docs/diploma
view thesis/tex/4-MasqmailsFuture.tex @ 297:39fffd8d1100
even more work on ch04 ... but it is quite finished now :-)
| author | meillo@marmaro.de |
|---|---|
| date | Sun, 18 Jan 2009 13:08:32 +0100 |
| parents | a014ce012053 |
| children | 0d88bf21e152 |
line source
1 \chapter{\masqmail's present and future}
3 This chapter identifies requirements for \masqmail\ which are compared against the current code to see what is already fulfilled and what is missing. Then the outstanding work is ordered by relevance and a list of tasks to do is created. The end of this chapter is the evaluation of the best development strategy to get the work done in order to achieve the requirements.
6 \section{The goal}
8 Before requirements can be identified and further development can be discussed, it is important to clearly specify the goal to achieve. This means: What shall \masqmail\ be like in, for instance, five years?
10 Should \masqmail\ become more specific to a more narrow niche or rather become more general and move a bit out of its niche? Or should it even become a totally general \MTA\ like \sendmail, \exim, \qmail, and \postfix?
12 Becoming completely general seems to be no choice because the competitors are too many and they are already too strong. It would require a strong base of developers and superior features to establish. There seems to be no need for another general purpose \MTA\ additional to those four programs. Thus the effort would most likely die a try. \person{Venema} stated ``It is becoming less and less likely that someone will write another full-featured Postfix or Sendmail \MTA\ \emph{from scratch} (100 kloc).'' \cite{venema:postfix-growth}. At least \masqmail\ is not going to try that.
14 \masqmail\ was intended to be a small ``real \MTA'' which covers the niche of managing the relay over several smart hosts. Small and resource friendly software is still important for workstations, home servers, and especially for embedded computers. Other software that focuses on the same niche is not known. Dial-up connections have become rare but mobile computers that move between different networks are popular. So, the niche is still present.
16 What has changed in general is the security that is needed for software. \person{Graff} and \person{van Wyk} describe the situation well: ``[I]n today's world, your software is likely to have to operate in a very hostile security environment.'' Additionally they say: ``By definition, mail software processes information from potentially untrusted sources. Therefore, mail software must be written with great care, even when it runs with user privileges and even when it does not talk directly to a network.'' \cite[page~33, page~90]{graff03}. As \masqmail\ is mail software and trusted environments become rare, it is best for \masqmail\ to become a secure \MTA.
18 In summary, the goal for \masqmail\ is to stay in the current niche with respect to modern usage scenarios and to become a secure \MTA.
25 \section{Requirements}
26 \label{sec:mta-requirements}
28 This section identifies the requirements for \masqmail\ to reach the above defined goal. Most of the requirements will apply to modern \MTA{}s in general.
32 \subsection{Functional requirements}
34 Functional requirements are about the function of the software. They define what the program can do and in what way.
35 %fixme: add ref
36 The requirements are named ``\NAME{RF}'' for ``requirement, functional''.
39 \paragraph{\RF1: Incoming and outgoing channels}
40 \sendmail-compatible \mta{}s must support at least two incoming channels: mail submitted using the \sendmail\ command, and mail received on a \NAME{TCP} port. Thus it is common to split the incoming channels into local and remote. This is done by \qmail\ and \postfix. The same way is \person{Hafiz}'s view \cite{hafiz05}.
42 \SMTP\ is the primary mail transport protocol today, but with the increasing need for new protocols (see section \ref{sec:what-will-be-important}) in mind, support for more than just \SMTP\ is good to have. New protocols will show up, maybe multiple protocols need to be supported then. This leads to multiple remote channels, one for each supported protocol as it was done in other \MTA{}s. Best would be interfaces to add further protocols as modules.
45 Outgoing mail is commonly either sent using \SMTP, piped into local commands (for example \path{uucp}), or delivered locally by appending to a mailbox. Outgoing channels are similar for \qmail, \postfix, and \name{sendmail X}: All of them have a module to send mail using \SMTP, and one for writing into a local mailbox.
47 %todo: is the def of MTA: transfer between machines, or transfer between users?
48 Local mail delivery is a job that uses root privilege to be able to switch to any user in order to write to his mailbox. It is possible to deliver without being root privilege, but delivery to user's home folders is not generally possible then. Thus even the modular \MTA{}s \qmail\ and \postfix\ use root privilege for it. As mail delivery to local users is \emph{not} included in the basic job of an \MTA{} and introduces a lot of new complexity, why should the \MTA\ bother? In order to keep the system simple, reduce privilege, and to have programs that do one job well, the local delivery job should be handed over to a specialist: the \NAME{MDA}. \NAME{MDA}s know about the various mailbox formats and are aware of the problems of concurrent write access and the like. Hence passing the message, and the responsibility for it, over to an \NAME{MDA} seems to be best.
50 This means an outgoing connection that pipes mail into local commands is required. To other outgoing channels applies what was already said about incoming channels.
52 \begin{figure}
53 \begin{center}
54 \includegraphics[scale=0.75]{img/mta-channels.eps}
55 \end{center}
56 \caption{Required incoming and outgoing channels}
57 \label{fig:mta-channels}
58 \end{figure}
60 An overview on in and outgoing channels required for an \MTA, gives figure \ref{fig:mta-channels}.
62 %fixme: write about submission (port 587)
67 \paragraph{\RF2: Mail queuing}
68 Mail queuing removes the need to deliver instantly as a message is received. The queue provides fail-safe storage of mails until they are delivered. Mail queues are probably used in all \mta{}s, even in some simple forwarders. The mail queue is essential for \masqmail, as \masqmail\ is used for non-permanent online connections. This means, mail must be queued until a online connection is available to send the message. This may be after a reboot. Hence the mail queue must provide persistence.
70 The mail queue and the module(s) to manage it are the central part of the whole system. This demands especially for robustness and reliability, as a failure here can lead to loosing mail. An \MTA\ takes over responsibility for mail in accepting it, hence loosing mail messages is absolutely to avoid. This covers any kind of crash situation too. The worst thing acceptable to happen is an already sent mail to be sent again.
75 \paragraph{\RF3: Header sanitizing}
76 Mail coming into the system often lacks important header lines. At least the required ones must be added by the \MTA. One example is the \texttt{Date:} header, another is the, not required but recommended, \texttt{Message-ID:} header. Apart from adding missing headers, rewriting headers is important too. Changing the locally known domain part of email addresses to globally known ones is an example. \masqmail\ needs to be able to rewrite the domain part dependent on the route used to send the message, to prevent messages to get classified as spam.
