Re: [PATCH] x86 idle: repair large-server 50-watt idle-power regression

["H. Peter Anvin" <hpa@zytor.com>]

CLFLUSH is only (guaranteed to be) ordered with respect to MFENCE according to the SDM.

Ingo Molnar <mingo@kernel.org> wrote:
>
>* Peter Zijlstra <peterz@infradead.org> wrote:
>
>> On Thu, Dec 19, 2013 at 06:25:35PM +0100, Peter Zijlstra wrote:
>> > On Thu, Dec 19, 2013 at 06:07:41PM +0100, Ingo Molnar wrote:
>> > > Likewise, having a barrier before the MONITOR looks sensible as
>well. 
>> > 
>> > I again have to disagree, one would expect monitor to flush all
>that is
>> > required to start the monitor -- and it actually does so. As is
>> > testified by this extra CLFLUSH being called a bug workaround.
>> 
>> SDM states that MONITOR is ordered like a LOAD, and a LOAD cannot 
>> pass a previous STORE to the same address.
>
>Yes ... but you could argue that CLFLUSH is neither a load nor a 
>store, it's a _cache sync_ operation, with its special ordering 
>properties.
>
>> That said; there's enough holes in there to swim a titanic through, 
>> seeing how MONITOR stares at an entire cacheline and LOAD/STORE 
>> order is specified on location, whatever that means.
>
>I think assuming that MONITOR is ordered as a load or better is a 
>pretty safe one (and in fact the Intel documentation seems to say so) 
>- I'd say MONITOR is in micro-code and essentially snoops on cache 
>events on that specific cache line, and loads the cache line on a 
>snoop hit?
>
>Btw., what state is the cache line after a MONITOR instruction, is it 
>loaded as shared, or as excusive? (exclusive would probably be better 
>for performance.)
>
>Thanks,
>
>	Ingo

-- 
Sent from my mobile phone.  Please pardon brevity and lack of formatting.