Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 13 Mar 2015 13:58:37 -0400
From:      John Baldwin <jhb@freebsd.org>
To:        Nathan Whitehorn <nwhitehorn@freebsd.org>
Cc:        'Andriy Gapon' <avg@freebsd.org>, freebsd-arch@freebsd.org
Subject:   Re: RFC: Simplfying hyperthreading distinctions
Message-ID:  <2058936.CdnVTM8mMb@ralph.baldwin.cx>
In-Reply-To: <54FA5EE9.4090305@freebsd.org>
References:  <1640664.8z9mx3EOQs@ralph.baldwin.cx> <54FA5EE9.4090305@freebsd.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Friday, March 06, 2015 06:14:01 PM Nathan Whitehorn wrote:
> 
> On 03/06/15 12:44, John Baldwin wrote:
> > Currently we go out of our way a bit to distinguish Pentium4-era
> > hyperthreading from more recent ("modern") hyperthreading.  I suspect that
> > this distinction probably results in confusion more than anything else.
> > Intel's documentation does not make near as broad a distinction as far as I
> > can tell.  Both types of SMT are called hyperthreading in the SDM for example.
> > However, we have the astonishing behavior that
> > 'machdep.hyperthreading_allowed' only affects "old" hyperthreads, but not
> > "new" ones.  We also try to be overly cute in our dmesg output by using HTT
> > for "old" hyperthreading, and SMT for "new" hyperthreading.  I propose the
> > following changes to simplify things a bit:
> >
> >    1) Call both "old" and "new" hyperthreading HTT in dmesg.
> >
> >    2) Change machdep.hyperthreading_allowed to apply to both new and old HTT.
> >       However, doing this means a POLA violation in that we would now disable
> >       modern HTT by default.  Balanced against re-enabling "old" HTT by default
> >       on an increasingly-shrinking pool of old hardware, I think the better
> >       approach here would be to also change the default to allow HTT.
> >
> >    3) Possibly add a different knob (or change the behavior of
> >       machdep.hyperthreading_allowed) to still bring up hyperthreads, but leave
> >       them out of the default cpuset (set 1).  This would allow those threads
> >       to be re-enabled dynamically at runtime by adjusting the mask on set 1.
> >       The original htt settings back when 'hyperthreading_allowed' was
> >       introduced actually permitted this via by adjusting 'machdep.hlt_cpus' at
> >       runtime.
> >
> > What do people think?
> 
> I'm fine with whatever naming, but if we're making new sysctls, 
> especially for the cpuset case, is there a reason to hide the behavior 
> under machdep? We support at least three non-x86 CPUs with SMT (POWER8, 
> Cell, and POWER5) and the relevant scheduling logic should be MI. At 
> least POWER8 supports 8 threads per core, so you might also want more 
> granularity than just "on" or "off".

3) would involve something new, yes, but my immediate concern is more
1) and 2) to make x86 more consistent.

I would certainly be fine with having an MI name for 3).  When I've
prototyped it before I did it in an MD SYSINIT, but I can perhaps make
use of the topology flags to do this in MI code instead.

-- 
John Baldwin



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?2058936.CdnVTM8mMb>