Currently, pidlists are reference counted from file open and release
methods.  This means that holding onto an open file may waste memory
and reads may return data which is very stale.  Both aren't critical
because pidlists are keyed and shared per namespace and, well, the
user isn't supposed to have large delay between open and reads.

cgroup is planned to be converted to use kernfs and it'd be best if we
can stick to just the seq_file operations - start, next, stop and
show.  This can be achieved by loading pidlist on demand from start
and release with time delay from stop, so that consecutive reads don't
end up reloading the pidlist on each iteration.  This would remove the
need for hooking into open and release while also avoiding issues with
holding onto pidlist for too long.

The previous patches implemented delayed release and restructured
pidlist handling so that pidlists can be loaded and released from
seq_file start / stop.  This patch actually moves pidlist load to
start and release to stop.

This means that pidlist is pinned only between start and stop and may
go away between two consecutive read calls if the two calls are apart
by more than CGROUP_PIDLIST_DESTROY_DELAY.  cgroup_pidlist_start()
thus can't re-use the stored cgroup_pid_list_open_file->pidlist
directly.  During start, it's only used as a hint indicating whether
this is the first start after open or not and pidlist is always looked
up or created.

pidlist_mutex locking and reference counting are moved out of
pidlist_array_load() so that pidlist_array_load() can perform lookup
and creation atomically.  While this enlarges the area covered by
pidlist_mutex, given how the lock is used, it's highly unlikely to be
noticeable.

v2: Refreshed on top of the updated "cgroup: introduce struct
    cgroup_pidlist_open_file".
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Li Zefan <lizefan@huawei.com>