After setting the right NUMA node for VMAP stack,
https://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git/commit/?id=75b5e0bf90bf
I found there are at least four places in the common code where
the cpu_to_node() is called before it is initialized:
0.) early_trace_init() in kernel/trace/trace.c
1.) sched_init() in kernel/sched/core.c
2.) init_sched_fair_class() in kernel/sched/fair.c
3.) workqueue_init_early() in kernel/workqueue.c
We cannot use early_cpu_to_node() for them, since early_cpu_to_node()
does not work for some ARCHs, such as x86, riscv, etc.
So we have to implement the arm64 specific cpu_to_node().
This patch
0.) introduces the __cpu_to_node function pointer,
and exports it for kernel modules.
1.) defines a macro cpu_to_node to override the
generic percpu implementation of cpu_to_node.
2.) __cpu_to_node is initialized with early_cpu_to_node() before
numa_node is initialized.
3.) __cpu_to_node is set to arm64_cpu_to_node() when numa_node is ready.
arm64_cpu_to_node() is a clone of the generic cpu_to_node().
Signed-off-by: Huang Shijie <[email protected]>
---
arch/arm64/include/asm/topology.h | 4 ++++
arch/arm64/kernel/smp.c | 21 ++++++++++++++++++++-
2 files changed, 24 insertions(+), 1 deletion(-)
diff --git a/arch/arm64/include/asm/topology.h b/arch/arm64/include/asm/topology.h
index a323b109b9c4..3e99b40b5f9f 100644
--- a/arch/arm64/include/asm/topology.h
+++ b/arch/arm64/include/asm/topology.h
@@ -12,6 +12,10 @@ int pcibus_to_node(struct pci_bus *bus);
cpu_all_mask : \
cpumask_of_node(pcibus_to_node(bus)))
+/* override generic percpu implementation of cpu_to_node */
+extern int (*__cpu_to_node)(int cpu);
+#define cpu_to_node(cpu) __cpu_to_node(cpu)
+
#endif /* CONFIG_NUMA */
#include <linux/arch_topology.h>
diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
index 4ced34f62dab..8a3bc101eaed 100644
--- a/arch/arm64/kernel/smp.c
+++ b/arch/arm64/kernel/smp.c
@@ -90,6 +90,21 @@ static struct irq_desc *ipi_desc[MAX_IPI] __ro_after_init;
static void ipi_setup(int cpu);
+int (*__cpu_to_node)(int cpu);
+EXPORT_SYMBOL(__cpu_to_node);
+
+#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
+static int arm64_cpu_to_node(int cpu)
+{
+ return per_cpu(numa_node, cpu);
+}
+#else
+static int arm64_cpu_to_node(int cpu)
+{
+ return 0;
+}
+#endif
+
#ifdef CONFIG_HOTPLUG_CPU
static void ipi_teardown(int cpu);
static int op_cpu_kill(unsigned int cpu);
@@ -613,6 +628,7 @@ static void __init acpi_parse_and_init_cpus(void)
for (i = 0; i < nr_cpu_ids; i++)
early_map_cpu_to_node(i, acpi_numa_get_nid(i));
+ __cpu_to_node = early_cpu_to_node;
}
#else
#define acpi_parse_and_init_cpus(...) do { } while (0)
@@ -674,6 +690,7 @@ static void __init of_parse_and_init_cpus(void)
next:
cpu_count++;
}
+ __cpu_to_node = early_cpu_to_node;
}
/*
@@ -733,7 +750,7 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
* secondary CPUs present.
*/
if (max_cpus == 0)
- return;
+ goto out;
/*
* Initialise the present map (which describes the set of CPUs
@@ -758,6 +775,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
set_cpu_present(cpu, true);
numa_store_cpu_info(cpu);
}
+out:
+ __cpu_to_node = arm64_cpu_to_node;
}
static const char *ipi_types[NR_IPI] __tracepoint_string = {
--
2.40.1
On Mon, Jan 15, 2024 at 05:59:31PM +0800, Huang Shijie wrote:
> After setting the right NUMA node for VMAP stack,
> https://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git/commit/?id=75b5e0bf90bf
>
> I found there are at least four places in the common code where
> the cpu_to_node() is called before it is initialized:
> 0.) early_trace_init() in kernel/trace/trace.c
> 1.) sched_init() in kernel/sched/core.c
> 2.) init_sched_fair_class() in kernel/sched/fair.c
> 3.) workqueue_init_early() in kernel/workqueue.c
>
> We cannot use early_cpu_to_node() for them, since early_cpu_to_node()
> does not work for some ARCHs, such as x86, riscv, etc.
I spot that x86 seems to have an implementation of early_cpu_to_node(); what's
wrong with it?
> So we have to implement the arm64 specific cpu_to_node().
Surely those early uses of cpu_to_node() are equally broken on those other
architectures, so why should this be specific to arm64?
> This patch
> 0.) introduces the __cpu_to_node function pointer,
> and exports it for kernel modules.
>
> 1.) defines a macro cpu_to_node to override the
> generic percpu implementation of cpu_to_node.
>
> 2.) __cpu_to_node is initialized with early_cpu_to_node() before
> numa_node is initialized.
>
> 3.) __cpu_to_node is set to arm64_cpu_to_node() when numa_node is ready.
> arm64_cpu_to_node() is a clone of the generic cpu_to_node().
I don't think this is the right approach. Regardlesss of anything else, we
shouldn't have a solution that only fixes arm64.
