Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * procarray.c
4 : * POSTGRES process array code.
5 : *
6 : *
7 : * This module maintains arrays of PGPROC substructures, as well as associated
8 : * arrays in ProcGlobal, for all active backends. Although there are several
9 : * uses for this, the principal one is as a means of determining the set of
10 : * currently running transactions.
11 : *
12 : * Because of various subtle race conditions it is critical that a backend
13 : * hold the correct locks while setting or clearing its xid (in
14 : * ProcGlobal->xids[]/MyProc->xid). See notes in
15 : * src/backend/access/transam/README.
16 : *
17 : * The process arrays now also include structures representing prepared
18 : * transactions. The xid and subxids fields of these are valid, as are the
19 : * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 : * at need by checking for pid == 0.
21 : *
22 : * During hot standby, we also keep a list of XIDs representing transactions
23 : * that are known to be running on the primary (or more precisely, were running
24 : * as of the current point in the WAL stream). This list is kept in the
25 : * KnownAssignedXids array, and is updated by watching the sequence of
26 : * arriving XIDs. This is necessary because if we leave those XIDs out of
27 : * snapshots taken for standby queries, then they will appear to be already
28 : * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 : * array represents standby processes, which by definition are not running
30 : * transactions that have XIDs.
31 : *
32 : * It is perhaps possible for a backend on the primary to terminate without
33 : * writing an abort record for its transaction. While that shouldn't really
34 : * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 : * ourselves by pruning the array when a valid list of running XIDs arrives.
36 : *
37 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/subtrans.h"
51 : #include "access/transam.h"
52 : #include "access/twophase.h"
53 : #include "access/xact.h"
54 : #include "access/xlogutils.h"
55 : #include "catalog/catalog.h"
56 : #include "catalog/pg_authid.h"
57 : #include "miscadmin.h"
58 : #include "pgstat.h"
59 : #include "postmaster/bgworker.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/procsignal.h"
64 : #include "storage/subsystems.h"
65 : #include "utils/acl.h"
66 : #include "utils/builtins.h"
67 : #include "utils/injection_point.h"
68 : #include "utils/lsyscache.h"
69 : #include "utils/rel.h"
70 : #include "utils/snapmgr.h"
71 : #include "utils/wait_event.h"
72 :
73 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
74 :
75 : /* Our shared memory area */
76 : typedef struct ProcArrayStruct
77 : {
78 : int numProcs; /* number of valid procs entries */
79 : int maxProcs; /* allocated size of procs array */
80 :
81 : /*
82 : * Known assigned XIDs handling
83 : */
84 : int maxKnownAssignedXids; /* allocated size of array */
85 : int numKnownAssignedXids; /* current # of valid entries */
86 : int tailKnownAssignedXids; /* index of oldest valid element */
87 : int headKnownAssignedXids; /* index of newest element, + 1 */
88 :
89 : /*
90 : * Highest subxid that has been removed from KnownAssignedXids array to
91 : * prevent overflow; or InvalidTransactionId if none. We track this for
92 : * similar reasons to tracking overflowing cached subxids in PGPROC
93 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
94 : * lock to read it.
95 : */
96 : TransactionId lastOverflowedXid;
97 :
98 : /* oldest xmin of any replication slot */
99 : TransactionId replication_slot_xmin;
100 : /* oldest catalog xmin of any replication slot */
101 : TransactionId replication_slot_catalog_xmin;
102 :
103 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
104 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
105 : } ProcArrayStruct;
106 :
107 : static void ProcArrayShmemRequest(void *arg);
108 : static void ProcArrayShmemInit(void *arg);
109 : static void ProcArrayShmemAttach(void *arg);
110 :
111 : static ProcArrayStruct *procArray;
112 :
113 : const struct ShmemCallbacks ProcArrayShmemCallbacks = {
114 : .request_fn = ProcArrayShmemRequest,
115 : .init_fn = ProcArrayShmemInit,
116 : .attach_fn = ProcArrayShmemAttach,
117 : };
118 :
119 : /*
120 : * State for the GlobalVisTest* family of functions. Those functions can
121 : * e.g. be used to decide if a deleted row can be removed without violating
122 : * MVCC semantics: If the deleted row's xmax is not considered to be running
123 : * by anyone, the row can be removed.
124 : *
125 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
126 : * cutoff XID while building a snapshot (looking at the frequently changing
127 : * xmins scales badly). Instead we compute two boundaries while building the
128 : * snapshot:
129 : *
130 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
131 : * definitely_needed are definitely still visible.
132 : *
133 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
134 : * definitely be removed
135 : *
136 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
137 : * && XID < definitely_needed), the boundaries can be recomputed (using
138 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
139 : * maintaining an accurate value all the time.
140 : *
141 : * As it is not cheap to compute accurate boundaries, we limit the number of
142 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
143 : *
144 : *
145 : * There are three backend lifetime instances of this struct, optimized for
146 : * different types of relations. As e.g. a normal user defined table in one
147 : * database is inaccessible to backends connected to another database, a test
148 : * specific to a relation can be more aggressive than a test for a shared
149 : * relation. Currently we track four different states:
150 : *
151 : * 1) GlobalVisSharedRels, which only considers an XID's
152 : * effects visible-to-everyone if neither snapshots in any database, nor a
153 : * replication slot's xmin, nor a replication slot's catalog_xmin might
154 : * still consider XID as running.
155 : *
156 : * 2) GlobalVisCatalogRels, which only considers an XID's
157 : * effects visible-to-everyone if neither snapshots in the current
158 : * database, nor a replication slot's xmin, nor a replication slot's
159 : * catalog_xmin might still consider XID as running.
160 : *
161 : * I.e. the difference to GlobalVisSharedRels is that
162 : * snapshot in other databases are ignored.
163 : *
164 : * 3) GlobalVisDataRels, which only considers an XID's
165 : * effects visible-to-everyone if neither snapshots in the current
166 : * database, nor a replication slot's xmin consider XID as running.
167 : *
168 : * I.e. the difference to GlobalVisCatalogRels is that
169 : * replication slot's catalog_xmin is not taken into account.
170 : *
171 : * 4) GlobalVisTempRels, which only considers the current session, as temp
172 : * tables are not visible to other sessions.
173 : *
174 : * GlobalVisTestFor(relation) returns the appropriate state
175 : * for the relation.
176 : *
177 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
178 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
179 : * prevent maybe_needed to become old enough after the GetSnapshotData()
180 : * call.
181 : *
182 : * The typedef is in the header.
183 : */
184 : struct GlobalVisState
185 : {
186 : /* XIDs >= are considered running by some backend */
187 : FullTransactionId definitely_needed;
188 :
189 : /* XIDs < are not considered to be running by any backend */
190 : FullTransactionId maybe_needed;
191 : };
192 :
193 : /*
194 : * Result of ComputeXidHorizons().
195 : */
196 : typedef struct ComputeXidHorizonsResult
197 : {
198 : /*
199 : * The value of TransamVariables->latestCompletedXid when
200 : * ComputeXidHorizons() held ProcArrayLock.
201 : */
202 : FullTransactionId latest_completed;
203 :
204 : /*
205 : * The same for procArray->replication_slot_xmin and
206 : * procArray->replication_slot_catalog_xmin.
207 : */
208 : TransactionId slot_xmin;
209 : TransactionId slot_catalog_xmin;
210 :
211 : /*
212 : * Oldest xid that any backend might still consider running. This needs to
213 : * include processes running VACUUM, in contrast to the normal visibility
214 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
215 : * determining visibility, but doesn't care about rows above its xmin to
216 : * be removed.
217 : *
218 : * This likely should only be needed to determine whether pg_subtrans can
219 : * be truncated. It currently includes the effects of replication slots,
220 : * for historical reasons. But that could likely be changed.
221 : */
222 : TransactionId oldest_considered_running;
223 :
224 : /*
225 : * Oldest xid for which deleted tuples need to be retained in shared
226 : * tables.
227 : *
228 : * This includes the effects of replication slots. If that's not desired,
229 : * look at shared_oldest_nonremovable_raw;
230 : */
231 : TransactionId shared_oldest_nonremovable;
232 :
233 : /*
234 : * Oldest xid that may be necessary to retain in shared tables. This is
235 : * the same as shared_oldest_nonremovable, except that is not affected by
236 : * replication slot's catalog_xmin.
237 : *
238 : * This is mainly useful to be able to send the catalog_xmin to upstream
239 : * streaming replication servers via hot_standby_feedback, so they can
240 : * apply the limit only when accessing catalog tables.
241 : */
242 : TransactionId shared_oldest_nonremovable_raw;
243 :
244 : /*
245 : * Oldest xid for which deleted tuples need to be retained in non-shared
246 : * catalog tables.
247 : */
248 : TransactionId catalog_oldest_nonremovable;
249 :
250 : /*
251 : * Oldest xid for which deleted tuples need to be retained in normal user
252 : * defined tables.
253 : */
254 : TransactionId data_oldest_nonremovable;
255 :
256 : /*
257 : * Oldest xid for which deleted tuples need to be retained in this
258 : * session's temporary tables.
259 : */
260 : TransactionId temp_oldest_nonremovable;
261 : } ComputeXidHorizonsResult;
262 :
263 : /*
264 : * Return value for GlobalVisHorizonKindForRel().
265 : */
266 : typedef enum GlobalVisHorizonKind
267 : {
268 : VISHORIZON_SHARED,
269 : VISHORIZON_CATALOG,
270 : VISHORIZON_DATA,
271 : VISHORIZON_TEMP,
272 : } GlobalVisHorizonKind;
273 :
274 : /*
275 : * Reason codes for KnownAssignedXidsCompress().
276 : */
277 : typedef enum KAXCompressReason
278 : {
279 : KAX_NO_SPACE, /* need to free up space at array end */
280 : KAX_PRUNE, /* we just pruned old entries */
281 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
282 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
283 : } KAXCompressReason;
284 :
285 : static PGPROC *allProcs;
286 :
287 : /*
288 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
289 : */
290 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
291 :
292 : /*
293 : * Bookkeeping for tracking emulated transactions in recovery
294 : */
295 :
296 : static TransactionId *KnownAssignedXids;
297 :
298 : static bool *KnownAssignedXidsValid;
299 :
300 : static TransactionId latestObservedXid = InvalidTransactionId;
301 :
302 : /*
303 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
304 : * the highest xid that might still be running that we don't have in
305 : * KnownAssignedXids.
306 : */
307 : static TransactionId standbySnapshotPendingXmin;
308 :
309 : /*
310 : * State for visibility checks on different types of relations. See struct
311 : * GlobalVisState for details. As shared, catalog, normal and temporary
312 : * relations can have different horizons, one such state exists for each.
313 : */
314 : static GlobalVisState GlobalVisSharedRels;
315 : static GlobalVisState GlobalVisCatalogRels;
316 : static GlobalVisState GlobalVisDataRels;
317 : static GlobalVisState GlobalVisTempRels;
318 :
319 : /*
320 : * This backend's RecentXmin at the last time the accurate xmin horizon was
321 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
322 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
323 : */
324 : static TransactionId ComputeXidHorizonsResultLastXmin;
325 :
326 : #ifdef XIDCACHE_DEBUG
327 :
328 : /* counters for XidCache measurement */
329 : static long xc_by_recent_xmin = 0;
330 : static long xc_by_known_xact = 0;
331 : static long xc_by_my_xact = 0;
332 : static long xc_by_latest_xid = 0;
333 : static long xc_by_main_xid = 0;
334 : static long xc_by_child_xid = 0;
335 : static long xc_by_known_assigned = 0;
336 : static long xc_no_overflow = 0;
337 : static long xc_slow_answer = 0;
338 :
339 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
340 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
341 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
342 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
343 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
344 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
345 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
346 : #define xc_no_overflow_inc() (xc_no_overflow++)
347 : #define xc_slow_answer_inc() (xc_slow_answer++)
348 :
349 : static void DisplayXidCache(void);
350 : #else /* !XIDCACHE_DEBUG */
351 :
352 : #define xc_by_recent_xmin_inc() ((void) 0)
353 : #define xc_by_known_xact_inc() ((void) 0)
354 : #define xc_by_my_xact_inc() ((void) 0)
355 : #define xc_by_latest_xid_inc() ((void) 0)
356 : #define xc_by_main_xid_inc() ((void) 0)
357 : #define xc_by_child_xid_inc() ((void) 0)
358 : #define xc_by_known_assigned_inc() ((void) 0)
359 : #define xc_no_overflow_inc() ((void) 0)
360 : #define xc_slow_answer_inc() ((void) 0)
361 : #endif /* XIDCACHE_DEBUG */
362 :
363 : /* Primitives for KnownAssignedXids array handling for standby */
364 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
365 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
366 : bool exclusive_lock);
367 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
368 : static bool KnownAssignedXidExists(TransactionId xid);
369 : static void KnownAssignedXidsRemove(TransactionId xid);
370 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
371 : TransactionId *subxids);
372 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
373 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
374 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
375 : TransactionId *xmin,
376 : TransactionId xmax);
377 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
378 : static void KnownAssignedXidsDisplay(int trace_level);
379 : static void KnownAssignedXidsReset(void);
380 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
381 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
382 : static void MaintainLatestCompletedXid(TransactionId latestXid);
383 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
384 :
385 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
386 : TransactionId xid);
387 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
388 :
389 : /*
390 : * Register the shared PGPROC array during postmaster startup.
391 : */
392 : static void
393 1255 : ProcArrayShmemRequest(void *arg)
394 : {
395 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
396 :
397 : /*
398 : * During Hot Standby processing we have a data structure called
399 : * KnownAssignedXids, created in shared memory. Local data structures are
400 : * also created in various backends during GetSnapshotData(),
401 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
402 : * main structures created in those functions must be identically sized,
403 : * since we may at times copy the whole of the data structures around. We
404 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
405 : *
406 : * Ideally we'd only create this structure if we were actually doing hot
407 : * standby in the current run, but we don't know that yet at the time
408 : * shared memory is being set up.
409 : */
410 : #define TOTAL_MAX_CACHED_SUBXIDS \
411 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
412 :
413 1255 : if (EnableHotStandby)
414 : {
415 1249 : ShmemRequestStruct(.name = "KnownAssignedXids",
416 : .size = mul_size(sizeof(TransactionId), TOTAL_MAX_CACHED_SUBXIDS),
417 : .ptr = (void **) &KnownAssignedXids,
418 : );
419 :
420 1249 : ShmemRequestStruct(.name = "KnownAssignedXidsValid",
421 : .size = mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
422 : .ptr = (void **) &KnownAssignedXidsValid,
423 : );
424 : }
425 :
426 : /* Register the ProcArray shared structure */
427 1255 : ShmemRequestStruct(.name = "Proc Array",
428 : .size = add_size(offsetof(ProcArrayStruct, pgprocnos),
429 : mul_size(sizeof(int), PROCARRAY_MAXPROCS)),
430 : .ptr = (void **) &procArray,
431 : );
432 1255 : }
433 :
434 : /*
435 : * Initialize the shared PGPROC array during postmaster startup.
436 : */
437 : static void
438 1252 : ProcArrayShmemInit(void *arg)
439 : {
440 1252 : procArray->numProcs = 0;
441 1252 : procArray->maxProcs = PROCARRAY_MAXPROCS;
442 1252 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
443 1252 : procArray->numKnownAssignedXids = 0;
444 1252 : procArray->tailKnownAssignedXids = 0;
445 1252 : procArray->headKnownAssignedXids = 0;
446 1252 : procArray->lastOverflowedXid = InvalidTransactionId;
447 1252 : procArray->replication_slot_xmin = InvalidTransactionId;
448 1252 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
449 1252 : TransamVariables->xactCompletionCount = 1;
450 :
451 1252 : allProcs = ProcGlobal->allProcs;
452 1252 : }
453 :
454 : static void
455 0 : ProcArrayShmemAttach(void *arg)
456 : {
457 0 : allProcs = ProcGlobal->allProcs;
458 0 : }
459 :
460 : /*
461 : * Add the specified PGPROC to the shared array.
462 : */
463 : void
464 20569 : ProcArrayAdd(PGPROC *proc)
465 : {
466 20569 : int pgprocno = GetNumberFromPGProc(proc);
467 20569 : ProcArrayStruct *arrayP = procArray;
468 : int index;
469 : int movecount;
470 :
471 : /* See ProcGlobal comment explaining why both locks are held */
472 20569 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
473 20569 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
474 :
475 20569 : if (arrayP->numProcs >= arrayP->maxProcs)
476 : {
477 : /*
478 : * Oops, no room. (This really shouldn't happen, since there is a
479 : * fixed supply of PGPROC structs too, and so we should have failed
480 : * earlier.)
481 : */
482 0 : ereport(FATAL,
483 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
484 : errmsg("sorry, too many clients already")));
485 : }
486 :
487 : /*
488 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
489 : * locality of references much better. This is useful while traversing the
490 : * ProcArray because there is an increased likelihood of finding the next
491 : * PGPROC structure in the cache.
492 : *
493 : * Since the occurrence of adding/removing a proc is much lower than the
494 : * access to the ProcArray itself, the overhead should be marginal
495 : */
496 52770 : for (index = 0; index < arrayP->numProcs; index++)
497 : {
498 47146 : int this_procno = arrayP->pgprocnos[index];
499 :
500 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
501 : Assert(allProcs[this_procno].pgxactoff == index);
502 :
503 : /* If we have found our right position in the array, break */
504 47146 : if (this_procno > pgprocno)
505 14945 : break;
506 : }
507 :
508 20569 : movecount = arrayP->numProcs - index;
509 20569 : memmove(&arrayP->pgprocnos[index + 1],
510 20569 : &arrayP->pgprocnos[index],
511 : movecount * sizeof(*arrayP->pgprocnos));
512 20569 : memmove(&ProcGlobal->xids[index + 1],
513 20569 : &ProcGlobal->xids[index],
514 : movecount * sizeof(*ProcGlobal->xids));
515 20569 : memmove(&ProcGlobal->subxidStates[index + 1],
516 20569 : &ProcGlobal->subxidStates[index],
517 : movecount * sizeof(*ProcGlobal->subxidStates));
518 20569 : memmove(&ProcGlobal->statusFlags[index + 1],
519 20569 : &ProcGlobal->statusFlags[index],
520 : movecount * sizeof(*ProcGlobal->statusFlags));
521 :
522 20569 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
523 20569 : proc->pgxactoff = index;
524 20569 : ProcGlobal->xids[index] = proc->xid;
525 20569 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
526 20569 : ProcGlobal->statusFlags[index] = proc->statusFlags;
527 :
528 20569 : arrayP->numProcs++;
529 :
530 : /* adjust pgxactoff for all following PGPROCs */
531 20569 : index++;
532 54735 : for (; index < arrayP->numProcs; index++)
533 : {
534 34166 : int procno = arrayP->pgprocnos[index];
535 :
536 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
537 : Assert(allProcs[procno].pgxactoff == index - 1);
538 :
539 34166 : allProcs[procno].pgxactoff = index;
540 : }
541 :
542 : /*
543 : * Release in reversed acquisition order, to reduce frequency of having to
544 : * wait for XidGenLock while holding ProcArrayLock.
545 : */
546 20569 : LWLockRelease(XidGenLock);
547 20569 : LWLockRelease(ProcArrayLock);
548 20569 : }
549 :
550 : /*
551 : * Remove the specified PGPROC from the shared array.
552 : *
553 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
554 : * array, and thus causing it to appear as "not running" anymore. In this
555 : * case we must advance latestCompletedXid. (This is essentially the same
556 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
557 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
558 : * twophase.c depends on the latter.)
559 : */
560 : void
561 20541 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
562 : {
563 20541 : ProcArrayStruct *arrayP = procArray;
564 : int myoff;
565 : int movecount;
566 :
567 : #ifdef XIDCACHE_DEBUG
568 : /* dump stats at backend shutdown, but not prepared-xact end */
569 : if (proc->pid != 0)
570 : DisplayXidCache();
571 : #endif
572 :
573 : /* See ProcGlobal comment explaining why both locks are held */
574 20541 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
575 20541 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
576 :
577 20541 : myoff = proc->pgxactoff;
578 :
579 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
580 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
581 :
582 20541 : if (TransactionIdIsValid(latestXid))
583 : {
584 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
585 :
586 : /* Advance global latestCompletedXid while holding the lock */
587 324 : MaintainLatestCompletedXid(latestXid);
588 :
589 : /* Same with xactCompletionCount */
590 324 : TransamVariables->xactCompletionCount++;
591 :
592 324 : ProcGlobal->xids[myoff] = InvalidTransactionId;
593 324 : ProcGlobal->subxidStates[myoff].overflowed = false;
594 324 : ProcGlobal->subxidStates[myoff].count = 0;
595 : }
596 : else
597 : {
598 : /* Shouldn't be trying to remove a live transaction here */
599 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
600 : }
601 :
602 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
603 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
604 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
605 :
606 20541 : ProcGlobal->statusFlags[myoff] = 0;
607 :
608 : /* Keep the PGPROC array sorted. See notes above */
609 20541 : movecount = arrayP->numProcs - myoff - 1;
610 20541 : memmove(&arrayP->pgprocnos[myoff],
611 20541 : &arrayP->pgprocnos[myoff + 1],
612 : movecount * sizeof(*arrayP->pgprocnos));
613 20541 : memmove(&ProcGlobal->xids[myoff],
614 20541 : &ProcGlobal->xids[myoff + 1],
615 : movecount * sizeof(*ProcGlobal->xids));
616 20541 : memmove(&ProcGlobal->subxidStates[myoff],
617 20541 : &ProcGlobal->subxidStates[myoff + 1],
618 : movecount * sizeof(*ProcGlobal->subxidStates));
619 20541 : memmove(&ProcGlobal->statusFlags[myoff],
620 20541 : &ProcGlobal->statusFlags[myoff + 1],
621 : movecount * sizeof(*ProcGlobal->statusFlags));
622 :
623 20541 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
624 20541 : arrayP->numProcs--;
625 :
626 : /*
627 : * Adjust pgxactoff of following procs for removed PGPROC (note that
628 : * numProcs already has been decremented).
629 : */
630 58979 : for (int index = myoff; index < arrayP->numProcs; index++)
631 : {
632 38438 : int procno = arrayP->pgprocnos[index];
633 :
634 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
635 : Assert(allProcs[procno].pgxactoff - 1 == index);
636 :
637 38438 : allProcs[procno].pgxactoff = index;
638 : }
639 :
640 : /*
641 : * Release in reversed acquisition order, to reduce frequency of having to
642 : * wait for XidGenLock while holding ProcArrayLock.
