Re: [PATCH] memory: samsung: exynos5422-dmc: propagate error from exynos5_counters_get()

[Lukasz Luba]

On 8/4/20 1:19 PM, Marek Szyprowski wrote:
> Hi Lukasz,
>
> On 04.08.2020 11:11, Lukasz Luba wrote:
> > Hi Marek,
> >
> > On 8/4/20 7:12 AM, Marek Szyprowski wrote:
> > > exynos5_counters_get() might fail with -EPROBE_DEFER if the driver for
> > > devfreq event counter is not yet probed. Propagate that error value to
> > > the caller to ensure that the exynos5422-dmc driver will be probed again
> > > when devfreq event contuner is available.
> > >
> > > This fixes boot hang if both exynos5422-dmc and exynos-ppmu drivers are
> > > compiled as modules.
> > >
> > > Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
> > > ---
> > >    drivers/memory/samsung/exynos5422-dmc.c | 2 +-
> > >    1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/memory/samsung/exynos5422-dmc.c
> > > b/drivers/memory/samsung/exynos5422-dmc.c
> > > index b9c7956e5031..639811a3eecb 100644
> > > --- a/drivers/memory/samsung/exynos5422-dmc.c
> > > +++ b/drivers/memory/samsung/exynos5422-dmc.c
> > > @@ -914,7 +914,7 @@ static int exynos5_dmc_get_status(struct device
> > > *dev,
> > >        } else {
> > >            ret = exynos5_counters_get(dmc, &load, &total);
> > >            if (ret < 0)
> > > -            return -EINVAL;
> > > +            return ret;
> > >              /* To protect from overflow, divide by 1024 */
> > >            stat->busy_time = load >> 10;
> >
> > Thank you for the patch, LGTM.
> > Some questions are still there, though. The function
> > exynos5_performance_counters_init() should capture that the counters
> > couldn't be enabled or set. So the functions:
> > exynos5_counters_enable_edev() and exynos5_counters_set_event()
> > must pass gently because devfreq device is registered.
> > Then devfreq checks device status, and reaches the state when
> > counters 'get' function returns that they are not ready...
> >
> > If that is a kind of 2-stage initialization, maybe we should add
> > another 'check' in the exynos5_performance_counters_init() and call
> > the devfreq_event_get_event() to make sure that we are ready to go,
> > otherwise return ret from that function (which is probably EPROBE_DEFER)
> > and not register the devfreq device.
>
> I've finally investigated this further and it turned out that the issue
> is elsewhere. The $subject patch can be discarded, as it doesn't fix
> anything. The -EPROBE_DEFER is properly returned by
> exynos5_performance_counters_init(), which redirects exynos5_dmc_probe()
> to remove_clocks label. This causes disabling mout_bpll/fout_bpll clocks
> what in turn *sometimes* causes boot hang. This random behavior mislead
> me that the $subject patch fixes the issue, but then longer tests
> revealed that it didn't change anything.

Really good investigation, great work Marek!

> It looks that the proper fix would be to keep fout_bpll enabled all the
> time.

Yes, I agree. I am looking for your next patch to test it then.

> > Marek do want to submit such patch or I should bake it and submit on top
> > of this patch?
> >
> > Could you tell me how I can reproduce this? Do you simply load one
> > module after another (exynos-ppmu than exynos5422-dmc) or in parallel?
>
> I've just boot zImage built from multi_v7_defconfig with modules
> installed. Modules are automatically loaded by udev during boot.

Thank you sharing this test procedure.

Regards,
Lukasz