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) 1996-2021, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/clog.h"
51 : #include "access/subtrans.h"
52 : #include "access/transam.h"
53 : #include "access/twophase.h"
54 : #include "access/xact.h"
55 : #include "access/xlogutils.h"
56 : #include "catalog/catalog.h"
57 : #include "catalog/pg_authid.h"
58 : #include "commands/dbcommands.h"
59 : #include "miscadmin.h"
60 : #include "pgstat.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/spin.h"
64 : #include "utils/acl.h"
65 : #include "utils/builtins.h"
66 : #include "utils/rel.h"
67 : #include "utils/snapmgr.h"
68 :
69 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
70 :
71 : /* Our shared memory area */
72 : typedef struct ProcArrayStruct
73 : {
74 : int numProcs; /* number of valid procs entries */
75 : int maxProcs; /* allocated size of procs array */
76 :
77 : /*
78 : * Known assigned XIDs handling
79 : */
80 : int maxKnownAssignedXids; /* allocated size of array */
81 : int numKnownAssignedXids; /* current # of valid entries */
82 : int tailKnownAssignedXids; /* index of oldest valid element */
83 : int headKnownAssignedXids; /* index of newest element, + 1 */
84 : slock_t known_assigned_xids_lck; /* protects head/tail pointers */
85 :
86 : /*
87 : * Highest subxid that has been removed from KnownAssignedXids array to
88 : * prevent overflow; or InvalidTransactionId if none. We track this for
89 : * similar reasons to tracking overflowing cached subxids in PGPROC
90 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
91 : * lock to read it.
92 : */
93 : TransactionId lastOverflowedXid;
94 :
95 : /* oldest xmin of any replication slot */
96 : TransactionId replication_slot_xmin;
97 : /* oldest catalog xmin of any replication slot */
98 : TransactionId replication_slot_catalog_xmin;
99 :
100 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
101 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
102 : } ProcArrayStruct;
103 :
104 : /*
105 : * State for the GlobalVisTest* family of functions. Those functions can
106 : * e.g. be used to decide if a deleted row can be removed without violating
107 : * MVCC semantics: If the deleted row's xmax is not considered to be running
108 : * by anyone, the row can be removed.
109 : *
110 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
111 : * cutoff XID while building a snapshot (looking at the frequently changing
112 : * xmins scales badly). Instead we compute two boundaries while building the
113 : * snapshot:
114 : *
115 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
116 : * definitely_needed are definitely still visible.
117 : *
118 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
119 : * definitely be removed
120 : *
121 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
122 : * && XID < definitely_needed), the boundaries can be recomputed (using
123 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
124 : * maintaining an accurate value all the time.
125 : *
126 : * As it is not cheap to compute accurate boundaries, we limit the number of
127 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
128 : *
129 : *
130 : * There are three backend lifetime instances of this struct, optimized for
131 : * different types of relations. As e.g. a normal user defined table in one
132 : * database is inaccessible to backends connected to another database, a test
133 : * specific to a relation can be more aggressive than a test for a shared
134 : * relation. Currently we track four different states:
135 : *
136 : * 1) GlobalVisSharedRels, which only considers an XID's
137 : * effects visible-to-everyone if neither snapshots in any database, nor a
138 : * replication slot's xmin, nor a replication slot's catalog_xmin might
139 : * still consider XID as running.
140 : *
141 : * 2) GlobalVisCatalogRels, which only considers an XID's
142 : * effects visible-to-everyone if neither snapshots in the current
143 : * database, nor a replication slot's xmin, nor a replication slot's
144 : * catalog_xmin might still consider XID as running.
145 : *
146 : * I.e. the difference to GlobalVisSharedRels is that
147 : * snapshot in other databases are ignored.
148 : *
149 : * 3) GlobalVisDataRels, which only considers an XID's
150 : * effects visible-to-everyone if neither snapshots in the current
151 : * database, nor a replication slot's xmin consider XID as running.
152 : *
153 : * I.e. the difference to GlobalVisCatalogRels is that
154 : * replication slot's catalog_xmin is not taken into account.
155 : *
156 : * 4) GlobalVisTempRels, which only considers the current session, as temp
157 : * tables are not visible to other sessions.
158 : *
159 : * GlobalVisTestFor(relation) returns the appropriate state
160 : * for the relation.
161 : *
162 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
163 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
164 : * prevent maybe_needed to become old enough after the GetSnapshotData()
165 : * call.
166 : *
167 : * The typedef is in the header.
168 : */
169 : struct GlobalVisState
170 : {
171 : /* XIDs >= are considered running by some backend */
172 : FullTransactionId definitely_needed;
173 :
174 : /* XIDs < are not considered to be running by any backend */
175 : FullTransactionId maybe_needed;
176 : };
177 :
178 : /*
179 : * Result of ComputeXidHorizons().
180 : */
181 : typedef struct ComputeXidHorizonsResult
182 : {
183 : /*
184 : * The value of ShmemVariableCache->latestCompletedXid when
185 : * ComputeXidHorizons() held ProcArrayLock.
186 : */
187 : FullTransactionId latest_completed;
188 :
189 : /*
190 : * The same for procArray->replication_slot_xmin and.
191 : * procArray->replication_slot_catalog_xmin.
192 : */
193 : TransactionId slot_xmin;
194 : TransactionId slot_catalog_xmin;
195 :
196 : /*
197 : * Oldest xid that any backend might still consider running. This needs to
198 : * include processes running VACUUM, in contrast to the normal visibility
199 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
200 : * determining visibility, but doesn't care about rows above its xmin to
201 : * be removed.
202 : *
203 : * This likely should only be needed to determine whether pg_subtrans can
204 : * be truncated. It currently includes the effects of replication slots,
205 : * for historical reasons. But that could likely be changed.
206 : */
207 : TransactionId oldest_considered_running;
208 :
209 : /*
210 : * Oldest xid for which deleted tuples need to be retained in shared
211 : * tables.
212 : *
213 : * This includes the effects of replication slots. If that's not desired,
214 : * look at shared_oldest_nonremovable_raw;
215 : */
216 : TransactionId shared_oldest_nonremovable;
217 :
218 : /*
219 : * Oldest xid that may be necessary to retain in shared tables. This is
220 : * the same as shared_oldest_nonremovable, except that is not affected by
221 : * replication slot's catalog_xmin.
222 : *
223 : * This is mainly useful to be able to send the catalog_xmin to upstream
224 : * streaming replication servers via hot_standby_feedback, so they can
225 : * apply the limit only when accessing catalog tables.
226 : */
227 : TransactionId shared_oldest_nonremovable_raw;
228 :
229 : /*
230 : * Oldest xid for which deleted tuples need to be retained in non-shared
231 : * catalog tables.
232 : */
233 : TransactionId catalog_oldest_nonremovable;
234 :
235 : /*
236 : * Oldest xid for which deleted tuples need to be retained in normal user
237 : * defined tables.
238 : */
239 : TransactionId data_oldest_nonremovable;
240 :
241 : /*
242 : * Oldest xid for which deleted tuples need to be retained in this
243 : * session's temporary tables.
244 : */
245 : TransactionId temp_oldest_nonremovable;
246 :
247 : } ComputeXidHorizonsResult;
248 :
249 : /*
250 : * Return value for GlobalVisHorizonKindForRel().
251 : */
252 : typedef enum GlobalVisHorizonKind
253 : {
254 : VISHORIZON_SHARED,
255 : VISHORIZON_CATALOG,
256 : VISHORIZON_DATA,
257 : VISHORIZON_TEMP
258 : } GlobalVisHorizonKind;
259 :
260 :
261 : static ProcArrayStruct *procArray;
262 :
263 : static PGPROC *allProcs;
264 :
265 : /*
266 : * Bookkeeping for tracking emulated transactions in recovery
267 : */
268 : static TransactionId *KnownAssignedXids;
269 : static bool *KnownAssignedXidsValid;
270 : static TransactionId latestObservedXid = InvalidTransactionId;
271 :
272 : /*
273 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
274 : * the highest xid that might still be running that we don't have in
275 : * KnownAssignedXids.
276 : */
277 : static TransactionId standbySnapshotPendingXmin;
278 :
279 : /*
280 : * State for visibility checks on different types of relations. See struct
281 : * GlobalVisState for details. As shared, catalog, normal and temporary
282 : * relations can have different horizons, one such state exists for each.
283 : */
284 : static GlobalVisState GlobalVisSharedRels;
285 : static GlobalVisState GlobalVisCatalogRels;
286 : static GlobalVisState GlobalVisDataRels;
287 : static GlobalVisState GlobalVisTempRels;
288 :
289 : /*
290 : * This backend's RecentXmin at the last time the accurate xmin horizon was
291 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
292 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
293 : */
294 : static TransactionId ComputeXidHorizonsResultLastXmin;
295 :
296 : #ifdef XIDCACHE_DEBUG
297 :
298 : /* counters for XidCache measurement */
299 : static long xc_by_recent_xmin = 0;
300 : static long xc_by_known_xact = 0;
301 : static long xc_by_my_xact = 0;
302 : static long xc_by_latest_xid = 0;
303 : static long xc_by_main_xid = 0;
304 : static long xc_by_child_xid = 0;
305 : static long xc_by_known_assigned = 0;
306 : static long xc_no_overflow = 0;
307 : static long xc_slow_answer = 0;
308 :
309 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
310 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
311 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
312 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
313 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
314 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
315 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
316 : #define xc_no_overflow_inc() (xc_no_overflow++)
317 : #define xc_slow_answer_inc() (xc_slow_answer++)
318 :
319 : static void DisplayXidCache(void);
320 : #else /* !XIDCACHE_DEBUG */
321 :
322 : #define xc_by_recent_xmin_inc() ((void) 0)
323 : #define xc_by_known_xact_inc() ((void) 0)
324 : #define xc_by_my_xact_inc() ((void) 0)
325 : #define xc_by_latest_xid_inc() ((void) 0)
326 : #define xc_by_main_xid_inc() ((void) 0)
327 : #define xc_by_child_xid_inc() ((void) 0)
328 : #define xc_by_known_assigned_inc() ((void) 0)
329 : #define xc_no_overflow_inc() ((void) 0)
330 : #define xc_slow_answer_inc() ((void) 0)
331 : #endif /* XIDCACHE_DEBUG */
332 :
333 : /* Primitives for KnownAssignedXids array handling for standby */
334 : static void KnownAssignedXidsCompress(bool force);
335 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
336 : bool exclusive_lock);
337 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
338 : static bool KnownAssignedXidExists(TransactionId xid);
339 : static void KnownAssignedXidsRemove(TransactionId xid);
340 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
341 : TransactionId *subxids);
342 : static void KnownAssignedXidsRemovePreceding(TransactionId xid);
343 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
344 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
345 : TransactionId *xmin,
346 : TransactionId xmax);
347 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
348 : static void KnownAssignedXidsDisplay(int trace_level);
349 : static void KnownAssignedXidsReset(void);
350 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
351 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
352 : static void MaintainLatestCompletedXid(TransactionId latestXid);
353 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
354 :
355 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
356 : TransactionId xid);
357 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
358 :
359 : /*
360 : * Report shared-memory space needed by CreateSharedProcArray.
361 : */
362 : Size
363 3842 : ProcArrayShmemSize(void)
364 : {
365 : Size size;
366 :
367 : /* Size of the ProcArray structure itself */
368 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
369 :
370 3842 : size = offsetof(ProcArrayStruct, pgprocnos);
371 3842 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
372 :
373 : /*
374 : * During Hot Standby processing we have a data structure called
375 : * KnownAssignedXids, created in shared memory. Local data structures are
376 : * also created in various backends during GetSnapshotData(),
377 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
378 : * main structures created in those functions must be identically sized,
379 : * since we may at times copy the whole of the data structures around. We
380 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
381 : *
382 : * Ideally we'd only create this structure if we were actually doing hot
383 : * standby in the current run, but we don't know that yet at the time
384 : * shared memory is being set up.
385 : */
386 : #define TOTAL_MAX_CACHED_SUBXIDS \
387 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
388 :
389 3842 : if (EnableHotStandby)
390 : {
391 3822 : size = add_size(size,
392 : mul_size(sizeof(TransactionId),
393 3822 : TOTAL_MAX_CACHED_SUBXIDS));
394 3822 : size = add_size(size,
395 3822 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
396 : }
397 :
398 3842 : return size;
399 : }
400 :
401 : /*
402 : * Initialize the shared PGPROC array during postmaster startup.
403 : */
404 : void
405 2894 : CreateSharedProcArray(void)
406 : {
407 : bool found;
408 :
409 : /* Create or attach to the ProcArray shared structure */
410 2894 : procArray = (ProcArrayStruct *)
411 2894 : ShmemInitStruct("Proc Array",
412 : add_size(offsetof(ProcArrayStruct, pgprocnos),
413 : mul_size(sizeof(int),
414 2894 : PROCARRAY_MAXPROCS)),
415 : &found);
416 :
417 2894 : if (!found)
418 : {
419 : /*
420 : * We're the first - initialize.
421 : */
422 2894 : procArray->numProcs = 0;
423 2894 : procArray->maxProcs = PROCARRAY_MAXPROCS;
424 2894 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
425 2894 : procArray->numKnownAssignedXids = 0;
426 2894 : procArray->tailKnownAssignedXids = 0;
427 2894 : procArray->headKnownAssignedXids = 0;
428 2894 : SpinLockInit(&procArray->known_assigned_xids_lck);
429 2894 : procArray->lastOverflowedXid = InvalidTransactionId;
430 2894 : procArray->replication_slot_xmin = InvalidTransactionId;
431 2894 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
432 2894 : ShmemVariableCache->xactCompletionCount = 1;
433 : }
434 :
435 2894 : allProcs = ProcGlobal->allProcs;
436 :
437 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
438 2894 : if (EnableHotStandby)
439 : {
440 2884 : KnownAssignedXids = (TransactionId *)
441 2884 : ShmemInitStruct("KnownAssignedXids",
442 : mul_size(sizeof(TransactionId),
443 2884 : TOTAL_MAX_CACHED_SUBXIDS),
444 : &found);
445 2884 : KnownAssignedXidsValid = (bool *)
446 2884 : ShmemInitStruct("KnownAssignedXidsValid",
447 2884 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
448 : &found);
449 : }
450 2894 : }
451 :
452 : /*
453 : * Add the specified PGPROC to the shared array.
454 : */
455 : void
456 16456 : ProcArrayAdd(PGPROC *proc)
457 : {
458 16456 : ProcArrayStruct *arrayP = procArray;
459 : int index;
460 : int movecount;
461 :
462 : /* See ProcGlobal comment explaining why both locks are held */
463 16456 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
464 16456 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
465 :
466 16456 : if (arrayP->numProcs >= arrayP->maxProcs)
467 : {
468 : /*
469 : * Oops, no room. (This really shouldn't happen, since there is a
470 : * fixed supply of PGPROC structs too, and so we should have failed
471 : * earlier.)
472 : */
473 0 : ereport(FATAL,
474 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
475 : errmsg("sorry, too many clients already")));
476 : }
477 :
478 : /*
479 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
480 : * locality of references much better. This is useful while traversing the
481 : * ProcArray because there is an increased likelihood of finding the next
482 : * PGPROC structure in the cache.
483 : *
484 : * Since the occurrence of adding/removing a proc is much lower than the
485 : * access to the ProcArray itself, the overhead should be marginal
486 : */
487 31780 : for (index = 0; index < arrayP->numProcs; index++)
488 : {
489 28756 : int procno PG_USED_FOR_ASSERTS_ONLY = arrayP->pgprocnos[index];
490 :
491 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
492 : Assert(allProcs[procno].pgxactoff == index);
493 :
494 : /* If we have found our right position in the array, break */
495 28756 : if (arrayP->pgprocnos[index] > proc->pgprocno)
496 13432 : break;
497 : }
498 :
499 16456 : movecount = arrayP->numProcs - index;
500 16456 : memmove(&arrayP->pgprocnos[index + 1],
501 16456 : &arrayP->pgprocnos[index],
502 : movecount * sizeof(*arrayP->pgprocnos));
503 16456 : memmove(&ProcGlobal->xids[index + 1],
504 16456 : &ProcGlobal->xids[index],
505 : movecount * sizeof(*ProcGlobal->xids));
506 16456 : memmove(&ProcGlobal->subxidStates[index + 1],
507 16456 : &ProcGlobal->subxidStates[index],
508 : movecount * sizeof(*ProcGlobal->subxidStates));
509 16456 : memmove(&ProcGlobal->statusFlags[index + 1],
510 16456 : &ProcGlobal->statusFlags[index],
511 : movecount * sizeof(*ProcGlobal->statusFlags));
512 :
513 16456 : arrayP->pgprocnos[index] = proc->pgprocno;
514 16456 : proc->pgxactoff = index;
515 16456 : ProcGlobal->xids[index] = proc->xid;
516 16456 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
517 16456 : ProcGlobal->statusFlags[index] = proc->statusFlags;
518 :
519 16456 : arrayP->numProcs++;
520 :
521 : /* adjust pgxactoff for all following PGPROCs */
522 16456 : index++;
523 48838 : for (; index < arrayP->numProcs; index++)
524 : {
525 32382 : int procno = arrayP->pgprocnos[index];
526 :
527 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
528 : Assert(allProcs[procno].pgxactoff == index - 1);
529 :
530 32382 : allProcs[procno].pgxactoff = index;
531 : }
532 :
533 : /*
534 : * Release in reversed acquisition order, to reduce frequency of having to
535 : * wait for XidGenLock while holding ProcArrayLock.
536 : */
537 16456 : LWLockRelease(XidGenLock);
538 16456 : LWLockRelease(ProcArrayLock);
539 16456 : }
540 :
541 : /*
542 : * Remove the specified PGPROC from the shared array.
543 : *
544 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
545 : * array, and thus causing it to appear as "not running" anymore. In this
546 : * case we must advance latestCompletedXid. (This is essentially the same
547 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
548 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
549 : * twophase.c depends on the latter.)
550 : */
551 : void
552 16408 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
553 : {
554 16408 : ProcArrayStruct *arrayP = procArray;
555 : int myoff;
556 : int movecount;
557 :
558 : #ifdef XIDCACHE_DEBUG
559 : /* dump stats at backend shutdown, but not prepared-xact end */
560 : if (proc->pid != 0)
561 : DisplayXidCache();
562 : #endif
563 :
564 : /* See ProcGlobal comment explaining why both locks are held */
565 16408 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
566 16408 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
567 :
568 16408 : myoff = proc->pgxactoff;
569 :
570 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
571 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
572 :
573 16408 : if (TransactionIdIsValid(latestXid))
574 : {
575 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
576 :
577 : /* Advance global latestCompletedXid while holding the lock */
578 652 : MaintainLatestCompletedXid(latestXid);
579 :
580 : /* Same with xactCompletionCount */
581 652 : ShmemVariableCache->xactCompletionCount++;
582 :
583 652 : ProcGlobal->xids[myoff] = InvalidTransactionId;
584 652 : ProcGlobal->subxidStates[myoff].overflowed = false;
585 652 : ProcGlobal->subxidStates[myoff].count = 0;
586 : }
587 : else
588 : {
589 : /* Shouldn't be trying to remove a live transaction here */
590 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
591 : }
592 :
593 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
594 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
595 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
596 :
597 16408 : ProcGlobal->statusFlags[myoff] = 0;
598 :
599 : /* Keep the PGPROC array sorted. See notes above */
600 16408 : movecount = arrayP->numProcs - myoff - 1;
601 16408 : memmove(&arrayP->pgprocnos[myoff],
602 16408 : &arrayP->pgprocnos[myoff + 1],
603 : movecount * sizeof(*arrayP->pgprocnos));
604 16408 : memmove(&ProcGlobal->xids[myoff],
605 16408 : &ProcGlobal->xids[myoff + 1],
606 : movecount * sizeof(*ProcGlobal->xids));
607 16408 : memmove(&ProcGlobal->subxidStates[myoff],
608 16408 : &ProcGlobal->subxidStates[myoff + 1],
609 : movecount * sizeof(*ProcGlobal->subxidStates));
610 16408 : memmove(&ProcGlobal->statusFlags[myoff],
611 16408 : &ProcGlobal->statusFlags[myoff + 1],
612 : movecount * sizeof(*ProcGlobal->statusFlags));
613 :
614 16408 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
615 16408 : arrayP->numProcs--;
616 :
617 : /*
618 : * Adjust pgxactoff of following procs for removed PGPROC (note that
619 : * numProcs already has been decremented).
620 : */
621 48532 : for (int index = myoff; index < arrayP->numProcs; index++)
622 : {
623 32124 : int procno = arrayP->pgprocnos[index];
624 :
625 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
626 : Assert(allProcs[procno].pgxactoff - 1 == index);
627 :
628 32124 : allProcs[procno].pgxactoff = index;
629 : }
630 :
631 : /*
632 : * Release in reversed acquisition order, to reduce frequency of having to
633 : * wait for XidGenLock while holding ProcArrayLock.
634 : */
635 16408 : LWLockRelease(XidGenLock);
636 16408 : LWLockRelease(ProcArrayLock);
637 16408 : }
638 :
639 :
640 : /*
641 : * ProcArrayEndTransaction -- mark a transaction as no longer running
642 : *
643 : * This is used interchangeably for commit and abort cases. The transaction
644 : * commit/abort must already be reported to WAL and pg_xact.
645 : *
646 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
647 : * latestXid is the latest Xid among the transaction's main XID and
648 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
649 : * the caller to pass latestXid, instead of computing it from the PGPROC's
650 : * contents, because the subxid information in the PGPROC might be
651 : * incomplete.)
652 : */
653 : void
654 728964 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
655 : {
656 728964 : if (TransactionIdIsValid(latestXid))
657 : {
658 : /*
659 : * We must lock ProcArrayLock while clearing our advertised XID, so
660 : * that we do not exit the set of "running" transactions while someone
661 : * else is taking a snapshot. See discussion in
662 : * src/backend/access/transam/README.
663 : */
664 : Assert(TransactionIdIsValid(proc->xid));
665 :
666 : /*
667 : * If we can immediately acquire ProcArrayLock, we clear our own XID
668 : * and release the lock. If not, use group XID clearing to improve
669 : * efficiency.
670 : */
671 454524 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
672 : {
673 454450 : ProcArrayEndTransactionInternal(proc, latestXid);
674 454450 : LWLockRelease(ProcArrayLock);
675 : }
676 : else
677 74 : ProcArrayGroupClearXid(proc, latestXid);
678 : }
679 : else
680 : {
681 : /*
682 : * If we have no XID, we don't need to lock, since we won't affect
683 : * anyone else's calculation of a snapshot. We might change their
684 : * estimate of global xmin, but that's OK.
