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) 1996-2020, 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/clog.h"
51 : #include "access/subtrans.h"
52 : #include "access/transam.h"
53 : #include "access/twophase.h"
54 : #include "access/xact.h"
55 : #include "access/xlog.h"
56 : #include "catalog/catalog.h"
57 : #include "catalog/pg_authid.h"
58 : #include "commands/dbcommands.h"
59 : #include "miscadmin.h"
60 : #include "pgstat.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/spin.h"
64 : #include "utils/acl.h"
65 : #include "utils/builtins.h"
66 : #include "utils/rel.h"
67 : #include "utils/snapmgr.h"
68 :
69 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
70 :
71 : /* Our shared memory area */
72 : typedef struct ProcArrayStruct
73 : {
74 : int numProcs; /* number of valid procs entries */
75 : int maxProcs; /* allocated size of procs array */
76 :
77 : /*
78 : * Known assigned XIDs handling
79 : */
80 : int maxKnownAssignedXids; /* allocated size of array */
81 : int numKnownAssignedXids; /* current # of valid entries */
82 : int tailKnownAssignedXids; /* index of oldest valid element */
83 : int headKnownAssignedXids; /* index of newest element, + 1 */
84 : slock_t known_assigned_xids_lck; /* protects head/tail pointers */
85 :
86 : /*
87 : * Highest subxid that has been removed from KnownAssignedXids array to
88 : * prevent overflow; or InvalidTransactionId if none. We track this for
89 : * similar reasons to tracking overflowing cached subxids in PGPROC
90 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
91 : * lock to read it.
92 : */
93 : TransactionId lastOverflowedXid;
94 :
95 : /* oldest xmin of any replication slot */
96 : TransactionId replication_slot_xmin;
97 : /* oldest catalog xmin of any replication slot */
98 : TransactionId replication_slot_catalog_xmin;
99 :
100 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
101 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
102 : } ProcArrayStruct;
103 :
104 : /*
105 : * State for the GlobalVisTest* family of functions. Those functions can
106 : * e.g. be used to decide if a deleted row can be removed without violating
107 : * MVCC semantics: If the deleted row's xmax is not considered to be running
108 : * by anyone, the row can be removed.
109 : *
110 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
111 : * cutoff XID while building a snapshot (looking at the frequently changing
112 : * xmins scales badly). Instead we compute two boundaries while building the
113 : * snapshot:
114 : *
115 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
116 : * definitely_needed are definitely still visible.
117 : *
118 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
119 : * definitely be removed
120 : *
121 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
122 : * && XID < definitely_needed), the boundaries can be recomputed (using
123 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
124 : * maintaining an accurate value all the time.
125 : *
126 : * As it is not cheap to compute accurate boundaries, we limit the number of
127 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
128 : *
129 : *
130 : * There are three backend lifetime instances of this struct, optimized for
131 : * different types of relations. As e.g. a normal user defined table in one
132 : * database is inaccessible to backends connected to another database, a test
133 : * specific to a relation can be more aggressive than a test for a shared
134 : * relation. Currently we track four different states:
135 : *
136 : * 1) GlobalVisSharedRels, which only considers an XID's
137 : * effects visible-to-everyone if neither snapshots in any database, nor a
138 : * replication slot's xmin, nor a replication slot's catalog_xmin might
139 : * still consider XID as running.
140 : *
141 : * 2) GlobalVisCatalogRels, which only considers an XID's
142 : * effects visible-to-everyone if neither snapshots in the current
143 : * database, nor a replication slot's xmin, nor a replication slot's
144 : * catalog_xmin might still consider XID as running.
145 : *
146 : * I.e. the difference to GlobalVisSharedRels is that
147 : * snapshot in other databases are ignored.
148 : *
149 : * 3) GlobalVisDataRels, which only considers an XID's
150 : * effects visible-to-everyone if neither snapshots in the current
151 : * database, nor a replication slot's xmin consider XID as running.
152 : *
153 : * I.e. the difference to GlobalVisCatalogRels is that
154 : * replication slot's catalog_xmin is not taken into account.
155 : *
156 : * 4) GlobalVisTempRels, which only considers the current session, as temp
157 : * tables are not visible to other sessions.
158 : *
159 : * GlobalVisTestFor(relation) returns the appropriate state
160 : * for the relation.
161 : *
162 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
163 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
164 : * prevent maybe_needed to become old enough after the GetSnapshotData()
165 : * call.
166 : *
167 : * The typedef is in the header.
168 : */
169 : struct GlobalVisState
170 : {
171 : /* XIDs >= are considered running by some backend */
172 : FullTransactionId definitely_needed;
173 :
174 : /* XIDs < are not considered to be running by any backend */
175 : FullTransactionId maybe_needed;
176 : };
177 :
178 : /*
179 : * Result of ComputeXidHorizons().
180 : */
181 : typedef struct ComputeXidHorizonsResult
182 : {
183 : /*
184 : * The value of ShmemVariableCache->latestCompletedXid when
185 : * ComputeXidHorizons() held ProcArrayLock.
186 : */
187 : FullTransactionId latest_completed;
188 :
189 : /*
190 : * The same for procArray->replication_slot_xmin and.
191 : * procArray->replication_slot_catalog_xmin.
192 : */
193 : TransactionId slot_xmin;
194 : TransactionId slot_catalog_xmin;
195 :
196 : /*
197 : * Oldest xid that any backend might still consider running. This needs to
198 : * include processes running VACUUM, in contrast to the normal visibility
199 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
200 : * determining visibility, but doesn't care about rows above its xmin to
201 : * be removed.
202 : *
203 : * This likely should only be needed to determine whether pg_subtrans can
204 : * be truncated. It currently includes the effects of replication slots,
205 : * for historical reasons. But that could likely be changed.
206 : */
207 : TransactionId oldest_considered_running;
208 :
209 : /*
210 : * Oldest xid for which deleted tuples need to be retained in shared
211 : * tables.
212 : *
213 : * This includes the effects of replication slots. If that's not desired,
214 : * look at shared_oldest_nonremovable_raw;
215 : */
216 : TransactionId shared_oldest_nonremovable;
217 :
218 : /*
219 : * Oldest xid that may be necessary to retain in shared tables. This is
220 : * the same as shared_oldest_nonremovable, except that is not affected by
221 : * replication slot's catalog_xmin.
222 : *
223 : * This is mainly useful to be able to send the catalog_xmin to upstream
224 : * streaming replication servers via hot_standby_feedback, so they can
225 : * apply the limit only when accessing catalog tables.
226 : */
227 : TransactionId shared_oldest_nonremovable_raw;
228 :
229 : /*
230 : * Oldest xid for which deleted tuples need to be retained in non-shared
231 : * catalog tables.
232 : */
233 : TransactionId catalog_oldest_nonremovable;
234 :
235 : /*
236 : * Oldest xid for which deleted tuples need to be retained in normal user
237 : * defined tables.
238 : */
239 : TransactionId data_oldest_nonremovable;
240 :
241 : /*
242 : * Oldest xid for which deleted tuples need to be retained in this
243 : * session's temporary tables.
244 : */
245 : TransactionId temp_oldest_nonremovable;
246 :
247 : } ComputeXidHorizonsResult;
248 :
249 :
250 : static ProcArrayStruct *procArray;
251 :
252 : static PGPROC *allProcs;
253 :
254 : /*
255 : * Bookkeeping for tracking emulated transactions in recovery
256 : */
257 : static TransactionId *KnownAssignedXids;
258 : static bool *KnownAssignedXidsValid;
259 : static TransactionId latestObservedXid = InvalidTransactionId;
260 :
261 : /*
262 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
263 : * the highest xid that might still be running that we don't have in
264 : * KnownAssignedXids.
265 : */
266 : static TransactionId standbySnapshotPendingXmin;
267 :
268 : /*
269 : * State for visibility checks on different types of relations. See struct
270 : * GlobalVisState for details. As shared, catalog, normal and temporary
271 : * relations can have different horizons, one such state exists for each.
272 : */
273 : static GlobalVisState GlobalVisSharedRels;
274 : static GlobalVisState GlobalVisCatalogRels;
275 : static GlobalVisState GlobalVisDataRels;
276 : static GlobalVisState GlobalVisTempRels;
277 :
278 : /*
279 : * This backend's RecentXmin at the last time the accurate xmin horizon was
280 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
281 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
282 : */
283 : static TransactionId ComputeXidHorizonsResultLastXmin;
284 :
285 : #ifdef XIDCACHE_DEBUG
286 :
287 : /* counters for XidCache measurement */
288 : static long xc_by_recent_xmin = 0;
289 : static long xc_by_known_xact = 0;
290 : static long xc_by_my_xact = 0;
291 : static long xc_by_latest_xid = 0;
292 : static long xc_by_main_xid = 0;
293 : static long xc_by_child_xid = 0;
294 : static long xc_by_known_assigned = 0;
295 : static long xc_no_overflow = 0;
296 : static long xc_slow_answer = 0;
297 :
298 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
299 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
300 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
301 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
302 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
303 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
304 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
305 : #define xc_no_overflow_inc() (xc_no_overflow++)
306 : #define xc_slow_answer_inc() (xc_slow_answer++)
307 :
308 : static void DisplayXidCache(void);
309 : #else /* !XIDCACHE_DEBUG */
310 :
311 : #define xc_by_recent_xmin_inc() ((void) 0)
312 : #define xc_by_known_xact_inc() ((void) 0)
313 : #define xc_by_my_xact_inc() ((void) 0)
314 : #define xc_by_latest_xid_inc() ((void) 0)
315 : #define xc_by_main_xid_inc() ((void) 0)
316 : #define xc_by_child_xid_inc() ((void) 0)
317 : #define xc_by_known_assigned_inc() ((void) 0)
318 : #define xc_no_overflow_inc() ((void) 0)
319 : #define xc_slow_answer_inc() ((void) 0)
320 : #endif /* XIDCACHE_DEBUG */
321 :
322 : /* Primitives for KnownAssignedXids array handling for standby */
323 : static void KnownAssignedXidsCompress(bool force);
324 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
325 : bool exclusive_lock);
326 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
327 : static bool KnownAssignedXidExists(TransactionId xid);
328 : static void KnownAssignedXidsRemove(TransactionId xid);
329 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
330 : TransactionId *subxids);
331 : static void KnownAssignedXidsRemovePreceding(TransactionId xid);
332 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
333 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
334 : TransactionId *xmin,
335 : TransactionId xmax);
336 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
337 : static void KnownAssignedXidsDisplay(int trace_level);
338 : static void KnownAssignedXidsReset(void);
339 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
340 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
341 : static void MaintainLatestCompletedXid(TransactionId latestXid);
342 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
343 :
344 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
345 : TransactionId xid);
346 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
347 :
348 : /*
349 : * Report shared-memory space needed by CreateSharedProcArray.
350 : */
351 : Size
352 2376 : ProcArrayShmemSize(void)
353 : {
354 : Size size;
355 :
356 : /* Size of the ProcArray structure itself */
357 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
358 :
359 2376 : size = offsetof(ProcArrayStruct, pgprocnos);
360 2376 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
361 :
362 : /*
363 : * During Hot Standby processing we have a data structure called
364 : * KnownAssignedXids, created in shared memory. Local data structures are
365 : * also created in various backends during GetSnapshotData(),
366 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
367 : * main structures created in those functions must be identically sized,
368 : * since we may at times copy the whole of the data structures around. We
369 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
370 : *
371 : * Ideally we'd only create this structure if we were actually doing hot
372 : * standby in the current run, but we don't know that yet at the time
373 : * shared memory is being set up.
374 : */
375 : #define TOTAL_MAX_CACHED_SUBXIDS \
376 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
377 :
378 2376 : if (EnableHotStandby)
379 : {
380 2376 : size = add_size(size,
381 : mul_size(sizeof(TransactionId),
382 2376 : TOTAL_MAX_CACHED_SUBXIDS));
383 2376 : size = add_size(size,
384 2376 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
385 : }
386 :
387 2376 : return size;
388 : }
389 :
390 : /*
391 : * Initialize the shared PGPROC array during postmaster startup.
392 : */
393 : void
394 2372 : CreateSharedProcArray(void)
395 : {
396 : bool found;
397 :
398 : /* Create or attach to the ProcArray shared structure */
399 2372 : procArray = (ProcArrayStruct *)
400 2372 : ShmemInitStruct("Proc Array",
401 : add_size(offsetof(ProcArrayStruct, pgprocnos),
402 : mul_size(sizeof(int),
403 2372 : PROCARRAY_MAXPROCS)),
404 : &found);
405 :
406 2372 : if (!found)
407 : {
408 : /*
409 : * We're the first - initialize.
410 : */
411 2372 : procArray->numProcs = 0;
412 2372 : procArray->maxProcs = PROCARRAY_MAXPROCS;
413 2372 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
414 2372 : procArray->numKnownAssignedXids = 0;
415 2372 : procArray->tailKnownAssignedXids = 0;
416 2372 : procArray->headKnownAssignedXids = 0;
417 2372 : SpinLockInit(&procArray->known_assigned_xids_lck);
418 2372 : procArray->lastOverflowedXid = InvalidTransactionId;
419 2372 : procArray->replication_slot_xmin = InvalidTransactionId;
420 2372 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
421 2372 : ShmemVariableCache->xactCompletionCount = 1;
422 : }
423 :
424 2372 : allProcs = ProcGlobal->allProcs;
425 :
426 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
427 2372 : if (EnableHotStandby)
428 : {
429 2372 : KnownAssignedXids = (TransactionId *)
430 2372 : ShmemInitStruct("KnownAssignedXids",
431 : mul_size(sizeof(TransactionId),
432 2372 : TOTAL_MAX_CACHED_SUBXIDS),
433 : &found);
434 2372 : KnownAssignedXidsValid = (bool *)
435 2372 : ShmemInitStruct("KnownAssignedXidsValid",
436 2372 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
437 : &found);
438 : }
439 2372 : }
440 :
441 : /*
442 : * Add the specified PGPROC to the shared array.
443 : */
444 : void
445 12002 : ProcArrayAdd(PGPROC *proc)
446 : {
447 12002 : ProcArrayStruct *arrayP = procArray;
448 : int index;
449 :
450 : /* See ProcGlobal comment explaining why both locks are held */
451 12002 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
452 12002 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
453 :
454 12002 : if (arrayP->numProcs >= arrayP->maxProcs)
455 : {
456 : /*
457 : * Oops, no room. (This really shouldn't happen, since there is a
458 : * fixed supply of PGPROC structs too, and so we should have failed
459 : * earlier.)
460 : */
461 0 : ereport(FATAL,
462 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
463 : errmsg("sorry, too many clients already")));
464 : }
465 :
466 : /*
467 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
468 : * locality of references much better. This is useful while traversing the
469 : * ProcArray because there is an increased likelihood of finding the next
470 : * PGPROC structure in the cache.
471 : *
472 : * Since the occurrence of adding/removing a proc is much lower than the
473 : * access to the ProcArray itself, the overhead should be marginal
474 : */
475 22446 : for (index = 0; index < arrayP->numProcs; index++)
476 : {
477 : /*
478 : * If we are the first PGPROC or if we have found our right position
479 : * in the array, break
480 : */
481 20192 : if ((arrayP->pgprocnos[index] == -1) || (arrayP->pgprocnos[index] > proc->pgprocno))
482 : break;
483 : }
484 :
485 12002 : memmove(&arrayP->pgprocnos[index + 1], &arrayP->pgprocnos[index],
486 12002 : (arrayP->numProcs - index) * sizeof(*arrayP->pgprocnos));
487 12002 : memmove(&ProcGlobal->xids[index + 1], &ProcGlobal->xids[index],
488 12002 : (arrayP->numProcs - index) * sizeof(*ProcGlobal->xids));
489 12002 : memmove(&ProcGlobal->subxidStates[index + 1], &ProcGlobal->subxidStates[index],
490 12002 : (arrayP->numProcs - index) * sizeof(*ProcGlobal->subxidStates));
491 12002 : memmove(&ProcGlobal->statusFlags[index + 1], &ProcGlobal->statusFlags[index],
492 12002 : (arrayP->numProcs - index) * sizeof(*ProcGlobal->statusFlags));
493 :
494 12002 : arrayP->pgprocnos[index] = proc->pgprocno;
495 12002 : ProcGlobal->xids[index] = proc->xid;
496 12002 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
497 12002 : ProcGlobal->statusFlags[index] = proc->statusFlags;
498 :
499 12002 : arrayP->numProcs++;
500 :
501 47540 : for (; index < arrayP->numProcs; index++)
502 : {
503 35538 : allProcs[arrayP->pgprocnos[index]].pgxactoff = index;
504 : }
505 :
506 : /*
507 : * Release in reversed acquisition order, to reduce frequency of having to
508 : * wait for XidGenLock while holding ProcArrayLock.
509 : */
510 12002 : LWLockRelease(XidGenLock);
511 12002 : LWLockRelease(ProcArrayLock);
512 12002 : }
513 :
514 : /*
515 : * Remove the specified PGPROC from the shared array.
516 : *
517 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
518 : * array, and thus causing it to appear as "not running" anymore. In this
519 : * case we must advance latestCompletedXid. (This is essentially the same
520 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
521 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
522 : * twophase.c depends on the latter.)
523 : */
524 : void
525 11976 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
526 : {
527 11976 : ProcArrayStruct *arrayP = procArray;
528 : int index;
529 :
530 : #ifdef XIDCACHE_DEBUG
531 : /* dump stats at backend shutdown, but not prepared-xact end */
532 : if (proc->pid != 0)
533 : DisplayXidCache();
534 : #endif
535 :
536 : /* See ProcGlobal comment explaining why both locks are held */
537 11976 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
538 11976 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
539 :
540 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[proc->pgxactoff]].pgxactoff == proc->pgxactoff);
541 :
542 11976 : if (TransactionIdIsValid(latestXid))
543 : {
544 : Assert(TransactionIdIsValid(ProcGlobal->xids[proc->pgxactoff]));
545 :
546 : /* Advance global latestCompletedXid while holding the lock */
547 70 : MaintainLatestCompletedXid(latestXid);
548 :
549 : /* Same with xactCompletionCount */
550 70 : ShmemVariableCache->xactCompletionCount++;
551 :
552 70 : ProcGlobal->xids[proc->pgxactoff] = 0;
553 70 : ProcGlobal->subxidStates[proc->pgxactoff].overflowed = false;
554 70 : ProcGlobal->subxidStates[proc->pgxactoff].count = 0;
555 : }
556 : else
557 : {
558 : /* Shouldn't be trying to remove a live transaction here */
559 : Assert(!TransactionIdIsValid(ProcGlobal->xids[proc->pgxactoff]));
560 : }
561 :
562 : Assert(TransactionIdIsValid(ProcGlobal->xids[proc->pgxactoff] == 0));
563 : Assert(TransactionIdIsValid(ProcGlobal->subxidStates[proc->pgxactoff].count == 0));
564 : Assert(TransactionIdIsValid(ProcGlobal->subxidStates[proc->pgxactoff].overflowed == false));
565 11976 : ProcGlobal->statusFlags[proc->pgxactoff] = 0;
566 :
567 22798 : for (index = 0; index < arrayP->numProcs; index++)
568 : {
569 22798 : if (arrayP->pgprocnos[index] == proc->pgprocno)
570 : {
571 : /* Keep the PGPROC array sorted. See notes above */
572 11976 : memmove(&arrayP->pgprocnos[index], &arrayP->pgprocnos[index + 1],
573 11976 : (arrayP->numProcs - index - 1) * sizeof(*arrayP->pgprocnos));
574 11976 : memmove(&ProcGlobal->xids[index], &ProcGlobal->xids[index + 1],
575 11976 : (arrayP->numProcs - index - 1) * sizeof(*ProcGlobal->xids));
576 11976 : memmove(&ProcGlobal->subxidStates[index], &ProcGlobal->subxidStates[index + 1],
577 11976 : (arrayP->numProcs - index - 1) * sizeof(*ProcGlobal->subxidStates));
578 11976 : memmove(&ProcGlobal->statusFlags[index], &ProcGlobal->statusFlags[index + 1],
579 11976 : (arrayP->numProcs - index - 1) * sizeof(*ProcGlobal->statusFlags));
580 :
581 11976 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
582 11976 : arrayP->numProcs--;
583 :
584 : /* adjust for removed PGPROC */
585 35174 : for (; index < arrayP->numProcs; index++)
586 23198 : allProcs[arrayP->pgprocnos[index]].pgxactoff--;
587 :
588 : /*
589 : * Release in reversed acquisition order, to reduce frequency of
590 : * having to wait for XidGenLock while holding ProcArrayLock.
591 : */
592 11976 : LWLockRelease(XidGenLock);
593 11976 : LWLockRelease(ProcArrayLock);
594 11976 : return;
595 : }
596 : }
597 :
598 : /* Oops */
599 0 : LWLockRelease(XidGenLock);
600 0 : LWLockRelease(ProcArrayLock);
601 :
602 0 : elog(LOG, "failed to find proc %p in ProcArray", proc);
603 : }
604 :
605 :
606 : /*
607 : * ProcArrayEndTransaction -- mark a transaction as no longer running
608 : *
609 : * This is used interchangeably for commit and abort cases. The transaction
610 : * commit/abort must already be reported to WAL and pg_xact.
611 : *
612 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
613 : * latestXid is the latest Xid among the transaction's main XID and
614 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
615 : * the caller to pass latestXid, instead of computing it from the PGPROC's
616 : * contents, because the subxid information in the PGPROC might be
617 : * incomplete.)
618 : */
619 : void
620 531756 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
621 : {
622 531756 : if (TransactionIdIsValid(latestXid))
623 : {
624 : /*
625 : * We must lock ProcArrayLock while clearing our advertised XID, so
626 : * that we do not exit the set of "running" transactions while someone
627 : * else is taking a snapshot. See discussion in
628 : * src/backend/access/transam/README.
629 : */
630 : Assert(TransactionIdIsValid(proc->xid));
631 :
632 : /*
633 : * If we can immediately acquire ProcArrayLock, we clear our own XID
634 : * and release the lock. If not, use group XID clearing to improve
635 : * efficiency.
