* apt-pkg/deb/dpkgpm.cc:

[ntk/apt.git] / doc / dpkg-tech.sgml

<!doctype debiandoc  PUBLIC  "-//DebianDoc//DTD DebianDoc//EN">
<book>
<title>dpkg technical manual</title>

<author>Tom Lees <email>tom@lpsg.demon.co.uk</email></author>
<version>$Id: dpkg-tech.sgml,v 1.3 2003/02/12 15:05:45 doogie Exp $</version>

<abstract>
This document describes the minimum necessary workings for the APT dselect
replacement. It gives an overall specification of what its external interface
must look like for compatibility, and also gives details of some internal
quirks.
</abstract>

<copyright>
Copyright &copy; Tom Lees, 1997.
<p>
APT and this document are free software; you can redistribute them and/or
modify them under the terms of the GNU General Public License as published
by the Free Software Foundation; either version 2 of the License, or (at your
option) any later version.

<p>
For more details, on Debian GNU/Linux systems, see the file
/usr/share/common-licenses/GPL for the full license.
</copyright>

<toc sect>

<chapt>Quick summary of dpkg's external interface
<sect id="control">Control files

<p>
The basic dpkg package control file supports the following major features:-

<list>
<item>5 types of dependencies:-
	<list>
	<item>Pre-Depends, which must be satisfied before a package may be
	unpacked
	<item>Depends, which must be satisfied before a package may be
	configured
	<item>Recommends, to specify a package which if not installed may
	severely limit the usefulness of the package
	<item>Suggests, to specify a package which may increase the
	productivity of the package
	<item>Conflicts, to specify a package which must NOT be installed
	in order for the package to be configured
	</list>
Each of these dependencies can specify a version and a depedency on that
version, for example "<= 0.5-1", "== 2.7.2-1", etc. The comparators available
are:-
	<list>
	<item>"&lt;&lt;" - less than
	<item>"&lt;=" - less than or equal to
	<item>"&gt;&gt;" - greater than
	<item>"&gt;=" - greater than or equal to
	<item>"==" - equal to
	</list>
<item>The concept of "virtual packages", which many other packages may provide,
using the Provides mechanism. An example of this is the "httpd" virtual package,
which all web servers should provide. Virtual package names may be used in
dependency headers. However, current policy is that virtual packages do not
support version numbers, so dependencies on virtual packages with versions
will always fail.
<item>Several other control fields, such as Package, Version, Description,
Section, Priority, etc., which are mainly for classification purposes. The
package name must consist entirely of lowercase characters, plus the characters
'+', '-', and '.'. Fields can extend across multiple lines - on the second
and subsequent lines, there is a space at the beginning instead of a field
name and a ':'. Empty lines must consist of the text " .", which will be
ignored, as will the initial space for other continuation lines. This feature
is usually only used in the Description field.
</list>

<sect>The dpkg status area

<p>
The "dpkg status area" is the term used to refer to the directory where dpkg
keeps its various status files (GNU would have you call it the dpkg shared
state directory). This is always, on Debian systems, /var/lib/dpkg. However,
the default directory name should not be hard-coded, but #define'd, so that
alteration is possible (it is available via configure in dpkg 1.4.0.9 and
above). Of course, in a library, code should be allowed to override the
default directory, but the default should be part of the library (so that
the user may change the dpkg admin dir simply by replacing the library).

