OMBUD
=====

Ombud (Swedish for proxy) is a simple network proxy which takes commands
from connecting clients. The commands have the form "ADDRESS:PORT\r\n",
when Ombud recieves a command it connects to the given address and port,
reads whatever is sent, caches it on disk, and relays it back to the
client.


BUILD AND REQUIREMENTS
----------------------
You need to run Linux 3.9 or later, this is because SO_REUSEPORT was
introduced in that release.

In order to build, first install build dependencies

    sudo apt-get install build-essential libssl-dev

then build by executing

    make


RUN
---
Execute Ombud with

    make run

This will start Ombud on the default port 8090. The cache is created in
$PWD/cache-ombud.

NB! The cache is removed from source directory during invokations of
make clean.

It is possible to give port number and number of processes as
*positional* command line arguments

    bin/ombud 8077 1  # port 8077, one (1) process

Quit by sending SIGINT, i.e. pressing Ctrl-C.


USAGE
-----
Connect with a network client to Ombud's port and send commands

    $ nc localhost 8090
    localhost:22
    SSH-2.0-OpenSSH_7.1p2 Debian-2

It is possible to give both alpha numerical and numerical hosts and
ports.


DESIGN
------
The main function is in the event loop, which handles client connections
and executes the client commands (connecting to services) and
caches/sends back the response. The event handling is based on epoll(7),
which is Linux specific, and non-blocking sockets.

The event loops are forked into their own processes, one process per CPU
core, and the listen sockets are setup with SO_REUSEPORT. This will
enable all processes to listen on the same port, and the kernel handles
distribution of incoming requests. (It is stated that SO_REUSEPORT will
distribute incoming requests uniformly among the processes, which was
stated to not be the case with threaded acceptors.  See Linux kernel
commit c617f39 for more information.)

    NB! Using SO_REUSEPORT is a little bit experimentation. If it
    doesn't work try running one process only.

When clients request data from address:port a cache lookup is performed.
On a cache hit the contents are sent to the client with sendfile(2),
which shuffles data from a file descriptor to a socket without leaving
kernel space. On a cache miss, Ombud connects to the given address and
port, reads the data, writes it to the cache, and finally relays it back
to the client.

The client connections are toggled between the states READ_CMD, which
reads commands from the client, and READ_REMOTE, which executes the
supplied client command (i.e. fetches data from the remote service or
the cache and relays back to the client).


ASSUMPTIONS
-----------
* Only IPv4, this application knows little to nothing about IPv6.

* Well formed client commands and connectable services are assumed,
  malformed requests and unconnectable hosts are silently dropped.

* Not a lot of data will be sent from the services that Ombud connects
  to, a static buffer size of 8 kB is deemed to be adequate for the use
  case. (This was a simplification instead of having to create a chunked
  buffer, which would have been used in a real world application.)

* Data downloaded from hosts is assumed to never change, hence the
  content in the cache never expires.

* There is no portability, Linux (3.9+) only. (I wanted to experiment
  with SO_REUSEPORT!)


POSSIBLE IMPROVEMENTS
---------------------
* Use libevent instead, for portability and also a nice framework for
  event driven programming. Given time I would have tried libuv from
  Joyent, it looks interesting.

* Evaluate if using thread pools for incoming connections, workers etc
  yields higher performance. Possibly trying a combination of processes,
  threads, and multiplexing.

* In order to make a more efficient cache, hostnames could be resolved
  to ip addresses before they are used as part of the key.

* Linking against OpenSSL. If the code is going to be released under the
  GPL, an OpenSSL linking exception is required in the license in order
  to be specific about the user's rights and responsibilities. It would
  of course be possible to use libgcrypt in this case instead,
  especially since SHA1() is only used because it is simple.

* Make cache directory and listen port configurable from the command
  line with getopt and possibly through a configuration file.

* Since this is a simple server it might be enough to store the cache in
  memory. Alternatively, store it in a ramdisk.

* More robust error handling, e.g. if cache storage breaks between
  succesful lookup and send.

* Of course the standard things: more documentation, configuration file,
  command line options (with getopt_long or some such) instead of fixed
  argument placing on the command line etc.

* Toggable DEBUG prints.

* Input data validation, little to none is done now. Dangerous!

* Possibly it clutters down the code casting between uint8_t and char
  all the time... The interfaces are clean though, the clutter comes
  from standard functions that use (char *).

* Tests. Unit tests with e.g. Google Test, functional tests etc. TDD is
  really nice. This time around though everything was built in a quick
  prototype fashion. (Maybe one could argue that the valgrind make
  target is a basic test for some good and bad input data.)

* Analyze Valgrind report and fix, or suppress, errors. Valgrind helped
  during development, it showed some errors that was fixed.

* Proper cleanup of sockets and allocated memory upon catching SIGINT.