Re: [PATCH] ipv4: only produce one record in fib_seq_show

[Andrew Morton <akpm@digeo.com>]

"David S. Miller" wrote:
> 
> ..
> kernel/timer.c's main data structures desperately want to be per-cpu
> or allocated at boot time also.  It, as has been noted often on this
> list, is actually more bloat than the ipv4 statistics stuff. :-)

I have a bunch of patches under test which do that.

The tricky bit is that at present, the per-cpu data for all 32
CPUs is always allocated.  We had to change that to only allocate
the secondary CPU's memory when the CPU is coming up.

This means that:

	for (i = 0; i < NR_CPUS; i++)
		play_with(per_cpu(something, i));

becomes a bug.  Because the per-cpu memory for not-possible
CPUs is not allocated.  I created a `for_each_possible_cpu(i)'
helper macro for that.

Another issue:

	/* this is basically for_each_possible_cpu() */
	for (i = 0; i < NR_CPUS; i++) {
		if (cpu_possible(i)) {
			play_with(per_cpu(something, i));
		}
	}

the above code will fail if it is run before smp_init(), because
smp_init() sets up cpu_callout_map() and the per-cpu memory.
We (Dipanker and I) fixed up all the callers.

The code works OK, on ia32.  But we think ia64 needs more work.

That's 128k saved.  There's another 128k-odd in a huge hashtable
in kernel/pid.c (even on uniprocessor) which needs a diet.  Also
80k or so in runqueues.  The runqueues are set up before smp_init(),
but I expect a cpu notifier and some reorg there will work.  Haven't
done sched.c yet.

I suspect we waste more memory than this in dynamically allocated
hashtables and mempools though.