643 : */
644 20541 : LWLockRelease(XidGenLock);
645 20541 : LWLockRelease(ProcArrayLock);
646 20541 : }
647 :
648 :
649 : /*
650 : * ProcArrayEndTransaction -- mark a transaction as no longer running
651 : *
652 : * This is used interchangeably for commit and abort cases. The transaction
653 : * commit/abort must already be reported to WAL and pg_xact.
654 : *
655 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
656 : * latestXid is the latest Xid among the transaction's main XID and
657 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
658 : * the caller to pass latestXid, instead of computing it from the PGPROC's
659 : * contents, because the subxid information in the PGPROC might be
660 : * incomplete.)
661 : */
662 : void
663 651558 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
664 : {
665 651558 : if (TransactionIdIsValid(latestXid))
666 : {
667 : /*
668 : * We must lock ProcArrayLock while clearing our advertised XID, so
669 : * that we do not exit the set of "running" transactions while someone
670 : * else is taking a snapshot. See discussion in
671 : * src/backend/access/transam/README.
672 : */
673 : Assert(TransactionIdIsValid(proc->xid));
674 :
675 : /*
676 : * If we can immediately acquire ProcArrayLock, we clear our own XID
677 : * and release the lock. If not, use group XID clearing to improve
678 : * efficiency.
679 : */
680 167879 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
681 : {
682 167709 : ProcArrayEndTransactionInternal(proc, latestXid);
683 167709 : LWLockRelease(ProcArrayLock);
684 : }
685 : else
686 170 : ProcArrayGroupClearXid(proc, latestXid);
687 : }
688 : else
689 : {
690 : /*
691 : * If we have no XID, we don't need to lock, since we won't affect
692 : * anyone else's calculation of a snapshot. We might change their
693 : * estimate of global xmin, but that's OK.
694 : */
695 : Assert(!TransactionIdIsValid(proc->xid));
696 : Assert(proc->subxidStatus.count == 0);
697 : Assert(!proc->subxidStatus.overflowed);
698 :
699 483679 : proc->vxid.lxid = InvalidLocalTransactionId;
700 483679 : proc->xmin = InvalidTransactionId;
701 :
702 : /* be sure this is cleared in abort */
703 483679 : proc->delayChkptFlags = 0;
704 :
705 : /* must be cleared with xid/xmin: */
706 : /* avoid unnecessarily dirtying shared cachelines */
707 483679 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
708 : {
709 : Assert(!LWLockHeldByMe(ProcArrayLock));
710 125924 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
711 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
712 125924 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
713 125924 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
714 125924 : LWLockRelease(ProcArrayLock);
715 : }
716 : }
717 651558 : }
718 :
719 : /*
720 : * Mark a write transaction as no longer running.
721 : *
722 : * We don't do any locking here; caller must handle that.
723 : */
724 : static inline void
725 167879 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
726 : {
727 167879 : int pgxactoff = proc->pgxactoff;
728 :
729 : /*
730 : * Note: we need exclusive lock here because we're going to change other
731 : * processes' PGPROC entries.
732 : */
733 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
734 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
735 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
736 :
737 167879 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
738 167879 : proc->xid = InvalidTransactionId;
739 167879 : proc->vxid.lxid = InvalidLocalTransactionId;
740 167879 : proc->xmin = InvalidTransactionId;
741 :
742 : /* be sure this is cleared in abort */
743 167879 : proc->delayChkptFlags = 0;
744 :
745 : /* must be cleared with xid/xmin: */
746 : /* avoid unnecessarily dirtying shared cachelines */
747 167879 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
748 : {
749 878 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
750 878 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
751 : }
752 :
753 : /* Clear the subtransaction-XID cache too while holding the lock */
754 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
755 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
756 167879 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
757 : {
758 616 : ProcGlobal->subxidStates[pgxactoff].count = 0;
759 616 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
760 616 : proc->subxidStatus.count = 0;
761 616 : proc->subxidStatus.overflowed = false;
762 : }
763 :
764 : /* Also advance global latestCompletedXid while holding the lock */
765 167879 : MaintainLatestCompletedXid(latestXid);
766 :
767 : /* Same with xactCompletionCount */
768 167879 : TransamVariables->xactCompletionCount++;
769 167879 : }
770 :
771 : /*
772 : * ProcArrayGroupClearXid -- group XID clearing
773 : *
774 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
775 : * commit time, add ourselves to a list of processes that need their XIDs
776 : * cleared. The first process to add itself to the list will acquire
777 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
778 : * on behalf of all group members. This avoids a great deal of contention
779 : * around ProcArrayLock when many processes are trying to commit at once,
780 : * since the lock need not be repeatedly handed off from one committing
781 : * process to the next.
782 : */
783 : static void
784 170 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
785 : {
786 170 : int pgprocno = GetNumberFromPGProc(proc);
787 170 : PROC_HDR *procglobal = ProcGlobal;
788 : uint32 nextidx;
789 : uint32 wakeidx;
790 :
791 : /* We should definitely have an XID to clear. */
792 : Assert(TransactionIdIsValid(proc->xid));
793 :
794 : /* Add ourselves to the list of processes needing a group XID clear. */
795 170 : proc->procArrayGroupMember = true;
796 170 : proc->procArrayGroupMemberXid = latestXid;
797 170 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
798 : while (true)
799 : {
800 170 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
801 :
802 170 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
803 : &nextidx,
804 : (uint32) pgprocno))
805 170 : break;
806 : }
807 :
808 : /*
809 : * If the list was not empty, the leader will clear our XID. It is
810 : * impossible to have followers without a leader because the first process
811 : * that has added itself to the list will always have nextidx as
812 : * INVALID_PROC_NUMBER.
813 : */
814 170 : if (nextidx != INVALID_PROC_NUMBER)
815 : {
816 8 : int extraWaits = 0;
817 :
818 : /* Sleep until the leader clears our XID. */
819 8 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
820 : for (;;)
821 : {
822 : /* acts as a read barrier */
823 8 : PGSemaphoreLock(proc->sem);
824 8 : if (!proc->procArrayGroupMember)
825 8 : break;
826 0 : extraWaits++;
827 : }
828 8 : pgstat_report_wait_end();
829 :
830 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
831 :
832 : /* Fix semaphore count for any absorbed wakeups */
833 8 : while (extraWaits-- > 0)
834 0 : PGSemaphoreUnlock(proc->sem);
835 8 : return;
836 : }
837 :
838 : /* We are the leader. Acquire the lock on behalf of everyone. */
839 162 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
840 :
841 : /*
842 : * Now that we've got the lock, clear the list of processes waiting for
843 : * group XID clearing, saving a pointer to the head of the list. Trying
844 : * to pop elements one at a time could lead to an ABA problem.
845 : */
846 162 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
847 : INVALID_PROC_NUMBER);
848 :
849 : /* Remember head of list so we can perform wakeups after dropping lock. */
850 162 : wakeidx = nextidx;
851 :
852 : /* Walk the list and clear all XIDs. */
853 332 : while (nextidx != INVALID_PROC_NUMBER)
854 : {
855 170 : PGPROC *nextproc = &allProcs[nextidx];
856 :
857 170 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
858 :
859 : /* Move to next proc in list. */
860 170 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
861 : }
862 :
863 : /* We're done with the lock now. */
864 162 : LWLockRelease(ProcArrayLock);
865 :
866 : /*
867 : * Now that we've released the lock, go back and wake everybody up. We
868 : * don't do this under the lock so as to keep lock hold times to a
869 : * minimum. The system calls we need to perform to wake other processes
870 : * up are probably much slower than the simple memory writes we did while
871 : * holding the lock.
872 : */
873 332 : while (wakeidx != INVALID_PROC_NUMBER)
874 : {
875 170 : PGPROC *nextproc = &allProcs[wakeidx];
876 :
877 170 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
878 170 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
879 :
880 : /* ensure all previous writes are visible before follower continues. */
881 170 : pg_write_barrier();
882 :
883 170 : nextproc->procArrayGroupMember = false;
884 :
885 170 : if (nextproc != MyProc)
886 8 : PGSemaphoreUnlock(nextproc->sem);
887 : }
888 : }
889 :
890 : /*
891 : * ProcArrayClearTransaction -- clear the transaction fields
892 : *
893 : * This is used after successfully preparing a 2-phase transaction. We are
894 : * not actually reporting the transaction's XID as no longer running --- it
895 : * will still appear as running because the 2PC's gxact is in the ProcArray
896 : * too. We just have to clear out our own PGPROC.
897 : */
898 : void
899 319 : ProcArrayClearTransaction(PGPROC *proc)
900 : {
901 : int pgxactoff;
902 :
903 : /*
904 : * Currently we need to lock ProcArrayLock exclusively here, as we
905 : * increment xactCompletionCount below. We also need it at least in shared
906 : * mode for pgproc->pgxactoff to stay the same below.
907 : *
908 : * We could however, as this action does not actually change anyone's view
909 : * of the set of running XIDs (our entry is duplicate with the gxact that
910 : * has already been inserted into the ProcArray), lower the lock level to
911 : * shared if we were to make xactCompletionCount an atomic variable. But
912 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
913 : * enough for this not to be the bottleneck. If it ever becomes a
914 : * bottleneck it may also be worth considering to combine this with the
915 : * subsequent ProcArrayRemove()
916 : */
917 319 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
918 :
919 319 : pgxactoff = proc->pgxactoff;
920 :
921 319 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
922 319 : proc->xid = InvalidTransactionId;
923 :
924 319 : proc->vxid.lxid = InvalidLocalTransactionId;
925 319 : proc->xmin = InvalidTransactionId;
926 :
927 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
928 : Assert(!proc->delayChkptFlags);
929 :
930 : /*
931 : * Need to increment completion count even though transaction hasn't
932 : * really committed yet. The reason for that is that GetSnapshotData()
933 : * omits the xid of the current transaction, thus without the increment we
934 : * otherwise could end up reusing the snapshot later. Which would be bad,
935 : * because it might not count the prepared transaction as running.
936 : */
937 319 : TransamVariables->xactCompletionCount++;
938 :
939 : /* Clear the subtransaction-XID cache too */
940 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
941 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
942 319 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
943 : {
944 107 : ProcGlobal->subxidStates[pgxactoff].count = 0;
945 107 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
946 107 : proc->subxidStatus.count = 0;
947 107 : proc->subxidStatus.overflowed = false;
948 : }
949 :
950 319 : LWLockRelease(ProcArrayLock);
951 319 : }
952 :
953 : /*
954 : * Update TransamVariables->latestCompletedXid to point to latestXid if
955 : * currently older.
956 : */
957 : static void
958 169059 : MaintainLatestCompletedXid(TransactionId latestXid)
959 : {
960 169059 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
961 :
962 : Assert(FullTransactionIdIsValid(cur_latest));
963 : Assert(!RecoveryInProgress());
964 : Assert(LWLockHeldByMe(ProcArrayLock));
965 :
966 169059 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
967 : {
968 152529 : TransamVariables->latestCompletedXid =
969 152529 : FullXidRelativeTo(cur_latest, latestXid);
970 : }
971 :
972 : Assert(IsBootstrapProcessingMode() ||
973 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
974 169059 : }
975 :
976 : /*
977 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
978 : */
979 : static void
980 24676 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
981 : {
982 24676 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
983 : FullTransactionId rel;
984 :
985 : Assert(AmStartupProcess() || !IsUnderPostmaster);
986 : Assert(LWLockHeldByMe(ProcArrayLock));
987 :
988 : /*
989 : * Need a FullTransactionId to compare latestXid with. Can't rely on
990 : * latestCompletedXid to be initialized in recovery. But in recovery it's
991 : * safe to access nextXid without a lock for the startup process.
992 : */
993 24676 : rel = TransamVariables->nextXid;
994 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
995 :
996 49230 : if (!FullTransactionIdIsValid(cur_latest) ||
997 24554 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
998 : {
999 19080 : TransamVariables->latestCompletedXid =
1000 19080 : FullXidRelativeTo(rel, latestXid);
1001 : }
1002 :
1003 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1004 24676 : }
1005 :
1006 : /*
1007 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1008 : *
1009 : * Remember up to where the startup process initialized the CLOG and subtrans
1010 : * so we can ensure it's initialized gaplessly up to the point where necessary
1011 : * while in recovery.
1012 : */
1013 : void
1014 122 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1015 : {
1016 : Assert(standbyState == STANDBY_INITIALIZED);
1017 : Assert(TransactionIdIsNormal(initializedUptoXID));
1018 :
1019 : /*
1020 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1021 : * to, so we can extend it from that point onwards in
1022 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1023 : * ProcArrayApplyRecoveryInfo().
1024 : */
1025 122 : latestObservedXid = initializedUptoXID;
1026 122 : TransactionIdRetreat(latestObservedXid);
1027 122 : }
1028 :
1029 : /*
1030 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1031 : *
1032 : * Takes us through 3 states: Initialized, Pending and Ready.
1033 : * Normal case is to go all the way to Ready straight away, though there
1034 : * are atypical cases where we need to take it in steps.
1035 : *
1036 : * Use the data about running transactions on the primary to create the initial
1037 : * state of KnownAssignedXids. We also use these records to regularly prune
1038 : * KnownAssignedXids because we know it is possible that some transactions
1039 : * with FATAL errors fail to write abort records, which could cause eventual
1040 : * overflow.
1041 : *
1042 : * See comments for LogStandbySnapshot().
1043 : */
1044 : void
1045 840 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1046 : {
1047 : TransactionId *xids;
1048 : TransactionId advanceNextXid;
1049 : int nxids;
1050 : int i;
1051 :
1052 : Assert(standbyState >= STANDBY_INITIALIZED);
1053 : Assert(TransactionIdIsValid(running->nextXid));
1054 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1055 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1056 :
1057 : /*
1058 : * Remove stale transactions, if any.
1059 : */
1060 840 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1061 :
1062 : /*
1063 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1064 : * because we will need it up to date for accessing two-phase transactions
1065 : * in StandbyReleaseOldLocks().
1066 : */
1067 840 : advanceNextXid = running->nextXid;
1068 840 : TransactionIdRetreat(advanceNextXid);
1069 840 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1070 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1071 :
1072 : /*
1073 : * Remove stale locks, if any.
1074 : */
1075 840 : StandbyReleaseOldLocks(running->oldestRunningXid);
1076 :
1077 : /*
1078 : * If our snapshot is already valid, nothing else to do...
1079 : */
1080 840 : if (standbyState == STANDBY_SNAPSHOT_READY)
1081 718 : return;
1082 :
1083 : /*
1084 : * If our initial RunningTransactionsData had an overflowed snapshot then
1085 : * we knew we were missing some subxids from our snapshot. If we continue
1086 : * to see overflowed snapshots then we might never be able to start up, so
1087 : * we make another test to see if our snapshot is now valid. We know that
1088 : * the missing subxids are equal to or earlier than nextXid. After we
1089 : * initialise we continue to apply changes during recovery, so once the
1090 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1091 : * know that we no longer have missing information and can mark the
1092 : * snapshot as valid.
1093 : */
1094 122 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1095 : {
1096 : /*
1097 : * If the snapshot isn't overflowed or if its empty we can reset our
1098 : * pending state and use this snapshot instead.
1099 : */
1100 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1101 : {
1102 : /*
1103 : * If we have already collected known assigned xids, we need to
1104 : * throw them away before we apply the recovery snapshot.
1105 : */
1106 0 : KnownAssignedXidsReset();
1107 0 : standbyState = STANDBY_INITIALIZED;
1108 : }
1109 : else
1110 : {
1111 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1112 : running->oldestRunningXid))
1113 : {
1114 0 : standbyState = STANDBY_SNAPSHOT_READY;
1115 0 : elog(DEBUG1,
1116 : "recovery snapshots are now enabled");
1117 : }
1118 : else
1119 0 : elog(DEBUG1,
1120 : "recovery snapshot waiting for non-overflowed snapshot or "
1121 : "until oldest active xid on standby is at least %u (now %u)",
1122 : standbySnapshotPendingXmin,
1123 : running->oldestRunningXid);
1124 0 : return;
1125 : }
1126 : }
1127 :
1128 : Assert(standbyState == STANDBY_INITIALIZED);
1129 :
1130 : /*
1131 : * NB: this can be reached at least twice, so make sure new code can deal
1132 : * with that.
1133 : */
1134 :
1135 : /*
1136 : * Nobody else is running yet, but take locks anyhow
1137 : */
1138 122 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1139 :
1140 : /*
1141 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1142 : * sort them first.
1143 : *
1144 : * Some of the new xids are top-level xids and some are subtransactions.
1145 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1146 : * aren't overflowed then all xids will fit in snapshot and so we don't
1147 : * need subtrans. If we later overflow, an xid assignment record will add
1148 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1149 : * don't have enough information to correctly update subtrans anyway.
1150 : */
1151 :
1152 : /*
1153 : * Allocate a temporary array to avoid modifying the array passed as
1154 : * argument.
1155 : */
1156 122 : xids = palloc_array(TransactionId, running->xcnt + running->subxcnt);
1157 :
1158 : /*
1159 : * Add to the temp array any xids which have not already completed.
1160 : */
1161 122 : nxids = 0;
1162 127 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1163 : {
1164 5 : TransactionId xid = running->xids[i];
1165 :
1166 : /*
1167 : * The running-xacts snapshot can contain xids that were still visible
1168 : * in the procarray when the snapshot was taken, but were already
1169 : * WAL-logged as completed. They're not running anymore, so ignore
1170 : * them.
1171 : */
1172 5 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1173 0 : continue;
1174 :
1175 5 : xids[nxids++] = xid;
1176 : }
1177 :
1178 122 : if (nxids > 0)
1179 : {
1180 5 : if (procArray->numKnownAssignedXids != 0)
1181 : {
1182 0 : LWLockRelease(ProcArrayLock);
1183 0 : elog(ERROR, "KnownAssignedXids is not empty");
1184 : }
1185 :
1186 : /*
1187 : * Sort the array so that we can add them safely into
1188 : * KnownAssignedXids.
1189 : *
1190 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1191 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1192 : * come from RUNNING_XACTS, which means there are only normal XIDs
1193 : * from the same epoch, so this is safe.
1194 : */
1195 5 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1196 :
1197 : /*
1198 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1199 : * snapshot may include duplicated xids because of prepared
1200 : * transactions, so ignore them.
1201 : */
1202 10 : for (i = 0; i < nxids; i++)
1203 : {
1204 5 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1205 : {
1206 0 : elog(DEBUG1,
1207 : "found duplicated transaction %u for KnownAssignedXids insertion",
1208 : xids[i]);
1209 0 : continue;
1210 : }
1211 5 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1212 : }
1213 :
1214 5 : KnownAssignedXidsDisplay(DEBUG3);
1215 : }
1216 :
1217 122 : pfree(xids);
1218 :
1219 : /*
1220 : * latestObservedXid is at least set to the point where SUBTRANS was
1221 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1222 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1223 : * subtrans from thereon, up to nextXid - 1.
1224 : *
1225 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1226 : * because we've just added xids to the known assigned xids machinery that
1227 : * haven't gone through RecordKnownAssignedTransactionId().
1228 : */
1229 : Assert(TransactionIdIsNormal(latestObservedXid));
1230 122 : TransactionIdAdvance(latestObservedXid);
1231 244 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1232 : {
1233 0 : ExtendSUBTRANS(latestObservedXid);
1234 0 : TransactionIdAdvance(latestObservedXid);
1235 : }
1236 122 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1237 :
1238 : /* ----------
1239 : * Now we've got the running xids we need to set the global values that
1240 : * are used to track snapshots as they evolve further.
1241 : *
1242 : * - latestCompletedXid which will be the xmax for snapshots
1243 : * - lastOverflowedXid which shows whether snapshots overflow
1244 : * - nextXid
1245 : *
1246 : * If the snapshot overflowed, then we still initialise with what we know,
1247 : * but the recovery snapshot isn't fully valid yet because we know there
1248 : * are some subxids missing. We don't know the specific subxids that are
1249 : * missing, so conservatively assume the last one is latestObservedXid.
1250 : * ----------
1251 : */
1252 122 : if (running->subxid_status == SUBXIDS_MISSING)
1253 : {
1254 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1255 :
1256 0 : standbySnapshotPendingXmin = latestObservedXid;
1257 0 : procArray->lastOverflowedXid = latestObservedXid;
1258 : }
1259 : else
1260 : {
1261 122 : standbyState = STANDBY_SNAPSHOT_READY;
1262 :
1263 122 : standbySnapshotPendingXmin = InvalidTransactionId;
1264 :
1265 : /*
1266 : * If the 'xids' array didn't include all subtransactions, we have to
1267 : * mark any snapshots taken as overflowed.
1268 : */
1269 122 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1270 25 : procArray->lastOverflowedXid = latestObservedXid;
1271 : else
1272 : {
1273 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1274 97 : procArray->lastOverflowedXid = InvalidTransactionId;
1275 : }
1276 : }
1277 :
1278 : /*
1279 : * If a transaction wrote a commit record in the gap between taking and
1280 : * logging the snapshot then latestCompletedXid may already be higher than
1281 : * the value from the snapshot, so check before we use the incoming value.
1282 : * It also might not yet be set at all.
1283 : */
1284 122 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1285 :
1286 : /*
1287 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1288 : * nobody can see it yet.
1289 : */
1290 :
1291 122 : LWLockRelease(ProcArrayLock);
1292 :
1293 122 : KnownAssignedXidsDisplay(DEBUG3);
1294 122 : if (standbyState == STANDBY_SNAPSHOT_READY)
1295 122 : elog(DEBUG1, "recovery snapshots are now enabled");
1296 : else
1297 0 : elog(DEBUG1,
1298 : "recovery snapshot waiting for non-overflowed snapshot or "
1299 : "until oldest active xid on standby is at least %u (now %u)",
1300 : standbySnapshotPendingXmin,
1301 : running->oldestRunningXid);
1302 : }
1303 :
1304 : /*
1305 : * ProcArrayApplyXidAssignment
1306 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1307 : */
1308 : void
1309 21 : ProcArrayApplyXidAssignment(TransactionId topxid,
1310 : int nsubxids, TransactionId *subxids)
1311 : {
1312 : TransactionId max_xid;
1313 : int i;
1314 :
1315 : Assert(standbyState >= STANDBY_INITIALIZED);
1316 :
1317 21 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1318 :
1319 : /*
1320 : * Mark all the subtransactions as observed.
1321 : *
1322 : * NOTE: This will fail if the subxid contains too many previously
1323 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1324 : * as the code stands, because xid-assignment records should never contain
1325 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1326 : */
1327 21 : RecordKnownAssignedTransactionIds(max_xid);
1328 :
1329 : /*
1330 : * Notice that we update pg_subtrans with the top-level xid, rather than
1331 : * the parent xid. This is a difference between normal processing and
1332 : * recovery, yet is still correct in all cases. The reason is that
1333 : * subtransaction commit is not marked in clog until commit processing, so
1334 : * all aborted subtransactions have already been clearly marked in clog.
