Re: [patch] add optimized generic percpu accessors

[Ingo Molnar <mingo@elte.hu>]

* Rusty Russell <rusty@rustcorp.com.au> wrote:

> On Thursday 15 January 2009 22:08:26 Tejun Heo wrote:
> > Hello,
> > 
> > Ingo Molnar wrote:
> > > The new ops are a pretty nice and clean solution i think.
> > > 
> > > Firstly, accessing the current CPU is the only safe shortcut anyway (there 
> > > is where we can do %fs/%gs / rip-relative addressing modes), and the 
> > > generic per_cpu() APIs dont really provide that guarantee for us. We might 
> > > be able to hook into __get_cpu_var() but those both require to be an 
> > > lvalue and are also relatively rarely used.
> > > 
> > > So introducing the new, rather straightforward APIs and using them 
> > > wherever they matter for performance is good. Your patchset already shaved 
> > > off an instruction from ordinary per_cpu() accesses, so it's all moving 
> > > rather close to the most-optimal situation already.
> > 
> > Yeah, I don't think we can do much better than those ops.  I have two
> > issues tho.
> > 
> > 1. percpu_and() is missing.  I added it for completeness's sake.
> > 
> > 2. The generic percpu_op() should be local to the cpu, so it should
> >    expand to...
> > 
> >    do { get_cpu_var(var) OP (val); put_cpu_var(var) } while (0)
> > 
> >    as the original x86_OP_percpu() did.  Right?
> > 
> > Thanks.
> 
> No no no no.  This is a crapload of infrastructure noone will use.
>
> Please just start by adding read_percpu like so (this won't apply since 
> there's lots of other per-cpu things going on, but you get the idea).
> 
> We don't need a whole set of operators for a handful of 
> non-arch-specific cases.  Reading a var is fairly common, the other ops 
> are diminishing returns and we already have local_t for some of these 
> cases (and we're reviewing that, too).

Actually, the percpu_add()/sub() ops are useful for statistics. (can be 
done without preempt disable/enable) percpu_write() is also obviously 
useful. The others are common arithmetic operators, for completeness of 
the API.

	Ingo