Re: [RFC] Avoid mutex starvation when optimistic spinning is disabled

[Imre Deak]

On to, 2016-07-21 at 15:29 -0700, Jason Low wrote:
> On Wed, 2016-07-20 at 14:37 -0400, Waiman Long wrote:
> > On 07/20/2016 12:39 AM, Jason Low wrote:
> > > On Tue, 2016-07-19 at 16:04 -0700, Jason Low wrote:
> > > > Hi Imre,
> > > > 
> > > > Here is a patch which prevents a thread from spending too much
> > > > "time"
> > > > waiting for a mutex in the !CONFIG_MUTEX_SPIN_ON_OWNER case.
> > > > 
> > > > Would you like to try this out and see if this addresses the
> > > > mutex
> > > > starvation issue you are seeing in your workload when
> > > > optimistic
> > > > spinning is disabled?
> > > Although it looks like it didn't take care of the 'lock stealing'
> > > case
> > > in the slowpath. Here is the updated fixed version:
> > > 
> > > ---
> > > Signed-off-by: Jason Low<jason.low2@xxxxxxx>
> > > ---
> > > Â include/linux/mutex.hÂÂ|ÂÂ2 ++
> > > Â kernel/locking/mutex.c | 65
> > > ++++++++++++++++++++++++++++++++++++++++++++------
> > > Â 2 files changed, 60 insertions(+), 7 deletions(-)
> > > 
> > > diff --git a/include/linux/mutex.h b/include/linux/mutex.h
> > > index 2cb7531..c1ca68d 100644
> > > --- a/include/linux/mutex.h
> > > +++ b/include/linux/mutex.h
> > > @@ -57,6 +57,8 @@ struct mutex {
> > > Â #endif
> > > Â #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
> > > ÂÂ	struct optimistic_spin_queue osq; /* Spinner MCS lock
> > > */
> > > +#else
> > > +	bool yield_to_waiter; /* Prevent starvation when
> > > spinning disabled */
> > > Â #endif
> > > Â #ifdef CONFIG_DEBUG_MUTEXES
> > > ÂÂ	void			*magic;
> > 
> > You don't need that on non-SMP system. So maybe you should put it
> > under 
> > #ifdef CONFIG_SMP block.
> 
> Right, maybe something like:
> 
> ÂÂÂÂ#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
> 	...
> 	...
> ÂÂÂÂ#elif !defined(CONFIG_SMP) /* If optimistic spinning disabled */
> ÂÂÂÂÂÂÂÂbool yield_to_waiter;
> ÂÂÂÂ#endif
> 
> > > @@ -556,7 +595,8 @@ __mutex_lock_common(struct mutex *lock, long
> > > state, unsigned int subclass,
> > > ÂÂ		Â* other waiters. We only attempt the xchg if
> > > the count is
> > > ÂÂ		Â* non-negative in order to avoid unnecessary
> > > xchg operations:
> > > ÂÂ		Â*/
> > > -		if (atomic_read(&lock->count)>= 0&&
> > > +		if ((!need_yield_to_waiter(lock) || loop>ÂÂ1)&&
> > > +		ÂÂÂÂatomic_read(&lock->count)>= 0&&
> > > ÂÂ		(atomic_xchg_acquire(&lock->count, -1) == 1))
> > > ÂÂ	
> > 
> > I think you need to reset the yield_to_waiter variable here when
> > loop > 
> > 1 instead of at the end of the loop.
> 
> So I think in the current state, only the top waiter would be able to
> both set and clear the yield_to_waiter variable anyway. However, I
> agree
> that this detail is not obvious and it would be better to reset the
> variable here when loop > 1 to make it more readable.

AFAICS an interruptible waiter behind the top waiter receiving a signal
and grabbing the lock could also reset yield_to_waiter incorrectly in
that way, increasing the top waiter's delay arbitrarily.

--Imre

> 
> > > 		break;
> > > 
> > > @@ -581,6 +621,7 @@ __mutex_lock_common(struct mutex *lock, long
> > > state, unsigned int subclass,
> > > ÂÂ		spin_unlock_mutex(&lock->wait_lock, flags);
> > > ÂÂ		schedule_preempt_disabled();
> > > ÂÂ		spin_lock_mutex(&lock->wait_lock, flags);
> > > +		do_yield_to_waiter(lock, loop);
> > > ÂÂ	}
> > > ÂÂ	__set_task_state(task, TASK_RUNNING);
> > > 
> > > @@ -590,6 +631,10 @@ __mutex_lock_common(struct mutex *lock, long
> > > state, unsigned int subclass,
> > > ÂÂ		atomic_set(&lock->count, 0);
> > > ÂÂ	debug_mutex_free_waiter(&waiter);
> > > 
> > > +#ifndef CONFIG_MUTEX_SPIN_ON_OWNER
> > > +	lock->yield_to_waiter = false;
> > > +#endif
> > > +
> > 
> > Maybe you should do the reset in an inline function instead.
> 
> Yes, this should be abstracted into a function like we do with
> do_yield_to_waiter().
> 
> 
> Jason
>