Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procarray.c
4 : * POSTGRES process array code.
5 : *
6 : *
7 : * This module maintains arrays of PGPROC substructures, as well as associated
8 : * arrays in ProcGlobal, for all active backends. Although there are several
9 : * uses for this, the principal one is as a means of determining the set of
10 : * currently running transactions.
11 : *
12 : * Because of various subtle race conditions it is critical that a backend
13 : * hold the correct locks while setting or clearing its xid (in
14 : * ProcGlobal->xids[]/MyProc->xid). See notes in
15 : * src/backend/access/transam/README.
16 : *
17 : * The process arrays now also include structures representing prepared
18 : * transactions. The xid and subxids fields of these are valid, as are the
19 : * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 : * at need by checking for pid == 0.
21 : *
22 : * During hot standby, we also keep a list of XIDs representing transactions
23 : * that are known to be running on the primary (or more precisely, were running
24 : * as of the current point in the WAL stream). This list is kept in the
25 : * KnownAssignedXids array, and is updated by watching the sequence of
26 : * arriving XIDs. This is necessary because if we leave those XIDs out of
27 : * snapshots taken for standby queries, then they will appear to be already
28 : * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 : * array represents standby processes, which by definition are not running
30 : * transactions that have XIDs.
31 : *
32 : * It is perhaps possible for a backend on the primary to terminate without
33 : * writing an abort record for its transaction. While that shouldn't really
34 : * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 : * ourselves by pruning the array when a valid list of running XIDs arrives.
36 : *
37 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/subtrans.h"
51 : #include "access/transam.h"
52 : #include "access/twophase.h"
53 : #include "access/xact.h"
54 : #include "access/xlogutils.h"
55 : #include "catalog/catalog.h"
56 : #include "catalog/pg_authid.h"
57 : #include "miscadmin.h"
58 : #include "pgstat.h"
59 : #include "postmaster/bgworker.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/procsignal.h"
64 : #include "utils/acl.h"
65 : #include "utils/builtins.h"
66 : #include "utils/injection_point.h"
67 : #include "utils/lsyscache.h"
68 : #include "utils/rel.h"
69 : #include "utils/snapmgr.h"
70 : #include "utils/wait_event.h"
71 :
72 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
73 :
74 : /* Our shared memory area */
75 : typedef struct ProcArrayStruct
76 : {
77 : int numProcs; /* number of valid procs entries */
78 : int maxProcs; /* allocated size of procs array */
79 :
80 : /*
81 : * Known assigned XIDs handling
82 : */
83 : int maxKnownAssignedXids; /* allocated size of array */
84 : int numKnownAssignedXids; /* current # of valid entries */
85 : int tailKnownAssignedXids; /* index of oldest valid element */
86 : int headKnownAssignedXids; /* index of newest element, + 1 */
87 :
88 : /*
89 : * Highest subxid that has been removed from KnownAssignedXids array to
90 : * prevent overflow; or InvalidTransactionId if none. We track this for
91 : * similar reasons to tracking overflowing cached subxids in PGPROC
92 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
93 : * lock to read it.
94 : */
95 : TransactionId lastOverflowedXid;
96 :
97 : /* oldest xmin of any replication slot */
98 : TransactionId replication_slot_xmin;
99 : /* oldest catalog xmin of any replication slot */
100 : TransactionId replication_slot_catalog_xmin;
101 :
102 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
103 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
104 : } ProcArrayStruct;
105 :
106 : /*
107 : * State for the GlobalVisTest* family of functions. Those functions can
108 : * e.g. be used to decide if a deleted row can be removed without violating
109 : * MVCC semantics: If the deleted row's xmax is not considered to be running
110 : * by anyone, the row can be removed.
111 : *
112 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
113 : * cutoff XID while building a snapshot (looking at the frequently changing
114 : * xmins scales badly). Instead we compute two boundaries while building the
115 : * snapshot:
116 : *
117 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
118 : * definitely_needed are definitely still visible.
119 : *
120 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
121 : * definitely be removed
122 : *
123 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
124 : * && XID < definitely_needed), the boundaries can be recomputed (using
125 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
126 : * maintaining an accurate value all the time.
127 : *
128 : * As it is not cheap to compute accurate boundaries, we limit the number of
129 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
130 : *
131 : *
132 : * There are three backend lifetime instances of this struct, optimized for
133 : * different types of relations. As e.g. a normal user defined table in one
134 : * database is inaccessible to backends connected to another database, a test
135 : * specific to a relation can be more aggressive than a test for a shared
136 : * relation. Currently we track four different states:
137 : *
138 : * 1) GlobalVisSharedRels, which only considers an XID's
139 : * effects visible-to-everyone if neither snapshots in any database, nor a
140 : * replication slot's xmin, nor a replication slot's catalog_xmin might
141 : * still consider XID as running.
142 : *
143 : * 2) GlobalVisCatalogRels, which only considers an XID's
144 : * effects visible-to-everyone if neither snapshots in the current
145 : * database, nor a replication slot's xmin, nor a replication slot's
146 : * catalog_xmin might still consider XID as running.
147 : *
148 : * I.e. the difference to GlobalVisSharedRels is that
149 : * snapshot in other databases are ignored.
150 : *
151 : * 3) GlobalVisDataRels, which only considers an XID's
152 : * effects visible-to-everyone if neither snapshots in the current
153 : * database, nor a replication slot's xmin consider XID as running.
154 : *
155 : * I.e. the difference to GlobalVisCatalogRels is that
156 : * replication slot's catalog_xmin is not taken into account.
157 : *
158 : * 4) GlobalVisTempRels, which only considers the current session, as temp
159 : * tables are not visible to other sessions.
160 : *
161 : * GlobalVisTestFor(relation) returns the appropriate state
162 : * for the relation.
163 : *
164 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
165 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
166 : * prevent maybe_needed to become old enough after the GetSnapshotData()
167 : * call.
168 : *
169 : * The typedef is in the header.
170 : */
171 : struct GlobalVisState
172 : {
173 : /* XIDs >= are considered running by some backend */
174 : FullTransactionId definitely_needed;
175 :
176 : /* XIDs < are not considered to be running by any backend */
177 : FullTransactionId maybe_needed;
178 : };
179 :
180 : /*
181 : * Result of ComputeXidHorizons().
182 : */
183 : typedef struct ComputeXidHorizonsResult
184 : {
185 : /*
186 : * The value of TransamVariables->latestCompletedXid when
187 : * ComputeXidHorizons() held ProcArrayLock.
188 : */
189 : FullTransactionId latest_completed;
190 :
191 : /*
192 : * The same for procArray->replication_slot_xmin and
193 : * procArray->replication_slot_catalog_xmin.
194 : */
195 : TransactionId slot_xmin;
196 : TransactionId slot_catalog_xmin;
197 :
198 : /*
199 : * Oldest xid that any backend might still consider running. This needs to
200 : * include processes running VACUUM, in contrast to the normal visibility
201 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
202 : * determining visibility, but doesn't care about rows above its xmin to
203 : * be removed.
204 : *
205 : * This likely should only be needed to determine whether pg_subtrans can
206 : * be truncated. It currently includes the effects of replication slots,
207 : * for historical reasons. But that could likely be changed.
208 : */
209 : TransactionId oldest_considered_running;
210 :
211 : /*
212 : * Oldest xid for which deleted tuples need to be retained in shared
213 : * tables.
214 : *
215 : * This includes the effects of replication slots. If that's not desired,
216 : * look at shared_oldest_nonremovable_raw;
217 : */
218 : TransactionId shared_oldest_nonremovable;
219 :
220 : /*
221 : * Oldest xid that may be necessary to retain in shared tables. This is
222 : * the same as shared_oldest_nonremovable, except that is not affected by
223 : * replication slot's catalog_xmin.
224 : *
225 : * This is mainly useful to be able to send the catalog_xmin to upstream
226 : * streaming replication servers via hot_standby_feedback, so they can
227 : * apply the limit only when accessing catalog tables.
228 : */
229 : TransactionId shared_oldest_nonremovable_raw;
230 :
231 : /*
232 : * Oldest xid for which deleted tuples need to be retained in non-shared
233 : * catalog tables.
234 : */
235 : TransactionId catalog_oldest_nonremovable;
236 :
237 : /*
238 : * Oldest xid for which deleted tuples need to be retained in normal user
239 : * defined tables.
240 : */
241 : TransactionId data_oldest_nonremovable;
242 :
243 : /*
244 : * Oldest xid for which deleted tuples need to be retained in this
245 : * session's temporary tables.
246 : */
247 : TransactionId temp_oldest_nonremovable;
248 : } ComputeXidHorizonsResult;
249 :
250 : /*
251 : * Return value for GlobalVisHorizonKindForRel().
252 : */
253 : typedef enum GlobalVisHorizonKind
254 : {
255 : VISHORIZON_SHARED,
256 : VISHORIZON_CATALOG,
257 : VISHORIZON_DATA,
258 : VISHORIZON_TEMP,
259 : } GlobalVisHorizonKind;
260 :
261 : /*
262 : * Reason codes for KnownAssignedXidsCompress().
263 : */
264 : typedef enum KAXCompressReason
265 : {
266 : KAX_NO_SPACE, /* need to free up space at array end */
267 : KAX_PRUNE, /* we just pruned old entries */
268 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
269 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
270 : } KAXCompressReason;
271 :
272 :
273 : static ProcArrayStruct *procArray;
274 :
275 : static PGPROC *allProcs;
276 :
277 : /*
278 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
279 : */
280 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
281 :
282 : /*
283 : * Bookkeeping for tracking emulated transactions in recovery
284 : */
285 : static TransactionId *KnownAssignedXids;
286 : static bool *KnownAssignedXidsValid;
287 : static TransactionId latestObservedXid = InvalidTransactionId;
288 :
289 : /*
290 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
291 : * the highest xid that might still be running that we don't have in
292 : * KnownAssignedXids.
293 : */
294 : static TransactionId standbySnapshotPendingXmin;
295 :
296 : /*
297 : * State for visibility checks on different types of relations. See struct
298 : * GlobalVisState for details. As shared, catalog, normal and temporary
299 : * relations can have different horizons, one such state exists for each.
300 : */
301 : static GlobalVisState GlobalVisSharedRels;
302 : static GlobalVisState GlobalVisCatalogRels;
303 : static GlobalVisState GlobalVisDataRels;
304 : static GlobalVisState GlobalVisTempRels;
305 :
306 : /*
307 : * This backend's RecentXmin at the last time the accurate xmin horizon was
308 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
309 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
310 : */
311 : static TransactionId ComputeXidHorizonsResultLastXmin;
312 :
313 : #ifdef XIDCACHE_DEBUG
314 :
315 : /* counters for XidCache measurement */
316 : static long xc_by_recent_xmin = 0;
317 : static long xc_by_known_xact = 0;
318 : static long xc_by_my_xact = 0;
319 : static long xc_by_latest_xid = 0;
320 : static long xc_by_main_xid = 0;
321 : static long xc_by_child_xid = 0;
322 : static long xc_by_known_assigned = 0;
323 : static long xc_no_overflow = 0;
324 : static long xc_slow_answer = 0;
325 :
326 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
327 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
328 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
329 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
330 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
331 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
332 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
333 : #define xc_no_overflow_inc() (xc_no_overflow++)
334 : #define xc_slow_answer_inc() (xc_slow_answer++)
335 :
336 : static void DisplayXidCache(void);
337 : #else /* !XIDCACHE_DEBUG */
338 :
339 : #define xc_by_recent_xmin_inc() ((void) 0)
340 : #define xc_by_known_xact_inc() ((void) 0)
341 : #define xc_by_my_xact_inc() ((void) 0)
342 : #define xc_by_latest_xid_inc() ((void) 0)
343 : #define xc_by_main_xid_inc() ((void) 0)
344 : #define xc_by_child_xid_inc() ((void) 0)
345 : #define xc_by_known_assigned_inc() ((void) 0)
346 : #define xc_no_overflow_inc() ((void) 0)
347 : #define xc_slow_answer_inc() ((void) 0)
348 : #endif /* XIDCACHE_DEBUG */
349 :
350 : /* Primitives for KnownAssignedXids array handling for standby */
351 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
352 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
353 : bool exclusive_lock);
354 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
355 : static bool KnownAssignedXidExists(TransactionId xid);
356 : static void KnownAssignedXidsRemove(TransactionId xid);
357 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
358 : TransactionId *subxids);
359 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
360 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
361 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
362 : TransactionId *xmin,
363 : TransactionId xmax);
364 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
365 : static void KnownAssignedXidsDisplay(int trace_level);
366 : static void KnownAssignedXidsReset(void);
367 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
368 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
369 : static void MaintainLatestCompletedXid(TransactionId latestXid);
370 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
371 :
372 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
373 : TransactionId xid);
374 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
375 :
376 : /*
377 : * Report shared-memory space needed by ProcArrayShmemInit
378 : */
379 : Size
380 2207 : ProcArrayShmemSize(void)
381 : {
382 : Size size;
383 :
384 : /* Size of the ProcArray structure itself */
385 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
386 :
387 2207 : size = offsetof(ProcArrayStruct, pgprocnos);
388 2207 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
389 :
390 : /*
391 : * During Hot Standby processing we have a data structure called
392 : * KnownAssignedXids, created in shared memory. Local data structures are
393 : * also created in various backends during GetSnapshotData(),
394 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
395 : * main structures created in those functions must be identically sized,
396 : * since we may at times copy the whole of the data structures around. We
397 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
398 : *
399 : * Ideally we'd only create this structure if we were actually doing hot
400 : * standby in the current run, but we don't know that yet at the time
401 : * shared memory is being set up.
402 : */
403 : #define TOTAL_MAX_CACHED_SUBXIDS \
404 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
405 :
406 2207 : if (EnableHotStandby)
407 : {
408 2195 : size = add_size(size,
409 : mul_size(sizeof(TransactionId),
410 2195 : TOTAL_MAX_CACHED_SUBXIDS));
411 2195 : size = add_size(size,
412 2195 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
413 : }
414 :
415 2207 : return size;
416 : }
417 :
418 : /*
419 : * Initialize the shared PGPROC array during postmaster startup.
420 : */
421 : void
422 1180 : ProcArrayShmemInit(void)
423 : {
424 : bool found;
425 :
426 : /* Create or attach to the ProcArray shared structure */
427 1180 : procArray = (ProcArrayStruct *)
428 1180 : ShmemInitStruct("Proc Array",
429 : add_size(offsetof(ProcArrayStruct, pgprocnos),
430 : mul_size(sizeof(int),
431 1180 : PROCARRAY_MAXPROCS)),
432 : &found);
433 :
434 1180 : if (!found)
435 : {
436 : /*
437 : * We're the first - initialize.
438 : */
439 1180 : procArray->numProcs = 0;
440 1180 : procArray->maxProcs = PROCARRAY_MAXPROCS;
441 1180 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
442 1180 : procArray->numKnownAssignedXids = 0;
443 1180 : procArray->tailKnownAssignedXids = 0;
444 1180 : procArray->headKnownAssignedXids = 0;
445 1180 : procArray->lastOverflowedXid = InvalidTransactionId;
446 1180 : procArray->replication_slot_xmin = InvalidTransactionId;
447 1180 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
448 1180 : TransamVariables->xactCompletionCount = 1;
449 : }
450 :
451 1180 : allProcs = ProcGlobal->allProcs;
452 :
453 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
454 1180 : if (EnableHotStandby)
455 : {
456 1174 : KnownAssignedXids = (TransactionId *)
457 1174 : ShmemInitStruct("KnownAssignedXids",
458 : mul_size(sizeof(TransactionId),
459 1174 : TOTAL_MAX_CACHED_SUBXIDS),
460 : &found);
461 1174 : KnownAssignedXidsValid = (bool *)
462 1174 : ShmemInitStruct("KnownAssignedXidsValid",
463 1174 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
464 : &found);
465 : }
466 1180 : }
467 :
468 : /*
469 : * Add the specified PGPROC to the shared array.
470 : */
471 : void
472 19788 : ProcArrayAdd(PGPROC *proc)
473 : {
474 19788 : int pgprocno = GetNumberFromPGProc(proc);
475 19788 : ProcArrayStruct *arrayP = procArray;
476 : int index;
477 : int movecount;
478 :
479 : /* See ProcGlobal comment explaining why both locks are held */
480 19788 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
481 19788 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
482 :
483 19788 : if (arrayP->numProcs >= arrayP->maxProcs)
484 : {
485 : /*
486 : * Oops, no room. (This really shouldn't happen, since there is a
487 : * fixed supply of PGPROC structs too, and so we should have failed
488 : * earlier.)
489 : */
490 0 : ereport(FATAL,
491 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
492 : errmsg("sorry, too many clients already")));
493 : }
494 :
495 : /*
496 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
497 : * locality of references much better. This is useful while traversing the
498 : * ProcArray because there is an increased likelihood of finding the next
499 : * PGPROC structure in the cache.
500 : *
501 : * Since the occurrence of adding/removing a proc is much lower than the
502 : * access to the ProcArray itself, the overhead should be marginal
503 : */
504 51115 : for (index = 0; index < arrayP->numProcs; index++)
505 : {
506 46023 : int this_procno = arrayP->pgprocnos[index];
507 :
508 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
509 : Assert(allProcs[this_procno].pgxactoff == index);
510 :
511 : /* If we have found our right position in the array, break */
512 46023 : if (this_procno > pgprocno)
513 14696 : break;
514 : }
515 :
516 19788 : movecount = arrayP->numProcs - index;
517 19788 : memmove(&arrayP->pgprocnos[index + 1],
518 19788 : &arrayP->pgprocnos[index],
519 : movecount * sizeof(*arrayP->pgprocnos));
520 19788 : memmove(&ProcGlobal->xids[index + 1],
521 19788 : &ProcGlobal->xids[index],
522 : movecount * sizeof(*ProcGlobal->xids));
523 19788 : memmove(&ProcGlobal->subxidStates[index + 1],
524 19788 : &ProcGlobal->subxidStates[index],
525 : movecount * sizeof(*ProcGlobal->subxidStates));
526 19788 : memmove(&ProcGlobal->statusFlags[index + 1],
527 19788 : &ProcGlobal->statusFlags[index],
528 : movecount * sizeof(*ProcGlobal->statusFlags));
529 :
530 19788 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
531 19788 : proc->pgxactoff = index;
532 19788 : ProcGlobal->xids[index] = proc->xid;
533 19788 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
534 19788 : ProcGlobal->statusFlags[index] = proc->statusFlags;
535 :
536 19788 : arrayP->numProcs++;
537 :
538 : /* adjust pgxactoff for all following PGPROCs */
539 19788 : index++;
540 53434 : for (; index < arrayP->numProcs; index++)
541 : {
542 33646 : int procno = arrayP->pgprocnos[index];
543 :
544 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
545 : Assert(allProcs[procno].pgxactoff == index - 1);
546 :
547 33646 : allProcs[procno].pgxactoff = index;
548 : }
549 :
550 : /*
551 : * Release in reversed acquisition order, to reduce frequency of having to
552 : * wait for XidGenLock while holding ProcArrayLock.
553 : */
554 19788 : LWLockRelease(XidGenLock);
555 19788 : LWLockRelease(ProcArrayLock);
556 19788 : }
557 :
558 : /*
559 : * Remove the specified PGPROC from the shared array.
560 : *
561 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
562 : * array, and thus causing it to appear as "not running" anymore. In this
563 : * case we must advance latestCompletedXid. (This is essentially the same
564 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
565 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
566 : * twophase.c depends on the latter.)
567 : */
568 : void
569 19759 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
570 : {
571 19759 : ProcArrayStruct *arrayP = procArray;
572 : int myoff;
573 : int movecount;
574 :
575 : #ifdef XIDCACHE_DEBUG
576 : /* dump stats at backend shutdown, but not prepared-xact end */
577 : if (proc->pid != 0)
578 : DisplayXidCache();
579 : #endif
580 :
581 : /* See ProcGlobal comment explaining why both locks are held */
582 19759 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
583 19759 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
584 :
585 19759 : myoff = proc->pgxactoff;
586 :
587 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
588 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
589 :
590 19759 : if (TransactionIdIsValid(latestXid))
591 : {
592 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
593 :
594 : /* Advance global latestCompletedXid while holding the lock */
595 313 : MaintainLatestCompletedXid(latestXid);
596 :
597 : /* Same with xactCompletionCount */
598 313 : TransamVariables->xactCompletionCount++;
599 :
600 313 : ProcGlobal->xids[myoff] = InvalidTransactionId;
601 313 : ProcGlobal->subxidStates[myoff].overflowed = false;
602 313 : ProcGlobal->subxidStates[myoff].count = 0;
603 : }
604 : else
605 : {
606 : /* Shouldn't be trying to remove a live transaction here */
607 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
608 : }
609 :
610 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
611 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
612 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
613 :
614 19759 : ProcGlobal->statusFlags[myoff] = 0;
615 :
616 : /* Keep the PGPROC array sorted. See notes above */
617 19759 : movecount = arrayP->numProcs - myoff - 1;
618 19759 : memmove(&arrayP->pgprocnos[myoff],
619 19759 : &arrayP->pgprocnos[myoff + 1],
620 : movecount * sizeof(*arrayP->pgprocnos));
621 19759 : memmove(&ProcGlobal->xids[myoff],
622 19759 : &ProcGlobal->xids[myoff + 1],
623 : movecount * sizeof(*ProcGlobal->xids));
624 19759 : memmove(&ProcGlobal->subxidStates[myoff],
625 19759 : &ProcGlobal->subxidStates[myoff + 1],
626 : movecount * sizeof(*ProcGlobal->subxidStates));
627 19759 : memmove(&ProcGlobal->statusFlags[myoff],
628 19759 : &ProcGlobal->statusFlags[myoff + 1],
629 : movecount * sizeof(*ProcGlobal->statusFlags));
630 :
631 19759 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
632 19759 : arrayP->numProcs--;
633 :
634 : /*
635 : * Adjust pgxactoff of following procs for removed PGPROC (note that
636 : * numProcs already has been decremented).
