Re: [PATCH v5 3/3] PM/runtime:Replace jiffies based accounting with ktime based accounting

[Rafael J. Wysocki]

On Fri, Jan 18, 2019 at 11:53 AM Vincent Guittot
<vincent.guittot@xxxxxxxxxx> wrote:
>
> On Fri, 18 Jan 2019 at 11:42, Vincent Guittot
> <vincent.guittot@xxxxxxxxxx> wrote:
> >
> > Hi Guenter,
> >
> > Le Thursday 17 Jan 2019 Ã 14:16:28 (-0800), Guenter Roeck a Ãcrit :
> > > On Fri, Dec 21, 2018 at 11:33:56AM +0100, Vincent Guittot wrote:
> > > > From: Thara Gopinath <thara.gopinath@xxxxxxxxxx>
> > > >
> > > > This patch replaces jiffies based accounting for runtime_active_time
> > > > and runtime_suspended_time with ktime base accounting. This makes the
> > > > runtime debug counters inline with genpd and other pm subsytems which
> > > > uses ktime based accounting.
> > > >
> > > > timekeeping is initialized before pm_runtime_init() so ktime_get() will
> > > > be ready before first call. In fact, timekeeping_init() is called early
> > > > in start_kernel() which is way before driver_init() (and that's when
> > > > devices can start to be initialized) called from rest_init() via
> > > > kernel_init_freeable() and do_basic_setup().
> > > >
> > > This is not (always) correct. My qemu "collie" boot test fails with this
> > > patch applied. Reverting the patch fixes the problem. Bisect log attached.
> > >
> >
> > Can you try the patch below ?
> > ktime_get_mono_fast_ns() has the advantage of being init with dummy clock so
> > it can be used at early_init.
>
> Another possibility would be delay the init of the gpiochip

Well, right.

Initializing devices before timekeeping doesn't feel particularly
robust from the design perspective.

How exactly does that happen?