Thread handling

[Evan Hunt]

On Tue, Aug 23, 2022 at 05:59:49PM +0000, Hamid Maadani wrote:
> I'm working on a MongoDB DLZ for bind9, and have a few questions about
> the multi-threaded version of named: 1. Is there a pre-processor flag
> I can use to determine if thread support is enabled? Can't seem to
> find it in configure.ac

All supported versions of BIND are now multi-threaded. Single-threaded
support was removed in 9.14, and everything older is now past end of
life.

If you're building for an older version, though, the flag is PTHREADS.

> 2. named does not fork itself, or create child processes, correct?

It does, actually, unless you run it with -g or -f.  That's how it puts
itself in the background, the function is named_os_daemonize().

> 3. when multi-threading is enabled, does named still call dlz_create
> only once when loading the shared object (dlopen)?

I believe so. That happens in dns_dlzcreate() when setting up the view
configuration, if I recall correctly, so it should only occur once in the
main thread.

> 4. when multi-threading is enabled, how is each request handled? are
> there multiple worker threads and requests passed to them? or would
> each request spin up a new thread? Trying to figure out how dlz_lookup
> is called in a multi-threaded version.

In named, there are multiple worker threads that handle queries.

The first step in query processing is to identify which local database,
if any, has data responsive to the query. If no local database does and
recursion is enabled, then we try the cache, and failing that, we recurse.
But if there *is* a local database that's authoritative for the name, then
we call dns_db_findext(), which is a front-end for the 'find' function in
the database implementation (whatever it may be). If the implementation
happens to be a DLZ, then it'll be calilng findext() in lib/dns/sdlz.c,
which in turn calls getnodedata(), which calls the DLZ's 'lookup' method.

In the DLZ module, there can be an arbitrary number of separate database
handles in use, and the query can pick one that isn't busy and use it to
send the query.

There are existing modules in the contrib tree (for example,
contrib/dlz/modules/ldap and contrib/dlz/modules/mysql) that have
examples of this you can use as a model.

(Should you come up with a better way to do this, I'd be happy to
see it. It's always seemed pretty clunky to me but I've never had
the necessary combination of time and brains to improve it.)

-- 
Evan Hunt -- each at isc.org
Internet Systems Consortium, Inc.