Config parsing (1/4): General principles
rra at stanford.edu
Thu May 10 16:34:30 UTC 2001
Okay, I'm hitting saturation point with trying to write more stuff down,
so I think it's time to kick all of this out for public comment. This is
the first message of four. The messages are:
1. General principles
2. Proposed syntax
3. Proposed semantics
4. Proposed starting interface
Of these, the syntax is the most fully fleshed out; the rest probably need
more details filled in. All of these are text files so far; some or all
of this documentation will turn into man pages and POD as things are
adopted and get finished.
Anyway, here we go. Comments very much welcome....
This file is documentation of the design principles that went into INN's
configuration file syntax, and some rationale for why those principles
1. All configuration files used by INN should have the same syntax.
This was the root reason why the project was taken on in the first
place; INN developed a proliferation of configuration files, all of
which had a slightly (or greatly) different syntax, forcing the
administrator to learn several different syntaxes and resulting in a
proliferation of parsers, all with their own little quirks.
2. Adding a new configuration file or a new set of configuration options
should not require writing a single line of code for syntax parsing.
Code that analyzes the semantics of the configuration will of course
be necessary, but absolutely no additional code to read files, parse
files, build configuration trees, or the like should be required.
This principle is intended to outlaw the proliferation of parsers;
INN should have a single configuration parser that everything uses.
It's also intended to outlaw syntactic differences between different
3. The syntax should look basically like the syntax of readers.conf,
incoming.conf, and innfeed.conf in INN 2.3. After extensive
discussion on the inn-workers mailing list, this seemed to be the
most generally popular syntax of the ones already used in INN, and
inventing a completely new syntax didn't appear likely to have gains
outweighing the effort involved. This syntax seemed sufficiently
general to represent all of the configuration information that INN
4. The parsing layer should *not* attempt to do semantic analysis of the
configuration; it should concern itself solely with syntax (or very
low-level semantics that are standard across all conceivable INN
configuration files). In particular, the parsing layer should not
know what parameters are valid, what groups are permitted, what types
the values for parameters should have, or what default values
This principle requires some additional explanation, since it is very
tempting to not do things this way. However, the more semantic
information the parser is aware of, the less general the parser is,
and it's very easy to paint oneself into a corner. In particular,
it's *not* a valid assumption that all clients of the parsing code
will want to reduce the configuration to a bunch of structs; this
happens to be true for most clients of inn.conf, for example, but
inndstart doesn't want the code needed to reduce everything to a
struct and set default values to necessarily be executed in a
Additionally, making the parser know more semantic information either
complicates (significantly) the parser interface or means that the
parser has to be modified when the semantics change. The latter is
not acceptable, and the parser interface should be as straightforward
as possible (to encourage all parts of INN to use it).
5. The result of a parse of the configuration file may be represented as
a tree of hash tables, where each hash table corresponds to a group
and each hash key corresponds to a parameter setting. (Note that
this does not assume that the underlying data structure is a hash
table, just that it has hash table-semantics, namely a collection of
key/value pairs with the keys presumed unique.)
6. Parameter values inherit via group nesting. In other words, if a
group is nested inside another group, all parameters defined in the
enclosing group are inherited by the nested group unless they're
explicitly overriden within the nested group. (This point and point
5 are to some degree just corollaries of point 3.)
7. The parsing library must permit writing as well as reading. It must
be possible for a program to read in a configuration file, modify
parameters, add and delete groups, and otherwise change the
configuration, and then write back out to disk a configuration file
that preserves those changes and still remains as faithful to the
original (possibly human-written) configuration file as possible.
(Ideally, this would extend to preserving comments, but that may be
too difficult to do and therefore isn't required.)
8. The parser must not limit the configuration arbitrarily. In
particular, unlimited length strings (within available memory) must
be supported for string values, and if allowable line length is
limited, line continuation must be supported everywhere that there's
any reasonable expectation that it might be necessary. One common
configuration parameter is a list of hosts or host wildmats that can
be almost arbitrarily long, and the syntax and parser must support
9. The parser should be reasonably efficient, enough so as to not cause
an annoying wait for command-line tools like sm and grephistory to
start. In general, though, efficiency in either time or memory is
not as high of a priority as readable, straightforward code; it's
safe to assume that configuration parsing is only done on startup and
at rare intervals and is not on any critical speed paths.
10. Error reporting is a must. It must be possible to clearly report
errors in the configuration files, including at minimum the file name
and line number where the error occurred.
11. The configuration parser should not trust its input, syntax-wise. It
must not segfault, infinitely loop, or otherwise explode on malformed
or broken input. And, as a related point, it's better to be
aggressively picky about syntax than to be lax and attempt to accept
minor violations. The intended configuration syntax is simple and
unambiguous, so it should be unnecessary to accept violations.
12. It must be possible to do comprehensive semantic checks of a
configuration file, including verifying that all provided parameters
are known ones, all parameter values have the correct type, group
types that are not expected to be repeated are not, and only expected
group types are used. This must *not* be done by the parser, but the
parser must provide sufficient hooks that the client program can do
this if it chooses.
Russ Allbery (rra at stanford.edu) <http://www.eyrie.org/~eagle/>
More information about the inn-workers