On Wed 26-10-16 11:10:44, Leizhen (ThunderTown) wrote:

On 2016/10/25 21:23, Michal Hocko wrote:

quoted

On Tue 25-10-16 10:59:17, Zhen Lei wrote:

quoted

If HAVE_MEMORYLESS_NODES is selected, and some memoryless numa nodes are
actually exist. The percpu variable areas and numa control blocks of that
memoryless numa nodes need to be allocated from the nearest available
node to improve performance.

Although memblock_alloc_try_nid and memblock_virt_alloc_try_nid try the
specified nid at the first time, but if that allocation failed it will
directly drop to use NUMA_NO_NODE. This mean any nodes maybe possible at
the second time.

To compatible the above old scene, I use a marco node_distance_ready to
control it. By default, the marco node_distance_ready is not defined in
any platforms, the above mentioned functions will work as normal as
before. Otherwise, they will try the nearest node first.

I am sorry but it is absolutely unclear to me _what_ is the motivation
of the patch. Is this a performance optimization, correctness issue or
something else? Could you please restate what is the problem, why do you
think it has to be fixed at memblock layer and describe what the actual
fix is please?

This is a performance optimization.

Do you have any numbers to back the improvements?

The problem is if some memoryless numa nodes are
actually exist, for example: there are total 4 nodes, 0,1,2,3, node 1 has no memory,
and the node distances is as below:
                    ---------board-------
		    |                   |
                    |                   |
                 socket0             socket1
                   / \                 / \
                  /   \               /   \
               node0 node1         node2 node3
distance[1][0] is nearer than distance[1][2] and distance[1][3]. CPUs on node1 access
the memory of node0 is faster than node2 or node3.

Linux defines a lot of percpu variables, each cpu has a copy of it and most of the time
only to access their own percpu area. In this example, we hope the percpu area of CPUs
on node1 allocated from node0. But without these patches, it's not sure that.

I am not familiar with the percpu allocator much so I might be
completely missig a point but why cannot this be solved in the percpu
allocator directly e.g. by using cpu_to_mem which should already be
memoryless aware.

Generating a new API while we have means to use an existing one sounds
just not right to me.
-- 
Michal Hocko
SUSE Labs