685 : */
686 : Assert(!TransactionIdIsValid(proc->xid));
687 : Assert(proc->subxidStatus.count == 0);
688 : Assert(!proc->subxidStatus.overflowed);
689 :
690 274440 : proc->lxid = InvalidLocalTransactionId;
691 274440 : proc->xmin = InvalidTransactionId;
692 274440 : proc->delayChkpt = false; /* be sure this is cleared in abort */
693 274440 : proc->recoveryConflictPending = false;
694 :
695 : /* must be cleared with xid/xmin: */
696 : /* avoid unnecessarily dirtying shared cachelines */
697 274440 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
698 : {
699 : Assert(!LWLockHeldByMe(ProcArrayLock));
700 57768 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
701 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
702 57768 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
703 57768 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
704 57768 : LWLockRelease(ProcArrayLock);
705 : }
706 : }
707 728964 : }
708 :
709 : /*
710 : * Mark a write transaction as no longer running.
711 : *
712 : * We don't do any locking here; caller must handle that.
713 : */
714 : static inline void
715 454524 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
716 : {
717 454524 : int pgxactoff = proc->pgxactoff;
718 :
719 : /*
720 : * Note: we need exclusive lock here because we're going to change other
721 : * processes' PGPROC entries.
722 : */
723 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
724 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
725 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
726 :
727 454524 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
728 454524 : proc->xid = InvalidTransactionId;
729 454524 : proc->lxid = InvalidLocalTransactionId;
730 454524 : proc->xmin = InvalidTransactionId;
731 454524 : proc->delayChkpt = false; /* be sure this is cleared in abort */
732 454524 : proc->recoveryConflictPending = false;
733 :
734 : /* must be cleared with xid/xmin: */
735 : /* avoid unnecessarily dirtying shared cachelines */
736 454524 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
737 : {
738 744 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
739 744 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
740 : }
741 :
742 : /* Clear the subtransaction-XID cache too while holding the lock */
743 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
744 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
745 454524 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
746 : {
747 768 : ProcGlobal->subxidStates[pgxactoff].count = 0;
748 768 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
749 768 : proc->subxidStatus.count = 0;
750 768 : proc->subxidStatus.overflowed = false;
751 : }
752 :
753 : /* Also advance global latestCompletedXid while holding the lock */
754 454524 : MaintainLatestCompletedXid(latestXid);
755 :
756 : /* Same with xactCompletionCount */
757 454524 : ShmemVariableCache->xactCompletionCount++;
758 454524 : }
759 :
760 : /*
761 : * ProcArrayGroupClearXid -- group XID clearing
762 : *
763 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
764 : * commit time, add ourselves to a list of processes that need their XIDs
765 : * cleared. The first process to add itself to the list will acquire
766 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
767 : * on behalf of all group members. This avoids a great deal of contention
768 : * around ProcArrayLock when many processes are trying to commit at once,
769 : * since the lock need not be repeatedly handed off from one committing
770 : * process to the next.
771 : */
772 : static void
773 74 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
774 : {
775 74 : PROC_HDR *procglobal = ProcGlobal;
776 : uint32 nextidx;
777 : uint32 wakeidx;
778 :
779 : /* We should definitely have an XID to clear. */
780 : Assert(TransactionIdIsValid(proc->xid));
781 :
782 : /* Add ourselves to the list of processes needing a group XID clear. */
783 74 : proc->procArrayGroupMember = true;
784 74 : proc->procArrayGroupMemberXid = latestXid;
785 74 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
786 : while (true)
787 : {
788 74 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
789 :
790 74 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
791 : &nextidx,
792 74 : (uint32) proc->pgprocno))
793 74 : break;
794 : }
795 :
796 : /*
797 : * If the list was not empty, the leader will clear our XID. It is
798 : * impossible to have followers without a leader because the first process
799 : * that has added itself to the list will always have nextidx as
800 : * INVALID_PGPROCNO.
801 : */
802 74 : if (nextidx != INVALID_PGPROCNO)
803 : {
804 4 : int extraWaits = 0;
805 :
806 : /* Sleep until the leader clears our XID. */
807 4 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
808 : for (;;)
809 : {
810 : /* acts as a read barrier */
811 4 : PGSemaphoreLock(proc->sem);
812 4 : if (!proc->procArrayGroupMember)
813 4 : break;
814 0 : extraWaits++;
815 : }
816 4 : pgstat_report_wait_end();
817 :
818 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PGPROCNO);
819 :
820 : /* Fix semaphore count for any absorbed wakeups */
821 4 : while (extraWaits-- > 0)
822 0 : PGSemaphoreUnlock(proc->sem);
823 4 : return;
824 : }
825 :
826 : /* We are the leader. Acquire the lock on behalf of everyone. */
827 70 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
828 :
829 : /*
830 : * Now that we've got the lock, clear the list of processes waiting for
831 : * group XID clearing, saving a pointer to the head of the list. Trying
832 : * to pop elements one at a time could lead to an ABA problem.
833 : */
834 70 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
835 : INVALID_PGPROCNO);
836 :
837 : /* Remember head of list so we can perform wakeups after dropping lock. */
838 70 : wakeidx = nextidx;
839 :
840 : /* Walk the list and clear all XIDs. */
841 144 : while (nextidx != INVALID_PGPROCNO)
842 : {
843 74 : PGPROC *nextproc = &allProcs[nextidx];
844 :
845 74 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
846 :
847 : /* Move to next proc in list. */
848 74 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
849 : }
850 :
851 : /* We're done with the lock now. */
852 70 : LWLockRelease(ProcArrayLock);
853 :
854 : /*
855 : * Now that we've released the lock, go back and wake everybody up. We
856 : * don't do this under the lock so as to keep lock hold times to a
857 : * minimum. The system calls we need to perform to wake other processes
858 : * up are probably much slower than the simple memory writes we did while
859 : * holding the lock.
860 : */
861 144 : while (wakeidx != INVALID_PGPROCNO)
862 : {
863 74 : PGPROC *nextproc = &allProcs[wakeidx];
864 :
865 74 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
866 74 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PGPROCNO);
867 :
868 : /* ensure all previous writes are visible before follower continues. */
869 74 : pg_write_barrier();
870 :
871 74 : nextproc->procArrayGroupMember = false;
872 :
873 74 : if (nextproc != MyProc)
874 4 : PGSemaphoreUnlock(nextproc->sem);
875 : }
876 : }
877 :
878 : /*
879 : * ProcArrayClearTransaction -- clear the transaction fields
880 : *
881 : * This is used after successfully preparing a 2-phase transaction. We are
882 : * not actually reporting the transaction's XID as no longer running --- it
883 : * will still appear as running because the 2PC's gxact is in the ProcArray
884 : * too. We just have to clear out our own PGPROC.
885 : */
886 : void
887 650 : ProcArrayClearTransaction(PGPROC *proc)
888 : {
889 : int pgxactoff;
890 :
891 : /*
892 : * Currently we need to lock ProcArrayLock exclusively here, as we
893 : * increment xactCompletionCount below. We also need it at least in shared
894 : * mode for pgproc->pgxactoff to stay the same below.
895 : *
896 : * We could however, as this action does not actually change anyone's view
897 : * of the set of running XIDs (our entry is duplicate with the gxact that
898 : * has already been inserted into the ProcArray), lower the lock level to
899 : * shared if we were to make xactCompletionCount an atomic variable. But
900 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
901 : * enough for this not to be the bottleneck. If it ever becomes a
902 : * bottleneck it may also be worth considering to combine this with the
903 : * subsequent ProcArrayRemove()
904 : */
905 650 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
906 :
907 650 : pgxactoff = proc->pgxactoff;
908 :
909 650 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
910 650 : proc->xid = InvalidTransactionId;
911 :
912 650 : proc->lxid = InvalidLocalTransactionId;
913 650 : proc->xmin = InvalidTransactionId;
914 650 : proc->recoveryConflictPending = false;
915 :
916 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
917 : Assert(!proc->delayChkpt);
918 :
919 : /*
920 : * Need to increment completion count even though transaction hasn't
921 : * really committed yet. The reason for that is that GetSnapshotData()
922 : * omits the xid of the current transaction, thus without the increment we
923 : * otherwise could end up reusing the snapshot later. Which would be bad,
924 : * because it might not count the prepared transaction as running.
925 : */
926 650 : ShmemVariableCache->xactCompletionCount++;
927 :
928 : /* Clear the subtransaction-XID cache too */
929 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
930 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
931 650 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
932 : {
933 260 : ProcGlobal->subxidStates[pgxactoff].count = 0;
934 260 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
935 260 : proc->subxidStatus.count = 0;
936 260 : proc->subxidStatus.overflowed = false;
937 : }
938 :
939 650 : LWLockRelease(ProcArrayLock);
940 650 : }
941 :
942 : /*
943 : * Update ShmemVariableCache->latestCompletedXid to point to latestXid if
944 : * currently older.
945 : */
946 : static void
947 456000 : MaintainLatestCompletedXid(TransactionId latestXid)
948 : {
949 456000 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
950 :
951 : Assert(FullTransactionIdIsValid(cur_latest));
952 : Assert(!RecoveryInProgress());
953 : Assert(LWLockHeldByMe(ProcArrayLock));
954 :
955 456000 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
956 : {
957 447218 : ShmemVariableCache->latestCompletedXid =
958 447218 : FullXidRelativeTo(cur_latest, latestXid);
959 : }
960 :
961 : Assert(IsBootstrapProcessingMode() ||
962 : FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
963 456000 : }
964 :
965 : /*
966 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
967 : */
968 : static void
969 434 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
970 : {
971 434 : FullTransactionId cur_latest = ShmemVariableCache->latestCompletedXid;
972 : FullTransactionId rel;
973 :
974 : Assert(AmStartupProcess() || !IsUnderPostmaster);
975 : Assert(LWLockHeldByMe(ProcArrayLock));
976 :
977 : /*
978 : * Need a FullTransactionId to compare latestXid with. Can't rely on
979 : * latestCompletedXid to be initialized in recovery. But in recovery it's
980 : * safe to access nextXid without a lock for the startup process.
981 : */
982 434 : rel = ShmemVariableCache->nextXid;
983 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
984 :
985 754 : if (!FullTransactionIdIsValid(cur_latest) ||
986 320 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
987 : {
988 422 : ShmemVariableCache->latestCompletedXid =
989 422 : FullXidRelativeTo(rel, latestXid);
990 : }
991 :
992 : Assert(FullTransactionIdIsNormal(ShmemVariableCache->latestCompletedXid));
993 434 : }
994 :
995 : /*
996 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
997 : *
998 : * Remember up to where the startup process initialized the CLOG and subtrans
999 : * so we can ensure it's initialized gaplessly up to the point where necessary
1000 : * while in recovery.
1001 : */
1002 : void
1003 114 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1004 : {
1005 : Assert(standbyState == STANDBY_INITIALIZED);
1006 : Assert(TransactionIdIsNormal(initializedUptoXID));
1007 :
1008 : /*
1009 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1010 : * to, so we can extend it from that point onwards in
1011 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1012 : * ProcArrayApplyRecoveryInfo().
1013 : */
1014 114 : latestObservedXid = initializedUptoXID;
1015 114 : TransactionIdRetreat(latestObservedXid);
1016 114 : }
1017 :
1018 : /*
1019 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1020 : *
1021 : * Takes us through 3 states: Initialized, Pending and Ready.
1022 : * Normal case is to go all the way to Ready straight away, though there
1023 : * are atypical cases where we need to take it in steps.
1024 : *
1025 : * Use the data about running transactions on the primary to create the initial
1026 : * state of KnownAssignedXids. We also use these records to regularly prune
1027 : * KnownAssignedXids because we know it is possible that some transactions
1028 : * with FATAL errors fail to write abort records, which could cause eventual
1029 : * overflow.
1030 : *
1031 : * See comments for LogStandbySnapshot().
1032 : */
1033 : void
1034 254 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1035 : {
1036 : TransactionId *xids;
1037 : int nxids;
1038 : int i;
1039 :
1040 : Assert(standbyState >= STANDBY_INITIALIZED);
1041 : Assert(TransactionIdIsValid(running->nextXid));
1042 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1043 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1044 :
1045 : /*
1046 : * Remove stale transactions, if any.
1047 : */
1048 254 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1049 :
1050 : /*
1051 : * Remove stale locks, if any.
1052 : */
1053 254 : StandbyReleaseOldLocks(running->oldestRunningXid);
1054 :
1055 : /*
1056 : * If our snapshot is already valid, nothing else to do...
1057 : */
1058 254 : if (standbyState == STANDBY_SNAPSHOT_READY)
1059 140 : return;
1060 :
1061 : /*
1062 : * If our initial RunningTransactionsData had an overflowed snapshot then
1063 : * we knew we were missing some subxids from our snapshot. If we continue
1064 : * to see overflowed snapshots then we might never be able to start up, so
1065 : * we make another test to see if our snapshot is now valid. We know that
1066 : * the missing subxids are equal to or earlier than nextXid. After we
1067 : * initialise we continue to apply changes during recovery, so once the
1068 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1069 : * know that we no longer have missing information and can mark the
1070 : * snapshot as valid.
1071 : */
1072 114 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1073 : {
1074 : /*
1075 : * If the snapshot isn't overflowed or if its empty we can reset our
1076 : * pending state and use this snapshot instead.
1077 : */
1078 0 : if (!running->subxid_overflow || running->xcnt == 0)
1079 : {
1080 : /*
1081 : * If we have already collected known assigned xids, we need to
1082 : * throw them away before we apply the recovery snapshot.
1083 : */
1084 0 : KnownAssignedXidsReset();
1085 0 : standbyState = STANDBY_INITIALIZED;
1086 : }
1087 : else
1088 : {
1089 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1090 : running->oldestRunningXid))
1091 : {
1092 0 : standbyState = STANDBY_SNAPSHOT_READY;
1093 0 : elog(trace_recovery(DEBUG1),
1094 : "recovery snapshots are now enabled");
1095 : }
1096 : else
1097 0 : elog(trace_recovery(DEBUG1),
1098 : "recovery snapshot waiting for non-overflowed snapshot or "
1099 : "until oldest active xid on standby is at least %u (now %u)",
1100 : standbySnapshotPendingXmin,
1101 : running->oldestRunningXid);
1102 0 : return;
1103 : }
1104 : }
1105 :
1106 : Assert(standbyState == STANDBY_INITIALIZED);
1107 :
1108 : /*
1109 : * NB: this can be reached at least twice, so make sure new code can deal
1110 : * with that.
1111 : */
1112 :
1113 : /*
1114 : * Nobody else is running yet, but take locks anyhow
1115 : */
1116 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1117 :
1118 : /*
1119 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1120 : * sort them first.
1121 : *
1122 : * Some of the new xids are top-level xids and some are subtransactions.
1123 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1124 : * aren't overflowed then all xids will fit in snapshot and so we don't
1125 : * need subtrans. If we later overflow, an xid assignment record will add
1126 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1127 : * don't have enough information to correctly update subtrans anyway.
1128 : */
1129 :
1130 : /*
1131 : * Allocate a temporary array to avoid modifying the array passed as
1132 : * argument.
1133 : */
1134 114 : xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt));
1135 :
1136 : /*
1137 : * Add to the temp array any xids which have not already completed.
1138 : */
1139 114 : nxids = 0;
1140 118 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1141 : {
1142 4 : TransactionId xid = running->xids[i];
1143 :
1144 : /*
1145 : * The running-xacts snapshot can contain xids that were still visible
1146 : * in the procarray when the snapshot was taken, but were already
1147 : * WAL-logged as completed. They're not running anymore, so ignore
1148 : * them.
1149 : */
1150 4 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1151 0 : continue;
1152 :
1153 4 : xids[nxids++] = xid;
1154 : }
1155 :
1156 114 : if (nxids > 0)
1157 : {
1158 4 : if (procArray->numKnownAssignedXids != 0)
1159 : {
1160 0 : LWLockRelease(ProcArrayLock);
1161 0 : elog(ERROR, "KnownAssignedXids is not empty");
1162 : }
1163 :
1164 : /*
1165 : * Sort the array so that we can add them safely into
1166 : * KnownAssignedXids.
1167 : */
1168 4 : qsort(xids, nxids, sizeof(TransactionId), xidComparator);
1169 :
1170 : /*
1171 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1172 : * snapshot may include duplicated xids because of prepared
1173 : * transactions, so ignore them.
1174 : */
1175 8 : for (i = 0; i < nxids; i++)
1176 : {
1177 4 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1178 : {
1179 0 : elog(DEBUG1,
1180 : "found duplicated transaction %u for KnownAssignedXids insertion",
1181 : xids[i]);
1182 0 : continue;
1183 : }
1184 4 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1185 : }
1186 :
1187 4 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1188 : }
1189 :
1190 114 : pfree(xids);
1191 :
1192 : /*
1193 : * latestObservedXid is at least set to the point where SUBTRANS was
1194 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1195 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1196 : * subtrans from thereon, up to nextXid - 1.
1197 : *
1198 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1199 : * because we've just added xids to the known assigned xids machinery that
1200 : * haven't gone through RecordKnownAssignedTransactionId().
1201 : */
1202 : Assert(TransactionIdIsNormal(latestObservedXid));
1203 114 : TransactionIdAdvance(latestObservedXid);
1204 114 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1205 : {
1206 0 : ExtendSUBTRANS(latestObservedXid);
1207 0 : TransactionIdAdvance(latestObservedXid);
1208 : }
1209 114 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1210 :
1211 : /* ----------
1212 : * Now we've got the running xids we need to set the global values that
1213 : * are used to track snapshots as they evolve further.
1214 : *
1215 : * - latestCompletedXid which will be the xmax for snapshots
1216 : * - lastOverflowedXid which shows whether snapshots overflow
1217 : * - nextXid
1218 : *
1219 : * If the snapshot overflowed, then we still initialise with what we know,
1220 : * but the recovery snapshot isn't fully valid yet because we know there
1221 : * are some subxids missing. We don't know the specific subxids that are
1222 : * missing, so conservatively assume the last one is latestObservedXid.
1223 : * ----------
1224 : */
1225 114 : if (running->subxid_overflow)
1226 : {
1227 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1228 :
1229 0 : standbySnapshotPendingXmin = latestObservedXid;
1230 0 : procArray->lastOverflowedXid = latestObservedXid;
1231 : }
1232 : else
1233 : {
1234 114 : standbyState = STANDBY_SNAPSHOT_READY;
1235 :
1236 114 : standbySnapshotPendingXmin = InvalidTransactionId;
1237 : }
1238 :
1239 : /*
1240 : * If a transaction wrote a commit record in the gap between taking and
1241 : * logging the snapshot then latestCompletedXid may already be higher than
1242 : * the value from the snapshot, so check before we use the incoming value.
1243 : * It also might not yet be set at all.
1244 : */
1245 114 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1246 :
1247 : /*
1248 : * NB: No need to increment ShmemVariableCache->xactCompletionCount here,
1249 : * nobody can see it yet.
1250 : */
1251 :
1252 114 : LWLockRelease(ProcArrayLock);
1253 :
1254 : /* ShmemVariableCache->nextXid must be beyond any observed xid. */
1255 114 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
1256 :
1257 : Assert(FullTransactionIdIsValid(ShmemVariableCache->nextXid));
1258 :
1259 114 : KnownAssignedXidsDisplay(trace_recovery(DEBUG3));
1260 114 : if (standbyState == STANDBY_SNAPSHOT_READY)
1261 114 : elog(trace_recovery(DEBUG1), "recovery snapshots are now enabled");
1262 : else
1263 0 : elog(trace_recovery(DEBUG1),
1264 : "recovery snapshot waiting for non-overflowed snapshot or "
1265 : "until oldest active xid on standby is at least %u (now %u)",
1266 : standbySnapshotPendingXmin,
1267 : running->oldestRunningXid);
1268 : }
1269 :
1270 : /*
1271 : * ProcArrayApplyXidAssignment
1272 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1273 : */
1274 : void
1275 40 : ProcArrayApplyXidAssignment(TransactionId topxid,
1276 : int nsubxids, TransactionId *subxids)
1277 : {
1278 : TransactionId max_xid;
1279 : int i;
1280 :
1281 : Assert(standbyState >= STANDBY_INITIALIZED);
1282 :
1283 40 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1284 :
1285 : /*
1286 : * Mark all the subtransactions as observed.
1287 : *
1288 : * NOTE: This will fail if the subxid contains too many previously
1289 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1290 : * as the code stands, because xid-assignment records should never contain
1291 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1292 : */
1293 40 : RecordKnownAssignedTransactionIds(max_xid);
1294 :
1295 : /*
1296 : * Notice that we update pg_subtrans with the top-level xid, rather than
1297 : * the parent xid. This is a difference between normal processing and
1298 : * recovery, yet is still correct in all cases. The reason is that
1299 : * subtransaction commit is not marked in clog until commit processing, so
1300 : * all aborted subtransactions have already been clearly marked in clog.
1301 : * As a result we are able to refer directly to the top-level
1302 : * transaction's state rather than skipping through all the intermediate
1303 : * states in the subtransaction tree. This should be the first time we
1304 : * have attempted to SubTransSetParent().
1305 : */
1306 2600 : for (i = 0; i < nsubxids; i++)
1307 2560 : SubTransSetParent(subxids[i], topxid);
1308 :
1309 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1310 40 : if (standbyState == STANDBY_INITIALIZED)
1311 0 : return;
1312 :
1313 : /*
1314 : * Uses same locking as transaction commit
1315 : */
1316 40 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1317 :
1318 : /*
1319 : * Remove subxids from known-assigned-xacts.
1320 : */
1321 40 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1322 :
1323 : /*
1324 : * Advance lastOverflowedXid to be at least the last of these subxids.
1325 : */
1326 40 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1327 40 : procArray->lastOverflowedXid = max_xid;
1328 :
1329 40 : LWLockRelease(ProcArrayLock);
1330 : }
1331 :
1332 : /*
1333 : * TransactionIdIsInProgress -- is given transaction running in some backend
1334 : *
1335 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1336 : * there are four possibilities for finding a running transaction:
1337 : *
1338 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1339 : * by looking at ProcGlobal->xids.
1340 : *
1341 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1342 : * We can find this out cheaply too.
1343 : *
1344 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1345 : * if the Xid is running on the primary.
