Broken autoconf mmap test

[Andy Koppe]

On 25 March 2011 10:17, Corinna Vinschen wrote:
>> If you guys think that cheating for autoconf on 64 bit systems is a good
>> idea, I'm willing to implement it.
>
> Here's a patch which works fine for me:
>
> Index: mmap.cc
> ===================================================================
> RCS file: /cvs/src/src/winsup/cygwin/mmap.cc,v
> retrieving revision 1.166
> diff -u -p -r1.166 mmap.cc
> --- mmap.cc     18 Mar 2011 13:56:56 -0000      1.166
> +++ mmap.cc     25 Mar 2011 10:17:09 -0000
> @@ -801,6 +801,38 @@ mmap64 (void *addr, size_t len, int prot
>       /* mmap /dev/zero is like MAP_ANONYMOUS. */
>       if (fh->get_device () == FH_ZERO)
>        flags |= MAP_ANONYMOUS;
> +
> +      /* The autoconf mmap test maps a file of size 1 byte.  It then tests
> +        every byte of the entire mapped page of 64K for 0-bytes since that's
> +        what POSIX requires.  The problem is, we can't create that mapping on
> +        64 bit systems.  The file mapping will be only a single page, 4K, and
> +        since 64 bit systems don't support the AT_ROUND_TO_PAGE flag, the
> +        remainder of the 64K slot will result in a SEGV when accessed.
> +
> +        So, what we do here is cheating for autoconf.  We know exactly what
> +        autoconf is doing.  The file is called conftest.txt, it has a size of
> +        1 byte, the requested mapping size is 64K, the requested protection
> +        is PROT_READ | PROT_WRITE, the mapping is MAP_SHARED, the offset is 0.
> +
> +        If all these requirements are given, we just return an anonymous map
> +        of 64K.  This will help to get over the autoconf test even on 64 bit
> +        systems. */

This should mention why cheating seems justified, i.e. the assumption
that "real" programs don't depend on being able to access past EOF.

Andy