Re: [RFC PATCH] mm: introduce N_LRU_MEMORY to distinguish betweennormal and movable memory

[Hanjun Guo]

On 2012/8/9 22:06, Christoph Lameter (Open Source) wrote:
> On Thu, 9 Aug 2012, Hanjun Guo wrote:
> 
>> Now, We have node masks for both N_NORMAL_MEMORY and
>> N_HIGH_MEMORY to distinguish between normal and highmem on platforms such as x86.
>> But we still don't have such a mechanism to distinguish between "normal" and "movable"
>> memory.
> 
> What is the exact difference that you want to establish?

Hi Christoph,
    Thanks for your comments very much!

We want to identify the node only has ZONE_MOVABLE memory.
for example:
	node 0: ZONE_DMA, ZONE_DMA32, ZONE_NORMAL--> N_LRU_MEMORY, N_NORMAL_MEMORY
	node 1: ZONE_MOVABLE			 --> N_LRU_MEMORY
thus, in SLUB allocator, will not allocate memory control structures for node1.

static int init_kmem_cache_nodes(struct kmem_cache *s)
{
	int node;

	for_each_node_state(node, N_NORMAL_MEMORY) { /* <-- skip nodes only has ZONE_MOVABLE memory */
		struct kmem_cache_node *n;

		if (slab_state == DOWN) {
			early_kmem_cache_node_alloc(node);
			continue;
		}
		n = kmem_cache_alloc_node(kmem_cache_node,
						GFP_KERNEL, node);

		...
	}
	...
}

> 
>> As suggested by Christoph Lameter in threads
>> http://marc.info/?l=linux-mm&m=134323057602484&w=2, we introduce N_LRU_MEMORY to
>> distinguish between "normal" and "movable" memory.
> 
> Well seems that I am having second thoughts about this. While is it true
> that current page migration can only move pages on the LRU there are
> already various mechanisms proposed and implemented that can move pages
> not on the LRU (like page table pages). Not sure if this is still a useful
> distinction to make. There is also the issue that segments from
> "N_LRU_MEMORY" may be allocated and then become not movable anymore.

Some kernel pagesïlike memmap pagesïusemap pages are still can not be
migrated.

> 
> For the slab case that you want to solve here you will need to know if the
> node has *only* movable memory and will never have any ZONE_NORMAL memory.
> If so then memory control structures for allocators that do not allow
> movable memory will not need to be allocated for these node. The node can
> be excluded from handling.

I think this is what we are trying to do in this patch.
did I miss something?

> 
> .
> 


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/