Re: [PATCH] mm: Only re-generate demotion targets when a numa node changes its N_CPU state

[Baolin Wang]

Hi Oscar,

On 3/10/2022 1:48 PM, Oscar Salvador wrote:
> On Thu, Mar 10, 2022 at 08:39:53AM +0800, Huang, Ying wrote:
> > We need to call set_migration_target_nodes() during system boot
> > somewhere, either here or in init_mm_internals().
>
> Hi Huang Ying,
>
> vmstat_cpu_online() already gets called during boot:
>
> static struct cpuhp_step cpuhp_hp_states[] = {
> ...
> 	#ifdef CONFIG_SMP
> 	[CPUHP_CREATE_THREADS]= {
> 		.name			= "threads:prepare",
> 		.startup.single		= smpboot_create_threads,
> 		.teardown.single	= NULL,
> 		.cant_stop		= true,
> 	},
> ...
>
> smpboot_create_threads
>   __smpboot_create_thread
>    smpboot_thread_fn
>     ht->thread_fn()
>      cpuhp_thread_fun
>       cpuhp_invoke_callback
>        vmstat_cpu_online
>
> That for every CPU that is brought up during boot.
>
> So unless I am missing something, I would say we are already covered
> there, right?

I've tested your patch, unfortunately it can not work. It already set 
the node with N_CPU state in init_cpu_node_state() earlier, so in 
vmstat_cpu_online() ,the sentence 'if(!node_state(cpu_to_node(cpu), 
N_CPU))' is always false, which means it will not call 
set_migration_target_nodes().

Like Huang Ying said, we should call set_migration_target_nodes() in 
migrate_on_reclaim_init() or init_mm_internals().