src/backend/access/transam/varsup.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * varsup.c
   4  *        postgres OID & XID variables support routines
   5  *
   6  * Copyright (c) 2000-2009, PostgreSQL Global Development Group
   7  *
   8  * IDENTIFICATION
   9  *        $PostgreSQL$
  10  *
  11  *-------------------------------------------------------------------------
  12  */
  13
  14 #include "postgres.h"
  15
  16 #include "access/clog.h"
  17 #include "access/subtrans.h"
  18 #include "access/transam.h"
  19 #include "miscadmin.h"
  20 #include "postmaster/autovacuum.h"
  21 #include "storage/pmsignal.h"
  22 #include "storage/proc.h"
  23 #include "utils/builtins.h"
  24
  25
  26 /* Number of OIDs to prefetch (preallocate) per XLOG write */
  27 #define VAR_OID_PREFETCH                8192
  28
  29 /* pointer to "variable cache" in shared memory (set up by shmem.c) */
  30 VariableCache ShmemVariableCache = NULL;
  31
  32
  33 /*
  34  * Allocate the next XID for my new transaction or subtransaction.
  35  *
  36  * The new XID is also stored into MyProc before returning.
  37  */
  38 TransactionId
  39 GetNewTransactionId(bool isSubXact)
  40 {
  41         TransactionId xid;
  42
  43         /*
  44          * During bootstrap initialization, we return the special bootstrap
  45          * transaction id.
  46          */
  47         if (IsBootstrapProcessingMode())
  48         {
  49                 Assert(!isSubXact);
  50                 MyProc->xid = BootstrapTransactionId;
  51                 return BootstrapTransactionId;
  52         }
  53
  54         LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
  55
  56         xid = ShmemVariableCache->nextXid;
  57
  58         /*----------
  59          * Check to see if it's safe to assign another XID.  This protects against
  60          * catastrophic data loss due to XID wraparound.  The basic rules are:
  61          *
  62          * If we're past xidVacLimit, start trying to force autovacuum cycles.
  63          * If we're past xidWarnLimit, start issuing warnings.
  64          * If we're past xidStopLimit, refuse to execute transactions, unless
  65          * we are running in a standalone backend (which gives an escape hatch
  66          * to the DBA who somehow got past the earlier defenses).
  67          *
  68          * Test is coded to fall out as fast as possible during normal operation,
  69          * ie, when the vac limit is set and we haven't violated it.
  70          *----------
  71          */
  72         if (TransactionIdFollowsOrEquals(xid, ShmemVariableCache->xidVacLimit) &&
  73                 TransactionIdIsValid(ShmemVariableCache->xidVacLimit))
  74         {
  75                 /*
  76                  * To avoid swamping the postmaster with signals, we issue the autovac
  77                  * request only once per 64K transaction starts.  This still gives
  78                  * plenty of chances before we get into real trouble.
  79                  */
  80                 if (IsUnderPostmaster && (xid % 65536) == 0)
  81                         SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
  82
  83                 if (IsUnderPostmaster &&
  84                  TransactionIdFollowsOrEquals(xid, ShmemVariableCache->xidStopLimit))
  85                         ereport(ERROR,
  86                                         (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
  87                                          errmsg("database is not accepting commands to avoid wraparound data loss in database \"%s\"",
  88                                                         NameStr(ShmemVariableCache->limit_datname)),
  89                                          errhint("Stop the postmaster and use a standalone backend to vacuum database \"%s\".\n"
  90                                                          "You might also need to commit or roll back old prepared transactions.",
  91                                                          NameStr(ShmemVariableCache->limit_datname))));
  92                 else if (TransactionIdFollowsOrEquals(xid, ShmemVariableCache->xidWarnLimit))
  93                         ereport(WARNING,
  94                         (errmsg("database \"%s\" must be vacuumed within %u transactions",
  95                                         NameStr(ShmemVariableCache->limit_datname),
  96                                         ShmemVariableCache->xidWrapLimit - xid),
  97                          errhint("To avoid a database shutdown, execute a database-wide VACUUM in \"%s\".\n"
  98                                          "You might also need to commit or roll back old prepared transactions.",
  99                                          NameStr(ShmemVariableCache->limit_datname))));
 100         }
 101
 102         /*
 103          * If we are allocating the first XID of a new page of the commit log,
 104          * zero out that commit-log page before returning. We must do this while
 105          * holding XidGenLock, else another xact could acquire and commit a later
 106          * XID before we zero the page.  Fortunately, a page of the commit log
 107          * holds 32K or more transactions, so we don't have to do this very often.
