lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20110715155916.GB2327@linux.vnet.ibm.com>
Date:	Fri, 15 Jul 2011 08:59:16 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Peter Zijlstra <a.p.zijlstra@...llo.nl>
Cc:	Ed Tomlinson <edt@....ca>, Steven Rostedt <rostedt@...dmis.org>,
	Sergey Senozhatsky <sergey.senozhatsky@...il.com>,
	Ingo Molnar <mingo@...e.hu>,
	Thomas Gleixner <tglx@...utronix.de>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Dipankar Sarma <dipankar@...ibm.com>,
	linux-kernel@...r.kernel.org
Subject: Re: INFO: possible circular locking dependency detected

On Fri, Jul 15, 2011 at 05:04:27PM +0200, Peter Zijlstra wrote:
> On Fri, 2011-07-15 at 07:36 -0700, Paul E. McKenney wrote:
> > On Fri, Jul 15, 2011 at 03:07:39PM +0200, Peter Zijlstra wrote:
> > > On Fri, 2011-07-15 at 05:42 -0700, Paul E. McKenney wrote:
> > > > On Fri, Jul 15, 2011 at 01:29:22PM +0200, Peter Zijlstra wrote:
> > > > > On Fri, 2011-07-15 at 07:05 -0400, Ed Tomlinson wrote:
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426] -> #1 (rcu_node_level_0){..-...}:
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff8108b7e5>] lock_acquire+0x95/0x140
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff8157808b>] _raw_spin_lock+0x3b/0x50
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff810ba797>] __rcu_read_unlock+0x197/0x2d0
> > 
> > Yow...  Looks like rcu_read_unlock_special() is being inlined
> > into __rcu_read_unlock().
> 
> Yeah, *cheer* for gcc being a smarty-pants.
> 
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff8103f2f5>] select_task_rq_fair+0x585/0xa80
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff8104633b>] try_to_wake_up+0x17b/0x360
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff81046575>] wake_up_process+0x15/0x20
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff810528f4>] irq_exit+0xb4/0x100
> > 
> > OK, so all the above stuff is in the context of an irq handler, right?
> 
> yep

OK...

> > In which case, why didn't the in_irq() check kick us out before we
> > had a chance to attempt to acquire any locks?
> 
> Because we're in irq_exit(), after decrementing preempt_count, so
> in_irq() returns false.

Can we delay decrementing preempt_count so that RCU has some chance
of actually working?

> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff8158197e>] smp_apic_timer_interrupt+0x6e/0x99
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff81580c53>] apic_timer_interrupt+0x13/0x20
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff810ba6e9>] __rcu_read_unlock+0xe9/0x2d0 
> > > > > > Jul 14 23:21:18 grover kernel: [  920.659426]        [<ffffffff814c20d4>] sock_def_readable+0x94/0xc0
> > > > > 
> > > > > Ed, are you perchance running with force_irqthreads?
> > 
> > Ah!  Would that mean that local_irq_save() gets mapped to locking?
> > Now -that- could be exciting!  ;-)
> 
> Nope, it simply makes the invoke_softirq() call in irq_exit() do an
> unconditional wakeup of ksoftirqd/# since there isn't an irq-tail to
> speak of.

That would be a bad thing for rcu_read_unlock_special() to call from
within a hardware irq handler.  :-/

