Re: [PATCH] net: make ip_rt_acct a normal percpu var

[Rusty Russell <rusty@rustcorp.com.au>]

On Thursday 20 November 2008 09:53:48 Eric Dumazet wrote:
> Rusty Russell a écrit :
> > On Thursday 20 November 2008 08:50:23 David Miller wrote:
> >> Do you really need this to forward some work you are doing?  If not
> >> can we just let sleeping dogs lie on this one? :)
> >
> > Yes, I have patches to convert the dynamic percpu data to use the same
> > mechanism as static percpu data.  Unfortunately we don't have a mechanism
> > for enlarging the percpu region (which is why this wasn't done earlier),
> > so we use a heuristic to figure out how much extra percpu region to
> > allocate at boot.
> >
> > And 4k makes this one of the Big Pigs in dynamic per-cpu allocations.
> >
> > (SNMP mibs are even worse, but that's a separate debate...)
> >
> > I can try to implement a bss-like DEFINE_PER_CPU_ZERO(), but it seems
> > silly to talk about tight boot loader size restrictions for SMP kernels.
>
> Then, if we really want to run 4096 cpus on a machine, we dont want to
> allocate 16 MBytes of memory for these ip_rt_acct counters, or even more
> for SNMP mibs.
>
> Maybe its time to design a new mechanism, to avoid the basic "one variable"
> shared by all cpus, and avoid the overkill "one separate variable for each
> cpu", and loop 4096 times to do the sum of this variable...

Per-node vars; no doubt we'll get there.  It might be worth having YA percpu 
counters implementation which does exactly this.  After the dynamic percpu 
changes and some local_* ops changes to allow use with dynamic percpu vars, it 
should be straightforward.

I don't think it's urgent: my concern is not with people who have 4096 cpus 
(but I do care about people with 2 cpus and CONFIG_NR_CPUS=4096).

Cheers,
Rusty.