Messages in this thread | | | From | "Rafael J. Wysocki" <> | Subject | Re: [PATCH 3/6 v4] cpufreq: tolerate inexact values when collecting stats | Date | Sat, 24 Nov 2012 11:05:09 +0100 |
| |
On Saturday, November 17, 2012 03:50:48 PM Borislav Petkov wrote: > On Tue, Nov 13, 2012 at 02:13:38PM -0500, Mark Langsdorf wrote: > > Although cpufreq_driver has a flag field, no part of cpufreq_driver > > is directly passed to the cpufreq_stat code. Only cpufreq_policy > > is. It's cleaner to do passes of the while loop than to copy the > > cpufreq_driver.flag field into cpufreq_policy and then store it again > > in cpufreq_stats. > > That maybe so but this newly added loop which is only Calxeda-relevant > is called in cpufreq_stat_notifier_trans, which is the frequency change > notifier call, AFAICT. > > So each cpufreq driver will be paying that small and needless penalty > now for nothing and on each frequency change. Which adds to the > kernel-wide bloat and we absolutely don't want that. > > So you probably need to find a slick way of detecting calxeda hw > somewhere along the init path of cpufreq_stats_init and set a > hw-specific flag instead of adding that cost to each driver.
Mark, I suppose you'd like me to take this series for v3.8, but the above comment from Boris has to be addressed for that.
Thanks, Rafael
-- I speak only for myself. Rafael J. Wysocki, Intel Open Source Technology Center.
| |