Re: [PATCH v2] x86: Reduce clock calibration time during slave cpu startup

[Matthew Garrett <mjg@redhat.com>]

On Fri, Aug 05, 2011 at 11:38:36PM +0200, Ingo Molnar wrote:

> Well, it still uses heuristics: it assumes frequency is the same when 
> the cpuid data tells us that two CPUs are on the same socket, right?

If we only assume that when we have a constant TSC then it's a pretty 
safe assumption - the delay loop will be calibrated against the TSC, and 
the TSC will be constant across the package regardless of what frequency 
the cores are actually running at.

> Cannot we directly see the frequency(ies) of the CPU, and decide 
> based on that whether the calibration result can be cached?

Kind of? We can look at the aperf/mperf ratio and compare that against 
the TSC to work out the frequency. But that only gives you meaningful 
results with a constant TSC, so since you're still depending on that you 
might as well skip the extra check.

The only time this should give problems is if someone produces a 
multi-core x86 with per-core (rather than per-package) TSC and runs the 
TSC at different frequencies per-core. Since that'd probably cost more 
than not doing it, I think we're probably safe.

-- 
Matthew Garrett | mjg59@srcf.ucam.org