Re: [PATCH] i386 tsc: remove xtime_lock'ing around cpufreq notifier

[Jeremy Fitzhardinge <jeremy@goop.org>]

Andi Kleen wrote:
> Even on real hardware it's also per CPU, although the errors
> are usually not big. At least the scheduler deals with that by
> only ever comparing time stamps from the same CPU.
>   

Well, it uses sched_clock to measure how long something has been asleep,
which is inherently non-per-cpu.  But it tries to keep a measure of the
skew between the various runqueue's sched_clocks, so the error doesn't
seem to get too large.

> If you have big deviations between CPUs then it might cause problems
> for non scheduler uses. I guess printk_clock is not critical, but
> it might be a little confusing.

They could be huge differences - unbounded, in fact.  It would make
printk fairly hard to interpret,  I would think.  The only benefit to
using sched_clock in printk is that if you're using it to work out the
startup latencies you won't be confused by stolen time.  But I think
that's a fairly small benefit compared to the disadvantage of not being
able to meaningfully compare the timestamps on two printk messages.

    J