Re: [PATCH] CPU frequency display in /proc/cpuinfo

[Andi Kleen]

> JACK needs to know the CPU speed in order to be able to use RDTSC for
> timing.  Yes that might be "broken" but gettimeofday() is simply not

It is broken then.  Was always broken, will be broken etc.

> fast enough for our use, we can't afford the overhead of thousands of
> system calls per second.  And until recently 99.999% of desktop machines
> had a monotonic TSC so this worked very well.

You're wrong. First if you say "monotonic" you don't understand the problem.
Monotonicity is only a small part of it.  The bigger one is just 
getting the current frequency and figuring out of if the information
is safe to use.

Chips where it doesn't work:

- Intel Prescott and derivatives (newer Pentium 4, newer Xeon, on
Celeron you are still lucky because they normally don't have speedstep): 
TSC runs at a different frequency than the current P state and P state often 
varies with speedstep [These are the most common desktop chips in the world, 
although not all have speedstep enabled. Many newer ones have though.]
- Intel P-M and earlier P4,P3 mobile chips, Athlon 64, Athlon XP-M,
Opteron etc: 
TSC varies with P state and user space has no way to get timely 
updates on P-state changes.
- VIA C3: I believe TSC stops in idle (at least it used to on
older versions, don't know if they fixed it)

-Andi

-
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/