Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procarray.c
4 : * POSTGRES process array code.
5 : *
6 : *
7 : * This module maintains arrays of PGPROC substructures, as well as associated
8 : * arrays in ProcGlobal, for all active backends. Although there are several
9 : * uses for this, the principal one is as a means of determining the set of
10 : * currently running transactions.
11 : *
12 : * Because of various subtle race conditions it is critical that a backend
13 : * hold the correct locks while setting or clearing its xid (in
14 : * ProcGlobal->xids[]/MyProc->xid). See notes in
15 : * src/backend/access/transam/README.
16 : *
17 : * The process arrays now also include structures representing prepared
18 : * transactions. The xid and subxids fields of these are valid, as are the
19 : * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 : * at need by checking for pid == 0.
21 : *
22 : * During hot standby, we also keep a list of XIDs representing transactions
23 : * that are known to be running on the primary (or more precisely, were running
24 : * as of the current point in the WAL stream). This list is kept in the
25 : * KnownAssignedXids array, and is updated by watching the sequence of
26 : * arriving XIDs. This is necessary because if we leave those XIDs out of
27 : * snapshots taken for standby queries, then they will appear to be already
28 : * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 : * array represents standby processes, which by definition are not running
30 : * transactions that have XIDs.
31 : *
32 : * It is perhaps possible for a backend on the primary to terminate without
33 : * writing an abort record for its transaction. While that shouldn't really
34 : * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 : * ourselves by pruning the array when a valid list of running XIDs arrives.
36 : *
37 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/subtrans.h"
51 : #include "access/transam.h"
52 : #include "access/twophase.h"
53 : #include "access/xact.h"
54 : #include "access/xlogutils.h"
55 : #include "catalog/catalog.h"
56 : #include "catalog/pg_authid.h"
57 : #include "commands/dbcommands.h"
58 : #include "miscadmin.h"
59 : #include "pgstat.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "utils/acl.h"
64 : #include "utils/builtins.h"
65 : #include "utils/rel.h"
66 : #include "utils/snapmgr.h"
67 :
68 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
69 :
70 : /* Our shared memory area */
71 : typedef struct ProcArrayStruct
72 : {
73 : int numProcs; /* number of valid procs entries */
74 : int maxProcs; /* allocated size of procs array */
75 :
76 : /*
77 : * Known assigned XIDs handling
78 : */
79 : int maxKnownAssignedXids; /* allocated size of array */
80 : int numKnownAssignedXids; /* current # of valid entries */
81 : int tailKnownAssignedXids; /* index of oldest valid element */
82 : int headKnownAssignedXids; /* index of newest element, + 1 */
83 :
84 : /*
85 : * Highest subxid that has been removed from KnownAssignedXids array to
86 : * prevent overflow; or InvalidTransactionId if none. We track this for
87 : * similar reasons to tracking overflowing cached subxids in PGPROC
88 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
89 : * lock to read it.
90 : */
91 : TransactionId lastOverflowedXid;
92 :
93 : /* oldest xmin of any replication slot */
94 : TransactionId replication_slot_xmin;
95 : /* oldest catalog xmin of any replication slot */
96 : TransactionId replication_slot_catalog_xmin;
97 :
98 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
99 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
100 : } ProcArrayStruct;
101 :
102 : /*
103 : * State for the GlobalVisTest* family of functions. Those functions can
104 : * e.g. be used to decide if a deleted row can be removed without violating
105 : * MVCC semantics: If the deleted row's xmax is not considered to be running
106 : * by anyone, the row can be removed.
107 : *
108 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
109 : * cutoff XID while building a snapshot (looking at the frequently changing
110 : * xmins scales badly). Instead we compute two boundaries while building the
111 : * snapshot:
112 : *
113 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
114 : * definitely_needed are definitely still visible.
115 : *
116 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
117 : * definitely be removed
118 : *
119 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
120 : * && XID < definitely_needed), the boundaries can be recomputed (using
121 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
122 : * maintaining an accurate value all the time.
123 : *
124 : * As it is not cheap to compute accurate boundaries, we limit the number of
125 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
126 : *
127 : *
128 : * There are three backend lifetime instances of this struct, optimized for
129 : * different types of relations. As e.g. a normal user defined table in one
130 : * database is inaccessible to backends connected to another database, a test
131 : * specific to a relation can be more aggressive than a test for a shared
132 : * relation. Currently we track four different states:
133 : *
134 : * 1) GlobalVisSharedRels, which only considers an XID's
135 : * effects visible-to-everyone if neither snapshots in any database, nor a
136 : * replication slot's xmin, nor a replication slot's catalog_xmin might
137 : * still consider XID as running.
138 : *
139 : * 2) GlobalVisCatalogRels, which only considers an XID's
140 : * effects visible-to-everyone if neither snapshots in the current
141 : * database, nor a replication slot's xmin, nor a replication slot's
142 : * catalog_xmin might still consider XID as running.
143 : *
144 : * I.e. the difference to GlobalVisSharedRels is that
145 : * snapshot in other databases are ignored.
146 : *
147 : * 3) GlobalVisDataRels, which only considers an XID's
148 : * effects visible-to-everyone if neither snapshots in the current
149 : * database, nor a replication slot's xmin consider XID as running.
150 : *
151 : * I.e. the difference to GlobalVisCatalogRels is that
152 : * replication slot's catalog_xmin is not taken into account.
153 : *
154 : * 4) GlobalVisTempRels, which only considers the current session, as temp
155 : * tables are not visible to other sessions.
156 : *
157 : * GlobalVisTestFor(relation) returns the appropriate state
158 : * for the relation.
159 : *
160 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
161 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
162 : * prevent maybe_needed to become old enough after the GetSnapshotData()
163 : * call.
164 : *
165 : * The typedef is in the header.
166 : */
167 : struct GlobalVisState
168 : {
169 : /* XIDs >= are considered running by some backend */
170 : FullTransactionId definitely_needed;
171 :
172 : /* XIDs < are not considered to be running by any backend */
173 : FullTransactionId maybe_needed;
174 : };
175 :
176 : /*
177 : * Result of ComputeXidHorizons().
178 : */
179 : typedef struct ComputeXidHorizonsResult
180 : {
181 : /*
182 : * The value of TransamVariables->latestCompletedXid when
183 : * ComputeXidHorizons() held ProcArrayLock.
184 : */
185 : FullTransactionId latest_completed;
186 :
187 : /*
188 : * The same for procArray->replication_slot_xmin and
189 : * procArray->replication_slot_catalog_xmin.
190 : */
191 : TransactionId slot_xmin;
192 : TransactionId slot_catalog_xmin;
193 :
194 : /*
195 : * Oldest xid that any backend might still consider running. This needs to
196 : * include processes running VACUUM, in contrast to the normal visibility
197 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
198 : * determining visibility, but doesn't care about rows above its xmin to
199 : * be removed.
200 : *
201 : * This likely should only be needed to determine whether pg_subtrans can
202 : * be truncated. It currently includes the effects of replication slots,
203 : * for historical reasons. But that could likely be changed.
204 : */
205 : TransactionId oldest_considered_running;
206 :
207 : /*
208 : * Oldest xid for which deleted tuples need to be retained in shared
209 : * tables.
210 : *
211 : * This includes the effects of replication slots. If that's not desired,
212 : * look at shared_oldest_nonremovable_raw;
213 : */
214 : TransactionId shared_oldest_nonremovable;
215 :
216 : /*
217 : * Oldest xid that may be necessary to retain in shared tables. This is
218 : * the same as shared_oldest_nonremovable, except that is not affected by
219 : * replication slot's catalog_xmin.
220 : *
221 : * This is mainly useful to be able to send the catalog_xmin to upstream
222 : * streaming replication servers via hot_standby_feedback, so they can
223 : * apply the limit only when accessing catalog tables.
224 : */
225 : TransactionId shared_oldest_nonremovable_raw;
226 :
227 : /*
228 : * Oldest xid for which deleted tuples need to be retained in non-shared
229 : * catalog tables.
230 : */
231 : TransactionId catalog_oldest_nonremovable;
232 :
233 : /*
234 : * Oldest xid for which deleted tuples need to be retained in normal user
235 : * defined tables.
236 : */
237 : TransactionId data_oldest_nonremovable;
238 :
239 : /*
240 : * Oldest xid for which deleted tuples need to be retained in this
241 : * session's temporary tables.
242 : */
243 : TransactionId temp_oldest_nonremovable;
244 : } ComputeXidHorizonsResult;
245 :
246 : /*
247 : * Return value for GlobalVisHorizonKindForRel().
248 : */
249 : typedef enum GlobalVisHorizonKind
250 : {
251 : VISHORIZON_SHARED,
252 : VISHORIZON_CATALOG,
253 : VISHORIZON_DATA,
254 : VISHORIZON_TEMP,
255 : } GlobalVisHorizonKind;
256 :
257 : /*
258 : * Reason codes for KnownAssignedXidsCompress().
259 : */
260 : typedef enum KAXCompressReason
261 : {
262 : KAX_NO_SPACE, /* need to free up space at array end */
263 : KAX_PRUNE, /* we just pruned old entries */
264 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
265 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
266 : } KAXCompressReason;
267 :
268 :
269 : static ProcArrayStruct *procArray;
270 :
271 : static PGPROC *allProcs;
272 :
273 : /*
274 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
275 : */
276 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
277 :
278 : /*
279 : * Bookkeeping for tracking emulated transactions in recovery
280 : */
281 : static TransactionId *KnownAssignedXids;
282 : static bool *KnownAssignedXidsValid;
283 : static TransactionId latestObservedXid = InvalidTransactionId;
284 :
285 : /*
286 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
287 : * the highest xid that might still be running that we don't have in
288 : * KnownAssignedXids.
289 : */
290 : static TransactionId standbySnapshotPendingXmin;
291 :
292 : /*
293 : * State for visibility checks on different types of relations. See struct
294 : * GlobalVisState for details. As shared, catalog, normal and temporary
295 : * relations can have different horizons, one such state exists for each.
296 : */
297 : static GlobalVisState GlobalVisSharedRels;
298 : static GlobalVisState GlobalVisCatalogRels;
299 : static GlobalVisState GlobalVisDataRels;
300 : static GlobalVisState GlobalVisTempRels;
301 :
302 : /*
303 : * This backend's RecentXmin at the last time the accurate xmin horizon was
304 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
305 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
306 : */
307 : static TransactionId ComputeXidHorizonsResultLastXmin;
308 :
309 : #ifdef XIDCACHE_DEBUG
310 :
311 : /* counters for XidCache measurement */
312 : static long xc_by_recent_xmin = 0;
313 : static long xc_by_known_xact = 0;
314 : static long xc_by_my_xact = 0;
315 : static long xc_by_latest_xid = 0;
316 : static long xc_by_main_xid = 0;
317 : static long xc_by_child_xid = 0;
318 : static long xc_by_known_assigned = 0;
319 : static long xc_no_overflow = 0;
320 : static long xc_slow_answer = 0;
321 :
322 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
323 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
324 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
325 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
326 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
327 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
328 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
329 : #define xc_no_overflow_inc() (xc_no_overflow++)
330 : #define xc_slow_answer_inc() (xc_slow_answer++)
331 :
332 : static void DisplayXidCache(void);
333 : #else /* !XIDCACHE_DEBUG */
334 :
335 : #define xc_by_recent_xmin_inc() ((void) 0)
336 : #define xc_by_known_xact_inc() ((void) 0)
337 : #define xc_by_my_xact_inc() ((void) 0)
338 : #define xc_by_latest_xid_inc() ((void) 0)
339 : #define xc_by_main_xid_inc() ((void) 0)
340 : #define xc_by_child_xid_inc() ((void) 0)
341 : #define xc_by_known_assigned_inc() ((void) 0)
342 : #define xc_no_overflow_inc() ((void) 0)
343 : #define xc_slow_answer_inc() ((void) 0)
344 : #endif /* XIDCACHE_DEBUG */
345 :
346 : /* Primitives for KnownAssignedXids array handling for standby */
347 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
348 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
349 : bool exclusive_lock);
350 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
351 : static bool KnownAssignedXidExists(TransactionId xid);
352 : static void KnownAssignedXidsRemove(TransactionId xid);
353 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
354 : TransactionId *subxids);
355 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
356 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
357 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
358 : TransactionId *xmin,
359 : TransactionId xmax);
360 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
361 : static void KnownAssignedXidsDisplay(int trace_level);
362 : static void KnownAssignedXidsReset(void);
363 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
364 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
365 : static void MaintainLatestCompletedXid(TransactionId latestXid);
366 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
367 :
368 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
369 : TransactionId xid);
370 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
371 :
372 : /*
373 : * Report shared-memory space needed by ProcArrayShmemInit
374 : */
375 : Size
376 3534 : ProcArrayShmemSize(void)
377 : {
378 : Size size;
379 :
380 : /* Size of the ProcArray structure itself */
381 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
382 :
383 3534 : size = offsetof(ProcArrayStruct, pgprocnos);
384 3534 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
385 :
386 : /*
387 : * During Hot Standby processing we have a data structure called
388 : * KnownAssignedXids, created in shared memory. Local data structures are
389 : * also created in various backends during GetSnapshotData(),
390 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
391 : * main structures created in those functions must be identically sized,
392 : * since we may at times copy the whole of the data structures around. We
393 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
394 : *
395 : * Ideally we'd only create this structure if we were actually doing hot
396 : * standby in the current run, but we don't know that yet at the time
397 : * shared memory is being set up.
398 : */
399 : #define TOTAL_MAX_CACHED_SUBXIDS \
400 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
401 :
402 3534 : if (EnableHotStandby)
403 : {
404 3514 : size = add_size(size,
405 : mul_size(sizeof(TransactionId),
406 3514 : TOTAL_MAX_CACHED_SUBXIDS));
407 3514 : size = add_size(size,
408 3514 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
409 : }
410 :
411 3534 : return size;
412 : }
413 :
414 : /*
415 : * Initialize the shared PGPROC array during postmaster startup.
416 : */
417 : void
418 1902 : ProcArrayShmemInit(void)
419 : {
420 : bool found;
421 :
422 : /* Create or attach to the ProcArray shared structure */
423 1902 : procArray = (ProcArrayStruct *)
424 1902 : ShmemInitStruct("Proc Array",
425 : add_size(offsetof(ProcArrayStruct, pgprocnos),
426 : mul_size(sizeof(int),
427 1902 : PROCARRAY_MAXPROCS)),
428 : &found);
429 :
430 1902 : if (!found)
431 : {
432 : /*
433 : * We're the first - initialize.
434 : */
435 1902 : procArray->numProcs = 0;
436 1902 : procArray->maxProcs = PROCARRAY_MAXPROCS;
437 1902 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
438 1902 : procArray->numKnownAssignedXids = 0;
439 1902 : procArray->tailKnownAssignedXids = 0;
440 1902 : procArray->headKnownAssignedXids = 0;
441 1902 : procArray->lastOverflowedXid = InvalidTransactionId;
442 1902 : procArray->replication_slot_xmin = InvalidTransactionId;
443 1902 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
444 1902 : TransamVariables->xactCompletionCount = 1;
445 : }
446 :
447 1902 : allProcs = ProcGlobal->allProcs;
448 :
449 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
450 1902 : if (EnableHotStandby)
451 : {
452 1892 : KnownAssignedXids = (TransactionId *)
453 1892 : ShmemInitStruct("KnownAssignedXids",
454 : mul_size(sizeof(TransactionId),
455 1892 : TOTAL_MAX_CACHED_SUBXIDS),
456 : &found);
457 1892 : KnownAssignedXidsValid = (bool *)
458 1892 : ShmemInitStruct("KnownAssignedXidsValid",
459 1892 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
460 : &found);
461 : }
462 1902 : }
463 :
464 : /*
465 : * Add the specified PGPROC to the shared array.
466 : */
467 : void
468 31892 : ProcArrayAdd(PGPROC *proc)
469 : {
470 31892 : int pgprocno = GetNumberFromPGProc(proc);
471 31892 : ProcArrayStruct *arrayP = procArray;
472 : int index;
473 : int movecount;
474 :
475 : /* See ProcGlobal comment explaining why both locks are held */
476 31892 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
477 31892 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
478 :
479 31892 : if (arrayP->numProcs >= arrayP->maxProcs)
480 : {
481 : /*
482 : * Oops, no room. (This really shouldn't happen, since there is a
483 : * fixed supply of PGPROC structs too, and so we should have failed
484 : * earlier.)
485 : */
486 0 : ereport(FATAL,
487 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
488 : errmsg("sorry, too many clients already")));
489 : }
490 :
491 : /*
492 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
493 : * locality of references much better. This is useful while traversing the
494 : * ProcArray because there is an increased likelihood of finding the next
495 : * PGPROC structure in the cache.
496 : *
497 : * Since the occurrence of adding/removing a proc is much lower than the
498 : * access to the ProcArray itself, the overhead should be marginal
499 : */
500 67912 : for (index = 0; index < arrayP->numProcs; index++)
501 : {
502 60330 : int this_procno = arrayP->pgprocnos[index];
503 :
504 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
505 : Assert(allProcs[this_procno].pgxactoff == index);
506 :
507 : /* If we have found our right position in the array, break */
508 60330 : if (this_procno > pgprocno)
509 24310 : break;
510 : }
511 :
512 31892 : movecount = arrayP->numProcs - index;
513 31892 : memmove(&arrayP->pgprocnos[index + 1],
514 31892 : &arrayP->pgprocnos[index],
515 : movecount * sizeof(*arrayP->pgprocnos));
516 31892 : memmove(&ProcGlobal->xids[index + 1],
517 31892 : &ProcGlobal->xids[index],
518 : movecount * sizeof(*ProcGlobal->xids));
519 31892 : memmove(&ProcGlobal->subxidStates[index + 1],
520 31892 : &ProcGlobal->subxidStates[index],
521 : movecount * sizeof(*ProcGlobal->subxidStates));
522 31892 : memmove(&ProcGlobal->statusFlags[index + 1],
523 31892 : &ProcGlobal->statusFlags[index],
524 : movecount * sizeof(*ProcGlobal->statusFlags));
525 :
526 31892 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
527 31892 : proc->pgxactoff = index;
528 31892 : ProcGlobal->xids[index] = proc->xid;
529 31892 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
530 31892 : ProcGlobal->statusFlags[index] = proc->statusFlags;
531 :
532 31892 : arrayP->numProcs++;
533 :
534 : /* adjust pgxactoff for all following PGPROCs */
535 31892 : index++;
536 91034 : for (; index < arrayP->numProcs; index++)
537 : {
538 59142 : int procno = arrayP->pgprocnos[index];
539 :
540 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
541 : Assert(allProcs[procno].pgxactoff == index - 1);
542 :
543 59142 : allProcs[procno].pgxactoff = index;
544 : }
545 :
546 : /*
547 : * Release in reversed acquisition order, to reduce frequency of having to
548 : * wait for XidGenLock while holding ProcArrayLock.
549 : */
550 31892 : LWLockRelease(XidGenLock);
551 31892 : LWLockRelease(ProcArrayLock);
552 31892 : }
553 :
554 : /*
555 : * Remove the specified PGPROC from the shared array.
556 : *
557 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
558 : * array, and thus causing it to appear as "not running" anymore. In this
559 : * case we must advance latestCompletedXid. (This is essentially the same
560 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
561 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
562 : * twophase.c depends on the latter.)
563 : */
564 : void
565 31844 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
566 : {
567 31844 : ProcArrayStruct *arrayP = procArray;
568 : int myoff;
569 : int movecount;
570 :
571 : #ifdef XIDCACHE_DEBUG
572 : /* dump stats at backend shutdown, but not prepared-xact end */
573 : if (proc->pid != 0)
574 : DisplayXidCache();
575 : #endif
576 :
577 : /* See ProcGlobal comment explaining why both locks are held */
578 31844 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
579 31844 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
580 :
581 31844 : myoff = proc->pgxactoff;
582 :
583 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
584 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
585 :
586 31844 : if (TransactionIdIsValid(latestXid))
587 : {
588 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
589 :
590 : /* Advance global latestCompletedXid while holding the lock */
591 760 : MaintainLatestCompletedXid(latestXid);
592 :
593 : /* Same with xactCompletionCount */
594 760 : TransamVariables->xactCompletionCount++;
595 :
596 760 : ProcGlobal->xids[myoff] = InvalidTransactionId;
597 760 : ProcGlobal->subxidStates[myoff].overflowed = false;
598 760 : ProcGlobal->subxidStates[myoff].count = 0;
599 : }
600 : else
601 : {
602 : /* Shouldn't be trying to remove a live transaction here */
603 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
604 : }
605 :
606 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
607 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
608 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
609 :
610 31844 : ProcGlobal->statusFlags[myoff] = 0;
611 :
612 : /* Keep the PGPROC array sorted. See notes above */
613 31844 : movecount = arrayP->numProcs - myoff - 1;
614 31844 : memmove(&arrayP->pgprocnos[myoff],
615 31844 : &arrayP->pgprocnos[myoff + 1],
616 : movecount * sizeof(*arrayP->pgprocnos));
617 31844 : memmove(&ProcGlobal->xids[myoff],
618 31844 : &ProcGlobal->xids[myoff + 1],
619 : movecount * sizeof(*ProcGlobal->xids));
620 31844 : memmove(&ProcGlobal->subxidStates[myoff],
621 31844 : &ProcGlobal->subxidStates[myoff + 1],
622 : movecount * sizeof(*ProcGlobal->subxidStates));
623 31844 : memmove(&ProcGlobal->statusFlags[myoff],
624 31844 : &ProcGlobal->statusFlags[myoff + 1],
625 : movecount * sizeof(*ProcGlobal->statusFlags));
626 :
627 31844 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
628 31844 : arrayP->numProcs--;
629 :
630 : /*
631 : * Adjust pgxactoff of following procs for removed PGPROC (note that
632 : * numProcs already has been decremented).
633 : */
634 94446 : for (int index = myoff; index < arrayP->numProcs; index++)
635 : {
636 62602 : int procno = arrayP->pgprocnos[index];
637 :
638 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
639 : Assert(allProcs[procno].pgxactoff - 1 == index);
640 :
641 62602 : allProcs[procno].pgxactoff = index;
642 : }
643 :
644 : /*
645 : * Release in reversed acquisition order, to reduce frequency of having to
646 : * wait for XidGenLock while holding ProcArrayLock.
