovacs; index++)
3808 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3809 :
3810 : /*
3811 : * Terminate all background workers for this database, if they have
3812 : * requested it (BGWORKER_INTERRUPTIBLE).
3813 : */
3814 7 : TerminateBackgroundWorkersForDatabase(databaseId);
3815 :
3816 : /* sleep, then try again */
3817 7 : pg_usleep(100 * 1000L); /* 100ms */
3818 : }
3819 :
3820 1 : return true; /* timed out, still conflicts */
3821 : }
3822 :
3823 : /*
3824 : * Terminate existing connections to the specified database. This routine
3825 : * is used by the DROP DATABASE command when user has asked to forcefully
3826 : * drop the database.
3827 : *
3828 : * The current backend is always ignored; it is caller's responsibility to
3829 : * check whether the current backend uses the given DB, if it's important.
3830 : *
3831 : * If the target database has a prepared transaction or permissions checks
3832 : * fail for a connection, this fails without terminating anything.
3833 : */
3834 : void
3835 1 : TerminateOtherDBBackends(Oid databaseId)
3836 : {
3837 1 : ProcArrayStruct *arrayP = procArray;
3838 1 : List *pids = NIL;
3839 1 : int nprepared = 0;
3840 : int i;
3841 :
3842 1 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3843 :
3844 4 : for (i = 0; i < procArray->numProcs; i++)
3845 : {
3846 3 : int pgprocno = arrayP->pgprocnos[i];
3847 3 : PGPROC *proc = &allProcs[pgprocno];
3848 :
3849 3 : if (proc->databaseId != databaseId)
3850 3 : continue;
3851 0 : if (proc == MyProc)
3852 0 : continue;
3853 :
3854 0 : if (proc->pid != 0)
3855 0 : pids = lappend_int(pids, proc->pid);
3856 : else
3857 0 : nprepared++;
3858 : }
3859 :
3860 1 : LWLockRelease(ProcArrayLock);
3861 :
3862 1 : if (nprepared > 0)
3863 0 : ereport(ERROR,
3864 : (errcode(ERRCODE_OBJECT_IN_USE),
3865 : errmsg("database \"%s\" is being used by prepared transactions",
3866 : get_database_name(databaseId)),
3867 : errdetail_plural("There is %d prepared transaction using the database.",
3868 : "There are %d prepared transactions using the database.",
3869 : nprepared,
3870 : nprepared)));
3871 :
3872 1 : if (pids)
3873 : {
3874 : ListCell *lc;
3875 :
3876 : /*
3877 : * Permissions checks relax the pg_terminate_backend checks in two
3878 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3879 : *
3880 : * - Accept terminating autovacuum workers, since DROP DATABASE
3881 : * without FORCE terminates them.
3882 : *
3883 : * - Accept terminating bgworkers. For bgworker authors, it's
3884 : * convenient to be able to recommend FORCE if a worker is blocking
3885 : * DROP DATABASE unexpectedly.
3886 : *
3887 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3888 : * "PID %d is not a PostgreSQL server process", because for us already
3889 : * finished session is not a problem.
3890 : */
3891 0 : foreach(lc, pids)
3892 : {
3893 0 : int pid = lfirst_int(lc);
3894 0 : PGPROC *proc = BackendPidGetProc(pid);
3895 :
3896 0 : if (proc != NULL)
3897 : {
3898 0 : if (superuser_arg(proc->roleId) && !superuser())
3899 0 : ereport(ERROR,
3900 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3901 : errmsg("permission denied to terminate process"),
3902 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3903 : "SUPERUSER", "SUPERUSER")));
3904 :
3905 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3906 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3907 0 : ereport(ERROR,
3908 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3909 : errmsg("permission denied to terminate process"),
3910 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3911 : "pg_signal_backend")));
3912 : }
3913 : }
3914 :
3915 : /*
3916 : * There's a race condition here: once we release the ProcArrayLock,
3917 : * it's possible for the session to exit before we issue kill. That
3918 : * race condition possibility seems too unlikely to worry about. See
3919 : * pg_signal_backend.
3920 : */
3921 0 : foreach(lc, pids)
3922 : {
3923 0 : int pid = lfirst_int(lc);
3924 0 : PGPROC *proc = BackendPidGetProc(pid);
3925 :
3926 0 : if (proc != NULL)
3927 : {
3928 : /*
3929 : * If we have setsid(), signal the backend's whole process
3930 : * group
3931 : */
3932 : #ifdef HAVE_SETSID
3933 0 : (void) kill(-pid, SIGTERM);
3934 : #else
3935 : (void) kill(pid, SIGTERM);
3936 : #endif
3937 : }
3938 : }
3939 : }
3940 1 : }
3941 :
3942 : /*
3943 : * ProcArraySetReplicationSlotXmin
3944 : *
3945 : * Install limits to future computations of the xmin horizon to prevent vacuum
3946 : * and HOT pruning from removing affected rows still needed by clients with
3947 : * replication slots.
3948 : */
3949 : void
3950 2661 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3951 : bool already_locked)
3952 : {
3953 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3954 :
3955 2661 : if (!already_locked)
3956 2138 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3957 :
3958 2661 : procArray->replication_slot_xmin = xmin;
3959 2661 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3960 :
3961 2661 : if (!already_locked)
3962 2138 : LWLockRelease(ProcArrayLock);
3963 :
3964 2661 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3965 : xmin, catalog_xmin);
3966 2661 : }
3967 :
3968 : /*
3969 : * ProcArrayGetReplicationSlotXmin
3970 : *
3971 : * Return the current slot xmin limits. That's useful to be able to remove
3972 : * data that's older than those limits.
3973 : */
3974 : void
3975 22 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3976 : TransactionId *catalog_xmin)
3977 : {
3978 22 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3979 :
3980 22 : if (xmin != NULL)
3981 0 : *xmin = procArray->replication_slot_xmin;
3982 :
3983 22 : if (catalog_xmin != NULL)
3984 22 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3985 :
3986 22 : LWLockRelease(ProcArrayLock);
3987 22 : }
3988 :
3989 : /*
3990 : * XidCacheRemoveRunningXids
3991 : *
3992 : * Remove a bunch of TransactionIds from the list of known-running
3993 : * subtransactions for my backend. Both the specified xid and those in
3994 : * the xids[] array (of length nxids) are removed from the subxids cache.
3995 : * latestXid must be the latest XID among the group.
