[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Declarative Diversions - Design Spec 2

Hey all,

The following is the second design spec for the project. I have addressed all of the issues that were brought up. If I missed anything or if you feel I haven't fully answered what you need to know let me know and I will fix it asap
A diversion is when it is possible to have dpkg not overwrite a file when it 
reinstalls the package it belongs to, and to have it put the file from the 
package somewhere else instead.
Declarative diversions involves a new control file with a declarative 
syntax which dpkg will parse and process directly as part of the package unpack 
and removal phases, eliminating the problems resulting from non-atomic 
handling of diversions.

There are a number of topics involved in implementing this kind of project
 * What syntax do we use for the new control file?
 * What dpkg does with the control file
 * How do we handle diversions to a non-existant directory?
 * How do we order unpacking a new package which adds a diversion?
 * How do we order unpacking a new package that removes a diversion?
 * How do we order removing a package which had a diversion?
 * How do we handle errors?
 * What happens to dpkg-divert?

Details - Control File Syntax
It will conform to RFC2822 style with the following format:
 * Divert-From:
 * Divert-To:
 * Blank lines and lines beginning with '#' will be comments
'Divert-To' will be optional and if it is ommitted then files being diverted 
will have their filename changed to 'file.distrib'
The above style has it's advantages, one of the main ones being that there is
no need to escape whitespace within filenames. It also allows for more commands 
to be added at a later date.

Details - Control File Handling
Within control.tar.gz the file should be named 'diversions'
This file is then copied to /var/lib/info/$package.diversions
We need to store the control file for the currently-installed package so that 
we can detect declarative diversions that have been removed from the control 
file between versions (On-disk version and the to-be-unpacked version).
If there is no previous version of a control file then all current diversions
within the package are left in place and any new ones declared within the 
control file are added.
If there is a previous version of the control file then the diversions 
listed within the file are compared with the diversions currently in place.
If a diversion exists that is no longer listed in the control file then it is 
removed. If there is a diversion listed in the control file that doesn't 
currently exist then it is added.

Details - Handling Diversions to non-existant directories
Diverting files to directories that don't exist can cause a number of problems. 
If the package does not 'own' the directory it may be left orphaned on
removal of the package
The package is responsible for ensuring the availability of the target 
directory in the unpack phase.
If a declarative diversion attempts to divert a file to a folder that neither 
exists already on the filesystem at unpack time, nor is shipped within 
the package, dpkg will return an error at unpack time.

Details - Ordering Requirements
 =>Unpacking a new package that adds a diversion
   *1. Check if package has a previous diversion control file
   *2. Copy control file to tmp directory
   *3. Calculate diversions to be added
   *4. Unpack folders
   *5. Add diversions
   *6. Unpack files
   *7. Copy control file to /var/lib/dpkg/info/$package.diversion
   *Performing diversions this way avoids trouble with creating folders and 
    accidentally overwriting the wrong version of files. 
   *Another major problem would be extracting a new version of the file and  
    overwriting the old one. 
   *This might cause the wrong version of the file to be diverted and the 
    package to break.
 =>Unpacking a new package that removes a diversion
   *1. Check if package has a previous diversion control file
   *2. Copy control file to tmp directory
   *3. Calculate diversions to be removed
   *4. Check if dropped diversion results in a file conflict and abort the 
       unpack if it does.
   *5. Unpack folders
   *6. Unpack files
   *7. Remove old files no longer needed in the package
   *8. Remove diversions
   *9. Copy control file to /var/lib/dpkg/info/$package.diversion
   *If a package has no previous diversion control file then no diversions 
    will be removed and existing diversions can only be removed using 
   *If removing a diversion results in a file conflict then dpkg will return 
    an error at unpack time before anything is unpacked and the unpack
    is aborted.
 =>Removing a package which had a diversion
   *1. Remove files
   *2. Remove diversions
   *3. Remove folders
   *4. Remove control file at /var/lib/dpkg/info/$package.diversion
   *This ensures that all files, diversions and folders are removed correctly
Details - Error handling
Errors in diversions will have to handled with a great deal of care due to
the fact that if they are not the package could be broken.
This means that a great deal of checks must be done to ensure that all the files
can be diverted properly before any actual diverting takes place. If they can't 
the package installation/update must be stopped and rolled back to avoid the 
package being installed incorrectly or broken.

Details - 'dpkg-divert'
When we impliment the new diversion method we should keep the current 
dpkg-divert. This allows maintainers to catch up with the new method without 
breaking their packages. It also allows maintainers to perform some operations 
that aren't support by the new method.
All current functionality of dpkg-divert will be left intact for use by 
sysadmins and for maintainers to catch up with the new system. The two 
systems will work together without directly conflicting with each other.

Example Usage #1
The file to be diverted is '/usr/share/foo'
It needs to be moved to '/usr/share/bar'
The syntax of the control file would be:
#Start File
Divert-From: /usr/share/foo
Divert-To: /usr/share/bar
#End File

Example Usage #2
In this example the maintainer doesn't want to move the file to any specific
The syntax of the control file would be:
#Start File
Divert-From: /usr/share/foo
#End File
This would divert the file to '/usr/share/foo.distrib'

RFC2822 Guide: 
 * http://www.faqs.org/rfcs/rfc2822.html
First Email Thread on Declarative Diversions (First Message in Thread):
 * http://lists.debian.org/debian-dpkg/2011/05/msg00102.html
Email thread on first design draft (First message in Thread):
 * http://lists.debian.org/debian-dpkg/2011/06/msg00019.html
Declarative Diversions Wiki:
 * http://wiki.debian.org/SummerOfCode2011/DeclarativeDiversions
My Blog for this Project: 
 * http://blog.sam-dunne.com

Sam Dunne

Reply to: