UDS decisions

Poul-Henning Kamp phk at phk.freebsd.dk
Tue Feb 13 13:33:11 UTC 2018


As you all know I've been agonizing about Unix Domain Sockets support,
but I've, finally, made up my mind, and came down on the side of
keeping them special, rather than generalize.

I'm going to explain the thinking in this email, for posterity and
so that you can squarely and fairly pin the blame on me at some
later date.

Tl;DR: just read the headings.

Objections ?

Poul-Henning

1. We will use bogo-IP numbers for client UDS connections
=========================================================

VCL variables {local|remote|client|server}.ip will return
VCL_IP(0.0.0.0:0) on UDS sessions.

If PROXY protocol is used, it will, as now, fill {client|server}.ip

Reasoning:
----------

UDS is a hack, and is used as a hack for hackish, but valid reasons.
That does not warrant a very expensive generalization of everything
Varnish knows about protocol endpoints and identities of other
parties.

This in particular because nobody cares one iota about the identity
at the other end of UDS connections and moreso because UDS does not
uniqely identify each UDS connection.

By using a bogo-IP number, we give people the ability to filter
UDS connections in ACLs.

We also avoid large-scale boiler-plate code changes that almost all
amount to "if (X->sa == NULL) ... else ..." where there is no sane
alternative but to invent a bogo-text for the NULL case.

0.0.0.0/8 is registered at IANA as "This host on this network" which
is where INADDR_ANY comes from, and will (these days) always be
changed to something different on real TCP/IP connections, so
there is no ambiguity.

2. We will not log the UDS path anywhere
========================================

Reasoning:
----------

UDS paths are typically longish, exact copies of the -a argument
and would thus be the same for all sessions on the same acceptor
socket, thus conveying no information.

This saves both VSL bandwidth and CPU time.

VSL::SessOpen already contains the socket name, defaulting to "a%d",
in field three, and if you need to distinguish between UDS connections,
you can simply name that -a argument.

If varnishncsa cannot log field 3 of SessOpen if so asked, somebody
should add that functionality.

3. We will not adapt struct suckaddr to UDS addresses
=====================================================

Reasoning:
----------

The reason we have struct suckaddr is primarily because TCP connections
keep copies of the connection endpoints in struct sess and
sockaddr_storage was a waste of space for that.

Given that we dont need to log the UDS path, we don't need to store it
in struct sess, so we don't need them in struct suckaddr either.

The only actual place we will need sockaddr_un is in the acceptor
and we only need one per UDS acceptor socket.


4. backend.ip also gets bogo-IP
===============================

Backend.ip will return VCL_IP(0.0.0.0:0) unless the director has a
real IP endpoint to report.

Reasoning:
----------

Directors were not supposed to be limited to IP connections and
backend.ip was, in hindsight, a mistake, but it is useful and
popular in VCL code.

Avoiding NULL checks (like under point 1) makes sense.

5. We will have a global, constant bogo-ip variable for this
============================================================

	extern const struct suckaddr bogo_ip;

Reasoning:
----------

For coding convenience. 

But do *NOT* simply compare the pointer, use VSA_Compare().


-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe
Never attribute to malice what can adequately be explained by incompetence.


More information about the varnish-dev mailing list