_________________________________________________________________
Design Notes On Fetchmail
These notes are for the benefit of future hackers and maintainers. The
following sections are both functional and narrative, read from
beginning to end.
History
A direct ancestor of the fetchmail program was originally authored
(under the name popclient) by Carl Harris <ceharris@mal.com>. I took
over development in June 1996 and subsequently renamed the program
`fetchmail' to reflect the addition of IMAP support. In early November
1996 Carl officially ended support for the last popclient versions.
Before accepting responsibility for the popclient sources from Carl, I
had investigated and used and tinkered with every other UNIX
remote-mail forwarder I could find, including fetchpop1.9,
PopTart-0.9.3, get-mail, gwpop, pimp-1.0, pop-perl5-1.2, popc,
popmail-1.6 and upop. My major goal was to get a header-rewrite
feature like fetchmail's working so I wouldn't have reply problems
anymore.
Despite having done a good bit of work on fetchpop1.9, when I found
popclient I quickly concluded that it offered the solidest base for
future development. I was convinced of this primarily by the presence
of multiple-protocol support. The competition didn't do
POP2/RPOP/APOP, and I was already having vague thoughts of maybe
adding IMAP. (This would advance two other goals: learn IMAP and get
comfortable writing TCP/IP client software.)
Until popclient 3.05 I was simply following out the implications of
Carl's basic design. He already had daemon.c in the distribution, and
I wanted daemon mode almost as badly as I wanted the header rewrite
feature. The other things I added were bug fixes or minor extensions.
After 3.1, when I put in SMTP-forwarding support (more about this
below) the nature of the project changed -- it became a
carefully-thought-out attempt to render obsolete every other program
in its class. The name change quickly followed.
The rewrite option
RFC 1123 stipulates that MTAs ought to canonicalize the addresses of
outgoing mail so that From:, To:, Cc:, Bcc: and other address headers
contain only fully qualified domain names. Failure to do so can break
the reply function on many mailers.
This problem only becomes obvious when a reply is generated on a
machine different from where the message was delivered. The two
machines will have different local username spaces, potentially
leading to misrouted mail.
Most MTAs (and sendmail in particular) do not canonicalize address
headers in this way (violating RFC 1123). Fetchmail therefore has to
do it. This is the first feature I added to the ancestral popclient.
Reorganization
The second thing I did reorganize and simplify popclient a lot. Carl
Harris's implementation was very sound, but exhibited a kind of
unnecessary complexity common to many C programmers. He treated the
code as central and the data structures as support for the code. As a
result, the code was beautiful but the data structure design ad-hoc
and rather ugly (at least to this old LISP hacker).
I was able to improve matters significantly by reorganizing most of
the program around the `query' data structure and eliminating a bunch
of global context. This especially simplified the main sequence in
fetchmail.c and was critical in enabling the daemon mode changes.
IMAP support and the method table
The next step was IMAP support. I initially wrote the IMAP code as a
generic query driver and a method table. The idea was to have all the
protocol-independent setup logic and flow of control in the driver,
and the protocol-specific stuff in the method table.
Once this worked, I rewrote the POP3 code to use the same
organization. The POP2 code kept its own driver for a couple more
releases, until I found sources of a POP2 server to test against (the
breed seems to be nearly extinct).
The purpose of this reorganization, of course, is to trivialize the
development of support for future protocols as much as possible. All
mail-retrieval protocols have to have pretty similar logical design by
the nature of the task. By abstracting out that common logic and its
interface to the rest of the program, both the common and
protocol-specific parts become easier to understand.
Furthermore, many kinds of new features can instantly be supported
across all protocols by modifying the one driver module.
Implications of smtp forwarding
The direction of the project changed radically when Harry Hochheiser
sent me his scratch code for forwarding fetched mail to the SMTP port.
I realized almost immediately that a reliable implementation of this
feature would make all the other delivery modes obsolete.
Why mess with all the complexity of configuring an MDA or setting up
lock-and-append on a mailbox when port 25 is guaranteed to be there on
any platform with TCP/IP support in the first place? Especially when
this means retrieved mail is guaranteed to look like normal sender-
initiated SMTP mail, which is really what we want anyway.
Clearly, the right thing to do was (1) hack SMTP forwarding support
into the generic driver, (2) make it the default mode, and (3)
eventually throw out all the other delivery modes.
I hesitated over step 3 for some time, fearing to upset long-time
popclient users dependent on the alternate delivery mechanisms. In
theory, they could immediately switch to .forward files or their
non-sendmail equivalents to get the same effects. In practice the
transition might have been messy.
