On Sun, Aug 14, 2022 at 02:41:54PM +0200, Borislav Petkov wrote:
> On Sat, Aug 13, 2022 at 01:27:40PM -0700, Linus Torvalds wrote:
> > On Sat, Aug 13, 2022 at 9:25 AM Borislav Petkov <bp@alien8.de> wrote:
> > >
> > > That task_struct.sighand is marked __rcu and thus noderef and sparse
> > > complains:
> > 
> > I think that RCU marking is misleading.
> > 
> > Doing a
> > 
> >         git grep -e '->sighand'
> > 
> > shows that we basically never treat that as some kind of RCU pointer.
> > 
> > Adding a
> > 
> >         grep -i rcu
> > 
> > to the above shows that we have a couple of places that do this
> > carefully, but they are the exception rather than the rule.
> > 
> > I think the issue is that "current->sighand" is always safe (and that
> > "me->sighand" is the same thing), and that sighand has RCU-delayed
> > freeing so that __lock_task_sighand() can safely try to take the lock
> > of another process' sighand.
> > 
> > And we have no real way to explain to sparse that *some* cases are
> > fine, others are not and need the sighand lock (after that careful
> > __lock_task_sighand thing).
> 
> Sounds to me like that sparse check was not such a good idea in the
> first place. Especially since the 0day bot is probably warning about all
> those cases where we try to lock ->sighand.
> 
> It was added by
> 
> 913292c97d75 ("sched.h: Annotate sighand_struct with __rcu")
> 
> Lemme add the involved parties to Cc.

If it is causing more trouble than it is worth, then I have not objection
to taking a different approach.

							Thanx, Paul