Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated

[Andi Kleen <ak@suse.de>]

[adding linux-arch]

On Sunday 29 January 2006 01:55, Andrew Morton wrote:
> Benjamin LaHaise <bcrl@kvack.org> wrote:
> > On Sat, Jan 28, 2006 at 01:28:20AM +0100, Eric Dumazet wrote:
> > > We might use atomic_long_t only (and no spinlocks)
> > > Something like this ?
> >
> > Erk, complex and slow...  Try using local_t instead, which is
> > substantially cheaper on the P4 as it doesn't use the lock prefix and act
> > as a memory barrier.  See asm/local.h.
>
> local_t isn't much use until we get rid of asm-generic/local.h.  Bloaty,
> racy with nested interrupts.

It is just implemented wrong. It should use 
local_irq_save()/local_irq_restore() instead. But my bigger problem
with local_t is these few architectures (IA64, PPC64) who implement it 
with atomic_t. This means we can't replace local statistics counters
with local_t because it would be regression for them. I haven't done the
benchmarks yet, but I suspect both IA64 and PPC64 really should
just turn off interrupts.

-Andi