Resolver in glibc 2.2

[Mark Kettenis <kettenis at wins dot uva dot nl>]

The resolver is a bit out of shape at the moment.  Ulrich already
mentioned that the per-thread _res variable doesn't really work right
now:

  http://sources.redhat.com/ml/libc-hacker/2000-06/msg00073.html

and there are some other issues too.  I've started to look into these
matters, and I'm planning on spending some more time on it, to bring
things in working order.

The primary goal will still be to provide a fully backwards compatible
libresolv.so for single-threaded programs, while providing the
necessary infrastructure for a thread-safe libnss_dns.

However, I think that the resolver in a multi-threaded program should
behave as if one had linked with the libbind from BIND 8.  This is
somewhat differently from how the resolver in glibc 2.2 behaves right
now:

 * res_init() won't affect the per-thread _res, but instead a static
   variable, just like the res_mkquery(), res_send(), etc.  Right now
   res_init() sets the per-thread _res, which isn't consistent.

 * _res will be per-thread, and will be properly initialized in
   multi-threaded programs (as if res_ninit() had been called on it).

I think this means that I can undo some of the ad-hoc patches that
have been applied over the last few months.

I'll also try to update to the latest and greatest BIND 8.2.3-T5B
(which is supposed to be a release candidate).

Mark