<p>
Dpkg keeps a variety of files in its status area. These are discussed later
on in this document, but a quick summary of the files is here:-

<list>
<item>available - this file contains a concatenation of control information
from all the packages which dpkg knows about. This is updated using the dpkg
commands "--update-avail &lt;file&gt;", "--merge-avail &lt;file&gt;", and
"--clear-avail".
<item>status - this file contains information on the following things for
every package:-
	<list>
	<item>Whether it is installed, not installed, unpacked, removed,
		failed configuration, or half-installed (deconfigured in
		favour of another package).
	<item>Whether it is selected as install, hold, remove, or purge.
	<item>If it is "ok" (no installation problems), or "not-ok".
	<item>It usually also contains the section and priority (so that
		dselect may classify packages not in available)
	<item>For packages which did not initially appear in the "available"
		file when they were installed, the other control information
		for them.
	</list>
   <p>
   The exact format for the "Status:" field is:
   <example>
      Status: Want Flag Status
   </example>
   Where <var>Want</> may be one of <em>unknown</>, <em>install</>,
   <em>hold</>, <em>deinstall</>, <em>purge</>. <var>Flag</>
   may be one of <em>ok</>, <em>reinstreq</>, <em>hold</>,
   <em>hold-reinstreq</>.
   <var>Status</> may be one of <em>not-installed</>, <em>unpacked</>, 
   <em>half-configured</>, <em>installed</>, <em>half-installed</>
   <em>config-files</>, <em>post-inst-failed</>, <em>removal-failed</>.
   The states are as follows:-
   <taglist>
     <tag>not-installed
     <item>No files are installed from the package, it has no config files
        left, it uninstalled cleanly if it ever was installed.
     <tag>unpacked
     <item>The basic files have been unpacked (and are listed in
        /var/lib/dpkg/info/[package].list. There are config files present,
        but the postinst script has _NOT_ been run.
     <tag>half-configured
     <item>The package was installed and unpacked, but the postinst script
        failed in some way.
     <tag>installed
     <item>All files for the package are installed, and the configuration
        was also successful.
     <tag>half-installed
     <item>An attempt was made to remove the packagem but there was a failure
        in the prerm script.
     <tag>config-files
     <item>The package was "removed", not "purged". The config files are left,
        but nothing else.
     <tag>post-inst-failed
     <item>Old name for half-configured. Do not use.
     <tag>removal-failed
     <item>Old name for half-installed. Do not use.
   </taglist>
   The two last items are only left in dpkg for compatibility - they are
   understood by it, but never written out in this form.

   <p>
   Please see the dpkg source code, <tt>lib/parshelp.c</tt>, 
   <em>statusinfos</>, <em>eflaginfos</> and <em>wantinfos</> for more 
   details.
   
<item>info - this directory contains files from the control archive of every
package currently installed. They are installed with a prefix of "&lt;packagename&gt;.".
In addition to this, it also contains a file called &lt;package&gt;.list for every
package, which contains a list of files. Note also that the control file is
not copied into here; it is instead found as part of status or available.
<item>methods - this directory is reserved for "method"-specific files - each
"method" has a subdirectory underneath this directory (or at least, it can
have). In addition, there is another subdirectory "mnt", where misc.
filesystems (floppies, CDROMs, etc.) are mounted.
<item>alternatives - directory used by the "update-alternatives" program. It
contains one file for each "alternatives" interface, which contains information
about all the needed symlinked files for each alternative.
<item>diversions - file used by the "dpkg-divert" program. Each diversion takes
three lines. The first is the package name (or ":" for user diversion), the
second the original filename, and the third the diverted filename.
<item>updates - directory used internally by dpkg. This is discussed later,
in the section <ref id="updates">.
<item>parts - temporary directory used by dpkg-split
</list>

<sect>The dpkg library files

<p>
These files are installed under /usr/lib/dpkg (usually), but
/usr/local/lib/dpkg is also a possibility (as Debian policy dictates). Under
this directory, there is a "methods" subdirectory. The methods subdirectory
in turn contains any number of subdirectories for each general method
processor (note that one set of method scripts can, and is, used for more than
one of the methods listed under dselect).

