On Wed, Sep 04, 2019 at 11:20:03AM -0700, Linus Torvalds wrote: > On Wed, Sep 4, 2019 at 9:33 AM Frederic Weisbecker <[email protected]> > wrote: > > > > I thought the point of these rcu_users was to be able to do: > > > > rcu_read_lock() > > p = rcu_dereference(task) > > if (!refcount_inc_not_zero(p->rcu_users)) { > > No. Because of the shared state, you can't do that from RCU context. > > But you *could* increase rcu_users from within the process context > itself (as long as you do it before the exit path, ie during normal > system call execution), or possibly while holding the tasklist_lock > and verifying that the task hasn't died yet. > > I'm not sure there is any sensible case for doing that, though. It > would have to have a similar pattern to the runqueue use, where you > add a new RCU lookup point for the task. I'm sure something like that > _could_ exist, I just can't think of any right now.
Yeah indeed, in fact I just hadn't read the patches entirely so I missed the point. I see now that the purpose of rcu_users is to postpone the RCU delayed put_task_struct() at least once we are done with both final schedule out and release_task(). Sorry for the noise :o)