636 : */
637 299392 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
638 : {
639 299318 : ProcArrayEndTransactionInternal(proc, latestXid);
640 299318 : LWLockRelease(ProcArrayLock);
641 : }
642 : else
643 74 : ProcArrayGroupClearXid(proc, latestXid);
644 : }
645 : else
646 : {
647 : /*
648 : * If we have no XID, we don't need to lock, since we won't affect
649 : * anyone else's calculation of a snapshot. We might change their
650 : * estimate of global xmin, but that's OK.
651 : */
652 : Assert(!TransactionIdIsValid(proc->xid));
653 : Assert(proc->subxidStatus.count == 0);
654 : Assert(!proc->subxidStatus.overflowed);
655 :
656 232364 : proc->lxid = InvalidLocalTransactionId;
657 232364 : proc->xmin = InvalidTransactionId;
658 232364 : proc->delayChkpt = false; /* be sure this is cleared in abort */
659 232364 : proc->recoveryConflictPending = false;
660 :
661 : /* must be cleared with xid/xmin: */
662 : /* avoid unnecessarily dirtying shared cachelines */
663 232364 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
664 : {
665 : Assert(!LWLockHeldByMe(ProcArrayLock));
666 48434 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
667 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
668 48434 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
669 48434 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
670 48434 : LWLockRelease(ProcArrayLock);
671 : }
672 : }
673 531756 : }
674 :
675 : /*
676 : * Mark a write transaction as no longer running.
677 : *
678 : * We don't do any locking here; caller must handle that.
679 : */
680 : static inline void
681 299392 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
682 : {
683 299392 : size_t pgxactoff = proc->pgxactoff;
684 :
685 : /*
686 : * Note: we need exclusive lock here because we're going to change other
687 : * processes' PGPROC entries.
688 : */
689 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
690 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
691 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
692 :
693 299392 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
694 299392 : proc->xid = InvalidTransactionId;
695 299392 : proc->lxid = InvalidLocalTransactionId;
696 299392 : proc->xmin = InvalidTransactionId;
697 299392 : proc->delayChkpt = false; /* be sure this is cleared in abort */
698 299392 : proc->recoveryConflictPending = false;
699 :
700 : /* must be cleared with xid/xmin: */
701 : /* avoid unnecessarily dirtying shared cachelines */
702 299392 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
703 : {
704 162 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
705 162 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
706 : }
707 :
708 : /* Clear the subtransaction-XID cache too while holding the lock */
709 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
710 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
711 299392 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
712 : {
713 422 : ProcGlobal->subxidStates[pgxactoff].count = 0;
714 422 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
715 422 : proc->subxidStatus.count = 0;
716 422 : proc->subxidStatus.overflowed = false;
717 : }
718 :
719 : /* Also advance global latestCompletedXid while holding the lock */
720 299392 : MaintainLatestCompletedXid(latestXid);
721 :
722 : /* Same with xactCompletionCount */
723 299392 : ShmemVariableCache->xactCompletionCount++;
724 299392 : }
725 :
726 : /*
727 : * ProcArrayGroupClearXid -- group XID clearing
728 : *
729 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
730 : * commit time, add ourselves to a list of processes that need their XIDs
731 : * cleared. The first process to add itself to the list will acquire
732 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
733 : * on behalf of all group members. This avoids a great deal of contention
734 : * around ProcArrayLock when many processes are trying to commit at once,
735 : * since the lock need not be repeatedly handed off from one committing
736 : * process to the next.
737 : */
738 : static void
739 74 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
740 : {
741 74 : PROC_HDR *procglobal = ProcGlobal;
742 : uint32 nextidx;
743 : uint32 wakeidx;
744 :
745 : /* We should definitely have an XID to clear. */
746 : Assert(TransactionIdIsValid(proc->xid));
747 :
748 : /* Add ourselves to the list of processes needing a group XID clear. */
749 74 : proc->procArrayGroupMember = true;
750 74 : proc->procArrayGroupMemberXid = latestXid;
751 74 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
752 : while (true)
753 : {
754 74 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
755 :
756 74 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
757 : &nextidx,
758 74 : (uint32) proc->pgprocno))
759 74 : break;
760 : }
761 :
762 : /*
763 : * If the list was not empty, the leader will clear our XID. It is
764 : * impossible to have followers without a leader because the first process
765 : * that has added itself to the list will always have nextidx as
766 : * INVALID_PGPROCNO.
767 : */
768 74 : if (nextidx != INVALID_PGPROCNO)
769 : {
770 10 : int extraWaits = 0;
771 :
772 : /* Sleep until the leader clears our XID. */
773 10 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
774 : for (;;)
775 : {
776 : /* acts as a read barrier */
777 10 : PGSemaphoreLock(proc->sem);
778 10 : if (!proc->procArrayGroupMember)
779 10 : break;
780 0 : extraWaits++;
781 : }
782 10 : pgstat_report_wait_end();
783 :
784 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PGPROCNO);
785 :
786 : /* Fix semaphore count for any absorbed wakeups */
787 10 : while (extraWaits-- > 0)
788 0 : PGSemaphoreUnlock(proc->sem);
789 10 : return;
790 : }
791 :
792 : /* We are the leader. Acquire the lock on behalf of everyone. */
793 64 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
794 :
795 : /*
796 : * Now that we've got the lock, clear the list of processes waiting for
797 : * group XID clearing, saving a pointer to the head of the list. Trying
798 : * to pop elements one at a time could lead to an ABA problem.
799 : */
800 64 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
801 : INVALID_PGPROCNO);
802 :
803 : /* Remember head of list so we can perform wakeups after dropping lock. */
804 64 : wakeidx = nextidx;
805 :
806 : /* Walk the list and clear all XIDs. */
807 138 : while (nextidx != INVALID_PGPROCNO)
808 : {
809 74 : PGPROC *proc = &allProcs[nextidx];
810 :
811 74 : ProcArrayEndTransactionInternal(proc, proc->procArrayGroupMemberXid);
812 :
813 : /* Move to next proc in list. */
814 74 : nextidx = pg_atomic_read_u32(&proc->procArrayGroupNext);
815 : }
816 :
817 : /* We're done with the lock now. */
818 64 : LWLockRelease(ProcArrayLock);
819 :
820 : /*
821 : * Now that we've released the lock, go back and wake everybody up. We
822 : * don't do this under the lock so as to keep lock hold times to a
823 : * minimum. The system calls we need to perform to wake other processes
824 : * up are probably much slower than the simple memory writes we did while
825 : * holding the lock.
826 : */
827 138 : while (wakeidx != INVALID_PGPROCNO)
828 : {
829 74 : PGPROC *proc = &allProcs[wakeidx];
830 :
831 74 : wakeidx = pg_atomic_read_u32(&proc->procArrayGroupNext);
832 74 : pg_atomic_write_u32(&proc->procArrayGroupNext, INVALID_PGPROCNO);
833 :
834 : /* ensure all previous writes are visible before follower continues. */
835 74 : pg_write_barrier();
836 :
837 74 : proc->procArrayGroupMember = false;
838 :
839 74 : if (proc != MyProc)
840 10 : PGSemaphoreUnlock(proc->sem);
841 : }
842 : }
843 :
844 : /*
845 : * ProcArrayClearTransaction -- clear the transaction fields
846 : *
847 : * This is used after successfully preparing a 2-phase transaction. We are
848 : * not actually reporting the transaction's XID as no longer running --- it
849 : * will still appear as running because the 2PC's gxact is in the ProcArray
850 : * too. We just have to clear out our own PGPROC.
851 : */
852 : void
853 66 : ProcArrayClearTransaction(PGPROC *proc)
854 : {
855 : size_t pgxactoff;
856 :
857 : /*
858 : * Currently we need to lock ProcArrayLock exclusively here, as we
859 : * increment xactCompletionCount below. We also need it at least in shared
860 : * mode for pgproc->pgxactoff to stay the same below.
861 : *
862 : * We could however, as this action does not actually change anyone's view
863 : * of the set of running XIDs (our entry is duplicate with the gxact that
864 : * has already been inserted into the ProcArray), lower the lock level to
865 : * shared if we were to make xactCompletionCount an atomic variable. But
866 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
867 : * enough for this not to be the bottleneck. If it ever becomes a
868 : * bottleneck it may also be worth considering to combine this with the
869 : * subsequent ProcArrayRemove()
870 : */
871 66 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
872 :
873 66 : pgxactoff = proc->pgxactoff;
874 :
875 66 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
876 66 : proc->xid = InvalidTransactionId;
877 :
878 66 : proc->lxid = InvalidLocalTransactionId;
879 66 : proc->xmin = InvalidTransactionId;
880 66 : proc->recoveryConflictPending = false;
881 :
882 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
883 : Assert(!proc->delayChkpt);
884 :
885 : /*
886 : * Need to increment completion count even though transaction hasn't
887 : * really committed yet. The reason for that is that GetSnapshotData()
888 : * omits the xid of the current transaction, thus without the increment we
889 : * otherwise could end up reusing the snapshot later. Which would be bad,
890 : * because it might not count the prepared transaction as running.
891 : */
892 66 : ShmemVariableCache->xactCompletionCount++;
893 :
894 : /* Clear the subtransaction-XID cache too */
895 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
896 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
897 66 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
898 : {
899 40 : ProcGlobal->subxidStates[pgxactoff].count = 0;
900 40 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
901 40 : proc->subxidStatus.count = 0;
902 40 : proc->subxidStatus.overflowed = false;
903 : }
904 :
905 66 : LWLockRelease(ProcArrayLock);
906 66 : }
907 :
908 : /*
909 : * Update ShmemVariableCache->latestCompletedXid to point to latestXid if
910 : * currently older.
911 : */
912 : static void
913 299846 : MaintainLatestCompletedXid(TransactionId latestXid)
914 : {
915 299846 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
916 :
917 : Assert(FullTransactionIdIsValid(cur_latest));
918 : Assert(!RecoveryInProgress());
919 : Assert(LWLockHeldByMe(ProcArrayLock));
920 :
921 299846 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
922 : {
923 291312 : ShmemVariableCache->latestCompletedXid =
924 291312 : FullXidRelativeTo(cur_latest, latestXid);
925 : }
926 :
927 : Assert(IsBootstrapProcessingMode() ||
928 : FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
929 299846 : }
930 :
931 : /*
932 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
933 : */
934 : static void
935 364 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
936 : {
937 364 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
938 : FullTransactionId rel;
939 :
940 : Assert(AmStartupProcess() || !IsUnderPostmaster);
941 : Assert(LWLockHeldByMe(ProcArrayLock));
942 :
943 : /*
944 : * Need a FullTransactionId to compare latestXid with. Can't rely on
945 : * latestCompletedXid to be initialized in recovery. But in recovery it's
946 : * safe to access nextXid without a lock for the startup process.
947 : */
948 364 : rel = ShmemVariableCache->nextXid;
949 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
950 :
951 642 : if (!FullTransactionIdIsValid(cur_latest) ||
952 278 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
953 : {
954 352 : ShmemVariableCache->latestCompletedXid =
955 352 : FullXidRelativeTo(rel, latestXid);
956 : }
957 :
958 : Assert(FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
959 364 : }
960 :
961 : /*
962 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
963 : *
964 : * Remember up to where the startup process initialized the CLOG and subtrans
965 : * so we can ensure it's initialized gaplessly up to the point where necessary
966 : * while in recovery.
967 : */
968 : void
969 86 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
970 : {
971 : Assert(standbyState == STANDBY_INITIALIZED);
972 : Assert(TransactionIdIsNormal(initializedUptoXID));
973 :
974 : /*
975 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
976 : * to, so we can extend it from that point onwards in
977 : * RecordKnownAssignedTransactionIds, and when we get consistent in
978 : * ProcArrayApplyRecoveryInfo().
979 : */
980 86 : latestObservedXid = initializedUptoXID;
981 86 : TransactionIdRetreat(latestObservedXid);
982 86 : }
983 :
984 : /*
985 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
986 : *
987 : * Takes us through 3 states: Initialized, Pending and Ready.
988 : * Normal case is to go all the way to Ready straight away, though there
989 : * are atypical cases where we need to take it in steps.
990 : *
991 : * Use the data about running transactions on the primary to create the initial
992 : * state of KnownAssignedXids. We also use these records to regularly prune
993 : * KnownAssignedXids because we know it is possible that some transactions
994 : * with FATAL errors fail to write abort records, which could cause eventual
995 : * overflow.
996 : *
997 : * See comments for LogStandbySnapshot().
998 : */
999 : void
1000 192 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1001 : {
1002 : TransactionId *xids;
1003 : int nxids;
1004 : int i;
1005 :
1006 : Assert(standbyState >= STANDBY_INITIALIZED);
1007 : Assert(TransactionIdIsValid(running->nextXid));
1008 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1009 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1010 :
1011 : /*
1012 : * Remove stale transactions, if any.
1013 : */
1014 192 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1015 :
1016 : /*
1017 : * Remove stale locks, if any.
1018 : */
1019 192 : StandbyReleaseOldLocks(running->oldestRunningXid);
1020 :
1021 : /*
1022 : * If our snapshot is already valid, nothing else to do...
1023 : */
1024 192 : if (standbyState == STANDBY_SNAPSHOT_READY)
1025 106 : return;
1026 :
1027 : /*
1028 : * If our initial RunningTransactionsData had an overflowed snapshot then
1029 : * we knew we were missing some subxids from our snapshot. If we continue
1030 : * to see overflowed snapshots then we might never be able to start up, so
1031 : * we make another test to see if our snapshot is now valid. We know that
1032 : * the missing subxids are equal to or earlier than nextXid. After we
1033 : * initialise we continue to apply changes during recovery, so once the
1034 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1035 : * know that we no longer have missing information and can mark the
1036 : * snapshot as valid.
1037 : */
1038 86 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1039 : {
1040 : /*
1041 : * If the snapshot isn't overflowed or if its empty we can reset our
1042 : * pending state and use this snapshot instead.
1043 : */
1044 0 : if (!running->subxid_overflow || running->xcnt == 0)
1045 : {
1046 : /*
1047 : * If we have already collected known assigned xids, we need to
1048 : * throw them away before we apply the recovery snapshot.
1049 : */
1050 0 : KnownAssignedXidsReset();
1051 0 : standbyState = STANDBY_INITIALIZED;
1052 : }
1053 : else
1054 : {
1055 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1056 : running->oldestRunningXid))
1057 : {
1058 0 : standbyState = STANDBY_SNAPSHOT_READY;
1059 0 : elog(trace_recovery(DEBUG1),
1060 : "recovery snapshots are now enabled");
1061 : }
1062 : else
1063 0 : elog(trace_recovery(DEBUG1),
1064 : "recovery snapshot waiting for non-overflowed snapshot or "
1065 : "until oldest active xid on standby is at least %u (now %u)",
1066 : standbySnapshotPendingXmin,
1067 : running->oldestRunningXid);
1068 0 : return;
1069 : }
1070 : }
1071 :
1072 : Assert(standbyState == STANDBY_INITIALIZED);
1073 :
1074 : /*
1075 : * NB: this can be reached at least twice, so make sure new code can deal
1076 : * with that.
1077 : */
1078 :
1079 : /*
1080 : * Nobody else is running yet, but take locks anyhow
1081 : */
1082 86 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1083 :
1084 : /*
1085 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1086 : * sort them first.
1087 : *
1088 : * Some of the new xids are top-level xids and some are subtransactions.
1089 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1090 : * aren't overflowed then all xids will fit in snapshot and so we don't
1091 : * need subtrans. If we later overflow, an xid assignment record will add
1092 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1093 : * don't have enough information to correctly update subtrans anyway.
1094 : */
1095 :
1096 : /*
1097 : * Allocate a temporary array to avoid modifying the array passed as
1098 : * argument.
1099 : */
1100 86 : xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt));
1101 :
1102 : /*
1103 : * Add to the temp array any xids which have not already completed.
1104 : */
1105 86 : nxids = 0;
1106 90 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1107 : {
1108 4 : TransactionId xid = running->xids[i];
1109 :
1110 : /*
1111 : * The running-xacts snapshot can contain xids that were still visible
1112 : * in the procarray when the snapshot was taken, but were already
1113 : * WAL-logged as completed. They're not running anymore, so ignore
1114 : * them.
1115 : */
1116 4 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1117 0 : continue;
1118 :
1119 4 : xids[nxids++] = xid;
1120 : }
1121 :
1122 86 : if (nxids > 0)
1123 : {
1124 4 : if (procArray->numKnownAssignedXids != 0)
1125 : {
1126 0 : LWLockRelease(ProcArrayLock);
1127 0 : elog(ERROR, "KnownAssignedXids is not empty");
1128 : }
1129 :
1130 : /*
1131 : * Sort the array so that we can add them safely into
1132 : * KnownAssignedXids.
1133 : */
1134 4 : qsort(xids, nxids, sizeof(TransactionId), xidComparator);
1135 :
1136 : /*
1137 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1138 : * snapshot may include duplicated xids because of prepared
1139 : * transactions, so ignore them.
1140 : */
1141 8 : for (i = 0; i < nxids; i++)
1142 : {
1143 4 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1144 : {
1145 0 : elog(DEBUG1,
1146 : "found duplicated transaction %u for KnownAssignedXids insertion",
1147 : xids[i]);
1148 0 : continue;
1149 : }
1150 4 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1151 : }
1152 :
1153 4 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1154 : }
1155 :
1156 86 : pfree(xids);
1157 :
1158 : /*
1159 : * latestObservedXid is at least set to the point where SUBTRANS was
1160 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1161 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1162 : * subtrans from thereon, up to nextXid - 1.
1163 : *
1164 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1165 : * because we've just added xids to the known assigned xids machinery that
1166 : * haven't gone through RecordKnownAssignedTransactionId().
1167 : */
1168 : Assert(TransactionIdIsNormal(latestObservedXid));
1169 86 : TransactionIdAdvance(latestObservedXid);
1170 86 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1171 : {
1172 0 : ExtendSUBTRANS(latestObservedXid);
1173 0 : TransactionIdAdvance(latestObservedXid);
1174 : }
1175 86 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1176 :
1177 : /* ----------
1178 : * Now we've got the running xids we need to set the global values that
1179 : * are used to track snapshots as they evolve further.
1180 : *
1181 : * - latestCompletedXid which will be the xmax for snapshots
1182 : * - lastOverflowedXid which shows whether snapshots overflow
1183 : * - nextXid
1184 : *
1185 : * If the snapshot overflowed, then we still initialise with what we know,
1186 : * but the recovery snapshot isn't fully valid yet because we know there
1187 : * are some subxids missing. We don't know the specific subxids that are
1188 : * missing, so conservatively assume the last one is latestObservedXid.
1189 : * ----------
1190 : */
1191 86 : if (running->subxid_overflow)
1192 : {
1193 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1194 :
1195 0 : standbySnapshotPendingXmin = latestObservedXid;
1196 0 : procArray->lastOverflowedXid = latestObservedXid;
1197 : }
1198 : else
1199 : {
1200 86 : standbyState = STANDBY_SNAPSHOT_READY;
1201 :
1202 86 : standbySnapshotPendingXmin = InvalidTransactionId;
1203 : }
1204 :
1205 : /*
1206 : * If a transaction wrote a commit record in the gap between taking and
1207 : * logging the snapshot then latestCompletedXid may already be higher than
1208 : * the value from the snapshot, so check before we use the incoming value.
1209 : * It also might not yet be set at all.
1210 : */
1211 86 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1212 :
1213 86 : LWLockRelease(ProcArrayLock);
1214 :
1215 : /* ShmemVariableCache->nextXid must be beyond any observed xid. */
1216 86 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
1217 :
1218 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
1219 :
1220 86 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1221 86 : if (standbyState == STANDBY_SNAPSHOT_READY)
1222 86 : elog(trace_recovery(DEBUG1), "recovery snapshots are now enabled");
1223 : else
1224 0 : elog(trace_recovery(DEBUG1),
1225 : "recovery snapshot waiting for non-overflowed snapshot or "
1226 : "until oldest active xid on standby is at least %u (now %u)",
1227 : standbySnapshotPendingXmin,
1228 : running->oldestRunningXid);
1229 : }
1230 :
1231 : /*
1232 : * ProcArrayApplyXidAssignment
1233 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1234 : */
1235 : void
1236 40 : ProcArrayApplyXidAssignment(TransactionId topxid,
1237 : int nsubxids, TransactionId *subxids)
1238 : {
1239 : TransactionId max_xid;
1240 : int i;
1241 :
1242 : Assert(standbyState >= STANDBY_INITIALIZED);
1243 :
1244 40 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1245 :
1246 : /*
1247 : * Mark all the subtransactions as observed.
1248 : *
1249 : * NOTE: This will fail if the subxid contains too many previously
1250 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1251 : * as the code stands, because xid-assignment records should never contain
1252 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1253 : */
1254 40 : RecordKnownAssignedTransactionIds(max_xid);
1255 :
1256 : /*
1257 : * Notice that we update pg_subtrans with the top-level xid, rather than
1258 : * the parent xid. This is a difference between normal processing and
1259 : * recovery, yet is still correct in all cases. The reason is that
1260 : * subtransaction commit is not marked in clog until commit processing, so
1261 : * all aborted subtransactions have already been clearly marked in clog.
1262 : * As a result we are able to refer directly to the top-level
1263 : * transaction's state rather than skipping through all the intermediate
1264 : * states in the subtransaction tree. This should be the first time we
1265 : * have attempted to SubTransSetParent().
1266 : */
1267 2600 : for (i = 0; i < nsubxids; i++)
1268 2560 : SubTransSetParent(subxids[i], topxid);
1269 :
1270 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1271 40 : if (standbyState == STANDBY_INITIALIZED)
1272 0 : return;
1273 :
1274 : /*
1275 : * Uses same locking as transaction commit
1276 : */
1277 40 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1278 :
1279 : /*
1280 : * Remove subxids from known-assigned-xacts.
