Re: [RFC][PATCH v2] cgroups: Run subsystem fork callback fromcgroup_post_fork()

[Frederic Weisbecker]

On Wed, Mar 07, 2012 at 05:22:38PM +0800, Li Zefan wrote:
> Frederic Weisbecker wrote:
> > On Thu, Mar 01, 2012 at 11:20:47AM +0800, Li Zefan wrote:
> >> ä 2012å03æ01æ 00:21, Frederic Weisbecker åé:
> >>> On Wed, Feb 29, 2012 at 07:55:00AM -0800, Mandeep Singh Baines wrote:
> >>>> Frederic Weisbecker (fweisbec@xxxxxxxxx) wrote:
> >>>>> When a user freezes a cgroup, the freezer sets the subsystem state
> >>>>> to CGROUP_FREEZING and then iterates over the tasks in the cgroup links.
> >>>>>
> >>>>> But there is a possible race here, although unlikely, if a task
> >>>>> forks and the parent is preempted between write_unlock(tasklist_lock)
> >>>>> and cgroup_post_fork(). If we freeze the cgroup while the parent
> >>>>
> >>>> So what if you moved cgroup_post_forks() a few lines up to be
> >>>> inside the tasklist_lock?
> >>>
> >>> It won't work. Consider this scenario:
> >>>
> >>> CPU 0                                     CPU 1
> >>>
> >>>                                        cgroup_fork_callbacks()
> >>>                                        write_lock(tasklist_lock)
> >>> try_to_freeze_cgroup() {               add child to task list etc...
> >>> 	cgroup_iter_start()
> >>>         freeze tasks                        
> >>>         cgroup_iter_end()
> >>> }                                      cgroup_post_fork()
> >>>                                        write_unlock(tasklist_lock)
> >>>
> >>> If this is not the first time we call cgroup_iter_start(), we won't go
> >>> through the whole tasklist, we simply iterate through the css set task links.
> >>>
> >>> Plus we try to avoid anything under tasklist_lock when possible.
> >>>
> >>
> >> Your patch won't close the race I'm afraid.
> >>
> >> // state will be set to FREEZING
> >> echo FROZEN > /cgroup/sub/freezer.state
> >>                                           write_lock(tasklist_lock)
> >>                                           add child to task list ...
> >> 					  write_unlock(tasklist_lock)
> >> // state will be updated to FROZEN
> >> cat /cgroup/sub/freezer.state
> >> 					  cgroup_post_fork()
> >> 					  ->freezer_fork()
> >>
> >> freezer_fork() will freeze the task only if the cgroup is in FREEZING
> >> state, and will BUG if the state is FROZEN.
> > 
> > Good point!
> > 
> >>
> >> We can fix freezer_fork(), but seems that requires we hold cgroup_mutex
> >> in that function(), which we don't like at all. Not to say your
> >> task_counter stuff..
> >>
> >> At this moment I don't see a solution without tasklist_lock involved,
> >> any better idea?
> > 
> > Ok, everything would be _much_ simpler if we were adding the css set task
> > link unconditionally.
> > 
> > Unfortunately this means acquiring a (global) lock unconditionally and
> > doing a list_add() on every fork. Although the lock could perhaps
> > be made per css set.
> > 
> > An idea could be to start the css set linking as soon as we create
> > the first subdirectory of a freezer cgroup. The root css set can't be
> > freezed and when we create the subdirectory we have no task there yet.
> > Due to the threadgroup_lock() that follows, none of the tasks that will be
> > attached there can be stuck in the middle of a fork so we are fine against
> > their css_set links: we know that when we attach a task to the cgroup of that
> > subdir, its css set link is set and we won't have any of the race we are describing.
> > 
> > How does that sound?
> > 
> 
> I'm confused..
> 
> The race you described is about freezing a cgroup while a task is forking,
> and it doesn't have something to do with attaching tasks manually and the
> enabling of css set links. So how can the race be fixed in the way you
> proposed?

You're right. I scratched my head so much that I got confused :-(
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/