637 : */
638 57297 : for (int index = myoff; index < arrayP->numProcs; index++)
639 : {
640 37538 : int procno = arrayP->pgprocnos[index];
641 :
642 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
643 : Assert(allProcs[procno].pgxactoff - 1 == index);
644 :
645 37538 : allProcs[procno].pgxactoff = index;
646 : }
647 :
648 : /*
649 : * Release in reversed acquisition order, to reduce frequency of having to
650 : * wait for XidGenLock while holding ProcArrayLock.
651 : */
652 19759 : LWLockRelease(XidGenLock);
653 19759 : LWLockRelease(ProcArrayLock);
654 19759 : }
655 :
656 :
657 : /*
658 : * ProcArrayEndTransaction -- mark a transaction as no longer running
659 : *
660 : * This is used interchangeably for commit and abort cases. The transaction
661 : * commit/abort must already be reported to WAL and pg_xact.
662 : *
663 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
664 : * latestXid is the latest Xid among the transaction's main XID and
665 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
666 : * the caller to pass latestXid, instead of computing it from the PGPROC's
667 : * contents, because the subxid information in the PGPROC might be
668 : * incomplete.)
669 : */
670 : void
671 617613 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
672 : {
673 617613 : if (TransactionIdIsValid(latestXid))
674 : {
675 : /*
676 : * We must lock ProcArrayLock while clearing our advertised XID, so
677 : * that we do not exit the set of "running" transactions while someone
678 : * else is taking a snapshot. See discussion in
679 : * src/backend/access/transam/README.
680 : */
681 : Assert(TransactionIdIsValid(proc->xid));
682 :
683 : /*
684 : * If we can immediately acquire ProcArrayLock, we clear our own XID
685 : * and release the lock. If not, use group XID clearing to improve
686 : * efficiency.
687 : */
688 159202 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
689 : {
690 159056 : ProcArrayEndTransactionInternal(proc, latestXid);
691 159056 : LWLockRelease(ProcArrayLock);
692 : }
693 : else
694 146 : ProcArrayGroupClearXid(proc, latestXid);
695 : }
696 : else
697 : {
698 : /*
699 : * If we have no XID, we don't need to lock, since we won't affect
700 : * anyone else's calculation of a snapshot. We might change their
701 : * estimate of global xmin, but that's OK.
702 : */
703 : Assert(!TransactionIdIsValid(proc->xid));
704 : Assert(proc->subxidStatus.count == 0);
705 : Assert(!proc->subxidStatus.overflowed);
706 :
707 458411 : proc->vxid.lxid = InvalidLocalTransactionId;
708 458411 : proc->xmin = InvalidTransactionId;
709 :
710 : /* be sure this is cleared in abort */
711 458411 : proc->delayChkptFlags = 0;
712 :
713 : /* must be cleared with xid/xmin: */
714 : /* avoid unnecessarily dirtying shared cachelines */
715 458411 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
716 : {
717 : Assert(!LWLockHeldByMe(ProcArrayLock));
718 118627 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
719 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
720 118627 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
721 118627 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
722 118627 : LWLockRelease(ProcArrayLock);
723 : }
724 : }
725 617613 : }
726 :
727 : /*
728 : * Mark a write transaction as no longer running.
729 : *
730 : * We don't do any locking here; caller must handle that.
731 : */
732 : static inline void
733 159202 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
734 : {
735 159202 : int pgxactoff = proc->pgxactoff;
736 :
737 : /*
738 : * Note: we need exclusive lock here because we're going to change other
739 : * processes' PGPROC entries.
740 : */
741 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
742 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
743 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
744 :
745 159202 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
746 159202 : proc->xid = InvalidTransactionId;
747 159202 : proc->vxid.lxid = InvalidLocalTransactionId;
748 159202 : proc->xmin = InvalidTransactionId;
749 :
750 : /* be sure this is cleared in abort */
751 159202 : proc->delayChkptFlags = 0;
752 :
753 : /* must be cleared with xid/xmin: */
754 : /* avoid unnecessarily dirtying shared cachelines */
755 159202 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
756 : {
757 819 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
758 819 : 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 159202 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
765 : {
766 595 : ProcGlobal->subxidStates[pgxactoff].count = 0;
767 595 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
768 595 : proc->subxidStatus.count = 0;
769 595 : proc->subxidStatus.overflowed = false;
770 : }
771 :
772 : /* Also advance global latestCompletedXid while holding the lock */
773 159202 : MaintainLatestCompletedXid(latestXid);
774 :
775 : /* Same with xactCompletionCount */
776 159202 : TransamVariables->xactCompletionCount++;
777 159202 : }
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
792 146 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
793 : {
794 146 : int pgprocno = GetNumberFromPGProc(proc);
795 146 : 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. */
803 146 : proc->procArrayGroupMember = true;
804 146 : proc->procArrayGroupMemberXid = latestXid;
805 146 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
806 : while (true)
807 : {
808 146 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
809 :
810 146 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
811 : &nextidx,
812 : (uint32) pgprocno))
813 146 : 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 : */
822 146 : if (nextidx != INVALID_PROC_NUMBER)
823 : {
824 8 : int extraWaits = 0;
825 :
826 : /* Sleep until the leader clears our XID. */
827 8 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
828 : for (;;)
829 : {
830 : /* acts as a read barrier */
831 8 : PGSemaphoreLock(proc->sem);
832 8 : if (!proc->procArrayGroupMember)
833 8 : break;
834 0 : extraWaits++;
835 : }
836 8 : 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 8 : while (extraWaits-- > 0)
842 0 : PGSemaphoreUnlock(proc->sem);
843 8 : return;
844 : }
845 :
846 : /* We are the leader. Acquire the lock on behalf of everyone. */
847 138 : 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 : */
854 138 : 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. */
858 138 : wakeidx = nextidx;
859 :
860 : /* Walk the list and clear all XIDs. */
861 284 : while (nextidx != INVALID_PROC_NUMBER)
862 : {
863 146 : PGPROC *nextproc = &allProcs[nextidx];
864 :
865 146 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
866 :
867 : /* Move to next proc in list. */
868 146 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
869 : }
870 :
871 : /* We're done with the lock now. */
872 138 : 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 : */
881 284 : while (wakeidx != INVALID_PROC_NUMBER)
882 : {
883 146 : PGPROC *nextproc = &allProcs[wakeidx];
884 :
885 146 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
886 146 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
887 :
888 : /* ensure all previous writes are visible before follower continues. */
889 146 : pg_write_barrier();
890 :
891 146 : nextproc->procArrayGroupMember = false;
892 :
893 146 : if (nextproc != MyProc)
894 8 : 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
907 309 : 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 : */
925 309 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
926 :
927 309 : pgxactoff = proc->pgxactoff;
928 :
929 309 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
930 309 : proc->xid = InvalidTransactionId;
931 :
932 309 : proc->vxid.lxid = InvalidLocalTransactionId;
933 309 : proc->xmin = InvalidTransactionId;
934 :
935 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
936 : Assert(!proc->delayChkptFlags);
937 :
938 : /*
939 : * Need to increment completion count even though transaction hasn't
940 : * really committed yet. The reason for that is that GetSnapshotData()
941 : * omits the xid of the current transaction, thus without the increment we
942 : * otherwise could end up reusing the snapshot later. Which would be bad,
943 : * because it might not count the prepared transaction as running.
944 : */
945 309 : TransamVariables->xactCompletionCount++;
946 :
947 : /* Clear the subtransaction-XID cache too */
948 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
949 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
950 309 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
951 : {
952 103 : ProcGlobal->subxidStates[pgxactoff].count = 0;
953 103 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
954 103 : proc->subxidStatus.count = 0;
955 103 : proc->subxidStatus.overflowed = false;
956 : }
957 :
958 309 : LWLockRelease(ProcArrayLock);
959 309 : }
960 :
961 : /*
962 : * Update TransamVariables->latestCompletedXid to point to latestXid if
963 : * currently older.
964 : */
965 : static void
966 160366 : MaintainLatestCompletedXid(TransactionId latestXid)
967 : {
968 160366 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
969 :
970 : Assert(FullTransactionIdIsValid(cur_latest));
971 : Assert(!RecoveryInProgress());
972 : Assert(LWLockHeldByMe(ProcArrayLock));
973 :
974 160366 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
975 : {
976 144611 : TransamVariables->latestCompletedXid =
977 144611 : FullXidRelativeTo(cur_latest, latestXid);
978 : }
979 :
980 : Assert(IsBootstrapProcessingMode() ||
981 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
982 160366 : }
983 :
984 : /*
985 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
986 : */
987 : static void
988 23944 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
989 : {
990 23944 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
991 : FullTransactionId rel;
992 :
993 : Assert(AmStartupProcess() || !IsUnderPostmaster);
994 : Assert(LWLockHeldByMe(ProcArrayLock));
995 :
996 : /*
997 : * Need a FullTransactionId to compare latestXid with. Can't rely on
998 : * latestCompletedXid to be initialized in recovery. But in recovery it's
999 : * safe to access nextXid without a lock for the startup process.
1000 : */
1001 23944 : rel = TransamVariables->nextXid;
1002 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1003 :
1004 47774 : if (!FullTransactionIdIsValid(cur_latest) ||
1005 23830 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1006 : {
1007 18430 : TransamVariables->latestCompletedXid =
1008 18430 : FullXidRelativeTo(rel, latestXid);
1009 : }
1010 :
1011 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1012 23944 : }
1013 :
1014 : /*
1015 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1016 : *
1017 : * Remember up to where the startup process initialized the CLOG and subtrans
1018 : * so we can ensure it's initialized gaplessly up to the point where necessary
1019 : * while in recovery.
1020 : */
1021 : void
1022 114 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1023 : {
1024 : Assert(standbyState == STANDBY_INITIALIZED);
1025 : Assert(TransactionIdIsNormal(initializedUptoXID));
1026 :
1027 : /*
1028 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1029 : * to, so we can extend it from that point onwards in
1030 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1031 : * ProcArrayApplyRecoveryInfo().
1032 : */
1033 114 : latestObservedXid = initializedUptoXID;
1034 114 : TransactionIdRetreat(latestObservedXid);
1035 114 : }
1036 :
1037 : /*
1038 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1039 : *
1040 : * Takes us through 3 states: Initialized, Pending and Ready.
1041 : * Normal case is to go all the way to Ready straight away, though there
1042 : * are atypical cases where we need to take it in steps.
1043 : *
1044 : * Use the data about running transactions on the primary to create the initial
1045 : * state of KnownAssignedXids. We also use these records to regularly prune
1046 : * KnownAssignedXids because we know it is possible that some transactions
1047 : * with FATAL errors fail to write abort records, which could cause eventual
1048 : * overflow.
1049 : *
1050 : * See comments for LogStandbySnapshot().
1051 : */
1052 : void
1053 826 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1054 : {
1055 : TransactionId *xids;
1056 : TransactionId advanceNextXid;
1057 : int nxids;
1058 : int i;
1059 :
1060 : Assert(standbyState >= STANDBY_INITIALIZED);
1061 : Assert(TransactionIdIsValid(running->nextXid));
1062 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1063 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1064 :
1065 : /*
1066 : * Remove stale transactions, if any.
1067 : */
1068 826 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1069 :
1070 : /*
1071 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1072 : * because we will need it up to date for accessing two-phase transactions
1073 : * in StandbyReleaseOldLocks().
1074 : */
1075 826 : advanceNextXid = running->nextXid;
1076 826 : TransactionIdRetreat(advanceNextXid);
1077 826 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1078 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1079 :
1080 : /*
1081 : * Remove stale locks, if any.
1082 : */
1083 826 : StandbyReleaseOldLocks(running->oldestRunningXid);
1084 :
1085 : /*
1086 : * If our snapshot is already valid, nothing else to do...
1087 : */
1088 826 : if (standbyState == STANDBY_SNAPSHOT_READY)
1089 712 : return;
1090 :
1091 : /*
1092 : * If our initial RunningTransactionsData had an overflowed snapshot then
1093 : * we knew we were missing some subxids from our snapshot. If we continue
1094 : * to see overflowed snapshots then we might never be able to start up, so
1095 : * we make another test to see if our snapshot is now valid. We know that
1096 : * the missing subxids are equal to or earlier than nextXid. After we
1097 : * initialise we continue to apply changes during recovery, so once the
1098 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1099 : * know that we no longer have missing information and can mark the
1100 : * snapshot as valid.
1101 : */
1102 114 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1103 : {
1104 : /*
1105 : * If the snapshot isn't overflowed or if its empty we can reset our
1106 : * pending state and use this snapshot instead.
1107 : */
1108 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1109 : {
1110 : /*
1111 : * If we have already collected known assigned xids, we need to
1112 : * throw them away before we apply the recovery snapshot.
1113 : */
1114 0 : KnownAssignedXidsReset();
1115 0 : standbyState = STANDBY_INITIALIZED;
1116 : }
1117 : else
1118 : {
1119 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1120 : running->oldestRunningXid))
1121 : {
1122 0 : standbyState = STANDBY_SNAPSHOT_READY;
1123 0 : elog(DEBUG1,
1124 : "recovery snapshots are now enabled");
1125 : }
1126 : else
1127 0 : elog(DEBUG1,
1128 : "recovery snapshot waiting for non-overflowed snapshot or "
1129 : "until oldest active xid on standby is at least %u (now %u)",
1130 : standbySnapshotPendingXmin,
1131 : running->oldestRunningXid);
1132 0 : return;
1133 : }
1134 : }
1135 :
1136 : Assert(standbyState == STANDBY_INITIALIZED);
1137 :
1138 : /*
1139 : * NB: this can be reached at least twice, so make sure new code can deal
1140 : * with that.
1141 : */
1142 :
1143 : /*
1144 : * Nobody else is running yet, but take locks anyhow
1145 : */
1146 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1147 :
1148 : /*
1149 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1150 : * sort them first.
1151 : *
1152 : * Some of the new xids are top-level xids and some are subtransactions.
1153 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1154 : * aren't overflowed then all xids will fit in snapshot and so we don't
1155 : * need subtrans. If we later overflow, an xid assignment record will add
1156 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1157 : * don't have enough information to correctly update subtrans anyway.
1158 : */
1159 :
1160 : /*
1161 : * Allocate a temporary array to avoid modifying the array passed as
1162 : * argument.
1163 : */
1164 114 : xids = palloc_array(TransactionId, running->xcnt + running->subxcnt);
1165 :
1166 : /*
1167 : * Add to the temp array any xids which have not already completed.
1168 : */
1169 114 : nxids = 0;
1170 119 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1171 : {
1172 5 : TransactionId xid = running->xids[i];
1173 :
1174 : /*
1175 : * The running-xacts snapshot can contain xids that were still visible
1176 : * in the procarray when the snapshot was taken, but were already
1177 : * WAL-logged as completed. They're not running anymore, so ignore
1178 : * them.
1179 : */
1180 5 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1181 0 : continue;
1182 :
1183 5 : xids[nxids++] = xid;
1184 : }
1185 :
1186 114 : if (nxids > 0)
1187 : {
1188 5 : if (procArray->numKnownAssignedXids != 0)
1189 : {
1190 0 : LWLockRelease(ProcArrayLock);
1191 0 : elog(ERROR, "KnownAssignedXids is not empty");
1192 : }
1193 :
1194 : /*
1195 : * Sort the array so that we can add them safely into
1196 : * KnownAssignedXids.
1197 : *
1198 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1199 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1200 : * come from RUNNING_XACTS, which means there are only normal XIDs
1201 : * from the same epoch, so this is safe.
1202 : */
1203 5 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1204 :
1205 : /*
1206 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1207 : * snapshot may include duplicated xids because of prepared
1208 : * transactions, so ignore them.
1209 : */
1210 10 : for (i = 0; i < nxids; i++)
1211 : {
1212 5 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1213 : {
1214 0 : elog(DEBUG1,
1215 : "found duplicated transaction %u for KnownAssignedXids insertion",
1216 : xids[i]);
1217 0 : continue;
1218 : }
1219 5 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1220 : }
1221 :
1222 5 : KnownAssignedXidsDisplay(DEBUG3);
1223 : }
1224 :
1225 114 : pfree(xids);
1226 :
1227 : /*
1228 : * latestObservedXid is at least set to the point where SUBTRANS was
1229 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1230 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1231 : * subtrans from thereon, up to nextXid - 1.
1232 : *
1233 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1234 : * because we've just added xids to the known assigned xids machinery that
1235 : * haven't gone through RecordKnownAssignedTransactionId().
1236 : */
1237 : Assert(TransactionIdIsNormal(latestObservedXid));
1238 114 : TransactionIdAdvance(latestObservedXid);
1239 228 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1240 : {
1241 0 : ExtendSUBTRANS(latestObservedXid);
1242 0 : TransactionIdAdvance(latestObservedXid);
1243 : }
1244 114 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1245 :
1246 : /* ----------
1247 : * Now we've got the running xids we need to set the global values that
1248 : * are used to track snapshots as they evolve further.
1249 : *
1250 : * - latestCompletedXid which will be the xmax for snapshots
1251 : * - lastOverflowedXid which shows whether snapshots overflow
1252 : * - nextXid
1253 : *
1254 : * If the snapshot overflowed, then we still initialise with what we know,
1255 : * but the recovery snapshot isn't fully valid yet because we know there
1256 : * are some subxids missing. We don't know the specific subxids that are
1257 : * missing, so conservatively assume the last one is latestObservedXid.
1258 : * ----------
1259 : */
1260 114 : if (running->subxid_status == SUBXIDS_MISSING)
1261 : {
1262 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1263 :
1264 0 : standbySnapshotPendingXmin = latestObservedXid;
1265 0 : procArray->lastOverflowedXid = latestObservedXid;
1266 : }
1267 : else
1268 : {
1269 114 : standbyState = STANDBY_SNAPSHOT_READY;
1270 :
1271 114 : standbySnapshotPendingXmin = InvalidTransactionId;
1272 :
1273 : /*
1274 : * If the 'xids' array didn't include all subtransactions, we have to
1275 : * mark any snapshots taken as overflowed.
1276 : */
1277 114 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1278 26 : procArray->lastOverflowedXid = latestObservedXid;
1279 : else
1280 : {
1281 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1282 88 : procArray->lastOverflowedXid = InvalidTransactionId;
1283 : }
1284 : }
1285 :
1286 : /*
1287 : * If a transaction wrote a commit record in the gap between taking and
1288 : * logging the snapshot then latestCompletedXid may already be higher than
1289 : * the value from the snapshot, so check before we use the incoming value.
1290 : * It also might not yet be set at all.
1291 : */
1292 114 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1293 :
1294 : /*
1295 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1296 : * nobody can see it yet.
1297 : */
1298 :
1299 114 : LWLockRelease(ProcArrayLock);
1300 :
1301 114 : KnownAssignedXidsDisplay(DEBUG3);
1302 114 : if (standbyState == STANDBY_SNAPSHOT_READY)
1303 114 : elog(DEBUG1, "recovery snapshots are now enabled");
1304 : else
1305 0 : elog(DEBUG1,
1306 : "recovery snapshot waiting for non-overflowed snapshot or "
1307 : "until oldest active xid on standby is at least %u (now %u)",
1308 : standbySnapshotPendingXmin,
1309 : running->oldestRunningXid);
1310 : }
1311 :
1312 : /*
1313 : * ProcArrayApplyXidAssignment
1314 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1315 : */
1316 : void
1317 21 : ProcArrayApplyXidAssignment(TransactionId topxid,
1318 : int nsubxids, TransactionId *subxids)
1319 : {
1320 : TransactionId max_xid;
1321 : int i;
1322 :
1323 : Assert(standbyState >= STANDBY_INITIALIZED);
1324 :
1325 21 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1326 :
1327 : /*
1328 : * Mark all the subtransactions as observed.
1329 : *
1330 : * NOTE: This will fail if the subxid contains too many previously
1331 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1332 : * as the code stands, because xid-assignment records should never contain
1333 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1334 : */
1335 21 : RecordKnownAssignedTransactionIds(max_xid);
1336 :
1337 : /*
1338 : * Notice that we update pg_subtrans with the top-level xid, rather than
1339 : * the parent xid. This is a difference between normal processing and
1340 : * recovery, yet is still correct in all cases. The reason is that
1341 : * subtransaction commit is not marked in clog until commit processing, so
1342 : * all aborted subtransactions have already been clearly marked in clog.
1343 : * As a result we are able to refer directly to the top-level
1344 : * transaction's state rather than skipping through all the intermediate
1345 : * states in the subtransaction tree. This should be the first time we
1346 : * have attempted to SubTransSetParent().
1347 : */
1348 1365 : for (i = 0; i < nsubxids; i++)
1349 1344 : SubTransSetParent(subxids[i], topxid);
1350 :
1351 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1352 21 : if (standbyState == STANDBY_INITIALIZED)
1353 0 : return;
1354 :
1355 : /*
1356 : * Uses same locking as transaction commit
1357 : */
1358 21 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1359 :
1360 : /*
1361 : * Remove subxids from known-assigned-xacts.
1362 : */
1363 21 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1364 :
1365 : /*
1366 : * Advance lastOverflowedXid to be at least the last of these subxids.
1367 : */
1368 21 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1369 21 : procArray->lastOverflowedXid = max_xid;
1370 :
1371 21 : LWLockRelease(ProcArrayLock);
1372 : }
1373 :
1374 : /*
1375 : * TransactionIdIsInProgress -- is given transaction running in some backend
1376 : *
1377 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1378 : * there are four possibilities for finding a running transaction:
1379 : *
1380 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1381 : * by looking at ProcGlobal->xids.
1382 : *
1383 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1384 : * We can find this out cheaply too.
1385 : *
1386 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1387 : * if the Xid is running on the primary.
1388 : *
1389 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1390 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1391 : * This is the slowest way, but sadly it has to be done always if the others
1392 : * failed, unless we see that the cached subxact sets are complete (none have
1393 : * overflowed).
