cgroup: use percpu refcnt for cgroup_subsys_states

cgroup: use percpu refcnt for cgroup_subsys_states

A css (cgroup_subsys_state) is how each cgroup is represented to a
controller.  As such, it can be used in hot paths across the various
subsystems different controllers are associated with.

One of the common operations is reference counting, which up until now
has been implemented using a global atomic counter and can have
significant adverse impact on scalability.  For example, css refcnt
can be gotten and put multiple times by blkcg for each IO request.
For highops configurations which try to do as much per-cpu as
possible, the global frequent refcnting can be very expensive.

In general, given the various and hugely diverse paths css's end up
being used from, we need to make it cheap and highly scalable.  In its
usage, css refcnting isn't very different from module refcnting.

This patch converts css refcnting to use the recently added
percpu_ref.  css_get/tryget/put() directly maps to the matching
percpu_ref operations and the deactivation logic is no longer
necessary as percpu_ref already has refcnt killing.

The only complication is that as the refcnt is per-cpu,
percpu_ref_kill() in itself doesn't ensure that further tryget
operations will fail, which we need to guarantee before invoking
->css_offline()'s.  This is resolved collecting kill confirmation
using percpu_ref_kill_and_confirm() and initiating the offline phase
of destruction after all css refcnt's are confirmed to be seen as
killed on all CPUs.  The previous patches already splitted destruction
into two phases, so percpu_ref_kill_and_confirm() can be hooked up
easily.

This patch removes css_refcnt() which is used for rcu dereference
sanity check in css_id().  While we can add a percpu refcnt API to ask
the same question, css_id() itself is scheduled to be removed fairly
soon, so let's not bother with it.  Just drop the sanity check and use
rcu_dereference_raw() instead.

v2: - init_cgroup_css() was calling percpu_ref_init() without checking
      the return value.  This causes two problems - the obvious lack
      of error handling and percpu_ref_init() being called from
      cgroup_init_subsys() before the allocators are up, which
      triggers warnings but doesn't cause actual problems as the
      refcnt isn't used for roots anyway.  Fix both by moving
      percpu_ref_init() to cgroup_create().

    - The base references were put too early by
      percpu_ref_kill_and_confirm() and cgroup_offline_fn() put the
      refs one extra time.  This wasn't noticeable because css's go
      through another RCU grace period before being freed.  Update
      cgroup_destroy_locked() to grab an extra reference before
      killing the refcnts.  This problem was noticed by Kent.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Kent Overstreet <koverstreet@google.com>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Vivek Goyal <vgoyal@redhat.com>
Cc: "Alasdair G. Kergon" <agk@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Mikulas Patocka <mpatocka@redhat.com>
Cc: Glauber Costa <glommer@gmail.com>