Re: [PATCH 6/6] sched_ext: idle: Introduce node-aware idle cpu kfunc helpers

[Andrea Righi <arighi@nvidia.com>]

On Fri, Feb 07, 2025 at 12:39:57PM -1000, Tejun Heo wrote:
> Hello,
> 
> On Fri, Feb 07, 2025 at 09:40:53PM +0100, Andrea Righi wrote:
> > +/**
> > + * scx_bpf_cpu_to_node - Return the NUMA node the given @cpu belongs to
> > + */
> > +__bpf_kfunc int scx_bpf_cpu_to_node(s32 cpu)
> 
> Maybe scx_bpf_cpu_node() to be in line with scx_bpf_task_cpu/cgroup()?

Ok, then maybe we can have scx_bpf_cpu_node() for the kfunc, that wraps
scx_cpu_node() for internal use.

> 
> > +{
> > +#ifdef CONFIG_NUMA
> > +	if (cpu < 0 || cpu >= nr_cpu_ids)
> > +		return -EINVAL;
> 
> Use ops_cpu_valid()? Otherwise, we can end up calling cpu_to_node() with an
> impossible CPU. Also, I don't think CPU -> node mapping function should be
> able to return an error value. It should just trigger ops error.

Ok.

> 
> > +
> > +	return idle_cpu_to_node(cpu);
> 
> This is contingent on scx_builtin_idle_per_node, right? It's confusing for
> CPU -> node mapping function to return NUMA_NO_NODE depending on an ops
> flag. Shouldn't this be a generic mapping function?

The idea is that BPF schedulers can use this kfunc to determine the right
idle cpumask to use, for example a typical usage could be:

  int node = scx_bpf_cpu_node(prev_cpu);
  s32 cpu = scx_bpf_pick_idle_cpu_in_node(p->cpus_ptr, node, SCX_PICK_IDLE_IN_NODE);

Or:

  int node = scx_bpf_cpu_node(prev_cpu);
  const struct cpumask *idle_cpumask = scx_bpf_get_idle_cpumask_node(node);

When SCX_OPS_BUILTIN_IDLE_PER_NODE is disabled, we need to point to the
global idle cpumask, that is identified by NUMA_NO_NODE, so this is why we
can return NUMA_NO_NODE fro scx_bpf_cpu_node().

Do you think we should make this more clear / document this better. Or do
you think we should use a different API?

> 
> > index 50e1499ae0935..caa1a80f9a60c 100644
> > --- a/tools/sched_ext/include/scx/compat.bpf.h
> > +++ b/tools/sched_ext/include/scx/compat.bpf.h
> > @@ -130,6 +130,25 @@ bool scx_bpf_dispatch_vtime_from_dsq___compat(struct bpf_iter_scx_dsq *it__iter,
> >  	 scx_bpf_now() :							\
> >  	 bpf_ktime_get_ns())
> >  
> > +#define __COMPAT_scx_bpf_cpu_to_node(cpu)					\
> > +	(bpf_ksym_exists(scx_bpf_cpu_to_node) ?					\
> > +	 scx_bpf_cpu_to_node(cpu) : 0)
> > +
> > +#define __COMPAT_scx_bpf_get_idle_cpumask_node(node)				\
> > +	(bpf_ksym_exists(scx_bpf_get_idle_cpumask_node) ?			\
> > +	 scx_bpf_get_idle_cpumask_node(node) :					\
> > +	 scx_bpf_get_idle_cpumask())						\
> > +
> > +#define __COMPAT_scx_bpf_get_idle_smtmask_node(node)				\
> > +	(bpf_ksym_exists(scx_bpf_get_idle_smtmask_node) ?			\
> > +	 scx_bpf_get_idle_smtmask_node(node) :					\
> > +	 scx_bpf_get_idle_smtmask())
> > +
> > +#define __COMPAT_scx_bpf_pick_idle_cpu_node(cpus_allowed, node, flags)		\
> > +	(bpf_ksym_exists(scx_bpf_pick_idle_cpu_node) ?				\
> > +	 scx_bpf_pick_idle_cpu_node(cpus_allowed, node, flags) :		\
> > +	 scx_bpf_pick_idle_cpu(cpus_allowed, flags))
> 
> Can you please document when these compat macros can be dropped? Also,
> shouldn't it also provide a compat macro for the new ops flag using
> __COMPAT_ENUM_OR_ZERO()? Otherwise, trying to load new binary using the new
> flag on an older kernel will fail, right?

Right. Will add that.

Thanks,
-Andrea