Re: [PATCH 08/10] sched_ext: idle: introduce SCX_PICK_IDLE_NODE

[Yury Norov <yury.norov@gmail.com>]

On Tue, Dec 24, 2024 at 09:37:35AM +0100, Andrea Righi wrote:
> On Mon, Dec 23, 2024 at 07:53:21PM -0800, Yury Norov wrote:
> > On Mon, Dec 23, 2024 at 06:48:48PM -0800, Yury Norov wrote:
> > > On Fri, Dec 20, 2024 at 04:11:40PM +0100, Andrea Righi wrote:
> > > > Introduce a flag to restrict the selection of an idle CPU to a specific
> > > > NUMA node.
> > > > 
> > > > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> > > > ---
> > > >  kernel/sched/ext.c      |  1 +
> > > >  kernel/sched/ext_idle.c | 11 +++++++++--
> > > >  2 files changed, 10 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> > > > index 143938e935f1..da5c15bd3c56 100644
> > > > --- a/kernel/sched/ext.c
> > > > +++ b/kernel/sched/ext.c
> > > > @@ -773,6 +773,7 @@ enum scx_deq_flags {
> > > >  
> > > >  enum scx_pick_idle_cpu_flags {
> > > >  	SCX_PICK_IDLE_CORE	= 1LLU << 0,	/* pick a CPU whose SMT siblings are also idle */
> > > > +	SCX_PICK_IDLE_NODE	= 1LLU << 1,	/* pick a CPU in the same target NUMA node */
> > > 
> > > SCX_FORCE_NODE or SCX_FIX_NODE?
> > > 
> > > >  };
> > > >  
> > > >  enum scx_kick_flags {
> > > > diff --git a/kernel/sched/ext_idle.c b/kernel/sched/ext_idle.c
> > > > index 444f2a15f1d4..013deaa08f12 100644
> > > > --- a/kernel/sched/ext_idle.c
> > > > +++ b/kernel/sched/ext_idle.c
> > > > @@ -199,6 +199,12 @@ static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, int node, u64 f
> > 
> > This function begins with:
> > 
> >  static s32 scx_pick_idle_cpu(const struct cpumask *cpus_allowed, int node, u64 flags)
> >  {
> >       nodemask_t hop_nodes = NODE_MASK_NONE;
> >       s32 cpu = -EBUSY;
> >  
> >       if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
> >               return pick_idle_cpu_from_node(cpus_allowed, NUMA_FLAT_NODE, flags);
> > 
> >       ...
> >  
> > So if I disable scx_builtin_idle_per_node and then call:
> > 
> >         scx_pick_idle_cpu(some_cpus, numa_node_id(), SCX_PICK_IDLE_NODE)
> > 
> > I may get a CPU from any non-local node, right? I think we need to honor user's
> > request:  
> > 
> >       if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node))
> >               return pick_idle_cpu_from_node(cpus_allowed,
> >                      flags & SCX_PICK_IDLE_NODE ? node :  NUMA_FLAT_NODE, flags);
> > 
> > That way the code will be coherent: if you enable idle cpumasks, you
> > will be able to follow all the NUMA hierarchy. If you disable them, at
> > least you honor user's request to return a CPU from a given node, if
> > he's very explicit about his intention.
> > 
> > You can be even nicer:
> > 
> >       if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
> >                 node = pick_idle_cpu_from_node(cpus, node, flags);
> >                 if (node == MAX_NUM_NODES && flags & SCX_PICK_IDLE_NODE == 0)
> >                         node = pick_idle_cpu_from_node(cpus, NUMA_FLAT_NODE, flags);
> > 
> >                 return node;
> >       }
> > 
> 
> Sorry, I'm not following, if scx_builtin_idle_per_node is disabled, we’re
> only tracking idle CPUs in a single NUMA_FLAT_NODE (which is node 0). All
> the other cpumasks are just empty, and we would always return -EBUSY if we
> honor the user request.

You're right. We can still do that like this:

       if (!static_branch_maybe(CONFIG_NUMA, &scx_builtin_idle_per_node)) {
                 cpumask_and(tmp, cpus, cpumask_of_node(node));
                 node = pick_idle_cpu_from_node(tmp, NUMA_FLAT_NODE, flags);
                 if (node == MAX_NUM_NODES && flags & SCX_PICK_IDLE_NODE == 0)
                         node = pick_idle_cpu_from_node(cpus, NUMA_FLAT_NODE, flags);
 
                 return node;
       }

But I'm not sure we need this complication. Maybe later...

> 
> Maybe we should just return an error if scx_builtin_idle_per_node is
> disabled and the user is requesting an idle CPU in a specific node?

The problem is that NUMA_FLAT_NODE is 0, and you can't distinguish it
from node #0. You can drop NUMA_FLAT_NODE and ask users to always
provide NUMA_NO_NODE if idle_per_node is disabled, or you can ignore
the node entirely. You just need to describe it explicitly.