Re: [git pull] more SLUB updates for 2.6.25

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Friday 08 February 2008 18:29, Eric Dumazet wrote:
> Nick Piggin a écrit :
> > On Friday 08 February 2008 13:13, Christoph Lameter wrote:
> >> are available in the git repository at:
> >>
> >>   git://git.kernel.org/pub/scm/linux/kernel/git/christoph/vm.git
> >> slub-linus
> >>
> >> (includes the cmpxchg_local fastpath since the cmpxchg_local work
> >> by Matheiu is in now, and the non atomic unlock by Nick. Verified that
> >> this is not doing any harm after some other patches had been removed.
> >
> > Ah, good. I think it is always a good thing to be able to remove atomics.
> > They place quite a bit of burden on the CPU, especially x86 where it also
> > has implicit memory ordering semantics (although x86 can speculatively
> > get around much of the problem, it's obviously worse than no restriction)
> >
> > Even if perhaps some cache coherency or timing quirk makes the non-atomic
> > version slower (all else being equal), then I'd still say that the non
> > atomic version should be preferred.
>
> What about IRQ masking then ?

I really did mean all else being equal. eg. "clear_bit" vs "__clear_bit".


> Many CPU pay high cost for cli/sti pair...

True, and many UP architectures have to implement atomic operations
with cli/sti pairs... so those are more reasons to use non-atomics.


> And SLAB/SLUB allocators, even if only used from process context, want to
> disable/re-enable interrupts...
>
> I understand kmalloc() want generic pools, but dedicated pools could avoid
> this cli/sti

Sure, I guess that would be possible. I've kind of toyed with doing
some cli/sti mitigation in the page allocator, but in that case I
found that it wasn't a win outside microbenchmarks: the cache
characteristics of the returned pages are just as important if not
more so than cli/sti costs (although that balance would change
depending on the CPU and workload I guess).

For slub yes you could do it with fewer downsides with process context
pools.

Is it possible instead for architectures where cli/sti is so expensive
to change their lowest level of irq handling to do this by setting and
clearing a soft flag somewhere? That's what I'd rather see, if possible.