1394 : *
1395 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1396 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1397 : * This buys back some concurrency (and we can't retrieve the main Xids from
1398 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1399 : */
1400 : bool
1401 13696294 : TransactionIdIsInProgress(TransactionId xid)
1402 : {
1403 : static TransactionId *xids = NULL;
1404 : static TransactionId *other_xids;
1405 : XidCacheStatus *other_subxidstates;
1406 13696294 : int nxids = 0;
1407 13696294 : ProcArrayStruct *arrayP = procArray;
1408 : TransactionId topxid;
1409 : TransactionId latestCompletedXid;
1410 : int mypgxactoff;
1411 : int numProcs;
1412 : int j;
1413 :
1414 : /*
1415 : * Don't bother checking a transaction older than RecentXmin; it could not
1416 : * possibly still be running. (Note: in particular, this guarantees that
1417 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1418 : * running.)
1419 : */
1420 13696294 : if (TransactionIdPrecedes(xid, RecentXmin))
1421 : {
1422 : xc_by_recent_xmin_inc();
1423 5959076 : return false;
1424 : }
1425 :
1426 : /*
1427 : * We may have just checked the status of this transaction, so if it is
1428 : * already known to be completed, we can fall out without any access to
1429 : * shared memory.
1430 : */
1431 7737218 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1432 : {
1433 : xc_by_known_xact_inc();
1434 1253816 : return false;
1435 : }
1436 :
1437 : /*
1438 : * Also, we can handle our own transaction (and subtransactions) without
1439 : * any access to shared memory.
1440 : */
1441 6483402 : if (TransactionIdIsCurrentTransactionId(xid))
1442 : {
1443 : xc_by_my_xact_inc();
1444 198024 : return true;
1445 : }
1446 :
1447 : /*
1448 : * If first time through, get workspace to remember main XIDs in. We
1449 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1450 : */
1451 6285378 : if (xids == NULL)
1452 : {
1453 : /*
1454 : * In hot standby mode, reserve enough space to hold all xids in the
1455 : * known-assigned list. If we later finish recovery, we no longer need
1456 : * the bigger array, but we don't bother to shrink it.
1457 : */
1458 1661 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1459 :
1460 1661 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1461 1661 : if (xids == NULL)
1462 0 : ereport(ERROR,
1463 : (errcode(ERRCODE_OUT_OF_MEMORY),
1464 : errmsg("out of memory")));
1465 : }
1466 :
1467 6285378 : other_xids = ProcGlobal->xids;
1468 6285378 : other_subxidstates = ProcGlobal->subxidStates;
1469 :
1470 6285378 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1471 :
1472 : /*
1473 : * Now that we have the lock, we can check latestCompletedXid; if the
1474 : * target Xid is after that, it's surely still running.
1475 : */
1476 6285378 : latestCompletedXid =
1477 6285378 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1478 6285378 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1479 : {
1480 1509118 : LWLockRelease(ProcArrayLock);
1481 : xc_by_latest_xid_inc();
1482 1509118 : return true;
1483 : }
1484 :
1485 : /* No shortcuts, gotta grovel through the array */
1486 4776260 : mypgxactoff = MyProc->pgxactoff;
1487 4776260 : numProcs = arrayP->numProcs;
1488 4952112 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1489 : {
1490 : int pgprocno;
1491 : PGPROC *proc;
1492 : TransactionId pxid;
1493 : int pxids;
1494 :
1495 : /* Ignore ourselves --- dealt with it above */
1496 4936785 : if (pgxactoff == mypgxactoff)
1497 16846 : continue;
1498 :
1499 : /* Fetch xid just once - see GetNewTransactionId */
1500 4919939 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1501 :
1502 4919939 : if (!TransactionIdIsValid(pxid))
1503 117393 : continue;
1504 :
1505 : /*
1506 : * Step 1: check the main Xid
1507 : */
1508 4802546 : if (TransactionIdEquals(pxid, xid))
1509 : {
1510 4760797 : LWLockRelease(ProcArrayLock);
1511 : xc_by_main_xid_inc();
1512 4760797 : return true;
1513 : }
1514 :
1515 : /*
1516 : * We can ignore main Xids that are younger than the target Xid, since
1517 : * the target could not possibly be their child.
1518 : */
1519 41749 : if (TransactionIdPrecedes(xid, pxid))
1520 19492 : continue;
1521 :
1522 : /*
1523 : * Step 2: check the cached child-Xids arrays
1524 : */
1525 22257 : pxids = other_subxidstates[pgxactoff].count;
1526 22257 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1527 22257 : pgprocno = arrayP->pgprocnos[pgxactoff];
1528 22257 : proc = &allProcs[pgprocno];
1529 40772 : for (j = pxids - 1; j >= 0; j--)
1530 : {
1531 : /* Fetch xid just once - see GetNewTransactionId */
1532 18651 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1533 :
1534 18651 : if (TransactionIdEquals(cxid, xid))
1535 : {
1536 136 : LWLockRelease(ProcArrayLock);
1537 : xc_by_child_xid_inc();
1538 136 : return true;
1539 : }
1540 : }
1541 :
1542 : /*
1543 : * Save the main Xid for step 4. We only need to remember main Xids
1544 : * that have uncached children. (Note: there is no race condition
1545 : * here because the overflowed flag cannot be cleared, only set, while
1546 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1547 : * worry about.)
1548 : */
1549 22121 : if (other_subxidstates[pgxactoff].overflowed)
1550 220 : xids[nxids++] = pxid;
1551 : }
1552 :
1553 : /*
1554 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1555 : * in the list must be treated as running.
1556 : */
1557 15327 : if (RecoveryInProgress())
1558 : {
1559 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1560 : Assert(nxids == 0);
1561 :
1562 1 : if (KnownAssignedXidExists(xid))
1563 : {
1564 0 : LWLockRelease(ProcArrayLock);
1565 : xc_by_known_assigned_inc();
1566 0 : return true;
1567 : }
1568 :
1569 : /*
1570 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1571 : * too. Fetch all xids from KnownAssignedXids that are lower than
1572 : * xid, since if xid is a subtransaction its parent will always have a
1573 : * lower value. Note we will collect both main and subXIDs here, but
1574 : * there's no help for it.
1575 : */
1576 1 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1577 0 : nxids = KnownAssignedXidsGet(xids, xid);
1578 : }
1579 :
1580 15327 : LWLockRelease(ProcArrayLock);
1581 :
1582 : /*
1583 : * If none of the relevant caches overflowed, we know the Xid is not
1584 : * running without even looking at pg_subtrans.
1585 : */
1586 15327 : if (nxids == 0)
1587 : {
1588 : xc_no_overflow_inc();
1589 15107 : cachedXidIsNotInProgress = xid;
1590 15107 : return false;
1591 : }
1592 :
1593 : /*
1594 : * Step 4: have to check pg_subtrans.
1595 : *
1596 : * At this point, we know it's either a subtransaction of one of the Xids
1597 : * in xids[], or it's not running. If it's an already-failed
1598 : * subtransaction, we want to say "not running" even though its parent may
1599 : * still be running. So first, check pg_xact to see if it's been aborted.
1600 : */
1601 : xc_slow_answer_inc();
1602 :
1603 220 : if (TransactionIdDidAbort(xid))
1604 : {
1605 0 : cachedXidIsNotInProgress = xid;
1606 0 : return false;
1607 : }
1608 :
1609 : /*
1610 : * It isn't aborted, so check whether the transaction tree it belongs to
1611 : * is still running (or, more precisely, whether it was running when we
1612 : * held ProcArrayLock).
1613 : */
1614 220 : topxid = SubTransGetTopmostTransaction(xid);
1615 : Assert(TransactionIdIsValid(topxid));
1616 377 : if (!TransactionIdEquals(topxid, xid) &&
1617 157 : pg_lfind32(topxid, xids, nxids))
1618 157 : return true;
1619 :
1620 63 : cachedXidIsNotInProgress = xid;
1621 63 : return false;
1622 : }
1623 :
1624 :
1625 : /*
1626 : * Determine XID horizons.
1627 : *
1628 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1629 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1630 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1631 : * GlobalVisState).
1632 : *
1633 : * See the definition of ComputeXidHorizonsResult for the various computed
1634 : * horizons.
1635 : *
1636 : * For VACUUM separate horizons (used to decide which deleted tuples must
1637 : * be preserved), for shared and non-shared tables are computed. For shared
1638 : * relations backends in all databases must be considered, but for non-shared
1639 : * relations that's not required, since only backends in my own database could
1640 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1641 : * VACUUMs because (a) they must be working on other tables, and (b) they
1642 : * don't need to do snapshot-based lookups.
1643 : *
1644 : * This also computes a horizon used to truncate pg_subtrans. For that
1645 : * backends in all databases have to be considered, and concurrently running
1646 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1647 : * accesses.
1648 : *
1649 : * Note: we include all currently running xids in the set of considered xids.
1650 : * This ensures that if a just-started xact has not yet set its snapshot,
1651 : * when it does set the snapshot it cannot set xmin less than what we compute.
1652 : * See notes in src/backend/access/transam/README.
1653 : *
1654 : * Note: despite the above, it's possible for the calculated values to move
1655 : * backwards on repeated calls. The calculated values are conservative, so
1656 : * that anything older is definitely not considered as running by anyone
1657 : * anymore, but the exact values calculated depend on a number of things. For
1658 : * example, if there are no transactions running in the current database, the
1659 : * horizon for normal tables will be latestCompletedXid. If a transaction
1660 : * begins after that, its xmin will include in-progress transactions in other
1661 : * databases that started earlier, so another call will return a lower value.
1662 : * Nonetheless it is safe to vacuum a table in the current database with the
1663 : * first result. There are also replication-related effects: a walsender
1664 : * process can set its xmin based on transactions that are no longer running
1665 : * on the primary but are still being replayed on the standby, thus possibly
1666 : * making the values go backwards. In this case there is a possibility that
1667 : * we lose data that the standby would like to have, but unless the standby
1668 : * uses a replication slot to make its xmin persistent there is little we can
1669 : * do about that --- data is only protected if the walsender runs continuously
1670 : * while queries are executed on the standby. (The Hot Standby code deals
1671 : * with such cases by failing standby queries that needed to access
1672 : * already-removed data, so there's no integrity bug.)
1673 : *
1674 : * Note: the approximate horizons (see definition of GlobalVisState) are
1675 : * updated by the computations done here. That's currently required for
1676 : * correctness and a small optimization. Without doing so it's possible that
1677 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1678 : * horizon than later when deciding which tuples can be removed - which the
1679 : * code doesn't expect (breaking HOT).
1680 : */
1681 : static void
1682 214267 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1683 : {
1684 214267 : ProcArrayStruct *arrayP = procArray;
1685 : TransactionId kaxmin;
1686 214267 : bool in_recovery = RecoveryInProgress();
1687 214267 : TransactionId *other_xids = ProcGlobal->xids;
1688 :
1689 : /* inferred after ProcArrayLock is released */
1690 214267 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1691 :
1692 214267 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1693 :
1694 214267 : h->latest_completed = TransamVariables->latestCompletedXid;
1695 :
1696 : /*
1697 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1698 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1699 : * and so protects us against overestimating the result due to future
1700 : * additions.
1701 : */
1702 : {
1703 : TransactionId initial;
1704 :
1705 214267 : initial = XidFromFullTransactionId(h->latest_completed);
1706 : Assert(TransactionIdIsValid(initial));
1707 214267 : TransactionIdAdvance(initial);
1708 :
1709 214267 : h->oldest_considered_running = initial;
1710 214267 : h->shared_oldest_nonremovable = initial;
1711 214267 : h->data_oldest_nonremovable = initial;
1712 :
1713 : /*
1714 : * Only modifications made by this backend affect the horizon for
1715 : * temporary relations. Instead of a check in each iteration of the
1716 : * loop over all PGPROCs it is cheaper to just initialize to the
1717 : * current top-level xid any.
1718 : *
1719 : * Without an assigned xid we could use a horizon as aggressive as
1720 : * GetNewTransactionId(), but we can get away with the much cheaper
1721 : * latestCompletedXid + 1: If this backend has no xid there, by
1722 : * definition, can't be any newer changes in the temp table than
1723 : * latestCompletedXid.
1724 : */
1725 214267 : if (TransactionIdIsValid(MyProc->xid))
1726 42307 : h->temp_oldest_nonremovable = MyProc->xid;
1727 : else
1728 171960 : h->temp_oldest_nonremovable = initial;
1729 : }
1730 :
1731 : /*
1732 : * Fetch slot horizons while ProcArrayLock is held - the
1733 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1734 : * the lock.
1735 : */
1736 214267 : h->slot_xmin = procArray->replication_slot_xmin;
1737 214267 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1738 :
1739 1403577 : for (int index = 0; index < arrayP->numProcs; index++)
1740 : {
1741 1189310 : int pgprocno = arrayP->pgprocnos[index];
1742 1189310 : PGPROC *proc = &allProcs[pgprocno];
1743 1189310 : int8 statusFlags = ProcGlobal->statusFlags[index];
1744 : TransactionId xid;
1745 : TransactionId xmin;
1746 :
1747 : /* Fetch xid just once - see GetNewTransactionId */
1748 1189310 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1749 1189310 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1750 :
1751 : /*
1752 : * Consider both the transaction's Xmin, and its Xid.
1753 : *
1754 : * We must check both because a transaction might have an Xmin but not
1755 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1756 : * some not-yet-set Xmin.
1757 : */
1758 1189310 : xmin = TransactionIdOlder(xmin, xid);
1759 :
1760 : /* if neither is set, this proc doesn't influence the horizon */
1761 1189310 : if (!TransactionIdIsValid(xmin))
1762 543818 : continue;
1763 :
1764 : /*
1765 : * Don't ignore any procs when determining which transactions might be
1766 : * considered running. While slots should ensure logical decoding
1767 : * backends are protected even without this check, it can't hurt to
1768 : * include them here as well..
1769 : */
1770 645492 : h->oldest_considered_running =
1771 645492 : TransactionIdOlder(h->oldest_considered_running, xmin);
1772 :
1773 : /*
1774 : * Skip over backends either vacuuming (which is ok with rows being
1775 : * removed, as long as pg_subtrans is not truncated) or doing logical
1776 : * decoding (which manages xmin separately, check below).
1777 : */
1778 645492 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1779 223724 : continue;
1780 :
1781 : /* shared tables need to take backends in all databases into account */
1782 421768 : h->shared_oldest_nonremovable =
1783 421768 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1784 :
1785 : /*
1786 : * Normally sessions in other databases are ignored for anything but
1787 : * the shared horizon.
1788 : *
1789 : * However, include them when MyDatabaseId is not (yet) set. A
1790 : * backend in the process of starting up must not compute a "too
1791 : * aggressive" horizon, otherwise we could end up using it to prune
1792 : * still-needed data away. If the current backend never connects to a
1793 : * database this is harmless, because data_oldest_nonremovable will
1794 : * never be utilized.
1795 : *
1796 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1797 : * be included. (This flag is used for hot standby feedback, which
1798 : * can't be tied to a specific database.)
1799 : *
1800 : * Also, while in recovery we cannot compute an accurate per-database
1801 : * horizon, as all xids are managed via the KnownAssignedXids
1802 : * machinery.
1803 : */
1804 421768 : if (proc->databaseId == MyDatabaseId ||
1805 20455 : MyDatabaseId == InvalidOid ||
1806 12159 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1807 : in_recovery)
1808 : {
1809 409611 : h->data_oldest_nonremovable =
1810 409611 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1811 : }
1812 : }
1813 :
1814 : /*
1815 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1816 : * after lock is released.
1817 : */
1818 214267 : if (in_recovery)
1819 373 : kaxmin = KnownAssignedXidsGetOldestXmin();
1820 :
1821 : /*
1822 : * No other information from shared state is needed, release the lock
1823 : * immediately. The rest of the computations can be done without a lock.
1824 : */
1825 214267 : LWLockRelease(ProcArrayLock);
1826 :
1827 214267 : if (in_recovery)
1828 : {
1829 373 : h->oldest_considered_running =
1830 373 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1831 373 : h->shared_oldest_nonremovable =
1832 373 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1833 373 : h->data_oldest_nonremovable =
1834 373 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1835 : /* temp relations cannot be accessed in recovery */
1836 : }
1837 :
1838 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1839 : h->shared_oldest_nonremovable));
1840 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1841 : h->data_oldest_nonremovable));
1842 :
1843 : /*
1844 : * Check whether there are replication slots requiring an older xmin.
1845 : */
1846 214267 : h->shared_oldest_nonremovable =
1847 214267 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1848 214267 : h->data_oldest_nonremovable =
1849 214267 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1850 :
1851 : /*
1852 : * The only difference between catalog / data horizons is that the slot's
1853 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1854 : * for logical decoding). Initialize with data horizon, and then back up
1855 : * further if necessary. Have to back up the shared horizon as well, since
1856 : * that also can contain catalogs.
1857 : */
1858 214267 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1859 214267 : h->shared_oldest_nonremovable =
1860 214267 : TransactionIdOlder(h->shared_oldest_nonremovable,
1861 : h->slot_catalog_xmin);
1862 214267 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1863 214267 : h->catalog_oldest_nonremovable =
1864 214267 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1865 : h->slot_catalog_xmin);
1866 :
1867 : /*
1868 : * It's possible that slots backed up the horizons further than
1869 : * oldest_considered_running. Fix.
1870 : */
1871 214267 : h->oldest_considered_running =
1872 214267 : TransactionIdOlder(h->oldest_considered_running,
1873 : h->shared_oldest_nonremovable);
1874 214267 : h->oldest_considered_running =
1875 214267 : TransactionIdOlder(h->oldest_considered_running,
1876 : h->catalog_oldest_nonremovable);
1877 214267 : h->oldest_considered_running =
1878 214267 : TransactionIdOlder(h->oldest_considered_running,
1879 : h->data_oldest_nonremovable);
1880 :
1881 : /*
1882 : * shared horizons have to be at least as old as the oldest visible in
1883 : * current db
1884 : */
1885 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1886 : h->data_oldest_nonremovable));
1887 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1888 : h->catalog_oldest_nonremovable));
1889 :
1890 : /*
1891 : * Horizons need to ensure that pg_subtrans access is still possible for
1892 : * the relevant backends.
1893 : */
1894 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1895 : h->shared_oldest_nonremovable));
1896 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1897 : h->catalog_oldest_nonremovable));
1898 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1899 : h->data_oldest_nonremovable));
1900 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1901 : h->temp_oldest_nonremovable));
1902 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1903 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1904 : h->slot_xmin));
1905 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1906 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1907 : h->slot_catalog_xmin));
1908 :
1909 : /* update approximate horizons with the computed horizons */
1910 214267 : GlobalVisUpdateApply(h);
1911 214267 : }
1912 :
1913 : /*
1914 : * Determine what kind of visibility horizon needs to be used for a
1915 : * relation. If rel is NULL, the most conservative horizon is used.
1916 : */
1917 : static inline GlobalVisHorizonKind
1918 18994704 : GlobalVisHorizonKindForRel(Relation rel)
1919 : {
1920 : /*
1921 : * Other relkinds currently don't contain xids, nor always the necessary
1922 : * logical decoding markers.
1923 : */
1924 : Assert(!rel ||
1925 : rel->rd_rel->relkind == RELKIND_RELATION ||
1926 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1927 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1928 :
1929 18994704 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1930 121818 : return VISHORIZON_SHARED;
1931 18872886 : else if (IsCatalogRelation(rel) ||
1932 15065994 : RelationIsAccessibleInLogicalDecoding(rel))
1933 3806896 : return VISHORIZON_CATALOG;
1934 15065990 : else if (!RELATION_IS_LOCAL(rel))
1935 14978107 : return VISHORIZON_DATA;
1936 : else
1937 87883 : return VISHORIZON_TEMP;
1938 : }
1939 :
1940 : /*
1941 : * Return the oldest XID for which deleted tuples must be preserved in the
1942 : * passed table.
1943 : *
1944 : * If rel is not NULL the horizon may be considerably more recent than
1945 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1946 : * that is correct (but not optimal) for all relations will be returned.
1947 : *
1948 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1949 : * the passed in table.
1950 : */
1951 : TransactionId
1952 146303 : GetOldestNonRemovableTransactionId(Relation rel)
1953 : {
1954 : ComputeXidHorizonsResult horizons;
1955 :
1956 146303 : ComputeXidHorizons(&horizons);
1957 :
1958 146303 : switch (GlobalVisHorizonKindForRel(rel))
1959 : {
1960 21022 : case VISHORIZON_SHARED:
1961 21022 : return horizons.shared_oldest_nonremovable;
1962 89575 : case VISHORIZON_CATALOG:
1963 89575 : return horizons.catalog_oldest_nonremovable;
1964 20575 : case VISHORIZON_DATA:
1965 20575 : return horizons.data_oldest_nonremovable;
1966 15131 : case VISHORIZON_TEMP:
1967 15131 : return horizons.temp_oldest_nonremovable;
1968 : }
1969 :
1970 : /* just to prevent compiler warnings */
1971 0 : return InvalidTransactionId;
1972 : }
1973 :
1974 : /*
1975 : * Return the oldest transaction id any currently running backend might still
1976 : * consider running. This should not be used for visibility / pruning
1977 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1978 : * decisions like up to where pg_subtrans can be truncated.
1979 : */
1980 : TransactionId
1981 1805 : GetOldestTransactionIdConsideredRunning(void)
1982 : {
1983 : ComputeXidHorizonsResult horizons;
1984 :
1985 1805 : ComputeXidHorizons(&horizons);
1986 :
1987 1805 : return horizons.oldest_considered_running;
1988 : }
1989 :
1990 : /*
1991 : * Return the visibility horizons for a hot standby feedback message.
1992 : */
1993 : void
1994 51 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1995 : {
1996 : ComputeXidHorizonsResult horizons;
1997 :
1998 51 : ComputeXidHorizons(&horizons);
1999 :
2000 : /*
2001 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2002 : * effect of replication slot's catalog_xmin. We want to send a separate
2003 : * feedback for the catalog horizon, so the primary can remove data table
2004 : * contents more aggressively.
2005 : */
2006 51 : *xmin = horizons.shared_oldest_nonremovable_raw;
2007 51 : *catalog_xmin = horizons.slot_catalog_xmin;
2008 51 : }
2009 :
2010 : /*
2011 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2012 : *
2013 : * We have to export this for use by snapmgr.c.