But when I did it (see the NEWS note on the great options massacre)
the benefits proved huge. The cruftiest parts of the driver code
vanished. Configuration got radically simpler -- no more grovelling
around for the system MDA and user's mailbox, no more worries about
whether the underlying OS supports file locking.
Also, the only way to lose mail vanished. If you specified localfolder
and the disk got full, your mail got lost. This can't happen with SMTP
forwarding because your SMTP listener won't return OK unless the
message can be spooled or processed.
Also, performance improved (though not so you'd notice it in a single
run). Another not insignificant benefit of this change was that the
manual page got a lot simpler.
Later, I had to bring --mda back in order to allow handling of some
obscure situations involving dynamic SLIP. But I found a much simpler
way to do it.
The moral? Don't hesitate to throw away superannuated features when
you can do it without loss of effectiveness. I tanked a couple I'd
added myself and have no regrets at all. As Saint-Exupery said,
"Perfection [in design] is achieved not when there is nothing more to
add, but rather when there is nothing more to take away." This program
isn't perfect, but it's trying.
The most-requested features that I will never add, and why not:
Password encryption in .fetchmailrc
The reason there's no facility to store passwords encrypted in the
.fetchmailrc file is because this doesn't actually add protection.
Anyone who's acquired the 0600 permissions needed to read your
.fetchmailrc file will be able to run fetchmail as you anyway -- and
if it's your password they're after, they'd be able to rip the
necessary decoder out of the fetchmail code itself to get it.
All .fetchmailrc encryption would do is give a false sense of security
to people who don't think very hard.
Truly concurrent queries to multiple hosts
Occasionally I get a request for this on "efficiency" grounds. These
people aren't thinking either. True concurrency would do nothing to
lessen fetchmail's total IP volume. The best it could possibly do is
change the usage profile to shorten the duration of the active part of
a poll cycle at the cost of increasing its demand on IP volume per
unit time.
If one could thread the protocol code so that fetchmail didn't block
on waiting for a protocol response, but rather switched to trying to
process another host query, one might get an efficiency gain (close to
constant loading at the single-host level).
Fortunately, I've only seldom seen a server that incurred significant
wait time on an individual response. I judge the gain from this not
worth the hideous complexity increase it would require in the code.
Multiple concurrent instances of fetchmail
Fetchmail locking is on a per-invoking-user because finer-grained
locks would be really hard to implement in a portable way. The problem
is that you don't want two fetchmails querying the same site for the
same remote user at the same time.
To handle this optimally, multiple fetchmails would have to associate
a system-wide semaphore with each active pair of a remote user and
host canonical address. A fetchmail would have to block until getting
this semaphore at the start of a query, and release it at the end of a
query.
This would be way too complicated to do just for an "it might be nice"
feature. Instead, you can run a single root fetchmail polling for
multiple users in either single-drop or multidrop mode.
The fundamental problem here is how an instance of fetchmail polling
host foo can assert that it's doing so in a way visible to all other
fetchmails. System V semaphores would be ideal for this purpose, but
they're not portable.
I've thought about this a lot and roughed up several designs. All are
complicated and fragile, with a bunch of the standard problems (what
happens if a fetchmail aborts before clearing its semaphore, and how
do we recover reliably?).
I'm just not satisfied that there's enough functional gain here to pay
for the large increase in complexity that adding these semaphores
would entail.
Multidrop and alias handling
I decided to add the multidrop support partly because some users were
clamoring for it, but mostly because I thought it would shake bugs out
of the single-drop code by forcing me to deal with addressing in full
generality. And so it proved.
There are two important aspects of the features for handling
multiple-drop aliases and mailing lists which future hackers should be
careful to preserve.
1. The logic path for single-recipient mailboxes doesn't involve
header parsing or DNS lookups at all. This is important -- it
means the code for the most common case can be much simpler and
more robust.
2. The multidrop handing does not rely on doing the equivalent of
passing the message to sendmail -oem -t. Instead, it explicitly
mines members of a specified set of local usernames out of the
header.
3. We do not attempt delivery to multidrop mailboxes in the presence
of DNS errors. Before each multidrop poll we probe DNS to see if
we have a nameserver handy. If not, the poll is skipped. If DNS
crashes during a poll, the error return from the next nameserver
lookup aborts message delivery and ends the poll. The daemon mode
will then quietly spin until DNS comes up again, at which point it
will resume delivering mail.
When I designed this support, I was terrified of doing anything that
could conceivably cause a mail loop (you should be too). That's why
the code as written can only append local names (never @-addresses) to
the recipients list.
