Re: [PATCH 7/9] [RFC] Support multiply-bindable cgroup subsystems

[Li Zefan <lizf@cn.fujitsu.com>]

Paul Menage wrote:
> [RFC] Support multiply-bindable cgroup subsystems
> 
> This patch allows a cgroup subsystem to be marked as bindable on
> multiple cgroup hierarchies independently, when declared in
> cgroup_subsys.h via MULTI_SUBSYS() rather than SUBSYS().
> 
> The state for such subsystems cannot be accessed directly from a
> task->cgroups (since there's no unique mapping for a task) but instead
> must be accessed via a particular control group object.
> 
> Multiply-bound subsystems are useful in cases where there's no direct
> correspondence between the cgroup configuration and some property of
> the kernel outside of the cgroups subsystem.  So this would not be
> applicable to e.g. the CFS cgroup, since there has to a unique mapping
> from a task to its CFS run queue.
> 
> As an example, the "debug" subsystem is marked multiply-bindable,
> since it has no state outside the cgroups framework itself.
> 

Great, this makes the debug subsystem more useful. Sometimes
I want to see some debug info in different hierarchies, but
I can't just because it can only be bound to one hierarchy.

> Example usage:
> 
> mount -t cgroup -o name=foo,debug,cpu cgroup /mnt1
> mount -t cgroup -o name=bar,debug,memory cgroup /mnt2
> 
> Open Issues:
> 
> - in the current version of this patch, mounting a cgroups hierarchy
>   with no options does *not* get you any of the multi-bindable
>   subsystems; possibly for consistency it should give you all of the
>   multi-bindable subsystems as well as all of the single-bindable
>   subsystems.
> 

Yeah, the latter is preferrable.

> - how can we avoid the checkpatch.pl errors due to creative use of
>   macros to generate enum names?

checkpatch.pl can sometimes generate false-positive, let's happily
ignore those "errors". ;)

But it whould be better if those macros can be handled in a cleaner way.