> > > > > Paul, what appears to be happening here is that some rcu_read_unlock()
> > > > > gets interrupted, possibly before calling rcu_read_unlock_special(),
> > > > > possibly not if the interrupt is itself the timer interrupt.
> > > > > 
> > > > > Supposing ->rcu_read_unlock_special is set before, any wakeup happening
> > > > > from an interrupt hitting __rcu_read_unlock():
> > > > > 
> > > > > void __rcu_read_unlock(void)
> > > > > {
> > > > >         struct task_struct *t = current;
> > > > >                 
> > > > >         barrier();  /* needed if we ever invoke rcu_read_unlock in rcutree.c */
> > > > >         --t->rcu_read_lock_nesting;
> > > > >         barrier();  /* decrement before load of ->rcu_read_unlock_special */
> > > > >         if (t->rcu_read_lock_nesting == 0 &&
> > > > >             unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
> > > > >                 rcu_read_unlock_special(t);
> > > > > #ifdef CONFIG_PROVE_LOCKING
> > > > >         WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0);
> > > > > #endif /* #ifdef CONFIG_PROVE_LOCKING */
> > > > > }
> > > > > 
> > > > > After --t->rcu_read_lock_nesting, but before calling
> > > > > rcu_read_unlock_special(), will trigger this lock inversion.
> > > > > 
> > > > > The alternative case, ->rcu_read_unlock_special is not set yet, it can
> > > > > be set if the interrupt hitting in that same spot above, is the timer
> > > > > interrupt, and the wakeup happens either from the softirq ran from the
> > > > > hard IRQ tail, or as I suspect here happens, the wakeup of ksoftirqd/#.
> > > 
> > > OK, so the latter case cannot happen (rcu_preempt_check_callbacks only
> > > sets NEED_QS when rcu_read_lock_nesting), we need two interrupts for
> > > this to happen.
> > > 
> > > rcu_read_lock()
> > > 
> > >  <IRQ>
> > >    |= RCU_READ_UNLOCK_NEED_QS
> > > 
> > > rcu_read_unlock()
> > >   __rcu_read_unlock()
> > >    --rcu_read_lock_nesting;
> > >      <IRQ>
> > > 	ttwu()
> > >           rcu_read_lock()
> > > 	  rcu_read_unlock()
> > > 	    rcu_read_unlock_special()
> > > 	      *BANG*
> > >    rcu_read_unlock_special()
> > 
> > The "*BANG*" indicating that the upper-level rcu_read_unlock_special()
> > might overwrite the lower-level rcu_read_unlock_special()'s attempt
> > to clear RCU_READ_UNLOCK_NEED_QS? 
> 
> No, the *BANG* being that we end up calling rcu_read_unlock_special()
> while holding scheduler locks, which is BAD(tm).

Well, it certainly is BAD(tm) if you guys continue to deprive
rcu_read_unlock_special() of the means of determining whether it is
being invoked from hardware irq handler context.

> >  (Which I believe, perhaps
> > incorrectly, to be prevented by the fact that all modifications to
> > ->rcu_read_unlock_special are carried out with irqs disabled on the
> > corresponding CPU, at least given no RCU_BOOST.)  The check for in_irq()
> > should prevent the from-irq rcu_read_unlock_special() from attempting
> > to acquire any locks.
> 
> Right, so in_irq() simply checks a few bits in preempt_count, which we
> just cleared due to being in irq_exit().

Right.  So how about delaying clearing those bits until after you get
done messing with the scheduler from hardware irq handler context?

> > Or am I missing the point of your example?
> 
> You were ;-)
> 
> > On the other points, to the extent that I have analyzed them so far:
> > 
> > 1.	If the task is preempted after the --rcu_read_lock_nesting,
> > 	it won't see it as being in an RCU read-side critical section,
> > 	so won't queue it.
> 
> Agreed.
> 
> > 2.	Of course, the task might have preempted earlier.  In this
> > 	case, the RCU_READ_UNLOCK_BLOCKED is already set, so we will be
> > 	invoking rcu_read_unlock_special() anyway.
> 
> Right, but
> 
> > 	If an interrupt happens between the decrement and the call to
> > 	rcu_read_unlock_special(), then, yes, the irq handler will also
> > 	call rcu_read_unlock_special() if it calls rcu_read_unlock(), but
> > 	the check for in_irq() will prevent the irq handler's invocation
> > 	of rcu_read_unlock_special() from acquiring any locks.
> 
> But in_irq() isn't sufficient for RCU usage after the hardirq ends, see
> irq_exit(). Also there's all of softirq to consider, that too can run
> and not get caught by in_irq().

Change the rules without adjusting the callers can in fact result in some
breakage.  ;-)

The bit about local_irq_save() and local_irq_restore() invoking the
scheduler is rather surprising -- is there a raw_ version that avoids
this?