1335 : * As a result we are able to refer directly to the top-level
1336 : * transaction's state rather than skipping through all the intermediate
1337 : * states in the subtransaction tree. This should be the first time we
1338 : * have attempted to SubTransSetParent().
1339 : */
1340 1365 : for (i = 0; i < nsubxids; i++)
1341 1344 : SubTransSetParent(subxids[i], topxid);
1342 :
1343 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1344 21 : if (standbyState == STANDBY_INITIALIZED)
1345 0 : return;
1346 :
1347 : /*
1348 : * Uses same locking as transaction commit
1349 : */
1350 21 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1351 :
1352 : /*
1353 : * Remove subxids from known-assigned-xacts.
1354 : */
1355 21 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1356 :
1357 : /*
1358 : * Advance lastOverflowedXid to be at least the last of these subxids.
1359 : */
1360 21 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1361 21 : procArray->lastOverflowedXid = max_xid;
1362 :
1363 21 : LWLockRelease(ProcArrayLock);
1364 : }
1365 :
1366 : /*
1367 : * TransactionIdIsInProgress -- is given transaction running in some backend
1368 : *
1369 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1370 : * there are four possibilities for finding a running transaction:
1371 : *
1372 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1373 : * by looking at ProcGlobal->xids.
1374 : *
1375 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1376 : * We can find this out cheaply too.
1377 : *
1378 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1379 : * if the Xid is running on the primary.
1380 : *
1381 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1382 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1383 : * This is the slowest way, but sadly it has to be done always if the others
1384 : * failed, unless we see that the cached subxact sets are complete (none have
1385 : * overflowed).
1386 : *
1387 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1388 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1389 : * This buys back some concurrency (and we can't retrieve the main Xids from
1390 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1391 : */
1392 : bool
1393 21246107 : TransactionIdIsInProgress(TransactionId xid)
1394 : {
1395 : static TransactionId *xids = NULL;
1396 : static TransactionId *other_xids;
1397 : XidCacheStatus *other_subxidstates;
1398 21246107 : int nxids = 0;
1399 21246107 : ProcArrayStruct *arrayP = procArray;
1400 : TransactionId topxid;
1401 : TransactionId latestCompletedXid;
1402 : int mypgxactoff;
1403 : int numProcs;
1404 : int j;
1405 :
1406 : /*
1407 : * Don't bother checking a transaction older than RecentXmin; it could not
1408 : * possibly still be running. (Note: in particular, this guarantees that
1409 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1410 : * running.)
1411 : */
1412 21246107 : if (TransactionIdPrecedes(xid, RecentXmin))
1413 : {
1414 : xc_by_recent_xmin_inc();
1415 13133655 : return false;
1416 : }
1417 :
1418 : /*
1419 : * We may have just checked the status of this transaction, so if it is
1420 : * already known to be completed, we can fall out without any access to
1421 : * shared memory.
1422 : */
1423 8112452 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1424 : {
1425 : xc_by_known_xact_inc();
1426 1422920 : return false;
1427 : }
1428 :
1429 : /*
1430 : * Also, we can handle our own transaction (and subtransactions) without
1431 : * any access to shared memory.
1432 : */
1433 6689532 : if (TransactionIdIsCurrentTransactionId(xid))
1434 : {
1435 : xc_by_my_xact_inc();
1436 221624 : return true;
1437 : }
1438 :
1439 : /*
1440 : * If first time through, get workspace to remember main XIDs in. We
1441 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1442 : */
1443 6467908 : if (xids == NULL)
1444 : {
1445 : /*
1446 : * In hot standby mode, reserve enough space to hold all xids in the
1447 : * known-assigned list. If we later finish recovery, we no longer need
1448 : * the bigger array, but we don't bother to shrink it.
1449 : */
1450 1752 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1451 :
1452 1752 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1453 1752 : if (xids == NULL)
1454 0 : ereport(ERROR,
1455 : (errcode(ERRCODE_OUT_OF_MEMORY),
1456 : errmsg("out of memory")));
1457 : }
1458 :
1459 6467908 : other_xids = ProcGlobal->xids;
1460 6467908 : other_subxidstates = ProcGlobal->subxidStates;
1461 :
1462 6467908 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1463 :
1464 : /*
1465 : * Now that we have the lock, we can check latestCompletedXid; if the
1466 : * target Xid is after that, it's surely still running.
1467 : */
1468 6467908 : latestCompletedXid =
1469 6467908 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1470 6467908 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1471 : {
1472 1509345 : LWLockRelease(ProcArrayLock);
1473 : xc_by_latest_xid_inc();
1474 1509345 : return true;
1475 : }
1476 :
1477 : /* No shortcuts, gotta grovel through the array */
1478 4958563 : mypgxactoff = MyProc->pgxactoff;
1479 4958563 : numProcs = arrayP->numProcs;
1480 5153128 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1481 : {
1482 : int pgprocno;
1483 : PGPROC *proc;
1484 : TransactionId pxid;
1485 : int pxids;
1486 :
1487 : /* Ignore ourselves --- dealt with it above */
1488 5136732 : if (pgxactoff == mypgxactoff)
1489 18369 : continue;
1490 :
1491 : /* Fetch xid just once - see GetNewTransactionId */
1492 5118363 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1493 :
1494 5118363 : if (!TransactionIdIsValid(pxid))
1495 128116 : continue;
1496 :
1497 : /*
1498 : * Step 1: check the main Xid
1499 : */
1500 4990247 : if (TransactionIdEquals(pxid, xid))
1501 : {
1502 4942033 : LWLockRelease(ProcArrayLock);
1503 : xc_by_main_xid_inc();
1504 4942033 : return true;
1505 : }
1506 :
1507 : /*
1508 : * We can ignore main Xids that are younger than the target Xid, since
1509 : * the target could not possibly be their child.
1510 : */
1511 48214 : if (TransactionIdPrecedes(xid, pxid))
1512 23021 : continue;
1513 :
1514 : /*
1515 : * Step 2: check the cached child-Xids arrays
1516 : */
1517 25193 : pxids = other_subxidstates[pgxactoff].count;
1518 25193 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1519 25193 : pgprocno = arrayP->pgprocnos[pgxactoff];
1520 25193 : proc = &allProcs[pgprocno];
1521 43676 : for (j = pxids - 1; j >= 0; j--)
1522 : {
1523 : /* Fetch xid just once - see GetNewTransactionId */
1524 18617 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1525 :
1526 18617 : if (TransactionIdEquals(cxid, xid))
1527 : {
1528 134 : LWLockRelease(ProcArrayLock);
1529 : xc_by_child_xid_inc();
1530 134 : return true;
1531 : }
1532 : }
1533 :
1534 : /*
1535 : * Save the main Xid for step 4. We only need to remember main Xids
1536 : * that have uncached children. (Note: there is no race condition
1537 : * here because the overflowed flag cannot be cleared, only set, while
1538 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1539 : * worry about.)
1540 : */
1541 25059 : if (other_subxidstates[pgxactoff].overflowed)
1542 219 : xids[nxids++] = pxid;
1543 : }
1544 :
1545 : /*
1546 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1547 : * in the list must be treated as running.
1548 : */
1549 16396 : if (RecoveryInProgress())
1550 : {
1551 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1552 : Assert(nxids == 0);
1553 :
1554 1 : if (KnownAssignedXidExists(xid))
1555 : {
1556 0 : LWLockRelease(ProcArrayLock);
1557 : xc_by_known_assigned_inc();
1558 0 : return true;
1559 : }
1560 :
1561 : /*
1562 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1563 : * too. Fetch all xids from KnownAssignedXids that are lower than
1564 : * xid, since if xid is a subtransaction its parent will always have a
1565 : * lower value. Note we will collect both main and subXIDs here, but
1566 : * there's no help for it.
1567 : */
1568 1 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1569 0 : nxids = KnownAssignedXidsGet(xids, xid);
1570 : }
1571 :
1572 16396 : LWLockRelease(ProcArrayLock);
1573 :
1574 : /*
1575 : * If none of the relevant caches overflowed, we know the Xid is not
1576 : * running without even looking at pg_subtrans.
1577 : */
1578 16396 : if (nxids == 0)
1579 : {
1580 : xc_no_overflow_inc();
1581 16177 : cachedXidIsNotInProgress = xid;
1582 16177 : return false;
1583 : }
1584 :
1585 : /*
1586 : * Step 4: have to check pg_subtrans.
1587 : *
1588 : * At this point, we know it's either a subtransaction of one of the Xids
1589 : * in xids[], or it's not running. If it's an already-failed
1590 : * subtransaction, we want to say "not running" even though its parent may
1591 : * still be running. So first, check pg_xact to see if it's been aborted.
1592 : */
1593 : xc_slow_answer_inc();
1594 :
1595 219 : if (TransactionIdDidAbort(xid))
1596 : {
1597 0 : cachedXidIsNotInProgress = xid;
1598 0 : return false;
1599 : }
1600 :
1601 : /*
1602 : * It isn't aborted, so check whether the transaction tree it belongs to
1603 : * is still running (or, more precisely, whether it was running when we
1604 : * held ProcArrayLock).
1605 : */
1606 219 : topxid = SubTransGetTopmostTransaction(xid);
1607 : Assert(TransactionIdIsValid(topxid));
1608 376 : if (!TransactionIdEquals(topxid, xid) &&
1609 157 : pg_lfind32(topxid, xids, nxids))
1610 157 : return true;
1611 :
1612 62 : cachedXidIsNotInProgress = xid;
1613 62 : return false;
1614 : }
1615 :
1616 :
1617 : /*
1618 : * Determine XID horizons.
1619 : *
1620 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1621 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1622 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1623 : * GlobalVisState).
1624 : *
1625 : * See the definition of ComputeXidHorizonsResult for the various computed
1626 : * horizons.
1627 : *
1628 : * For VACUUM separate horizons (used to decide which deleted tuples must
1629 : * be preserved), for shared and non-shared tables are computed. For shared
1630 : * relations backends in all databases must be considered, but for non-shared
1631 : * relations that's not required, since only backends in my own database could
1632 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1633 : * VACUUMs because (a) they must be working on other tables, and (b) they
1634 : * don't need to do snapshot-based lookups.
1635 : *
1636 : * This also computes a horizon used to truncate pg_subtrans. For that
1637 : * backends in all databases have to be considered, and concurrently running
1638 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1639 : * accesses.
1640 : *
1641 : * Note: we include all currently running xids in the set of considered xids.
1642 : * This ensures that if a just-started xact has not yet set its snapshot,
1643 : * when it does set the snapshot it cannot set xmin less than what we compute.
1644 : * See notes in src/backend/access/transam/README.
1645 : *
1646 : * Note: despite the above, it's possible for the calculated values to move
1647 : * backwards on repeated calls. The calculated values are conservative, so
1648 : * that anything older is definitely not considered as running by anyone
1649 : * anymore, but the exact values calculated depend on a number of things. For
1650 : * example, if there are no transactions running in the current database, the
1651 : * horizon for normal tables will be latestCompletedXid. If a transaction
1652 : * begins after that, its xmin will include in-progress transactions in other
1653 : * databases that started earlier, so another call will return a lower value.
1654 : * Nonetheless it is safe to vacuum a table in the current database with the
1655 : * first result. There are also replication-related effects: a walsender
1656 : * process can set its xmin based on transactions that are no longer running
1657 : * on the primary but are still being replayed on the standby, thus possibly
1658 : * making the values go backwards. In this case there is a possibility that
1659 : * we lose data that the standby would like to have, but unless the standby
1660 : * uses a replication slot to make its xmin persistent there is little we can
1661 : * do about that --- data is only protected if the walsender runs continuously
1662 : * while queries are executed on the standby. (The Hot Standby code deals
1663 : * with such cases by failing standby queries that needed to access
1664 : * already-removed data, so there's no integrity bug.)
1665 : *
1666 : * Note: the approximate horizons (see definition of GlobalVisState) are
1667 : * updated by the computations done here. That's currently required for
1668 : * correctness and a small optimization. Without doing so it's possible that
1669 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1670 : * horizon than later when deciding which tuples can be removed - which the
1671 : * code doesn't expect (breaking HOT).
1672 : */
1673 : static void
1674 242441 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1675 : {
1676 242441 : ProcArrayStruct *arrayP = procArray;
1677 : TransactionId kaxmin;
1678 242441 : bool in_recovery = RecoveryInProgress();
1679 242441 : TransactionId *other_xids = ProcGlobal->xids;
1680 :
1681 : /* inferred after ProcArrayLock is released */
1682 242441 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1683 :
1684 242441 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1685 :
1686 242441 : h->latest_completed = TransamVariables->latestCompletedXid;
1687 :
1688 : /*
1689 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1690 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1691 : * and so protects us against overestimating the result due to future
1692 : * additions.
1693 : */
1694 : {
1695 : TransactionId initial;
1696 :
1697 242441 : initial = XidFromFullTransactionId(h->latest_completed);
1698 : Assert(TransactionIdIsValid(initial));
1699 242441 : TransactionIdAdvance(initial);
1700 :
1701 242441 : h->oldest_considered_running = initial;
1702 242441 : h->shared_oldest_nonremovable = initial;
1703 242441 : h->data_oldest_nonremovable = initial;
1704 :
1705 : /*
1706 : * Only modifications made by this backend affect the horizon for
1707 : * temporary relations. Instead of a check in each iteration of the
1708 : * loop over all PGPROCs it is cheaper to just initialize to the
1709 : * current top-level xid any.
1710 : *
1711 : * Without an assigned xid we could use a horizon as aggressive as
1712 : * GetNewTransactionId(), but we can get away with the much cheaper
1713 : * latestCompletedXid + 1: If this backend has no xid there, by
1714 : * definition, can't be any newer changes in the temp table than
1715 : * latestCompletedXid.
1716 : */
1717 242441 : if (TransactionIdIsValid(MyProc->xid))
1718 49810 : h->temp_oldest_nonremovable = MyProc->xid;
1719 : else
1720 192631 : h->temp_oldest_nonremovable = initial;
1721 : }
1722 :
1723 : /*
1724 : * Fetch slot horizons while ProcArrayLock is held - the
1725 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1726 : * the lock.
1727 : */
1728 242441 : h->slot_xmin = procArray->replication_slot_xmin;
1729 242441 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1730 :
1731 1553789 : for (int index = 0; index < arrayP->numProcs; index++)
1732 : {
1733 1311348 : int pgprocno = arrayP->pgprocnos[index];
1734 1311348 : PGPROC *proc = &allProcs[pgprocno];
1735 1311348 : int8 statusFlags = ProcGlobal->statusFlags[index];
1736 : TransactionId xid;
1737 : TransactionId xmin;
1738 :
1739 : /* Fetch xid just once - see GetNewTransactionId */
1740 1311348 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1741 1311348 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1742 :
1743 : /*
1744 : * Consider both the transaction's Xmin, and its Xid.
1745 : *
1746 : * We must check both because a transaction might have an Xmin but not
1747 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1748 : * some not-yet-set Xmin.
1749 : */
1750 1311348 : xmin = TransactionIdOlder(xmin, xid);
1751 :
1752 : /* if neither is set, this proc doesn't influence the horizon */
1753 1311348 : if (!TransactionIdIsValid(xmin))
1754 612358 : continue;
1755 :
1756 : /*
1757 : * Don't ignore any procs when determining which transactions might be
1758 : * considered running. While slots should ensure logical decoding
1759 : * backends are protected even without this check, it can't hurt to
1760 : * include them here as well..
1761 : */
1762 698990 : h->oldest_considered_running =
1763 698990 : TransactionIdOlder(h->oldest_considered_running, xmin);
1764 :
1765 : /*
1766 : * Skip over backends either vacuuming (which is ok with rows being
1767 : * removed, as long as pg_subtrans is not truncated) or doing logical
1768 : * decoding (which manages xmin separately, check below).
1769 : */
1770 698990 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1771 239098 : continue;
1772 :
1773 : /* shared tables need to take backends in all databases into account */
1774 459892 : h->shared_oldest_nonremovable =
1775 459892 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1776 :
1777 : /*
1778 : * Normally sessions in other databases are ignored for anything but
1779 : * the shared horizon.
1780 : *
1781 : * However, include them when MyDatabaseId is not (yet) set. A
1782 : * backend in the process of starting up must not compute a "too
1783 : * aggressive" horizon, otherwise we could end up using it to prune
1784 : * still-needed data away. If the current backend never connects to a
1785 : * database this is harmless, because data_oldest_nonremovable will
1786 : * never be utilized.
1787 : *
1788 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1789 : * be included. (This flag is used for hot standby feedback, which
1790 : * can't be tied to a specific database.)
1791 : *
1792 : * Also, while in recovery we cannot compute an accurate per-database
1793 : * horizon, as all xids are managed via the KnownAssignedXids
1794 : * machinery.
1795 : */
1796 459892 : if (proc->databaseId == MyDatabaseId ||
1797 22368 : MyDatabaseId == InvalidOid ||
1798 12527 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1799 : in_recovery)
1800 : {
1801 447370 : h->data_oldest_nonremovable =
1802 447370 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1803 : }
1804 : }
1805 :
1806 : /*
1807 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1808 : * after lock is released.
1809 : */
1810 242441 : if (in_recovery)
1811 383 : kaxmin = KnownAssignedXidsGetOldestXmin();
1812 :
1813 : /*
1814 : * No other information from shared state is needed, release the lock
1815 : * immediately. The rest of the computations can be done without a lock.
1816 : */
1817 242441 : LWLockRelease(ProcArrayLock);
1818 :
1819 242441 : if (in_recovery)
1820 : {
1821 383 : h->oldest_considered_running =
1822 383 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1823 383 : h->shared_oldest_nonremovable =
1824 383 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1825 383 : h->data_oldest_nonremovable =
1826 383 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1827 : /* temp relations cannot be accessed in recovery */
1828 : }
1829 :
1830 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1831 : h->shared_oldest_nonremovable));
1832 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1833 : h->data_oldest_nonremovable));
1834 :
1835 : /*
1836 : * Check whether there are replication slots requiring an older xmin.
1837 : */
1838 242441 : h->shared_oldest_nonremovable =
1839 242441 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1840 242441 : h->data_oldest_nonremovable =
1841 242441 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1842 :
1843 : /*
1844 : * The only difference between catalog / data horizons is that the slot's
1845 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1846 : * for logical decoding). Initialize with data horizon, and then back up
1847 : * further if necessary. Have to back up the shared horizon as well, since
1848 : * that also can contain catalogs.
1849 : */
1850 242441 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1851 242441 : h->shared_oldest_nonremovable =
1852 242441 : TransactionIdOlder(h->shared_oldest_nonremovable,
1853 : h->slot_catalog_xmin);
1854 242441 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1855 242441 : h->catalog_oldest_nonremovable =
1856 242441 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1857 : h->slot_catalog_xmin);
1858 :
1859 : /*
1860 : * It's possible that slots backed up the horizons further than
1861 : * oldest_considered_running. Fix.
1862 : */
1863 242441 : h->oldest_considered_running =
1864 242441 : TransactionIdOlder(h->oldest_considered_running,
1865 : h->shared_oldest_nonremovable);
1866 242441 : h->oldest_considered_running =
1867 242441 : TransactionIdOlder(h->oldest_considered_running,
1868 : h->catalog_oldest_nonremovable);
1869 242441 : h->oldest_considered_running =
1870 242441 : TransactionIdOlder(h->oldest_considered_running,
1871 : h->data_oldest_nonremovable);
1872 :
1873 : /*
1874 : * shared horizons have to be at least as old as the oldest visible in
1875 : * current db
1876 : */
1877 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1878 : h->data_oldest_nonremovable));
1879 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1880 : h->catalog_oldest_nonremovable));
1881 :
1882 : /*
1883 : * Horizons need to ensure that pg_subtrans access is still possible for
1884 : * the relevant backends.
1885 : */
1886 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1887 : h->shared_oldest_nonremovable));
1888 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1889 : h->catalog_oldest_nonremovable));
1890 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1891 : h->data_oldest_nonremovable));
1892 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1893 : h->temp_oldest_nonremovable));
1894 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1895 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1896 : h->slot_xmin));
1897 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1898 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1899 : h->slot_catalog_xmin));
1900 :
1901 : /* update approximate horizons with the computed horizons */
1902 242441 : GlobalVisUpdateApply(h);
1903 242441 : }
1904 :
1905 : /*
1906 : * Determine what kind of visibility horizon needs to be used for a
1907 : * relation. If rel is NULL, the most conservative horizon is used.
1908 : */
1909 : static inline GlobalVisHorizonKind
1910 20545262 : GlobalVisHorizonKindForRel(Relation rel)
1911 : {
1912 : /*
1913 : * Other relkinds currently don't contain xids, nor always the necessary
1914 : * logical decoding markers.
1915 : */
1916 : Assert(!rel ||
1917 : rel->rd_rel->relkind == RELKIND_RELATION ||
1918 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1919 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1920 :
1921 20545262 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1922 169844 : return VISHORIZON_SHARED;
1923 20375418 : else if (IsCatalogRelation(rel) ||
1924 15577838 : RelationIsAccessibleInLogicalDecoding(rel))
1925 4797585 : return VISHORIZON_CATALOG;
1926 15577833 : else if (!RELATION_IS_LOCAL(rel))
1927 15434101 : return VISHORIZON_DATA;
1928 : else
1929 143732 : return VISHORIZON_TEMP;
1930 : }
1931 :
1932 : /*
1933 : * Return the oldest XID for which deleted tuples must be preserved in the
1934 : * passed table.
1935 : *
1936 : * If rel is not NULL the horizon may be considerably more recent than
1937 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1938 : * that is correct (but not optimal) for all relations will be returned.
1939 : *
1940 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1941 : * the passed in table.
1942 : */
1943 : TransactionId
1944 154693 : GetOldestNonRemovableTransactionId(Relation rel)
1945 : {
1946 : ComputeXidHorizonsResult horizons;
1947 :
1948 154693 : ComputeXidHorizons(&horizons);
1949 :
1950 154693 : switch (GlobalVisHorizonKindForRel(rel))
1951 : {
1952 22249 : case VISHORIZON_SHARED:
1953 22249 : return horizons.shared_oldest_nonremovable;
1954 95177 : case VISHORIZON_CATALOG:
1955 95177 : return horizons.catalog_oldest_nonremovable;
1956 21432 : case VISHORIZON_DATA:
1957 21432 : return horizons.data_oldest_nonremovable;
1958 15835 : case VISHORIZON_TEMP:
1959 15835 : return horizons.temp_oldest_nonremovable;
1960 : }
1961 :
1962 : /* just to prevent compiler warnings */
1963 0 : return InvalidTransactionId;
1964 : }
1965 :
1966 : /*
1967 : * Return the oldest transaction id any currently running backend might still
1968 : * consider running. This should not be used for visibility / pruning
1969 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1970 : * decisions like up to where pg_subtrans can be truncated.
