1 /*-------------------------------------------------------------------------
4 * postgres OID & XID variables support routines
6 * Copyright (c) 2000-2019, PostgreSQL Global Development Group
9 * src/backend/access/transam/varsup.c
11 *-------------------------------------------------------------------------
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"
30 /* Number of OIDs to prefetch (preallocate) per XLOG write */
31 #define VAR_OID_PREFETCH 8192
33 /* pointer to "variable cache" in shared memory (set up by shmem.c) */
34 VariableCache ShmemVariableCache = NULL;
38 * Allocate the next FullTransactionId for a new transaction or
41 * The new XID is also stored into MyPgXact before returning.
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.
49 GetNewTransactionId(bool isSubXact)
51 FullTransactionId full_xid;
55 * Workers synchronize transaction state at the beginning of each parallel
56 * operation, so we can't account for new XIDs after that point.
58 if (IsInParallelMode())
59 elog(ERROR, "cannot assign TransactionIds during a parallel operation");
62 * During bootstrap initialization, we return the special bootstrap
65 if (IsBootstrapProcessingMode())
68 MyPgXact->xid = BootstrapTransactionId;
69 return FullTransactionIdFromEpochAndXid(0, BootstrapTransactionId);
72 /* safety check, we should never get this far in a HS standby */
73 if (RecoveryInProgress())
74 elog(ERROR, "cannot assign TransactionIds during recovery");
76 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
78 full_xid = ShmemVariableCache->nextFullXid;
79 xid = XidFromFullTransactionId(full_xid);
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:
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).
91 * Note that this coding also appears in GetNewMultiXactId.
94 if (TransactionIdFollowsOrEquals(xid, ShmemVariableCache->xidVacLimit))
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.
103 TransactionId xidWarnLimit = ShmemVariableCache->xidWarnLimit;
104 TransactionId xidStopLimit = ShmemVariableCache->xidStopLimit;
105 TransactionId xidWrapLimit = ShmemVariableCache->xidWrapLimit;
106 Oid oldest_datoid = ShmemVariableCache->oldestXidDB;
108 LWLockRelease(XidGenLock);
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.
115 if (IsUnderPostmaster && (xid % 65536) == 0)
116 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
118 if (IsUnderPostmaster &&
119 TransactionIdFollowsOrEquals(xid, xidStopLimit))
121 char *oldest_datname = get_database_name(oldest_datoid);
123 /* complain even if that DB has disappeared */
126 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
127 errmsg("database is not accepting commands to avoid wraparound data loss in database \"%s\"",
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.")));
133 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
134 errmsg("database is not accepting commands to avoid wraparound data loss in database with OID %u",
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.")));
139 else if (TransactionIdFollowsOrEquals(xid, xidWarnLimit))
141 char *oldest_datname = get_database_name(oldest_datoid);
143 /* complain even if that DB has disappeared */
146 (errmsg("database \"%s\" must be vacuumed within %u transactions",
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.")));
153 (errmsg("database with OID %u must be vacuumed within %u transactions",
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.")));
160 /* Re-acquire lock and start over */
161 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
162 xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
166 * If we are allocating the first XID of a new page of the commit log,
167 * zero out that commit-log page before returning. We must do this while
168 * holding XidGenLock, else another xact could acquire and commit a later
169 * XID before we zero the page. Fortunately, a page of the commit log
170 * holds 32K or more transactions, so we don't have to do this very often.
172 * Extend pg_subtrans and pg_commit_ts too.
179 * Now advance the nextFullXid counter. This must not happen until after
180 * we have successfully completed ExtendCLOG() --- if that routine fails,
181 * we want the next incoming transaction to try it again. We cannot
182 * assign more XIDs until there is CLOG space for them.
184 FullTransactionIdAdvance(&ShmemVariableCache->nextFullXid);
187 * We must store the new XID into the shared ProcArray before releasing
188 * XidGenLock. This ensures that every active XID older than
189 * latestCompletedXid is present in the ProcArray, which is essential for
190 * correct OldestXmin tracking; see src/backend/access/transam/README.
192 * Note that readers of PGXACT xid fields should be careful to fetch the
193 * value only once, rather than assume they can read a value multiple
194 * times and get the same answer each time. Note we are assuming that
195 * TransactionId and int fetch/store are atomic.
197 * The same comments apply to the subxact xid count and overflow fields.