> > 3.	It is possible that the task is preempted after the
> > 	--rcu_read_lock_nesting, in which case the task won't be queued.
> > 	Of course the task might already be queued if there was an
> > 	earlier preemption during this same RCU read-side critical
> > 	section, in which case #2 applies.
> > 
> > 	In other words, a preemption in __rcu_read_unlock() after the
> > 	--rcu_read_lock_nesting has no effect on RCU state: either the
> > 	task was already marked RCU_READ_UNLOCK_BLOCKED, or it wasn't.
> > 	Either way, rcu_note_context_switch() does not see this task as
> > 	being in an RCU read-side critical section.
> > 
> > So what am I missing here?
> 
>  $task			IRQ		SoftIRQ
> 
>  rcu_read_lock()
> 
>  /* do stuff */
> 
>  <preempt> |= UNLOCK_BLOCKED
> 
>  rcu_read_unlock()
>    --t->rcu_read_lock_nesting
> 
> 			irq_enter();
> 			/* do stuff, don't use RCU */
> 			irq_exit();
> 			  sub_preempt_count(IRQ_EXIT_OFFSET);
> 			  invoke_softirq()

Why can't we exchange the order of the above two so that RCU correctly
avoids messing with the scheduler if called from hardware interrupt
context?

> 
> 					ttwu();
> 					  spin_lock_irq(&pi->lock)
> 					  rcu_read_lock();
> 					  /* do stuff */
> 					  rcu_read_unlock();
> 					    rcu_read_unlock_special()
> 					      rcu_report_exp_rnp()
> 					        ttwu()
> 					          spin_lock_irq(&pi->lock) /* deadlock */
> 					      
> 
>    rcu_read_unlock_special(t);
> 
> Ed can simply trigger this 'easy' because invoke_softirq() immediately
> does a ttwu() of ksoftirqd/# instead of doing the in-place softirq stuff
> first, but even without that the above happens.

An easily reproduced bug is certainly a nice change of pace...

> Something like the below _might_ fix it..

Maybe, but how does tglx make PREEMPT_RT work in this case?  The problem
is that PREEMPT_RT allows ksoftirqd to be preempted, and thus allows it
to be RCU priority boosted.

							Thanx, Paul

> ---
>  kernel/rcutree_plugin.h |    2 +-
>  kernel/softirq.c        |   12 ++++++++++--
>  2 files changed, 11 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/rcutree_plugin.h b/kernel/rcutree_plugin.h
> index 14dc7dd..373c9c8 100644
> --- a/kernel/rcutree_plugin.h
> +++ b/kernel/rcutree_plugin.h
> @@ -309,7 +309,7 @@ static void rcu_read_unlock_special(struct task_struct *t)
>  	}
>  
>  	/* Hardware IRQ handlers cannot block. */
> -	if (in_irq()) {
> +	if (in_irq() || in_serving_softirq()) {
>  		local_irq_restore(flags);
>  		return;
>  	}
> diff --git a/kernel/softirq.c b/kernel/softirq.c
> index 40cf63d..fca82c3 100644
> --- a/kernel/softirq.c
> +++ b/kernel/softirq.c
> @@ -315,16 +315,24 @@ static inline void invoke_softirq(void)
>  {
>  	if (!force_irqthreads)
>  		__do_softirq();
> -	else
> +	else {
> +		__local_bh_disable((unsigned long)__builtin_return_address(0),
> +				SOFTIRQ_OFFSET);
>  		wakeup_softirqd();
> +		__local_bh_enable(SOFTIRQ_OFFSET);
> +	}
>  }
>  #else
>  static inline void invoke_softirq(void)
>  {
>  	if (!force_irqthreads)
>  		do_softirq();
> -	else
> +	else {
> +		__local_bh_disable((unsigned long)__builtin_return_address(0),
> +				SOFTIRQ_OFFSET);
>  		wakeup_softirqd();
> +		__local_bh_enable(SOFTIRQ_OFFSET);
> +	}
>  }
>  #endif
>  
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