Now that we've rearranged relation open to get a lock before touching
authorTom Lane
Fri, 18 Aug 2006 16:09:13 +0000 (16:09 +0000)
committerTom Lane
Fri, 18 Aug 2006 16:09:13 +0000 (16:09 +0000)
the rel, it's easy to get rid of the narrow race-condition window that
used to exist in VACUUM and CLUSTER.  Did some minor code-beautification
work in the same area, too.

src/backend/access/heap/heapam.c
src/backend/commands/analyze.c
src/backend/commands/cluster.c
src/backend/commands/lockcmds.c
src/backend/commands/vacuum.c
src/backend/storage/lmgr/lmgr.c
src/include/access/heapam.h
src/include/storage/lmgr.h

index f29407e8e5df1f4ab2b9e3e468d8a7ebbba2a0f2..759f0b1f13b05726a83e538a615609dbb2c9832f 100644 (file)
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/access/heap/heapam.c,v 1.218 2006/07/31 20:08:59 tgl Exp $
+ *   $PostgreSQL: pgsql/src/backend/access/heap/heapam.c,v 1.219 2006/08/18 16:09:08 tgl Exp $
  *
  *
  * INTERFACE ROUTINES
@@ -53,6 +53,7 @@
 #include "utils/inval.h"
 #include "utils/lsyscache.h"
 #include "utils/relcache.h"
+#include "utils/syscache.h"
 
 
 static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
@@ -702,14 +703,56 @@ relation_open(Oid relationId, LOCKMODE lockmode)
 }
 
 /* ----------------
- *     conditional_relation_open - open with option not to wait
+ *     try_relation_open - open any relation by relation OID
  *
- *     As above, but if nowait is true, then throw an error rather than
- *     waiting when the lock is not immediately obtainable.
+ *     Same as relation_open, except return NULL instead of failing
+ *     if the relation does not exist.
  * ----------------
  */
 Relation
-conditional_relation_open(Oid relationId, LOCKMODE lockmode, bool nowait)
+try_relation_open(Oid relationId, LOCKMODE lockmode)
+{
+   Relation    r;
+
+   Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
+
+   /* Get the lock first */
+   if (lockmode != NoLock)
+       LockRelationOid(relationId, lockmode);
+
+   /*
+    * Now that we have the lock, probe to see if the relation really
+    * exists or not.
+    */
+   if (!SearchSysCacheExists(RELOID,
+                             ObjectIdGetDatum(relationId),
+                             0, 0, 0))
+   {
+       /* Release useless lock */
+       if (lockmode != NoLock)
+           UnlockRelationOid(relationId, lockmode);
+
+       return NULL;
+   }
+
+   /* Should be safe to do a relcache load */
+   r = RelationIdGetRelation(relationId);
+
+   if (!RelationIsValid(r))
+       elog(ERROR, "could not open relation with OID %u", relationId);
+
+   return r;
+}
+
+/* ----------------
+ *     relation_open_nowait - open but don't wait for lock
+ *
+ *     Same as relation_open, except throw an error instead of waiting
+ *     when the requested lock is not immediately obtainable.
+ * ----------------
+ */
+Relation
+relation_open_nowait(Oid relationId, LOCKMODE lockmode)
 {
    Relation    r;
 
@@ -718,27 +761,22 @@ conditional_relation_open(Oid relationId, LOCKMODE lockmode, bool nowait)
    /* Get the lock before trying to open the relcache entry */
    if (lockmode != NoLock)
    {
-       if (nowait)
+       if (!ConditionalLockRelationOid(relationId, lockmode))
        {
-           if (!ConditionalLockRelationOid(relationId, lockmode))
-           {
-               /* try to throw error by name; relation could be deleted... */
-               char   *relname = get_rel_name(relationId);
-
-               if (relname)
-                   ereport(ERROR,
-                           (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
-                            errmsg("could not obtain lock on relation \"%s\"",
-                                   relname)));
-               else
-                   ereport(ERROR,
-                           (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
-                            errmsg("could not obtain lock on relation with OID %u",
-                                   relationId)));
-           }
+           /* try to throw error by name; relation could be deleted... */
+           char   *relname = get_rel_name(relationId);
+
+           if (relname)
+               ereport(ERROR,
+                       (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
+                        errmsg("could not obtain lock on relation \"%s\"",
+                               relname)));
+           else
+               ereport(ERROR,
+                       (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
+                        errmsg("could not obtain lock on relation with OID %u",
+                               relationId)));
        }
-       else
-           LockRelationOid(relationId, lockmode);
    }
 