1281 : */
1282 40 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1283 :
1284 : /*
1285 : * Advance lastOverflowedXid to be at least the last of these subxids.
1286 : */
1287 40 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1288 40 : procArray->lastOverflowedXid = max_xid;
1289 :
1290 40 : LWLockRelease(ProcArrayLock);
1291 : }
1292 :
1293 : /*
1294 : * TransactionIdIsInProgress -- is given transaction running in some backend
1295 : *
1296 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1297 : * there are four possibilities for finding a running transaction:
1298 : *
1299 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1300 : * by looking at ProcGlobal->xids.
1301 : *
1302 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1303 : * We can find this out cheaply too.
1304 : *
1305 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1306 : * if the Xid is running on the primary.
1307 : *
1308 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1309 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1310 : * This is the slowest way, but sadly it has to be done always if the others
1311 : * failed, unless we see that the cached subxact sets are complete (none have
1312 : * overflowed).
1313 : *
1314 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1315 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1316 : * This buys back some concurrency (and we can't retrieve the main Xids from
1317 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1318 : */
1319 : bool
1320 11087690 : TransactionIdIsInProgress(TransactionId xid)
1321 : {
1322 : static TransactionId *xids = NULL;
1323 : static TransactionId *other_xids;
1324 : XidCacheStatus *other_subxidstates;
1325 11087690 : int nxids = 0;
1326 11087690 : ProcArrayStruct *arrayP = procArray;
1327 : TransactionId topxid;
1328 : TransactionId latestCompletedXid;
1329 : int mypgxactoff;
1330 : size_t numProcs;
1331 : int j;
1332 :
1333 : /*
1334 : * Don't bother checking a transaction older than RecentXmin; it could not
1335 : * possibly still be running. (Note: in particular, this guarantees that
1336 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1337 : * running.)
1338 : */
1339 11087690 : if (TransactionIdPrecedes(xid, RecentXmin))
1340 : {
1341 : xc_by_recent_xmin_inc();
1342 10007562 : return false;
1343 : }
1344 :
1345 : /*
1346 : * We may have just checked the status of this transaction, so if it is
1347 : * already known to be completed, we can fall out without any access to
1348 : * shared memory.
1349 : */
1350 1080128 : if (TransactionIdIsKnownCompleted(xid))
1351 : {
1352 : xc_by_known_xact_inc();
1353 933956 : return false;
1354 : }
1355 :
1356 : /*
1357 : * Also, we can handle our own transaction (and subtransactions) without
1358 : * any access to shared memory.
1359 : */
1360 146172 : if (TransactionIdIsCurrentTransactionId(xid))
1361 : {
1362 : xc_by_my_xact_inc();
1363 128536 : return true;
1364 : }
1365 :
1366 : /*
1367 : * If first time through, get workspace to remember main XIDs in. We
1368 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1369 : */
1370 17636 : if (xids == NULL)
1371 : {
1372 : /*
1373 : * In hot standby mode, reserve enough space to hold all xids in the
1374 : * known-assigned list. If we later finish recovery, we no longer need
1375 : * the bigger array, but we don't bother to shrink it.
1376 : */
1377 678 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1378 :
1379 678 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1380 678 : if (xids == NULL)
1381 0 : ereport(ERROR,
1382 : (errcode(ERRCODE_OUT_OF_MEMORY),
1383 : errmsg("out of memory")));
1384 : }
1385 :
1386 17636 : other_xids = ProcGlobal->xids;
1387 17636 : other_subxidstates = ProcGlobal->subxidStates;
1388 :
1389 17636 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1390 :
1391 : /*
1392 : * Now that we have the lock, we can check latestCompletedXid; if the
1393 : * target Xid is after that, it's surely still running.
1394 : */
1395 17636 : latestCompletedXid =
1396 17636 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
1397 17636 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1398 : {
1399 6488 : LWLockRelease(ProcArrayLock);
1400 : xc_by_latest_xid_inc();
1401 6488 : return true;
1402 : }
1403 :
1404 : /* No shortcuts, gotta grovel through the array */
1405 11148 : mypgxactoff = MyProc->pgxactoff;
1406 11148 : numProcs = arrayP->numProcs;
1407 78716 : for (size_t pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1408 : {
1409 : int pgprocno;
1410 : PGPROC *proc;
1411 : TransactionId pxid;
1412 : int pxids;
1413 :
1414 : /* Ignore ourselves --- dealt with it above */
1415 70074 : if (pgxactoff == mypgxactoff)
1416 9268 : continue;
1417 :
1418 : /* Fetch xid just once - see GetNewTransactionId */
1419 60806 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1420 :
1421 60806 : if (!TransactionIdIsValid(pxid))
1422 42394 : continue;
1423 :
1424 : /*
1425 : * Step 1: check the main Xid
1426 : */
1427 18412 : if (TransactionIdEquals(pxid, xid))
1428 : {
1429 1866 : LWLockRelease(ProcArrayLock);
1430 : xc_by_main_xid_inc();
1431 1866 : return true;
1432 : }
1433 :
1434 : /*
1435 : * We can ignore main Xids that are younger than the target Xid, since
1436 : * the target could not possibly be their child.
1437 : */
1438 16546 : if (TransactionIdPrecedes(xid, pxid))
1439 4594 : continue;
1440 :
1441 : /*
1442 : * Step 2: check the cached child-Xids arrays
1443 : */
1444 11952 : pxids = other_subxidstates[pgxactoff].count;
1445 11952 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1446 11952 : pgprocno = arrayP->pgprocnos[pgxactoff];
1447 11952 : proc = &allProcs[pgprocno];
1448 55038 : for (j = pxids - 1; j >= 0; j--)
1449 : {
1450 : /* Fetch xid just once - see GetNewTransactionId */
1451 43726 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1452 :
1453 43726 : if (TransactionIdEquals(cxid, xid))
1454 : {
1455 640 : LWLockRelease(ProcArrayLock);
1456 : xc_by_child_xid_inc();
1457 640 : return true;
1458 : }
1459 : }
1460 :
1461 : /*
1462 : * Save the main Xid for step 4. We only need to remember main Xids
1463 : * that have uncached children. (Note: there is no race condition
1464 : * here because the overflowed flag cannot be cleared, only set, while
1465 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1466 : * worry about.)
1467 : */
1468 11312 : if (other_subxidstates[pgxactoff].overflowed)
1469 310 : xids[nxids++] = pxid;
1470 : }
1471 :
1472 : /*
1473 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1474 : * in the list must be treated as running.
1475 : */
1476 8642 : if (RecoveryInProgress())
1477 : {
1478 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1479 : Assert(nxids == 0);
1480 :
1481 0 : if (KnownAssignedXidExists(xid))
1482 : {
1483 0 : LWLockRelease(ProcArrayLock);
1484 : xc_by_known_assigned_inc();
1485 0 : return true;
1486 : }
1487 :
1488 : /*
1489 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1490 : * too. Fetch all xids from KnownAssignedXids that are lower than
1491 : * xid, since if xid is a subtransaction its parent will always have a
1492 : * lower value. Note we will collect both main and subXIDs here, but
1493 : * there's no help for it.
1494 : */
1495 0 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1496 0 : nxids = KnownAssignedXidsGet(xids, xid);
1497 : }
1498 :
1499 8642 : LWLockRelease(ProcArrayLock);
1500 :
1501 : /*
1502 : * If none of the relevant caches overflowed, we know the Xid is not
1503 : * running without even looking at pg_subtrans.
1504 : */
1505 8642 : if (nxids == 0)
1506 : {
1507 : xc_no_overflow_inc();
1508 8332 : return false;
1509 : }
1510 :
1511 : /*
1512 : * Step 4: have to check pg_subtrans.
1513 : *
1514 : * At this point, we know it's either a subtransaction of one of the Xids
1515 : * in xids[], or it's not running. If it's an already-failed
1516 : * subtransaction, we want to say "not running" even though its parent may
1517 : * still be running. So first, check pg_xact to see if it's been aborted.
1518 : */
1519 : xc_slow_answer_inc();
1520 :
1521 310 : if (TransactionIdDidAbort(xid))
1522 0 : return false;
1523 :
1524 : /*
1525 : * It isn't aborted, so check whether the transaction tree it belongs to
1526 : * is still running (or, more precisely, whether it was running when we
1527 : * held ProcArrayLock).
1528 : */
1529 310 : topxid = SubTransGetTopmostTransaction(xid);
1530 : Assert(TransactionIdIsValid(topxid));
1531 310 : if (!TransactionIdEquals(topxid, xid))
1532 : {
1533 310 : for (int i = 0; i < nxids; i++)
1534 : {
1535 310 : if (TransactionIdEquals(xids[i], topxid))
1536 310 : return true;
1537 : }
1538 : }
1539 :
1540 0 : return false;
1541 : }
1542 :
1543 : /*
1544 : * TransactionIdIsActive -- is xid the top-level XID of an active backend?
1545 : *
1546 : * This differs from TransactionIdIsInProgress in that it ignores prepared
1547 : * transactions, as well as transactions running on the primary if we're in
1548 : * hot standby. Also, we ignore subtransactions since that's not needed
1549 : * for current uses.
1550 : */
1551 : bool
1552 0 : TransactionIdIsActive(TransactionId xid)
1553 : {
1554 0 : bool result = false;
1555 0 : ProcArrayStruct *arrayP = procArray;
1556 0 : TransactionId *other_xids = ProcGlobal->xids;
1557 : int i;
1558 :
1559 : /*
1560 : * Don't bother checking a transaction older than RecentXmin; it could not
1561 : * possibly still be running.
1562 : */
1563 0 : if (TransactionIdPrecedes(xid, RecentXmin))
1564 0 : return false;
1565 :
1566 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1567 :
1568 0 : for (i = 0; i < arrayP->numProcs; i++)
1569 : {
1570 0 : int pgprocno = arrayP->pgprocnos[i];
1571 0 : PGPROC *proc = &allProcs[pgprocno];
1572 : TransactionId pxid;
1573 :
1574 : /* Fetch xid just once - see GetNewTransactionId */
1575 0 : pxid = UINT32_ACCESS_ONCE(other_xids[i]);
1576 :
1577 0 : if (!TransactionIdIsValid(pxid))
1578 0 : continue;
1579 :
1580 0 : if (proc->pid == 0)
1581 0 : continue; /* ignore prepared transactions */
1582 :
1583 0 : if (TransactionIdEquals(pxid, xid))
1584 : {
1585 0 : result = true;
1586 0 : break;
1587 : }
1588 : }
1589 :
1590 0 : LWLockRelease(ProcArrayLock);
1591 :
1592 0 : return result;
1593 : }
1594 :
1595 :
1596 : /*
1597 : * Determine XID horizons.
1598 : *
1599 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1600 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1601 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1602 : * GlobalVisState).
1603 : *
1604 : * See the definition of ComputedXidHorizonsResult for the various computed
1605 : * horizons.
1606 : *
1607 : * For VACUUM separate horizons (used to decide which deleted tuples must
1608 : * be preserved), for shared and non-shared tables are computed. For shared
1609 : * relations backends in all databases must be considered, but for non-shared
1610 : * relations that's not required, since only backends in my own database could
1611 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1612 : * VACUUMs because (a) they must be working on other tables, and (b) they
1613 : * don't need to do snapshot-based lookups.
1614 : *
1615 : * This also computes a horizon used to truncate pg_subtrans. For that
1616 : * backends in all databases have to be considered, and concurrently running
1617 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1618 : * accesses.
1619 : *
1620 : * Note: we include all currently running xids in the set of considered xids.
1621 : * This ensures that if a just-started xact has not yet set its snapshot,
1622 : * when it does set the snapshot it cannot set xmin less than what we compute.
1623 : * See notes in src/backend/access/transam/README.
1624 : *
1625 : * Note: despite the above, it's possible for the calculated values to move
1626 : * backwards on repeated calls. The calculated values are conservative, so
1627 : * that anything older is definitely not considered as running by anyone
1628 : * anymore, but the exact values calculated depend on a number of things. For
1629 : * example, if there are no transactions running in the current database, the
1630 : * horizon for normal tables will be latestCompletedXid. If a transaction
1631 : * begins after that, its xmin will include in-progress transactions in other
1632 : * databases that started earlier, so another call will return a lower value.
1633 : * Nonetheless it is safe to vacuum a table in the current database with the
1634 : * first result. There are also replication-related effects: a walsender
1635 : * process can set its xmin based on transactions that are no longer running
1636 : * on the primary but are still being replayed on the standby, thus possibly
1637 : * making the values go backwards. In this case there is a possibility that
1638 : * we lose data that the standby would like to have, but unless the standby
1639 : * uses a replication slot to make its xmin persistent there is little we can
1640 : * do about that --- data is only protected if the walsender runs continuously
1641 : * while queries are executed on the standby. (The Hot Standby code deals
1642 : * with such cases by failing standby queries that needed to access
1643 : * already-removed data, so there's no integrity bug.) The computed values
1644 : * are also adjusted with vacuum_defer_cleanup_age, so increasing that setting
1645 : * on the fly is another easy way to make horizons move backwards, with no
1646 : * consequences for data integrity.
1647 : *
1648 : * Note: the approximate horizons (see definition of GlobalVisState) are
1649 : * updated by the computations done here. That's currently required for
1650 : * correctness and a small optimization. Without doing so it's possible that
1651 : * heap vacuum's call to heap_page_prune() uses a more conservative horizon
1652 : * than later when deciding which tuples can be removed - which the code
1653 : * doesn't expect (breaking HOT).
1654 : */
1655 : static void
1656 219158 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1657 : {
1658 219158 : ProcArrayStruct *arrayP = procArray;
1659 : TransactionId kaxmin;
1660 219158 : bool in_recovery = RecoveryInProgress();
1661 219158 : TransactionId *other_xids = ProcGlobal->xids;
1662 :
1663 : /* inferred after ProcArrayLock is released */
1664 219158 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1665 :
1666 219158 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1667 :
1668 219158 : h->latest_completed = ShmemVariableCache->latestCompletedXid;
1669 :
1670 : /*
1671 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1672 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1673 : * and so protects us against overestimating the result due to future
1674 : * additions.
1675 : */
1676 : {
1677 : TransactionId initial;
1678 :
1679 219158 : initial = XidFromFullTransactionId(h->latest_completed);
1680 : Assert(TransactionIdIsValid(initial));
1681 219158 : TransactionIdAdvance(initial);
1682 :
1683 219158 : h->oldest_considered_running = initial;
1684 219158 : h->shared_oldest_nonremovable = initial;
1685 219158 : h->data_oldest_nonremovable = initial;
1686 :
1687 : /*
1688 : * Only modifications made by this backend affect the horizon for
1689 : * temporary relations. Instead of a check in each iteration of the
1690 : * loop over all PGPROCs it is cheaper to just initialize to the
1691 : * current top-level xid any.
1692 : *
1693 : * Without an assigned xid we could use a horizon as aggressive as
1694 : * ReadNewTransactionid(), but we can get away with the much cheaper
1695 : * latestCompletedXid + 1: If this backend has no xid there, by
1696 : * definition, can't be any newer changes in the temp table than
1697 : * latestCompletedXid.
1698 : */
1699 219158 : if (TransactionIdIsValid(MyProc->xid))
1700 51218 : h->temp_oldest_nonremovable = MyProc->xid;
1701 : else
1702 167940 : h->temp_oldest_nonremovable = initial;
1703 : }
1704 :
1705 : /*
1706 : * Fetch slot horizons while ProcArrayLock is held - the
1707 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1708 : * the lock.
1709 : */
1710 219158 : h->slot_xmin = procArray->replication_slot_xmin;
1711 219158 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1712 :
1713 756558 : for (int index = 0; index < arrayP->numProcs; index++)
1714 : {
1715 537400 : int pgprocno = arrayP->pgprocnos[index];
1716 537400 : PGPROC *proc = &allProcs[pgprocno];
1717 537400 : int8 statusFlags = ProcGlobal->statusFlags[index];
1718 : TransactionId xid;
1719 : TransactionId xmin;
1720 :
1721 : /* Fetch xid just once - see GetNewTransactionId */
1722 537400 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1723 537400 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1724 :
1725 : /*
1726 : * Consider both the transaction's Xmin, and its Xid.
1727 : *
1728 : * We must check both because a transaction might have an Xmin but not
1729 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1730 : * some not-yet-set Xmin.
1731 : */
1732 537400 : xmin = TransactionIdOlder(xmin, xid);
1733 :
1734 : /* if neither is set, this proc doesn't influence the horizon */
1735 537400 : if (!TransactionIdIsValid(xmin))
1736 237582 : continue;
1737 :
1738 : /*
1739 : * Don't ignore any procs when determining which transactions might be
1740 : * considered running. While slots should ensure logical decoding
1741 : * backends are protected even without this check, it can't hurt to
1742 : * include them here as well..
1743 : */
1744 299818 : h->oldest_considered_running =
1745 299818 : TransactionIdOlder(h->oldest_considered_running, xmin);
1746 :
1747 : /*
1748 : * Skip over backends either vacuuming (which is ok with rows being
1749 : * removed, as long as pg_subtrans is not truncated) or doing logical
1750 : * decoding (which manages xmin separately, check below).
1751 : */
1752 299818 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1753 50060 : continue;
1754 :
1755 : /* shared tables need to take backends in all database into account */
1756 249758 : h->shared_oldest_nonremovable =
1757 249758 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1758 :
1759 : /*
1760 : * Normally queries in other databases are ignored for anything but
1761 : * the shared horizon. But in recovery we cannot compute an accurate
1762 : * per-database horizon as all xids are managed via the
1763 : * KnownAssignedXids machinery.
1764 : *
1765 : * Be careful to compute a pessimistic value when MyDatabaseId is not
1766 : * set. If this is a backend in the process of starting up, we may not
1767 : * use a "too aggressive" horizon (otherwise we could end up using it
1768 : * to prune still needed data away). If the current backend never
1769 : * connects to a database that is harmless, because
1770 : * data_oldest_nonremovable will never be utilized.
1771 : */
1772 249758 : if (in_recovery ||
1773 249720 : MyDatabaseId == InvalidOid || proc->databaseId == MyDatabaseId ||
1774 204 : proc->databaseId == 0) /* always include WalSender */
1775 : {
1776 249708 : h->data_oldest_nonremovable =
1777 249708 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1778 : }
1779 : }
1780 :
1781 : /*
1782 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1783 : * after lock is released.
1784 : */
1785 219158 : if (in_recovery)
1786 88 : kaxmin = KnownAssignedXidsGetOldestXmin();
1787 :
1788 : /*
1789 : * No other information from shared state is needed, release the lock
1790 : * immediately. The rest of the computations can be done without a lock.
1791 : */
1792 219158 : LWLockRelease(ProcArrayLock);
1793 :
1794 219158 : if (in_recovery)
1795 : {
1796 88 : h->oldest_considered_running =
1797 88 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1798 88 : h->shared_oldest_nonremovable =
1799 88 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1800 88 : h->data_oldest_nonremovable =
1801 88 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1802 : /* temp relations cannot be accessed in recovery */
1803 : }
1804 : else
1805 : {
1806 : /*
1807 : * Compute the cutoff XID by subtracting vacuum_defer_cleanup_age.
1808 : *
1809 : * vacuum_defer_cleanup_age provides some additional "slop" for the
1810 : * benefit of hot standby queries on standby servers. This is quick
1811 : * and dirty, and perhaps not all that useful unless the primary has a
1812 : * predictable transaction rate, but it offers some protection when
1813 : * there's no walsender connection. Note that we are assuming
1814 : * vacuum_defer_cleanup_age isn't large enough to cause wraparound ---
1815 : * so guc.c should limit it to no more than the xidStopLimit threshold
1816 : * in varsup.c. Also note that we intentionally don't apply
1817 : * vacuum_defer_cleanup_age on standby servers.
1818 : */
1819 219070 : h->oldest_considered_running =
1820 219070 : TransactionIdRetreatedBy(h->oldest_considered_running,
1821 : vacuum_defer_cleanup_age);
1822 219070 : h->shared_oldest_nonremovable =
1823 219070 : TransactionIdRetreatedBy(h->shared_oldest_nonremovable,
1824 : vacuum_defer_cleanup_age);
1825 219070 : h->data_oldest_nonremovable =
1826 219070 : TransactionIdRetreatedBy(h->data_oldest_nonremovable,
1827 : vacuum_defer_cleanup_age);
1828 : /* defer doesn't apply to temp relations */
1829 : }
1830 :
1831 : /*
1832 : * Check whether there are replication slots requiring an older xmin.
1833 : */
1834 219158 : h->shared_oldest_nonremovable =
1835 219158 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1836 219158 : h->data_oldest_nonremovable =
1837 219158 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1838 :
1839 : /*
1840 : * The only difference between catalog / data horizons is that the slot's
1841 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1842 : * for logical decoding). Initialize with data horizon, and then back up
1843 : * further if necessary. Have to back up the shared horizon as well, since
1844 : * that also can contain catalogs.
1845 : */
1846 219158 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1847 219158 : h->shared_oldest_nonremovable =
1848 219158 : TransactionIdOlder(h->shared_oldest_nonremovable,
1849 : h->slot_catalog_xmin);
1850 219158 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1851 219158 : h->catalog_oldest_nonremovable =
1852 219158 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1853 : h->slot_catalog_xmin);
1854 :
1855 : /*
1856 : * It's possible that slots / vacuum_defer_cleanup_age backed up the
1857 : * horizons further than oldest_considered_running. Fix.
1858 : */
1859 219158 : h->oldest_considered_running =
1860 219158 : TransactionIdOlder(h->oldest_considered_running,
1861 : h->shared_oldest_nonremovable);
1862 219158 : h->oldest_considered_running =
1863 219158 : TransactionIdOlder(h->oldest_considered_running,
1864 : h->catalog_oldest_nonremovable);
1865 219158 : h->oldest_considered_running =
1866 219158 : TransactionIdOlder(h->oldest_considered_running,
1867 : h->data_oldest_nonremovable);
1868 :
1869 : /*
1870 : * shared horizons have to be at least as old as the oldest visible in
1871 : * current db
1872 : */
1873 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1874 : h->data_oldest_nonremovable));
1875 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1876 : h->catalog_oldest_nonremovable));
1877 :
1878 : /*
1879 : * Horizons need to ensure that pg_subtrans access is still possible for
1880 : * the relevant backends.