3996 : */
3997 : void
3998 860 : XidCacheRemoveRunningXids(TransactionId xid,
3999 : int nxids, const TransactionId *xids,
4000 : TransactionId latestXid)
4001 : {
4002 : int i,
4003 : j;
4004 : XidCacheStatus *mysubxidstat;
4005 :
4006 : Assert(TransactionIdIsValid(xid));
4007 :
4008 : /*
4009 : * We must hold ProcArrayLock exclusively in order to remove transactions
4010 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
4011 : * possible this could be relaxed since we know this routine is only used
4012 : * to abort subtransactions, but pending closer analysis we'd best be
4013 : * conservative.
4014 : *
4015 : * Note that we do not have to be careful about memory ordering of our own
4016 : * reads wrt. GetNewTransactionId() here - only this process can modify
4017 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
4018 : * careful about our own writes being well ordered.
4019 : */
4020 860 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4021 :
4022 860 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4023 :
4024 : /*
4025 : * Under normal circumstances xid and xids[] will be in increasing order,
4026 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4027 : * behavior when removing a lot of xids.
4028 : */
4029 892 : for (i = nxids - 1; i >= 0; i--)
4030 : {
4031 32 : TransactionId anxid = xids[i];
4032 :
4033 32 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4034 : {
4035 32 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4036 : {
4037 32 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4038 32 : pg_write_barrier();
4039 32 : mysubxidstat->count--;
4040 32 : MyProc->subxidStatus.count--;
4041 32 : break;
4042 : }
4043 : }
4044 :
4045 : /*
4046 : * Ordinarily we should have found it, unless the cache has
4047 : * overflowed. However it's also possible for this routine to be
4048 : * invoked multiple times for the same subtransaction, in case of an
4049 : * error during AbortSubTransaction. So instead of Assert, emit a
4050 : * debug warning.
4051 : */
4052 32 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4053 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4054 : }
4055 :
4056 924 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4057 : {
4058 923 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4059 : {
4060 859 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4061 859 : pg_write_barrier();
4062 859 : mysubxidstat->count--;
4063 859 : MyProc->subxidStatus.count--;
4064 859 : break;
4065 : }
4066 : }
4067 : /* Ordinarily we should have found it, unless the cache has overflowed */
4068 860 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4069 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4070 :
4071 : /* Also advance global latestCompletedXid while holding the lock */
4072 860 : MaintainLatestCompletedXid(latestXid);
4073 :
4074 : /* ... and xactCompletionCount */
4075 860 : TransamVariables->xactCompletionCount++;
4076 :
4077 860 : LWLockRelease(ProcArrayLock);
4078 860 : }
4079 :
4080 : #ifdef XIDCACHE_DEBUG
4081 :
4082 : /*
4083 : * Print stats about effectiveness of XID cache
4084 : */
4085 : static void
4086 : DisplayXidCache(void)
4087 : {
4088 : fprintf(stderr,
4089 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4090 : xc_by_recent_xmin,
4091 : xc_by_known_xact,
4092 : xc_by_my_xact,
4093 : xc_by_latest_xid,
4094 : xc_by_main_xid,
4095 : xc_by_child_xid,
4096 : xc_by_known_assigned,
4097 : xc_no_overflow,
4098 : xc_slow_answer);
4099 : }
4100 : #endif /* XIDCACHE_DEBUG */
4101 :
4102 : /*
4103 : * If rel != NULL, return test state appropriate for relation, otherwise
4104 : * return state usable for all relations. The latter may consider XIDs as
4105 : * not-yet-visible-to-everyone that a state for a specific relation would
4106 : * already consider visible-to-everyone.
4107 : *
4108 : * This needs to be called while a snapshot is active or registered, otherwise
4109 : * there are wraparound and other dangers.
4110 : *
4111 : * See comment for GlobalVisState for details.
4112 : */
4113 : GlobalVisState *
4114 20427811 : GlobalVisTestFor(Relation rel)
4115 : {
4116 20427811 : GlobalVisState *state = NULL;
4117 :
4118 : /* XXX: we should assert that a snapshot is pushed or registered */
4119 : Assert(RecentXmin);
4120 :
4121 20427811 : switch (GlobalVisHorizonKindForRel(rel))
4122 : {
4123 144177 : case VISHORIZON_SHARED:
4124 144177 : state = &GlobalVisSharedRels;
4125 144177 : break;
4126 4781638 : case VISHORIZON_CATALOG:
4127 4781638 : state = &GlobalVisCatalogRels;
4128 4781638 : break;
4129 15394253 : case VISHORIZON_DATA:
4130 15394253 : state = &GlobalVisDataRels;
4131 15394253 : break;
4132 107743 : case VISHORIZON_TEMP:
4133 107743 : state = &GlobalVisTempRels;
4134 107743 : break;
4135 : }
4136 :
4137 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4138 : FullTransactionIdIsValid(state->maybe_needed));
4139 :
4140 20427811 : return state;
4141 : }
4142 :
4143 : /*
4144 : * Return true if it's worth updating the accurate maybe_needed boundary.
4145 : *
4146 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4147 : * repeatedly do so when there is a low likelihood of it being beneficial.
4148 : *
4149 : * The current heuristic is that we update only if RecentXmin has changed
4150 : * since the last update. If the oldest currently running transaction has not
4151 : * finished, it is unlikely that recomputing the horizon would be useful.
4152 : */
4153 : static bool
4154 732609 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4155 : {
4156 : /* hasn't been updated yet */
4157 732609 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4158 16666 : return true;
4159 :
4160 : /*
4161 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4162 : * unlikely to be beneficial to refresh boundaries.
4163 : */
4164 715943 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4165 : state->definitely_needed))
4166 0 : return false;
4167 :
4168 : /* does the last snapshot built have a different xmin? */
4169 715943 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4170 : }
4171 :
4172 : static void
4173 242499 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4174 : {
4175 : GlobalVisSharedRels.maybe_needed =
4176 242499 : FullXidRelativeTo(horizons->latest_completed,
4177 : horizons->shared_oldest_nonremovable);
4178 : GlobalVisCatalogRels.maybe_needed =
4179 242499 : FullXidRelativeTo(horizons->latest_completed,
4180 : horizons->catalog_oldest_nonremovable);
4181 : GlobalVisDataRels.maybe_needed =
4182 242499 : FullXidRelativeTo(horizons->latest_completed,
4183 : horizons->data_oldest_nonremovable);
4184 : GlobalVisTempRels.maybe_needed =
4185 242499 : FullXidRelativeTo(horizons->latest_completed,
4186 : horizons->temp_oldest_nonremovable);
4187 :
4188 : /*
4189 : * In longer running transactions it's possible that transactions we
4190 : * previously needed to treat as running aren't around anymore. So update
4191 : * definitely_needed to not be earlier than maybe_needed.