Why can't we mandate an early_cpu_to_node(), and have the other architectures
implement that?
Why can't we change cpu_to_node() to automatically do the right thing?
Mark.
>
> Signed-off-by: Huang Shijie <[email protected]>
> ---
> arch/arm64/include/asm/topology.h | 4 ++++
> arch/arm64/kernel/smp.c | 21 ++++++++++++++++++++-
> 2 files changed, 24 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm64/include/asm/topology.h b/arch/arm64/include/asm/topology.h
> index a323b109b9c4..3e99b40b5f9f 100644
> --- a/arch/arm64/include/asm/topology.h
> +++ b/arch/arm64/include/asm/topology.h
> @@ -12,6 +12,10 @@ int pcibus_to_node(struct pci_bus *bus);
> cpu_all_mask : \
> cpumask_of_node(pcibus_to_node(bus)))
>
> +/* override generic percpu implementation of cpu_to_node */
> +extern int (*__cpu_to_node)(int cpu);
> +#define cpu_to_node(cpu) __cpu_to_node(cpu)
> +
> #endif /* CONFIG_NUMA */
>
> #include <linux/arch_topology.h>
> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> index 4ced34f62dab..8a3bc101eaed 100644
> --- a/arch/arm64/kernel/smp.c
> +++ b/arch/arm64/kernel/smp.c
> @@ -90,6 +90,21 @@ static struct irq_desc *ipi_desc[MAX_IPI] __ro_after_init;
>
> static void ipi_setup(int cpu);
>
> +int (*__cpu_to_node)(int cpu);
> +EXPORT_SYMBOL(__cpu_to_node);
> +
> +#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
> +static int arm64_cpu_to_node(int cpu)
> +{
> + return per_cpu(numa_node, cpu);
> +}
> +#else
> +static int arm64_cpu_to_node(int cpu)
> +{
> + return 0;
> +}
> +#endif
> +
> #ifdef CONFIG_HOTPLUG_CPU
> static void ipi_teardown(int cpu);
> static int op_cpu_kill(unsigned int cpu);
> @@ -613,6 +628,7 @@ static void __init acpi_parse_and_init_cpus(void)
>
> for (i = 0; i < nr_cpu_ids; i++)
> early_map_cpu_to_node(i, acpi_numa_get_nid(i));
> + __cpu_to_node = early_cpu_to_node;
> }
> #else
> #define acpi_parse_and_init_cpus(...) do { } while (0)
> @@ -674,6 +690,7 @@ static void __init of_parse_and_init_cpus(void)
> next:
> cpu_count++;
> }
> + __cpu_to_node = early_cpu_to_node;
> }
>
> /*
> @@ -733,7 +750,7 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> * secondary CPUs present.
> */
> if (max_cpus == 0)
> - return;
> + goto out;
>
> /*
> * Initialise the present map (which describes the set of CPUs
> @@ -758,6 +775,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> set_cpu_present(cpu, true);
> numa_store_cpu_info(cpu);
> }
> +out:
> + __cpu_to_node = arm64_cpu_to_node;
> }
>
> static const char *ipi_types[NR_IPI] __tracepoint_string = {
> --
> 2.40.1
>
Hi Mark,
在 2024/1/15 19:44, Mark Rutland 写道:
> On Mon, Jan 15, 2024 at 05:59:31PM +0800, Huang Shijie wrote:
>> After setting the right NUMA node for VMAP stack,
>> https://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux.git/commit/?id=75b5e0bf90bf
>>
>> I found there are at least four places in the common code where
>> the cpu_to_node() is called before it is initialized:
>> 0.) early_trace_init() in kernel/trace/trace.c
>> 1.) sched_init() in kernel/sched/core.c
>> 2.) init_sched_fair_class() in kernel/sched/fair.c
>> 3.) workqueue_init_early() in kernel/workqueue.c
>>
>> We cannot use early_cpu_to_node() for them, since early_cpu_to_node()
>> does not work for some ARCHs, such as x86, riscv, etc.
> I spot that x86 seems to have an implementation of early_cpu_to_node(); what's
> wrong with it?
Yes, you are right. I check the code again, x86 has its own
early_cpu_to_node().
>
>> So we have to implement the arm64 specific cpu_to_node().
> Surely those early uses of cpu_to_node() are equally broken on those other
yes. some ARCHs also has the same issue. But I am not sure if all ARCHs
support the NUMA.
> architectures, so why should this be specific to arm64?
I just want to fix the arm64 first. :)
But if you think we should fix the common code firstly, I am okay.
>
>> This patch
>> 0.) introduces the __cpu_to_node function pointer,
>> and exports it for kernel modules.
>>
>> 1.) defines a macro cpu_to_node to override the
>> generic percpu implementation of cpu_to_node.
>>
>> 2.) __cpu_to_node is initialized with early_cpu_to_node() before
>> numa_node is initialized.
>>
>> 3.) __cpu_to_node is set to arm64_cpu_to_node() when numa_node is ready.
>> arm64_cpu_to_node() is a clone of the generic cpu_to_node().
> I don't think this is the right approach. Regardlesss of anything else, we
> shouldn't have a solution that only fixes arm64.
>
> Why can't we mandate an early_cpu_to_node(), and have the other architectures
> implement that?
>
> Why can't we change cpu_to_node() to automatically do the right thing?
ok, I will change the common code.
Thanks
Huang Shijie