78 Generating the envelope is a related job. The envelope specifies the actual recipient of the mail, no matter what the \texttt{To:}, \texttt{Cc:}, and \texttt{Bcc:} headers contain. Multiple recipients lead to multiple different envelopes, containing all the same mail message.
83 \paragraph{\RF4: Aliasing}
84 Email addresses can have aliases, thus they need to be expanded. Aliases can be of different kind: another local user, a remote user, a list containing local and remote users, or a command. Most important are the aliases in the \path{aliases} file, usually located at \path{/etc/aliases}. Addresses expanding to lists of users lead to more envelopes. Aliases changing the recipient's domain part may require a different route to be used.
89 \paragraph{\RF5: Route management}
90 One key feature of \masqmail\ is its ability to send mail out over different routes. The online state defines the active route to be used. A specific route may not be suited for all messages, thus these messages are hold back until a suiting route is active. For more information on this concept see section \ref{sec:masqmail-routes}.
95 \paragraph{\RF6: Authentication}
96 One thing to avoid is being an \name{open relay}. Open relays allow to relay mail from everywhere to everywhere. This is a source of spam. The solution is restricting relay\footnote{Relaying is passing mail, that is not from and not for the own system, through it.} access. It may be also wanted to refuse all connections to the \MTA\ except ones from a specific set of hosts.
98 Several ways to restrict access are available. The most simple one is restriction by the \NAME{IP} address. No extra complexity is added this way, but the \NAME{IP} addresses have to be static or within known ranges. This approach is often used to allow relaying for local nets. The access check can be done by the \MTA\ or by a guard (e.g.\ \NAME{TCP} \name{Wrappers}) before. The main advantage here is the minimal setup and maintainence work needed. This kind of access restriction is important to be implemented.
100 This authentication based on \NAME{IP} addresses is impossible in situations where hosts with changing \NAME{IP} addresses, that are not part of a known subnet, need access. Then a authentication mechanism based on some \emph{secret} is required. Three common approaches exist:
101 \begin{enumerate}
102 \item \SMTP-after-\NAME{POP}: Uses authentication on the \NAME{POP} protocol to permit incoming \SMTP\ connections for a limited time afterwards. The variant \SMTP-after-\NAME{IMAP} exists too.
103 \item \SMTP\ authentication: An extension to \SMTP. It allows to request authentication before mail is accepted. Here no helper protocols are needed.
104 \item Certificates: The identity of a user or a host is confirmed by certificates that are signed by trusted authorities. Certificates are closely related to encryption, they do normally satisfy both needs: \NAME{SSL} tunnels encrypt the data transmission and allow to identify the remote user/host by his certificate.
105 \end{enumerate}
106 At least one of the secret-based mechanisms should be supported.
110 \paragraph{\RF7: Encryption}
111 Electronic mail is vulnerable to sniffing attacks, because in generic \SMTP\ all data transfer is unencrypted. The message's body, the header, and envelope are all unencrypted, but also authentication dialogs that transfer plain text passwords (e.g.\ \NAME{PLAIN} and \NAME{LOGIN}). Hence encryption is throughout important.
113 The common way to encrypt \SMTP\ dialogs is using \name{Transport Layer Security} (short: \TLS, the successor of \NAME{SSL}). \TLS\ encrypts the datagrams of the \name{transport layer}. This means it works below the application protocols and can be used with any of them \citeweb{wikipedia:tls}.
115 Using secure tunnels that are provided by external programs, should be prefered over including encryption into the application, because the application needs not to bother with encryption then. Outgoing \SMTP\ connections can get encrypted using a secure tunnel, created by an external application (like \name{openssl}). But incoming connections can not use external secure tunnels, because the remote \NAME{IP} address is hidden then; all connections would appear to come from localhost instead. Figure \ref{fig:stunnel} depicts the situation of using an application like \name{stunnel} for incoming connections. The connection to port 25 comes from localhost and this information reaches the \MTA. Authentication based on \NAME{IP} addresses and many spam prevention methods are useless then.
117 \begin{figure}
118 \begin{center}
119 \includegraphics[scale=0.75]{img/stunnel.eps}
120 \end{center}
121 \caption{Using \name{stunnel} for incoming connections}
122 \label{fig:stunnel}
123 \end{figure}
125 To provide encrypted incoming channels, the \MTA\ could implement encryption and listen on a port that is dedicated to encrypted \SMTP\ (\NAME{SMTPS}). This approach would be possible, but it is deprecated in favor for \NAME{STARTTLS}. \RFC3207 ``\SMTP\ Service Extension for Secure \SMTP\ over Transport Layer Security'' shows this by not mentioning \NAME{SMTPS} on port 465. Also port 465 is not even reserved for \NAME{SMTPS} anymore \citeweb{iana:port-numbers}.
127 \NAME{STARTTLS}---defined in \RFC2487---is what \RFC3207 recommends to use for secure \SMTP. The connection then goes over port 25 (or the submission port 587), but gets encrypted as the \NAME{STARTTLS} keyword is issued. Email depends on compatibility---only encryption methods that client and server support can be used. Hence it is best to act after the recommendations of the \RFC\ documents. This means \NAME{STARTTLS} encryption should be supported for incoming and for outgoing connections.
131 \paragraph{\RF8: Spam handling}
132 Spam is a major threat nowadays, but it is a war that is hard to win. The goal is to provide state-of-the-art spam protection, but not more (see section \ref{sec:swot-analysis}).
134 As spam is, by increasing the amount of mail messages, not just a nuisance for end users, but also for the infrastructure---the \mta{}s---they need to protect themselves.
136 Filtering spam can be done by either refusing spam during the \SMTP\ dialog or by checking for spam after the mail was accepted and queued. Both ways have advantages and disadvantages, so modern \MTA{}s use them in combination.
138 Spam is identified by the results of a set of checks. Static rules, querying databases (\NAME{DNS} blacklists \cite{cole07} \cite{levine08}), requesting special client behavior (\name{greylisting} \cite{harris03}, \name{hashcash} \cite{back02}), or statistical analysis (\name{bayesian filters} \cite{graham02}) are checks that may be used. Running more checks leads to better results, but takes more system resources and more time.