4192 : */
4193 : GlobalVisSharedRels.definitely_needed =
4194 242499 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4195 : GlobalVisSharedRels.definitely_needed);
4196 : GlobalVisCatalogRels.definitely_needed =
4197 242499 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4198 : GlobalVisCatalogRels.definitely_needed);
4199 : GlobalVisDataRels.definitely_needed =
4200 242499 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4201 : GlobalVisDataRels.definitely_needed);
4202 242499 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4203 :
4204 242499 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4205 242499 : }
4206 :
4207 : /*
4208 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4209 : * using ComputeXidHorizons().
4210 : */
4211 : static void
4212 85835 : GlobalVisUpdate(void)
4213 : {
4214 : ComputeXidHorizonsResult horizons;
4215 :
4216 : /* updates the horizons as a side-effect */
4217 85835 : ComputeXidHorizons(&horizons);
4218 85835 : }
4219 :
4220 : /*
4221 : * Return true if no snapshot still considers fxid to be running.
4222 : *
4223 : * The state passed needs to have been initialized for the relation fxid is
4224 : * from (NULL is also OK), otherwise the result may not be correct.
4225 : *
4226 : * If allow_update is false, the GlobalVisState boundaries will not be updated
4227 : * even if it would otherwise be beneficial. This is useful for callers that
4228 : * do not want GlobalVisState to advance at all, for example because they need
4229 : * a conservative answer based on the current boundaries.
4230 : *
4231 : * See comment for GlobalVisState for details.
4232 : */
4233 : bool
4234 16107533 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4235 : FullTransactionId fxid,
4236 : bool allow_update)
4237 : {
4238 : /*
4239 : * If fxid is older than maybe_needed bound, it definitely is visible to
4240 : * everyone.
4241 : */
4242 16107533 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4243 6236319 : return true;
4244 :
4245 : /*
4246 : * If fxid is >= definitely_needed bound, it is very likely to still be
4247 : * considered running.
4248 : */
4249 9871214 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4250 9138605 : return false;
4251 :
4252 : /*
4253 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4254 : * might not exist a snapshot considering fxid running. If it makes sense,
4255 : * update boundaries and recheck.
4256 : */
4257 732609 : if (allow_update && GlobalVisTestShouldUpdate(state))
4258 : {
4259 85835 : GlobalVisUpdate();
4260 :
4261 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4262 :
4263 85835 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4264 : }
4265 : else
4266 646774 : return false;
4267 : }
4268 :
4269 : /*
4270 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4271 : *
4272 : * It is crucial that this only gets called for xids from a source that
4273 : * protects against xid wraparounds (e.g. from a table and thus protected by
4274 : * relfrozenxid).
4275 : */
4276 : bool
4277 16107138 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid,
4278 : bool allow_update)
4279 : {
4280 : FullTransactionId fxid;
4281 :
4282 : /*
4283 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4284 : * because we know the xid has to, at the very least, be between
4285 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4286 : * lock to determine either, we can just compare with
4287 : * state->definitely_needed, which was based on those value at the time
4288 : * the current snapshot was built.
4289 : */
4290 16107138 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4291 :
4292 16107138 : return GlobalVisTestIsRemovableFullXid(state, fxid, allow_update);
4293 : }
4294 :
4295 : /*
4296 : * Wrapper around GlobalVisTestIsRemovableXid() for use when examining live
4297 : * tuples. Returns true if the given XID may be considered running by at least
4298 : * one snapshot.
4299 : *
4300 : * This function alone is insufficient to determine tuple visibility; callers
4301 : * must also consider the XID's commit status. Its purpose is purely semantic:
4302 : * when applied to live tuples, GlobalVisTestIsRemovableXid() is checking
4303 : * whether the inserting transaction is still considered running, not whether
4304 : * the tuple is removable. Live tuples are, by definition, not removable, but
4305 : * the snapshot criteria for "transaction still running" are identical to
4306 : * those used for removal XIDs.
4307 : *
4308 : * If allow_update is true, the GlobalVisState boundaries may be updated. If
4309 : * it is false, they definitely will not be updated.
4310 : *
4311 : * See the comment above GlobalVisTestIsRemovable[Full]Xid() for details on
4312 : * the required preconditions for calling this function.
4313 : */
4314 : bool
4315 78704 : GlobalVisTestXidConsideredRunning(GlobalVisState *state, TransactionId xid,
4316 : bool allow_update)
4317 : {
4318 78704 : return !GlobalVisTestIsRemovableXid(state, xid, allow_update);
4319 : }
4320 :
4321 : /*
4322 : * Convenience wrapper around GlobalVisTestFor() and
4323 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4324 : */
4325 : bool
4326 395 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4327 : {
4328 : GlobalVisState *state;
4329 :
4330 395 : state = GlobalVisTestFor(rel);
4331 :
4332 395 : return GlobalVisTestIsRemovableFullXid(state, fxid, true);
4333 : }
4334 :
4335 : /*
4336 : * Convenience wrapper around GlobalVisTestFor() and
4337 : * GlobalVisTestIsRemovableXid(), see their comments.
4338 : */
4339 : bool
4340 8 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4341 : {
4342 : GlobalVisState *state;
4343 :
4344 8 : state = GlobalVisTestFor(rel);
4345 :
4346 8 : return GlobalVisTestIsRemovableXid(state, xid, true);
4347 : }
4348 :
4349 : /*
4350 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4351 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4352 : *
4353 : * Be very careful about when to use this function. It can only safely be used
4354 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4355 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4356 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4357 : * the xid has to be within that range), or if xid is from the procarray and
4358 : * prevents xid wraparound that way.
