Re: [PATCH] tasks: Extract rcu_users out of union

From: David Vernet
Date: Thu Feb 16 2023 - 13:40:07 EST


On Thu, Feb 16, 2023 at 09:04:59AM +0100, Oleg Nesterov wrote:
> I won't argue with this patch, but I can't understand the changelog...
>
> On 02/15, David Vernet wrote:
> >
> > Similarly, in sched_ext, schedulers are using integer pids to remember
> > tasks, and then looking them up with find_task_by_pid_ns(). This is
> > slow, error prone, and adds complexity. It would be more convenient and
> > performant if BPF schedulers could instead store tasks directly in maps,
> > and then leverage RCU to ensure they can be safely accessed with low
> > overhead.
>
> To simplify, suppose we have
>
> int global_pid;
>
> void func(void)
> {
> rcu_read_lock();
> task = find_task_by_pid(global_pid);
> do_something(task);
> rcu_read_unlock();
> }
>
> Could you explain how exactly can this patch help to turn global_pid into
> "task_struct *" ? Why do you need to increment task->rcu_users ?

If you're not persisting the task in a map / data structure, then I
agree that find_task_by_pid_ns() is likely sufficient. What we want to
be able to do is something like this:

void func(void)
{
rcu_read_lock();
task = peek_next_task();
if (task)
do_something(task);
rcu_read_unlock();
}

In such an example, we could be peeking into a statically allocated
circular queue, and want to be able to ensure that a task we look at
from the top is protected with rcu. The general mechanics would be that
a task is inserted with a refcount_inc_not_zero(), and when it's
removed, we do a put_task_struct_rcu_user().

Does that make sense?

>
> > a task that's successfully looked
> > up in e.g. the pid_list with find_task_by_pid_ns(), can always have a
> > 'usage' reference acquired on them, as it's guaranteed to be >
> > 0 until after the next gp.
>
> Yes. So it seems you need another key-to-task_struct map with rcu-safe
> lookup/get and thus the add() method needs inc_not_zero(task->rcu_users) ?

Yes, exactly.

Thanks for taking a look at the patch.

- David