647 : */
648 31844 : LWLockRelease(XidGenLock);
649 31844 : LWLockRelease(ProcArrayLock);
650 31844 : }
651 :
652 :
653 : /*
654 : * ProcArrayEndTransaction -- mark a transaction as no longer running
655 : *
656 : * This is used interchangeably for commit and abort cases. The transaction
657 : * commit/abort must already be reported to WAL and pg_xact.
658 : *
659 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
660 : * latestXid is the latest Xid among the transaction's main XID and
661 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
662 : * the caller to pass latestXid, instead of computing it from the PGPROC's
663 : * contents, because the subxid information in the PGPROC might be
664 : * incomplete.)
665 : */
666 : void
667 747754 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
668 : {
669 747754 : if (TransactionIdIsValid(latestXid))
670 : {
671 : /*
672 : * We must lock ProcArrayLock while clearing our advertised XID, so
673 : * that we do not exit the set of "running" transactions while someone
674 : * else is taking a snapshot. See discussion in
675 : * src/backend/access/transam/README.
676 : */
677 : Assert(TransactionIdIsValid(proc->xid));
678 :
679 : /*
680 : * If we can immediately acquire ProcArrayLock, we clear our own XID
681 : * and release the lock. If not, use group XID clearing to improve
682 : * efficiency.
683 : */
684 241186 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
685 : {
686 240854 : ProcArrayEndTransactionInternal(proc, latestXid);
687 240854 : LWLockRelease(ProcArrayLock);
688 : }
689 : else
690 332 : ProcArrayGroupClearXid(proc, latestXid);
691 : }
692 : else
693 : {
694 : /*
695 : * If we have no XID, we don't need to lock, since we won't affect
696 : * anyone else's calculation of a snapshot. We might change their
697 : * estimate of global xmin, but that's OK.
698 : */
699 : Assert(!TransactionIdIsValid(proc->xid));
700 : Assert(proc->subxidStatus.count == 0);
701 : Assert(!proc->subxidStatus.overflowed);
702 :
703 506568 : proc->vxid.lxid = InvalidLocalTransactionId;
704 506568 : proc->xmin = InvalidTransactionId;
705 :
706 : /* be sure this is cleared in abort */
707 506568 : proc->delayChkptFlags = 0;
708 :
709 506568 : proc->recoveryConflictPending = false;
710 :
711 : /* must be cleared with xid/xmin: */
712 : /* avoid unnecessarily dirtying shared cachelines */
713 506568 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
714 : {
715 : Assert(!LWLockHeldByMe(ProcArrayLock));
716 97920 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
717 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
718 97920 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
719 97920 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
720 97920 : LWLockRelease(ProcArrayLock);
721 : }
722 : }
723 747754 : }
724 :
725 : /*
726 : * Mark a write transaction as no longer running.
727 : *
728 : * We don't do any locking here; caller must handle that.
729 : */
730 : static inline void
731 241186 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
732 : {
733 241186 : int pgxactoff = proc->pgxactoff;
734 :
735 : /*
736 : * Note: we need exclusive lock here because we're going to change other
737 : * processes' PGPROC entries.
738 : */
739 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
740 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
741 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
742 :
743 241186 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
744 241186 : proc->xid = InvalidTransactionId;
745 241186 : proc->vxid.lxid = InvalidLocalTransactionId;
746 241186 : proc->xmin = InvalidTransactionId;
747 :
748 : /* be sure this is cleared in abort */
749 241186 : proc->delayChkptFlags = 0;
750 :
751 241186 : proc->recoveryConflictPending = false;
752 :
753 : /* must be cleared with xid/xmin: */
754 : /* avoid unnecessarily dirtying shared cachelines */
755 241186 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
756 : {
757 1170 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
758 1170 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
759 : }
760 :
761 : /* Clear the subtransaction-XID cache too while holding the lock */
762 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
763 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
764 241186 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
765 : {
766 1138 : ProcGlobal->subxidStates[pgxactoff].count = 0;
767 1138 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
768 1138 : proc->subxidStatus.count = 0;
769 1138 : proc->subxidStatus.overflowed = false;
770 : }
771 :
772 : /* Also advance global latestCompletedXid while holding the lock */
773 241186 : MaintainLatestCompletedXid(latestXid);
774 :
775 : /* Same with xactCompletionCount */
776 241186 : TransamVariables->xactCompletionCount++;
777 241186 : }
778 :
779 : /*
780 : * ProcArrayGroupClearXid -- group XID clearing
781 : *
782 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
783 : * commit time, add ourselves to a list of processes that need their XIDs
784 : * cleared. The first process to add itself to the list will acquire
785 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
786 : * on behalf of all group members. This avoids a great deal of contention
787 : * around ProcArrayLock when many processes are trying to commit at once,
788 : * since the lock need not be repeatedly handed off from one committing
789 : * process to the next.
790 : */
791 : static void
792 332 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
793 : {
794 332 : int pgprocno = GetNumberFromPGProc(proc);
795 332 : PROC_HDR *procglobal = ProcGlobal;
796 : uint32 nextidx;
797 : uint32 wakeidx;
798 :
799 : /* We should definitely have an XID to clear. */
800 : Assert(TransactionIdIsValid(proc->xid));
801 :
802 : /* Add ourselves to the list of processes needing a group XID clear. */
803 332 : proc->procArrayGroupMember = true;
804 332 : proc->procArrayGroupMemberXid = latestXid;
805 332 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
806 : while (true)
807 : {
808 334 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
809 :
810 334 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
811 : &nextidx,
812 : (uint32) pgprocno))
813 332 : break;
814 : }
815 :
816 : /*
817 : * If the list was not empty, the leader will clear our XID. It is
818 : * impossible to have followers without a leader because the first process
819 : * that has added itself to the list will always have nextidx as
820 : * INVALID_PROC_NUMBER.
821 : */
822 332 : if (nextidx != INVALID_PROC_NUMBER)
823 : {
824 20 : int extraWaits = 0;
825 :
826 : /* Sleep until the leader clears our XID. */
827 20 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
828 : for (;;)
829 : {
830 : /* acts as a read barrier */
831 20 : PGSemaphoreLock(proc->sem);
832 20 : if (!proc->procArrayGroupMember)
833 20 : break;
834 0 : extraWaits++;
835 : }
836 20 : pgstat_report_wait_end();
837 :
838 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
839 :
840 : /* Fix semaphore count for any absorbed wakeups */
841 20 : while (extraWaits-- > 0)
842 0 : PGSemaphoreUnlock(proc->sem);
843 20 : return;
844 : }
845 :
846 : /* We are the leader. Acquire the lock on behalf of everyone. */
847 312 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
848 :
849 : /*
850 : * Now that we've got the lock, clear the list of processes waiting for
851 : * group XID clearing, saving a pointer to the head of the list. Trying
852 : * to pop elements one at a time could lead to an ABA problem.
853 : */
854 312 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
855 : INVALID_PROC_NUMBER);
856 :
857 : /* Remember head of list so we can perform wakeups after dropping lock. */
858 312 : wakeidx = nextidx;
859 :
860 : /* Walk the list and clear all XIDs. */
861 644 : while (nextidx != INVALID_PROC_NUMBER)
862 : {
863 332 : PGPROC *nextproc = &allProcs[nextidx];
864 :
865 332 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
866 :
867 : /* Move to next proc in list. */
868 332 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
869 : }
870 :
871 : /* We're done with the lock now. */
872 312 : LWLockRelease(ProcArrayLock);
873 :
874 : /*
875 : * Now that we've released the lock, go back and wake everybody up. We
876 : * don't do this under the lock so as to keep lock hold times to a
877 : * minimum. The system calls we need to perform to wake other processes
878 : * up are probably much slower than the simple memory writes we did while
879 : * holding the lock.
880 : */
881 644 : while (wakeidx != INVALID_PROC_NUMBER)
882 : {
883 332 : PGPROC *nextproc = &allProcs[wakeidx];
884 :
885 332 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
886 332 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
887 :
888 : /* ensure all previous writes are visible before follower continues. */
889 332 : pg_write_barrier();
890 :
891 332 : nextproc->procArrayGroupMember = false;
892 :
893 332 : if (nextproc != MyProc)
894 20 : PGSemaphoreUnlock(nextproc->sem);
895 : }
896 : }
897 :
898 : /*
899 : * ProcArrayClearTransaction -- clear the transaction fields
900 : *
901 : * This is used after successfully preparing a 2-phase transaction. We are
902 : * not actually reporting the transaction's XID as no longer running --- it
903 : * will still appear as running because the 2PC's gxact is in the ProcArray
904 : * too. We just have to clear out our own PGPROC.
905 : */
906 : void
907 744 : ProcArrayClearTransaction(PGPROC *proc)
908 : {
909 : int pgxactoff;
910 :
911 : /*
912 : * Currently we need to lock ProcArrayLock exclusively here, as we
913 : * increment xactCompletionCount below. We also need it at least in shared
914 : * mode for pgproc->pgxactoff to stay the same below.
915 : *
916 : * We could however, as this action does not actually change anyone's view
917 : * of the set of running XIDs (our entry is duplicate with the gxact that
918 : * has already been inserted into the ProcArray), lower the lock level to
919 : * shared if we were to make xactCompletionCount an atomic variable. But
920 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
921 : * enough for this not to be the bottleneck. If it ever becomes a
922 : * bottleneck it may also be worth considering to combine this with the
923 : * subsequent ProcArrayRemove()
924 : */
925 744 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
926 :
927 744 : pgxactoff = proc->pgxactoff;
928 :
929 744 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
930 744 : proc->xid = InvalidTransactionId;
931 :
932 744 : proc->vxid.lxid = InvalidLocalTransactionId;
933 744 : proc->xmin = InvalidTransactionId;
934 744 : proc->recoveryConflictPending = false;
935 :
936 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
937 : Assert(!proc->delayChkptFlags);
938 :
939 : /*
940 : * Need to increment completion count even though transaction hasn't
941 : * really committed yet. The reason for that is that GetSnapshotData()
942 : * omits the xid of the current transaction, thus without the increment we
943 : * otherwise could end up reusing the snapshot later. Which would be bad,
944 : * because it might not count the prepared transaction as running.
945 : */
946 744 : TransamVariables->xactCompletionCount++;
947 :
948 : /* Clear the subtransaction-XID cache too */
949 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
950 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
951 744 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
952 : {
953 296 : ProcGlobal->subxidStates[pgxactoff].count = 0;
954 296 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
955 296 : proc->subxidStatus.count = 0;
956 296 : proc->subxidStatus.overflowed = false;
957 : }
958 :
959 744 : LWLockRelease(ProcArrayLock);
960 744 : }
961 :
962 : /*
963 : * Update TransamVariables->latestCompletedXid to point to latestXid if
964 : * currently older.
965 : */
966 : static void
967 243244 : MaintainLatestCompletedXid(TransactionId latestXid)
968 : {
969 243244 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
970 :
971 : Assert(FullTransactionIdIsValid(cur_latest));
972 : Assert(!RecoveryInProgress());
973 : Assert(LWLockHeldByMe(ProcArrayLock));
974 :
975 243244 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
976 : {
977 221676 : TransamVariables->latestCompletedXid =
978 221676 : FullXidRelativeTo(cur_latest, latestXid);
979 : }
980 :
981 : Assert(IsBootstrapProcessingMode() ||
982 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
983 243244 : }
984 :
985 : /*
986 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
987 : */
988 : static void
989 40360 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
990 : {
991 40360 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
992 : FullTransactionId rel;
993 :
994 : Assert(AmStartupProcess() || !IsUnderPostmaster);
995 : Assert(LWLockHeldByMe(ProcArrayLock));
996 :
997 : /*
998 : * Need a FullTransactionId to compare latestXid with. Can't rely on
999 : * latestCompletedXid to be initialized in recovery. But in recovery it's
1000 : * safe to access nextXid without a lock for the startup process.
1001 : */
1002 40360 : rel = TransamVariables->nextXid;
1003 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1004 :
1005 80522 : if (!FullTransactionIdIsValid(cur_latest) ||
1006 40162 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1007 : {
1008 32032 : TransamVariables->latestCompletedXid =
1009 32032 : FullXidRelativeTo(rel, latestXid);
1010 : }
1011 :
1012 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1013 40360 : }
1014 :
1015 : /*
1016 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1017 : *
1018 : * Remember up to where the startup process initialized the CLOG and subtrans
1019 : * so we can ensure it's initialized gaplessly up to the point where necessary
1020 : * while in recovery.
1021 : */
1022 : void
1023 198 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1024 : {
1025 : Assert(standbyState == STANDBY_INITIALIZED);
1026 : Assert(TransactionIdIsNormal(initializedUptoXID));
1027 :
1028 : /*
1029 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1030 : * to, so we can extend it from that point onwards in
1031 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1032 : * ProcArrayApplyRecoveryInfo().
1033 : */
1034 198 : latestObservedXid = initializedUptoXID;
1035 198 : TransactionIdRetreat(latestObservedXid);
1036 198 : }
1037 :
1038 : /*
1039 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1040 : *
1041 : * Takes us through 3 states: Initialized, Pending and Ready.
1042 : * Normal case is to go all the way to Ready straight away, though there
1043 : * are atypical cases where we need to take it in steps.
1044 : *
1045 : * Use the data about running transactions on the primary to create the initial
1046 : * state of KnownAssignedXids. We also use these records to regularly prune
1047 : * KnownAssignedXids because we know it is possible that some transactions
1048 : * with FATAL errors fail to write abort records, which could cause eventual
1049 : * overflow.
1050 : *
1051 : * See comments for LogStandbySnapshot().
1052 : */
1053 : void
1054 870 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1055 : {
1056 : TransactionId *xids;
1057 : TransactionId advanceNextXid;
1058 : int nxids;
1059 : int i;
1060 :
1061 : Assert(standbyState >= STANDBY_INITIALIZED);
1062 : Assert(TransactionIdIsValid(running->nextXid));
1063 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1064 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1065 :
1066 : /*
1067 : * Remove stale transactions, if any.
1068 : */
1069 870 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1070 :
1071 : /*
1072 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1073 : * because we will need it up to date for accessing two-phase transactions
1074 : * in StandbyReleaseOldLocks().
1075 : */
1076 870 : advanceNextXid = running->nextXid;
1077 870 : TransactionIdRetreat(advanceNextXid);
1078 870 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1079 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1080 :
1081 : /*
1082 : * Remove stale locks, if any.
1083 : */
1084 870 : StandbyReleaseOldLocks(running->oldestRunningXid);
1085 :
1086 : /*
1087 : * If our snapshot is already valid, nothing else to do...
1088 : */
1089 870 : if (standbyState == STANDBY_SNAPSHOT_READY)
1090 672 : return;
1091 :
1092 : /*
1093 : * If our initial RunningTransactionsData had an overflowed snapshot then
1094 : * we knew we were missing some subxids from our snapshot. If we continue
1095 : * to see overflowed snapshots then we might never be able to start up, so
1096 : * we make another test to see if our snapshot is now valid. We know that
1097 : * the missing subxids are equal to or earlier than nextXid. After we
1098 : * initialise we continue to apply changes during recovery, so once the
1099 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1100 : * know that we no longer have missing information and can mark the
1101 : * snapshot as valid.
1102 : */
1103 198 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1104 : {
1105 : /*
1106 : * If the snapshot isn't overflowed or if its empty we can reset our
1107 : * pending state and use this snapshot instead.
1108 : */
1109 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1110 : {
1111 : /*
1112 : * If we have already collected known assigned xids, we need to
1113 : * throw them away before we apply the recovery snapshot.
1114 : */
1115 0 : KnownAssignedXidsReset();
1116 0 : standbyState = STANDBY_INITIALIZED;
1117 : }
1118 : else
1119 : {
1120 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1121 : running->oldestRunningXid))
1122 : {
1123 0 : standbyState = STANDBY_SNAPSHOT_READY;
1124 0 : elog(DEBUG1,
1125 : "recovery snapshots are now enabled");
1126 : }
1127 : else
1128 0 : elog(DEBUG1,
1129 : "recovery snapshot waiting for non-overflowed snapshot or "
1130 : "until oldest active xid on standby is at least %u (now %u)",
1131 : standbySnapshotPendingXmin,
1132 : running->oldestRunningXid);
1133 0 : return;
1134 : }
1135 : }
1136 :
1137 : Assert(standbyState == STANDBY_INITIALIZED);
1138 :
1139 : /*
1140 : * NB: this can be reached at least twice, so make sure new code can deal
1141 : * with that.
1142 : */
1143 :
1144 : /*
1145 : * Nobody else is running yet, but take locks anyhow
1146 : */
1147 198 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1148 :
1149 : /*
1150 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1151 : * sort them first.
1152 : *
1153 : * Some of the new xids are top-level xids and some are subtransactions.
1154 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1155 : * aren't overflowed then all xids will fit in snapshot and so we don't
1156 : * need subtrans. If we later overflow, an xid assignment record will add
1157 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1158 : * don't have enough information to correctly update subtrans anyway.
1159 : */
1160 :
1161 : /*
1162 : * Allocate a temporary array to avoid modifying the array passed as
1163 : * argument.
1164 : */
1165 198 : xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt));
1166 :
1167 : /*
1168 : * Add to the temp array any xids which have not already completed.
1169 : */
1170 198 : nxids = 0;
1171 206 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1172 : {
1173 8 : TransactionId xid = running->xids[i];
1174 :
1175 : /*
1176 : * The running-xacts snapshot can contain xids that were still visible
1177 : * in the procarray when the snapshot was taken, but were already
1178 : * WAL-logged as completed. They're not running anymore, so ignore
1179 : * them.
1180 : */
1181 8 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1182 0 : continue;
1183 :
1184 8 : xids[nxids++] = xid;
1185 : }
1186 :
1187 198 : if (nxids > 0)
1188 : {
1189 8 : if (procArray->numKnownAssignedXids != 0)
1190 : {
1191 0 : LWLockRelease(ProcArrayLock);
1192 0 : elog(ERROR, "KnownAssignedXids is not empty");
1193 : }
1194 :
1195 : /*
1196 : * Sort the array so that we can add them safely into
1197 : * KnownAssignedXids.
1198 : *
1199 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1200 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1201 : * come from RUNNING_XACTS, which means there are only normal XIDs
1202 : * from the same epoch, so this is safe.
1203 : */
1204 8 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1205 :
1206 : /*
1207 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1208 : * snapshot may include duplicated xids because of prepared
1209 : * transactions, so ignore them.
1210 : */
1211 16 : for (i = 0; i < nxids; i++)
1212 : {
1213 8 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1214 : {
1215 0 : elog(DEBUG1,
1216 : "found duplicated transaction %u for KnownAssignedXids insertion",
1217 : xids[i]);
1218 0 : continue;
1219 : }
1220 8 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1221 : }
1222 :
1223 8 : KnownAssignedXidsDisplay(DEBUG3);
1224 : }
1225 :
1226 198 : pfree(xids);
1227 :
1228 : /*
1229 : * latestObservedXid is at least set to the point where SUBTRANS was
1230 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1231 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1232 : * subtrans from thereon, up to nextXid - 1.
1233 : *
1234 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1235 : * because we've just added xids to the known assigned xids machinery that
1236 : * haven't gone through RecordKnownAssignedTransactionId().
1237 : */
1238 : Assert(TransactionIdIsNormal(latestObservedXid));
1239 198 : TransactionIdAdvance(latestObservedXid);
1240 198 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1241 : {
1242 0 : ExtendSUBTRANS(latestObservedXid);
1243 0 : TransactionIdAdvance(latestObservedXid);
1244 : }
1245 198 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1246 :
1247 : /* ----------
1248 : * Now we've got the running xids we need to set the global values that
1249 : * are used to track snapshots as they evolve further.
1250 : *
1251 : * - latestCompletedXid which will be the xmax for snapshots
1252 : * - lastOverflowedXid which shows whether snapshots overflow
1253 : * - nextXid
1254 : *
1255 : * If the snapshot overflowed, then we still initialise with what we know,
1256 : * but the recovery snapshot isn't fully valid yet because we know there
1257 : * are some subxids missing. We don't know the specific subxids that are
1258 : * missing, so conservatively assume the last one is latestObservedXid.
1259 : * ----------
1260 : */
1261 198 : if (running->subxid_status == SUBXIDS_MISSING)
1262 : {
1263 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1264 :
1265 0 : standbySnapshotPendingXmin = latestObservedXid;
1266 0 : procArray->lastOverflowedXid = latestObservedXid;
1267 : }
1268 : else
1269 : {
1270 198 : standbyState = STANDBY_SNAPSHOT_READY;
1271 :
1272 198 : standbySnapshotPendingXmin = InvalidTransactionId;
1273 :
1274 : /*
1275 : * If the 'xids' array didn't include all subtransactions, we have to
1276 : * mark any snapshots taken as overflowed.
1277 : */
1278 198 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1279 48 : procArray->lastOverflowedXid = latestObservedXid;
1280 : else
1281 : {
1282 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1283 150 : procArray->lastOverflowedXid = InvalidTransactionId;
1284 : }
1285 : }
1286 :
1287 : /*
1288 : * If a transaction wrote a commit record in the gap between taking and
1289 : * logging the snapshot then latestCompletedXid may already be higher than
1290 : * the value from the snapshot, so check before we use the incoming value.
1291 : * It also might not yet be set at all.
1292 : */
1293 198 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1294 :
1295 : /*
1296 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1297 : * nobody can see it yet.
1298 : */
1299 :
1300 198 : LWLockRelease(ProcArrayLock);
1301 :
1302 198 : KnownAssignedXidsDisplay(DEBUG3);
1303 198 : if (standbyState == STANDBY_SNAPSHOT_READY)
1304 198 : elog(DEBUG1, "recovery snapshots are now enabled");
1305 : else
1306 0 : elog(DEBUG1,
1307 : "recovery snapshot waiting for non-overflowed snapshot or "
1308 : "until oldest active xid on standby is at least %u (now %u)",
1309 : standbySnapshotPendingXmin,
1310 : running->oldestRunningXid);
1311 : }
1312 :
1313 : /*
1314 : * ProcArrayApplyXidAssignment
1315 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1316 : */
1317 : void
1318 42 : ProcArrayApplyXidAssignment(TransactionId topxid,
1319 : int nsubxids, TransactionId *subxids)
1320 : {
1321 : TransactionId max_xid;
1322 : int i;
1323 :
1324 : Assert(standbyState >= STANDBY_INITIALIZED);
1325 :
1326 42 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1327 :
1328 : /*
1329 : * Mark all the subtransactions as observed.
