[PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node

[PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node

[Feng Tang]

'reuseport_bpf_numa' case failed on a platform with CXL memory:

	#./reuseport_bpf_numa
	---- IPv4 UDP ----
	send node 0, receive socket 0
	./reuseport_bpf_numa: failed to pin to node: Invalid argument

The root cause is that platform has 2 numa nodes: node 0 has both
cpu and memory, while node 1 is a CXL node which only has memory,
which caused numa_run_on_node() failure.

Add check to skip cpuless numa node for the numa binding test.

Signed-off-by: Feng Tang <feng.tang@linux.alibaba.com>
---
 .../selftests/net/reuseport_bpf_numa.c        | 31 +++++++++++++++++--
 1 file changed, 28 insertions(+), 3 deletions(-)

diff --git a/tools/testing/selftests/net/reuseport_bpf_numa.c b/tools/testing/selftests/net/reuseport_bpf_numa.c
index 8ec52fc5ef41..197dbfc475f6 100644
--- a/tools/testing/selftests/net/reuseport_bpf_numa.c
+++ b/tools/testing/selftests/net/reuseport_bpf_numa.c
@@ -104,13 +104,35 @@ static void attach_bpf(int fd)
 	close(bpf_fd);
 }
 
-static void send_from_node(int node_id, int family, int proto)
+static inline bool is_cpuless_node(int node_id)
+{
+	struct bitmask *cpumask;
+	bool ret = false;
+
+	cpumask = numa_allocate_cpumask();
+
+	numa_node_to_cpus(node_id, cpumask);
+	if (!numa_bitmask_weight(cpumask))
+		ret = true;
+
+	numa_bitmask_free(cpumask);
+	return ret;
+}
+
+/*
+ * return -1 on node_id pointing to a cpuless node (like a CXL node),
+ * return 0 on successful numa setup
+ */
+static int send_from_node(int node_id, int family, int proto)
 {
 	struct sockaddr_storage saddr, daddr;
 	struct sockaddr_in  *saddr4, *daddr4;
 	struct sockaddr_in6 *saddr6, *daddr6;
 	int fd;
 
+	if (is_cpuless_node(node_id))
+		return -1;
+
 	switch (family) {
 	case AF_INET:
 		saddr4 = (struct sockaddr_in *)&saddr;
@@ -155,6 +177,7 @@ static void send_from_node(int node_id, int family, int proto)
 		error(1, errno, "failed to send message");
 
 	close(fd);
+	return 0;
 }
 
 static
@@ -213,7 +236,8 @@ static void test(int *rcv_fd, int len, int family, int proto)
 	for (node = 0; node < len; ++node) {
 		if (!numa_bitmask_isbitset(numa_nodes_ptr, node))
 			continue;
-		send_from_node(node, family, proto);
+		if (send_from_node(node, family, proto))
+			continue;
 		receive_on_node(rcv_fd, len, epfd, node, proto);
 	}
 
@@ -221,7 +245,8 @@ static void test(int *rcv_fd, int len, int family, int proto)
 	for (node = len - 1; node >= 0; --node) {
 		if (!numa_bitmask_isbitset(numa_nodes_ptr, node))
 			continue;
-		send_from_node(node, family, proto);
+		if (send_from_node(node, family, proto))
+			continue;
 		receive_on_node(rcv_fd, len, epfd, node, proto);
 	}
 
-- 
2.39.5 (Apple Git-154)

Re: [PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [Medium] Missing NULL check for cpumask allocation in is_cpuless_node()
- [Medium] Missing return value check for numa_node_to_cpus() leading to masked errors
--

--- Patch [1]: [PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node ---
commit 806931cbf2ad9438843e4d30f69b4887e641f7d2
Author: Feng Tang <feng.tang@linux.alibaba.com>

selftests: net: reuseport_bpf_numa: consider cpuless numa node

This patch addresses a failure in the 'reuseport_bpf_numa' selftest on
platforms with CXL memory where one node is cpuless. It introduces an
is_cpuless_node() check to skip memory-only NUMA nodes during the binding
test.

> diff --git a/tools/testing/selftests/net/reuseport_bpf_numa.c b/tools/testing/selftests/net/reuseport_bpf_numa.c
> index 8ec52fc5ef41b..197dbfc475f65 100644
> --- a/tools/testing/selftests/net/reuseport_bpf_numa.c
> +++ b/tools/testing/selftests/net/reuseport_bpf_numa.c
[ ... ]
> @@ -104,13 +104,35 @@ static void attach_bpf(int fd)
>  	close(bpf_fd);
>  }
>  
> -static void send_from_node(int node_id, int family, int proto)
> +static inline bool is_cpuless_node(int node_id)
> +{
> +	struct bitmask *cpumask;
> +	bool ret = false;
> +
> +	cpumask = numa_allocate_cpumask();
> +
> +	numa_node_to_cpus(node_id, cpumask);

[Severity: Medium]
Will this crash if numa_allocate_cpumask() fails to allocate memory?

It looks like cpumask could be NULL here, which is passed directly to
numa_node_to_cpus() and numa_bitmask_weight() without validation.