140 Doing some basic checks during the \SMTP\ dialog seems to be a must \cite[page~25]{eisentraut05}. Including them into the \MTA\ makes them fast to avoid \SMTP\ dialog timeouts. For modularity and reusability reasons internal interfaces to specialized modules seem to be best.
142 More detailed checks after the message is queued should be done using external scanners. Interfaces to invoke them need to be defined. (See also the remarks about \name{amavis} in the next section.)
148 \paragraph{\RF9: Malware handling}
149 Related to spam is malicious content (short: \name{malware}) like viruses, worms, trojan horses. They, in contrast to spam, do not affect the \MTA\ itself, as they are in the mail's body. \MTA{}s searching for malware is equal to real world's post offices opening letters to check if they contain something that could harm the recipient. This is not a mail transport job. But by many people the \MTA\ which is responsible for the recipient is seen to be at a good position to do this work, so it is often done there.
151 In any way should malware checking be performed by external programs that may be invoked by the \mta. But \NAME{MDA}s are better points to invoke content scanners.
153 A popular email filter framework is \name{amavis} which integrates various spam and malware scanners. The common setup includes a receiving \MTA\ which sends it to \name{amavis} using \SMTP, \name{amavis} processes the mail and sends it then to a second \MTA\ that does the outgoing transfer. Having interfaces to such scanners is nice to have, though. (This setup with two \MTA\ instances is discussed in more detail in section \ref{sec:double-mta-setup}).
157 \paragraph{\RF10: Archiving}
158 Mail archiving and auditability become more important as email establishes as technology for serious business communication. It is also a must for companies in many countries. << SOX >> %fixme: cite SOX
159 The ability to archive verbatim copies of every mail coming into and every mail going out of the system, with relation between them, appears to be a goal to achieve.
161 \postfix\ for example has a \texttt{always\_bcc} feature, to send a copy of every outgoing mail to a definable recipient. At least this functionality should be given, although a more complete approach, like \qmail\ provides, is preferable. \qmail\ is able to save copies of all sent and received messages and additionally complete \SMTP\ dialogs \cite[page~12]{sill02}.
167 \subsection{Non-functional requirements}
169 Here follows a list of non-functional requirements for \masqmail. These requirements specify the quality properties of software. The list is based on \person{Hafiz} \cite[page~2]{hafiz05}, with inspiration from \person{Spinellis} \cite[page~6]{spinellis06} and \person{Kan} \cite{kan03}.
170 %fixme: refer to ch01 and ch02
171 These non-functional requirements are named ``\NAME{RG}'' for ``requirement, general''.
174 \paragraph{\RG1: Security}
175 \MTA{}s are critical points for computer security, as they are accessible from external networks. They must be secured with high effort. Properties like the need for high privilege level, from outside influenced work load, work on unsafe data, and demand for reliability, increase the need for security. This is best done by modularization, also called \name{compartementalization}, as described in section \ref{sec:discussion-mta-arch}.
177 \masqmail\ needs to be secure enough for its target field of operation. \masqmail\ is targeted to workstations and private networks, with explicit warning to not use it on permanent online hosts \citeweb{masqmail:homepage2}. But as non-permanent online connections and trustable environments become rare, \masqmail's security should be so good, that it is usable with permanent online connections and in unsafe environments. For example should mails with bad content not break \masqmail.
180 \paragraph{\RG2: Reliability}
181 Reliability is the second essential quality property for an \MTA. Mail for which the \MTA\ took responsibility must never get lost while it is within the \MTA{}s responsibility. The \MTA\ must not be \emph{the cause} of any mail loss, no matter what happens. Unreliable \mta{}s are of no value. However, as the mail transport infrastructure are distributed systems, one of the communication partners or the transport medium may crash at any time during mail tranfer. Thus reliability is needed for mail transfer communication too.
183 The goal is to transfer exactly one copy of the message. \person{Tanenbaum} evaluates the situation and comes to the conclusion that ``in general, there is no way to arrange this.'' \cite[pages~377--379]{tanenbaum02}. Only strategies where now mail gets lost are acceptable; he identifies three of them, but one generates more duplicates than the others, so two strategies remain. (1) The client always reissues the transfer; the server first sends an acknowledgement, then handles the transfer. (2) The client reissues the transfer only if no acknowledgement was received; the server first handles the transfer and sends the acknowledgement afterwards. The first strategy does not need acknowledgements at all, however, it will lose mail if the second transfer fails too.
185 Hence, mail transfer between two processes must use the strategy: The client reissues if it receives no acknowledgement; the server first handles the message and then sends the acknowledgement. This strategy only leads to duplicates if a crash happens in the time between the message is fully transfered to the server and the acknowlegement is received by the client. No mail will get lost.
188 \paragraph{\RG3: Robustness}
189 Being robust means handling errors properly. Small errors may get corrected, large errors may kill a process. Killed processes should restarted automatically and lead to a clean state again. Log messages should be written in every case. Robust software does not need a special environment, it creates a friendly environment itself. \person{Raymond}'s \name{Rule of Robustness} and his \name{Rule of Repair} are good descriptions \cite[pages~18--21]{raymond03}.
192 \paragraph{\RG4: Extendability}
193 \masqmail's architecture needs to be extendable, to allow new features to be added afterwards. The reason for this need are changing requirements. New requirements appear, like more efficient mail transfer of large messages or a final solution for spam problem. Extendability is the ability of software to include new function with little work.
196 \paragraph{\RG5: Maintainability}
197 Maintaining software takes much time and effort. \person{Spinellis} guesses ``40\,\% to 70\,\% of the effort that goes into a software system is expended after the system is written first time.'' \cite[page~1]{spinellis03}. This work is called \emph{maintaining}. Hence making software good to maintain will ease work afterwards.
200 \paragraph{\RG6: Testability}
201 Good testability make maintenance easier too, because functionality is directly verifiable when changes are done, thus removing uncertainty. Modularized software makes testing easier, because parts can be tested without external influences. \person{Spinellis} sees testability as a sub-quality of maintainability.
204 \paragraph{\RG7: Performance}
205 Also called ``efficiency''. Efficient software requires few time and few resources. The merge of communication hardware and its move from service providers to homes and to mobile devices, demand smaller and more resource-friendly software. The amount of mail will be lower even if much more mail will be sent, thus time performance is less important. \masqmail\ is not a program to be used on large servers, but on small devices. Thus more important for \masqmail\ will be energy and heat saving, maybe also system resources.
