lkml.org 
[lkml]   [2013]   [Dec]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] x86 idle: repair large-server 50-watt idle-power regression

* H. Peter Anvin <hpa@zytor.com> wrote:

> On 12/19/2013 09:36 AM, Peter Zijlstra wrote:
> > On Thu, Dec 19, 2013 at 06:25:35PM +0100, Peter Zijlstra wrote:
> >> That said, I would find it very strange indeed if a CLFLUSH doesn't also
> >> flush the store buffer.
> >
> > OK, it explicitly states it does not do that and you indeed need
> > an mfence before the clflush.
>
> So, MONITOR is defined to be ordered as a load, which I think should
> be adequate, but I really wonder if we should have mfence on both
> sides of clflush. This now is up to 9 bytes, and perhaps pushing it
> a bit with how much we would be willing to patch out.
>
> On the other hand - the CLFLUSH seems to have worked well enough by
> itself, and this is all probabilistic anyway, so perhaps we should
> just leave the naked CLFLUSH in and not worry about it unless
> measurements say otherwise?

So I think the window of breakage was rather large here, and since it
seems to trigger on rare types of hardware I think we'd be better off
by erring on the side of robustness this time around ...

This is the 'go to idle' path, which isn't as time-critical as the
'get out of idle' code path.

Thanks,

Ingo


\
 
 \ /
  Last update: 2013-12-19 20:01    [W:0.301 / U:0.104 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site