1971 : */
1972 : TransactionId
1973 1916 : GetOldestTransactionIdConsideredRunning(void)
1974 : {
1975 : ComputeXidHorizonsResult horizons;
1976 :
1977 1916 : ComputeXidHorizons(&horizons);
1978 :
1979 1916 : return horizons.oldest_considered_running;
1980 : }
1981 :
1982 : /*
1983 : * Return the visibility horizons for a hot standby feedback message.
1984 : */
1985 : void
1986 54 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1987 : {
1988 : ComputeXidHorizonsResult horizons;
1989 :
1990 54 : ComputeXidHorizons(&horizons);
1991 :
1992 : /*
1993 : * Don't want to use shared_oldest_nonremovable here, as that contains the
1994 : * effect of replication slot's catalog_xmin. We want to send a separate
1995 : * feedback for the catalog horizon, so the primary can remove data table
1996 : * contents more aggressively.
1997 : */
1998 54 : *xmin = horizons.shared_oldest_nonremovable_raw;
1999 54 : *catalog_xmin = horizons.slot_catalog_xmin;
2000 54 : }
2001 :
2002 : /*
2003 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2004 : *
2005 : * We have to export this for use by snapmgr.c.
2006 : */
2007 : int
2008 38287 : GetMaxSnapshotXidCount(void)
2009 : {
2010 38287 : return procArray->maxProcs;
2011 : }
2012 :
2013 : /*
2014 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2015 : *
2016 : * We have to export this for use by snapmgr.c.
2017 : */
2018 : int
2019 38075 : GetMaxSnapshotSubxidCount(void)
2020 : {
2021 38075 : return TOTAL_MAX_CACHED_SUBXIDS;
2022 : }
2023 :
2024 : /*
2025 : * Helper function for GetSnapshotData() that checks if the bulk of the
2026 : * visibility information in the snapshot is still valid. If so, it updates
2027 : * the fields that need to change and returns true. Otherwise it returns
2028 : * false.
2029 : *
2030 : * This very likely can be evolved to not need ProcArrayLock held (at very
2031 : * least in the case we already hold a snapshot), but that's for another day.
2032 : */
2033 : static bool
2034 2692824 : GetSnapshotDataReuse(Snapshot snapshot)
2035 : {
2036 : uint64 curXactCompletionCount;
2037 :
2038 : Assert(LWLockHeldByMe(ProcArrayLock));
2039 :
2040 2692824 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2041 38058 : return false;
2042 :
2043 2654766 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2044 2654766 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2045 467383 : return false;
2046 :
2047 : /*
2048 : * If the current xactCompletionCount is still the same as it was at the
2049 : * time the snapshot was built, we can be sure that rebuilding the
2050 : * contents of the snapshot the hard way would result in the same snapshot
2051 : * contents:
2052 : *
2053 : * As explained in transam/README, the set of xids considered running by
2054 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2055 : * contents only depend on transactions with xids and xactCompletionCount
2056 : * is incremented whenever a transaction with an xid finishes (while
2057 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2058 : * ensures we would detect if the snapshot would have changed.
2059 : *
2060 : * As the snapshot contents are the same as it was before, it is safe to
2061 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2062 : * visible under the snapshot could already have been removed (that'd
2063 : * require the set of running transactions to change) and it fulfills the
2064 : * requirement that concurrent GetSnapshotData() calls yield the same
2065 : * xmin.
2066 : */
2067 2187383 : if (!TransactionIdIsValid(MyProc->xmin))
2068 817516 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2069 :
2070 2187383 : RecentXmin = snapshot->xmin;
2071 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2072 :
2073 2187383 : snapshot->curcid = GetCurrentCommandId(false);
2074 2187383 : snapshot->active_count = 0;
2075 2187383 : snapshot->regd_count = 0;
2076 2187383 : snapshot->copied = false;
2077 :
2078 2187383 : return true;
2079 : }
2080 :
2081 : /*
2082 : * GetSnapshotData -- returns information about running transactions.
2083 : *
2084 : * The returned snapshot includes xmin (lowest still-running xact ID),
2085 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2086 : * in the range xmin <= xid < xmax. It is used as follows:
2087 : * All xact IDs < xmin are considered finished.
2088 : * All xact IDs >= xmax are considered still running.
2089 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2090 : * it is considered running or not.
2091 : * This ensures that the set of transactions seen as "running" by the
2092 : * current xact will not change after it takes the snapshot.
2093 : *
2094 : * All running top-level XIDs are included in the snapshot, except for lazy
2095 : * VACUUM processes. We also try to include running subtransaction XIDs,
2096 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2097 : * information may not be available. If we find any overflowed subxid arrays,
2098 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2099 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2100 : *
2101 : * We also update the following backend-global variables:
2102 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2103 : * current transaction (this is the same as MyProc->xmin).
2104 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2105 : * older than this are known not running any more.
2106 : *
2107 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2108 : * for the benefit of the GlobalVisTest* family of functions.
2109 : *
2110 : * Note: this function should probably not be called with an argument that's
2111 : * not statically allocated (see xip allocation below).
2112 : */
2113 : Snapshot
2114 2692824 : GetSnapshotData(Snapshot snapshot)
2115 : {
2116 2692824 : ProcArrayStruct *arrayP = procArray;
2117 2692824 : TransactionId *other_xids = ProcGlobal->xids;
2118 : TransactionId xmin;
2119 : TransactionId xmax;
2120 2692824 : int count = 0;
2121 2692824 : int subcount = 0;
2122 2692824 : bool suboverflowed = false;
2123 : FullTransactionId latest_completed;
2124 : TransactionId oldestxid;
2125 : int mypgxactoff;
2126 : TransactionId myxid;
2127 : uint64 curXactCompletionCount;
2128 :
2129 2692824 : TransactionId replication_slot_xmin = InvalidTransactionId;
2130 2692824 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2131 :
2132 : Assert(snapshot != NULL);
2133 :
2134 : /*
2135 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2136 : * be sufficient. But it seems better to do the malloc while not holding
2137 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2138 : * more subxip storage than is probably needed.
2139 : *
2140 : * This does open a possibility for avoiding repeated malloc/free: since
2141 : * maxProcs does not change at runtime, we can simply reuse the previous
2142 : * xip arrays if any. (This relies on the fact that all callers pass
2143 : * static SnapshotData structs.)
2144 : */
2145 2692824 : if (snapshot->xip == NULL)
2146 : {
2147 : /*
2148 : * First call for this snapshot. Snapshot is same size whether or not
2149 : * we are in recovery, see later comments.
2150 : */
2151 38050 : snapshot->xip = (TransactionId *)
2152 38050 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2153 38050 : if (snapshot->xip == NULL)
2154 0 : ereport(ERROR,
2155 : (errcode(ERRCODE_OUT_OF_MEMORY),
2156 : errmsg("out of memory")));
2157 : Assert(snapshot->subxip == NULL);
2158 38050 : snapshot->subxip = (TransactionId *)
2159 38050 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2160 38050 : if (snapshot->subxip == NULL)
2161 0 : ereport(ERROR,
2162 : (errcode(ERRCODE_OUT_OF_MEMORY),
2163 : errmsg("out of memory")));
2164 : }
2165 :
2166 : /*
2167 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2168 : * going to set MyProc->xmin.
2169 : */
2170 2692824 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2171 :
2172 2692824 : if (GetSnapshotDataReuse(snapshot))
2173 : {
2174 2187383 : LWLockRelease(ProcArrayLock);
2175 2187383 : return snapshot;
2176 : }
2177 :
2178 505441 : latest_completed = TransamVariables->latestCompletedXid;
2179 505441 : mypgxactoff = MyProc->pgxactoff;
2180 505441 : myxid = other_xids[mypgxactoff];
2181 : Assert(myxid == MyProc->xid);
2182 :
2183 505441 : oldestxid = TransamVariables->oldestXid;
2184 505441 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2185 :
2186 : /* xmax is always latestCompletedXid + 1 */
2187 505441 : xmax = XidFromFullTransactionId(latest_completed);
2188 505441 : TransactionIdAdvance(xmax);
2189 : Assert(TransactionIdIsNormal(xmax));
2190 :
2191 : /* initialize xmin calculation with xmax */
2192 505441 : xmin = xmax;
2193 :
2194 : /* take own xid into account, saves a check inside the loop */
2195 505441 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2196 35357 : xmin = myxid;
2197 :
2198 505441 : snapshot->takenDuringRecovery = RecoveryInProgress();
2199 :
2200 505441 : if (!snapshot->takenDuringRecovery)
2201 : {
2202 503865 : int numProcs = arrayP->numProcs;
2203 503865 : TransactionId *xip = snapshot->xip;
2204 503865 : int *pgprocnos = arrayP->pgprocnos;
2205 503865 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2206 503865 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2207 :
2208 : /*
2209 : * First collect set of pgxactoff/xids that need to be included in the
2210 : * snapshot.
2211 : */
2212 4472930 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2213 : {
2214 : /* Fetch xid just once - see GetNewTransactionId */
2215 3969065 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2216 : uint8 statusFlags;
2217 :
2218 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2219 :
2220 : /*
2221 : * If the transaction has no XID assigned, we can skip it; it
2222 : * won't have sub-XIDs either.
2223 : */
2224 3969065 : if (likely(xid == InvalidTransactionId))
2225 3186468 : continue;
2226 :
2227 : /*
2228 : * We don't include our own XIDs (if any) in the snapshot. It
2229 : * needs to be included in the xmin computation, but we did so
2230 : * outside the loop.
2231 : */
2232 782597 : if (pgxactoff == mypgxactoff)
2233 70058 : continue;
2234 :
2235 : /*
2236 : * The only way we are able to get here with a non-normal xid is
2237 : * during bootstrap - with this backend using
2238 : * BootstrapTransactionId. But the above test should filter that
2239 : * out.
2240 : */
2241 : Assert(TransactionIdIsNormal(xid));
2242 :
2243 : /*
2244 : * If the XID is >= xmax, we can skip it; such transactions will
2245 : * be treated as running anyway (and any sub-XIDs will also be >=
2246 : * xmax).
2247 : */
2248 712539 : if (!NormalTransactionIdPrecedes(xid, xmax))
2249 210235 : continue;
2250 :
2251 : /*
2252 : * Skip over backends doing logical decoding which manages xmin
2253 : * separately (check below) and ones running LAZY VACUUM.
2254 : */
2255 502304 : statusFlags = allStatusFlags[pgxactoff];
2256 502304 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2257 12 : continue;
2258 :
2259 502292 : if (NormalTransactionIdPrecedes(xid, xmin))
2260 314211 : xmin = xid;
2261 :
2262 : /* Add XID to snapshot. */
2263 502292 : xip[count++] = xid;
2264 :
2265 : /*
2266 : * Save subtransaction XIDs if possible (if we've already
2267 : * overflowed, there's no point). Note that the subxact XIDs must
2268 : * be later than their parent, so no need to check them against
2269 : * xmin. We could filter against xmax, but it seems better not to
2270 : * do that much work while holding the ProcArrayLock.
2271 : *
2272 : * The other backend can add more subxids concurrently, but cannot
2273 : * remove any. Hence it's important to fetch nxids just once.
2274 : * Should be safe to use memcpy, though. (We needn't worry about
2275 : * missing any xids added concurrently, because they must postdate
2276 : * xmax.)
2277 : *
2278 : * Again, our own XIDs are not included in the snapshot.
2279 : */
2280 502292 : if (!suboverflowed)
2281 : {
2282 :
2283 502288 : if (subxidStates[pgxactoff].overflowed)
2284 616 : suboverflowed = true;
2285 : else
2286 : {
2287 501672 : int nsubxids = subxidStates[pgxactoff].count;
2288 :
2289 501672 : if (nsubxids > 0)
2290 : {
2291 8370 : int pgprocno = pgprocnos[pgxactoff];
2292 8370 : PGPROC *proc = &allProcs[pgprocno];
2293 :
2294 8370 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2295 :
2296 8370 : memcpy(snapshot->subxip + subcount,
2297 8370 : proc->subxids.xids,
2298 : nsubxids * sizeof(TransactionId));
2299 8370 : subcount += nsubxids;
2300 : }
2301 : }
2302 : }
2303 : }
2304 : }
2305 : else
2306 : {
2307 : /*
2308 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2309 : *
2310 : * We store all xids directly into subxip[]. Here's why:
2311 : *
2312 : * In recovery we don't know which xids are top-level and which are
2313 : * subxacts, a design choice that greatly simplifies xid processing.
2314 : *
2315 : * It seems like we would want to try to put xids into xip[] only, but
2316 : * that is fairly small. We would either need to make that bigger or
2317 : * to increase the rate at which we WAL-log xid assignment; neither is
2318 : * an appealing choice.
2319 : *
2320 : * We could try to store xids into xip[] first and then into subxip[]
2321 : * if there are too many xids. That only works if the snapshot doesn't
2322 : * overflow because we do not search subxip[] in that case. A simpler
2323 : * way is to just store all xids in the subxip array because this is
2324 : * by far the bigger array. We just leave the xip array empty.
2325 : *
2326 : * Either way we need to change the way XidInMVCCSnapshot() works
2327 : * depending upon when the snapshot was taken, or change normal
2328 : * snapshot processing so it matches.
2329 : *
2330 : * Note: It is possible for recovery to end before we finish taking
2331 : * the snapshot, and for newly assigned transaction ids to be added to
2332 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2333 : * those newly added transaction ids would be filtered away, so we
2334 : * need not be concerned about them.
2335 : */
2336 1576 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2337 : xmax);
2338 :
2339 1576 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2340 8 : suboverflowed = true;
2341 : }
2342 :
2343 :
2344 : /*
2345 : * Fetch into local variable while ProcArrayLock is held - the
2346 : * LWLockRelease below is a barrier, ensuring this happens inside the
2347 : * lock.
2348 : */
2349 505441 : replication_slot_xmin = procArray->replication_slot_xmin;
2350 505441 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2351 :
2352 505441 : if (!TransactionIdIsValid(MyProc->xmin))
2353 273408 : MyProc->xmin = TransactionXmin = xmin;
2354 :
2355 505441 : LWLockRelease(ProcArrayLock);
2356 :
2357 : /* maintain state for GlobalVis* */
2358 : {
2359 : TransactionId def_vis_xid;
2360 : TransactionId def_vis_xid_data;
2361 : FullTransactionId def_vis_fxid;
2362 : FullTransactionId def_vis_fxid_data;
2363 : FullTransactionId oldestfxid;
2364 :
2365 : /*
2366 : * Converting oldestXid is only safe when xid horizon cannot advance,
2367 : * i.e. holding locks. While we don't hold the lock anymore, all the
2368 : * necessary data has been gathered with lock held.
2369 : */
2370 505441 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2371 :
2372 : /* Check whether there's a replication slot requiring an older xmin. */
2373 : def_vis_xid_data =
2374 505441 : TransactionIdOlder(xmin, replication_slot_xmin);
2375 :
2376 : /*
2377 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2378 : * if older than this xid.
2379 : */
2380 505441 : def_vis_xid = def_vis_xid_data;
2381 :
2382 : /*
2383 : * Check whether there's a replication slot requiring an older catalog
2384 : * xmin.
2385 : */
2386 : def_vis_xid =
2387 505441 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2388 :
2389 505441 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2390 505441 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2391 :
2392 : /*
2393 : * Check if we can increase upper bound. As a previous
2394 : * GlobalVisUpdate() might have computed more aggressive values, don't
2395 : * overwrite them if so.
2396 : */
2397 : GlobalVisSharedRels.definitely_needed =
2398 505441 : FullTransactionIdNewer(def_vis_fxid,
2399 : GlobalVisSharedRels.definitely_needed);
2400 : GlobalVisCatalogRels.definitely_needed =
2401 505441 : FullTransactionIdNewer(def_vis_fxid,
2402 : GlobalVisCatalogRels.definitely_needed);
2403 : GlobalVisDataRels.definitely_needed =
2404 505441 : FullTransactionIdNewer(def_vis_fxid_data,
2405 : GlobalVisDataRels.definitely_needed);
2406 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2407 505441 : if (TransactionIdIsNormal(myxid))
2408 : GlobalVisTempRels.definitely_needed =
2409 69944 : FullXidRelativeTo(latest_completed, myxid);
2410 : else
2411 : {
2412 435497 : GlobalVisTempRels.definitely_needed = latest_completed;
2413 435497 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2414 : }
2415 :
2416 : /*
2417 : * Check if we know that we can initialize or increase the lower
2418 : * bound. Currently the only cheap way to do so is to use
2419 : * TransamVariables->oldestXid as input.
2420 : *
2421 : * We should definitely be able to do better. We could e.g. put a
2422 : * global lower bound value into TransamVariables.
2423 : */
2424 : GlobalVisSharedRels.maybe_needed =
2425 505441 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2426 : oldestfxid);
2427 : GlobalVisCatalogRels.maybe_needed =
2428 505441 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2429 : oldestfxid);
2430 : GlobalVisDataRels.maybe_needed =
2431 505441 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2432 : oldestfxid);
2433 : /* accurate value known */
2434 505441 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2435 : }
2436 :
2437 505441 : RecentXmin = xmin;
2438 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2439 :
2440 505441 : snapshot->xmin = xmin;
2441 505441 : snapshot->xmax = xmax;
2442 505441 : snapshot->xcnt = count;
2443 505441 : snapshot->subxcnt = subcount;
2444 505441 : snapshot->suboverflowed = suboverflowed;
2445 505441 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2446 :
2447 505441 : snapshot->curcid = GetCurrentCommandId(false);
2448 :
2449 : /*
2450 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2451 : * not copied in persistent memory.
2452 : */
2453 505441 : snapshot->active_count = 0;
2454 505441 : snapshot->regd_count = 0;
2455 505441 : snapshot->copied = false;
2456 :
2457 505441 : return snapshot;
2458 : }
2459 :
2460 : /*
2461 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2462 : *
2463 : * This is called when installing a snapshot imported from another
2464 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2465 : * check that the source transaction is still running, and we'd better do
2466 : * that atomically with installing the new xmin.
2467 : *
2468 : * Returns true if successful, false if source xact is no longer running.
2469 : */
2470 : bool
2471 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2472 : VirtualTransactionId *sourcevxid)
2473 : {
2474 16 : bool result = false;
2475 16 : ProcArrayStruct *arrayP = procArray;
2476 : int index;
2477 :
2478 : Assert(TransactionIdIsNormal(xmin));
2479 16 : if (!sourcevxid)
2480 0 : return false;
2481 :
2482 : /* Get lock so source xact can't end while we're doing this */
2483 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2484 :
2485 : /*
2486 : * Find the PGPROC entry of the source transaction. (This could use
2487 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2488 : * performance critical.)
2489 : */
2490 16 : for (index = 0; index < arrayP->numProcs; index++)
2491 : {
2492 16 : int pgprocno = arrayP->pgprocnos[index];
2493 16 : PGPROC *proc = &allProcs[pgprocno];
2494 16 : int statusFlags = ProcGlobal->statusFlags[index];
2495 : TransactionId xid;
2496 :
2497 : /* Ignore procs running LAZY VACUUM */
2498 16 : if (statusFlags & PROC_IN_VACUUM)
2499 0 : continue;
2500 :
2501 : /* We are only interested in the specific virtual transaction. */
2502 16 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2503 0 : continue;
2504 16 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2505 0 : continue;
2506 :
2507 : /*
2508 : * We check the transaction's database ID for paranoia's sake: if it's
2509 : * in another DB then its xmin does not cover us. Caller should have
2510 : * detected this already, so we just treat any funny cases as
2511 : * "transaction not found".
2512 : */
2513 16 : if (proc->databaseId != MyDatabaseId)
2514 0 : continue;
2515 :
2516 : /*
2517 : * Likewise, let's just make real sure its xmin does cover us.
2518 : */
2519 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2520 16 : if (!TransactionIdIsNormal(xid) ||
2521 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2522 0 : continue;
2523 :
2524 : /*
2525 : * We're good. Install the new xmin. As in GetSnapshotData, set
2526 : * TransactionXmin too. (Note that because snapmgr.c called
2527 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2528 : * we don't check that.)
2529 : */
2530 16 : MyProc->xmin = TransactionXmin = xmin;
2531 :
2532 16 : result = true;
2533 16 : break;
2534 : }
2535 :
2536 16 : LWLockRelease(ProcArrayLock);
2537 :
2538 16 : return result;
2539 : }
2540 :
2541 : /*
2542 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2543 : *
2544 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2545 : * PGPROC of the transaction from which we imported the snapshot, rather than
2546 : * an XID.
2547 : *
2548 : * Note that this function also copies statusFlags from the source `proc` in
2549 : * order to avoid the case where MyProc's xmin needs to be skipped for
2550 : * computing xid horizon.
2551 : *
2552 : * Returns true if successful, false if source xact is no longer running.
2553 : */
2554 : bool
2555 2231 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2556 : {
2557 2231 : bool result = false;
2558 : TransactionId xid;
2559 :
2560 : Assert(TransactionIdIsNormal(xmin));
2561 : Assert(proc != NULL);
2562 :
2563 : /*
2564 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2565 : */
2566 2231 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2567 :
2568 : /*
2569 : * Be certain that the referenced PGPROC has an advertised xmin which is
2570 : * no later than the one we're installing, so that the system-wide xmin
2571 : * can't go backwards. Also, make sure it's running in the same database,
2572 : * so that the per-database xmin cannot go backwards.
2573 : */
2574 2231 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2575 2231 : if (proc->databaseId == MyDatabaseId &&
2576 2231 : TransactionIdIsNormal(xid) &&
2577 2231 : TransactionIdPrecedesOrEquals(xid, xmin))
2578 : {
2579 : /*
2580 : * Install xmin and propagate the statusFlags that affect how the
2581 : * value is interpreted by vacuum.
2582 : */
2583 2231 : MyProc->xmin = TransactionXmin = xmin;
2584 2231 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2585 2231 : (proc->statusFlags & PROC_XMIN_FLAGS);
2586 2231 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2587 :
2588 2231 : result = true;
2589 : }
2590 :
2591 2231 : LWLockRelease(ProcArrayLock);
2592 :
2593 2231 : return result;
2594 : }
2595 :
2596 : /*
2597 : * GetRunningTransactionData -- returns information about running transactions.
2598 : *
2599 : * Similar to GetSnapshotData but returns more information. We include
2600 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2601 : * prepared transactions.
2602 : *
2603 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2604 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2605 : * array until the caller has WAL-logged this snapshot, and releases the
2606 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2607 : * lock is released.
