Messages in this thread | | | Date | Thu, 15 Jan 2009 19:39:42 +0100 | From | Ingo Molnar <> | Subject | Re: [PATCH] percpu: add optimized generic percpu accessors |
| |
* Andrew Morton <akpm@linux-foundation.org> wrote:
> On Thu, 15 Jan 2009 19:02:59 +0100 Ingo Molnar <mingo@elte.hu> wrote: > > > > > * Andrew Morton <akpm@linux-foundation.org> wrote: > > > > > On Thu, 15 Jan 2009 22:23:19 +0900 Tejun Heo <tj@kernel.org> wrote: > > > > > > > --- a/include/asm-generic/percpu.h > > > > +++ b/include/asm-generic/percpu.h > > > > @@ -80,4 +80,56 @@ extern void setup_per_cpu_areas(void); > > > > #define DECLARE_PER_CPU(type, name) extern PER_CPU_ATTRIBUTES \ > > > > __typeof__(type) per_cpu_var(name) > > > > > > > > +/* > > > > + * Optional methods for optimized non-lvalue per-cpu variable access. > > > > + * > > > > + * @var can be a percpu variable or a field of it and its size should > > > > + * equal char, int or long. percpu_read() evaluates to a lvalue and > > > > + * all others to void. > > > > + * > > > > + * These operations are guaranteed to be atomic w.r.t. preemption. > > > > + * The generic versions use plain get/put_cpu_var(). Archs are > > > > + * encouraged to implement single-instruction alternatives which don't > > > > + * require preemption protection. > > > > + */ > > > > +#ifndef percpu_read > > > > +# define percpu_read(var) \ > > > > + ({ \ > > > > + typeof(per_cpu_var(var)) __tmp_var__; \ > > > > + __tmp_var__ = get_cpu_var(var); \ > > > > + put_cpu_var(var); \ > > > > + __tmp_var__; \ > > > > + }) > > > > +#endif > > > > > > I wonder if the preempt_disable()/preempt_enable() in here actually > > > does anything useful on any architecture. > > > > Provides "this is IRQ safe" > > ? > > > and "this is preempt safe" semantics. > > Of course. But do any architectures actually _need_ that for a single > read?
not for a read i guess - but for the other ops like add/and/or/xor.
> Maybe. And if so, they can interpose their arch-specific > implementation. But if the generic version is optimal for them, they > wouldn't need to..
we cannot turn the generic ops into a single instruction so arch methods are preferable no matter how thick or thin the generic version is. But i agree that the optimization you suggest could be done.
Ingo
| |