207 As performance improvements are in contrast to many other quality properties (reliability, maintainability, usability, capability \cite[page~5]{kan03}), jeopardizing these to gain some more performance should not be done. \person{Kernighan} and \person{Pike} state clear: ``[T]he first principle of optimization is \emph{don't}.''\cite[page~165]{kernighan99}. Simplicity and clearness are of higher value.
211 \paragraph{\RG8: Availability}
212 Availability is important for server programs. They must stay operational by blocking \name{denial of service} attacks and the like.
215 \paragraph{\RG9: Portability}
216 Source code that compiles and runs on various operation systems is called portable. Portability can be achieved by using standard features of the programming language and common libraries. Basic rules to achieve portable code are defined by \person{Kernighan} and \person{Pike} \cite{kernighan99}. Portable code lets software spread faster. Portability among the various flavors of \unix\ systems is a goal, because these systems are the ones \MTA{}s run on usually. No special care needs to be taken for non-\unix\ platforms.
220 \paragraph{\RG10: Usability}
221 Usability, not mentioned by \person{Hafiz} (he focuses on architecture) but by \person{Spinellis} and \person{Kan}, is a property very important from the user's point of view. Software with bad usability is rarely used, no matter how good it is. If substitutes with better usability exist, the user will switch to one of them. Here, usability includes setting up and configuring; and the term ``users'' includes administrators. Having \mta{}s on home servers and workstations requires easy and standardized configuration. The common setups should be configurable with little action by the user. Complex configuration should be possible, but focused must be the most common form of configuration: choosing one of several common setups.
223 << masqmail as portable app? >>
227 \subsection{Architecture}
228 \label{sec:discussion-mta-arch}
230 %todo: what's this section to do with requirements?
232 \masqmail's current architecture is monolithic like \sendmail's and \exim's. But more than the other two, is it one block of interweaved code. \exim\ has a highly structured code with many internal interfaces, a good example is the one for authentication ``modules''. %fixme: add ref
233 \sendmail\ provides now, with its \name{milter} interface, standardized connection channels to external modules.
234 \masqmail\ has none of them; it is what \sendmail\ was in the beginning: a single large block.
236 Figure \ref{fig:masqmail-arch} is a call graph generated from \masqmail's source code, excluding logging functions. It gives a impression of how interweaved the internals are. There are no compartments existent.
237 %fixme: what is included, what not?
239 \begin{figure}
240 \begin{center}
241 \vspace*{2ex}
242 %\includegraphics[scale=0.75]{img/callgraph.eps}
243 \includegraphics[scale=0.75]{img/masqmail-3-omitlog5.eps}
244 \end{center}
245 \caption{Internal structure of \masqmail, showed by a call graph. (Logging functions are excluded.)}
246 %fixme: what else is excluded
247 \label{fig:masqmail-arch}
248 \end{figure}
250 \sendmail\ improved its old architecture by adding the milter interface, to include further functionality by invoking external programs. \exim\ was designed, and is carefully maintained, with a modular-like code structure in mind. \qmail\ started from scratch with a ``security-first'' approach, \postfix\ improved on it, and \name{sendmail X}/\name{MeTA1} tries to adopt the best of \qmail\ and \postfix\ to completely replace the old \sendmail\ architecture. \person{Hafiz} describes this evolution of \mta\ architecture very well \cite{hafiz05}.
252 Every one of these programs is more modular, or became more modular over time, than \masqmail\ is. Modern requirements like spam protection and future requirements like---probably---the use of new mail transport protocols demand for modular designs in order to keep the software simple. Simplicity is a key property for security. ``the essence of security engineering is to build systems that are as simple as possible.'' \cite[page 45]{graff03}.
254 \person{Hafiz} agrees: ``The goal of making software secure can be better achieved by making the design simple and easier to understand and verify.'' \cite[page 64]{hafiz05}. He identifies the security of \qmail\ to come from it's \name{compartmentalization}, which goes hand in hand with modularity:
255 \begin{quote}
256 A perfect example is the contrast between the feature envy early \sendmail\ architecture implemented as one process and the simple, modular architecture of \qmail. The security of \qmail\ comes from its compartmentalized simple processes that perform one task only and are therefore testable for security.
257 \hfill\cite[page 64]{hafiz05}
258 \end{quote}
259 Equal does \person{Dent} see the situation for \postfix: ``The modular architecture of Postfix forms the basis for much of its security.'' \cite[page 7]{dent04}
261 Modularity is also needed to satisfy modern \MTA\ requirements, in providing a clear interface to add functionality without increasing the overall complexity much.
263 Modularity is no direct requirement, but a goal that has positive influence on important requirements like security, testability, extendability, maintainability, and not least simplicity. These quality properties then, on their part, make achieving the functional requirements easier.
265 Hence, aspiration for modularity, by compartmentalization, improves the overall quality and function of the software. It can be seen as an architectural requirement for a secure and modern \MTA.
267 %fixme: explain: why are compartments and interfaces so good?
277 \section{Fulfilled requirements}
278 \label{sec:fulfilled-requirements}
280 Here follows a description of how far the requirements are already fulfilled by \masqmail.
283 \paragraph{\RF1: In/out channels}
284 The incoming and outgoing channels that \masqmail\ already has are the ones required for an \MTA{}s at the moment. They are depicted in figure \ref{fig:masqmail-in-out} on page \pageref{fig:masqmail-in-out}.
285 Support for other protocols seems not to be necessary at the moment, although new protocols and mailing concepts are likely to appear (see section \ref{sec:electronic-mail}).
286 Today, other protocols are not needed, so \masqmail\ is regarded to fulfill \RF1.
287 But as \masqmail\ has no support for adding further protocols, delaying the work to support them until they are widely used, appears to be the best strategy anyway.
289 << smtp submission >> %fixme
291 \paragraph{\RF2: Queueing}
292 One single mail queue is used in \masqmail; it satisfies all current requirements.
294 << persistence: DB >> %fixme
296 \paragraph{\RF3: Header sanitizing}
297 The envelope and mail headers are generated when the mail is put into the queue. The requirements are fulfilled.
299 \paragraph{\RF4: Aliasing}
300 Aliasing is done on delivery. All common kinds of aliases in the global aliases file are supported. \name{.forward} aliasing is not, but this is less common and seldom used.
302 \paragraph{\RF5: Route management}
303 Setting of the route to use is done on delivery. Headers can get rewritten a second time then. This part does provide all the functionality required.