2608 : *
2609 : * The returned data structure is statically allocated; caller should not
2610 : * modify it, and must not assume it is valid past the next call.
2611 : *
2612 : * This is never executed during recovery so there is no need to look at
2613 : * KnownAssignedXids.
2614 : *
2615 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2616 : * may return entries with duplicated TransactionId values coming from
2617 : * transaction finishing to prepare. Nothing is done about duplicated
2618 : * entries here to not hold on ProcArrayLock more than necessary.
2619 : *
2620 : * We don't worry about updating other counters, we want to keep this as
2621 : * simple as possible and leave GetSnapshotData() as the primary code for
2622 : * that bookkeeping.
2623 : *
2624 : * Note that if any transaction has overflowed its cached subtransactions
2625 : * then there is no real need include any subtransactions.
2626 : */
2627 : RunningTransactions
2628 1525 : GetRunningTransactionData(void)
2629 : {
2630 : /* result workspace */
2631 : static RunningTransactionsData CurrentRunningXactsData;
2632 :
2633 1525 : ProcArrayStruct *arrayP = procArray;
2634 1525 : TransactionId *other_xids = ProcGlobal->xids;
2635 1525 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2636 : TransactionId latestCompletedXid;
2637 : TransactionId oldestRunningXid;
2638 : TransactionId oldestDatabaseRunningXid;
2639 : TransactionId *xids;
2640 : int index;
2641 : int count;
2642 : int subcount;
2643 : bool suboverflowed;
2644 :
2645 : Assert(!RecoveryInProgress());
2646 :
2647 : /*
2648 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2649 : * be sufficient. But it seems better to do the malloc while not holding
2650 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2651 : * more subxip storage than is probably needed.
2652 : *
2653 : * Should only be allocated in bgwriter, since only ever executed during
2654 : * checkpoints.
2655 : */
2656 1525 : if (CurrentRunningXacts->xids == NULL)
2657 : {
2658 : /*
2659 : * First call
2660 : */
2661 611 : CurrentRunningXacts->xids = (TransactionId *)
2662 611 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2663 611 : if (CurrentRunningXacts->xids == NULL)
2664 0 : ereport(ERROR,
2665 : (errcode(ERRCODE_OUT_OF_MEMORY),
2666 : errmsg("out of memory")));
2667 : }
2668 :
2669 1525 : xids = CurrentRunningXacts->xids;
2670 :
2671 1525 : count = subcount = 0;
2672 1525 : suboverflowed = false;
2673 :
2674 : /*
2675 : * Ensure that no xids enter or leave the procarray while we obtain
2676 : * snapshot.
2677 : */
2678 1525 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2679 1525 : LWLockAcquire(XidGenLock, LW_SHARED);
2680 :
2681 1525 : latestCompletedXid =
2682 1525 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2683 1525 : oldestDatabaseRunningXid = oldestRunningXid =
2684 1525 : XidFromFullTransactionId(TransamVariables->nextXid);
2685 :
2686 : /*
2687 : * Spin over procArray collecting all xids
2688 : */
2689 7564 : for (index = 0; index < arrayP->numProcs; index++)
2690 : {
2691 : TransactionId xid;
2692 :
2693 : /* Fetch xid just once - see GetNewTransactionId */
2694 6039 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2695 :
2696 : /*
2697 : * We don't need to store transactions that don't have a TransactionId
2698 : * yet because they will not show as running on a standby server.
2699 : */
2700 6039 : if (!TransactionIdIsValid(xid))
2701 5118 : continue;
2702 :
2703 : /*
2704 : * Be careful not to exclude any xids before calculating the values of
2705 : * oldestRunningXid and suboverflowed, since these are used to clean
2706 : * up transaction information held on standbys.
2707 : */
2708 921 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2709 626 : oldestRunningXid = xid;
2710 :
2711 : /*
2712 : * Also, update the oldest running xid within the current database. As
2713 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2714 : * TransactionId comparison first.
2715 : */
2716 921 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2717 : {
2718 921 : int pgprocno = arrayP->pgprocnos[index];
2719 921 : PGPROC *proc = &allProcs[pgprocno];
2720 :
2721 921 : if (proc->databaseId == MyDatabaseId)
2722 236 : oldestDatabaseRunningXid = xid;
2723 : }
2724 :
2725 921 : if (ProcGlobal->subxidStates[index].overflowed)
2726 0 : suboverflowed = true;
2727 :
2728 : /*
2729 : * If we wished to exclude xids this would be the right place for it.
2730 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2731 : * but they do during truncation at the end when they get the lock and
2732 : * truncate, so it is not much of a problem to include them if they
2733 : * are seen and it is cleaner to include them.
2734 : */
2735 :
2736 921 : xids[count++] = xid;
2737 : }
2738 :
2739 : /*
2740 : * Spin over procArray collecting all subxids, but only if there hasn't
2741 : * been a suboverflow.
2742 : */
2743 1525 : if (!suboverflowed)
2744 : {
2745 1525 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2746 :
2747 7564 : for (index = 0; index < arrayP->numProcs; index++)
2748 : {
2749 6039 : int pgprocno = arrayP->pgprocnos[index];
2750 6039 : PGPROC *proc = &allProcs[pgprocno];
2751 : int nsubxids;
2752 :
2753 : /*
2754 : * Save subtransaction XIDs. Other backends can't add or remove
2755 : * entries while we're holding XidGenLock.
2756 : */
2757 6039 : nsubxids = other_subxidstates[index].count;
2758 6039 : if (nsubxids > 0)
2759 : {
2760 : /* barrier not really required, as XidGenLock is held, but ... */
2761 17 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2762 :
2763 17 : memcpy(&xids[count], proc->subxids.xids,
2764 : nsubxids * sizeof(TransactionId));
2765 17 : count += nsubxids;
2766 17 : subcount += nsubxids;
2767 :
2768 : /*
2769 : * Top-level XID of a transaction is always less than any of
2770 : * its subxids, so we don't need to check if any of the
2771 : * subxids are smaller than oldestRunningXid
2772 : */
2773 : }
2774 : }
2775 : }
2776 :
2777 : /*
2778 : * It's important *not* to include the limits set by slots here because
2779 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2780 : * were to be included here the initial value could never increase because
2781 : * of a circular dependency where slots only increase their limits when
2782 : * running xacts increases oldestRunningXid and running xacts only
2783 : * increases if slots do.
2784 : */
2785 :
2786 1525 : CurrentRunningXacts->xcnt = count - subcount;
2787 1525 : CurrentRunningXacts->subxcnt = subcount;
2788 1525 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2789 1525 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2790 1525 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2791 1525 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2792 1525 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2793 :
2794 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2795 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2796 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2797 :
2798 : /* We don't release the locks here, the caller is responsible for that */
2799 :
2800 1525 : return CurrentRunningXacts;
2801 : }
2802 :
2803 : /*
2804 : * GetOldestActiveTransactionId()
2805 : *
2806 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2807 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2808 : *
2809 : * If allDbs is true, we look at all databases, though there is no need to
2810 : * include WALSender since this has no effect on hot standby conflicts. If
2811 : * allDbs is false, skip processes attached to other databases.
2812 : *
2813 : * This is never executed during recovery so there is no need to look at
2814 : * KnownAssignedXids.
2815 : *
2816 : * We don't worry about updating other counters, we want to keep this as
2817 : * simple as possible and leave GetSnapshotData() as the primary code for
2818 : * that bookkeeping.
2819 : *
2820 : * inCommitOnly indicates getting the oldestActiveXid among the transactions
2821 : * in the commit critical section.
2822 : */
2823 : TransactionId
2824 12471 : GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2825 : {
2826 12471 : ProcArrayStruct *arrayP = procArray;
2827 12471 : TransactionId *other_xids = ProcGlobal->xids;
2828 : TransactionId oldestRunningXid;
2829 : int index;
2830 :
2831 : Assert(!RecoveryInProgress());
2832 :
2833 : /*
2834 : * Read nextXid, as the upper bound of what's still active.
2835 : *
2836 : * Reading a TransactionId is atomic, but we must grab the lock to make
2837 : * sure that all XIDs < nextXid are already present in the proc array (or
2838 : * have already completed), when we spin over it.
2839 : */
2840 12471 : LWLockAcquire(XidGenLock, LW_SHARED);
2841 12471 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2842 12471 : LWLockRelease(XidGenLock);
2843 :
2844 : /*
2845 : * Spin over procArray collecting all xids and subxids.
2846 : */
2847 12471 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2848 74108 : for (index = 0; index < arrayP->numProcs; index++)
2849 : {
2850 : TransactionId xid;
2851 61637 : int pgprocno = arrayP->pgprocnos[index];
2852 61637 : PGPROC *proc = &allProcs[pgprocno];
2853 :
2854 : /* Fetch xid just once - see GetNewTransactionId */
2855 61637 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2856 :
2857 61637 : if (!TransactionIdIsNormal(xid))
2858 49364 : continue;
2859 :
2860 12273 : if (inCommitOnly &&
2861 11377 : (proc->delayChkptFlags & DELAY_CHKPT_IN_COMMIT) == 0)
2862 11124 : continue;
2863 :
2864 1149 : if (!allDbs && proc->databaseId != MyDatabaseId)
2865 0 : continue;
2866 :
2867 1149 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2868 849 : oldestRunningXid = xid;
2869 :
2870 : /*
2871 : * Top-level XID of a transaction is always less than any of its
2872 : * subxids, so we don't need to check if any of the subxids are
2873 : * smaller than oldestRunningXid
2874 : */
2875 : }
2876 12471 : LWLockRelease(ProcArrayLock);
2877 :
2878 12471 : return oldestRunningXid;
2879 : }
2880 :
2881 : /*
2882 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2883 : *
2884 : * Returns the oldest xid that we can guarantee not to have been affected by
2885 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2886 : * transaction aborted. Note that the value can (and most of the time will) be
2887 : * much more conservative than what really has been affected by vacuum, but we
2888 : * currently don't have better data available.
2889 : *
2890 : * This is useful to initialize the cutoff xid after which a new changeset
2891 : * extraction replication slot can start decoding changes.
2892 : *
2893 : * Must be called with ProcArrayLock held either shared or exclusively,
2894 : * although most callers will want to use exclusive mode since it is expected
2895 : * that the caller will immediately use the xid to peg the xmin horizon.
2896 : */
2897 : TransactionId
2898 737 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2899 : {
2900 737 : ProcArrayStruct *arrayP = procArray;
2901 : TransactionId oldestSafeXid;
2902 : int index;
2903 737 : bool recovery_in_progress = RecoveryInProgress();
2904 :
2905 : Assert(LWLockHeldByMe(ProcArrayLock));
2906 :
2907 : /*
2908 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2909 : * running. If no transaction with xid were running concurrently a new xid
2910 : * could influence the RecentXmin et al.
2911 : *
2912 : * We initialize the computation to nextXid since that's guaranteed to be
2913 : * a safe, albeit pessimal, value.
2914 : */
2915 737 : LWLockAcquire(XidGenLock, LW_SHARED);
2916 737 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2917 :
2918 : /*
2919 : * If there's already a slot pegging the xmin horizon, we can start with
2920 : * that value, it's guaranteed to be safe since it's computed by this
2921 : * routine initially and has been enforced since. We can always use the
2922 : * slot's general xmin horizon, but the catalog horizon is only usable
2923 : * when only catalog data is going to be looked at.
2924 : */
2925 977 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2926 240 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2927 : oldestSafeXid))
2928 12 : oldestSafeXid = procArray->replication_slot_xmin;
2929 :
2930 737 : if (catalogOnly &&
2931 368 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2932 77 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2933 : oldestSafeXid))
2934 27 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2935 :
2936 : /*
2937 : * If we're not in recovery, we walk over the procarray and collect the
2938 : * lowest xid. Since we're called with ProcArrayLock held and have
2939 : * acquired XidGenLock, no entries can vanish concurrently, since
2940 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2941 : * with ProcArrayLock held.
2942 : *
2943 : * In recovery we can't lower the safe value besides what we've computed
2944 : * above, so we'll have to wait a bit longer there. We unfortunately can
2945 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2946 : * machinery can miss values and return an older value than is safe.
2947 : */
2948 737 : if (!recovery_in_progress)
2949 : {
2950 703 : TransactionId *other_xids = ProcGlobal->xids;
2951 :
2952 : /*
2953 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2954 : */
2955 3615 : for (index = 0; index < arrayP->numProcs; index++)
2956 : {
2957 : TransactionId xid;
2958 :
2959 : /* Fetch xid just once - see GetNewTransactionId */
2960 2912 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2961 :
2962 2912 : if (!TransactionIdIsNormal(xid))
2963 2903 : continue;
2964 :
2965 9 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2966 8 : oldestSafeXid = xid;
2967 : }
2968 : }
2969 :
2970 737 : LWLockRelease(XidGenLock);
2971 :
2972 737 : return oldestSafeXid;
2973 : }
2974 :
2975 : /*
2976 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2977 : * delaying checkpoint because they have critical actions in progress.
2978 : *
2979 : * Constructs an array of VXIDs of transactions that are currently in commit
2980 : * critical sections, as shown by having specified delayChkptFlags bits set
2981 : * in their PGPROC.
2982 : *
2983 : * Returns a palloc'd array that should be freed by the caller.
2984 : * *nvxids is the number of valid entries.
2985 : *
2986 : * Note that because backends set or clear delayChkptFlags without holding any
2987 : * lock, the result is somewhat indeterminate, but we don't really care. Even
2988 : * in a multiprocessor with delayed writes to shared memory, it should be
2989 : * certain that setting of delayChkptFlags will propagate to shared memory
2990 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
2991 : * delayChkptFlags if it's already inserted its commit record. Whether it
2992 : * takes a little while for clearing of delayChkptFlags to propagate is
2993 : * unimportant for correctness.
2994 : */
2995 : VirtualTransactionId *
2996 3476 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
2997 : {
2998 : VirtualTransactionId *vxids;
2999 3476 : ProcArrayStruct *arrayP = procArray;
3000 3476 : int count = 0;
3001 : int index;
3002 :
3003 : Assert(type != 0);
3004 :
3005 : /* allocate what's certainly enough result space */
3006 3476 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3007 :
3008 3476 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3009 :
3010 11393 : for (index = 0; index < arrayP->numProcs; index++)
3011 : {
3012 7917 : int pgprocno = arrayP->pgprocnos[index];
3013 7917 : PGPROC *proc = &allProcs[pgprocno];
3014 :
3015 7917 : if ((proc->delayChkptFlags & type) != 0)
3016 : {
3017 : VirtualTransactionId vxid;
3018 :
3019 37 : GET_VXID_FROM_PGPROC(vxid, *proc);
3020 37 : if (VirtualTransactionIdIsValid(vxid))
3021 37 : vxids[count++] = vxid;
3022 : }
3023 : }
3024 :
3025 3476 : LWLockRelease(ProcArrayLock);
3026 :
3027 3476 : *nvxids = count;
3028 3476 : return vxids;
3029 : }
3030 :
3031 : /*
3032 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3033 : *
3034 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3035 : * of the specified VXIDs are still in critical sections of code.
3036 : *
3037 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3038 : * those numbers should be small enough for it not to be a problem.
3039 : */
3040 : bool
3041 34 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3042 : {
3043 34 : bool result = false;
3044 34 : ProcArrayStruct *arrayP = procArray;
3045 : int index;
3046 :
3047 : Assert(type != 0);
3048 :
3049 34 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3050 :
3051 400 : for (index = 0; index < arrayP->numProcs; index++)
3052 : {
3053 367 : int pgprocno = arrayP->pgprocnos[index];
3054 367 : PGPROC *proc = &allProcs[pgprocno];
3055 : VirtualTransactionId vxid;
3056 :
3057 367 : GET_VXID_FROM_PGPROC(vxid, *proc);
3058 :
3059 367 : if ((proc->delayChkptFlags & type) != 0 &&
3060 4 : VirtualTransactionIdIsValid(vxid))
3061 : {
3062 : int i;
3063 :
3064 9 : for (i = 0; i < nvxids; i++)
3065 : {
3066 6 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3067 : {
3068 1 : result = true;
3069 1 : break;
3070 : }
3071 : }
3072 4 : if (result)
3073 1 : break;
3074 : }
3075 : }
3076 :
3077 34 : LWLockRelease(ProcArrayLock);
3078 :
3079 34 : return result;
3080 : }
3081 :
3082 : /*
3083 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3084 : *
3085 : * The result may be out of date arbitrarily quickly, so the caller
3086 : * must be careful about how this information is used. NULL is
3087 : * returned if the backend is not active.
3088 : */
3089 : PGPROC *
3090 735 : ProcNumberGetProc(ProcNumber procNumber)
3091 : {
3092 : PGPROC *result;
3093 :
3094 735 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3095 1 : return NULL;
3096 734 : result = GetPGProcByNumber(procNumber);
3097 :
3098 734 : if (result->pid == 0)
3099 3 : return NULL;
3100 :
3101 731 : return result;
3102 : }
3103 :
3104 : /*
3105 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3106 : *
3107 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3108 : * result may be out of date arbitrarily quickly, so the caller must be
3109 : * careful about how this information is used.
3110 : */
3111 : void
3112 8213 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3113 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3114 : {
3115 : PGPROC *proc;
3116 :
3117 8213 : *xid = InvalidTransactionId;
3118 8213 : *xmin = InvalidTransactionId;
3119 8213 : *nsubxid = 0;
3120 8213 : *overflowed = false;
3121 :
3122 8213 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3123 0 : return;
3124 8213 : proc = GetPGProcByNumber(procNumber);
3125 :
3126 : /* Need to lock out additions/removals of backends */
3127 8213 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3128 :
3129 8213 : if (proc->pid != 0)
3130 : {
3131 8213 : *xid = proc->xid;
3132 8213 : *xmin = proc->xmin;
3133 8213 : *nsubxid = proc->subxidStatus.count;
3134 8213 : *overflowed = proc->subxidStatus.overflowed;
3135 : }
3136 :
3137 8213 : LWLockRelease(ProcArrayLock);
3138 : }
3139 :
3140 : /*
3141 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3142 : *
3143 : * Returns NULL if not found. Note that it is up to the caller to be
3144 : * sure that the question remains meaningful for long enough for the
3145 : * answer to be used ...
3146 : */
3147 : PGPROC *
3148 9114 : BackendPidGetProc(int pid)
3149 : {
3150 : PGPROC *result;
3151 :
3152 9114 : if (pid == 0) /* never match dummy PGPROCs */
3153 4 : return NULL;
3154 :
3155 9110 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3156 :
3157 9110 : result = BackendPidGetProcWithLock(pid);
3158 :
3159 9110 : LWLockRelease(ProcArrayLock);
3160 :
3161 9110 : return result;
3162 : }
3163 :
3164 : /*
3165 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3166 : *
3167 : * Same as above, except caller must be holding ProcArrayLock. The found
3168 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3169 : */
3170 : PGPROC *
3171 10838 : BackendPidGetProcWithLock(int pid)
3172 : {
3173 10838 : PGPROC *result = NULL;
3174 10838 : ProcArrayStruct *arrayP = procArray;
3175 : int index;
3176 :
3177 10838 : if (pid == 0) /* never match dummy PGPROCs */
3178 0 : return NULL;
3179 :
3180 41500 : for (index = 0; index < arrayP->numProcs; index++)
3181 : {
3182 37643 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3183 :
3184 37643 : if (proc->pid == pid)
3185 : {
3186 6981 : result = proc;
3187 6981 : break;
3188 : }
3189 : }
3190 :
3191 10838 : return result;
3192 : }
3193 :
3194 : /*
3195 : * BackendXidGetPid -- get a backend's pid given its XID
3196 : *
3197 : * Returns 0 if not found or it's a prepared transaction. Note that
3198 : * it is up to the caller to be sure that the question remains
3199 : * meaningful for long enough for the answer to be used ...
3200 : *
3201 : * Only main transaction Ids are considered. This function is mainly
3202 : * useful for determining what backend owns a lock.
3203 : *
3204 : * Beware that not every xact has an XID assigned. However, as long as you
3205 : * only call this using an XID found on disk, you're safe.
3206 : */
3207 : int
3208 30 : BackendXidGetPid(TransactionId xid)
3209 : {
3210 30 : int result = 0;
3211 30 : ProcArrayStruct *arrayP = procArray;
3212 30 : TransactionId *other_xids = ProcGlobal->xids;
3213 : int index;
3214 :
3215 30 : if (xid == InvalidTransactionId) /* never match invalid xid */
3216 0 : return 0;
3217 :
3218 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3219 :
3220 92 : for (index = 0; index < arrayP->numProcs; index++)
3221 : {
3222 84 : if (other_xids[index] == xid)
3223 : {
3224 22 : int pgprocno = arrayP->pgprocnos[index];
3225 22 : PGPROC *proc = &allProcs[pgprocno];
3226 :
3227 22 : result = proc->pid;
3228 22 : break;
3229 : }
3230 : }
3231 :
3232 30 : LWLockRelease(ProcArrayLock);
3233 :
3234 30 : return result;
3235 : }
3236 :
3237 : /*
3238 : * IsBackendPid -- is a given pid a running backend
3239 : *
3240 : * This is not called by the backend, but is called by external modules.
3241 : */
3242 : bool
3243 2 : IsBackendPid(int pid)
3244 : {
3245 2 : return (BackendPidGetProc(pid) != NULL);
3246 : }
3247 :
3248 :
3249 : /*
3250 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3251 : *
3252 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3253 : *
3254 : * The arguments allow filtering the set of VXIDs returned. Our own process
3255 : * is always skipped. In addition:
3256 : * If limitXmin is not InvalidTransactionId, skip processes with
3257 : * xmin > limitXmin.
3258 : * If excludeXmin0 is true, skip processes with xmin = 0.
3259 : * If allDbs is false, skip processes attached to other databases.
3260 : * If excludeVacuum isn't zero, skip processes for which
3261 : * (statusFlags & excludeVacuum) is not zero.
3262 : *
3263 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3264 : * allow skipping backends whose oldest live snapshot is no older than
3265 : * some snapshot we have. Since we examine the procarray with only shared
3266 : * lock, there are race conditions: a backend could set its xmin just after
3267 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3268 : * other backend could have set its xmin *before* we look. We know however
3269 : * that such a backend must have held shared ProcArrayLock overlapping our
3270 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3271 : * any snapshot the other backend is taking concurrently with our scan cannot
3272 : * consider any transactions as still running that we think are committed
3273 : * (since backends must hold ProcArrayLock exclusive to commit).
