RE: [PATCH] idle using PNI monitor/mwait

[Nakajima, Jun]

That's right. If we have a lot of high-contention locks in the kernel,
we need to fix the code first, to get benefits for the other
architectures. 

"mwait" granularity (64-byte, for example) is given by the cpuid
instruction, and we did not use it because 1) it's unlikely that the
other fields of the task structure are modified when it's idle, 2) the
processor needs to check the flag after mwait anyway, to avoid waking up
with a false signal caused by other break events (i.e. mwait is a hint).

Jun

> -----Original Message-----
> From: Linus Torvalds [mailto:[e-mail]]
> Sent: Tuesday, July 08, 2003 4:34 PM
> To: Nakajima, Jun
> Cc: [e-mail]; Saxena, Sunil; Mallick, Asit K;
> Pallipadi, Venkatesh
> Subject: Re: [PATCH] idle using PNI monitor/mwait
> 
> 
> On Tue, 8 Jul 2003, Nakajima, Jun wrote:
> >
> > Attached is a patch that enables PNI (Prescott New Instructions)
> > monitor/mwait in kernel idle (opcodes are now public). Basically
MWAIT
> > is similar to hlt, but you can avoid IPI to wake up the processor
> > waiting. A write (by another processor) to the address range
specified
> > by MONITOR would wake up the processor waiting on MWAIT.
> 
> How about spinlocks? Does it make sense to make the contention code
use
> mwait too, or are the latencies too high? Not that we have a lot of
> high-contention locks any more, so maybe it doesn't much matter.
> 
> Also, wasn't there some flag to set the "mwait" granularity? I don't
see
> anything like that in the patch..
> 
> 		Linus