Re: [uml-devel] [patch 02/12] uml: cpu_relax fix

[Blaisorblade <blaisorblade@yahoo.it>]

On Wednesday 23 March 2005 18:09, Bodo Stroesser wrote:
> blaisorblade@yahoo.it wrote:
> > Use rep_nop instead of barrier for cpu_relax, following $(SUBARCH)'s
> > doing that (i.e. i386 and x86_64).
>
> IIRC, Jeff had the idea, to use sched_yield() for this (from a discussion
> on #uml).
Hmm, makes sense, but this is to benchmark well... I remember from early 
discussions on 2.6 scheduler that using sched_yield might decrease 
performance (IIRC starve the calling application).

Also, that call should be put inside the idle loop, not for cpu_relax, which 
is very different, since it is used (for instance) in kernel/spinlock.c for 
spinlocks, and in such things. The "Pause" opcode is explicitly recommended 
(by Intel manuals, I don't recall why) for things like spinlock loops, and 
using yield there would be bad.

> S390 does something similar using a special DIAG-opcode that 
> gives permission to zVM, that another Guest might run.

> On a host running many UMLs, this might improve performance.
>
> So, I would like to have the small patch below (it's not tested, just an
> idea).

-- 
Paolo Giarrusso, aka Blaisorblade
Linux registered user n. 292729
http://www.user-mode-linux.org/~blaisorblade