Skip to contentSkip to conversation

mm: fix NULL NODE_DATA dereference for memoryless nodes on boot

PATCH Superseded Ming Lei 2 participants 3 messages Feb 22

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Ming Lei

author
#0 lore raw

PATCH 1/1 mm: fix NULL NODE_DATA dereference for memoryless nodes on boot

Commit d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map") moved free_area_init() from setup_arch() to mm_core_init_early(), which runs after setup_arch() returns.

This changed the ordering relative to init_cpu_to_node() on x86. Before the commit, free_area_init() ran during paging_init() (called from setup_arch()) *before* init_cpu_to_node(). After the commit, it runs *after* init_cpu_to_node().

On machines with memoryless NUMA nodes (e.g., node 0 has CPUs but no memory), this causes a NULL pointer dereference:

1. numa_register_nodes() skips memoryless nodes: no alloc_node_data() and no node_set_online() for them.
2. init_cpu_to_node() sets memoryless nodes online (they have CPUs) but does not allocate NODE_DATA.
3. free_area_init() checks "if (!node_online(nid))" to decide whether to call alloc_offline_node_data(). Since the memoryless node is now online, the allocation is skipped, leaving NODE_DATA(nid) == NULL.
4. The immediate "pgdat = NODE_DATA(nid)" dereferences NULL.

The crash happens before console_init(), so no output is visible without earlyprintk. With earlyprintk enabled, the following panic is observed:

 BUG: unable to handle page fault for address: 000000000002a1e0
 Oops: Oops: 0000 [#1] SMP NOPTI
 RIP: 0010:free_area_init_node+0x3a/0x540
 Call Trace:
  <TASK>
  free_area_init+0x331/0x4e0
  start_kernel+0x69/0x4a0
  x86_64_start_reservations+0x24/0x30
  x86_64_start_kernel+0x125/0x130
  common_startup_64+0x13e/0x148
  </TASK>
 Kernel panic - not syncing: Attempted to kill the idle task!

Fix this by checking "if (!NODE_DATA(nid))" instead of
"if (!node_online(nid))". This directly tests whether the per-node data structure needs to be allocated, regardless of the node's online status.

Cc Mike Rapoport (Microsoft) <[email protected]>
Fixes d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map")
---
 mm/mm_init.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/mm_init.c +1 -1
index 61d983d23f55..9d63cab36204 100644
@@ -1896,7 +1896,7 @@ static void __init free_area_init(void)
 	for_each_node(nid) {
 		pg_data_t *pgdat;
 
-		if (!node_online(nid))
+		if (!NODE_DATA(nid))
 			alloc_offline_node_data(nid);
 
 		pgdat = NODE_DATA(nid);
Signature · 1 line
2.52.0
maintainer
#1 lore raw

Hi,

Quoted 37 lines · Ming Lei

Commit d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map") moved free_area_init() from setup_arch() to mm_core_init_early(), which runs after setup_arch() returns.

This changed the ordering relative to init_cpu_to_node() on x86. Before the commit, free_area_init() ran during paging_init() (called from setup_arch()) *before* init_cpu_to_node(). After the commit, it runs *after* init_cpu_to_node().

On machines with memoryless NUMA nodes (e.g., node 0 has CPUs but no memory), this causes a NULL pointer dereference:

1. numa_register_nodes() skips memoryless nodes: no alloc_node_data() and no node_set_online() for them.
2. init_cpu_to_node() sets memoryless nodes online (they have CPUs) but does not allocate NODE_DATA.
3. free_area_init() checks "if (!node_online(nid))" to decide whether to call alloc_offline_node_data(). Since the memoryless node is now online, the allocation is skipped, leaving NODE_DATA(nid) == NULL.
4. The immediate "pgdat = NODE_DATA(nid)" dereferences NULL.

The crash happens before console_init(), so no output is visible without earlyprintk. With earlyprintk enabled, the following panic is observed:

 BUG: unable to handle page fault for address: 000000000002a1e0
 Oops: Oops: 0000 [#1] SMP NOPTI
 RIP: 0010:free_area_init_node+0x3a/0x540
 Call Trace:
  <TASK>
  free_area_init+0x331/0x4e0
  start_kernel+0x69/0x4a0
  x86_64_start_reservations+0x24/0x30
  x86_64_start_kernel+0x125/0x130
  common_startup_64+0x13e/0x148
  </TASK>
 Kernel panic - not syncing: Attempted to kill the idle task!
Quoted 3 lines

Fix this by checking "if (!NODE_DATA(nid))" instead of
"if (!node_online(nid))". This directly tests whether the per-node data structure needs to be allocated, regardless of the node's online status.

