Re: 2.6.26-rc4: RIP find_pid_ns+0x6b/0xa0

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, May 27, 2008 at 08:03:03AM -0700, Linus Torvalds wrote:
> 
> 
> On Tue, 27 May 2008, Oleg Nesterov wrote:
> 
> > On 05/27, Alexey Dobriyan wrote:
> > >
> > > PREEMPT_RCU is in use, again.
> 
> I do wonder if PREEMPT_RCU is broken.

I never stop wondering that...

> > > 0xffffffff802447cb is in find_pid_ns (kernel/pid.c:297).
> > > 292             struct hlist_node *elem;
> > > 293             struct upid *pnr;
> > > 294
> > > 295             hlist_for_each_entry_rcu(pnr, elem,
> > > 296                             &pid_hash[pid_hashfn(nr, ns)], pid_chain)
> > > 297                     if (pnr->nr == nr && pnr->ns == ns)
> 
> > > general protection fault: 0000 [2] PREEMPT SMP DEBUG_PAGEALLOC
> > > RDX: 6b6b6b6b6b6b6b6b RSI: ffffffff80566760 RDI: 0000000000003cef
> 
> That repeated 0x6b is POISON_FREE, and the code is
> 
> 	cmp    -0x10(%rdx),%edi
> 
> which is the load of "pnr->nr". So 'pnr' has been free'd.
> 
> On Tue, 27 May 2008, Oleg Nesterov wrote:
> > 
> > Is this reproducible?
> > 
> > In theory find_pid() is not safe without rcu_read_lock() if CONFIG_PREEMPT_RCU.
> > But we have a lot of "read_lock(tasklist_lock) + find_pid()", this was legal
> > and documented. It was actually broken, but happened to work because read_lock()
> > implied rcu_read_lock().
> > 
> > Could you look at
> > 
> > 	[PATCH] fix tasklist + find_pid() with CONFIG_PREEMPT_RCU
> > 	http://marc.info/?t=120162615300012
> > 
> > ?
> > 
> > I am not sure this is the actual reason though, the race is very unlikely.
> 
> That is a *very* unlikely race, especially as that bad_fork_free_pid case 
> would only happen if pid_ns_prepare_proc() fails. And if it fails, it's 
> still very unlikely to hit, I think.
> 
> That said, it does smell like a bug. But I *really* would be much much 
> happier if even SRCU at least waited for a grace period, so that it would 
> always be safe to just disable preemption for a "rcu_read_lock()". That 
> way, things that take spinlocks are safe even with SRCU.

SRCU does wait for all CPUs to schedule, and thus already waits for all
pre-existing non-preemptable code sequences to finish on all CPUs.

> Paul? How hard would it be to make preemptable RCU just honor that classic 
> RCU behavior?

Hmmm...  Might not be too hard, I will look into this.  Should just be
another stage in the rcu_try_flip state machine, along with a few of
the changes already in the queue for call_rcu_sched().

But this will only help until preemptible spinlocks arrive, right?

							Thanx, Paul