Re: [PATCH 1/8] mm/memory-tiers: consolidate memory type dedup into mt_get_memory_type()

[David Hildenbrand (Arm)]

On 3/21/26 16:03, Gregory Price wrote:
> Replace per-driver memory type list infrastructure with a single
> mt_get_memory_type(adist) that deduplicates against the global
> default_memory_types list under memory_tier_lock.
> 
> The per-driver lists (mutex + list_head + find/put wrappers) provided
> dedup within a single driver, but not across drivers or with the core.
> Since the number of distinct adist values is bounded and types on
> default_memory_types are never freed anyway, the per-driver cleanup
> on module unload was not useful.
> 
> Add MEMTIER_DEFAULT_LOWTIER_ADISTANCE to replace the default DAX
> adistance, since it was really used as a standin for all kmem hotplugged
> memory.  This at least makes the default tier relationship clearer to
> other drivers and they can see where to put their memory in relation to
> the default lower tier.

Very confusing code.

What's the purpose of kref_get/kref_put if "the types on
default_memory_types are never freed anyway" ?

IIUC, init_node_memory_type() is always called with one of the types
obtained through mt_get_memory_type(). And that one always gives us one
from the default_memory_types list, no?

Why does mt_find_alloc_memory_type() now even consume a list, if there
really only is &default_memory_types?


Is there even a need to expose the "struct memory_dev_type" to kmem,
when all it does is effectively passing a reference to
init_node_memory_type() / clear_node_memory_type().

IOW, couldn't init_node_memory_type()/clear_node_memory_type() just
consume the "adist", and lookup the memory type itself?

All you'd need is some way in the driver to verify that there is a
memory type for the given adist, as some kind of prepare step.

Alternatively, let init_node_memory_type() return an error to get a
clean interface? :)

[...]

>  		mapped++;
>  	}
> +	data->mtype = mtype;
>  
>  	dev_set_drvdata(dev, data);
>  
> @@ -253,7 +233,7 @@ static void dev_dax_kmem_remove(struct dev_dax *dev_dax)
>  		 * for that. This implies this reference will be around
>  		 * till next reboot.
>  		 */
> -		clear_node_memory_type(node, NULL);
> +		clear_node_memory_type(node, data->mtype);
>  	}

Likely clear_node_memory_type() can now be taught to not handle
"!memtype" anymore?


-- 
Cheers,

David