1346 : *
1347 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1348 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1349 : * This is the slowest way, but sadly it has to be done always if the others
1350 : * failed, unless we see that the cached subxact sets are complete (none have
1351 : * overflowed).
1352 : *
1353 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1354 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1355 : * This buys back some concurrency (and we can't retrieve the main Xids from
1356 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1357 : */
1358 : bool
1359 10765240 : TransactionIdIsInProgress(TransactionId xid)
1360 : {
1361 : static TransactionId *xids = NULL;
1362 : static TransactionId *other_xids;
1363 : XidCacheStatus *other_subxidstates;
1364 10765240 : int nxids = 0;
1365 10765240 : ProcArrayStruct *arrayP = procArray;
1366 : TransactionId topxid;
1367 : TransactionId latestCompletedXid;
1368 : int mypgxactoff;
1369 : int numProcs;
1370 : int j;
1371 :
1372 : /*
1373 : * Don't bother checking a transaction older than RecentXmin; it could not
1374 : * possibly still be running. (Note: in particular, this guarantees that
1375 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1376 : * running.)
1377 : */
1378 10765240 : if (TransactionIdPrecedes(xid, RecentXmin))
1379 : {
1380 : xc_by_recent_xmin_inc();
1381 9319020 : return false;
1382 : }
1383 :
1384 : /*
1385 : * We may have just checked the status of this transaction, so if it is
1386 : * already known to be completed, we can fall out without any access to
1387 : * shared memory.
1388 : */
1389 1446220 : if (TransactionIdIsKnownCompleted(xid))
1390 : {
1391 : xc_by_known_xact_inc();
1392 990764 : return false;
1393 : }
1394 :
1395 : /*
1396 : * Also, we can handle our own transaction (and subtransactions) without
1397 : * any access to shared memory.
1398 : */
1399 455456 : if (TransactionIdIsCurrentTransactionId(xid))
1400 : {
1401 : xc_by_my_xact_inc();
1402 430492 : return true;
1403 : }
1404 :
1405 : /*
1406 : * If first time through, get workspace to remember main XIDs in. We
1407 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1408 : */
1409 24964 : if (xids == NULL)
1410 : {
1411 : /*
1412 : * In hot standby mode, reserve enough space to hold all xids in the
1413 : * known-assigned list. If we later finish recovery, we no longer need
1414 : * the bigger array, but we don't bother to shrink it.
1415 : */
1416 810 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1417 :
1418 810 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1419 810 : if (xids == NULL)
1420 0 : ereport(ERROR,
1421 : (errcode(ERRCODE_OUT_OF_MEMORY),
1422 : errmsg("out of memory")));
1423 : }
1424 :
1425 24964 : other_xids = ProcGlobal->xids;
1426 24964 : other_subxidstates = ProcGlobal->subxidStates;
1427 :
1428 24964 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1429 :
1430 : /*
1431 : * Now that we have the lock, we can check latestCompletedXid; if the
1432 : * target Xid is after that, it's surely still running.
1433 : */
1434 24964 : latestCompletedXid =
1435 24964 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
1436 24964 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1437 : {
1438 8402 : LWLockRelease(ProcArrayLock);
1439 : xc_by_latest_xid_inc();
1440 8402 : return true;
1441 : }
1442 :
1443 : /* No shortcuts, gotta grovel through the array */
1444 16562 : mypgxactoff = MyProc->pgxactoff;
1445 16562 : numProcs = arrayP->numProcs;
1446 113856 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1447 : {
1448 : int pgprocno;
1449 : PGPROC *proc;
1450 : TransactionId pxid;
1451 : int pxids;
1452 :
1453 : /* Ignore ourselves --- dealt with it above */
1454 101502 : if (pgxactoff == mypgxactoff)
1455 13398 : continue;
1456 :
1457 : /* Fetch xid just once - see GetNewTransactionId */
1458 88104 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1459 :
1460 88104 : if (!TransactionIdIsValid(pxid))
1461 62686 : continue;
1462 :
1463 : /*
1464 : * Step 1: check the main Xid
1465 : */
1466 25418 : if (TransactionIdEquals(pxid, xid))
1467 : {
1468 3184 : LWLockRelease(ProcArrayLock);
1469 : xc_by_main_xid_inc();
1470 3184 : return true;
1471 : }
1472 :
1473 : /*
1474 : * We can ignore main Xids that are younger than the target Xid, since
1475 : * the target could not possibly be their child.
1476 : */
1477 22234 : if (TransactionIdPrecedes(xid, pxid))
1478 5754 : continue;
1479 :
1480 : /*
1481 : * Step 2: check the cached child-Xids arrays
1482 : */
1483 16480 : pxids = other_subxidstates[pgxactoff].count;
1484 16480 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1485 16480 : pgprocno = arrayP->pgprocnos[pgxactoff];
1486 16480 : proc = &allProcs[pgprocno];
1487 80576 : for (j = pxids - 1; j >= 0; j--)
1488 : {
1489 : /* Fetch xid just once - see GetNewTransactionId */
1490 65120 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1491 :
1492 65120 : if (TransactionIdEquals(cxid, xid))
1493 : {
1494 1024 : LWLockRelease(ProcArrayLock);
1495 : xc_by_child_xid_inc();
1496 1024 : return true;
1497 : }
1498 : }
1499 :
1500 : /*
1501 : * Save the main Xid for step 4. We only need to remember main Xids
1502 : * that have uncached children. (Note: there is no race condition
1503 : * here because the overflowed flag cannot be cleared, only set, while
1504 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1505 : * worry about.)
1506 : */
1507 15456 : if (other_subxidstates[pgxactoff].overflowed)
1508 310 : xids[nxids++] = pxid;
1509 : }
1510 :
1511 : /*
1512 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1513 : * in the list must be treated as running.
1514 : */
1515 12354 : if (RecoveryInProgress())
1516 : {
1517 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1518 : Assert(nxids == 0);
1519 :
1520 0 : if (KnownAssignedXidExists(xid))
1521 : {
1522 0 : LWLockRelease(ProcArrayLock);
1523 : xc_by_known_assigned_inc();
1524 0 : return true;
1525 : }
1526 :
1527 : /*
1528 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1529 : * too. Fetch all xids from KnownAssignedXids that are lower than
1530 : * xid, since if xid is a subtransaction its parent will always have a
1531 : * lower value. Note we will collect both main and subXIDs here, but
1532 : * there's no help for it.
1533 : */
1534 0 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1535 0 : nxids = KnownAssignedXidsGet(xids, xid);
1536 : }
1537 :
1538 12354 : LWLockRelease(ProcArrayLock);
1539 :
1540 : /*
1541 : * If none of the relevant caches overflowed, we know the Xid is not
1542 : * running without even looking at pg_subtrans.
1543 : */
1544 12354 : if (nxids == 0)
1545 : {
1546 : xc_no_overflow_inc();
1547 12044 : return false;
1548 : }
1549 :
1550 : /*
1551 : * Step 4: have to check pg_subtrans.
1552 : *
1553 : * At this point, we know it's either a subtransaction of one of the Xids
1554 : * in xids[], or it's not running. If it's an already-failed
1555 : * subtransaction, we want to say "not running" even though its parent may
1556 : * still be running. So first, check pg_xact to see if it's been aborted.
1557 : */
1558 : xc_slow_answer_inc();
1559 :
1560 310 : if (TransactionIdDidAbort(xid))
1561 0 : return false;
1562 :
1563 : /*
1564 : * It isn't aborted, so check whether the transaction tree it belongs to
1565 : * is still running (or, more precisely, whether it was running when we
1566 : * held ProcArrayLock).
1567 : */
1568 310 : topxid = SubTransGetTopmostTransaction(xid);
1569 : Assert(TransactionIdIsValid(topxid));
1570 310 : if (!TransactionIdEquals(topxid, xid))
1571 : {
1572 310 : for (int i = 0; i < nxids; i++)
1573 : {
1574 310 : if (TransactionIdEquals(xids[i], topxid))
1575 310 : return true;
1576 : }
1577 : }
1578 :
1579 0 : return false;
1580 : }
1581 :
1582 : /*
1583 : * TransactionIdIsActive -- is xid the top-level XID of an active backend?
1584 : *
1585 : * This differs from TransactionIdIsInProgress in that it ignores prepared
1586 : * transactions, as well as transactions running on the primary if we're in
1587 : * hot standby. Also, we ignore subtransactions since that's not needed
1588 : * for current uses.
1589 : */
1590 : bool
1591 0 : TransactionIdIsActive(TransactionId xid)
1592 : {
1593 0 : bool result = false;
1594 0 : ProcArrayStruct *arrayP = procArray;
1595 0 : TransactionId *other_xids = ProcGlobal->xids;
1596 : int i;
1597 :
1598 : /*
1599 : * Don't bother checking a transaction older than RecentXmin; it could not
1600 : * possibly still be running.
1601 : */
1602 0 : if (TransactionIdPrecedes(xid, RecentXmin))
1603 0 : return false;
1604 :
1605 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1606 :
1607 0 : for (i = 0; i < arrayP->numProcs; i++)
1608 : {
1609 0 : int pgprocno = arrayP->pgprocnos[i];
1610 0 : PGPROC *proc = &allProcs[pgprocno];
1611 : TransactionId pxid;
1612 :
1613 : /* Fetch xid just once - see GetNewTransactionId */
1614 0 : pxid = UINT32_ACCESS_ONCE(other_xids[i]);
1615 :
1616 0 : if (!TransactionIdIsValid(pxid))
1617 0 : continue;
1618 :
1619 0 : if (proc->pid == 0)
1620 0 : continue; /* ignore prepared transactions */
1621 :
1622 0 : if (TransactionIdEquals(pxid, xid))
1623 : {
1624 0 : result = true;
1625 0 : break;
1626 : }
1627 : }
1628 :
1629 0 : LWLockRelease(ProcArrayLock);
1630 :
1631 0 : return result;
1632 : }
1633 :
1634 :
1635 : /*
1636 : * Determine XID horizons.
1637 : *
1638 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1639 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1640 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1641 : * GlobalVisState).
1642 : *
1643 : * See the definition of ComputeXidHorizonsResult for the various computed
1644 : * horizons.
1645 : *
1646 : * For VACUUM separate horizons (used to decide which deleted tuples must
1647 : * be preserved), for shared and non-shared tables are computed. For shared
1648 : * relations backends in all databases must be considered, but for non-shared
1649 : * relations that's not required, since only backends in my own database could
1650 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1651 : * VACUUMs because (a) they must be working on other tables, and (b) they
1652 : * don't need to do snapshot-based lookups. Similarly, for the non-catalog
1653 : * horizon, we can ignore CREATE INDEX CONCURRENTLY and REINDEX CONCURRENTLY
1654 : * when they are working on non-partial, non-expressional indexes, for the
1655 : * same reasons and because they can't run in transaction blocks. (They are
1656 : * not possible to ignore for catalogs, because CIC and RC do some catalog
1657 : * operations.) Do note that this means that CIC and RC must use a lock level
1658 : * that conflicts with VACUUM.
1659 : *
1660 : * This also computes a horizon used to truncate pg_subtrans. For that
1661 : * backends in all databases have to be considered, and concurrently running
1662 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1663 : * accesses.
1664 : *
1665 : * Note: we include all currently running xids in the set of considered xids.
1666 : * This ensures that if a just-started xact has not yet set its snapshot,
1667 : * when it does set the snapshot it cannot set xmin less than what we compute.
1668 : * See notes in src/backend/access/transam/README.
1669 : *
1670 : * Note: despite the above, it's possible for the calculated values to move
1671 : * backwards on repeated calls. The calculated values are conservative, so
1672 : * that anything older is definitely not considered as running by anyone
1673 : * anymore, but the exact values calculated depend on a number of things. For
1674 : * example, if there are no transactions running in the current database, the
1675 : * horizon for normal tables will be latestCompletedXid. If a transaction
1676 : * begins after that, its xmin will include in-progress transactions in other
1677 : * databases that started earlier, so another call will return a lower value.
1678 : * Nonetheless it is safe to vacuum a table in the current database with the
1679 : * first result. There are also replication-related effects: a walsender
1680 : * process can set its xmin based on transactions that are no longer running
1681 : * on the primary but are still being replayed on the standby, thus possibly
1682 : * making the values go backwards. In this case there is a possibility that
1683 : * we lose data that the standby would like to have, but unless the standby
1684 : * uses a replication slot to make its xmin persistent there is little we can
1685 : * do about that --- data is only protected if the walsender runs continuously
1686 : * while queries are executed on the standby. (The Hot Standby code deals
1687 : * with such cases by failing standby queries that needed to access
1688 : * already-removed data, so there's no integrity bug.) The computed values
1689 : * are also adjusted with vacuum_defer_cleanup_age, so increasing that setting
1690 : * on the fly is another easy way to make horizons move backwards, with no
1691 : * consequences for data integrity.
1692 : *
1693 : * Note: the approximate horizons (see definition of GlobalVisState) are
1694 : * updated by the computations done here. That's currently required for
1695 : * correctness and a small optimization. Without doing so it's possible that
1696 : * heap vacuum's call to heap_page_prune() uses a more conservative horizon
1697 : * than later when deciding which tuples can be removed - which the code
1698 : * doesn't expect (breaking HOT).
1699 : */
1700 : static void
1701 314222 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1702 : {
1703 314222 : ProcArrayStruct *arrayP = procArray;
1704 : TransactionId kaxmin;
1705 314222 : bool in_recovery = RecoveryInProgress();
1706 314222 : TransactionId *other_xids = ProcGlobal->xids;
1707 :
1708 314222 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1709 :
1710 314222 : h->latest_completed = ShmemVariableCache->latestCompletedXid;
1711 :
1712 : /*
1713 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1714 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1715 : * and so protects us against overestimating the result due to future
1716 : * additions.
1717 : */
1718 : {
1719 : TransactionId initial;
1720 :
1721 314222 : initial = XidFromFullTransactionId(h->latest_completed);
1722 : Assert(TransactionIdIsValid(initial));
1723 314222 : TransactionIdAdvance(initial);
1724 :
1725 314222 : h->oldest_considered_running = initial;
1726 314222 : h->shared_oldest_nonremovable = initial;
1727 314222 : h->catalog_oldest_nonremovable = initial;
1728 314222 : h->data_oldest_nonremovable = initial;
1729 :
1730 : /*
1731 : * Only modifications made by this backend affect the horizon for
1732 : * temporary relations. Instead of a check in each iteration of the
1733 : * loop over all PGPROCs it is cheaper to just initialize to the
1734 : * current top-level xid any.
1735 : *
1736 : * Without an assigned xid we could use a horizon as aggressive as
1737 : * ReadNewTransactionid(), but we can get away with the much cheaper
1738 : * latestCompletedXid + 1: If this backend has no xid there, by
1739 : * definition, can't be any newer changes in the temp table than
1740 : * latestCompletedXid.
1741 : */
1742 314222 : if (TransactionIdIsValid(MyProc->xid))
1743 78756 : h->temp_oldest_nonremovable = MyProc->xid;
1744 : else
1745 235466 : h->temp_oldest_nonremovable = initial;
1746 : }
1747 :
1748 : /*
1749 : * Fetch slot horizons while ProcArrayLock is held - the
1750 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1751 : * the lock.
1752 : */
1753 314222 : h->slot_xmin = procArray->replication_slot_xmin;
1754 314222 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1755 :
1756 993560 : for (int index = 0; index < arrayP->numProcs; index++)
1757 : {
1758 679338 : int pgprocno = arrayP->pgprocnos[index];
1759 679338 : PGPROC *proc = &allProcs[pgprocno];
1760 679338 : int8 statusFlags = ProcGlobal->statusFlags[index];
1761 : TransactionId xid;
1762 : TransactionId xmin;
1763 :
1764 : /* Fetch xid just once - see GetNewTransactionId */
1765 679338 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1766 679338 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1767 :
1768 : /*
1769 : * Consider both the transaction's Xmin, and its Xid.
1770 : *
1771 : * We must check both because a transaction might have an Xmin but not
1772 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1773 : * some not-yet-set Xmin.
1774 : */
1775 679338 : xmin = TransactionIdOlder(xmin, xid);
1776 :
1777 : /* if neither is set, this proc doesn't influence the horizon */
1778 679338 : if (!TransactionIdIsValid(xmin))
1779 270620 : continue;
1780 :
1781 : /*
1782 : * Don't ignore any procs when determining which transactions might be
1783 : * considered running. While slots should ensure logical decoding
1784 : * backends are protected even without this check, it can't hurt to
1785 : * include them here as well..
1786 : */
1787 408718 : h->oldest_considered_running =
1788 408718 : TransactionIdOlder(h->oldest_considered_running, xmin);
1789 :
1790 : /*
1791 : * Skip over backends either vacuuming (which is ok with rows being
1792 : * removed, as long as pg_subtrans is not truncated) or doing logical
1793 : * decoding (which manages xmin separately, check below).
1794 : */
1795 408718 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1796 58874 : continue;
1797 :
1798 : /* shared tables need to take backends in all databases into account */
1799 349844 : h->shared_oldest_nonremovable =
1800 349844 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1801 :
1802 : /*
1803 : * Normally queries in other databases are ignored for anything but
1804 : * the shared horizon. But in recovery we cannot compute an accurate
1805 : * per-database horizon as all xids are managed via the
1806 : * KnownAssignedXids machinery.
1807 : *
1808 : * Be careful to compute a pessimistic value when MyDatabaseId is not
1809 : * set. If this is a backend in the process of starting up, we may not
1810 : * use a "too aggressive" horizon (otherwise we could end up using it
1811 : * to prune still needed data away). If the current backend never
1812 : * connects to a database that is harmless, because
1813 : * data_oldest_nonremovable will never be utilized.
1814 : */
1815 349844 : if (in_recovery ||
1816 349744 : MyDatabaseId == InvalidOid || proc->databaseId == MyDatabaseId ||
1817 254 : proc->databaseId == 0) /* always include WalSender */
1818 : {
1819 : /*
1820 : * We can ignore this backend if it's running CREATE INDEX
1821 : * CONCURRENTLY or REINDEX CONCURRENTLY on a "safe" index -- but
1822 : * only on vacuums of user-defined tables.
1823 : */
1824 349770 : if (!(statusFlags & PROC_IN_SAFE_IC))
1825 349568 : h->data_oldest_nonremovable =
1826 349568 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1827 :
1828 : /* Catalog tables need to consider all backends in this db */
1829 349770 : h->catalog_oldest_nonremovable =
1830 349770 : TransactionIdOlder(h->catalog_oldest_nonremovable, xmin);
1831 :
1832 : }
1833 : }
1834 :
1835 : /* catalog horizon should never be later than data */
1836 : Assert(TransactionIdPrecedesOrEquals(h->catalog_oldest_nonremovable,
1837 : h->data_oldest_nonremovable));
1838 :
1839 : /*
1840 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1841 : * after lock is released.
1842 : */
1843 314222 : if (in_recovery)
1844 160 : kaxmin = KnownAssignedXidsGetOldestXmin();
1845 :
1846 : /*
1847 : * No other information from shared state is needed, release the lock
1848 : * immediately. The rest of the computations can be done without a lock.
1849 : */
1850 314222 : LWLockRelease(ProcArrayLock);
1851 :
1852 314222 : if (in_recovery)
1853 : {
1854 160 : h->oldest_considered_running =
1855 160 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1856 160 : h->shared_oldest_nonremovable =
1857 160 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1858 160 : h->data_oldest_nonremovable =
1859 160 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1860 160 : h->catalog_oldest_nonremovable =
1861 160 : TransactionIdOlder(h->catalog_oldest_nonremovable, kaxmin);
1862 : /* temp relations cannot be accessed in recovery */
1863 : }
1864 : else
1865 : {
1866 : /*
1867 : * Compute the cutoff XID by subtracting vacuum_defer_cleanup_age.
1868 : *
1869 : * vacuum_defer_cleanup_age provides some additional "slop" for the
1870 : * benefit of hot standby queries on standby servers. This is quick
1871 : * and dirty, and perhaps not all that useful unless the primary has a
1872 : * predictable transaction rate, but it offers some protection when
1873 : * there's no walsender connection. Note that we are assuming
1874 : * vacuum_defer_cleanup_age isn't large enough to cause wraparound ---
1875 : * so guc.c should limit it to no more than the xidStopLimit threshold
1876 : * in varsup.c. Also note that we intentionally don't apply
1877 : * vacuum_defer_cleanup_age on standby servers.
1878 : */
1879 314062 : h->oldest_considered_running =
1880 314062 : TransactionIdRetreatedBy(h->oldest_considered_running,
1881 : vacuum_defer_cleanup_age);
1882 314062 : h->shared_oldest_nonremovable =
1883 314062 : TransactionIdRetreatedBy(h->shared_oldest_nonremovable,
1884 : vacuum_defer_cleanup_age);
1885 314062 : h->data_oldest_nonremovable =
1886 314062 : TransactionIdRetreatedBy(h->data_oldest_nonremovable,
1887 : vacuum_defer_cleanup_age);
1888 314062 : h->catalog_oldest_nonremovable =
1889 314062 : TransactionIdRetreatedBy(h->catalog_oldest_nonremovable,
1890 : vacuum_defer_cleanup_age);
1891 : /* defer doesn't apply to temp relations */
1892 : }
1893 :
1894 : /*
1895 : * Check whether there are replication slots requiring an older xmin.
1896 : */
1897 314222 : h->shared_oldest_nonremovable =
1898 314222 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1899 314222 : h->data_oldest_nonremovable =
1900 314222 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1901 :
1902 : /*
1903 : * The only difference between catalog / data horizons is that the slot's
1904 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1905 : * for logical decoding). Initialize with data horizon, and then back up
1906 : * further if necessary. Have to back up the shared horizon as well, since
1907 : * that also can contain catalogs.
1908 : */
1909 314222 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1910 314222 : h->shared_oldest_nonremovable =
1911 314222 : TransactionIdOlder(h->shared_oldest_nonremovable,
1912 : h->slot_catalog_xmin);
1913 314222 : h->catalog_oldest_nonremovable =
1914 314222 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1915 : h->slot_xmin);
1916 314222 : h->catalog_oldest_nonremovable =
1917 314222 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1918 : h->slot_catalog_xmin);
1919 :
1920 : /*
1921 : * It's possible that slots / vacuum_defer_cleanup_age backed up the
1922 : * horizons further than oldest_considered_running. Fix.
