Exfiltrating log data using syslog

Written by hannes
Classified under: mirageosprotocollogging
Published: 2016-11-05 (last updated: 2016-11-06)

It has been a while since my last entry... I've been busy working on too many projects in parallel, and was also travelling on several continents. I hope to get back to a biweekly cycle.

What is syslog?

According to Wikipedia, syslog is a standard for message logging. Syslog permits separation of the software which generates, stores, reports, and analyses the message. A syslog message contains at least a timestamp, a facility, and a severity. It was initially specified in RFC 3164, though usage predates this RFC.

For a unikernel, which likely won't have any persistent storage, syslog is a way to emit log messages (HTTP access log, debug messages, ...) via the network, and defer the persistency problem to some other service.

Lots of programming languages have logger libraries, which reflect the different severity of syslog roughly as log levels (debug, informational, warning, error, fatal). So does OCaml since the beginning of 2016, there is the Logs library which separates log message generation from reporting: the closure producing the log string is only evaluated if there is a reporter which needs to send it out. Additionally, the reporter can extend the message with the log source name, a timestamp, etc.

The Logs library is slowly getting adopted by the MirageOS community (you can see an incomplete list here), there are reporters available which integrate into Apple System Log, Windows event log, and also for MirageOS console. There is a command-line argument interface to set the log levels of your individual sources, which is pretty neat. For debugging and running on Unix, console output is usually sufficient, but for production usage having a console in some screen or tmux or dumped to a file is usually annoying.

Gladly there was already the syslog-message library, which encodes and decodes syslog messages from the wire format to a typed representation. I plugged those together and implemented a reporter. The simplest one emits each log message via UDP to a log collector. All reporters contain a socket and handle socket errors themselves (trying to recover) - your application (or unikernel) shouldn't fail just because the log collector is currently offline.

The setup for Unix is straightforward:

Logs.set_reporter (udp_reporter (Unix.inet_addr_of_string "127.0.0.1") ())

It will report all log messages (you have to set the Log level yourself, defaults to warning) to your local syslog. You might have already listening a collector on your host, look in netstat -an for UDP port 514 (and in your /etc/syslog.conf to see where log messages are routed to).

You can even do this from the OCaml toplevel (after opam install logs-syslog):

$ utop
# #require "logs-syslog.unix";;
# Logs.set_reporter (Logs_syslog_unix.udp_reporter (Unix.inet_addr_of_string "127.0.0.1") ());;
# Logs.app (fun m -> m "hello, syslog world");;

I configured my syslog to have all informational messages routed to /var/log/info.log, you can also try Logs.err (fun m -> m "err");; and look into your /var/log/messages.

This is a good first step, but we want more: on the one side integration into MirageOS, and a more reliable log stream (what about authentication and encryption?). I'll cover both topics in the rest of this article.

Reliable syslog

The old BSD syslog RFC is obsoleted by RFC 5424, which describes a new wire format, and also a transport over TCP, and TLS in a subsequent RFC. Unfortunately the syslog-message library does not yet support the new format (which supports user-defined structured data (key/value fields), and unicode encoding), but I'm sure one day it will.

Another competing syslog RFC 3195 uses XML encoding, but I have not bothered to look deeper into that one.

I implemented both the transport via TCP and via TLS. There are various solutions used for framing (as described in RFC 6587): either prepend a decimal encoded length (also specified in RFC6524, but obviously violates streaming characteristics: the log source needs to have the full message in memory before sending it out), or have a special delimiter between messages (0 byte, line feed, CR LN, a custom byte sequence).

The TLS reporter uses our TLS library written entirely in OCaml, and requires mutual authentication, both the log reporter has a private key and certificate, and the log collector needs to present a certificate chain rooted in a provided CA certificate.

Logs supports synchronous and asynchronous logging (where the latter is the default, please read the note on synchronous logging). In logs-syslog this behaviour is not altered. There is no buffer or queue and single writer task to emit log messages, but a mutex and error recovery which tries to reconnect once for each log message (of course only if there is not already a working connection). It is still not clear to me what the desired behaviour should be, but when introducing buffers I'd loose the synchronous logging (or will have to write rather intricate code).

To rewrap, logs-syslog implements the old BSD syslog protocol via UDP, TCP, and TLS. There are reporters available using only the Caml Unix module (dependency-free!), using Lwt (also lwt-tls, and using MirageOS interface (also TLS). The code size is below 500 lines in total.

MirageOS syslog in production

As collector I use syslog-ng, which is capable of receiving both the new and the old syslog messages on all three transports. The configuration snippet for a BSD syslog TLS collector is as following:

source s_tls {
  tcp(port(6514)
      tls(peer-verify(require-trusted)
          cert-file("/etc/log/server.pem")
          key-file("/etc/log/server.key")
          ca-dir("/etc/log/certs"))); };

destination d_tls { file("/var/log/ng-tls.log"); };

log { source(s_tls); destination(d_tls); };

The "/etc/log/certs" directory contains the CA certificates, together with links to their hashes (with a 0 appended: ln -s cacert.pem `openssl x509 -noout -hash -in cacert.pem`.0). I used certify to generate the CA infrastructure (CA cert, a server certificate for syslog-ng, and a client certificate for my MirageOS unikernel).

I added the boilerplate code to this blog software, surely this should be massaged and moved up the stack, thus it is easily available for other MirageOS unikernels. It is running since a week like a charm (already collected 700KB of HTTP access log), and feels much better than previous ad-hoc solutions to exfiltrate log data.

The downside of syslog is obviously that it only works when the network is up -- thus it does not work while booting, or when a persistent network failure occured.

Code is on GitHub, documentation is online, released in opam. A version of logs-syslog supporting the current development version of MirageOS is available in the dev branch.

I'm interested in feedback, either via twitter or as an issue on the data repository on GitHub.