[Intel-gfx] [PATCH] [RFC] kernel/cpu: Use lockref for online CPU reference counting
joonas.lahtinen at linux.intel.com
Tue Feb 16 08:49:36 UTC 2016
On ma, 2016-02-15 at 18:06 +0100, Peter Zijlstra wrote:
> On Mon, Feb 15, 2016 at 03:17:55PM +0100, Peter Zijlstra wrote:
> > On Mon, Feb 15, 2016 at 02:36:43PM +0200, Joonas Lahtinen wrote:
> > > Instead of implementing a custom locked reference counting, use lockref.
> > >
> > > Current implementation leads to a deadlock splat on Intel SKL platforms
> > > when lockdep debugging is enabled.
> > >
> > > This is due to few of CPUfreq drivers (including Intel P-state) having this;
> > > policy->rwsem is locked during driver initialization and the functions called
> > > during init that actually apply CPU limits use get_online_cpus (because they
> > > have other calling paths too), which will briefly lock cpu_hotplug.lock to
> > > increase cpu_hotplug.refcount.
> > >
> > > On later calling path, when doing a suspend, when cpu_hotplug_begin() is called
> > > in disable_nonboot_cpus(), callbacks to CPUfreq functions get called after,
> > > which will lock policy->rwsem and cpu_hotplug.lock is already held by
> > > cpu_hotplug_begin() and we do have a potential deadlock scenario reported by
> > > our CI system (though it is a very unlikely one). See the Bugzilla link for more
> > > details.
> > I've been meaning to change the thing into a percpu-rwsem, I just
> > haven't had time to look into the lockdep splat that generated.
> The below has plenty lockdep issues because percpu-rwsem is
> reader-writer fair (like the regular rwsem), so it does throw up a fair
> number of very icky issues.
I originally thought of implementing this more similar to what you
specify, but then I came across a discussion in the mailing list where
it was NAKed adding more members to task_struct;
Adding proper recursion (the way my initial implementation was going)
got ugly without modifying task_struct because get_online_cpus() is a
speed critical code path.
So I'm all for fixing the current code in a different way if that will
then be merged.
> If at all possible, I'd really rather fix those and have a 'saner'
> hotplug lock, rather than muddle on with open-coded horror lock we have
Open Source Technology Center
More information about the Intel-gfx