[Devel] [RFC rh7] ve: cgroups -- Allow to attach non-self into ve cgroups

[Vladimir Davydov]

On Tue, Jun 16, 2015 at 06:19:44PM +0300, Kirill Tkhai wrote:
> This patch brings a couple of problems. The first one is if we're setting
> a ve cgroup for a forking task, it's possible the parent and the child
> fall into different cgroups. For example, parent is inside CT, while
> child is in ve0.
> 
> The second problem is that it is racy with all places, where we're doing
> get_exec_env() and hoping on a stable result. There are a lot of places,
> we use it, so every place need too be checked.
> 
> The first problem can be solved using cgroup_postfork() method, while
> the second is worse. The number of get_exec_env() is big, and it tends
> to increase.  We could check all of them, but an universal solutions
> will be better. But it seems there is no easy universal solution.
> 
> We can implement a task lock and use get_exec_env() under it. 
> We can use something like rcu_read_{lock,unlock} braces around get_exec_env(),
> and use synchronize_rcu() in cgroup_postfork().

Alternatively, we could set task_ve in a task_work. See
kernel/bc/beancounter.c for an example.