Re: [PATCH 4/4] sched_ext: Introduce NUMA aware idle cpu kfunc helpers

[Andrea Righi <arighi@nvidia.com>]

On Wed, Dec 11, 2024 at 09:43:26AM -0800, Yury Norov wrote:
> On Mon, Dec 09, 2024 at 11:40:58AM +0100, Andrea Righi wrote:
> > Add the following kfunc's to provide scx schedulers direct access to
> > per-node idle cpumasks information:
> > 
> >  const struct cpumask *scx_bpf_get_idle_cpumask_node(int node)
> >  const struct cpumask *scx_bpf_get_idle_smtmask_node(int node)
> >  s32 scx_bpf_pick_idle_cpu_node(int node,
> >                                 const cpumask_t *cpus_allowed, u64 flags)
> >  int scx_bpf_cpu_to_node(s32 cpu)
> > 
> > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> > ---
> >  kernel/sched/ext.c                       | 96 +++++++++++++++++++++++-
> >  tools/sched_ext/include/scx/common.bpf.h |  4 +
> >  tools/sched_ext/include/scx/compat.bpf.h | 19 +++++
> >  3 files changed, 117 insertions(+), 2 deletions(-)
> > 
> > diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> > index d0d57323bcfc..ea7cc481782c 100644
> > --- a/kernel/sched/ext.c
> > +++ b/kernel/sched/ext.c
> > @@ -433,6 +433,7 @@ struct sched_ext_ops {
> >  	 * - scx_bpf_select_cpu_dfl()
> >  	 * - scx_bpf_test_and_clear_cpu_idle()
> >  	 * - scx_bpf_pick_idle_cpu()
> > +	 * - scx_bpf_pick_idle_cpu_node()
> >  	 *
> >  	 * The user also must implement ops.select_cpu() as the default
> >  	 * implementation relies on scx_bpf_select_cpu_dfl().
> > @@ -955,6 +956,8 @@ 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;
> >  
> > +	if (node < 0 || node >= num_possible_nodes())
> > +		return NULL;
> 
> 1. This sanity should go before the check above.
> 2. In-kernel users don't need to do sanity checks. BPF users should,
>    but for them you need to move it in BPF wrapper.
> 3. -1 is a valid parameter, means NUMA_NO_NODE. 

Ok, but what would you return with NUMA_NO_NODE, in theory we should return
a global system-wide cpumask, that doesn't exist with the per-node
cpumasks. Maybe just return cpu_none_mask? That's what I've done in the
next version, that seems safer than returning NULL.

> 
> >  	return idle_masks[node]->cpu;
> >  }
> >  
> > @@ -963,6 +966,8 @@ static struct cpumask *get_idle_smtmask_node(int node)
> >  	if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
> >  		return idle_masks[0]->smt;
> >  
> > +	if (node < 0 || node >= num_possible_nodes())
> > +		return NULL;
> >  	return idle_masks[node]->smt;
> >  }
> >  
> > @@ -7469,6 +7474,16 @@ __bpf_kfunc u32 scx_bpf_nr_cpu_ids(void)
> >  	return nr_cpu_ids;
> >  }
> >  
> > +/**
> > + * scx_bpf_cpu_to_node - Return the NUMA node the given @cpu belongs to
> > + */
> > +__bpf_kfunc int scx_bpf_cpu_to_node(s32 cpu)
> > +{
> > +	if (cpu < 0 || cpu >= nr_cpu_ids)
> > +		return -EINVAL;
> > +	return cpu_to_node(cpu);
> > +}
> 
> I believe this wrapper should be declared somewhere in
> kernel/sched/topology.c, and better be a separate patch.

Maybe kernel/bpf/helpers.c? And name it bpf_cpu_to_node()?

> 
> > +
> >  /**
> >   * scx_bpf_get_possible_cpumask - Get a referenced kptr to cpu_possible_mask
> >   */
> > @@ -7499,11 +7514,32 @@ __bpf_kfunc void scx_bpf_put_cpumask(const struct cpumask *cpumask)
> >  	 */
> >  }
> >  
> > +/**
> > + * scx_bpf_get_idle_cpumask_node - Get a referenced kptr to the idle-tracking
> > + * per-CPU cpumask of a target NUMA node.
> > + *
> > + * Returns an empty cpumask if idle tracking is not enabled, if @node is not
> > + * valid, or running on a UP kernel.
> > + */
> > +__bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask_node(int node)
> > +{
> > +	if (!static_branch_likely(&scx_builtin_idle_enabled)) {
> > +		scx_ops_error("built-in idle tracking is disabled");
> > +		return cpu_none_mask;
> > +	}
> > +	if (!static_branch_likely(&scx_builtin_idle_per_node)) {
> > +		scx_ops_error("per-node idle tracking is disabled");
> > +		return cpu_none_mask;
> > +	}
> 
> Nub question: is it possible that scx_builtin_idle_per_node is enable,
> but scx_builtin_idle_enabled not? From my naive perspective, we can't
> enable per-node idle masks without enabling general idle masks. Or I
> mislead it?

