Re: [PATCH] procfs: fix missing RCU protection when reading real_parent in do_task_stat()

[Oleg Nesterov <oleg@redhat.com>]

On 01/27, Mateusz Guzik wrote:
>
> On Tue, Jan 27, 2026 at 06:25:25PM +0100, Oleg Nesterov wrote:
> > On 01/27, alexjlzheng@gmail.com wrote:
> > > --- a/fs/proc/array.c
> > > +++ b/fs/proc/array.c
> > > @@ -528,7 +528,9 @@ static int do_task_stat(struct seq_file *m, struct pid_namespace *ns,
> > >  		}
> > >
> > >  		sid = task_session_nr_ns(task, ns);
> > > -		ppid = task_tgid_nr_ns(task->real_parent, ns);
> > > +		rcu_read_lock();
> > > +		ppid = task_tgid_nr_ns(rcu_dereference(task->real_parent), ns);
> > > +		rcu_read_unlock();
> >
> > But this can't really help. If task->real_parent has already exited and
> > it was reaped, then it is actually "Too late!" for rcu_read_lock().
> >
> > Please use task_ppid_nr_ns() which does the necessary pid_alive() check.

Ah, I was wrong, I forgot about lock_task_sighand(task). So in this case
pid_alive() is not necessary, and the patch is fine.

But unless you have a strong opinion, I'd still suggest to use
task_ppid_nr_ns(), see below.

> Suppose it fits the time window between the current parent exiting and
> the task being reassigned to init. Then you transiently see 0 as the pid,
> instead of 1 (or whatever). This reads like a bug to me.

But we can't avoid this. Without tasklist_lock even

 	task_tgid_nr_ns(current->real_parent, ns);

can return zero if we race with reparenting. If ->real_parent is reaped
right after we read the ->real_parent pointer, it has no pids. See
__unhash_process() -> detach_pid().

> It probably should do precisely the same thing proposed in this patch,
> as in:
> 	rcu_read_lock();
> 	ppid = task_tgid_nr_ns(rcu_dereference(task->real_parent), ns);
> 	rcu_read_unlock();

No, task_ppid_nr_ns(tsk) does need the pid_alive() check. If tsk exits,
tsk->real_parent points to nowhere, rcu_read_lock() can't help.

This all needs cleanups. ->real_parent and ->group_leader need the helpers
(probably with some CONFIG_PROVE_RCU checks) and they should be moved to
signal_struct.

So far I have only sent some trivial initial cleanups/preparations, see
https://lore.kernel.org/all/aXY_h8i78n6yD9JY@redhat.com/

I'll try to do the next step this week. If I have time ;) I am on a
forced PTO caused by renovations in our apartment.

Oleg.