1/*-------------------------------------------------------------------------
2 *
3 * varsup.c
4 * postgres OID & XID variables support routines
5 *
6 * Copyright (c) 2000-2019, PostgreSQL Global Development Group
7 *
8 * IDENTIFICATION
9 * src/backend/access/transam/varsup.c
10 *
11 *-------------------------------------------------------------------------
12 */
13
14#include "postgres.h"
15
16#include "access/clog.h"
17#include "access/commit_ts.h"
18#include "access/subtrans.h"
19#include "access/transam.h"
20#include "access/xact.h"
21#include "access/xlog.h"
22#include "commands/dbcommands.h"
23#include "miscadmin.h"
24#include "postmaster/autovacuum.h"
25#include "storage/pmsignal.h"
26#include "storage/proc.h"
27#include "utils/syscache.h"
28
29
30/* Number of OIDs to prefetch (preallocate) per XLOG write */
31#define VAR_OID_PREFETCH 8192
32
33/* pointer to "variable cache" in shared memory (set up by shmem.c) */
34VariableCache ShmemVariableCache = NULL;
35
36
37/*
38 * Allocate the next FullTransactionId for a new transaction or
39 * subtransaction.
40 *
41 * The new XID is also stored into MyPgXact before returning.
42 *
43 * Note: when this is called, we are actually already inside a valid
44 * transaction, since XIDs are now not allocated until the transaction
45 * does something. So it is safe to do a database lookup if we want to
46 * issue a warning about XID wrap.
47 */
48FullTransactionId
49GetNewTransactionId(bool isSubXact)
50{
51 FullTransactionId full_xid;
52 TransactionId xid;
53
54 /*
55 * Workers synchronize transaction state at the beginning of each parallel
56 * operation, so we can't account for new XIDs after that point.
57 */
58 if (IsInParallelMode())
59 elog(ERROR, "cannot assign TransactionIds during a parallel operation");
60
61 /*
62 * During bootstrap initialization, we return the special bootstrap
63 * transaction id.
64 */
65 if (IsBootstrapProcessingMode())
66 {
67 Assert(!isSubXact);
68 MyPgXact->xid = BootstrapTransactionId;
69 return FullTransactionIdFromEpochAndXid(0, BootstrapTransactionId);
70 }
71
72 /* safety check, we should never get this far in a HS standby */
73 if (RecoveryInProgress())
74 elog(ERROR, "cannot assign TransactionIds during recovery");
75
76 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
77
78 full_xid = ShmemVariableCache->nextFullXid;
79 xid = XidFromFullTransactionId(full_xid);
80
81 /*----------
82 * Check to see if it's safe to assign another XID. This protects against
83 * catastrophic data loss due to XID wraparound. The basic rules are:
84 *
85 * If we're past xidVacLimit, start trying to force autovacuum cycles.
86 * If we're past xidWarnLimit, start issuing warnings.
87 * If we're past xidStopLimit, refuse to execute transactions, unless
88 * we are running in single-user mode (which gives an escape hatch
89 * to the DBA who somehow got past the earlier defenses).
90 *
91 * Note that this coding also appears in GetNewMultiXactId.
92 *----------
93 */
94 if (TransactionIdFollowsOrEquals(xid, ShmemVariableCache->xidVacLimit))
95 {
96 /*
97 * For safety's sake, we release XidGenLock while sending signals,
98 * warnings, etc. This is not so much because we care about
99 * preserving concurrency in this situation, as to avoid any
100 * possibility of deadlock while doing get_database_name(). First,
101 * copy all the shared values we'll need in this path.
102 */
103 TransactionId xidWarnLimit = ShmemVariableCache->xidWarnLimit;
104 TransactionId xidStopLimit = ShmemVariableCache->xidStopLimit;
105 TransactionId xidWrapLimit = ShmemVariableCache->xidWrapLimit;
106 Oid oldest_datoid = ShmemVariableCache->oldestXidDB;
107
108 LWLockRelease(XidGenLock);
109
110 /*
111 * To avoid swamping the postmaster with signals, we issue the autovac
112 * request only once per 64K transaction starts. This still gives
113 * plenty of chances before we get into real trouble.
