On Monday 10 Sep 2018 at 10:55:43 (+0200), Rafael J. Wysocki wrote:
> On Mon, Sep 10, 2018 at 10:24 AM Quentin Perret <quentin.per...@arm.com>
> wrote:
> >
> > Hi Rafael,
> >
> > On Sunday 09 Sep 2018 at 22:13:52 (+0200), Rafael J. Wysocki wrote:
> > > On Fri, Sep 7, 2018 at 5:29 PM Quentin Perret <quentin.per...@arm.com>
> > > wrote:
> > > > On Friday 07 Sep 2018 at 10:52:01 (+0200), Rafael J. Wysocki wrote:
> > > > > Well, why don't you implement it as something like "if the governor
> > > > > changes
> > > > > from sugov to something else (or the other way around), call this
> > > > > function
> > > > > from the scheduler"?
> > > >
> > > > I just gave it a try and ended up with the diff below. It's basically
> > > > the exact same patch with a direct function call instead of a notifier.
> > > > (I also tried the sugov_start/stop thing I keep mentioning but it is
> > > > more complex, so let's see if the simplest solution could work first).
> > > >
> > > > What do you think ?
> > >
> > > This generally works for me from the cpufreq perspective, but I would
> > > add "cpufreq" to the name of the new function, that is call it
> > > something like sched_cpufreq_governor_change().
> >
> > Ok, no problem.
> >
> > > Also do you really need the extra work item? Governor changes are
> > > carried out in process context anyway.
> >
> > Ah, good point, I can remove that. I just tried and got the following
> > lock warning on boot, though:
> >
> > [ 2.518684] ============================================
> > [ 2.523942] WARNING: possible recursive locking detected
> > [ 2.529200] 4.18.0-rc6-00086-g940e7a9fd5ec #10 Not tainted
> > [ 2.534630] --------------------------------------------
> > [ 2.539888] kworker/2:3/1349 is trying to acquire lock:
> > [ 2.545059] (____ptrval____) (cpu_hotplug_lock.rw_sem){++++}, at:
> > rebuild_sched_domains_locked+0x2c/0x598
> > [ 2.554559]
> > [ 2.554559] but task is already holding lock:
> > [ 2.560332] (____ptrval____) (cpu_hotplug_lock.rw_sem){++++}, at:
> > cpufreq_register_driver+0x80/0x1d0
> > [ 2.569396]
> > [ 2.569396] other info that might help us debug this:
> > [ 2.575858] Possible unsafe locking scenario:
> > [ 2.575858]
> > [ 2.581717] CPU0
> > [ 2.584135] ----
> > [ 2.586553] lock(cpu_hotplug_lock.rw_sem);
> > [ 2.590785] lock(cpu_hotplug_lock.rw_sem);
> > [ 2.595017]
> > [ 2.595017] *** DEADLOCK ***
> > [ 2.595017]
> > [ 2.600877] May be due to missing lock nesting notation
> >
> > That seems to happen because cpufreq_register_driver() calls
> > cpus_read_lock(), which is then called again by rebuild_sched_domains()
> > down the line. So it might just be a missing lock nesting notation as
> > the warning suggests ?
> >
> > I'll have a look.
>
> It only is nested in the _register_driver() code path, otherwise it may not
> be.
Right.
> Using the work item may be the most straightforward way to deal with
> that, but then I would add a comment to explain what's up.
Indeed, rw_sems don't seem to be appropriate for nested locking:
https://elixir.bootlin.com/linux/latest/source/include/linux/rwsem.h#L156
I'll stick a comment explaining that for now, unless I find a better
idea than using a work item in the meantime.
Thanks,
Quentin
