Source-Changes-D archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: CVS commit: src/sys/arch/x86/x86

On Mon, Aug 08, 2011 at 02:29:36PM +0300, Jukka Ruohonen wrote:
> We should not really trust ACPI/FADT here. See acpicpu(4) how this is
> derived from the actual CPU information. Additionally, I suggested decreasing
> the quality of tsc(9) based on this information a long time ago, but joerg@
> had concerns about this.

The basic concern here is that TSC and LAPIC timer are an order of
magnitude faster than the alternatives. IIRC TSC and LAPIC on my old
Core 2 laptop needed around 100ns +- 20% and the HPET around 1200ns for
gettimeofday (e.g. including system call overhead).

I forgot whether the phase shift correction for LAPIC was ever
implemented, but with it, there is a whole range of machines where LAPIC
is stable and TSC isn't. More importantly, if the LAPIC is stopped, you
are hosed anyway since it also drives hardclock. The performance
difference is quite small though, so dropping TSC's quality below LAPIC
wouldn't be too bad.


Home | Main Index | Thread Index | Old Index