1330 : *
1331 : * NOTE: This will fail if the subxid contains too many previously
1332 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1333 : * as the code stands, because xid-assignment records should never contain
1334 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1335 : */
1336 42 : RecordKnownAssignedTransactionIds(max_xid);
1337 :
1338 : /*
1339 : * Notice that we update pg_subtrans with the top-level xid, rather than
1340 : * the parent xid. This is a difference between normal processing and
1341 : * recovery, yet is still correct in all cases. The reason is that
1342 : * subtransaction commit is not marked in clog until commit processing, so
1343 : * all aborted subtransactions have already been clearly marked in clog.
1344 : * As a result we are able to refer directly to the top-level
1345 : * transaction's state rather than skipping through all the intermediate
1346 : * states in the subtransaction tree. This should be the first time we
1347 : * have attempted to SubTransSetParent().
1348 : */
1349 2730 : for (i = 0; i < nsubxids; i++)
1350 2688 : SubTransSetParent(subxids[i], topxid);
1351 :
1352 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1353 42 : if (standbyState == STANDBY_INITIALIZED)
1354 0 : return;
1355 :
1356 : /*
1357 : * Uses same locking as transaction commit
1358 : */
1359 42 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1360 :
1361 : /*
1362 : * Remove subxids from known-assigned-xacts.
1363 : */
1364 42 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1365 :
1366 : /*
1367 : * Advance lastOverflowedXid to be at least the last of these subxids.
1368 : */
1369 42 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1370 42 : procArray->lastOverflowedXid = max_xid;
1371 :
1372 42 : LWLockRelease(ProcArrayLock);
1373 : }
1374 :
1375 : /*
1376 : * TransactionIdIsInProgress -- is given transaction running in some backend
1377 : *
1378 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1379 : * there are four possibilities for finding a running transaction:
1380 : *
1381 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1382 : * by looking at ProcGlobal->xids.
1383 : *
1384 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1385 : * We can find this out cheaply too.
1386 : *
1387 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1388 : * if the Xid is running on the primary.
1389 : *
1390 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1391 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1392 : * This is the slowest way, but sadly it has to be done always if the others
1393 : * failed, unless we see that the cached subxact sets are complete (none have
1394 : * overflowed).
1395 : *
1396 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1397 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1398 : * This buys back some concurrency (and we can't retrieve the main Xids from
1399 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1400 : */
1401 : bool
1402 15415006 : TransactionIdIsInProgress(TransactionId xid)
1403 : {
1404 : static TransactionId *xids = NULL;
1405 : static TransactionId *other_xids;
1406 : XidCacheStatus *other_subxidstates;
1407 15415006 : int nxids = 0;
1408 15415006 : ProcArrayStruct *arrayP = procArray;
1409 : TransactionId topxid;
1410 : TransactionId latestCompletedXid;
1411 : int mypgxactoff;
1412 : int numProcs;
1413 : int j;
1414 :
1415 : /*
1416 : * Don't bother checking a transaction older than RecentXmin; it could not
1417 : * possibly still be running. (Note: in particular, this guarantees that
1418 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1419 : * running.)
1420 : */
1421 15415006 : if (TransactionIdPrecedes(xid, RecentXmin))
1422 : {
1423 : xc_by_recent_xmin_inc();
1424 8570642 : return false;
1425 : }
1426 :
1427 : /*
1428 : * We may have just checked the status of this transaction, so if it is
1429 : * already known to be completed, we can fall out without any access to
1430 : * shared memory.
1431 : */
1432 6844364 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1433 : {
1434 : xc_by_known_xact_inc();
1435 1789962 : return false;
1436 : }
1437 :
1438 : /*
1439 : * Also, we can handle our own transaction (and subtransactions) without
1440 : * any access to shared memory.
1441 : */
1442 5054402 : if (TransactionIdIsCurrentTransactionId(xid))
1443 : {
1444 : xc_by_my_xact_inc();
1445 388844 : return true;
1446 : }
1447 :
1448 : /*
1449 : * If first time through, get workspace to remember main XIDs in. We
1450 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1451 : */
1452 4665558 : if (xids == NULL)
1453 : {
1454 : /*
1455 : * In hot standby mode, reserve enough space to hold all xids in the
1456 : * known-assigned list. If we later finish recovery, we no longer need
1457 : * the bigger array, but we don't bother to shrink it.
1458 : */
1459 1718 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1460 :
1461 1718 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1462 1718 : if (xids == NULL)
1463 0 : ereport(ERROR,
1464 : (errcode(ERRCODE_OUT_OF_MEMORY),
1465 : errmsg("out of memory")));
1466 : }
1467 :
1468 4665558 : other_xids = ProcGlobal->xids;
1469 4665558 : other_subxidstates = ProcGlobal->subxidStates;
1470 :
1471 4665558 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1472 :
1473 : /*
1474 : * Now that we have the lock, we can check latestCompletedXid; if the
1475 : * target Xid is after that, it's surely still running.
1476 : */
1477 4665558 : latestCompletedXid =
1478 4665558 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1479 4665558 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1480 : {
1481 69478 : LWLockRelease(ProcArrayLock);
1482 : xc_by_latest_xid_inc();
1483 69478 : return true;
1484 : }
1485 :
1486 : /* No shortcuts, gotta grovel through the array */
1487 4596080 : mypgxactoff = MyProc->pgxactoff;
1488 4596080 : numProcs = arrayP->numProcs;
1489 4775936 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1490 : {
1491 : int pgprocno;
1492 : PGPROC *proc;
1493 : TransactionId pxid;
1494 : int pxids;
1495 :
1496 : /* Ignore ourselves --- dealt with it above */
1497 4757082 : if (pgxactoff == mypgxactoff)
1498 20986 : continue;
1499 :
1500 : /* Fetch xid just once - see GetNewTransactionId */
1501 4736096 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1502 :
1503 4736096 : if (!TransactionIdIsValid(pxid))
1504 102554 : continue;
1505 :
1506 : /*
1507 : * Step 1: check the main Xid
1508 : */
1509 4633542 : if (TransactionIdEquals(pxid, xid))
1510 : {
1511 4576962 : LWLockRelease(ProcArrayLock);
1512 : xc_by_main_xid_inc();
1513 4576962 : return true;
1514 : }
1515 :
1516 : /*
1517 : * We can ignore main Xids that are younger than the target Xid, since
1518 : * the target could not possibly be their child.
1519 : */
1520 56580 : if (TransactionIdPrecedes(xid, pxid))
1521 25248 : continue;
1522 :
1523 : /*
1524 : * Step 2: check the cached child-Xids arrays
1525 : */
1526 31332 : pxids = other_subxidstates[pgxactoff].count;
1527 31332 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1528 31332 : pgprocno = arrayP->pgprocnos[pgxactoff];
1529 31332 : proc = &allProcs[pgprocno];
1530 63124 : for (j = pxids - 1; j >= 0; j--)
1531 : {
1532 : /* Fetch xid just once - see GetNewTransactionId */
1533 32056 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1534 :
1535 32056 : if (TransactionIdEquals(cxid, xid))
1536 : {
1537 264 : LWLockRelease(ProcArrayLock);
1538 : xc_by_child_xid_inc();
1539 264 : return true;
1540 : }
1541 : }
1542 :
1543 : /*
1544 : * Save the main Xid for step 4. We only need to remember main Xids
1545 : * that have uncached children. (Note: there is no race condition
1546 : * here because the overflowed flag cannot be cleared, only set, while
1547 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1548 : * worry about.)
1549 : */
1550 31068 : if (other_subxidstates[pgxactoff].overflowed)
1551 354 : xids[nxids++] = pxid;
1552 : }
1553 :
1554 : /*
1555 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1556 : * in the list must be treated as running.
1557 : */
1558 18854 : if (RecoveryInProgress())
1559 : {
1560 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1561 : Assert(nxids == 0);
1562 :
1563 2 : if (KnownAssignedXidExists(xid))
1564 : {
1565 0 : LWLockRelease(ProcArrayLock);
1566 : xc_by_known_assigned_inc();
1567 0 : return true;
1568 : }
1569 :
1570 : /*
1571 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1572 : * too. Fetch all xids from KnownAssignedXids that are lower than
1573 : * xid, since if xid is a subtransaction its parent will always have a
1574 : * lower value. Note we will collect both main and subXIDs here, but
1575 : * there's no help for it.
1576 : */
1577 2 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1578 0 : nxids = KnownAssignedXidsGet(xids, xid);
1579 : }
1580 :
1581 18854 : LWLockRelease(ProcArrayLock);
1582 :
1583 : /*
1584 : * If none of the relevant caches overflowed, we know the Xid is not
1585 : * running without even looking at pg_subtrans.
1586 : */
1587 18854 : if (nxids == 0)
1588 : {
1589 : xc_no_overflow_inc();
1590 18500 : cachedXidIsNotInProgress = xid;
1591 18500 : return false;
1592 : }
1593 :
1594 : /*
1595 : * Step 4: have to check pg_subtrans.
1596 : *
1597 : * At this point, we know it's either a subtransaction of one of the Xids
1598 : * in xids[], or it's not running. If it's an already-failed
1599 : * subtransaction, we want to say "not running" even though its parent may
1600 : * still be running. So first, check pg_xact to see if it's been aborted.
1601 : */
1602 : xc_slow_answer_inc();
1603 :
1604 354 : if (TransactionIdDidAbort(xid))
1605 : {
1606 0 : cachedXidIsNotInProgress = xid;
1607 0 : return false;
1608 : }
1609 :
1610 : /*
1611 : * It isn't aborted, so check whether the transaction tree it belongs to
1612 : * is still running (or, more precisely, whether it was running when we
1613 : * held ProcArrayLock).
1614 : */
1615 354 : topxid = SubTransGetTopmostTransaction(xid);
1616 : Assert(TransactionIdIsValid(topxid));
1617 668 : if (!TransactionIdEquals(topxid, xid) &&
1618 314 : pg_lfind32(topxid, xids, nxids))
1619 314 : return true;
1620 :
1621 40 : cachedXidIsNotInProgress = xid;
1622 40 : return false;
1623 : }
1624 :
1625 : /*
1626 : * TransactionIdIsActive -- is xid the top-level XID of an active backend?
1627 : *
1628 : * This differs from TransactionIdIsInProgress in that it ignores prepared
1629 : * transactions, as well as transactions running on the primary if we're in
1630 : * hot standby. Also, we ignore subtransactions since that's not needed
1631 : * for current uses.
1632 : */
1633 : bool
1634 0 : TransactionIdIsActive(TransactionId xid)
1635 : {
1636 0 : bool result = false;
1637 0 : ProcArrayStruct *arrayP = procArray;
1638 0 : TransactionId *other_xids = ProcGlobal->xids;
1639 : int i;
1640 :
1641 : /*
1642 : * Don't bother checking a transaction older than RecentXmin; it could not
1643 : * possibly still be running.
1644 : */
1645 0 : if (TransactionIdPrecedes(xid, RecentXmin))
1646 0 : return false;
1647 :
1648 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1649 :
1650 0 : for (i = 0; i < arrayP->numProcs; i++)
1651 : {
1652 0 : int pgprocno = arrayP->pgprocnos[i];
1653 0 : PGPROC *proc = &allProcs[pgprocno];
1654 : TransactionId pxid;
1655 :
1656 : /* Fetch xid just once - see GetNewTransactionId */
1657 0 : pxid = UINT32_ACCESS_ONCE(other_xids[i]);
1658 :
1659 0 : if (!TransactionIdIsValid(pxid))
1660 0 : continue;
1661 :
1662 0 : if (proc->pid == 0)
1663 0 : continue; /* ignore prepared transactions */
1664 :
1665 0 : if (TransactionIdEquals(pxid, xid))
1666 : {
1667 0 : result = true;
1668 0 : break;
1669 : }
1670 : }
1671 :
1672 0 : LWLockRelease(ProcArrayLock);
1673 :
1674 0 : return result;
1675 : }
1676 :
1677 :
1678 : /*
1679 : * Determine XID horizons.
1680 : *
1681 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1682 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1683 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1684 : * GlobalVisState).
1685 : *
1686 : * See the definition of ComputeXidHorizonsResult for the various computed
1687 : * horizons.
1688 : *
1689 : * For VACUUM separate horizons (used to decide which deleted tuples must
1690 : * be preserved), for shared and non-shared tables are computed. For shared
1691 : * relations backends in all databases must be considered, but for non-shared
1692 : * relations that's not required, since only backends in my own database could
1693 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1694 : * VACUUMs because (a) they must be working on other tables, and (b) they
1695 : * don't need to do snapshot-based lookups.
1696 : *
1697 : * This also computes a horizon used to truncate pg_subtrans. For that
1698 : * backends in all databases have to be considered, and concurrently running
1699 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1700 : * accesses.
1701 : *
1702 : * Note: we include all currently running xids in the set of considered xids.
1703 : * This ensures that if a just-started xact has not yet set its snapshot,
1704 : * when it does set the snapshot it cannot set xmin less than what we compute.
1705 : * See notes in src/backend/access/transam/README.
1706 : *
1707 : * Note: despite the above, it's possible for the calculated values to move
1708 : * backwards on repeated calls. The calculated values are conservative, so
1709 : * that anything older is definitely not considered as running by anyone
1710 : * anymore, but the exact values calculated depend on a number of things. For
1711 : * example, if there are no transactions running in the current database, the
1712 : * horizon for normal tables will be latestCompletedXid. If a transaction
1713 : * begins after that, its xmin will include in-progress transactions in other
1714 : * databases that started earlier, so another call will return a lower value.
1715 : * Nonetheless it is safe to vacuum a table in the current database with the
1716 : * first result. There are also replication-related effects: a walsender
1717 : * process can set its xmin based on transactions that are no longer running
1718 : * on the primary but are still being replayed on the standby, thus possibly
1719 : * making the values go backwards. In this case there is a possibility that
1720 : * we lose data that the standby would like to have, but unless the standby
1721 : * uses a replication slot to make its xmin persistent there is little we can
1722 : * do about that --- data is only protected if the walsender runs continuously
1723 : * while queries are executed on the standby. (The Hot Standby code deals
1724 : * with such cases by failing standby queries that needed to access
1725 : * already-removed data, so there's no integrity bug.)
1726 : *
1727 : * Note: the approximate horizons (see definition of GlobalVisState) are
1728 : * updated by the computations done here. That's currently required for
1729 : * correctness and a small optimization. Without doing so it's possible that
1730 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1731 : * horizon than later when deciding which tuples can be removed - which the
1732 : * code doesn't expect (breaking HOT).
1733 : */
1734 : static void
1735 259044 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1736 : {
1737 259044 : ProcArrayStruct *arrayP = procArray;
1738 : TransactionId kaxmin;
1739 259044 : bool in_recovery = RecoveryInProgress();
1740 259044 : TransactionId *other_xids = ProcGlobal->xids;
1741 :
1742 : /* inferred after ProcArrayLock is released */
1743 259044 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1744 :
1745 259044 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1746 :
1747 259044 : h->latest_completed = TransamVariables->latestCompletedXid;
1748 :
1749 : /*
1750 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1751 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1752 : * and so protects us against overestimating the result due to future
1753 : * additions.
1754 : */
1755 : {
1756 : TransactionId initial;
1757 :
1758 259044 : initial = XidFromFullTransactionId(h->latest_completed);
1759 : Assert(TransactionIdIsValid(initial));
1760 259044 : TransactionIdAdvance(initial);
1761 :
1762 259044 : h->oldest_considered_running = initial;
1763 259044 : h->shared_oldest_nonremovable = initial;
1764 259044 : h->data_oldest_nonremovable = initial;
1765 :
1766 : /*
1767 : * Only modifications made by this backend affect the horizon for
1768 : * temporary relations. Instead of a check in each iteration of the
1769 : * loop over all PGPROCs it is cheaper to just initialize to the
1770 : * current top-level xid any.
1771 : *
1772 : * Without an assigned xid we could use a horizon as aggressive as
1773 : * GetNewTransactionId(), but we can get away with the much cheaper
1774 : * latestCompletedXid + 1: If this backend has no xid there, by
1775 : * definition, can't be any newer changes in the temp table than
1776 : * latestCompletedXid.
1777 : */
1778 259044 : if (TransactionIdIsValid(MyProc->xid))
1779 61758 : h->temp_oldest_nonremovable = MyProc->xid;
1780 : else
1781 197286 : h->temp_oldest_nonremovable = initial;
1782 : }
1783 :
1784 : /*
1785 : * Fetch slot horizons while ProcArrayLock is held - the
1786 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1787 : * the lock.
1788 : */
1789 259044 : h->slot_xmin = procArray->replication_slot_xmin;
1790 259044 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1791 :
1792 1401004 : for (int index = 0; index < arrayP->numProcs; index++)
1793 : {
1794 1141960 : int pgprocno = arrayP->pgprocnos[index];
1795 1141960 : PGPROC *proc = &allProcs[pgprocno];
1796 1141960 : int8 statusFlags = ProcGlobal->statusFlags[index];
1797 : TransactionId xid;
1798 : TransactionId xmin;
1799 :
1800 : /* Fetch xid just once - see GetNewTransactionId */
1801 1141960 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1802 1141960 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1803 :
1804 : /*
1805 : * Consider both the transaction's Xmin, and its Xid.
1806 : *
1807 : * We must check both because a transaction might have an Xmin but not
1808 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1809 : * some not-yet-set Xmin.
1810 : */
1811 1141960 : xmin = TransactionIdOlder(xmin, xid);
1812 :
1813 : /* if neither is set, this proc doesn't influence the horizon */
1814 1141960 : if (!TransactionIdIsValid(xmin))
1815 579520 : continue;
1816 :
1817 : /*
1818 : * Don't ignore any procs when determining which transactions might be
1819 : * considered running. While slots should ensure logical decoding
1820 : * backends are protected even without this check, it can't hurt to
1821 : * include them here as well..
1822 : */
1823 562440 : h->oldest_considered_running =
1824 562440 : TransactionIdOlder(h->oldest_considered_running, xmin);
1825 :
1826 : /*
1827 : * Skip over backends either vacuuming (which is ok with rows being
1828 : * removed, as long as pg_subtrans is not truncated) or doing logical
1829 : * decoding (which manages xmin separately, check below).
1830 : */
1831 562440 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1832 104060 : continue;
1833 :
1834 : /* shared tables need to take backends in all databases into account */
1835 458380 : h->shared_oldest_nonremovable =
1836 458380 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1837 :
1838 : /*
1839 : * Normally sessions in other databases are ignored for anything but
1840 : * the shared horizon.
1841 : *
1842 : * However, include them when MyDatabaseId is not (yet) set. A
1843 : * backend in the process of starting up must not compute a "too
1844 : * aggressive" horizon, otherwise we could end up using it to prune
1845 : * still-needed data away. If the current backend never connects to a
1846 : * database this is harmless, because data_oldest_nonremovable will
1847 : * never be utilized.
1848 : *
1849 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1850 : * be included. (This flag is used for hot standby feedback, which
1851 : * can't be tied to a specific database.)
1852 : *
1853 : * Also, while in recovery we cannot compute an accurate per-database
1854 : * horizon, as all xids are managed via the KnownAssignedXids
1855 : * machinery.
1856 : */
1857 458380 : if (proc->databaseId == MyDatabaseId ||
1858 32046 : MyDatabaseId == InvalidOid ||
1859 21766 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1860 : in_recovery)
1861 : {
1862 436614 : h->data_oldest_nonremovable =
1863 436614 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1864 : }
1865 : }
1866 :
1867 : /*
1868 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1869 : * after lock is released.
1870 : */
1871 259044 : if (in_recovery)
1872 682 : kaxmin = KnownAssignedXidsGetOldestXmin();
1873 :
1874 : /*
1875 : * No other information from shared state is needed, release the lock
1876 : * immediately. The rest of the computations can be done without a lock.
1877 : */
1878 259044 : LWLockRelease(ProcArrayLock);
1879 :
1880 259044 : if (in_recovery)
1881 : {
1882 682 : h->oldest_considered_running =
1883 682 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1884 682 : h->shared_oldest_nonremovable =
1885 682 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1886 682 : h->data_oldest_nonremovable =
1887 682 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1888 : /* temp relations cannot be accessed in recovery */
1889 : }
1890 :
1891 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1892 : h->shared_oldest_nonremovable));
1893 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1894 : h->data_oldest_nonremovable));
1895 :
1896 : /*
1897 : * Check whether there are replication slots requiring an older xmin.
1898 : */
1899 259044 : h->shared_oldest_nonremovable =
1900 259044 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1901 259044 : h->data_oldest_nonremovable =
1902 259044 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1903 :
1904 : /*
1905 : * The only difference between catalog / data horizons is that the slot's
1906 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1907 : * for logical decoding). Initialize with data horizon, and then back up
1908 : * further if necessary. Have to back up the shared horizon as well, since
1909 : * that also can contain catalogs.
1910 : */
1911 259044 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1912 259044 : h->shared_oldest_nonremovable =
1913 259044 : TransactionIdOlder(h->shared_oldest_nonremovable,
1914 : h->slot_catalog_xmin);
1915 259044 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1916 259044 : h->catalog_oldest_nonremovable =
1917 259044 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1918 : h->slot_catalog_xmin);
1919 :
1920 : /*
1921 : * It's possible that slots backed up the horizons further than
1922 : * oldest_considered_running. Fix.
1923 : */
1924 259044 : h->oldest_considered_running =
1925 259044 : TransactionIdOlder(h->oldest_considered_running,
1926 : h->shared_oldest_nonremovable);
1927 259044 : h->oldest_considered_running =
1928 259044 : TransactionIdOlder(h->oldest_considered_running,
1929 : h->catalog_oldest_nonremovable);
1930 259044 : h->oldest_considered_running =
1931 259044 : TransactionIdOlder(h->oldest_considered_running,
1932 : h->data_oldest_nonremovable);
1933 :
1934 : /*
1935 : * shared horizons have to be at least as old as the oldest visible in
1936 : * current db
1937 : */
1938 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1939 : h->data_oldest_nonremovable));
1940 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1941 : h->catalog_oldest_nonremovable));
1942 :
1943 : /*
1944 : * Horizons need to ensure that pg_subtrans access is still possible for
1945 : * the relevant backends.
