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

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

On Thursday 20 November 2008 09:47:54 David Miller wrote:
> We make a big fuss (rightly) about a few hundred bytes and this sucker
> is FOUR KILOBYTES.

Sure, but this is just trading 4k runtime for 4k static, ie. it's just 
unhiding it.  That's very different from adding random bloat.

> Really for the time being I'd rather see this converted to a
> num_possible_cpus() sized normal kzalloc() and direct indexing.  I
> don't want the networking to bloat up the main kernel image by so
> much.

Nooooooo! :)

First, num_possible_cpus() is the wrong answer (if there are holes in 
cpu_possible_map).  Second, there's little point having percpu infrastructure 
which people avoid.  Third, that memory not going to be numa-aware.  Fourth, 
the dynamic percpu version is fewer instructions (with dynamic percpu 
patches).

I've spent an hour trying to implement DEFINE_PER_CPU_ZERO(), but AFAICT it 
can't be done (gas rejects "b" as a section attribute) :(

So I've dropped the patch, but I wonder if this 4k of stats should be here at 
all.

Thanks,
Rusty.