1923 : */
1924 314222 : h->oldest_considered_running =
1925 314222 : TransactionIdOlder(h->oldest_considered_running,
1926 : h->shared_oldest_nonremovable);
1927 314222 : h->oldest_considered_running =
1928 314222 : TransactionIdOlder(h->oldest_considered_running,
1929 : h->catalog_oldest_nonremovable);
1930 314222 : h->oldest_considered_running =
1931 314222 : TransactionIdOlder(h->oldest_considered_running,
1932 : h->data_oldest_nonremovable);
1933 :
1934 : /*
1935 : * shared horizons have to be at least as old as the oldest visible in
1936 : * current db
1937 : */
1938 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1939 : h->data_oldest_nonremovable));
1940 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1941 : h->catalog_oldest_nonremovable));
1942 :
1943 : /*
1944 : * Horizons need to ensure that pg_subtrans access is still possible for
1945 : * the relevant backends.
1946 : */
1947 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1948 : h->shared_oldest_nonremovable));
1949 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1950 : h->catalog_oldest_nonremovable));
1951 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1952 : h->data_oldest_nonremovable));
1953 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1954 : h->temp_oldest_nonremovable));
1955 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1956 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1957 : h->slot_xmin));
1958 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1959 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1960 : h->slot_catalog_xmin));
1961 :
1962 : /* update approximate horizons with the computed horizons */
1963 314222 : GlobalVisUpdateApply(h);
1964 314222 : }
1965 :
1966 : /*
1967 : * Determine what kind of visibility horizon needs to be used for a
1968 : * relation. If rel is NULL, the most conservative horizon is used.
1969 : */
1970 : static inline GlobalVisHorizonKind
1971 19269042 : GlobalVisHorizonKindForRel(Relation rel)
1972 : {
1973 : /*
1974 : * Other relkkinds currently don't contain xids, nor always the necessary
1975 : * logical decoding markers.
1976 : */
1977 : Assert(!rel ||
1978 : rel->rd_rel->relkind == RELKIND_RELATION ||
1979 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1980 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1981 :
1982 19269042 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1983 182732 : return VISHORIZON_SHARED;
1984 19086310 : else if (IsCatalogRelation(rel) ||
1985 15256150 : RelationIsAccessibleInLogicalDecoding(rel))
1986 3830162 : return VISHORIZON_CATALOG;
1987 15256148 : else if (!RELATION_IS_LOCAL(rel))
1988 15177678 : return VISHORIZON_DATA;
1989 : else
1990 78470 : return VISHORIZON_TEMP;
1991 : }
1992 :
1993 : /*
1994 : * Return the oldest XID for which deleted tuples must be preserved in the
1995 : * passed table.
1996 : *
1997 : * If rel is not NULL the horizon may be considerably more recent than
1998 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1999 : * that is correct (but not optimal) for all relations will be returned.
2000 : *
2001 : * This is used by VACUUM to decide which deleted tuples must be preserved in
2002 : * the passed in table.
2003 : */
2004 : TransactionId
2005 126066 : GetOldestNonRemovableTransactionId(Relation rel)
2006 : {
2007 : ComputeXidHorizonsResult horizons;
2008 :
2009 126066 : ComputeXidHorizons(&horizons);
2010 :
2011 126066 : switch (GlobalVisHorizonKindForRel(rel))
2012 : {
2013 20606 : case VISHORIZON_SHARED:
2014 20606 : return horizons.shared_oldest_nonremovable;
2015 72934 : case VISHORIZON_CATALOG:
2016 72934 : return horizons.catalog_oldest_nonremovable;
2017 16618 : case VISHORIZON_DATA:
2018 16618 : return horizons.data_oldest_nonremovable;
2019 15908 : case VISHORIZON_TEMP:
2020 15908 : return horizons.temp_oldest_nonremovable;
2021 : }
2022 :
2023 : /* just to prevent compiler warnings */
2024 0 : return InvalidTransactionId;
2025 : }
2026 :
2027 : /*
2028 : * Return the oldest transaction id any currently running backend might still
2029 : * consider running. This should not be used for visibility / pruning
2030 : * determinations (see GetOldestNonRemovableTransactionId()), but for
2031 : * decisions like up to where pg_subtrans can be truncated.
2032 : */
2033 : TransactionId
2034 4096 : GetOldestTransactionIdConsideredRunning(void)
2035 : {
2036 : ComputeXidHorizonsResult horizons;
2037 :
2038 4096 : ComputeXidHorizons(&horizons);
2039 :
2040 4096 : return horizons.oldest_considered_running;
2041 : }
2042 :
2043 : /*
2044 : * Return the visibility horizons for a hot standby feedback message.
2045 : */
2046 : void
2047 18 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
2048 : {
2049 : ComputeXidHorizonsResult horizons;
2050 :
2051 18 : ComputeXidHorizons(&horizons);
2052 :
2053 : /*
2054 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2055 : * effect of replication slot's catalog_xmin. We want to send a separate
2056 : * feedback for the catalog horizon, so the primary can remove data table
2057 : * contents more aggressively.
2058 : */
2059 18 : *xmin = horizons.shared_oldest_nonremovable_raw;
2060 18 : *catalog_xmin = horizons.slot_catalog_xmin;
2061 18 : }
2062 :
2063 : /*
2064 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2065 : *
2066 : * We have to export this for use by snapmgr.c.
2067 : */
2068 : int
2069 32332 : GetMaxSnapshotXidCount(void)
2070 : {
2071 32332 : return procArray->maxProcs;
2072 : }
2073 :
2074 : /*
2075 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2076 : *
2077 : * We have to export this for use by snapmgr.c.
2078 : */
2079 : int
2080 32164 : GetMaxSnapshotSubxidCount(void)
2081 : {
2082 32164 : return TOTAL_MAX_CACHED_SUBXIDS;
2083 : }
2084 :
2085 : /*
2086 : * Initialize old_snapshot_threshold specific parts of a newly build snapshot.
2087 : */
2088 : static void
2089 3101748 : GetSnapshotDataInitOldSnapshot(Snapshot snapshot)
2090 : {
2091 3101748 : if (!OldSnapshotThresholdActive())
2092 : {
2093 : /*
2094 : * If not using "snapshot too old" feature, fill related fields with
2095 : * dummy values that don't require any locking.
2096 : */
2097 3077788 : snapshot->lsn = InvalidXLogRecPtr;
2098 3077788 : snapshot->whenTaken = 0;
2099 : }
2100 : else
2101 : {
2102 : /*
2103 : * Capture the current time and WAL stream location in case this
2104 : * snapshot becomes old enough to need to fall back on the special
2105 : * "old snapshot" logic.
2106 : */
2107 23960 : snapshot->lsn = GetXLogInsertRecPtr();
2108 23960 : snapshot->whenTaken = GetSnapshotCurrentTimestamp();
2109 23960 : MaintainOldSnapshotTimeMapping(snapshot->whenTaken, snapshot->xmin);
2110 : }
2111 3101748 : }
2112 :
2113 : /*
2114 : * Helper function for GetSnapshotData() that checks if the bulk of the
2115 : * visibility information in the snapshot is still valid. If so, it updates
2116 : * the fields that need to change and returns true. Otherwise it returns
2117 : * false.
2118 : *
2119 : * This very likely can be evolved to not need ProcArrayLock held (at very
2120 : * least in the case we already hold a snapshot), but that's for another day.
2121 : */
2122 : static bool
2123 3101748 : GetSnapshotDataReuse(Snapshot snapshot)
2124 : {
2125 : uint64 curXactCompletionCount;
2126 :
2127 : Assert(LWLockHeldByMe(ProcArrayLock));
2128 :
2129 3101748 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2130 32140 : return false;
2131 :
2132 3069608 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2133 3069608 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2134 949982 : return false;
2135 :
2136 : /*
2137 : * If the current xactCompletionCount is still the same as it was at the
2138 : * time the snapshot was built, we can be sure that rebuilding the
2139 : * contents of the snapshot the hard way would result in the same snapshot
2140 : * contents:
2141 : *
2142 : * As explained in transam/README, the set of xids considered running by
2143 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2144 : * contents only depend on transactions with xids and xactCompletionCount
2145 : * is incremented whenever a transaction with an xid finishes (while
2146 : * holding ProcArrayLock) exclusively). Thus the xactCompletionCount check
2147 : * ensures we would detect if the snapshot would have changed.
2148 : *
2149 : * As the snapshot contents are the same as it was before, it is safe to
2150 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2151 : * visible under the snapshot could already have been removed (that'd
2152 : * require the set of running transactions to change) and it fulfills the
2153 : * requirement that concurrent GetSnapshotData() calls yield the same
2154 : * xmin.
2155 : */
2156 2119626 : if (!TransactionIdIsValid(MyProc->xmin))
2157 887682 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2158 :
2159 2119626 : RecentXmin = snapshot->xmin;
2160 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2161 :
2162 2119626 : snapshot->curcid = GetCurrentCommandId(false);
2163 2119626 : snapshot->active_count = 0;
2164 2119626 : snapshot->regd_count = 0;
2165 2119626 : snapshot->copied = false;
2166 :
2167 2119626 : GetSnapshotDataInitOldSnapshot(snapshot);
2168 :
2169 2119626 : return true;
2170 : }
2171 :
2172 : /*
2173 : * GetSnapshotData -- returns information about running transactions.
2174 : *
2175 : * The returned snapshot includes xmin (lowest still-running xact ID),
2176 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2177 : * in the range xmin <= xid < xmax. It is used as follows:
2178 : * All xact IDs < xmin are considered finished.
2179 : * All xact IDs >= xmax are considered still running.
2180 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2181 : * it is considered running or not.
2182 : * This ensures that the set of transactions seen as "running" by the
2183 : * current xact will not change after it takes the snapshot.
2184 : *
2185 : * All running top-level XIDs are included in the snapshot, except for lazy
2186 : * VACUUM processes. We also try to include running subtransaction XIDs,
2187 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2188 : * information may not be available. If we find any overflowed subxid arrays,
2189 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2190 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()
2191 : * in heapam_visibility.c).
2192 : *
2193 : * We also update the following backend-global variables:
2194 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2195 : * current transaction (this is the same as MyProc->xmin).
2196 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2197 : * older than this are known not running any more.
2198 : *
2199 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2200 : * for the benefit of the GlobalVisTest* family of functions.
2201 : *
2202 : * Note: this function should probably not be called with an argument that's
2203 : * not statically allocated (see xip allocation below).
2204 : */
2205 : Snapshot
2206 3101748 : GetSnapshotData(Snapshot snapshot)
2207 : {
2208 3101748 : ProcArrayStruct *arrayP = procArray;
2209 3101748 : TransactionId *other_xids = ProcGlobal->xids;
2210 : TransactionId xmin;
2211 : TransactionId xmax;
2212 3101748 : int count = 0;
2213 3101748 : int subcount = 0;
2214 3101748 : bool suboverflowed = false;
2215 : FullTransactionId latest_completed;
2216 : TransactionId oldestxid;
2217 : int mypgxactoff;
2218 : TransactionId myxid;
2219 : uint64 curXactCompletionCount;
2220 :
2221 3101748 : TransactionId replication_slot_xmin = InvalidTransactionId;
2222 3101748 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2223 :
2224 : Assert(snapshot != NULL);
2225 :
2226 : /*
2227 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2228 : * be sufficient. But it seems better to do the malloc while not holding
2229 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2230 : * more subxip storage than is probably needed.
2231 : *
2232 : * This does open a possibility for avoiding repeated malloc/free: since
2233 : * maxProcs does not change at runtime, we can simply reuse the previous
2234 : * xip arrays if any. (This relies on the fact that all callers pass
2235 : * static SnapshotData structs.)
2236 : */
2237 3101748 : if (snapshot->xip == NULL)
2238 : {
2239 : /*
2240 : * First call for this snapshot. Snapshot is same size whether or not
2241 : * we are in recovery, see later comments.
2242 : */
2243 32140 : snapshot->xip = (TransactionId *)
2244 32140 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2245 32140 : if (snapshot->xip == NULL)
2246 0 : ereport(ERROR,
2247 : (errcode(ERRCODE_OUT_OF_MEMORY),
2248 : errmsg("out of memory")));
2249 : Assert(snapshot->subxip == NULL);
2250 32140 : snapshot->subxip = (TransactionId *)
2251 32140 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2252 32140 : if (snapshot->subxip == NULL)
2253 0 : ereport(ERROR,
2254 : (errcode(ERRCODE_OUT_OF_MEMORY),
2255 : errmsg("out of memory")));
2256 : }
2257 :
2258 : /*
2259 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2260 : * going to set MyProc->xmin.
2261 : */
2262 3101748 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2263 :
2264 3101748 : if (GetSnapshotDataReuse(snapshot))
2265 : {
2266 2119626 : LWLockRelease(ProcArrayLock);
2267 2119626 : return snapshot;
2268 : }
2269 :
2270 982122 : latest_completed = ShmemVariableCache->latestCompletedXid;
2271 982122 : mypgxactoff = MyProc->pgxactoff;
2272 982122 : myxid = other_xids[mypgxactoff];
2273 : Assert(myxid == MyProc->xid);
2274 :
2275 982122 : oldestxid = ShmemVariableCache->oldestXid;
2276 982122 : curXactCompletionCount = ShmemVariableCache->xactCompletionCount;
2277 :
2278 : /* xmax is always latestCompletedXid + 1 */
2279 982122 : xmax = XidFromFullTransactionId(latest_completed);
2280 982122 : TransactionIdAdvance(xmax);
2281 : Assert(TransactionIdIsNormal(xmax));
2282 :
2283 : /* initialize xmin calculation with xmax */
2284 982122 : xmin = xmax;
2285 :
2286 : /* take own xid into account, saves a check inside the loop */
2287 982122 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2288 12780 : xmin = myxid;
2289 :
2290 982122 : snapshot->takenDuringRecovery = RecoveryInProgress();
2291 :
2292 982122 : if (!snapshot->takenDuringRecovery)
2293 : {
2294 981284 : int numProcs = arrayP->numProcs;
2295 981284 : TransactionId *xip = snapshot->xip;
2296 981284 : int *pgprocnos = arrayP->pgprocnos;
2297 981284 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2298 981284 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2299 :
2300 : /*
2301 : * First collect set of pgxactoff/xids that need to be included in the
2302 : * snapshot.
2303 : */
2304 3539134 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2305 : {
2306 : /* Fetch xid just once - see GetNewTransactionId */
2307 2557850 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2308 : uint8 statusFlags;
2309 :
2310 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2311 :
2312 : /*
2313 : * If the transaction has no XID assigned, we can skip it; it
2314 : * won't have sub-XIDs either.
2315 : */
2316 2557850 : if (likely(xid == InvalidTransactionId))
2317 2223042 : continue;
2318 :
2319 : /*
2320 : * We don't include our own XIDs (if any) in the snapshot. It
2321 : * needs to be includeded in the xmin computation, but we did so
2322 : * outside the loop.
2323 : */
2324 334808 : if (pgxactoff == mypgxactoff)
2325 79628 : continue;
2326 :
2327 : /*
2328 : * The only way we are able to get here with a non-normal xid is
2329 : * during bootstrap - with this backend using
2330 : * BootstrapTransactionId. But the above test should filter that
2331 : * out.
2332 : */
2333 : Assert(TransactionIdIsNormal(xid));
2334 :
2335 : /*
2336 : * If the XID is >= xmax, we can skip it; such transactions will
2337 : * be treated as running anyway (and any sub-XIDs will also be >=
2338 : * xmax).
2339 : */
2340 255180 : if (!NormalTransactionIdPrecedes(xid, xmax))
2341 53336 : continue;
2342 :
2343 : /*
2344 : * Skip over backends doing logical decoding which manages xmin
2345 : * separately (check below) and ones running LAZY VACUUM.
2346 : */
2347 201844 : statusFlags = allStatusFlags[pgxactoff];
2348 201844 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2349 0 : continue;
2350 :
2351 201844 : if (NormalTransactionIdPrecedes(xid, xmin))
2352 138264 : xmin = xid;
2353 :
2354 : /* Add XID to snapshot. */
2355 201844 : xip[count++] = xid;
2356 :
2357 : /*
2358 : * Save subtransaction XIDs if possible (if we've already
2359 : * overflowed, there's no point). Note that the subxact XIDs must
2360 : * be later than their parent, so no need to check them against
2361 : * xmin. We could filter against xmax, but it seems better not to
2362 : * do that much work while holding the ProcArrayLock.
2363 : *
2364 : * The other backend can add more subxids concurrently, but cannot
2365 : * remove any. Hence it's important to fetch nxids just once.
2366 : * Should be safe to use memcpy, though. (We needn't worry about
2367 : * missing any xids added concurrently, because they must postdate
2368 : * xmax.)
2369 : *
2370 : * Again, our own XIDs are not included in the snapshot.
2371 : */
2372 201844 : if (!suboverflowed)
2373 : {
2374 :
2375 201844 : if (subxidStates[pgxactoff].overflowed)
2376 32 : suboverflowed = true;
2377 : else
2378 : {
2379 201812 : int nsubxids = subxidStates[pgxactoff].count;
2380 :
2381 201812 : if (nsubxids > 0)
2382 : {
2383 3428 : int pgprocno = pgprocnos[pgxactoff];
2384 3428 : PGPROC *proc = &allProcs[pgprocno];
2385 :
2386 3428 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2387 :
2388 3428 : memcpy(snapshot->subxip + subcount,
2389 3428 : (void *) proc->subxids.xids,
2390 : nsubxids * sizeof(TransactionId));
2391 3428 : subcount += nsubxids;
2392 : }
2393 : }
2394 : }
2395 : }
2396 : }
2397 : else
2398 : {
2399 : /*
2400 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2401 : *
2402 : * We store all xids directly into subxip[]. Here's why:
2403 : *
2404 : * In recovery we don't know which xids are top-level and which are
2405 : * subxacts, a design choice that greatly simplifies xid processing.
2406 : *
2407 : * It seems like we would want to try to put xids into xip[] only, but
2408 : * that is fairly small. We would either need to make that bigger or
2409 : * to increase the rate at which we WAL-log xid assignment; neither is
2410 : * an appealing choice.
2411 : *
2412 : * We could try to store xids into xip[] first and then into subxip[]
2413 : * if there are too many xids. That only works if the snapshot doesn't
2414 : * overflow because we do not search subxip[] in that case. A simpler
2415 : * way is to just store all xids in the subxact array because this is
2416 : * by far the bigger array. We just leave the xip array empty.
2417 : *
2418 : * Either way we need to change the way XidInMVCCSnapshot() works
2419 : * depending upon when the snapshot was taken, or change normal
2420 : * snapshot processing so it matches.
2421 : *
2422 : * Note: It is possible for recovery to end before we finish taking
2423 : * the snapshot, and for newly assigned transaction ids to be added to
2424 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2425 : * those newly added transaction ids would be filtered away, so we
2426 : * need not be concerned about them.
2427 : */
2428 838 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2429 : xmax);
2430 :
2431 838 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2432 8 : suboverflowed = true;
2433 : }
2434 :
2435 :
2436 : /*
2437 : * Fetch into local variable while ProcArrayLock is held - the
2438 : * LWLockRelease below is a barrier, ensuring this happens inside the
2439 : * lock.
2440 : */
2441 982122 : replication_slot_xmin = procArray->replication_slot_xmin;
2442 982122 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2443 :
2444 982122 : if (!TransactionIdIsValid(MyProc->xmin))
2445 504708 : MyProc->xmin = TransactionXmin = xmin;
2446 :
2447 982122 : LWLockRelease(ProcArrayLock);
2448 :
2449 : /* maintain state for GlobalVis* */
2450 : {
2451 : TransactionId def_vis_xid;
2452 : TransactionId def_vis_xid_data;
2453 : FullTransactionId def_vis_fxid;
2454 : FullTransactionId def_vis_fxid_data;
2455 : FullTransactionId oldestfxid;
2456 :
2457 : /*
2458 : * Converting oldestXid is only safe when xid horizon cannot advance,
2459 : * i.e. holding locks. While we don't hold the lock anymore, all the
2460 : * necessary data has been gathered with lock held.
2461 : */
2462 982122 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2463 :
2464 : /* apply vacuum_defer_cleanup_age */
2465 : def_vis_xid_data =
2466 982122 : TransactionIdRetreatedBy(xmin, vacuum_defer_cleanup_age);
2467 :
2468 : /* Check whether there's a replication slot requiring an older xmin. */
2469 : def_vis_xid_data =
2470 982122 : TransactionIdOlder(def_vis_xid_data, replication_slot_xmin);
2471 :
2472 : /*
2473 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2474 : * if older than this xid.
2475 : */
2476 982122 : def_vis_xid = def_vis_xid_data;
2477 :
2478 : /*
2479 : * Check whether there's a replication slot requiring an older catalog
2480 : * xmin.
2481 : */
2482 : def_vis_xid =
2483 982122 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2484 :
2485 982122 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2486 982122 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2487 :
2488 : /*
2489 : * Check if we can increase upper bound. As a previous
2490 : * GlobalVisUpdate() might have computed more aggressive values, don't
2491 : * overwrite them if so.
2492 : */
2493 : GlobalVisSharedRels.definitely_needed =
2494 982122 : FullTransactionIdNewer(def_vis_fxid,
2495 : GlobalVisSharedRels.definitely_needed);
2496 : GlobalVisCatalogRels.definitely_needed =
2497 982122 : FullTransactionIdNewer(def_vis_fxid,
2498 : GlobalVisCatalogRels.definitely_needed);
2499 : GlobalVisDataRels.definitely_needed =
2500 982122 : FullTransactionIdNewer(def_vis_fxid_data,
2501 : GlobalVisDataRels.definitely_needed);
2502 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2503 982122 : if (TransactionIdIsNormal(myxid))
2504 : GlobalVisTempRels.definitely_needed =
2505 78664 : FullXidRelativeTo(latest_completed, myxid);
2506 : else
2507 : {
2508 903458 : GlobalVisTempRels.definitely_needed = latest_completed;
2509 903458 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2510 : }
2511 :
2512 : /*
2513 : * Check if we know that we can initialize or increase the lower
2514 : * bound. Currently the only cheap way to do so is to use
2515 : * ShmemVariableCache->oldestXid as input.
2516 : *
2517 : * We should definitely be able to do better. We could e.g. put a
2518 : * global lower bound value into ShmemVariableCache.
