Re: [PATCH v10 5/5] usb: dwc3: add rockchip specific glue layer

[Felipe Balbi]

Hi,

"William.wu" <William.wu@xxxxxxxxxxxxxx> writes:
>> William Wu <william.wu@xxxxxxxxxxxxxx> writes:
>>> Add rockchip specific glue layer to support USB3 Peripheral mode
>>> and Host mode on rockchip platforms (e.g. rk3399).
>>>
>>> The DesignWare USB3 integrated in rockchip SoCs is a configurable
>>> IP Core which can be instantiated as Dual-Role Device (DRD), Host
>>> Only (XHCI) and Peripheral Only configurations.
>>>
>>> We use extcon notifier to manage usb cable detection and mode switch.
>>> Enable DWC3 PM runtime auto suspend to allow core enter runtime_suspend
>>> if USB cable is dettached. For host mode, it requires to keep whole
>>> DWC3 controller in reset state to hold pipe power state in P2 before
>>> initializing PHY. And it need to reconfigure USB PHY interface of DWC3
>>> core after deassert DWC3 controller reset.
>>>
>>> The current driver supports Host only and Peripheral Only well, for
>>> now, we will add support for OTG after we have it all stabilized.
>>>
>>> Signed-off-by: William Wu <william.wu@xxxxxxxxxxxxxx>
>>> ---
>>> Changes in v10:
>>> - fix building error
>>>
>>> Changes in v9:
>>> - Add a new dwc3-rockchip.c driver for rk3399, rather than use dwc3-of-simple.c
>>>
>>>   drivers/usb/dwc3/Kconfig         |   9 +
>>>   drivers/usb/dwc3/Makefile        |   1 +
>>>   drivers/usb/dwc3/core.c          |   2 +-
>>>   drivers/usb/dwc3/core.h          |   1 +
>>>   drivers/usb/dwc3/dwc3-rockchip.c | 441 +++++++++++++++++++++++++++++++++++++++
>> William, if you need to touch core dwc3 to introduce a glue layer,
>> you're doing it wrong.
>
> Sorryï I realized that it's not better to touch core dwc3 in a specific 
> glue layer.
> I touched dwc3 in glue layer, because I want to support dual-role modeï and
> according to our dwc3 IP and usb3 PHY IP designï it need to reinit dwc3 
> core
> whenever  usb cable attached.
>
> Anyway, it's wrong to do that.:-[
>
>>> diff --git a/drivers/usb/dwc3/core.c b/drivers/usb/dwc3/core.c
>>> index e887b38..3da6215 100644
>>> --- a/drivers/usb/dwc3/core.c
>>> +++ b/drivers/usb/dwc3/core.c
>>> @@ -405,7 +405,7 @@ static void dwc3_cache_hwparams(struct dwc3 *dwc)
>>>    * initialized. The PHY interfaces and the PHYs get initialized together with
>>>    * the core in dwc3_core_init.
>>>    */
>>> -static int dwc3_phy_setup(struct dwc3 *dwc)
>>> +int dwc3_phy_setup(struct dwc3 *dwc)
>> there's no way I'll let this slip through the cracks :-)
>
> Why I need  dwc3_phy_setup in  glue layer is because usb3 IP design
> in rockchip platform. If dwc3 works on host mode, it requires to put
> dwc3 controller in reset state before usb3 phy initializingïand after
> deassert reset,  we need to reconfigure UTMI+ PHY interface because
> our dwc3 core can't configure PHY interface correctly.
>
> Thank you for give me a chance to explain it.:-)
>
>>
>>> diff --git a/drivers/usb/dwc3/dwc3-rockchip.c b/drivers/usb/dwc3/dwc3-rockchip.c
>>> new file mode 100644
>>> index 0000000..eeae1a9
>>> --- /dev/null
>>> +++ b/drivers/usb/dwc3/dwc3-rockchip.c
>>> @@ -0,0 +1,441 @@
>> [...]
>>
>>> +static int dwc3_rockchip_probe(struct platform_device *pdev)
>>> +{
>>> +	struct dwc3_rockchip	*rockchip;
>>> +	struct device		*dev = &pdev->dev;
>>> +	struct device_node	*np = dev->of_node, *child;
>>> +	struct platform_device	*child_pdev;
>>> +
>>> +	unsigned int		count;
>>> +	int			ret;
>>> +	int			i;
>>> +
>>> +	rockchip = devm_kzalloc(dev, sizeof(*rockchip), GFP_KERNEL);
>>> +
>>> +	if (!rockchip)
>>> +		return -ENOMEM;
>>> +
>>> +	count = of_clk_get_parent_count(np);
>>> +	if (!count)
>>> +		return -ENOENT;
>>> +
>>> +	rockchip->num_clocks = count;
>>> +
>>> +	rockchip->clks = devm_kcalloc(dev, rockchip->num_clocks,
>>> +				      sizeof(struct clk *), GFP_KERNEL);
>>> +	if (!rockchip->clks)
>>> +		return -ENOMEM;
>>> +
>>> +	platform_set_drvdata(pdev, rockchip);
>>> +
>>> +	rockchip->dev = dev;