    /* The relcache does all the real work... */
@@ -753,7 +791,7 @@ conditional_relation_open(Oid relationId, LOCKMODE lockmode, bool nowait)
 /* ----------------
  *     relation_openrv - open any relation specified by a RangeVar
  *
- *     As above, but the relation is specified by a RangeVar.
+ *     Same as relation_open, but the relation is specified by a RangeVar.
  * ----------------
  */
 Relation
index 34a38b20e297deafc29f2374e06764e486a110fb..2930eacb503d0841cb39f5ace4700505c5685d1a 100644 (file)
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/commands/analyze.c,v 1.96 2006/07/14 14:52:18 momjian Exp $
+ *   $PostgreSQL: pgsql/src/backend/commands/analyze.c,v 1.97 2006/08/18 16:09:08 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -129,20 +129,16 @@ analyze_rel(Oid relid, VacuumStmt *vacstmt)
    CHECK_FOR_INTERRUPTS();
 
    /*
-    * Race condition -- if the pg_class tuple has gone away since the last
-    * time we saw it, we don't need to process it.
+    * Open the relation, getting only a read lock on it.  If the rel has
+    * been dropped since we last saw it, we don't need to process it.
     */
-   if (!SearchSysCacheExists(RELOID,
-                             ObjectIdGetDatum(relid),
-                             0, 0, 0))
+   onerel = try_relation_open(relid, AccessShareLock);
+   if (!onerel)
        return;
 
    /*
-    * Open the class, getting only a read lock on it, and check permissions.
-    * Permissions check should match vacuum's check!
+    * Check permissions --- this should match vacuum's check!
     */
-   onerel = relation_open(relid, AccessShareLock);
-
    if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
          (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
    {
index f0cbaefd60efb9be48b2ce81ffeb641c5650dd29..bd2301dd62e95c4762324f57932535995b344d0b 100644 (file)
@@ -11,7 +11,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/commands/cluster.c,v 1.152 2006/07/31 20:09:00 tgl Exp $
+ *   $PostgreSQL: pgsql/src/backend/commands/cluster.c,v 1.153 2006/08/18 16:09:08 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -239,6 +239,18 @@ cluster_rel(RelToCluster *rvtc, bool recheck)
    /* Check for user-requested abort. */
    CHECK_FOR_INTERRUPTS();
 
+   /*
+    * We grab exclusive access to the target rel and index for the duration
+    * of the transaction.  (This is redundant for the single-transaction
+    * case, since cluster() already did it.)  The index lock is taken inside
+    * check_index_is_clusterable.
+    */
+   OldHeap = try_relation_open(rvtc->tableOid, AccessExclusiveLock);
+
+   /* If the table has gone away, we can skip processing it */
+   if (!OldHeap)
+       return;
+
    /*
     * Since we may open a new transaction for each relation, we have to check
     * that the relation still is what we think it is.
@@ -252,20 +264,23 @@ cluster_rel(RelToCluster *rvtc, bool recheck)
        HeapTuple   tuple;
        Form_pg_index indexForm;
 
+       /* Check that the user still owns the relation */
+       if (!pg_class_ownercheck(rvtc->tableOid, GetUserId()))
+       {
+           relation_close(OldHeap, AccessExclusiveLock);
+           return;
+       }
+
        /*
-        * Check if the relation and index still exist before opening them
+        * Check that the index still exists
         */
        if (!SearchSysCacheExists(RELOID,
-                                 ObjectIdGetDatum(rvtc->tableOid),
-                                 0, 0, 0) ||
-           !SearchSysCacheExists(RELOID,
                                  ObjectIdGetDatum(rvtc->indexOid),
                                  0, 0, 0))
+       {
+           relation_close(OldHeap, AccessExclusiveLock);
            return;
-
-       /* Check that the user still owns the relation */
-       if (!pg_class_ownercheck(rvtc->tableOid, GetUserId()))
-           return;
+       }
 
        /*
         * Check that the index is still the one with indisclustered set.
@@ -273,25 +288,21 @@ cluster_rel(RelToCluster *rvtc, bool recheck)
        tuple = SearchSysCache(INDEXRELID,
                               ObjectIdGetDatum(rvtc->indexOid),
                               0, 0, 0);
-       if (!HeapTupleIsValid(tuple))
-           return;             /* could have gone away... */
+       if (!HeapTupleIsValid(tuple))       /* probably can't happen */
+       {
+           relation_close(OldHeap, AccessExclusiveLock);
+           return;
+       }
        indexForm = (Form_pg_index) GETSTRUCT(tuple);
        if (!indexForm->indisclustered)
        {
            ReleaseSysCache(tuple);
+           relation_close(OldHeap, AccessExclusiveLock);
            return;
        }
        ReleaseSysCache(tuple);
    }
 