305 \paragraph{\RF6: Authentication}
306 Static authentication, based on \NAME{IP} addresses, can be achieved with \person{Venema}'s \NAME{TCP} \name{Wrapper} \cite{venema92}, by editing the \path{hosts.allow} and \path{hosts.deny} files. This is only relevant to authenticate host that try to submit mail into the system. Dynamic (secret-based) \SMTP\ authentication is already supported in form of \NAME{SMTP-AUTH} and \SMTP-after-\NAME{POP}, but only for outgoing connections. For incoming connections, only address-based authentication is supported.
308 \paragraph{\RF7: Encryption}
309 Similar is the situation for encryption which is also only available for outgoing channels; here a wrapper application like \name{openssl} is needed. This creates a secure tunnel to send mail trough, but state-of-the-art is using \NAME{STARTTLS}, which is not supported. For incoming channels, no encryption is available. The only possible setup to provide encryption of incoming channels is using an application like \name{stunnel} to translate between the secure connection to the remote host and the \MTA. Unfortunately, this suffers from the problem explained in section \ref{sec:FIXME} and figure \ref{fig:stunnel}. Anyway, this would still be no \NAME{STARTTLS} support.
311 \paragraph{\RF8: Spam handling}
312 \masqmail\ nowadays does not provide special support for spam filtering. Spam prevention by not accepting spam during the \SMTP\ dialog is not possible at all. Spam filtering is only possible by using two \masqmail\ instances with an external spam filter inbetween. The mail flow is from the receiving \MTA\ instance, which accepts mail, to the filter application that processes and possible modifies it, to the second \MTA\ which is responsible for further delivery of the mail. This is a concept that works in general. And it is a good concept in principle to separate work with clear interfaces. But the need of two instances of the same \MTA (each for only half of the job) with doubled setup, is more a work-around. Best is to have this data flow respected in the \MTA\ design, like in \postfix. But the more important part of spam handling, for sure, is done during the \SMTP\ dialog in completely refusing unwanted mail.
314 \paragraph{\RF9: Malware handling}
315 For malware handling applies nearly the same, except all checks are done after mail is accepted. So the possible setup is the same with the two \MTA\ instances and the filter inbetween. \masqmail\ does support such a setup, but not in a nice way.
317 \paragraph{\RF10: Archiving}
318 There is currently no way of archiving every message going through \masqmail.
322 \paragraph{\RG1: Security}
323 \masqmail's current security is bad. However, it seems acceptable for using \masqmail\ on workstations and private networks, if the environment is trustable and \masqmail\ is protected against remote attackers. In environments where untrusted components or persons have access to \masqmail, its security is too low.
324 Its author states it ``is not designed to'' such usage \citeweb{masqmail:homepage2}. This is a clear indicator for being careful. Issues like high memory consumption, low performance, and denial-of-service attacks---things not regarded by design---may cause serious problems. In any way, is a security report missing that confirms \masqmail's security level.
326 \masqmail\ uses conditional compilation to exclude unneeded functionality from the executable at complile time. Excluding code means excluding all bugs and weaknesses within this code too. Excluding unused code is a good concept to improve security.
328 \paragraph{\RG2: Reliability}
329 Similar is its reliability not good enough. Situations where only one part of sent message was removed from the queue, and the other part remained as garbage, showed off \citeweb{debian:bug245882}. Problems with large mail and small bandwidth were also reported \citeweb{debian:bug216226}. Fortunately, lost email was no big problem yet, but \person{Kurth} warns:
330 \begin{quote}
331 There may still be serious bugs in [masqmail], so mail might get lost. But in the nearly two years of its existence so far there was only one time a bug which caused mail retrieved via pop3 to be lost in rare circumstances.
332 \hfill\citeweb{masqmail:homepage2}
333 \end{quote}
334 In summary: Current reliability needs to be improved.
335 %fixme: state machine
337 \masqmail\ uses the filesytem to store the queue, storing the queue in a databases might improve the reliability through better persistence. %fixme
339 \paragraph{\RG3: Robustness}
340 The logging behavior of \masqmail\ is good, although it does not cover all problem situations. For example, if the queue directory is world writeable by accident (or as action of an intruder), any user can remove messages from the queue or replace them with own ones. \masqmail\ does not even write a debug message in this case. The origin of this problem, however, is \masqmail's trust in its environment.
341 %todo: rule of robustness, rule of repair
343 \paragraph{\RG4: Extendability}
344 \masqmail's extendability is very poor. This is a general problem of monolithic software, but can thus be provided with high effort. \exim\ is an example for good extendability in a monolithic program.
346 \paragraph{\RG5: Maintainability}
347 The maintainability of \masqmail\ is equivalent to other software of similar kind. Missing modularity and therefore more complexity makes the maintainer's work harder. Conditional compilation might be good for security, but \name{ifdef}s scattered throughout the source code is a pain for maintainability. In summary is \masqmail's maintainability bearable, like in average Free Software projects.
351 \paragraph{\RG6: Testability}
352 The testability suffers from missing modularity. Testing program parts is hard to do. Nevertheless, it is done by compiling parts of the source to special test programs. %fixme: what are the names? what do they test?
354 This kind of testing is only clean-room testing, so .... %fixme
356 \paragraph{\RG7: Performance}
357 The performance---efficiency---of \masqmail\ is good enough for its target field of operation, where this is a minor goal.
359 \paragraph{\RG8: Availability}
360 This applies equal to availability. Hence no further work needs to be done her.
362 \paragraph{\RG9: Portability}
363 The code's portability is good with view on \unix-like operation systems. At least \name{Debian}, \name{Red Hat}, \NAME{SUSE}, \name{Slackware}, \name{Free}\NAME{BSD}, \name{Open}\NAME{BSD}, and \name{Net}\NAME{BSD} are reported to be able to compile and run \masqmail\ \citeweb{masqmail:homepage2}. Special requirements for the underlying file system are not known. Therefore, the portability is already good.
366 \paragraph{\RG10: Usability}
367 The usability, from the administrator's point of view, is very good. \masqmail\ was developed to suite a specific, limited job---its configuration does perfect match. The user's view does not reach to the \MTA, as it is hidden behind the \name{mail user agent}. But configuration could be eased more, by providing configuration generators to be able to use \masqmail\ right ``out of the box'' after running one of several configuration scripts for common setups. This would improve \masqmail's usability for not technical educated people.
