Re: [PATCH v2 1/2] rtmutex Real-Time Linux: Fixing kernel BUG at kernel/locking/rtmutex.c:997!

[Mike Galbraith]

On Tue, 2015-04-07 at 12:29 +0200, Peter Zijlstra wrote:
> On Tue, Apr 07, 2015 at 07:09:43AM +0200, Mike Galbraith wrote:
> > On Mon, 2015-04-06 at 21:59 -0400, Steven Rostedt wrote:
> > > 
> > > We really should have a rt_spin_trylock_in_irq() and not have the
> > > below if conditional.
> > > 
> > > The paths that will be executed in hard irq context are static. 
> > > They
> > > should be labeled as such.
> > 
> > +/*
> > + * Special purpose for locks taken in interrupt context: Take and 
> > hold
> > + * ->wait_lock lest PI catching us with our fingers in the cookie 
> > jar.
> > + * Do NOT abuse.
> > + */
> > +int __lockfunc rt_spin_trylock_in_irq(spinlock_t *lock)
> > +{
> > +       struct task_struct *owner;
> > +       if (!raw_spin_trylock(&lock->lock.wait_lock))
> > +               return 0;
> > +       owner = idle_task(raw_smp_processor_id());
> > +       if (!(rt_mutex_cmpxchg(&lock->lock, NULL, owner))) {
> > +               raw_spin_unlock(&lock->lock.wait_lock);
> > +               return 0;
> > +       }
> > +       spin_acquire(&lock->dep_map, 0, 1, _RET_IP_);
> > +       return 1;
> > +}
> > +
> > +/* ONLY for use with rt_spin_trylock_in_irq(), do NOT abuse. */
> > +void __lockfunc rt_spin_trylock_in_irq_unlock(spinlock_t *lock)
> > +{
> > +       struct task_struct *owner = 
> > idle_task(raw_smp_processor_id());
> > +       /* NOTE: we always pass in '1' for nested, for simplicity 
> > */
> > +       spin_release(&lock->dep_map, 1, _RET_IP_);
> > +       BUG_ON(!(rt_mutex_cmpxchg(&lock->lock, owner, NULL)));
> > +       raw_spin_unlock(&lock->lock.wait_lock);
> > +}
> > +
> 
> Can someone explain this braindamage? You should _NOT_ take mutexes 
> in
> hardirq context.

No.. really? ;-)

If you have a spot where it'd be nice to do that despite it being 
somewhat (koff).. discouraged shall we say, you have to do something 
funky.  Thomas had a patch to not raise sirq unconditionally for -rt 
to let nohz_full work, but it needs a lock that's converted to an 
rtmutex in -rt, and which doesn't want to be un-converted.  Ergo, get 
funky.

> And if its an irq thread, then the irq thread _IS_ the right owner, 
> the
> thread needs to be boosted by waiters.

No irq thread.

> The idle thread cannot ever be owner of a mutex, that's complete and
> utter bullshit.

Not if you want to hide current from the deadlock detector lest it get 
upset and make box go boom.

        -Mike
--
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/