memleak in bitset_realloc()

[Minh Van Nguyen]

This is part of the larger project at #7656. See ​this thread for some background information.

The relevant module is

sage/misc/bitset.pxi

and the function concerned is bitset_realloc(). A memleak occurs when we have a bitset of positive size (or capacity) and then reallocate the size to be zero. To pick up the memleak, I first loaded Sage 4.6.1.alpha1 under valgrind and quit Sage without doing any computation. The resulting memleak summary is:

==26274== LEAK SUMMARY:
==26274==    definitely lost: 80 bytes in 3 blocks.
==26274==    indirectly lost: 240 bytes in 10 blocks.
==26274==      possibly lost: 562,324 bytes in 1,305 blocks.
==26274==    still reachable: 60,547,159 bytes in 41,676 blocks.
==26274==         suppressed: 0 bytes in 0 blocks.
==26274== Reachable blocks (those to which a pointer was found) are not shown.
==26274== To see them, rerun with: --leak-check=full --show-reachable=yes

I then applied the following patch to the Sage library, see test.patch​:

module_list.py

@@ -978 +978 @@
     Extension('sage.misc.session',
               sources = ['sage/misc/session.pyx']),
 
+    Extension('sage.misc.test',
+              sources = ['sage/misc/test.pyx']),
+
     ################################
     ##
     ## sage.modular

new file sage/misc/test.pyx

@@ +1 @@
+include "../ext/stdsage.pxi"
+include "bitset_pxd.pxi"
+include "bitset.pxi"
+
+def test():
+    cdef bitset_t a
+    bitset_init(a, 10)
+    try:
+        bitset_realloc(a, 0)
+    except MemoryError:
+        pass
+    bitset_free(a)

I loaded Sage under valgrind again and performed the following computation:

sage: from sage.misc.test import test
sage: test()

which resulted in the following memleak summary:

==16502== LEAK SUMMARY:
==16502==    definitely lost: 88 bytes in 4 blocks.
==16502==    indirectly lost: 240 bytes in 10 blocks.
==16502==      possibly lost: 563,732 bytes in 1,308 blocks.
==16502==    still reachable: 60,550,541 bytes in 41,691 blocks.
==16502==         suppressed: 0 bytes in 0 blocks.
==16502== Reachable blocks (those to which a pointer was found) are not shown.
==16502== To see them, rerun with: --leak-check=full --show-reachable=yes

Notice that despite the function test() explicitly freeing memory, the report on "definitely lost" shows that we leaked an extra 8 bytes (compare the 80 and 88 above in the first and second memleak summaries, respectively). Finally, I applied the following patch to the Sage library:

sage/misc/bitset.pxi

@@ -52 +52 @@
     cdef unsigned long size_old = bits.size
     if size_old == size: return 0
     bits.limbs = (size - 1)/(8*sizeof(unsigned long)) + 1
-    bits.bits = <unsigned long*>sage_realloc(bits.bits, bits.limbs * sizeof(unsigned long))
-    if bits.bits == NULL:
+    tmp = <unsigned long*>sage_realloc(bits.bits, bits.limbs * sizeof(unsigned long))
+    if tmp != NULL:
+        bits.bits = tmp
+    else:
         bits.limbs = limbs_old
         raise MemoryError
     bits.size = size

I loaded Sage under valgrind a third time and performed the following computation, same as above:

sage: from sage.misc.test import test
sage: test()

The corresponding memleak summary is:

==22029== LEAK SUMMARY:
==22029==    definitely lost: 80 bytes in 3 blocks.
==22029==    indirectly lost: 240 bytes in 10 blocks.
==22029==      possibly lost: 564,252 bytes in 1,309 blocks.
==22029==    still reachable: 60,550,168 bytes in 41,696 blocks.
==22029==         suppressed: 0 bytes in 0 blocks.
==22029== Reachable blocks (those to which a pointer was found) are not shown.
==22029== To see them, rerun with: --leak-check=full --show-reachable=yes

Notice that the summary for "definitely lost" now shows 80 bytes, the same as for the first memleak summary above. Thus the patch memleak.patch fixes the memleak in sage.misc.bitset.bitset_realloc.

Apply:

1. trac-10287_memleak.patch​