Re:Re: [BUG] x86/smpboot: WARN_ON in set_cpu_sibling_map triggered by numa=fake=2

[w15303746062]

From: Mingyu Wang <25181214217@xxxxxxxxxxxxxxxxx>

Hi,

First, apologies for missing LKML in the initial report. I have added it to the Cc list for this reply.

Thank you for the patch.

I have applied it and run multiple test cycles in my fuzzing environment using the `numa=fake=2` parameter. I can confirm that the patch completely resolves the original boot panic. The system now boots reliably every time. 

Additionally, I carefully checked the logs and did not observe any new warnings being triggered, including the new `WARN_ON_ONCE(x86_has_fake_numa)` you added in `arch_sched_node_distance()`.

Please feel free to add:

Tested-by: Mingyu Wang <25181214217@xxxxxxxxxxxxxxxxx>

Thanks again for the quick and elegant fix!

Best regards,
Mingyu


At 2026-05-04 15:48:30, "Peter Zijlstra" <peterz@xxxxxxxxxxxxx> wrote:
>
>You forgot to Cc lkml.
>
>On Wed, Apr 29, 2026 at 11:41:25AM +0800, w15303746062@xxxxxxx wrote:
>> Hi x86 maintainers,
>> 
>> While fuzzing the v7.0 kernel, I encountered a persistent WARNING in
>> `set_cpu_sibling_map()` during boot when using the `numa=fake=2`
>> command-line parameter.
>> 
>> The issue appears to be a logic gap in the topology consistency check.
>> When `numa=fake=N` is used, it artificially divides a single physical
>> package into multiple software NUMA nodes. However, the existing check
>> in `set_cpu_sibling_map()` does not account for this fake NUMA state:
>> 
>>     if (match_pkg(c, o) && !topology_same_node(c, o))
>>         WARN_ON_ONCE(topology_num_nodes_per_package() == 1);
>> 
>> Since `numa=fake` forces `!topology_same_node(c, o)` to be true for
>> CPUs on the same package, the WARN_ON_ONCE is falsely triggered. With
>> `panic_on_warn=1` enabled in many fuzzing and testing environments,
>> this leads to an early boot panic.
>
>*groan*, so ideally we'd just rip out the whole fake numa stuff
>entirely, and let people use VMs to fake topology in a consistent
>manner.
>
>I'm assuming something like so helps?
>
>diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
>index 294a8ea60298..2447b0317f7b 100644
>--- a/arch/x86/kernel/smpboot.c
>+++ b/arch/x86/kernel/smpboot.c
>@@ -468,6 +468,8 @@ static int x86_cluster_flags(void)
> }
> #endif
> 
>+static bool x86_has_fake_numa = false;
>+
> static struct sched_domain_topology_level x86_topology[] = {
> 	SDTL_INIT(tl_smt_mask, cpu_smt_flags, SMT),
> #ifdef CONFIG_SCHED_CLUSTER
>@@ -489,7 +491,7 @@ static void __init build_sched_topology(void)
> 	 * PKG domain since the NUMA domains will auto-magically create the
> 	 * right spanning domains based on the SLIT.
> 	 */
>-	if (topology_num_nodes_per_package() > 1) {
>+	if (topology_num_nodes_per_package() > 1 || x86_has_fake_numa) {
> 		unsigned int pkgdom = ARRAY_SIZE(x86_topology) - 2;
> 
> 		memset(&x86_topology[pkgdom], 0, sizeof(x86_topology[pkgdom]));
>@@ -662,6 +664,13 @@ int arch_sched_node_distance(int from, int to)
> 		    topology_num_nodes_per_package() < 3)
> 			return d;
> 
>+		/*
>+		 * XXX someone needs to figure out what, if anything can be
>+		 * done here.
>+		 */
>+		if (WARN_ON_ONCE(x86_has_fake_numa))
>+			return d;
>+
> 		/*
> 		 * Handle SNC-3 asymmetries.
> 		 */
>@@ -694,8 +703,10 @@ void set_cpu_sibling_map(int cpu)
> 	for_each_cpu(i, cpu_sibling_setup_mask) {
> 		o = &cpu_data(i);
> 
>-		if (match_pkg(c, o) && !topology_same_node(c, o))
>-			WARN_ON_ONCE(topology_num_nodes_per_package() == 1);
>+		if (match_pkg(c, o) && !topology_same_node(c, o)) {
>+			if (topology_num_nodes_per_package() == 1)
>+				x86_has_fake_numa = true;
>+		}
> 
> 		if ((i == cpu) || (has_smt && match_smt(c, o)))
> 			link_mask(topology_sibling_cpumask, cpu, i);