Re: [bug-gnulib] mmapping of /dev/zero always fails on darwin

[Bruno Haible]

Peter O'Gorman wrote:
> As you can see from this little snippet of gcc configure, you can't mmap 
> /dev/zero on darwin.
> 
>     AC_CACHE_CHECK([whether mmap from /dev/zero works],
>    gcc_cv_func_mmap_dev_zero,
>    [# Add a system to this blacklist if it has mmap() but /dev/zero
>     # does not exist, or if mmapping /dev/zero does not give anonymous
>     # zeroed pages with both the following properties:
>     # 1. If you map N consecutive pages in with one call, and then
>     #    unmap any subset of those pages, the pages that were not
>     #    explicitly unmapped remain accessible.
>     # 2. If you map two adjacent blocks of memory and then unmap them
>     #    both at once, they must both go away.
>     # Systems known to be in this category are Windows (all variants),
>     # VMS, and Darwin.
>     case "$host_os" in
>       vms* | cygwin* | pe | mingw* | darwin* | ultrix* | hpux10* | hpux11.00)
>          gcc_cv_func_mmap_dev_zero=no ;;
>       *)
>          gcc_cv_func_mmap_dev_zero=yes;;
>     esac])

But gnulib's pagealign_alloc is not doing case 1 and not doing case 2.
It calls mmap() on N pages at once, and then ultimately munmap() on the
same N pages at once.

Also, GNU clisp uses mmap() with more complicated cases than this, including
mprotect() and partial munmap()s, and it works fine on Darwin.

> I just tried to use cvs-1.12.12 on Mac OS X, and it always dies. I traced 
> the problem to pagealign_alloc.c, which attempts to mmap /dev/zero.

How did you track that down to pagealign_alloc.c?

I'd rather guess that mmap() on MacOS X works fine but either
  - pagealign_alloc.c has a bug, or
  - CVS has some memory related bugs, and you don't happen to see them
    when pagealign_alloc uses malloc() instead of mmap() - because of the
    way malloc() internally does its bookkeeping.

Can you exclude these?

Bruno