atom feed4 messages in org.freebsd.freebsd-bugsRe: kern/145385: [cpu] Logical proces...
FromSent OnAttachments
Garrett CooperNov 17, 2010 4:39 pm 
Garrett CooperNov 18, 2010 4:09 pm 
Garrett CooperNov 18, 2010 7:09 pm 
Andriy GaponNov 19, 2010 1:09 am 
Subject:Re: kern/145385: [cpu] Logical processor cannot be disabled for some SMT-enabled Intel procs
From:Garrett Cooper (gcoo@freebsd.org)
Date:Nov 18, 2010 4:09:44 pm
List:org.freebsd.freebsd-bugs

The following reply was made to PR kern/145385; it has been noted by GNATS.

From: Garrett Cooper <gcoo@freebsd.org> To: Andriy Gapon <av@freebsd.org> Cc: bug-@freebsd.org Subject: Re: kern/145385: [cpu] Logical processor cannot be disabled for some SMT-enabled Intel procs Date: Thu, 18 Nov 2010 16:09:46 -0800

On Wed, Nov 17, 2010 at 4:34 PM, Garrett Cooper <gcoo@freebsd.org> wrote= :

On Thu, Oct 7, 2010 at 10:03 AM, Andriy Gapon <av@freebsd.org> wrote:

Here's a patch to remove halted logical processors from root cpu set (cp= uset number zero) and consequently all child sets: http://people.freebsd.org/~avg/cpu-halt.diff

Please note that unhalting CPUs will add them to cpuset zero, but will n= ot (re-)add them cpusets of the running processes. =A0So additional action = will be required from system administrator if e would like existing processes to= make use of unhalted CPUs.

Also, interrupts that are bound to halted CPUs are not rebound on halt, = and so will be delivered to halted CPUs. =A0This should not be a problem for ty= pical environments, but would be nice to fix anyway.

=A0 =A0Sorry for taking so long to get back on this item. The patch looks ok as far as setting the CPUSET is concerned (setting machdep.hlt_logical_cpus=3D1 works on an r710 with 8 logical procs as SCHED_ULE schedules all of the tasks on the non-logical SMT cores, not the logical ones, and the patch properly detects that there aren't any logical processors on a Core2 Quad I have at home), but it's missing a key case where I go from...

machdep.hlt_logical_cpus: 1 -> 0

=A0 =A0... it should reset CPUSETZERO. =A0 =A0The overall CPU topology should probably be cached and restored when reset, or at least CPUSETZERO should be reset (probably the simpler and cleaner route to go). =A0 =A0After that all that's required is work for i386 and I'd vote for a commit of the patch. Thanks for the hard work Andriy :)!

Just to illustrate, here's how I reproduced the successful behavior (and the problem):

1. Execute the following script:

#!/bin/sh i=3D0 ncpus=3D`sysctl -n kern.smp.cpus` while [ $i -lt $ncpus ] ; do while : ; do : ; done & : $(( i +=3D 1 )) done

2. Execute sysctl machdep.hlt_logical_cpus=3D1 . 3. Look at cpuset info; example:

Before:

# for i in `ps aux | grep gcooper | grep -v grep | awk '{ print $2 }'` ; do cpuset -g -p $i ; done pid 2179 mask: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 pid 2180 mask: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15

After:

# for i in `ps aux | grep gcooper | grep -v grep | awk '{ print $2 }'` ; do cpuset -g -p $i ; done pid 2179 mask: 0, 2, 4, 6, 8, 10, 12, 14 pid 2180 mask: 0, 2, 4, 6, 8, 10, 12, 14 # sysctl machdep.hlt_logical_cpus=3D0 machdep.hlt_logical_cpus: 1 -> 0 # for i in `ps aux | grep gcooper | grep -v grep | awk '{ print $2 }'` ; do cpuset -g -p $i ; done pid 2179 mask: 0, 2, 4, 6, 8, 10, 12, 14 pid 2180 mask: 0, 2, 4, 6, 8, 10, 12, 14

I'm taking the patch and running with it now to try and get it working 100%= .

_______________________________________________ free@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-bugs To unsubscribe, send any mail to "free@freebsd.org"