[PATCH] cgroup: Convert synchronize_rcu to call_rcu in cgroup_attach_task

[Li Zefan]

Paul Menage wrote:
> On Sun, Nov 21, 2010 at 8:06 PM, Colin Cross <ccross at android.com> wrote:
>> The synchronize_rcu call in cgroup_attach_task can be very
>> expensive.  All fastpath accesses to task->cgroups that expect
>> task->cgroups not to change already use task_lock() or
>> cgroup_lock() to protect against updates, and, in cgroup.c,
>> only the CGROUP_DEBUG files have RCU read-side critical
>> sections.
> 
> I definitely agree with the goal of using lighter-weight
> synchronization than the current synchronize_rcu() call. However,
> there are definitely some subtleties to worry about in this code.
> 
> One of the reasons originally for the current synchronization was to
> avoid the case of calling subsystem destroy() callbacks while there
> could still be threads with RCU references to the subsystem state. The
> fact that synchronize_rcu() was called within a cgroup_mutex critical
> section meant that an rmdir (or any other significant cgrooup
> management action) couldn't possibly start until any RCU read sections
> were done.
> 
> I suspect that when we moved a lot of the cgroup teardown code from
> cgroup_rmdir() to cgroup_diput() (which also has a synchronize_rcu()
> call in it) this restriction could have been eased, but I think I left
> it as it was mostly out of paranoia that I was missing/forgetting some
> crucial reason for keeping it in place.
> 
> I'd suggest trying the following approach, which I suspect is similar
> to what you were suggesting in your last email
> 
> 1) make find_existing_css_set ignore css_set objects with a zero refcount
> 2) change __put_css_set to be simply
> 
> if (atomic_dec_and_test(&cg->refcount)) {
>   call_rcu(&cg->rcu_head, free_css_set_rcu);
> }

If we do this, it's not anymore safe to use get_css_set(), which just
increments the refcount without checking if it's zero.

> 
> 3) move the rest of __put_css_set into a delayed work struct that's
> scheduled by free_css_set_rcu
> 
> 4) Get rid of the taskexit parameter - I think we can do that via a
> simple flag that indicates whether any task has ever been moved into
> the cgroup.
> 
> 5) Put extra checks in cgroup_rmdir() such that if it tries to remove
> a cgroup that has a non-zero refcount, it scans the cgroup's css_sets
> list - if it finds only zero-refcount entries, then wait (via
> synchronize_rcu() or some other appropriate means, maybe reusing the
> CGRP_WAIT_ON_RMDIR mechanism?) until the css_set objects have been
> fully cleaned up and the cgroup's refcounts have been released.
> Otherwise the operation of moving the last thread out of a cgroup and
> immediately deleting the cgroup would very likely fail with an EBUSY
>