Enable kdump mechanism from kdump-tool instead of kexec-tools
The kdump mechanism from the kdump-tool package is more flexible than its kexec-tool equivalent. It allows for multiple dumps to be collected, is the mechanism used in upstream Debian and has a configuration file that let the sysadmin controls some parameters.
It is currently functional on Ubuntu but the documented way of gathering a kernel dump is to use the 0_kdump initscript delivered by kexec-tools which package the kernel dump file as an Apport bundle. While this solution is sufficient on Desktops, it becomes restrictive on server/cloud installs.
Rationale: Current kexec-tool kernel dump capture mechanism is too limitative in a server/cloud environment
Goal: Decide on the feasibility of using kdump-tool as default for server/cloud installs
Blueprint information
- Status:
- Not started
- Approver:
- Dave Walker
- Priority:
- Medium
- Drafter:
- Ubuntu Server
- Direction:
- Approved
- Assignee:
- Louis Bouchard
- Definition:
- Approved
- Series goal:
- Accepted for raring
- Implementation:
- Unknown
- Milestone target:
- ubuntu-13.04-feature-freeze
- Started by
- Completed by
Whiteboard
User Stories:
Risks:
Test Plans:
Release Note:
----------
The kdump mechanism from the kdump-tool package is more flexible than its kexec-tool equivalent. It allows for multiple dumps to be collected, is the mechanism used in upstream Debian and has a configuration file that let the sysadmin controls some parameters.
It is currently functional on Ubuntu but the documented way of gathering a kernel dump is to use the 0_kdump initscript delivered by kexec-tools which package the kernel dump file as an Apport bundle. While this solution is sufficient on Desktops, it becomes restrictive on server/cloud installs.
Rationale: Current kexec-tool kernel dump capture mechanism is too limitative in a server/cloud environment. Because you only capture last crash; people want to change mkdumpfile options, have to hack scripts vs. conf files.
We might need to consider different use cases for desktop/server if using kdump-tool. But this might not be bad since we can have a similar setup as kexec-tool.
linux-crashdump
- needs to be changed to install kdump-tool
- does it need to install 'crash'? we might not do analysis on same machine
Goal: Decide on the feasibility of using kdump-tool as default for server/cloud installs
Issues :
what is the current story around kdump with UEFI ?
- currently disabled if using secure boot
- but will be re-enabled when we can use signed kexec kernels
Is the "server only" requirement valid or should we use kdump accross the board ?
Is usage of apport hooks for kernel dumps useful and/or valid ?
Need to be considerate of
- whoopise integration (how is this turned on for servers?)
- private data reporting (private daisy instance?)
Can server/cloud users upload large amounts of crash data?; need to consider sensitive environments, etc.
Discussion points:
Right now start with x86, long term ensure it works on ARM.
consider how we can reproduce apport behavor with kdump-tools during discussion.
need to determine use cases for desktop/server for kdump-tool configurations
Notes:
MIR is not required for kdump-tools as its source package (makedumpfile) is already in main
A simple kexec fails on ARM (Calxeda/Highbank) as of linux 3.8.0-17.27
Work Items
Work items:
[louis] - kdump-tools needs to be MIR / file a bug: DONE
[smb] - discuss linuxcrashdump changes on ubuntu-devel / ubuntu-kernel: DONE
[smb] - linuxcrashdump changed to install kdump-tool (after discussion): DONE
[arges] - write better documentation for kdump-tool: DONE
[dannf] - validate/invalidate with kdump-tools / mkdumpfile on ARM: DONE
[louis] - Integrate netdump support: POSTPONED
[med] - netdump charm to aggregate dumps in a cloud: POSTPONED
Dependency tree
* Blueprints in grey have been implemented.