 108          *
 109          * Extend pg_subtrans too.
 110          */
 111         ExtendCLOG(xid);
 112         ExtendSUBTRANS(xid);
 113
 114         /*
 115          * Now advance the nextXid counter.  This must not happen until after we
 116          * have successfully completed ExtendCLOG() --- if that routine fails, we
 117          * want the next incoming transaction to try it again.  We cannot assign
 118          * more XIDs until there is CLOG space for them.
 119          */
 120         TransactionIdAdvance(ShmemVariableCache->nextXid);
 121
 122         /*
 123          * We must store the new XID into the shared ProcArray before releasing
 124          * XidGenLock.  This ensures that every active XID older than
 125          * latestCompletedXid is present in the ProcArray, which is essential for
 126          * correct OldestXmin tracking; see src/backend/access/transam/README.
 127          *
 128          * XXX by storing xid into MyProc without acquiring ProcArrayLock, we are
 129          * relying on fetch/store of an xid to be atomic, else other backends
 130          * might see a partially-set xid here.  But holding both locks at once
 131          * would be a nasty concurrency hit.  So for now, assume atomicity.
 132          *
 133          * Note that readers of PGPROC xid fields should be careful to fetch the
 134          * value only once, rather than assume they can read a value multiple
 135          * times and get the same answer each time.
 136          *
 137          * The same comments apply to the subxact xid count and overflow fields.
 138          *
 139          * A solution to the atomic-store problem would be to give each PGPROC its
 140          * own spinlock used only for fetching/storing that PGPROC's xid and
 141          * related fields.
 142          *
 143          * If there's no room to fit a subtransaction XID into PGPROC, set the
 144          * cache-overflowed flag instead.  This forces readers to look in
 145          * pg_subtrans to map subtransaction XIDs up to top-level XIDs. There is a
 146          * race-condition window, in that the new XID will not appear as running
 147          * until its parent link has been placed into pg_subtrans. However, that
 148          * will happen before anyone could possibly have a reason to inquire about
 149          * the status of the XID, so it seems OK.  (Snapshots taken during this
 150          * window *will* include the parent XID, so they will deliver the correct
 151          * answer later on when someone does have a reason to inquire.)
 152          */
 153         {
 154                 /*
 155                  * Use volatile pointer to prevent code rearrangement; other backends
 156                  * could be examining my subxids info concurrently, and we don't want
 157                  * them to see an invalid intermediate state, such as incrementing
 158                  * nxids before filling the array entry.  Note we are assuming that
 159                  * TransactionId and int fetch/store are atomic.
 160                  */
 161                 volatile PGPROC *myproc = MyProc;
 162
 163                 if (!isSubXact)
 164                         myproc->xid = xid;
 165                 else
 166                 {
 167                         int                     nxids = myproc->subxids.nxids;
 168
 169                         if (nxids < PGPROC_MAX_CACHED_SUBXIDS)
 170                         {
 171                                 myproc->subxids.xids[nxids] = xid;
 172                                 myproc->subxids.nxids = nxids + 1;
 173                         }
 174                         else
 175                                 myproc->subxids.overflowed = true;
 176                 }
 177         }
 178
 179         LWLockRelease(XidGenLock);
 180
 181         return xid;
 182 }
 183
 184 /*
 185  * Read nextXid but don't allocate it.
 186  */
 187 TransactionId
 188 ReadNewTransactionId(void)
 189 {
 190         TransactionId xid;
 191
 192         LWLockAcquire(XidGenLock, LW_SHARED);
 193         xid = ShmemVariableCache->nextXid;
 194         LWLockRelease(XidGenLock);
 195
 196         return xid;
 197 }
 198
 199 /*
 200  * Determine the last safe XID to allocate given the currently oldest
 201  * datfrozenxid (ie, the oldest XID that might exist in any database
 202  * of our cluster).