The code in mxget.c is nasty, no two ways about it. But it's utterly
necessary, there are a lot of MX pointers out there. It really ought
to be a (documented!) entry point in the bind library.
DNS error handling
Fetchmail's behavior on DNS errors is to suppress forwarding and
deletion of the individual message that each occurs in, leaving it
queued on the server for retrieval on a subsequent poll. The
assumption is that DNS errors are transient, due to temporary server
outages.
Unfortunately this means that if a DNS error is permanent a message
can be perpetually stuck in the server mailbox. We've had a couple bug
reports of this kind due to subtle RFC822 parsing errors in the
fetchmail code that resulted in impossible things getting passed to
the DNS lookup routines.
Alternative ways to handle the problem: ignore DNS errors (treating
them as a non-match on the mailserver domain), or forward messages
with errors to fetchmail's invoking user in addition to any other
recipients. These would fit an assumption that DNS lookup errors are
likely to be permanent problems associated with an address.
IPv6 and IPSEC
The IPv6 support patches are really more protocol-family independence
patches. Because of this, in most places, "ports" (numbers) have been
replaced with "services" (strings, that may be digits). This allows us
to run with certain protocols that use strings as "service names"
where we in the IP world think of port numbers. Someday we'll plumb
strings all over and then, if inet6 is not enabled, do a
getservbyname() down in SocketOpen. The IPv6 support patches use
getaddrinfo(), which is a POSIX p1003.1g mandated function. So, in the
not too distant future, we'll zap the ifdefs and just let autoconf
check for getaddrinfo. IPv6 support comes pretty much automatically
once you have protocol family independence.
Craig Metz used his inner_connect() function to handle most of the
connect work. This is a nonstandard function not likely to ever exist
in a system's libc, but we can just include that source file if the
day comes when we want to support IPv6 without the inet6-apps library.
It just makes life easier.
Internationalization
Internationalization is handled using GNU gettext (see the file
ABOUT_NLS in the source distribution). This places some minor
constraints on the code.
Strings that must be subject to translation should be wrapped with _()
or N_() -- the former in functuib arguments, the latter in static
initializers and other non-function-argument contexts.
To test the support
Checklist for Adding Options
Adding a control option is not complicated in principle, but there are
a lot of fiddly details in the process. You'll need to do the
following minimum steps.
* Add a field to represent the control in struct run, struct query,
or struct hostdata.
* Go to rcfile_y.y. Add the token to the grammar. Don't forget the
%token declaration.
* Pick an actual string to declare the option in the .fetchmailrc
file. Add the token to rcfile_l.
* Pick a long-form option name, and a one-letter short option if any
are left. Go to options.c. Pick a new LA_ value. Hack the
longoptions table to set up the association. Hack the big switch
statement to set the option. Hack the `?' message to describe it.
* If the default is nonzero, set it in def_opts near the top of
load_params in fetchmail.c.
* Add code to dump the option value in fetchmail.c:dump_params.
* Add proper FLAG_MERGE actions in fetchmail.c's optmerge()
function. (If it's a global option, add an override at the end of
load_params.
* Document the option in fetchmail.man. This will require at least
two changes; one to the collected table of options, and one full
text description of the option.
* Add the new token and a brief description to the header comment of
sample.rcfile.
* Hack fetchmailconf to configure it. Bump the fetchmailconf
version.
* Hack conf.c to dump the option so we won't have a version-skew
problem.
* Add an entry to NEWS.
There may be other things you have to do in the way of logic, of
course.
Lessons learned
1. Server-side state is essential
The person(s) responsible for removing LAST from POP3 deserve to
suffer. Without it, a client has no way to know which messages in a
box have been read by other means, such as an MUA running on the
server.
The POP3 UID feature described in RFC1725 to replace LAST is
insufficient. The only problem it solves is tracking which messages
have been read by this client -- and even that requires tricky,
fragile implementation.
The underlying lesson is that maintaining accessible server-side
`seen' state bits associated with Status headers is indispensible in a
Unix/RFC822 mail server protocol. IMAP gets this right.
2. Readable text protocol transactions are a Good Thing
A nice thing about the general class of text-based protocols that
SMTP, POP2, POP3, and IMAP belongs to is that client/server
transactions are easy to watch and transaction code correspondingly
easy to debug. Given a decent layer of socket utility functions (which
Carl provided) it's easy to write protocol engines and not hard to
show that they're working correctly.
This is an advantage not to be despised! Because of it, this project
has been interesting and fun -- no serious or persistent bugs, no long
hours spent looking for subtle pathologies.
