(2013/01/04 6:35), Tejun Heo wrote:

Hello, guys.

This is the second attempt at decoupling cpuset locking from cgroup
core.  Changes from the last take[L] are

* cpuset-drop-async_rebuild_sched_domains.patch moved from 0007 to
   0009.  This reordering makes cpu hotplug handling async first and
   removes the temporary cyclic locking dependency.

* 0006-cpuset-cleanup-cpuset-_can-_attach.patch no longer converts
   cpumask_var_t to cpumask_t as per Rusty Russell.

* 0008-cpuset-don-t-nest-cgroup_mutex-inside-get_online_cpu.patch now
   synchronously rebuilds sched domains from cpu hotplug callback.
   This fixes various issues caused by confused scheduler puttings
   tasks into a dead cpu including the RCU stall problem reported by Li
   Zefan.

Original patchset description follows.

Depending on cgroup core locking - cgroup_mutex - is messy and makes
cgroup prone to locking dependency problems.  The current code already
has lock dependency loop - memcg nests get_online_cpus() inside
cgroup_mutex.  cpuset the other way around.

Regardless of the locking details, whatever is protecting cgroup has
inherently to be something outer to most other locking constructs.
cgroup calls into a lot of major subsystems which in turn have to
perform subsystem-specific locking.  Trying to nest cgroup
synchronization inside other locks isn't something which can work
well.

cgroup now has enough API to allow subsystems to implement their own
locking and cgroup_mutex is scheduled to be made private to cgroup
core.  This patchset makes cpuset implement its own locking instead of
relying on cgroup_mutex.

cpuset is rather nasty in this respect.  Some of it seems to have come
from the implementation history - cgroup core grew out of cpuset - but
big part stems from cpuset's need to migrate tasks to an ancestor
cgroup when an hotunplug event makes a cpuset empty (w/o any cpu or
memory).

This patchset decouples cpuset locking from cgroup_mutex.  After the
patchset, cpuset uses cpuset-specific cpuset_mutex instead of
cgroup_mutex.  This also removes the lockdep warning triggered during
cpu offlining (see 0009).

Note that this leaves memcg as the only external user of cgroup_mutex.
Michal, Kame, can you guys please convert memcg to use its own locking
too?

Okay...but If Costa has a new version of his patch, I'd like to see it.
I'm sorry if I missed his new patches for removing cgroup_lock.

Thanks,
-Kame

Thanks,
-Kame