199 * Use of a write barrier prevents dangerous code rearrangement in this
200 * function; other backends could otherwise e.g. be examining my subxids
201 * info concurrently, and we don't want them to see an invalid
202 * intermediate state, such as an incremented nxids before the array entry
205 * Other processes that read nxids should do so before reading xids
206 * elements with a pg_read_barrier() in between, so that they can be sure
207 * not to read an uninitialized array element; see
208 * src/backend/storage/lmgr/README.barrier.
210 * If there's no room to fit a subtransaction XID into PGPROC, set the
211 * cache-overflowed flag instead. This forces readers to look in
212 * pg_subtrans to map subtransaction XIDs up to top-level XIDs. There is a
213 * race-condition window, in that the new XID will not appear as running
214 * until its parent link has been placed into pg_subtrans. However, that
215 * will happen before anyone could possibly have a reason to inquire about
216 * the status of the XID, so it seems OK. (Snapshots taken during this
217 * window *will* include the parent XID, so they will deliver the correct
218 * answer later on when someone does have a reason to inquire.)
221 MyPgXact->xid = xid; /* LWLockRelease acts as barrier */
224 int nxids = MyPgXact->nxids;
226 if (nxids < PGPROC_MAX_CACHED_SUBXIDS)
228 MyProc->subxids.xids[nxids] = xid;
230 MyPgXact->nxids = nxids + 1;
233 MyPgXact->overflowed = true;
236 LWLockRelease(XidGenLock);
242 * Read nextFullXid but don't allocate it.
245 ReadNextFullTransactionId(void)
247 FullTransactionId fullXid;
249 LWLockAcquire(XidGenLock, LW_SHARED);
250 fullXid = ShmemVariableCache->nextFullXid;
251 LWLockRelease(XidGenLock);
257 * Advance nextFullXid to the value after a given xid. The epoch is inferred.
258 * This must only be called during recovery or from two-phase start-up code.
261 AdvanceNextFullTransactionIdPastXid(TransactionId xid)
263 FullTransactionId newNextFullXid;
264 TransactionId next_xid;
268 * It is safe to read nextFullXid without a lock, because this is only
269 * called from the startup process or single-process mode, meaning that no
270 * other process can modify it.
272 Assert(AmStartupProcess() || !IsUnderPostmaster);
274 /* Fast return if this isn't an xid high enough to move the needle. */
275 next_xid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
276 if (!TransactionIdFollowsOrEquals(xid, next_xid))
280 * Compute the FullTransactionId that comes after the given xid. To do
281 * this, we preserve the existing epoch, but detect when we've wrapped
282 * into a new epoch. This is necessary because WAL records and 2PC state
283 * currently contain 32 bit xids. The wrap logic is safe in those cases
284 * because the span of active xids cannot exceed one epoch at any given
285 * point in the WAL stream.
287 TransactionIdAdvance(xid);
288 epoch = EpochFromFullTransactionId(ShmemVariableCache->nextFullXid);
289 if (unlikely(xid < next_xid))
291 newNextFullXid = FullTransactionIdFromEpochAndXid(epoch, xid);
294 * We still need to take a lock to modify the value when there are
295 * concurrent readers.
297 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
298 ShmemVariableCache->nextFullXid = newNextFullXid;
299 LWLockRelease(XidGenLock);
303 * Advance the cluster-wide value for the oldest valid clog entry.
305 * We must acquire CLogTruncationLock to advance the oldestClogXid. It's not
306 * necessary to hold the lock during the actual clog truncation, only when we
307 * advance the limit, as code looking up arbitrary xids is required to hold
308 * CLogTruncationLock from when it tests oldestClogXid through to when it
309 * completes the clog lookup.
312 AdvanceOldestClogXid(TransactionId oldest_datfrozenxid)
314 LWLockAcquire(CLogTruncationLock, LW_EXCLUSIVE);
315 if (TransactionIdPrecedes(ShmemVariableCache->oldestClogXid,
316 oldest_datfrozenxid))
318 ShmemVariableCache->oldestClogXid = oldest_datfrozenxid;
320 LWLockRelease(CLogTruncationLock);
324 * Determine the last safe XID to allocate using the currently oldest
325 * datfrozenxid (ie, the oldest XID that might exist in any database
326 * of our cluster), and the OID of the (or a) database with that value.
329 SetTransactionIdLimit(TransactionId oldest_datfrozenxid, Oid oldest_datoid)
331 TransactionId xidVacLimit;
332 TransactionId xidWarnLimit;
333 TransactionId xidStopLimit;
334 TransactionId xidWrapLimit;
335 TransactionId curXid;
337 Assert(TransactionIdIsNormal(oldest_datfrozenxid));
340 * The place where we actually get into deep trouble is halfway around
341 * from the oldest potentially-existing XID. (This calculation is
342 * probably off by one or two counts, because the special XIDs reduce the
343 * size of the loop a little bit. But we throw in plenty of slop below,
344 * so it doesn't matter.)