1881 : */
1882 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1883 : h->shared_oldest_nonremovable));
1884 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1885 : h->catalog_oldest_nonremovable));
1886 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1887 : h->data_oldest_nonremovable));
1888 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1889 : h->temp_oldest_nonremovable));
1890 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1891 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1892 : h->slot_xmin));
1893 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1894 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1895 : h->slot_catalog_xmin));
1896 :
1897 : /* update approximate horizons with the computed horizons */
1898 219158 : GlobalVisUpdateApply(h);
1899 219158 : }
1900 :
1901 : /*
1902 : * Return the oldest XID for which deleted tuples must be preserved in the
1903 : * passed table.
1904 : *
1905 : * If rel is not NULL the horizon may be considerably more recent than
1906 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1907 : * that is correct (but not optimal) for all relations will be returned.
1908 : *
1909 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1910 : * the passed in table.
1911 : */
1912 : TransactionId
1913 108858 : GetOldestNonRemovableTransactionId(Relation rel)
1914 : {
1915 : ComputeXidHorizonsResult horizons;
1916 :
1917 108858 : ComputeXidHorizons(&horizons);
1918 :
1919 : /* select horizon appropriate for relation */
1920 108858 : if (rel == NULL || rel->rd_rel->relisshared)
1921 18600 : return horizons.shared_oldest_nonremovable;
1922 90258 : else if (RelationIsAccessibleInLogicalDecoding(rel))
1923 128 : return horizons.catalog_oldest_nonremovable;
1924 90130 : else if (RELATION_IS_LOCAL(rel))
1925 14612 : return horizons.temp_oldest_nonremovable;
1926 : else
1927 75518 : return horizons.data_oldest_nonremovable;
1928 : }
1929 :
1930 : /*
1931 : * Return the oldest transaction id any currently running backend might still
1932 : * consider running. This should not be used for visibility / pruning
1933 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1934 : * decisions like up to where pg_subtrans can be truncated.
1935 : */
1936 : TransactionId
1937 3344 : GetOldestTransactionIdConsideredRunning(void)
1938 : {
1939 : ComputeXidHorizonsResult horizons;
1940 :
1941 3344 : ComputeXidHorizons(&horizons);
1942 :
1943 3344 : return horizons.oldest_considered_running;
1944 : }
1945 :
1946 : /*
1947 : * Return the visibility horizons for a hot standby feedback message.
1948 : */
1949 : void
1950 18 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1951 : {
1952 : ComputeXidHorizonsResult horizons;
1953 :
1954 18 : ComputeXidHorizons(&horizons);
1955 :
1956 : /*
1957 : * Don't want to use shared_oldest_nonremovable here, as that contains the
1958 : * effect of replication slot's catalog_xmin. We want to send a separate
1959 : * feedback for the catalog horizon, so the primary can remove data table
1960 : * contents more aggressively.
1961 : */
1962 18 : *xmin = horizons.shared_oldest_nonremovable_raw;
1963 18 : *catalog_xmin = horizons.slot_catalog_xmin;
1964 18 : }
1965 :
1966 : /*
1967 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
1968 : *
1969 : * We have to export this for use by snapmgr.c.
1970 : */
1971 : int
1972 24458 : GetMaxSnapshotXidCount(void)
1973 : {
1974 24458 : return procArray->maxProcs;
1975 : }
1976 :
1977 : /*
1978 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
1979 : *
1980 : * We have to export this for use by snapmgr.c.
1981 : */
1982 : int
1983 24346 : GetMaxSnapshotSubxidCount(void)
1984 : {
1985 24346 : return TOTAL_MAX_CACHED_SUBXIDS;
1986 : }
1987 :
1988 : /*
1989 : * Initialize old_snapshot_threshold specific parts of a newly build snapshot.
1990 : */
1991 : static void
1992 2280738 : GetSnapshotDataInitOldSnapshot(Snapshot snapshot)
1993 : {
1994 2280738 : if (!OldSnapshotThresholdActive())
1995 : {
1996 : /*
1997 : * If not using "snapshot too old" feature, fill related fields with
1998 : * dummy values that don't require any locking.
1999 : */
2000 2256836 : snapshot->lsn = InvalidXLogRecPtr;
2001 2256836 : snapshot->whenTaken = 0;
2002 : }
2003 : else
2004 : {
2005 : /*
2006 : * Capture the current time and WAL stream location in case this
2007 : * snapshot becomes old enough to need to fall back on the special
2008 : * "old snapshot" logic.
2009 : */
2010 23902 : snapshot->lsn = GetXLogInsertRecPtr();
2011 23902 : snapshot->whenTaken = GetSnapshotCurrentTimestamp();
2012 23902 : MaintainOldSnapshotTimeMapping(snapshot->whenTaken, snapshot->xmin);
2013 : }
2014 2280738 : }
2015 :
2016 : /*
2017 : * Helper function for GetSnapshotData() that checks if the bulk of the
2018 : * visibility information in the snapshot is still valid. If so, it updates
2019 : * the fields that need to change and returns true. Otherwise it returns
2020 : * false.
2021 : *
2022 : * This very likely can be evolved to not need ProcArrayLock held (at very
2023 : * least in the case we already hold a snapshot), but that's for another day.
2024 : */
2025 : static bool
2026 2280738 : GetSnapshotDataReuse(Snapshot snapshot)
2027 : {
2028 : uint64 curXactCompletionCount;
2029 :
2030 : Assert(LWLockHeldByMe(ProcArrayLock));
2031 :
2032 2280738 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2033 24322 : return false;
2034 :
2035 2256416 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2036 2256416 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2037 651146 : return false;
2038 :
2039 : /*
2040 : * If the current xactCompletionCount is still the same as it was at the
2041 : * time the snapshot was built, we can be sure that rebuilding the
2042 : * contents of the snapshot the hard way would result in the same snapshot
2043 : * contents:
2044 : *
2045 : * As explained in transam/README, the set of xids considered running by
2046 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2047 : * contents only depend on transactions with xids and xactCompletionCount
2048 : * is incremented whenever a transaction with an xid finishes (while
2049 : * holding ProcArrayLock) exclusively). Thus the xactCompletionCount check
2050 : * ensures we would detect if the snapshot would have changed.
2051 : *
2052 : * As the snapshot contents are the same as it was before, it is is safe
2053 : * to re-enter the snapshot's xmin into the PGPROC array. None of the rows
2054 : * visible under the snapshot could already have been removed (that'd
2055 : * require the set of running transactions to change) and it fulfills the
2056 : * requirement that concurrent GetSnapshotData() calls yield the same
2057 : * xmin.
2058 : */
2059 1605270 : if (!TransactionIdIsValid(MyProc->xmin))
2060 673794 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2061 :
2062 1605270 : RecentXmin = snapshot->xmin;
2063 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2064 :
2065 1605270 : snapshot->curcid = GetCurrentCommandId(false);
2066 1605270 : snapshot->active_count = 0;
2067 1605270 : snapshot->regd_count = 0;
2068 1605270 : snapshot->copied = false;
2069 :
2070 1605270 : GetSnapshotDataInitOldSnapshot(snapshot);
2071 :
2072 1605270 : return true;
2073 : }
2074 :
2075 : /*
2076 : * GetSnapshotData -- returns information about running transactions.
2077 : *
2078 : * The returned snapshot includes xmin (lowest still-running xact ID),
2079 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2080 : * in the range xmin <= xid < xmax. It is used as follows:
2081 : * All xact IDs < xmin are considered finished.
2082 : * All xact IDs >= xmax are considered still running.
2083 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2084 : * it is considered running or not.
2085 : * This ensures that the set of transactions seen as "running" by the
2086 : * current xact will not change after it takes the snapshot.
2087 : *
2088 : * All running top-level XIDs are included in the snapshot, except for lazy
2089 : * VACUUM processes. We also try to include running subtransaction XIDs,
2090 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2091 : * information may not be available. If we find any overflowed subxid arrays,
2092 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2093 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()
2094 : * in heapam_visibility.c).
2095 : *
2096 : * We also update the following backend-global variables:
2097 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2098 : * current transaction (this is the same as MyProc->xmin).
2099 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2100 : * older than this are known not running any more.
2101 : *
2102 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2103 : * for the benefit of theGlobalVisTest* family of functions.
2104 : *
2105 : * Note: this function should probably not be called with an argument that's
2106 : * not statically allocated (see xip allocation below).
2107 : */
2108 : Snapshot
2109 2280738 : GetSnapshotData(Snapshot snapshot)
2110 : {
2111 2280738 : ProcArrayStruct *arrayP = procArray;
2112 2280738 : TransactionId *other_xids = ProcGlobal->xids;
2113 : TransactionId xmin;
2114 : TransactionId xmax;
2115 2280738 : size_t count = 0;
2116 2280738 : int subcount = 0;
2117 2280738 : bool suboverflowed = false;
2118 : FullTransactionId latest_completed;
2119 : TransactionId oldestxid;
2120 : int mypgxactoff;
2121 : TransactionId myxid;
2122 : uint64 curXactCompletionCount;
2123 :
2124 2280738 : TransactionId replication_slot_xmin = InvalidTransactionId;
2125 2280738 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2126 :
2127 : Assert(snapshot != NULL);
2128 :
2129 : /*
2130 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2131 : * be sufficient. But it seems better to do the malloc while not holding
2132 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2133 : * more subxip storage than is probably needed.
2134 : *
2135 : * This does open a possibility for avoiding repeated malloc/free: since
2136 : * maxProcs does not change at runtime, we can simply reuse the previous
2137 : * xip arrays if any. (This relies on the fact that all callers pass
2138 : * static SnapshotData structs.)
2139 : */
2140 2280738 : if (snapshot->xip == NULL)
2141 : {
2142 : /*
2143 : * First call for this snapshot. Snapshot is same size whether or not
2144 : * we are in recovery, see later comments.
2145 : */
2146 24322 : snapshot->xip = (TransactionId *)
2147 24322 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2148 24322 : if (snapshot->xip == NULL)
2149 0 : ereport(ERROR,
2150 : (errcode(ERRCODE_OUT_OF_MEMORY),
2151 : errmsg("out of memory")));
2152 : Assert(snapshot->subxip == NULL);
2153 24322 : snapshot->subxip = (TransactionId *)
2154 24322 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2155 24322 : if (snapshot->subxip == NULL)
2156 0 : ereport(ERROR,
2157 : (errcode(ERRCODE_OUT_OF_MEMORY),
2158 : errmsg("out of memory")));
2159 : }
2160 :
2161 : /*
2162 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2163 : * going to set MyProc->xmin.
2164 : */
2165 2280738 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2166 :
2167 2280738 : if (GetSnapshotDataReuse(snapshot))
2168 : {
2169 1605270 : LWLockRelease(ProcArrayLock);
2170 1605270 : return snapshot;
2171 : }
2172 :
2173 675468 : latest_completed = ShmemVariableCache->latestCompletedXid;
2174 675468 : mypgxactoff = MyProc->pgxactoff;
2175 675468 : myxid = other_xids[mypgxactoff];
2176 : Assert(myxid == MyProc->xid);
2177 :
2178 675468 : oldestxid = ShmemVariableCache->oldestXid;
2179 675468 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2180 :
2181 : /* xmax is always latestCompletedXid + 1 */
2182 675468 : xmax = XidFromFullTransactionId(latest_completed);
2183 675468 : TransactionIdAdvance(xmax);
2184 : Assert(TransactionIdIsNormal(xmax));
2185 :
2186 : /* initialize xmin calculation with xmax */
2187 675468 : xmin = xmax;
2188 :
2189 : /* take own xid into account, saves a check inside the loop */
2190 675468 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2191 13610 : xmin = myxid;
2192 :
2193 675468 : snapshot->takenDuringRecovery = RecoveryInProgress();
2194 :
2195 675468 : if (!snapshot->takenDuringRecovery)
2196 : {
2197 674938 : size_t numProcs = arrayP->numProcs;
2198 674938 : TransactionId *xip = snapshot->xip;
2199 674938 : int *pgprocnos = arrayP->pgprocnos;
2200 674938 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2201 674938 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2202 :
2203 : /*
2204 : * First collect set of pgxactoff/xids that need to be included in the
2205 : * snapshot.
2206 : */
2207 2773352 : for (size_t pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2208 : {
2209 : /* Fetch xid just once - see GetNewTransactionId */
2210 2098414 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2211 : uint8 statusFlags;
2212 :
2213 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2214 :
2215 : /*
2216 : * If the transaction has no XID assigned, we can skip it; it
2217 : * won't have sub-XIDs either.
2218 : */
2219 2098414 : if (likely(xid == InvalidTransactionId))
2220 1788984 : continue;
2221 :
2222 : /*
2223 : * We don't include our own XIDs (if any) in the snapshot. It
2224 : * needs to be includeded in the xmin computation, but we did so
2225 : * outside the loop.
2226 : */
2227 309430 : if (pgxactoff == mypgxactoff)
2228 61072 : continue;
2229 :
2230 : /*
2231 : * The only way we are able to get here with a non-normal xid
2232 : * is during bootstrap - with this backend using
2233 : * BootstrapTransactionId. But the above test should filter
2234 : * that out.
2235 : */
2236 : Assert(TransactionIdIsNormal(xid));
2237 :
2238 : /*
2239 : * If the XID is >= xmax, we can skip it; such transactions will
2240 : * be treated as running anyway (and any sub-XIDs will also be >=
2241 : * xmax).
2242 : */
2243 248358 : if (!NormalTransactionIdPrecedes(xid, xmax))
2244 53814 : continue;
2245 :
2246 : /*
2247 : * Skip over backends doing logical decoding which manages xmin
2248 : * separately (check below) and ones running LAZY VACUUM.
2249 : */
2250 194544 : statusFlags = allStatusFlags[pgxactoff];
2251 194544 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2252 4 : continue;
2253 :
2254 194540 : if (NormalTransactionIdPrecedes(xid, xmin))
2255 137552 : xmin = xid;
2256 :
2257 : /* Add XID to snapshot. */
2258 194540 : xip[count++] = xid;
2259 :
2260 : /*
2261 : * Save subtransaction XIDs if possible (if we've already
2262 : * overflowed, there's no point). Note that the subxact XIDs must
2263 : * be later than their parent, so no need to check them against
2264 : * xmin. We could filter against xmax, but it seems better not to
2265 : * do that much work while holding the ProcArrayLock.
2266 : *
2267 : * The other backend can add more subxids concurrently, but cannot
2268 : * remove any. Hence it's important to fetch nxids just once.
2269 : * Should be safe to use memcpy, though. (We needn't worry about
2270 : * missing any xids added concurrently, because they must postdate
2271 : * xmax.)
2272 : *
2273 : * Again, our own XIDs are not included in the snapshot.
2274 : */
2275 194540 : if (!suboverflowed)
2276 : {
2277 :
2278 194540 : if (subxidStates[pgxactoff].overflowed)
2279 32 : suboverflowed = true;
2280 : else
2281 : {
2282 194508 : int nsubxids = subxidStates[pgxactoff].count;
2283 :
2284 194508 : if (nsubxids > 0)
2285 : {
2286 2048 : int pgprocno = pgprocnos[pgxactoff];
2287 2048 : PGPROC *proc = &allProcs[pgprocno];
2288 :
2289 2048 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2290 :
2291 4096 : memcpy(snapshot->subxip + subcount,
2292 2048 : (void *) proc->subxids.xids,
2293 : nsubxids * sizeof(TransactionId));
2294 2048 : subcount += nsubxids;
2295 : }
2296 : }
2297 : }
2298 : }
2299 : }
2300 : else
2301 : {
2302 : /*
2303 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2304 : *
2305 : * We store all xids directly into subxip[]. Here's why:
2306 : *
2307 : * In recovery we don't know which xids are top-level and which are
2308 : * subxacts, a design choice that greatly simplifies xid processing.
2309 : *
2310 : * It seems like we would want to try to put xids into xip[] only, but
2311 : * that is fairly small. We would either need to make that bigger or
2312 : * to increase the rate at which we WAL-log xid assignment; neither is
2313 : * an appealing choice.
2314 : *
2315 : * We could try to store xids into xip[] first and then into subxip[]
2316 : * if there are too many xids. That only works if the snapshot doesn't
2317 : * overflow because we do not search subxip[] in that case. A simpler
2318 : * way is to just store all xids in the subxact array because this is
2319 : * by far the bigger array. We just leave the xip array empty.
2320 : *
2321 : * Either way we need to change the way XidInMVCCSnapshot() works
2322 : * depending upon when the snapshot was taken, or change normal
2323 : * snapshot processing so it matches.
2324 : *
2325 : * Note: It is possible for recovery to end before we finish taking
2326 : * the snapshot, and for newly assigned transaction ids to be added to
2327 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2328 : * those newly added transaction ids would be filtered away, so we
2329 : * need not be concerned about them.
2330 : */
2331 530 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2332 : xmax);
2333 :
2334 530 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2335 8 : suboverflowed = true;
2336 : }
2337 :
2338 :
2339 : /*
2340 : * Fetch into local variable while ProcArrayLock is held - the
2341 : * LWLockRelease below is a barrier, ensuring this happens inside the
2342 : * lock.
2343 : */
2344 675468 : replication_slot_xmin = procArray->replication_slot_xmin;
2345 675468 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2346 :
2347 675468 : if (!TransactionIdIsValid(MyProc->xmin))
2348 341726 : MyProc->xmin = TransactionXmin = xmin;
2349 :
2350 675468 : LWLockRelease(ProcArrayLock);
2351 :
2352 : /* maintain state for GlobalVis* */
2353 : {
2354 : TransactionId def_vis_xid;
2355 : TransactionId def_vis_xid_data;
2356 : FullTransactionId def_vis_fxid;
2357 : FullTransactionId def_vis_fxid_data;
2358 : FullTransactionId oldestfxid;
2359 :
2360 : /*
2361 : * Converting oldestXid is only safe when xid horizon cannot advance,
2362 : * i.e. holding locks. While we don't hold the lock anymore, all the
2363 : * necessary data has been gathered with lock held.
2364 : */
2365 675468 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2366 :
2367 : /* apply vacuum_defer_cleanup_age */
2368 : def_vis_xid_data =
2369 675468 : TransactionIdRetreatedBy(xmin, vacuum_defer_cleanup_age);
2370 :
2371 : /* Check whether there's a replication slot requiring an older xmin. */
2372 : def_vis_xid_data =
2373 675468 : TransactionIdOlder(def_vis_xid_data, replication_slot_xmin);
2374 :
2375 : /*
2376 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2377 : * if older than this xid.
2378 : */
2379 675468 : def_vis_xid = def_vis_xid_data;
2380 :
2381 : /*
2382 : * Check whether there's a replication slot requiring an older catalog
2383 : * xmin.
2384 : */
2385 : def_vis_xid =
2386 675468 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2387 :
2388 675468 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2389 675468 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2390 :
2391 : /*
2392 : * Check if we can increase upper bound. As a previous
2393 : * GlobalVisUpdate() might have computed more aggressive values, don't
2394 : * overwrite them if so.
2395 : */
2396 : GlobalVisSharedRels.definitely_needed =
2397 675468 : FullTransactionIdNewer(def_vis_fxid,
2398 : GlobalVisSharedRels.definitely_needed);
2399 : GlobalVisCatalogRels.definitely_needed =
2400 675468 : FullTransactionIdNewer(def_vis_fxid,
2401 : GlobalVisCatalogRels.definitely_needed);
2402 : GlobalVisDataRels.definitely_needed =
2403 675468 : FullTransactionIdNewer(def_vis_fxid_data,
2404 : GlobalVisDataRels.definitely_needed);
2405 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2406 675468 : if (TransactionIdIsNormal(myxid))
2407 : GlobalVisTempRels.definitely_needed =
2408 60280 : FullXidRelativeTo(latest_completed, myxid);
2409 : else
2410 : {
2411 615188 : GlobalVisTempRels.definitely_needed = latest_completed;
2412 615188 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2413 : }
2414 :
2415 : /*
2416 : * Check if we know that we can initialize or increase the lower
2417 : * bound. Currently the only cheap way to do so is to use
2418 : * ShmemVariableCache->oldestXid as input.
2419 : *
2420 : * We should definitely be able to do better. We could e.g. put a
2421 : * global lower bound value into ShmemVariableCache.
2422 : */
2423 : GlobalVisSharedRels.maybe_needed =
2424 675468 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2425 : oldestfxid);
2426 : GlobalVisCatalogRels.maybe_needed =
2427 675468 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2428 : oldestfxid);
2429 : GlobalVisDataRels.maybe_needed =
2430 675468 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2431 : oldestfxid);
2432 : /* accurate value known */
2433 675468 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2434 : }
2435 :
2436 675468 : RecentXmin = xmin;
2437 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2438 :
2439 675468 : snapshot->xmin = xmin;
2440 675468 : snapshot->xmax = xmax;
2441 675468 : snapshot->xcnt = count;
2442 675468 : snapshot->subxcnt = subcount;
2443 675468 : snapshot->suboverflowed = suboverflowed;
2444 675468 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2445 :
2446 675468 : snapshot->curcid = GetCurrentCommandId(false);
2447 :
2448 : /*
2449 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2450 : * not copied in persistent memory.
2451 : */
2452 675468 : snapshot->active_count = 0;
2453 675468 : snapshot->regd_count = 0;
2454 675468 : snapshot->copied = false;
2455 :
2456 675468 : GetSnapshotDataInitOldSnapshot(snapshot);
2457 :
2458 675468 : return snapshot;
2459 : }
2460 :
2461 : /*
2462 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2463 : *
2464 : * This is called when installing a snapshot imported from another
2465 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2466 : * check that the source transaction is still running, and we'd better do
2467 : * that atomically with installing the new xmin.
2468 : *
2469 : * Returns true if successful, false if source xact is no longer running.
