for_each_node_numadist() can lead to hard lockups on kernels built
without CONFIG_NUMA. For instance, the following was triggered by
sched_ext:

  watchdog: CPU5: Watchdog detected hard LOCKUP on cpu 5
  ...
  RIP: 0010:_find_first_and_bit+0x8/0x60
  ...
  Call Trace:
  <TASK>
   cpumask_any_and_distribute+0x49/0x80
   pick_idle_cpu_in_node+0xcf/0x140
   scx_bpf_pick_idle_cpu_node+0xaa/0x110
   bpf_prog_16ee5b1f077af006_pick_idle_cpu+0x57f/0x5de
   bpf_prog_df2ce5cfac58ce09_bpfland_select_cpu+0x37/0xf4
   bpf__sched_ext_ops_select_cpu+0x4b/0xb3

This happens because nearest_node_nodemask() always returns NUMA_NO_NODE
(-1) when CONFIG_NUMA is disabled, causing the loop to never terminate,
as the condition node >= MAX_NUMNODES is never satisfied.

Prevent this by handling NUMA_NO_NODE explicitly in the exit condition.

Fixes: f09177ca5f242 ("sched/topology: Introduce for_each_node_numadist() iterator")
Signed-off-by: Andrea Righi <arighi@nvidia.com>
---
 include/linux/topology.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/topology.h b/include/linux/topology.h
index cd6b4bdc9cfd3..095cda6dbf041 100644
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -310,7 +310,7 @@ sched_numa_hop_mask(unsigned int node, unsigned int hops)
 #define for_each_node_numadist(node, unvisited)					\
 	for (int __start = (node),						\
 	     (node) = nearest_node_nodemask((__start), &(unvisited));		\
-	     (node) < MAX_NUMNODES;						\
+	     (node) < MAX_NUMNODES && (node) != NUMA_NO_NODE;			\
 	     node_clear((node), (unvisited)),					\
 	     (node) = nearest_node_nodemask((__start), &(unvisited)))
 
-- 
2.49.0

On Tue, Jun 03, 2025 at 10:04:02AM +0200, Andrea Righi wrote:
> for_each_node_numadist() can lead to hard lockups on kernels built
> without CONFIG_NUMA. For instance, the following was triggered by
> sched_ext:
> 
>   watchdog: CPU5: Watchdog detected hard LOCKUP on cpu 5
>   ...
>   RIP: 0010:_find_first_and_bit+0x8/0x60
>   ...
>   Call Trace:
>   <TASK>
>    cpumask_any_and_distribute+0x49/0x80
>    pick_idle_cpu_in_node+0xcf/0x140
>    scx_bpf_pick_idle_cpu_node+0xaa/0x110
>    bpf_prog_16ee5b1f077af006_pick_idle_cpu+0x57f/0x5de
>    bpf_prog_df2ce5cfac58ce09_bpfland_select_cpu+0x37/0xf4
>    bpf__sched_ext_ops_select_cpu+0x4b/0xb3
> 
> This happens because nearest_node_nodemask() always returns NUMA_NO_NODE
> (-1) when CONFIG_NUMA is disabled, causing the loop to never terminate,
> as the condition node >= MAX_NUMNODES is never satisfied.
> 
> Prevent this by handling NUMA_NO_NODE explicitly in the exit condition.
> 
> Fixes: f09177ca5f242 ("sched/topology: Introduce for_each_node_numadist() iterator")
> Signed-off-by: Andrea Righi <arighi@nvidia.com>
> ---
>  include/linux/topology.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/include/linux/topology.h b/include/linux/topology.h
> index cd6b4bdc9cfd3..095cda6dbf041 100644
> --- a/include/linux/topology.h
> +++ b/include/linux/topology.h
> @@ -310,7 +310,7 @@ sched_numa_hop_mask(unsigned int node, unsigned int hops)
>  #define for_each_node_numadist(node, unvisited)					\
>  	for (int __start = (node),						\
>  	     (node) = nearest_node_nodemask((__start), &(unvisited));		\
> -	     (node) < MAX_NUMNODES;						\
> +	     (node) < MAX_NUMNODES && (node) != NUMA_NO_NODE;			\
>  	     node_clear((node), (unvisited)),					\
>  	     (node) = nearest_node_nodemask((__start), &(unvisited)))

When NUMA is enabled, you add an extra conditional which is never the
true.

When disabled, I think this macro should not be available, or more
likely have a stub implementation, similar to for_each_node_mask()

Thanks,
Yury

Hi Yuri,

On Wed, Jun 04, 2025 at 10:13:21AM -0400, Yury Norov wrote:
> On Tue, Jun 03, 2025 at 10:04:02AM +0200, Andrea Righi wrote:
> > for_each_node_numadist() can lead to hard lockups on kernels built
> > without CONFIG_NUMA. For instance, the following was triggered by
> > sched_ext:
> > 
> >   watchdog: CPU5: Watchdog detected hard LOCKUP on cpu 5
> >   ...
> >   RIP: 0010:_find_first_and_bit+0x8/0x60
> >   ...
> >   Call Trace:
> >   <TASK>
> >    cpumask_any_and_distribute+0x49/0x80
> >    pick_idle_cpu_in_node+0xcf/0x140
> >    scx_bpf_pick_idle_cpu_node+0xaa/0x110
> >    bpf_prog_16ee5b1f077af006_pick_idle_cpu+0x57f/0x5de
> >    bpf_prog_df2ce5cfac58ce09_bpfland_select_cpu+0x37/0xf4
> >    bpf__sched_ext_ops_select_cpu+0x4b/0xb3
> > 
> > This happens because nearest_node_nodemask() always returns NUMA_NO_NODE
> > (-1) when CONFIG_NUMA is disabled, causing the loop to never terminate,
> > as the condition node >= MAX_NUMNODES is never satisfied.
> > 
> > Prevent this by handling NUMA_NO_NODE explicitly in the exit condition.
> > 
> > Fixes: f09177ca5f242 ("sched/topology: Introduce for_each_node_numadist() iterator")
> > Signed-off-by: Andrea Righi <arighi@nvidia.com>
> > ---
> >  include/linux/topology.h | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/include/linux/topology.h b/include/linux/topology.h
> > index cd6b4bdc9cfd3..095cda6dbf041 100644
> > --- a/include/linux/topology.h
> > +++ b/include/linux/topology.h
> > @@ -310,7 +310,7 @@ sched_numa_hop_mask(unsigned int node, unsigned int hops)
> >  #define for_each_node_numadist(node, unvisited)					\
> >  	for (int __start = (node),						\
> >  	     (node) = nearest_node_nodemask((__start), &(unvisited));		\
> > -	     (node) < MAX_NUMNODES;						\
> > +	     (node) < MAX_NUMNODES && (node) != NUMA_NO_NODE;			\
> >  	     node_clear((node), (unvisited)),					\
> >  	     (node) = nearest_node_nodemask((__start), &(unvisited)))
> 
> When NUMA is enabled, you add an extra conditional which is never the
> true.
> 
> When disabled, I think this macro should not be available, or more
> likely have a stub implementation, similar to for_each_node_mask()

Makes sense, I like the idea of having a stub implementation, I'll send a
v2 with that.

Thanks!
-Andrea