On Thu, 12 Jun 2014, Paul E. McKenney wrote:
> On Thu, Jun 12, 2014 at 07:28:44PM +0200, Oleg Nesterov wrote:
> > On 06/11, Paul E. McKenney wrote:
> > >
> > > On Wed, Jun 11, 2014 at 07:59:34PM +0200, Oleg Nesterov wrote:
> > > > On 06/11, Paul E. McKenney wrote:
> > > > >
> > > > > I was thinking of ->boost_completion as the way to solve it easily,
> > > > > but
> > > > > what did you have in mind?
> > > >
> > > > I meant, rcu_boost() could probably just do "mtx->owner = t", we know
> > > > that
> > > > it was unlocked by us and nobody else can use it until we set
> > > > t->rcu_boost_mutex.
> > >
> > > My concern with this is that rcu_read_unlock_special() could
> > > hypothetically
> > > get preempted (either by kernel or hypervisor), so that it might be a long
> > > time until it makes its reference. But maybe that reference would be
> > > harmless in this case.
> >
> > Confused... Not sure I understand what did you mean, and certainly I do not
> > understand how this connects to the proxy-locking method.
> >
> > Could you explain?
>
> Here is the hypothetical sequence of events, which cannot happen unless
> the CPU releasing the lock accesses the structure after releasing
> the lock:
>
> CPU 0 CPU 1 (booster)
>
> releases boost_mutex
>
> acquires boost_mutex
> releases boost_mutex
> post-release boost_mutex access?
> Loops to next task to boost
> proxy-locks boost_mutex
>
> post-release boost_mutex access:
> confused due to proxy-lock
> operation?
>
> Now maybe this ends up being safe, but it sure feels like an accident
> waiting to happen. Some bright developer comes up with a super-fast
> handoff, and blam, RCU priority boosting takes it in the shorts. ;-)
> In contrast, using the completion prevents this.
>
> > > > And if we move it into rcu_node, then we can probably kill
> > > > ->rcu_boost_mutex,
> > > > rcu_read_unlock_special() could check rnp->boost_mutex->owner ==
> > > > current.
> > >
> > > If this was anywhere near a hot code path, I would be sorely tempted.
> >
> > Ah, but I didn't mean perfomance. I think it is always good to try to remove
> > something from task_struct, it is huge. I do not mean sizeof() in the first
> > place, the very fact that I can hardly understand the purpose of a half of
> > its
> > members makes me sad ;)
>
> Now -that- just might make a huge amount of sense! Let's see...
>
> o We hold the rcu_node structure's ->lock when checking the owner
> (looks like rt_mutex_owner() is the right API).
>
> o We hold the rcu_node structure's ->lock when doing
> rt_mutex_init_proxy_locked().
>
> o We -don't- hold ->lock when releasing the rt_mutex, but that
> should be OK: The owner is releasing it, and it is going to
> not-owned, so no other task can possibly see ownership moving
> to/from them.
>
> o The rcu_node structure grows a bit, but not enough to worry
> about, and on most systems, the decrease in task_struct size
> will more than outweigh the increase in rcu_node size.
>
> Looks quite promising, how about the following? (Hey, it builds, so it
> must be correct, right?)
True. Why should we have users if we would test the crap we produce?
Just FYI, I have a patch pending which makes the release safe.
http://marc.info/?l=linux-kernel&m=140251240630730&w=2
Thanks,
tglx
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
