Re: [PATCH] Fixing directly deferencing a __rcu pointer warning

[Oleg Nesterov]

On 10/26, Peter Zijlstra wrote:
>
> >  kernel/fork.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 10917c3e1f03..802b7bbe3d92 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -2369,7 +2369,7 @@ __latent_entropy struct task_struct *copy_process(
> >
> >  	retval = -EAGAIN;
> >  	if (is_rlimit_overlimit(task_ucounts(p), UCOUNT_RLIMIT_NPROC, rlimit(RLIMIT_NPROC))) {
> > -		if (p->real_cred->user != INIT_USER &&
> > +		if (rcu_dereference(p->real_cred)->user != INIT_USER &&
> >  		    !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN))
> >  			goto bad_fork_cleanup_count;
> >  	}
>
> This seems entirely misguided and only makes the code more confusing.
>
> AFAICT at this point @p is not life, we're constructing the new task,
> but it's not yet published, therefore no concurrency possible.
> Additionally we're not actually in an RCU critical section afaict.
>
> > @@ -2692,7 +2692,7 @@ __latent_entropy struct task_struct *copy_process(
> >  			 */
> >  			p->signal->has_child_subreaper = p->real_parent->signal->has_child_subreaper ||
> >  							 p->real_parent->signal->is_child_subreaper;
> > -			list_add_tail(&p->sibling, &p->real_parent->children);
> > +			list_add_tail(&p->sibling, &(rcu_dereference(p->real_parent)->children));
> >  			list_add_tail_rcu(&p->tasks, &init_task.tasks);
> >  			attach_pid(p, PIDTYPE_TGID);
> >  			attach_pid(p, PIDTYPE_PGID);
>
> As to the real_parent, we hold the tasklist lock, which is the write
> side lock for parent stuff, so rcu dereference is pointless here.

Agreed.

Plus I don't think this change is correct, iiuc rcu_dereference() will trigger
the run-time "suspicious rcu_dereference_check() usage" warning, it is called
without rcu_read_lock().

Oleg.