2519 : */
2520 : GlobalVisSharedRels.maybe_needed =
2521 982122 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2522 : oldestfxid);
2523 : GlobalVisCatalogRels.maybe_needed =
2524 982122 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2525 : oldestfxid);
2526 : GlobalVisDataRels.maybe_needed =
2527 982122 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2528 : oldestfxid);
2529 : /* accurate value known */
2530 982122 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2531 : }
2532 :
2533 982122 : RecentXmin = xmin;
2534 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2535 :
2536 982122 : snapshot->xmin = xmin;
2537 982122 : snapshot->xmax = xmax;
2538 982122 : snapshot->xcnt = count;
2539 982122 : snapshot->subxcnt = subcount;
2540 982122 : snapshot->suboverflowed = suboverflowed;
2541 982122 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2542 :
2543 982122 : snapshot->curcid = GetCurrentCommandId(false);
2544 :
2545 : /*
2546 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2547 : * not copied in persistent memory.
2548 : */
2549 982122 : snapshot->active_count = 0;
2550 982122 : snapshot->regd_count = 0;
2551 982122 : snapshot->copied = false;
2552 :
2553 982122 : GetSnapshotDataInitOldSnapshot(snapshot);
2554 :
2555 982122 : return snapshot;
2556 : }
2557 :
2558 : /*
2559 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2560 : *
2561 : * This is called when installing a snapshot imported from another
2562 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2563 : * check that the source transaction is still running, and we'd better do
2564 : * that atomically with installing the new xmin.
2565 : *
2566 : * Returns true if successful, false if source xact is no longer running.
2567 : */
2568 : bool
2569 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2570 : VirtualTransactionId *sourcevxid)
2571 : {
2572 16 : bool result = false;
2573 16 : ProcArrayStruct *arrayP = procArray;
2574 : int index;
2575 :
2576 : Assert(TransactionIdIsNormal(xmin));
2577 16 : if (!sourcevxid)
2578 0 : return false;
2579 :
2580 : /* Get lock so source xact can't end while we're doing this */
2581 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2582 :
2583 38 : for (index = 0; index < arrayP->numProcs; index++)
2584 : {
2585 38 : int pgprocno = arrayP->pgprocnos[index];
2586 38 : PGPROC *proc = &allProcs[pgprocno];
2587 38 : int statusFlags = ProcGlobal->statusFlags[index];
2588 : TransactionId xid;
2589 :
2590 : /* Ignore procs running LAZY VACUUM */
2591 38 : if (statusFlags & PROC_IN_VACUUM)
2592 0 : continue;
2593 :
2594 : /* We are only interested in the specific virtual transaction. */
2595 38 : if (proc->backendId != sourcevxid->backendId)
2596 22 : continue;
2597 16 : if (proc->lxid != sourcevxid->localTransactionId)
2598 0 : continue;
2599 :
2600 : /*
2601 : * We check the transaction's database ID for paranoia's sake: if it's
2602 : * in another DB then its xmin does not cover us. Caller should have
2603 : * detected this already, so we just treat any funny cases as
2604 : * "transaction not found".
2605 : */
2606 16 : if (proc->databaseId != MyDatabaseId)
2607 0 : continue;
2608 :
2609 : /*
2610 : * Likewise, let's just make real sure its xmin does cover us.
2611 : */
2612 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2613 16 : if (!TransactionIdIsNormal(xid) ||
2614 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2615 0 : continue;
2616 :
2617 : /*
2618 : * We're good. Install the new xmin. As in GetSnapshotData, set
2619 : * TransactionXmin too. (Note that because snapmgr.c called
2620 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2621 : * we don't check that.)
2622 : */
2623 16 : MyProc->xmin = TransactionXmin = xmin;
2624 :
2625 16 : result = true;
2626 16 : break;
2627 : }
2628 :
2629 16 : LWLockRelease(ProcArrayLock);
2630 :
2631 16 : return result;
2632 : }
2633 :
2634 : /*
2635 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2636 : *
2637 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2638 : * PGPROC of the transaction from which we imported the snapshot, rather than
2639 : * an XID.
2640 : *
2641 : * Note that this function also copies statusFlags from the source `proc` in
2642 : * order to avoid the case where MyProc's xmin needs to be skipped for
2643 : * computing xid horizon.
2644 : *
2645 : * Returns true if successful, false if source xact is no longer running.
2646 : */
2647 : bool
2648 1868 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2649 : {
2650 1868 : bool result = false;
2651 : TransactionId xid;
2652 :
2653 : Assert(TransactionIdIsNormal(xmin));
2654 : Assert(proc != NULL);
2655 :
2656 : /*
2657 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2658 : */
2659 1868 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2660 :
2661 : /*
2662 : * Be certain that the referenced PGPROC has an advertised xmin which is
2663 : * no later than the one we're installing, so that the system-wide xmin
2664 : * can't go backwards. Also, make sure it's running in the same database,
2665 : * so that the per-database xmin cannot go backwards.
2666 : */
2667 1868 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2668 1868 : if (proc->databaseId == MyDatabaseId &&
2669 1868 : TransactionIdIsNormal(xid) &&
2670 1868 : TransactionIdPrecedesOrEquals(xid, xmin))
2671 : {
2672 : /* Install xmin */
2673 1868 : MyProc->xmin = TransactionXmin = xmin;
2674 :
2675 : /* Flags being copied must be valid copy-able flags. */
2676 : Assert((proc->statusFlags & (~PROC_COPYABLE_FLAGS)) == 0);
2677 1868 : MyProc->statusFlags = proc->statusFlags;
2678 1868 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2679 :
2680 1868 : result = true;
2681 : }
2682 :
2683 1868 : LWLockRelease(ProcArrayLock);
2684 :
2685 1868 : return result;
2686 : }
2687 :
2688 : /*
2689 : * GetRunningTransactionData -- returns information about running transactions.
2690 : *
2691 : * Similar to GetSnapshotData but returns more information. We include
2692 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2693 : * prepared transactions.
2694 : *
2695 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2696 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2697 : * array until the caller has WAL-logged this snapshot, and releases the
2698 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2699 : * lock is released.
2700 : *
2701 : * The returned data structure is statically allocated; caller should not
2702 : * modify it, and must not assume it is valid past the next call.
2703 : *
2704 : * This is never executed during recovery so there is no need to look at
2705 : * KnownAssignedXids.
2706 : *
2707 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2708 : * may return entries with duplicated TransactionId values coming from
2709 : * transaction finishing to prepare. Nothing is done about duplicated
2710 : * entries here to not hold on ProcArrayLock more than necessary.
2711 : *
2712 : * We don't worry about updating other counters, we want to keep this as
2713 : * simple as possible and leave GetSnapshotData() as the primary code for
2714 : * that bookkeeping.
2715 : *
2716 : * Note that if any transaction has overflowed its cached subtransactions
2717 : * then there is no real need include any subtransactions.
2718 : */
2719 : RunningTransactions
2720 2974 : GetRunningTransactionData(void)
2721 : {
2722 : /* result workspace */
2723 : static RunningTransactionsData CurrentRunningXactsData;
2724 :
2725 2974 : ProcArrayStruct *arrayP = procArray;
2726 2974 : TransactionId *other_xids = ProcGlobal->xids;
2727 2974 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2728 : TransactionId latestCompletedXid;
2729 : TransactionId oldestRunningXid;
2730 : TransactionId *xids;
2731 : int index;
2732 : int count;
2733 : int subcount;
2734 : bool suboverflowed;
2735 :
2736 : Assert(!RecoveryInProgress());
2737 :
2738 : /*
2739 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2740 : * be sufficient. But it seems better to do the malloc while not holding
2741 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2742 : * more subxip storage than is probably needed.
2743 : *
2744 : * Should only be allocated in bgwriter, since only ever executed during
2745 : * checkpoints.
2746 : */
2747 2974 : if (CurrentRunningXacts->xids == NULL)
2748 : {
2749 : /*
2750 : * First call
2751 : */
2752 1056 : CurrentRunningXacts->xids = (TransactionId *)
2753 1056 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2754 1056 : if (CurrentRunningXacts->xids == NULL)
2755 0 : ereport(ERROR,
2756 : (errcode(ERRCODE_OUT_OF_MEMORY),
2757 : errmsg("out of memory")));
2758 : }
2759 :
2760 2974 : xids = CurrentRunningXacts->xids;
2761 :
2762 2974 : count = subcount = 0;
2763 2974 : suboverflowed = false;
2764 :
2765 : /*
2766 : * Ensure that no xids enter or leave the procarray while we obtain
2767 : * snapshot.
2768 : */
2769 2974 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2770 2974 : LWLockAcquire(XidGenLock, LW_SHARED);
2771 :
2772 2974 : latestCompletedXid =
2773 2974 : XidFromFullTransactionId(ShmemVariableCache->latestCompletedXid);
2774 2974 : oldestRunningXid =
2775 2974 : XidFromFullTransactionId(ShmemVariableCache->nextXid);
2776 :
2777 : /*
2778 : * Spin over procArray collecting all xids
2779 : */
2780 8604 : for (index = 0; index < arrayP->numProcs; index++)
2781 : {
2782 : TransactionId xid;
2783 :
2784 : /* Fetch xid just once - see GetNewTransactionId */
2785 5630 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2786 :
2787 : /*
2788 : * We don't need to store transactions that don't have a TransactionId
2789 : * yet because they will not show as running on a standby server.
2790 : */
2791 5630 : if (!TransactionIdIsValid(xid))
2792 3260 : continue;
2793 :
2794 : /*
2795 : * Be careful not to exclude any xids before calculating the values of
2796 : * oldestRunningXid and suboverflowed, since these are used to clean
2797 : * up transaction information held on standbys.
2798 : */
2799 2370 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2800 2362 : oldestRunningXid = xid;
2801 :
2802 2370 : if (ProcGlobal->subxidStates[index].overflowed)
2803 4 : suboverflowed = true;
2804 :
2805 : /*
2806 : * If we wished to exclude xids this would be the right place for it.
2807 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2808 : * but they do during truncation at the end when they get the lock and
2809 : * truncate, so it is not much of a problem to include them if they
2810 : * are seen and it is cleaner to include them.
2811 : */
2812 :
2813 2370 : xids[count++] = xid;
2814 : }
2815 :
2816 : /*
2817 : * Spin over procArray collecting all subxids, but only if there hasn't
2818 : * been a suboverflow.
2819 : */
2820 2974 : if (!suboverflowed)
2821 : {
2822 2970 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2823 :
2824 8592 : for (index = 0; index < arrayP->numProcs; index++)
2825 : {
2826 5622 : int pgprocno = arrayP->pgprocnos[index];
2827 5622 : PGPROC *proc = &allProcs[pgprocno];
2828 : int nsubxids;
2829 :
2830 : /*
2831 : * Save subtransaction XIDs. Other backends can't add or remove
2832 : * entries while we're holding XidGenLock.
2833 : */
2834 5622 : nsubxids = other_subxidstates[index].count;
2835 5622 : if (nsubxids > 0)
2836 : {
2837 : /* barrier not really required, as XidGenLock is held, but ... */
2838 8 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2839 :
2840 8 : memcpy(&xids[count], (void *) proc->subxids.xids,
2841 : nsubxids * sizeof(TransactionId));
2842 8 : count += nsubxids;
2843 8 : subcount += nsubxids;
2844 :
2845 : /*
2846 : * Top-level XID of a transaction is always less than any of
2847 : * its subxids, so we don't need to check if any of the
2848 : * subxids are smaller than oldestRunningXid
2849 : */
2850 : }
2851 : }
2852 : }
2853 :
2854 : /*
2855 : * It's important *not* to include the limits set by slots here because
2856 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2857 : * were to be included here the initial value could never increase because
2858 : * of a circular dependency where slots only increase their limits when
2859 : * running xacts increases oldestRunningXid and running xacts only
2860 : * increases if slots do.
2861 : */
2862 :
2863 2974 : CurrentRunningXacts->xcnt = count - subcount;
2864 2974 : CurrentRunningXacts->subxcnt = subcount;
2865 2974 : CurrentRunningXacts->subxid_overflow = suboverflowed;
2866 2974 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2867 2974 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2868 2974 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2869 :
2870 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2871 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2872 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2873 :
2874 : /* We don't release the locks here, the caller is responsible for that */
2875 :
2876 2974 : return CurrentRunningXacts;
2877 : }
2878 :
2879 : /*
2880 : * GetOldestActiveTransactionId()
2881 : *
2882 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2883 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2884 : * We look at all databases, though there is no need to include WALSender
2885 : * since this has no effect on hot standby conflicts.
2886 : *
2887 : * This is never executed during recovery so there is no need to look at
2888 : * KnownAssignedXids.
2889 : *
2890 : * We don't worry about updating other counters, we want to keep this as
2891 : * simple as possible and leave GetSnapshotData() as the primary code for
2892 : * that bookkeeping.
2893 : */
2894 : TransactionId
2895 2482 : GetOldestActiveTransactionId(void)
2896 : {
2897 2482 : ProcArrayStruct *arrayP = procArray;
2898 2482 : TransactionId *other_xids = ProcGlobal->xids;
2899 : TransactionId oldestRunningXid;
2900 : int index;
2901 :
2902 : Assert(!RecoveryInProgress());
2903 :
2904 : /*
2905 : * Read nextXid, as the upper bound of what's still active.
2906 : *
2907 : * Reading a TransactionId is atomic, but we must grab the lock to make
2908 : * sure that all XIDs < nextXid are already present in the proc array (or
2909 : * have already completed), when we spin over it.
2910 : */
2911 2482 : LWLockAcquire(XidGenLock, LW_SHARED);
2912 2482 : oldestRunningXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2913 2482 : LWLockRelease(XidGenLock);
2914 :
2915 : /*
2916 : * Spin over procArray collecting all xids and subxids.
2917 : */
2918 2482 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2919 6074 : for (index = 0; index < arrayP->numProcs; index++)
2920 : {
2921 : TransactionId xid;
2922 :
2923 : /* Fetch xid just once - see GetNewTransactionId */
2924 3592 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2925 :
2926 3592 : if (!TransactionIdIsNormal(xid))
2927 1258 : continue;
2928 :
2929 2334 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2930 2320 : oldestRunningXid = xid;
2931 :
2932 : /*
2933 : * Top-level XID of a transaction is always less than any of its
2934 : * subxids, so we don't need to check if any of the subxids are
2935 : * smaller than oldestRunningXid
2936 : */
2937 : }
2938 2482 : LWLockRelease(ProcArrayLock);
2939 :
2940 2482 : return oldestRunningXid;
2941 : }
2942 :
2943 : /*
2944 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2945 : *
2946 : * Returns the oldest xid that we can guarantee not to have been affected by
2947 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2948 : * transaction aborted. Note that the value can (and most of the time will) be
2949 : * much more conservative than what really has been affected by vacuum, but we
2950 : * currently don't have better data available.
2951 : *
2952 : * This is useful to initialize the cutoff xid after which a new changeset
2953 : * extraction replication slot can start decoding changes.
2954 : *
2955 : * Must be called with ProcArrayLock held either shared or exclusively,
2956 : * although most callers will want to use exclusive mode since it is expected
2957 : * that the caller will immediately use the xid to peg the xmin horizon.
2958 : */
2959 : TransactionId
2960 434 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2961 : {
2962 434 : ProcArrayStruct *arrayP = procArray;
2963 : TransactionId oldestSafeXid;
2964 : int index;
2965 434 : bool recovery_in_progress = RecoveryInProgress();
2966 :
2967 : Assert(LWLockHeldByMe(ProcArrayLock));
2968 :
2969 : /*
2970 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2971 : * running. If no transaction with xid were running concurrently a new xid
2972 : * could influence the RecentXmin et al.
2973 : *
2974 : * We initialize the computation to nextXid since that's guaranteed to be
2975 : * a safe, albeit pessimal, value.
2976 : */
2977 434 : LWLockAcquire(XidGenLock, LW_SHARED);
2978 434 : oldestSafeXid = XidFromFullTransactionId(ShmemVariableCache->nextXid);
2979 :
2980 : /*
2981 : * If there's already a slot pegging the xmin horizon, we can start with
2982 : * that value, it's guaranteed to be safe since it's computed by this
2983 : * routine initially and has been enforced since. We can always use the
2984 : * slot's general xmin horizon, but the catalog horizon is only usable
2985 : * when only catalog data is going to be looked at.
2986 : */
2987 436 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2988 2 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2989 : oldestSafeXid))
2990 2 : oldestSafeXid = procArray->replication_slot_xmin;
2991 :
2992 434 : if (catalogOnly &&
2993 312 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2994 54 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2995 : oldestSafeXid))
2996 20 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2997 :
2998 : /*
2999 : * If we're not in recovery, we walk over the procarray and collect the
3000 : * lowest xid. Since we're called with ProcArrayLock held and have
3001 : * acquired XidGenLock, no entries can vanish concurrently, since
3002 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
3003 : * with ProcArrayLock held.
3004 : *
3005 : * In recovery we can't lower the safe value besides what we've computed
3006 : * above, so we'll have to wait a bit longer there. We unfortunately can
3007 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
3008 : * machinery can miss values and return an older value than is safe.
3009 : */
3010 434 : if (!recovery_in_progress)
3011 : {
3012 434 : TransactionId *other_xids = ProcGlobal->xids;
3013 :
3014 : /*
3015 : * Spin over procArray collecting min(ProcGlobal->xids[i])
3016 : */
3017 2154 : for (index = 0; index < arrayP->numProcs; index++)
3018 : {
3019 : TransactionId xid;
3020 :
3021 : /* Fetch xid just once - see GetNewTransactionId */
3022 1720 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
3023 :
3024 1720 : if (!TransactionIdIsNormal(xid))
3025 1712 : continue;
3026 :
3027 8 : if (TransactionIdPrecedes(xid, oldestSafeXid))
3028 8 : oldestSafeXid = xid;
3029 : }
3030 : }
3031 :
3032 434 : LWLockRelease(XidGenLock);
3033 :
3034 434 : return oldestSafeXid;
3035 : }
3036 :
3037 : /*
3038 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
3039 : * delaying checkpoint because they have critical actions in progress.
3040 : *
3041 : * Constructs an array of VXIDs of transactions that are currently in commit
3042 : * critical sections, as shown by having delayChkpt set in their PGPROC.
3043 : *
3044 : * Returns a palloc'd array that should be freed by the caller.
3045 : * *nvxids is the number of valid entries.
3046 : *
3047 : * Note that because backends set or clear delayChkpt without holding any lock,
3048 : * the result is somewhat indeterminate, but we don't really care. Even in
3049 : * a multiprocessor with delayed writes to shared memory, it should be certain
3050 : * that setting of delayChkpt will propagate to shared memory when the backend
3051 : * takes a lock, so we cannot fail to see a virtual xact as delayChkpt if
3052 : * it's already inserted its commit record. Whether it takes a little while
3053 : * for clearing of delayChkpt to propagate is unimportant for correctness.
3054 : */
3055 : VirtualTransactionId *
3056 4100 : GetVirtualXIDsDelayingChkpt(int *nvxids)
3057 : {
3058 : VirtualTransactionId *vxids;
3059 4100 : ProcArrayStruct *arrayP = procArray;
3060 4100 : int count = 0;
3061 : int index;
3062 :
3063 : /* allocate what's certainly enough result space */
3064 : vxids = (VirtualTransactionId *)
3065 4100 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3066 :
3067 4100 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3068 :
3069 9036 : for (index = 0; index < arrayP->numProcs; index++)
3070 : {
3071 4936 : int pgprocno = arrayP->pgprocnos[index];
3072 4936 : PGPROC *proc = &allProcs[pgprocno];
3073 :
3074 4936 : if (proc->delayChkpt)
3075 : {
3076 : VirtualTransactionId vxid;
3077 :
3078 4 : GET_VXID_FROM_PGPROC(vxid, *proc);
3079 4 : if (VirtualTransactionIdIsValid(vxid))
3080 4 : vxids[count++] = vxid;
3081 : }
3082 : }
3083 :
3084 4100 : LWLockRelease(ProcArrayLock);
3085 :
3086 4100 : *nvxids = count;
3087 4100 : return vxids;
3088 : }
3089 :
3090 : /*
3091 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3092 : *
3093 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3094 : * of the specified VXIDs are still in critical sections of code.
3095 : *
3096 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3097 : * those numbers should be small enough for it not to be a problem.
3098 : */
3099 : bool
3100 4 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids)
3101 : {
3102 4 : bool result = false;
3103 4 : ProcArrayStruct *arrayP = procArray;
3104 : int index;
3105 :
3106 4 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3107 :
3108 44 : for (index = 0; index < arrayP->numProcs; index++)
3109 : {
3110 40 : int pgprocno = arrayP->pgprocnos[index];
3111 40 : PGPROC *proc = &allProcs[pgprocno];
3112 : VirtualTransactionId vxid;
3113 :
3114 40 : GET_VXID_FROM_PGPROC(vxid, *proc);
3115 :
3116 40 : if (proc->delayChkpt && VirtualTransactionIdIsValid(vxid))
3117 : {
3118 : int i;
3119 :
3120 0 : for (i = 0; i < nvxids; i++)
3121 : {
3122 0 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3123 : {
3124 0 : result = true;
3125 0 : break;
3126 : }
3127 : }
3128 0 : if (result)
3129 0 : break;
3130 : }
3131 : }
3132 :
3133 4 : LWLockRelease(ProcArrayLock);
3134 :
3135 4 : return result;
3136 : }
3137 :
3138 : /*
3139 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3140 : *
3141 : * Returns NULL if not found. Note that it is up to the caller to be
3142 : * sure that the question remains meaningful for long enough for the
3143 : * answer to be used ...
3144 : */
3145 : PGPROC *
3146 7358 : BackendPidGetProc(int pid)
3147 : {
3148 : PGPROC *result;
3149 :
3150 7358 : if (pid == 0) /* never match dummy PGPROCs */
3151 0 : return NULL;
3152 :
3153 7358 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3154 :
3155 7358 : result = BackendPidGetProcWithLock(pid);
3156 :
3157 7358 : LWLockRelease(ProcArrayLock);
3158 :
3159 7358 : return result;
3160 : }
3161 :
3162 : /*
3163 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3164 : *
3165 : * Same as above, except caller must be holding ProcArrayLock. The found
3166 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3167 : */
3168 : PGPROC *
3169 21556 : BackendPidGetProcWithLock(int pid)
3170 : {
3171 21556 : PGPROC *result = NULL;
3172 21556 : ProcArrayStruct *arrayP = procArray;
3173 : int index;
3174 :
3175 21556 : if (pid == 0) /* never match dummy PGPROCs */
3176 0 : return NULL;
3177 :
3178 59560 : for (index = 0; index < arrayP->numProcs; index++)
3179 : {
3180 56654 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3181 :
3182 56654 : if (proc->pid == pid)
3183 : {
3184 18650 : result = proc;
3185 18650 : break;
3186 : }
3187 : }
3188 :
3189 21556 : return result;
3190 : }
3191 :
3192 : /*
3193 : * BackendXidGetPid -- get a backend's pid given its XID
3194 : *
3195 : * Returns 0 if not found or it's a prepared transaction. Note that
3196 : * it is up to the caller to be sure that the question remains
3197 : * meaningful for long enough for the answer to be used ...