2014 : */
2015 : int
2016 37009 : GetMaxSnapshotXidCount(void)
2017 : {
2018 37009 : return procArray->maxProcs;
2019 : }
2020 :
2021 : /*
2022 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2023 : *
2024 : * We have to export this for use by snapmgr.c.
2025 : */
2026 : int
2027 36803 : GetMaxSnapshotSubxidCount(void)
2028 : {
2029 36803 : return TOTAL_MAX_CACHED_SUBXIDS;
2030 : }
2031 :
2032 : /*
2033 : * Helper function for GetSnapshotData() that checks if the bulk of the
2034 : * visibility information in the snapshot is still valid. If so, it updates
2035 : * the fields that need to change and returns true. Otherwise it returns
2036 : * false.
2037 : *
2038 : * This very likely can be evolved to not need ProcArrayLock held (at very
2039 : * least in the case we already hold a snapshot), but that's for another day.
2040 : */
2041 : static bool
2042 2949284 : GetSnapshotDataReuse(Snapshot snapshot)
2043 : {
2044 : uint64 curXactCompletionCount;
2045 :
2046 : Assert(LWLockHeldByMe(ProcArrayLock));
2047 :
2048 2949284 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2049 36786 : return false;
2050 :
2051 2912498 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2052 2912498 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2053 445276 : return false;
2054 :
2055 : /*
2056 : * If the current xactCompletionCount is still the same as it was at the
2057 : * time the snapshot was built, we can be sure that rebuilding the
2058 : * contents of the snapshot the hard way would result in the same snapshot
2059 : * contents:
2060 : *
2061 : * As explained in transam/README, the set of xids considered running by
2062 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2063 : * contents only depend on transactions with xids and xactCompletionCount
2064 : * is incremented whenever a transaction with an xid finishes (while
2065 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2066 : * ensures we would detect if the snapshot would have changed.
2067 : *
2068 : * As the snapshot contents are the same as it was before, it is safe to
2069 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2070 : * visible under the snapshot could already have been removed (that'd
2071 : * require the set of running transactions to change) and it fulfills the
2072 : * requirement that concurrent GetSnapshotData() calls yield the same
2073 : * xmin.
2074 : */
2075 2467222 : if (!TransactionIdIsValid(MyProc->xmin))
2076 768849 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2077 :
2078 2467222 : RecentXmin = snapshot->xmin;
2079 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2080 :
2081 2467222 : snapshot->curcid = GetCurrentCommandId(false);
2082 2467222 : snapshot->active_count = 0;
2083 2467222 : snapshot->regd_count = 0;
2084 2467222 : snapshot->copied = false;
2085 :
2086 2467222 : return true;
2087 : }
2088 :
2089 : /*
2090 : * GetSnapshotData -- returns information about running transactions.
2091 : *
2092 : * The returned snapshot includes xmin (lowest still-running xact ID),
2093 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2094 : * in the range xmin <= xid < xmax. It is used as follows:
2095 : * All xact IDs < xmin are considered finished.
2096 : * All xact IDs >= xmax are considered still running.
2097 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2098 : * it is considered running or not.
2099 : * This ensures that the set of transactions seen as "running" by the
2100 : * current xact will not change after it takes the snapshot.
2101 : *
2102 : * All running top-level XIDs are included in the snapshot, except for lazy
2103 : * VACUUM processes. We also try to include running subtransaction XIDs,
2104 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2105 : * information may not be available. If we find any overflowed subxid arrays,
2106 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2107 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2108 : *
2109 : * We also update the following backend-global variables:
2110 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2111 : * current transaction (this is the same as MyProc->xmin).
2112 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2113 : * older than this are known not running any more.
2114 : *
2115 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2116 : * for the benefit of the GlobalVisTest* family of functions.
2117 : *
2118 : * Note: this function should probably not be called with an argument that's
2119 : * not statically allocated (see xip allocation below).
2120 : */
2121 : Snapshot
2122 2949284 : GetSnapshotData(Snapshot snapshot)
2123 : {
2124 2949284 : ProcArrayStruct *arrayP = procArray;
2125 2949284 : TransactionId *other_xids = ProcGlobal->xids;
2126 : TransactionId xmin;
2127 : TransactionId xmax;
2128 2949284 : int count = 0;
2129 2949284 : int subcount = 0;
2130 2949284 : bool suboverflowed = false;
2131 : FullTransactionId latest_completed;
2132 : TransactionId oldestxid;
2133 : int mypgxactoff;
2134 : TransactionId myxid;
2135 : uint64 curXactCompletionCount;
2136 :
2137 2949284 : TransactionId replication_slot_xmin = InvalidTransactionId;
2138 2949284 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2139 :
2140 : Assert(snapshot != NULL);
2141 :
2142 : /*
2143 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2144 : * be sufficient. But it seems better to do the malloc while not holding
2145 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2146 : * more subxip storage than is probably needed.
2147 : *
2148 : * This does open a possibility for avoiding repeated malloc/free: since
2149 : * maxProcs does not change at runtime, we can simply reuse the previous
2150 : * xip arrays if any. (This relies on the fact that all callers pass
2151 : * static SnapshotData structs.)
2152 : */
2153 2949284 : if (snapshot->xip == NULL)
2154 : {
2155 : /*
2156 : * First call for this snapshot. Snapshot is same size whether or not
2157 : * we are in recovery, see later comments.
2158 : */
2159 36778 : snapshot->xip = (TransactionId *)
2160 36778 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2161 36778 : if (snapshot->xip == NULL)
2162 0 : ereport(ERROR,
2163 : (errcode(ERRCODE_OUT_OF_MEMORY),
2164 : errmsg("out of memory")));
2165 : Assert(snapshot->subxip == NULL);
2166 36778 : snapshot->subxip = (TransactionId *)
2167 36778 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2168 36778 : if (snapshot->subxip == NULL)
2169 0 : ereport(ERROR,
2170 : (errcode(ERRCODE_OUT_OF_MEMORY),
2171 : errmsg("out of memory")));
2172 : }
2173 :
2174 : /*
2175 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2176 : * going to set MyProc->xmin.
2177 : */
2178 2949284 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2179 :
2180 2949284 : if (GetSnapshotDataReuse(snapshot))
2181 : {
2182 2467222 : LWLockRelease(ProcArrayLock);
2183 2467222 : return snapshot;
2184 : }
2185 :
2186 482062 : latest_completed = TransamVariables->latestCompletedXid;
2187 482062 : mypgxactoff = MyProc->pgxactoff;
2188 482062 : myxid = other_xids[mypgxactoff];
2189 : Assert(myxid == MyProc->xid);
2190 :
2191 482062 : oldestxid = TransamVariables->oldestXid;
2192 482062 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2193 :
2194 : /* xmax is always latestCompletedXid + 1 */
2195 482062 : xmax = XidFromFullTransactionId(latest_completed);
2196 482062 : TransactionIdAdvance(xmax);
2197 : Assert(TransactionIdIsNormal(xmax));
2198 :
2199 : /* initialize xmin calculation with xmax */
2200 482062 : xmin = xmax;
2201 :
2202 : /* take own xid into account, saves a check inside the loop */
2203 482062 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2204 32660 : xmin = myxid;
2205 :
2206 482062 : snapshot->takenDuringRecovery = RecoveryInProgress();
2207 :
2208 482062 : if (!snapshot->takenDuringRecovery)
2209 : {
2210 480774 : int numProcs = arrayP->numProcs;
2211 480774 : TransactionId *xip = snapshot->xip;
2212 480774 : int *pgprocnos = arrayP->pgprocnos;
2213 480774 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2214 480774 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2215 :
2216 : /*
2217 : * First collect set of pgxactoff/xids that need to be included in the
2218 : * snapshot.
2219 : */
2220 4338826 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2221 : {
2222 : /* Fetch xid just once - see GetNewTransactionId */
2223 3858052 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2224 : uint8 statusFlags;
2225 :
2226 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2227 :
2228 : /*
2229 : * If the transaction has no XID assigned, we can skip it; it
2230 : * won't have sub-XIDs either.
2231 : */
2232 3858052 : if (likely(xid == InvalidTransactionId))
2233 3107772 : continue;
2234 :
2235 : /*
2236 : * We don't include our own XIDs (if any) in the snapshot. It
2237 : * needs to be included in the xmin computation, but we did so
2238 : * outside the loop.
2239 : */
2240 750280 : if (pgxactoff == mypgxactoff)
2241 63165 : continue;
2242 :
2243 : /*
2244 : * The only way we are able to get here with a non-normal xid is
2245 : * during bootstrap - with this backend using
2246 : * BootstrapTransactionId. But the above test should filter that
2247 : * out.
2248 : */
2249 : Assert(TransactionIdIsNormal(xid));
2250 :
2251 : /*
2252 : * If the XID is >= xmax, we can skip it; such transactions will
2253 : * be treated as running anyway (and any sub-XIDs will also be >=
2254 : * xmax).
2255 : */
2256 687115 : if (!NormalTransactionIdPrecedes(xid, xmax))
2257 206511 : continue;
2258 :
2259 : /*
2260 : * Skip over backends doing logical decoding which manages xmin
2261 : * separately (check below) and ones running LAZY VACUUM.
2262 : */
2263 480604 : statusFlags = allStatusFlags[pgxactoff];
2264 480604 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2265 76 : continue;
2266 :
2267 480528 : if (NormalTransactionIdPrecedes(xid, xmin))
2268 292644 : xmin = xid;
2269 :
2270 : /* Add XID to snapshot. */
2271 480528 : xip[count++] = xid;
2272 :
2273 : /*
2274 : * Save subtransaction XIDs if possible (if we've already
2275 : * overflowed, there's no point). Note that the subxact XIDs must
2276 : * be later than their parent, so no need to check them against
2277 : * xmin. We could filter against xmax, but it seems better not to
2278 : * do that much work while holding the ProcArrayLock.
2279 : *
2280 : * The other backend can add more subxids concurrently, but cannot
2281 : * remove any. Hence it's important to fetch nxids just once.
2282 : * Should be safe to use memcpy, though. (We needn't worry about
2283 : * missing any xids added concurrently, because they must postdate
2284 : * xmax.)
2285 : *
2286 : * Again, our own XIDs are not included in the snapshot.
2287 : */
2288 480528 : if (!suboverflowed)
2289 : {
2290 :
2291 480524 : if (subxidStates[pgxactoff].overflowed)
2292 803 : suboverflowed = true;
2293 : else
2294 : {
2295 479721 : int nsubxids = subxidStates[pgxactoff].count;
2296 :
2297 479721 : if (nsubxids > 0)
2298 : {
2299 6092 : int pgprocno = pgprocnos[pgxactoff];
2300 6092 : PGPROC *proc = &allProcs[pgprocno];
2301 :
2302 6092 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2303 :
2304 6092 : memcpy(snapshot->subxip + subcount,
2305 6092 : proc->subxids.xids,
2306 : nsubxids * sizeof(TransactionId));
2307 6092 : subcount += nsubxids;
2308 : }
2309 : }
2310 : }
2311 : }
2312 : }
2313 : else
2314 : {
2315 : /*
2316 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2317 : *
2318 : * We store all xids directly into subxip[]. Here's why:
2319 : *
2320 : * In recovery we don't know which xids are top-level and which are
2321 : * subxacts, a design choice that greatly simplifies xid processing.
2322 : *
2323 : * It seems like we would want to try to put xids into xip[] only, but
2324 : * that is fairly small. We would either need to make that bigger or
2325 : * to increase the rate at which we WAL-log xid assignment; neither is
2326 : * an appealing choice.
2327 : *
2328 : * We could try to store xids into xip[] first and then into subxip[]
2329 : * if there are too many xids. That only works if the snapshot doesn't
2330 : * overflow because we do not search subxip[] in that case. A simpler
2331 : * way is to just store all xids in the subxip array because this is
2332 : * by far the bigger array. We just leave the xip array empty.
2333 : *
2334 : * Either way we need to change the way XidInMVCCSnapshot() works
2335 : * depending upon when the snapshot was taken, or change normal
2336 : * snapshot processing so it matches.
2337 : *
2338 : * Note: It is possible for recovery to end before we finish taking
2339 : * the snapshot, and for newly assigned transaction ids to be added to
2340 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2341 : * those newly added transaction ids would be filtered away, so we
2342 : * need not be concerned about them.
2343 : */
2344 1288 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2345 : xmax);
2346 :
2347 1288 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2348 6 : suboverflowed = true;
2349 : }
2350 :
2351 :
2352 : /*
2353 : * Fetch into local variable while ProcArrayLock is held - the
2354 : * LWLockRelease below is a barrier, ensuring this happens inside the
2355 : * lock.
2356 : */
2357 482062 : replication_slot_xmin = procArray->replication_slot_xmin;
2358 482062 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2359 :
2360 482062 : if (!TransactionIdIsValid(MyProc->xmin))
2361 262197 : MyProc->xmin = TransactionXmin = xmin;
2362 :
2363 482062 : LWLockRelease(ProcArrayLock);
2364 :
2365 : /* maintain state for GlobalVis* */
2366 : {
2367 : TransactionId def_vis_xid;
2368 : TransactionId def_vis_xid_data;
2369 : FullTransactionId def_vis_fxid;
2370 : FullTransactionId def_vis_fxid_data;
2371 : FullTransactionId oldestfxid;
2372 :
2373 : /*
2374 : * Converting oldestXid is only safe when xid horizon cannot advance,
2375 : * i.e. holding locks. While we don't hold the lock anymore, all the
2376 : * necessary data has been gathered with lock held.
2377 : */
2378 482062 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2379 :
2380 : /* Check whether there's a replication slot requiring an older xmin. */
2381 : def_vis_xid_data =
2382 482062 : TransactionIdOlder(xmin, replication_slot_xmin);
2383 :
2384 : /*
2385 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2386 : * if older than this xid.
2387 : */
2388 482062 : def_vis_xid = def_vis_xid_data;
2389 :
2390 : /*
2391 : * Check whether there's a replication slot requiring an older catalog
2392 : * xmin.
2393 : */
2394 : def_vis_xid =
2395 482062 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2396 :
2397 482062 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2398 482062 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2399 :
2400 : /*
2401 : * Check if we can increase upper bound. As a previous
2402 : * GlobalVisUpdate() might have computed more aggressive values, don't
2403 : * overwrite them if so.
2404 : */
2405 : GlobalVisSharedRels.definitely_needed =
2406 482062 : FullTransactionIdNewer(def_vis_fxid,
2407 : GlobalVisSharedRels.definitely_needed);
2408 : GlobalVisCatalogRels.definitely_needed =
2409 482062 : FullTransactionIdNewer(def_vis_fxid,
2410 : GlobalVisCatalogRels.definitely_needed);
2411 : GlobalVisDataRels.definitely_needed =
2412 482062 : FullTransactionIdNewer(def_vis_fxid_data,
2413 : GlobalVisDataRels.definitely_needed);
2414 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2415 482062 : if (TransactionIdIsNormal(myxid))
2416 : GlobalVisTempRels.definitely_needed =
2417 63063 : FullXidRelativeTo(latest_completed, myxid);
2418 : else
2419 : {
2420 418999 : GlobalVisTempRels.definitely_needed = latest_completed;
2421 418999 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2422 : }
2423 :
2424 : /*
2425 : * Check if we know that we can initialize or increase the lower
2426 : * bound. Currently the only cheap way to do so is to use
2427 : * TransamVariables->oldestXid as input.
2428 : *
2429 : * We should definitely be able to do better. We could e.g. put a
2430 : * global lower bound value into TransamVariables.
2431 : */
2432 : GlobalVisSharedRels.maybe_needed =
2433 482062 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2434 : oldestfxid);
2435 : GlobalVisCatalogRels.maybe_needed =
2436 482062 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2437 : oldestfxid);
2438 : GlobalVisDataRels.maybe_needed =
2439 482062 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2440 : oldestfxid);
2441 : /* accurate value known */
2442 482062 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2443 : }
2444 :
2445 482062 : RecentXmin = xmin;
2446 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2447 :
2448 482062 : snapshot->xmin = xmin;
2449 482062 : snapshot->xmax = xmax;
2450 482062 : snapshot->xcnt = count;
2451 482062 : snapshot->subxcnt = subcount;
2452 482062 : snapshot->suboverflowed = suboverflowed;
2453 482062 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2454 :
2455 482062 : snapshot->curcid = GetCurrentCommandId(false);
2456 :
2457 : /*
2458 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2459 : * not copied in persistent memory.
2460 : */
2461 482062 : snapshot->active_count = 0;
2462 482062 : snapshot->regd_count = 0;
2463 482062 : snapshot->copied = false;
2464 :
2465 482062 : return snapshot;
2466 : }
2467 :
2468 : /*
2469 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2470 : *
2471 : * This is called when installing a snapshot imported from another
2472 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2473 : * check that the source transaction is still running, and we'd better do
2474 : * that atomically with installing the new xmin.
2475 : *
2476 : * Returns true if successful, false if source xact is no longer running.
2477 : */
2478 : bool
2479 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2480 : VirtualTransactionId *sourcevxid)
2481 : {
2482 16 : bool result = false;
2483 16 : ProcArrayStruct *arrayP = procArray;
2484 : int index;
2485 :
2486 : Assert(TransactionIdIsNormal(xmin));
2487 16 : if (!sourcevxid)
2488 0 : return false;
2489 :
2490 : /* Get lock so source xact can't end while we're doing this */
2491 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2492 :
2493 : /*
2494 : * Find the PGPROC entry of the source transaction. (This could use
2495 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2496 : * performance critical.)
2497 : */
2498 16 : for (index = 0; index < arrayP->numProcs; index++)
2499 : {
2500 16 : int pgprocno = arrayP->pgprocnos[index];
2501 16 : PGPROC *proc = &allProcs[pgprocno];
2502 16 : int statusFlags = ProcGlobal->statusFlags[index];
2503 : TransactionId xid;
2504 :
2505 : /* Ignore procs running LAZY VACUUM */
2506 16 : if (statusFlags & PROC_IN_VACUUM)
2507 0 : continue;
2508 :
2509 : /* We are only interested in the specific virtual transaction. */
2510 16 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2511 0 : continue;
2512 16 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2513 0 : continue;
2514 :
2515 : /*
2516 : * We check the transaction's database ID for paranoia's sake: if it's
2517 : * in another DB then its xmin does not cover us. Caller should have
2518 : * detected this already, so we just treat any funny cases as
2519 : * "transaction not found".
2520 : */
2521 16 : if (proc->databaseId != MyDatabaseId)
2522 0 : continue;
2523 :
2524 : /*
2525 : * Likewise, let's just make real sure its xmin does cover us.
2526 : */
2527 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2528 16 : if (!TransactionIdIsNormal(xid) ||
2529 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2530 0 : continue;
2531 :
2532 : /*
2533 : * We're good. Install the new xmin. As in GetSnapshotData, set
2534 : * TransactionXmin too. (Note that because snapmgr.c called
2535 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2536 : * we don't check that.)
2537 : */
2538 16 : MyProc->xmin = TransactionXmin = xmin;
2539 :
2540 16 : result = true;
2541 16 : break;
2542 : }
2543 :
2544 16 : LWLockRelease(ProcArrayLock);
2545 :
2546 16 : return result;
2547 : }
2548 :
2549 : /*
2550 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2551 : *
2552 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2553 : * PGPROC of the transaction from which we imported the snapshot, rather than
2554 : * an XID.
2555 : *
2556 : * Note that this function also copies statusFlags from the source `proc` in
2557 : * order to avoid the case where MyProc's xmin needs to be skipped for
2558 : * computing xid horizon.
2559 : *
2560 : * Returns true if successful, false if source xact is no longer running.
2561 : */
2562 : bool
2563 2210 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2564 : {
2565 2210 : bool result = false;
2566 : TransactionId xid;
2567 :
2568 : Assert(TransactionIdIsNormal(xmin));
2569 : Assert(proc != NULL);
2570 :
2571 : /*
2572 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2573 : */
2574 2210 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2575 :
2576 : /*
2577 : * Be certain that the referenced PGPROC has an advertised xmin which is
2578 : * no later than the one we're installing, so that the system-wide xmin
2579 : * can't go backwards. Also, make sure it's running in the same database,
2580 : * so that the per-database xmin cannot go backwards.
2581 : */
2582 2210 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2583 2210 : if (proc->databaseId == MyDatabaseId &&
2584 2210 : TransactionIdIsNormal(xid) &&
2585 2210 : TransactionIdPrecedesOrEquals(xid, xmin))
2586 : {
2587 : /*
2588 : * Install xmin and propagate the statusFlags that affect how the
2589 : * value is interpreted by vacuum.
2590 : */
2591 2210 : MyProc->xmin = TransactionXmin = xmin;
2592 2210 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2593 2210 : (proc->statusFlags & PROC_XMIN_FLAGS);
2594 2210 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2595 :
2596 2210 : result = true;
2597 : }
2598 :
2599 2210 : LWLockRelease(ProcArrayLock);
2600 :
2601 2210 : return result;
2602 : }
2603 :
2604 : /*
2605 : * GetRunningTransactionData -- returns information about running transactions.
2606 : *
2607 : * Similar to GetSnapshotData but returns more information. We include
2608 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2609 : * prepared transactions.
2610 : *
2611 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2612 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2613 : * array until the caller has WAL-logged this snapshot, and releases the
2614 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2615 : * lock is released.
2616 : *
2617 : * The returned data structure is statically allocated; caller should not
2618 : * modify it, and must not assume it is valid past the next call.
2619 : *
2620 : * This is never executed during recovery so there is no need to look at
2621 : * KnownAssignedXids.
2622 : *
2623 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2624 : * may return entries with duplicated TransactionId values coming from
2625 : * transaction finishing to prepare. Nothing is done about duplicated
2626 : * entries here to not hold on ProcArrayLock more than necessary.
2627 : *
2628 : * We don't worry about updating other counters, we want to keep this as
2629 : * simple as possible and leave GetSnapshotData() as the primary code for
2630 : * that bookkeeping.
