Messages in this thread | | | Date | Wed, 19 Jun 2013 14:14:54 +0800 | From | Tang Chen <> | Subject | Re: [Part2 PATCH v4 08/15] x86, numa: Save nid when reserve memory into memblock.reserved[]. |
| |
Hi Vasilis,
Thanks for reviewing. :)
On 06/19/2013 12:57 AM, Vasilis Liaskovitis wrote: ...... > > However, patches 21,22 of part1 and all part3 patches increase kernel usage > of local node memory by putting pagetables local to those nodes. Are these > pagetable pages accounted in part2's memblock_kernel_nodemask? It looks like
No, they are not. What I wanted to acheve was that the local pagetable pages are transparent to users. For a movable node (all memory is hotpluggable), seeing from users level, they think all the node's memory is not used by the kernel. Actually pagetable pages are used by the kernel, but users don't know it, and they don't care about it.
And also, memblock_kernel_nodemask is only used at very early time. When the system is up, it is useless.
This is just my approach for this problem. It is not good enough, and we can improve it.
> part1 and part3 of these patchsets contradict or make the goal of part2 more > difficult to achieve. (I will send more comments for part3 separately). >
I think allocating pagetable to local node really makes thing a little more difficult than before. But I also think Yinghai's work is reasonable because it helps to improve the performance.
What I am thinking now is to allocate things like pagetable pages to local device. (Seems I mentioned this before.)
If a node has more than one memory device, and all the pagetable pages are allocated in one device. Then this device cannot be hot-removed unless all the other memory devices are hot-removed.
So I think allocating pagetable pages to local device is more reasonable. But as you said, this could be more complex.
Thanks. :)
| |