> +	if (!numa_bitmask_weight(cpumask))
> +		ret = true;

[Severity: Medium]
What happens if numa_node_to_cpus() fails?

It can return a negative error value, leaving the cpumask unmodified with
zero bits set. Because the return value is not checked,
numa_bitmask_weight(cpumask) will evaluate to 0. Does this cause the function
to return true and silently skip testing on nodes where an error occurs,
rather than reporting the infrastructure issue?

> +
> +	numa_bitmask_free(cpumask);
> +	return ret;
> +}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260730113040.75718-1-feng.tang@linux.alibaba.com?part=1

Re: [PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node

[Feng Tang]

Thanks for the review!

Add Andi

On Thu, Jul 30, 2026 at 11:37:37AM +0000, sashiko-bot@kernel.org wrote:
> Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
> - [Medium] Missing NULL check for cpumask allocation in is_cpuless_node()
> - [Medium] Missing return value check for numa_node_to_cpus() leading to masked errors
> --
> 
> --- Patch [1]: [PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node ---
> commit 806931cbf2ad9438843e4d30f69b4887e641f7d2
> Author: Feng Tang <feng.tang@linux.alibaba.com>
> 
> selftests: net: reuseport_bpf_numa: consider cpuless numa node
> 
> This patch addresses a failure in the 'reuseport_bpf_numa' selftest on
> platforms with CXL memory where one node is cpuless. It introduces an
> is_cpuless_node() check to skip memory-only NUMA nodes during the binding
> test.
> 
> > diff --git a/tools/testing/selftests/net/reuseport_bpf_numa.c b/tools/testing/selftests/net/reuseport_bpf_numa.c
> > index 8ec52fc5ef41b..197dbfc475f65 100644
> > --- a/tools/testing/selftests/net/reuseport_bpf_numa.c
> > +++ b/tools/testing/selftests/net/reuseport_bpf_numa.c
> [ ... ]
> > @@ -104,13 +104,35 @@ static void attach_bpf(int fd)
> >  	close(bpf_fd);
> >  }
> >  
> > -static void send_from_node(int node_id, int family, int proto)
> > +static inline bool is_cpuless_node(int node_id)
> > +{
> > +	struct bitmask *cpumask;
> > +	bool ret = false;
> > +
> > +	cpumask = numa_allocate_cpumask();
> > +
> > +	numa_node_to_cpus(node_id, cpumask);
> 
> [Severity: Medium]
> Will this crash if numa_allocate_cpumask() fails to allocate memory?
> 
> It looks like cpumask could be NULL here, which is passed directly to
> numa_node_to_cpus() and numa_bitmask_weight() without validation.
 
Good point. 

If the allocation fail, will return 'false' to let the code go the
original flow.


> > +	if (!numa_bitmask_weight(cpumask))
> > +		ret = true;
> 
> [Severity: Medium]
> What happens if numa_node_to_cpus() fails?
> 
> It can return a negative error value, leaving the cpumask unmodified with
> zero bits set. Because the return value is not checked,
> numa_bitmask_weight(cpumask) will evaluate to 0. Does this cause the function
> to return true and silently skip testing on nodes where an error occurs,
> rather than reporting the infrastructure issue?

From the code logic (with above change), the 'node_id' and 'cpumask' have
both been validated, and this numa_node_to_cpus() is very unlikely to fail
here upon reading libnuma source code.

Thanks,
Feng

> > +
> > +	numa_bitmask_free(cpumask);
> > +	return ret;
> > +}
> 
> -- 
> Sashiko AI review · https://sashiko.dev/#/patchset/20260730113040.75718-1-feng.tang@linux.alibaba.com?part=1

Re: [PATCH] selftests: net: reuseport_bpf_numa: consider cpuless numa node

[Andi Kleen]

> > Will this crash if numa_allocate_cpumask() fails to allocate memory?
> > 
> > It looks like cpumask could be NULL here, which is passed directly to
> > numa_node_to_cpus() and numa_bitmask_weight() without validation.
>  
> Good point. 
> 
> If the allocation fail, will return 'false' to let the code go the
> original flow.


While it's technically true, checking for out of memory in ring 3
(or at least other checking than exit(1) on error) is usually 
useless. Most of the time when this would fail you're as likely
to just get killed by the OOM killer, as well as the harness that
called you. Maybe it could happen without that with some very tight
ulimit settings, but with that not much else will work anyways.
Also these paths are basically never exercised so likely buggy
when non trivial.

For libnuma if you set numa_exit_on_error to 1 it will always
exiting after printing an error. That includes the allocation
functions.



> > > +		ret = true;
> > 
> > [Severity: Medium]
> > What happens if numa_node_to_cpus() fails?
> > 
> > It can return a negative error value, leaving the cpumask unmodified with
> > zero bits set. Because the return value is not checked,
> > numa_bitmask_weight(cpumask) will evaluate to 0. Does this cause the function
> > to return true and silently skip testing on nodes where an error occurs,
> > rather than reporting the infrastructure issue?
> 
> From the code logic (with above change), the 'node_id' and 'cpumask' have
> both been validated, and this numa_node_to_cpus() is very unlikely to fail
> here upon reading libnuma source code.

Yes it's unlikely to fail.

-Andi