Re: writing a cpufreq driver

[Dave Jones <davej@redhat.com>]

On Fri, Sep 22, 2006 at 09:39:54AM -0500, Ryan Underwood wrote:
 > 
 > On Fri, Sep 22, 2006 at 09:13:18AM -0500, Ryan Underwood wrote:
 > > 
 > > I have a question, how critical is it that I have exactly the correct
 > > CPU MHz for the scaling to work properly?  Unfortunately, for the slow
 > > speed there is no way for me to know what it actually is on a particular
 > > system due to limitations of the firmware, so the best I could do is a
 > > RDTSC timing to try to figure it out.  I wonder if skew between the
 > > actual CPU speed and what cpufreq thinks it is would be the source of
 > > the problem.
 > 
 > Hmm, this might have been the problem.  On my machine, the fast and slow
 > MHz (as measured by my TSC routine) are the same.  However, the machine
 > is noticeably slower at the 'slow' firmware setting.  I thought maybe
 > they were disabling L1 cache, but that is not the case.  The CPU is a
 > Pentium VRT 120MHZ so I'm not sure what is going on.... maybe down
 > clocking the system bus or memory or inserting wait states somewhere.
 > 
 > In any case, since the fast and slow MHZ are the same, cpufreq does
 > nothing.  But cpufreq should still do "something" in this case, since
 > something else that my driver does is to disable the L2 cache when going
 > to the slow state, which is a decent power savings even on my laptop
 > that doesn't have an actual slow MHZ.  But since the fast and slow mhz
 > are the same, cpufreq performs no actions.

*shrug*, I'm out of ideas. I looked over your code, and it seems to
be correctly doing notifications when it performs a transition,
which should be taking care of the timing issues for you.

Puzzlings.

	Dave