114 */
115 if (IsUnderPostmaster && (xid % 65536) == 0)
116 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
117
118 if (IsUnderPostmaster &&
119 TransactionIdFollowsOrEquals(xid, xidStopLimit))
120 {
121 char *oldest_datname = get_database_name(oldest_datoid);
122
123 /* complain even if that DB has disappeared */
124 if (oldest_datname)
125 ereport(ERROR,
126 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
127 errmsg("database is not accepting commands to avoid wraparound data loss in database \"%s\"",
128 oldest_datname),
129 errhint("Stop the postmaster and vacuum that database in single-user mode.\n"
130 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
131 else
132 ereport(ERROR,
133 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
134 errmsg("database is not accepting commands to avoid wraparound data loss in database with OID %u",
135 oldest_datoid),
136 errhint("Stop the postmaster and vacuum that database in single-user mode.\n"
137 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
138 }
139 else if (TransactionIdFollowsOrEquals(xid, xidWarnLimit))
140 {
141 char *oldest_datname = get_database_name(oldest_datoid);
142
143 /* complain even if that DB has disappeared */
144 if (oldest_datname)
145 ereport(WARNING,
146 (errmsg("database \"%s\" must be vacuumed within %u transactions",
147 oldest_datname,
148 xidWrapLimit - xid),
149 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
150 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
151 else
152 ereport(WARNING,
153 (errmsg("database with OID %u must be vacuumed within %u transactions",
154 oldest_datoid,
155 xidWrapLimit - xid),
156 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
157 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
158 }
159
160 /* Re-acquire lock and start over */
161 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
162 full_xid = ShmemVariableCache->nextFullXid;
163 xid = XidFromFullTransactionId(full_xid);
164 }
165
166 /*
167 * If we are allocating the first XID of a new page of the commit log,
168 * zero out that commit-log page before returning. We must do this while
169 * holding XidGenLock, else another xact could acquire and commit a later
170 * XID before we zero the page. Fortunately, a page of the commit log
171 * holds 32K or more transactions, so we don't have to do this very often.
172 *
173 * Extend pg_subtrans and pg_commit_ts too.
174 */
175 ExtendCLOG(xid);
176 ExtendCommitTs(xid);
177 ExtendSUBTRANS(xid);
178
179 /*
180 * Now advance the nextFullXid counter. This must not happen until after
181 * we have successfully completed ExtendCLOG() --- if that routine fails,
182 * we want the next incoming transaction to try it again. We cannot
183 * assign more XIDs until there is CLOG space for them.
184 */
185 FullTransactionIdAdvance(&ShmemVariableCache->nextFullXid);
186
187 /*
188 * We must store the new XID into the shared ProcArray before releasing
189 * XidGenLock. This ensures that every active XID older than
190 * latestCompletedXid is present in the ProcArray, which is essential for
191 * correct OldestXmin tracking; see src/backend/access/transam/README.
192 *
193 * Note that readers of PGXACT xid fields should be careful to fetch the
194 * value only once, rather than assume they can read a value multiple
195 * times and get the same answer each time. Note we are assuming that
196 * TransactionId and int fetch/store are atomic.
197 *
198 * The same comments apply to the subxact xid count and overflow fields.
199 *
200 * Use of a write barrier prevents dangerous code rearrangement in this
201 * function; other backends could otherwise e.g. be examining my subxids
202 * info concurrently, and we don't want them to see an invalid
203 * intermediate state, such as an incremented nxids before the array entry
204 * is filled.
205 *
206 * Other processes that read nxids should do so before reading xids
207 * elements with a pg_read_barrier() in between, so that they can be sure
208 * not to read an uninitialized array element; see
209 * src/backend/storage/lmgr/README.barrier.
210 *
211 * If there's no room to fit a subtransaction XID into PGPROC, set the
212 * cache-overflowed flag instead. This forces readers to look in
213 * pg_subtrans to map subtransaction XIDs up to top-level XIDs. There is a
214 * race-condition window, in that the new XID will not appear as running
215 * until its parent link has been placed into pg_subtrans. However, that
216 * will happen before anyone could possibly have a reason to inquire about
217 * the status of the XID, so it seems OK. (Snapshots taken during this
218 * window *will* include the parent XID, so they will deliver the correct
219 * answer later on when someone does have a reason to inquire.)
220 */
221 if (!isSubXact)
222 MyPgXact->xid = xid; /* LWLockRelease acts as barrier */
223 else
224 {
225 int nxids = MyPgXact->nxids;
226
227 if (nxids < PGPROC_MAX_CACHED_SUBXIDS)
228 {
229 MyProc->subxids.xids[nxids] = xid;
230 pg_write_barrier();
231 MyPgXact->nxids = nxids + 1;
232 }
233 else
234 MyPgXact->overflowed = true;
235 }
236
237 LWLockRelease(XidGenLock);
238
239 return full_xid;
240}
241
242/*
243 * Read nextFullXid but don't allocate it.
