Re: [PATCH 3/3] nohz: Fix iowait overcounting if iowait task migrates

[Denys Vlasenko]

On 04/24/2014 10:19 AM, Hidetoshi Seto wrote:
> As I already mentioned in previous discussion with Peter, I have
> concern here that this change might have impact on performance.
> Especially in case if system is a kind of io-busy box, originally
> there may be no iowait time (and possibly also no idle time).
> For such case this change adds extra execution cost to manage
> value of iowait_exittime which might not used.

Everything has some cost. Correctness usually trumps a few extra
locked bus cycles.

Wouldn't it be nice if we'd know whether anyone even needs
the stats? If no one going to read /proc/stat on the box,
there's no point in going to all the trouble to maintain the counters...

> And if we successfully found a way to get the iowait_exittime
> within reasonable negligible cheap cost, then why we don't use
> it for NOHZ=n kernels too? 

Kernels without NOHZ maintain the counters based on timer interrupt
sampling. It should still work fine.

> As Frederic already pointed, seqcount must be better choice.

Yes, I'm switching to seqcounts.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/