Re: Quick review of -rt RCU-related patches
From: Paul E. McKenney
Date: Tue Oct 04 2011 - 19:56:15 EST
On Wed, Oct 05, 2011 at 01:27:15AM +0200, Thomas Gleixner wrote:
> On Tue, 4 Oct 2011, Paul E. McKenney wrote:
> > On Wed, Oct 05, 2011 at 12:05:47AM +0200, Thomas Gleixner wrote:
> > > > This means that might_sleep() will never complain about
> > > > blocking in an RCU read-side critical section. I guess that
> > > > this is necessary, though it would be better to have some
> > > > way to complain for general sleeping (e.g., waiting on
> > > > network receive) as opposed to blocking on a lock that is
> > > > subject to priority inheritance.
> > >
> > > Well, there's always a remaining problem. We need that stuff fully
> > > preemptible on rt. Any ideas ?
> > Not yet. We would have to classify context switches into two groups:
> > 1. Preemptions or blocking waiting for sleeping spinlocks.
> > 2. Everything else.
> > Given that classification, it would be straightforward: prohibit
> > group #2 context switches while in RCU-preempt read-side critical
> > sections. I know, easy for me to say! ;-)
> Well, you know the preemtible regions of RT, it basically boils down
> to #1 - except that it differs a bit from vanilla that locks and bh
> stuff does not prevent preemption.
> If RCU can deal with that, then #2 is a non issue :)
Ah! I thought you were asking for something that checked for people
violating -rt's rules for context switches within RCU-preempt read-side
> > > > rcu-disable-the-rcu-bh-stuff-for-rt.patch
> > > >
> > > > This implements RCU-bh in terms of RCU-preempt, but disables
> > > > BH around the resulting RCU-preempt read-side critical section.
> > > > May be vulnerable to network-based denial-of-service attacks,
> > > > which could OOM a system with this patch.
> > > >
> > > > The implementation of rcu_read_lock_bh_held() is weak, but OK. In
> > > > an ideal world, it would also complain if not local_bh_disable().
> > >
> > > Well, I dropped that after our IRC conversation, but we still need to
> > > have some extra debugging for RT to diagnose situations where we break
> > > those rcu_bh assumptions. That _bh rcu stuff should have never been
> > > there, we'd rather should drop the softirq processing back to
> > > ksoftirqd in such an overload case (in mainline) and voluntary
> > > schedule away from ksoftirqd until the situation is resolved.
> > >
> > > I consider rcu_bh a bandaid for the nasty implict semantics of BH
> > > processing and I'm looking really forward to Peters analysis of the
> > > modern cpu local BKL constructs at RTLWS.
> > >
> > > The patch stemmed from an earlier discussion about getting rid of
> > > those special rcu_bh semantics even in mainline for the sake of not
> > > making a special case for a completely over(ab)used construct which
> > > originates from the historically grown softirq behaviour. I think that
> > > keeping the special cased rcu_bh stuff around is just taking any
> > > incentive away from moving that whole softirq processing into a
> > > scheduler controllable environment (i.e. threaded interrupts).
> > Between -rt and the guys pushing packets, I can tell that this is going
> > to be a fun one. ;-)
> We'll see. At some point they'll find out that a thread context will
> make their life easier simply because the locking maze can be
> distangled. That's why I'm ranting at the special "foster _bh" rcu
> abominations, which keep them thinking that going down that road is
> actually a good thing.
> > I will see if I can come up with a way to make that patch safe to
> > apply. Might not be all that hard.
Easier now that "git clone" works for me once more. Re-installed my
system, but my notes from doing it last time were incomplete. They
are better now. ;-)
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/