strace of the following program on Linux x86 (4096 bytes/page) shows that there
are lots of redundant calls to mremap() (old size and new size are equal).  The
bug is present in glibc up to and including 2.6.1 (bug report Version field
doesn't have 2.6).

#include <stdlib.h>

int
main()
{
  const size_t size = 256 * 1024 + 4050;
  void *p;
  int i;

  p = malloc(size);
  for (i = 0; i < 100; ++i)
    p = realloc(p, size + i);

  return 0;
}


The following patch fixes the problem (comment line is copied from the similar
check in _int_realloc()):

--- malloc/malloc.c-orig        2007-09-26 13:28:36.000000000 +0400
+++ malloc/malloc.c     2007-09-26 12:57:41.000000000 +0400
@@ -3503,6 +3503,10 @@ mremap_chunk(p, new_size) mchunkptr p; s
   /* Note the extra SIZE_SZ overhead as in mmap_chunk(). */
   new_size = (new_size + offset + SIZE_SZ + page_mask) & ~page_mask;

+  /* don't need to remap if still within same page */
+  if (size + offset == new_size)
+    return p;
+
   cp = (char *)mremap((char *)p - offset, size + offset, new_size,
                       MREMAP_MAYMOVE);