Re: [PATCH 3/3] mm/memory_hotplug: Refactor hotadd_init_pgdat and try_online_node

[David Hildenbrand <david@redhat.com>]

On 07.03.22 16:07, Oscar Salvador wrote:
> Since we pre-allocate all nodes now, hotadd_init_pgdat() does
> not need to return a pgdat struct, as that was meant for
> __try_online_node() to check whether the node was succesfully
> allocated.
> 
> Also get rid of the __ref as all functions hotadd_init_pgdat()
> calls fall within the same section.
> 
> Also try to make more clear the return codes from __try_online_node().
> __try_online_node() can return either 0, 1 or -errno (the latter not really
> as the BUG_ON() would catch it before we return) but depending on the caller
> that has different meanings.
> For add_memory_resource(), when __try_online_node() returns non-zero,
> it means that the node was already allocated and it does not need to bring
> it up. It is fine not to check for -errno values because we do not
> get to call register_one_node() when !set_node_online.
> For those who call try_online_node(), so set_node_online is true, a value
> other than zero means a failure (e.g: cpu_up() or find_and_online_cpu_nid()).
> 
> Signed-off-by: Oscar Salvador <osalvador@suse.de>
> ---
>  mm/memory_hotplug.c | 35 ++++++++++++++---------------------
>  1 file changed, 14 insertions(+), 21 deletions(-)
> 
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index 07cece9e22e4..5c92ac81a399 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -1161,8 +1161,7 @@ static void reset_node_present_pages(pg_data_t *pgdat)
>  	pgdat->node_present_pages = 0;
>  }
>  
> -/* we are OK calling __meminit stuff here - we have CONFIG_MEMORY_HOTPLUG */
> -static pg_data_t __ref *hotadd_init_pgdat(int nid)
> +static void hotadd_init_pgdat(int nid)
>  {
>  	struct pglist_data *pgdat = NODE_DATA(nid);
>  
> @@ -1182,8 +1181,6 @@ static pg_data_t __ref *hotadd_init_pgdat(int nid)
>  	 * to access not-initialized zonelist, build here.
>  	 */
>  	build_all_zonelists(pgdat);
> -
> -	return pgdat;
>  }
>  
>  /*
> @@ -1193,31 +1190,27 @@ static pg_data_t __ref *hotadd_init_pgdat(int nid)
>   * called by cpu_up() to online a node without onlined memory.
>   *
>   * Returns:
> - * 1 -> a new node has been allocated
> - * 0 -> the node is already online
> - * -ENOMEM -> the node could not be allocated
> + * 1 -> The node has been initialized.
> + * 0 -> Either the node was already online, or we succesfully registered a new
> + *      one.
> + * -errno -> register_one_node() failed.
>   */
>  static int __try_online_node(int nid, bool set_node_online)
>  {
> -	pg_data_t *pgdat;
> -	int ret = 1;
> +	int ret;
>  
>  	if (node_online(nid))
>  		return 0;
>  
> -	pgdat = hotadd_init_pgdat(nid);
> -	if (!pgdat) {
> -		pr_err("Cannot online node %d due to NULL pgdat\n", nid);
> -		ret = -ENOMEM;
> -		goto out;
> -	}
> +	hotadd_init_pgdat(nid);
> +
> +	if (!set_node_online)
> +		return 1;
> +
> +	node_set_online(nid);
> +	ret = register_one_node(nid);
> +	BUG_ON(ret);
>  
> -	if (set_node_online) {
> -		node_set_online(nid);
> -		ret = register_one_node(nid);
> -		BUG_ON(ret);
> -	}
> -out:
>  	return ret;

BUG_ON(ret);
return ret;

hm? This will never return :)

So either leave the old code flow or "return 1;" I'd actually prefer the
old code flow to then "return 1;" here:

if (set_node_online) {
	node_set_online(nid);
	BUG_ON(register_one_node(nid));
}
return 1;

We can then let go of "ret".

-- 
Thanks,

David / dhildenb