summaryrefslogtreecommitdiff
path: root/include/my_bitmap.h
diff options
context:
space:
mode:
authorJon Olav Hauglid <jon.hauglid@oracle.com>2011-02-16 16:26:19 +0100
committerJon Olav Hauglid <jon.hauglid@oracle.com>2011-02-16 16:26:19 +0100
commit6ad0c9b16daccb64ec34dffc8a21e73ca9c6073d (patch)
treeaa203e6df8aca6f3e7ded96d19bf13beff263790 /include/my_bitmap.h
parent9c89cca5e364310928bde10287a56128378c107c (diff)
downloadmariadb-git-6ad0c9b16daccb64ec34dffc8a21e73ca9c6073d.tar.gz
Bug #11752069 (former bug 43152)
Assertion `bitmap_is_set_all(&table->s->all_set)' failed in handler::ha_reset This assertion could be triggered if two connections simultaneously executed two bitmap test functions on the same bitmap. For example, the assertion could be triggered if one connection executed UPDATE while a second connection executed SELECT on the same table. Even if bitmap test functions have read-only semantics and have const bitmaps as parameter, several of them modified the internal state of the bitmap. With interleaved execution of two such functions it was possible for one function to modify the state of the same bitmap that the other function had just modified. This lead to an inconsistent state and could trigger the assert. Internally the bitmap uses 32 bit words for storage. Since bitmaps can contain any number of bits, the last word in the bitmap may not be fully used. A 32 bit mask is maintained where a bit is set if the corresponding bit in the last bitmap word is unused. The problem was that several test functions applied this mask to the last word. Sometimes the mask was negated and used to zero out the remainder of the last word and sometimes the mask was used as-is to fill the remainder of the last word with 1's. This meant that if a function first used the negated mask and another function then used the mask as-is (or vice-versa), the first function would then get the wrong result. This patch fixes the problem by changing the implementation of 9 bitmap functions that modified the bitmap state even if the bitmap was declared const. These functions now preserve the internal state of the bitmap. This makes it possible for two connections to concurrently execute two of these functions on the same bitmap without issues. The patch also removes dead testing code from my_bitmap.c. These tests have already been moved to unittest/mysys/bitmap-t.c. Existing test coverage of my_bitmap has been extended. No MTR test case added as this would require adding several sync points to the bitmap functions. The patch has been tested with a non-deterministic test case posted on the bug report. include/my_bit.h: Removed my_count_bits_ushort() which is not needed anymore. Added my_count_bits_uint32(). unittest/mysys/bitmap-t.c: Extended test coverage of my_bitmap.
Diffstat (limited to 'include/my_bitmap.h')
-rw-r--r--include/my_bitmap.h9
1 files changed, 5 insertions, 4 deletions
diff --git a/include/my_bitmap.h b/include/my_bitmap.h
index ab69b2d671d..42f985c8918 100644
--- a/include/my_bitmap.h
+++ b/include/my_bitmap.h
@@ -1,4 +1,4 @@
-/* Copyright (C) 2000 MySQL AB
+/* Copyright (c) 2000, 2011, Oracle and/or its affiliates. All rights reserved.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@@ -149,9 +149,10 @@ bitmap_is_set(const MY_BITMAP *map,uint bit)
static inline my_bool bitmap_cmp(const MY_BITMAP *map1, const MY_BITMAP *map2)
{
- *(map1)->last_word_ptr|= (map1)->last_word_mask;
- *(map2)->last_word_ptr|= (map2)->last_word_mask;
- return memcmp((map1)->bitmap, (map2)->bitmap, 4*no_words_in_map((map1)))==0;
+ if (memcmp(map1->bitmap, map2->bitmap, 4*(no_words_in_map(map1)-1)) != 0)
+ return FALSE;
+ return ((*map1->last_word_ptr | map1->last_word_mask) ==
+ (*map2->last_word_ptr | map2->last_word_mask));
}
#define bitmap_clear_all(MAP) \