373 \section{Work to do}
375 After the requirements for modern \mta{}s were identified in section \ref{sec:mta-requirements} and \masqmail's features were set against them in section \ref{sec:fulfilled-requirements}, here the work that is left to do is identified. Table \ref{tab:requirements} lists all requirements with importance and the work needed to achieve them. The column ``Focus'' shows the attention a work task should get. The focus depends on the task's importance and the amount of work it includes.
377 \begin{table}
378 \begin{center}
379 \input{tbl/requirements.tbl}
380 \end{center}
381 \caption{Importance of and pending work for requirements}
382 \label{tab:requirements}
383 \end{table}
385 The importance is ranked from `-{}-' (not important) to `++' (very important). The pending work is ranked from `-{}-' (nothing) to `++' (very much). Large work tasks with high importance need to receive much attention, they need to be in focus. In contrast should small low importance work receive few attention. Here the focus for a task is calculated by summing up the importance and the pending work with equal weight. Normally, tasks with high focus are the ones of high priority and should be done first.
387 The functional requirements that receive highest attention are \RF6: authentication, \RF7: encryption, and \RF8: spam handling. Of the non-functional requirements, \RG1: security, \RG2: reliability, and \RG4: Extendability, rank highest.
389 These tasks are presented in more detail in an list of work tasks now. The list is sorted by focus and then by importance.
392 \subsubsection*{\TODO1: Encryption (\RF7)}
393 Encryption is chosen for number one as it is essential to provide privacy. Encryption by using \NAME{STARTTLS} is definitely needed and should be added first. Encrypted data transfer is hardly possible without support for it.
396 \subsubsection*{\TODO2: Authentication (\RF6)}
397 Authentication of incoming \SMTP\ connections is also needed and should be added second. It is important to restrict access and to prevent relaying. For workstations and local networks, it has only medium importance and address-based authentication is sufficient in most times. But secret-based authentication is mandatory to receive mail from the Internet. Additionally it is a guard against spam.
400 \subsubsection*{\TODO3: Security (\RG1)}
401 \masqmail's security is bad, thus the program is forced into a limited field of operation. This field of operation even shrinks as security becomes more important and networking and interaction increases. Save and trusted environment become rare. Thus improving security is an important thing to do. The focus should be on adding compartments to split \masqmail\ into separate modules. (See section \ref{sec:discussion-mta-arch}.) Further more should \masqmail's security be tested throughout to get a definitive view how good it really is and where the weak spots are.
404 \subsubsection*{\TODO4: Reliability (\RG2)}
405 Reliability is also to improve. It is a key quality property for an \MTA, and not good enough in \masqmail. Reliability is strong related to the queue, thus improvements there are favorable. Applying ideas of \name{crash-only software} \cite{candea03} will be a good step. \person{Candea} and \person{Fox} see in killing the process the best way to stop a running program. Doing so inevitably demands for good reliability of the queue, and the startup process inevitably demands for good recovery. The critical situations for reliability are nothing special anymore, they are common. Hence they are regulary tested and will definately work.
408 \subsubsection*{\TODO5: Spam handling (\RF8)}
409 As authentication can be a guard against spam, filter facilities have lower priority. But basic spam filtering and interfaces for external tools should be implemented in future. Configuration guides for a setup using the approach of two \masqmail\ instances with a spam scanner inbetween should be written. And at least a basic kind of spam prevention during the \SMTP\ dialog should be implemented.
412 \subsubsection*{\TODO6: Extendability (\RG4)}
413 \masqmail\ lacks an interface to plug in modules with additional functionality. There exists no add-on or module system. The code is only separated by function to the various source files. Some functional parts can be included or excluded by conditional compilation. But the \name{ifdef}s are scattered through all the code. This situation needs to be improved by collecting related function into single places that interact through clear interfaces with other parts. Also should these interfaces allow efficient adding of further functionality.
424 \section{Ways for further development}
426 Knowing what needs to be done is only one part, the other is deciding \emph{how} to do it by focusing on a global development strategy.
429 \subsection{Possibilities}
431 Futher development of software can always go three different ways:
432 \begin{enumerate}
433 \item[S1:] Improve the current code base.
434 \item[S2:] Add wrappers or interposition filters.
435 \item[S3:] Redesign the software from scratch and rebuild it.
436 \end{enumerate}
438 The first two strategies base on the available source code, and can be applied in combination. The third strategy splits from the old code base and starts over again. Wrappers and interposition filters would be outright included into a new architecture; they are a subset of a new design. Also parts of existing code can be used in a new design if appropriate.
441 The requirements are now regarded each on its own, and are linked to the development strategy that is prefered to reach each specific requirement. If some requirement is well achievable by using different strategies then it is linked to all of them. Implementing encryption (\TODO1) and authentication (\TODO2), for example, are limited to a narrow region in the code. Such features are addable to the current code base without much problem. In contrast can quality properties like reliability (\TODO4), extendability (\TODO6), and maintainability hardly be added to code afterwards---if at all. Security (\TODO3) is addable in a new design, of course, but also with wrappers or interposition filters.
443 This linking of strategies to the requirements is shown in table \ref{tab:strategies}. The requirements are ordered by their focus.
445 \begin{table}
446 \begin{center}
447 \input{tbl/strategies.tbl}
448 \end{center}
449 \caption{Development strategies and their suitability for requirements}
450 \label{tab:strategies}
451 \end{table}
454 Next, the best strategy for further development needs to be discovered. Therefore a score for each strategy is obtained now by summing up the focus points of each requirement for which a strategy is prefered. Herefore only positive focus points are regarded, with each plus symbol counting one. Requirements with negative focus are not regareded because they are already or nearly reached, but the view here is on outstanding work. %(Respecting negative focus points leads to a similar result.)
456 Strategy 1 (Improve current code) has a score of 9 points. Strategy 2 (Wrappers and interposition filters) has a score of 7 points. Strategy 3 (A new design) scores on top with 17 points. \St1 and \St2 can be used in combination; the combined score is 13 points. Thus strategy 3 ranges first, followed by the combination of strategy 1 and 2.
458 This leads to the conclusion, that S3 (A new design) is probably the best strategy for further development. But this result respects only the view on requirements and their relevance. Other factors like development effort and risks are important to think about too. These issues are discussed in the following sections, comparing \St3 against the combination \St1+2.