3274 : */
3275 : VirtualTransactionId *
3276 513 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3277 : bool allDbs, int excludeVacuum,
3278 : int *nvxids)
3279 : {
3280 : VirtualTransactionId *vxids;
3281 513 : ProcArrayStruct *arrayP = procArray;
3282 513 : int count = 0;
3283 : int index;
3284 :
3285 : /* allocate what's certainly enough result space */
3286 513 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3287 :
3288 513 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3289 :
3290 3223 : for (index = 0; index < arrayP->numProcs; index++)
3291 : {
3292 2710 : int pgprocno = arrayP->pgprocnos[index];
3293 2710 : PGPROC *proc = &allProcs[pgprocno];
3294 2710 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3295 :
3296 2710 : if (proc == MyProc)
3297 513 : continue;
3298 :
3299 2197 : if (excludeVacuum & statusFlags)
3300 23 : continue;
3301 :
3302 2174 : if (allDbs || proc->databaseId == MyDatabaseId)
3303 : {
3304 : /* Fetch xmin just once - might change on us */
3305 1069 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3306 :
3307 1069 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3308 637 : continue;
3309 :
3310 : /*
3311 : * InvalidTransactionId precedes all other XIDs, so a proc that
3312 : * hasn't set xmin yet will not be rejected by this test.
3313 : */
3314 864 : if (!TransactionIdIsValid(limitXmin) ||
3315 432 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3316 : {
3317 : VirtualTransactionId vxid;
3318 :
3319 396 : GET_VXID_FROM_PGPROC(vxid, *proc);
3320 396 : if (VirtualTransactionIdIsValid(vxid))
3321 396 : vxids[count++] = vxid;
3322 : }
3323 : }
3324 : }
3325 :
3326 513 : LWLockRelease(ProcArrayLock);
3327 :
3328 513 : *nvxids = count;
3329 513 : return vxids;
3330 : }
3331 :
3332 : /*
3333 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3334 : *
3335 : * Usage is limited to conflict resolution during recovery on standby servers.
3336 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3337 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3338 : * determine a safe snapshotConflictHorizon value.
3339 : *
3340 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3341 : * so we're not worried if they have a snapshot or not, nor does it really
3342 : * matter what type of lock we hold. Caller must avoid calling here with
3343 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3344 : * during original execution, since that actually indicates that there is
3345 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3346 : * convention for InvalidTransactionId values is the opposite of our own!).
3347 : *
3348 : * All callers that are checking xmins always now supply a valid and useful
3349 : * value for limitXmin. The limitXmin is always lower than the lowest
3350 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3351 : * because we only care about cleanup records that are cleaning up tuple
3352 : * versions from committed transactions. In that case they will only occur
3353 : * at the point where the record is less than the lowest running xid. That
3354 : * allows us to say that if any backend takes a snapshot concurrently with
3355 : * us then the conflict assessment made here would never include the snapshot
3356 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3357 : * concurrent snapshots when limitXmin is valid. We might think about adding
3358 : * Assert(limitXmin < lowest(KnownAssignedXids))
3359 : * but that would not be true in the case of FATAL errors lagging in array,
3360 : * but we already know those are bogus anyway, so we skip that test.
3361 : *
3362 : * If dbOid is valid we skip backends attached to other databases.
3363 : *
3364 : * Be careful to *not* pfree the result from this function. We reuse
3365 : * this array sufficiently often that we use malloc for the result.
3366 : */
3367 : VirtualTransactionId *
3368 16460 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3369 : {
3370 : static VirtualTransactionId *vxids;
3371 16460 : ProcArrayStruct *arrayP = procArray;
3372 16460 : int count = 0;
3373 : int index;
3374 :
3375 : /*
3376 : * If first time through, get workspace to remember main XIDs in. We
3377 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3378 : * result space, remembering room for a terminator.
3379 : */
3380 16460 : if (vxids == NULL)
3381 : {
3382 19 : vxids = (VirtualTransactionId *)
3383 19 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3384 19 : if (vxids == NULL)
3385 0 : ereport(ERROR,
3386 : (errcode(ERRCODE_OUT_OF_MEMORY),
3387 : errmsg("out of memory")));
3388 : }
3389 :
3390 16460 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3391 :
3392 16690 : for (index = 0; index < arrayP->numProcs; index++)
3393 : {
3394 230 : int pgprocno = arrayP->pgprocnos[index];
3395 230 : PGPROC *proc = &allProcs[pgprocno];
3396 :
3397 : /* Exclude prepared transactions */
3398 230 : if (proc->pid == 0)
3399 0 : continue;
3400 :
3401 230 : if (!OidIsValid(dbOid) ||
3402 224 : proc->databaseId == dbOid)
3403 : {
3404 : /* Fetch xmin just once - can't change on us, but good coding */
3405 17 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3406 :
3407 : /*
3408 : * We ignore an invalid pxmin because this means that backend has
3409 : * no snapshot currently. We hold a Share lock to avoid contention
3410 : * with users taking snapshots. That is not a problem because the
3411 : * current xmin is always at least one higher than the latest
3412 : * removed xid, so any new snapshot would never conflict with the
3413 : * test here.
3414 : */
3415 17 : if (!TransactionIdIsValid(limitXmin) ||
3416 3 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3417 : {
3418 : VirtualTransactionId vxid;
3419 :
3420 2 : GET_VXID_FROM_PGPROC(vxid, *proc);
3421 2 : if (VirtualTransactionIdIsValid(vxid))
3422 2 : vxids[count++] = vxid;
3423 : }
3424 : }
3425 : }
3426 :
3427 16460 : LWLockRelease(ProcArrayLock);
3428 :
3429 : /* add the terminator */
3430 16460 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3431 16460 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3432 :
3433 16460 : return vxids;
3434 : }
3435 :
3436 : /*
3437 : * SignalRecoveryConflict -- signal that a process is blocking recovery
3438 : *
3439 : * The 'pid' is redundant with 'proc', but it acts as a cross-check to
3440 : * detect process had exited and the PGPROC entry was reused for a different
3441 : * process.
3442 : *
3443 : * Returns true if the process was signaled, or false if not found.
3444 : */
3445 : bool
3446 5 : SignalRecoveryConflict(PGPROC *proc, pid_t pid, RecoveryConflictReason reason)
3447 : {
3448 5 : bool found = false;
3449 :
3450 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3451 :
3452 : /*
3453 : * Kill the pid if it's still here. If not, that's what we wanted so
3454 : * ignore any errors.
3455 : */
3456 5 : if (proc->pid == pid)
3457 : {
3458 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3459 :
3460 : /* wake up the process */
3461 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, GetNumberFromPGProc(proc));
3462 5 : found = true;
3463 : }
3464 :
3465 5 : LWLockRelease(ProcArrayLock);
3466 :
3467 5 : return found;
3468 : }
3469 :
3470 : /*
3471 : * SignalRecoveryConflictWithVirtualXID -- signal that a VXID is blocking recovery
3472 : *
3473 : * Like SignalRecoveryConflict, but the target is identified by VXID
3474 : */
3475 : bool
3476 5 : SignalRecoveryConflictWithVirtualXID(VirtualTransactionId vxid, RecoveryConflictReason reason)
3477 : {
3478 5 : ProcArrayStruct *arrayP = procArray;
3479 : int index;
3480 5 : pid_t pid = 0;
3481 :
3482 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3483 :
3484 5 : for (index = 0; index < arrayP->numProcs; index++)
3485 : {
3486 5 : int pgprocno = arrayP->pgprocnos[index];
3487 5 : PGPROC *proc = &allProcs[pgprocno];
3488 : VirtualTransactionId procvxid;
3489 :
3490 5 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3491 :
3492 5 : if (procvxid.procNumber == vxid.procNumber &&
3493 5 : procvxid.localTransactionId == vxid.localTransactionId)
3494 : {
3495 5 : pid = proc->pid;
3496 5 : if (pid != 0)
3497 : {
3498 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3499 :
3500 : /*
3501 : * Kill the pid if it's still here. If not, that's what we
3502 : * wanted so ignore any errors.
3503 : */
3504 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, vxid.procNumber);
3505 : }
3506 5 : break;
3507 : }
3508 : }
3509 :
3510 5 : LWLockRelease(ProcArrayLock);
3511 :
3512 5 : return pid != 0;
3513 : }
3514 :
3515 : /*
3516 : * SignalRecoveryConflictWithDatabase -- signal backends using specified database
3517 : *
3518 : * Like SignalRecoveryConflict, but signals all backends using the database.
3519 : */
3520 : void
3521 10 : SignalRecoveryConflictWithDatabase(Oid databaseid, RecoveryConflictReason reason)
3522 : {
3523 10 : ProcArrayStruct *arrayP = procArray;
3524 : int index;
3525 :
3526 : /* tell all backends to die */
3527 10 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3528 :
3529 20 : for (index = 0; index < arrayP->numProcs; index++)
3530 : {
3531 10 : int pgprocno = arrayP->pgprocnos[index];
3532 10 : PGPROC *proc = &allProcs[pgprocno];
3533 :
3534 10 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3535 : {
3536 : VirtualTransactionId procvxid;
3537 : pid_t pid;
3538 :
3539 10 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3540 :
3541 10 : pid = proc->pid;
3542 10 : if (pid != 0)
3543 : {
3544 10 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3545 :
3546 : /*
3547 : * Kill the pid if it's still here. If not, that's what we
3548 : * wanted so ignore any errors.
3549 : */
3550 10 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, procvxid.procNumber);
3551 : }
3552 : }
3553 : }
3554 :
3555 10 : LWLockRelease(ProcArrayLock);
3556 10 : }
3557 :
3558 : /*
3559 : * MinimumActiveBackends --- count backends (other than myself) that are
3560 : * in active transactions. Return true if the count exceeds the
3561 : * minimum threshold passed. This is used as a heuristic to decide if
3562 : * a pre-XLOG-flush delay is worthwhile during commit.
3563 : *
3564 : * Do not count backends that are blocked waiting for locks, since they are
3565 : * not going to get to run until someone else commits.
3566 : */
3567 : bool
3568 0 : MinimumActiveBackends(int min)
3569 : {
3570 0 : ProcArrayStruct *arrayP = procArray;
3571 0 : int count = 0;
3572 : int index;
3573 :
3574 : /* Quick short-circuit if no minimum is specified */
3575 0 : if (min == 0)
3576 0 : return true;
3577 :
3578 : /*
3579 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3580 : * bogus, but since we are only testing fields for zero or nonzero, it
3581 : * should be OK. The result is only used for heuristic purposes anyway...
3582 : */
3583 0 : for (index = 0; index < arrayP->numProcs; index++)
3584 : {
3585 0 : int pgprocno = arrayP->pgprocnos[index];
3586 0 : PGPROC *proc = &allProcs[pgprocno];
3587 :
3588 : /*
3589 : * Since we're not holding a lock, need to be prepared to deal with
3590 : * garbage, as someone could have incremented numProcs but not yet
3591 : * filled the structure.
3592 : *
3593 : * If someone just decremented numProcs, 'proc' could also point to a
3594 : * PGPROC entry that's no longer in the array. It still points to a
3595 : * PGPROC struct, though, because freed PGPROC entries just go to the
3596 : * free list and are recycled. Its contents are nonsense in that case,
3597 : * but that's acceptable for this function.
3598 : */
3599 0 : if (pgprocno == -1)
3600 0 : continue; /* do not count deleted entries */
3601 0 : if (proc == MyProc)
3602 0 : continue; /* do not count myself */
3603 0 : if (proc->xid == InvalidTransactionId)
3604 0 : continue; /* do not count if no XID assigned */
3605 0 : if (proc->pid == 0)
3606 0 : continue; /* do not count prepared xacts */
3607 0 : if (proc->waitLock != NULL)
3608 0 : continue; /* do not count if blocked on a lock */
3609 0 : count++;
3610 0 : if (count >= min)
3611 0 : break;
3612 : }
3613 :
3614 0 : return count >= min;
3615 : }
3616 :
3617 : /*
3618 : * CountDBBackends --- count backends that are using specified database
3619 : */
3620 : int
3621 16 : CountDBBackends(Oid databaseid)
3622 : {
3623 16 : ProcArrayStruct *arrayP = procArray;
3624 16 : int count = 0;
3625 : int index;
3626 :
3627 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3628 :
3629 23 : for (index = 0; index < arrayP->numProcs; index++)
3630 : {
3631 7 : int pgprocno = arrayP->pgprocnos[index];
3632 7 : PGPROC *proc = &allProcs[pgprocno];
3633 :
3634 7 : if (proc->pid == 0)
3635 0 : continue; /* do not count prepared xacts */
3636 7 : if (!OidIsValid(databaseid) ||
3637 7 : proc->databaseId == databaseid)
3638 2 : count++;
3639 : }
3640 :
3641 16 : LWLockRelease(ProcArrayLock);
3642 :
3643 16 : return count;
3644 : }
3645 :
3646 : /*
3647 : * CountDBConnections --- counts database backends (only regular backends)
3648 : */
3649 : int
3650 0 : CountDBConnections(Oid databaseid)
3651 : {
3652 0 : ProcArrayStruct *arrayP = procArray;
3653 0 : int count = 0;
3654 : int index;
3655 :
3656 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3657 :
3658 0 : for (index = 0; index < arrayP->numProcs; index++)
3659 : {
3660 0 : int pgprocno = arrayP->pgprocnos[index];
3661 0 : PGPROC *proc = &allProcs[pgprocno];
3662 :
3663 0 : if (proc->pid == 0)
3664 0 : continue; /* do not count prepared xacts */
3665 0 : if (proc->backendType != B_BACKEND)
3666 0 : continue; /* count only regular backend processes */
3667 0 : if (!OidIsValid(databaseid) ||
3668 0 : proc->databaseId == databaseid)
3669 0 : count++;
3670 : }
3671 :
3672 0 : LWLockRelease(ProcArrayLock);
3673 :
3674 0 : return count;
3675 : }
3676 :
3677 : /*
3678 : * CountUserBackends --- count backends that are used by specified user
3679 : * (only regular backends, not any type of background worker)
3680 : */
3681 : int
3682 0 : CountUserBackends(Oid roleid)
3683 : {
3684 0 : ProcArrayStruct *arrayP = procArray;
3685 0 : int count = 0;
3686 : int index;
3687 :
3688 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3689 :
3690 0 : for (index = 0; index < arrayP->numProcs; index++)
3691 : {
3692 0 : int pgprocno = arrayP->pgprocnos[index];
3693 0 : PGPROC *proc = &allProcs[pgprocno];
3694 :
3695 0 : if (proc->pid == 0)
3696 0 : continue; /* do not count prepared xacts */
3697 0 : if (proc->backendType != B_BACKEND)
3698 0 : continue; /* count only regular backend processes */
3699 0 : if (proc->roleId == roleid)
3700 0 : count++;
3701 : }
3702 :
3703 0 : LWLockRelease(ProcArrayLock);
3704 :
3705 0 : return count;
3706 : }
3707 :
3708 : /*
3709 : * CountOtherDBBackends -- check for other backends running in the given DB
3710 : *
3711 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3712 : * for them to exit (or 0.3s for testing purposes). Autovacuum backends are
3713 : * encouraged to exit early by sending them SIGTERM, but normal user backends
3714 : * are just waited for. If background workers connected to this database are
3715 : * marked as interruptible, they are terminated.
3716 : *
3717 : * The current backend is always ignored; it is caller's responsibility to
3718 : * check whether the current backend uses the given DB, if it's important.
3719 : *
3720 : * Returns true if there are (still) other backends in the DB, false if not.
3721 : * Also, *nbackends and *nprepared are set to the number of other backends
3722 : * and prepared transactions in the DB, respectively.
3723 : *
3724 : * This function is used to interlock DROP DATABASE and related commands
3725 : * against there being any active backends in the target DB --- dropping the
3726 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3727 : * detect here the possibility of a newly-started backend that is trying to
3728 : * connect to the doomed database, so additional interlocking is needed during
3729 : * backend startup. The caller should normally hold an exclusive lock on the
3730 : * target DB before calling this, which is one reason we mustn't wait
3731 : * indefinitely.
3732 : */
3733 : bool
3734 503 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3735 : {
3736 503 : ProcArrayStruct *arrayP = procArray;
3737 :
3738 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3739 : int autovac_pids[MAXAUTOVACPIDS];
3740 :
3741 : /*
3742 : * Retry up to 50 times with 100ms between attempts (max 5s total). Can be
3743 : * reduced to 3 attempts (max 0.3s total) to speed up tests.
3744 : */
3745 503 : int ntries = 50;
3746 :
3747 : #ifdef USE_INJECTION_POINTS
3748 503 : if (IS_INJECTION_POINT_ATTACHED("procarray-reduce-count"))
3749 1 : ntries = 3;
3750 : #endif
3751 :
3752 510 : for (int tries = 0; tries < ntries; tries++)
3753 : {
3754 509 : int nautovacs = 0;
3755 509 : bool found = false;
3756 : int index;
3757 :
3758 509 : CHECK_FOR_INTERRUPTS();
3759 :
3760 509 : *nbackends = *nprepared = 0;
3761 :
3762 509 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3763 :
3764 1921 : for (index = 0; index < arrayP->numProcs; index++)
3765 : {
3766 1412 : int pgprocno = arrayP->pgprocnos[index];
3767 1412 : PGPROC *proc = &allProcs[pgprocno];
3768 1412 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3769 :
3770 1412 : if (proc->databaseId != databaseId)
3771 1290 : continue;
3772 122 : if (proc == MyProc)
3773 115 : continue;
3774 :
3775 7 : found = true;
3776 :
3777 7 : if (proc->pid == 0)
3778 0 : (*nprepared)++;
3779 : else
3780 : {
3781 7 : (*nbackends)++;
3782 7 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3783 : nautovacs < MAXAUTOVACPIDS)
3784 0 : autovac_pids[nautovacs++] = proc->pid;
3785 : }
3786 : }
3787 :
3788 509 : LWLockRelease(ProcArrayLock);
3789 :
3790 509 : if (!found)
3791 502 : return false; /* no conflicting backends, so done */
3792 :
3793 : /*
3794 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3795 : * postpone this step until after the loop because we don't want to
3796 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3797 : * block kill() inside the kernel...
3798 : */
3799 7 : for (index = 0; index < nautovacs; index++)
3800 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3801 :
3802 : /*
3803 : * Terminate all background workers for this database, if they have
3804 : * requested it (BGWORKER_INTERRUPTIBLE).
3805 : */
3806 7 : TerminateBackgroundWorkersForDatabase(databaseId);
3807 :
3808 : /* sleep, then try again */
3809 7 : pg_usleep(100 * 1000L); /* 100ms */
3810 : }
3811 :
3812 1 : return true; /* timed out, still conflicts */
3813 : }
3814 :
3815 : /*
3816 : * Terminate existing connections to the specified database. This routine
3817 : * is used by the DROP DATABASE command when user has asked to forcefully
3818 : * drop the database.
3819 : *
3820 : * The current backend is always ignored; it is caller's responsibility to
3821 : * check whether the current backend uses the given DB, if it's important.
3822 : *
3823 : * If the target database has a prepared transaction or permissions checks
3824 : * fail for a connection, this fails without terminating anything.
3825 : */
3826 : void
3827 1 : TerminateOtherDBBackends(Oid databaseId)
3828 : {
3829 1 : ProcArrayStruct *arrayP = procArray;
3830 1 : List *pids = NIL;
3831 1 : int nprepared = 0;
3832 : int i;
3833 :
3834 1 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3835 :
3836 4 : for (i = 0; i < procArray->numProcs; i++)
3837 : {
3838 3 : int pgprocno = arrayP->pgprocnos[i];
3839 3 : PGPROC *proc = &allProcs[pgprocno];
3840 :
3841 3 : if (proc->databaseId != databaseId)
3842 3 : continue;
3843 0 : if (proc == MyProc)
3844 0 : continue;
3845 :
3846 0 : if (proc->pid != 0)
3847 0 : pids = lappend_int(pids, proc->pid);
3848 : else
3849 0 : nprepared++;
3850 : }
3851 :
3852 1 : LWLockRelease(ProcArrayLock);
3853 :
3854 1 : if (nprepared > 0)
3855 0 : ereport(ERROR,
3856 : (errcode(ERRCODE_OBJECT_IN_USE),
3857 : errmsg("database \"%s\" is being used by prepared transactions",
3858 : get_database_name(databaseId)),
3859 : errdetail_plural("There is %d prepared transaction using the database.",
3860 : "There are %d prepared transactions using the database.",
3861 : nprepared,
3862 : nprepared)));
3863 :
3864 1 : if (pids)
3865 : {
3866 : ListCell *lc;
3867 :
3868 : /*
3869 : * Permissions checks relax the pg_terminate_backend checks in two
3870 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3871 : *
3872 : * - Accept terminating autovacuum workers, since DROP DATABASE
3873 : * without FORCE terminates them.
3874 : *
3875 : * - Accept terminating bgworkers. For bgworker authors, it's
3876 : * convenient to be able to recommend FORCE if a worker is blocking
3877 : * DROP DATABASE unexpectedly.
3878 : *
3879 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3880 : * "PID %d is not a PostgreSQL server process", because for us already
3881 : * finished session is not a problem.
3882 : */
3883 0 : foreach(lc, pids)
3884 : {
3885 0 : int pid = lfirst_int(lc);
3886 0 : PGPROC *proc = BackendPidGetProc(pid);
3887 :
3888 0 : if (proc != NULL)
3889 : {
3890 0 : if (superuser_arg(proc->roleId) && !superuser())
3891 0 : ereport(ERROR,
3892 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3893 : errmsg("permission denied to terminate process"),
3894 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3895 : "SUPERUSER", "SUPERUSER")));
3896 :
3897 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3898 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3899 0 : ereport(ERROR,
3900 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3901 : errmsg("permission denied to terminate process"),
3902 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3903 : "pg_signal_backend")));
3904 : }
3905 : }
3906 :
3907 : /*
3908 : * There's a race condition here: once we release the ProcArrayLock,
3909 : * it's possible for the session to exit before we issue kill. That
3910 : * race condition possibility seems too unlikely to worry about. See
3911 : * pg_signal_backend.
3912 : */
3913 0 : foreach(lc, pids)
3914 : {
3915 0 : int pid = lfirst_int(lc);
3916 0 : PGPROC *proc = BackendPidGetProc(pid);
3917 :
3918 0 : if (proc != NULL)
3919 : {
3920 : /*
3921 : * If we have setsid(), signal the backend's whole process
3922 : * group
3923 : */
3924 : #ifdef HAVE_SETSID
3925 0 : (void) kill(-pid, SIGTERM);
3926 : #else
3927 : (void) kill(pid, SIGTERM);
3928 : #endif
3929 : }
3930 : }
3931 : }
3932 1 : }
3933 :
3934 : /*
3935 : * ProcArraySetReplicationSlotXmin
3936 : *
3937 : * Install limits to future computations of the xmin horizon to prevent vacuum
3938 : * and HOT pruning from removing affected rows still needed by clients with
3939 : * replication slots.