1946 : */
1947 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1948 : h->shared_oldest_nonremovable));
1949 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1950 : h->catalog_oldest_nonremovable));
1951 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1952 : h->data_oldest_nonremovable));
1953 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1954 : h->temp_oldest_nonremovable));
1955 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1956 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1957 : h->slot_xmin));
1958 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1959 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1960 : h->slot_catalog_xmin));
1961 :
1962 : /* update approximate horizons with the computed horizons */
1963 259044 : GlobalVisUpdateApply(h);
1964 259044 : }
1965 :
1966 : /*
1967 : * Determine what kind of visibility horizon needs to be used for a
1968 : * relation. If rel is NULL, the most conservative horizon is used.
1969 : */
1970 : static inline GlobalVisHorizonKind
1971 27703910 : GlobalVisHorizonKindForRel(Relation rel)
1972 : {
1973 : /*
1974 : * Other relkinds currently don't contain xids, nor always the necessary
1975 : * logical decoding markers.
1976 : */
1977 : Assert(!rel ||
1978 : rel->rd_rel->relkind == RELKIND_RELATION ||
1979 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1980 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1981 :
1982 27703910 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1983 189606 : return VISHORIZON_SHARED;
1984 27514304 : else if (IsCatalogRelation(rel) ||
1985 22703454 : RelationIsAccessibleInLogicalDecoding(rel))
1986 4810858 : return VISHORIZON_CATALOG;
1987 22703446 : else if (!RELATION_IS_LOCAL(rel))
1988 22596952 : return VISHORIZON_DATA;
1989 : else
1990 106494 : return VISHORIZON_TEMP;
1991 : }
1992 :
1993 : /*
1994 : * Return the oldest XID for which deleted tuples must be preserved in the
1995 : * passed table.
1996 : *
1997 : * If rel is not NULL the horizon may be considerably more recent than
1998 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1999 : * that is correct (but not optimal) for all relations will be returned.
2000 : *
2001 : * This is used by VACUUM to decide which deleted tuples must be preserved in
2002 : * the passed in table.
2003 : */
2004 : TransactionId
2005 153328 : GetOldestNonRemovableTransactionId(Relation rel)
2006 : {
2007 : ComputeXidHorizonsResult horizons;
2008 :
2009 153328 : ComputeXidHorizons(&horizons);
2010 :
2011 153328 : switch (GlobalVisHorizonKindForRel(rel))
2012 : {
2013 21324 : case VISHORIZON_SHARED:
2014 21324 : return horizons.shared_oldest_nonremovable;
2015 79662 : case VISHORIZON_CATALOG:
2016 79662 : return horizons.catalog_oldest_nonremovable;
2017 29508 : case VISHORIZON_DATA:
2018 29508 : return horizons.data_oldest_nonremovable;
2019 22834 : case VISHORIZON_TEMP:
2020 22834 : return horizons.temp_oldest_nonremovable;
2021 : }
2022 :
2023 : /* just to prevent compiler warnings */
2024 0 : return InvalidTransactionId;
2025 : }
2026 :
2027 : /*
2028 : * Return the oldest transaction id any currently running backend might still
2029 : * consider running. This should not be used for visibility / pruning
2030 : * determinations (see GetOldestNonRemovableTransactionId()), but for
2031 : * decisions like up to where pg_subtrans can be truncated.
2032 : */
2033 : TransactionId
2034 2418 : GetOldestTransactionIdConsideredRunning(void)
2035 : {
2036 : ComputeXidHorizonsResult horizons;
2037 :
2038 2418 : ComputeXidHorizons(&horizons);
2039 :
2040 2418 : return horizons.oldest_considered_running;
2041 : }
2042 :
2043 : /*
2044 : * Return the visibility horizons for a hot standby feedback message.
2045 : */
2046 : void
2047 86 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
2048 : {
2049 : ComputeXidHorizonsResult horizons;
2050 :
2051 86 : ComputeXidHorizons(&horizons);
2052 :
2053 : /*
2054 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2055 : * effect of replication slot's catalog_xmin. We want to send a separate
2056 : * feedback for the catalog horizon, so the primary can remove data table
2057 : * contents more aggressively.
2058 : */
2059 86 : *xmin = horizons.shared_oldest_nonremovable_raw;
2060 86 : *catalog_xmin = horizons.slot_catalog_xmin;
2061 86 : }
2062 :
2063 : /*
2064 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2065 : *
2066 : * We have to export this for use by snapmgr.c.
2067 : */
2068 : int
2069 63490 : GetMaxSnapshotXidCount(void)
2070 : {
2071 63490 : return procArray->maxProcs;
2072 : }
2073 :
2074 : /*
2075 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2076 : *
2077 : * We have to export this for use by snapmgr.c.
2078 : */
2079 : int
2080 63162 : GetMaxSnapshotSubxidCount(void)
2081 : {
2082 63162 : return TOTAL_MAX_CACHED_SUBXIDS;
2083 : }
2084 :
2085 : /*
2086 : * Helper function for GetSnapshotData() that checks if the bulk of the
2087 : * visibility information in the snapshot is still valid. If so, it updates
2088 : * the fields that need to change and returns true. Otherwise it returns
2089 : * false.
2090 : *
2091 : * This very likely can be evolved to not need ProcArrayLock held (at very
2092 : * least in the case we already hold a snapshot), but that's for another day.
2093 : */
2094 : static bool
2095 3779964 : GetSnapshotDataReuse(Snapshot snapshot)
2096 : {
2097 : uint64 curXactCompletionCount;
2098 :
2099 : Assert(LWLockHeldByMe(ProcArrayLock));
2100 :
2101 3779964 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2102 63132 : return false;
2103 :
2104 3716832 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2105 3716832 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2106 630414 : return false;
2107 :
2108 : /*
2109 : * If the current xactCompletionCount is still the same as it was at the
2110 : * time the snapshot was built, we can be sure that rebuilding the
2111 : * contents of the snapshot the hard way would result in the same snapshot
2112 : * contents:
2113 : *
2114 : * As explained in transam/README, the set of xids considered running by
2115 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2116 : * contents only depend on transactions with xids and xactCompletionCount
2117 : * is incremented whenever a transaction with an xid finishes (while
2118 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2119 : * ensures we would detect if the snapshot would have changed.
2120 : *
2121 : * As the snapshot contents are the same as it was before, it is safe to
2122 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2123 : * visible under the snapshot could already have been removed (that'd
2124 : * require the set of running transactions to change) and it fulfills the
2125 : * requirement that concurrent GetSnapshotData() calls yield the same
2126 : * xmin.
2127 : */
2128 3086418 : if (!TransactionIdIsValid(MyProc->xmin))
2129 1091178 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2130 :
2131 3086418 : RecentXmin = snapshot->xmin;
2132 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2133 :
2134 3086418 : snapshot->curcid = GetCurrentCommandId(false);
2135 3086418 : snapshot->active_count = 0;
2136 3086418 : snapshot->regd_count = 0;
2137 3086418 : snapshot->copied = false;
2138 3086418 : snapshot->lsn = InvalidXLogRecPtr;
2139 3086418 : snapshot->whenTaken = 0;
2140 :
2141 3086418 : return true;
2142 : }
2143 :
2144 : /*
2145 : * GetSnapshotData -- returns information about running transactions.
2146 : *
2147 : * The returned snapshot includes xmin (lowest still-running xact ID),
2148 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2149 : * in the range xmin <= xid < xmax. It is used as follows:
2150 : * All xact IDs < xmin are considered finished.
2151 : * All xact IDs >= xmax are considered still running.
2152 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2153 : * it is considered running or not.
2154 : * This ensures that the set of transactions seen as "running" by the
2155 : * current xact will not change after it takes the snapshot.
2156 : *
2157 : * All running top-level XIDs are included in the snapshot, except for lazy
2158 : * VACUUM processes. We also try to include running subtransaction XIDs,
2159 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2160 : * information may not be available. If we find any overflowed subxid arrays,
2161 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2162 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2163 : *
2164 : * We also update the following backend-global variables:
2165 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2166 : * current transaction (this is the same as MyProc->xmin).
2167 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2168 : * older than this are known not running any more.
2169 : *
2170 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2171 : * for the benefit of the GlobalVisTest* family of functions.
2172 : *
2173 : * Note: this function should probably not be called with an argument that's
2174 : * not statically allocated (see xip allocation below).
2175 : */
2176 : Snapshot
2177 3779964 : GetSnapshotData(Snapshot snapshot)
2178 : {
2179 3779964 : ProcArrayStruct *arrayP = procArray;
2180 3779964 : TransactionId *other_xids = ProcGlobal->xids;
2181 : TransactionId xmin;
2182 : TransactionId xmax;
2183 3779964 : int count = 0;
2184 3779964 : int subcount = 0;
2185 3779964 : bool suboverflowed = false;
2186 : FullTransactionId latest_completed;
2187 : TransactionId oldestxid;
2188 : int mypgxactoff;
2189 : TransactionId myxid;
2190 : uint64 curXactCompletionCount;
2191 :
2192 3779964 : TransactionId replication_slot_xmin = InvalidTransactionId;
2193 3779964 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2194 :
2195 : Assert(snapshot != NULL);
2196 :
2197 : /*
2198 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2199 : * be sufficient. But it seems better to do the malloc while not holding
2200 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2201 : * more subxip storage than is probably needed.
2202 : *
2203 : * This does open a possibility for avoiding repeated malloc/free: since
2204 : * maxProcs does not change at runtime, we can simply reuse the previous
2205 : * xip arrays if any. (This relies on the fact that all callers pass
2206 : * static SnapshotData structs.)
2207 : */
2208 3779964 : if (snapshot->xip == NULL)
2209 : {
2210 : /*
2211 : * First call for this snapshot. Snapshot is same size whether or not
2212 : * we are in recovery, see later comments.
2213 : */
2214 63114 : snapshot->xip = (TransactionId *)
2215 63114 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2216 63114 : if (snapshot->xip == NULL)
2217 0 : ereport(ERROR,
2218 : (errcode(ERRCODE_OUT_OF_MEMORY),
2219 : errmsg("out of memory")));
2220 : Assert(snapshot->subxip == NULL);
2221 63114 : snapshot->subxip = (TransactionId *)
2222 63114 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2223 63114 : if (snapshot->subxip == NULL)
2224 0 : ereport(ERROR,
2225 : (errcode(ERRCODE_OUT_OF_MEMORY),
2226 : errmsg("out of memory")));
2227 : }
2228 :
2229 : /*
2230 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2231 : * going to set MyProc->xmin.
2232 : */
2233 3779964 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2234 :
2235 3779964 : if (GetSnapshotDataReuse(snapshot))
2236 : {
2237 3086418 : LWLockRelease(ProcArrayLock);
2238 3086418 : return snapshot;
2239 : }
2240 :
2241 693546 : latest_completed = TransamVariables->latestCompletedXid;
2242 693546 : mypgxactoff = MyProc->pgxactoff;
2243 693546 : myxid = other_xids[mypgxactoff];
2244 : Assert(myxid == MyProc->xid);
2245 :
2246 693546 : oldestxid = TransamVariables->oldestXid;
2247 693546 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2248 :
2249 : /* xmax is always latestCompletedXid + 1 */
2250 693546 : xmax = XidFromFullTransactionId(latest_completed);
2251 693546 : TransactionIdAdvance(xmax);
2252 : Assert(TransactionIdIsNormal(xmax));
2253 :
2254 : /* initialize xmin calculation with xmax */
2255 693546 : xmin = xmax;
2256 :
2257 : /* take own xid into account, saves a check inside the loop */
2258 693546 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2259 43360 : xmin = myxid;
2260 :
2261 693546 : snapshot->takenDuringRecovery = RecoveryInProgress();
2262 :
2263 693546 : if (!snapshot->takenDuringRecovery)
2264 : {
2265 691384 : int numProcs = arrayP->numProcs;
2266 691384 : TransactionId *xip = snapshot->xip;
2267 691384 : int *pgprocnos = arrayP->pgprocnos;
2268 691384 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2269 691384 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2270 :
2271 : /*
2272 : * First collect set of pgxactoff/xids that need to be included in the
2273 : * snapshot.
2274 : */
2275 4934612 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2276 : {
2277 : /* Fetch xid just once - see GetNewTransactionId */
2278 4243228 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2279 : uint8 statusFlags;
2280 :
2281 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2282 :
2283 : /*
2284 : * If the transaction has no XID assigned, we can skip it; it
2285 : * won't have sub-XIDs either.
2286 : */
2287 4243228 : if (likely(xid == InvalidTransactionId))
2288 3306052 : continue;
2289 :
2290 : /*
2291 : * We don't include our own XIDs (if any) in the snapshot. It
2292 : * needs to be included in the xmin computation, but we did so
2293 : * outside the loop.
2294 : */
2295 937176 : if (pgxactoff == mypgxactoff)
2296 85400 : continue;
2297 :
2298 : /*
2299 : * The only way we are able to get here with a non-normal xid is
2300 : * during bootstrap - with this backend using
2301 : * BootstrapTransactionId. But the above test should filter that
2302 : * out.
2303 : */
2304 : Assert(TransactionIdIsNormal(xid));
2305 :
2306 : /*
2307 : * If the XID is >= xmax, we can skip it; such transactions will
2308 : * be treated as running anyway (and any sub-XIDs will also be >=
2309 : * xmax).
2310 : */
2311 851776 : if (!NormalTransactionIdPrecedes(xid, xmax))
2312 193872 : continue;
2313 :
2314 : /*
2315 : * Skip over backends doing logical decoding which manages xmin
2316 : * separately (check below) and ones running LAZY VACUUM.
2317 : */
2318 657904 : statusFlags = allStatusFlags[pgxactoff];
2319 657904 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2320 58 : continue;
2321 :
2322 657846 : if (NormalTransactionIdPrecedes(xid, xmin))
2323 394338 : xmin = xid;
2324 :
2325 : /* Add XID to snapshot. */
2326 657846 : xip[count++] = xid;
2327 :
2328 : /*
2329 : * Save subtransaction XIDs if possible (if we've already
2330 : * overflowed, there's no point). Note that the subxact XIDs must
2331 : * be later than their parent, so no need to check them against
2332 : * xmin. We could filter against xmax, but it seems better not to
2333 : * do that much work while holding the ProcArrayLock.
2334 : *
2335 : * The other backend can add more subxids concurrently, but cannot
2336 : * remove any. Hence it's important to fetch nxids just once.
2337 : * Should be safe to use memcpy, though. (We needn't worry about
2338 : * missing any xids added concurrently, because they must postdate
2339 : * xmax.)
2340 : *
2341 : * Again, our own XIDs are not included in the snapshot.
2342 : */
2343 657846 : if (!suboverflowed)
2344 : {
2345 :
2346 657838 : if (subxidStates[pgxactoff].overflowed)
2347 548 : suboverflowed = true;
2348 : else
2349 : {
2350 657290 : int nsubxids = subxidStates[pgxactoff].count;
2351 :
2352 657290 : if (nsubxids > 0)
2353 : {
2354 11140 : int pgprocno = pgprocnos[pgxactoff];
2355 11140 : PGPROC *proc = &allProcs[pgprocno];
2356 :
2357 11140 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2358 :
2359 11140 : memcpy(snapshot->subxip + subcount,
2360 11140 : proc->subxids.xids,
2361 : nsubxids * sizeof(TransactionId));
2362 11140 : subcount += nsubxids;
2363 : }
2364 : }
2365 : }
2366 : }
2367 : }
2368 : else
2369 : {
2370 : /*
2371 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2372 : *
2373 : * We store all xids directly into subxip[]. Here's why:
2374 : *
2375 : * In recovery we don't know which xids are top-level and which are
2376 : * subxacts, a design choice that greatly simplifies xid processing.
2377 : *
2378 : * It seems like we would want to try to put xids into xip[] only, but
2379 : * that is fairly small. We would either need to make that bigger or
2380 : * to increase the rate at which we WAL-log xid assignment; neither is
2381 : * an appealing choice.
2382 : *
2383 : * We could try to store xids into xip[] first and then into subxip[]
2384 : * if there are too many xids. That only works if the snapshot doesn't
2385 : * overflow because we do not search subxip[] in that case. A simpler
2386 : * way is to just store all xids in the subxip array because this is
2387 : * by far the bigger array. We just leave the xip array empty.
2388 : *
2389 : * Either way we need to change the way XidInMVCCSnapshot() works
2390 : * depending upon when the snapshot was taken, or change normal
2391 : * snapshot processing so it matches.
2392 : *
2393 : * Note: It is possible for recovery to end before we finish taking
2394 : * the snapshot, and for newly assigned transaction ids to be added to
2395 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2396 : * those newly added transaction ids would be filtered away, so we
2397 : * need not be concerned about them.
2398 : */
2399 2162 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2400 : xmax);
2401 :
2402 2162 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2403 12 : suboverflowed = true;
2404 : }
2405 :
2406 :
2407 : /*
2408 : * Fetch into local variable while ProcArrayLock is held - the
2409 : * LWLockRelease below is a barrier, ensuring this happens inside the
2410 : * lock.
2411 : */
2412 693546 : replication_slot_xmin = procArray->replication_slot_xmin;
2413 693546 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2414 :
2415 693546 : if (!TransactionIdIsValid(MyProc->xmin))
2416 351332 : MyProc->xmin = TransactionXmin = xmin;
2417 :
2418 693546 : LWLockRelease(ProcArrayLock);
2419 :
2420 : /* maintain state for GlobalVis* */
2421 : {
2422 : TransactionId def_vis_xid;
2423 : TransactionId def_vis_xid_data;
2424 : FullTransactionId def_vis_fxid;
2425 : FullTransactionId def_vis_fxid_data;
2426 : FullTransactionId oldestfxid;
2427 :
2428 : /*
2429 : * Converting oldestXid is only safe when xid horizon cannot advance,
2430 : * i.e. holding locks. While we don't hold the lock anymore, all the
2431 : * necessary data has been gathered with lock held.
2432 : */
2433 693546 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2434 :
2435 : /* Check whether there's a replication slot requiring an older xmin. */
2436 : def_vis_xid_data =
2437 693546 : TransactionIdOlder(xmin, replication_slot_xmin);
2438 :
2439 : /*
2440 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2441 : * if older than this xid.
2442 : */
2443 693546 : def_vis_xid = def_vis_xid_data;
2444 :
2445 : /*
2446 : * Check whether there's a replication slot requiring an older catalog
2447 : * xmin.
2448 : */
2449 : def_vis_xid =
2450 693546 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2451 :
2452 693546 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2453 693546 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2454 :
2455 : /*
2456 : * Check if we can increase upper bound. As a previous
2457 : * GlobalVisUpdate() might have computed more aggressive values, don't
2458 : * overwrite them if so.
2459 : */
2460 : GlobalVisSharedRels.definitely_needed =
2461 693546 : FullTransactionIdNewer(def_vis_fxid,
2462 : GlobalVisSharedRels.definitely_needed);
2463 : GlobalVisCatalogRels.definitely_needed =
2464 693546 : FullTransactionIdNewer(def_vis_fxid,
2465 : GlobalVisCatalogRels.definitely_needed);
2466 : GlobalVisDataRels.definitely_needed =
2467 693546 : FullTransactionIdNewer(def_vis_fxid_data,
2468 : GlobalVisDataRels.definitely_needed);
2469 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2470 693546 : if (TransactionIdIsNormal(myxid))
2471 : GlobalVisTempRels.definitely_needed =
2472 85220 : FullXidRelativeTo(latest_completed, myxid);
2473 : else
2474 : {
2475 608326 : GlobalVisTempRels.definitely_needed = latest_completed;
2476 608326 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2477 : }
2478 :
2479 : /*
2480 : * Check if we know that we can initialize or increase the lower
2481 : * bound. Currently the only cheap way to do so is to use
2482 : * TransamVariables->oldestXid as input.
2483 : *
2484 : * We should definitely be able to do better. We could e.g. put a
2485 : * global lower bound value into TransamVariables.
2486 : */
2487 : GlobalVisSharedRels.maybe_needed =
2488 693546 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2489 : oldestfxid);
2490 : GlobalVisCatalogRels.maybe_needed =
2491 693546 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2492 : oldestfxid);
2493 : GlobalVisDataRels.maybe_needed =
2494 693546 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2495 : oldestfxid);
2496 : /* accurate value known */
2497 693546 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2498 : }
2499 :
2500 693546 : RecentXmin = xmin;
2501 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2502 :
2503 693546 : snapshot->xmin = xmin;
2504 693546 : snapshot->xmax = xmax;
2505 693546 : snapshot->xcnt = count;
2506 693546 : snapshot->subxcnt = subcount;
2507 693546 : snapshot->suboverflowed = suboverflowed;
2508 693546 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2509 :
2510 693546 : snapshot->curcid = GetCurrentCommandId(false);
2511 :
2512 : /*
2513 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2514 : * not copied in persistent memory.
2515 : */
2516 693546 : snapshot->active_count = 0;
2517 693546 : snapshot->regd_count = 0;
2518 693546 : snapshot->copied = false;
2519 693546 : snapshot->lsn = InvalidXLogRecPtr;
2520 693546 : snapshot->whenTaken = 0;
2521 :
2522 693546 : return snapshot;
2523 : }
2524 :
2525 : /*
2526 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2527 : *
2528 : * This is called when installing a snapshot imported from another
2529 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2530 : * check that the source transaction is still running, and we'd better do
2531 : * that atomically with installing the new xmin.
2532 : *
2533 : * Returns true if successful, false if source xact is no longer running.
2534 : */
2535 : bool
2536 32 : ProcArrayInstallImportedXmin(TransactionId xmin,
2537 : VirtualTransactionId *sourcevxid)
2538 : {
2539 32 : bool result = false;
2540 32 : ProcArrayStruct *arrayP = procArray;
2541 : int index;
2542 :
2543 : Assert(TransactionIdIsNormal(xmin));
2544 32 : if (!sourcevxid)
2545 0 : return false;
2546 :
2547 : /* Get lock so source xact can't end while we're doing this */
2548 32 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2549 :
2550 : /*
2551 : * Find the PGPROC entry of the source transaction. (This could use
2552 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2553 : * performance critical.)
