Re: [PATCH] LoongArch: Fix cpu hotplug issue

From: maobibo
Date: Mon Oct 14 2024 - 03:21:38 EST


Huacai,

On 2024/10/14 下午3:05, Huacai Chen wrote:
Hi, Bibo,

I'm a little confused, so please correct me if I'm wrong.

On Mon, Oct 14, 2024 at 2:33 PM Bibo Mao <maobibo@xxxxxxxxxxx> wrote:

On LoongArch system, there are two places to set cpu numa node. One
is in arch specified function smp_prepare_boot_cpu(), the other is
in generic function early_numa_node_init(). The latter will overwrite
the numa node information.

However for hot-added cpu, cpu_logical_map() fails to its physical
cpuid at beginning since it is not enabled in ACPI MADT table. So
function early_cpu_to_node() also fails to get its numa node for
hot-added cpu, and generic function early_numa_node_init() will
overwrite incorrect numa node.
For hot-added cpus, we will call acpi_map_cpu() -->
acpi_map_cpu2node() --> set_cpuid_to_node(), and set_cpuid_to_node()
operates on __cpuid_to_node[]. So I think early_cpu_to_node() should
be correct?

__cpuid_to_node[] is correct which is physical cpuid to numa node, however cpu_logical_map(cpu) is not set. It fails to get physical cpuid
from logic cpu.

int early_cpu_to_node(int cpu)
{
int physid = cpu_logical_map(cpu);

<<<<<<<<<<< Here physid is -1.

if (physid < 0)
return NUMA_NO_NODE;

return __cpuid_to_node[physid];
}

Regards
Bibo Mao

Huacai


Here static array __cpu_to_node and api set_early_cpu_to_node()
is added, so that early_cpu_to_node is consistent with function
cpu_to_node() for hot-added cpu.

Signed-off-by: Bibo Mao <maobibo@xxxxxxxxxxx>
---
arch/loongarch/include/asm/numa.h | 2 ++
arch/loongarch/kernel/numa.c | 10 +++++++++-
arch/loongarch/kernel/smp.c | 1 +
3 files changed, 12 insertions(+), 1 deletion(-)

diff --git a/arch/loongarch/include/asm/numa.h b/arch/loongarch/include/asm/numa.h
index b5f9de9f102e..e8e6fcfb006a 100644
--- a/arch/loongarch/include/asm/numa.h
+++ b/arch/loongarch/include/asm/numa.h
@@ -50,6 +50,7 @@ static inline void set_cpuid_to_node(int cpuid, s16 node)
}

extern int early_cpu_to_node(int cpu);
+extern void set_early_cpu_to_node(int cpu, s16 node);

#else

@@ -57,6 +58,7 @@ static inline void early_numa_add_cpu(int cpuid, s16 node) { }
static inline void numa_add_cpu(unsigned int cpu) { }
static inline void numa_remove_cpu(unsigned int cpu) { }
static inline void set_cpuid_to_node(int cpuid, s16 node) { }
+static inline void set_early_cpu_to_node(int cpu, s16 node) { }

static inline int early_cpu_to_node(int cpu)
{
diff --git a/arch/loongarch/kernel/numa.c b/arch/loongarch/kernel/numa.c
index 84fe7f854820..62508aace644 100644
--- a/arch/loongarch/kernel/numa.c
+++ b/arch/loongarch/kernel/numa.c
@@ -34,6 +34,9 @@ static struct numa_meminfo numa_meminfo;
cpumask_t cpus_on_node[MAX_NUMNODES];
cpumask_t phys_cpus_on_node[MAX_NUMNODES];
EXPORT_SYMBOL(cpus_on_node);
+static s16 __cpu_to_node[NR_CPUS] = {
+ [0 ... CONFIG_NR_CPUS - 1] = NUMA_NO_NODE
+};

/*
* apicid, cpu, node mappings
@@ -117,11 +120,16 @@ int early_cpu_to_node(int cpu)
int physid = cpu_logical_map(cpu);

if (physid < 0)
- return NUMA_NO_NODE;
+ return __cpu_to_node[cpu];

return __cpuid_to_node[physid];
}

+void set_early_cpu_to_node(int cpu, s16 node)
+{
+ __cpu_to_node[cpu] = node;
+}
+
void __init early_numa_add_cpu(int cpuid, s16 node)
{
int cpu = __cpu_number_map[cpuid];
diff --git a/arch/loongarch/kernel/smp.c b/arch/loongarch/kernel/smp.c
index 9afc2d8b3414..998668be858c 100644
--- a/arch/loongarch/kernel/smp.c
+++ b/arch/loongarch/kernel/smp.c
@@ -512,6 +512,7 @@ void __init smp_prepare_boot_cpu(void)
set_cpu_numa_node(cpu, node);
else {
set_cpu_numa_node(cpu, rr_node);
+ set_early_cpu_to_node(cpu, rr_node);
rr_node = next_node_in(rr_node, node_online_map);
}
}

base-commit: 6485cf5ea253d40d507cd71253c9568c5470cd27
--
2.39.3