3940 : */
3941 : void
3942 2647 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3943 : bool already_locked)
3944 : {
3945 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3946 :
3947 2647 : if (!already_locked)
3948 2131 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3949 :
3950 2647 : procArray->replication_slot_xmin = xmin;
3951 2647 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3952 :
3953 2647 : if (!already_locked)
3954 2131 : LWLockRelease(ProcArrayLock);
3955 :
3956 2647 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3957 : xmin, catalog_xmin);
3958 2647 : }
3959 :
3960 : /*
3961 : * ProcArrayGetReplicationSlotXmin
3962 : *
3963 : * Return the current slot xmin limits. That's useful to be able to remove
3964 : * data that's older than those limits.
3965 : */
3966 : void
3967 22 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3968 : TransactionId *catalog_xmin)
3969 : {
3970 22 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3971 :
3972 22 : if (xmin != NULL)
3973 0 : *xmin = procArray->replication_slot_xmin;
3974 :
3975 22 : if (catalog_xmin != NULL)
3976 22 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3977 :
3978 22 : LWLockRelease(ProcArrayLock);
3979 22 : }
3980 :
3981 : /*
3982 : * XidCacheRemoveRunningXids
3983 : *
3984 : * Remove a bunch of TransactionIds from the list of known-running
3985 : * subtransactions for my backend. Both the specified xid and those in
3986 : * the xids[] array (of length nxids) are removed from the subxids cache.
3987 : * latestXid must be the latest XID among the group.
3988 : */
3989 : void
3990 856 : XidCacheRemoveRunningXids(TransactionId xid,
3991 : int nxids, const TransactionId *xids,
3992 : TransactionId latestXid)
3993 : {
3994 : int i,
3995 : j;
3996 : XidCacheStatus *mysubxidstat;
3997 :
3998 : Assert(TransactionIdIsValid(xid));
3999 :
4000 : /*
4001 : * We must hold ProcArrayLock exclusively in order to remove transactions
4002 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
4003 : * possible this could be relaxed since we know this routine is only used
4004 : * to abort subtransactions, but pending closer analysis we'd best be
4005 : * conservative.
4006 : *
4007 : * Note that we do not have to be careful about memory ordering of our own
4008 : * reads wrt. GetNewTransactionId() here - only this process can modify
4009 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
4010 : * careful about our own writes being well ordered.
4011 : */
4012 856 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4013 :
4014 856 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4015 :
4016 : /*
4017 : * Under normal circumstances xid and xids[] will be in increasing order,
4018 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4019 : * behavior when removing a lot of xids.
4020 : */
4021 887 : for (i = nxids - 1; i >= 0; i--)
4022 : {
4023 31 : TransactionId anxid = xids[i];
4024 :
4025 31 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4026 : {
4027 31 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4028 : {
4029 31 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4030 31 : pg_write_barrier();
4031 31 : mysubxidstat->count--;
4032 31 : MyProc->subxidStatus.count--;
4033 31 : break;
4034 : }
4035 : }
4036 :
4037 : /*
4038 : * Ordinarily we should have found it, unless the cache has
4039 : * overflowed. However it's also possible for this routine to be
4040 : * invoked multiple times for the same subtransaction, in case of an
4041 : * error during AbortSubTransaction. So instead of Assert, emit a
4042 : * debug warning.
4043 : */
4044 31 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4045 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4046 : }
4047 :
4048 920 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4049 : {
4050 919 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4051 : {
4052 855 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4053 855 : pg_write_barrier();
4054 855 : mysubxidstat->count--;
4055 855 : MyProc->subxidStatus.count--;
4056 855 : break;
4057 : }
4058 : }
4059 : /* Ordinarily we should have found it, unless the cache has overflowed */
4060 856 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4061 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4062 :
4063 : /* Also advance global latestCompletedXid while holding the lock */
4064 856 : MaintainLatestCompletedXid(latestXid);
4065 :
4066 : /* ... and xactCompletionCount */
4067 856 : TransamVariables->xactCompletionCount++;
4068 :
4069 856 : LWLockRelease(ProcArrayLock);
4070 856 : }
4071 :
4072 : #ifdef XIDCACHE_DEBUG
4073 :
4074 : /*
4075 : * Print stats about effectiveness of XID cache
4076 : */
4077 : static void
4078 : DisplayXidCache(void)
4079 : {
4080 : fprintf(stderr,
4081 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4082 : xc_by_recent_xmin,
4083 : xc_by_known_xact,
4084 : xc_by_my_xact,
4085 : xc_by_latest_xid,
4086 : xc_by_main_xid,
4087 : xc_by_child_xid,
4088 : xc_by_known_assigned,
4089 : xc_no_overflow,
4090 : xc_slow_answer);
4091 : }
4092 : #endif /* XIDCACHE_DEBUG */
4093 :
4094 : /*
4095 : * If rel != NULL, return test state appropriate for relation, otherwise
4096 : * return state usable for all relations. The latter may consider XIDs as
4097 : * not-yet-visible-to-everyone that a state for a specific relation would
4098 : * already consider visible-to-everyone.
4099 : *
4100 : * This needs to be called while a snapshot is active or registered, otherwise
4101 : * there are wraparound and other dangers.
4102 : *
4103 : * See comment for GlobalVisState for details.
4104 : */
4105 : GlobalVisState *
4106 20390569 : GlobalVisTestFor(Relation rel)
4107 : {
4108 20390569 : GlobalVisState *state = NULL;
4109 :
4110 : /* XXX: we should assert that a snapshot is pushed or registered */
4111 : Assert(RecentXmin);
4112 :
4113 20390569 : switch (GlobalVisHorizonKindForRel(rel))
4114 : {
4115 147595 : case VISHORIZON_SHARED:
4116 147595 : state = &GlobalVisSharedRels;
4117 147595 : break;
4118 4702408 : case VISHORIZON_CATALOG:
4119 4702408 : state = &GlobalVisCatalogRels;
4120 4702408 : break;
4121 15412669 : case VISHORIZON_DATA:
4122 15412669 : state = &GlobalVisDataRels;
4123 15412669 : break;
4124 127897 : case VISHORIZON_TEMP:
4125 127897 : state = &GlobalVisTempRels;
4126 127897 : break;
4127 : }
4128 :
4129 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4130 : FullTransactionIdIsValid(state->maybe_needed));
4131 :
4132 20390569 : return state;
4133 : }
4134 :
4135 : /*
4136 : * Return true if it's worth updating the accurate maybe_needed boundary.
4137 : *
4138 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4139 : * repeatedly do so when there is a low likelihood of it being beneficial.
4140 : *
4141 : * The current heuristic is that we update only if RecentXmin has changed
4142 : * since the last update. If the oldest currently running transaction has not
4143 : * finished, it is unlikely that recomputing the horizon would be useful.
4144 : */
4145 : static bool
4146 724249 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4147 : {
4148 : /* hasn't been updated yet */
4149 724249 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4150 16679 : return true;
4151 :
4152 : /*
4153 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4154 : * unlikely to be beneficial to refresh boundaries.
4155 : */
4156 707570 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4157 : state->definitely_needed))
4158 0 : return false;
4159 :
4160 : /* does the last snapshot built have a different xmin? */
4161 707570 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4162 : }
4163 :
4164 : static void
4165 242441 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4166 : {
4167 : GlobalVisSharedRels.maybe_needed =
4168 242441 : FullXidRelativeTo(horizons->latest_completed,
4169 : horizons->shared_oldest_nonremovable);
4170 : GlobalVisCatalogRels.maybe_needed =
4171 242441 : FullXidRelativeTo(horizons->latest_completed,
4172 : horizons->catalog_oldest_nonremovable);
4173 : GlobalVisDataRels.maybe_needed =
4174 242441 : FullXidRelativeTo(horizons->latest_completed,
4175 : horizons->data_oldest_nonremovable);
4176 : GlobalVisTempRels.maybe_needed =
4177 242441 : FullXidRelativeTo(horizons->latest_completed,
4178 : horizons->temp_oldest_nonremovable);
4179 :
4180 : /*
4181 : * In longer running transactions it's possible that transactions we
4182 : * previously needed to treat as running aren't around anymore. So update
4183 : * definitely_needed to not be earlier than maybe_needed.
4184 : */
4185 : GlobalVisSharedRels.definitely_needed =
4186 242441 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4187 : GlobalVisSharedRels.definitely_needed);
4188 : GlobalVisCatalogRels.definitely_needed =
4189 242441 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4190 : GlobalVisCatalogRels.definitely_needed);
4191 : GlobalVisDataRels.definitely_needed =
4192 242441 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4193 : GlobalVisDataRels.definitely_needed);
4194 242441 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4195 :
4196 242441 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4197 242441 : }
4198 :
4199 : /*
4200 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4201 : * using ComputeXidHorizons().
4202 : */
4203 : static void
4204 85778 : GlobalVisUpdate(void)
4205 : {
4206 : ComputeXidHorizonsResult horizons;
4207 :
4208 : /* updates the horizons as a side-effect */
4209 85778 : ComputeXidHorizons(&horizons);
4210 85778 : }
4211 :
4212 : /*
4213 : * Return true if no snapshot still considers fxid to be running.
4214 : *
4215 : * The state passed needs to have been initialized for the relation fxid is
4216 : * from (NULL is also OK), otherwise the result may not be correct.
4217 : *
4218 : * If allow_update is false, the GlobalVisState boundaries will not be updated
4219 : * even if it would otherwise be beneficial. This is useful for callers that
4220 : * do not want GlobalVisState to advance at all, for example because they need
4221 : * a conservative answer based on the current boundaries.
4222 : *
4223 : * See comment for GlobalVisState for details.
4224 : */
4225 : bool
4226 16054887 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4227 : FullTransactionId fxid,
4228 : bool allow_update)
4229 : {
4230 : /*
4231 : * If fxid is older than maybe_needed bound, it definitely is visible to
4232 : * everyone.
4233 : */
4234 16054887 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4235 6188744 : return true;
4236 :
4237 : /*
4238 : * If fxid is >= definitely_needed bound, it is very likely to still be
4239 : * considered running.
4240 : */
4241 9866143 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4242 9141894 : return false;
4243 :
4244 : /*
4245 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4246 : * might not exist a snapshot considering fxid running. If it makes sense,
4247 : * update boundaries and recheck.
4248 : */
4249 724249 : if (allow_update && GlobalVisTestShouldUpdate(state))
4250 : {
4251 85778 : GlobalVisUpdate();
4252 :
4253 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4254 :
4255 85778 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4256 : }
4257 : else
4258 638471 : return false;
4259 : }
4260 :
4261 : /*
4262 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4263 : *
4264 : * It is crucial that this only gets called for xids from a source that
4265 : * protects against xid wraparounds (e.g. from a table and thus protected by
4266 : * relfrozenxid).
4267 : */
4268 : bool
4269 16054391 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid,
4270 : bool allow_update)
4271 : {
4272 : FullTransactionId fxid;
4273 :
4274 : /*
4275 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4276 : * because we know the xid has to, at the very least, be between
4277 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4278 : * lock to determine either, we can just compare with
4279 : * state->definitely_needed, which was based on those value at the time
4280 : * the current snapshot was built.
4281 : */
4282 16054391 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4283 :
4284 16054391 : return GlobalVisTestIsRemovableFullXid(state, fxid, allow_update);
4285 : }
4286 :
4287 : /*
4288 : * Wrapper around GlobalVisTestIsRemovableXid() for use when examining live
4289 : * tuples. Returns true if the given XID may be considered running by at least
4290 : * one snapshot.
4291 : *
4292 : * This function alone is insufficient to determine tuple visibility; callers
4293 : * must also consider the XID's commit status. Its purpose is purely semantic:
4294 : * when applied to live tuples, GlobalVisTestIsRemovableXid() is checking
4295 : * whether the inserting transaction is still considered running, not whether
4296 : * the tuple is removable. Live tuples are, by definition, not removable, but
4297 : * the snapshot criteria for "transaction still running" are identical to
4298 : * those used for removal XIDs.
4299 : *
4300 : * If allow_update is true, the GlobalVisState boundaries may be updated. If
4301 : * it is false, they definitely will not be updated.
4302 : *
4303 : * See the comment above GlobalVisTestIsRemovable[Full]Xid() for details on
4304 : * the required preconditions for calling this function.
4305 : */
4306 : bool
4307 78852 : GlobalVisTestXidConsideredRunning(GlobalVisState *state, TransactionId xid,
4308 : bool allow_update)
4309 : {
4310 78852 : return !GlobalVisTestIsRemovableXid(state, xid, allow_update);
4311 : }
4312 :
4313 : /*
4314 : * Convenience wrapper around GlobalVisTestFor() and
4315 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4316 : */
4317 : bool
4318 496 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4319 : {
4320 : GlobalVisState *state;
4321 :
4322 496 : state = GlobalVisTestFor(rel);
4323 :
4324 496 : return GlobalVisTestIsRemovableFullXid(state, fxid, true);
4325 : }
4326 :
4327 : /*
4328 : * Convenience wrapper around GlobalVisTestFor() and
4329 : * GlobalVisTestIsRemovableXid(), see their comments.
4330 : */
4331 : bool
4332 8 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4333 : {
4334 : GlobalVisState *state;
4335 :
4336 8 : state = GlobalVisTestFor(rel);
4337 :
4338 8 : return GlobalVisTestIsRemovableXid(state, xid, true);
4339 : }
4340 :
4341 : /*
4342 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4343 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4344 : *
4345 : * Be very careful about when to use this function. It can only safely be used
4346 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4347 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4348 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4349 : * the xid has to be within that range), or if xid is from the procarray and
4350 : * prevents xid wraparound that way.
4351 : */
4352 : static inline FullTransactionId
4353 18782031 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4354 : {
4355 18782031 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4356 :
4357 : Assert(TransactionIdIsValid(xid));
4358 : Assert(TransactionIdIsValid(rel_xid));
4359 :
4360 : /* not guaranteed to find issues, but likely to catch mistakes */
4361 : AssertTransactionIdInAllowableRange(xid);
4362 :
4363 37564062 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4364 18782031 : + (int32) (xid - rel_xid));
4365 : }
4366 :
4367 :
4368 : /* ----------------------------------------------
4369 : * KnownAssignedTransactionIds sub-module
4370 : * ----------------------------------------------
4371 : */
4372 :
4373 : /*
4374 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4375 : * running on the primary at the current point in WAL. These XIDs must be
4376 : * treated as running by standby transactions, even though they are not in
4377 : * the standby server's PGPROC array.
4378 : *
4379 : * We record all XIDs that we know have been assigned. That includes all the
4380 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4381 : * been assigned. We can deduce the existence of unobserved XIDs because we
4382 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4383 : * list expands as new XIDs are observed or inferred, and contracts when
4384 : * transaction completion records arrive.
4385 : *
4386 : * During hot standby we do not fret too much about the distinction between
4387 : * top-level XIDs and subtransaction XIDs. We store both together in the
4388 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4389 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4390 : * doesn't care about the distinction either. Subtransaction XIDs are
4391 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4392 : * links are *not* maintained (which does not affect visibility).
4393 : *
4394 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4395 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4396 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4397 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4398 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4399 : * and then remove them from KnownAssignedXids. This prevents overflow of
4400 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4401 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4402 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4403 : * though it should be complete for top-level XIDs; this is the same situation
4404 : * that holds with respect to the PGPROC entries in normal running.
4405 : *
4406 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4407 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4408 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4409 : * As long as that is within the range of interesting XIDs, we have to assume
4410 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4411 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4412 : * subXID arrives - that is not an error.)
4413 : *
4414 : * Should a backend on primary somehow disappear before it can write an abort
4415 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4416 : * aborted but we think they were running; the distinction is irrelevant
4417 : * because either way any changes done by the transaction are not visible to
4418 : * backends in the standby. We prune KnownAssignedXids when
4419 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4420 : * array due to such dead XIDs.
4421 : */
4422 :
4423 : /*
4424 : * RecordKnownAssignedTransactionIds
4425 : * Record the given XID in KnownAssignedXids, as well as any preceding
4426 : * unobserved XIDs.
4427 : *
4428 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4429 : * associated with a transaction. Must be called for each record after we
4430 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4431 : *
4432 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4433 : */
4434 : void
4435 2604151 : RecordKnownAssignedTransactionIds(TransactionId xid)
4436 : {
4437 : Assert(standbyState >= STANDBY_INITIALIZED);
4438 : Assert(TransactionIdIsValid(xid));
4439 : Assert(TransactionIdIsValid(latestObservedXid));
4440 :
4441 2604151 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4442 : xid, latestObservedXid);
4443 :
4444 : /*
4445 : * When a newly observed xid arrives, it is frequently the case that it is
4446 : * *not* the next xid in sequence. When this occurs, we must treat the
4447 : * intervening xids as running also.
4448 : */
4449 2604151 : if (TransactionIdFollows(xid, latestObservedXid))
4450 : {
4451 : TransactionId next_expected_xid;
4452 :
4453 : /*
4454 : * Extend subtrans like we do in GetNewTransactionId() during normal
4455 : * operation using individual extend steps. Note that we do not need
4456 : * to extend clog since its extensions are WAL logged.
4457 : *
4458 : * This part has to be done regardless of standbyState since we
4459 : * immediately start assigning subtransactions to their toplevel
4460 : * transactions.
4461 : */
4462 24477 : next_expected_xid = latestObservedXid;
4463 49721 : while (TransactionIdPrecedes(next_expected_xid, xid))
4464 : {
4465 25244 : TransactionIdAdvance(next_expected_xid);
4466 25244 : ExtendSUBTRANS(next_expected_xid);
4467 : }
4468 : Assert(next_expected_xid == xid);
4469 :
4470 : /*
4471 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4472 : * more to do since we don't track assigned xids yet.
4473 : */
4474 24477 : if (standbyState <= STANDBY_INITIALIZED)
4475 : {
4476 0 : latestObservedXid = xid;
4477 0 : return;
4478 : }
4479 :
4480 : /*
4481 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4482 : */
4483 24477 : next_expected_xid = latestObservedXid;
4484 24477 : TransactionIdAdvance(next_expected_xid);
4485 24477 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4486 :
4487 : /*
4488 : * Now we can advance latestObservedXid
4489 : */
4490 24477 : latestObservedXid = xid;
4491 :
4492 : /* TransamVariables->nextXid must be beyond any observed xid */
4493 24477 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4494 : }
4495 : }
4496 :
4497 : /*
4498 : * ExpireTreeKnownAssignedTransactionIds
4499 : * Remove the given XIDs from KnownAssignedXids.
4500 : *
4501 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4502 : */
4503 : void
4504 23714 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4505 : TransactionId *subxids, TransactionId max_xid)
4506 : {
4507 : Assert(standbyState >= STANDBY_INITIALIZED);
4508 :
4509 : /*
4510 : * Uses same locking as transaction commit
4511 : */
4512 23714 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4513 :
4514 23714 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4515 :
4516 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4517 23714 : MaintainLatestCompletedXidRecovery(max_xid);
4518 :
4519 : /* ... and xactCompletionCount */
4520 23714 : TransamVariables->xactCompletionCount++;
4521 :
4522 23714 : LWLockRelease(ProcArrayLock);
4523 23714 : }
4524 :
4525 : /*
4526 : * ExpireAllKnownAssignedTransactionIds
4527 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4528 : */
4529 : void
4530 122 : ExpireAllKnownAssignedTransactionIds(void)
4531 : {
4532 : FullTransactionId latestXid;
4533 :
4534 122 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4535 122 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4536 :
4537 : /* Reset latestCompletedXid to nextXid - 1 */
4538 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
4539 122 : latestXid = TransamVariables->nextXid;
4540 122 : FullTransactionIdRetreat(&latestXid);
4541 122 : TransamVariables->latestCompletedXid = latestXid;
4542 :
4543 : /*
4544 : * Any transactions that were in-progress were effectively aborted, so
4545 : * advance xactCompletionCount.
4546 : */
4547 122 : TransamVariables->xactCompletionCount++;
4548 :
4549 : /*
4550 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4551 : * the call of this function. But do this for unification with what
4552 : * ExpireOldKnownAssignedTransactionIds() do.
4553 : */
4554 122 : procArray->lastOverflowedXid = InvalidTransactionId;
4555 122 : LWLockRelease(ProcArrayLock);
4556 122 : }
4557 :
4558 : /*
4559 : * ExpireOldKnownAssignedTransactionIds
4560 : * Remove KnownAssignedXids entries preceding the given XID and
4561 : * potentially reset lastOverflowedXid.
4562 : */
4563 : void
4564 840 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4565 : {
4566 : TransactionId latestXid;
4567 :
4568 840 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4569 :
4570 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4571 840 : latestXid = xid;
4572 840 : TransactionIdRetreat(latestXid);
4573 840 : MaintainLatestCompletedXidRecovery(latestXid);
4574 :
4575 : /* ... and xactCompletionCount */
4576 840 : TransamVariables->xactCompletionCount++;
4577 :
4578 : /*
4579 : * Reset lastOverflowedXid if we know all transactions that have been
4580 : * possibly running are being gone. Not doing so could cause an incorrect
4581 : * lastOverflowedXid value, which makes extra snapshots be marked as
4582 : * suboverflowed.
4583 : */
4584 840 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4585 830 : procArray->lastOverflowedXid = InvalidTransactionId;
4586 840 : KnownAssignedXidsRemovePreceding(xid);
4587 840 : LWLockRelease(ProcArrayLock);
4588 840 : }
4589 :
4590 : /*
4591 : * KnownAssignedTransactionIdsIdleMaintenance
4592 : * Opportunistically do maintenance work when the startup process
4593 : * is about to go idle.
4594 : */
4595 : void
4596 16177 : KnownAssignedTransactionIdsIdleMaintenance(void)
4597 : {
4598 16177 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4599 16177 : }
4600 :
4601 :
4602 : /*
4603 : * Private module functions to manipulate KnownAssignedXids
4604 : *
4605 : * There are 5 main uses of the KnownAssignedXids data structure:
4606 : *
4607 : * * backends taking snapshots - all valid XIDs need to be copied out
4608 : * * backends seeking to determine presence of a specific XID
4609 : * * startup process adding new known-assigned XIDs
4610 : * * startup process removing specific XIDs as transactions end
4611 : * * startup process pruning array when special WAL records arrive
4612 : *
4613 : * This data structure is known to be a hot spot during Hot Standby, so we
4614 : * go to some lengths to make these operations as efficient and as concurrent
4615 : * as possible.
