Most rootwrap filters are overly permissive, and allow running commands as root with no additional filtering on the arguments given, and little sanitization is done of the input commands.
Additionally, maintaining command filters is difficult. We have found that unused filters are left in place after they should have been removed.
Finally, rootwrap cannot easily express precise semantics of the use cases of privileged commands.
Before attempting to complete a privileged task with best practice, the task should be analyzed to see if the task can be completed in an unprivileged fashion. For example, using service accounts with file permissions and ownership for access instead of running a command as root.
Additionally, files can be owned by the project group instead of the root group. These cases are preferred solutions where an architectural change is preferred, and using privileged access can introduce security issues.
Rootwrap currently has several deficiencies:
Instead, a better approach is multi-faceted:
These guidelines will allow better semantic filtering of arguments being passed for each task. For example, if we need to mount an image or device in a specific filesystem sub-tree the caller does not need to pass the full path of the mount point. Additionally, each external command or library will have its own interpretation of arguments, and we need to be able to sanitize those in task-specific and use-specific ways.
Auditing the filtering and sanitization code is necessary as well, and this needs to be easier to do so it can be done as needed. Sharing common tasks will allow easier auditing as well as contribute to code centralization and re-use.
Within the OpenStack project, understanding how to better limit the usage of general commands is needed. For example, DeleteLink should not be able to delete an arbitrary path. Discussions are converging around a privileged daemon, similar to the one recently proposed for neutron. This would give several advantages including better SELinux and AppArmor profiles as the necessary privileges would be clearly stated in the code. Another area this would help is in system calls, both direct and external. Unfortunately, this would seemingly contradict the last bullet outlined above to build a better approach. It is, however, necessary to secure the OpenStack project.
The alternative is to accept that projects (such as nova-compute) will need to run as root, and as such will need to be fully audited - including rootwrap filters. SELinux and AppArmor profiles will be the responsibility of the deployer to create and maintain.
All calls to rootwrap, or project-specific interfaces to rootwrap, should be migrated to interfaces in a project specific privileged task module. This will still call rootwrap but will allow for better sanitization, identification of what can be consolidated into shared code, and will allow easier migration to a future solution.
The current codebase should be audited to determine if there are any specific places that can be re-architected to avoid running tasks as root. A Bandit plugin that looked for the use of rootwrap would allow easy identification of the code that needs to be audited.
Better documentation around how to write safe filters for rootwrap would allow for developers to become better educated, but would also give reviewers a citable document to link against in Gerrit.
The neutron privilege separation daemon should be supported to build experience and understand best practices around such an implementation.
When a final solution is determined the implementation should be audited regularly to ensure it is being used correctly and the results are as expected.