346 xidWrapLimit = oldest_datfrozenxid + (MaxTransactionId >> 1);
347 if (xidWrapLimit < FirstNormalTransactionId)
348 xidWrapLimit += FirstNormalTransactionId;
351 * We'll refuse to continue assigning XIDs in interactive mode once we get
352 * within 1M transactions of data loss. This leaves lots of room for the
353 * DBA to fool around fixing things in a standalone backend, while not
354 * being significant compared to total XID space. (Note that since
355 * vacuuming requires one transaction per table cleaned, we had better be
356 * sure there's lots of XIDs left...)
358 xidStopLimit = xidWrapLimit - 1000000;
359 if (xidStopLimit < FirstNormalTransactionId)
360 xidStopLimit -= FirstNormalTransactionId;
363 * We'll start complaining loudly when we get within 10M transactions of
364 * the stop point. This is kind of arbitrary, but if you let your gas
365 * gauge get down to 1% of full, would you be looking for the next gas
366 * station? We need to be fairly liberal about this number because there
367 * are lots of scenarios where most transactions are done by automatic
368 * clients that won't pay attention to warnings. (No, we're not gonna make
369 * this configurable. If you know enough to configure it, you know enough
370 * to not get in this kind of trouble in the first place.)
372 xidWarnLimit = xidStopLimit - 10000000;
373 if (xidWarnLimit < FirstNormalTransactionId)
374 xidWarnLimit -= FirstNormalTransactionId;
377 * We'll start trying to force autovacuums when oldest_datfrozenxid gets
378 * to be more than autovacuum_freeze_max_age transactions old.
380 * Note: guc.c ensures that autovacuum_freeze_max_age is in a sane range,
381 * so that xidVacLimit will be well before xidWarnLimit.
383 * Note: autovacuum_freeze_max_age is a PGC_POSTMASTER parameter so that
384 * we don't have to worry about dealing with on-the-fly changes in its
385 * value. It doesn't look practical to update shared state from a GUC
386 * assign hook (too many processes would try to execute the hook,
387 * resulting in race conditions as well as crashes of those not connected
388 * to shared memory). Perhaps this can be improved someday. See also
389 * SetMultiXactIdLimit.
391 xidVacLimit = oldest_datfrozenxid + autovacuum_freeze_max_age;
392 if (xidVacLimit < FirstNormalTransactionId)
393 xidVacLimit += FirstNormalTransactionId;
395 /* Grab lock for just long enough to set the new limit values */
396 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
397 ShmemVariableCache->oldestXid = oldest_datfrozenxid;
398 ShmemVariableCache->xidVacLimit = xidVacLimit;
399 ShmemVariableCache->xidWarnLimit = xidWarnLimit;
400 ShmemVariableCache->xidStopLimit = xidStopLimit;
401 ShmemVariableCache->xidWrapLimit = xidWrapLimit;
402 ShmemVariableCache->oldestXidDB = oldest_datoid;
403 curXid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
404 LWLockRelease(XidGenLock);
408 (errmsg("transaction ID wrap limit is %u, limited by database with OID %u",
409 xidWrapLimit, oldest_datoid)));
412 * If past the autovacuum force point, immediately signal an autovac
413 * request. The reason for this is that autovac only processes one
414 * database per invocation. Once it's finished cleaning up the oldest
415 * database, it'll call here, and we'll signal the postmaster to start
416 * another iteration immediately if there are still any old databases.
418 if (TransactionIdFollowsOrEquals(curXid, xidVacLimit) &&
419 IsUnderPostmaster && !InRecovery)
420 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
422 /* Give an immediate warning if past the wrap warn point */
423 if (TransactionIdFollowsOrEquals(curXid, xidWarnLimit) && !InRecovery)
425 char *oldest_datname;
428 * We can be called when not inside a transaction, for example during
429 * StartupXLOG(). In such a case we cannot do database access, so we
430 * must just report the oldest DB's OID.
432 * Note: it's also possible that get_database_name fails and returns
433 * NULL, for example because the database just got dropped. We'll
434 * still warn, even though the warning might now be unnecessary.