3. IMAP is a Good Thing.
If there were a standard IMAP equivalent of the POP3 APOP validation,
POP3 would be completely obsolete.
4. SMTP is the Right Thing
In retrospect it seems clear that this program (and others like it)
should have been designed to forward via SMTP from the beginning. This
lesson may be applicable to other Unix programs that now call the
local MDA/MTA as a program.
5. Syntactic noise can be your friend
The optional `noise' keywords in the rc file syntax started out as a
late-night experiment. The English-like syntax they allow is
considerably more readable than the traditional terse keyword-value
pairs you get when you strip them all out. I think there may be a
wider lesson here.
Motivation and validation
It is truly written: the best hacks start out as personal solutions to
the author's everyday problems, and spread because the problem turns
out to be typical for a large class of users. So it was with Carl
Harris and the ancestral popclient, and so with me and fetchmail.
It's gratifying that fetchmail has become so popular. Until just
before 1.9 I was designing strictly to my own taste. The multi-drop
mailbox support and the new --limit option were the first features to
go in that I didn't need myself.
By 1.9, four months after I started hacking on popclient and a month
after the first fetchmail release, there were literally a hundred
people on the fetchmail-friends contact list. That's pretty powerful
motivation. And they were a good crowd, too, sending fixes and
intelligent bug reports in volume. A user population like that is a
gift from the gods, and this is my expression of gratitude.
The beta testers didn't know it at the time, but they were also the
subjects of a sociological experiment. The results are described in my
paper, The Cathedral And The Bazaar.
Credits
Special thanks go to Carl Harris, who built a good solid code base and
then tolerated me hacking it out of recognition. And to Harry
Hochheiser, who gave me the idea of the SMTP-forwarding delivery mode.
Other significant contributors to the code have included Dave
Bodenstab (error.c code and --syslog), George Sipe (--monitor and
--interface), Gordon Matzigkeit (netrc.c), Al Longyear (UIDL support),
Chris Hanson (Kerberos V4 support), and Craig Metz (OPIE, IPv6,
IPSEC).
Conclusion
At this point, the fetchmail code appears to be pretty stable. It will
probably undergo substantial change only if and when support for a new
retrieval protocol or authentication method is added.
Relevant RFCS
Not all of these describe standards explicitly used in fetchmail, but
they all shaped the design in one way or another.
RFC821
SMTP protocol
RFC822
Mail header format
RFC937
Post Office Protocol - Version 2
RFC974
MX routing
RFC976
UUCP mail format
RFC1081
Post Office Protocol - Version 3
RFC1123
Host requirements (modifies 821, 822, and 974)
RFC1176
Interactive Mail Access Protocol - Version 2
RFC1203
Interactive Mail Access Protocol - Version 3
RFC1225
Post Office Protocol - Version 3
RFC1344
Implications of MIME for Internet Mail Gateways
RFC1413
Identification server
RFC1428
Transition of Internet Mail from Just-Send-8 to 8-bit SMTP/MIME
RFC1460
Post Office Protocol - Version 3
RFC1521
MIME: Multipurpose Internet Mail Extensions
RFC1869
SMTP Service Extensions (ESMTP spec)
RFC1652
SMTP Service Extension for 8bit-MIMEtransport
RFC1725
Post Office Protocol - Version 3
RFC1730
Interactive Mail Access Protocol - Version 4
RFC1731
IMAP4 Authentication Mechanisms
RFC1732
IMAP4 Compatibility With IMAP2 And IMAP2bis
RFC1734
POP3 AUTHentication command
RFC1870
SMTP Service Extension for Message Size Declaration
RFC1891
SMTP Service Extension for Delivery Status Notifications
RFC1892
The Multipart/Report Content Type for the Reporting of Mail
System Administrative Messages
RFC1894
An Extensible Message Format for Delivery Status Notifications
RFC1893
Enhanced Mail System Status Codes
RFC1894
An Extensible Message Format for Delivery Status Notifications
RFC1938
A One-Time Password System
RFC1939
Post Office Protocol - Version 3
RFC1985
SMTP Service Extension for Remote Message Queue Starting
RFC2033
Local Mail Transfer Protocol
RFC2060
Internet Message Access Protocol - Version 4rev1
RFC2061
IMAP4 Compatibility With IMAP2bis
RFC2062
Internet Message Access Protocol - Obsolete Syntax
RFC2449
IMAP/POP AUTHorize Extension for Simple Challenge/Response
RFC2449
POP3 Extension Mechanism
_________________________________________________________________
Eric S. Raymond <esr@snark.thyrsus.com>