Re: [PATCH v12 07/13] task_isolation: add debug boot flag

[Peter Zijlstra]

On Thu, May 19, 2016 at 10:42:39AM -0400, Chris Metcalf wrote:

> >>>>+	rcu_read_lock();
> >>>>+	p = cpu_curr(cpu);

Here @cpu can schedule, hit TASK_DEAD and do put_task_struct() and
kfree() the task.

> >>>>+	get_task_struct(p);

And here we then do a use-after-free.

> >>>>+	rcu_read_unlock();
> >>>>+	task_isolation_debug_task(cpu, p);
> >>>>+	put_task_struct(p);

> So, I think what you're saying is that there is a race between when we
> read per_cpu(runqueues, cpu).curr, and when we increment the
> p->usage value in the task, and that the RCU read lock doesn't help
> with that? 

Yep, as per the above.

> My impression was that by being the ".curr" task, we are
> guaranteed that it hasn't gone through do_exit() yet, and thus we
> benefit from an RCU guarantee around being able to validly dereference
> the pointer, i.e. it hasn't yet been freed and so dereferencing is safe.

Nope... the only way to avoid this from happening is taking @cpu's
rq->lock to prevent the remote CPU from scheduling.

> I don't see how grabbing the ->curr from the runqueue is any more
> fragile from an RCU perspective than grabbing the task from the pid in
> kill_pid_info().

The whole pid data structure is RCU managed, rq->curr is not.

> Anyway, whatever more clarity you can offer me, or suggestions for
> APIs to use are welcome.

The API proposed in the discussion below..

> >See also the discussion around:
> >
> >lkml.kernel.org/r/20160518170218.GY3192@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> 
> This makes me wonder if I should use rcu_dereference(&cpu_curr(p))
> just for clarity, though I think it's just as correct either way.

Nope, that's just as broken.

So the 'simple' thing is:

	struct rq *rq = cpu_rq(cpu);
	struct task_struct *task;

	raw_spin_lock_irq(&rq->lock);
	task = rq->curr;
	get_task_struct(task);
	raw_spin_unlock_irq(&rq->lock);

Because by holding rq->lock, the remote CPU cannot schedule and the
current task _must_ still be valid.

And note; the above can result in a task which already has PF_EXITING
set.

The complex thing is described in the linked thread and will likely make
your head hurt.