2631 : *
2632 : * Note that if any transaction has overflowed its cached subtransactions
2633 : * then there is no real need include any subtransactions.
2634 : */
2635 : RunningTransactions
2636 1467 : GetRunningTransactionData(void)
2637 : {
2638 : /* result workspace */
2639 : static RunningTransactionsData CurrentRunningXactsData;
2640 :
2641 1467 : ProcArrayStruct *arrayP = procArray;
2642 1467 : TransactionId *other_xids = ProcGlobal->xids;
2643 1467 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2644 : TransactionId latestCompletedXid;
2645 : TransactionId oldestRunningXid;
2646 : TransactionId oldestDatabaseRunningXid;
2647 : TransactionId *xids;
2648 : int index;
2649 : int count;
2650 : int subcount;
2651 : bool suboverflowed;
2652 :
2653 : Assert(!RecoveryInProgress());
2654 :
2655 : /*
2656 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2657 : * be sufficient. But it seems better to do the malloc while not holding
2658 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2659 : * more subxip storage than is probably needed.
2660 : *
2661 : * Should only be allocated in bgwriter, since only ever executed during
2662 : * checkpoints.
2663 : */
2664 1467 : if (CurrentRunningXacts->xids == NULL)
2665 : {
2666 : /*
2667 : * First call
2668 : */
2669 588 : CurrentRunningXacts->xids = (TransactionId *)
2670 588 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2671 588 : if (CurrentRunningXacts->xids == NULL)
2672 0 : ereport(ERROR,
2673 : (errcode(ERRCODE_OUT_OF_MEMORY),
2674 : errmsg("out of memory")));
2675 : }
2676 :
2677 1467 : xids = CurrentRunningXacts->xids;
2678 :
2679 1467 : count = subcount = 0;
2680 1467 : suboverflowed = false;
2681 :
2682 : /*
2683 : * Ensure that no xids enter or leave the procarray while we obtain
2684 : * snapshot.
2685 : */
2686 1467 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2687 1467 : LWLockAcquire(XidGenLock, LW_SHARED);
2688 :
2689 1467 : latestCompletedXid =
2690 1467 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2691 1467 : oldestDatabaseRunningXid = oldestRunningXid =
2692 1467 : XidFromFullTransactionId(TransamVariables->nextXid);
2693 :
2694 : /*
2695 : * Spin over procArray collecting all xids
2696 : */
2697 7311 : for (index = 0; index < arrayP->numProcs; index++)
2698 : {
2699 : TransactionId xid;
2700 :
2701 : /* Fetch xid just once - see GetNewTransactionId */
2702 5844 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2703 :
2704 : /*
2705 : * We don't need to store transactions that don't have a TransactionId
2706 : * yet because they will not show as running on a standby server.
2707 : */
2708 5844 : if (!TransactionIdIsValid(xid))
2709 4990 : continue;
2710 :
2711 : /*
2712 : * Be careful not to exclude any xids before calculating the values of
2713 : * oldestRunningXid and suboverflowed, since these are used to clean
2714 : * up transaction information held on standbys.
2715 : */
2716 854 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2717 576 : oldestRunningXid = xid;
2718 :
2719 : /*
2720 : * Also, update the oldest running xid within the current database. As
2721 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2722 : * TransactionId comparison first.
2723 : */
2724 854 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2725 : {
2726 854 : int pgprocno = arrayP->pgprocnos[index];
2727 854 : PGPROC *proc = &allProcs[pgprocno];
2728 :
2729 854 : if (proc->databaseId == MyDatabaseId)
2730 211 : oldestDatabaseRunningXid = xid;
2731 : }
2732 :
2733 854 : if (ProcGlobal->subxidStates[index].overflowed)
2734 2 : suboverflowed = true;
2735 :
2736 : /*
2737 : * If we wished to exclude xids this would be the right place for it.
2738 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2739 : * but they do during truncation at the end when they get the lock and
2740 : * truncate, so it is not much of a problem to include them if they
2741 : * are seen and it is cleaner to include them.
2742 : */
2743 :
2744 854 : xids[count++] = xid;
2745 : }
2746 :
2747 : /*
2748 : * Spin over procArray collecting all subxids, but only if there hasn't
2749 : * been a suboverflow.
2750 : */
2751 1467 : if (!suboverflowed)
2752 : {
2753 1465 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2754 :
2755 7303 : for (index = 0; index < arrayP->numProcs; index++)
2756 : {
2757 5838 : int pgprocno = arrayP->pgprocnos[index];
2758 5838 : PGPROC *proc = &allProcs[pgprocno];
2759 : int nsubxids;
2760 :
2761 : /*
2762 : * Save subtransaction XIDs. Other backends can't add or remove
2763 : * entries while we're holding XidGenLock.
2764 : */
2765 5838 : nsubxids = other_subxidstates[index].count;
2766 5838 : if (nsubxids > 0)
2767 : {
2768 : /* barrier not really required, as XidGenLock is held, but ... */
2769 18 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2770 :
2771 18 : memcpy(&xids[count], proc->subxids.xids,
2772 : nsubxids * sizeof(TransactionId));
2773 18 : count += nsubxids;
2774 18 : subcount += nsubxids;
2775 :
2776 : /*
2777 : * Top-level XID of a transaction is always less than any of
2778 : * its subxids, so we don't need to check if any of the
2779 : * subxids are smaller than oldestRunningXid
2780 : */
2781 : }
2782 : }
2783 : }
2784 :
2785 : /*
2786 : * It's important *not* to include the limits set by slots here because
2787 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2788 : * were to be included here the initial value could never increase because
2789 : * of a circular dependency where slots only increase their limits when
2790 : * running xacts increases oldestRunningXid and running xacts only
2791 : * increases if slots do.
2792 : */
2793 :
2794 1467 : CurrentRunningXacts->xcnt = count - subcount;
2795 1467 : CurrentRunningXacts->subxcnt = subcount;
2796 1467 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2797 1467 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2798 1467 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2799 1467 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2800 1467 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2801 :
2802 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2803 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2804 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2805 :
2806 : /* We don't release the locks here, the caller is responsible for that */
2807 :
2808 1467 : return CurrentRunningXacts;
2809 : }
2810 :
2811 : /*
2812 : * GetOldestActiveTransactionId()
2813 : *
2814 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2815 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2816 : *
2817 : * If allDbs is true, we look at all databases, though there is no need to
2818 : * include WALSender since this has no effect on hot standby conflicts. If
2819 : * allDbs is false, skip processes attached to other databases.
2820 : *
2821 : * This is never executed during recovery so there is no need to look at
2822 : * KnownAssignedXids.
2823 : *
2824 : * We don't worry about updating other counters, we want to keep this as
2825 : * simple as possible and leave GetSnapshotData() as the primary code for
2826 : * that bookkeeping.
2827 : *
2828 : * inCommitOnly indicates getting the oldestActiveXid among the transactions
2829 : * in the commit critical section.
2830 : */
2831 : TransactionId
2832 5630 : GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2833 : {
2834 5630 : ProcArrayStruct *arrayP = procArray;
2835 5630 : TransactionId *other_xids = ProcGlobal->xids;
2836 : TransactionId oldestRunningXid;
2837 : int index;
2838 :
2839 : Assert(!RecoveryInProgress());
2840 :
2841 : /*
2842 : * Read nextXid, as the upper bound of what's still active.
2843 : *
2844 : * Reading a TransactionId is atomic, but we must grab the lock to make
2845 : * sure that all XIDs < nextXid are already present in the proc array (or
2846 : * have already completed), when we spin over it.
2847 : */
2848 5630 : LWLockAcquire(XidGenLock, LW_SHARED);
2849 5630 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2850 5630 : LWLockRelease(XidGenLock);
2851 :
2852 : /*
2853 : * Spin over procArray collecting all xids and subxids.
2854 : */
2855 5630 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2856 32786 : for (index = 0; index < arrayP->numProcs; index++)
2857 : {
2858 : TransactionId xid;
2859 27156 : int pgprocno = arrayP->pgprocnos[index];
2860 27156 : PGPROC *proc = &allProcs[pgprocno];
2861 :
2862 : /* Fetch xid just once - see GetNewTransactionId */
2863 27156 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2864 :
2865 27156 : if (!TransactionIdIsNormal(xid))
2866 21860 : continue;
2867 :
2868 5296 : if (inCommitOnly &&
2869 4459 : (proc->delayChkptFlags & DELAY_CHKPT_IN_COMMIT) == 0)
2870 4115 : continue;
2871 :
2872 1181 : if (!allDbs && proc->databaseId != MyDatabaseId)
2873 0 : continue;
2874 :
2875 1181 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2876 909 : oldestRunningXid = xid;
2877 :
2878 : /*
2879 : * Top-level XID of a transaction is always less than any of its
2880 : * subxids, so we don't need to check if any of the subxids are
2881 : * smaller than oldestRunningXid
2882 : */
2883 : }
2884 5630 : LWLockRelease(ProcArrayLock);
2885 :
2886 5630 : return oldestRunningXid;
2887 : }
2888 :
2889 : /*
2890 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2891 : *
2892 : * Returns the oldest xid that we can guarantee not to have been affected by
2893 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2894 : * transaction aborted. Note that the value can (and most of the time will) be
2895 : * much more conservative than what really has been affected by vacuum, but we
2896 : * currently don't have better data available.
2897 : *
2898 : * This is useful to initialize the cutoff xid after which a new changeset
2899 : * extraction replication slot can start decoding changes.
2900 : *
2901 : * Must be called with ProcArrayLock held either shared or exclusively,
2902 : * although most callers will want to use exclusive mode since it is expected
2903 : * that the caller will immediately use the xid to peg the xmin horizon.
2904 : */
2905 : TransactionId
2906 723 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2907 : {
2908 723 : ProcArrayStruct *arrayP = procArray;
2909 : TransactionId oldestSafeXid;
2910 : int index;
2911 723 : bool recovery_in_progress = RecoveryInProgress();
2912 :
2913 : Assert(LWLockHeldByMe(ProcArrayLock));
2914 :
2915 : /*
2916 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2917 : * running. If no transaction with xid were running concurrently a new xid
2918 : * could influence the RecentXmin et al.
2919 : *
2920 : * We initialize the computation to nextXid since that's guaranteed to be
2921 : * a safe, albeit pessimal, value.
2922 : */
2923 723 : LWLockAcquire(XidGenLock, LW_SHARED);
2924 723 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2925 :
2926 : /*
2927 : * If there's already a slot pegging the xmin horizon, we can start with
2928 : * that value, it's guaranteed to be safe since it's computed by this
2929 : * routine initially and has been enforced since. We can always use the
2930 : * slot's general xmin horizon, but the catalog horizon is only usable
2931 : * when only catalog data is going to be looked at.
2932 : */
2933 953 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2934 230 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2935 : oldestSafeXid))
2936 10 : oldestSafeXid = procArray->replication_slot_xmin;
2937 :
2938 723 : if (catalogOnly &&
2939 366 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2940 77 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2941 : oldestSafeXid))
2942 30 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2943 :
2944 : /*
2945 : * If we're not in recovery, we walk over the procarray and collect the
2946 : * lowest xid. Since we're called with ProcArrayLock held and have
2947 : * acquired XidGenLock, no entries can vanish concurrently, since
2948 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2949 : * with ProcArrayLock held.
2950 : *
2951 : * In recovery we can't lower the safe value besides what we've computed
2952 : * above, so we'll have to wait a bit longer there. We unfortunately can
2953 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2954 : * machinery can miss values and return an older value than is safe.
2955 : */
2956 723 : if (!recovery_in_progress)
2957 : {
2958 690 : TransactionId *other_xids = ProcGlobal->xids;
2959 :
2960 : /*
2961 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2962 : */
2963 3513 : for (index = 0; index < arrayP->numProcs; index++)
2964 : {
2965 : TransactionId xid;
2966 :
2967 : /* Fetch xid just once - see GetNewTransactionId */
2968 2823 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2969 :
2970 2823 : if (!TransactionIdIsNormal(xid))
2971 2816 : continue;
2972 :
2973 7 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2974 6 : oldestSafeXid = xid;
2975 : }
2976 : }
2977 :
2978 723 : LWLockRelease(XidGenLock);
2979 :
2980 723 : return oldestSafeXid;
2981 : }
2982 :
2983 : /*
2984 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2985 : * delaying checkpoint because they have critical actions in progress.
2986 : *
2987 : * Constructs an array of VXIDs of transactions that are currently in commit
2988 : * critical sections, as shown by having specified delayChkptFlags bits set
2989 : * in their PGPROC.
2990 : *
2991 : * Returns a palloc'd array that should be freed by the caller.
2992 : * *nvxids is the number of valid entries.
2993 : *
2994 : * Note that because backends set or clear delayChkptFlags without holding any
2995 : * lock, the result is somewhat indeterminate, but we don't really care. Even
2996 : * in a multiprocessor with delayed writes to shared memory, it should be
2997 : * certain that setting of delayChkptFlags will propagate to shared memory
2998 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
2999 : * delayChkptFlags if it's already inserted its commit record. Whether it
3000 : * takes a little while for clearing of delayChkptFlags to propagate is
3001 : * unimportant for correctness.
3002 : */
3003 : VirtualTransactionId *
3004 3260 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3005 : {
3006 : VirtualTransactionId *vxids;
3007 3260 : ProcArrayStruct *arrayP = procArray;
3008 3260 : int count = 0;
3009 : int index;
3010 :
3011 : Assert(type != 0);
3012 :
3013 : /* allocate what's certainly enough result space */
3014 3260 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3015 :
3016 3260 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3017 :
3018 10885 : for (index = 0; index < arrayP->numProcs; index++)
3019 : {
3020 7625 : int pgprocno = arrayP->pgprocnos[index];
3021 7625 : PGPROC *proc = &allProcs[pgprocno];
3022 :
3023 7625 : if ((proc->delayChkptFlags & type) != 0)
3024 : {
3025 : VirtualTransactionId vxid;
3026 :
3027 35 : GET_VXID_FROM_PGPROC(vxid, *proc);
3028 35 : if (VirtualTransactionIdIsValid(vxid))
3029 35 : vxids[count++] = vxid;
3030 : }
3031 : }
3032 :
3033 3260 : LWLockRelease(ProcArrayLock);
3034 :
3035 3260 : *nvxids = count;
3036 3260 : return vxids;
3037 : }
3038 :
3039 : /*
3040 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3041 : *
3042 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3043 : * of the specified VXIDs are still in critical sections of code.
3044 : *
3045 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3046 : * those numbers should be small enough for it not to be a problem.
3047 : */
3048 : bool
3049 31 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3050 : {
3051 31 : bool result = false;
3052 31 : ProcArrayStruct *arrayP = procArray;
3053 : int index;
3054 :
3055 : Assert(type != 0);
3056 :
3057 31 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3058 :
3059 349 : for (index = 0; index < arrayP->numProcs; index++)
3060 : {
3061 318 : int pgprocno = arrayP->pgprocnos[index];
3062 318 : PGPROC *proc = &allProcs[pgprocno];
3063 : VirtualTransactionId vxid;
3064 :
3065 318 : GET_VXID_FROM_PGPROC(vxid, *proc);
3066 :
3067 318 : if ((proc->delayChkptFlags & type) != 0 &&
3068 7 : VirtualTransactionIdIsValid(vxid))
3069 : {
3070 : int i;
3071 :
3072 17 : for (i = 0; i < nvxids; i++)
3073 : {
3074 10 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3075 : {
3076 0 : result = true;
3077 0 : break;
3078 : }
3079 : }
3080 7 : if (result)
3081 0 : break;
3082 : }
3083 : }
3084 :
3085 31 : LWLockRelease(ProcArrayLock);
3086 :
3087 31 : return result;
3088 : }
3089 :
3090 : /*
3091 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3092 : *
3093 : * The result may be out of date arbitrarily quickly, so the caller
3094 : * must be careful about how this information is used. NULL is
3095 : * returned if the backend is not active.
3096 : */
3097 : PGPROC *
3098 649 : ProcNumberGetProc(ProcNumber procNumber)
3099 : {
3100 : PGPROC *result;
3101 :
3102 649 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3103 1 : return NULL;
3104 648 : result = GetPGProcByNumber(procNumber);
3105 :
3106 648 : if (result->pid == 0)
3107 3 : return NULL;
3108 :
3109 645 : return result;
3110 : }
3111 :
3112 : /*
3113 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3114 : *
3115 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3116 : * result may be out of date arbitrarily quickly, so the caller must be
3117 : * careful about how this information is used.
3118 : */
3119 : void
3120 10342 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3121 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3122 : {
3123 : PGPROC *proc;
3124 :
3125 10342 : *xid = InvalidTransactionId;
3126 10342 : *xmin = InvalidTransactionId;
3127 10342 : *nsubxid = 0;
3128 10342 : *overflowed = false;
3129 :
3130 10342 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3131 0 : return;
3132 10342 : proc = GetPGProcByNumber(procNumber);
3133 :
3134 : /* Need to lock out additions/removals of backends */
3135 10342 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3136 :
3137 10342 : if (proc->pid != 0)
3138 : {
3139 10342 : *xid = proc->xid;
3140 10342 : *xmin = proc->xmin;
3141 10342 : *nsubxid = proc->subxidStatus.count;
3142 10342 : *overflowed = proc->subxidStatus.overflowed;
3143 : }
3144 :
3145 10342 : LWLockRelease(ProcArrayLock);
3146 : }
3147 :
3148 : /*
3149 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3150 : *
3151 : * Returns NULL if not found. Note that it is up to the caller to be
3152 : * sure that the question remains meaningful for long enough for the
3153 : * answer to be used ...
3154 : */
3155 : PGPROC *
3156 10973 : BackendPidGetProc(int pid)
3157 : {
3158 : PGPROC *result;
3159 :
3160 10973 : if (pid == 0) /* never match dummy PGPROCs */
3161 4 : return NULL;
3162 :
3163 10969 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3164 :
3165 10969 : result = BackendPidGetProcWithLock(pid);
3166 :
3167 10969 : LWLockRelease(ProcArrayLock);
3168 :
3169 10969 : return result;
3170 : }
3171 :
3172 : /*
3173 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3174 : *
3175 : * Same as above, except caller must be holding ProcArrayLock. The found
3176 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3177 : */
3178 : PGPROC *
3179 12541 : BackendPidGetProcWithLock(int pid)
3180 : {
3181 12541 : PGPROC *result = NULL;
3182 12541 : ProcArrayStruct *arrayP = procArray;
3183 : int index;
3184 :
3185 12541 : if (pid == 0) /* never match dummy PGPROCs */
3186 0 : return NULL;
3187 :
3188 50233 : for (index = 0; index < arrayP->numProcs; index++)
3189 : {
3190 44776 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3191 :
3192 44776 : if (proc->pid == pid)
3193 : {
3194 7084 : result = proc;
3195 7084 : break;
3196 : }
3197 : }
3198 :
3199 12541 : return result;
3200 : }
3201 :
3202 : /*
3203 : * BackendXidGetPid -- get a backend's pid given its XID
3204 : *
3205 : * Returns 0 if not found or it's a prepared transaction. Note that
3206 : * it is up to the caller to be sure that the question remains
3207 : * meaningful for long enough for the answer to be used ...
3208 : *
3209 : * Only main transaction Ids are considered. This function is mainly
3210 : * useful for determining what backend owns a lock.
3211 : *
3212 : * Beware that not every xact has an XID assigned. However, as long as you
3213 : * only call this using an XID found on disk, you're safe.
3214 : */
3215 : int
3216 30 : BackendXidGetPid(TransactionId xid)
3217 : {
3218 30 : int result = 0;
3219 30 : ProcArrayStruct *arrayP = procArray;
3220 30 : TransactionId *other_xids = ProcGlobal->xids;
3221 : int index;
3222 :
3223 30 : if (xid == InvalidTransactionId) /* never match invalid xid */
3224 0 : return 0;
3225 :
3226 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3227 :
3228 92 : for (index = 0; index < arrayP->numProcs; index++)
3229 : {
3230 84 : if (other_xids[index] == xid)
3231 : {
3232 22 : int pgprocno = arrayP->pgprocnos[index];
3233 22 : PGPROC *proc = &allProcs[pgprocno];
3234 :
3235 22 : result = proc->pid;
3236 22 : break;
3237 : }
3238 : }
3239 :
3240 30 : LWLockRelease(ProcArrayLock);
3241 :
3242 30 : return result;
3243 : }
3244 :
3245 : /*
3246 : * IsBackendPid -- is a given pid a running backend
3247 : *
3248 : * This is not called by the backend, but is called by external modules.
3249 : */
3250 : bool
3251 2 : IsBackendPid(int pid)
3252 : {
3253 2 : return (BackendPidGetProc(pid) != NULL);
3254 : }
3255 :
3256 :
3257 : /*
3258 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3259 : *
3260 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3261 : *
3262 : * The arguments allow filtering the set of VXIDs returned. Our own process
3263 : * is always skipped. In addition:
3264 : * If limitXmin is not InvalidTransactionId, skip processes with
3265 : * xmin > limitXmin.
3266 : * If excludeXmin0 is true, skip processes with xmin = 0.
3267 : * If allDbs is false, skip processes attached to other databases.
3268 : * If excludeVacuum isn't zero, skip processes for which
3269 : * (statusFlags & excludeVacuum) is not zero.
3270 : *
3271 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3272 : * allow skipping backends whose oldest live snapshot is no older than
3273 : * some snapshot we have. Since we examine the procarray with only shared
3274 : * lock, there are race conditions: a backend could set its xmin just after
3275 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3276 : * other backend could have set its xmin *before* we look. We know however
3277 : * that such a backend must have held shared ProcArrayLock overlapping our
3278 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3279 : * any snapshot the other backend is taking concurrently with our scan cannot
3280 : * consider any transactions as still running that we think are committed
3281 : * (since backends must hold ProcArrayLock exclusive to commit).