<p>
The following files may be found in each of these subdirectories:-

<list>
<item>names - One line per method, two-digit priority to appear on menu
at beginning, followed by a space, the name, and then another space and the
short description.
<item>desc.&lt;name&gt; - Contains the long description displayed by dselect
when the cursor is put over the &lt;name&gt; method.
<item>setup - Script or program which sets up the initial values to be used
by this method. Called with first argument as the status area directory
(/var/lib/dpkg), second argument as the name of the method (as in the directory
name), and the third argument as the option (as in the names file).
<item>install - Script/program called when the "install" option of dselect is
run with this method. Same arguments as for setup.
<item>update - Script/program called when the "update" option of dselect is
run. Same arguments as for setup/install.
</list>

<sect>The "dpkg" command-line utility

<sect1>"Documented" command-line interfaces

<p>
As yet unwritten. You can refer to the other manuals for now. See
<manref name="dpkg" section="8">.

<sect1>Environment variables which dpkg responds to

<p>
<list>
<item>DPKG_NO_TSTP - if set to a non-null value, this variable causes dpkg to
run a child shell process instead of sending itself a SIGTSTP, when the user
selects to background the dpkg process when it asks about conffiles.
<item>SHELL - used to determine which shell to run in the case when
DPKG_NO_TSTP is set.
<item>CC - used as the C compiler to call to determine the target architecture.
The default is "gcc".
<item>PATH - dpkg checks that it can find at least the following files in the
path when it wants to run package installation scripts, and gives an error if
it cannot find all of them:-
	<list>
	<item>ldconfig
	<item>start-stop-daemon
	<item>install-info
	<item>update-rc.d
	</list>
</list>

<sect1>Assertions

<p>
The dpkg utility itself is required for quite a number of packages, even if
they have been installed with a tool totally separate from dpkg. The reason for
this is that some packages, in their pre-installation scripts, check that your
version of dpkg supports certain features. This was broken from the start, and
it should have actually been a control file header "Dpkg-requires", or similar.
What happens is that the configuration scripts will abort or continue according
to the exit code of a call to dpkg, which will stop them from being wrongly
configured.

<p>
These special command-line options, which simply return as true or false are
all prefixed with "--assert-". Here is a list of them (without the prefix):-

<list>
<item>support-predepends - Returns success or failure according to whether
a version of dpkg which supports predepends properly (1.1.0 or above) is
installed, according to the database.
<item>working-epoch - Return success or failure according to whether a version
of dpkg which supports epochs in version properly (1.4.0.7 or above) is
installed, according to the database.
</list>

<p>
Both these options check the status database to see what version of the "dpkg"
package is installed, and check it against a known working version.

<sect1>--predep-package

<p>
This strange option is described as follows in the source code:

<example>
/* Print a single package which:
 *  (a) is the target of one or more relevant predependencies.
 *  (b) has itself no unsatisfied pre-dependencies.
 * If such a package is present output is the Packages file entry,
 *  which can be massaged as appropriate.
 * Exit status:
 *  0 = a package printed, OK
 *  1 = no suitable package available
 *  2 = error
 */
</example>

<p>
On further inspection of the source code, it appears that what is does is
this:-

<list>
<item>Looks at the packages in the database which are selected as "install",
and are installed.
<item>It then looks at the Pre-Depends information for each of these packages
from the available file. When it find a package for which any of the
pre-dependencies are not satisfied, it breaks from the loop through the packages.
<item>It then looks through the unsatisfied pre-dependencies, and looks for
packages which would satisfy this pre-dependency, stopping on the first it
finds. If it finds none, it bombs out with an error.
<item>It then continues this for every dependency of the initial package.
</list>

Eventually, it writes out the record of all the packages to satisfy the
pre-dependencies. This is used by the disk method to make sure that its
dependency ordering is correct. What happens is that all pre-depending
packages are first installed, then it runs dpkg -iGROEB on the directory,
which installs in the order package files are found. Since pre-dependencies
mean that a package may not even be unpacked unless they are satisfied, it is
necessary to do this (usually, since all the package files are unpacked in one
phase, the configured in another, this is not needed).

