Re: [PATCHSET] cgroup: simplify cgroup removal path

[Glauber Costa <glommer@parallels.com>]

On 10/31/2012 08:22 AM, Tejun Heo wrote:
> Hello, guys.
> 
> cgroup removal path is quite ugly.  A lot of the ugliness comes from
> the weird design which allows ->pre_destroy() to fail and the feature
> to drain existing CSS reference counts before committing to removal.
> Both mean that it should be possible to roll-back cgroup destruction
> after some or all ->pre_destroy() invocations.
> 
> This weird design has never really worked.  To list a couple examples.
> 
>  * Some ->pre_destroy() implementations aren't side-effect free.
>    Roll-back happens after a lot of state is already lost.
> 
>  * Some ->pre_destroy() implementations (naturally) assume that the
>    cgroup being destroyed would stay quiescent between successful
>    ->pre_destroy() and its destruction.  Unfortunately, any operation
>    can happen inbetween and the cgroup could be in a very different
>    state by the time it actually gets destroyed.
> 
> It's just such an unusual design which unnecessarily contains weird
> code path combinations which are tricky to hit, reproduce and expect.
> Moreover, the design's deficiencies attracts kludges on top as
> workarounds and we end up with stuff like cgroup_exclude_rmdir() and
> cgroup_release_and_wakeup_rmdir() which really make me want to cry.
> 
> Now that memcg has moved away from failable ->pre_destroy(), we can do
> away with all these.  I tested some basic operations and some corner
> cases but am still a bit scared.  Would love to get acks from Li and
> memcg people.
> 
> This patchset contains the following eight patches.
> 
>  0001-cgroup-kill-cgroup_subsys-__DEPRECATED_clear_css_ref.patch
>  0002-cgroup-kill-CSS_REMOVED.patch
>  0003-cgroup-use-cgroup_lock_live_group-parent-in-cgroup_c.patch
>  0004-cgroup-deactivate-CSS-s-and-mark-cgroup-dead-before-.patch
>  0005-cgroup-remove-CGRP_WAIT_ON_RMDIR-cgroup_exclude_rmdi.patch
>  0006-memcg-make-mem_cgroup_reparent_charges-non-failing.patch
>  0007-hugetlb-do-not-fail-in-hugetlb_cgroup_pre_destroy.patch
>  0008-cgroup-make-pre_destroy-return-void.patch
> 
> 0001-0002 remove now unused ->pre_destroy() failure handling and do
> follow-up simplification.
> 
> 0003-0004 update removal path such that each ->pre_destroy() is
> guaranteed to be invoked once per removal and the cgroup being
> destroyed stays quiescent until destruction is complete.
> 
> 0005 removes the scary CGRP_WAIT_ON_RMDIR mechanism.
> 
> 0006-0008 are follow-up clean-ups.  0006 and 0007 are from Michal's
> patchset[1].
> 
> This patchset is on top of
> 
>   v3.6 (a0d271cbfe)
> + [1] the first three patches of
>       "memcg/cgroup: do not fail fail on pre_destroy callbacks" patchset
> 
> and available in the following git branch.
> 
>  git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git review-cgroup-rmdir-updates
> 
> Thanks.
> 
>  block/blk-cgroup.c     |    3 
>  include/linux/cgroup.h |   41 -------
>  kernel/cgroup.c        |  256 +++++++++++--------------------------------------
>  mm/hugetlb_cgroup.c    |   11 --
>  mm/memcontrol.c        |   51 +--------
>  5 files changed, 75 insertions(+), 287 deletions(-)


The patches are quite straightforward, and you are basically throwing
useless code away...

The only think that drew my attention is that you are changing the
local_irq_save callsite to local_irq_disable. It shouldn't be a problem,
since this is never expected to be called in interrupt context.

Still... it makes me wonder if that disabled-interrupt block is still
needed? According to the changelogs, it was introduced in e7c5ec919 for
the css_tryget mechanism. But css_tryget itself will never scan
subsystems, so if we can no longer fail, we should be able to just ditch
it. Unless I am missing something