244 */
245FullTransactionId
246ReadNextFullTransactionId(void)
247{
248 FullTransactionId fullXid;
249
250 LWLockAcquire(XidGenLock, LW_SHARED);
251 fullXid = ShmemVariableCache->nextFullXid;
252 LWLockRelease(XidGenLock);
253
254 return fullXid;
255}
256
257/*
258 * Advance nextFullXid to the value after a given xid. The epoch is inferred.
259 * This must only be called during recovery or from two-phase start-up code.
260 */
261void
262AdvanceNextFullTransactionIdPastXid(TransactionId xid)
263{
264 FullTransactionId newNextFullXid;
265 TransactionId next_xid;
266 uint32 epoch;
267
268 /*
269 * It is safe to read nextFullXid without a lock, because this is only
270 * called from the startup process or single-process mode, meaning that no
271 * other process can modify it.
272 */
273 Assert(AmStartupProcess() || !IsUnderPostmaster);
274
275 /* Fast return if this isn't an xid high enough to move the needle. */
276 next_xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
277 if (!TransactionIdFollowsOrEquals(xid, next_xid))
278 return;
279
280 /*
281 * Compute the FullTransactionId that comes after the given xid. To do
282 * this, we preserve the existing epoch, but detect when we've wrapped
283 * into a new epoch. This is necessary because WAL records and 2PC state
284 * currently contain 32 bit xids. The wrap logic is safe in those cases
285 * because the span of active xids cannot exceed one epoch at any given
286 * point in the WAL stream.
287 */
288 TransactionIdAdvance(xid);
289 epoch = EpochFromFullTransactionId(ShmemVariableCache->nextFullXid);
290 if (unlikely(xid < next_xid))
291 ++epoch;
292 newNextFullXid = FullTransactionIdFromEpochAndXid(epoch, xid);
293
294 /*
295 * We still need to take a lock to modify the value when there are
296 * concurrent readers.
297 */
298 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
299 ShmemVariableCache->nextFullXid = newNextFullXid;
300 LWLockRelease(XidGenLock);
301}
302
303/*
304 * Advance the cluster-wide value for the oldest valid clog entry.
305 *
306 * We must acquire CLogTruncationLock to advance the oldestClogXid. It's not
307 * necessary to hold the lock during the actual clog truncation, only when we
308 * advance the limit, as code looking up arbitrary xids is required to hold
309 * CLogTruncationLock from when it tests oldestClogXid through to when it
310 * completes the clog lookup.
311 */
312void
313AdvanceOldestClogXid(TransactionId oldest_datfrozenxid)
314{
315 LWLockAcquire(CLogTruncationLock, LW_EXCLUSIVE);
316 if (TransactionIdPrecedes(ShmemVariableCache->oldestClogXid,
317 oldest_datfrozenxid))
318 {
319 ShmemVariableCache->oldestClogXid = oldest_datfrozenxid;
320 }
321 LWLockRelease(CLogTruncationLock);
322}
323
324/*
325 * Determine the last safe XID to allocate using the currently oldest
326 * datfrozenxid (ie, the oldest XID that might exist in any database
327 * of our cluster), and the OID of the (or a) database with that value.
328 */
329void
330SetTransactionIdLimit(TransactionId oldest_datfrozenxid, Oid oldest_datoid)
331{
332 TransactionId xidVacLimit;
333 TransactionId xidWarnLimit;
334 TransactionId xidStopLimit;
335 TransactionId xidWrapLimit;
336 TransactionId curXid;
337
338 Assert(TransactionIdIsNormal(oldest_datfrozenxid));
339
340 /*
341 * The place where we actually get into deep trouble is halfway around
342 * from the oldest potentially-existing XID. (This calculation is
343 * probably off by one or two counts, because the special XIDs reduce the
344 * size of the loop a little bit. But we throw in plenty of slop below,
345 * so it doesn't matter.)
346 */
347 xidWrapLimit = oldest_datfrozenxid + (MaxTransactionId >> 1);
348 if (xidWrapLimit < FirstNormalTransactionId)
349 xidWrapLimit += FirstNormalTransactionId;
350
351 /*
352 * We'll refuse to continue assigning XIDs in interactive mode once we get
353 * within 1M transactions of data loss. This leaves lots of room for the
354 * DBA to fool around fixing things in a standalone backend, while not
355 * being significant compared to total XID space. (Note that since
356 * vacuuming requires one transaction per table cleaned, we had better be
357 * sure there's lots of XIDs left...)