2470 : */
2471 : bool
2472 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2473 : VirtualTransactionId *sourcevxid)
2474 : {
2475 16 : bool result = false;
2476 16 : ProcArrayStruct *arrayP = procArray;
2477 : int index;
2478 :
2479 : Assert(TransactionIdIsNormal(xmin));
2480 16 : if (!sourcevxid)
2481 0 : return false;
2482 :
2483 : /* Get lock so source xact can't end while we're doing this */
2484 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2485 :
2486 32 : for (index = 0; index < arrayP->numProcs; index++)
2487 : {
2488 32 : int pgprocno = arrayP->pgprocnos[index];
2489 32 : PGPROC *proc = &allProcs[pgprocno];
2490 32 : int statusFlags = ProcGlobal->statusFlags[index];
2491 : TransactionId xid;
2492 :
2493 : /* Ignore procs running LAZY VACUUM */
2494 32 : if (statusFlags & PROC_IN_VACUUM)
2495 0 : continue;
2496 :
2497 : /* We are only interested in the specific virtual transaction. */
2498 32 : if (proc->backendId != sourcevxid->backendId)
2499 16 : continue;
2500 16 : if (proc->lxid != sourcevxid->localTransactionId)
2501 0 : continue;
2502 :
2503 : /*
2504 : * We check the transaction's database ID for paranoia's sake: if it's
2505 : * in another DB then its xmin does not cover us. Caller should have
2506 : * detected this already, so we just treat any funny cases as
2507 : * "transaction not found".
2508 : */
2509 16 : if (proc->databaseId != MyDatabaseId)
2510 0 : continue;
2511 :
2512 : /*
2513 : * Likewise, let's just make real sure its xmin does cover us.
2514 : */
2515 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2516 16 : if (!TransactionIdIsNormal(xid) ||
2517 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2518 0 : continue;
2519 :
2520 : /*
2521 : * We're good. Install the new xmin. As in GetSnapshotData, set
2522 : * TransactionXmin too. (Note that because snapmgr.c called
2523 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2524 : * we don't check that.)
2525 : */
2526 16 : MyProc->xmin = TransactionXmin = xmin;
2527 :
2528 16 : result = true;
2529 16 : break;
2530 : }
2531 :
2532 16 : LWLockRelease(ProcArrayLock);
2533 :
2534 16 : return result;
2535 : }
2536 :
2537 : /*
2538 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2539 : *
2540 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2541 : * PGPROC of the transaction from which we imported the snapshot, rather than
2542 : * an XID.
2543 : *
2544 : * Returns true if successful, false if source xact is no longer running.
2545 : */
2546 : bool
2547 1722 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2548 : {
2549 1722 : bool result = false;
2550 : TransactionId xid;
2551 :
2552 : Assert(TransactionIdIsNormal(xmin));
2553 : Assert(proc != NULL);
2554 :
2555 : /* Get lock so source xact can't end while we're doing this */
2556 1722 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2557 :
2558 : /*
2559 : * Be certain that the referenced PGPROC has an advertised xmin which is
2560 : * no later than the one we're installing, so that the system-wide xmin
2561 : * can't go backwards. Also, make sure it's running in the same database,
2562 : * so that the per-database xmin cannot go backwards.
2563 : */
2564 1722 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2565 1722 : if (proc->databaseId == MyDatabaseId &&
2566 1722 : TransactionIdIsNormal(xid) &&
2567 1722 : TransactionIdPrecedesOrEquals(xid, xmin))
2568 : {
2569 1722 : MyProc->xmin = TransactionXmin = xmin;
2570 1722 : result = true;
2571 : }
2572 :
2573 1722 : LWLockRelease(ProcArrayLock);
2574 :
2575 1722 : return result;
2576 : }
2577 :
2578 : /*
2579 : * GetRunningTransactionData -- returns information about running transactions.
2580 : *
2581 : * Similar to GetSnapshotData but returns more information. We include
2582 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2583 : * prepared transactions.
2584 : *
2585 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2586 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2587 : * array until the caller has WAL-logged this snapshot, and releases the
2588 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2589 : * lock is released.
2590 : *
2591 : * The returned data structure is statically allocated; caller should not
2592 : * modify it, and must not assume it is valid past the next call.
2593 : *
2594 : * This is never executed during recovery so there is no need to look at
2595 : * KnownAssignedXids.
2596 : *
2597 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2598 : * may return entries with duplicated TransactionId values coming from
2599 : * transaction finishing to prepare. Nothing is done about duplicated
2600 : * entries here to not hold on ProcArrayLock more than necessary.
2601 : *
2602 : * We don't worry about updating other counters, we want to keep this as
2603 : * simple as possible and leave GetSnapshotData() as the primary code for
2604 : * that bookkeeping.
2605 : *
2606 : * Note that if any transaction has overflowed its cached subtransactions
2607 : * then there is no real need include any subtransactions.
2608 : */
2609 : RunningTransactions
2610 2450 : GetRunningTransactionData(void)
2611 : {
2612 : /* result workspace */
2613 : static RunningTransactionsData CurrentRunningXactsData;
2614 :
2615 2450 : ProcArrayStruct *arrayP = procArray;
2616 2450 : TransactionId *other_xids = ProcGlobal->xids;
2617 2450 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2618 : TransactionId latestCompletedXid;
2619 : TransactionId oldestRunningXid;
2620 : TransactionId *xids;
2621 : int index;
2622 : int count;
2623 : int subcount;
2624 : bool suboverflowed;
2625 :
2626 : Assert(!RecoveryInProgress());
2627 :
2628 : /*
2629 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2630 : * be sufficient. But it seems better to do the malloc while not holding
2631 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2632 : * more subxip storage than is probably needed.
2633 : *
2634 : * Should only be allocated in bgwriter, since only ever executed during
2635 : * checkpoints.
2636 : */
2637 2450 : if (CurrentRunningXacts->xids == NULL)
2638 : {
2639 : /*
2640 : * First call
2641 : */
2642 850 : CurrentRunningXacts->xids = (TransactionId *)
2643 850 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2644 850 : if (CurrentRunningXacts->xids == NULL)
2645 0 : ereport(ERROR,
2646 : (errcode(ERRCODE_OUT_OF_MEMORY),
2647 : errmsg("out of memory")));
2648 : }
2649 :
2650 2450 : xids = CurrentRunningXacts->xids;
2651 :
2652 2450 : count = subcount = 0;
2653 2450 : suboverflowed = false;
2654 :
2655 : /*
2656 : * Ensure that no xids enter or leave the procarray while we obtain
2657 : * snapshot.
2658 : */
2659 2450 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2660 2450 : LWLockAcquire(XidGenLock, LW_SHARED);
2661 :
2662 2450 : latestCompletedXid =
2663 2450 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
2664 2450 : oldestRunningXid =
2665 2450 : XidFromFullTransactionId(ShmemVariableCache->nextXid);
2666 :
2667 : /*
2668 : * Spin over procArray collecting all xids
2669 : */
2670 7098 : for (index = 0; index < arrayP->numProcs; index++)
2671 : {
2672 : TransactionId xid;
2673 :
2674 : /* Fetch xid just once - see GetNewTransactionId */
2675 4648 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2676 :
2677 : /*
2678 : * We don't need to store transactions that don't have a TransactionId
2679 : * yet because they will not show as running on a standby server.
2680 : */
2681 4648 : if (!TransactionIdIsValid(xid))
2682 2628 : continue;
2683 :
2684 : /*
2685 : * Be careful not to exclude any xids before calculating the values of
2686 : * oldestRunningXid and suboverflowed, since these are used to clean
2687 : * up transaction information held on standbys.
2688 : */
2689 2020 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2690 2008 : oldestRunningXid = xid;
2691 :
2692 2020 : if (ProcGlobal->subxidStates[index].overflowed)
2693 4 : suboverflowed = true;
2694 :
2695 : /*
2696 : * If we wished to exclude xids this would be the right place for it.
2697 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2698 : * but they do during truncation at the end when they get the lock and
2699 : * truncate, so it is not much of a problem to include them if they
2700 : * are seen and it is cleaner to include them.
2701 : */
2702 :
2703 2020 : xids[count++] = xid;
2704 : }
2705 :
2706 : /*
2707 : * Spin over procArray collecting all subxids, but only if there hasn't
2708 : * been a suboverflow.
2709 : */
2710 2450 : if (!suboverflowed)
2711 : {
2712 2446 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2713 :
2714 7086 : for (index = 0; index < arrayP->numProcs; index++)
2715 : {
2716 4640 : int pgprocno = arrayP->pgprocnos[index];
2717 4640 : PGPROC *proc = &allProcs[pgprocno];
2718 : int nsubxids;
2719 :
2720 : /*
2721 : * Save subtransaction XIDs. Other backends can't add or remove
2722 : * entries while we're holding XidGenLock.
2723 : */
2724 4640 : nsubxids = other_subxidstates[index].count;
2725 4640 : if (nsubxids > 0)
2726 : {
2727 : /* barrier not really required, as XidGenLock is held, but ... */
2728 12 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2729 :
2730 12 : memcpy(&xids[count], (void *) proc->subxids.xids,
2731 : nsubxids * sizeof(TransactionId));
2732 12 : count += nsubxids;
2733 12 : subcount += nsubxids;
2734 :
2735 : /*
2736 : * Top-level XID of a transaction is always less than any of
2737 : * its subxids, so we don't need to check if any of the
2738 : * subxids are smaller than oldestRunningXid
2739 : */
2740 : }
2741 : }
2742 : }
2743 :
2744 : /*
2745 : * It's important *not* to include the limits set by slots here because
2746 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2747 : * were to be included here the initial value could never increase because
2748 : * of a circular dependency where slots only increase their limits when
2749 : * running xacts increases oldestRunningXid and running xacts only
2750 : * increases if slots do.
2751 : */
2752 :
2753 2450 : CurrentRunningXacts->xcnt = count - subcount;
2754 2450 : CurrentRunningXacts->subxcnt = subcount;
2755 2450 : CurrentRunningXacts->subxid_overflow = suboverflowed;
2756 2450 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2757 2450 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2758 2450 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2759 :
2760 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2761 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2762 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2763 :
2764 : /* We don't release the locks here, the caller is responsible for that */
2765 :
2766 2450 : return CurrentRunningXacts;
2767 : }
2768 :
2769 : /*
2770 : * GetOldestActiveTransactionId()
2771 : *
2772 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2773 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2774 : * We look at all databases, though there is no need to include WALSender
2775 : * since this has no effect on hot standby conflicts.
2776 : *
2777 : * This is never executed during recovery so there is no need to look at
2778 : * KnownAssignedXids.
2779 : *
2780 : * We don't worry about updating other counters, we want to keep this as
2781 : * simple as possible and leave GetSnapshotData() as the primary code for
2782 : * that bookkeeping.
2783 : */
2784 : TransactionId
2785 2066 : GetOldestActiveTransactionId(void)
2786 : {
2787 2066 : ProcArrayStruct *arrayP = procArray;
2788 2066 : TransactionId *other_xids = ProcGlobal->xids;
2789 : TransactionId oldestRunningXid;
2790 : int index;
2791 :
2792 : Assert(!RecoveryInProgress());
2793 :
2794 : /*
2795 : * Read nextXid, as the upper bound of what's still active.
2796 : *
2797 : * Reading a TransactionId is atomic, but we must grab the lock to make
2798 : * sure that all XIDs < nextXid are already present in the proc array (or
2799 : * have already completed), when we spin over it.
2800 : */
2801 2066 : LWLockAcquire(XidGenLock, LW_SHARED);
2802 2066 : oldestRunningXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2803 2066 : LWLockRelease(XidGenLock);
2804 :
2805 : /*
2806 : * Spin over procArray collecting all xids and subxids.
2807 : */
2808 2066 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2809 5070 : for (index = 0; index < arrayP->numProcs; index++)
2810 : {
2811 : TransactionId xid;
2812 :
2813 : /* Fetch xid just once - see GetNewTransactionId */
2814 3004 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2815 :
2816 3004 : if (!TransactionIdIsNormal(xid))
2817 1038 : continue;
2818 :
2819 1966 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2820 1958 : oldestRunningXid = xid;
2821 :
2822 : /*
2823 : * Top-level XID of a transaction is always less than any of its
2824 : * subxids, so we don't need to check if any of the subxids are
2825 : * smaller than oldestRunningXid
2826 : */
2827 : }
2828 2066 : LWLockRelease(ProcArrayLock);
2829 :
2830 2066 : return oldestRunningXid;
2831 : }
2832 :
2833 : /*
2834 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2835 : *
2836 : * Returns the oldest xid that we can guarantee not to have been affected by
2837 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2838 : * transaction aborted. Note that the value can (and most of the time will) be
2839 : * much more conservative than what really has been affected by vacuum, but we
2840 : * currently don't have better data available.
2841 : *
2842 : * This is useful to initialize the cutoff xid after which a new changeset
2843 : * extraction replication slot can start decoding changes.
2844 : *
2845 : * Must be called with ProcArrayLock held either shared or exclusively,
2846 : * although most callers will want to use exclusive mode since it is expected
2847 : * that the caller will immediately use the xid to peg the xmin horizon.
2848 : */
2849 : TransactionId
2850 330 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2851 : {
2852 330 : ProcArrayStruct *arrayP = procArray;
2853 : TransactionId oldestSafeXid;
2854 : int index;
2855 330 : bool recovery_in_progress = RecoveryInProgress();
2856 :
2857 : Assert(LWLockHeldByMe(ProcArrayLock));
2858 :
2859 : /*
2860 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2861 : * running. If no transaction with xid were running concurrently a new xid
2862 : * could influence the RecentXmin et al.
2863 : *
2864 : * We initialize the computation to nextXid since that's guaranteed to be
2865 : * a safe, albeit pessimal, value.
2866 : */
2867 330 : LWLockAcquire(XidGenLock, LW_SHARED);
2868 330 : oldestSafeXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2869 :
2870 : /*
2871 : * If there's already a slot pegging the xmin horizon, we can start with
2872 : * that value, it's guaranteed to be safe since it's computed by this
2873 : * routine initially and has been enforced since. We can always use the
2874 : * slot's general xmin horizon, but the catalog horizon is only usable
2875 : * when only catalog data is going to be looked at.
2876 : */
2877 332 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2878 2 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2879 : oldestSafeXid))
2880 2 : oldestSafeXid = procArray->replication_slot_xmin;
2881 :
2882 330 : if (catalogOnly &&
2883 250 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2884 40 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2885 : oldestSafeXid))
2886 12 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2887 :
2888 : /*
2889 : * If we're not in recovery, we walk over the procarray and collect the
2890 : * lowest xid. Since we're called with ProcArrayLock held and have
2891 : * acquired XidGenLock, no entries can vanish concurrently, since
2892 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2893 : * with ProcArrayLock held.
2894 : *
2895 : * In recovery we can't lower the safe value besides what we've computed
2896 : * above, so we'll have to wait a bit longer there. We unfortunately can
2897 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2898 : * machinery can miss values and return an older value than is safe.
2899 : */
2900 330 : if (!recovery_in_progress)
2901 : {
2902 330 : TransactionId *other_xids = ProcGlobal->xids;
2903 :
2904 : /*
2905 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2906 : */
2907 1644 : for (index = 0; index < arrayP->numProcs; index++)
2908 : {
2909 : TransactionId xid;
2910 :
2911 : /* Fetch xid just once - see GetNewTransactionId */
2912 1314 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2913 :
2914 1314 : if (!TransactionIdIsNormal(xid))
2915 1310 : continue;
2916 :
2917 4 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2918 4 : oldestSafeXid = xid;
2919 : }
2920 : }
2921 :
2922 330 : LWLockRelease(XidGenLock);
2923 :
2924 330 : return oldestSafeXid;
2925 : }
2926 :
2927 : /*
2928 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2929 : * delaying checkpoint because they have critical actions in progress.
2930 : *
2931 : * Constructs an array of VXIDs of transactions that are currently in commit
2932 : * critical sections, as shown by having delayChkpt set in their PGPROC.
2933 : *
2934 : * Returns a palloc'd array that should be freed by the caller.
2935 : * *nvxids is the number of valid entries.
2936 : *
2937 : * Note that because backends set or clear delayChkpt without holding any lock,
2938 : * the result is somewhat indeterminate, but we don't really care. Even in
2939 : * a multiprocessor with delayed writes to shared memory, it should be certain
2940 : * that setting of delayChkpt will propagate to shared memory when the backend
2941 : * takes a lock, so we cannot fail to see a virtual xact as delayChkpt if
2942 : * it's already inserted its commit record. Whether it takes a little while
2943 : * for clearing of delayChkpt to propagate is unimportant for correctness.
2944 : */
2945 : VirtualTransactionId *
2946 3426 : GetVirtualXIDsDelayingChkpt(int *nvxids)
2947 : {
2948 : VirtualTransactionId *vxids;
2949 3426 : ProcArrayStruct *arrayP = procArray;
2950 3426 : int count = 0;
2951 : int index;
2952 :
2953 : /* allocate what's certainly enough result space */
2954 : vxids = (VirtualTransactionId *)
2955 3426 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
2956 :
2957 3426 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2958 :
2959 7422 : for (index = 0; index < arrayP->numProcs; index++)
2960 : {
2961 3996 : int pgprocno = arrayP->pgprocnos[index];
2962 3996 : PGPROC *proc = &allProcs[pgprocno];
2963 :
2964 3996 : if (proc->delayChkpt)
2965 : {
2966 : VirtualTransactionId vxid;
2967 :
2968 4 : GET_VXID_FROM_PGPROC(vxid, *proc);
2969 4 : if (VirtualTransactionIdIsValid(vxid))
2970 4 : vxids[count++] = vxid;
2971 : }
2972 : }
2973 :
2974 3426 : LWLockRelease(ProcArrayLock);
2975 :
2976 3426 : *nvxids = count;
2977 3426 : return vxids;
2978 : }
2979 :
2980 : /*
2981 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
2982 : *
2983 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
2984 : * of the specified VXIDs are still in critical sections of code.
2985 : *
2986 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
2987 : * those numbers should be small enough for it not to be a problem.
2988 : */
2989 : bool
2990 4 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids)
2991 : {
2992 4 : bool result = false;
2993 4 : ProcArrayStruct *arrayP = procArray;
2994 : int index;
2995 :
2996 4 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2997 :
2998 28 : for (index = 0; index < arrayP->numProcs; index++)
2999 : {
3000 24 : int pgprocno = arrayP->pgprocnos[index];
3001 24 : PGPROC *proc = &allProcs[pgprocno];
3002 : VirtualTransactionId vxid;
3003 :
3004 24 : GET_VXID_FROM_PGPROC(vxid, *proc);
3005 :
3006 24 : if (proc->delayChkpt && VirtualTransactionIdIsValid(vxid))
3007 : {
3008 : int i;
3009 :
3010 0 : for (i = 0; i < nvxids; i++)
3011 : {
3012 0 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3013 : {
3014 0 : result = true;
3015 0 : break;
3016 : }
3017 : }
3018 0 : if (result)
3019 0 : break;
3020 : }
3021 : }
3022 :
3023 4 : LWLockRelease(ProcArrayLock);
3024 :
3025 4 : return result;
3026 : }
3027 :
3028 : /*
3029 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3030 : *
3031 : * Returns NULL if not found. Note that it is up to the caller to be
3032 : * sure that the question remains meaningful for long enough for the
3033 : * answer to be used ...
3034 : */
3035 : PGPROC *
3036 2996 : BackendPidGetProc(int pid)
3037 : {
3038 : PGPROC *result;
3039 :
3040 2996 : if (pid == 0) /* never match dummy PGPROCs */
3041 0 : return NULL;
3042 :
3043 2996 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3044 :
3045 2996 : result = BackendPidGetProcWithLock(pid);
3046 :
3047 2996 : LWLockRelease(ProcArrayLock);
3048 :
3049 2996 : return result;
3050 : }
3051 :
3052 : /*
3053 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3054 : *
3055 : * Same as above, except caller must be holding ProcArrayLock. The found
3056 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3057 : */
3058 : PGPROC *
3059 18654 : BackendPidGetProcWithLock(int pid)
3060 : {
3061 18654 : PGPROC *result = NULL;
3062 18654 : ProcArrayStruct *arrayP = procArray;
3063 : int index;
3064 :
3065 18654 : if (pid == 0) /* never match dummy PGPROCs */
3066 0 : return NULL;
3067 :
3068 46722 : for (index = 0; index < arrayP->numProcs; index++)
3069 : {
3070 45528 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3071 :
3072 45528 : if (proc->pid == pid)
3073 : {
3074 17460 : result = proc;
3075 17460 : break;
3076 : }
3077 : }
3078 :
3079 18654 : return result;
3080 : }
3081 :
3082 : /*
3083 : * BackendXidGetPid -- get a backend's pid given its XID
3084 : *
3085 : * Returns 0 if not found or it's a prepared transaction. Note that
3086 : * it is up to the caller to be sure that the question remains
3087 : * meaningful for long enough for the answer to be used ...
3088 : *
3089 : * Only main transaction Ids are considered. This function is mainly
3090 : * useful for determining what backend owns a lock.
3091 : *
3092 : * Beware that not every xact has an XID assigned. However, as long as you
3093 : * only call this using an XID found on disk, you're safe.
3094 : */
3095 : int
3096 0 : BackendXidGetPid(TransactionId xid)
3097 : {
3098 0 : int result = 0;
3099 0 : ProcArrayStruct *arrayP = procArray;
3100 0 : TransactionId *other_xids = ProcGlobal->xids;
3101 : int index;
3102 :
3103 0 : if (xid == InvalidTransactionId) /* never match invalid xid */
3104 0 : return 0;
3105 :
3106 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3107 :
3108 0 : for (index = 0; index < arrayP->numProcs; index++)
3109 : {
3110 0 : int pgprocno = arrayP->pgprocnos[index];
3111 0 : PGPROC *proc = &allProcs[pgprocno];
3112 :
3113 0 : if (other_xids[index] == xid)
3114 : {
3115 0 : result = proc->pid;
3116 0 : break;
3117 : }
3118 : }
3119 :
3120 0 : LWLockRelease(ProcArrayLock);
3121 :
3122 0 : return result;
3123 : }
3124 :
3125 : /*
3126 : * IsBackendPid -- is a given pid a running backend
3127 : *
3128 : * This is not called by the backend, but is called by external modules.