2554 : */
2555 32 : for (index = 0; index < arrayP->numProcs; index++)
2556 : {
2557 32 : int pgprocno = arrayP->pgprocnos[index];
2558 32 : PGPROC *proc = &allProcs[pgprocno];
2559 32 : int statusFlags = ProcGlobal->statusFlags[index];
2560 : TransactionId xid;
2561 :
2562 : /* Ignore procs running LAZY VACUUM */
2563 32 : if (statusFlags & PROC_IN_VACUUM)
2564 0 : continue;
2565 :
2566 : /* We are only interested in the specific virtual transaction. */
2567 32 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2568 0 : continue;
2569 32 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2570 0 : continue;
2571 :
2572 : /*
2573 : * We check the transaction's database ID for paranoia's sake: if it's
2574 : * in another DB then its xmin does not cover us. Caller should have
2575 : * detected this already, so we just treat any funny cases as
2576 : * "transaction not found".
2577 : */
2578 32 : if (proc->databaseId != MyDatabaseId)
2579 0 : continue;
2580 :
2581 : /*
2582 : * Likewise, let's just make real sure its xmin does cover us.
2583 : */
2584 32 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2585 32 : if (!TransactionIdIsNormal(xid) ||
2586 32 : !TransactionIdPrecedesOrEquals(xid, xmin))
2587 0 : continue;
2588 :
2589 : /*
2590 : * We're good. Install the new xmin. As in GetSnapshotData, set
2591 : * TransactionXmin too. (Note that because snapmgr.c called
2592 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2593 : * we don't check that.)
2594 : */
2595 32 : MyProc->xmin = TransactionXmin = xmin;
2596 :
2597 32 : result = true;
2598 32 : break;
2599 : }
2600 :
2601 32 : LWLockRelease(ProcArrayLock);
2602 :
2603 32 : return result;
2604 : }
2605 :
2606 : /*
2607 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2608 : *
2609 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2610 : * PGPROC of the transaction from which we imported the snapshot, rather than
2611 : * an XID.
2612 : *
2613 : * Note that this function also copies statusFlags from the source `proc` in
2614 : * order to avoid the case where MyProc's xmin needs to be skipped for
2615 : * computing xid horizon.
2616 : *
2617 : * Returns true if successful, false if source xact is no longer running.
2618 : */
2619 : bool
2620 3056 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2621 : {
2622 3056 : bool result = false;
2623 : TransactionId xid;
2624 :
2625 : Assert(TransactionIdIsNormal(xmin));
2626 : Assert(proc != NULL);
2627 :
2628 : /*
2629 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2630 : */
2631 3056 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2632 :
2633 : /*
2634 : * Be certain that the referenced PGPROC has an advertised xmin which is
2635 : * no later than the one we're installing, so that the system-wide xmin
2636 : * can't go backwards. Also, make sure it's running in the same database,
2637 : * so that the per-database xmin cannot go backwards.
2638 : */
2639 3056 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2640 3056 : if (proc->databaseId == MyDatabaseId &&
2641 3056 : TransactionIdIsNormal(xid) &&
2642 3056 : TransactionIdPrecedesOrEquals(xid, xmin))
2643 : {
2644 : /*
2645 : * Install xmin and propagate the statusFlags that affect how the
2646 : * value is interpreted by vacuum.
2647 : */
2648 3056 : MyProc->xmin = TransactionXmin = xmin;
2649 3056 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2650 3056 : (proc->statusFlags & PROC_XMIN_FLAGS);
2651 3056 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2652 :
2653 3056 : result = true;
2654 : }
2655 :
2656 3056 : LWLockRelease(ProcArrayLock);
2657 :
2658 3056 : return result;
2659 : }
2660 :
2661 : /*
2662 : * GetRunningTransactionData -- returns information about running transactions.
2663 : *
2664 : * Similar to GetSnapshotData but returns more information. We include
2665 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2666 : * prepared transactions.
2667 : *
2668 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2669 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2670 : * array until the caller has WAL-logged this snapshot, and releases the
2671 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2672 : * lock is released.
2673 : *
2674 : * The returned data structure is statically allocated; caller should not
2675 : * modify it, and must not assume it is valid past the next call.
2676 : *
2677 : * This is never executed during recovery so there is no need to look at
2678 : * KnownAssignedXids.
2679 : *
2680 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2681 : * may return entries with duplicated TransactionId values coming from
2682 : * transaction finishing to prepare. Nothing is done about duplicated
2683 : * entries here to not hold on ProcArrayLock more than necessary.
2684 : *
2685 : * We don't worry about updating other counters, we want to keep this as
2686 : * simple as possible and leave GetSnapshotData() as the primary code for
2687 : * that bookkeeping.
2688 : *
2689 : * Note that if any transaction has overflowed its cached subtransactions
2690 : * then there is no real need include any subtransactions.
2691 : */
2692 : RunningTransactions
2693 1906 : GetRunningTransactionData(void)
2694 : {
2695 : /* result workspace */
2696 : static RunningTransactionsData CurrentRunningXactsData;
2697 :
2698 1906 : ProcArrayStruct *arrayP = procArray;
2699 1906 : TransactionId *other_xids = ProcGlobal->xids;
2700 1906 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2701 : TransactionId latestCompletedXid;
2702 : TransactionId oldestRunningXid;
2703 : TransactionId oldestDatabaseRunningXid;
2704 : TransactionId *xids;
2705 : int index;
2706 : int count;
2707 : int subcount;
2708 : bool suboverflowed;
2709 :
2710 : Assert(!RecoveryInProgress());
2711 :
2712 : /*
2713 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2714 : * be sufficient. But it seems better to do the malloc while not holding
2715 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2716 : * more subxip storage than is probably needed.
2717 : *
2718 : * Should only be allocated in bgwriter, since only ever executed during
2719 : * checkpoints.
2720 : */
2721 1906 : if (CurrentRunningXacts->xids == NULL)
2722 : {
2723 : /*
2724 : * First call
2725 : */
2726 934 : CurrentRunningXacts->xids = (TransactionId *)
2727 934 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2728 934 : if (CurrentRunningXacts->xids == NULL)
2729 0 : ereport(ERROR,
2730 : (errcode(ERRCODE_OUT_OF_MEMORY),
2731 : errmsg("out of memory")));
2732 : }
2733 :
2734 1906 : xids = CurrentRunningXacts->xids;
2735 :
2736 1906 : count = subcount = 0;
2737 1906 : suboverflowed = false;
2738 :
2739 : /*
2740 : * Ensure that no xids enter or leave the procarray while we obtain
2741 : * snapshot.
2742 : */
2743 1906 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2744 1906 : LWLockAcquire(XidGenLock, LW_SHARED);
2745 :
2746 1906 : latestCompletedXid =
2747 1906 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2748 1906 : oldestDatabaseRunningXid = oldestRunningXid =
2749 1906 : XidFromFullTransactionId(TransamVariables->nextXid);
2750 :
2751 : /*
2752 : * Spin over procArray collecting all xids
2753 : */
2754 9414 : for (index = 0; index < arrayP->numProcs; index++)
2755 : {
2756 : TransactionId xid;
2757 :
2758 : /* Fetch xid just once - see GetNewTransactionId */
2759 7508 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2760 :
2761 : /*
2762 : * We don't need to store transactions that don't have a TransactionId
2763 : * yet because they will not show as running on a standby server.
2764 : */
2765 7508 : if (!TransactionIdIsValid(xid))
2766 6024 : continue;
2767 :
2768 : /*
2769 : * Be careful not to exclude any xids before calculating the values of
2770 : * oldestRunningXid and suboverflowed, since these are used to clean
2771 : * up transaction information held on standbys.
2772 : */
2773 1484 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2774 1046 : oldestRunningXid = xid;
2775 :
2776 : /*
2777 : * Also, update the oldest running xid within the current database. As
2778 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2779 : * TransactionId comparison first.
2780 : */
2781 1484 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2782 : {
2783 1484 : int pgprocno = arrayP->pgprocnos[index];
2784 1484 : PGPROC *proc = &allProcs[pgprocno];
2785 :
2786 1484 : if (proc->databaseId == MyDatabaseId)
2787 376 : oldestDatabaseRunningXid = xid;
2788 : }
2789 :
2790 1484 : if (ProcGlobal->subxidStates[index].overflowed)
2791 4 : suboverflowed = true;
2792 :
2793 : /*
2794 : * If we wished to exclude xids this would be the right place for it.
2795 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2796 : * but they do during truncation at the end when they get the lock and
2797 : * truncate, so it is not much of a problem to include them if they
2798 : * are seen and it is cleaner to include them.
2799 : */
2800 :
2801 1484 : xids[count++] = xid;
2802 : }
2803 :
2804 : /*
2805 : * Spin over procArray collecting all subxids, but only if there hasn't
2806 : * been a suboverflow.
2807 : */
2808 1906 : if (!suboverflowed)
2809 : {
2810 1902 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2811 :
2812 9402 : for (index = 0; index < arrayP->numProcs; index++)
2813 : {
2814 7500 : int pgprocno = arrayP->pgprocnos[index];
2815 7500 : PGPROC *proc = &allProcs[pgprocno];
2816 : int nsubxids;
2817 :
2818 : /*
2819 : * Save subtransaction XIDs. Other backends can't add or remove
2820 : * entries while we're holding XidGenLock.
2821 : */
2822 7500 : nsubxids = other_subxidstates[index].count;
2823 7500 : if (nsubxids > 0)
2824 : {
2825 : /* barrier not really required, as XidGenLock is held, but ... */
2826 50 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2827 :
2828 50 : memcpy(&xids[count], proc->subxids.xids,
2829 : nsubxids * sizeof(TransactionId));
2830 50 : count += nsubxids;
2831 50 : subcount += nsubxids;
2832 :
2833 : /*
2834 : * Top-level XID of a transaction is always less than any of
2835 : * its subxids, so we don't need to check if any of the
2836 : * subxids are smaller than oldestRunningXid
2837 : */
2838 : }
2839 : }
2840 : }
2841 :
2842 : /*
2843 : * It's important *not* to include the limits set by slots here because
2844 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2845 : * were to be included here the initial value could never increase because
2846 : * of a circular dependency where slots only increase their limits when
2847 : * running xacts increases oldestRunningXid and running xacts only
2848 : * increases if slots do.
2849 : */
2850 :
2851 1906 : CurrentRunningXacts->xcnt = count - subcount;
2852 1906 : CurrentRunningXacts->subxcnt = subcount;
2853 1906 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2854 1906 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2855 1906 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2856 1906 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2857 1906 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2858 :
2859 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2860 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2861 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2862 :
2863 : /* We don't release the locks here, the caller is responsible for that */
2864 :
2865 1906 : return CurrentRunningXacts;
2866 : }
2867 :
2868 : /*
2869 : * GetOldestActiveTransactionId()
2870 : *
2871 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2872 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2873 : * We look at all databases, though there is no need to include WALSender
2874 : * since this has no effect on hot standby conflicts.
2875 : *
2876 : * This is never executed during recovery so there is no need to look at
2877 : * KnownAssignedXids.
2878 : *
2879 : * We don't worry about updating other counters, we want to keep this as
2880 : * simple as possible and leave GetSnapshotData() as the primary code for
2881 : * that bookkeeping.
2882 : */
2883 : TransactionId
2884 1000 : GetOldestActiveTransactionId(void)
2885 : {
2886 1000 : ProcArrayStruct *arrayP = procArray;
2887 1000 : TransactionId *other_xids = ProcGlobal->xids;
2888 : TransactionId oldestRunningXid;
2889 : int index;
2890 :
2891 : Assert(!RecoveryInProgress());
2892 :
2893 : /*
2894 : * Read nextXid, as the upper bound of what's still active.
2895 : *
2896 : * Reading a TransactionId is atomic, but we must grab the lock to make
2897 : * sure that all XIDs < nextXid are already present in the proc array (or
2898 : * have already completed), when we spin over it.
2899 : */
2900 1000 : LWLockAcquire(XidGenLock, LW_SHARED);
2901 1000 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2902 1000 : LWLockRelease(XidGenLock);
2903 :
2904 : /*
2905 : * Spin over procArray collecting all xids and subxids.
2906 : */
2907 1000 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2908 4874 : for (index = 0; index < arrayP->numProcs; index++)
2909 : {
2910 : TransactionId xid;
2911 :
2912 : /* Fetch xid just once - see GetNewTransactionId */
2913 3874 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2914 :
2915 3874 : if (!TransactionIdIsNormal(xid))
2916 2466 : continue;
2917 :
2918 1408 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2919 984 : oldestRunningXid = xid;
2920 :
2921 : /*
2922 : * Top-level XID of a transaction is always less than any of its
2923 : * subxids, so we don't need to check if any of the subxids are
2924 : * smaller than oldestRunningXid
2925 : */
2926 : }
2927 1000 : LWLockRelease(ProcArrayLock);
2928 :
2929 1000 : return oldestRunningXid;
2930 : }
2931 :
2932 : /*
2933 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2934 : *
2935 : * Returns the oldest xid that we can guarantee not to have been affected by
2936 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2937 : * transaction aborted. Note that the value can (and most of the time will) be
2938 : * much more conservative than what really has been affected by vacuum, but we
2939 : * currently don't have better data available.
2940 : *
2941 : * This is useful to initialize the cutoff xid after which a new changeset
2942 : * extraction replication slot can start decoding changes.
2943 : *
2944 : * Must be called with ProcArrayLock held either shared or exclusively,
2945 : * although most callers will want to use exclusive mode since it is expected
2946 : * that the caller will immediately use the xid to peg the xmin horizon.
2947 : */
2948 : TransactionId
2949 1168 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2950 : {
2951 1168 : ProcArrayStruct *arrayP = procArray;
2952 : TransactionId oldestSafeXid;
2953 : int index;
2954 1168 : bool recovery_in_progress = RecoveryInProgress();
2955 :
2956 : Assert(LWLockHeldByMe(ProcArrayLock));
2957 :
2958 : /*
2959 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2960 : * running. If no transaction with xid were running concurrently a new xid
2961 : * could influence the RecentXmin et al.
2962 : *
2963 : * We initialize the computation to nextXid since that's guaranteed to be
2964 : * a safe, albeit pessimal, value.
2965 : */
2966 1168 : LWLockAcquire(XidGenLock, LW_SHARED);
2967 1168 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2968 :
2969 : /*
2970 : * If there's already a slot pegging the xmin horizon, we can start with
2971 : * that value, it's guaranteed to be safe since it's computed by this
2972 : * routine initially and has been enforced since. We can always use the
2973 : * slot's general xmin horizon, but the catalog horizon is only usable
2974 : * when only catalog data is going to be looked at.
2975 : */
2976 1536 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2977 368 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2978 : oldestSafeXid))
2979 20 : oldestSafeXid = procArray->replication_slot_xmin;
2980 :
2981 1168 : if (catalogOnly &&
2982 608 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2983 128 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2984 : oldestSafeXid))
2985 50 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2986 :
2987 : /*
2988 : * If we're not in recovery, we walk over the procarray and collect the
2989 : * lowest xid. Since we're called with ProcArrayLock held and have
2990 : * acquired XidGenLock, no entries can vanish concurrently, since
2991 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2992 : * with ProcArrayLock held.
2993 : *
2994 : * In recovery we can't lower the safe value besides what we've computed
2995 : * above, so we'll have to wait a bit longer there. We unfortunately can
2996 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2997 : * machinery can miss values and return an older value than is safe.
2998 : */
2999 1168 : if (!recovery_in_progress)
3000 : {
3001 1114 : TransactionId *other_xids = ProcGlobal->xids;
3002 :
3003 : /*
3004 : * Spin over procArray collecting min(ProcGlobal->xids[i])
3005 : */
3006 5592 : for (index = 0; index < arrayP->numProcs; index++)
3007 : {
3008 : TransactionId xid;
3009 :
3010 : /* Fetch xid just once - see GetNewTransactionId */
3011 4478 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
3012 :
3013 4478 : if (!TransactionIdIsNormal(xid))
3014 4462 : continue;
3015 :
3016 16 : if (TransactionIdPrecedes(xid, oldestSafeXid))
3017 14 : oldestSafeXid = xid;
3018 : }
3019 : }
3020 :
3021 1168 : LWLockRelease(XidGenLock);
3022 :
3023 1168 : return oldestSafeXid;
3024 : }
3025 :
3026 : /*
3027 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
3028 : * delaying checkpoint because they have critical actions in progress.
3029 : *
3030 : * Constructs an array of VXIDs of transactions that are currently in commit
3031 : * critical sections, as shown by having specified delayChkptFlags bits set
3032 : * in their PGPROC.
3033 : *
3034 : * Returns a palloc'd array that should be freed by the caller.
3035 : * *nvxids is the number of valid entries.
3036 : *
3037 : * Note that because backends set or clear delayChkptFlags without holding any
3038 : * lock, the result is somewhat indeterminate, but we don't really care. Even
3039 : * in a multiprocessor with delayed writes to shared memory, it should be
3040 : * certain that setting of delayChkptFlags will propagate to shared memory
3041 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
3042 : * delayChkptFlags if it's already inserted its commit record. Whether it
3043 : * takes a little while for clearing of delayChkptFlags to propagate is
3044 : * unimportant for correctness.
3045 : */
3046 : VirtualTransactionId *
3047 4212 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3048 : {
3049 : VirtualTransactionId *vxids;
3050 4212 : ProcArrayStruct *arrayP = procArray;
3051 4212 : int count = 0;
3052 : int index;
3053 :
3054 : Assert(type != 0);
3055 :
3056 : /* allocate what's certainly enough result space */
3057 : vxids = (VirtualTransactionId *)
3058 4212 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3059 :
3060 4212 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3061 :
3062 13224 : for (index = 0; index < arrayP->numProcs; index++)
3063 : {
3064 9012 : int pgprocno = arrayP->pgprocnos[index];
3065 9012 : PGPROC *proc = &allProcs[pgprocno];
3066 :
3067 9012 : if ((proc->delayChkptFlags & type) != 0)
3068 : {
3069 : VirtualTransactionId vxid;
3070 :
3071 50 : GET_VXID_FROM_PGPROC(vxid, *proc);
3072 50 : if (VirtualTransactionIdIsValid(vxid))
3073 50 : vxids[count++] = vxid;
3074 : }
3075 : }
3076 :
3077 4212 : LWLockRelease(ProcArrayLock);
3078 :
3079 4212 : *nvxids = count;
3080 4212 : return vxids;
3081 : }
3082 :
3083 : /*
3084 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3085 : *
3086 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3087 : * of the specified VXIDs are still in critical sections of code.
3088 : *
3089 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3090 : * those numbers should be small enough for it not to be a problem.
3091 : */
3092 : bool
3093 58 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3094 : {
3095 58 : bool result = false;
3096 58 : ProcArrayStruct *arrayP = procArray;
3097 : int index;
3098 :
3099 : Assert(type != 0);
3100 :
3101 58 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3102 :
3103 486 : for (index = 0; index < arrayP->numProcs; index++)
3104 : {
3105 446 : int pgprocno = arrayP->pgprocnos[index];
3106 446 : PGPROC *proc = &allProcs[pgprocno];
3107 : VirtualTransactionId vxid;
3108 :
3109 446 : GET_VXID_FROM_PGPROC(vxid, *proc);
3110 :
3111 446 : if ((proc->delayChkptFlags & type) != 0 &&
3112 52 : VirtualTransactionIdIsValid(vxid))
3113 : {
3114 : int i;
3115 :
3116 102 : for (i = 0; i < nvxids; i++)
3117 : {
3118 68 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3119 : {
3120 18 : result = true;
3121 18 : break;
3122 : }
3123 : }
3124 52 : if (result)
3125 18 : break;
3126 : }
3127 : }
3128 :
3129 58 : LWLockRelease(ProcArrayLock);
3130 :
3131 58 : return result;
3132 : }
3133 :
3134 : /*
3135 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3136 : *
3137 : * The result may be out of date arbitrarily quickly, so the caller
3138 : * must be careful about how this information is used. NULL is
3139 : * returned if the backend is not active.
3140 : */
3141 : PGPROC *
3142 1056 : ProcNumberGetProc(ProcNumber procNumber)
3143 : {
3144 : PGPROC *result;
3145 :
3146 1056 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3147 2 : return NULL;
3148 1054 : result = GetPGProcByNumber(procNumber);
3149 :
3150 1054 : if (result->pid == 0)
3151 10 : return NULL;
3152 :
3153 1044 : return result;
3154 : }
3155 :
3156 : /*
3157 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3158 : *
3159 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3160 : * result may be out of date arbitrarily quickly, so the caller must be
3161 : * careful about how this information is used.
3162 : */
3163 : void
3164 11450 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3165 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3166 : {
3167 : PGPROC *proc;
3168 :
3169 11450 : *xid = InvalidTransactionId;
3170 11450 : *xmin = InvalidTransactionId;
3171 11450 : *nsubxid = 0;
3172 11450 : *overflowed = false;
3173 :
3174 11450 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3175 0 : return;
3176 11450 : proc = GetPGProcByNumber(procNumber);
3177 :
3178 : /* Need to lock out additions/removals of backends */
3179 11450 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3180 :
3181 11450 : if (proc->pid != 0)
3182 : {
3183 11450 : *xid = proc->xid;
3184 11450 : *xmin = proc->xmin;
3185 11450 : *nsubxid = proc->subxidStatus.count;
3186 11450 : *overflowed = proc->subxidStatus.overflowed;
3187 : }
3188 :
3189 11450 : LWLockRelease(ProcArrayLock);
3190 : }
3191 :
3192 : /*
3193 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3194 : *
3195 : * Returns NULL if not found. Note that it is up to the caller to be
3196 : * sure that the question remains meaningful for long enough for the
3197 : * answer to be used ...
3198 : */
3199 : PGPROC *
3200 14184 : BackendPidGetProc(int pid)
3201 : {
3202 : PGPROC *result;
3203 :
3204 14184 : if (pid == 0) /* never match dummy PGPROCs */
3205 0 : return NULL;
3206 :
3207 14184 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3208 :
3209 14184 : result = BackendPidGetProcWithLock(pid);
3210 :
3211 14184 : LWLockRelease(ProcArrayLock);
3212 :
3213 14184 : return result;
3214 : }
3215 :
3216 : /*
3217 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3218 : *
3219 : * Same as above, except caller must be holding ProcArrayLock. The found
3220 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3221 : */
3222 : PGPROC *
3223 17668 : BackendPidGetProcWithLock(int pid)
3224 : {
3225 17668 : PGPROC *result = NULL;
3226 17668 : ProcArrayStruct *arrayP = procArray;
3227 : int index;
3228 :
3229 17668 : if (pid == 0) /* never match dummy PGPROCs */
3230 0 : return NULL;
3231 :
3232 60454 : for (index = 0; index < arrayP->numProcs; index++)
3233 : {
3234 56110 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3235 :
3236 56110 : if (proc->pid == pid)
3237 : {
3238 13324 : result = proc;
3239 13324 : break;
3240 : }
3241 : }
3242 :
3243 17668 : return result;
3244 : }
3245 :
3246 : /*
3247 : * BackendXidGetPid -- get a backend's pid given its XID
3248 : *
3249 : * Returns 0 if not found or it's a prepared transaction. Note that
3250 : * it is up to the caller to be sure that the question remains
3251 : * meaningful for long enough for the answer to be used ...