-   /*
-    * We grab exclusive access to the target rel and index for the duration
-    * of the transaction.  (This is redundant for the single- transaction
-    * case, since cluster() already did it.)  The index lock is taken inside
-    * check_index_is_clusterable.
-    */
-   OldHeap = heap_open(rvtc->tableOid, AccessExclusiveLock);
-
    /* Check index is valid to cluster on */
    check_index_is_clusterable(OldHeap, rvtc->indexOid, recheck);
 
index a84b1b2cdede25f10df9e850e2c6e1b27f78ce38..a5e10b793e6af6b3ad765a842aa96cc32a17d45c 100644 (file)
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/commands/lockcmds.c,v 1.14 2006/03/05 15:58:24 momjian Exp $
+ *   $PostgreSQL: pgsql/src/backend/commands/lockcmds.c,v 1.15 2006/08/18 16:09:08 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -59,7 +59,10 @@ LockTableCommand(LockStmt *lockstmt)
            aclcheck_error(aclresult, ACL_KIND_CLASS,
                           get_rel_name(reloid));
 
-       rel = conditional_relation_open(reloid, lockstmt->mode, lockstmt->nowait);
+       if (lockstmt->nowait)
+           rel = relation_open_nowait(reloid, lockstmt->mode);
+       else
+           rel = relation_open(reloid, lockstmt->mode);
 
        /* Currently, we only allow plain tables to be locked */
        if (rel->rd_rel->relkind != RELKIND_RELATION)
index c21be2f4788e4b67b82925b12a8d3f9e0aa19de2..8c1df23d9b094a03742668cdc4323e42cd92f3d1 100644 (file)
@@ -13,7 +13,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.337 2006/07/31 20:09:00 tgl Exp $
+ *   $PostgreSQL: pgsql/src/backend/commands/vacuum.c,v 1.338 2006/08/18 16:09:08 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -1041,31 +1041,12 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
        MyProc->inVacuum = true;
    }
 
-   /*
-    * Tell the cache replacement strategy that vacuum is causing all
-    * following IO
-    */
-   StrategyHintVacuum(true);
-
    /*
     * Check for user-requested abort.  Note we want this to be inside a
     * transaction, so xact.c doesn't issue useless WARNING.
     */
    CHECK_FOR_INTERRUPTS();
 
-   /*
-    * Race condition -- if the pg_class tuple has gone away since the last
-    * time we saw it, we don't need to vacuum it.
-    */
-   if (!SearchSysCacheExists(RELOID,
-                             ObjectIdGetDatum(relid),
-                             0, 0, 0))
-   {
-       StrategyHintVacuum(false);
-       CommitTransactionCommand();
-       return;
-   }
-
    /*
     * Determine the type of lock we want --- hard exclusive lock for a FULL
     * vacuum, but just ShareUpdateExclusiveLock for concurrent vacuum. Either
@@ -1074,7 +1055,21 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
    lmode = vacstmt->full ? AccessExclusiveLock : ShareUpdateExclusiveLock;
 
    /*
-    * Open the class, get an appropriate lock on it, and check permissions.
+    * Open the relation and get the appropriate lock on it.
+    *
+    * There's a race condition here: the rel may have gone away since
+    * the last time we saw it.  If so, we don't need to vacuum it.
+    */
+   onerel = try_relation_open(relid, lmode);
+
+   if (!onerel)
+   {
+       CommitTransactionCommand();
+       return;
+   }
+
+   /*
+    * Check permissions.
     *
     * We allow the user to vacuum a table if he is superuser, the table
     * owner, or the database owner (but in the latter case, only if it's not
@@ -1083,8 +1078,6 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
     * Note we choose to treat permissions failure as a WARNING and keep
     * trying to vacuum the rest of the DB --- is this appropriate?
     */
-   onerel = relation_open(relid, lmode);
-
    if (!(pg_class_ownercheck(RelationGetRelid(onerel), GetUserId()) ||
          (pg_database_ownercheck(MyDatabaseId, GetUserId()) && !onerel->rd_rel->relisshared)))
    {
@@ -1092,7 +1085,6 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
                (errmsg("skipping \"%s\" --- only table or database owner can vacuum it",
                        RelationGetRelationName(onerel))));
        relation_close(onerel, lmode);
-       StrategyHintVacuum(false);
        CommitTransactionCommand();
        return;
    }
@@ -1107,7 +1099,6 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
                (errmsg("skipping \"%s\" --- cannot vacuum indexes, views, or special system tables",
                        RelationGetRelationName(onerel))));
        relation_close(onerel, lmode);
-       StrategyHintVacuum(false);
        CommitTransactionCommand();
        return;
    }
@@ -1122,7 +1113,6 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
    if (isOtherTempNamespace(RelationGetNamespace(onerel)))
    {
        relation_close(onerel, lmode);
-       StrategyHintVacuum(false);
        CommitTransactionCommand();
        return;         /* assume no long-lived data in temp tables */
    }
@@ -1145,6 +1135,12 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
     */
    toast_relid = onerel->rd_rel->reltoastrelid;
 
