Problem reported by Kevin Cernekee in:
http://lists.gnu.org/archive/html/bug-gnulib/2015-02/msg00112.html
* doc/glibc-functions/getdtablesize.texi (getdtablesize): Mention bug.
* lib/getdtablesize.c (getdtablesize): Don't fall back on _SC_OPEN_MAX.
Instead, just use getrlimit, taking care to avoid Cygwin bug.
dup2, fcntl: cross-compile better for Android
Problem reported by Kevin Cernekee in:
http://lists.gnu.org/archive/html/bug-gnulib/2015-02/msg00092.html
* doc/glibc-functions/getdtablesize.texi (getdtablesize):
Mention that getdtablesize doesn't work on Android.
* lib/getdtablesize.c: Use getrlimit substitute only if
getdtablesize is declared. This should suffice for Cygwin
while not breaking Android.
* m4/dup2.m4 (gl_FUNC_DUP2):
* m4/fcntl.m4 (gl_FUNC_FCNTL):
Prefer sysconf (_SC_OPEN_MAX) to getdtablesize, as the former is
standardized but the latter is not, and sysconf works on Android.
* m4/getdtablesize.m4 (gl_FUNC_GETDTABLESIZE):
Also check that getdtablesize is declared.
This removes the need for a special case for Android.
Cygwin 1.7.25 has a bug that even though it claims to support
RLIMIT_NOFILE inheritance, there is no enforcement of the soft
limit, and getdtablesize() automatically grows until it reaches
the hard limit which cannot be changed by setrlimit(). Best is
to just treat things as an invariant limit, as several other
modules assume that getdtablesize() will not change without an
intervening setrlimit().
* m4/getdtablesize.m4 (gl_FUNC_GETDTABLESIZE): Detect problem.
* modules/getdtablesize (configure.ac): Build replacement.
* m4/unistd_h.m4 (gl_UNISTD_H_DEFAULTS): Set up a witness.
* modules/unistd (Makefile.am): Expose the witness.
* lib/unistd.in.h (getdtablesize): Declare replacement.
* lib/getdtablesize.c (rpl_getdtablesize): Work around it.
* tests/test-getdtablesize.c (main): Test it.
* doc/glibc-functions/getdtablesize.texi (getdtablesize): Document it.
Signed-off-by: Eric Blake <eblake@redhat.com>
