Re: [PATCH] [15/58] i386: Rewrite sched_clock

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Andi Kleen (ak@suse.de) wrote:
> 
> > So if, after this, we run tsc_sched_clock() with an unstable TSC, we
> > read a last_val containing the interrupt's MSB and the last_val LSB. It
> > can particularity hurt if we are around a 32 bits overflow, because time
> > could "jump" forward of about 1.43 seconds on a 3 GHz system.
> > 
> > So I guess we need synchronization on the fast path, and therefore using
> > cmpxchg_local on x86_64
> 
> On x86-64 the 64bit write is atomic against interrupts.
> 
> You're right 32bit has a problem though. I'm not too happy about 
> cmpxchg though because that wouldn't work on some CPUs.
> 

Which CPUs ? 386 ? do they even have a cycle counter ?

Please have a look at my reply to this email for the x86_64 problematic
case. This problematic case also applies to i386 (non atomicity of tsc
read vs write to memory).

> I wonder if we can just get away with using a 32bit value on i386.
> Just for the purpose of keeping the value monotonic it should be good
> enough. Will think about it.
> 

Yes, it could work. In this case you have to be aware that the 32 LSBs
of the TSC will overflow every ~ 1s and, in order to be sure to be able
to detect the overflow, you have to do at least 1 TSC read per second
(more precisely per 32 LSB overflow period).

> Thanks for the review.
> 

YW,

Mathieu

> -Andi
> 

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68