3198 : *
3199 : * Only main transaction Ids are considered. This function is mainly
3200 : * useful for determining what backend owns a lock.
3201 : *
3202 : * Beware that not every xact has an XID assigned. However, as long as you
3203 : * only call this using an XID found on disk, you're safe.
3204 : */
3205 : int
3206 0 : BackendXidGetPid(TransactionId xid)
3207 : {
3208 0 : int result = 0;
3209 0 : ProcArrayStruct *arrayP = procArray;
3210 0 : TransactionId *other_xids = ProcGlobal->xids;
3211 : int index;
3212 :
3213 0 : if (xid == InvalidTransactionId) /* never match invalid xid */
3214 0 : return 0;
3215 :
3216 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3217 :
3218 0 : for (index = 0; index < arrayP->numProcs; index++)
3219 : {
3220 0 : int pgprocno = arrayP->pgprocnos[index];
3221 0 : PGPROC *proc = &allProcs[pgprocno];
3222 :
3223 0 : if (other_xids[index] == xid)
3224 : {
3225 0 : result = proc->pid;
3226 0 : break;
3227 : }
3228 : }
3229 :
3230 0 : LWLockRelease(ProcArrayLock);
3231 :
3232 0 : return result;
3233 : }
3234 :
3235 : /*
3236 : * IsBackendPid -- is a given pid a running backend
3237 : *
3238 : * This is not called by the backend, but is called by external modules.
3239 : */
3240 : bool
3241 4 : IsBackendPid(int pid)
3242 : {
3243 4 : return (BackendPidGetProc(pid) != NULL);
3244 : }
3245 :
3246 :
3247 : /*
3248 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3249 : *
3250 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3251 : *
3252 : * The arguments allow filtering the set of VXIDs returned. Our own process
3253 : * is always skipped. In addition:
3254 : * If limitXmin is not InvalidTransactionId, skip processes with
3255 : * xmin > limitXmin.
3256 : * If excludeXmin0 is true, skip processes with xmin = 0.
3257 : * If allDbs is false, skip processes attached to other databases.
3258 : * If excludeVacuum isn't zero, skip processes for which
3259 : * (statusFlags & excludeVacuum) is not zero.
3260 : *
3261 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3262 : * allow skipping backends whose oldest live snapshot is no older than
3263 : * some snapshot we have. Since we examine the procarray with only shared
3264 : * lock, there are race conditions: a backend could set its xmin just after
3265 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3266 : * other backend could have set its xmin *before* we look. We know however
3267 : * that such a backend must have held shared ProcArrayLock overlapping our
3268 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3269 : * any snapshot the other backend is taking concurrently with our scan cannot
3270 : * consider any transactions as still running that we think are committed
3271 : * (since backends must hold ProcArrayLock exclusive to commit).
3272 : */
3273 : VirtualTransactionId *
3274 410 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3275 : bool allDbs, int excludeVacuum,
3276 : int *nvxids)
3277 : {
3278 : VirtualTransactionId *vxids;
3279 410 : ProcArrayStruct *arrayP = procArray;
3280 410 : int count = 0;
3281 : int index;
3282 :
3283 : /* allocate what's certainly enough result space */
3284 : vxids = (VirtualTransactionId *)
3285 410 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3286 :
3287 410 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3288 :
3289 2144 : for (index = 0; index < arrayP->numProcs; index++)
3290 : {
3291 1734 : int pgprocno = arrayP->pgprocnos[index];
3292 1734 : PGPROC *proc = &allProcs[pgprocno];
3293 1734 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3294 :
3295 1734 : if (proc == MyProc)
3296 410 : continue;
3297 :
3298 1324 : if (excludeVacuum & statusFlags)
3299 0 : continue;
3300 :
3301 1324 : if (allDbs || proc->databaseId == MyDatabaseId)
3302 : {
3303 : /* Fetch xmin just once - might change on us */
3304 504 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3305 :
3306 504 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3307 356 : continue;
3308 :
3309 : /*
3310 : * InvalidTransactionId precedes all other XIDs, so a proc that
3311 : * hasn't set xmin yet will not be rejected by this test.
3312 : */
3313 296 : if (!TransactionIdIsValid(limitXmin) ||
3314 148 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3315 : {
3316 : VirtualTransactionId vxid;
3317 :
3318 140 : GET_VXID_FROM_PGPROC(vxid, *proc);
3319 140 : if (VirtualTransactionIdIsValid(vxid))
3320 140 : vxids[count++] = vxid;
3321 : }
3322 : }
3323 : }
3324 :
3325 410 : LWLockRelease(ProcArrayLock);
3326 :
3327 410 : *nvxids = count;
3328 410 : return vxids;
3329 : }
3330 :
3331 : /*
3332 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3333 : *
3334 : * Usage is limited to conflict resolution during recovery on standby servers.
3335 : * limitXmin is supplied as either latestRemovedXid, or InvalidTransactionId
3336 : * in cases where we cannot accurately determine a value for latestRemovedXid.
3337 : *
3338 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3339 : * so we're not worried if they have a snapshot or not, nor does it really
3340 : * matter what type of lock we hold.
3341 : *
3342 : * All callers that are checking xmins always now supply a valid and useful
3343 : * value for limitXmin. The limitXmin is always lower than the lowest
3344 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3345 : * because we only care about cleanup records that are cleaning up tuple
3346 : * versions from committed transactions. In that case they will only occur
3347 : * at the point where the record is less than the lowest running xid. That
3348 : * allows us to say that if any backend takes a snapshot concurrently with
3349 : * us then the conflict assessment made here would never include the snapshot
3350 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3351 : * concurrent snapshots when limitXmin is valid. We might think about adding
3352 : * Assert(limitXmin < lowest(KnownAssignedXids))
3353 : * but that would not be true in the case of FATAL errors lagging in array,
3354 : * but we already know those are bogus anyway, so we skip that test.
3355 : *
3356 : * If dbOid is valid we skip backends attached to other databases.
3357 : *
3358 : * Be careful to *not* pfree the result from this function. We reuse
3359 : * this array sufficiently often that we use malloc for the result.
3360 : */
3361 : VirtualTransactionId *
3362 1156 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3363 : {
3364 : static VirtualTransactionId *vxids;
3365 1156 : ProcArrayStruct *arrayP = procArray;
3366 1156 : int count = 0;
3367 : int index;
3368 :
3369 : /*
3370 : * If first time through, get workspace to remember main XIDs in. We
3371 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3372 : * result space, remembering room for a terminator.
3373 : */
3374 1156 : if (vxids == NULL)
3375 : {
3376 20 : vxids = (VirtualTransactionId *)
3377 20 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3378 20 : if (vxids == NULL)
3379 0 : ereport(ERROR,
3380 : (errcode(ERRCODE_OUT_OF_MEMORY),
3381 : errmsg("out of memory")));
3382 : }
3383 :
3384 1156 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3385 :
3386 1184 : for (index = 0; index < arrayP->numProcs; index++)
3387 : {
3388 28 : int pgprocno = arrayP->pgprocnos[index];
3389 28 : PGPROC *proc = &allProcs[pgprocno];
3390 :
3391 : /* Exclude prepared transactions */
3392 28 : if (proc->pid == 0)
3393 0 : continue;
3394 :
3395 28 : if (!OidIsValid(dbOid) ||
3396 20 : proc->databaseId == dbOid)
3397 : {
3398 : /* Fetch xmin just once - can't change on us, but good coding */
3399 8 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3400 :
3401 : /*
3402 : * We ignore an invalid pxmin because this means that backend has
3403 : * no snapshot currently. We hold a Share lock to avoid contention
3404 : * with users taking snapshots. That is not a problem because the
3405 : * current xmin is always at least one higher than the latest
3406 : * removed xid, so any new snapshot would never conflict with the
3407 : * test here.
3408 : */
3409 8 : if (!TransactionIdIsValid(limitXmin) ||
3410 0 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3411 : {
3412 : VirtualTransactionId vxid;
3413 :
3414 0 : GET_VXID_FROM_PGPROC(vxid, *proc);
3415 0 : if (VirtualTransactionIdIsValid(vxid))
3416 0 : vxids[count++] = vxid;
3417 : }
3418 : }
3419 : }
3420 :
3421 1156 : LWLockRelease(ProcArrayLock);
3422 :
3423 : /* add the terminator */
3424 1156 : vxids[count].backendId = InvalidBackendId;
3425 1156 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3426 :
3427 1156 : return vxids;
3428 : }
3429 :
3430 : /*
3431 : * CancelVirtualTransaction - used in recovery conflict processing
3432 : *
3433 : * Returns pid of the process signaled, or 0 if not found.
3434 : */
3435 : pid_t
3436 0 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3437 : {
3438 0 : return SignalVirtualTransaction(vxid, sigmode, true);
3439 : }
3440 :
3441 : pid_t
3442 0 : SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode,
3443 : bool conflictPending)
3444 : {
3445 0 : ProcArrayStruct *arrayP = procArray;
3446 : int index;
3447 0 : pid_t pid = 0;
3448 :
3449 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3450 :
3451 0 : for (index = 0; index < arrayP->numProcs; index++)
3452 : {
3453 0 : int pgprocno = arrayP->pgprocnos[index];
3454 0 : PGPROC *proc = &allProcs[pgprocno];
3455 : VirtualTransactionId procvxid;
3456 :
3457 0 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3458 :
3459 0 : if (procvxid.backendId == vxid.backendId &&
3460 0 : procvxid.localTransactionId == vxid.localTransactionId)
3461 : {
3462 0 : proc->recoveryConflictPending = conflictPending;
3463 0 : pid = proc->pid;
3464 0 : if (pid != 0)
3465 : {
3466 : /*
3467 : * Kill the pid if it's still here. If not, that's what we
3468 : * wanted so ignore any errors.
3469 : */
3470 0 : (void) SendProcSignal(pid, sigmode, vxid.backendId);
3471 : }
3472 0 : break;
3473 : }
3474 : }
3475 :
3476 0 : LWLockRelease(ProcArrayLock);
3477 :
3478 0 : return pid;
3479 : }
3480 :
3481 : /*
3482 : * MinimumActiveBackends --- count backends (other than myself) that are
3483 : * in active transactions. Return true if the count exceeds the
3484 : * minimum threshold passed. This is used as a heuristic to decide if
3485 : * a pre-XLOG-flush delay is worthwhile during commit.
3486 : *
3487 : * Do not count backends that are blocked waiting for locks, since they are
3488 : * not going to get to run until someone else commits.
3489 : */
3490 : bool
3491 0 : MinimumActiveBackends(int min)
3492 : {
3493 0 : ProcArrayStruct *arrayP = procArray;
3494 0 : int count = 0;
3495 : int index;
3496 :
3497 : /* Quick short-circuit if no minimum is specified */
3498 0 : if (min == 0)
3499 0 : return true;
3500 :
3501 : /*
3502 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3503 : * bogus, but since we are only testing fields for zero or nonzero, it
3504 : * should be OK. The result is only used for heuristic purposes anyway...
3505 : */
3506 0 : for (index = 0; index < arrayP->numProcs; index++)
3507 : {
3508 0 : int pgprocno = arrayP->pgprocnos[index];
3509 0 : PGPROC *proc = &allProcs[pgprocno];
3510 :
3511 : /*
3512 : * Since we're not holding a lock, need to be prepared to deal with
3513 : * garbage, as someone could have incremented numProcs but not yet
3514 : * filled the structure.
3515 : *
3516 : * If someone just decremented numProcs, 'proc' could also point to a
3517 : * PGPROC entry that's no longer in the array. It still points to a
3518 : * PGPROC struct, though, because freed PGPROC entries just go to the
3519 : * free list and are recycled. Its contents are nonsense in that case,
3520 : * but that's acceptable for this function.
3521 : */
3522 0 : if (pgprocno == -1)
3523 0 : continue; /* do not count deleted entries */
3524 0 : if (proc == MyProc)
3525 0 : continue; /* do not count myself */
3526 0 : if (proc->xid == InvalidTransactionId)
3527 0 : continue; /* do not count if no XID assigned */
3528 0 : if (proc->pid == 0)
3529 0 : continue; /* do not count prepared xacts */
3530 0 : if (proc->waitLock != NULL)
3531 0 : continue; /* do not count if blocked on a lock */
3532 0 : count++;
3533 0 : if (count >= min)
3534 0 : break;
3535 : }
3536 :
3537 0 : return count >= min;
3538 : }
3539 :
3540 : /*
3541 : * CountDBBackends --- count backends that are using specified database
3542 : */
3543 : int
3544 2 : CountDBBackends(Oid databaseid)
3545 : {
3546 2 : ProcArrayStruct *arrayP = procArray;
3547 2 : int count = 0;
3548 : int index;
3549 :
3550 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3551 :
3552 2 : for (index = 0; index < arrayP->numProcs; index++)
3553 : {
3554 0 : int pgprocno = arrayP->pgprocnos[index];
3555 0 : PGPROC *proc = &allProcs[pgprocno];
3556 :
3557 0 : if (proc->pid == 0)
3558 0 : continue; /* do not count prepared xacts */
3559 0 : if (!OidIsValid(databaseid) ||
3560 0 : proc->databaseId == databaseid)
3561 0 : count++;
3562 : }
3563 :
3564 2 : LWLockRelease(ProcArrayLock);
3565 :
3566 2 : return count;
3567 : }
3568 :
3569 : /*
3570 : * CountDBConnections --- counts database backends ignoring any background
3571 : * worker processes
3572 : */
3573 : int
3574 0 : CountDBConnections(Oid databaseid)
3575 : {
3576 0 : ProcArrayStruct *arrayP = procArray;
3577 0 : int count = 0;
3578 : int index;
3579 :
3580 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3581 :
3582 0 : for (index = 0; index < arrayP->numProcs; index++)
3583 : {
3584 0 : int pgprocno = arrayP->pgprocnos[index];
3585 0 : PGPROC *proc = &allProcs[pgprocno];
3586 :
3587 0 : if (proc->pid == 0)
3588 0 : continue; /* do not count prepared xacts */
3589 0 : if (proc->isBackgroundWorker)
3590 0 : continue; /* do not count background workers */
3591 0 : if (!OidIsValid(databaseid) ||
3592 0 : proc->databaseId == databaseid)
3593 0 : count++;
3594 : }
3595 :
3596 0 : LWLockRelease(ProcArrayLock);
3597 :
3598 0 : return count;
3599 : }
3600 :
3601 : /*
3602 : * CancelDBBackends --- cancel backends that are using specified database
3603 : */
3604 : void
3605 0 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3606 : {
3607 0 : ProcArrayStruct *arrayP = procArray;
3608 : int index;
3609 :
3610 : /* tell all backends to die */
3611 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3612 :
3613 0 : for (index = 0; index < arrayP->numProcs; index++)
3614 : {
3615 0 : int pgprocno = arrayP->pgprocnos[index];
3616 0 : PGPROC *proc = &allProcs[pgprocno];
3617 :
3618 0 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3619 : {
3620 : VirtualTransactionId procvxid;
3621 : pid_t pid;
3622 :
3623 0 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3624 :
3625 0 : proc->recoveryConflictPending = conflictPending;
3626 0 : pid = proc->pid;
3627 0 : if (pid != 0)
3628 : {
3629 : /*
3630 : * Kill the pid if it's still here. If not, that's what we
3631 : * wanted so ignore any errors.
3632 : */
3633 0 : (void) SendProcSignal(pid, sigmode, procvxid.backendId);
3634 : }
3635 : }
3636 : }
3637 :
3638 0 : LWLockRelease(ProcArrayLock);
3639 0 : }
3640 :
3641 : /*
3642 : * CountUserBackends --- count backends that are used by specified user
3643 : */
3644 : int
3645 0 : CountUserBackends(Oid roleid)
3646 : {
3647 0 : ProcArrayStruct *arrayP = procArray;
3648 0 : int count = 0;
3649 : int index;
3650 :
3651 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3652 :
3653 0 : for (index = 0; index < arrayP->numProcs; index++)
3654 : {
3655 0 : int pgprocno = arrayP->pgprocnos[index];
3656 0 : PGPROC *proc = &allProcs[pgprocno];
3657 :
3658 0 : if (proc->pid == 0)
3659 0 : continue; /* do not count prepared xacts */
3660 0 : if (proc->isBackgroundWorker)
3661 0 : continue; /* do not count background workers */
3662 0 : if (proc->roleId == roleid)
3663 0 : count++;
3664 : }
3665 :
3666 0 : LWLockRelease(ProcArrayLock);
3667 :
3668 0 : return count;
3669 : }
3670 :
3671 : /*
3672 : * CountOtherDBBackends -- check for other backends running in the given DB
3673 : *
3674 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3675 : * for them to exit. Autovacuum backends are encouraged to exit early by
3676 : * sending them SIGTERM, but normal user backends are just waited for.
3677 : *
3678 : * The current backend is always ignored; it is caller's responsibility to
3679 : * check whether the current backend uses the given DB, if it's important.
3680 : *
3681 : * Returns true if there are (still) other backends in the DB, false if not.
3682 : * Also, *nbackends and *nprepared are set to the number of other backends
3683 : * and prepared transactions in the DB, respectively.
3684 : *
3685 : * This function is used to interlock DROP DATABASE and related commands
3686 : * against there being any active backends in the target DB --- dropping the
3687 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3688 : * detect here the possibility of a newly-started backend that is trying to
3689 : * connect to the doomed database, so additional interlocking is needed during
3690 : * backend startup. The caller should normally hold an exclusive lock on the
3691 : * target DB before calling this, which is one reason we mustn't wait
3692 : * indefinitely.
3693 : */
3694 : bool
3695 1274 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3696 : {
3697 1274 : ProcArrayStruct *arrayP = procArray;
3698 :
3699 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3700 : int autovac_pids[MAXAUTOVACPIDS];
3701 : int tries;
3702 :
3703 : /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3704 1274 : for (tries = 0; tries < 50; tries++)
3705 : {
3706 1274 : int nautovacs = 0;
3707 1274 : bool found = false;
3708 : int index;
3709 :
3710 1274 : CHECK_FOR_INTERRUPTS();
3711 :
3712 1274 : *nbackends = *nprepared = 0;
3713 :
3714 1274 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3715 :
3716 3150 : for (index = 0; index < arrayP->numProcs; index++)
3717 : {
3718 1876 : int pgprocno = arrayP->pgprocnos[index];
3719 1876 : PGPROC *proc = &allProcs[pgprocno];
3720 1876 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3721 :
3722 1876 : if (proc->databaseId != databaseId)
3723 912 : continue;
3724 964 : if (proc == MyProc)
3725 964 : continue;
3726 :
3727 0 : found = true;
3728 :
3729 0 : if (proc->pid == 0)
3730 0 : (*nprepared)++;
3731 : else
3732 : {
3733 0 : (*nbackends)++;
3734 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3735 : nautovacs < MAXAUTOVACPIDS)
3736 0 : autovac_pids[nautovacs++] = proc->pid;
3737 : }
3738 : }
3739 :
3740 1274 : LWLockRelease(ProcArrayLock);
3741 :
3742 1274 : if (!found)
3743 1274 : return false; /* no conflicting backends, so done */
3744 :
3745 : /*
3746 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3747 : * postpone this step until after the loop because we don't want to
3748 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3749 : * block kill() inside the kernel...
3750 : */
3751 0 : for (index = 0; index < nautovacs; index++)
3752 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3753 :
3754 : /* sleep, then try again */
3755 0 : pg_usleep(100 * 1000L); /* 100ms */
3756 : }
3757 :
3758 0 : return true; /* timed out, still conflicts */
3759 : }
3760 :
3761 : /*
3762 : * Terminate existing connections to the specified database. This routine
3763 : * is used by the DROP DATABASE command when user has asked to forcefully
3764 : * drop the database.
3765 : *
3766 : * The current backend is always ignored; it is caller's responsibility to
3767 : * check whether the current backend uses the given DB, if it's important.
3768 : *
3769 : * It doesn't allow to terminate the connections even if there is a one
3770 : * backend with the prepared transaction in the target database.
3771 : */
3772 : void
3773 2 : TerminateOtherDBBackends(Oid databaseId)
3774 : {
3775 2 : ProcArrayStruct *arrayP = procArray;
3776 2 : List *pids = NIL;
3777 2 : int nprepared = 0;
3778 : int i;
3779 :
3780 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3781 :
3782 8 : for (i = 0; i < procArray->numProcs; i++)
3783 : {
3784 6 : int pgprocno = arrayP->pgprocnos[i];
3785 6 : PGPROC *proc = &allProcs[pgprocno];
3786 :
3787 6 : if (proc->databaseId != databaseId)
3788 6 : continue;
3789 0 : if (proc == MyProc)
3790 0 : continue;
3791 :
3792 0 : if (proc->pid != 0)
3793 0 : pids = lappend_int(pids, proc->pid);
3794 : else
3795 0 : nprepared++;
3796 : }
3797 :
3798 2 : LWLockRelease(ProcArrayLock);
3799 :
3800 2 : if (nprepared > 0)
3801 0 : ereport(ERROR,
3802 : (errcode(ERRCODE_OBJECT_IN_USE),
3803 : errmsg("database \"%s\" is being used by prepared transactions",
3804 : get_database_name(databaseId)),
3805 : errdetail_plural("There is %d prepared transaction using the database.",
3806 : "There are %d prepared transactions using the database.",
3807 : nprepared,
3808 : nprepared)));
3809 :
3810 2 : if (pids)
3811 : {
3812 : ListCell *lc;
3813 :
3814 : /*
3815 : * Check whether we have the necessary rights to terminate other
3816 : * sessions. We don't terminate any session until we ensure that we
3817 : * have rights on all the sessions to be terminated. These checks are
3818 : * the same as we do in pg_terminate_backend.
3819 : *
3820 : * In this case we don't raise some warnings - like "PID %d is not a
3821 : * PostgreSQL server process", because for us already finished session
3822 : * is not a problem.
