Re: [tip:x86/apic] x86: Add NumaChip support

[Ingo Molnar <mingo@elte.hu>]

* Kevin Winchester <kjwinchester@gmail.com> wrote:

> On 6 December 2011 01:50, Ingo Molnar <mingo@elte.hu> wrote:
> >
> > * Steffen Persvold <sp@numascale.com> wrote:
> >
> >> On 12/5/2011 21:31, Ingo Molnar wrote:
> >> []
> >> > There's a new build failure caused by this commit:
> >>
> >> It's this commit actually :
> >>
> >> Commit-ID:  64be4c1c2428e148de6081af235e2418e6a66dda
> >> Gitweb:     http://git.kernel.org/tip/64be4c1c2428e148de6081af235e2418e6a66dda
> >> Author:     Daniel J Blueman <daniel@numascale-asia.com>
> >> AuthorDate: Mon, 5 Dec 2011 16:20:37 +0800
> >> Committer:  Ingo Molnar <mingo@elte.hu>
> >> CommitDate: Mon, 5 Dec 2011 17:17:21 +0100
> >>
> >> x86: Add x86_init platform override to fix up NUMA core numbering
> >>
> >> >
> >> >   arch/x86/kernel/cpu/common.c:1149:2: error: ‘struct cpuinfo_x86’ has no member named ‘phys_proc_id’
> >> >
> >> > Note that the config is a !SMP one - that's probably the key
> >> > detail.
> >>
> >> Yes, phys_proc_id in struct cpuinfo_x86 is covered within "ifdef CONFIG_SMP".
> >>
> >> >
> >> > Please send a delta fix that resolves this bug. You can find the
> >> > current lineup in:
> >> >
> >> >    git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git
> >> >
> >>
> >> Will do, should be fairly simple. Something in the lines of the attached should work (tested).
> >>
> >> Note: I used "ifdef CONFIG_NUMA" simply because it doesn't make sense in a non-numa configuration even with SMP enabled. Besides, the only place where it is called right now is in kernel/cpu/amd.c:srat_detect_node() within the "CONFIG_NUMA" protected part.
> >>
> >> Cheers,
> >> --
> >> Steffen Persvold, Chief Architect NumaChip
> >> Numascale AS - www.numascale.com
> >> Tel: +47 92 49 25 54 Skype: spersvold
> >
> >> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
> >> index ad4da45..a70bd5b 100644
> >> --- a/arch/x86/kernel/cpu/common.c
> >> +++ b/arch/x86/kernel/cpu/common.c
> >> @@ -1146,7 +1146,9 @@ static void dbg_restore_debug_regs(void)
> >>   */
> >>  void __cpuinit x86_default_fixup_cpu_id(struct cpuinfo_x86 *c, int node)
> >>  {
> >> +#ifdef CONFIG_NUMA
> >>       pr_err("NUMA core number %d differs from configured core number %d\n", node, c->phys_proc_id);
> >> +#endif
> >>  }
> >
> > Yep, this should do the trick - i'll need a Signed-off-by line
> > to apply it.
> >
> > Ideally i think c->phys_proc_id should should be available
> > regardless of CONFIG_SMP or CONFIG_NUMA considerations - but
> > that would be a wider change. (feel free to have a shot at it
> > though, in addition to the fix above)
> >
> 
> If Steffen does not plan to do this additional cleanup, I 
> would give it a try.  You would likely prefer the changes 
> against -tip, correct?

On a second thought, the !SMP block in processor.h::cpuinfo_x86 
is pretty self-contained and making it unconditional would 
increase UP kernel size by 4x5==20 bytes.

I have not checked how many further simplifications this allows 
- if it's a really nice cleanup then i guess we could do it and 
keep the all-zeroes-and-ones default value on UP.

The fields *do* make sense on UP as well.

So it's a "try and see how it ends up" thing.

Thanks,

	Ingo