Date:   Fri, 16 Feb 2018 12:10:18 +0800
From:   Chen-Yu Tsai <wens@...e.org>
To:     Philipp Rossak <embed3d@...il.com>
Cc:     Maxime Ripard <maxime.ripard@...tlin.com>,
        Alessandro Zummo <a.zummo@...ertech.it>,
        alexandre.belloni@...tlin.com,
        linux-kernel <linux-kernel@...r.kernel.org>,
        linux-sunxi <linux-sunxi@...glegroups.com>,
        linux-rtc@...r.kernel.org
Subject: Re: [PATCH v3] rtc: ac100: Fix ac100 determine rate bug

On Fri, Feb 16, 2018 at 1:53 AM, Philipp Rossak <embed3d@...il.com> wrote:
>
>
> On 15.02.2018 15:11, Maxime Ripard wrote:
>>
>> On Wed, Feb 14, 2018 at 02:56:12PM +0100, Philipp Rossak wrote:
>>>
>>> This patch fixes a bug, that prevents the Allwinner A83T and the A80
>>> from a successful boot.
>>>
>>> The bug is there since v4.16-rc1 and appeared after the clk branch was
>>> merged.
>>
>>
>> Out of curiosity, which patch has introduced this? I couldn't find any
>> obvious match.
>>
>
> I wasn't also n

To be honest, I'm not sure why this is hitting you and not me.
I have both A83T boards that have assigned-clock-rates set for
the ac100 clock outputs for WiFi. I have them running 4.16-rc1
and have not seen this. The device tree patches that add these
are in 4.15.