3282 : */
3283 : VirtualTransactionId *
3284 476 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3285 : bool allDbs, int excludeVacuum,
3286 : int *nvxids)
3287 : {
3288 : VirtualTransactionId *vxids;
3289 476 : ProcArrayStruct *arrayP = procArray;
3290 476 : int count = 0;
3291 : int index;
3292 :
3293 : /* allocate what's certainly enough result space */
3294 476 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3295 :
3296 476 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3297 :
3298 2856 : for (index = 0; index < arrayP->numProcs; index++)
3299 : {
3300 2380 : int pgprocno = arrayP->pgprocnos[index];
3301 2380 : PGPROC *proc = &allProcs[pgprocno];
3302 2380 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3303 :
3304 2380 : if (proc == MyProc)
3305 476 : continue;
3306 :
3307 1904 : if (excludeVacuum & statusFlags)
3308 13 : continue;
3309 :
3310 1891 : if (allDbs || proc->databaseId == MyDatabaseId)
3311 : {
3312 : /* Fetch xmin just once - might change on us */
3313 867 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3314 :
3315 867 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3316 567 : continue;
3317 :
3318 : /*
3319 : * InvalidTransactionId precedes all other XIDs, so a proc that
3320 : * hasn't set xmin yet will not be rejected by this test.
3321 : */
3322 600 : if (!TransactionIdIsValid(limitXmin) ||
3323 300 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3324 : {
3325 : VirtualTransactionId vxid;
3326 :
3327 287 : GET_VXID_FROM_PGPROC(vxid, *proc);
3328 287 : if (VirtualTransactionIdIsValid(vxid))
3329 287 : vxids[count++] = vxid;
3330 : }
3331 : }
3332 : }
3333 :
3334 476 : LWLockRelease(ProcArrayLock);
3335 :
3336 476 : *nvxids = count;
3337 476 : return vxids;
3338 : }
3339 :
3340 : /*
3341 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3342 : *
3343 : * Usage is limited to conflict resolution during recovery on standby servers.
3344 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3345 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3346 : * determine a safe snapshotConflictHorizon value.
3347 : *
3348 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3349 : * so we're not worried if they have a snapshot or not, nor does it really
3350 : * matter what type of lock we hold. Caller must avoid calling here with
3351 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3352 : * during original execution, since that actually indicates that there is
3353 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3354 : * convention for InvalidTransactionId values is the opposite of our own!).
3355 : *
3356 : * All callers that are checking xmins always now supply a valid and useful
3357 : * value for limitXmin. The limitXmin is always lower than the lowest
3358 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3359 : * because we only care about cleanup records that are cleaning up tuple
3360 : * versions from committed transactions. In that case they will only occur
3361 : * at the point where the record is less than the lowest running xid. That
3362 : * allows us to say that if any backend takes a snapshot concurrently with
3363 : * us then the conflict assessment made here would never include the snapshot
3364 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3365 : * concurrent snapshots when limitXmin is valid. We might think about adding
3366 : * Assert(limitXmin < lowest(KnownAssignedXids))
3367 : * but that would not be true in the case of FATAL errors lagging in array,
3368 : * but we already know those are bogus anyway, so we skip that test.
3369 : *
3370 : * If dbOid is valid we skip backends attached to other databases.
3371 : *
3372 : * Be careful to *not* pfree the result from this function. We reuse
3373 : * this array sufficiently often that we use malloc for the result.
3374 : */
3375 : VirtualTransactionId *
3376 15350 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3377 : {
3378 : static VirtualTransactionId *vxids;
3379 15350 : ProcArrayStruct *arrayP = procArray;
3380 15350 : int count = 0;
3381 : int index;
3382 :
3383 : /*
3384 : * If first time through, get workspace to remember main XIDs in. We
3385 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3386 : * result space, remembering room for a terminator.
3387 : */
3388 15350 : if (vxids == NULL)
3389 : {
3390 19 : vxids = (VirtualTransactionId *)
3391 19 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3392 19 : if (vxids == NULL)
3393 0 : ereport(ERROR,
3394 : (errcode(ERRCODE_OUT_OF_MEMORY),
3395 : errmsg("out of memory")));
3396 : }
3397 :
3398 15350 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3399 :
3400 15499 : for (index = 0; index < arrayP->numProcs; index++)
3401 : {
3402 149 : int pgprocno = arrayP->pgprocnos[index];
3403 149 : PGPROC *proc = &allProcs[pgprocno];
3404 :
3405 : /* Exclude prepared transactions */
3406 149 : if (proc->pid == 0)
3407 0 : continue;
3408 :
3409 149 : if (!OidIsValid(dbOid) ||
3410 143 : proc->databaseId == dbOid)
3411 : {
3412 : /* Fetch xmin just once - can't change on us, but good coding */
3413 18 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3414 :
3415 : /*
3416 : * We ignore an invalid pxmin because this means that backend has
3417 : * no snapshot currently. We hold a Share lock to avoid contention
3418 : * with users taking snapshots. That is not a problem because the
3419 : * current xmin is always at least one higher than the latest
3420 : * removed xid, so any new snapshot would never conflict with the
3421 : * test here.
3422 : */
3423 18 : if (!TransactionIdIsValid(limitXmin) ||
3424 3 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3425 : {
3426 : VirtualTransactionId vxid;
3427 :
3428 2 : GET_VXID_FROM_PGPROC(vxid, *proc);
3429 2 : if (VirtualTransactionIdIsValid(vxid))
3430 2 : vxids[count++] = vxid;
3431 : }
3432 : }
3433 : }
3434 :
3435 15350 : LWLockRelease(ProcArrayLock);
3436 :
3437 : /* add the terminator */
3438 15350 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3439 15350 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3440 :
3441 15350 : return vxids;
3442 : }
3443 :
3444 : /*
3445 : * SignalRecoveryConflict -- signal that a process is blocking recovery
3446 : *
3447 : * The 'pid' is redundant with 'proc', but it acts as a cross-check to
3448 : * detect process had exited and the PGPROC entry was reused for a different
3449 : * process.
3450 : *
3451 : * Returns true if the process was signaled, or false if not found.
3452 : */
3453 : bool
3454 5 : SignalRecoveryConflict(PGPROC *proc, pid_t pid, RecoveryConflictReason reason)
3455 : {
3456 5 : bool found = false;
3457 :
3458 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3459 :
3460 : /*
3461 : * Kill the pid if it's still here. If not, that's what we wanted so
3462 : * ignore any errors.
3463 : */
3464 5 : if (proc->pid == pid)
3465 : {
3466 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3467 :
3468 : /* wake up the process */
3469 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, GetNumberFromPGProc(proc));
3470 5 : found = true;
3471 : }
3472 :
3473 5 : LWLockRelease(ProcArrayLock);
3474 :
3475 5 : return found;
3476 : }
3477 :
3478 : /*
3479 : * SignalRecoveryConflictWithVirtualXID -- signal that a VXID is blocking recovery
3480 : *
3481 : * Like SignalRecoveryConflict, but the target is identified by VXID
3482 : */
3483 : bool
3484 5 : SignalRecoveryConflictWithVirtualXID(VirtualTransactionId vxid, RecoveryConflictReason reason)
3485 : {
3486 5 : ProcArrayStruct *arrayP = procArray;
3487 : int index;
3488 5 : pid_t pid = 0;
3489 :
3490 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3491 :
3492 5 : for (index = 0; index < arrayP->numProcs; index++)
3493 : {
3494 5 : int pgprocno = arrayP->pgprocnos[index];
3495 5 : PGPROC *proc = &allProcs[pgprocno];
3496 : VirtualTransactionId procvxid;
3497 :
3498 5 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3499 :
3500 5 : if (procvxid.procNumber == vxid.procNumber &&
3501 5 : procvxid.localTransactionId == vxid.localTransactionId)
3502 : {
3503 5 : pid = proc->pid;
3504 5 : if (pid != 0)
3505 : {
3506 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3507 :
3508 : /*
3509 : * Kill the pid if it's still here. If not, that's what we
3510 : * wanted so ignore any errors.
3511 : */
3512 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, vxid.procNumber);
3513 : }
3514 5 : break;
3515 : }
3516 : }
3517 :
3518 5 : LWLockRelease(ProcArrayLock);
3519 :
3520 5 : return pid != 0;
3521 : }
3522 :
3523 : /*
3524 : * SignalRecoveryConflictWithDatabase -- signal backends using specified database
3525 : *
3526 : * Like SignalRecoveryConflict, but signals all backends using the database.
3527 : */
3528 : void
3529 9 : SignalRecoveryConflictWithDatabase(Oid databaseid, RecoveryConflictReason reason)
3530 : {
3531 9 : ProcArrayStruct *arrayP = procArray;
3532 : int index;
3533 :
3534 : /* tell all backends to die */
3535 9 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3536 :
3537 18 : for (index = 0; index < arrayP->numProcs; index++)
3538 : {
3539 9 : int pgprocno = arrayP->pgprocnos[index];
3540 9 : PGPROC *proc = &allProcs[pgprocno];
3541 :
3542 9 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3543 : {
3544 : VirtualTransactionId procvxid;
3545 : pid_t pid;
3546 :
3547 9 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3548 :
3549 9 : pid = proc->pid;
3550 9 : if (pid != 0)
3551 : {
3552 9 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3553 :
3554 : /*
3555 : * Kill the pid if it's still here. If not, that's what we
3556 : * wanted so ignore any errors.
3557 : */
3558 9 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, procvxid.procNumber);
3559 : }
3560 : }
3561 : }
3562 :
3563 9 : LWLockRelease(ProcArrayLock);
3564 9 : }
3565 :
3566 : /*
3567 : * MinimumActiveBackends --- count backends (other than myself) that are
3568 : * in active transactions. Return true if the count exceeds the
3569 : * minimum threshold passed. This is used as a heuristic to decide if
3570 : * a pre-XLOG-flush delay is worthwhile during commit.
3571 : *
3572 : * Do not count backends that are blocked waiting for locks, since they are
3573 : * not going to get to run until someone else commits.
3574 : */
3575 : bool
3576 0 : MinimumActiveBackends(int min)
3577 : {
3578 0 : ProcArrayStruct *arrayP = procArray;
3579 0 : int count = 0;
3580 : int index;
3581 :
3582 : /* Quick short-circuit if no minimum is specified */
3583 0 : if (min == 0)
3584 0 : return true;
3585 :
3586 : /*
3587 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3588 : * bogus, but since we are only testing fields for zero or nonzero, it
3589 : * should be OK. The result is only used for heuristic purposes anyway...
3590 : */
3591 0 : for (index = 0; index < arrayP->numProcs; index++)
3592 : {
3593 0 : int pgprocno = arrayP->pgprocnos[index];
3594 0 : PGPROC *proc = &allProcs[pgprocno];
3595 :
3596 : /*
3597 : * Since we're not holding a lock, need to be prepared to deal with
3598 : * garbage, as someone could have incremented numProcs but not yet
3599 : * filled the structure.
3600 : *
3601 : * If someone just decremented numProcs, 'proc' could also point to a
3602 : * PGPROC entry that's no longer in the array. It still points to a
3603 : * PGPROC struct, though, because freed PGPROC entries just go to the
3604 : * free list and are recycled. Its contents are nonsense in that case,
3605 : * but that's acceptable for this function.
3606 : */
3607 0 : if (pgprocno == -1)
3608 0 : continue; /* do not count deleted entries */
3609 0 : if (proc == MyProc)
3610 0 : continue; /* do not count myself */
3611 0 : if (proc->xid == InvalidTransactionId)
3612 0 : continue; /* do not count if no XID assigned */
3613 0 : if (proc->pid == 0)
3614 0 : continue; /* do not count prepared xacts */
3615 0 : if (proc->waitLock != NULL)
3616 0 : continue; /* do not count if blocked on a lock */
3617 0 : count++;
3618 0 : if (count >= min)
3619 0 : break;
3620 : }
3621 :
3622 0 : return count >= min;
3623 : }
3624 :
3625 : /*
3626 : * CountDBBackends --- count backends that are using specified database
3627 : */
3628 : int
3629 16 : CountDBBackends(Oid databaseid)
3630 : {
3631 16 : ProcArrayStruct *arrayP = procArray;
3632 16 : int count = 0;
3633 : int index;
3634 :
3635 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3636 :
3637 24 : for (index = 0; index < arrayP->numProcs; index++)
3638 : {
3639 8 : int pgprocno = arrayP->pgprocnos[index];
3640 8 : PGPROC *proc = &allProcs[pgprocno];
3641 :
3642 8 : if (proc->pid == 0)
3643 0 : continue; /* do not count prepared xacts */
3644 8 : if (!OidIsValid(databaseid) ||
3645 8 : proc->databaseId == databaseid)
3646 2 : count++;
3647 : }
3648 :
3649 16 : LWLockRelease(ProcArrayLock);
3650 :
3651 16 : return count;
3652 : }
3653 :
3654 : /*
3655 : * CountDBConnections --- counts database backends (only regular backends)
3656 : */
3657 : int
3658 0 : CountDBConnections(Oid databaseid)
3659 : {
3660 0 : ProcArrayStruct *arrayP = procArray;
3661 0 : int count = 0;
3662 : int index;
3663 :
3664 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3665 :
3666 0 : for (index = 0; index < arrayP->numProcs; index++)
3667 : {
3668 0 : int pgprocno = arrayP->pgprocnos[index];
3669 0 : PGPROC *proc = &allProcs[pgprocno];
3670 :
3671 0 : if (proc->pid == 0)
3672 0 : continue; /* do not count prepared xacts */
3673 0 : if (proc->backendType != B_BACKEND)
3674 0 : continue; /* count only regular backend processes */
3675 0 : if (!OidIsValid(databaseid) ||
3676 0 : proc->databaseId == databaseid)
3677 0 : count++;
3678 : }
3679 :
3680 0 : LWLockRelease(ProcArrayLock);
3681 :
3682 0 : return count;
3683 : }
3684 :
3685 : /*
3686 : * CountUserBackends --- count backends that are used by specified user
3687 : * (only regular backends, not any type of background worker)
3688 : */
3689 : int
3690 0 : CountUserBackends(Oid roleid)
3691 : {
3692 0 : ProcArrayStruct *arrayP = procArray;
3693 0 : int count = 0;
3694 : int index;
3695 :
3696 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3697 :
3698 0 : for (index = 0; index < arrayP->numProcs; index++)
3699 : {
3700 0 : int pgprocno = arrayP->pgprocnos[index];
3701 0 : PGPROC *proc = &allProcs[pgprocno];
3702 :
3703 0 : if (proc->pid == 0)
3704 0 : continue; /* do not count prepared xacts */
3705 0 : if (proc->backendType != B_BACKEND)
3706 0 : continue; /* count only regular backend processes */
3707 0 : if (proc->roleId == roleid)
3708 0 : count++;
3709 : }
3710 :
3711 0 : LWLockRelease(ProcArrayLock);
3712 :
3713 0 : return count;
3714 : }
3715 :
3716 : /*
3717 : * CountOtherDBBackends -- check for other backends running in the given DB
3718 : *
3719 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3720 : * for them to exit (or 0.3s for testing purposes). Autovacuum backends are
3721 : * encouraged to exit early by sending them SIGTERM, but normal user backends
3722 : * are just waited for. If background workers connected to this database are
3723 : * marked as interruptible, they are terminated.
3724 : *
3725 : * The current backend is always ignored; it is caller's responsibility to
3726 : * check whether the current backend uses the given DB, if it's important.
3727 : *
3728 : * Returns true if there are (still) other backends in the DB, false if not.
3729 : * Also, *nbackends and *nprepared are set to the number of other backends
3730 : * and prepared transactions in the DB, respectively.
3731 : *
3732 : * This function is used to interlock DROP DATABASE and related commands
3733 : * against there being any active backends in the target DB --- dropping the
3734 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3735 : * detect here the possibility of a newly-started backend that is trying to
3736 : * connect to the doomed database, so additional interlocking is needed during
3737 : * backend startup. The caller should normally hold an exclusive lock on the
3738 : * target DB before calling this, which is one reason we mustn't wait
3739 : * indefinitely.
3740 : */
3741 : bool
3742 487 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3743 : {
3744 487 : ProcArrayStruct *arrayP = procArray;
3745 :
3746 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3747 : int autovac_pids[MAXAUTOVACPIDS];
3748 :
3749 : /*
3750 : * Retry up to 50 times with 100ms between attempts (max 5s total). Can be
3751 : * reduced to 3 attempts (max 0.3s total) to speed up tests.
3752 : */
3753 487 : int ntries = 50;
3754 :
3755 : #ifdef USE_INJECTION_POINTS
3756 487 : if (IS_INJECTION_POINT_ATTACHED("procarray-reduce-count"))
3757 1 : ntries = 3;
3758 : #endif
3759 :
3760 494 : for (int tries = 0; tries < ntries; tries++)
3761 : {
3762 493 : int nautovacs = 0;
3763 493 : bool found = false;
3764 : int index;
3765 :
3766 493 : CHECK_FOR_INTERRUPTS();
3767 :
3768 493 : *nbackends = *nprepared = 0;
3769 :
3770 493 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3771 :
3772 1894 : for (index = 0; index < arrayP->numProcs; index++)
3773 : {
3774 1401 : int pgprocno = arrayP->pgprocnos[index];
3775 1401 : PGPROC *proc = &allProcs[pgprocno];
3776 1401 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3777 :
3778 1401 : if (proc->databaseId != databaseId)
3779 1291 : continue;
3780 110 : if (proc == MyProc)
3781 103 : continue;
3782 :
3783 7 : found = true;
3784 :
3785 7 : if (proc->pid == 0)
3786 0 : (*nprepared)++;
3787 : else
3788 : {
3789 7 : (*nbackends)++;
3790 7 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3791 : nautovacs < MAXAUTOVACPIDS)
3792 0 : autovac_pids[nautovacs++] = proc->pid;
3793 : }
3794 : }
3795 :
3796 493 : LWLockRelease(ProcArrayLock);
3797 :
3798 493 : if (!found)
3799 486 : return false; /* no conflicting backends, so done */
3800 :
3801 : /*
3802 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3803 : * postpone this step until after the loop because we don't want to
3804 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3805 : * block kill() inside the kernel...
3806 : */
3807 7 : for (index = 0; index < nautovacs; index++)
3808 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3809 :
3810 : /*
3811 : * Terminate all background workers for this database, if they have
3812 : * requested it (BGWORKER_INTERRUPTIBLE).
3813 : */
3814 7 : TerminateBackgroundWorkersForDatabase(databaseId);
3815 :
3816 : /* sleep, then try again */
3817 7 : pg_usleep(100 * 1000L); /* 100ms */
3818 : }
3819 :
3820 1 : return true; /* timed out, still conflicts */
3821 : }
3822 :
3823 : /*
3824 : * Terminate existing connections to the specified database. This routine
3825 : * is used by the DROP DATABASE command when user has asked to forcefully
3826 : * drop the database.
3827 : *
3828 : * The current backend is always ignored; it is caller's responsibility to
3829 : * check whether the current backend uses the given DB, if it's important.
3830 : *
3831 : * If the target database has a prepared transaction or permissions checks
3832 : * fail for a connection, this fails without terminating anything.
3833 : */
3834 : void
3835 1 : TerminateOtherDBBackends(Oid databaseId)
3836 : {
3837 1 : ProcArrayStruct *arrayP = procArray;
3838 1 : List *pids = NIL;
3839 1 : int nprepared = 0;
3840 : int i;
3841 :
3842 1 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3843 :
3844 4 : for (i = 0; i < procArray->numProcs; i++)
3845 : {
3846 3 : int pgprocno = arrayP->pgprocnos[i];
3847 3 : PGPROC *proc = &allProcs[pgprocno];
3848 :
3849 3 : if (proc->databaseId != databaseId)
3850 3 : continue;
3851 0 : if (proc == MyProc)
3852 0 : continue;
3853 :
3854 0 : if (proc->pid != 0)
3855 0 : pids = lappend_int(pids, proc->pid);
3856 : else
3857 0 : nprepared++;
3858 : }
3859 :
3860 1 : LWLockRelease(ProcArrayLock);
3861 :
3862 1 : if (nprepared > 0)
3863 0 : ereport(ERROR,
3864 : (errcode(ERRCODE_OBJECT_IN_USE),
3865 : errmsg("database \"%s\" is being used by prepared transactions",
3866 : get_database_name(databaseId)),
3867 : errdetail_plural("There is %d prepared transaction using the database.",
3868 : "There are %d prepared transactions using the database.",
3869 : nprepared,
3870 : nprepared)));
3871 :
3872 1 : if (pids)
3873 : {
3874 : ListCell *lc;
3875 :
3876 : /*
3877 : * Permissions checks relax the pg_terminate_backend checks in two
3878 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3879 : *
3880 : * - Accept terminating autovacuum workers, since DROP DATABASE
3881 : * without FORCE terminates them.
3882 : *
3883 : * - Accept terminating bgworkers. For bgworker authors, it's
3884 : * convenient to be able to recommend FORCE if a worker is blocking
3885 : * DROP DATABASE unexpectedly.
3886 : *
3887 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3888 : * "PID %d is not a PostgreSQL server process", because for us already
3889 : * finished session is not a problem.
3890 : */
3891 0 : foreach(lc, pids)
3892 : {
3893 0 : int pid = lfirst_int(lc);
3894 0 : PGPROC *proc = BackendPidGetProc(pid);
3895 :
3896 0 : if (proc != NULL)
3897 : {
3898 0 : if (superuser_arg(proc->roleId) && !superuser())
3899 0 : ereport(ERROR,
3900 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3901 : errmsg("permission denied to terminate process"),
3902 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3903 : "SUPERUSER", "SUPERUSER")));
3904 :
3905 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3906 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3907 0 : ereport(ERROR,
3908 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3909 : errmsg("permission denied to terminate process"),
3910 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3911 : "pg_signal_backend")));
3912 : }
3913 : }
3914 :
3915 : /*
3916 : * There's a race condition here: once we release the ProcArrayLock,
3917 : * it's possible for the session to exit before we issue kill. That
3918 : * race condition possibility seems too unlikely to worry about. See
3919 : * pg_signal_backend.
