Messages in this thread | | | Date | Tue, 22 Sep 2009 10:32:20 +0100 | From | Mel Gorman <> | Subject | Re: [PATCH 1/3] powerpc: Allocate per-cpu areas for node IDs for SLQB to use as per-node areas |
| |
On Tue, Sep 22, 2009 at 09:01:55AM +0900, Tejun Heo wrote: > Mel Gorman wrote: > > diff --git a/arch/powerpc/kernel/setup_64.c b/arch/powerpc/kernel/setup_64.c > > index 1f68160..a5f52d4 100644 > > --- a/arch/powerpc/kernel/setup_64.c > > +++ b/arch/powerpc/kernel/setup_64.c > > @@ -588,6 +588,26 @@ void __init setup_per_cpu_areas(void) > > paca[i].data_offset = ptr - __per_cpu_start; > > memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start); > > } > > +#ifdef CONFIG_SLQB > > + /* > > + * SLQB abuses DEFINE_PER_CPU to setup a per-node area. This trick > > + * assumes that ever node ID will have a CPU of that ID to match. > > + * On systems with memoryless nodes, this may not hold true. Hence, > > + * we take a second pass initialising a "per-cpu" area for node-ids > > + * that SLQB can use > > + */ > > + for_each_node_state(i, N_NORMAL_MEMORY) { > > + > > + /* Skip node IDs that a valid CPU id exists for */ > > + if (paca[i].data_offset) > > + continue; > > + > > + ptr = alloc_bootmem_pages_node(NODE_DATA(cpu_to_node(i)), size); > > + > > + paca[i].data_offset = ptr - __per_cpu_start; > > + memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start); > > + } > > +#endif /* CONFIG_SLQB */ > > } > > #endif > > Eh... I don't know. This seems too hacky to me.
I've come around to this opinion as well. There are probably too many other architectures and corners where this is gotten wrong and a more fundamental fix is needed for SLQB to be able to use this hack.
> Why not just > allocate pointer array of MAX_NUMNODES and allocate per-node memory > there?
That's what patch 1 from V1 did. I'll make it Patch 1 for V3.
> This will be slightly more expensive but I doubt it will be > noticeable. The only extra overhead is the cachline footprint for the > extra array. >
I'll compare the vmlinux's to quantify the exact penalty but basically, I don't think it can be avoided at this point.
Thanks.
-- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab
| |