In theory a BPF scheduler could set SCX_OPS_BUILTIN_IDLE_PER_NODE (without
SCX_OPS_KEEP_BUILTIN_IDLE) in .flags while implementing ops.update_idle().

In this way we would have scx_builtin_idle_enabled==false and
scx_builtin_idle_per_node==true, which doesn't make much sense, so we
should probably handle this case in validate_ops() and trigger an error.

Good catch!

> 
> > +
> > +	return get_idle_cpumask_node(node) ? : cpu_none_mask;
> > +}
> >  /**
> >   * scx_bpf_get_idle_cpumask - Get a referenced kptr to the idle-tracking
> >   * per-CPU cpumask of the current NUMA node.
> >   *
> > - * Returns NULL if idle tracking is not enabled, or running on a UP kernel.
> > + * Returns an emtpy cpumask if idle tracking is not enabled, or running on a UP
> > + * kernel.
> >   */
> >  __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(void)
> >  {
> > @@ -7515,12 +7551,35 @@ __bpf_kfunc const struct cpumask *scx_bpf_get_idle_cpumask(void)
> >  	return get_curr_idle_cpumask();
> >  }
> >  
> > +/**
> > + * scx_bpf_get_idle_smtmask_node - Get a referenced kptr to the idle-tracking,
> > + * per-physical-core cpumask of a target NUMA node. Can be used to determine
> > + * if an entire physical core is free.
> > + *
> > + * Returns an empty cpumask if idle tracking is not enabled, if @node is not
> > + * valid, or running on a UP kernel.
> > + */
> > +__bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask_node(int node)
> > +{
> > +	if (!static_branch_likely(&scx_builtin_idle_enabled)) {
> > +		scx_ops_error("built-in idle tracking is disabled");
> > +		return cpu_none_mask;
> > +	}
> > +	if (!static_branch_likely(&scx_builtin_idle_per_node)) {
> > +		scx_ops_error("per-node idle tracking is disabled");
> > +		return cpu_none_mask;
> > +	}
> 
> Can you add vertical spacing between blocks?

You mean a blank between the two blocks, right?

Anyway, ...

> 
> Also, because you use this construction more than once, I think it
> makes sense to make it a helper.

With a proper error check in validate_ops() we can just get rid of the
scx_builtin_idle_enabled block and simply check scx_builtin_idle_per_node.

> 
> > +
> > +	return get_idle_smtmask_node(node) ? : cpu_none_mask;
> > +}
> > +
> >  /**
> >   * scx_bpf_get_idle_smtmask - Get a referenced kptr to the idle-tracking,
> >   * per-physical-core cpumask of the current NUMA node. Can be used to determine
> >   * if an entire physical core is free.
> >   *
> > - * Returns NULL if idle tracking is not enabled, or running on a UP kernel.
> > + * Returns an empty cumask if idle tracking is not enabled, or running on a UP
> > + * kernel.
> >   */
> >  __bpf_kfunc const struct cpumask *scx_bpf_get_idle_smtmask(void)
> >  {
> > @@ -7569,6 +7628,35 @@ __bpf_kfunc bool scx_bpf_test_and_clear_cpu_idle(s32 cpu)
> >  		return false;
> >  }
> >  
> > +/**
> > + * scx_bpf_pick_idle_cpu_node - Pick and claim an idle cpu from a NUMA node
> > + * @node: target NUMA node
> > + * @cpus_allowed: Allowed cpumask
> > + * @flags: %SCX_PICK_IDLE_CPU_* flags
> > + *
> > + * Pick and claim an idle cpu in @cpus_allowed from the NUMA node @node.
> > + * Returns the picked idle cpu number on success. -%EBUSY if no matching cpu
> > + * was found.
> > + *
> > + * Unavailable if ops.update_idle() is implemented and
> > + * %SCX_OPS_KEEP_BUILTIN_IDLE is not set or if %SCX_OPS_KEEP_BUILTIN_IDLE is
> > + * not set.
> > + */
> > +__bpf_kfunc s32 scx_bpf_pick_idle_cpu_node(int node, const struct cpumask *cpus_allowed,
> > +				      u64 flags)
> > +{
> 
> Sanity checks here?

Indeed, thanks!

-Andrea