3823 : */
3824 0 : foreach(lc, pids)
3825 : {
3826 0 : int pid = lfirst_int(lc);
3827 0 : PGPROC *proc = BackendPidGetProc(pid);
3828 :
3829 0 : if (proc != NULL)
3830 : {
3831 : /* Only allow superusers to signal superuser-owned backends. */
3832 0 : if (superuser_arg(proc->roleId) && !superuser())
3833 0 : ereport(ERROR,
3834 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3835 : errmsg("must be a superuser to terminate superuser process")));
3836 :
3837 : /* Users can signal backends they have role membership in. */
3838 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3839 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3840 0 : ereport(ERROR,
3841 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3842 : errmsg("must be a member of the role whose process is being terminated or member of pg_signal_backend")));
3843 : }
3844 : }
3845 :
3846 : /*
3847 : * There's a race condition here: once we release the ProcArrayLock,
3848 : * it's possible for the session to exit before we issue kill. That
3849 : * race condition possibility seems too unlikely to worry about. See
3850 : * pg_signal_backend.
3851 : */
3852 0 : foreach(lc, pids)
3853 : {
3854 0 : int pid = lfirst_int(lc);
3855 0 : PGPROC *proc = BackendPidGetProc(pid);
3856 :
3857 0 : if (proc != NULL)
3858 : {
3859 : /*
3860 : * If we have setsid(), signal the backend's whole process
3861 : * group
3862 : */
3863 : #ifdef HAVE_SETSID
3864 0 : (void) kill(-pid, SIGTERM);
3865 : #else
3866 : (void) kill(pid, SIGTERM);
3867 : #endif
3868 : }
3869 : }
3870 : }
3871 2 : }
3872 :
3873 : /*
3874 : * ProcArraySetReplicationSlotXmin
3875 : *
3876 : * Install limits to future computations of the xmin horizon to prevent vacuum
3877 : * and HOT pruning from removing affected rows still needed by clients with
3878 : * replication slots.
3879 : */
3880 : void
3881 3130 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3882 : bool already_locked)
3883 : {
3884 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3885 :
3886 3130 : if (!already_locked)
3887 2696 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3888 :
3889 3130 : procArray->replication_slot_xmin = xmin;
3890 3130 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3891 :
3892 3130 : if (!already_locked)
3893 2696 : LWLockRelease(ProcArrayLock);
3894 3130 : }
3895 :
3896 : /*
3897 : * ProcArrayGetReplicationSlotXmin
3898 : *
3899 : * Return the current slot xmin limits. That's useful to be able to remove
3900 : * data that's older than those limits.
3901 : */
3902 : void
3903 40 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3904 : TransactionId *catalog_xmin)
3905 : {
3906 40 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3907 :
3908 40 : if (xmin != NULL)
3909 0 : *xmin = procArray->replication_slot_xmin;
3910 :
3911 40 : if (catalog_xmin != NULL)
3912 40 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3913 :
3914 40 : LWLockRelease(ProcArrayLock);
3915 40 : }
3916 :
3917 : /*
3918 : * XidCacheRemoveRunningXids
3919 : *
3920 : * Remove a bunch of TransactionIds from the list of known-running
3921 : * subtransactions for my backend. Both the specified xid and those in
3922 : * the xids[] array (of length nxids) are removed from the subxids cache.
3923 : * latestXid must be the latest XID among the group.
3924 : */
3925 : void
3926 824 : XidCacheRemoveRunningXids(TransactionId xid,
3927 : int nxids, const TransactionId *xids,
3928 : TransactionId latestXid)
3929 : {
3930 : int i,
3931 : j;
3932 : XidCacheStatus *mysubxidstat;
3933 :
3934 : Assert(TransactionIdIsValid(xid));
3935 :
3936 : /*
3937 : * We must hold ProcArrayLock exclusively in order to remove transactions
3938 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
3939 : * possible this could be relaxed since we know this routine is only used
3940 : * to abort subtransactions, but pending closer analysis we'd best be
3941 : * conservative.
3942 : *
3943 : * Note that we do not have to be careful about memory ordering of our own
3944 : * reads wrt. GetNewTransactionId() here - only this process can modify
3945 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3946 : * careful about our own writes being well ordered.
3947 : */
3948 824 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3949 :
3950 824 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
3951 :
3952 : /*
3953 : * Under normal circumstances xid and xids[] will be in increasing order,
3954 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
3955 : * behavior when removing a lot of xids.
3956 : */
3957 878 : for (i = nxids - 1; i >= 0; i--)
3958 : {
3959 54 : TransactionId anxid = xids[i];
3960 :
3961 54 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
3962 : {
3963 54 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
3964 : {
3965 54 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
3966 54 : pg_write_barrier();
3967 54 : mysubxidstat->count--;
3968 54 : MyProc->subxidStatus.count--;
3969 54 : break;
3970 : }
3971 : }
3972 :
3973 : /*
3974 : * Ordinarily we should have found it, unless the cache has
3975 : * overflowed. However it's also possible for this routine to be
3976 : * invoked multiple times for the same subtransaction, in case of an
3977 : * error during AbortSubTransaction. So instead of Assert, emit a
3978 : * debug warning.
3979 : */
3980 54 : if (j < 0 && !MyProc->subxidStatus.overflowed)
3981 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
3982 : }
3983 :
3984 824 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
3985 : {
3986 824 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
3987 : {
3988 824 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
3989 824 : pg_write_barrier();
3990 824 : mysubxidstat->count--;
3991 824 : MyProc->subxidStatus.count--;
3992 824 : break;
3993 : }
3994 : }
3995 : /* Ordinarily we should have found it, unless the cache has overflowed */
3996 824 : if (j < 0 && !MyProc->subxidStatus.overflowed)
3997 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
3998 :
3999 : /* Also advance global latestCompletedXid while holding the lock */
4000 824 : MaintainLatestCompletedXid(latestXid);
4001 :
4002 : /* ... and xactCompletionCount */
4003 824 : ShmemVariableCache->xactCompletionCount++;
4004 :
4005 824 : LWLockRelease(ProcArrayLock);
4006 824 : }
4007 :
4008 : #ifdef XIDCACHE_DEBUG
4009 :
4010 : /*
4011 : * Print stats about effectiveness of XID cache
4012 : */
4013 : static void
4014 : DisplayXidCache(void)
4015 : {
4016 : fprintf(stderr,
4017 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4018 : xc_by_recent_xmin,
4019 : xc_by_known_xact,
4020 : xc_by_my_xact,
4021 : xc_by_latest_xid,
4022 : xc_by_main_xid,
4023 : xc_by_child_xid,
4024 : xc_by_known_assigned,
4025 : xc_no_overflow,
4026 : xc_slow_answer);
4027 : }
4028 : #endif /* XIDCACHE_DEBUG */
4029 :
4030 : /*
4031 : * If rel != NULL, return test state appropriate for relation, otherwise
4032 : * return state usable for all relations. The latter may consider XIDs as
4033 : * not-yet-visible-to-everyone that a state for a specific relation would
4034 : * already consider visible-to-everyone.
4035 : *
4036 : * This needs to be called while a snapshot is active or registered, otherwise
4037 : * there are wraparound and other dangers.
4038 : *
4039 : * See comment for GlobalVisState for details.
4040 : */
4041 : GlobalVisState *
4042 19142976 : GlobalVisTestFor(Relation rel)
4043 : {
4044 19142976 : GlobalVisState *state = NULL;
4045 :
4046 : /* XXX: we should assert that a snapshot is pushed or registered */
4047 : Assert(RecentXmin);
4048 :
4049 19142976 : switch (GlobalVisHorizonKindForRel(rel))
4050 : {
4051 162126 : case VISHORIZON_SHARED:
4052 162126 : state = &GlobalVisSharedRels;
4053 162126 : break;
4054 3757228 : case VISHORIZON_CATALOG:
4055 3757228 : state = &GlobalVisCatalogRels;
4056 3757228 : break;
4057 15161060 : case VISHORIZON_DATA:
4058 15161060 : state = &GlobalVisDataRels;
4059 15161060 : break;
4060 62562 : case VISHORIZON_TEMP:
4061 62562 : state = &GlobalVisTempRels;
4062 62562 : break;
4063 : }
4064 :
4065 19142976 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4066 : FullTransactionIdIsValid(state->maybe_needed));
4067 :
4068 19142976 : return state;
4069 : }
4070 :
4071 : /*
4072 : * Return true if it's worth updating the accurate maybe_needed boundary.
4073 : *
4074 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4075 : * repeatedly do so when there is a low likelihood of it being beneficial.
4076 : *
4077 : * The current heuristic is that we update only if RecentXmin has changed
4078 : * since the last update. If the oldest currently running transaction has not
4079 : * finished, it is unlikely that recomputing the horizon would be useful.
4080 : */
4081 : static bool
4082 586322 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4083 : {
4084 : /* hasn't been updated yet */
4085 586322 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4086 9488 : return true;
4087 :
4088 : /*
4089 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4090 : * unlikely to be beneficial to refresh boundaries.
4091 : */
4092 576834 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4093 : state->definitely_needed))
4094 72 : return false;
4095 :
4096 : /* does the last snapshot built have a different xmin? */
4097 576762 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4098 : }
4099 :
4100 : static void
4101 314222 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4102 : {
4103 : GlobalVisSharedRels.maybe_needed =
4104 314222 : FullXidRelativeTo(horizons->latest_completed,
4105 : horizons->shared_oldest_nonremovable);
4106 : GlobalVisCatalogRels.maybe_needed =
4107 314222 : FullXidRelativeTo(horizons->latest_completed,
4108 : horizons->catalog_oldest_nonremovable);
4109 : GlobalVisDataRels.maybe_needed =
4110 314222 : FullXidRelativeTo(horizons->latest_completed,
4111 : horizons->data_oldest_nonremovable);
4112 : GlobalVisTempRels.maybe_needed =
4113 314222 : FullXidRelativeTo(horizons->latest_completed,
4114 : horizons->temp_oldest_nonremovable);
4115 :
4116 : /*
4117 : * In longer running transactions it's possible that transactions we
4118 : * previously needed to treat as running aren't around anymore. So update
4119 : * definitely_needed to not be earlier than maybe_needed.
4120 : */
4121 : GlobalVisSharedRels.definitely_needed =
4122 314222 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4123 : GlobalVisSharedRels.definitely_needed);
4124 : GlobalVisCatalogRels.definitely_needed =
4125 314222 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4126 : GlobalVisCatalogRels.definitely_needed);
4127 : GlobalVisDataRels.definitely_needed =
4128 314222 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4129 : GlobalVisDataRels.definitely_needed);
4130 314222 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4131 :
4132 314222 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4133 314222 : }
4134 :
4135 : /*
4136 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4137 : * using ComputeXidHorizons().
4138 : */
4139 : static void
4140 184042 : GlobalVisUpdate(void)
4141 : {
4142 : ComputeXidHorizonsResult horizons;
4143 :
4144 : /* updates the horizons as a side-effect */
4145 184042 : ComputeXidHorizons(&horizons);
4146 184042 : }
4147 :
4148 : /*
4149 : * Return true if no snapshot still considers fxid to be running.
4150 : *
4151 : * The state passed needs to have been initialized for the relation fxid is
4152 : * from (NULL is also OK), otherwise the result may not be correct.
4153 : *
4154 : * See comment for GlobalVisState for details.
4155 : */
4156 : bool
4157 13715552 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4158 : FullTransactionId fxid)
4159 : {
4160 : /*
4161 : * If fxid is older than maybe_needed bound, it definitely is visible to
4162 : * everyone.
4163 : */
4164 13715552 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4165 4079810 : return true;
4166 :
4167 : /*
4168 : * If fxid is >= definitely_needed bound, it is very likely to still be
4169 : * considered running.
4170 : */
4171 9635742 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4172 9049512 : return false;
4173 :
4174 : /*
4175 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4176 : * might not exist a snapshot considering fxid running. If it makes sense,
4177 : * update boundaries and recheck.
4178 : */
4179 586230 : if (GlobalVisTestShouldUpdate(state))
4180 : {
4181 184022 : GlobalVisUpdate();
4182 :
4183 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4184 :
4185 184022 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4186 : }
4187 : else
4188 402208 : return false;
4189 : }
4190 :
4191 : /*
4192 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4193 : *
4194 : * It is crucial that this only gets called for xids from a source that
4195 : * protects against xid wraparounds (e.g. from a table and thus protected by
4196 : * relfrozenxid).
4197 : */
4198 : bool
4199 13713882 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4200 : {
4201 : FullTransactionId fxid;
4202 :
4203 : /*
4204 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4205 : * because we know the xid has to, at the very least, be between
4206 : * [oldestXid, nextFullXid), i.e. within 2 billion of xid. To avoid taking
4207 : * a lock to determine either, we can just compare with
4208 : * state->definitely_needed, which was based on those value at the time
4209 : * the current snapshot was built.
4210 : */
4211 13713882 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4212 :
4213 13713882 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4214 : }
4215 :
4216 : /*
4217 : * Return FullTransactionId below which all transactions are not considered
4218 : * running anymore.
4219 : *
4220 : * Note: This is less efficient than testing with
4221 : * GlobalVisTestIsRemovableFullXid as it likely requires building an accurate
4222 : * cutoff, even in the case all the XIDs compared with the cutoff are outside
4223 : * [maybe_needed, definitely_needed).
4224 : */
4225 : FullTransactionId
4226 92 : GlobalVisTestNonRemovableFullHorizon(GlobalVisState *state)
4227 : {
4228 : /* acquire accurate horizon if not already done */
4229 92 : if (GlobalVisTestShouldUpdate(state))
4230 20 : GlobalVisUpdate();
4231 :
4232 92 : return state->maybe_needed;
4233 : }
4234 :
4235 : /* Convenience wrapper around GlobalVisTestNonRemovableFullHorizon */
4236 : TransactionId
4237 92 : GlobalVisTestNonRemovableHorizon(GlobalVisState *state)
4238 : {
4239 : FullTransactionId cutoff;
4240 :
4241 92 : cutoff = GlobalVisTestNonRemovableFullHorizon(state);
4242 :
4243 92 : return XidFromFullTransactionId(cutoff);
4244 : }
4245 :
4246 : /*
4247 : * Convenience wrapper around GlobalVisTestFor() and
4248 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4249 : */
4250 : bool
4251 1670 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4252 : {
4253 : GlobalVisState *state;
4254 :
4255 1670 : state = GlobalVisTestFor(rel);
4256 :
4257 1670 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4258 : }
4259 :
4260 : /*
4261 : * Convenience wrapper around GlobalVisTestFor() and
4262 : * GlobalVisTestIsRemovableXid(), see their comments.
4263 : */
4264 : bool
4265 8 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4266 : {
4267 : GlobalVisState *state;
4268 :
4269 8 : state = GlobalVisTestFor(rel);
4270 :
4271 8 : return GlobalVisTestIsRemovableXid(state, xid);
4272 : }
4273 :
4274 : /*
4275 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4276 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4277 : *
4278 : * Be very careful about when to use this function. It can only safely be used
4279 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4280 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4281 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4282 : * the xid has to be within that range), or if xid is from the procarray and
4283 : * prevents xid wraparound that way.
4284 : */
4285 : static inline FullTransactionId
4286 18443440 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4287 : {
4288 18443440 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4289 :
4290 : Assert(TransactionIdIsValid(xid));
4291 : Assert(TransactionIdIsValid(rel_xid));
4292 :
4293 : /* not guaranteed to find issues, but likely to catch mistakes */
4294 : AssertTransactionIdInAllowableRange(xid);
4295 :
4296 36886880 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4297 18443440 : + (int32) (xid - rel_xid));
4298 : }
4299 :
4300 :
4301 : /* ----------------------------------------------
4302 : * KnownAssignedTransactionIds sub-module
4303 : * ----------------------------------------------
4304 : */
4305 :
4306 : /*
4307 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4308 : * running on the primary at the current point in WAL. These XIDs must be
4309 : * treated as running by standby transactions, even though they are not in
4310 : * the standby server's PGPROC array.
4311 : *
4312 : * We record all XIDs that we know have been assigned. That includes all the
4313 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4314 : * been assigned. We can deduce the existence of unobserved XIDs because we
4315 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4316 : * list expands as new XIDs are observed or inferred, and contracts when
4317 : * transaction completion records arrive.
4318 : *
4319 : * During hot standby we do not fret too much about the distinction between
4320 : * top-level XIDs and subtransaction XIDs. We store both together in the
4321 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4322 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4323 : * doesn't care about the distinction either. Subtransaction XIDs are
4324 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4325 : * links are *not* maintained (which does not affect visibility).
4326 : *
4327 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4328 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4329 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4330 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4331 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4332 : * and then remove them from KnownAssignedXids. This prevents overflow of
4333 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4334 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4335 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4336 : * though it should be complete for top-level XIDs; this is the same situation
4337 : * that holds with respect to the PGPROC entries in normal running.
4338 : *
4339 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4340 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4341 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4342 : * As long as that is within the range of interesting XIDs, we have to assume
4343 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4344 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4345 : * subXID arrives - that is not an error.)
4346 : *
4347 : * Should a backend on primary somehow disappear before it can write an abort
4348 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4349 : * aborted but we think they were running; the distinction is irrelevant
4350 : * because either way any changes done by the transaction are not visible to
4351 : * backends in the standby. We prune KnownAssignedXids when
4352 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4353 : * array due to such dead XIDs.
4354 : */
4355 :
4356 : /*
4357 : * RecordKnownAssignedTransactionIds
4358 : * Record the given XID in KnownAssignedXids, as well as any preceding
4359 : * unobserved XIDs.
4360 : *
4361 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4362 : * associated with a transaction. Must be called for each record after we
4363 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4364 : *
4365 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4366 : */
4367 : void
4368 224550 : RecordKnownAssignedTransactionIds(TransactionId xid)
4369 : {
4370 : Assert(standbyState >= STANDBY_INITIALIZED);
4371 : Assert(TransactionIdIsValid(xid));
4372 : Assert(TransactionIdIsValid(latestObservedXid));
4373 :
4374 224550 : elog(trace_recovery(DEBUG4), "record known xact %u latestObservedXid %u",
4375 : xid, latestObservedXid);
4376 :
4377 : /*
4378 : * When a newly observed xid arrives, it is frequently the case that it is
4379 : * *not* the next xid in sequence. When this occurs, we must treat the
4380 : * intervening xids as running also.
4381 : */
4382 224550 : if (TransactionIdFollows(xid, latestObservedXid))
4383 : {
4384 : TransactionId next_expected_xid;
4385 :
4386 : /*
4387 : * Extend subtrans like we do in GetNewTransactionId() during normal
4388 : * operation using individual extend steps. Note that we do not need
4389 : * to extend clog since its extensions are WAL logged.
4390 : *
4391 : * This part has to be done regardless of standbyState since we
4392 : * immediately start assigning subtransactions to their toplevel
4393 : * transactions.
4394 : */
4395 3276 : next_expected_xid = latestObservedXid;
4396 6560 : while (TransactionIdPrecedes(next_expected_xid, xid))
4397 : {
4398 3284 : TransactionIdAdvance(next_expected_xid);
4399 3284 : ExtendSUBTRANS(next_expected_xid);
4400 : }
4401 : Assert(next_expected_xid == xid);
4402 :
4403 : /*
4404 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4405 : * more to do since we don't track assigned xids yet.
4406 : */
4407 3276 : if (standbyState <= STANDBY_INITIALIZED)
4408 : {
4409 8 : latestObservedXid = xid;
4410 8 : return;
4411 : }
4412 :
4413 : /*
4414 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4415 : */
4416 3268 : next_expected_xid = latestObservedXid;
4417 3268 : TransactionIdAdvance(next_expected_xid);
4418 3268 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4419 :
4420 : /*
4421 : * Now we can advance latestObservedXid
4422 : */
4423 3268 : latestObservedXid = xid;
4424 :
4425 : /* ShmemVariableCache->nextXid must be beyond any observed xid */
4426 3268 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4427 : }
4428 : }
4429 :
4430 : /*
4431 : * ExpireTreeKnownAssignedTransactionIds
4432 : * Remove the given XIDs from KnownAssignedXids.
4433 : *
4434 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4435 : */
4436 : void
4437 320 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4438 : TransactionId *subxids, TransactionId max_xid)
4439 : {
4440 : Assert(standbyState >= STANDBY_INITIALIZED);
4441 :
4442 : /*
4443 : * Uses same locking as transaction commit
4444 : */
4445 320 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4446 :
4447 320 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4448 :
4449 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4450 320 : MaintainLatestCompletedXidRecovery(max_xid);
4451 :
4452 : /* ... and xactCompletionCount */
4453 320 : ShmemVariableCache->xactCompletionCount++;
4454 :
4455 320 : LWLockRelease(ProcArrayLock);
4456 320 : }
4457 :
4458 : /*
4459 : * ExpireAllKnownAssignedTransactionIds
4460 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4461 : */
4462 : void
4463 114 : ExpireAllKnownAssignedTransactionIds(void)
4464 : {
4465 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4466 114 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4467 :
4468 : /*
4469 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4470 : * the call of this function. But do this for unification with what
4471 : * ExpireOldKnownAssignedTransactionIds() do.
4472 : */
4473 114 : procArray->lastOverflowedXid = InvalidTransactionId;
4474 114 : LWLockRelease(ProcArrayLock);
4475 114 : }
4476 :
4477 : /*
4478 : * ExpireOldKnownAssignedTransactionIds
4479 : * Remove KnownAssignedXids entries preceding the given XID and
4480 : * potentially reset lastOverflowedXid.
4481 : */
4482 : void
4483 254 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4484 : {
4485 254 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4486 :
4487 : /*
4488 : * Reset lastOverflowedXid if we know all transactions that have been
4489 : * possibly running are being gone. Not doing so could cause an incorrect
4490 : * lastOverflowedXid value, which makes extra snapshots be marked as
4491 : * suboverflowed.
4492 : */
4493 254 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4494 246 : procArray->lastOverflowedXid = InvalidTransactionId;
4495 254 : KnownAssignedXidsRemovePreceding(xid);
4496 254 : LWLockRelease(ProcArrayLock);
4497 254 : }
4498 :
4499 :
4500 : /*
4501 : * Private module functions to manipulate KnownAssignedXids
4502 : *
4503 : * There are 5 main uses of the KnownAssignedXids data structure:
4504 : *
4505 : * * backends taking snapshots - all valid XIDs need to be copied out
4506 : * * backends seeking to determine presence of a specific XID
4507 : * * startup process adding new known-assigned XIDs
4508 : * * startup process removing specific XIDs as transactions end
4509 : * * startup process pruning array when special WAL records arrive
4510 : *
4511 : * This data structure is known to be a hot spot during Hot Standby, so we
4512 : * go to some lengths to make these operations as efficient and as concurrent
4513 : * as possible.