358 */
359 xidStopLimit = xidWrapLimit - 1000000;
360 if (xidStopLimit < FirstNormalTransactionId)
361 xidStopLimit -= FirstNormalTransactionId;
362
363 /*
364 * We'll start complaining loudly when we get within 10M transactions of
365 * the stop point. This is kind of arbitrary, but if you let your gas
366 * gauge get down to 1% of full, would you be looking for the next gas
367 * station? We need to be fairly liberal about this number because there
368 * are lots of scenarios where most transactions are done by automatic
369 * clients that won't pay attention to warnings. (No, we're not gonna make
370 * this configurable. If you know enough to configure it, you know enough
371 * to not get in this kind of trouble in the first place.)
372 */
373 xidWarnLimit = xidStopLimit - 10000000;
374 if (xidWarnLimit < FirstNormalTransactionId)
375 xidWarnLimit -= FirstNormalTransactionId;
376
377 /*
378 * We'll start trying to force autovacuums when oldest_datfrozenxid gets
379 * to be more than autovacuum_freeze_max_age transactions old.
380 *
381 * Note: guc.c ensures that autovacuum_freeze_max_age is in a sane range,
382 * so that xidVacLimit will be well before xidWarnLimit.
383 *
384 * Note: autovacuum_freeze_max_age is a PGC_POSTMASTER parameter so that
385 * we don't have to worry about dealing with on-the-fly changes in its
386 * value. It doesn't look practical to update shared state from a GUC
387 * assign hook (too many processes would try to execute the hook,
388 * resulting in race conditions as well as crashes of those not connected
389 * to shared memory). Perhaps this can be improved someday. See also
390 * SetMultiXactIdLimit.
391 */
392 xidVacLimit = oldest_datfrozenxid + autovacuum_freeze_max_age;
393 if (xidVacLimit < FirstNormalTransactionId)
394 xidVacLimit += FirstNormalTransactionId;
395
396 /* Grab lock for just long enough to set the new limit values */
397 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
398 ShmemVariableCache->oldestXid = oldest_datfrozenxid;
399 ShmemVariableCache->xidVacLimit = xidVacLimit;
400 ShmemVariableCache->xidWarnLimit = xidWarnLimit;
401 ShmemVariableCache->xidStopLimit = xidStopLimit;
402 ShmemVariableCache->xidWrapLimit = xidWrapLimit;
403 ShmemVariableCache->oldestXidDB = oldest_datoid;
404 curXid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
405 LWLockRelease(XidGenLock);
406
407 /* Log the info */
408 ereport(DEBUG1,
409 (errmsg("transaction ID wrap limit is %u, limited by database with OID %u",
410 xidWrapLimit, oldest_datoid)));
411
412 /*
413 * If past the autovacuum force point, immediately signal an autovac
414 * request. The reason for this is that autovac only processes one
415 * database per invocation. Once it's finished cleaning up the oldest
416 * database, it'll call here, and we'll signal the postmaster to start
417 * another iteration immediately if there are still any old databases.
418 */
419 if (TransactionIdFollowsOrEquals(curXid, xidVacLimit) &&
420 IsUnderPostmaster && !InRecovery)
421 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
422
423 /* Give an immediate warning if past the wrap warn point */
424 if (TransactionIdFollowsOrEquals(curXid, xidWarnLimit) && !InRecovery)
425 {
426 char *oldest_datname;
427
428 /*
429 * We can be called when not inside a transaction, for example during
430 * StartupXLOG(). In such a case we cannot do database access, so we
431 * must just report the oldest DB's OID.
432 *
433 * Note: it's also possible that get_database_name fails and returns
434 * NULL, for example because the database just got dropped. We'll
435 * still warn, even though the warning might now be unnecessary.
436 */
437 if (IsTransactionState())
438 oldest_datname = get_database_name(oldest_datoid);
439 else
440 oldest_datname = NULL;
441
442 if (oldest_datname)
443 ereport(WARNING,
444 (errmsg("database \"%s\" must be vacuumed within %u transactions",
445 oldest_datname,
446 xidWrapLimit - curXid),
447 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
448 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
449 else
450 ereport(WARNING,
451 (errmsg("database with OID %u must be vacuumed within %u transactions",
452 oldest_datoid,
453 xidWrapLimit - curXid),
454 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
455 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
456 }
457}
458
459
460/*
461 * ForceTransactionIdLimitUpdate -- does the XID wrap-limit data need updating?
