Re: Question about sched_setaffinity()

["Paul E. McKenney" <paulmck@linux.ibm.com>]

On Fri, May 10, 2019 at 02:08:19PM +0200, Peter Zijlstra wrote:
> On Thu, May 09, 2019 at 12:36:25PM -0700, Paul E. McKenney wrote:
> > I forward-ported the relevant patches from -rcu and placed them on -rcu
> > branch peterz.2019.05.09a, and this is what produced the output above.
> > 
> > Any other debugging thoughts?
> > 
> > Or, if you wish, you can reproduce by running the following:
> > 
> > nohup tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 8 --duration 2 --configs "TRIVIAL" --bootargs "trace_event=sched:sched_switch,sched:sched_wakeup ftrace=function_graph ftrace_graph_filter=sched_setaffinity,migration_cpu_stop" --kconfig "CONFIG_FUNCTION_TRACER=y CONFIG_FUNCTION_GRAPH_TRACER=y"
> > 
> > This gets me the following summary output:
> > 
> >  --- Thu May  9 12:08:31 PDT 2019 Test summary:
> >  Results directory: /home/git/linux-2.6-tip/tools/testing/selftests/rcutorture/res/2019.05.09-12:08:31
> >  tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 8 --duration 2 --configs TRIVIAL --bootargs trace_event=sched:sched_switch,sched:sched_wakeup ftrace=function_graph ftrace_graph_filter=sched_setaffinity,migration_cpu_stop --kconfig CONFIG_FUNCTION_TRACER=y CONFIG_FUNCTION_GRAPH_TRACER=y
> >  TRIVIAL ------- 2177 GPs (18.1417/s) [trivial: g0 f0x0 ]
> >  :CONFIG_HOTPLUG_CPU: improperly set
> >  WARNING: BAD SEQ 2176:2176 last:2177 version 4
> >      /home/git/linux-2.6-tip/tools/testing/selftests/rcutorture/res/2019.05.09-12:08:31/TRIVIAL/console.log
> >      WARNING: Assertion failure in /home/git/linux-2.6-tip/tools/testing/selftests/rcutorture/res/2019.05.09-12:08:31/TRIVIAL/console.log
> >      WARNING: Summary: Warnings: 1 Bugs: 1 Call Traces: 5 Stalls: 8
> 
> So I could reproduce...
> 
> I must first complain about your scripts; it does "make mrproper" on the
> source tree every time you run it, this is not appreciated. For one, it
> deletes my 'tags' file.

This is because it builds in a different directory, and "make O=/path"
complains if you don't have the source directory pristine.

But there really is no longer any reason to build in a different
directory, I suppose.  This is a largish change, but working on it.

> Getting it to not rebuild the whole kernel every time wasn't easy
> either.

You trust "make" far more than I do!  I am thinking of adding a
"--trust-make" argument that suppresses the "make clean".  Maybe if
I grow to trust "make" in the fulness of time, I can remove the "make
clean" entirely.  But given ccache, and given the duration of the typical
rcutorture run, and given that there are multiple rcutorture scenarios
each with a different .config, this hasn't been a priority.  The build
step is already omitted for repeated runs.

> Aside from that it seems to 'work'.
> 
> The below trace explain the issue. Some Paul person did it, see below.
> It's broken per construction :-)

*facepalm*  Hence the very strange ->cpus_allowed mask.  I really
should have figured that one out.

The fix is straightforward.  I just added "rcutorture.shuffle_interval=0"
to the TRIVIAL.boot file, which stops rcutorture from shuffling its
kthreads around.

Please accept my apologies for the hassle, and thank you for tracking
this down!!!

							Thanx, Paul