4514 : *
4515 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4516 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4517 : * in general TransactionIdPrecedes would not provide a total order, but
4518 : * we know that the entries present at any instant should not extend across
4519 : * a large enough fraction of XID space to wrap around (the primary would
4520 : * shut down for fear of XID wrap long before that happens). So it's OK to
4521 : * use TransactionIdPrecedes as a binary-search comparator.
4522 : *
4523 : * It's cheap to maintain the sortedness during insertions, since new known
4524 : * XIDs are always reported in XID order; we just append them at the right.
4525 : *
4526 : * To keep individual deletions cheap, we need to allow gaps in the array.
4527 : * This is implemented by marking array elements as valid or invalid using
4528 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4529 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4530 : * XID entry itself. This preserves the property that the XID entries are
4531 : * sorted, so we can do binary searches easily. Periodically we compress
4532 : * out the unused entries; that's much cheaper than having to compress the
4533 : * array immediately on every deletion.
4534 : *
4535 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4536 : * are those with indexes tail <= i < head; items outside this subscript range
4537 : * have unspecified contents. When head reaches the end of the array, we
4538 : * force compression of unused entries rather than wrapping around, since
4539 : * allowing wraparound would greatly complicate the search logic. We maintain
4540 : * an explicit tail pointer so that pruning of old XIDs can be done without
4541 : * immediately moving the array contents. In most cases only a small fraction
4542 : * of the array contains valid entries at any instant.
4543 : *
4544 : * Although only the startup process can ever change the KnownAssignedXids
4545 : * data structure, we still need interlocking so that standby backends will
4546 : * not observe invalid intermediate states. The convention is that backends
4547 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4548 : * the array, the startup process must hold ProcArrayLock exclusively, for
4549 : * the usual transactional reasons (compare commit/abort of a transaction
4550 : * during normal running). Compressing unused entries out of the array
4551 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4552 : * them into slots to the right of the head pointer and then advance the head
4553 : * pointer. This wouldn't require any lock at all, except that on machines
4554 : * with weak memory ordering we need to be careful that other processors
4555 : * see the array element changes before they see the head pointer change.
4556 : * We handle this by using a spinlock to protect reads and writes of the
4557 : * head/tail pointers. (We could dispense with the spinlock if we were to
4558 : * create suitable memory access barrier primitives and use those instead.)
4559 : * The spinlock must be taken to read or write the head/tail pointers unless
4560 : * the caller holds ProcArrayLock exclusively.
4561 : *
4562 : * Algorithmic analysis:
4563 : *
4564 : * If we have a maximum of M slots, with N XIDs currently spread across
4565 : * S elements then we have N <= S <= M always.
4566 : *
4567 : * * Adding a new XID is O(1) and needs little locking (unless compression
4568 : * must happen)
4569 : * * Compressing the array is O(S) and requires exclusive lock
4570 : * * Removing an XID is O(logS) and requires exclusive lock
4571 : * * Taking a snapshot is O(S) and requires shared lock
4572 : * * Checking for an XID is O(logS) and requires shared lock
4573 : *
4574 : * In comparison, using a hash table for KnownAssignedXids would mean that
4575 : * taking snapshots would be O(M). If we can maintain S << M then the
4576 : * sorted array technique will deliver significantly faster snapshots.
4577 : * If we try to keep S too small then we will spend too much time compressing,
4578 : * so there is an optimal point for any workload mix. We use a heuristic to
4579 : * decide when to compress the array, though trimming also helps reduce
4580 : * frequency of compressing. The heuristic requires us to track the number of
4581 : * currently valid XIDs in the array.
4582 : */
4583 :
4584 :
4585 : /*
4586 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4587 : * array, removing any gaps.
4588 : *
4589 : * A compression step is forced if "force" is true, otherwise we do it
4590 : * only if a heuristic indicates it's a good time to do it.
4591 : *
4592 : * Caller must hold ProcArrayLock in exclusive mode.
4593 : */
4594 : static void
4595 614 : KnownAssignedXidsCompress(bool force)
4596 : {
4597 614 : ProcArrayStruct *pArray = procArray;
4598 : int head,
4599 : tail;
4600 : int compress_index;
4601 : int i;
4602 :
4603 : /* no spinlock required since we hold ProcArrayLock exclusively */
4604 614 : head = pArray->headKnownAssignedXids;
4605 614 : tail = pArray->tailKnownAssignedXids;
4606 :
4607 614 : if (!force)
4608 : {
4609 : /*
4610 : * If we can choose how much to compress, use a heuristic to avoid
4611 : * compressing too often or not often enough.
4612 : *
4613 : * Heuristic is if we have a large enough current spread and less than
4614 : * 50% of the elements are currently in use, then compress. This
4615 : * should ensure we compress fairly infrequently. We could compress
4616 : * less often though the virtual array would spread out more and
4617 : * snapshots would become more expensive.
4618 : */
4619 614 : int nelements = head - tail;
4620 :
4621 614 : if (nelements < 4 * PROCARRAY_MAXPROCS ||
4622 16 : nelements < 2 * pArray->numKnownAssignedXids)
4623 598 : return;
4624 : }
4625 :
4626 : /*
4627 : * We compress the array by reading the valid values from tail to head,
4628 : * re-aligning data to 0th element.
4629 : */
4630 16 : compress_index = 0;
4631 2208 : for (i = tail; i < head; i++)
4632 : {
4633 2192 : if (KnownAssignedXidsValid[i])
4634 : {
4635 16 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4636 16 : KnownAssignedXidsValid[compress_index] = true;
4637 16 : compress_index++;
4638 : }
4639 : }
4640 :
4641 16 : pArray->tailKnownAssignedXids = 0;
4642 16 : pArray->headKnownAssignedXids = compress_index;
4643 : }
4644 :
4645 : /*
4646 : * Add xids into KnownAssignedXids at the head of the array.
4647 : *
4648 : * xids from from_xid to to_xid, inclusive, are added to the array.
4649 : *
4650 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4651 : * exclusive mode, so we need no extra locking here. Else caller holds no
4652 : * lock, so we need to be sure we maintain sufficient interlocks against
4653 : * concurrent readers. (Only the startup process ever calls this, so no need
4654 : * to worry about concurrent writers.)
4655 : */
4656 : static void
4657 3272 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4658 : bool exclusive_lock)
4659 : {
4660 3272 : ProcArrayStruct *pArray = procArray;
4661 : TransactionId next_xid;
4662 : int head,
4663 : tail;
4664 : int nxids;
4665 : int i;
4666 :
4667 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4668 :
4669 : /*
4670 : * Calculate how many array slots we'll need. Normally this is cheap; in
4671 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4672 : * way.
4673 : */
4674 3272 : if (to_xid >= from_xid)
4675 3272 : nxids = to_xid - from_xid + 1;
4676 : else
4677 : {
4678 0 : nxids = 1;
4679 0 : next_xid = from_xid;
4680 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4681 : {
4682 0 : nxids++;
4683 0 : TransactionIdAdvance(next_xid);
4684 : }
4685 : }
4686 :
4687 : /*
4688 : * Since only the startup process modifies the head/tail pointers, we
4689 : * don't need a lock to read them here.
4690 : */
4691 3272 : head = pArray->headKnownAssignedXids;
4692 3272 : tail = pArray->tailKnownAssignedXids;
4693 :
4694 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4695 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4696 :
4697 : /*
4698 : * Verify that insertions occur in TransactionId sequence. Note that even
4699 : * if the last existing element is marked invalid, it must still have a
4700 : * correctly sequenced XID value.
4701 : */
4702 6238 : if (head > tail &&
4703 2966 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4704 : {
4705 0 : KnownAssignedXidsDisplay(LOG);
4706 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4707 : }
4708 :
4709 : /*
4710 : * If our xids won't fit in the remaining space, compress out free space
4711 : */
4712 3272 : if (head + nxids > pArray->maxKnownAssignedXids)
4713 : {
4714 : /* must hold lock to compress */
4715 0 : if (!exclusive_lock)
4716 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4717 :
4718 0 : KnownAssignedXidsCompress(true);
4719 :
4720 0 : head = pArray->headKnownAssignedXids;
4721 : /* note: we no longer care about the tail pointer */
4722 :
4723 0 : if (!exclusive_lock)
4724 0 : LWLockRelease(ProcArrayLock);
4725 :
4726 : /*
4727 : * If it still won't fit then we're out of memory
4728 : */
4729 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4730 0 : elog(ERROR, "too many KnownAssignedXids");
4731 : }
4732 :
4733 : /* Now we can insert the xids into the space starting at head */
4734 3272 : next_xid = from_xid;
4735 6552 : for (i = 0; i < nxids; i++)
4736 : {
4737 3280 : KnownAssignedXids[head] = next_xid;
4738 3280 : KnownAssignedXidsValid[head] = true;
4739 3280 : TransactionIdAdvance(next_xid);
4740 3280 : head++;
4741 : }
4742 :
4743 : /* Adjust count of number of valid entries */
4744 3272 : pArray->numKnownAssignedXids += nxids;
4745 :
4746 : /*
4747 : * Now update the head pointer. We use a spinlock to protect this
4748 : * pointer, not because the update is likely to be non-atomic, but to
4749 : * ensure that other processors see the above array updates before they
4750 : * see the head pointer change.
4751 : *
4752 : * If we're holding ProcArrayLock exclusively, there's no need to take the
4753 : * spinlock.
4754 : */
4755 3272 : if (exclusive_lock)
4756 4 : pArray->headKnownAssignedXids = head;
4757 : else
4758 : {
4759 3268 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4760 3268 : pArray->headKnownAssignedXids = head;
4761 3268 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4762 : }
4763 3272 : }
4764 :
4765 : /*
4766 : * KnownAssignedXidsSearch
4767 : *
4768 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4769 : * Returns true if it was found, false if not.
4770 : *
4771 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4772 : * Exclusive lock must be held for remove = true.
4773 : */
4774 : static bool
4775 5430 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4776 : {
4777 5430 : ProcArrayStruct *pArray = procArray;
4778 : int first,
4779 : last;
4780 : int head;
4781 : int tail;
4782 5430 : int result_index = -1;
4783 :
4784 5430 : if (remove)
4785 : {
4786 : /* we hold ProcArrayLock exclusively, so no need for spinlock */
4787 5430 : tail = pArray->tailKnownAssignedXids;
4788 5430 : head = pArray->headKnownAssignedXids;
4789 : }
4790 : else
4791 : {
4792 : /* take spinlock to ensure we see up-to-date array contents */
4793 0 : SpinLockAcquire(&pArray->known_assigned_xids_lck);
4794 0 : tail = pArray->tailKnownAssignedXids;
4795 0 : head = pArray->headKnownAssignedXids;
4796 0 : SpinLockRelease(&pArray->known_assigned_xids_lck);
4797 : }
4798 :
4799 : /*
4800 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4801 : * array here, since even invalid entries will contain sorted XIDs.
4802 : */
4803 5430 : first = tail;
4804 5430 : last = head - 1;
4805 28298 : while (first <= last)
4806 : {
4807 : int mid_index;
4808 : TransactionId mid_xid;
4809 :
4810 25984 : mid_index = (first + last) / 2;
4811 25984 : mid_xid = KnownAssignedXids[mid_index];
4812 :
4813 25984 : if (xid == mid_xid)
4814 : {
4815 3116 : result_index = mid_index;
4816 3116 : break;
4817 : }
4818 22868 : else if (TransactionIdPrecedes(xid, mid_xid))
4819 16070 : last = mid_index - 1;
4820 : else
4821 6798 : first = mid_index + 1;
4822 : }
4823 :
4824 5430 : if (result_index < 0)
4825 2314 : return false; /* not in array */
4826 :
4827 3116 : if (!KnownAssignedXidsValid[result_index])
4828 0 : return false; /* in array, but invalid */
4829 :
4830 3116 : if (remove)
4831 : {
4832 3116 : KnownAssignedXidsValid[result_index] = false;
4833 :
4834 3116 : pArray->numKnownAssignedXids--;
4835 : Assert(pArray->numKnownAssignedXids >= 0);
4836 :
4837 : /*
4838 : * If we're removing the tail element then advance tail pointer over
4839 : * any invalid elements. This will speed future searches.
4840 : */
4841 3116 : if (result_index == tail)
4842 : {
4843 544 : tail++;
4844 812 : while (tail < head && !KnownAssignedXidsValid[tail])
4845 268 : tail++;
4846 544 : if (tail >= head)
4847 : {
4848 : /* Array is empty, so we can reset both pointers */
4849 292 : pArray->headKnownAssignedXids = 0;
4850 292 : pArray->tailKnownAssignedXids = 0;
4851 : }
4852 : else
4853 : {
4854 252 : pArray->tailKnownAssignedXids = tail;
4855 : }
4856 : }
4857 : }
4858 :
4859 3116 : return true;
4860 : }
4861 :
4862 : /*
4863 : * Is the specified XID present in KnownAssignedXids[]?
4864 : *
4865 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4866 : */
4867 : static bool
4868 0 : KnownAssignedXidExists(TransactionId xid)
4869 : {
4870 : Assert(TransactionIdIsValid(xid));
4871 :
4872 0 : return KnownAssignedXidsSearch(xid, false);
4873 : }
4874 :
4875 : /*
4876 : * Remove the specified XID from KnownAssignedXids[].
4877 : *
4878 : * Caller must hold ProcArrayLock in exclusive mode.
4879 : */
4880 : static void
4881 5430 : KnownAssignedXidsRemove(TransactionId xid)
4882 : {
4883 : Assert(TransactionIdIsValid(xid));
4884 :
4885 5430 : elog(trace_recovery(DEBUG4), "remove KnownAssignedXid %u", xid);
4886 :
4887 : /*
4888 : * Note: we cannot consider it an error to remove an XID that's not
4889 : * present. We intentionally remove subxact IDs while processing
4890 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4891 : * removed again when the top-level xact commits or aborts.
4892 : *
4893 : * It might be possible to track such XIDs to distinguish this case from
4894 : * actual errors, but it would be complicated and probably not worth it.
4895 : * So, just ignore the search result.
4896 : */
4897 5430 : (void) KnownAssignedXidsSearch(xid, true);
4898 5430 : }
4899 :
4900 : /*
4901 : * KnownAssignedXidsRemoveTree
4902 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4903 : *
4904 : * Caller must hold ProcArrayLock in exclusive mode.
4905 : */
4906 : static void
4907 360 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
4908 : TransactionId *subxids)
4909 : {
4910 : int i;
4911 :
4912 360 : if (TransactionIdIsValid(xid))
4913 320 : KnownAssignedXidsRemove(xid);
4914 :
4915 5470 : for (i = 0; i < nsubxids; i++)
4916 5110 : KnownAssignedXidsRemove(subxids[i]);
4917 :
4918 : /* Opportunistically compress the array */
4919 360 : KnownAssignedXidsCompress(false);
4920 360 : }
4921 :
4922 : /*
4923 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
4924 : * then clear the whole table.
4925 : *
4926 : * Caller must hold ProcArrayLock in exclusive mode.
4927 : */
4928 : static void
4929 368 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
4930 : {
4931 368 : ProcArrayStruct *pArray = procArray;
4932 368 : int count = 0;
4933 : int head,
4934 : tail,
4935 : i;
4936 :
4937 368 : if (!TransactionIdIsValid(removeXid))
4938 : {
4939 114 : elog(trace_recovery(DEBUG4), "removing all KnownAssignedXids");
4940 114 : pArray->numKnownAssignedXids = 0;
4941 114 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
4942 114 : return;
4943 : }
4944 :
4945 254 : elog(trace_recovery(DEBUG4), "prune KnownAssignedXids to %u", removeXid);
4946 :
4947 : /*
4948 : * Mark entries invalid starting at the tail. Since array is sorted, we
4949 : * can stop as soon as we reach an entry >= removeXid.
4950 : */
4951 254 : tail = pArray->tailKnownAssignedXids;
4952 254 : head = pArray->headKnownAssignedXids;
4953 :
4954 254 : for (i = tail; i < head; i++)
4955 : {
4956 36 : if (KnownAssignedXidsValid[i])
4957 : {
4958 36 : TransactionId knownXid = KnownAssignedXids[i];
4959 :
4960 36 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
4961 36 : break;
4962 :
4963 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
4964 : {
4965 0 : KnownAssignedXidsValid[i] = false;
4966 0 : count++;
4967 : }
4968 : }
4969 : }
4970 :
4971 254 : pArray->numKnownAssignedXids -= count;
4972 : Assert(pArray->numKnownAssignedXids >= 0);
4973 :
4974 : /*
4975 : * Advance the tail pointer if we've marked the tail item invalid.
4976 : */
4977 254 : for (i = tail; i < head; i++)
4978 : {
4979 36 : if (KnownAssignedXidsValid[i])
4980 36 : break;
4981 : }
4982 254 : if (i >= head)
4983 : {
4984 : /* Array is empty, so we can reset both pointers */
4985 218 : pArray->headKnownAssignedXids = 0;
4986 218 : pArray->tailKnownAssignedXids = 0;
4987 : }
4988 : else
4989 : {
4990 36 : pArray->tailKnownAssignedXids = i;
4991 : }
4992 :
4993 : /* Opportunistically compress the array */
4994 254 : KnownAssignedXidsCompress(false);
4995 : }
4996 :
4997 : /*
4998 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
4999 : * We filter out anything >= xmax.
5000 : *
5001 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5002 : * that array is large enough.
5003 : *
5004 : * Caller must hold ProcArrayLock in (at least) shared mode.
5005 : */
5006 : static int
5007 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5008 : {
5009 0 : TransactionId xtmp = InvalidTransactionId;
5010 :
5011 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5012 : }
5013 :
5014 : /*
5015 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5016 : * we reduce *xmin to the lowest xid value seen if not already lower.
5017 : *
5018 : * Caller must hold ProcArrayLock in (at least) shared mode.
5019 : */
5020 : static int
5021 838 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5022 : TransactionId xmax)
5023 : {
5024 838 : int count = 0;
5025 : int head,
5026 : tail;
5027 : int i;
5028 :
5029 : /*
5030 : * Fetch head just once, since it may change while we loop. We can stop
5031 : * once we reach the initially seen head, since we are certain that an xid
5032 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5033 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5034 : * anyway.
5035 : *
5036 : * Must take spinlock to ensure we see up-to-date array contents.
5037 : */
5038 838 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
5039 838 : tail = procArray->tailKnownAssignedXids;
5040 838 : head = procArray->headKnownAssignedXids;
5041 838 : SpinLockRelease(&procArray->known_assigned_xids_lck);
5042 :
5043 862 : for (i = tail; i < head; i++)
5044 : {
5045 : /* Skip any gaps in the array */
5046 142 : if (KnownAssignedXidsValid[i])
5047 : {
5048 124 : TransactionId knownXid = KnownAssignedXids[i];
5049 :
5050 : /*
5051 : * Update xmin if required. Only the first XID need be checked,
5052 : * since the array is sorted.
5053 : */
5054 248 : if (count == 0 &&
5055 124 : TransactionIdPrecedes(knownXid, *xmin))
5056 6 : *xmin = knownXid;
5057 :
5058 : /*
5059 : * Filter out anything >= xmax, again relying on sorted property
5060 : * of array.
5061 : */
5062 248 : if (TransactionIdIsValid(xmax) &&
5063 124 : TransactionIdFollowsOrEquals(knownXid, xmax))
5064 118 : break;
5065 :
5066 : /* Add knownXid into output array */
5067 6 : xarray[count++] = knownXid;
5068 : }
5069 : }
5070 :
5071 838 : return count;
5072 : }
5073 :
5074 : /*
5075 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5076 : * if nothing there.
5077 : */
5078 : static TransactionId
5079 160 : KnownAssignedXidsGetOldestXmin(void)
5080 : {
5081 : int head,
5082 : tail;
5083 : int i;
5084 :
5085 : /*
5086 : * Fetch head just once, since it may change while we loop.
5087 : */
5088 160 : SpinLockAcquire(&procArray->known_assigned_xids_lck);
5089 160 : tail = procArray->tailKnownAssignedXids;
5090 160 : head = procArray->headKnownAssignedXids;
5091 160 : SpinLockRelease(&procArray->known_assigned_xids_lck);
5092 :
5093 160 : for (i = tail; i < head; i++)
5094 : {
5095 : /* Skip any gaps in the array */
5096 6 : if (KnownAssignedXidsValid[i])
5097 6 : return KnownAssignedXids[i];
5098 : }
5099 :
5100 154 : return InvalidTransactionId;
5101 : }
5102 :
5103 : /*
5104 : * Display KnownAssignedXids to provide debug trail
5105 : *
5106 : * Currently this is only called within startup process, so we need no
5107 : * special locking.
5108 : *
5109 : * Note this is pretty expensive, and much of the expense will be incurred
5110 : * even if the elog message will get discarded. It's not currently called
5111 : * in any performance-critical places, however, so no need to be tenser.
5112 : */
5113 : static void
5114 118 : KnownAssignedXidsDisplay(int trace_level)
5115 : {
5116 118 : ProcArrayStruct *pArray = procArray;
5117 : StringInfoData buf;
5118 : int head,
5119 : tail,
5120 : i;
5121 118 : int nxids = 0;
5122 :
5123 118 : tail = pArray->tailKnownAssignedXids;
5124 118 : head = pArray->headKnownAssignedXids;
5125 :
5126 118 : initStringInfo(&buf);
5127 :
5128 126 : for (i = tail; i < head; i++)
5129 : {
5130 8 : if (KnownAssignedXidsValid[i])
5131 : {
5132 8 : nxids++;
5133 8 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5134 : }
5135 : }
5136 :
5137 118 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5138 : nxids,
5139 : pArray->numKnownAssignedXids,
5140 : pArray->tailKnownAssignedXids,
5141 : pArray->headKnownAssignedXids,
5142 : buf.data);
5143 :
5144 118 : pfree(buf.data);
5145 118 : }
5146 :
5147 : /*
5148 : * KnownAssignedXidsReset
5149 : * Resets KnownAssignedXids to be empty
5150 : */
5151 : static void
5152 0 : KnownAssignedXidsReset(void)
5153 : {
5154 0 : ProcArrayStruct *pArray = procArray;
5155 :
5156 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5157 :
5158 0 : pArray->numKnownAssignedXids = 0;
5159 0 : pArray->tailKnownAssignedXids = 0;
5160 0 : pArray->headKnownAssignedXids = 0;
5161 :
5162 0 : LWLockRelease(ProcArrayLock);
5163 0 : }
|
