lkml.org 
[lkml]   [2008]   [Sep]   [26]   [last100]   RSS Feed
Views: [more markup]  [less markup]  [headers]  [forward] 
 
Messages in this thread
/
DateFri, 26 Sep 2008 09:02:17 +0200
FromIngo Molnar <>
SubjectRe: [PATCH] x86: TSC resync
* Michael Davidson <md@google.com> wrote:

> This patch is a slightly cleaned up version of one which has been in 
> use at for some time now at Google.
> 
> It uses an HPET based time source to resynchronize the TSC on systems 
> where it would otherwise be unsynchronized - eg early AMD Opteron 
> based systems where the TSC rate drifts when going in and out of the 
> C1E halt state.
> 
> While the approach is quite crude it has been effective for systems 
> where user space code relies on the TSC advancing at a constant rate 
> and being reasonably well synchronized between CPUs. The skew between 
> TSC's on different processors as seen from user space is typically 
> less than +/- 1000 clock cycles which has proved to be sufficient for 
> the applications that we care about.

hm, this patch syncs the TSCs every 20 seconds. That is enough to sync 
up AMD CPUs where the TSC slows down _slightly_ (at 10 ppm per second or 
so) when it's in HLT.

How does it behave in face of 'TSC stops' systems - systems with C2/C3 
sleeps? Basically all modern CPUs that save power are affected by that: 
the TSC get brutally cut when idle - almost all modern power saving 
laptop, desktop and server CPUs.

Also, what does it do in face of cpufreq-affected TSCs? That too is a 
large category of systems. (but most currently shipping CPUs fortunately 
have a cpufreq-invariant TSC already.)

> I don't expect this patch to be of much general interest, but if you 
> happen to be unlucky enough to have a system where the TSC is not 
> synchronized across CPUs and user space code which relies on the 
> assumption that it is, then this patch may be useful.
> 
> Signed-off-by: Michael Davidson <md@google.com>

This actually looks pretty interesting IMO, and the code looks simple, 
clean and straightforward enough - but it might not be enough to be a 
generic solution.

	Ingo


\
 
 \ /
  Last update: 2008-09-26 09:05    [from the cache]
©2003-2008