436 if (IsTransactionState())
437 oldest_datname = get_database_name(oldest_datoid);
439 oldest_datname = NULL;
443 (errmsg("database \"%s\" must be vacuumed within %u transactions",
445 xidWrapLimit - curXid),
446 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
447 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
450 (errmsg("database with OID %u must be vacuumed within %u transactions",
452 xidWrapLimit - curXid),
453 errhint("To avoid a database shutdown, execute a database-wide VACUUM in that database.\n"
454 "You might also need to commit or roll back old prepared transactions, or drop stale replication slots.")));
460 * ForceTransactionIdLimitUpdate -- does the XID wrap-limit data need updating?
462 * We primarily check whether oldestXidDB is valid. The cases we have in
463 * mind are that that database was dropped, or the field was reset to zero
464 * by pg_resetwal. In either case we should force recalculation of the
465 * wrap limit. Also do it if oldestXid is old enough to be forcing
466 * autovacuums or other actions; this ensures we update our state as soon
467 * as possible once extra overhead is being incurred.
470 ForceTransactionIdLimitUpdate(void)
472 TransactionId nextXid;
473 TransactionId xidVacLimit;
474 TransactionId oldestXid;
477 /* Locking is probably not really necessary, but let's be careful */
478 LWLockAcquire(XidGenLock, LW_SHARED);
479 nextXid = XidFromFullTransactionId(ShmemVariableCache->nextFullXid);
480 xidVacLimit = ShmemVariableCache->xidVacLimit;
481 oldestXid = ShmemVariableCache->oldestXid;
482 oldestXidDB = ShmemVariableCache->oldestXidDB;
483 LWLockRelease(XidGenLock);
485 if (!TransactionIdIsNormal(oldestXid))
486 return true; /* shouldn't happen, but just in case */
487 if (!TransactionIdIsValid(xidVacLimit))
488 return true; /* this shouldn't happen anymore either */
489 if (TransactionIdFollowsOrEquals(nextXid, xidVacLimit))
490 return true; /* past VacLimit, don't delay updating */
491 if (!SearchSysCacheExists1(DATABASEOID, ObjectIdGetDatum(oldestXidDB)))
492 return true; /* could happen, per comments above */
498 * GetNewObjectId -- allocate a new OID
500 * OIDs are generated by a cluster-wide counter. Since they are only 32 bits
501 * wide, counter wraparound will occur eventually, and therefore it is unwise
502 * to assume they are unique unless precautions are taken to make them so.
503 * Hence, this routine should generally not be used directly. The only direct
504 * callers should be GetNewOidWithIndex() and GetNewRelFileNode() in
512 /* safety check, we should never get this far in a HS standby */
513 if (RecoveryInProgress())
514 elog(ERROR, "cannot assign OIDs during recovery");
516 LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
519 * Check for wraparound of the OID counter. We *must* not return 0
520 * (InvalidOid), and in normal operation we mustn't return anything below
521 * FirstNormalObjectId since that range is reserved for initdb (see
522 * IsCatalogClass()). Note we are relying on unsigned comparison.
524 * During initdb, we start the OID generator at FirstBootstrapObjectId, so
525 * we only wrap if before that point when in bootstrap or standalone mode.
526 * The first time through this routine after normal postmaster start, the
527 * counter will be forced up to FirstNormalObjectId. This mechanism
528 * leaves the OIDs between FirstBootstrapObjectId and FirstNormalObjectId
529 * available for automatic assignment during initdb, while ensuring they
530 * will never conflict with user-assigned OIDs.
532 if (ShmemVariableCache->nextOid < ((Oid) FirstNormalObjectId))
534 if (IsPostmasterEnvironment)
536 /* wraparound, or first post-initdb assignment, in normal mode */
537 ShmemVariableCache->nextOid = FirstNormalObjectId;
538 ShmemVariableCache->oidCount = 0;
542 /* we may be bootstrapping, so don't enforce the full range */
543 if (ShmemVariableCache->nextOid < ((Oid) FirstBootstrapObjectId))
545 /* wraparound in standalone mode (unlikely but possible) */
546 ShmemVariableCache->nextOid = FirstNormalObjectId;
547 ShmemVariableCache->oidCount = 0;
552 /* If we run out of logged for use oids then we must log more */
553 if (ShmemVariableCache->oidCount == 0)
555 XLogPutNextOid(ShmemVariableCache->nextOid + VAR_OID_PREFETCH);
556 ShmemVariableCache->oidCount = VAR_OID_PREFETCH;
559 result = ShmemVariableCache->nextOid;
561 (ShmemVariableCache->nextOid)++;
562 (ShmemVariableCache->oidCount)--;
564 LWLockRelease(OidGenLock);