I believe that this change is fine for !x86 as well, but it deserves a sentence in the commit log.

Quoted 12 lines
Cc Mike Rapoport (Microsoft) <[email protected]>
Fixes d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map")
---
 mm/mm_init.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/mm_init.c +0 -0
index 61d983d23f55..9d63cab36204 100644

mm/mm_init.c static void __init free_area_init(void) · line 1900

Quoted 6 lines
mm/mm_init.c +1 -1
@@ -1896,7 +1896,7 @@ static void __init free_area_init(void)
 	for_each_node(nid) {
 		pg_data_t *pgdat;
 
-		if (!node_online(nid))
+		if (!NODE_DATA(nid))
 			alloc_offline_node_data(nid);

A comment that says that if an architecture didn't allocate node data, we presume that the node is memoryless and offline would be nice here.

Quoted 5 lines
mm/mm_init.c +0 -0
 		pgdat = NODE_DATA(nid);
Signature · 1 line
2.52.0
Signature · 2 lines
Sincerely yours,
Mike.
author
#2 lore raw
Quoted 62 lines · Mike Rapoport

Hi,

Quoted 37 lines · Ming Lei

Commit d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map") moved free_area_init() from setup_arch() to mm_core_init_early(), which runs after setup_arch() returns.

This changed the ordering relative to init_cpu_to_node() on x86. Before the commit, free_area_init() ran during paging_init() (called from setup_arch()) *before* init_cpu_to_node(). After the commit, it runs *after* init_cpu_to_node().

On machines with memoryless NUMA nodes (e.g., node 0 has CPUs but no memory), this causes a NULL pointer dereference:

1. numa_register_nodes() skips memoryless nodes: no alloc_node_data() and no node_set_online() for them.
2. init_cpu_to_node() sets memoryless nodes online (they have CPUs) but does not allocate NODE_DATA.
3. free_area_init() checks "if (!node_online(nid))" to decide whether to call alloc_offline_node_data(). Since the memoryless node is now online, the allocation is skipped, leaving NODE_DATA(nid) == NULL.
4. The immediate "pgdat = NODE_DATA(nid)" dereferences NULL.

The crash happens before console_init(), so no output is visible without earlyprintk. With earlyprintk enabled, the following panic is observed:

 BUG: unable to handle page fault for address: 000000000002a1e0
 Oops: Oops: 0000 [#1] SMP NOPTI
 RIP: 0010:free_area_init_node+0x3a/0x540
 Call Trace:
  <TASK>
  free_area_init+0x331/0x4e0
  start_kernel+0x69/0x4a0
  x86_64_start_reservations+0x24/0x30
  x86_64_start_kernel+0x125/0x130
  common_startup_64+0x13e/0x148
  </TASK>
 Kernel panic - not syncing: Attempted to kill the idle task!
Quoted 3 lines

Fix this by checking "if (!NODE_DATA(nid))" instead of
"if (!node_online(nid))". This directly tests whether the per-node data structure needs to be allocated, regardless of the node's online status.

I believe that this change is fine for !x86 as well, but it deserves a sentence in the commit log.

Quoted 15 lines
Cc Mike Rapoport (Microsoft) <[email protected]>
Fixes d49004c5f0c1 ("arch, mm: consolidate initialization of nodes, zones and memory map")
---
 mm/mm_init.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/mm_init.c +0 -0
index 61d983d23f55..9d63cab36204 100644
@@ -1896,7 +1896,7 @@ static void __init free_area_init(void)
 	for_each_node(nid) {
 		pg_data_t *pgdat;
 
Quoted 6 lines

mm/mm_init.c static void __init free_area_init(void) · line 1900

Quoted 3 lines
mm/mm_init.c +1 -1
-		if (!node_online(nid))
+		if (!NODE_DATA(nid))
 			alloc_offline_node_data(nid);

A comment that says that if an architecture didn't allocate node data, we presume that the node is memoryless and offline would be nice here.

Hi Mike,

All are addressed in V2:

https://lore.kernel.org/linux-mm/[email protected]/

But miss to Cc you, sorry...

Thanks,
Ming