Re: [PATCH 3/4] sched_ext: Introduce SCX_OPS_NODE_BUILTIN_IDLE

[Andrea Righi <arighi@nvidia.com>]

On Wed, Dec 11, 2024 at 10:21:49AM -0800, Yury Norov wrote:
...
> > +	/*
> > +	 * Check if we need to enable per-node cpumasks.
> > +	 */
> > +	if (ops->flags & SCX_OPS_BUILTIN_IDLE_PER_NODE)
> > +		static_branch_enable_cpuslocked(&scx_builtin_idle_per_node);
> > +	else
> > +		static_branch_disable_cpuslocked(&scx_builtin_idle_per_node);
> >  }
> 
> The patch that introduces the flag should go the very first in the series,
> but should unconditionally disable scx_builtin_idle_per_node.

Ack, that's a good idea.

> 
> The following patches should add all the machinery you need. The machinery
> should be conditional on the scx_builtin_idle_per_node, i.e. disabled for
> a while.
> 
> Doing that, you'll be able to introduce your functionality as a whole:
> 
>    static struct cpumask *get_idle_cpumask_node(int node)
>    {
>    	if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
>    		return idle_masks[0]->cpu;
>    
>    	return idle_masks[node]->cpu;
>    }
> 
> Much better than patching just introduced code, right?

Agreed.

> 
> The very last patch should only be a chunk that enables scx_builtin_idle_per_node
> based on SCX_OPS_BUILTIN_IDLE_PER_NODE.
> 
> This way, when your feature will get merged, from git-bisect perspective
> it will be enabled atomically by the very last patch, but those interested
> in internals will have nice coherent history.

Makes sense, I'll refactor this in the next version, thanks!

-Andrea