]> granicus.if.org Git - postgresql/commitdiff
Make the overflow guards in ExecChooseHashTableSize be more protective.
authorTom Lane <tgl@sss.pgh.pa.us>
Fri, 30 Oct 2009 20:58:45 +0000 (20:58 +0000)
committerTom Lane <tgl@sss.pgh.pa.us>
Fri, 30 Oct 2009 20:58:45 +0000 (20:58 +0000)
The original coding ensured nbuckets and nbatch didn't exceed INT_MAX,
which while not insane on its own terms did nothing to protect subsequent
code like "palloc(nbatch * sizeof(BufFile *))".  Since enormous join size
estimates might well be planner error rather than reality, it seems best
to constrain the initial sizes to be not more than work_mem/sizeof(pointer),
thus ensuring the allocated arrays don't exceed work_mem.  We will allow
nbatch to get bigger than that during subsequent ExecHashIncreaseNumBatches
calls, but we should still guard against integer overflow in those palloc
requests.  Per bug #5145 from Bernt Marius Johnsen.

Although the given test case only seems to fail back to 8.2, previous
releases have variants of this issue, so patch all supported branches.

src/backend/executor/nodeHash.c

index 2e6a77d89241baf03818e8833cdc96f784d17952..16b84d0f5ad3392ddd3f6738dac61a683ce45fb1 100644 (file)
@@ -8,7 +8,7 @@
  *
  *
  * IDENTIFICATION
- *       $PostgreSQL: pgsql/src/backend/executor/nodeHash.c,v 1.122 2009/09/27 21:10:53 tgl Exp $
+ *       $PostgreSQL: pgsql/src/backend/executor/nodeHash.c,v 1.123 2009/10/30 20:58:45 tgl Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -394,6 +394,7 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
        double          inner_rel_bytes;
        long            hash_table_bytes;
        long            skew_table_bytes;
+       long            max_pointers;
        int                     nbatch;
        int                     nbuckets;
        int                     i;
@@ -435,17 +436,18 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
        {
                skew_table_bytes = hash_table_bytes * SKEW_WORK_MEM_PERCENT / 100;
 
-               *num_skew_mcvs = skew_table_bytes / (
-               /* size of a hash tuple */
-                                                                                        tupsize +
-               /* worst-case size of skewBucket[] per MCV */
+               /*----------
+                * Divisor is:
+                * size of a hash tuple +
+                * worst-case size of skewBucket[] per MCV +
+                * size of skewBucketNums[] entry +
+                * size of skew bucket struct itself
+                *----------
+                */
+               *num_skew_mcvs = skew_table_bytes / (tupsize +
                                                                                         (8 * sizeof(HashSkewBucket *)) +
-               /* size of skewBucketNums[] entry */
                                                                                         sizeof(int) +
-               /* size of skew bucket struct itself */
-                                                                                        SKEW_BUCKET_OVERHEAD
-                       );
-
+                                                                                        SKEW_BUCKET_OVERHEAD);
                if (*num_skew_mcvs > 0)
                        hash_table_bytes -= skew_table_bytes;
        }
@@ -455,8 +457,13 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
        /*
         * Set nbuckets to achieve an average bucket load of NTUP_PER_BUCKET when
         * memory is filled.  Set nbatch to the smallest power of 2 that appears
-        * sufficient.
+        * sufficient.  The Min() steps limit the results so that the pointer
+        * arrays we'll try to allocate do not exceed work_mem.
         */
+       max_pointers = (work_mem * 1024L) / sizeof(void *);
+       /* also ensure we avoid integer overflow in nbatch and nbuckets */
+       max_pointers = Min(max_pointers, INT_MAX / 2);
+
        if (inner_rel_bytes > hash_table_bytes)
        {
                /* We'll need multiple batches */
@@ -465,11 +472,11 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
                int                     minbatch;
 
                lbuckets = (hash_table_bytes / tupsize) / NTUP_PER_BUCKET;
-               lbuckets = Min(lbuckets, INT_MAX / 2);
+               lbuckets = Min(lbuckets, max_pointers);
                nbuckets = (int) lbuckets;
 
                dbatch = ceil(inner_rel_bytes / hash_table_bytes);
-               dbatch = Min(dbatch, INT_MAX / 2);
+               dbatch = Min(dbatch, max_pointers);
                minbatch = (int) dbatch;
                nbatch = 2;
                while (nbatch < minbatch)
@@ -481,7 +488,7 @@ ExecChooseHashTableSize(double ntuples, int tupwidth, bool useskew,
                double          dbuckets;
 
                dbuckets = ceil(ntuples / NTUP_PER_BUCKET);
-               dbuckets = Min(dbuckets, INT_MAX / 2);
+               dbuckets = Min(dbuckets, max_pointers);
                nbuckets = (int) dbuckets;
 
                nbatch = 1;
@@ -555,7 +562,7 @@ ExecHashIncreaseNumBatches(HashJoinTable hashtable)
                return;
 
        /* safety check to avoid overflow */
-       if (oldnbatch > INT_MAX / 2)
+       if (oldnbatch > Min(INT_MAX / 2, MaxAllocSize / (sizeof(void *) * 2)))
                return;
 
        nbatch = oldnbatch * 2;
@@ -1033,9 +1040,9 @@ ExecHashBuildSkewHash(HashJoinTable hashtable, Hash *node, int mcvsToUse)
                 * will be at least one null entry, so searches will always
                 * terminate.)
                 *
-                * Note: this code could fail if mcvsToUse exceeds INT_MAX/8, but that
-                * is not currently possible since we limit pg_statistic entries to
-                * much less than that.
+                * Note: this code could fail if mcvsToUse exceeds INT_MAX/8 or
+                * MaxAllocSize/sizeof(void *)/8, but that is not currently possible
+                * since we limit pg_statistic entries to much less than that.
                 */
                nbuckets = 2;
                while (nbuckets <= mcvsToUse)