+   /*
+    * Tell the cache replacement strategy that vacuum is causing all
+    * following IO
+    */
+   StrategyHintVacuum(true);
+
    /*
     * Do the actual work --- either FULL or "lazy" vacuum
     */
@@ -1153,13 +1149,14 @@ vacuum_rel(Oid relid, VacuumStmt *vacstmt, char expected_relkind)
    else
        lazy_vacuum_rel(onerel, vacstmt);
 
+   StrategyHintVacuum(false);
+
    /* all done with this class, but hold lock until commit */
    relation_close(onerel, NoLock);
 
    /*
     * Complete the transaction and free all temporary memory used.
     */
-   StrategyHintVacuum(false);
    CommitTransactionCommand();
 
    /*
index ad7ee3e6013621aaf528e0e3f7b0adefc801cdb4..4442d3909f04d978ba906624a9c2b1cc14633574 100644 (file)
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *   $PostgreSQL: pgsql/src/backend/storage/lmgr/lmgr.c,v 1.86 2006/07/31 20:09:05 tgl Exp $
+ *   $PostgreSQL: pgsql/src/backend/storage/lmgr/lmgr.c,v 1.87 2006/08/18 16:09:09 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -128,8 +128,8 @@ ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode)
 /*
  *     UnlockRelationId
  *
- * Note: we don't supply UnlockRelationOid since it's normally easy for
- * callers to provide the LockRelId info from a relcache entry.
+ * Unlock, given a LockRelId.  This is preferred over UnlockRelationOid
+ * for speed reasons.
  */
 void
 UnlockRelationId(LockRelId *relid, LOCKMODE lockmode)
@@ -141,6 +141,21 @@ UnlockRelationId(LockRelId *relid, LOCKMODE lockmode)
    LockRelease(&tag, lockmode, false);
 }
 
+/*
+ *     UnlockRelationOid
+ *
+ * Unlock, given only a relation Oid.  Use UnlockRelationId if you can.
+ */
+void
+UnlockRelationOid(Oid relid, LOCKMODE lockmode)
+{
+   LOCKTAG     tag;
+
+   SetLocktagRelationOid(&tag, relid);
+
+   LockRelease(&tag, lockmode, false);
+}
+
 /*
  *     LockRelation
  *
index a2ca20bddd9ca9b933da018102faca2106bb5aaa..772a6588faefbeb7d4ba36527c0f97ea51ed2b73 100644 (file)
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2006, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/access/heapam.h,v 1.114 2006/07/03 22:45:39 tgl Exp $
+ * $PostgreSQL: pgsql/src/include/access/heapam.h,v 1.115 2006/08/18 16:09:10 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -129,7 +129,8 @@ typedef enum
 } LockTupleMode;
 
 extern Relation relation_open(Oid relationId, LOCKMODE lockmode);
-extern Relation conditional_relation_open(Oid relationId, LOCKMODE lockmode, bool nowait);
+extern Relation try_relation_open(Oid relationId, LOCKMODE lockmode);
+extern Relation relation_open_nowait(Oid relationId, LOCKMODE lockmode);
 extern Relation relation_openrv(const RangeVar *relation, LOCKMODE lockmode);
 extern void relation_close(Relation relation, LOCKMODE lockmode);
 
index d0f9ba2b9c646748a66e6749501405cb611ef7d2..ea7a1bf948f0942815c3835098a5255241d7b7c7 100644 (file)
@@ -7,7 +7,7 @@
  * Portions Copyright (c) 1996-2006, PostgreSQL Global Development Group
  * Portions Copyright (c) 1994, Regents of the University of California
  *
- * $PostgreSQL: pgsql/src/include/storage/lmgr.h,v 1.55 2006/07/31 20:09:05 tgl Exp $
+ * $PostgreSQL: pgsql/src/include/storage/lmgr.h,v 1.56 2006/08/18 16:09:13 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -24,6 +24,7 @@ extern void RelationInitLockInfo(Relation relation);
 extern void LockRelationOid(Oid relid, LOCKMODE lockmode);
 extern bool ConditionalLockRelationOid(Oid relid, LOCKMODE lockmode);
 extern void UnlockRelationId(LockRelId *relid, LOCKMODE lockmode);
+extern void UnlockRelationOid(Oid relid, LOCKMODE lockmode);
 
 extern void LockRelation(Relation relation, LOCKMODE lockmode);
 extern bool ConditionalLockRelation(Relation relation, LOCKMODE lockmode);