 203  */
 204 void
 205 SetTransactionIdLimit(TransactionId oldest_datfrozenxid,
 206                                           Name oldest_datname)
 207 {
 208         TransactionId xidVacLimit;
 209         TransactionId xidWarnLimit;
 210         TransactionId xidStopLimit;
 211         TransactionId xidWrapLimit;
 212         TransactionId curXid;
 213
 214         Assert(TransactionIdIsNormal(oldest_datfrozenxid));
 215
 216         /*
 217          * The place where we actually get into deep trouble is halfway around
 218          * from the oldest potentially-existing XID.  (This calculation is
 219          * probably off by one or two counts, because the special XIDs reduce the
 220          * size of the loop a little bit.  But we throw in plenty of slop below,
 221          * so it doesn't matter.)
 222          */
 223         xidWrapLimit = oldest_datfrozenxid + (MaxTransactionId >> 1);
 224         if (xidWrapLimit < FirstNormalTransactionId)
 225                 xidWrapLimit += FirstNormalTransactionId;
 226
 227         /*
 228          * We'll refuse to continue assigning XIDs in interactive mode once we get
 229          * within 1M transactions of data loss.  This leaves lots of room for the
 230          * DBA to fool around fixing things in a standalone backend, while not
 231          * being significant compared to total XID space. (Note that since
 232          * vacuuming requires one transaction per table cleaned, we had better be
 233          * sure there's lots of XIDs left...)
 234          */
 235         xidStopLimit = xidWrapLimit - 1000000;
 236         if (xidStopLimit < FirstNormalTransactionId)
 237                 xidStopLimit -= FirstNormalTransactionId;
 238
 239         /*
 240          * We'll start complaining loudly when we get within 10M transactions of
 241          * the stop point.      This is kind of arbitrary, but if you let your gas
 242          * gauge get down to 1% of full, would you be looking for the next gas
 243          * station?  We need to be fairly liberal about this number because there
 244          * are lots of scenarios where most transactions are done by automatic
 245          * clients that won't pay attention to warnings. (No, we're not gonna make
 246          * this configurable.  If you know enough to configure it, you know enough
 247          * to not get in this kind of trouble in the first place.)
 248          */
 249         xidWarnLimit = xidStopLimit - 10000000;
 250         if (xidWarnLimit < FirstNormalTransactionId)
 251                 xidWarnLimit -= FirstNormalTransactionId;
 252
 253         /*
 254          * We'll start trying to force autovacuums when oldest_datfrozenxid gets
 255          * to be more than autovacuum_freeze_max_age transactions old.
 256          *
 257          * Note: guc.c ensures that autovacuum_freeze_max_age is in a sane range,
 258          * so that xidVacLimit will be well before xidWarnLimit.
 259          *
 260          * Note: autovacuum_freeze_max_age is a PGC_POSTMASTER parameter so that
 261          * we don't have to worry about dealing with on-the-fly changes in its
 262          * value.  It doesn't look practical to update shared state from a GUC
 263          * assign hook (too many processes would try to execute the hook,
 264          * resulting in race conditions as well as crashes of those not connected
 265          * to shared memory).  Perhaps this can be improved someday.
 266          */
 267         xidVacLimit = oldest_datfrozenxid + autovacuum_freeze_max_age;
 268         if (xidVacLimit < FirstNormalTransactionId)
 269                 xidVacLimit += FirstNormalTransactionId;
 270
 271         /* Grab lock for just long enough to set the new limit values */
 272         LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
 273         ShmemVariableCache->oldestXid = oldest_datfrozenxid;
 274         ShmemVariableCache->xidVacLimit = xidVacLimit;
 275         ShmemVariableCache->xidWarnLimit = xidWarnLimit;
 276         ShmemVariableCache->xidStopLimit = xidStopLimit;
 277         ShmemVariableCache->xidWrapLimit = xidWrapLimit;
 278         namecpy(&ShmemVariableCache->limit_datname, oldest_datname);
 279         curXid = ShmemVariableCache->nextXid;
 280         LWLockRelease(XidGenLock);
 281
 282         /* Log the info */
 283         ereport(DEBUG1,
 284            (errmsg("transaction ID wrap limit is %u, limited by database \"%s\"",
 285                            xidWrapLimit, NameStr(*oldest_datname))));
 286
 287         /*
 288          * If past the autovacuum force point, immediately signal an autovac
 289          * request.  The reason for this is that autovac only processes one
 290          * database per invocation.  Once it's finished cleaning up the oldest
 291          * database, it'll call here, and we'll signal the postmaster to start
 292          * another iteration immediately if there are still any old databases.
