Re: [PATCH RFC] rtmutex: Permit rt_mutex_unlock() to be invoked withirqs disabled

[Arnaud Lacombe]

Hi,

On Sun, Jul 24, 2011 at 11:56 AM, Paul E. McKenney
<paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> On Sun, Jul 24, 2011 at 11:00:41AM +0200, Thomas Gleixner wrote:
>> On Sat, 23 Jul 2011, Paul E. McKenney wrote:
>> > On Sun, Jul 24, 2011 at 02:05:13AM +0200, Thomas Gleixner wrote:
>> > > On Sun, 24 Jul 2011, Thomas Gleixner wrote:
>> > > > > > Thomas, I'm inclined to merge this, any objections?
>> > > > >
>> > > > > FWIW, it has been passing tests here.
>> > > >
>> > > > If it's only the unlock path, I'm fine with that change.
>> > > >
>> > > > Acked-by-me
>> > >
>> > > Hrmpft. That's requiring all places to take the lock irq safe. Not
>> > > really amused. For -RT that's a hotpath and we can really do without
>> > > the irq fiddling there. That needs a bit more thought.
>> >
>> > Indeed...  If I make only some of the lock acquisitions irq safe, lockdep
>> > will yell at me.  And rightfully so, as that could result in deadlock.
>> >
>> > So, what did you have in mind?
>>
>> Have no real good idea yet for this. Could you grab rt and check
>> whether you can observe any impact when the patch is applied?
>
> Hmmm, wait a minute...  There might be a way to do this with zero
> impact on the fastpath, given that I am allocating an rt_mutex on
> the stack that is used only by RCU priority boosting, and that only
> rt_mutex_init_proxy_locked(), rt_mutex_lock(), and rt_mutex_unlock()
> are used.
>
> So I could do the following:
>
> o       Use lockdep_set_class_and_name() to make the ->wait_lock()
>        field of my rt_mutex have a separate lockdep class.  I guess
>        I should allocate a global variable for lock_class_key
>        rather than allocating it on the stack.  ;-)
>
> o       Make all calls from RCU priority boosting to rt_mutex_lock()
>        and rt_mutex_unlock() have irqs disabled.
>
> o       Make __rt_mutex_slowlock() do the following when sleeping:
>
>        raw_spin_unlock(&lock->wait_lock);
>
>        debug_rt_mutex_print_deadlock(waiter);
>
>        {
>                int was_disabled = irqs_disabled();
>
>                if (was_disabled)
>                        local_irq_enable();
>
FWIW, the final construct you opted for in -next:

    if (was_disabled = irqs_disabled())
        local_irq_enable();

triggers:

/linux/linux/kernel/rtmutex.c: In function '__rt_mutex_slowlock':
/linux/linux/kernel/rtmutex.c:605:3: warning: suggest parentheses
around assignment used as truth value

 - Arnaud

>                schedule_rt_mutex(lock);
>
>                if (was_disabled)
>                        local_irq_disable();
>
>        }
>
>        raw_spin_lock(&lock->wait_lock);
>        set_current_state(state);
>
> Would that work reasonably?
>
>                                                        Thanx, Paul
> --
> 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/
>
--
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/