Re: [patch] Make prof_counter use per-cpu areas patch 1/4 -- x86

[Andrew Morton]

Ravikiran G Thirumalai <[e-mail]> wrote:
>
> Even cpu_possible does not seem to be setup this early.  Seems like 
> reinitialisation of prof_counter/prof_multiplier et al is redundant.
> Here's a newer patch which removes this initialisation at smp_boot_cpus.
> Works fine for me (tested same on a 4 way with difft profiling multipliers..
> LOC interrupts seem to fire at the right intervals).

Things still look a bit fishy to me.

In apic.c:

	int prof_multiplier[NR_CPUS] = { 1, };
	int prof_old_multiplier[NR_CPUS] = { 1, };
	DEFINE_PER_CPU(int, prof_counter) = 1;

This means that all the prof_counter values are set to 1, but the multiplier
arrays have 1 in the zeroeth entry, and zero in the remaining entries.

The zero multipliers remain in place until someone runs
setup_profiling_timer().

One approach would be to initialise all members:

	int prof_multiplier[NR_CPUS] = { [ 0 ... NR_CPUS-1 ] = 1 };

But I think it would be better to put the multipliers into per-cpu memory as
well.  Something like:

struct profiling_info {
	int multiplier;
	int old_multiplier;
	int counter;
};

DEFINE_PER_CPU(struct profiling_info, profiling_info) = {1, 1, 1};

Perhaps?

Also bits and pieces of the profiling code seem to be randomly splattered all
over the place.  Consolidating it all into the one .c file would be nice.