462 *
463 * We primarily check whether oldestXidDB is valid. The cases we have in
464 * mind are that that database was dropped, or the field was reset to zero
465 * by pg_resetwal. In either case we should force recalculation of the
466 * wrap limit. Also do it if oldestXid is old enough to be forcing
467 * autovacuums or other actions; this ensures we update our state as soon
468 * as possible once extra overhead is being incurred.
469 */
470bool
471ForceTransactionIdLimitUpdate(void)
472{
473 TransactionId nextXid;
474 TransactionId xidVacLimit;
475 TransactionId oldestXid;
476 Oid oldestXidDB;
477
478 /* Locking is probably not really necessary, but let's be careful */
479 LWLockAcquire(XidGenLock, LW_SHARED);
480 nextXid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
481 xidVacLimit = ShmemVariableCache->xidVacLimit;
482 oldestXid = ShmemVariableCache->oldestXid;
483 oldestXidDB = ShmemVariableCache->oldestXidDB;
484 LWLockRelease(XidGenLock);
485
486 if (!TransactionIdIsNormal(oldestXid))
487 return true; /* shouldn't happen, but just in case */
488 if (!TransactionIdIsValid(xidVacLimit))
489 return true; /* this shouldn't happen anymore either */
490 if (TransactionIdFollowsOrEquals(nextXid, xidVacLimit))
491 return true; /* past VacLimit, don't delay updating */
492 if (!SearchSysCacheExists1(DATABASEOID, ObjectIdGetDatum(oldestXidDB)))
493 return true; /* could happen, per comments above */
494 return false;
495}
496
497
498/*
499 * GetNewObjectId -- allocate a new OID
500 *
501 * OIDs are generated by a cluster-wide counter. Since they are only 32 bits
502 * wide, counter wraparound will occur eventually, and therefore it is unwise
503 * to assume they are unique unless precautions are taken to make them so.
504 * Hence, this routine should generally not be used directly. The only direct
505 * callers should be GetNewOidWithIndex() and GetNewRelFileNode() in
506 * catalog/catalog.c.
507 */
508Oid
509GetNewObjectId(void)
510{
511 Oid result;
512
513 /* safety check, we should never get this far in a HS standby */
514 if (RecoveryInProgress())
515 elog(ERROR, "cannot assign OIDs during recovery");
516
517 LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
518
519 /*
520 * Check for wraparound of the OID counter. We *must* not return 0
521 * (InvalidOid), and in normal operation we mustn't return anything below
522 * FirstNormalObjectId since that range is reserved for initdb (see
523 * IsCatalogRelationOid()). Note we are relying on unsigned comparison.
524 *
525 * During initdb, we start the OID generator at FirstBootstrapObjectId, so
526 * we only wrap if before that point when in bootstrap or standalone mode.
527 * The first time through this routine after normal postmaster start, the
528 * counter will be forced up to FirstNormalObjectId. This mechanism
529 * leaves the OIDs between FirstBootstrapObjectId and FirstNormalObjectId
530 * available for automatic assignment during initdb, while ensuring they
531 * will never conflict with user-assigned OIDs.
532 */
533 if (ShmemVariableCache->nextOid < ((Oid) FirstNormalObjectId))
534 {
535 if (IsPostmasterEnvironment)
536 {
537 /* wraparound, or first post-initdb assignment, in normal mode */
538 ShmemVariableCache->nextOid = FirstNormalObjectId;
539 ShmemVariableCache->oidCount = 0;
540 }
541 else
542 {
543 /* we may be bootstrapping, so don't enforce the full range */
544 if (ShmemVariableCache->nextOid < ((Oid) FirstBootstrapObjectId))
545 {
546 /* wraparound in standalone mode (unlikely but possible) */
547 ShmemVariableCache->nextOid = FirstNormalObjectId;
548 ShmemVariableCache->oidCount = 0;
549 }
550 }
551 }
552
553 /* If we run out of logged for use oids then we must log more */
554 if (ShmemVariableCache->oidCount == 0)
555 {
556 XLogPutNextOid(ShmemVariableCache->nextOid + VAR_OID_PREFETCH);
557 ShmemVariableCache->oidCount = VAR_OID_PREFETCH;
558 }
559
560 result = ShmemVariableCache->nextOid;
561
562 (ShmemVariableCache->nextOid)++;
563 (ShmemVariableCache->oidCount)--;
564
565 LWLockRelease(OidGenLock);
566
567 return result;
568}
569