<chapt>dpkg-deb and .deb file internals

<p>
This chapter describes the internals to the "dpkg-deb" tool, which is used
by "dpkg" as a back-end. dpkg-deb has its own tar extraction functions, which
is the source of many problems, as it does not support long filenames, using
extension blocks.

<sect>The .deb archive format

<p>
The main principal of the new-format Debian archive (I won't describe the old
format - for that have a look at deb-old.5), is that the archive really is
an archive - as used by "ar" and friends. However, dpkg-deb uses this format
internally, rather than calling "ar". Inside this archive, there are usually
the folowing members:-

<list>
<item>debian-binary
<item>control.tar.gz
<item>data.tar.gz
</list>

<p>
The debian-binary member consists simply of the string "2.0", indicating the
format version. control.tar.gz contains the control files (and scripts), and
the data.tar.gz contains the actual files to populate the filesystem with.
Both tarfiles extract straight into the current directory. Information on the
tar formats can be found in the GNU tar info page. Since dpkg-deb calls
"tar -cf" to build packages, the Debian packages use the GNU extensions.

<sect>The dpkg-deb command-line 

<p>
dpkg-deb documents itself thoroughly with its '--help' command-line option.
However, I am including a reference to these for completeness. dpkg-deb
supports the following options:-

<list>
<item>--build (-b) &lt;dir&gt; - builds a .deb archive, takes a directory which
contains all the files as an argument. Note that the directory
&lt;dir&gt;/DEBIAN will be packed separately into the control archive.
<item>--contents (-c) &lt;debfile&gt; - Lists the contents of ther "data.tar.gz"
member.
<item>--control (-e) &lt;debfile&gt; - Extracts the control archive into a
directory called DEBIAN. Alternatively, with another argument, it will extract
it into a different directory.
<item>--info (-I) &lt;debfile&gt; - Prints the contents of the "control" file
in the control archive to stdout. Alternatively, giving it other arguments will
cause it to print the contents of those files instead.
<item>--field (-f) &lt;debfile&gt; &lt;field&gt; ... - Prints any number of
fields from the "control" file. Giving it extra arguments limits the fields it
prints to only those specified. With no command-line arguments other than a
filename, it is equivalent to -I and just the .deb filename.
<item>--extract (-x) &lt;debfile&gt; &lt;dir&gt; - Extracts the data archive
of a debian package under the directory &lt;dir&gt;.
<item>--vextract (-X) &lt;debfile&gt; &lt;dir&gt; - Same as --extract, except
it is equivalent of giving tar the '-v' option - it prints the filenames as
it extracts them.
<item>--fsys-tarfile &lt;debfile&gt; - This option outputs a gunzip'd version
of data.tar.gz to stdout.
<item>--new - sets the archive format to be used to the new Debian format
<item>--old - sets the archive format to be used to the old Debian format
<item>--debug - Tells dpkg-deb to produce debugging output
<item>--nocheck - Tells dpkg-deb not to check the sanity of the control file
<item>--help (-h) - Gives a help message
<item>--version - Shows the version number
<item>--licence/--license (UK/US spellings) - Shows a brief outline of the GPL
</list>

<sect1>Internal checks used by dpkg-deb when building packages

<p>
Here is a list of the internal checks used by dpkg-deb when building packages.
It is in the order they are done.