3252 : *
3253 : * Only main transaction Ids are considered. This function is mainly
3254 : * useful for determining what backend owns a lock.
3255 : *
3256 : * Beware that not every xact has an XID assigned. However, as long as you
3257 : * only call this using an XID found on disk, you're safe.
3258 : */
3259 : int
3260 60 : BackendXidGetPid(TransactionId xid)
3261 : {
3262 60 : int result = 0;
3263 60 : ProcArrayStruct *arrayP = procArray;
3264 60 : TransactionId *other_xids = ProcGlobal->xids;
3265 : int index;
3266 :
3267 60 : if (xid == InvalidTransactionId) /* never match invalid xid */
3268 0 : return 0;
3269 :
3270 60 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3271 :
3272 184 : for (index = 0; index < arrayP->numProcs; index++)
3273 : {
3274 168 : if (other_xids[index] == xid)
3275 : {
3276 44 : int pgprocno = arrayP->pgprocnos[index];
3277 44 : PGPROC *proc = &allProcs[pgprocno];
3278 :
3279 44 : result = proc->pid;
3280 44 : break;
3281 : }
3282 : }
3283 :
3284 60 : LWLockRelease(ProcArrayLock);
3285 :
3286 60 : return result;
3287 : }
3288 :
3289 : /*
3290 : * IsBackendPid -- is a given pid a running backend
3291 : *
3292 : * This is not called by the backend, but is called by external modules.
3293 : */
3294 : bool
3295 4 : IsBackendPid(int pid)
3296 : {
3297 4 : return (BackendPidGetProc(pid) != NULL);
3298 : }
3299 :
3300 :
3301 : /*
3302 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3303 : *
3304 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3305 : *
3306 : * The arguments allow filtering the set of VXIDs returned. Our own process
3307 : * is always skipped. In addition:
3308 : * If limitXmin is not InvalidTransactionId, skip processes with
3309 : * xmin > limitXmin.
3310 : * If excludeXmin0 is true, skip processes with xmin = 0.
3311 : * If allDbs is false, skip processes attached to other databases.
3312 : * If excludeVacuum isn't zero, skip processes for which
3313 : * (statusFlags & excludeVacuum) is not zero.
3314 : *
3315 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3316 : * allow skipping backends whose oldest live snapshot is no older than
3317 : * some snapshot we have. Since we examine the procarray with only shared
3318 : * lock, there are race conditions: a backend could set its xmin just after
3319 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3320 : * other backend could have set its xmin *before* we look. We know however
3321 : * that such a backend must have held shared ProcArrayLock overlapping our
3322 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3323 : * any snapshot the other backend is taking concurrently with our scan cannot
3324 : * consider any transactions as still running that we think are committed
3325 : * (since backends must hold ProcArrayLock exclusive to commit).
3326 : */
3327 : VirtualTransactionId *
3328 712 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3329 : bool allDbs, int excludeVacuum,
3330 : int *nvxids)
3331 : {
3332 : VirtualTransactionId *vxids;
3333 712 : ProcArrayStruct *arrayP = procArray;
3334 712 : int count = 0;
3335 : int index;
3336 :
3337 : /* allocate what's certainly enough result space */
3338 : vxids = (VirtualTransactionId *)
3339 712 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3340 :
3341 712 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3342 :
3343 4242 : for (index = 0; index < arrayP->numProcs; index++)
3344 : {
3345 3530 : int pgprocno = arrayP->pgprocnos[index];
3346 3530 : PGPROC *proc = &allProcs[pgprocno];
3347 3530 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3348 :
3349 3530 : if (proc == MyProc)
3350 712 : continue;
3351 :
3352 2818 : if (excludeVacuum & statusFlags)
3353 24 : continue;
3354 :
3355 2794 : if (allDbs || proc->databaseId == MyDatabaseId)
3356 : {
3357 : /* Fetch xmin just once - might change on us */
3358 1222 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3359 :
3360 1222 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3361 698 : continue;
3362 :
3363 : /*
3364 : * InvalidTransactionId precedes all other XIDs, so a proc that
3365 : * hasn't set xmin yet will not be rejected by this test.
3366 : */
3367 1048 : if (!TransactionIdIsValid(limitXmin) ||
3368 524 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3369 : {
3370 : VirtualTransactionId vxid;
3371 :
3372 498 : GET_VXID_FROM_PGPROC(vxid, *proc);
3373 498 : if (VirtualTransactionIdIsValid(vxid))
3374 498 : vxids[count++] = vxid;
3375 : }
3376 : }
3377 : }
3378 :
3379 712 : LWLockRelease(ProcArrayLock);
3380 :
3381 712 : *nvxids = count;
3382 712 : return vxids;
3383 : }
3384 :
3385 : /*
3386 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3387 : *
3388 : * Usage is limited to conflict resolution during recovery on standby servers.
3389 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3390 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3391 : * determine a safe snapshotConflictHorizon value.
3392 : *
3393 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3394 : * so we're not worried if they have a snapshot or not, nor does it really
3395 : * matter what type of lock we hold. Caller must avoid calling here with
3396 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3397 : * during original execution, since that actually indicates that there is
3398 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3399 : * convention for InvalidTransactionId values is the opposite of our own!).
3400 : *
3401 : * All callers that are checking xmins always now supply a valid and useful
3402 : * value for limitXmin. The limitXmin is always lower than the lowest
3403 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3404 : * because we only care about cleanup records that are cleaning up tuple
3405 : * versions from committed transactions. In that case they will only occur
3406 : * at the point where the record is less than the lowest running xid. That
3407 : * allows us to say that if any backend takes a snapshot concurrently with
3408 : * us then the conflict assessment made here would never include the snapshot
3409 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3410 : * concurrent snapshots when limitXmin is valid. We might think about adding
3411 : * Assert(limitXmin < lowest(KnownAssignedXids))
3412 : * but that would not be true in the case of FATAL errors lagging in array,
3413 : * but we already know those are bogus anyway, so we skip that test.
3414 : *
3415 : * If dbOid is valid we skip backends attached to other databases.
3416 : *
3417 : * Be careful to *not* pfree the result from this function. We reuse
3418 : * this array sufficiently often that we use malloc for the result.
3419 : */
3420 : VirtualTransactionId *
3421 18784 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3422 : {
3423 : static VirtualTransactionId *vxids;
3424 18784 : ProcArrayStruct *arrayP = procArray;
3425 18784 : int count = 0;
3426 : int index;
3427 :
3428 : /*
3429 : * If first time through, get workspace to remember main XIDs in. We
3430 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3431 : * result space, remembering room for a terminator.
3432 : */
3433 18784 : if (vxids == NULL)
3434 : {
3435 40 : vxids = (VirtualTransactionId *)
3436 40 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3437 40 : if (vxids == NULL)
3438 0 : ereport(ERROR,
3439 : (errcode(ERRCODE_OUT_OF_MEMORY),
3440 : errmsg("out of memory")));
3441 : }
3442 :
3443 18784 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3444 :
3445 19078 : for (index = 0; index < arrayP->numProcs; index++)
3446 : {
3447 294 : int pgprocno = arrayP->pgprocnos[index];
3448 294 : PGPROC *proc = &allProcs[pgprocno];
3449 :
3450 : /* Exclude prepared transactions */
3451 294 : if (proc->pid == 0)
3452 0 : continue;
3453 :
3454 294 : if (!OidIsValid(dbOid) ||
3455 282 : proc->databaseId == dbOid)
3456 : {
3457 : /* Fetch xmin just once - can't change on us, but good coding */
3458 32 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3459 :
3460 : /*
3461 : * We ignore an invalid pxmin because this means that backend has
3462 : * no snapshot currently. We hold a Share lock to avoid contention
3463 : * with users taking snapshots. That is not a problem because the
3464 : * current xmin is always at least one higher than the latest
3465 : * removed xid, so any new snapshot would never conflict with the
3466 : * test here.
3467 : */
3468 32 : if (!TransactionIdIsValid(limitXmin) ||
3469 6 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3470 : {
3471 : VirtualTransactionId vxid;
3472 :
3473 4 : GET_VXID_FROM_PGPROC(vxid, *proc);
3474 4 : if (VirtualTransactionIdIsValid(vxid))
3475 4 : vxids[count++] = vxid;
3476 : }
3477 : }
3478 : }
3479 :
3480 18784 : LWLockRelease(ProcArrayLock);
3481 :
3482 : /* add the terminator */
3483 18784 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3484 18784 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3485 :
3486 18784 : return vxids;
3487 : }
3488 :
3489 : /*
3490 : * CancelVirtualTransaction - used in recovery conflict processing
3491 : *
3492 : * Returns pid of the process signaled, or 0 if not found.
3493 : */
3494 : pid_t
3495 6 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3496 : {
3497 6 : return SignalVirtualTransaction(vxid, sigmode, true);
3498 : }
3499 :
3500 : pid_t
3501 10 : SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode,
3502 : bool conflictPending)
3503 : {
3504 10 : ProcArrayStruct *arrayP = procArray;
3505 : int index;
3506 10 : pid_t pid = 0;
3507 :
3508 10 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3509 :
3510 10 : for (index = 0; index < arrayP->numProcs; index++)
3511 : {
3512 10 : int pgprocno = arrayP->pgprocnos[index];
3513 10 : PGPROC *proc = &allProcs[pgprocno];
3514 : VirtualTransactionId procvxid;
3515 :
3516 10 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3517 :
3518 10 : if (procvxid.procNumber == vxid.procNumber &&
3519 10 : procvxid.localTransactionId == vxid.localTransactionId)
3520 : {
3521 10 : proc->recoveryConflictPending = conflictPending;
3522 10 : pid = proc->pid;
3523 10 : if (pid != 0)
3524 : {
3525 : /*
3526 : * Kill the pid if it's still here. If not, that's what we
3527 : * wanted so ignore any errors.
3528 : */
3529 10 : (void) SendProcSignal(pid, sigmode, vxid.procNumber);
3530 : }
3531 10 : break;
3532 : }
3533 : }
3534 :
3535 10 : LWLockRelease(ProcArrayLock);
3536 :
3537 10 : return pid;
3538 : }
3539 :
3540 : /*
3541 : * MinimumActiveBackends --- count backends (other than myself) that are
3542 : * in active transactions. Return true if the count exceeds the
3543 : * minimum threshold passed. This is used as a heuristic to decide if
3544 : * a pre-XLOG-flush delay is worthwhile during commit.
3545 : *
3546 : * Do not count backends that are blocked waiting for locks, since they are
3547 : * not going to get to run until someone else commits.
3548 : */
3549 : bool
3550 0 : MinimumActiveBackends(int min)
3551 : {
3552 0 : ProcArrayStruct *arrayP = procArray;
3553 0 : int count = 0;
3554 : int index;
3555 :
3556 : /* Quick short-circuit if no minimum is specified */
3557 0 : if (min == 0)
3558 0 : return true;
3559 :
3560 : /*
3561 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3562 : * bogus, but since we are only testing fields for zero or nonzero, it
3563 : * should be OK. The result is only used for heuristic purposes anyway...
3564 : */
3565 0 : for (index = 0; index < arrayP->numProcs; index++)
3566 : {
3567 0 : int pgprocno = arrayP->pgprocnos[index];
3568 0 : PGPROC *proc = &allProcs[pgprocno];
3569 :
3570 : /*
3571 : * Since we're not holding a lock, need to be prepared to deal with
3572 : * garbage, as someone could have incremented numProcs but not yet
3573 : * filled the structure.
3574 : *
3575 : * If someone just decremented numProcs, 'proc' could also point to a
3576 : * PGPROC entry that's no longer in the array. It still points to a
3577 : * PGPROC struct, though, because freed PGPROC entries just go to the
3578 : * free list and are recycled. Its contents are nonsense in that case,
3579 : * but that's acceptable for this function.
3580 : */
3581 0 : if (pgprocno == -1)
3582 0 : continue; /* do not count deleted entries */
3583 0 : if (proc == MyProc)
3584 0 : continue; /* do not count myself */
3585 0 : if (proc->xid == InvalidTransactionId)
3586 0 : continue; /* do not count if no XID assigned */
3587 0 : if (proc->pid == 0)
3588 0 : continue; /* do not count prepared xacts */
3589 0 : if (proc->waitLock != NULL)
3590 0 : continue; /* do not count if blocked on a lock */
3591 0 : count++;
3592 0 : if (count >= min)
3593 0 : break;
3594 : }
3595 :
3596 0 : return count >= min;
3597 : }
3598 :
3599 : /*
3600 : * CountDBBackends --- count backends that are using specified database
3601 : */
3602 : int
3603 30 : CountDBBackends(Oid databaseid)
3604 : {
3605 30 : ProcArrayStruct *arrayP = procArray;
3606 30 : int count = 0;
3607 : int index;
3608 :
3609 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3610 :
3611 44 : for (index = 0; index < arrayP->numProcs; index++)
3612 : {
3613 14 : int pgprocno = arrayP->pgprocnos[index];
3614 14 : PGPROC *proc = &allProcs[pgprocno];
3615 :
3616 14 : if (proc->pid == 0)
3617 0 : continue; /* do not count prepared xacts */
3618 14 : if (!OidIsValid(databaseid) ||
3619 14 : proc->databaseId == databaseid)
3620 4 : count++;
3621 : }
3622 :
3623 30 : LWLockRelease(ProcArrayLock);
3624 :
3625 30 : return count;
3626 : }
3627 :
3628 : /*
3629 : * CountDBConnections --- counts database backends ignoring any background
3630 : * worker processes
3631 : */
3632 : int
3633 0 : CountDBConnections(Oid databaseid)
3634 : {
3635 0 : ProcArrayStruct *arrayP = procArray;
3636 0 : int count = 0;
3637 : int index;
3638 :
3639 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3640 :
3641 0 : for (index = 0; index < arrayP->numProcs; index++)
3642 : {
3643 0 : int pgprocno = arrayP->pgprocnos[index];
3644 0 : PGPROC *proc = &allProcs[pgprocno];
3645 :
3646 0 : if (proc->pid == 0)
3647 0 : continue; /* do not count prepared xacts */
3648 0 : if (proc->isBackgroundWorker)
3649 0 : continue; /* do not count background workers */
3650 0 : if (!OidIsValid(databaseid) ||
3651 0 : proc->databaseId == databaseid)
3652 0 : count++;
3653 : }
3654 :
3655 0 : LWLockRelease(ProcArrayLock);
3656 :
3657 0 : return count;
3658 : }
3659 :
3660 : /*
3661 : * CancelDBBackends --- cancel backends that are using specified database
3662 : */
3663 : void
3664 20 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3665 : {
3666 20 : ProcArrayStruct *arrayP = procArray;
3667 : int index;
3668 :
3669 : /* tell all backends to die */
3670 20 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3671 :
3672 40 : for (index = 0; index < arrayP->numProcs; index++)
3673 : {
3674 20 : int pgprocno = arrayP->pgprocnos[index];
3675 20 : PGPROC *proc = &allProcs[pgprocno];
3676 :
3677 20 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3678 : {
3679 : VirtualTransactionId procvxid;
3680 : pid_t pid;
3681 :
3682 20 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3683 :
3684 20 : proc->recoveryConflictPending = conflictPending;
3685 20 : pid = proc->pid;
3686 20 : if (pid != 0)
3687 : {
3688 : /*
3689 : * Kill the pid if it's still here. If not, that's what we
3690 : * wanted so ignore any errors.
3691 : */
3692 20 : (void) SendProcSignal(pid, sigmode, procvxid.procNumber);
3693 : }
3694 : }
3695 : }
3696 :
3697 20 : LWLockRelease(ProcArrayLock);
3698 20 : }
3699 :
3700 : /*
3701 : * CountUserBackends --- count backends that are used by specified user
3702 : */
3703 : int
3704 0 : CountUserBackends(Oid roleid)
3705 : {
3706 0 : ProcArrayStruct *arrayP = procArray;
3707 0 : int count = 0;
3708 : int index;
3709 :
3710 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3711 :
3712 0 : for (index = 0; index < arrayP->numProcs; index++)
3713 : {
3714 0 : int pgprocno = arrayP->pgprocnos[index];
3715 0 : PGPROC *proc = &allProcs[pgprocno];
3716 :
3717 0 : if (proc->pid == 0)
3718 0 : continue; /* do not count prepared xacts */
3719 0 : if (proc->isBackgroundWorker)
3720 0 : continue; /* do not count background workers */
3721 0 : if (proc->roleId == roleid)
3722 0 : count++;
3723 : }
3724 :
3725 0 : LWLockRelease(ProcArrayLock);
3726 :
3727 0 : return count;
3728 : }
3729 :
3730 : /*
3731 : * CountOtherDBBackends -- check for other backends running in the given DB
3732 : *
3733 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3734 : * for them to exit. Autovacuum backends are encouraged to exit early by
3735 : * sending them SIGTERM, but normal user backends are just waited for.
3736 : *
3737 : * The current backend is always ignored; it is caller's responsibility to
3738 : * check whether the current backend uses the given DB, if it's important.
3739 : *
3740 : * Returns true if there are (still) other backends in the DB, false if not.
3741 : * Also, *nbackends and *nprepared are set to the number of other backends
3742 : * and prepared transactions in the DB, respectively.
3743 : *
3744 : * This function is used to interlock DROP DATABASE and related commands
3745 : * against there being any active backends in the target DB --- dropping the
3746 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3747 : * detect here the possibility of a newly-started backend that is trying to
3748 : * connect to the doomed database, so additional interlocking is needed during
3749 : * backend startup. The caller should normally hold an exclusive lock on the
3750 : * target DB before calling this, which is one reason we mustn't wait
3751 : * indefinitely.
3752 : */
3753 : bool
3754 736 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3755 : {
3756 736 : ProcArrayStruct *arrayP = procArray;
3757 :
3758 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3759 : int autovac_pids[MAXAUTOVACPIDS];
3760 : int tries;
3761 :
3762 : /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3763 736 : for (tries = 0; tries < 50; tries++)
3764 : {
3765 736 : int nautovacs = 0;
3766 736 : bool found = false;
3767 : int index;
3768 :
3769 736 : CHECK_FOR_INTERRUPTS();
3770 :
3771 736 : *nbackends = *nprepared = 0;
3772 :
3773 736 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3774 :
3775 2702 : for (index = 0; index < arrayP->numProcs; index++)
3776 : {
3777 1966 : int pgprocno = arrayP->pgprocnos[index];
3778 1966 : PGPROC *proc = &allProcs[pgprocno];
3779 1966 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3780 :
3781 1966 : if (proc->databaseId != databaseId)
3782 1794 : continue;
3783 172 : if (proc == MyProc)
3784 172 : continue;
3785 :
3786 0 : found = true;
3787 :
3788 0 : if (proc->pid == 0)
3789 0 : (*nprepared)++;
3790 : else
3791 : {
3792 0 : (*nbackends)++;
3793 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3794 : nautovacs < MAXAUTOVACPIDS)
3795 0 : autovac_pids[nautovacs++] = proc->pid;
3796 : }
3797 : }
3798 :
3799 736 : LWLockRelease(ProcArrayLock);
3800 :
3801 736 : if (!found)
3802 736 : return false; /* no conflicting backends, so done */
3803 :
3804 : /*
3805 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3806 : * postpone this step until after the loop because we don't want to
3807 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3808 : * block kill() inside the kernel...
3809 : */
3810 0 : for (index = 0; index < nautovacs; index++)
3811 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3812 :
3813 : /* sleep, then try again */
3814 0 : pg_usleep(100 * 1000L); /* 100ms */
3815 : }
3816 :
3817 0 : return true; /* timed out, still conflicts */
3818 : }
3819 :
3820 : /*
3821 : * Terminate existing connections to the specified database. This routine
3822 : * is used by the DROP DATABASE command when user has asked to forcefully
3823 : * drop the database.
3824 : *
3825 : * The current backend is always ignored; it is caller's responsibility to
3826 : * check whether the current backend uses the given DB, if it's important.
3827 : *
3828 : * If the target database has a prepared transaction or permissions checks
3829 : * fail for a connection, this fails without terminating anything.
3830 : */
3831 : void
3832 2 : TerminateOtherDBBackends(Oid databaseId)
3833 : {
3834 2 : ProcArrayStruct *arrayP = procArray;
3835 2 : List *pids = NIL;
3836 2 : int nprepared = 0;
3837 : int i;
3838 :
3839 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3840 :
3841 8 : for (i = 0; i < procArray->numProcs; i++)
3842 : {
3843 6 : int pgprocno = arrayP->pgprocnos[i];
3844 6 : PGPROC *proc = &allProcs[pgprocno];
3845 :
3846 6 : if (proc->databaseId != databaseId)
3847 6 : continue;
3848 0 : if (proc == MyProc)
3849 0 : continue;
3850 :
3851 0 : if (proc->pid != 0)
3852 0 : pids = lappend_int(pids, proc->pid);
3853 : else
3854 0 : nprepared++;
3855 : }
3856 :
3857 2 : LWLockRelease(ProcArrayLock);
3858 :
3859 2 : if (nprepared > 0)
3860 0 : ereport(ERROR,
3861 : (errcode(ERRCODE_OBJECT_IN_USE),
3862 : errmsg("database \"%s\" is being used by prepared transactions",
3863 : get_database_name(databaseId)),
3864 : errdetail_plural("There is %d prepared transaction using the database.",
3865 : "There are %d prepared transactions using the database.",
3866 : nprepared,
3867 : nprepared)));
3868 :
3869 2 : if (pids)
3870 : {
3871 : ListCell *lc;
3872 :
3873 : /*
3874 : * Permissions checks relax the pg_terminate_backend checks in two
3875 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3876 : *
3877 : * - Accept terminating autovacuum workers, since DROP DATABASE
3878 : * without FORCE terminates them.
