Messages in this thread |  | | Date | Mon, 18 Feb 2008 12:16:14 -0800 | | From | "Yinghai Lu" <> | | Subject | Re: [PATCH] x86_64: remove never used nodenumer in pda |
| |
On Feb 18, 2008 4:23 AM, Andi Kleen <andi@firstfloor.org> wrote: > Yinghai Lu <Yinghai.Lu@Sun.COM> writes: > > > we don't need copy too. already have x86_cpu_to_node_map > > That's a regression (probably from Mike's patches?). Until recently it was > used. > > The reason the node number was put in there is that it generates far > shorter code to just fetch the local node number from the PDA than to > first go through a array lookup from the cpu number. It also saves a > costly cache line miss on the array if you're unlucky. > > It is far better to fix it than to remove it. > > I know Mike/Christoph want to get rid of the PDA and make per cpu data > as efficient as the PDA. If that happens the right fix is to create > a new per CPU data variable for the node number again. > > Here's a quick patch (tested on kvm with numa emulation only) > > It should be ok because PDA is set up early and > the early node is always 0 and there is a 0 in there > at early boot. > > Saves about 1.6k of text on a vmlinux here.
that is because of the inline...
static inline int early_cpu_to_node(int cpu) { int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
if (cpu_to_node_map) return cpu_to_node_map[cpu]; else if (per_cpu_offset(cpu)) return per_cpu(x86_cpu_to_node_map, cpu); else return NUMA_NO_NODE; }
should use per_cpu(x86_cpu_to_node_map, cpu) instead. and limited using early_cpu_to_node etc.
YH
|  |