<list>
<item>First, the output Debian archive argument, if it is given, is checked
using stat. If it is a directory, an internal flag is set. This check is only
made if the archive name is specified explicitly on the command-line. If the
argument was not given, the default is the directory name, with ".deb"
appended.
<item>Next, the control file is checked, unless the --nocheck flag was
specified on the command-line. dpkg-deb will bomb out if the second argument
to --build was a directory, and --nocheck was specified. Note that dpkg-deb
will not be able to determine the name of the package in this case. In the
control file, the following things are checked:-
	<list>
	<item>The package name is checked to see if it contains any invalid
	characters (see <ref id="control"> for this).
	<item>The priority field is checked to see if it uses standard values,
	and user-defined values are warned against. However, note that this
	check is now redundant, since the control file no longer contains
	the priority - the changes file now does this.
	<item>The control file fields are then checked against the standard
	list of fields which appear in control files, and any "user-defined"
	fields are reported as warnings.
	<item>dpkg-deb then checks that the control file contains a valid
	version number.
	</list>
<item>After this, in the case where a directory was specified to build the
.deb file in, the filename is created as "directory/pkg_ver.deb" or
"directory/pkg_ver_arch.deb", depending on whether the control file contains
an architecture field.
<item>Next, dpkg-deb checks for the &lt;dir&gt;/DEBIAN directory. It complains
if it doesn't exist, or if it has permissions &lt; 0755, or &gt; 0775.
<item>It then checks that all the files in this subdir are either symlinks
or plain files, and have permissions between 0555 and 0775.
<item>The conffiles file is then checked to see if the filenames are too
long. Warnings are produced for each that is. After this, it checks that
the package provides initial copies of each of these conffiles, and that
they are all plain files.
</list>

<chapt>dpkg internals

<p>
This chapter describes the internals of dpkg itself. Although the low-level
formats are quite simple, what dpkg does in certain cases often does not
make sense.

<sect id="updates">Updates

<p>
This describes the /var/lib/dpkg/updates directory. The function of this
directory is somewhat strange, and seems only to be used internally. A function
called cleanupdates is called whenever the database is scanned. This function
in turn uses <manref name="scandir" section="3">, to sort the files in this
directory. Files who names do not consist entirely of digits are discarded.
dpkg also causes a fatal error if any of the filenames are different lengths.

<p>
After having scanned the directory, dpkg in turn parses each file the same way
it parses the status file (they are sorted by the scandir to be in numerical
order). After having done this, it then writes the status information back
to the "status" file, and removes all the "updates" files.

<p>
These files are created internally by dpkg's "checkpoint" function, and are
cleaned up when dpkg exits cleanly.

<p>
Juding by the use of the updates directory I would call it a Journal. Inorder
to effeciently ensure the complete integrity of the status file dpkg will
"checkpoint" or journal all of it's activities in the updates directory. By
merging the contents of the updates directory (in order!!) against the 
original status file it can get the precise current state of the system,
even in the event of a system failure while dpkg is running.

<p> 
The other option would be to sync-rewrite the status file after each 
operation, which would kill performance.

<p>
It is very important that any program that uses the status file abort if
the updates directory is not empty! The user should be informed to run dpkg
manually (what options though??) to correct the situation.

<sect>What happens when dpkg reads the database

<p>
First, the status file is read. This gives dpkg an initial idea of the packages
that are there. Next, the updates files are read in, overriding the status
file, and if necessary, the status file is re-written, and updates files are
removed. Finally, the available file is read. The available file is read
with flags which preclude dpkg from updating any status information from it,
though - installed version, etc., and is also told to record that the packages
it reads this time are available, not installed.

<p>
More information on updates is given above.

<sect>How dpkg compares version numbers

<p>
Version numbers consist of three parts: the epoch, the upstream version, and
the Debian revision. Dpkg compares these parts in that order. If the epochs
are different, it returns immediately, and so on.

<p>
However, the important part is how it compares the versions which are
essentially stored as just strings. These are compared in two distinct parts:
those consisting of numerical characters (which are evaluated, and then
compared), and those consisting of other characters. When comparing
non-numerical parts, they are compared as the character values (ASCII), but
non-alphabetical characters are considered "greater than" alphabetical ones.
Also note that longer strings (after excluding differences where numerical
values are equal) are considered "greater than" shorter ones.

<p>
Here are a few examples of how these rules apply:-

<example>
15 > 10
0010 == 10

d.r > dsr
32.d.r == 0032.d.r
d.rnr < d.rnrn
</example>

</book>