Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 30 Mar 2018 16:35:01 -0400
From: Rich Felker <dalias@...c.org>
To: musl@...ts.openwall.com
Subject: Re: [PATCH] resolver: only exit the search path loop there
 are a positive number of results given

On Fri, Mar 30, 2018 at 02:44:44PM -0500, William Pitcock wrote:
> Hello,
> 
> On Fri, Mar 30, 2018 at 2:35 PM, Rich Felker <dalias@...c.org> wrote:
> > On Fri, Mar 30, 2018 at 02:19:48PM -0500, William Pitcock wrote:
> >> Hello,
> >>
> >> On Fri, Mar 30, 2018 at 2:14 PM, Rich Felker <dalias@...c.org> wrote:
> >> > On Fri, Mar 30, 2018 at 06:52:25PM +0000, William Pitcock wrote:
> >> >> In the event of no results being given by any of the lookup modules, EAI_NONAME will still
> >> >> be thrown.
> >> >>
> >> >> This is intended to mitigate problems that occur when zones are hosted by weird DNS servers,
> >> >> such as the one Cloudflare have implemented, and appear in the search path.
> >> >> ---
> >> >>  src/network/lookup_name.c | 2 +-
> >> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >> >>
> >> >> diff --git a/src/network/lookup_name.c b/src/network/lookup_name.c
> >> >> index 209c20f0..b068bb92 100644
> >> >> --- a/src/network/lookup_name.c
> >> >> +++ b/src/network/lookup_name.c
> >> >> @@ -202,7 +202,7 @@ static int name_from_dns_search(struct address buf[static MAXADDRS], char canon[
> >> >>                       memcpy(canon+l+1, p, z-p);
> >> >>                       canon[z-p+1+l] = 0;
> >> >>                       int cnt = name_from_dns(buf, canon, canon, family, &conf);
> >> >> -                     if (cnt) return cnt;
> >> >> +                     if (cnt > 0) return cnt;
> >> >>               }
> >> >>       }
> >> >
> >> > This patch is incorrect, and the reason should be an FAQ item if it's
> >> > not already. Only a return value of 0 means that the requested name
> >> > does not exist and that it's permissible to continue search. Other
> >> > nonpositive return values indicate either that the name does exist but
> >> > does not have a record of the quested type, or that a transient error
> >> > occurred, making it impossible to determine whether the search can be
> >> > continued and thus requiring the error to be reported to the caller.
> >> > Anything else results in one or both of the following bugs:
> >> >
> >> > - Nondeterministically returning different results for the same query
> >> >   depending on transient unavailability of the nameservers to answer
> >> >   on time.
> >> >
> >> > - Returning inconsistent results (for different search components)
> >> >   depending on whether AF_INET, AF_INET6, or AF_UNSPEC was requested.
> >> >
> >> > I'm aware that at least rancher-dns and Cloudflare's nameservers have
> >> > had bugs related to this issue. I'm not sure what the status on
> >> > getting them fixed is, and for Cloudflare I don't know exactly what it
> >> > is they're doing wrong or why. But I do know the problem is that
> >> > they're returning semantically incorrect dns replies.
> >>
> >> Kubernetes imposes a default search path with the cluster domain last, so:
> >>
> >>   - local.prod.svc.whatever
> >>   - prod.svc.whatever
> >>   - svc.whatever
> >>   - yourdomain.com
> >>
> >> The cloudflare issue is that they send SUCCESS code with 0 replies,
> >> which causes musl to error when it hits the yourdomain.com.
> >
> > Yes, that makes sense. Do you know why they're doing it? If they
> > refuse to fix it, the only clean fix I know is a local proxy
> > configured to fix the records for the specific broken domains you care
> > about. But of course that's not convenient.
> 
> My contacts at cloudflare indicate that their environment depends on
> this behaviour, so they have no interest in fixing it.
> 
> A local proxy isn't going to be workable, because most people are
> going to just say "but Debian or Fedora doesn't require this," and
> then just go use a glibc distribution.
> 
> There is a talk in a few weeks at Kubecon (the Kubernetes conference),
> explicitly titled "Don't Use Alpine If You Care About DNS."  The talk
> largely centers around how musl's overly strict behaviour makes Alpine
> a bad choice for "the real world."  I would like to turn this into a
> story where we can announce that Alpine 3.8 mitigates this problem
> instead, doing such will be good for both Alpine and the musl
> ecosystem as a whole, as it is defanging a point of possible FUD.

I understand that this is a frustrating position to be in, where
someone who (from your description so far, at least) seems like a
troll, is planning to publicly malign your distro. I don't think
rushing quick, wrong workarounds helps. Rather it makes it look like
you knew there was a problem and only acted on it because someone
threatened to make you look bad in public. The reality is that
Cloudflare is returning wrong DNS results that aren't suitable for the
search settings k8s users want to use.

I think a better response would be demonstrating why the glibc
behavior they're relying on is buggy. This should be possible by
configuring two search domains:

	ex1.example.com
	ex2.example.com

and adding the following records to them:

	foo.ex1.example.com.	A	10.0.0.1
	foo.ex2.example.com.	A	192.168.0.1
	foo.ex2.example.com.	AAAA	fd00::1

Then, try glibc's getaddrinfo for the name "foo" with AF_INET,
AF_INET6, and AF_UNSPEC. The results I expect you'll see are:

	AF_INET:	10.0.0.1
	AF_INET6:	192.168.0.1, fd00::1
	AF_UNSPEC:	10.0.0.1 (possibly also fd00::1 but doubtful)

That is, *which fqdn you see results for* is depending on which
address family you made the query with. This is clearly wrong.

With musl you'll consistently see results for foo.ex1.example.com:

	AF_INET:	10.0.0.1
	AF_INET6:	no results
	AF_UNSPEC:	10.0.0.1

since it's the first definition of foo in the serch domains.

> >> Do you have any suggestions on a mitigation which would be more
> >> palatable?  We need to ship a mitigation for this in Alpine 3.8
> >> regardless.  I would much rather carry a patch that is upstreamable,
> >> but I am quite willing to carry one that isn't, in order to solve this
> >> problem.
> >
> > A theoretically-non-horrible (but somewhat costly) solution is to
> > always query both A and AAAA, rather than only doing it for AF_UNSPEC.
> > Then if you see a reply with 0 (total, between both) records, you can
> > opt to interpret that the same way as NxDomain without breaking
> > consistency properties. If Cloudflare refuses to fix the bug, maybe we
> > should consider adding an _option_ (in the resolv.conf options line)
> > to do this. I don't think it should be the default behavior because it
> > mildly slows down lookups, especially if you have nontrivial packet
> > loss since probability of failure is now 1-(1-p)²=2p-p² rather than p
> > (where p is the packet loss rate).
> 
> It seems to me we could just send ANY and filter out the records we
> don't care about.  This is what I did with charybdis's asynchronous
> DNS resolver when it had a similar problem.  What are your thoughts on
> that?

If ANY requests actually worked we would have done this years ago.
They don't. Not only is the ANY result unlikely to fit in a UDP reply
(and lacking any reason to include the records you need if it doesn't
fit); A and AAAA queries are also special in that they resolve CNAMEs
and automatically include the A or AAAA records for the name the CNAME
points to in the reply. With ANY you'd have to reimplement the whole
CNAME logic as a second (slow) query from the stub resolver rather
than having the recursive resolver give you the answer it already knew
when you made the first query.

Rich


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.