3920 : */
3921 0 : foreach(lc, pids)
3922 : {
3923 0 : int pid = lfirst_int(lc);
3924 0 : PGPROC *proc = BackendPidGetProc(pid);
3925 :
3926 0 : if (proc != NULL)
3927 : {
3928 : /*
3929 : * If we have setsid(), signal the backend's whole process
3930 : * group
3931 : */
3932 : #ifdef HAVE_SETSID
3933 0 : (void) kill(-pid, SIGTERM);
3934 : #else
3935 : (void) kill(pid, SIGTERM);
3936 : #endif
3937 : }
3938 : }
3939 : }
3940 1 : }
3941 :
3942 : /*
3943 : * ProcArraySetReplicationSlotXmin
3944 : *
3945 : * Install limits to future computations of the xmin horizon to prevent vacuum
3946 : * and HOT pruning from removing affected rows still needed by clients with
3947 : * replication slots.
3948 : */
3949 : void
3950 2555 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3951 : bool already_locked)
3952 : {
3953 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3954 :
3955 2555 : if (!already_locked)
3956 2047 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3957 :
3958 2555 : procArray->replication_slot_xmin = xmin;
3959 2555 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3960 :
3961 2555 : if (!already_locked)
3962 2047 : LWLockRelease(ProcArrayLock);
3963 :
3964 2555 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3965 : xmin, catalog_xmin);
3966 2555 : }
3967 :
3968 : /*
3969 : * ProcArrayGetReplicationSlotXmin
3970 : *
3971 : * Return the current slot xmin limits. That's useful to be able to remove
3972 : * data that's older than those limits.
3973 : */
3974 : void
3975 22 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3976 : TransactionId *catalog_xmin)
3977 : {
3978 22 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3979 :
3980 22 : if (xmin != NULL)
3981 0 : *xmin = procArray->replication_slot_xmin;
3982 :
3983 22 : if (catalog_xmin != NULL)
3984 22 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3985 :
3986 22 : LWLockRelease(ProcArrayLock);
3987 22 : }
3988 :
3989 : /*
3990 : * XidCacheRemoveRunningXids
3991 : *
3992 : * Remove a bunch of TransactionIds from the list of known-running
3993 : * subtransactions for my backend. Both the specified xid and those in
3994 : * the xids[] array (of length nxids) are removed from the subxids cache.
3995 : * latestXid must be the latest XID among the group.
3996 : */
3997 : void
3998 851 : XidCacheRemoveRunningXids(TransactionId xid,
3999 : int nxids, const TransactionId *xids,
4000 : TransactionId latestXid)
4001 : {
4002 : int i,
4003 : j;
4004 : XidCacheStatus *mysubxidstat;
4005 :
4006 : Assert(TransactionIdIsValid(xid));
4007 :
4008 : /*
4009 : * We must hold ProcArrayLock exclusively in order to remove transactions
4010 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
4011 : * possible this could be relaxed since we know this routine is only used
4012 : * to abort subtransactions, but pending closer analysis we'd best be
4013 : * conservative.
4014 : *
4015 : * Note that we do not have to be careful about memory ordering of our own
4016 : * reads wrt. GetNewTransactionId() here - only this process can modify
4017 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
4018 : * careful about our own writes being well ordered.
4019 : */
4020 851 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4021 :
4022 851 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4023 :
4024 : /*
4025 : * Under normal circumstances xid and xids[] will be in increasing order,
4026 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4027 : * behavior when removing a lot of xids.
4028 : */
4029 883 : for (i = nxids - 1; i >= 0; i--)
4030 : {
4031 32 : TransactionId anxid = xids[i];
4032 :
4033 32 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4034 : {
4035 32 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4036 : {
4037 32 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4038 32 : pg_write_barrier();
4039 32 : mysubxidstat->count--;
4040 32 : MyProc->subxidStatus.count--;
4041 32 : break;
4042 : }
4043 : }
4044 :
4045 : /*
4046 : * Ordinarily we should have found it, unless the cache has
4047 : * overflowed. However it's also possible for this routine to be
4048 : * invoked multiple times for the same subtransaction, in case of an
4049 : * error during AbortSubTransaction. So instead of Assert, emit a
4050 : * debug warning.
4051 : */
4052 32 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4053 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4054 : }
4055 :
4056 915 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4057 : {
4058 914 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4059 : {
4060 850 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4061 850 : pg_write_barrier();
4062 850 : mysubxidstat->count--;
4063 850 : MyProc->subxidStatus.count--;
4064 850 : break;
4065 : }
4066 : }
4067 : /* Ordinarily we should have found it, unless the cache has overflowed */
4068 851 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4069 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4070 :
4071 : /* Also advance global latestCompletedXid while holding the lock */
4072 851 : MaintainLatestCompletedXid(latestXid);
4073 :
4074 : /* ... and xactCompletionCount */
4075 851 : TransamVariables->xactCompletionCount++;
4076 :
4077 851 : LWLockRelease(ProcArrayLock);
4078 851 : }
4079 :
4080 : #ifdef XIDCACHE_DEBUG
4081 :
4082 : /*
4083 : * Print stats about effectiveness of XID cache
4084 : */
4085 : static void
4086 : DisplayXidCache(void)
4087 : {
4088 : fprintf(stderr,
4089 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4090 : xc_by_recent_xmin,
4091 : xc_by_known_xact,
4092 : xc_by_my_xact,
4093 : xc_by_latest_xid,
4094 : xc_by_main_xid,
4095 : xc_by_child_xid,
4096 : xc_by_known_assigned,
4097 : xc_no_overflow,
4098 : xc_slow_answer);
4099 : }
4100 : #endif /* XIDCACHE_DEBUG */
4101 :
4102 : /*
4103 : * If rel != NULL, return test state appropriate for relation, otherwise
4104 : * return state usable for all relations. The latter may consider XIDs as
4105 : * not-yet-visible-to-everyone that a state for a specific relation would
4106 : * already consider visible-to-everyone.
4107 : *
4108 : * This needs to be called while a snapshot is active or registered, otherwise
4109 : * there are wraparound and other dangers.
4110 : *
4111 : * See comment for GlobalVisState for details.
4112 : */
4113 : GlobalVisState *
4114 18848401 : GlobalVisTestFor(Relation rel)
4115 : {
4116 18848401 : GlobalVisState *state = NULL;
4117 :
4118 : /* XXX: we should assert that a snapshot is pushed or registered */
4119 : Assert(RecentXmin);
4120 :
4121 18848401 : switch (GlobalVisHorizonKindForRel(rel))
4122 : {
4123 100796 : case VISHORIZON_SHARED:
4124 100796 : state = &GlobalVisSharedRels;
4125 100796 : break;
4126 3717321 : case VISHORIZON_CATALOG:
4127 3717321 : state = &GlobalVisCatalogRels;
4128 3717321 : break;
4129 14957532 : case VISHORIZON_DATA:
4130 14957532 : state = &GlobalVisDataRels;
4131 14957532 : break;
4132 72752 : case VISHORIZON_TEMP:
4133 72752 : state = &GlobalVisTempRels;
4134 72752 : break;
4135 : }
4136 :
4137 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4138 : FullTransactionIdIsValid(state->maybe_needed));
4139 :
4140 18848401 : return state;
4141 : }
4142 :
4143 : /*
4144 : * Return true if it's worth updating the accurate maybe_needed boundary.
4145 : *
4146 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4147 : * repeatedly do so when there is a low likelihood of it being beneficial.
4148 : *
4149 : * The current heuristic is that we update only if RecentXmin has changed
4150 : * since the last update. If the oldest currently running transaction has not
4151 : * finished, it is unlikely that recomputing the horizon would be useful.
4152 : */
4153 : static bool
4154 510796 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4155 : {
4156 : /* hasn't been updated yet */
4157 510796 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4158 9920 : return true;
4159 :
4160 : /*
4161 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4162 : * unlikely to be beneficial to refresh boundaries.
4163 : */
4164 500876 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4165 : state->definitely_needed))
4166 0 : return false;
4167 :
4168 : /* does the last snapshot built have a different xmin? */
4169 500876 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4170 : }
4171 :
4172 : static void
4173 214267 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4174 : {
4175 : GlobalVisSharedRels.maybe_needed =
4176 214267 : FullXidRelativeTo(horizons->latest_completed,
4177 : horizons->shared_oldest_nonremovable);
4178 : GlobalVisCatalogRels.maybe_needed =
4179 214267 : FullXidRelativeTo(horizons->latest_completed,
4180 : horizons->catalog_oldest_nonremovable);
4181 : GlobalVisDataRels.maybe_needed =
4182 214267 : FullXidRelativeTo(horizons->latest_completed,
4183 : horizons->data_oldest_nonremovable);
4184 : GlobalVisTempRels.maybe_needed =
4185 214267 : FullXidRelativeTo(horizons->latest_completed,
4186 : horizons->temp_oldest_nonremovable);
4187 :
4188 : /*
4189 : * In longer running transactions it's possible that transactions we
4190 : * previously needed to treat as running aren't around anymore. So update
4191 : * definitely_needed to not be earlier than maybe_needed.
4192 : */
4193 : GlobalVisSharedRels.definitely_needed =
4194 214267 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4195 : GlobalVisSharedRels.definitely_needed);
4196 : GlobalVisCatalogRels.definitely_needed =
4197 214267 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4198 : GlobalVisCatalogRels.definitely_needed);
4199 : GlobalVisDataRels.definitely_needed =
4200 214267 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4201 : GlobalVisDataRels.definitely_needed);
4202 214267 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4203 :
4204 214267 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4205 214267 : }
4206 :
4207 : /*
4208 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4209 : * using ComputeXidHorizons().
4210 : */
4211 : static void
4212 66108 : GlobalVisUpdate(void)
4213 : {
4214 : ComputeXidHorizonsResult horizons;
4215 :
4216 : /* updates the horizons as a side-effect */
4217 66108 : ComputeXidHorizons(&horizons);
4218 66108 : }
4219 :
4220 : /*
4221 : * Return true if no snapshot still considers fxid to be running.
4222 : *
4223 : * The state passed needs to have been initialized for the relation fxid is
4224 : * from (NULL is also OK), otherwise the result may not be correct.
4225 : *
4226 : * See comment for GlobalVisState for details.
4227 : */
4228 : bool
4229 12420821 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4230 : FullTransactionId fxid)
4231 : {
4232 : /*
4233 : * If fxid is older than maybe_needed bound, it definitely is visible to
4234 : * everyone.
4235 : */
4236 12420821 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4237 3118222 : return true;
4238 :
4239 : /*
4240 : * If fxid is >= definitely_needed bound, it is very likely to still be
4241 : * considered running.
4242 : */
4243 9302599 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4244 8791803 : return false;
4245 :
4246 : /*
4247 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4248 : * might not exist a snapshot considering fxid running. If it makes sense,
4249 : * update boundaries and recheck.
4250 : */
4251 510796 : if (GlobalVisTestShouldUpdate(state))
4252 : {
4253 66108 : GlobalVisUpdate();
4254 :
4255 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4256 :
4257 66108 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4258 : }
4259 : else
4260 444688 : return false;
4261 : }
4262 :
4263 : /*
4264 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4265 : *
4266 : * It is crucial that this only gets called for xids from a source that
4267 : * protects against xid wraparounds (e.g. from a table and thus protected by
4268 : * relfrozenxid).
4269 : */
4270 : bool
4271 12419841 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4272 : {
4273 : FullTransactionId fxid;
4274 :
4275 : /*
4276 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4277 : * because we know the xid has to, at the very least, be between
4278 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4279 : * lock to determine either, we can just compare with
4280 : * state->definitely_needed, which was based on those value at the time
4281 : * the current snapshot was built.
4282 : */
4283 12419841 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4284 :
4285 12419841 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4286 : }
4287 :
4288 : /*
4289 : * Convenience wrapper around GlobalVisTestFor() and
4290 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4291 : */
4292 : bool
4293 980 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4294 : {
4295 : GlobalVisState *state;
4296 :
4297 980 : state = GlobalVisTestFor(rel);
4298 :
4299 980 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4300 : }
4301 :
4302 : /*
4303 : * Convenience wrapper around GlobalVisTestFor() and
4304 : * GlobalVisTestIsRemovableXid(), see their comments.
4305 : */
4306 : bool
4307 8 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4308 : {
4309 : GlobalVisState *state;
4310 :
4311 8 : state = GlobalVisTestFor(rel);
4312 :
4313 8 : return GlobalVisTestIsRemovableXid(state, xid);
4314 : }
4315 :
4316 : /*
4317 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4318 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4319 : *
4320 : * Be very careful about when to use this function. It can only safely be used
4321 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4322 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4323 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4324 : * the xid has to be within that range), or if xid is from the procarray and
4325 : * prevents xid wraparound that way.
4326 : */
4327 : static inline FullTransactionId
4328 14949199 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4329 : {
4330 14949199 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4331 :
4332 : Assert(TransactionIdIsValid(xid));
4333 : Assert(TransactionIdIsValid(rel_xid));
4334 :
4335 : /* not guaranteed to find issues, but likely to catch mistakes */
4336 : AssertTransactionIdInAllowableRange(xid);
4337 :
4338 29898398 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4339 14949199 : + (int32) (xid - rel_xid));
4340 : }
4341 :
4342 :
4343 : /* ----------------------------------------------
4344 : * KnownAssignedTransactionIds sub-module
4345 : * ----------------------------------------------
4346 : */
4347 :
4348 : /*
4349 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4350 : * running on the primary at the current point in WAL. These XIDs must be
4351 : * treated as running by standby transactions, even though they are not in
4352 : * the standby server's PGPROC array.
4353 : *
4354 : * We record all XIDs that we know have been assigned. That includes all the
4355 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4356 : * been assigned. We can deduce the existence of unobserved XIDs because we
4357 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4358 : * list expands as new XIDs are observed or inferred, and contracts when
4359 : * transaction completion records arrive.
4360 : *
4361 : * During hot standby we do not fret too much about the distinction between
4362 : * top-level XIDs and subtransaction XIDs. We store both together in the
4363 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4364 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4365 : * doesn't care about the distinction either. Subtransaction XIDs are
4366 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4367 : * links are *not* maintained (which does not affect visibility).
4368 : *
4369 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4370 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4371 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4372 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4373 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4374 : * and then remove them from KnownAssignedXids. This prevents overflow of
4375 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4376 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4377 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4378 : * though it should be complete for top-level XIDs; this is the same situation
4379 : * that holds with respect to the PGPROC entries in normal running.
4380 : *
4381 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4382 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4383 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4384 : * As long as that is within the range of interesting XIDs, we have to assume
4385 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4386 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4387 : * subXID arrives - that is not an error.)
4388 : *
4389 : * Should a backend on primary somehow disappear before it can write an abort
4390 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4391 : * aborted but we think they were running; the distinction is irrelevant
4392 : * because either way any changes done by the transaction are not visible to
4393 : * backends in the standby. We prune KnownAssignedXids when
4394 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4395 : * array due to such dead XIDs.
4396 : */
4397 :
4398 : /*
4399 : * RecordKnownAssignedTransactionIds
4400 : * Record the given XID in KnownAssignedXids, as well as any preceding
4401 : * unobserved XIDs.
4402 : *
4403 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4404 : * associated with a transaction. Must be called for each record after we
4405 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4406 : *
4407 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4408 : */
4409 : void
4410 2549575 : RecordKnownAssignedTransactionIds(TransactionId xid)
4411 : {
4412 : Assert(standbyState >= STANDBY_INITIALIZED);
4413 : Assert(TransactionIdIsValid(xid));
4414 : Assert(TransactionIdIsValid(latestObservedXid));
4415 :
4416 2549575 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4417 : xid, latestObservedXid);
4418 :
4419 : /*
4420 : * When a newly observed xid arrives, it is frequently the case that it is
4421 : * *not* the next xid in sequence. When this occurs, we must treat the
4422 : * intervening xids as running also.
4423 : */
4424 2549575 : if (TransactionIdFollows(xid, latestObservedXid))
4425 : {
4426 : TransactionId next_expected_xid;
4427 :
4428 : /*
4429 : * Extend subtrans like we do in GetNewTransactionId() during normal
4430 : * operation using individual extend steps. Note that we do not need
4431 : * to extend clog since its extensions are WAL logged.
4432 : *
4433 : * This part has to be done regardless of standbyState since we
4434 : * immediately start assigning subtransactions to their toplevel
4435 : * transactions.
4436 : */
4437 23801 : next_expected_xid = latestObservedXid;
4438 48334 : while (TransactionIdPrecedes(next_expected_xid, xid))
4439 : {
4440 24533 : TransactionIdAdvance(next_expected_xid);
4441 24533 : ExtendSUBTRANS(next_expected_xid);
4442 : }
4443 : Assert(next_expected_xid == xid);
4444 :
4445 : /*
4446 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4447 : * more to do since we don't track assigned xids yet.
4448 : */
4449 23801 : if (standbyState <= STANDBY_INITIALIZED)
4450 : {
4451 0 : latestObservedXid = xid;
4452 0 : return;
4453 : }
4454 :
4455 : /*
4456 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4457 : */
4458 23801 : next_expected_xid = latestObservedXid;
4459 23801 : TransactionIdAdvance(next_expected_xid);
4460 23801 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4461 :
4462 : /*
4463 : * Now we can advance latestObservedXid
4464 : */
4465 23801 : latestObservedXid = xid;
4466 :
4467 : /* TransamVariables->nextXid must be beyond any observed xid */
4468 23801 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4469 : }
4470 : }
4471 :
4472 : /*
4473 : * ExpireTreeKnownAssignedTransactionIds
4474 : * Remove the given XIDs from KnownAssignedXids.
4475 : *
4476 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4477 : */
4478 : void
4479 23004 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4480 : TransactionId *subxids, TransactionId max_xid)
4481 : {
4482 : Assert(standbyState >= STANDBY_INITIALIZED);
4483 :
4484 : /*
4485 : * Uses same locking as transaction commit
4486 : */
4487 23004 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4488 :
4489 23004 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4490 :
4491 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4492 23004 : MaintainLatestCompletedXidRecovery(max_xid);
4493 :
4494 : /* ... and xactCompletionCount */
4495 23004 : TransamVariables->xactCompletionCount++;
4496 :
4497 23004 : LWLockRelease(ProcArrayLock);
4498 23004 : }
4499 :
4500 : /*
4501 : * ExpireAllKnownAssignedTransactionIds
4502 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4503 : */
4504 : void
4505 114 : ExpireAllKnownAssignedTransactionIds(void)
4506 : {
4507 : FullTransactionId latestXid;
4508 :
4509 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4510 114 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4511 :
4512 : /* Reset latestCompletedXid to nextXid - 1 */
4513 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
4514 114 : latestXid = TransamVariables->nextXid;
4515 114 : FullTransactionIdRetreat(&latestXid);
4516 114 : TransamVariables->latestCompletedXid = latestXid;
4517 :
4518 : /*
4519 : * Any transactions that were in-progress were effectively aborted, so
4520 : * advance xactCompletionCount.
4521 : */
4522 114 : TransamVariables->xactCompletionCount++;
4523 :
4524 : /*
4525 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4526 : * the call of this function. But do this for unification with what
4527 : * ExpireOldKnownAssignedTransactionIds() do.
4528 : */
4529 114 : procArray->lastOverflowedXid = InvalidTransactionId;
4530 114 : LWLockRelease(ProcArrayLock);
4531 114 : }
4532 :
4533 : /*
4534 : * ExpireOldKnownAssignedTransactionIds
4535 : * Remove KnownAssignedXids entries preceding the given XID and
4536 : * potentially reset lastOverflowedXid.
4537 : */
4538 : void
4539 826 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4540 : {
4541 : TransactionId latestXid;
4542 :
4543 826 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4544 :
4545 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4546 826 : latestXid = xid;
4547 826 : TransactionIdRetreat(latestXid);
4548 826 : MaintainLatestCompletedXidRecovery(latestXid);
4549 :
4550 : /* ... and xactCompletionCount */
4551 826 : TransamVariables->xactCompletionCount++;
4552 :
4553 : /*
4554 : * Reset lastOverflowedXid if we know all transactions that have been
4555 : * possibly running are being gone. Not doing so could cause an incorrect
4556 : * lastOverflowedXid value, which makes extra snapshots be marked as
4557 : * suboverflowed.
4558 : */
4559 826 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4560 819 : procArray->lastOverflowedXid = InvalidTransactionId;
4561 826 : KnownAssignedXidsRemovePreceding(xid);
4562 826 : LWLockRelease(ProcArrayLock);
4563 826 : }
4564 :
4565 : /*
4566 : * KnownAssignedTransactionIdsIdleMaintenance
4567 : * Opportunistically do maintenance work when the startup process
4568 : * is about to go idle.
4569 : */
4570 : void
4571 15030 : KnownAssignedTransactionIdsIdleMaintenance(void)
4572 : {
4573 15030 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4574 15030 : }
4575 :
4576 :
4577 : /*
4578 : * Private module functions to manipulate KnownAssignedXids
4579 : *
4580 : * There are 5 main uses of the KnownAssignedXids data structure:
4581 : *
4582 : * * backends taking snapshots - all valid XIDs need to be copied out
4583 : * * backends seeking to determine presence of a specific XID
4584 : * * startup process adding new known-assigned XIDs
4585 : * * startup process removing specific XIDs as transactions end
4586 : * * startup process pruning array when special WAL records arrive
4587 : *
4588 : * This data structure is known to be a hot spot during Hot Standby, so we
4589 : * go to some lengths to make these operations as efficient and as concurrent
4590 : * as possible.
4591 : *
4592 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4593 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4594 : * in general TransactionIdPrecedes would not provide a total order, but
4595 : * we know that the entries present at any instant should not extend across
4596 : * a large enough fraction of XID space to wrap around (the primary would
4597 : * shut down for fear of XID wrap long before that happens). So it's OK to
4598 : * use TransactionIdPrecedes as a binary-search comparator.
4599 : *
4600 : * It's cheap to maintain the sortedness during insertions, since new known
4601 : * XIDs are always reported in XID order; we just append them at the right.
4602 : *
4603 : * To keep individual deletions cheap, we need to allow gaps in the array.