4359 : */
4360 : static inline FullTransactionId
4361 18828268 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4362 : {
4363 18828268 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4364 :
4365 : Assert(TransactionIdIsValid(xid));
4366 : Assert(TransactionIdIsValid(rel_xid));
4367 :
4368 : /* not guaranteed to find issues, but likely to catch mistakes */
4369 : AssertTransactionIdInAllowableRange(xid);
4370 :
4371 37656536 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4372 18828268 : + (int32) (xid - rel_xid));
4373 : }
4374 :
4375 :
4376 : /* ----------------------------------------------
4377 : * KnownAssignedTransactionIds sub-module
4378 : * ----------------------------------------------
4379 : */
4380 :
4381 : /*
4382 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4383 : * running on the primary at the current point in WAL. These XIDs must be
4384 : * treated as running by standby transactions, even though they are not in
4385 : * the standby server's PGPROC array.
4386 : *
4387 : * We record all XIDs that we know have been assigned. That includes all the
4388 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4389 : * been assigned. We can deduce the existence of unobserved XIDs because we
4390 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4391 : * list expands as new XIDs are observed or inferred, and contracts when
4392 : * transaction completion records arrive.
4393 : *
4394 : * During hot standby we do not fret too much about the distinction between
4395 : * top-level XIDs and subtransaction XIDs. We store both together in the
4396 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4397 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4398 : * doesn't care about the distinction either. Subtransaction XIDs are
4399 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4400 : * links are *not* maintained (which does not affect visibility).
4401 : *
4402 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4403 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4404 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4405 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4406 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4407 : * and then remove them from KnownAssignedXids. This prevents overflow of
4408 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4409 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4410 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4411 : * though it should be complete for top-level XIDs; this is the same situation
4412 : * that holds with respect to the PGPROC entries in normal running.
4413 : *
4414 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4415 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4416 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4417 : * As long as that is within the range of interesting XIDs, we have to assume
4418 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4419 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4420 : * subXID arrives - that is not an error.)
4421 : *
4422 : * Should a backend on primary somehow disappear before it can write an abort
4423 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4424 : * aborted but we think they were running; the distinction is irrelevant
4425 : * because either way any changes done by the transaction are not visible to
4426 : * backends in the standby. We prune KnownAssignedXids when
4427 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4428 : * array due to such dead XIDs.
4429 : */
4430 :
4431 : /*
4432 : * RecordKnownAssignedTransactionIds
4433 : * Record the given XID in KnownAssignedXids, as well as any preceding
4434 : * unobserved XIDs.
4435 : *
4436 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4437 : * associated with a transaction. Must be called for each record after we
4438 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4439 : *
4440 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4441 : */
4442 : void
4443 2611892 : RecordKnownAssignedTransactionIds(TransactionId xid)
4444 : {
4445 : Assert(standbyState >= STANDBY_INITIALIZED);
4446 : Assert(TransactionIdIsValid(xid));
4447 : Assert(TransactionIdIsValid(latestObservedXid));
4448 :
4449 2611892 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4450 : xid, latestObservedXid);
4451 :
4452 : /*
4453 : * When a newly observed xid arrives, it is frequently the case that it is
4454 : * *not* the next xid in sequence. When this occurs, we must treat the
4455 : * intervening xids as running also.
4456 : */
4457 2611892 : if (TransactionIdFollows(xid, latestObservedXid))
4458 : {
4459 : TransactionId next_expected_xid;
4460 :
4461 : /*
4462 : * Extend subtrans like we do in GetNewTransactionId() during normal
4463 : * operation using individual extend steps. Note that we do not need
4464 : * to extend clog since its extensions are WAL logged.
4465 : *
4466 : * This part has to be done regardless of standbyState since we
4467 : * immediately start assigning subtransactions to their toplevel
4468 : * transactions.
4469 : */
4470 24585 : next_expected_xid = latestObservedXid;
4471 49943 : while (TransactionIdPrecedes(next_expected_xid, xid))
4472 : {
4473 25358 : TransactionIdAdvance(next_expected_xid);
4474 25358 : ExtendSUBTRANS(next_expected_xid);
4475 : }
4476 : Assert(next_expected_xid == xid);
4477 :
4478 : /*
4479 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4480 : * more to do since we don't track assigned xids yet.
4481 : */
4482 24585 : if (standbyState <= STANDBY_INITIALIZED)
4483 : {
4484 0 : latestObservedXid = xid;
4485 0 : return;
4486 : }
4487 :
4488 : /*
4489 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4490 : */
4491 24585 : next_expected_xid = latestObservedXid;
4492 24585 : TransactionIdAdvance(next_expected_xid);
4493 24585 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4494 :
4495 : /*
4496 : * Now we can advance latestObservedXid
4497 : */
4498 24585 : latestObservedXid = xid;
4499 :
4500 : /* TransamVariables->nextXid must be beyond any observed xid */
4501 24585 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4502 : }
4503 : }
4504 :
4505 : /*
4506 : * ExpireTreeKnownAssignedTransactionIds
4507 : * Remove the given XIDs from KnownAssignedXids.
4508 : *
4509 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4510 : */
4511 : void
4512 23829 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4513 : TransactionId *subxids, TransactionId max_xid)
4514 : {
4515 : Assert(standbyState >= STANDBY_INITIALIZED);
4516 :
4517 : /*
4518 : * Uses same locking as transaction commit
4519 : */
4520 23829 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4521 :
4522 23829 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4523 :
4524 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4525 23829 : MaintainLatestCompletedXidRecovery(max_xid);
4526 :
4527 : /* ... and xactCompletionCount */
4528 23829 : TransamVariables->xactCompletionCount++;
4529 :
4530 23829 : LWLockRelease(ProcArrayLock);
4531 23829 : }
4532 :
4533 : /*
4534 : * ExpireAllKnownAssignedTransactionIds
4535 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4536 : */
4537 : void
4538 123 : ExpireAllKnownAssignedTransactionIds(void)
4539 : {
4540 : FullTransactionId latestXid;
4541 :
4542 123 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4543 123 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4544 :
4545 : /* Reset latestCompletedXid to nextXid - 1 */
4546 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
4547 123 : latestXid = TransamVariables->nextXid;
4548 123 : FullTransactionIdRetreat(&latestXid);
4549 123 : TransamVariables->latestCompletedXid = latestXid;
4550 :
4551 : /*
4552 : * Any transactions that were in-progress were effectively aborted, so
4553 : * advance xactCompletionCount.
4554 : */
4555 123 : TransamVariables->xactCompletionCount++;
4556 :
4557 : /*
4558 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4559 : * the call of this function. But do this for unification with what
4560 : * ExpireOldKnownAssignedTransactionIds() do.
