On 6/23/22 08:49, Michal Koutný wrote:
When we migrate a task between two cgroups, one of the checks is a verification that we can modify task's scheduler settings (cap_task_setscheduler()). An implicit migration occurs also when enabling a controller on the unified hierarchy (think of parent to child migration). The aforementioned check may be problematic if the caller of the migration (enabling a controller) has no permissions over migrated tasks. For instance, user's cgroup that ends up running a process of a different user. Although cgroup permissions are configured favorably, the enablement fails due to the foreign process [1]. Change the behavior by relaxing the permissions check on the unified hierarchy (or in v2 mode). This is in accordance with unified hierarchy attachment behavior when permissions of the source to target cgroups are decisive whereas the migrated task is opaque (for contrast, see more restrictive check in __cgroup1_procs_write()). [1] https://github.com/systemd/systemd/issues/18293#issuecomment-831205649 Reasons for RFC: 1) The unified hierarchy attachment behavior -- is that the right/consented model that migrated objects don't matter? 2) If 1) is true, have I missed any danger in allowing cpuset'ing a possibly privileged processes?
That could be an issue.
2.2) cpuset may be in v2 mode even on v1 hierarchy with different migration control rules (but checking migratee's creds in v1 eliminates effect of this patch). 3) Alternative approach would be to allow cpuset migrations only when nothing effectively changes (which is the case for parent->child migration upon controller enablement).
What do you mean by nothing effectively changes?
4) This is just idea draft, not tested in the real case.
Since the check is done on a taskset level, if only one of the tasks in the taskset fails, the whole taskset fails. Maybe we should consider an option for task based migration. So all the tasks that can be migrated will be migrated and the rests will be left behind in the original cpuset. Just a thought.
Cheers, Longman