Re: [PATCH 6/6] Makes procs file writable to move all threads by tgid at once

["Serge E. Hallyn" <serue@us.ibm.com>]

Quoting Benjamin Blum (bblum@google.com):
> On Mon, Aug 3, 2009 at 1:54 PM, Serge E. Hallyn<serue@us.ibm.com> wrote:
> > Quoting Ben Blum (bblum@google.com):
> > ...
> >> +static int cgroup_task_migrate(struct cgroup *cgrp, struct cgroup *oldcgrp,
> >> +                            struct task_struct *tsk, int guarantee)
> >> +{
> >> +     struct css_set *oldcg;
> >> +     struct css_set *newcg;
> >> +
> >> +     /*
> >> +      * get old css_set. we need to take task_lock and refcount it, because
> >> +      * an exiting task can change its css_set to init_css_set and drop its
> >> +      * old one without taking cgroup_mutex.
> >> +      */
> >> +     task_lock(tsk);
> >> +     oldcg = tsk->cgroups;
> >> +     get_css_set(oldcg);
> >> +     task_unlock(tsk);
> >> +     /*
> >> +      * locate or allocate a new css_set for this task. 'guarantee' tells
> >> +      * us whether or not we are sure that a new css_set already exists;
> >> +      * in that case, we are not allowed to fail, as we won't need malloc.
> >> +      */
> >> +     if (guarantee) {
> >> +             /*
> >> +              * our caller promises us that the css_set we want already
> >> +              * exists, so we use find_existing_css_set directly.
> >> +              */
> >> +             struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT];
> >> +             read_lock(&css_set_lock);
> >> +             newcg = find_existing_css_set(oldcg, cgrp, template);
> >> +             BUG_ON(!newcg);
> >> +             get_css_set(newcg);
> >> +             read_unlock(&css_set_lock);
> >> +     } else {
> >> +             might_sleep();
> >
> > So cgroup_task_migrate() might sleep, but
> >
> > ...
> >
> >
> >> +     down_write(&leader->cgroup_fork_mutex);
> >> +     rcu_read_lock();
> >> +     list_for_each_entry_rcu(tsk, &leader->thread_group, thread_group) {
> >> +             /* leave current thread as it is if it's already there */
> >> +             oldcgrp = task_cgroup(tsk, subsys_id);
> >> +             if (cgrp == oldcgrp)
> >> +                     continue;
> >> +             /* we don't care whether these threads are exiting */
> >> +             retval = cgroup_task_migrate(cgrp, oldcgrp, tsk, 1);
> >
> > Here it is called under rcu_read_lock().
> 
> You'll notice the fourth argument, which tells cgroup_task_migrate

Hmmm, in my defense one would notice it more readily if the caller used a
meaningful #define instead of '1'.

> whether the css_set is guaranteed or not. If we say we've already got
> it covered, the might_sleep section doesn't happen.
> 
> >> -void cgroup_fork(struct task_struct *child)
> >> +void cgroup_fork(struct task_struct *child, int clone_flags)
> >>  {
> >> +     if (clone_flags & CLONE_THREAD)
> >> +             down_read(&current->group_leader->cgroup_fork_mutex);
> >> +     else
> >> +             init_rwsem(&child->cgroup_fork_mutex);
> >
> > I'm also worried about the overhead here on what should be a
> > fast case, CLONE_THREAD.  Have you done any benchmarking of
> > one thread spawning a bunch of others?
> 
> Should be strictly better as this is making the rwsem local to the
> threadgroup - at least in comparison to the previous edition of this
> patch which had it as a global lock.
> 
> > What *exactly* is it we are protecting with cgroup_fork_mutex?
> > 'fork' (as the name implies) is not a good answer, since we should be
> > protecting data, not code.  If it is solely tsk->cgroups, then perhaps
> > we should in fact try switching to (s?)rcu.  Then cgroup_fork() could
> > just do rcu_read_lock, while cgroup_task_migrate() would make the change
> > under a spinlock (to protect against concurrent cgroup_task_migrate()s),
> > and using rcu_assign_pointer to let cgroup_fork() see consistent data
> > either before or after the update...  That might mean that any checks done
> > before completing the migrate which involve the # of tasks might become
> > invalidated before the migration completes?  Seems acceptable (since
> > it'll be a small overcharge at most and can be quickly remedied).
> 
> You'll notice where the rwsem is released - not until cgroup_post_fork
> or cgroup_fork_failed. It doesn't just protect the tsk->cgroups
> pointer, but rather guarantees atomicity between adjusting
> tsk->cgroups and attaching it to the cgroups lists with respect to the
> critical section in attach_proc. If you've a better name for the lock
> for such a race condition, do suggest.

No the name is pretty accurate - it's the lock itself I'm objecting
to.  Maybe it's the best we can do, though.

thanks,
-serge