Re: [PATCH V2 3/8] soc: qcom-geni-se: Add interconnect support to fix earlycon crash

[Akash Asthana <akashast@codeaurora.org>]

Hi Matthias,

On 3/17/2020 11:59 PM, Matthias Kaehlcke wrote:
> Hi Akash,
>
> On Tue, Mar 17, 2020 at 04:27:47PM +0530, Akash Asthana wrote:
>> Hi Matthias,
>>
>> On 3/14/2020 2:14 AM, Matthias Kaehlcke wrote:
>>> Hi Akash,
>>>
>>> On Fri, Mar 13, 2020 at 06:42:09PM +0530, Akash Asthana wrote:
>>>> V1 patch@https://patchwork.kernel.org/patch/11386469/ caused SC7180 system
>>>> to reset at boot time.
>>> The v1 patch isn't relevant in the commit message, please just describe the
>>> problem. Also the crash only occurs when earlycon is used.
>> ok
>>>> As QUP core clock is shared among all the SE drivers present on particular
>>>> QUP wrapper, the reset seen is due to earlycon usage after QUP core clock
>>>> is put to 0 from other SE drivers before real console comes up.
>>>>
>>>> As earlycon can't vote for it's QUP core need, to fix this add ICC
>>>> support to common/QUP wrapper driver and put vote for QUP core from
>>>> probe on behalf of earlycon and remove vote during sys suspend.
>>> Only removing the vote on suspend isn't ideal, the system might never get
>>> suspended. That said I don't have a really good alternative suggestion.
>>>
>>> One thing you could possibly do is to launch a delayed work, check
>>> console_device() every second or so and remove the vote when it returns
>>> non-NULL. Not claiming this would be a great solution ...
>>>
>>> The cleanest solution might be a notifier when the early console is
>>> unregistered, it seems somewhat over-engineered though ... Then again
>>> other (future) uart drivers with interconnect support might run into
>>> the same problem.
>> We are hitting this problem because QUP core clocks are shared among all the
>> SE driver present in particular QUP wrapper, if other HW controllers has
>> similar architecture we will hit this issue.
>>
>> How about if we expose an API from common driver(geni-se) for putting QUP
>> core BW vote to 0.
>>
>> We call this from console probe just after uart_add_one_port call (console
>> resources are enabled as part of this call) to put core quota to 0 on behalf
>> of earlyconsole?
>  From my notes from earlier debugging I have doubts this would work:
>
>    There is a short window where the early console and the 'real' console coexist:
>
>    [    3.858122] printk: console [ttyMSM0] enabled
>    [    3.875692] printk: bootconsole [qcom_geni0] disabled
>
>    The reset probably occurs when the early console tries to write, but the ICC
>    is effectively disabled because ttyMSM0 and the other geni ports are runtime
>    suspended.

Code flow from console driver probe(qcom_geni_serial.c)

uart_add_one_port--->uart_configure_port--->{ 1) uart_change_pm(enable 
console resources)  2)register_console(boot to real console switch 
happens here)}

Console resources are not disabled from anywhere before the switch 
happens completely. I meant to say until we saw below logs.

[    3.875692] printk: bootconsole [qcom_geni0] disabled

I think the board reset issue cannot occur during the window where early 
console and 'real' console coexist.

I have validated proposed solution by me, it is working fine.

Currently voting is done for every QUP and not only to which earlycon is 
connect, with the above approach we can't remove vote from other QUPs.

However we can limit voting only to earlycon QUP by removing 
interconnect from DT node of other QUPs.

I am not sure how clean is this solution.