4604 : * This is implemented by marking array elements as valid or invalid using
4605 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4606 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4607 : * XID entry itself. This preserves the property that the XID entries are
4608 : * sorted, so we can do binary searches easily. Periodically we compress
4609 : * out the unused entries; that's much cheaper than having to compress the
4610 : * array immediately on every deletion.
4611 : *
4612 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4613 : * are those with indexes tail <= i < head; items outside this subscript range
4614 : * have unspecified contents. When head reaches the end of the array, we
4615 : * force compression of unused entries rather than wrapping around, since
4616 : * allowing wraparound would greatly complicate the search logic. We maintain
4617 : * an explicit tail pointer so that pruning of old XIDs can be done without
4618 : * immediately moving the array contents. In most cases only a small fraction
4619 : * of the array contains valid entries at any instant.
4620 : *
4621 : * Although only the startup process can ever change the KnownAssignedXids
4622 : * data structure, we still need interlocking so that standby backends will
4623 : * not observe invalid intermediate states. The convention is that backends
4624 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4625 : * the array, the startup process must hold ProcArrayLock exclusively, for
4626 : * the usual transactional reasons (compare commit/abort of a transaction
4627 : * during normal running). Compressing unused entries out of the array
4628 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4629 : * them into slots to the right of the head pointer and then advance the head
4630 : * pointer. This doesn't require any lock at all, but on machines with weak
4631 : * memory ordering, we need to be careful that other processors see the array
4632 : * element changes before they see the head pointer change. We handle this by
4633 : * using memory barriers when reading or writing the head/tail pointers (unless
4634 : * the caller holds ProcArrayLock exclusively).
4635 : *
4636 : * Algorithmic analysis:
4637 : *
4638 : * If we have a maximum of M slots, with N XIDs currently spread across
4639 : * S elements then we have N <= S <= M always.
4640 : *
4641 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4642 : * happen)
4643 : * * Compressing the array is O(S) and requires exclusive lock
4644 : * * Removing an XID is O(logS) and requires exclusive lock
4645 : * * Taking a snapshot is O(S) and requires shared lock
4646 : * * Checking for an XID is O(logS) and requires shared lock
4647 : *
4648 : * In comparison, using a hash table for KnownAssignedXids would mean that
4649 : * taking snapshots would be O(M). If we can maintain S << M then the
4650 : * sorted array technique will deliver significantly faster snapshots.
4651 : * If we try to keep S too small then we will spend too much time compressing,
4652 : * so there is an optimal point for any workload mix. We use a heuristic to
4653 : * decide when to compress the array, though trimming also helps reduce
4654 : * frequency of compressing. The heuristic requires us to track the number of
4655 : * currently valid XIDs in the array (N). Except in special cases, we'll
4656 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4657 : * taking a snapshot to be O(N), which it would have to be anyway.
4658 : */
4659 :
4660 :
4661 : /*
4662 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4663 : * array, removing any gaps.
4664 : *
4665 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4666 : * we do it only if a heuristic indicates it's a good time to do it.
4667 : *
4668 : * Compression requires holding ProcArrayLock in exclusive mode.
4669 : * Caller must pass haveLock = true if it already holds the lock.
4670 : */
4671 : static void
4672 38881 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4673 : {
4674 38881 : ProcArrayStruct *pArray = procArray;
4675 : int head,
4676 : tail,
4677 : nelements;
4678 : int compress_index;
4679 : int i;
4680 :
4681 : /* Counters for compression heuristics */
4682 : static unsigned int transactionEndsCounter;
4683 : static TimestampTz lastCompressTs;
4684 :
4685 : /* Tuning constants */
4686 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4687 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4688 :
4689 : /*
4690 : * Since only the startup process modifies the head/tail pointers, we
4691 : * don't need a lock to read them here.
4692 : */
4693 38881 : head = pArray->headKnownAssignedXids;
4694 38881 : tail = pArray->tailKnownAssignedXids;
4695 38881 : nelements = head - tail;
4696 :
4697 : /*
4698 : * If we can choose whether to compress, use a heuristic to avoid
4699 : * compressing too often or not often enough. "Compress" here simply
4700 : * means moving the values to the beginning of the array, so it is not as
4701 : * complex or costly as typical data compression algorithms.
4702 : */
4703 38881 : if (nelements == pArray->numKnownAssignedXids)
4704 : {
4705 : /*
4706 : * When there are no gaps between head and tail, don't bother to
4707 : * compress, except in the KAX_NO_SPACE case where we must compress to
4708 : * create some space after the head.
4709 : */
4710 21036 : if (reason != KAX_NO_SPACE)
4711 21036 : return;
4712 : }
4713 17845 : else if (reason == KAX_TRANSACTION_END)
4714 : {
4715 : /*
4716 : * Consider compressing only once every so many commits. Frequency
4717 : * determined by benchmarks.
4718 : */
4719 12716 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4720 12607 : return;
4721 :
4722 : /*
4723 : * Furthermore, compress only if the used part of the array is less
4724 : * than 50% full (see comments above).
4725 : */
4726 109 : if (nelements < 2 * pArray->numKnownAssignedXids)
4727 9 : return;
4728 : }
4729 5129 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4730 : {
4731 : /*
4732 : * We're about to go idle for lack of new WAL, so we might as well
4733 : * compress. But not too often, to avoid ProcArray lock contention
4734 : * with readers.
4735 : */
4736 5014 : if (lastCompressTs != 0)
4737 : {
4738 : TimestampTz compress_after;
4739 :
4740 5013 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4741 : KAX_COMPRESS_IDLE_INTERVAL);
4742 5013 : if (GetCurrentTimestamp() < compress_after)
4743 4979 : return;
4744 : }
4745 : }
4746 :
4747 : /* Need to compress, so get the lock if we don't have it. */
4748 250 : if (!haveLock)
4749 35 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4750 :
4751 : /*
4752 : * We compress the array by reading the valid values from tail to head,
4753 : * re-aligning data to 0th element.
4754 : */
4755 250 : compress_index = 0;
4756 8604 : for (i = tail; i < head; i++)
4757 : {
4758 8354 : if (KnownAssignedXidsValid[i])
4759 : {
4760 864 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4761 864 : KnownAssignedXidsValid[compress_index] = true;
4762 864 : compress_index++;
4763 : }
4764 : }
4765 : Assert(compress_index == pArray->numKnownAssignedXids);
4766 :
4767 250 : pArray->tailKnownAssignedXids = 0;
4768 250 : pArray->headKnownAssignedXids = compress_index;
4769 :
4770 250 : if (!haveLock)
4771 35 : LWLockRelease(ProcArrayLock);
4772 :
4773 : /* Update timestamp for maintenance. No need to hold lock for this. */
4774 250 : lastCompressTs = GetCurrentTimestamp();
4775 : }
4776 :
4777 : /*
4778 : * Add xids into KnownAssignedXids at the head of the array.
4779 : *
4780 : * xids from from_xid to to_xid, inclusive, are added to the array.
4781 : *
4782 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4783 : * exclusive mode, so we need no extra locking here. Else caller holds no
4784 : * lock, so we need to be sure we maintain sufficient interlocks against
4785 : * concurrent readers. (Only the startup process ever calls this, so no need
4786 : * to worry about concurrent writers.)
4787 : */
4788 : static void
4789 23806 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4790 : bool exclusive_lock)
4791 : {
4792 23806 : ProcArrayStruct *pArray = procArray;
4793 : TransactionId next_xid;
4794 : int head,
4795 : tail;
4796 : int nxids;
4797 : int i;
4798 :
4799 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4800 :
4801 : /*
4802 : * Calculate how many array slots we'll need. Normally this is cheap; in
4803 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4804 : * way.
4805 : */
4806 23806 : if (to_xid >= from_xid)
4807 23806 : nxids = to_xid - from_xid + 1;
4808 : else
4809 : {
4810 0 : nxids = 1;
4811 0 : next_xid = from_xid;
4812 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4813 : {
4814 0 : nxids++;
4815 0 : TransactionIdAdvance(next_xid);
4816 : }
4817 : }
4818 :
4819 : /*
4820 : * Since only the startup process modifies the head/tail pointers, we
4821 : * don't need a lock to read them here.
4822 : */
4823 23806 : head = pArray->headKnownAssignedXids;
4824 23806 : tail = pArray->tailKnownAssignedXids;
4825 :
4826 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4827 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4828 :
4829 : /*
4830 : * Verify that insertions occur in TransactionId sequence. Note that even
4831 : * if the last existing element is marked invalid, it must still have a
4832 : * correctly sequenced XID value.
4833 : */
4834 38022 : if (head > tail &&
4835 14216 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4836 : {
4837 0 : KnownAssignedXidsDisplay(LOG);
4838 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4839 : }
4840 :
4841 : /*
4842 : * If our xids won't fit in the remaining space, compress out free space
4843 : */
4844 23806 : if (head + nxids > pArray->maxKnownAssignedXids)
4845 : {
4846 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4847 :
4848 0 : head = pArray->headKnownAssignedXids;
4849 : /* note: we no longer care about the tail pointer */
4850 :
4851 : /*
4852 : * If it still won't fit then we're out of memory
4853 : */
4854 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4855 0 : elog(ERROR, "too many KnownAssignedXids");
4856 : }
4857 :
4858 : /* Now we can insert the xids into the space starting at head */
4859 23806 : next_xid = from_xid;
4860 48344 : for (i = 0; i < nxids; i++)
4861 : {
4862 24538 : KnownAssignedXids[head] = next_xid;
4863 24538 : KnownAssignedXidsValid[head] = true;
4864 24538 : TransactionIdAdvance(next_xid);
4865 24538 : head++;
4866 : }
4867 :
4868 : /* Adjust count of number of valid entries */
4869 23806 : pArray->numKnownAssignedXids += nxids;
4870 :
4871 : /*
4872 : * Now update the head pointer. We use a write barrier to ensure that
4873 : * other processors see the above array updates before they see the head
4874 : * pointer change. The barrier isn't required if we're holding
4875 : * ProcArrayLock exclusively.
4876 : */
4877 23806 : if (!exclusive_lock)
4878 23801 : pg_write_barrier();
4879 :
4880 23806 : pArray->headKnownAssignedXids = head;
4881 23806 : }
4882 :
4883 : /*
4884 : * KnownAssignedXidsSearch
4885 : *
4886 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4887 : * Returns true if it was found, false if not.
4888 : *
4889 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4890 : * Exclusive lock must be held for remove = true.
4891 : */
4892 : static bool
4893 25673 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4894 : {
4895 25673 : ProcArrayStruct *pArray = procArray;
4896 : int first,
4897 : last;
4898 : int head;
4899 : int tail;
4900 25673 : int result_index = -1;
4901 :
4902 25673 : tail = pArray->tailKnownAssignedXids;
4903 25673 : head = pArray->headKnownAssignedXids;
4904 :
4905 : /*
4906 : * Only the startup process removes entries, so we don't need the read
4907 : * barrier in that case.
4908 : */
4909 25673 : if (!remove)
4910 1 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4911 :
4912 : /*
4913 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4914 : * array here, since even invalid entries will contain sorted XIDs.
4915 : */
4916 25673 : first = tail;
4917 25673 : last = head - 1;
4918 80939 : while (first <= last)
4919 : {
4920 : int mid_index;
4921 : TransactionId mid_xid;
4922 :
4923 79753 : mid_index = (first + last) / 2;
4924 79753 : mid_xid = KnownAssignedXids[mid_index];
4925 :
4926 79753 : if (xid == mid_xid)
4927 : {
4928 24487 : result_index = mid_index;
4929 24487 : break;
4930 : }
4931 55266 : else if (TransactionIdPrecedes(xid, mid_xid))
4932 12401 : last = mid_index - 1;
4933 : else
4934 42865 : first = mid_index + 1;
4935 : }
4936 :
4937 25673 : if (result_index < 0)
4938 1186 : return false; /* not in array */
4939 :
4940 24487 : if (!KnownAssignedXidsValid[result_index])
4941 35 : return false; /* in array, but invalid */
4942 :
4943 24452 : if (remove)
4944 : {
4945 24452 : KnownAssignedXidsValid[result_index] = false;
4946 :
4947 24452 : pArray->numKnownAssignedXids--;
4948 : Assert(pArray->numKnownAssignedXids >= 0);
4949 :
4950 : /*
4951 : * If we're removing the tail element then advance tail pointer over
4952 : * any invalid elements. This will speed future searches.
4953 : */
4954 24452 : if (result_index == tail)
4955 : {
4956 10806 : tail++;
4957 16962 : while (tail < head && !KnownAssignedXidsValid[tail])
4958 6156 : tail++;
4959 10806 : if (tail >= head)
4960 : {
4961 : /* Array is empty, so we can reset both pointers */
4962 9580 : pArray->headKnownAssignedXids = 0;
4963 9580 : pArray->tailKnownAssignedXids = 0;
4964 : }
4965 : else
4966 : {
4967 1226 : pArray->tailKnownAssignedXids = tail;
4968 : }
4969 : }
4970 : }
4971 :
4972 24452 : return true;
4973 : }
4974 :
4975 : /*
4976 : * Is the specified XID present in KnownAssignedXids[]?
4977 : *
4978 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4979 : */
4980 : static bool
4981 1 : KnownAssignedXidExists(TransactionId xid)
4982 : {
4983 : Assert(TransactionIdIsValid(xid));
4984 :
4985 1 : return KnownAssignedXidsSearch(xid, false);
4986 : }
4987 :
4988 : /*
4989 : * Remove the specified XID from KnownAssignedXids[].
4990 : *
4991 : * Caller must hold ProcArrayLock in exclusive mode.
4992 : */
4993 : static void
4994 25672 : KnownAssignedXidsRemove(TransactionId xid)
4995 : {
4996 : Assert(TransactionIdIsValid(xid));
4997 :
4998 25672 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
4999 :
5000 : /*
5001 : * Note: we cannot consider it an error to remove an XID that's not
5002 : * present. We intentionally remove subxact IDs while processing
5003 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
5004 : * removed again when the top-level xact commits or aborts.
5005 : *
5006 : * It might be possible to track such XIDs to distinguish this case from
5007 : * actual errors, but it would be complicated and probably not worth it.
5008 : * So, just ignore the search result.
5009 : */
5010 25672 : (void) KnownAssignedXidsSearch(xid, true);
5011 25672 : }
5012 :
5013 : /*
5014 : * KnownAssignedXidsRemoveTree
5015 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
5016 : *
5017 : * Caller must hold ProcArrayLock in exclusive mode.
5018 : */
5019 : static void
5020 23025 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
5021 : TransactionId *subxids)
5022 : {
5023 : int i;
5024 :
5025 23025 : if (TransactionIdIsValid(xid))
5026 23004 : KnownAssignedXidsRemove(xid);
5027 :
5028 25693 : for (i = 0; i < nsubxids; i++)
5029 2668 : KnownAssignedXidsRemove(subxids[i]);
5030 :
5031 : /* Opportunistically compress the array */
5032 23025 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5033 23025 : }
5034 :
5035 : /*
5036 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5037 : * then clear the whole table.
5038 : *
5039 : * Caller must hold ProcArrayLock in exclusive mode.
5040 : */
5041 : static void
5042 940 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5043 : {
5044 940 : ProcArrayStruct *pArray = procArray;
5045 940 : int count = 0;
5046 : int head,
5047 : tail,
5048 : i;
5049 :
5050 940 : if (!TransactionIdIsValid(removeXid))
5051 : {
5052 114 : elog(DEBUG4, "removing all KnownAssignedXids");
5053 114 : pArray->numKnownAssignedXids = 0;
5054 114 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5055 114 : return;
5056 : }
5057 :
5058 826 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5059 :
5060 : /*
5061 : * Mark entries invalid starting at the tail. Since array is sorted, we
5062 : * can stop as soon as we reach an entry >= removeXid.
5063 : */
5064 826 : tail = pArray->tailKnownAssignedXids;
5065 826 : head = pArray->headKnownAssignedXids;
5066 :
5067 826 : for (i = tail; i < head; i++)
5068 : {
5069 195 : if (KnownAssignedXidsValid[i])
5070 : {
5071 195 : TransactionId knownXid = KnownAssignedXids[i];
5072 :
5073 195 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5074 195 : break;
5075 :
5076 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5077 : {
5078 0 : KnownAssignedXidsValid[i] = false;
5079 0 : count++;
5080 : }
5081 : }
5082 : }
5083 :
5084 826 : pArray->numKnownAssignedXids -= count;
5085 : Assert(pArray->numKnownAssignedXids >= 0);
5086 :
5087 : /*
5088 : * Advance the tail pointer if we've marked the tail item invalid.
5089 : */
5090 826 : for (i = tail; i < head; i++)
5091 : {
5092 195 : if (KnownAssignedXidsValid[i])
5093 195 : break;
5094 : }
5095 826 : if (i >= head)
5096 : {
5097 : /* Array is empty, so we can reset both pointers */
5098 631 : pArray->headKnownAssignedXids = 0;
5099 631 : pArray->tailKnownAssignedXids = 0;
5100 : }
5101 : else
5102 : {
5103 195 : pArray->tailKnownAssignedXids = i;
5104 : }
5105 :
5106 : /* Opportunistically compress the array */
5107 826 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5108 : }
5109 :
5110 : /*
5111 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5112 : * We filter out anything >= xmax.
5113 : *
5114 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5115 : * that array is large enough.
5116 : *
5117 : * Caller must hold ProcArrayLock in (at least) shared mode.
5118 : */
5119 : static int
5120 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5121 : {
5122 0 : TransactionId xtmp = InvalidTransactionId;
5123 :
5124 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5125 : }
5126 :
5127 : /*
5128 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5129 : * we reduce *xmin to the lowest xid value seen if not already lower.
5130 : *
5131 : * Caller must hold ProcArrayLock in (at least) shared mode.
5132 : */
5133 : static int
5134 1288 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5135 : TransactionId xmax)
5136 : {
5137 1288 : int count = 0;
5138 : int head,
5139 : tail;
5140 : int i;
5141 :
5142 : /*
5143 : * Fetch head just once, since it may change while we loop. We can stop
5144 : * once we reach the initially seen head, since we are certain that an xid
5145 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5146 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5147 : * anyway.
5148 : */
5149 1288 : tail = procArray->tailKnownAssignedXids;
5150 1288 : head = procArray->headKnownAssignedXids;
5151 :
5152 1288 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5153 :
5154 1312 : for (i = tail; i < head; i++)
5155 : {
5156 : /* Skip any gaps in the array */
5157 147 : if (KnownAssignedXidsValid[i])
5158 : {
5159 137 : TransactionId knownXid = KnownAssignedXids[i];
5160 :
5161 : /*
5162 : * Update xmin if required. Only the first XID need be checked,
5163 : * since the array is sorted.
5164 : */
5165 274 : if (count == 0 &&
5166 137 : TransactionIdPrecedes(knownXid, *xmin))
5167 14 : *xmin = knownXid;
5168 :
5169 : /*
5170 : * Filter out anything >= xmax, again relying on sorted property
5171 : * of array.
5172 : */
5173 274 : if (TransactionIdIsValid(xmax) &&
5174 137 : TransactionIdFollowsOrEquals(knownXid, xmax))
5175 123 : break;
5176 :
5177 : /* Add knownXid into output array */
5178 14 : xarray[count++] = knownXid;
5179 : }
5180 : }
5181 :
5182 1288 : return count;
5183 : }
5184 :
5185 : /*
5186 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5187 : * if nothing there.
5188 : */
5189 : static TransactionId
5190 373 : KnownAssignedXidsGetOldestXmin(void)
5191 : {
5192 : int head,
5193 : tail;
5194 : int i;
5195 :
5196 : /*
5197 : * Fetch head just once, since it may change while we loop.
5198 : */
5199 373 : tail = procArray->tailKnownAssignedXids;
5200 373 : head = procArray->headKnownAssignedXids;
5201 :
5202 373 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5203 :
5204 373 : for (i = tail; i < head; i++)
5205 : {
5206 : /* Skip any gaps in the array */
5207 147 : if (KnownAssignedXidsValid[i])
5208 147 : return KnownAssignedXids[i];
5209 : }
5210 :
5211 226 : return InvalidTransactionId;
5212 : }
5213 :
5214 : /*
5215 : * Display KnownAssignedXids to provide debug trail
5216 : *
5217 : * Currently this is only called within startup process, so we need no
5218 : * special locking.
5219 : *
5220 : * Note this is pretty expensive, and much of the expense will be incurred
5221 : * even if the elog message will get discarded. It's not currently called
5222 : * in any performance-critical places, however, so no need to be tenser.
5223 : */
5224 : static void
5225 119 : KnownAssignedXidsDisplay(int trace_level)
5226 : {
5227 119 : ProcArrayStruct *pArray = procArray;
5228 : StringInfoData buf;
5229 : int head,
5230 : tail,
5231 : i;
5232 119 : int nxids = 0;
5233 :
5234 119 : tail = pArray->tailKnownAssignedXids;
5235 119 : head = pArray->headKnownAssignedXids;
5236 :
5237 119 : initStringInfo(&buf);
5238 :
5239 129 : for (i = tail; i < head; i++)
5240 : {
5241 10 : if (KnownAssignedXidsValid[i])
5242 : {
5243 10 : nxids++;
5244 10 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5245 : }
5246 : }
5247 :
5248 119 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5249 : nxids,
5250 : pArray->numKnownAssignedXids,
5251 : pArray->tailKnownAssignedXids,
5252 : pArray->headKnownAssignedXids,
5253 : buf.data);
5254 :
5255 119 : pfree(buf.data);
5256 119 : }
5257 :
5258 : /*
5259 : * KnownAssignedXidsReset
5260 : * Resets KnownAssignedXids to be empty
5261 : */
5262 : static void
5263 0 : KnownAssignedXidsReset(void)
5264 : {
5265 0 : ProcArrayStruct *pArray = procArray;
5266 :
5267 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5268 :
5269 0 : pArray->numKnownAssignedXids = 0;
5270 0 : pArray->tailKnownAssignedXids = 0;
5271 0 : pArray->headKnownAssignedXids = 0;
5272 :
5273 0 : LWLockRelease(ProcArrayLock);
5274 0 : }
|