4561 : */
4562 123 : procArray->lastOverflowedXid = InvalidTransactionId;
4563 123 : LWLockRelease(ProcArrayLock);
4564 123 : }
4565 :
4566 : /*
4567 : * ExpireOldKnownAssignedTransactionIds
4568 : * Remove KnownAssignedXids entries preceding the given XID and
4569 : * potentially reset lastOverflowedXid.
4570 : */
4571 : void
4572 852 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4573 : {
4574 : TransactionId latestXid;
4575 :
4576 852 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4577 :
4578 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4579 852 : latestXid = xid;
4580 852 : TransactionIdRetreat(latestXid);
4581 852 : MaintainLatestCompletedXidRecovery(latestXid);
4582 :
4583 : /* ... and xactCompletionCount */
4584 852 : TransamVariables->xactCompletionCount++;
4585 :
4586 : /*
4587 : * Reset lastOverflowedXid if we know all transactions that have been
4588 : * possibly running are being gone. Not doing so could cause an incorrect
4589 : * lastOverflowedXid value, which makes extra snapshots be marked as
4590 : * suboverflowed.
4591 : */
4592 852 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4593 846 : procArray->lastOverflowedXid = InvalidTransactionId;
4594 852 : KnownAssignedXidsRemovePreceding(xid);
4595 852 : LWLockRelease(ProcArrayLock);
4596 852 : }
4597 :
4598 : /*
4599 : * KnownAssignedTransactionIdsIdleMaintenance
4600 : * Opportunistically do maintenance work when the startup process
4601 : * is about to go idle.
4602 : */
4603 : void
4604 15720 : KnownAssignedTransactionIdsIdleMaintenance(void)
4605 : {
4606 15720 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4607 15720 : }
4608 :
4609 :
4610 : /*
4611 : * Private module functions to manipulate KnownAssignedXids
4612 : *
4613 : * There are 5 main uses of the KnownAssignedXids data structure:
4614 : *
4615 : * * backends taking snapshots - all valid XIDs need to be copied out
4616 : * * backends seeking to determine presence of a specific XID
4617 : * * startup process adding new known-assigned XIDs
4618 : * * startup process removing specific XIDs as transactions end
4619 : * * startup process pruning array when special WAL records arrive
4620 : *
4621 : * This data structure is known to be a hot spot during Hot Standby, so we
4622 : * go to some lengths to make these operations as efficient and as concurrent
4623 : * as possible.
4624 : *
4625 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4626 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4627 : * in general TransactionIdPrecedes would not provide a total order, but
4628 : * we know that the entries present at any instant should not extend across
4629 : * a large enough fraction of XID space to wrap around (the primary would
4630 : * shut down for fear of XID wrap long before that happens). So it's OK to
4631 : * use TransactionIdPrecedes as a binary-search comparator.
4632 : *
4633 : * It's cheap to maintain the sortedness during insertions, since new known
4634 : * XIDs are always reported in XID order; we just append them at the right.
4635 : *
4636 : * To keep individual deletions cheap, we need to allow gaps in the array.
4637 : * This is implemented by marking array elements as valid or invalid using
4638 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4639 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4640 : * XID entry itself. This preserves the property that the XID entries are
4641 : * sorted, so we can do binary searches easily. Periodically we compress
4642 : * out the unused entries; that's much cheaper than having to compress the
4643 : * array immediately on every deletion.
4644 : *
4645 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4646 : * are those with indexes tail <= i < head; items outside this subscript range
4647 : * have unspecified contents. When head reaches the end of the array, we
4648 : * force compression of unused entries rather than wrapping around, since
4649 : * allowing wraparound would greatly complicate the search logic. We maintain
4650 : * an explicit tail pointer so that pruning of old XIDs can be done without
4651 : * immediately moving the array contents. In most cases only a small fraction
4652 : * of the array contains valid entries at any instant.
4653 : *
4654 : * Although only the startup process can ever change the KnownAssignedXids
4655 : * data structure, we still need interlocking so that standby backends will
4656 : * not observe invalid intermediate states. The convention is that backends
4657 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4658 : * the array, the startup process must hold ProcArrayLock exclusively, for
4659 : * the usual transactional reasons (compare commit/abort of a transaction
4660 : * during normal running). Compressing unused entries out of the array
4661 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4662 : * them into slots to the right of the head pointer and then advance the head
4663 : * pointer. This doesn't require any lock at all, but on machines with weak
4664 : * memory ordering, we need to be careful that other processors see the array
4665 : * element changes before they see the head pointer change. We handle this by
4666 : * using memory barriers when reading or writing the head/tail pointers (unless
4667 : * the caller holds ProcArrayLock exclusively).
4668 : *
4669 : * Algorithmic analysis:
4670 : *
4671 : * If we have a maximum of M slots, with N XIDs currently spread across
4672 : * S elements then we have N <= S <= M always.
4673 : *
4674 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4675 : * happen)
4676 : * * Compressing the array is O(S) and requires exclusive lock
4677 : * * Removing an XID is O(logS) and requires exclusive lock
4678 : * * Taking a snapshot is O(S) and requires shared lock
4679 : * * Checking for an XID is O(logS) and requires shared lock
4680 : *
4681 : * In comparison, using a hash table for KnownAssignedXids would mean that
4682 : * taking snapshots would be O(M). If we can maintain S << M then the
4683 : * sorted array technique will deliver significantly faster snapshots.
4684 : * If we try to keep S too small then we will spend too much time compressing,
4685 : * so there is an optimal point for any workload mix. We use a heuristic to
4686 : * decide when to compress the array, though trimming also helps reduce
4687 : * frequency of compressing. The heuristic requires us to track the number of
4688 : * currently valid XIDs in the array (N). Except in special cases, we'll
4689 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4690 : * taking a snapshot to be O(N), which it would have to be anyway.
4691 : */
4692 :
4693 :
4694 : /*
4695 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4696 : * array, removing any gaps.
4697 : *
4698 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4699 : * we do it only if a heuristic indicates it's a good time to do it.
4700 : *
4701 : * Compression requires holding ProcArrayLock in exclusive mode.
4702 : * Caller must pass haveLock = true if it already holds the lock.
4703 : */
4704 : static void
4705 40422 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4706 : {
4707 40422 : ProcArrayStruct *pArray = procArray;
4708 : int head,
4709 : tail,
4710 : nelements;
4711 : int compress_index;
4712 : int i;
4713 :
4714 : /* Counters for compression heuristics */
4715 : static unsigned int transactionEndsCounter;
4716 : static TimestampTz lastCompressTs;
4717 :
4718 : /* Tuning constants */
4719 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4720 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4721 :
4722 : /*
4723 : * Since only the startup process modifies the head/tail pointers, we
4724 : * don't need a lock to read them here.