>
>>>> Signed-off-by: Akash Asthana <akashast@codeaurora.org>
>>>> Reported-by: Matthias Kaehlcke <mka@chromium.org>
>>>> ---
>>>>    drivers/soc/qcom/qcom-geni-se.c | 41 +++++++++++++++++++++++++++++++++++++++++
>>>>    1 file changed, 41 insertions(+)
>>>>
>>>> diff --git a/drivers/soc/qcom/qcom-geni-se.c b/drivers/soc/qcom/qcom-geni-se.c
>>>> index 7d622ea..d244dfc 100644
>>>> --- a/drivers/soc/qcom/qcom-geni-se.c
>>>> +++ b/drivers/soc/qcom/qcom-geni-se.c
>>>> @@ -90,6 +90,7 @@ struct geni_wrapper {
>>>>    	struct device *dev;
>>>>    	void __iomem *base;
>>>>    	struct clk_bulk_data ahb_clks[NUM_AHB_CLKS];
>>>> +	struct icc_path *icc_path_geni_to_core;
>>>>    };
>>>>    #define QUP_HW_VER_REG			0x4
>>>> @@ -747,11 +748,50 @@ static int geni_se_probe(struct platform_device *pdev)
>>>>    		}
>>>>    	}
>>>> +#ifdef CONFIG_SERIAL_EARLYCON
>>>> +	wrapper->icc_path_geni_to_core = devm_of_icc_get(dev, "qup-core");
>>>> +	if (IS_ERR(wrapper->icc_path_geni_to_core))
>>>> +		return PTR_ERR(wrapper->icc_path_geni_to_core);
>>>> +	/*
>>>> +	 * Put minmal BW request on core clocks on behalf of early console.
>>>> +	 * The vote will be removed in suspend call.
>>>> +	 */
>>>> +	ret = icc_set_bw(wrapper->icc_path_geni_to_core, Bps_to_icc(1000),
>>>> +			Bps_to_icc(1000));
>>>> +	if (ret) {
>>>> +		dev_err(&pdev->dev, "%s: ICC BW voting failed for core\n",
>>>> +			__func__);
>>>> +		return ret;
>>>> +	}
>>> What is ugly about this is that it's done for every QUP, not only the one
>>> with the early console. Again, I don't have a good solution for it, maybe
>>> it's a limitation we have to live with :(
>> There is one more limitation from QUP core side. Core clocks for both the
>> QUP wrapper runs at same speed.
>>
>> core2x_1 = core2x_2 = max(core2x_1, core2x_2);
>>
>> So with above limitation and if we are removing early con vote from Core
>> when real console comes up. It doesn't matter whether it's done for every
>> QUP or the only with early console.
> it's still sorta ugly at an abstraction level, but it seem we have to be
> pragmatic here.

How about if we limit voting only to earlycon QUP by removing 
interconnect from DT node of other QUPs.?

>
>>>> +#endif
>>>> +
>>>>    	dev_set_drvdata(dev, wrapper);
>>>>    	dev_dbg(dev, "GENI SE Driver probed\n");
>>>>    	return devm_of_platform_populate(dev);
>>>>    }
>>>> +static int __maybe_unused geni_se_sys_suspend(struct device *dev)
>>>> +{
>>>> +	struct geni_wrapper *wrapper = dev_get_drvdata(dev);
>>>> +	int ret;
>>>> +
>>>> +#ifdef CONFIG_SERIAL_EARLYCON
>>>> +	ret = icc_set_bw(wrapper->icc_path_geni_to_core, 0, 0);
>>> I think you only want to do this on the first suspend.
>> Ok, I can add that logic using global variable.
>>> Do we need to handle the case where no 'real' console is configured?
>>> In this case the early console would be active forever and setting
>>> the bandwidths to 0 might cause a similar crash than the one you are
>>> trying to fix. Not sure if that's a real world use case, but wanted to
>>> mention it. Maybe this is an argument of the notifier approach?
>> We can't support earlycon without real console.
>>
>> As earlyconsole doesn't do any kind of resource enablement(SE clocks,
>> pinctrl, etc) it assumes that resources are already enabled from previous
>> stages.
>>
>> So if real console doesn't come up no one will vote for that SE clock, and
>> it will be disabled from clk late_init call which will result into
>> un-clocked access.
> Ok, IIUC what you are saying is that earlycon can't work on its own after geni
> initialization. Because it clearly can work before (otherwise what would be
> its purpose?), supposedly because the bootloader configures the necessary
> bits.

If there is no real console then earlyconsole should work until clk 
driver disables unvoted SE clocks.

Yes, early console depends on bootloader to enable neccessary resources 
for it. I guess currently Coreboot console driver is doing that.

>
> In any case the bottom line is that earlycon requires a real console to be
> configured and there is no need to handle the corner case I brought up.

Yes that is correct. We can't support earlyconsole for long without 
having a real console.


Thanks,

Akash

-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project