Re: [PATCH] netfilter: save the hash of the tuple in the original direction for latter use

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 17 août 2010 à 16:46 +0800, Changli Gao a écrit :
> On Tue, Aug 17, 2010 at 4:30 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> >
> > Three variables ?
> >
> > static atomic_t rnd __read_mostly;
> >
> > if (unlikely(!atomic_read(&rnd))) {
> >        unsigned int val;
> >
> >        get_random_bytes(&val, sizeof(val));
> >        if (!val)
> >                val = 1;
> >        atomic_cmpxchg(&rnd, 0, val);
> > }
> >
> 
> 
> Good idea. However, atomic_t is a volatile variable, and it prevent
> ILP. I think maybe it hurts the likely case. cmpxchg() is an atomic
> operations, so is the bellow code better?
> 

I am not sure what you mean by ILP.

On x86, reading twice same memory location is faster than
reading it and store in temp variable (on stack)
reading from stack

>  static unsigned long rnd __read_mostly;
> 
>  if (unlikely(!rnd)) {
>         unsigned long val;
> 
>         get_random_bytes(&val, sizeof(val));
>         if (!val)
>                 val = 1;
>         cmpxchg(&rnd, 0, val);
>  }
> 
> Thanks.
> 

I am not sure we must use a long (we really need 4 bytes only), and last
time I tried to use cmpxchg(), I was being told it was not available on
all arches.

But seeing it used in kernel/pid.c, maybe its not true anymore (that is,
__HAVE_ARCH_CMPXCHG is always defined to 1)

Since its a recent change (in kernel/pid.c), I would wait a bit and see
if an arch maintainer complains ;)