Skip to contentSkip to conversation

x86/numa: Initialize __apicid_to_node in dummy_numa_init() to prevent OOB

PATCH Superseded Mouse Zhang 2 participants 3 messages Jan 16 – Jan 23

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/numa: Initialize __apicid_to_node in dummy_numa_init() to prevent OOB

When NUMA is disabled (numa=off) or the BIOS does not provide an SRAT table, the kernel falls back to dummy_numa_init(). This function fakes a single node (Node 0) and maps all memory to it, but it leaves the __apicid_to_node[] mapping table uninitialized (filled with NUMA_NO_NODE).

This leads to a potential out-of-bounds access in srat_detect_node() and other topology-related code. Specifically, when numa_cpu_node() returns NUMA_NO_NODE, some code paths attempt to use cpu_llc_id as a fallback for the node ID. On modern systems with large APIC IDs, the cpu_llc_id (derived from APIC ID) can exceed MAX_NUMNODES. Using this invalid ID in functions like node_online(node) causes memory corruption or kernel panic.

Fix this by explicitly mapping all unassigned APIC IDs to Node 0 in dummy_numa_init(). This ensures that numa_cpu_node() consistently returns Node 0 in non-NUMA environments, avoiding dangerous fallbacks and keeping the mapping consistent with the fake Node 0.

---
 arch/x86/mm/numa.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/mm/numa.c +7 -0
index 7a97327140df..673815d95974 100644
@@ -212,6 +212,13 @@ static int __init dummy_numa_init(void)
 	node_set(0, numa_nodes_parsed);
 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
 
+	/* Map all unassociated APIC IDs to the fake node 0 */
+	unsigned int apicid;
+	for (apicid = 0; apicid < MAX_LOCAL_APIC; apicid++) {
+		if (__apicid_to_node[apicid] == NUMA_NO_NODE)
+			 __apicid_to_node[apicid] = 0;
+        }
+
 	return 0;
 }
 
Signature · 1 line
2.51.0
#1 lore raw

PATCH v2 1/1 x86/numa: Initialize __apicid_to_node in dummy_numa_init() to prevent OOB

When NUMA is disabled (numa=off) or the BIOS does not provide an SRAT table, the kernel falls back to dummy_numa_init(). This function fakes a single node (Node 0) and maps all memory to it, but it leaves the __apicid_to_node[] mapping table uninitialized (filled with NUMA_NO_NODE).

This leads to a potential out-of-bounds access in srat_detect_node() and other topology-related code. Specifically, when numa_cpu_node() returns NUMA_NO_NODE, some code paths attempt to use cpu_llc_id as a fallback for the node ID. On modern systems with large APIC IDs, the cpu_llc_id (derived from APIC ID) can exceed MAX_NUMNODES. Using this invalid ID in functions like node_online(node) causes memory corruption or kernel panic.

Fix this by explicitly mapping all unassigned APIC IDs to Node 0 in dummy_numa_init(). This ensures that numa_cpu_node() consistently returns Node 0 in non-NUMA environments, avoiding dangerous fallbacks and keeping the mapping consistent with the fake Node 0.

---
Fix bad format.

 arch/x86/mm/numa.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/mm/numa.c +8 -0
index 7a97327140df..f78b3ff7a67e 100644
@@ -212,6 +212,14 @@ static int __init dummy_numa_init(void)
 	node_set(0, numa_nodes_parsed);
 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
 
+	/* Map all unassociated APIC IDs to the fake node 0 */
+	unsigned int apicid;
+
+	for (apicid = 0; apicid < MAX_LOCAL_APIC; apicid++) {
+		if (__apicid_to_node[apicid] == NUMA_NO_NODE)
+			__apicid_to_node[apicid] = 0;
+	}
+
 	return 0;
 }
 
Signature · 1 line
2.51.0
#2 lore raw

Hi Mouse Zhang,

Quoted 5 lines · Mouse Zhang

When NUMA is disabled (numa=off) or the BIOS does not provide an SRAT table, the kernel falls back to dummy_numa_init(). This function fakes a single node (Node 0) and maps all memory to it, but it leaves the __apicid_to_node[] mapping table uninitialized (filled with NUMA_NO_NODE).

I agree about setting '__apicid_to_node[apicid] = 0' in dummy_numa_init(). Both x86_acpi_numa_init() and amd_numa_init() initialize __apicid_to_node[]. I think dummy_numa_init() should initialize __apicid_to_node[] too.

Quoted 24 lines

This leads to a potential out-of-bounds access in srat_detect_node() and other topology-related code. Specifically, when numa_cpu_node() returns NUMA_NO_NODE, some code paths attempt to use cpu_llc_id as a fallback for the node ID. On modern systems with large APIC IDs, the cpu_llc_id (derived from APIC ID) can exceed MAX_NUMNODES. Using this invalid ID in functions like node_online(node) causes memory corruption or kernel panic.

Fix this by explicitly mapping all unassigned APIC IDs to Node 0 in dummy_numa_init(). This ensures that numa_cpu_node() consistently returns Node 0 in non-NUMA environments, avoiding dangerous fallbacks and keeping the mapping consistent with the fake Node 0.

---
Fix bad format.

  arch/x86/mm/numa.c | 8 ++++++++
  1 file changed, 8 insertions(+)
arch/x86/mm/numa.c +0 -0
index 7a97327140df..f78b3ff7a67e 100644

arch/x86/mm/numa.c static int __init dummy_numa_init(void) · line 219

Quoted 8 lines
arch/x86/mm/numa.c +5 -0
@@ -212,6 +212,14 @@ static int __init dummy_numa_init(void)
  	node_set(0, numa_nodes_parsed);
  	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
 
+	/* Map all unassociated APIC IDs to the fake node 0 */
+	unsigned int apicid;
+
+	for (apicid = 0; apicid < MAX_LOCAL_APIC; apicid++) {
+		if (__apicid_to_node[apicid] == NUMA_NO_NODE)

Is numa_cpu_node() better? Do we need to check this conditional at all?

arch/x86/mm/numa.c static int __init dummy_numa_init(void) · line 220

Quoted 1 line
arch/x86/mm/numa.c +1 -0
+			__apicid_to_node[apicid] = 0;

set_apicid_to_node(apicid, 0) looks better. x86_acpi_numa_init() and amd_numa_init() use set_apicid_to_node().

arch/x86/mm/numa.c static int __init dummy_numa_init(void) · line 225

Quoted 5 lines
arch/x86/mm/numa.c +2 -0
+	}
+
  	return 0;
  }
 

Thanks,
-Yi