3129 : */
3130 : bool
3131 4 : IsBackendPid(int pid)
3132 : {
3133 4 : return (BackendPidGetProc(pid) != NULL);
3134 : }
3135 :
3136 :
3137 : /*
3138 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3139 : *
3140 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3141 : *
3142 : * The arguments allow filtering the set of VXIDs returned. Our own process
3143 : * is always skipped. In addition:
3144 : * If limitXmin is not InvalidTransactionId, skip processes with
3145 : * xmin > limitXmin.
3146 : * If excludeXmin0 is true, skip processes with xmin = 0.
3147 : * If allDbs is false, skip processes attached to other databases.
3148 : * If excludeVacuum isn't zero, skip processes for which
3149 : * (statusFlags & excludeVacuum) is not zero.
3150 : *
3151 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3152 : * allow skipping backends whose oldest live snapshot is no older than
3153 : * some snapshot we have. Since we examine the procarray with only shared
3154 : * lock, there are race conditions: a backend could set its xmin just after
3155 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3156 : * other backend could have set its xmin *before* we look. We know however
3157 : * that such a backend must have held shared ProcArrayLock overlapping our
3158 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3159 : * any snapshot the other backend is taking concurrently with our scan cannot
3160 : * consider any transactions as still running that we think are committed
3161 : * (since backends must hold ProcArrayLock exclusive to commit).
3162 : */
3163 : VirtualTransactionId *
3164 206 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3165 : bool allDbs, int excludeVacuum,
3166 : int *nvxids)
3167 : {
3168 : VirtualTransactionId *vxids;
3169 206 : ProcArrayStruct *arrayP = procArray;
3170 206 : int count = 0;
3171 : int index;
3172 :
3173 : /* allocate what's certainly enough result space */
3174 : vxids = (VirtualTransactionId *)
3175 206 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3176 :
3177 206 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3178 :
3179 1084 : for (index = 0; index < arrayP->numProcs; index++)
3180 : {
3181 878 : int pgprocno = arrayP->pgprocnos[index];
3182 878 : PGPROC *proc = &allProcs[pgprocno];
3183 878 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3184 :
3185 878 : if (proc == MyProc)
3186 206 : continue;
3187 :
3188 672 : if (excludeVacuum & statusFlags)
3189 0 : continue;
3190 :
3191 672 : if (allDbs || proc->databaseId == MyDatabaseId)
3192 : {
3193 : /* Fetch xmin just once - might change on us */
3194 260 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3195 :
3196 260 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3197 138 : continue;
3198 :
3199 : /*
3200 : * InvalidTransactionId precedes all other XIDs, so a proc that
3201 : * hasn't set xmin yet will not be rejected by this test.
3202 : */
3203 244 : if (!TransactionIdIsValid(limitXmin) ||
3204 122 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3205 : {
3206 : VirtualTransactionId vxid;
3207 :
3208 114 : GET_VXID_FROM_PGPROC(vxid, *proc);
3209 114 : if (VirtualTransactionIdIsValid(vxid))
3210 114 : vxids[count++] = vxid;
3211 : }
3212 : }
3213 : }
3214 :
3215 206 : LWLockRelease(ProcArrayLock);
3216 :
3217 206 : *nvxids = count;
3218 206 : return vxids;
3219 : }
3220 :
3221 : /*
3222 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3223 : *
3224 : * Usage is limited to conflict resolution during recovery on standby servers.
3225 : * limitXmin is supplied as either latestRemovedXid, or InvalidTransactionId
3226 : * in cases where we cannot accurately determine a value for latestRemovedXid.
3227 : *
3228 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3229 : * so we're not worried if they have a snapshot or not, nor does it really
3230 : * matter what type of lock we hold.
3231 : *
3232 : * All callers that are checking xmins always now supply a valid and useful
3233 : * value for limitXmin. The limitXmin is always lower than the lowest
3234 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3235 : * because we only care about cleanup records that are cleaning up tuple
3236 : * versions from committed transactions. In that case they will only occur
3237 : * at the point where the record is less than the lowest running xid. That
3238 : * allows us to say that if any backend takes a snapshot concurrently with
3239 : * us then the conflict assessment made here would never include the snapshot
3240 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3241 : * concurrent snapshots when limitXmin is valid. We might think about adding
3242 : * Assert(limitXmin < lowest(KnownAssignedXids))
3243 : * but that would not be true in the case of FATAL errors lagging in array,
3244 : * but we already know those are bogus anyway, so we skip that test.
3245 : *
3246 : * If dbOid is valid we skip backends attached to other databases.
3247 : *
3248 : * Be careful to *not* pfree the result from this function. We reuse
3249 : * this array sufficiently often that we use malloc for the result.
3250 : */
3251 : VirtualTransactionId *
3252 1258 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3253 : {
3254 : static VirtualTransactionId *vxids;
3255 1258 : ProcArrayStruct *arrayP = procArray;
3256 1258 : int count = 0;
3257 : int index;
3258 :
3259 : /*
3260 : * If first time through, get workspace to remember main XIDs in. We
3261 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3262 : * result space, remembering room for a terminator.
3263 : */
3264 1258 : if (vxids == NULL)
3265 : {
3266 20 : vxids = (VirtualTransactionId *)
3267 20 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3268 20 : if (vxids == NULL)
3269 0 : ereport(ERROR,
3270 : (errcode(ERRCODE_OUT_OF_MEMORY),
3271 : errmsg("out of memory")));
3272 : }
3273 :
3274 1258 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3275 :
3276 1284 : for (index = 0; index < arrayP->numProcs; index++)
3277 : {
3278 26 : int pgprocno = arrayP->pgprocnos[index];
3279 26 : PGPROC *proc = &allProcs[pgprocno];
3280 :
3281 : /* Exclude prepared transactions */
3282 26 : if (proc->pid == 0)
3283 0 : continue;
3284 :
3285 26 : if (!OidIsValid(dbOid) ||
3286 18 : proc->databaseId == dbOid)
3287 : {
3288 : /* Fetch xmin just once - can't change on us, but good coding */
3289 8 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3290 :
3291 : /*
3292 : * We ignore an invalid pxmin because this means that backend has
3293 : * no snapshot currently. We hold a Share lock to avoid contention
3294 : * with users taking snapshots. That is not a problem because the
3295 : * current xmin is always at least one higher than the latest
3296 : * removed xid, so any new snapshot would never conflict with the
3297 : * test here.
3298 : */
3299 8 : if (!TransactionIdIsValid(limitXmin) ||
3300 0 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3301 : {
3302 : VirtualTransactionId vxid;
3303 :
3304 0 : GET_VXID_FROM_PGPROC(vxid, *proc);
3305 0 : if (VirtualTransactionIdIsValid(vxid))
3306 0 : vxids[count++] = vxid;
3307 : }
3308 : }
3309 : }
3310 :
3311 1258 : LWLockRelease(ProcArrayLock);
3312 :
3313 : /* add the terminator */
3314 1258 : vxids[count].backendId = InvalidBackendId;
3315 1258 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3316 :
3317 1258 : return vxids;
3318 : }
3319 :
3320 : /*
3321 : * CancelVirtualTransaction - used in recovery conflict processing
3322 : *
3323 : * Returns pid of the process signaled, or 0 if not found.
3324 : */
3325 : pid_t
3326 0 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3327 : {
3328 0 : ProcArrayStruct *arrayP = procArray;
3329 : int index;
3330 0 : pid_t pid = 0;
3331 :
3332 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3333 :
3334 0 : for (index = 0; index < arrayP->numProcs; index++)
3335 : {
3336 0 : int pgprocno = arrayP->pgprocnos[index];
3337 0 : PGPROC *proc = &allProcs[pgprocno];
3338 : VirtualTransactionId procvxid;
3339 :
3340 0 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3341 :
3342 0 : if (procvxid.backendId == vxid.backendId &&
3343 0 : procvxid.localTransactionId == vxid.localTransactionId)
3344 : {
3345 0 : proc->recoveryConflictPending = true;
3346 0 : pid = proc->pid;
3347 0 : if (pid != 0)
3348 : {
3349 : /*
3350 : * Kill the pid if it's still here. If not, that's what we
3351 : * wanted so ignore any errors.
3352 : */
3353 0 : (void) SendProcSignal(pid, sigmode, vxid.backendId);
3354 : }
3355 0 : break;
3356 : }
3357 : }
3358 :
3359 0 : LWLockRelease(ProcArrayLock);
3360 :
3361 0 : return pid;
3362 : }
3363 :
3364 : /*
3365 : * MinimumActiveBackends --- count backends (other than myself) that are
3366 : * in active transactions. Return true if the count exceeds the
3367 : * minimum threshold passed. This is used as a heuristic to decide if
3368 : * a pre-XLOG-flush delay is worthwhile during commit.
3369 : *
3370 : * Do not count backends that are blocked waiting for locks, since they are
3371 : * not going to get to run until someone else commits.
3372 : */
3373 : bool
3374 0 : MinimumActiveBackends(int min)
3375 : {
3376 0 : ProcArrayStruct *arrayP = procArray;
3377 0 : int count = 0;
3378 : int index;
3379 :
3380 : /* Quick short-circuit if no minimum is specified */
3381 0 : if (min == 0)
3382 0 : return true;
3383 :
3384 : /*
3385 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3386 : * bogus, but since we are only testing fields for zero or nonzero, it
3387 : * should be OK. The result is only used for heuristic purposes anyway...
3388 : */
3389 0 : for (index = 0; index < arrayP->numProcs; index++)
3390 : {
3391 0 : int pgprocno = arrayP->pgprocnos[index];
3392 0 : PGPROC *proc = &allProcs[pgprocno];
3393 :
3394 : /*
3395 : * Since we're not holding a lock, need to be prepared to deal with
3396 : * garbage, as someone could have incremented numProcs but not yet
3397 : * filled the structure.
3398 : *
3399 : * If someone just decremented numProcs, 'proc' could also point to a
3400 : * PGPROC entry that's no longer in the array. It still points to a
3401 : * PGPROC struct, though, because freed PGPROC entries just go to the
3402 : * free list and are recycled. Its contents are nonsense in that case,
3403 : * but that's acceptable for this function.
3404 : */
3405 0 : if (pgprocno == -1)
3406 0 : continue; /* do not count deleted entries */
3407 0 : if (proc == MyProc)
3408 0 : continue; /* do not count myself */
3409 0 : if (proc->xid == InvalidTransactionId)
3410 0 : continue; /* do not count if no XID assigned */
3411 0 : if (proc->pid == 0)
3412 0 : continue; /* do not count prepared xacts */
3413 0 : if (proc->waitLock != NULL)
3414 0 : continue; /* do not count if blocked on a lock */
3415 0 : count++;
3416 0 : if (count >= min)
3417 0 : break;
3418 : }
3419 :
3420 0 : return count >= min;
3421 : }
3422 :
3423 : /*
3424 : * CountDBBackends --- count backends that are using specified database
3425 : */
3426 : int
3427 2 : CountDBBackends(Oid databaseid)
3428 : {
3429 2 : ProcArrayStruct *arrayP = procArray;
3430 2 : int count = 0;
3431 : int index;
3432 :
3433 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3434 :
3435 2 : for (index = 0; index < arrayP->numProcs; index++)
3436 : {
3437 0 : int pgprocno = arrayP->pgprocnos[index];
3438 0 : PGPROC *proc = &allProcs[pgprocno];
3439 :
3440 0 : if (proc->pid == 0)
3441 0 : continue; /* do not count prepared xacts */
3442 0 : if (!OidIsValid(databaseid) ||
3443 0 : proc->databaseId == databaseid)
3444 0 : count++;
3445 : }
3446 :
3447 2 : LWLockRelease(ProcArrayLock);
3448 :
3449 2 : return count;
3450 : }
3451 :
3452 : /*
3453 : * CountDBConnections --- counts database backends ignoring any background
3454 : * worker processes
3455 : */
3456 : int
3457 0 : CountDBConnections(Oid databaseid)
3458 : {
3459 0 : ProcArrayStruct *arrayP = procArray;
3460 0 : int count = 0;
3461 : int index;
3462 :
3463 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3464 :
3465 0 : for (index = 0; index < arrayP->numProcs; index++)
3466 : {
3467 0 : int pgprocno = arrayP->pgprocnos[index];
3468 0 : PGPROC *proc = &allProcs[pgprocno];
3469 :
3470 0 : if (proc->pid == 0)
3471 0 : continue; /* do not count prepared xacts */
3472 0 : if (proc->isBackgroundWorker)
3473 0 : continue; /* do not count background workers */
3474 0 : if (!OidIsValid(databaseid) ||
3475 0 : proc->databaseId == databaseid)
3476 0 : count++;
3477 : }
3478 :
3479 0 : LWLockRelease(ProcArrayLock);
3480 :
3481 0 : return count;
3482 : }
3483 :
3484 : /*
3485 : * CancelDBBackends --- cancel backends that are using specified database
3486 : */
3487 : void
3488 0 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3489 : {
3490 0 : ProcArrayStruct *arrayP = procArray;
3491 : int index;
3492 :
3493 : /* tell all backends to die */
3494 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3495 :
3496 0 : for (index = 0; index < arrayP->numProcs; index++)
3497 : {
3498 0 : int pgprocno = arrayP->pgprocnos[index];
3499 0 : PGPROC *proc = &allProcs[pgprocno];
3500 :
3501 0 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3502 : {
3503 : VirtualTransactionId procvxid;
3504 : pid_t pid;
3505 :
3506 0 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3507 :
3508 0 : proc->recoveryConflictPending = conflictPending;
3509 0 : pid = proc->pid;
3510 0 : if (pid != 0)
3511 : {
3512 : /*
3513 : * Kill the pid if it's still here. If not, that's what we
3514 : * wanted so ignore any errors.
3515 : */
3516 0 : (void) SendProcSignal(pid, sigmode, procvxid.backendId);
3517 : }
3518 : }
3519 : }
3520 :
3521 0 : LWLockRelease(ProcArrayLock);
3522 0 : }
3523 :
3524 : /*
3525 : * CountUserBackends --- count backends that are used by specified user
3526 : */
3527 : int
3528 0 : CountUserBackends(Oid roleid)
3529 : {
3530 0 : ProcArrayStruct *arrayP = procArray;
3531 0 : int count = 0;
3532 : int index;
3533 :
3534 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3535 :
3536 0 : for (index = 0; index < arrayP->numProcs; index++)
3537 : {
3538 0 : int pgprocno = arrayP->pgprocnos[index];
3539 0 : PGPROC *proc = &allProcs[pgprocno];
3540 :
3541 0 : if (proc->pid == 0)
3542 0 : continue; /* do not count prepared xacts */
3543 0 : if (proc->isBackgroundWorker)
3544 0 : continue; /* do not count background workers */
3545 0 : if (proc->roleId == roleid)
3546 0 : count++;
3547 : }
3548 :
3549 0 : LWLockRelease(ProcArrayLock);
3550 :
3551 0 : return count;
3552 : }
3553 :
3554 : /*
3555 : * CountOtherDBBackends -- check for other backends running in the given DB
3556 : *
3557 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3558 : * for them to exit. Autovacuum backends are encouraged to exit early by
3559 : * sending them SIGTERM, but normal user backends are just waited for.
3560 : *
3561 : * The current backend is always ignored; it is caller's responsibility to
3562 : * check whether the current backend uses the given DB, if it's important.
3563 : *
3564 : * Returns true if there are (still) other backends in the DB, false if not.
3565 : * Also, *nbackends and *nprepared are set to the number of other backends
3566 : * and prepared transactions in the DB, respectively.
3567 : *
3568 : * This function is used to interlock DROP DATABASE and related commands
3569 : * against there being any active backends in the target DB --- dropping the
3570 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3571 : * detect here the possibility of a newly-started backend that is trying to
3572 : * connect to the doomed database, so additional interlocking is needed during
3573 : * backend startup. The caller should normally hold an exclusive lock on the
3574 : * target DB before calling this, which is one reason we mustn't wait
3575 : * indefinitely.
3576 : */
3577 : bool
3578 1062 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3579 : {
3580 1062 : ProcArrayStruct *arrayP = procArray;
3581 :
3582 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3583 : int autovac_pids[MAXAUTOVACPIDS];
3584 : int tries;
3585 :
3586 : /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3587 1062 : for (tries = 0; tries < 50; tries++)
3588 : {
3589 1062 : int nautovacs = 0;
3590 1062 : bool found = false;
3591 : int index;
3592 :
3593 1062 : CHECK_FOR_INTERRUPTS();
3594 :
3595 1062 : *nbackends = *nprepared = 0;
3596 :
3597 1062 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3598 :
3599 2652 : for (index = 0; index < arrayP->numProcs; index++)
3600 : {
3601 1590 : int pgprocno = arrayP->pgprocnos[index];
3602 1590 : PGPROC *proc = &allProcs[pgprocno];
3603 1590 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3604 :
3605 1590 : if (proc->databaseId != databaseId)
3606 798 : continue;
3607 792 : if (proc == MyProc)
3608 792 : continue;
3609 :
3610 0 : found = true;
3611 :
3612 0 : if (proc->pid == 0)
3613 0 : (*nprepared)++;
3614 : else
3615 : {
3616 0 : (*nbackends)++;
3617 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3618 : nautovacs < MAXAUTOVACPIDS)
3619 0 : autovac_pids[nautovacs++] = proc->pid;
3620 : }
3621 : }
3622 :
3623 1062 : LWLockRelease(ProcArrayLock);
3624 :
3625 1062 : if (!found)
3626 1062 : return false; /* no conflicting backends, so done */
3627 :
3628 : /*
3629 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3630 : * postpone this step until after the loop because we don't want to
3631 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3632 : * block kill() inside the kernel...
3633 : */
3634 0 : for (index = 0; index < nautovacs; index++)
3635 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3636 :
3637 : /* sleep, then try again */
3638 0 : pg_usleep(100 * 1000L); /* 100ms */
3639 : }
3640 :
3641 0 : return true; /* timed out, still conflicts */
3642 : }
3643 :
3644 : /*
3645 : * Terminate existing connections to the specified database. This routine
3646 : * is used by the DROP DATABASE command when user has asked to forcefully
3647 : * drop the database.
3648 : *
3649 : * The current backend is always ignored; it is caller's responsibility to
3650 : * check whether the current backend uses the given DB, if it's important.
3651 : *
3652 : * It doesn't allow to terminate the connections even if there is a one
3653 : * backend with the prepared transaction in the target database.
3654 : */
3655 : void
3656 2 : TerminateOtherDBBackends(Oid databaseId)
3657 : {
3658 2 : ProcArrayStruct *arrayP = procArray;
3659 2 : List *pids = NIL;
3660 2 : int nprepared = 0;
3661 : int i;
3662 :
3663 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3664 :
3665 8 : for (i = 0; i < procArray->numProcs; i++)
3666 : {
3667 6 : int pgprocno = arrayP->pgprocnos[i];
3668 6 : PGPROC *proc = &allProcs[pgprocno];
3669 :
3670 6 : if (proc->databaseId != databaseId)
3671 6 : continue;
3672 0 : if (proc == MyProc)
3673 0 : continue;
3674 :
3675 0 : if (proc->pid != 0)
3676 0 : pids = lappend_int(pids, proc->pid);
3677 : else
3678 0 : nprepared++;
3679 : }
3680 :
3681 2 : LWLockRelease(ProcArrayLock);
3682 :
3683 2 : if (nprepared > 0)
3684 0 : ereport(ERROR,
3685 : (errcode(ERRCODE_OBJECT_IN_USE),
3686 : errmsg("database \"%s\" is being used by prepared transactions",
3687 : get_database_name(databaseId)),
3688 : errdetail_plural("There is %d prepared transaction using the database.",
3689 : "There are %d prepared transactions using the database.",
3690 : nprepared,
3691 : nprepared)));
3692 :
3693 2 : if (pids)
3694 : {
3695 : ListCell *lc;
3696 :
3697 : /*
3698 : * Check whether we have the necessary rights to terminate other
3699 : * sessions. We don't terminate any session until we ensure that we
3700 : * have rights on all the sessions to be terminated. These checks are
3701 : * the same as we do in pg_terminate_backend.
3702 : *
3703 : * In this case we don't raise some warnings - like "PID %d is not a
3704 : * PostgreSQL server process", because for us already finished session
3705 : * is not a problem.
3706 : */
3707 0 : foreach(lc, pids)
3708 : {
3709 0 : int pid = lfirst_int(lc);
3710 0 : PGPROC *proc = BackendPidGetProc(pid);
3711 :
3712 0 : if (proc != NULL)
3713 : {
3714 : /* Only allow superusers to signal superuser-owned backends. */
3715 0 : if (superuser_arg(proc->roleId) && !superuser())
3716 0 : ereport(ERROR,
3717 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3718 : errmsg("must be a superuser to terminate superuser process")));
3719 :
3720 : /* Users can signal backends they have role membership in. */
3721 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3722 0 : !has_privs_of_role(GetUserId(), DEFAULT_ROLE_SIGNAL_BACKENDID))
3723 0 : ereport(ERROR,
3724 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3725 : errmsg("must be a member of the role whose process is being terminated or member of pg_signal_backend")));
3726 : }
3727 : }
3728 :
3729 : /*
3730 : * There's a race condition here: once we release the ProcArrayLock,
3731 : * it's possible for the session to exit before we issue kill. That
3732 : * race condition possibility seems too unlikely to worry about. See
3733 : * pg_signal_backend.
