Re: [RFC 00/15] x86_64: Optimize percpu accesses

[H. Peter Anvin]

Jeremy Fitzhardinge wrote:
> H. Peter Anvin wrote:
> > 1. it means pda references are invalid if their offsets are ever more 
> > than CONFIG_PHYSICAL_BASE (which I do not think is likely, but still...)
>
> Why?
>
> As an aside, could we solve the problems by making CONFIG_PHYSICAL_BASE 
> 0 - putting the percpu variables as the first thing in the kernel - and 
> relocating on load?  That would avoid having to make a special PT_LOAD 
> segment at 0.  Hm, would that result in the pda and the boot params 
> getting mushed together?

CONFIG_PHYSICAL_START rather.  And no, it can't be zero!  Realistically 
we should make it 16 MB by default (currently 2 MB), to keep the DMA 
zone clear.

Either way, I really suspect that the right thing to do is to use 
negative offsets, with the possible exception of a handful of things (40 
bytes or less, perhaps like current) which can get small positive 
offsets and end up in the "super hot" cacheline.

The sucky part is that I don't believe GNU ld has native support for a 
"hanging down" section (one which has a fixed endpoint rather than a 
starting point), so it requires extra magic around the link (or finding 
some way to do it with linker script functions.)  Let me see if I can 
cook up something in linker script that would actually work.

	-hpa

--
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/