3879 : *
3880 : * - Accept terminating bgworkers. For bgworker authors, it's
3881 : * convenient to be able to recommend FORCE if a worker is blocking
3882 : * DROP DATABASE unexpectedly.
3883 : *
3884 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3885 : * "PID %d is not a PostgreSQL server process", because for us already
3886 : * finished session is not a problem.
3887 : */
3888 0 : foreach(lc, pids)
3889 : {
3890 0 : int pid = lfirst_int(lc);
3891 0 : PGPROC *proc = BackendPidGetProc(pid);
3892 :
3893 0 : if (proc != NULL)
3894 : {
3895 0 : if (superuser_arg(proc->roleId) && !superuser())
3896 0 : ereport(ERROR,
3897 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3898 : errmsg("permission denied to terminate process"),
3899 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3900 : "SUPERUSER", "SUPERUSER")));
3901 :
3902 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3903 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3904 0 : ereport(ERROR,
3905 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3906 : errmsg("permission denied to terminate process"),
3907 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3908 : "pg_signal_backend")));
3909 : }
3910 : }
3911 :
3912 : /*
3913 : * There's a race condition here: once we release the ProcArrayLock,
3914 : * it's possible for the session to exit before we issue kill. That
3915 : * race condition possibility seems too unlikely to worry about. See
3916 : * pg_signal_backend.
3917 : */
3918 0 : foreach(lc, pids)
3919 : {
3920 0 : int pid = lfirst_int(lc);
3921 0 : PGPROC *proc = BackendPidGetProc(pid);
3922 :
3923 0 : if (proc != NULL)
3924 : {
3925 : /*
3926 : * If we have setsid(), signal the backend's whole process
3927 : * group
3928 : */
3929 : #ifdef HAVE_SETSID
3930 0 : (void) kill(-pid, SIGTERM);
3931 : #else
3932 : (void) kill(pid, SIGTERM);
3933 : #endif
3934 : }
3935 : }
3936 : }
3937 2 : }
3938 :
3939 : /*
3940 : * ProcArraySetReplicationSlotXmin
3941 : *
3942 : * Install limits to future computations of the xmin horizon to prevent vacuum
3943 : * and HOT pruning from removing affected rows still needed by clients with
3944 : * replication slots.
3945 : */
3946 : void
3947 4072 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3948 : bool already_locked)
3949 : {
3950 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3951 :
3952 4072 : if (!already_locked)
3953 3248 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3954 :
3955 4072 : procArray->replication_slot_xmin = xmin;
3956 4072 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3957 :
3958 4072 : if (!already_locked)
3959 3248 : LWLockRelease(ProcArrayLock);
3960 :
3961 4072 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3962 : xmin, catalog_xmin);
3963 4072 : }
3964 :
3965 : /*
3966 : * ProcArrayGetReplicationSlotXmin
3967 : *
3968 : * Return the current slot xmin limits. That's useful to be able to remove
3969 : * data that's older than those limits.
3970 : */
3971 : void
3972 44 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3973 : TransactionId *catalog_xmin)
3974 : {
3975 44 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3976 :
3977 44 : if (xmin != NULL)
3978 0 : *xmin = procArray->replication_slot_xmin;
3979 :
3980 44 : if (catalog_xmin != NULL)
3981 44 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3982 :
3983 44 : LWLockRelease(ProcArrayLock);
3984 44 : }
3985 :
3986 : /*
3987 : * XidCacheRemoveRunningXids
3988 : *
3989 : * Remove a bunch of TransactionIds from the list of known-running
3990 : * subtransactions for my backend. Both the specified xid and those in
3991 : * the xids[] array (of length nxids) are removed from the subxids cache.
3992 : * latestXid must be the latest XID among the group.
3993 : */
3994 : void
3995 1298 : XidCacheRemoveRunningXids(TransactionId xid,
3996 : int nxids, const TransactionId *xids,
3997 : TransactionId latestXid)
3998 : {
3999 : int i,
4000 : j;
4001 : XidCacheStatus *mysubxidstat;
4002 :
4003 : Assert(TransactionIdIsValid(xid));
4004 :
4005 : /*
4006 : * We must hold ProcArrayLock exclusively in order to remove transactions
4007 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
4008 : * possible this could be relaxed since we know this routine is only used
4009 : * to abort subtransactions, but pending closer analysis we'd best be
4010 : * conservative.
4011 : *
4012 : * Note that we do not have to be careful about memory ordering of our own
4013 : * reads wrt. GetNewTransactionId() here - only this process can modify
4014 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
4015 : * careful about our own writes being well ordered.
4016 : */
4017 1298 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4018 :
4019 1298 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4020 :
4021 : /*
4022 : * Under normal circumstances xid and xids[] will be in increasing order,
4023 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4024 : * behavior when removing a lot of xids.
4025 : */
4026 1358 : for (i = nxids - 1; i >= 0; i--)
4027 : {
4028 60 : TransactionId anxid = xids[i];
4029 :
4030 60 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4031 : {
4032 60 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4033 : {
4034 60 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4035 60 : pg_write_barrier();
4036 60 : mysubxidstat->count--;
4037 60 : MyProc->subxidStatus.count--;
4038 60 : break;
4039 : }
4040 : }
4041 :
4042 : /*
4043 : * Ordinarily we should have found it, unless the cache has
4044 : * overflowed. However it's also possible for this routine to be
4045 : * invoked multiple times for the same subtransaction, in case of an
4046 : * error during AbortSubTransaction. So instead of Assert, emit a
4047 : * debug warning.
4048 : */
4049 60 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4050 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4051 : }
4052 :
4053 1426 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4054 : {
4055 1424 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4056 : {
4057 1296 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4058 1296 : pg_write_barrier();
4059 1296 : mysubxidstat->count--;
4060 1296 : MyProc->subxidStatus.count--;
4061 1296 : break;
4062 : }
4063 : }
4064 : /* Ordinarily we should have found it, unless the cache has overflowed */
4065 1298 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4066 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4067 :
4068 : /* Also advance global latestCompletedXid while holding the lock */
4069 1298 : MaintainLatestCompletedXid(latestXid);
4070 :
4071 : /* ... and xactCompletionCount */
4072 1298 : TransamVariables->xactCompletionCount++;
4073 :
4074 1298 : LWLockRelease(ProcArrayLock);
4075 1298 : }
4076 :
4077 : #ifdef XIDCACHE_DEBUG
4078 :
4079 : /*
4080 : * Print stats about effectiveness of XID cache
4081 : */
4082 : static void
4083 : DisplayXidCache(void)
4084 : {
4085 : fprintf(stderr,
4086 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4087 : xc_by_recent_xmin,
4088 : xc_by_known_xact,
4089 : xc_by_my_xact,
4090 : xc_by_latest_xid,
4091 : xc_by_main_xid,
4092 : xc_by_child_xid,
4093 : xc_by_known_assigned,
4094 : xc_no_overflow,
4095 : xc_slow_answer);
4096 : }
4097 : #endif /* XIDCACHE_DEBUG */
4098 :
4099 : /*
4100 : * If rel != NULL, return test state appropriate for relation, otherwise
4101 : * return state usable for all relations. The latter may consider XIDs as
4102 : * not-yet-visible-to-everyone that a state for a specific relation would
4103 : * already consider visible-to-everyone.
4104 : *
4105 : * This needs to be called while a snapshot is active or registered, otherwise
4106 : * there are wraparound and other dangers.
4107 : *
4108 : * See comment for GlobalVisState for details.
4109 : */
4110 : GlobalVisState *
4111 27550582 : GlobalVisTestFor(Relation rel)
4112 : {
4113 27550582 : GlobalVisState *state = NULL;
4114 :
4115 : /* XXX: we should assert that a snapshot is pushed or registered */
4116 : Assert(RecentXmin);
4117 :
4118 27550582 : switch (GlobalVisHorizonKindForRel(rel))
4119 : {
4120 168282 : case VISHORIZON_SHARED:
4121 168282 : state = &GlobalVisSharedRels;
4122 168282 : break;
4123 4731196 : case VISHORIZON_CATALOG:
4124 4731196 : state = &GlobalVisCatalogRels;
4125 4731196 : break;
4126 22567444 : case VISHORIZON_DATA:
4127 22567444 : state = &GlobalVisDataRels;
4128 22567444 : break;
4129 83660 : case VISHORIZON_TEMP:
4130 83660 : state = &GlobalVisTempRels;
4131 83660 : break;
4132 : }
4133 :
4134 27550582 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4135 : FullTransactionIdIsValid(state->maybe_needed));
4136 :
4137 27550582 : return state;
4138 : }
4139 :
4140 : /*
4141 : * Return true if it's worth updating the accurate maybe_needed boundary.
4142 : *
4143 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4144 : * repeatedly do so when there is a low likelihood of it being beneficial.
4145 : *
4146 : * The current heuristic is that we update only if RecentXmin has changed
4147 : * since the last update. If the oldest currently running transaction has not
4148 : * finished, it is unlikely that recomputing the horizon would be useful.
4149 : */
4150 : static bool
4151 706702 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4152 : {
4153 : /* hasn't been updated yet */
4154 706702 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4155 18022 : return true;
4156 :
4157 : /*
4158 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4159 : * unlikely to be beneficial to refresh boundaries.
4160 : */
4161 688680 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4162 : state->definitely_needed))
4163 0 : return false;
4164 :
4165 : /* does the last snapshot built have a different xmin? */
4166 688680 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4167 : }
4168 :
4169 : static void
4170 259044 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4171 : {
4172 : GlobalVisSharedRels.maybe_needed =
4173 259044 : FullXidRelativeTo(horizons->latest_completed,
4174 : horizons->shared_oldest_nonremovable);
4175 : GlobalVisCatalogRels.maybe_needed =
4176 259044 : FullXidRelativeTo(horizons->latest_completed,
4177 : horizons->catalog_oldest_nonremovable);
4178 : GlobalVisDataRels.maybe_needed =
4179 259044 : FullXidRelativeTo(horizons->latest_completed,
4180 : horizons->data_oldest_nonremovable);
4181 : GlobalVisTempRels.maybe_needed =
4182 259044 : FullXidRelativeTo(horizons->latest_completed,
4183 : horizons->temp_oldest_nonremovable);
4184 :
4185 : /*
4186 : * In longer running transactions it's possible that transactions we
4187 : * previously needed to treat as running aren't around anymore. So update
4188 : * definitely_needed to not be earlier than maybe_needed.
4189 : */
4190 : GlobalVisSharedRels.definitely_needed =
4191 259044 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4192 : GlobalVisSharedRels.definitely_needed);
4193 : GlobalVisCatalogRels.definitely_needed =
4194 259044 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4195 : GlobalVisCatalogRels.definitely_needed);
4196 : GlobalVisDataRels.definitely_needed =
4197 259044 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4198 : GlobalVisDataRels.definitely_needed);
4199 259044 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4200 :
4201 259044 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4202 259044 : }
4203 :
4204 : /*
4205 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4206 : * using ComputeXidHorizons().
4207 : */
4208 : static void
4209 103212 : GlobalVisUpdate(void)
4210 : {
4211 : ComputeXidHorizonsResult horizons;
4212 :
4213 : /* updates the horizons as a side-effect */
4214 103212 : ComputeXidHorizons(&horizons);
4215 103212 : }
4216 :
4217 : /*
4218 : * Return true if no snapshot still considers fxid to be running.
4219 : *
4220 : * The state passed needs to have been initialized for the relation fxid is
4221 : * from (NULL is also OK), otherwise the result may not be correct.
4222 : *
4223 : * See comment for GlobalVisState for details.
4224 : */
4225 : bool
4226 18196442 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4227 : FullTransactionId fxid)
4228 : {
4229 : /*
4230 : * If fxid is older than maybe_needed bound, it definitely is visible to
4231 : * everyone.
4232 : */
4233 18196442 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4234 4421276 : return true;
4235 :
4236 : /*
4237 : * If fxid is >= definitely_needed bound, it is very likely to still be
4238 : * considered running.
4239 : */
4240 13775166 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4241 13068464 : return false;
4242 :
4243 : /*
4244 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4245 : * might not exist a snapshot considering fxid running. If it makes sense,
4246 : * update boundaries and recheck.
4247 : */
4248 706702 : if (GlobalVisTestShouldUpdate(state))
4249 : {
4250 103212 : GlobalVisUpdate();
4251 :
4252 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4253 :
4254 103212 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4255 : }
4256 : else
4257 603490 : return false;
4258 : }
4259 :
4260 : /*
4261 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4262 : *
4263 : * It is crucial that this only gets called for xids from a source that
4264 : * protects against xid wraparounds (e.g. from a table and thus protected by
4265 : * relfrozenxid).
4266 : */
4267 : bool
4268 18196144 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4269 : {
4270 : FullTransactionId fxid;
4271 :
4272 : /*
4273 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4274 : * because we know the xid has to, at the very least, be between
4275 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4276 : * lock to determine either, we can just compare with
4277 : * state->definitely_needed, which was based on those value at the time
4278 : * the current snapshot was built.
4279 : */
4280 18196144 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4281 :
4282 18196144 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4283 : }
4284 :
4285 : /*
4286 : * Convenience wrapper around GlobalVisTestFor() and
4287 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4288 : */
4289 : bool
4290 298 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4291 : {
4292 : GlobalVisState *state;
4293 :
4294 298 : state = GlobalVisTestFor(rel);
4295 :
4296 298 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4297 : }
4298 :
4299 : /*
4300 : * Convenience wrapper around GlobalVisTestFor() and
4301 : * GlobalVisTestIsRemovableXid(), see their comments.
4302 : */
4303 : bool
4304 12 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4305 : {
4306 : GlobalVisState *state;
4307 :
4308 12 : state = GlobalVisTestFor(rel);
4309 :
4310 12 : return GlobalVisTestIsRemovableXid(state, xid);
4311 : }
4312 :
4313 : /*
4314 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4315 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4316 : *
4317 : * Be very careful about when to use this function. It can only safely be used
4318 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4319 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4320 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4321 : * the xid has to be within that range), or if xid is from the procarray and
4322 : * prevents xid wraparound that way.
4323 : */
4324 : static inline FullTransactionId
4325 21651886 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4326 : {
4327 21651886 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4328 :
4329 : Assert(TransactionIdIsValid(xid));
4330 : Assert(TransactionIdIsValid(rel_xid));
4331 :
4332 : /* not guaranteed to find issues, but likely to catch mistakes */
4333 : AssertTransactionIdInAllowableRange(xid);
4334 :
4335 43303772 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4336 21651886 : + (int32) (xid - rel_xid));
4337 : }
4338 :
4339 :
4340 : /* ----------------------------------------------
4341 : * KnownAssignedTransactionIds sub-module
4342 : * ----------------------------------------------
4343 : */
4344 :
4345 : /*
4346 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4347 : * running on the primary at the current point in WAL. These XIDs must be
4348 : * treated as running by standby transactions, even though they are not in
4349 : * the standby server's PGPROC array.
4350 : *
4351 : * We record all XIDs that we know have been assigned. That includes all the
4352 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4353 : * been assigned. We can deduce the existence of unobserved XIDs because we
4354 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4355 : * list expands as new XIDs are observed or inferred, and contracts when
4356 : * transaction completion records arrive.
4357 : *
4358 : * During hot standby we do not fret too much about the distinction between
4359 : * top-level XIDs and subtransaction XIDs. We store both together in the
4360 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4361 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4362 : * doesn't care about the distinction either. Subtransaction XIDs are
4363 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4364 : * links are *not* maintained (which does not affect visibility).
4365 : *
4366 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4367 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4368 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4369 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4370 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4371 : * and then remove them from KnownAssignedXids. This prevents overflow of
4372 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4373 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4374 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4375 : * though it should be complete for top-level XIDs; this is the same situation
4376 : * that holds with respect to the PGPROC entries in normal running.
4377 : *
4378 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4379 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4380 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4381 : * As long as that is within the range of interesting XIDs, we have to assume
4382 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4383 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4384 : * subXID arrives - that is not an error.)
4385 : *
4386 : * Should a backend on primary somehow disappear before it can write an abort
4387 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4388 : * aborted but we think they were running; the distinction is irrelevant
4389 : * because either way any changes done by the transaction are not visible to
4390 : * backends in the standby. We prune KnownAssignedXids when
4391 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4392 : * array due to such dead XIDs.
4393 : */
4394 :
4395 : /*
4396 : * RecordKnownAssignedTransactionIds
4397 : * Record the given XID in KnownAssignedXids, as well as any preceding
4398 : * unobserved XIDs.
4399 : *
4400 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4401 : * associated with a transaction. Must be called for each record after we
4402 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4403 : *
4404 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4405 : */
4406 : void
4407 4765468 : RecordKnownAssignedTransactionIds(TransactionId xid)
4408 : {
4409 : Assert(standbyState >= STANDBY_INITIALIZED);
4410 : Assert(TransactionIdIsValid(xid));
4411 : Assert(TransactionIdIsValid(latestObservedXid));
4412 :
4413 4765468 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4414 : xid, latestObservedXid);
4415 :
4416 : /*
4417 : * When a newly observed xid arrives, it is frequently the case that it is
4418 : * *not* the next xid in sequence. When this occurs, we must treat the
4419 : * intervening xids as running also.
4420 : */
4421 4765468 : if (TransactionIdFollows(xid, latestObservedXid))
4422 : {
4423 : TransactionId next_expected_xid;
4424 :
4425 : /*
4426 : * Extend subtrans like we do in GetNewTransactionId() during normal
4427 : * operation using individual extend steps. Note that we do not need
4428 : * to extend clog since its extensions are WAL logged.
4429 : *
4430 : * This part has to be done regardless of standbyState since we
4431 : * immediately start assigning subtransactions to their toplevel
4432 : * transactions.
4433 : */
4434 42352 : next_expected_xid = latestObservedXid;
4435 85576 : while (TransactionIdPrecedes(next_expected_xid, xid))
4436 : {
4437 43224 : TransactionIdAdvance(next_expected_xid);
4438 43224 : ExtendSUBTRANS(next_expected_xid);
4439 : }
4440 : Assert(next_expected_xid == xid);
4441 :
4442 : /*
4443 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4444 : * more to do since we don't track assigned xids yet.
4445 : */
4446 42352 : if (standbyState <= STANDBY_INITIALIZED)
4447 : {
4448 0 : latestObservedXid = xid;
4449 0 : return;
4450 : }
4451 :
4452 : /*
4453 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4454 : */
4455 42352 : next_expected_xid = latestObservedXid;
4456 42352 : TransactionIdAdvance(next_expected_xid);
4457 42352 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4458 :
4459 : /*
4460 : * Now we can advance latestObservedXid
4461 : */
4462 42352 : latestObservedXid = xid;
4463 :
4464 : /* TransamVariables->nextXid must be beyond any observed xid */
4465 42352 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4466 : }
4467 : }
4468 :
4469 : /*
4470 : * ExpireTreeKnownAssignedTransactionIds
4471 : * Remove the given XIDs from KnownAssignedXids.
4472 : *
4473 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4474 : */
4475 : void
4476 40162 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4477 : TransactionId *subxids, TransactionId max_xid)
4478 : {
4479 : Assert(standbyState >= STANDBY_INITIALIZED);
4480 :
4481 : /*
4482 : * Uses same locking as transaction commit
4483 : */
4484 40162 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4485 :
4486 40162 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4487 :
4488 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4489 40162 : MaintainLatestCompletedXidRecovery(max_xid);
4490 :
4491 : /* ... and xactCompletionCount */
4492 40162 : TransamVariables->xactCompletionCount++;
4493 :
4494 40162 : LWLockRelease(ProcArrayLock);
4495 40162 : }
4496 :
4497 : /*
4498 : * ExpireAllKnownAssignedTransactionIds
4499 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4500 : */
4501 : void
4502 198 : ExpireAllKnownAssignedTransactionIds(void)
4503 : {
4504 198 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4505 198 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4506 :
4507 : /*
4508 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4509 : * the call of this function. But do this for unification with what
4510 : * ExpireOldKnownAssignedTransactionIds() do.
4511 : */
4512 198 : procArray->lastOverflowedXid = InvalidTransactionId;
4513 198 : LWLockRelease(ProcArrayLock);
4514 198 : }
4515 :
4516 : /*
4517 : * ExpireOldKnownAssignedTransactionIds
4518 : * Remove KnownAssignedXids entries preceding the given XID and
4519 : * potentially reset lastOverflowedXid.
4520 : */
4521 : void
4522 870 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4523 : {
4524 870 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4525 :
4526 : /*
4527 : * Reset lastOverflowedXid if we know all transactions that have been
4528 : * possibly running are being gone. Not doing so could cause an incorrect
4529 : * lastOverflowedXid value, which makes extra snapshots be marked as
4530 : * suboverflowed.
4531 : */
4532 870 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4533 854 : procArray->lastOverflowedXid = InvalidTransactionId;
4534 870 : KnownAssignedXidsRemovePreceding(xid);
4535 870 : LWLockRelease(ProcArrayLock);
4536 870 : }
4537 :
4538 : /*
4539 : * KnownAssignedTransactionIdsIdleMaintenance
4540 : * Opportunistically do maintenance work when the startup process
4541 : * is about to go idle.
4542 : */
4543 : void
4544 4276 : KnownAssignedTransactionIdsIdleMaintenance(void)
4545 : {
4546 4276 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4547 4276 : }
4548 :
4549 :
4550 : /*
4551 : * Private module functions to manipulate KnownAssignedXids
4552 : *
4553 : * There are 5 main uses of the KnownAssignedXids data structure:
4554 : *
4555 : * * backends taking snapshots - all valid XIDs need to be copied out
4556 : * * backends seeking to determine presence of a specific XID
4557 : * * startup process adding new known-assigned XIDs
4558 : * * startup process removing specific XIDs as transactions end
4559 : * * startup process pruning array when special WAL records arrive
4560 : *
4561 : * This data structure is known to be a hot spot during Hot Standby, so we
4562 : * go to some lengths to make these operations as efficient and as concurrent
4563 : * as possible.
4564 : *
4565 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4566 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4567 : * in general TransactionIdPrecedes would not provide a total order, but
4568 : * we know that the entries present at any instant should not extend across
4569 : * a large enough fraction of XID space to wrap around (the primary would
4570 : * shut down for fear of XID wrap long before that happens). So it's OK to
4571 : * use TransactionIdPrecedes as a binary-search comparator.