3734 : */
3735 0 : foreach(lc, pids)
3736 : {
3737 0 : int pid = lfirst_int(lc);
3738 0 : PGPROC *proc = BackendPidGetProc(pid);
3739 :
3740 0 : if (proc != NULL)
3741 : {
3742 : /*
3743 : * If we have setsid(), signal the backend's whole process
3744 : * group
3745 : */
3746 : #ifdef HAVE_SETSID
3747 0 : (void) kill(-pid, SIGTERM);
3748 : #else
3749 : (void) kill(pid, SIGTERM);
3750 : #endif
3751 : }
3752 : }
3753 : }
3754 2 : }
3755 :
3756 : /*
3757 : * ProcArraySetReplicationSlotXmin
3758 : *
3759 : * Install limits to future computations of the xmin horizon to prevent vacuum
3760 : * and HOT pruning from removing affected rows still needed by clients with
3761 : * replication slots.
3762 : */
3763 : void
3764 2484 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3765 : bool already_locked)
3766 : {
3767 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3768 :
3769 2484 : if (!already_locked)
3770 2154 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3771 :
3772 2484 : procArray->replication_slot_xmin = xmin;
3773 2484 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3774 :
3775 2484 : if (!already_locked)
3776 2154 : LWLockRelease(ProcArrayLock);
3777 2484 : }
3778 :
3779 : /*
3780 : * ProcArrayGetReplicationSlotXmin
3781 : *
3782 : * Return the current slot xmin limits. That's useful to be able to remove
3783 : * data that's older than those limits.
3784 : */
3785 : void
3786 40 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3787 : TransactionId *catalog_xmin)
3788 : {
3789 40 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3790 :
3791 40 : if (xmin != NULL)
3792 0 : *xmin = procArray->replication_slot_xmin;
3793 :
3794 40 : if (catalog_xmin != NULL)
3795 40 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3796 :
3797 40 : LWLockRelease(ProcArrayLock);
3798 40 : }
3799 :
3800 : /*
3801 : * XidCacheRemoveRunningXids
3802 : *
3803 : * Remove a bunch of TransactionIds from the list of known-running
3804 : * subtransactions for my backend. Both the specified xid and those in
3805 : * the xids[] array (of length nxids) are removed from the subxids cache.
3806 : * latestXid must be the latest XID among the group.
3807 : */
3808 : void
3809 384 : XidCacheRemoveRunningXids(TransactionId xid,
3810 : int nxids, const TransactionId *xids,
3811 : TransactionId latestXid)
3812 : {
3813 : int i,
3814 : j;
3815 : XidCacheStatus *mysubxidstat;
3816 :
3817 : Assert(TransactionIdIsValid(xid));
3818 :
3819 : /*
3820 : * We must hold ProcArrayLock exclusively in order to remove transactions
3821 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
3822 : * possible this could be relaxed since we know this routine is only used
3823 : * to abort subtransactions, but pending closer analysis we'd best be
3824 : * conservative.
3825 : *
3826 : * Note that we do not have to be careful about memory ordering of our own
3827 : * reads wrt. GetNewTransactionId() here - only this process can modify
3828 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3829 : * careful about our own writes being well ordered.
3830 : */
3831 384 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3832 :
3833 384 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
3834 :
3835 : /*
3836 : * Under normal circumstances xid and xids[] will be in increasing order,
3837 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
3838 : * behavior when removing a lot of xids.
3839 : */
3840 438 : for (i = nxids - 1; i >= 0; i--)
3841 : {
3842 54 : TransactionId anxid = xids[i];
3843 :
3844 54 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
3845 : {
3846 54 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
3847 : {
3848 54 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
3849 54 : pg_write_barrier();
3850 54 : mysubxidstat->count--;
3851 54 : MyProc->subxidStatus.count--;
3852 54 : break;
3853 : }
3854 : }
3855 :
3856 : /*
3857 : * Ordinarily we should have found it, unless the cache has
3858 : * overflowed. However it's also possible for this routine to be
3859 : * invoked multiple times for the same subtransaction, in case of an
3860 : * error during AbortSubTransaction. So instead of Assert, emit a
3861 : * debug warning.
3862 : */
3863 54 : if (j < 0 && !MyProc->subxidStatus.overflowed)
3864 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
3865 : }
3866 :
3867 384 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
3868 : {
3869 384 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
3870 : {
3871 384 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
3872 384 : pg_write_barrier();
3873 384 : mysubxidstat->count--;
3874 384 : MyProc->subxidStatus.count--;
3875 384 : break;
3876 : }
3877 : }
3878 : /* Ordinarily we should have found it, unless the cache has overflowed */
3879 384 : if (j < 0 && !MyProc->subxidStatus.overflowed)
3880 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
3881 :
3882 : /* Also advance global latestCompletedXid while holding the lock */
3883 384 : MaintainLatestCompletedXid(latestXid);
3884 :
3885 384 : LWLockRelease(ProcArrayLock);
3886 384 : }
3887 :
3888 : #ifdef XIDCACHE_DEBUG
3889 :
3890 : /*
3891 : * Print stats about effectiveness of XID cache
3892 : */
3893 : static void
3894 : DisplayXidCache(void)
3895 : {
3896 : fprintf(stderr,
3897 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
3898 : xc_by_recent_xmin,
3899 : xc_by_known_xact,
3900 : xc_by_my_xact,
3901 : xc_by_latest_xid,
3902 : xc_by_main_xid,
3903 : xc_by_child_xid,
3904 : xc_by_known_assigned,
3905 : xc_no_overflow,
3906 : xc_slow_answer);
3907 : }
3908 : #endif /* XIDCACHE_DEBUG */
3909 :
3910 : /*
3911 : * If rel != NULL, return test state appropriate for relation, otherwise
3912 : * return state usable for all relations. The latter may consider XIDs as
3913 : * not-yet-visible-to-everyone that a state for a specific relation would
3914 : * already consider visible-to-everyone.
3915 : *
3916 : * This needs to be called while a snapshot is active or registered, otherwise
3917 : * there are wraparound and other dangers.
3918 : *
3919 : * See comment for GlobalVisState for details.
3920 : */
3921 : GlobalVisState *
3922 18226754 : GlobalVisTestFor(Relation rel)
3923 : {
3924 : bool need_shared;
3925 : bool need_catalog;
3926 : GlobalVisState *state;
3927 :
3928 : /* XXX: we should assert that a snapshot is pushed or registered */
3929 : Assert(RecentXmin);
3930 :
3931 18226754 : if (!rel)
3932 5684 : need_shared = need_catalog = true;
3933 : else
3934 : {
3935 : /*
3936 : * Other kinds currently don't contain xids, nor always the necessary
3937 : * logical decoding markers.
3938 : */
3939 : Assert(rel->rd_rel->relkind == RELKIND_RELATION ||
3940 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
3941 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
3942 :
3943 18221070 : need_shared = rel->rd_rel->relisshared || RecoveryInProgress();
3944 18221070 : need_catalog = IsCatalogRelation(rel) || RelationIsAccessibleInLogicalDecoding(rel);
3945 : }
3946 :
3947 18226754 : if (need_shared)
3948 166912 : state = &GlobalVisSharedRels;
3949 18059842 : else if (need_catalog)
3950 3192934 : state = &GlobalVisCatalogRels;
3951 14866908 : else if (RELATION_IS_LOCAL(rel))
3952 22862 : state = &GlobalVisTempRels;
3953 : else
3954 14844046 : state = &GlobalVisDataRels;
3955 :
3956 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
3957 : FullTransactionIdIsValid(state->maybe_needed));
3958 :
3959 18226754 : return state;
3960 : }
3961 :
3962 : /*
3963 : * Return true if it's worth updating the accurate maybe_needed boundary.
3964 : *
3965 : * As it is somewhat expensive to determine xmin horizons, we don't want to
3966 : * repeatedly do so when there is a low likelihood of it being beneficial.
3967 : *
3968 : * The current heuristic is that we update only if RecentXmin has changed
3969 : * since the last update. If the oldest currently running transaction has not
3970 : * finished, it is unlikely that recomputing the horizon would be useful.
3971 : */
3972 : static bool
3973 510838 : GlobalVisTestShouldUpdate(GlobalVisState *state)
3974 : {
3975 : /* hasn't been updated yet */
3976 510838 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
3977 7496 : return true;
3978 :
3979 : /*
3980 : * If the maybe_needed/definitely_needed boundaries are the same, it's
3981 : * unlikely to be beneficial to refresh boundaries.
3982 : */
3983 503342 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
3984 : state->definitely_needed))
3985 46 : return false;
3986 :
3987 : /* does the last snapshot built have a different xmin? */
3988 503296 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
3989 : }
3990 :
3991 : static void
3992 219158 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
3993 : {
3994 : GlobalVisSharedRels.maybe_needed =
3995 219158 : FullXidRelativeTo(horizons->latest_completed,
3996 : horizons->shared_oldest_nonremovable);
3997 : GlobalVisCatalogRels.maybe_needed =
3998 219158 : FullXidRelativeTo(horizons->latest_completed,
3999 : horizons->catalog_oldest_nonremovable);
4000 : GlobalVisDataRels.maybe_needed =
4001 219158 : FullXidRelativeTo(horizons->latest_completed,
4002 : horizons->data_oldest_nonremovable);
4003 : GlobalVisTempRels.maybe_needed =
4004 219158 : FullXidRelativeTo(horizons->latest_completed,
4005 : horizons->temp_oldest_nonremovable);
4006 :
4007 : /*
4008 : * In longer running transactions it's possible that transactions we
4009 : * previously needed to treat as running aren't around anymore. So update
4010 : * definitely_needed to not be earlier than maybe_needed.
4011 : */
4012 : GlobalVisSharedRels.definitely_needed =
4013 219158 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4014 : GlobalVisSharedRels.definitely_needed);
4015 : GlobalVisCatalogRels.definitely_needed =
4016 219158 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4017 : GlobalVisCatalogRels.definitely_needed);
4018 : GlobalVisDataRels.definitely_needed =
4019 219158 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4020 : GlobalVisDataRels.definitely_needed);
4021 219158 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4022 :
4023 219158 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4024 219158 : }
4025 :
4026 : /*
4027 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4028 : * using ComputeXidHorizons().
4029 : */
4030 : static void
4031 106938 : GlobalVisUpdate(void)
4032 : {
4033 : ComputeXidHorizonsResult horizons;
4034 :
4035 : /* updates the horizons as a side-effect */
4036 106938 : ComputeXidHorizons(&horizons);
4037 106938 : }
4038 :
4039 : /*
4040 : * Return true if no snapshot still considers fxid to be running.
4041 : *
4042 : * The state passed needs to have been initialized for the relation fxid is
4043 : * from (NULL is also OK), otherwise the result may not be correct.
4044 : *
4045 : * See comment for GlobalVisState for details.
4046 : */
4047 : bool
4048 12442242 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4049 : FullTransactionId fxid)
4050 : {
4051 : /*
4052 : * If fxid is older than maybe_needed bound, it definitely is visible to
4053 : * everyone.
4054 : */
4055 12442242 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4056 3267100 : return true;
4057 :
4058 : /*
4059 : * If fxid is >= definitely_needed bound, it is very likely to still be
4060 : * considered running.
4061 : */
4062 9175142 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4063 8664370 : return false;
4064 :
4065 : /*
4066 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4067 : * might not exist a snapshot considering fxid running. If it makes sense,
4068 : * update boundaries and recheck.
4069 : */
4070 510772 : if (GlobalVisTestShouldUpdate(state))
4071 : {
4072 106918 : GlobalVisUpdate();
4073 :
4074 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4075 :
4076 106918 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4077 : }
4078 : else
4079 403854 : return false;
4080 : }
4081 :
4082 : /*
4083 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4084 : *
4085 : * It is crucial that this only gets called for xids from a source that
4086 : * protects against xid wraparounds (e.g. from a table and thus protected by
4087 : * relfrozenxid).
4088 : */
4089 : bool
4090 12442242 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4091 : {
4092 : FullTransactionId fxid;
4093 :
4094 : /*
4095 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4096 : * because we know the xid has to, at the very least, be between
4097 : * [oldestXid, nextFullXid), i.e. within 2 billion of xid. To avoid taking
4098 : * a lock to determine either, we can just compare with
4099 : * state->definitely_needed, which was based on those value at the time
4100 : * the current snapshot was built.
4101 : */
4102 12442242 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4103 :
4104 12442242 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4105 : }
4106 :
4107 : /*
4108 : * Return FullTransactionId below which all transactions are not considered
4109 : * running anymore.
4110 : *
4111 : * Note: This is less efficient than testing with
4112 : * GlobalVisTestIsRemovableFullXid as it likely requires building an accurate
4113 : * cutoff, even in the case all the XIDs compared with the cutoff are outside
4114 : * [maybe_needed, definitely_needed).
4115 : */
4116 : FullTransactionId
4117 66 : GlobalVisTestNonRemovableFullHorizon(GlobalVisState *state)
4118 : {
4119 : /* acquire accurate horizon if not already done */
4120 66 : if (GlobalVisTestShouldUpdate(state))
4121 20 : GlobalVisUpdate();
4122 :
4123 66 : return state->maybe_needed;
4124 : }
4125 :
4126 : /* Convenience wrapper around GlobalVisTestNonRemovableFullHorizon */
4127 : TransactionId
4128 66 : GlobalVisTestNonRemovableHorizon(GlobalVisState *state)
4129 : {
4130 : FullTransactionId cutoff;
4131 :
4132 66 : cutoff = GlobalVisTestNonRemovableFullHorizon(state);
4133 :
4134 66 : return XidFromFullTransactionId(cutoff);
4135 : }
4136 :
4137 : /*
4138 : * Convenience wrapper around GlobalVisTestFor() and
4139 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4140 : */
4141 : bool
4142 0 : GlobalVisIsRemovableFullXid(Relation rel, FullTransactionId fxid)
4143 : {
4144 : GlobalVisState *state;
4145 :
4146 0 : state = GlobalVisTestFor(rel);
4147 :
4148 0 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4149 : }
4150 :
4151 : /*
4152 : * Convenience wrapper around GlobalVisTestFor() and
4153 : * GlobalVisTestIsRemovableXid(), see their comments.
4154 : */
4155 : bool
4156 1922 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4157 : {
4158 : GlobalVisState *state;
4159 :
4160 1922 : state = GlobalVisTestFor(rel);
4161 :
4162 1922 : return GlobalVisTestIsRemovableXid(state, xid);
4163 : }
4164 :
4165 : /*
4166 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4167 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4168 : *
4169 : * Be very careful about when to use this function. It can only safely be used
4170 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4171 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4172 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4173 : * the xid has to be within that range), or if xid is from the procarray and
4174 : * prevents xid wraparound that way.
4175 : */
4176 : static inline FullTransactionId
4177 15697222 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4178 : {
4179 15697222 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4180 :
4181 : Assert(TransactionIdIsValid(xid));
4182 : Assert(TransactionIdIsValid(rel_xid));
4183 :
4184 : /* not guaranteed to find issues, but likely to catch mistakes */
4185 : AssertTransactionIdInAllowableRange(xid);
4186 :
4187 47091666 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4188 15697222 : + (int32) (xid - rel_xid));
4189 : }
4190 :
4191 :
4192 : /* ----------------------------------------------
4193 : * KnownAssignedTransactionIds sub-module
4194 : * ----------------------------------------------
4195 : */
4196 :
4197 : /*
4198 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4199 : * running on the primary at the current point in WAL. These XIDs must be
4200 : * treated as running by standby transactions, even though they are not in
4201 : * the standby server's PGPROC array.
4202 : *
4203 : * We record all XIDs that we know have been assigned. That includes all the
4204 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4205 : * been assigned. We can deduce the existence of unobserved XIDs because we
4206 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4207 : * list expands as new XIDs are observed or inferred, and contracts when
4208 : * transaction completion records arrive.
4209 : *
4210 : * During hot standby we do not fret too much about the distinction between
4211 : * top-level XIDs and subtransaction XIDs. We store both together in the
4212 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4213 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4214 : * doesn't care about the distinction either. Subtransaction XIDs are
4215 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4216 : * links are *not* maintained (which does not affect visibility).
4217 : *
4218 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4219 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4220 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4221 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4222 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4223 : * and then remove them from KnownAssignedXids. This prevents overflow of
4224 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4225 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4226 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4227 : * though it should be complete for top-level XIDs; this is the same situation
4228 : * that holds with respect to the PGPROC entries in normal running.
4229 : *
4230 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4231 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4232 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4233 : * As long as that is within the range of interesting XIDs, we have to assume
4234 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4235 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4236 : * subXID arrives - that is not an error.)
4237 : *
4238 : * Should a backend on primary somehow disappear before it can write an abort
4239 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4240 : * aborted but we think they were running; the distinction is irrelevant
4241 : * because either way any changes done by the transaction are not visible to
4242 : * backends in the standby. We prune KnownAssignedXids when
4243 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4244 : * array due to such dead XIDs.
4245 : */
4246 :
4247 : /*
4248 : * RecordKnownAssignedTransactionIds
4249 : * Record the given XID in KnownAssignedXids, as well as any preceding
4250 : * unobserved XIDs.
4251 : *
4252 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4253 : * associated with a transaction. Must be called for each record after we
4254 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4255 : *
4256 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4257 : */
4258 : void
4259 201122 : RecordKnownAssignedTransactionIds(TransactionId xid)
4260 : {
4261 : Assert(standbyState >= STANDBY_INITIALIZED);
4262 : Assert(TransactionIdIsValid(xid));
4263 : Assert(TransactionIdIsValid(latestObservedXid));
4264 :
4265 201122 : elog(trace_recovery(DEBUG4), "record known xact %u latestObservedXid %u",
4266 : xid, latestObservedXid);
4267 :
4268 : /*
4269 : * When a newly observed xid arrives, it is frequently the case that it is
4270 : * *not* the next xid in sequence. When this occurs, we must treat the
4271 : * intervening xids as running also.
4272 : */
4273 201122 : if (TransactionIdFollows(xid, latestObservedXid))
4274 : {
4275 : TransactionId next_expected_xid;
4276 :
4277 : /*
4278 : * Extend subtrans like we do in GetNewTransactionId() during normal
4279 : * operation using individual extend steps. Note that we do not need
4280 : * to extend clog since its extensions are WAL logged.
4281 : *
4282 : * This part has to be done regardless of standbyState since we
4283 : * immediately start assigning subtransactions to their toplevel
4284 : * transactions.
4285 : */
4286 3232 : next_expected_xid = latestObservedXid;
4287 6472 : while (TransactionIdPrecedes(next_expected_xid, xid))
4288 : {
4289 3240 : TransactionIdAdvance(next_expected_xid);
4290 3240 : ExtendSUBTRANS(next_expected_xid);
4291 : }
4292 : Assert(next_expected_xid == xid);
4293 :
4294 : /*
4295 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4296 : * more to do since we don't track assigned xids yet.
4297 : */
4298 3232 : if (standbyState <= STANDBY_INITIALIZED)
4299 : {
4300 4 : latestObservedXid = xid;
4301 4 : return;
4302 : }
4303 :
4304 : /*
4305 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4306 : */
4307 3228 : next_expected_xid = latestObservedXid;
4308 3228 : TransactionIdAdvance(next_expected_xid);
4309 3228 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4310 :
4311 : /*
4312 : * Now we can advance latestObservedXid
4313 : */
4314 3228 : latestObservedXid = xid;
4315 :
4316 : /* ShmemVariableCache->nextXid must be beyond any observed xid */
4317 3228 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4318 3228 : next_expected_xid = latestObservedXid;
4319 3228 : TransactionIdAdvance(next_expected_xid);
4320 : }
4321 : }
4322 :
4323 : /*
4324 : * ExpireTreeKnownAssignedTransactionIds
4325 : * Remove the given XIDs from KnownAssignedXids.
4326 : *
4327 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4328 : */
4329 : void
4330 278 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4331 : TransactionId *subxids, TransactionId max_xid)
4332 : {
4333 : Assert(standbyState >= STANDBY_INITIALIZED);
4334 :
4335 : /*
4336 : * Uses same locking as transaction commit
4337 : */
4338 278 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4339 :
4340 278 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4341 :
4342 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4343 278 : MaintainLatestCompletedXidRecovery(max_xid);
4344 :
4345 : /* ... and xactCompletionCount */
4346 278 : ShmemVariableCache->xactCompletionCount++;
4347 :
4348 278 : LWLockRelease(ProcArrayLock);
4349 278 : }
4350 :
4351 : /*
4352 : * ExpireAllKnownAssignedTransactionIds
4353 : * Remove all entries in KnownAssignedXids
4354 : */
4355 : void
4356 50 : ExpireAllKnownAssignedTransactionIds(void)
4357 : {
4358 50 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4359 50 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4360 50 : LWLockRelease(ProcArrayLock);
4361 50 : }
4362 :
4363 : /*
4364 : * ExpireOldKnownAssignedTransactionIds
4365 : * Remove KnownAssignedXids entries preceding the given XID
4366 : */
4367 : void
4368 192 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4369 : {
4370 192 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4371 192 : KnownAssignedXidsRemovePreceding(xid);
4372 192 : LWLockRelease(ProcArrayLock);
4373 192 : }
4374 :
4375 :
4376 : /*
4377 : * Private module functions to manipulate KnownAssignedXids
4378 : *
4379 : * There are 5 main uses of the KnownAssignedXids data structure:
4380 : *
4381 : * * backends taking snapshots - all valid XIDs need to be copied out
4382 : * * backends seeking to determine presence of a specific XID
4383 : * * startup process adding new known-assigned XIDs
4384 : * * startup process removing specific XIDs as transactions end
4385 : * * startup process pruning array when special WAL records arrive
4386 : *
4387 : * This data structure is known to be a hot spot during Hot Standby, so we
4388 : * go to some lengths to make these operations as efficient and as concurrent
4389 : * as possible.
4390 : *
4391 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4392 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4393 : * in general TransactionIdPrecedes would not provide a total order, but
4394 : * we know that the entries present at any instant should not extend across
4395 : * a large enough fraction of XID space to wrap around (the primary would
4396 : * shut down for fear of XID wrap long before that happens). So it's OK to
4397 : * use TransactionIdPrecedes as a binary-search comparator.
