Re: [PATCH 03/10] threadgroup: extend threadgroup_lock() to cover exit and exec

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sun, Nov 13, 2011 at 05:44:32PM +0100, Frederic Weisbecker wrote:
> On Tue, Nov 01, 2011 at 04:46:26PM -0700, Tejun Heo wrote:
> > threadgroup_lock() protected only protected against new addition to
> > the threadgroup, which was inherently somewhat incomplete and
> > problematic for its only user cgroup.  On-going migration could race
> > against exec and exit leading to interesting problems - the symmetry
> > between various attach methods, task exiting during method execution,
> > ->exit() racing against attach methods, migrating task switching basic
> > properties during exec and so on.
> > 
> > This patch extends threadgroup_lock() such that it protects against
> > all three threadgroup altering operations - fork, exit and exec.  For
> > exit, threadgroup_change_begin/end() calls are added to exit path.
> > For exec, threadgroup_[un]lock() are updated to also grab and release
> > cred_guard_mutex.
> > 
> > With this change, threadgroup_lock() guarantees that the target
> > threadgroup will remain stable - no new task will be added, no new
> > PF_EXITING will be set and exec won't happen.
> > 
> > The next patch will update cgroup so that it can take full advantage
> > of this change.
> 
> I don't want to nitpick really, but IMHO the races involved in exit and exec
> are too different, specific and complicated on their own to be solved in a
> single one patch. This should be split in two things.
> 
> The specific problems and their fix need to be described more in detail
> in the changelog because the issues are very tricky.
> 
> The exec case:
> 
> IIUC, the race in exec is about the group leader that can be changed
> to become the exec'ing thread, making while_each_thread() unsafe.
> We also have other things happening there like all the other threads
> in the group that get killed, but that should be handled by the threadgroup_change_begin()
> you put in the exit path.
> The old leader is also killed but release_task() -> __unhash_process() is called
> for it manually from the exec path. However this thread too should be covered by your
> synchronisation in exit().
> 
> So after your protection in the exit path, the only thing to protect against in exec
> is that group_leader that can change concurrently. But I may be missing something in the picture.
> Also note this is currently protected by the tasklist readlock. Cred guard mutex is
> certainly better, I just don't remember if you remove the tasklist lock in a
> further patch.

Ah recalling what Ben Blum said, we also need the leader to stay stable because it
is excpected to be passed in ->can_attach(), ->attach(), ->cancel_attach(), ...
Although that's going to change after your patches that pass a flex array.