>
>>> You can find the shortend trace below:
>>>
>>> Unable to handle kernel NULL pointer dereference at virtual address
>>> 00000000
>>> pgd = (ptrval)
>>> [00000000] *pgd=00000000
>>> Internal error: Oops: 5 [#1] SMP ARM
>>> Modules linked in:
>>> CPU: 0 PID: 49 Comm: kworker/0:1 Not tainted 4.15.0-10190-gb89e32ccd1be
>>> #2
>>> Hardware name: Allwinner sun8i Family
>>> Workqueue: events deferred_probe_work_func
>>> PC is at clk_hw_get_rate+0x0/0x34
>>> LR is at ac100_clkout_determine_rate+0x48/0x19c
>>>
>>> [ ... ]
>>>
>>> (clk_hw_get_rate) from (ac100_clkout_determine_rate+0x48/0x19c)
>>> (ac100_clkout_determine_rate) from  (clk_core_set_rate_nolock+0x3c/0x1a0)
>>> (clk_core_set_rate_nolock) from (clk_set_rate+0x30/0x88)
>>> (clk_set_rate) from (of_clk_set_defaults+0x200/0x364)
>>> (of_clk_set_defaults) from (platform_drv_probe+0x18/0xb0)
>>>
>>> To fix that bug, we first check if the return of the
>>> clk_hw_get_parent_by_index is non zero. If it is zero we skip that
>>> clock parent.
>>>
>>> The BUG report could be found here: https://lkml.org/lkml/2018/2/10/198
>>>
>>> Fixes: 04940631b8d2 ("rtc: ac100: Add clk output support")
>>>
>>> Signed-off-by: Philipp Rossak <embed3d@...il.com>
>>> ---
>>>
>>> Changes in v3:
>>>         * add information when the bug appeared
>>>         * make the comment more clear
>>> Changes in v2:
>>>         * add tag Fixes: ... to commit message
>>>         * add comment to if statement why we are doing this check
>>>
>>>   drivers/rtc/rtc-ac100.c | 19 ++++++++++++++++++-
>>>   1 file changed, 18 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/rtc/rtc-ac100.c b/drivers/rtc/rtc-ac100.c
>>> index 8ff9dc3fe5bf..2412aa2e8399 100644
>>> --- a/drivers/rtc/rtc-ac100.c
>>> +++ b/drivers/rtc/rtc-ac100.c
>>> @@ -183,7 +183,24 @@ static int ac100_clkout_determine_rate(struct clk_hw
>>> *hw,
>>>         for (i = 0; i < num_parents; i++) {
>>>                 struct clk_hw *parent = clk_hw_get_parent_by_index(hw,
>>> i);
>>> -               unsigned long tmp, prate = clk_hw_get_rate(parent);
>>> +               unsigned long tmp, prate;
>>> +
>>> +               /*
>>> +                * The clock has two parents, one is a fixed clock which
>>> is
>>> +                * internally registered by the ac100 driver. The other
>>> parent
>>> +                * is a clock from the codec side of the chip, which we
>>> +                * properly declare and reference in the devicetree and
>>> is
>>> +                * not implemented in any driver right now.
>>> +                * If the clock core looks for the parent of that second
>>> +                * missing clock, it can't one that is registered and
>>> +                * returns NULL.
>>> +                * Thus we need to check if the parent exists before
>>> +                * we get the parent rate.
>>> +                */
>>> +               if (!parent)
>>> +                       continue;
>>
>>
>> I'm sorry, but I still don't get it. When you register that clock, you
>> will give it two parents. Why would that change during the life of the
>> clock?
>>
>> This really looks like a workaround rather than an actual fix.
>>
>> Maxime
>>
> I agree this is more a workaround!
> A proper solution/fix would be to define the devicetree correct like this:
>
> diff --git a/arch/arm/boot/dts/sun8i-a83t-bananapi-m3.dts
> b/arch/arm/boot/dts/sun8i-a83t-bananapi-m3.dts
> index 6550bf0e594b..6f56d429f17e 100644
> --- a/arch/arm/boot/dts/sun8i-a83t-bananapi-m3.dts
> +++ b/arch/arm/boot/dts/sun8i-a83t-bananapi-m3.dts
> @@ -175,11 +175,18 @@
>                         compatible = "x-powers,ac100-rtc";
>                         interrupt-parent = <&r_intc>;
>                         interrupts = <0 IRQ_TYPE_LEVEL_LOW>;
> -                       clocks = <&ac100_codec>;
> +                       clocks = <&ac100_rtc_32k>;
>                         #clock-cells = <1>;
>                         clock-output-names = "cko1_rtc",
>                                              "cko2_rtc",
>                                              "cko3_rtc";
> +
> +                       ac100_rtc_32k: rtc-32k-oscillator {
> +                               compatible = "fixed-clock";
> +                               #clock-cells = <0>;
> +                               clock-frequency  = <32768>;
> +                               clock-output-names = "ac100-rtc-32k";
> +                       };
>                 };
>         };
>  };
>
> What do you think about that solution?

That's not quite right either. As I mentioned before, the
RTC block has two clock inputs, one 4MHz signal from the
codec block, and one 32.768 kHz signal from an external
crystal. The original device tree binding describes the
first one, and the 32.768 kHz clock was registered by
the RTC driver internally.

If you're going to add the crystal clock, you still need
to keep the codec one. Note that this does not fix what
Maxime is asking you. I've already provided an explanation:

The clock core allows registering clocks with not-yet-existing
clock parents. Parents are matches by string names. If no
clock by that name is registered yet, the clock core simply
orphans the new clock if the unregistered parent is its
current parent or simply ignores that parent if its not the
current parent. This is entirely valid and is what we are
counting on here, as we haven't implemented the codec-side
driver.

Note we also sort of depend on this behavior for sunxi-ng
clocks, where in some cases we get circular dependencies
between clock blocks.

BTW, since this is mostly related to the clk subsystem,
please CC the clk maintainers as well.

>
> I already tested it and it looks like it is working.

There is no point in testing things if it doesn't match
the hardware. Yes your code works and doesn't crash. No
it does not represent what the hardware is actually doing.

Regards
ChenYu

Powered by blists - more mailing lists