When the number of CPUs is fewer than the number of memory nodes,
some memory nodes may not be properly initialized because they are
not added to numa_nodes_parsed during memory parsing.

In of_numa_parse_memory_nodes(), after successfully adding a memory
block via numa_add_memblk(), the corresponding node ID should be
marked as parsed. However, the current implementation in numa_add_memblk()
only adds the memory block to numa_meminfo but fails to update
numa_nodes_parsed, leaving some nodes uninitialized.

During boot in a QEMU-emulated ARM64 NUMA environment, the kernel
panics when free_area_init() attempts to access NODE_DATA() for
memory nodes that were uninitialized.

[    0.000000] Call trace:
[    0.000000]  free_area_init+0x620/0x106c (P)
[    0.000000]  bootmem_init+0x110/0x1dc
[    0.000000]  setup_arch+0x278/0x60c
[    0.000000]  start_kernel+0x70/0x748
[    0.000000]  __primary_switched+0x88/0x90

Cc: stable@vger.kernel.org
Fixes: 87482708210f ("mm: introduce numa_memblks")
Signed-off-by: Yin Tirui <yintirui@huawei.com>
---
 mm/numa_memblks.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/mm/numa_memblks.c b/mm/numa_memblks.c
index 541a99c4071a..1dfe434e71b5 100644
--- a/mm/numa_memblks.c
+++ b/mm/numa_memblks.c
@@ -197,7 +197,13 @@ static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
  */
 int __init numa_add_memblk(int nid, u64 start, u64 end)
 {
-	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
+	int ret;
+
+	ret = numa_add_memblk_to(nid, start, end, &numa_meminfo);
+	if (!ret)
+		node_set(nid, numa_nodes_parsed);
+
+	return ret;
 }
 
 /**
-- 
2.43.0

On 2025/8/15 18:30, Yin Tirui wrote:
> When the number of CPUs is fewer than the number of memory nodes,
> some memory nodes may not be properly initialized because they are
> not added to numa_nodes_parsed during memory parsing.
> 
> In of_numa_parse_memory_nodes(), after successfully adding a memory
> block via numa_add_memblk(), the corresponding node ID should be
> marked as parsed. However, the current implementation in numa_add_memblk()
> only adds the memory block to numa_meminfo but fails to update
> numa_nodes_parsed, leaving some nodes uninitialized.

The issue is not existed in ACPI NUMA,

     acpi_parse_memory_affinity
       numa_add_memblk
       node_set(node, numa_nodes_parsed);

so we only need to fix of_numa_parse_memory_nodes(),
> 
> During boot in a QEMU-emulated ARM64 NUMA environment, the kernel
> panics when free_area_init() attempts to access NODE_DATA() for
> memory nodes that were uninitialized.
> 
> [    0.000000] Call trace:
> [    0.000000]  free_area_init+0x620/0x106c (P)
> [    0.000000]  bootmem_init+0x110/0x1dc
> [    0.000000]  setup_arch+0x278/0x60c
> [    0.000000]  start_kernel+0x70/0x748
> [    0.000000]  __primary_switched+0x88/0x90
> 
> Cc: stable@vger.kernel.org
> Fixes: 87482708210f ("mm: introduce numa_memblks")

I think the Fixes tag should be commit 767507654c22
（"arch_numa: switch over to numa_memblks").


> Signed-off-by: Yin Tirui <yintirui@huawei.com>
> ---
>   mm/numa_memblks.c | 8 +++++++-
>   1 file changed, 7 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/numa_memblks.c b/mm/numa_memblks.c
> index 541a99c4071a..1dfe434e71b5 100644
> --- a/mm/numa_memblks.c
> +++ b/mm/numa_memblks.c
> @@ -197,7 +197,13 @@ static void __init numa_move_tail_memblk(struct numa_meminfo *dst, int idx,
>    */
>   int __init numa_add_memblk(int nid, u64 start, u64 end)
>   {
> -	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
> +	int ret;
> +
> +	ret = numa_add_memblk_to(nid, start, end, &numa_meminfo);
> +	if (!ret)
> +		node_set(nid, numa_nodes_parsed);
> +
> +	return ret;
>   }
>   
>   /**