powerpc kernel flavour

Jeremy Kerr jeremy.kerr at canonical.com
Tue Dec 7 02:20:40 UTC 2010


Hi Tim,

> I'm all for dropping flavours, but I have no idea how many G4s there are
> still running, and if 3.5% has a noticeable impact.

Ben, any idea about general UP ppc32 usage? I'm guessing we're only looking at 
the pmac platforms here.

> Doesn't x86 rewrite locks at early boot time depending on UP or SMP? Is
> that something we could do for powerpc ?

There's no spinlock-patching on powerpc currently. There has been some 
interest in doing this, but I don't think this is a high priority for the ppc 
folks; most of the current development is on SMP machines.

> Though its likely the dependency on CONFIG_PPC32=y goes deeper then just
> spinlocks.

From a quick look, these seem to be the major runtime differences between 
running a single-CPU machine with CONFIG_SMP and !CONFIG_SMP:

 * accessing the paca (ie, per-cpu info) is slightly more expensive

 * tlb teardown is more complex

 * mm code locking required - eg mmu_hash_lock avoided on !CONFIG_SMP

 * no lazy FP/VMX switching - on !SMP, reg state is only saved if another
   process uses FP/VMX instructions (although we still only restore state
   when required).

 * higher memory usage for per-cpu vars (but NR_CPUS=4, so this is
   pretty small)

 * larger code size (7.3M vs 7.6M), larger bss (484k vs 751k).

[if there's anything I've missed, please add]

Cheers,


Jeremy




More information about the kernel-team mailing list