This reverts commit 196ebd4f.
It was buggy, and I can't simply recover the intended effect.

A down-side is that validation can now modify the validated RRset
on success.  I checked all transitive call sites that it's OK.
The change is pretty simple; I just hand-tested it a bit with faketime.

It's fairly easy to keep keep compatible with both 2.8 and 2.9,
so I'd go for that for now, as it may be practical.

As kresd works now, typically we do not know whether these records are
bogus, as with +cd we do not attempt validation.  Still, it's possible
that we have those records in cache from an occasion without +cd, in
which case we know they're bogus and this regression happened.

The potential impact of this issue seems minimal.

Somehow I didn't notice this field when adding ::add_selected.
We probably never put anything into answer's ADDITIONAL,
so noone's noticed a problem until now.

We currently allow modules to produce a packet beforehand,
but the possibility was rather implicit in this part of code.
In particular, we might possible trigger a false alarm in
assert(answer->current <= KNOT_ANSWER)

Basically all use cases get renamed to *_E(), except one in utils/kr_cache_gc.
It was perhaps confusing that this name only applied to 'E' entries.

This reverts commit c3a754dd, reversing
changes made to 06d01bf0.

I've have accidentally merged incorrect branch, let's try to fix that by
reverting incorrect version of GC and merging the correct one.

Don't stash a packet with mismatching QNAME+QTYPE.
When receiving an NXDOMAIN or NODATA packet in an insecure zone,
it would get cached with KR_RANK_INSECURE regardless of mismatch
in QNAME.  If the 0x20 pattern was preserved in the fake QNAME,
such packet would then be used to answer queries with matching QNAME,
even if there's no proof that this QNAME is insecure.

Cache uses dname_lf for keys, i.e. zero bytes serve as separators
between labels.  Therefore having a zero inside label could masquerade
for QNAME that does have label separators instead of these zeros.
That doesn't seem really exploitable in practice, as standard registries
won't allow such labels, so I can't see any possible attack that would
"cross border" of these registries, e.g. attacking anything inside
example.org without any cooperation from its owner (or org or root).

Abnormally terminated processes might leave stale reader locks,
and this is required to clear them.

Basically all use cases get renamed to *_E(), except one in utils/kr_cache_gc.
It was perhaps confusing that this name only applied to 'E' entries.

We use packet cache instead.  Also do the same on some kinds of weird
RRsets, as even there some caching should be better than none at all.
This "incidentally" works around all known cases of DVE-2018-0003.

Returning void isn't allowed, apparently.

Mainly reduce duplication of state and request->state,
and slightly reorganize the code.

This is a nitpick, basically.

This was almost all done during the IETF 104 hackathon.

I don't know any particular case where the bit-test would succeed and
equality wouldn't, but it seems better to take no chances.

and slightly better error reporting.

Thanks to changes in this branch the functions are called with correctly
typed parameters already, so these weird casts can be deleted.

This should have been a part of 176b1c28, but I forgot it somehow.

Theoretically a lua module could have used them, but I see no motivation
for that, as they have much more convenient ways inside lua.

... without light userdata and yet it should be efficient.
Also with checkout layer, but I have no nice way of testing that one.

On many places we've been assuming that there's only a single worker,
but we still often didn't utilize the property well.  To get the pointer
we used various ways, all even untyped:
 - __worker global variable in lua
 - uv_default_loop()->data
 - kr_request::daemon_context

Now we instead simply define a global typed pointer the_worker.

Nitpick: also worker_{init,deinit}() are reordered to correspond
to the order of the fields, etc.

There's no support for the "dns" and "tls" kinds (for now).

Tested briefly:
> modules = { 'http' }
> net.listen('/tmp/kresd-socket', nil, { kind = 'webmgmt' })
$ socat TCP-LISTEN:5555,reuseaddr,fork UNIX-CLIENT:/tmp/kresd-socket
$ xdg-open http://127.0.0.1:5555

This is yet another logically equivalent change after 270d9964.
Now it's written in a way that expresses the original intention more
clearly and without copy&paste or long lines.

It seems easiest to verify by inspecting the meaning of the code
*separartely* for the two cases, based on condition:
cur_addr_score < KR_NS_TIMEOUT

The complication is that we need to work with addresses and
just file-descriptors passed from some parent process.
The former approach lead to logical duplication of some steps;
now we add a step converting addresses to file-descriptors.
Thanks to that we always do bind() without touching libuv,
so the problem with forking disappears :-)

We run meson with -std=gnu11, but apparently some compiler still
complained about it.  Unfortunately it wouldn't be easy to use
standard C11 in this case.