 293          */
 294         if (TransactionIdFollowsOrEquals(curXid, xidVacLimit) &&
 295                 IsUnderPostmaster)
 296                 SendPostmasterSignal(PMSIGNAL_START_AUTOVAC_LAUNCHER);
 297
 298         /* Give an immediate warning if past the wrap warn point */
 299         if (TransactionIdFollowsOrEquals(curXid, xidWarnLimit))
 300                 ereport(WARNING,
 301                    (errmsg("database \"%s\" must be vacuumed within %u transactions",
 302                                    NameStr(*oldest_datname),
 303                                    xidWrapLimit - curXid),
 304                         errhint("To avoid a database shutdown, execute a database-wide VACUUM in \"%s\".\n"
 305                                         "You might also need to commit or roll back old prepared transactions.",
 306                                         NameStr(*oldest_datname))));
 307 }
 308
 309
 310 /*
 311  * GetNewObjectId -- allocate a new OID
 312  *
 313  * OIDs are generated by a cluster-wide counter.  Since they are only 32 bits
 314  * wide, counter wraparound will occur eventually, and therefore it is unwise
 315  * to assume they are unique unless precautions are taken to make them so.
 316  * Hence, this routine should generally not be used directly.  The only
 317  * direct callers should be GetNewOid() and GetNewRelFileNode() in
 318  * catalog/catalog.c.
 319  */
 320 Oid
 321 GetNewObjectId(void)
 322 {
 323         Oid                     result;
 324
 325         LWLockAcquire(OidGenLock, LW_EXCLUSIVE);
 326
 327         /*
 328          * Check for wraparound of the OID counter.  We *must* not return 0
 329          * (InvalidOid); and as long as we have to check that, it seems a good
 330          * idea to skip over everything below FirstNormalObjectId too. (This
 331          * basically just avoids lots of collisions with bootstrap-assigned OIDs
 332          * right after a wrap occurs, so as to avoid a possibly large number of
 333          * iterations in GetNewOid.)  Note we are relying on unsigned comparison.
 334          *
 335          * During initdb, we start the OID generator at FirstBootstrapObjectId, so
 336          * we only enforce wrapping to that point when in bootstrap or standalone
 337          * mode.  The first time through this routine after normal postmaster
 338          * start, the counter will be forced up to FirstNormalObjectId. This
 339          * mechanism leaves the OIDs between FirstBootstrapObjectId and
 340          * FirstNormalObjectId available for automatic assignment during initdb,
 341          * while ensuring they will never conflict with user-assigned OIDs.
 342          */
 343         if (ShmemVariableCache->nextOid < ((Oid) FirstNormalObjectId))
 344         {
 345                 if (IsPostmasterEnvironment)
 346                 {
 347                         /* wraparound in normal environment */
 348                         ShmemVariableCache->nextOid = FirstNormalObjectId;
 349                         ShmemVariableCache->oidCount = 0;
 350                 }
 351                 else
 352                 {
 353                         /* we may be bootstrapping, so don't enforce the full range */
 354                         if (ShmemVariableCache->nextOid < ((Oid) FirstBootstrapObjectId))
 355                         {
 356                                 /* wraparound in standalone environment? */
 357                                 ShmemVariableCache->nextOid = FirstBootstrapObjectId;
 358                                 ShmemVariableCache->oidCount = 0;
 359                         }
 360                 }
 361         }
 362
 363         /* If we run out of logged for use oids then we must log more */
 364         if (ShmemVariableCache->oidCount == 0)
 365         {
 366                 XLogPutNextOid(ShmemVariableCache->nextOid + VAR_OID_PREFETCH);
 367                 ShmemVariableCache->oidCount = VAR_OID_PREFETCH;
 368         }
 369
 370         result = ShmemVariableCache->nextOid;
 371
 372         (ShmemVariableCache->nextOid)++;
 373         (ShmemVariableCache->oidCount)--;
 374
 375         LWLockRelease(OidGenLock);
 376
 377         return result;
 378 }