4616 : *
4617 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4618 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4619 : * in general TransactionIdPrecedes would not provide a total order, but
4620 : * we know that the entries present at any instant should not extend across
4621 : * a large enough fraction of XID space to wrap around (the primary would
4622 : * shut down for fear of XID wrap long before that happens). So it's OK to
4623 : * use TransactionIdPrecedes as a binary-search comparator.
4624 : *
4625 : * It's cheap to maintain the sortedness during insertions, since new known
4626 : * XIDs are always reported in XID order; we just append them at the right.
4627 : *
4628 : * To keep individual deletions cheap, we need to allow gaps in the array.
4629 : * This is implemented by marking array elements as valid or invalid using
4630 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4631 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4632 : * XID entry itself. This preserves the property that the XID entries are
4633 : * sorted, so we can do binary searches easily. Periodically we compress
4634 : * out the unused entries; that's much cheaper than having to compress the
4635 : * array immediately on every deletion.
4636 : *
4637 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4638 : * are those with indexes tail <= i < head; items outside this subscript range
4639 : * have unspecified contents. When head reaches the end of the array, we
4640 : * force compression of unused entries rather than wrapping around, since
4641 : * allowing wraparound would greatly complicate the search logic. We maintain
4642 : * an explicit tail pointer so that pruning of old XIDs can be done without
4643 : * immediately moving the array contents. In most cases only a small fraction
4644 : * of the array contains valid entries at any instant.
4645 : *
4646 : * Although only the startup process can ever change the KnownAssignedXids
4647 : * data structure, we still need interlocking so that standby backends will
4648 : * not observe invalid intermediate states. The convention is that backends
4649 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4650 : * the array, the startup process must hold ProcArrayLock exclusively, for
4651 : * the usual transactional reasons (compare commit/abort of a transaction
4652 : * during normal running). Compressing unused entries out of the array
4653 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4654 : * them into slots to the right of the head pointer and then advance the head
4655 : * pointer. This doesn't require any lock at all, but on machines with weak
4656 : * memory ordering, we need to be careful that other processors see the array
4657 : * element changes before they see the head pointer change. We handle this by
4658 : * using memory barriers when reading or writing the head/tail pointers (unless
4659 : * the caller holds ProcArrayLock exclusively).
4660 : *
4661 : * Algorithmic analysis:
4662 : *
4663 : * If we have a maximum of M slots, with N XIDs currently spread across
4664 : * S elements then we have N <= S <= M always.
4665 : *
4666 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4667 : * happen)
4668 : * * Compressing the array is O(S) and requires exclusive lock
4669 : * * Removing an XID is O(logS) and requires exclusive lock
4670 : * * Taking a snapshot is O(S) and requires shared lock
4671 : * * Checking for an XID is O(logS) and requires shared lock
4672 : *
4673 : * In comparison, using a hash table for KnownAssignedXids would mean that
4674 : * taking snapshots would be O(M). If we can maintain S << M then the
4675 : * sorted array technique will deliver significantly faster snapshots.
4676 : * If we try to keep S too small then we will spend too much time compressing,
4677 : * so there is an optimal point for any workload mix. We use a heuristic to
4678 : * decide when to compress the array, though trimming also helps reduce
4679 : * frequency of compressing. The heuristic requires us to track the number of
4680 : * currently valid XIDs in the array (N). Except in special cases, we'll
4681 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4682 : * taking a snapshot to be O(N), which it would have to be anyway.
4683 : */
4684 :
4685 :
4686 : /*
4687 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4688 : * array, removing any gaps.
4689 : *
4690 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4691 : * we do it only if a heuristic indicates it's a good time to do it.
4692 : *
4693 : * Compression requires holding ProcArrayLock in exclusive mode.
4694 : * Caller must pass haveLock = true if it already holds the lock.
4695 : */
4696 : static void
4697 40752 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4698 : {
4699 40752 : ProcArrayStruct *pArray = procArray;
4700 : int head,
4701 : tail,
4702 : nelements;
4703 : int compress_index;
4704 : int i;
4705 :
4706 : /* Counters for compression heuristics */
4707 : static unsigned int transactionEndsCounter;
4708 : static TimestampTz lastCompressTs;
4709 :
4710 : /* Tuning constants */
4711 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4712 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4713 :
4714 : /*
4715 : * Since only the startup process modifies the head/tail pointers, we
4716 : * don't need a lock to read them here.
4717 : */
4718 40752 : head = pArray->headKnownAssignedXids;
4719 40752 : tail = pArray->tailKnownAssignedXids;
4720 40752 : nelements = head - tail;
4721 :
4722 : /*
4723 : * If we can choose whether to compress, use a heuristic to avoid
4724 : * compressing too often or not often enough. "Compress" here simply
4725 : * means moving the values to the beginning of the array, so it is not as
4726 : * complex or costly as typical data compression algorithms.
4727 : */
4728 40752 : if (nelements == pArray->numKnownAssignedXids)
4729 : {
4730 : /*
4731 : * When there are no gaps between head and tail, don't bother to
4732 : * compress, except in the KAX_NO_SPACE case where we must compress to
4733 : * create some space after the head.
4734 : */
4735 22470 : if (reason != KAX_NO_SPACE)
4736 22470 : return;
4737 : }
4738 18282 : else if (reason == KAX_TRANSACTION_END)
4739 : {
4740 : /*
4741 : * Consider compressing only once every so many commits. Frequency
4742 : * determined by benchmarks.
4743 : */
4744 12880 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4745 12770 : return;
4746 :
4747 : /*
4748 : * Furthermore, compress only if the used part of the array is less
4749 : * than 50% full (see comments above).
4750 : */
4751 110 : if (nelements < 2 * pArray->numKnownAssignedXids)
4752 5 : return;
4753 : }
4754 5402 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4755 : {
4756 : /*
4757 : * We're about to go idle for lack of new WAL, so we might as well
4758 : * compress. But not too often, to avoid ProcArray lock contention
4759 : * with readers.
4760 : */
4761 5283 : if (lastCompressTs != 0)
4762 : {
4763 : TimestampTz compress_after;
4764 :
4765 5283 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4766 : KAX_COMPRESS_IDLE_INTERVAL);
4767 5283 : if (GetCurrentTimestamp() < compress_after)
4768 5248 : return;
4769 : }
4770 : }
4771 :
4772 : /* Need to compress, so get the lock if we don't have it. */
4773 259 : if (!haveLock)
4774 35 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4775 :
4776 : /*
4777 : * We compress the array by reading the valid values from tail to head,
4778 : * re-aligning data to 0th element.
4779 : */
4780 259 : compress_index = 0;
4781 8662 : for (i = tail; i < head; i++)
4782 : {
4783 8403 : if (KnownAssignedXidsValid[i])
4784 : {
4785 921 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4786 921 : KnownAssignedXidsValid[compress_index] = true;
4787 921 : compress_index++;
4788 : }
4789 : }
4790 : Assert(compress_index == pArray->numKnownAssignedXids);
4791 :
4792 259 : pArray->tailKnownAssignedXids = 0;
4793 259 : pArray->headKnownAssignedXids = compress_index;
4794 :
4795 259 : if (!haveLock)
4796 35 : LWLockRelease(ProcArrayLock);
4797 :
4798 : /* Update timestamp for maintenance. No need to hold lock for this. */
4799 259 : lastCompressTs = GetCurrentTimestamp();
4800 : }
4801 :
4802 : /*
4803 : * Add xids into KnownAssignedXids at the head of the array.
4804 : *
4805 : * xids from from_xid to to_xid, inclusive, are added to the array.
4806 : *
4807 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4808 : * exclusive mode, so we need no extra locking here. Else caller holds no
4809 : * lock, so we need to be sure we maintain sufficient interlocks against
4810 : * concurrent readers. (Only the startup process ever calls this, so no need
4811 : * to worry about concurrent writers.)
4812 : */
4813 : static void
4814 24482 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4815 : bool exclusive_lock)
4816 : {
4817 24482 : ProcArrayStruct *pArray = procArray;
4818 : TransactionId next_xid;
4819 : int head,
4820 : tail;
4821 : int nxids;
4822 : int i;
4823 :
4824 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4825 :
4826 : /*
4827 : * Calculate how many array slots we'll need. Normally this is cheap; in
4828 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4829 : * way.
4830 : */
4831 24482 : if (to_xid >= from_xid)
4832 24482 : nxids = to_xid - from_xid + 1;
4833 : else
4834 : {
4835 0 : nxids = 1;
4836 0 : next_xid = from_xid;
4837 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4838 : {
4839 0 : nxids++;
4840 0 : TransactionIdAdvance(next_xid);
4841 : }
4842 : }
4843 :
4844 : /*
4845 : * Since only the startup process modifies the head/tail pointers, we
4846 : * don't need a lock to read them here.
4847 : */
4848 24482 : head = pArray->headKnownAssignedXids;
4849 24482 : tail = pArray->tailKnownAssignedXids;
4850 :
4851 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4852 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4853 :
4854 : /*
4855 : * Verify that insertions occur in TransactionId sequence. Note that even
4856 : * if the last existing element is marked invalid, it must still have a
4857 : * correctly sequenced XID value.
4858 : */
4859 38826 : if (head > tail &&
4860 14344 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4861 : {
4862 0 : KnownAssignedXidsDisplay(LOG);
4863 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4864 : }
4865 :
4866 : /*
4867 : * If our xids won't fit in the remaining space, compress out free space
4868 : */
4869 24482 : if (head + nxids > pArray->maxKnownAssignedXids)
4870 : {
4871 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4872 :
4873 0 : head = pArray->headKnownAssignedXids;
4874 : /* note: we no longer care about the tail pointer */
4875 :
4876 : /*
4877 : * If it still won't fit then we're out of memory
4878 : */
4879 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4880 0 : elog(ERROR, "too many KnownAssignedXids");
4881 : }
4882 :
4883 : /* Now we can insert the xids into the space starting at head */
4884 24482 : next_xid = from_xid;
4885 49731 : for (i = 0; i < nxids; i++)
4886 : {
4887 25249 : KnownAssignedXids[head] = next_xid;
4888 25249 : KnownAssignedXidsValid[head] = true;
4889 25249 : TransactionIdAdvance(next_xid);
4890 25249 : head++;
4891 : }
4892 :
4893 : /* Adjust count of number of valid entries */
4894 24482 : pArray->numKnownAssignedXids += nxids;
4895 :
4896 : /*
4897 : * Now update the head pointer. We use a write barrier to ensure that
4898 : * other processors see the above array updates before they see the head
4899 : * pointer change. The barrier isn't required if we're holding
4900 : * ProcArrayLock exclusively.
4901 : */
4902 24482 : if (!exclusive_lock)
4903 24477 : pg_write_barrier();
4904 :
4905 24482 : pArray->headKnownAssignedXids = head;
4906 24482 : }
4907 :
4908 : /*
4909 : * KnownAssignedXidsSearch
4910 : *
4911 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4912 : * Returns true if it was found, false if not.
4913 : *
4914 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4915 : * Exclusive lock must be held for remove = true.
4916 : */
4917 : static bool
4918 26384 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4919 : {
4920 26384 : ProcArrayStruct *pArray = procArray;
4921 : int first,
4922 : last;
4923 : int head;
4924 : int tail;
4925 26384 : int result_index = -1;
4926 :
4927 26384 : tail = pArray->tailKnownAssignedXids;
4928 26384 : head = pArray->headKnownAssignedXids;
4929 :
4930 : /*
4931 : * Only the startup process removes entries, so we don't need the read
4932 : * barrier in that case.
4933 : */
4934 26384 : if (!remove)
4935 1 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4936 :
4937 : /*
4938 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4939 : * array here, since even invalid entries will contain sorted XIDs.
4940 : */
4941 26384 : first = tail;
4942 26384 : last = head - 1;
4943 82228 : while (first <= last)
4944 : {
4945 : int mid_index;
4946 : TransactionId mid_xid;
4947 :
4948 81016 : mid_index = (first + last) / 2;
4949 81016 : mid_xid = KnownAssignedXids[mid_index];
4950 :
4951 81016 : if (xid == mid_xid)
4952 : {
4953 25172 : result_index = mid_index;
4954 25172 : break;
4955 : }
4956 55844 : else if (TransactionIdPrecedes(xid, mid_xid))
4957 12355 : last = mid_index - 1;
4958 : else
4959 43489 : first = mid_index + 1;
4960 : }
4961 :
4962 26384 : if (result_index < 0)
4963 1212 : return false; /* not in array */
4964 :
4965 25172 : if (!KnownAssignedXidsValid[result_index])
4966 9 : return false; /* in array, but invalid */
4967 :
4968 25163 : if (remove)
4969 : {
4970 25163 : KnownAssignedXidsValid[result_index] = false;
4971 :
4972 25163 : pArray->numKnownAssignedXids--;
4973 : Assert(pArray->numKnownAssignedXids >= 0);
4974 :
4975 : /*
4976 : * If we're removing the tail element then advance tail pointer over
4977 : * any invalid elements. This will speed future searches.
4978 : */
4979 25163 : if (result_index == tail)
4980 : {
4981 11327 : tail++;
4982 17681 : while (tail < head && !KnownAssignedXidsValid[tail])
4983 6354 : tail++;
4984 11327 : if (tail >= head)
4985 : {
4986 : /* Array is empty, so we can reset both pointers */
4987 10128 : pArray->headKnownAssignedXids = 0;
4988 10128 : pArray->tailKnownAssignedXids = 0;
4989 : }
4990 : else
4991 : {
4992 1199 : pArray->tailKnownAssignedXids = tail;
4993 : }
4994 : }
4995 : }
4996 :
4997 25163 : return true;
4998 : }
4999 :
5000 : /*
5001 : * Is the specified XID present in KnownAssignedXids[]?
5002 : *
5003 : * Caller must hold ProcArrayLock in shared or exclusive mode.
5004 : */
5005 : static bool
5006 1 : KnownAssignedXidExists(TransactionId xid)
5007 : {
5008 : Assert(TransactionIdIsValid(xid));
5009 :
5010 1 : return KnownAssignedXidsSearch(xid, false);
5011 : }
5012 :
5013 : /*
5014 : * Remove the specified XID from KnownAssignedXids[].
5015 : *
5016 : * Caller must hold ProcArrayLock in exclusive mode.
5017 : */
5018 : static void
5019 26383 : KnownAssignedXidsRemove(TransactionId xid)
5020 : {
5021 : Assert(TransactionIdIsValid(xid));
5022 :
5023 26383 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
5024 :
5025 : /*
5026 : * Note: we cannot consider it an error to remove an XID that's not
5027 : * present. We intentionally remove subxact IDs while processing
5028 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
5029 : * removed again when the top-level xact commits or aborts.
5030 : *
5031 : * It might be possible to track such XIDs to distinguish this case from
5032 : * actual errors, but it would be complicated and probably not worth it.
5033 : * So, just ignore the search result.
5034 : */
5035 26383 : (void) KnownAssignedXidsSearch(xid, true);
5036 26383 : }
5037 :
5038 : /*
5039 : * KnownAssignedXidsRemoveTree
5040 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
5041 : *
5042 : * Caller must hold ProcArrayLock in exclusive mode.
5043 : */
5044 : static void
5045 23735 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
5046 : TransactionId *subxids)
5047 : {
5048 : int i;
5049 :
5050 23735 : if (TransactionIdIsValid(xid))
5051 23714 : KnownAssignedXidsRemove(xid);
5052 :
5053 26404 : for (i = 0; i < nsubxids; i++)
5054 2669 : KnownAssignedXidsRemove(subxids[i]);
5055 :
5056 : /* Opportunistically compress the array */
5057 23735 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5058 23735 : }
5059 :
5060 : /*
5061 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5062 : * then clear the whole table.
5063 : *
5064 : * Caller must hold ProcArrayLock in exclusive mode.
5065 : */
5066 : static void
5067 962 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5068 : {
5069 962 : ProcArrayStruct *pArray = procArray;
5070 962 : int count = 0;
5071 : int head,
5072 : tail,
5073 : i;
5074 :
5075 962 : if (!TransactionIdIsValid(removeXid))
5076 : {
5077 122 : elog(DEBUG4, "removing all KnownAssignedXids");
5078 122 : pArray->numKnownAssignedXids = 0;
5079 122 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5080 122 : return;
5081 : }
5082 :
5083 840 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5084 :
5085 : /*
5086 : * Mark entries invalid starting at the tail. Since array is sorted, we
5087 : * can stop as soon as we reach an entry >= removeXid.
5088 : */
5089 840 : tail = pArray->tailKnownAssignedXids;
5090 840 : head = pArray->headKnownAssignedXids;
5091 :
5092 840 : for (i = tail; i < head; i++)
5093 : {
5094 205 : if (KnownAssignedXidsValid[i])
5095 : {
5096 205 : TransactionId knownXid = KnownAssignedXids[i];
5097 :
5098 205 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5099 205 : break;
5100 :
5101 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5102 : {
5103 0 : KnownAssignedXidsValid[i] = false;
5104 0 : count++;
5105 : }
5106 : }
5107 : }
5108 :
5109 840 : pArray->numKnownAssignedXids -= count;
5110 : Assert(pArray->numKnownAssignedXids >= 0);
5111 :
5112 : /*
5113 : * Advance the tail pointer if we've marked the tail item invalid.
5114 : */
5115 840 : for (i = tail; i < head; i++)
5116 : {
5117 205 : if (KnownAssignedXidsValid[i])
5118 205 : break;
5119 : }
5120 840 : if (i >= head)
5121 : {
5122 : /* Array is empty, so we can reset both pointers */
5123 635 : pArray->headKnownAssignedXids = 0;
5124 635 : pArray->tailKnownAssignedXids = 0;
5125 : }
5126 : else
5127 : {
5128 205 : pArray->tailKnownAssignedXids = i;
5129 : }
5130 :
5131 : /* Opportunistically compress the array */
5132 840 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5133 : }
5134 :
5135 : /*
5136 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5137 : * We filter out anything >= xmax.
5138 : *
5139 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5140 : * that array is large enough.
5141 : *
5142 : * Caller must hold ProcArrayLock in (at least) shared mode.
5143 : */
5144 : static int
5145 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5146 : {
5147 0 : TransactionId xtmp = InvalidTransactionId;
5148 :
5149 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5150 : }
5151 :
5152 : /*
5153 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5154 : * we reduce *xmin to the lowest xid value seen if not already lower.
5155 : *
5156 : * Caller must hold ProcArrayLock in (at least) shared mode.
5157 : */
5158 : static int
5159 1576 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5160 : TransactionId xmax)
5161 : {
5162 1576 : int count = 0;
5163 : int head,
5164 : tail;
5165 : int i;
5166 :
5167 : /*
5168 : * Fetch head just once, since it may change while we loop. We can stop
5169 : * once we reach the initially seen head, since we are certain that an xid
5170 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5171 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5172 : * anyway.
5173 : */
5174 1576 : tail = procArray->tailKnownAssignedXids;
5175 1576 : head = procArray->headKnownAssignedXids;
5176 :
5177 1576 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5178 :
5179 1607 : for (i = tail; i < head; i++)
5180 : {
5181 : /* Skip any gaps in the array */
5182 161 : if (KnownAssignedXidsValid[i])
5183 : {
5184 148 : TransactionId knownXid = KnownAssignedXids[i];
5185 :
5186 : /*
5187 : * Update xmin if required. Only the first XID need be checked,
5188 : * since the array is sorted.
5189 : */
5190 296 : if (count == 0 &&
5191 148 : TransactionIdPrecedes(knownXid, *xmin))
5192 18 : *xmin = knownXid;
5193 :
5194 : /*
5195 : * Filter out anything >= xmax, again relying on sorted property
5196 : * of array.
5197 : */
5198 296 : if (TransactionIdIsValid(xmax) &&
5199 148 : TransactionIdFollowsOrEquals(knownXid, xmax))
5200 130 : break;
5201 :
5202 : /* Add knownXid into output array */
5203 18 : xarray[count++] = knownXid;
5204 : }
5205 : }
5206 :
5207 1576 : return count;
5208 : }
5209 :
5210 : /*
5211 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5212 : * if nothing there.
5213 : */
5214 : static TransactionId
5215 383 : KnownAssignedXidsGetOldestXmin(void)
5216 : {
5217 : int head,
5218 : tail;
5219 : int i;
5220 :
5221 : /*
5222 : * Fetch head just once, since it may change while we loop.
5223 : */
5224 383 : tail = procArray->tailKnownAssignedXids;
5225 383 : head = procArray->headKnownAssignedXids;
5226 :
5227 383 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5228 :
5229 383 : for (i = tail; i < head; i++)
5230 : {
5231 : /* Skip any gaps in the array */
5232 148 : if (KnownAssignedXidsValid[i])
5233 148 : return KnownAssignedXids[i];
5234 : }
5235 :
5236 235 : return InvalidTransactionId;
5237 : }
5238 :
5239 : /*
5240 : * Display KnownAssignedXids to provide debug trail
5241 : *
5242 : * Currently this is only called within startup process, so we need no
5243 : * special locking.
5244 : *
5245 : * Note this is pretty expensive, and much of the expense will be incurred
5246 : * even if the elog message will get discarded. It's not currently called
5247 : * in any performance-critical places, however, so no need to be tenser.
5248 : */
5249 : static void
5250 127 : KnownAssignedXidsDisplay(int trace_level)
5251 : {
5252 127 : ProcArrayStruct *pArray = procArray;
5253 : StringInfoData buf;
5254 : int head,
5255 : tail,
5256 : i;
5257 127 : int nxids = 0;
5258 :
5259 127 : tail = pArray->tailKnownAssignedXids;
5260 127 : head = pArray->headKnownAssignedXids;
5261 :
5262 127 : initStringInfo(&buf);
5263 :
5264 137 : for (i = tail; i < head; i++)
5265 : {
5266 10 : if (KnownAssignedXidsValid[i])
5267 : {
5268 10 : nxids++;
5269 10 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5270 : }
5271 : }
5272 :
5273 127 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5274 : nxids,
5275 : pArray->numKnownAssignedXids,
5276 : pArray->tailKnownAssignedXids,
5277 : pArray->headKnownAssignedXids,
5278 : buf.data);
5279 :
5280 127 : pfree(buf.data);
5281 127 : }
5282 :
5283 : /*
5284 : * KnownAssignedXidsReset
5285 : * Resets KnownAssignedXids to be empty
5286 : */
5287 : static void
5288 0 : KnownAssignedXidsReset(void)
5289 : {
5290 0 : ProcArrayStruct *pArray = procArray;
5291 :
5292 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5293 :
5294 0 : pArray->numKnownAssignedXids = 0;
5295 0 : pArray->tailKnownAssignedXids = 0;
5296 0 : pArray->headKnownAssignedXids = 0;
5297 :
5298 0 : LWLockRelease(ProcArrayLock);
5299 0 : }
|