4398 : *
4399 : * It's cheap to maintain the sortedness during insertions, since new known
4400 : * XIDs are always reported in XID order; we just append them at the right.
4401 : *
4402 : * To keep individual deletions cheap, we need to allow gaps in the array.
4403 : * This is implemented by marking array elements as valid or invalid using
4404 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4405 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4406 : * XID entry itself. This preserves the property that the XID entries are
4407 : * sorted, so we can do binary searches easily. Periodically we compress
4408 : * out the unused entries; that's much cheaper than having to compress the
4409 : * array immediately on every deletion.
4410 : *
4411 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4412 : * are those with indexes tail <= i < head; items outside this subscript range
4413 : * have unspecified contents. When head reaches the end of the array, we
4414 : * force compression of unused entries rather than wrapping around, since
4415 : * allowing wraparound would greatly complicate the search logic. We maintain
4416 : * an explicit tail pointer so that pruning of old XIDs can be done without
4417 : * immediately moving the array contents. In most cases only a small fraction
4418 : * of the array contains valid entries at any instant.
4419 : *
4420 : * Although only the startup process can ever change the KnownAssignedXids
4421 : * data structure, we still need interlocking so that standby backends will
4422 : * not observe invalid intermediate states. The convention is that backends
4423 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4424 : * the array, the startup process must hold ProcArrayLock exclusively, for
4425 : * the usual transactional reasons (compare commit/abort of a transaction
4426 : * during normal running). Compressing unused entries out of the array
4427 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4428 : * them into slots to the right of the head pointer and then advance the head
4429 : * pointer. This wouldn't require any lock at all, except that on machines
4430 : * with weak memory ordering we need to be careful that other processors
4431 : * see the array element changes before they see the head pointer change.
4432 : * We handle this by using a spinlock to protect reads and writes of the
4433 : * head/tail pointers. (We could dispense with the spinlock if we were to
4434 : * create suitable memory access barrier primitives and use those instead.)
4435 : * The spinlock must be taken to read or write the head/tail pointers unless
4436 : * the caller holds ProcArrayLock exclusively.
4437 : *
4438 : * Algorithmic analysis:
4439 : *
4440 : * If we have a maximum of M slots, with N XIDs currently spread across
4441 : * S elements then we have N <= S <= M always.
4442 : *
4443 : * * Adding a new XID is O(1) and needs little locking (unless compression
4444 : * must happen)
4445 : * * Compressing the array is O(S) and requires exclusive lock
4446 : * * Removing an XID is O(logS) and requires exclusive lock
4447 : * * Taking a snapshot is O(S) and requires shared lock
4448 : * * Checking for an XID is O(logS) and requires shared lock
4449 : *
4450 : * In comparison, using a hash table for KnownAssignedXids would mean that
4451 : * taking snapshots would be O(M). If we can maintain S << M then the
4452 : * sorted array technique will deliver significantly faster snapshots.
4453 : * If we try to keep S too small then we will spend too much time compressing,
4454 : * so there is an optimal point for any workload mix. We use a heuristic to
4455 : * decide when to compress the array, though trimming also helps reduce
4456 : * frequency of compressing. The heuristic requires us to track the number of
4457 : * currently valid XIDs in the array.
4458 : */
4459 :
4460 :
4461 : /*
4462 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4463 : * array, removing any gaps.
4464 : *
4465 : * A compression step is forced if "force" is true, otherwise we do it
4466 : * only if a heuristic indicates it's a good time to do it.
4467 : *
4468 : * Caller must hold ProcArrayLock in exclusive mode.
4469 : */
4470 : static void
4471 510 : KnownAssignedXidsCompress(bool force)
4472 : {
4473 510 : ProcArrayStruct *pArray = procArray;
4474 : int head,
4475 : tail;
4476 : int compress_index;
4477 : int i;
4478 :
4479 : /* no spinlock required since we hold ProcArrayLock exclusively */
4480 510 : head = pArray->headKnownAssignedXids;
4481 510 : tail = pArray->tailKnownAssignedXids;
4482 :
4483 510 : if (!force)
4484 : {
4485 : /*
4486 : * If we can choose how much to compress, use a heuristic to avoid
4487 : * compressing too often or not often enough.
4488 : *
4489 : * Heuristic is if we have a large enough current spread and less than
4490 : * 50% of the elements are currently in use, then compress. This
4491 : * should ensure we compress fairly infrequently. We could compress
4492 : * less often though the virtual array would spread out more and
4493 : * snapshots would become more expensive.
4494 : */
4495 510 : int nelements = head - tail;
4496 :
4497 510 : if (nelements < 4 * PROCARRAY_MAXPROCS ||
4498 16 : nelements < 2 * pArray->numKnownAssignedXids)
4499 494 : return;
4500 : }
4501 :
4502 : /*
4503 : * We compress the array by reading the valid values from tail to head,
4504 : * re-aligning data to 0th element.
4505 : */
4506 16 : compress_index = 0;
4507 2208 : for (i = tail; i < head; i++)
4508 : {
4509 2192 : if (KnownAssignedXidsValid[i])
4510 : {
4511 16 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4512 16 : KnownAssignedXidsValid[compress_index] = true;
4513 16 : compress_index++;
4514 : }
4515 : }
4516 :
4517 16 : pArray->tailKnownAssignedXids = 0;
4518 16 : pArray->headKnownAssignedXids = compress_index;
4519 : }
4520 :
4521 : /*
4522 : * Add xids into KnownAssignedXids at the head of the array.
4523 : *
4524 : * xids from from_xid to to_xid, inclusive, are added to the array.
4525 : *
4526 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4527 : * exclusive mode, so we need no extra locking here. Else caller holds no
4528 : * lock, so we need to be sure we maintain sufficient interlocks against
4529 : * concurrent readers. (Only the startup process ever calls this, so no need
4530 : * to worry about concurrent writers.)
4531 : */
4532 : static void
4533 3232 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4534 : bool exclusive_lock)
4535 : {
4536 3232 : ProcArrayStruct *pArray = procArray;
4537 : TransactionId next_xid;
4538 : int head,
4539 : tail;
4540 : int nxids;
4541 : int i;
4542 :
4543 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4544 :
4545 : /*
4546 : * Calculate how many array slots we'll need. Normally this is cheap; in
4547 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4548 : * way.
4549 : */
4550 3232 : if (to_xid >= from_xid)
4551 3232 : nxids = to_xid - from_xid + 1;
4552 : else
4553 : {
4554 0 : nxids = 1;
4555 0 : next_xid = from_xid;
4556 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4557 : {
4558 0 : nxids++;
4559 0 : TransactionIdAdvance(next_xid);
4560 : }
4561 : }
4562 :
4563 : /*
4564 : * Since only the startup process modifies the head/tail pointers, we
4565 : * don't need a lock to read them here.
4566 : */
4567 3232 : head = pArray->headKnownAssignedXids;
4568 3232 : tail = pArray->tailKnownAssignedXids;
4569 :
4570 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4571 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4572 :
4573 : /*
4574 : * Verify that insertions occur in TransactionId sequence. Note that even
4575 : * if the last existing element is marked invalid, it must still have a
4576 : * correctly sequenced XID value.
4577 : */
4578 6198 : if (head > tail &&
4579 2966 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4580 : {
4581 0 : KnownAssignedXidsDisplay(LOG);
4582 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4583 : }
4584 :
4585 : /*
4586 : * If our xids won't fit in the remaining space, compress out free space
4587 : */
4588 3232 : if (head + nxids > pArray->maxKnownAssignedXids)
4589 : {
4590 : /* must hold lock to compress */
4591 0 : if (!exclusive_lock)
4592 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4593 :
4594 0 : KnownAssignedXidsCompress(true);
4595 :
4596 0 : head = pArray->headKnownAssignedXids;
4597 : /* note: we no longer care about the tail pointer */
4598 :
4599 0 : if (!exclusive_lock)
4600 0 : LWLockRelease(ProcArrayLock);
4601 :
4602 : /*
4603 : * If it still won't fit then we're out of memory
4604 : */
4605 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4606 0 : elog(ERROR, "too many KnownAssignedXids");
4607 : }
4608 :
4609 : /* Now we can insert the xids into the space starting at head */
4610 3232 : next_xid = from_xid;
4611 6472 : for (i = 0; i < nxids; i++)
4612 : {
4613 3240 : KnownAssignedXids[head] = next_xid;
4614 3240 : KnownAssignedXidsValid[head] = true;
4615 3240 : TransactionIdAdvance(next_xid);
4616 3240 : head++;
4617 : }
4618 :
4619 : /* Adjust count of number of valid entries */
4620 3232 : pArray->numKnownAssignedXids += nxids;
4621 :
4622 : /*
4623 : * Now update the head pointer. We use a spinlock to protect this
4624 : * pointer, not because the update is likely to be non-atomic, but to
4625 : * ensure that other processors see the above array updates before they
4626 : * see the head pointer change.
4627 : *
4628 : * If we're holding ProcArrayLock exclusively, there's no need to take the
4629 : * spinlock.
4630 : */
4631 3232 : if (exclusive_lock)
4632 4 : pArray->headKnownAssignedXids = head;
4633 : else
4634 : {
4635 3228 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4636 3228 : pArray->headKnownAssignedXids = head;
4637 3228 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4638 : }
4639 3232 : }
4640 :
4641 : /*
4642 : * KnownAssignedXidsSearch
4643 : *
4644 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4645 : * Returns true if it was found, false if not.
4646 : *
4647 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4648 : * Exclusive lock must be held for remove = true.
4649 : */
4650 : static bool
4651 5388 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4652 : {
4653 5388 : ProcArrayStruct *pArray = procArray;
4654 : int first,
4655 : last;
4656 : int head;
4657 : int tail;
4658 5388 : int result_index = -1;
4659 :
4660 5388 : if (remove)
4661 : {
4662 : /* we hold ProcArrayLock exclusively, so no need for spinlock */
4663 5388 : tail = pArray->tailKnownAssignedXids;
4664 5388 : head = pArray->headKnownAssignedXids;
4665 : }
4666 : else
4667 : {
4668 : /* take spinlock to ensure we see up-to-date array contents */
4669 0 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4670 0 : tail = pArray->tailKnownAssignedXids;
4671 0 : head = pArray->headKnownAssignedXids;
4672 0 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4673 : }
4674 :
4675 : /*
4676 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4677 : * array here, since even invalid entries will contain sorted XIDs.
4678 : */
4679 5388 : first = tail;
4680 5388 : last = head - 1;
4681 28256 : while (first <= last)
4682 : {
4683 : int mid_index;
4684 : TransactionId mid_xid;
4685 :
4686 25946 : mid_index = (first + last) / 2;
4687 25946 : mid_xid = KnownAssignedXids[mid_index];
4688 :
4689 25946 : if (xid == mid_xid)
4690 : {
4691 3078 : result_index = mid_index;
4692 3078 : break;
4693 : }
4694 22868 : else if (TransactionIdPrecedes(xid, mid_xid))
4695 16070 : last = mid_index - 1;
4696 : else
4697 6798 : first = mid_index + 1;
4698 : }
4699 :
4700 5388 : if (result_index < 0)
4701 2310 : return false; /* not in array */
4702 :
4703 3078 : if (!KnownAssignedXidsValid[result_index])
4704 0 : return false; /* in array, but invalid */
4705 :
4706 3078 : if (remove)
4707 : {
4708 3078 : KnownAssignedXidsValid[result_index] = false;
4709 :
4710 3078 : pArray->numKnownAssignedXids--;
4711 : Assert(pArray->numKnownAssignedXids >= 0);
4712 :
4713 : /*
4714 : * If we're removing the tail element then advance tail pointer over
4715 : * any invalid elements. This will speed future searches.
4716 : */
4717 3078 : if (result_index == tail)
4718 : {
4719 506 : tail++;
4720 774 : while (tail < head && !KnownAssignedXidsValid[tail])
4721 268 : tail++;
4722 506 : if (tail >= head)
4723 : {
4724 : /* Array is empty, so we can reset both pointers */
4725 254 : pArray->headKnownAssignedXids = 0;
4726 254 : pArray->tailKnownAssignedXids = 0;
4727 : }
4728 : else
4729 : {
4730 252 : pArray->tailKnownAssignedXids = tail;
4731 : }
4732 : }
4733 : }
4734 :
4735 3078 : return true;
4736 : }
4737 :
4738 : /*
4739 : * Is the specified XID present in KnownAssignedXids[]?
4740 : *
4741 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4742 : */
4743 : static bool
4744 0 : KnownAssignedXidExists(TransactionId xid)
4745 : {
4746 : Assert(TransactionIdIsValid(xid));
4747 :
4748 0 : return KnownAssignedXidsSearch(xid, false);
4749 : }
4750 :
4751 : /*
4752 : * Remove the specified XID from KnownAssignedXids[].
4753 : *
4754 : * Caller must hold ProcArrayLock in exclusive mode.
4755 : */
4756 : static void
4757 5388 : KnownAssignedXidsRemove(TransactionId xid)
4758 : {
4759 : Assert(TransactionIdIsValid(xid));
4760 :
4761 5388 : elog(trace_recovery(DEBUG4), "remove KnownAssignedXid %u", xid);
4762 :
4763 : /*
4764 : * Note: we cannot consider it an error to remove an XID that's not
4765 : * present. We intentionally remove subxact IDs while processing
4766 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4767 : * removed again when the top-level xact commits or aborts.
4768 : *
4769 : * It might be possible to track such XIDs to distinguish this case from
4770 : * actual errors, but it would be complicated and probably not worth it.
4771 : * So, just ignore the search result.
4772 : */
4773 5388 : (void) KnownAssignedXidsSearch(xid, true);
4774 5388 : }
4775 :
4776 : /*
4777 : * KnownAssignedXidsRemoveTree
4778 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4779 : *
4780 : * Caller must hold ProcArrayLock in exclusive mode.
4781 : */
4782 : static void
4783 318 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
4784 : TransactionId *subxids)
4785 : {
4786 : int i;
4787 :
4788 318 : if (TransactionIdIsValid(xid))
4789 278 : KnownAssignedXidsRemove(xid);
4790 :
4791 5428 : for (i = 0; i < nsubxids; i++)
4792 5110 : KnownAssignedXidsRemove(subxids[i]);
4793 :
4794 : /* Opportunistically compress the array */
4795 318 : KnownAssignedXidsCompress(false);
4796 318 : }
4797 :
4798 : /*
4799 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
4800 : * then clear the whole table.
4801 : *
4802 : * Caller must hold ProcArrayLock in exclusive mode.
4803 : */
4804 : static void
4805 242 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
4806 : {
4807 242 : ProcArrayStruct *pArray = procArray;
4808 242 : int count = 0;
4809 : int head,
4810 : tail,
4811 : i;
4812 :
4813 242 : if (!TransactionIdIsValid(removeXid))
4814 : {
4815 50 : elog(trace_recovery(DEBUG4), "removing all KnownAssignedXids");
4816 50 : pArray->numKnownAssignedXids = 0;
4817 50 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
4818 50 : return;
4819 : }
4820 :
4821 192 : elog(trace_recovery(DEBUG4), "prune KnownAssignedXids to %u", removeXid);
4822 :
4823 : /*
4824 : * Mark entries invalid starting at the tail. Since array is sorted, we
4825 : * can stop as soon as we reach an entry >= removeXid.
4826 : */
4827 192 : tail = pArray->tailKnownAssignedXids;
4828 192 : head = pArray->headKnownAssignedXids;
4829 :
4830 192 : for (i = tail; i < head; i++)
4831 : {
4832 36 : if (KnownAssignedXidsValid[i])
4833 : {
4834 36 : TransactionId knownXid = KnownAssignedXids[i];
4835 :
4836 36 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
4837 36 : break;
4838 :
4839 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
4840 : {
4841 0 : KnownAssignedXidsValid[i] = false;
4842 0 : count++;
4843 : }
4844 : }
4845 : }
4846 :
4847 192 : pArray->numKnownAssignedXids -= count;
4848 : Assert(pArray->numKnownAssignedXids >= 0);
4849 :
4850 : /*
4851 : * Advance the tail pointer if we've marked the tail item invalid.
4852 : */
4853 192 : for (i = tail; i < head; i++)
4854 : {
4855 36 : if (KnownAssignedXidsValid[i])
4856 36 : break;
4857 : }
4858 192 : if (i >= head)
4859 : {
4860 : /* Array is empty, so we can reset both pointers */
4861 156 : pArray->headKnownAssignedXids = 0;
4862 156 : pArray->tailKnownAssignedXids = 0;
4863 : }
4864 : else
4865 : {
4866 36 : pArray->tailKnownAssignedXids = i;
4867 : }
4868 :
4869 : /* Opportunistically compress the array */
4870 192 : KnownAssignedXidsCompress(false);
4871 : }
4872 :
4873 : /*
4874 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
4875 : * We filter out anything >= xmax.
4876 : *
4877 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
4878 : * that array is large enough.
4879 : *
4880 : * Caller must hold ProcArrayLock in (at least) shared mode.
4881 : */
4882 : static int
4883 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
4884 : {
4885 0 : TransactionId xtmp = InvalidTransactionId;
4886 :
4887 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
4888 : }
4889 :
4890 : /*
4891 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
4892 : * we reduce *xmin to the lowest xid value seen if not already lower.
4893 : *
4894 : * Caller must hold ProcArrayLock in (at least) shared mode.
4895 : */
4896 : static int
4897 530 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
4898 : TransactionId xmax)
4899 : {
4900 530 : int count = 0;
4901 : int head,
4902 : tail;
4903 : int i;
4904 :
4905 : /*
4906 : * Fetch head just once, since it may change while we loop. We can stop
4907 : * once we reach the initially seen head, since we are certain that an xid
4908 : * cannot enter and then leave the array while we hold ProcArrayLock. We
4909 : * might miss newly-added xids, but they should be >= xmax so irrelevant
4910 : * anyway.
4911 : *
4912 : * Must take spinlock to ensure we see up-to-date array contents.
4913 : */
4914 530 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
4915 530 : tail = procArray->tailKnownAssignedXids;
4916 530 : head = procArray->headKnownAssignedXids;
4917 530 : SpinLockRelease(&procArray->known_assigned_xids_lck);
4918 :
4919 554 : for (i = tail; i < head; i++)
4920 : {
4921 : /* Skip any gaps in the array */
4922 110 : if (KnownAssignedXidsValid[i])
4923 : {
4924 92 : TransactionId knownXid = KnownAssignedXids[i];
4925 :
4926 : /*
4927 : * Update xmin if required. Only the first XID need be checked,
4928 : * since the array is sorted.
4929 : */
4930 184 : if (count == 0 &&
4931 92 : TransactionIdPrecedes(knownXid, *xmin))
4932 6 : *xmin = knownXid;
4933 :
4934 : /*
4935 : * Filter out anything >= xmax, again relying on sorted property
4936 : * of array.
4937 : */
4938 184 : if (TransactionIdIsValid(xmax) &&
4939 92 : TransactionIdFollowsOrEquals(knownXid, xmax))
4940 86 : break;
4941 :
4942 : /* Add knownXid into output array */
4943 6 : xarray[count++] = knownXid;
4944 : }
4945 : }
4946 :
4947 530 : return count;
4948 : }
4949 :
4950 : /*
4951 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
4952 : * if nothing there.
4953 : */
4954 : static TransactionId
4955 88 : KnownAssignedXidsGetOldestXmin(void)
4956 : {
4957 : int head,
4958 : tail;
4959 : int i;
4960 :
4961 : /*
4962 : * Fetch head just once, since it may change while we loop.
4963 : */
4964 88 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
4965 88 : tail = procArray->tailKnownAssignedXids;
4966 88 : head = procArray->headKnownAssignedXids;
4967 88 : SpinLockRelease(&procArray->known_assigned_xids_lck);
4968 :
4969 88 : for (i = tail; i < head; i++)
4970 : {
4971 : /* Skip any gaps in the array */
4972 6 : if (KnownAssignedXidsValid[i])
4973 6 : return KnownAssignedXids[i];
4974 : }
4975 :
4976 82 : return InvalidTransactionId;
4977 : }
4978 :
4979 : /*
4980 : * Display KnownAssignedXids to provide debug trail
4981 : *
4982 : * Currently this is only called within startup process, so we need no
4983 : * special locking.
4984 : *
4985 : * Note this is pretty expensive, and much of the expense will be incurred
4986 : * even if the elog message will get discarded. It's not currently called
4987 : * in any performance-critical places, however, so no need to be tenser.
4988 : */
4989 : static void
4990 90 : KnownAssignedXidsDisplay(int trace_level)
4991 : {
4992 90 : ProcArrayStruct *pArray = procArray;
4993 : StringInfoData buf;
4994 : int head,
4995 : tail,
4996 : i;
4997 90 : int nxids = 0;
4998 :
4999 90 : tail = pArray->tailKnownAssignedXids;
5000 90 : head = pArray->headKnownAssignedXids;
5001 :
5002 90 : initStringInfo(&buf);
5003 :
5004 98 : for (i = tail; i < head; i++)
5005 : {
5006 8 : if (KnownAssignedXidsValid[i])
5007 : {
5008 8 : nxids++;
5009 8 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5010 : }
5011 : }
5012 :
5013 90 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5014 : nxids,
5015 : pArray->numKnownAssignedXids,
5016 : pArray->tailKnownAssignedXids,
5017 : pArray->headKnownAssignedXids,
5018 : buf.data);
5019 :
5020 90 : pfree(buf.data);
5021 90 : }
5022 :
5023 : /*
5024 : * KnownAssignedXidsReset
5025 : * Resets KnownAssignedXids to be empty
5026 : */
5027 : static void
5028 0 : KnownAssignedXidsReset(void)
5029 : {
5030 0 : ProcArrayStruct *pArray = procArray;
5031 :
5032 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5033 :
5034 0 : pArray->numKnownAssignedXids = 0;
5035 0 : pArray->tailKnownAssignedXids = 0;
5036 0 : pArray->headKnownAssignedXids = 0;
5037 :
5038 0 : LWLockRelease(ProcArrayLock);
5039 0 : }
|
