Re: [PATCH 2/2] cgroup: Walk task list under tasklist_lock in cgroup_enable_task_cg_list

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Feb 22, 2012 at 01:55:28AM +0100, Frederic Weisbecker wrote:
> On Tue, Feb 21, 2012 at 02:23:43PM -0800, Mandeep Singh Baines wrote:
> > Frederic Weisbecker (fweisbec@gmail.com) wrote:
> > > Walking through the tasklist in cgroup_enable_task_cg_list() inside
> > > an RCU read side critical section is not enough because:
> > > 
> > > - RCU is not (yet) safe against while_each_thread()
> > > 
> > > - If we use only RCU, a forking task that has passed cgroup_post_fork()
> > >   without seeing use_task_css_set_links == 1 is not guaranteed to have
> > >   its child immediately visible in the tasklist if we walk through it
> > >   remotely with RCU. In this case it will be missing in its css_set's
> > >   task list.
> > > 
> > > Thus we need to traverse the list (unfortunately) under the
> > > tasklist_lock. It makes us safe against while_each_thread() and also
> > > make sure we see all forked task that have been added to the tasklist.
> > > 
> > > As a secondary effect, reading and writing use_task_css_set_links are
> > > now well ordered against tasklist traversing and modification. The new
> > > layout is:
> > > 
> > > CPU 0                                      CPU 1
> > > 
> > > use_task_css_set_links = 1                write_lock(tasklist_lock)
> > > read_lock(tasklist_lock)                  add task to tasklist
> > > do_each_thread() {                        write_unlock(tasklist_lock)
> > > 	add thread to css set links       if (use_task_css_set_links)
> > > } while_each_thread()                         add thread to css set links
> > > read_unlock(tasklist_lock)
> > > 
> > > If CPU 0 traverse the list after the task has been added to the tasklist
> > > then it is correctly added to the css set links. OTOH if CPU 0 traverse
> > > the tasklist before the new task had the opportunity to be added to the
> > > tasklist because it was too early in the fork process, then CPU 1
> > > catches up and add the task to the css set links after it added the task
> > > to the tasklist. The right value of use_task_css_set_links is guaranteed
> > > to be visible from CPU 1 due to the LOCK/UNLOCK implicit barrier properties:
> > > the read_unlock on CPU 0 makes the write on use_task_css_set_links happening
> > > and the write_lock on CPU 1 make the read of use_task_css_set_links that comes
> > > afterward to return the correct value.
> > > 
> > > Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
> > > Cc: Tejun Heo <tj@kernel.org>
> > > Cc: Li Zefan <lizf@cn.fujitsu.com>
> > > Cc: Mandeep Singh Baines <msb@chromium.org>
> > 
> > Reviewed-by: Mandeep Singh Baines <msb@chromium.org>
> > 
> > Sorry for being late. My feedback is really just comments.
> > 
> > > Cc: Oleg Nesterov <oleg@redhat.com>
> > > Cc: Andrew Morton <akpm@linux-foundation.org>
> > > Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > > ---
> > >  kernel/cgroup.c |   20 ++++++++++++++++++++
> > >  1 files changed, 20 insertions(+), 0 deletions(-)
> > > 
> > > diff --git a/kernel/cgroup.c b/kernel/cgroup.c
> > > index 6e4eb43..c6877fe 100644
> > > --- a/kernel/cgroup.c
> > > +++ b/kernel/cgroup.c
> > > @@ -2707,6 +2707,14 @@ static void cgroup_enable_task_cg_lists(void)
> > >  	struct task_struct *p, *g;
> > >  	write_lock(&css_set_lock);
> > 
> > You might want to re-test use_task_css_set_links once you have the lock
> > in order to avoid an unnecessary do_each_thread()/while_each_thread() in
> > case you race between reading the value and entering the loop. This is
> > a potential optimization in a rare case so maybe not worth the LOC.
> 
> Makes sense. I'll do that in a seperate patch.
> 
> > 
> > >  	use_task_css_set_links = 1;
> > > +	/*
> > > +	 * We need tasklist_lock because RCU is not safe against
> > > +	 * while_each_thread(). Besides, a forking task that has passed
> > > +	 * cgroup_post_fork() without seeing use_task_css_set_links = 1
> > > +	 * is not guaranteed to have its child immediately visible in the
> > > +	 * tasklist if we walk through it with RCU.
> > > +	 */
> > 
> > Maybe add TODO to remove the lock once do_each_thread()/while_each_thread()
> > is made rcu safe. On a large system, it could take a while to iterate
> > over every thread in the system. Thats a long time to hold a spinlock.
> > But it only happens once so probably not that big a deal.
> 
> I think that even if while_each_thread() was RCU safe, that wouldn't
> work here.
> 
> Unless I'm mistaken, we have no guarantee that a remote list_add_rcu()
> is immediately visible by the local CPU if it walks the list under
> rcu_read_lock() only.

Indeed, the guarantee is instead that -if- a reader encounters a newly
added list element, then that reader will see any initialization of that
list element carried out prior to the list_add_rcu().

Memory barriers are about ordering, not about making memory writes
visible faster.

							Thanx, Paul

> Consider that ordering scenario:
> 
> CPU 0                                      CPU 1
> ---------------                        --------------
> 
> fork() {
> 	write_lock(tasklist_lock);
>         add child to tasklist
> 	write_unlock(tasklist_lock);
>         cgroup_post_fork()
> }
> 
>                                          cgroup_enable_task_cg_lists() {
>                                                 rcu_read_lock();
>                                                 do_each_thread() {
>                                                         ..... <-- find child ?
>                                                 } while_each_thread()
>                                                 rcu_read_unlock()
> 
> 
> We have no guarantee here that the write on CPU 0 will be visible
> in time to CPU 1.
> 
> But may be I misunderstood the ordering and committing guarantees with RCU.
> Perhaps Paul can confirm or correct me.
> 
> Paul?
>