Re: [RFC PATCH 07/11] sched: Add proxy execution

[Peter Zijlstra]

On Mon, Oct 03, 2022 at 09:44:57PM +0000, Connor O'Brien wrote:
> @@ -703,15 +707,28 @@ __mutex_lock_common(struct mutex *lock, unsigned int state, unsigned int subclas
>  			break;
>  
>  		if (first) {
> +			bool acquired;
> +
> +			/*
> +			 * XXX connoro: mutex_optimistic_spin() can schedule, so

I'm thinking it might be better to remove that schedule point for
PROXY_EXEC instead of doing this.

Getting preempted right after acquiring a lock is less of a problem with
PE on than it is without it.

> +			 * we need to release these locks before calling it.
> +			 * This needs refactoring though b/c currently we take
> +			 * the locks earlier than necessary when proxy exec is
> +			 * disabled and release them unnecessarily when it's
> +			 * enabled. At a minimum, need to verify that releasing
> +			 * blocked_lock here doesn't create any races.
> +			 */
> +			raw_spin_unlock(&current->blocked_lock);
> +			raw_spin_unlock_irqrestore(&lock->wait_lock, flags);
>  			trace_contention_begin(lock, LCB_F_MUTEX | LCB_F_SPIN);
> -			if (mutex_optimistic_spin(lock, ww_ctx, &waiter))
> +			acquired = mutex_optimistic_spin(lock, ww_ctx, &waiter);
> +			raw_spin_lock_irqsave(&lock->wait_lock, flags);
> +			raw_spin_lock(&current->blocked_lock);
> +			if (acquired)
>  				break;
>  			trace_contention_begin(lock, LCB_F_MUTEX);
>  		}