4572 : *
4573 : * It's cheap to maintain the sortedness during insertions, since new known
4574 : * XIDs are always reported in XID order; we just append them at the right.
4575 : *
4576 : * To keep individual deletions cheap, we need to allow gaps in the array.
4577 : * This is implemented by marking array elements as valid or invalid using
4578 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4579 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4580 : * XID entry itself. This preserves the property that the XID entries are
4581 : * sorted, so we can do binary searches easily. Periodically we compress
4582 : * out the unused entries; that's much cheaper than having to compress the
4583 : * array immediately on every deletion.
4584 : *
4585 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4586 : * are those with indexes tail <= i < head; items outside this subscript range
4587 : * have unspecified contents. When head reaches the end of the array, we
4588 : * force compression of unused entries rather than wrapping around, since
4589 : * allowing wraparound would greatly complicate the search logic. We maintain
4590 : * an explicit tail pointer so that pruning of old XIDs can be done without
4591 : * immediately moving the array contents. In most cases only a small fraction
4592 : * of the array contains valid entries at any instant.
4593 : *
4594 : * Although only the startup process can ever change the KnownAssignedXids
4595 : * data structure, we still need interlocking so that standby backends will
4596 : * not observe invalid intermediate states. The convention is that backends
4597 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4598 : * the array, the startup process must hold ProcArrayLock exclusively, for
4599 : * the usual transactional reasons (compare commit/abort of a transaction
4600 : * during normal running). Compressing unused entries out of the array
4601 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4602 : * them into slots to the right of the head pointer and then advance the head
4603 : * pointer. This doesn't require any lock at all, but on machines with weak
4604 : * memory ordering, we need to be careful that other processors see the array
4605 : * element changes before they see the head pointer change. We handle this by
4606 : * using memory barriers when reading or writing the head/tail pointers (unless
4607 : * the caller holds ProcArrayLock exclusively).
4608 : *
4609 : * Algorithmic analysis:
4610 : *
4611 : * If we have a maximum of M slots, with N XIDs currently spread across
4612 : * S elements then we have N <= S <= M always.
4613 : *
4614 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4615 : * happen)
4616 : * * Compressing the array is O(S) and requires exclusive lock
4617 : * * Removing an XID is O(logS) and requires exclusive lock
4618 : * * Taking a snapshot is O(S) and requires shared lock
4619 : * * Checking for an XID is O(logS) and requires shared lock
4620 : *
4621 : * In comparison, using a hash table for KnownAssignedXids would mean that
4622 : * taking snapshots would be O(M). If we can maintain S << M then the
4623 : * sorted array technique will deliver significantly faster snapshots.
4624 : * If we try to keep S too small then we will spend too much time compressing,
4625 : * so there is an optimal point for any workload mix. We use a heuristic to
4626 : * decide when to compress the array, though trimming also helps reduce
4627 : * frequency of compressing. The heuristic requires us to track the number of
4628 : * currently valid XIDs in the array (N). Except in special cases, we'll
4629 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4630 : * taking a snapshot to be O(N), which it would have to be anyway.
4631 : */
4632 :
4633 :
4634 : /*
4635 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4636 : * array, removing any gaps.
4637 : *
4638 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4639 : * we do it only if a heuristic indicates it's a good time to do it.
4640 : *
4641 : * Compression requires holding ProcArrayLock in exclusive mode.
4642 : * Caller must pass haveLock = true if it already holds the lock.
4643 : */
4644 : static void
4645 45350 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4646 : {
4647 45350 : ProcArrayStruct *pArray = procArray;
4648 : int head,
4649 : tail,
4650 : nelements;
4651 : int compress_index;
4652 : int i;
4653 :
4654 : /* Counters for compression heuristics */
4655 : static unsigned int transactionEndsCounter;
4656 : static TimestampTz lastCompressTs;
4657 :
4658 : /* Tuning constants */
4659 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4660 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4661 :
4662 : /*
4663 : * Since only the startup process modifies the head/tail pointers, we
4664 : * don't need a lock to read them here.
4665 : */
4666 45350 : head = pArray->headKnownAssignedXids;
4667 45350 : tail = pArray->tailKnownAssignedXids;
4668 45350 : nelements = head - tail;
4669 :
4670 : /*
4671 : * If we can choose whether to compress, use a heuristic to avoid
4672 : * compressing too often or not often enough. "Compress" here simply
4673 : * means moving the values to the beginning of the array, so it is not as
4674 : * complex or costly as typical data compression algorithms.
4675 : */
4676 45350 : if (nelements == pArray->numKnownAssignedXids)
4677 : {
4678 : /*
4679 : * When there are no gaps between head and tail, don't bother to
4680 : * compress, except in the KAX_NO_SPACE case where we must compress to
4681 : * create some space after the head.
4682 : */
4683 19068 : if (reason != KAX_NO_SPACE)
4684 19068 : return;
4685 : }
4686 26282 : else if (reason == KAX_TRANSACTION_END)
4687 : {
4688 : /*
4689 : * Consider compressing only once every so many commits. Frequency
4690 : * determined by benchmarks.
4691 : */
4692 24812 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4693 24600 : return;
4694 :
4695 : /*
4696 : * Furthermore, compress only if the used part of the array is less
4697 : * than 50% full (see comments above).
4698 : */
4699 212 : if (nelements < 2 * pArray->numKnownAssignedXids)
4700 12 : return;
4701 : }
4702 1470 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4703 : {
4704 : /*
4705 : * We're about to go idle for lack of new WAL, so we might as well
4706 : * compress. But not too often, to avoid ProcArray lock contention
4707 : * with readers.
4708 : */
4709 1270 : if (lastCompressTs != 0)
4710 : {
4711 : TimestampTz compress_after;
4712 :
4713 1270 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4714 : KAX_COMPRESS_IDLE_INTERVAL);
4715 1270 : if (GetCurrentTimestamp() < compress_after)
4716 1246 : return;
4717 : }
4718 : }
4719 :
4720 : /* Need to compress, so get the lock if we don't have it. */
4721 424 : if (!haveLock)
4722 24 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4723 :
4724 : /*
4725 : * We compress the array by reading the valid values from tail to head,
4726 : * re-aligning data to 0th element.
4727 : */
4728 424 : compress_index = 0;
4729 16698 : for (i = tail; i < head; i++)
4730 : {
4731 16274 : if (KnownAssignedXidsValid[i])
4732 : {
4733 1776 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4734 1776 : KnownAssignedXidsValid[compress_index] = true;
4735 1776 : compress_index++;
4736 : }
4737 : }
4738 : Assert(compress_index == pArray->numKnownAssignedXids);
4739 :
4740 424 : pArray->tailKnownAssignedXids = 0;
4741 424 : pArray->headKnownAssignedXids = compress_index;
4742 :
4743 424 : if (!haveLock)
4744 24 : LWLockRelease(ProcArrayLock);
4745 :
4746 : /* Update timestamp for maintenance. No need to hold lock for this. */
4747 424 : lastCompressTs = GetCurrentTimestamp();
4748 : }
4749 :
4750 : /*
4751 : * Add xids into KnownAssignedXids at the head of the array.
4752 : *
4753 : * xids from from_xid to to_xid, inclusive, are added to the array.
4754 : *
4755 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4756 : * exclusive mode, so we need no extra locking here. Else caller holds no
4757 : * lock, so we need to be sure we maintain sufficient interlocks against
4758 : * concurrent readers. (Only the startup process ever calls this, so no need
4759 : * to worry about concurrent writers.)
4760 : */
4761 : static void
4762 42360 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4763 : bool exclusive_lock)
4764 : {
4765 42360 : ProcArrayStruct *pArray = procArray;
4766 : TransactionId next_xid;
4767 : int head,
4768 : tail;
4769 : int nxids;
4770 : int i;
4771 :
4772 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4773 :
4774 : /*
4775 : * Calculate how many array slots we'll need. Normally this is cheap; in
4776 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4777 : * way.
4778 : */
4779 42360 : if (to_xid >= from_xid)
4780 42360 : nxids = to_xid - from_xid + 1;
4781 : else
4782 : {
4783 0 : nxids = 1;
4784 0 : next_xid = from_xid;
4785 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4786 : {
4787 0 : nxids++;
4788 0 : TransactionIdAdvance(next_xid);
4789 : }
4790 : }
4791 :
4792 : /*
4793 : * Since only the startup process modifies the head/tail pointers, we
4794 : * don't need a lock to read them here.
4795 : */
4796 42360 : head = pArray->headKnownAssignedXids;
4797 42360 : tail = pArray->tailKnownAssignedXids;
4798 :
4799 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4800 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4801 :
4802 : /*
4803 : * Verify that insertions occur in TransactionId sequence. Note that even
4804 : * if the last existing element is marked invalid, it must still have a
4805 : * correctly sequenced XID value.
4806 : */
4807 71158 : if (head > tail &&
4808 28798 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4809 : {
4810 0 : KnownAssignedXidsDisplay(LOG);
4811 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4812 : }
4813 :
4814 : /*
4815 : * If our xids won't fit in the remaining space, compress out free space
4816 : */
4817 42360 : if (head + nxids > pArray->maxKnownAssignedXids)
4818 : {
4819 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4820 :
4821 0 : head = pArray->headKnownAssignedXids;
4822 : /* note: we no longer care about the tail pointer */
4823 :
4824 : /*
4825 : * If it still won't fit then we're out of memory
4826 : */
4827 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4828 0 : elog(ERROR, "too many KnownAssignedXids");
4829 : }
4830 :
4831 : /* Now we can insert the xids into the space starting at head */
4832 42360 : next_xid = from_xid;
4833 85592 : for (i = 0; i < nxids; i++)
4834 : {
4835 43232 : KnownAssignedXids[head] = next_xid;
4836 43232 : KnownAssignedXidsValid[head] = true;
4837 43232 : TransactionIdAdvance(next_xid);
4838 43232 : head++;
4839 : }
4840 :
4841 : /* Adjust count of number of valid entries */
4842 42360 : pArray->numKnownAssignedXids += nxids;
4843 :
4844 : /*
4845 : * Now update the head pointer. We use a write barrier to ensure that
4846 : * other processors see the above array updates before they see the head
4847 : * pointer change. The barrier isn't required if we're holding
4848 : * ProcArrayLock exclusively.
4849 : */
4850 42360 : if (!exclusive_lock)
4851 42352 : pg_write_barrier();
4852 :
4853 42360 : pArray->headKnownAssignedXids = head;
4854 42360 : }
4855 :
4856 : /*
4857 : * KnownAssignedXidsSearch
4858 : *
4859 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4860 : * Returns true if it was found, false if not.
4861 : *
4862 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4863 : * Exclusive lock must be held for remove = true.
4864 : */
4865 : static bool
4866 45502 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4867 : {
4868 45502 : ProcArrayStruct *pArray = procArray;
4869 : int first,
4870 : last;
4871 : int head;
4872 : int tail;
4873 45502 : int result_index = -1;
4874 :
4875 45502 : tail = pArray->tailKnownAssignedXids;
4876 45502 : head = pArray->headKnownAssignedXids;
4877 :
4878 : /*
4879 : * Only the startup process removes entries, so we don't need the read
4880 : * barrier in that case.
4881 : */
4882 45502 : if (!remove)
4883 2 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4884 :
4885 : /*
4886 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4887 : * array here, since even invalid entries will contain sorted XIDs.
4888 : */
4889 45502 : first = tail;
4890 45502 : last = head - 1;
4891 156796 : while (first <= last)
4892 : {
4893 : int mid_index;
4894 : TransactionId mid_xid;
4895 :
4896 154406 : mid_index = (first + last) / 2;
4897 154406 : mid_xid = KnownAssignedXids[mid_index];
4898 :
4899 154406 : if (xid == mid_xid)
4900 : {
4901 43112 : result_index = mid_index;
4902 43112 : break;
4903 : }
4904 111294 : else if (TransactionIdPrecedes(xid, mid_xid))
4905 24096 : last = mid_index - 1;
4906 : else
4907 87198 : first = mid_index + 1;
4908 : }
4909 :
4910 45502 : if (result_index < 0)
4911 2390 : return false; /* not in array */
4912 :
4913 43112 : if (!KnownAssignedXidsValid[result_index])
4914 50 : return false; /* in array, but invalid */
4915 :
4916 43062 : if (remove)
4917 : {
4918 43062 : KnownAssignedXidsValid[result_index] = false;
4919 :
4920 43062 : pArray->numKnownAssignedXids--;
4921 : Assert(pArray->numKnownAssignedXids >= 0);
4922 :
4923 : /*
4924 : * If we're removing the tail element then advance tail pointer over
4925 : * any invalid elements. This will speed future searches.
4926 : */
4927 43062 : if (result_index == tail)
4928 : {
4929 16338 : tail++;
4930 28564 : while (tail < head && !KnownAssignedXidsValid[tail])
4931 12226 : tail++;
4932 16338 : if (tail >= head)
4933 : {
4934 : /* Array is empty, so we can reset both pointers */
4935 13544 : pArray->headKnownAssignedXids = 0;
4936 13544 : pArray->tailKnownAssignedXids = 0;
4937 : }
4938 : else
4939 : {
4940 2794 : pArray->tailKnownAssignedXids = tail;
4941 : }
4942 : }
4943 : }
4944 :
4945 43062 : return true;
4946 : }
4947 :
4948 : /*
4949 : * Is the specified XID present in KnownAssignedXids[]?
4950 : *
4951 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4952 : */
4953 : static bool
4954 2 : KnownAssignedXidExists(TransactionId xid)
4955 : {
4956 : Assert(TransactionIdIsValid(xid));
4957 :
4958 2 : return KnownAssignedXidsSearch(xid, false);
4959 : }
4960 :
4961 : /*
4962 : * Remove the specified XID from KnownAssignedXids[].
4963 : *
4964 : * Caller must hold ProcArrayLock in exclusive mode.
4965 : */
4966 : static void
4967 45500 : KnownAssignedXidsRemove(TransactionId xid)
4968 : {
4969 : Assert(TransactionIdIsValid(xid));
4970 :
4971 45500 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
4972 :
4973 : /*
4974 : * Note: we cannot consider it an error to remove an XID that's not
4975 : * present. We intentionally remove subxact IDs while processing
4976 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4977 : * removed again when the top-level xact commits or aborts.
4978 : *
4979 : * It might be possible to track such XIDs to distinguish this case from
4980 : * actual errors, but it would be complicated and probably not worth it.
4981 : * So, just ignore the search result.
4982 : */
4983 45500 : (void) KnownAssignedXidsSearch(xid, true);
4984 45500 : }
4985 :
4986 : /*
4987 : * KnownAssignedXidsRemoveTree
4988 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4989 : *
4990 : * Caller must hold ProcArrayLock in exclusive mode.
4991 : */
4992 : static void
4993 40204 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
4994 : TransactionId *subxids)
4995 : {
4996 : int i;
4997 :
4998 40204 : if (TransactionIdIsValid(xid))
4999 40162 : KnownAssignedXidsRemove(xid);
5000 :
5001 45542 : for (i = 0; i < nsubxids; i++)
5002 5338 : KnownAssignedXidsRemove(subxids[i]);
5003 :
5004 : /* Opportunistically compress the array */
5005 40204 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5006 40204 : }
5007 :
5008 : /*
5009 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5010 : * then clear the whole table.
5011 : *
5012 : * Caller must hold ProcArrayLock in exclusive mode.
5013 : */
5014 : static void
5015 1068 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5016 : {
5017 1068 : ProcArrayStruct *pArray = procArray;
5018 1068 : int count = 0;
5019 : int head,
5020 : tail,
5021 : i;
5022 :
5023 1068 : if (!TransactionIdIsValid(removeXid))
5024 : {
5025 198 : elog(DEBUG4, "removing all KnownAssignedXids");
5026 198 : pArray->numKnownAssignedXids = 0;
5027 198 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5028 198 : return;
5029 : }
5030 :
5031 870 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5032 :
5033 : /*
5034 : * Mark entries invalid starting at the tail. Since array is sorted, we
5035 : * can stop as soon as we reach an entry >= removeXid.
5036 : */
5037 870 : tail = pArray->tailKnownAssignedXids;
5038 870 : head = pArray->headKnownAssignedXids;
5039 :
5040 870 : for (i = tail; i < head; i++)
5041 : {
5042 392 : if (KnownAssignedXidsValid[i])
5043 : {
5044 392 : TransactionId knownXid = KnownAssignedXids[i];
5045 :
5046 392 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5047 392 : break;
5048 :
5049 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5050 : {
5051 0 : KnownAssignedXidsValid[i] = false;
5052 0 : count++;
5053 : }
5054 : }
5055 : }
5056 :
5057 870 : pArray->numKnownAssignedXids -= count;
5058 : Assert(pArray->numKnownAssignedXids >= 0);
5059 :
5060 : /*
5061 : * Advance the tail pointer if we've marked the tail item invalid.
5062 : */
5063 870 : for (i = tail; i < head; i++)
5064 : {
5065 392 : if (KnownAssignedXidsValid[i])
5066 392 : break;
5067 : }
5068 870 : if (i >= head)
5069 : {
5070 : /* Array is empty, so we can reset both pointers */
5071 478 : pArray->headKnownAssignedXids = 0;
5072 478 : pArray->tailKnownAssignedXids = 0;
5073 : }
5074 : else
5075 : {
5076 392 : pArray->tailKnownAssignedXids = i;
5077 : }
5078 :
5079 : /* Opportunistically compress the array */
5080 870 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5081 : }
5082 :
5083 : /*
5084 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5085 : * We filter out anything >= xmax.
5086 : *
5087 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5088 : * that array is large enough.
5089 : *
5090 : * Caller must hold ProcArrayLock in (at least) shared mode.
5091 : */
5092 : static int
5093 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5094 : {
5095 0 : TransactionId xtmp = InvalidTransactionId;
5096 :
5097 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5098 : }
5099 :
5100 : /*
5101 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5102 : * we reduce *xmin to the lowest xid value seen if not already lower.
5103 : *
5104 : * Caller must hold ProcArrayLock in (at least) shared mode.
5105 : */
5106 : static int
5107 2162 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5108 : TransactionId xmax)
5109 : {
5110 2162 : int count = 0;
5111 : int head,
5112 : tail;
5113 : int i;
5114 :
5115 : /*
5116 : * Fetch head just once, since it may change while we loop. We can stop
5117 : * once we reach the initially seen head, since we are certain that an xid
5118 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5119 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5120 : * anyway.
5121 : */
5122 2162 : tail = procArray->tailKnownAssignedXids;
5123 2162 : head = procArray->headKnownAssignedXids;
5124 :
5125 2162 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5126 :
5127 2210 : for (i = tail; i < head; i++)
5128 : {
5129 : /* Skip any gaps in the array */
5130 178 : if (KnownAssignedXidsValid[i])
5131 : {
5132 162 : TransactionId knownXid = KnownAssignedXids[i];
5133 :
5134 : /*
5135 : * Update xmin if required. Only the first XID need be checked,
5136 : * since the array is sorted.
5137 : */
5138 324 : if (count == 0 &&
5139 162 : TransactionIdPrecedes(knownXid, *xmin))
5140 32 : *xmin = knownXid;
5141 :
5142 : /*
5143 : * Filter out anything >= xmax, again relying on sorted property
5144 : * of array.
5145 : */
5146 324 : if (TransactionIdIsValid(xmax) &&
5147 162 : TransactionIdFollowsOrEquals(knownXid, xmax))
5148 130 : break;
5149 :
5150 : /* Add knownXid into output array */
5151 32 : xarray[count++] = knownXid;
5152 : }
5153 : }
5154 :
5155 2162 : return count;
5156 : }
5157 :
5158 : /*
5159 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5160 : * if nothing there.
5161 : */
5162 : static TransactionId
5163 682 : KnownAssignedXidsGetOldestXmin(void)
5164 : {
5165 : int head,
5166 : tail;
5167 : int i;
5168 :
5169 : /*
5170 : * Fetch head just once, since it may change while we loop.
5171 : */
5172 682 : tail = procArray->tailKnownAssignedXids;
5173 682 : head = procArray->headKnownAssignedXids;
5174 :
5175 682 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5176 :
5177 682 : for (i = tail; i < head; i++)
5178 : {
5179 : /* Skip any gaps in the array */
5180 288 : if (KnownAssignedXidsValid[i])
5181 288 : return KnownAssignedXids[i];
5182 : }
5183 :
5184 394 : return InvalidTransactionId;
5185 : }
5186 :
5187 : /*
5188 : * Display KnownAssignedXids to provide debug trail
5189 : *
5190 : * Currently this is only called within startup process, so we need no
5191 : * special locking.
5192 : *
5193 : * Note this is pretty expensive, and much of the expense will be incurred
5194 : * even if the elog message will get discarded. It's not currently called
5195 : * in any performance-critical places, however, so no need to be tenser.
5196 : */
5197 : static void
5198 206 : KnownAssignedXidsDisplay(int trace_level)
5199 : {
5200 206 : ProcArrayStruct *pArray = procArray;
5201 : StringInfoData buf;
5202 : int head,
5203 : tail,
5204 : i;
5205 206 : int nxids = 0;
5206 :
5207 206 : tail = pArray->tailKnownAssignedXids;
5208 206 : head = pArray->headKnownAssignedXids;
5209 :
5210 206 : initStringInfo(&buf);
5211 :
5212 222 : for (i = tail; i < head; i++)
5213 : {
5214 16 : if (KnownAssignedXidsValid[i])
5215 : {
5216 16 : nxids++;
5217 16 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5218 : }
5219 : }
5220 :
5221 206 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5222 : nxids,
5223 : pArray->numKnownAssignedXids,
5224 : pArray->tailKnownAssignedXids,
5225 : pArray->headKnownAssignedXids,
5226 : buf.data);
5227 :
5228 206 : pfree(buf.data);
5229 206 : }
5230 :
5231 : /*
5232 : * KnownAssignedXidsReset
5233 : * Resets KnownAssignedXids to be empty
5234 : */
5235 : static void
5236 0 : KnownAssignedXidsReset(void)
5237 : {
5238 0 : ProcArrayStruct *pArray = procArray;
5239 :
5240 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5241 :
5242 0 : pArray->numKnownAssignedXids = 0;
5243 0 : pArray->tailKnownAssignedXids = 0;
5244 0 : pArray->headKnownAssignedXids = 0;
5245 :
5246 0 : LWLockRelease(ProcArrayLock);
5247 0 : }
|
