Re: [RFC 2/2] x86, tsc: Enable clock for ealry printk timestamp

[Feng Tang]

Hi Peter and Petr,

Thanks for your suggestions, will try to find a cleaner and less hacky way,
and it may take some time as dealing with all kinds of TSC is tricky :)

- Feng

On Thu, May 31, 2018 at 05:52:10PM +0200, Peter Zijlstra wrote:
> On Thu, May 31, 2018 at 03:55:42PM +0200, Petr Mladek wrote:
> > I wonder if we could get some cleaner integration into the timer and
> > printk code.
> 
> Yes, these patches are particularly horrific..
> 
> There were some earlier patches by Pavel Tatashin, which attempted do
> get things running earlier.
> 
>   http://lkml.kernel.org/r/20180209211143.16215-1-pasha.tatashin@xxxxxxxxxx
> 
> I'm not entirely happy with that, but I never did get around to
> reviewing that last version :-( In particuarly, now that you made me
> look, I dislike his patch 6 almost as much as these patches.
> 
> The idea was to get regular sched_clock() running earlier, not to botch
> some early_sched_clock() into it.
> 
> Basically run calibrate_tsc() earlier (like _waaay_ earlier, it doesn't
> rely on anything other than CPUID) and if you have a recent part (with
> exception of SKX) you'll get a usable tsc rate (and TSC_RELIABLE) and
> things will work.
> 
> If you have a dodgy part (sorry SKX), you'll just have to live with
> sched_clock starting late(r).
> 
> Do not cobble things on the side, try and get the normal things running
> earlier.