Re: [patch 3/4] net: Percpufy frequently used variables -- proto.sockets_allocated

[Eric Dumazet <dada1@cosmosbay.com>]

Ravikiran G Thirumalai a écrit :
> On Sat, Jan 28, 2006 at 01:35:03AM +0100, Eric Dumazet wrote:
> > Eric Dumazet a écrit :
> > > Andrew Morton a écrit :
> > > > Eric Dumazet <dada1@cosmosbay.com> wrote:
> > >
> > > long percpu_counter_read_accurate(struct percpu_counter *fbc)
> > > {
> > > 	long res = 0;
> > > 	int cpu;
> > > 	atomic_long_t *pcount;
> > >
> > > 	for_each_cpu(cpu) {
> > > 		pcount = per_cpu_ptr(fbc->counters, cpu);
> > > 		/* dont dirty cache line if not necessary */
> > > 		if (atomic_long_read(pcount))
> > > 			res += atomic_long_xchg(pcount, 0);
> 				--------------------------->  (A)
> > > 	}
>
> > 	atomic_long_add(res, &fbc->count);
> 				--------------------------->  (B)
> > 	res = atomic_long_read(&fbc->count);
> >
> > > 	return res;
> > > }
>
> The read is still theoritically FBC_BATCH * NR_CPUS inaccurate no?
> What happens when other cpus update  their local counters at (A) and (B)?

Well, after  atomic_read(&some_atomic) or percpu_counter_read_accurate(&s) the 
 'value' you got may be inaccurate by whatever... You cannot 'freeze the 
atomic / percpu_counter and gets its accurate value'. All you can do is 
'atomically add/remove some value from this counter (atomic or percpu_counter))

See percpu_counter_read_accurate() as an attempt to collect all local offset 
(and atomically set them to 0) and atomically reajust the central fbc->count 
to with the sum of all these offsets. Kind of a consolidation that might be 
driven by a user request (read /proc/sys/...) or a periodic timer.



> (I am hoping we don't need percpu_counter_read_accurate anywhere yet and
> this is just demo ;).  I certainly don't want to go on all cpus for read / 
> add cmpxchg on the write path for the proto counters that started this 
> discussion)

As pointed by Andrew (If you decode carefully its short sentences :) ), all 
you really need is atomic_long_xchg() (only in slow path), and 
atomic_long_{read/add} in fast path)


Eric

struct percpu_counter {
	atomic_long_t count;
	atomic_long_t *counters;
};

#ifdef CONFIG_SMP
void percpu_counter_mod(struct percpu_counter *fbc, long amount)
{
        long new;
	atomic_long_t *pcount;

	pcount = per_cpu_ptr(fbc->counters, get_cpu());

        new = atomic_long_read(pcount) + amount;
 
	if (new >= FBC_BATCH || new <= -FBC_BATCH) {
                new = atomic_long_xchg(pcount, 0) + amount;
                if (new)
                        atomic_long_add(new, &fbc->count);
	} else
		atomic_long_add(amount, pcount);

	put_cpu();
}
EXPORT_SYMBOL(percpu_counter_mod);

long percpu_counter_read_accurate(struct percpu_counter *fbc)
{
	long res = 0;
	int cpu;
	atomic_long_t *pcount;

	for_each_cpu(cpu) {
		pcount = per_cpu_ptr(fbc->counters, cpu);
		/* dont dirty cache line if not necessary */
		if (atomic_long_read(pcount))
			res += atomic_long_xchg(pcount, 0);
	}
        atomic_long_add(res, &fbc->count);
        return atomic_long_read(&fbc->count);
}
EXPORT_SYMBOL(percpu_counter_read_accurate);
#endif /* CONFIG_SMP */