4725 : */
4726 40422 : head = pArray->headKnownAssignedXids;
4727 40422 : tail = pArray->tailKnownAssignedXids;
4728 40422 : nelements = head - tail;
4729 :
4730 : /*
4731 : * If we can choose whether to compress, use a heuristic to avoid
4732 : * compressing too often or not often enough. "Compress" here simply
4733 : * means moving the values to the beginning of the array, so it is not as
4734 : * complex or costly as typical data compression algorithms.
4735 : */
4736 40422 : if (nelements == pArray->numKnownAssignedXids)
4737 : {
4738 : /*
4739 : * When there are no gaps between head and tail, don't bother to
4740 : * compress, except in the KAX_NO_SPACE case where we must compress to
4741 : * create some space after the head.
4742 : */
4743 22213 : if (reason != KAX_NO_SPACE)
4744 22213 : return;
4745 : }
4746 18209 : else if (reason == KAX_TRANSACTION_END)
4747 : {
4748 : /*
4749 : * Consider compressing only once every so many commits. Frequency
4750 : * determined by benchmarks.
4751 : */
4752 13246 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4753 13133 : return;
4754 :
4755 : /*
4756 : * Furthermore, compress only if the used part of the array is less
4757 : * than 50% full (see comments above).
4758 : */
4759 113 : if (nelements < 2 * pArray->numKnownAssignedXids)
4760 8 : return;
4761 : }
4762 4963 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4763 : {
4764 : /*
4765 : * We're about to go idle for lack of new WAL, so we might as well
4766 : * compress. But not too often, to avoid ProcArray lock contention
4767 : * with readers.
4768 : */
4769 4841 : if (lastCompressTs != 0)
4770 : {
4771 : TimestampTz compress_after;
4772 :
4773 4840 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4774 : KAX_COMPRESS_IDLE_INTERVAL);
4775 4840 : if (GetCurrentTimestamp() < compress_after)
4776 4811 : return;
4777 : }
4778 : }
4779 :
4780 : /* Need to compress, so get the lock if we don't have it. */
4781 257 : if (!haveLock)
4782 30 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4783 :
4784 : /*
4785 : * We compress the array by reading the valid values from tail to head,
4786 : * re-aligning data to 0th element.
4787 : */
4788 257 : compress_index = 0;
4789 8790 : for (i = tail; i < head; i++)
4790 : {
4791 8533 : if (KnownAssignedXidsValid[i])
4792 : {
4793 927 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4794 927 : KnownAssignedXidsValid[compress_index] = true;
4795 927 : compress_index++;
4796 : }
4797 : }
4798 : Assert(compress_index == pArray->numKnownAssignedXids);
4799 :
4800 257 : pArray->tailKnownAssignedXids = 0;
4801 257 : pArray->headKnownAssignedXids = compress_index;
4802 :
4803 257 : if (!haveLock)
4804 30 : LWLockRelease(ProcArrayLock);
4805 :
4806 : /* Update timestamp for maintenance. No need to hold lock for this. */
4807 257 : lastCompressTs = GetCurrentTimestamp();
4808 : }
4809 :
4810 : /*
4811 : * Add xids into KnownAssignedXids at the head of the array.
4812 : *
4813 : * xids from from_xid to to_xid, inclusive, are added to the array.
4814 : *
4815 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4816 : * exclusive mode, so we need no extra locking here. Else caller holds no
4817 : * lock, so we need to be sure we maintain sufficient interlocks against
4818 : * concurrent readers. (Only the startup process ever calls this, so no need
4819 : * to worry about concurrent writers.)
4820 : */
4821 : static void
4822 24590 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4823 : bool exclusive_lock)
4824 : {
4825 24590 : ProcArrayStruct *pArray = procArray;
4826 : TransactionId next_xid;
4827 : int head,
4828 : tail;
4829 : int nxids;
4830 : int i;
4831 :
4832 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4833 :
4834 : /*
4835 : * Calculate how many array slots we'll need. Normally this is cheap; in
4836 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4837 : * way.
4838 : */
4839 24590 : if (to_xid >= from_xid)
4840 24590 : nxids = to_xid - from_xid + 1;
4841 : else
4842 : {
4843 0 : nxids = 1;
4844 0 : next_xid = from_xid;
4845 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4846 : {
4847 0 : nxids++;
4848 0 : TransactionIdAdvance(next_xid);
4849 : }
4850 : }
4851 :
4852 : /*
4853 : * Since only the startup process modifies the head/tail pointers, we
4854 : * don't need a lock to read them here.
4855 : */
4856 24590 : head = pArray->headKnownAssignedXids;
4857 24590 : tail = pArray->tailKnownAssignedXids;
4858 :
4859 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4860 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4861 :
4862 : /*
4863 : * Verify that insertions occur in TransactionId sequence. Note that even
4864 : * if the last existing element is marked invalid, it must still have a
4865 : * correctly sequenced XID value.
4866 : */
4867 39443 : if (head > tail &&
4868 14853 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4869 : {
4870 0 : KnownAssignedXidsDisplay(LOG);
4871 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4872 : }
4873 :
4874 : /*
4875 : * If our xids won't fit in the remaining space, compress out free space
4876 : */
4877 24590 : if (head + nxids > pArray->maxKnownAssignedXids)
4878 : {
4879 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4880 :
4881 0 : head = pArray->headKnownAssignedXids;
4882 : /* note: we no longer care about the tail pointer */
4883 :
4884 : /*
4885 : * If it still won't fit then we're out of memory
4886 : */
4887 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4888 0 : elog(ERROR, "too many KnownAssignedXids");
4889 : }
4890 :
4891 : /* Now we can insert the xids into the space starting at head */
4892 24590 : next_xid = from_xid;
4893 49953 : for (i = 0; i < nxids; i++)
4894 : {
4895 25363 : KnownAssignedXids[head] = next_xid;
4896 25363 : KnownAssignedXidsValid[head] = true;
4897 25363 : TransactionIdAdvance(next_xid);
4898 25363 : head++;
4899 : }
4900 :
4901 : /* Adjust count of number of valid entries */
4902 24590 : pArray->numKnownAssignedXids += nxids;
4903 :
4904 : /*
4905 : * Now update the head pointer. We use a write barrier to ensure that
4906 : * other processors see the above array updates before they see the head
4907 : * pointer change. The barrier isn't required if we're holding
4908 : * ProcArrayLock exclusively.
