On 20.09.16 19:32:34, Hanjun Guo wrote:
On 09/20/2016 06:43 PM, Robert Richter wrote:
Instead we need to make sure the set_*numa_node() functions are called
earlier before secondary cpus are booted. My suggested change for that
is this:
diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
index d93d43352504..952365c2f100 100644
--- a/arch/arm64/kernel/smp.c
+++ b/arch/arm64/kernel/smp.c
@@ -204,7 +204,6 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
static void smp_store_cpu_info(unsigned int cpuid)
{
store_cpu_topology(cpuid);
- numa_store_cpu_info(cpuid);
}
/*
@@ -719,6 +718,7 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
continue;
set_cpu_present(cpu, true);
+ numa_store_cpu_info(cpu);
}
}
We tried a similar approach which add numa_store_cpu_info() in
early_map_cpu_to_node(), and remove it from smp_store_cpu_info,
but didn't work for us, we will try your approach to see if works.
Calling it in early_map_cpu_to_node() is probably too early since
setup_node_to_cpumask_map() is called in numa_init() afterwards
overwriting it again.
Actually, early_map_cpu_to_node() is used to temporarily store the
mapping until it can be setup in numa_store_cpu_info().