>>> +	rockchip->edev = NULL;
>>> +
>>> +	pm_runtime_set_active(dev);
>>> +	pm_runtime_enable(dev);
>>> +	ret = pm_runtime_get_sync(dev);
>>> +	if (ret < 0) {
>>> +		dev_err(dev, "get_sync failed with err %d\n", ret);
>>> +		goto err1;
>>> +	}
>>> +
>>> +	for (i = 0; i < rockchip->num_clocks; i++) {
>>> +		struct clk	*clk;
>>> +
>>> +		clk = of_clk_get(np, i);
>>> +		if (IS_ERR(clk)) {
>>> +			while (--i >= 0)
>>> +				clk_put(rockchip->clks[i]);
>>> +			ret = PTR_ERR(clk);
>>> +
>>> +			goto err1;
>>> +		}
>>> +
>>> +		ret = clk_prepare_enable(clk);
>>> +		if (ret < 0) {
>>> +			while (--i >= 0) {
>>> +				clk_disable_unprepare(rockchip->clks[i]);
>>> +				clk_put(rockchip->clks[i]);
>>> +			}
>>> +			clk_put(clk);
>>> +
>>> +			goto err1;
>>> +		}
>>> +
>>> +		rockchip->clks[i] = clk;
>>> +	}
>>> +
>>> +	rockchip->otg_rst = devm_reset_control_get(dev, "usb3-otg");
>>> +	if (IS_ERR(rockchip->otg_rst)) {
>>> +		dev_err(dev, "could not get reset controller\n");
>>> +		ret = PTR_ERR(rockchip->otg_rst);
>>> +		goto err2;
>>> +	}
>>> +
>>> +	ret = dwc3_rockchip_extcon_register(rockchip);
>>> +	if (ret < 0)
>>> +		goto err2;
>>> +
>>> +	child = of_get_child_by_name(np, "dwc3");
>>> +	if (!child) {
>>> +		dev_err(dev, "failed to find dwc3 core node\n");
>>> +		ret = -ENODEV;
>>> +		goto err3;
>>> +	}
>>> +
>>> +	/* Allocate and initialize the core */
>>> +	ret = of_platform_populate(np, NULL, NULL, dev);
>>> +	if (ret) {
>>> +		dev_err(dev, "failed to create dwc3 core\n");
>>> +		goto err3;
>>> +	}
>>> +
>>> +	child_pdev = of_find_device_by_node(child);
>>> +	if (!child_pdev) {
>>> +		dev_err(dev, "failed to find dwc3 core device\n");
>>> +		ret = -ENODEV;
>>> +		goto err4;
>>> +	}
>>> +
>>> +	rockchip->dwc = platform_get_drvdata(child_pdev);
>> No! You will *NOT* the core struct device. Don't even try to come up
>> with tricks like this.
>>
>> Let's do this: introduce a glue layer that gets peripheral-only
>> working. Remove PM for now, too. Start with something simple, get the
>> bare minimum working upstream and add more stuff as you go.
>>
>> Trying to do everything in one patch just makes it much more likely for
>> your patch to be NAKed. What you're doing here is bypassing all the
>> layering we've built. That won't work very well. The only thing you'll
>> get is for your patches to continue to be NAKed.
>>
>> Avoid the tricks and abuses. Just because you _can_ do it somehow, it
>> doesn't mean you _should_ do it :-)
>>
>> Your best option right now, is to remove PM and dual-role support and a
>> minimal glue layer supported.
>>
>> In fact, all you *really* need is to add a compatible to
>> dwc3-of-simple.c. That should be enough to get your dwc3 working. Don't
>> do anything more than that. For dual-role and PM, we can add it
>> generically to dwc3-of-simple.c when all pieces fall into place.
>>
> Ah, thanks very much for your kind explanation. I think I just only need
> to add a compatible to dwc3-of-simple.cïfor nowï and I have tested
> my dwc3ï it worked well on peripheral only mode and host only mode
> without PM. Further, if dwc3-of-simple.c adds generic handling of dual-role
> and PM, I can improve our dwc3 feature.:-)

that's my point exactly. We can add more support generically so that
other platforms can benefit from the work. PM should be simple for
dwc3-of-simple.c. Dual-role will take a little more effort. In almost
there actually. There are a few missing pieces but it should be doable
(hopefully) within the next two major releases.

Your integration is no different than other companies' using DWC3 in
dual-role setup. For example TI's DWC3 have the same requirements as you
do, so it makes sense to add it straight to dwc3-core. Roger Quadros
(now in Cc) has been working on dual-role for TI's platforms and we've
been discussing about how to add missing pieces generically. Perhaps
you'd want to join the discussion.

-- 
balbi

Attachment: signature.asc
Description: PGP signature