4909 : */
4910 24590 : if (!exclusive_lock)
4911 24585 : pg_write_barrier();
4912 :
4913 24590 : pArray->headKnownAssignedXids = head;
4914 24590 : }
4915 :
4916 : /*
4917 : * KnownAssignedXidsSearch
4918 : *
4919 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4920 : * Returns true if it was found, false if not.
4921 : *
4922 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4923 : * Exclusive lock must be held for remove = true.
4924 : */
4925 : static bool
4926 26499 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4927 : {
4928 26499 : ProcArrayStruct *pArray = procArray;
4929 : int first,
4930 : last;
4931 : int head;
4932 : int tail;
4933 26499 : int result_index = -1;
4934 :
4935 26499 : tail = pArray->tailKnownAssignedXids;
4936 26499 : head = pArray->headKnownAssignedXids;
4937 :
4938 : /*
4939 : * Only the startup process removes entries, so we don't need the read
4940 : * barrier in that case.
4941 : */
4942 26499 : if (!remove)
4943 1 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4944 :
4945 : /*
4946 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4947 : * array here, since even invalid entries will contain sorted XIDs.
4948 : */
4949 26499 : first = tail;
4950 26499 : last = head - 1;
4951 83131 : while (first <= last)
4952 : {
4953 : int mid_index;
4954 : TransactionId mid_xid;
4955 :
4956 81922 : mid_index = (first + last) / 2;
4957 81922 : mid_xid = KnownAssignedXids[mid_index];
4958 :
4959 81922 : if (xid == mid_xid)
4960 : {
4961 25290 : result_index = mid_index;
4962 25290 : break;
4963 : }
4964 56632 : else if (TransactionIdPrecedes(xid, mid_xid))
4965 12639 : last = mid_index - 1;
4966 : else
4967 43993 : first = mid_index + 1;
4968 : }
4969 :
4970 26499 : if (result_index < 0)
4971 1209 : return false; /* not in array */
4972 :
4973 25290 : if (!KnownAssignedXidsValid[result_index])
4974 13 : return false; /* in array, but invalid */
4975 :
4976 25277 : if (remove)
4977 : {
4978 25277 : KnownAssignedXidsValid[result_index] = false;
4979 :
4980 25277 : pArray->numKnownAssignedXids--;
4981 : Assert(pArray->numKnownAssignedXids >= 0);
4982 :
4983 : /*
4984 : * If we're removing the tail element then advance tail pointer over
4985 : * any invalid elements. This will speed future searches.
4986 : */
4987 25277 : if (result_index == tail)
4988 : {
4989 11109 : tail++;
4990 17671 : while (tail < head && !KnownAssignedXidsValid[tail])
4991 6562 : tail++;
4992 11109 : if (tail >= head)
4993 : {
4994 : /* Array is empty, so we can reset both pointers */
4995 9727 : pArray->headKnownAssignedXids = 0;
4996 9727 : pArray->tailKnownAssignedXids = 0;
4997 : }
4998 : else
4999 : {
5000 1382 : pArray->tailKnownAssignedXids = tail;
5001 : }
5002 : }
5003 : }
5004 :
5005 25277 : return true;
5006 : }
5007 :
5008 : /*
5009 : * Is the specified XID present in KnownAssignedXids[]?
5010 : *
5011 : * Caller must hold ProcArrayLock in shared or exclusive mode.
5012 : */
5013 : static bool
5014 1 : KnownAssignedXidExists(TransactionId xid)
5015 : {
5016 : Assert(TransactionIdIsValid(xid));
5017 :
5018 1 : return KnownAssignedXidsSearch(xid, false);
5019 : }
5020 :
5021 : /*
5022 : * Remove the specified XID from KnownAssignedXids[].
5023 : *
5024 : * Caller must hold ProcArrayLock in exclusive mode.
5025 : */
5026 : static void
5027 26498 : KnownAssignedXidsRemove(TransactionId xid)
5028 : {
5029 : Assert(TransactionIdIsValid(xid));
5030 :
5031 26498 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
5032 :
5033 : /*
5034 : * Note: we cannot consider it an error to remove an XID that's not
5035 : * present. We intentionally remove subxact IDs while processing
5036 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
5037 : * removed again when the top-level xact commits or aborts.
5038 : *
5039 : * It might be possible to track such XIDs to distinguish this case from
5040 : * actual errors, but it would be complicated and probably not worth it.
5041 : * So, just ignore the search result.
5042 : */
5043 26498 : (void) KnownAssignedXidsSearch(xid, true);
5044 26498 : }
5045 :
5046 : /*
5047 : * KnownAssignedXidsRemoveTree
5048 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
5049 : *
5050 : * Caller must hold ProcArrayLock in exclusive mode.
5051 : */
5052 : static void
5053 23850 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
5054 : TransactionId *subxids)
5055 : {
5056 : int i;
5057 :
5058 23850 : if (TransactionIdIsValid(xid))
5059 23829 : KnownAssignedXidsRemove(xid);
5060 :
5061 26519 : for (i = 0; i < nsubxids; i++)
5062 2669 : KnownAssignedXidsRemove(subxids[i]);
5063 :
5064 : /* Opportunistically compress the array */
5065 23850 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5066 23850 : }
5067 :
5068 : /*
5069 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5070 : * then clear the whole table.
5071 : *
5072 : * Caller must hold ProcArrayLock in exclusive mode.
5073 : */
5074 : static void
5075 975 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5076 : {
5077 975 : ProcArrayStruct *pArray = procArray;
5078 975 : int count = 0;
5079 : int head,
5080 : tail,
5081 : i;
5082 :
5083 975 : if (!TransactionIdIsValid(removeXid))
5084 : {
5085 123 : elog(DEBUG4, "removing all KnownAssignedXids");
5086 123 : pArray->numKnownAssignedXids = 0;
5087 123 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5088 123 : return;
5089 : }
5090 :
5091 852 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5092 :
5093 : /*
5094 : * Mark entries invalid starting at the tail. Since array is sorted, we
5095 : * can stop as soon as we reach an entry >= removeXid.
