Categories: Development
There's nothing fundamentally broken with current pkgutils, however quite many ideas developed throughout the last couple of years. This page tries to provide a high level summary of those to get an overview
Note that the features listed here are not necessarily going to go into into pkgutils, so don't be afraid. The developers are destined to keep the original CRUX goals of simplicity by design, however the requirements have changed quite a bit over the years, so some changes are strongly wanted.
Currently, the package database contains the package name, version, and stores the files which belong to a package. There are several use cases where more information could be interesting to have, although not necessarily everyone will need it. The idea here (suggested by Per and discussed at CRUXCon 2004) was to introduce a generic mechanism called "attributes" which are integrated in a binary package, and can then at will be moved into the package database; this allows to satisfy the requirements for the following use cases I consider relevant for CRUX:
use case | attribute |
---|---|
plain source based crux | name,version,release,files |
binary desktop crux | name,vers,rel,deps,desc,alias |
embedded binary crux | name,vers,rel,files |
server crux | name,vers,rel,files,alias,md5sum |
It's important to make sure the users can't shoot himself too easily, so configuring the required attributes at compile time might be a good approach.
Note that attributes are simply key value pairs, and therefore pkgutils doesn't have to be changed to introduce new attributes (unless it internally depends on their values, of course).
Random meta data could be added to packages with hooks, allowing users to build customized packages if needed. Examples for this are host, build user, build time, maintainer, packager, collection (ugh, where might this package come from) etc.
Shipping the meta data with the package seems inevitable for many use cases; this would then also support *-install and *-remove scripts
We considered adding a magical /ATTRIBUTES directory to the packages, where attribute data is stored; in the worst case tar -C / would just create this directory, while keeping the advantage of having a very simple package format (easy to construct by hand if needed).
When pkgutils was originally written, only scripts would use the pkgdb. Nowadays, things are a bit different: tools like prt-get parse the pkgdb, and several tools call pkgutils as a process to obtain infomation. Exporting an API to do things like pkgadd(path), or pkgdb querying would help application developers.
ports and prt-get both support diff'ing and listing, however ports sources all Pkgfiles (which is a potential security problem) and doesn't cope very well with duplicate ports. Mark Rosenstand has suggested to merge ports into prt-get, which is probably a reasonable request for users mostly using prt-get to maintain their system; it would allow to do things like caching on updates etc. Another scenario could be to keep ports for the networking part, but get rid of the listing and diff functionality. Finally, a sync command could be added to prt-get to either call ports or the drivers directly.
prt-get contains some code to handle dependencies, or parse Pkgfiles. It also has code for aliasing which is definitely in the wrong place there. The same holds true for locking, which currently only affects prt-get, but not pkgrm (i.e. when removing a package with pkgrm, it's not unlocked).
Also, caching you be done right.
This is one of the features which is probably overdesign, but having a common API for pkgdb access would allow to introduce a DB abstraction layer, allowing people to play with different backends-
pkginfo -o and pkgadd -u are current broken WRT symlinks
Mark Rosenstand mentions that packages could use "needs:" and "wants:" for required/optional dependencies, and that if (and only if) we want to separate build time deps from run time deps, it might be a good idea to prefix them, like:
# Needs: curl b:scons b:doxygen # Wants: libgnomecanvas
My (Johannes') take: I don't think the separation is needed right now (and Mark agrees with that :-)), however the prefixing is pretty clean, and since adding meta data to packages would allow to design alternative flavours of CRUX which use binary packages, separation of the two would be a nice thing.
Allow users to omit installation of certain files/directories based on patterns, like this:
INSTALL ^usr/info.*$ NO INSTALL ^usr/share/bash-completion/*$ YES
In combination with this, our package guidelines should require maintainers to install those files in question to make sure people actually get them. This concept is not meant to replace the no-junk rule, but to solve cases where there's no clearly superious solution (info pages or KDE doc for example).
It should be considered whether this needs to be on a per port basis (i.e. install info pages for selected ports only); one possibility would be to extend the above syntax to accept two regexes:
INSTALL .* ^usr/info.*$ NO INSTALL grub ^usr/info.*$ YES
(last rule matches)
Currently, pkgadd runs ldconfig after installing libraries, which is very useful. There are different cases which could benefit from such hooks, like:
The question is whether the concept of running hooks from pkgadd should be extracted, such that you define the following in pkgadd.conf:
HOOK ^*.\.so handle_shared_lib HOOK ^/lib/modules/*.\.ko handle_depmod
The handle scripts would be in a predefined location, like /etc/pkgadd-hooks. This way, packages can provide such hooks, but we still let the user decide whether he/she wants to use them. Also, if you don't like this, you can always use INSTALL ^etc/pkgadd-hooks.* NO
:-).
Per wanted to rewrite pkgadd in C, mainly to not link against libstdc++ anymore; porting to http://www.exactcode.de/embeddedSTL/ might be an alternative. Rewriting pkgmk was not planned as far as I (Johannes) know.
Han rewrote everything in sh.
This page lists lots of shiny new ideas, and the one thing we should always ask ourselves is "are we building the next emerge clone?". Every feature we add means added complexity and cost of maintenance. Let's not forget that while looking at the nice use cases, especially those which have no direct customer (for example the binary package CRUX flavours).
Brief listing of discussion points
I'm adding this section since several CRUX developers have different ideas where this should go to, and since I tried to keep the former section free from personal feelings.
I think pkgutils are not broken right now.
My goals are:
I believe that