Le jeudi 28 avril 2011 Ã 10:22 -0500, Christoph Lameter a Ãcrit :
> On Thu, 28 Apr 2011, Tejun Heo wrote:
>
> > It seems like we can split hairs all day long about the similarities
> > and differences with atomics, so let's forget about atomics for now.
>
> Ok good idea. So you accept that per cpu counters are inevitably fuzzy and
> that the result cannot be relied upon in the same way as on atomics?
>
> > I don't like any update having possibility of causing @batch jumps in
> > _sum() result. That severely limits the usefulness of hugely
> > expensive _sum() and the ability to scale @batch. Not everything in
> > the world is vmstat. Think about other _CURRENT_ use cases in
> > filesystems.
>
> Of course you can cause jumps > batch. You are looping over 8 cpus and
> have done cpu 1 and 2 already. Then cpu 1 adds batch-1 to the local per
> cpu counter. cpu 2 increments its per cpu counter. When _sum returns we
> have deviation of batch.
>
> In the worst case the deviation can increase to (NR_CPUS - 1) * (batch
> -1). That is for the current code!!!
>
> The hugely expensive _sum() is IMHO pretty useless given the above. It is
> a function that is called with the *hope* of getting a more accurate
> result.
>
> The only way to reduce the fuzziness is by reducing batch. But then you
> dont need the _sum function.
>
> Either that or you need additional external synchronization by the
> subsystem that prevents updates.
>
>
We could add a seqcount (shared), and increment it each time one cpu
changes global count.
_sum() could get an additional parameter, the max number of allowed
changes during _sum() run.
If _sum() notices seqcount was changed too much, restart the loop.
s64 __percpu_counter_sum(struct percpu_counter *fbc, unsigned maxfuzzy)
{
s64 ret;
unsigned int oldseq, newseq;
int cpu;
restart:
oldseq = fbc->seqcount;
smp_rmb();
ret = fbc->count;
for_each_online_cpu(cpu) {
s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
ret += *pcount;
}
smp_rmb()
newseq = fbc->count;
if (newseq - oldseq >= maxfuzzy)
goto restart;
return ret;
}
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>