Re: [PATCHSET cgroup/for-3.12] cgroup: use cgroup_subsys_state as the primary subsystem interface handle

[Li Zefan <lizefan@huawei.com>]

On 2013/8/2 5:49, Tejun Heo wrote:
> Hello,
> 
> Currently, struct cgroup * is used as the main interface handle
> between cgroup core and its subsystems, which works but is a bit
> clunky because subsystems usually care much more about css's
> (cgroup_subsys_state) a lot more than cgroups, which is natural as a
> css is the intersection between a cgroup and a subsystem.
> 
> In addition to being a bit clunky, dealing with cgroups directly pose
> a bit of trouble for the planned unified hierarchy support on two
> fronts.  First, most iterations become subsystem dependent as task
> membership is affected by which subtree has the specific subsystem
> enabled and thus require specifying which subsystem the iteration is
> for, which is automatically achieved if the interfaces deal with css's
> instead of cgroups.
> 
> Second, as css's may be created, attached, detached and destroyed
> dynamically multiple times across the lifetime of a given cgroup as
> they're enabled and disabled, which makes cgroup -> css mapping much
> more difficult to synchronize.  Giving out cgroup to subsystems and
> then requiring them to take the extra steps to deal with their css's
> coming and going dynamically is a lot more fragile than cgroup core
> proper handling it internally and giving out the resulting css's to
> subsystems.
> 
> So, this patchset converts all cgroup subsystem APIs to deal with
> css's instead of cgroups.  The patchset is fairly large but most of
> the conversions, while being tedious, aren't complex.  At the end of
> series, subsystems no longer make cgroup -> css mapping themselves and
> cgroup_css() - formerly cgroup_subsys_state() - is made internal to
> cgroup core proper.
> 
> This is a rather large update to the interface and likely to play as a
> barrier when porting commits, which is inconvenient but also provides
> an opportunity to clean up the API where we can as doing so won't
> significantly raise the level of inconvenience.  As such, this
> patchset contains some API cleanups and I'll probably follow up with
> further API updates that I've been meaning to do and, if you have some
> good idea to clean up cgroup internal API, this probably is a good
> time to submit it.
> 
> This patchset contains the following 23 patches.
> 
>  0001-cgroup-s-cgroup_subsys_state-cgroup_css-s-task_subsy.patch
>  0002-cpuset-drop-const-qualifiers-from-struct-cpuset-inst.patch
>  0003-netprio_cgroup-pass-around-css-instead-of-cgroup-and.patch
>  0004-hugetlb_cgroup-pass-around-hugetlb_cgroup-instead-of.patch
>  0005-cgroup-add-subsystem-pointer-to-cgroup_subsys_state.patch
>  0006-cgroup-add-update-accessors-which-obtain-subsys-spec.patch
>  0007-cgroup-add-css_parent.patch
>  0008-cgroup-pass-around-cgroup_subsys_state-instead-of-cg.patch
>  0009-cgroup-add-subsys-backlink-pointer-to-cftype.patch
>  0010-cgroup-pin-cgroup_subsys_state-when-opening-a-cgroup.patch
>  0011-cgroup-add-cgroup-dummy_css.patch
>  0012-cgroup-pass-around-cgroup_subsys_state-instead-of-cg.patch
>  0013-cgroup-convert-cgroup_next_sibling-to-cgroup_next_ch.patch
>  0014-cgroup-always-use-cgroup_next_child-to-walk-the-chil.patch
>  0015-cgroup-make-hierarchy-iterators-deal-with-cgroup_sub.patch
>  0016-cgroup-relocate-cgroup_advance_iter.patch
>  0017-cgroup-rename-cgroup_iter-to-cgroup_task_iter.patch
>  0018-cgroup-make-cgroup_task_iter-remember-the-cgroup-bei.patch
>  0019-cgroup-remove-struct-cgroup_scanner.patch
>  0020-cgroup-make-task-iterators-deal-with-cgroup_subsys_s.patch
>  0021-cgroup-make-cftype-un-register_event-deal-with-cgrou.patch
>  0022-cgroup-make-cgroup_taskset-deal-with-cgroup_subsys_s.patch
>  0023-cgroup-unexport-cgroup_css.patch
> 

Looks good to me!

Acked-by: Li Zefan <lizefan@huawei.com>