Re: [PATCH 5/6] locking/rwsem: Optimize slowpath/sleeping

[Peter Zijlstra <peterz@infradead.org>]

On Sun, Jan 25, 2015 at 11:36:08PM -0800, Davidlohr Bueso wrote:
> When blocking , we incur in multiple barriers when setting the
> task's uninterruptable state. This is particularly bad when the
> lock keeps getting stolen from the task trying to acquire the sem.
> These changes propose delaying setting the task's new state until
> we are sure that calling schedule is inevitable.
> 
> This implies that we do the trylock and active check (both basically
> ->counter checks) as TASK_RUNNING. For the trylock we hold the wait
> lock with interrupts disabled, so no risk there. And for the active
> check, the window for which we could get interrupted is quite small
> and makes no tangible difference.
> 
> This patch increases Unixbench's 'execl' throughput by 25% on a 40
> core machine.
> 
> Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
> ---
>  kernel/locking/rwsem-xadd.c | 20 +++++++++++++++-----
>  1 file changed, 15 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
> index 18a50da..88b3468 100644
> --- a/kernel/locking/rwsem-xadd.c
> +++ b/kernel/locking/rwsem-xadd.c
> @@ -459,17 +459,27 @@ struct rw_semaphore __sched *rwsem_down_write_failed(struct rw_semaphore *sem)
>  		count = rwsem_atomic_update(RWSEM_WAITING_BIAS, sem);
>  
>  	/* wait until we successfully acquire the lock */
>  	while (true) {
>  		if (rwsem_try_write_lock(count, sem))
>  			break;
> +
> +		__set_current_state(TASK_UNINTERRUPTIBLE);
>  		raw_spin_unlock_irq(&sem->wait_lock);
>  
> +		/*
> +		 * When there are active locks after we wake up,
> +		 * the lock was probably stolen from us. Thus,
> +		 * go immediately back to sleep and avoid taking
> +		 * the wait_lock.
> +		 */
> +		while (true) {
>  			schedule();
> +
> +			count = READ_ONCE(sem->count);
> +			if (!(count & RWSEM_ACTIVE_MASK))
> +				break;
> +			__set_current_state(TASK_UNINTERRUPTIBLE);
> +		}

So its late and I'm not seeing it; why is this safe? How will we not
miss the wakeup that makes condition true?

>  
>  		raw_spin_lock_irq(&sem->wait_lock);
>  	}