467 \subsection{Discussion}
470 \subsubsection*{Quality improvements}
472 Most quality properties can hardly be added to a software afterwards. Hence, if reliability, extendability, or maintainability shall be improved, a redesign of \masqmail\ is the best way to take. The wish to improve quality inevitably point towards a modular architecture. Modularity with internal and external interfaces is highly prefered from the architectural point of view (see section \ref{sec:discussion-mta-arch}). The need for further features, especially ones that require changes in \masqmail's structure, support the decision for a new design too. Hence a rewrite is enfavored if \masqmail\ should become a modern \MTA, with good quality properties.
476 \subsubsection*{Security}
478 Similar is the situation for security. Security comes from good design, explain \person{Graff} and \person{van Wyk}:
479 \begin{quote}
480 Good design is the sword and shield of the security-conscious developer. Sound design defends your application from subversion or misuse, protecting your network and the information on it from internal and external attacks alike. It also provides a safe foundation for future extensions and maintenance of the software.
481 %
482 %Bad design makes life easier for attackers and harder for the good guys, especially if it contributes to a false sends of security while obscuring pertinent failings.
483 \hfill\cite[page 55]{graff03}
484 \end{quote}
486 They also suggest to add wrappers and interposition filters \emph{around} applications, but more as repair techniques if it is not possible to design security \emph{into} a software the first way \cite[pages~71--72]{graff03}.
488 \person{Hafiz} adds: ``The major idea is that security cannot be retrofitted \emph{into} an architecture.'' \cite[page 64]{hafiz05} (emphasisis added).
493 \subsubsection*{Effort estimation}
495 Although a strategy might lead to the best result, one may choose another one if the required effort is too high. The effort for a redesign and rebuild is estimated now.
497 \person{Wheeler}'s program \name{sloccount} calculates following estimations for \masqmail's code base as of version 0.2.21 (excluding library code):
499 \codeinput{input/masqmail-sloccount.txt}
501 The development costs in money are not relevant for a \freesw\ project with volunteer developers, but the development time is. About 24 man-months are estimated. The current code base was written almost completely by \person{Oliver Kurth} within four years in his spare time. This means he needed around twice as much time. Of course, he programmed as a volunteer developer not as an employee with eight work-hours per day.
503 Given the assumptions that (1) an equal amount of code needs to be produced for a new designed \masqmail, (2) a third of existing code can be reused plus concepts and knowledge, and (3) development speed is like \person{Kurth}'s. Then it would take between two and three years for one programmer to produce a redesigned new \masqmail\ with the same features that \masqmail\ now has. Less time would be needed if a simpler architecture allows faster development, better testing, and less bugs. Of course more developers would speed it up too.
508 \subsubsection*{Risks}
510 The gained result might still overwights the development effort. But risks are something more to consider.
512 A redesign and rewrite of software from scratch is hard. It takes time to design a new architecture, which then must prove that it is as good as expected. As well is much time and work needed to implement the design, test it, fix bugs, and so on. If flaws in the design appear during prototype implementation, it is necessary to start again.
514 Such a redesign can fail at many points and it is for long unclear if the result is really better than the code that is already existent. Even if the new code is working like expected, it is still not matured.
516 One thing is clear: Doing a redesign and rebuild \emph{is} a risky decision.
520 \subsubsection*{Existing code is precious}
522 If a new design needs much effort and additionally is a risk, what about the existing code base then?
524 Adding new functionality to an existing code base seems to be a secure and cheap strategy. The existing code is known to work and features can often be added in small increments. Risks like wasted effort if a new design fails are hardly existent. And the faults in the current design are already made and most probably fixed.
526 And functionality that is hard to add incrementally into the application, like support for new protocols, may be addable by ``translation programs'' to the outside. \masqmail\ can be secured to a huge amount by guarding it with wrappers that block attackers. Spam and malware scanners can be included by running two instances of \masqmail. All those methods base on the current code which they can indirectly improve.
528 The required effort is probably under one third of a new design and work directly shows results. These are strong arguments against a new design.
533 \subsubsection*{Repairing}
535 Besides these advantages of existing code, one must not forget that further work on it is often repair work. Small bug fixes are not the problem, but adding something for which the software originally was not designed for are problems. Such work often destroys the clear concepts of the software, especially in interweaved monolithic code.
537 Repair strategies are useful, but only in the short-time view and in times of trouble. If the future is bright, however, one does best by investing into a software. As shown in section \ref{sec:market-analysis-conclusion}, the future for \MTA{}s is bright, currently. This means it is time to invest into a redesigning to build up a more modern product.
539 In the author's view is \masqmail\ already needing a redesign since about 2003 when the old design was still quite suitable \dots\ it already delayed too long.
541 %Clinging to much to existing code will be no help, it is an indicator for fear. Having the courage to through bad code away to make it better, shows the view forward.
543 Anyway, further development on base of current code needs to improve the quality properties too. Some quality requirements can be satisfied by adding wrappers or interposition filters from the outside. For those is the development effort approximately equal to a solution with a new design. But for adding quality requirements like extendability or maintainability which affect the source code throughout, the effort does increase with exponential rate as development proceeds. In case these properties get not improved, development will likely come to a dead end sooner or later.
549 \subsubsection*{A guard against dead ends}
551 A new design does protect against such dead ends.
553 Changing requirements are one possible dead end if the software does not evolve with them. A famous example is \sendmail, which had an almost monopoly for a long time. But when security became important \sendmail\ was only repaired instead of removing the problem sources---its unsecure design. Thus security problems reappeared and over the years \sendmail's market share shrinked as more secure \MTA{}s became available. %fixme: declined ??
554 \sendmail's reaction to the new requirements, in form of \name{sendmail X} and \name{MeTA1}, came much to late---the users already switched to other \MTA{}s. Redesigning a software as requirements change helps keeping it alive. % add quote: ``one thing surely remains: change'' (something like that)
556 Another danger is the dead end of complexity which is likely to appear by constantly working on the same code base. It is even more likely if the code base has a monolithic architecture. A good example for simplicity is \qmail\ which consists of small independent modules, each with only about one thousand lines of code. %fixme: proof
557 Such simple code makes it obvious to understand what it does. The \name{suckless} project \citeweb{suckless.org} for example advertizes such a philosophy of small and simple software by following the thoughts of the \unix\ inventors \cite{kernighan84} \cite{kernighan99}. Simple, small, and clear code avoids complexity and is thus also a strong prequisite for security.
