cpuset: don't nest cgroup_mutex inside get_online_cpus()

CPU / memory hotplug path currently grabs cgroup_mutex from hotplug
event notifications.  We want to separate cpuset locking from cgroup
core and make cgroup_mutex outer to hotplug synchronization so that,
among other things, mechanisms which depend on get_online_cpus() can
be used from cgroup callbacks.  In general, we want to keep
cgroup_mutex the outermost lock to minimize locking interactions among
different controllers.

Convert cpuset_handle_hotplug() to cpuset_hotplug_workfn() and
schedule it from the hotplug notifications.  As the function can
already handle multiple mixed events without any input, converting it
to a work function is mostly trivial; however, one complication is
that cpuset_update_active_cpus() needs to update sched domains
synchronously to reflect an offlined cpu to avoid confusing the
scheduler.  This is worked around by falling back to the the default
single sched domain synchronously before scheduling the actual hotplug
work.  This makes sched domain rebuilt twice per CPU hotplug event but
the operation isn't that heavy and a lot of the second operation would
be noop for systems w/ single sched domain, which is the common case.

This decouples cpuset hotplug handling from the notification callbacks
and there can be an arbitrary delay between the actual event and
updates to cpusets.  Scheduler and mm can handle it fine but moving
tasks out of an empty cpuset may race against writes to the cpuset
restoring execution resources which can lead to confusing behavior.
Flush hotplug work item from cpuset_write_resmask() to avoid such
confusions.

v2: Synchronous sched domain rebuilding using the fallback sched
    domain added.  This fixes various issues caused by confused
    scheduler putting tasks on a dead CPU, including the one reported
    by Li Zefan.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Li Zefan <lizefan@huawei.com>