timekeeping_adjust may set mult to 0

timekeeping_adjust may set mult to 0

[David Engraf]

Hi,

I have encountered a problem when a linux system uses a clocksource with 
mult = 1 and shift = 0 (clocksource cycle = nanoseconds). It may happen 
that the function timekeeping_adjust reduces the value of mult to 0 when 
error is lower than the interval [1].
As soon as timekeeper.mult is 0, ktime_get will no longer work because 
it uses timekeeping_get_ns which converts the cycle to nanoseconds with 
mult as 0 and the system clocksource returns always 0.

Best regards
David Engraf


[1] http://lxr.linux.no/linux+v3.0.62/kernel/time/timekeeping.c#L821

Re: timekeeping_adjust may set mult to 0

[John Stultz]

On Mon, Feb 11, 2013 at 7:48 AM, David Engraf <david.engraf@sysgo.com> wrote:
> I have encountered a problem when a linux system uses a clocksource with
> mult = 1 and shift = 0 (clocksource cycle = nanoseconds). It may happen that
> the function timekeeping_adjust reduces the value of mult to 0 when error is
> lower than the interval [1].
> As soon as timekeeper.mult is 0, ktime_get will no longer work because it
> uses timekeeping_get_ns which converts the cycle to nanoseconds with mult as
> 0 and the system clocksource returns always 0.

So you *don't* want to use shift=0, since that kills the ability for
the frequency adjustment code to do anything, as you've found.

Instead of calculating the clocksource mult/shift pair yourself,
please use clocksource_register_hz/khz().

I'm hoping to kill off the open clocksource_register() call soon, to
avoid this sort of confusion. Sorry for the trouble.

thanks
-john

Re: timekeeping_adjust may set mult to 0

[David Engraf]

Am 15.02.2013 23:34, schrieb John Stultz:
> On Mon, Feb 11, 2013 at 7:48 AM, David Engraf <david.engraf@sysgo.com> wrote:
>> I have encountered a problem when a linux system uses a clocksource with
>> mult = 1 and shift = 0 (clocksource cycle = nanoseconds). It may happen that
>> the function timekeeping_adjust reduces the value of mult to 0 when error is
>> lower than the interval [1].
>> As soon as timekeeper.mult is 0, ktime_get will no longer work because it
>> uses timekeeping_get_ns which converts the cycle to nanoseconds with mult as
>> 0 and the system clocksource returns always 0.
>
> So you *don't* want to use shift=0, since that kills the ability for
> the frequency adjustment code to do anything, as you've found.

The problem is not shift=0 it's mult=1. The frequency adjustment code 
may increase/decrease mult and mult=0 will not work.

> Instead of calculating the clocksource mult/shift pair yourself,
> please use clocksource_register_hz/khz().

Thanks, I will try it.

Best regards
- David

> I'm hoping to kill off the open clocksource_register() call soon, to
> avoid this sort of confusion. Sorry for the trouble.
>
> thanks
> -john
>