563 \subsubsection*{Modularity}
565 The avoidence of dead ends is essential for further development on current code too. Hence it is mandatory to refactor the existing code base sooner or later. Most important is the intention to modularize it, as it improves many quality requirements, eases further development, and essentially improves securtiy.
567 One example how modular structure makes it easy to add further functionality: \person{Sill} describes that integrating the \name{amavis} filter framework into the \qmail\ system can be done by renaming the \name{qmail-queue} module to \name{qmail-queue-real} and renaming the \name{amavis} to \name{qmail-queue} \cite[section~12.7.1]{sill02}. Nothing more in the \qmail\ system needs to be changed. This is a very admirable approach, but only possible in a modular system that consists of independent executables.
569 This thesis showed several times that modularity is the key property for good software design. This property can hardly be retrofitted into software. Hence development on base of current code will need a throughout restructuring too to modularize the source code. Thus a new design is similar to such a throughout refactoring, except without depending on current code.
576 \subsubsection*{Function versus quality}
578 Remarkable is the distribution of functional and non-functional requirements to the strategies. The strategies for current code (\St1+2) have a functional to non-functional ratio of 10 to 3. The new design strategy (\St3) has a ratio of 5 to 12.
580 This leaves current code to be better suited for adding functionality, and a new design to be better suited for quality improvements. Both strategies need to improve function as well as quality, but the difference determines the focus of the strategy.
582 Easier work is likely to be done earlier in Free Software projects than hard work. Thus by choosing \St1+2 volunteer developers tend to implement function first and delay quality improvements, no matter what the suggested order is. \St3 in contrast would support %fixme: beguenstigen
583 early quality improvements and later function improvements. This is real-life experience in Free Software development.
590 \subsubsection*{Break Even}
592 It is important to keep the time dimension in mind. This includes the separation into a short-time and a long-time view. The short-time view shall cover between two and four years. The long-time view is the following time. % fixme: find sources!
594 In the short-time view, the effort for improving the existing code is much smaller than the effort for a new design plus improvements. But to have similar quality properties at the end of the short-time frame, a \masqmail\ that is based on current code will probably require nearly as much effort as a new designed \masqmail\ will take. For all further development afterwards, the new design will scale well while the old code will require exponentiel more work.
596 In the long-time view, a restructuring for modularity is necessary anyway. The question is, when to do it: Right at the start in a new design, or later in some restructuring.
598 %fixme: define exactly, be clear: what does break even here mean
602 \subsubsection*{The problem with ``good enough''}
604 The decision for later restructuring is problematic. Functionality is often more wanted than quality, so further function is prefered over better quality, as quality is still ``good enough''. But it might be still ``good enough'' the next time, and the time after that one, and so on.
606 Quality improving is no popular work but it is required to avoid dead ends. As more code increases quality and modularity improvement work, it is better to do it early. Afterwards all further development profits from it.
608 Also if some design is bad one should never hesitate to erase it and rebuild it in a sane way.
610 However, making such a cut is hard, especially if the bad design is still ``good enough''.
616 \subsubsection*{Good software, good feelings}
618 One last argument shall be added. It is more common for Free Software but can be seen in non-free software too.
620 Free Software ``sells'' if it has a good user base. Although \qmail\ is somehow outdated and its author has released no new version since about 10 years, \qmail\ still has a very strong user base and community.
622 Good concepts, sound design, and a sane philosophy gives users good feelings for the software and faith in it. They become interested in using it and to contribute. In contrast does constantly repairing and reappearing weaknesses leave a bad feeling.
624 The motivation most volunteer developers have is their wish of doing good work to create software of value. Projects that follow admireable plans towards a good product will motivate volunteers to help with it. More helpers can get the 2,5 man-years for a new design in less absolute time done. Additionally is a good developers base the best start for a good user base, and users define a software's value.
633 \section{Result}
635 This chapter identified the requirements and the outstanding work to achieve them. Their importance and the required work on them lead to a focus ranking amoung the requirements, which resulted in a list of tasks to do. Afterwards possible development strategies to control the work process were compared and discussed.
637 Strategy 3 (A new design) is slightly prefered over the combination of strategy 1 (Improve existing code) and 2 (Add wrappers and interposition filters) in regard of the requirements.
639 The discussion afterwards did generally support the new design strategy. But some arguments stand against it. These are:
641 \begin{enumerate}
642 \item The development time and effort
643 \item The time delay until new features can be added
644 \item The risks for failure
645 \end{enumerate}
647 The first two arguments are only relevant for the short-time view, because both will become \emph{support arguments} for the new design, once the Break Even point is reached.
649 The third argument, the risks, remain. There are risk in every investion. Taking no risks means remaining the same, means drifting towards a dead end in a world that does change.
652 With respect to the current situation, the suggested further development plan for \masqmail\ is splitted into a short-time plan and a long-time plan:
654 \begin{enumerate}
655 \item The short-time plan: Add the most needed features, being encryption, authentication, and security wrappers, to the current code base.
656 \item The long-time plan: Design a new architecture that satisfies the modern requirements especially the quality requirements.
657 \end{enumerate}
659 The background thought is to first do the most needed stuff on the existing code to keep %fixme: erhalten
660 it usable. This satisfies the urgent needs and removes the time pressure from the development of the new design. After this is done, a new designed \masqmail\ should be developed from scratch. This is the work for the future. It shall, after it is usable and throughout tested, supersede the old \masqmail.
662 The basic idea is, regularly developing a new design from scratch while the current version is still in use and gets repaired. Hence a modern design will inherit an old one in regular intervals. This is a very future-proove concept that combines the best of both worlds. The price to pay is only the increased work which gets covered %fixme: uebernommen
663 by volunteers that \emph{want} to do it.
667 %fixme: move that sentence to the beginning of the next chapter?
668 The following chapter describes approaches and techniques for the work on the current code base, and it introduces ideas and plans for a new, modern \MTA\ design the next generation of \masqmail.
673 %A program's structure is primary its architecture. Which is the most influencing design decision, and has the greatest impact on the program's future capabilities. The architecture defines what the program can do, and how it can be used. If the architecture does not fit to the requirements, development will reach a dead end \dots\ further work then will make everything worse. The only good solution then is to change the architecture, which, sadly but most likely, means a redesign from scratch.
675 %This plan is similar to the change from \sendmail\ to \name{sendmail X}/\name{MeTA1}, except the \sendmail\ change was much too late.