5096 : */
5097 852 : tail = pArray->tailKnownAssignedXids;
5098 852 : head = pArray->headKnownAssignedXids;
5099 :
5100 852 : for (i = tail; i < head; i++)
5101 : {
5102 216 : if (KnownAssignedXidsValid[i])
5103 : {
5104 216 : TransactionId knownXid = KnownAssignedXids[i];
5105 :
5106 216 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5107 216 : break;
5108 :
5109 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5110 : {
5111 0 : KnownAssignedXidsValid[i] = false;
5112 0 : count++;
5113 : }
5114 : }
5115 : }
5116 :
5117 852 : pArray->numKnownAssignedXids -= count;
5118 : Assert(pArray->numKnownAssignedXids >= 0);
5119 :
5120 : /*
5121 : * Advance the tail pointer if we've marked the tail item invalid.
5122 : */
5123 852 : for (i = tail; i < head; i++)
5124 : {
5125 216 : if (KnownAssignedXidsValid[i])
5126 216 : break;
5127 : }
5128 852 : if (i >= head)
5129 : {
5130 : /* Array is empty, so we can reset both pointers */
5131 636 : pArray->headKnownAssignedXids = 0;
5132 636 : pArray->tailKnownAssignedXids = 0;
5133 : }
5134 : else
5135 : {
5136 216 : pArray->tailKnownAssignedXids = i;
5137 : }
5138 :
5139 : /* Opportunistically compress the array */
5140 852 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5141 : }
5142 :
5143 : /*
5144 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5145 : * We filter out anything >= xmax.
5146 : *
5147 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5148 : * that array is large enough.
5149 : *
5150 : * Caller must hold ProcArrayLock in (at least) shared mode.
5151 : */
5152 : static int
5153 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5154 : {
5155 0 : TransactionId xtmp = InvalidTransactionId;
5156 :
5157 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5158 : }
5159 :
5160 : /*
5161 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5162 : * we reduce *xmin to the lowest xid value seen if not already lower.
5163 : *
5164 : * Caller must hold ProcArrayLock in (at least) shared mode.
5165 : */
5166 : static int
5167 1601 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5168 : TransactionId xmax)
5169 : {
5170 1601 : int count = 0;
5171 : int head,
5172 : tail;
5173 : int i;
5174 :
5175 : /*
5176 : * Fetch head just once, since it may change while we loop. We can stop
5177 : * once we reach the initially seen head, since we are certain that an xid
5178 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5179 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5180 : * anyway.
5181 : */
5182 1601 : tail = procArray->tailKnownAssignedXids;
5183 1601 : head = procArray->headKnownAssignedXids;
5184 :
5185 1601 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5186 :
5187 1633 : for (i = tail; i < head; i++)
5188 : {
5189 : /* Skip any gaps in the array */
5190 135 : if (KnownAssignedXidsValid[i])
5191 : {
5192 122 : TransactionId knownXid = KnownAssignedXids[i];
5193 :
5194 : /*
5195 : * Update xmin if required. Only the first XID need be checked,
5196 : * since the array is sorted.
5197 : */
5198 244 : if (count == 0 &&
5199 122 : TransactionIdPrecedes(knownXid, *xmin))
5200 19 : *xmin = knownXid;
5201 :
5202 : /*
5203 : * Filter out anything >= xmax, again relying on sorted property
5204 : * of array.
5205 : */
5206 244 : if (TransactionIdIsValid(xmax) &&
5207 122 : TransactionIdFollowsOrEquals(knownXid, xmax))
5208 103 : break;
5209 :
5210 : /* Add knownXid into output array */
5211 19 : xarray[count++] = knownXid;
5212 : }
5213 : }
5214 :
5215 1601 : return count;
5216 : }
5217 :
5218 : /*
5219 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5220 : * if nothing there.
5221 : */
5222 : static TransactionId
5223 396 : KnownAssignedXidsGetOldestXmin(void)
5224 : {
5225 : int head,
5226 : tail;
5227 : int i;
5228 :
5229 : /*
5230 : * Fetch head just once, since it may change while we loop.
5231 : */
5232 396 : tail = procArray->tailKnownAssignedXids;
5233 396 : head = procArray->headKnownAssignedXids;
5234 :
5235 396 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5236 :
5237 396 : for (i = tail; i < head; i++)
5238 : {
5239 : /* Skip any gaps in the array */
5240 148 : if (KnownAssignedXidsValid[i])
5241 148 : return KnownAssignedXids[i];
5242 : }
5243 :
5244 248 : return InvalidTransactionId;
5245 : }
5246 :
5247 : /*
5248 : * Display KnownAssignedXids to provide debug trail
5249 : *
5250 : * Currently this is only called within startup process, so we need no
5251 : * special locking.
5252 : *
5253 : * Note this is pretty expensive, and much of the expense will be incurred
5254 : * even if the elog message will get discarded. It's not currently called
5255 : * in any performance-critical places, however, so no need to be tenser.
5256 : */
5257 : static void
5258 128 : KnownAssignedXidsDisplay(int trace_level)
5259 : {
5260 128 : ProcArrayStruct *pArray = procArray;
5261 : StringInfoData buf;
5262 : int head,
5263 : tail,
5264 : i;
5265 128 : int nxids = 0;
5266 :
5267 128 : tail = pArray->tailKnownAssignedXids;
5268 128 : head = pArray->headKnownAssignedXids;
5269 :
5270 128 : initStringInfo(&buf);
5271 :
5272 138 : for (i = tail; i < head; i++)
5273 : {
5274 10 : if (KnownAssignedXidsValid[i])
5275 : {
5276 10 : nxids++;
5277 10 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5278 : }
5279 : }
5280 :
5281 128 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5282 : nxids,
5283 : pArray->numKnownAssignedXids,
5284 : pArray->tailKnownAssignedXids,
5285 : pArray->headKnownAssignedXids,
5286 : buf.data);
5287 :
5288 128 : pfree(buf.data);
5289 128 : }
5290 :
5291 : /*
5292 : * KnownAssignedXidsReset
5293 : * Resets KnownAssignedXids to be empty
5294 : */
5295 : static void
5296 0 : KnownAssignedXidsReset(void)
5297 : {
5298 0 : ProcArrayStruct *pArray = procArray;
5299 :
5300 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5301 :
5302 0 : pArray->numKnownAssignedXids = 0;
5303 0 : pArray->tailKnownAssignedXids = 0;
5304 0 : pArray->headKnownAssignedXids = 0;
5305 :
5306 0 : LWLockRelease(ProcArrayLock);
5307 0 : }
|
