RE: [RFC PATCH v2 04/15] usb:cdns3: Driver initialization code.

[Pawel Laszczak]

Hi,

>On 04/12/18 10:50, Peter Chen wrote:
>>>> + * Cadence USBSS DRD Driver.
>>>> + *
>>>> + * Copyright (C) 2018 Cadence.
>>>> + *
>>>> + * Author: Peter Chen <peter.chen@xxxxxxx>
>>>> + *         Pawel Laszczak <pawell@xxxxxxxxxxx>
>>>> + */
>>>> +
>>>> +#include <linux/module.h>
>>>> +#include <linux/kernel.h>
>>>> +#include <linux/platform_device.h>
>>>> +#include <linux/interrupt.h>
>>>> +#include <linux/io.h>
>>>> +#include <linux/pm_runtime.h>
>>>> +
>>>> +#include "gadget.h"
>>>> +#include "core.h"
>>>> +
>>>> +static inline struct cdns3_role_driver *cdns3_get_current_role_driver(struct cdns3 *cdns)
>>>> +{
>>>> +     WARN_ON(cdns->role >= CDNS3_ROLE_END || !cdns->roles[cdns->role]);
>>>> +     return cdns->roles[cdns->role];
>>>> +}
>>>> +
>>>> +static inline int cdns3_role_start(struct cdns3 *cdns, enum cdns3_roles role)
>>>> +{
>>>> +     int ret;
>>>> +
>>>> +     if (role >= CDNS3_ROLE_END)
>>>
>>> WARN_ON()?
>>>
>>>> +             return 0;
>>>> +
>>>> +     if (!cdns->roles[role])
>>>> +             return -ENXIO;
>>>> +
>>>> +     mutex_lock(&cdns->mutex);
>>>> +     cdns->role = role;
>>>> +     ret = cdns->roles[role]->start(cdns);
>>>> +     mutex_unlock(&cdns->mutex);
>>>> +     return ret;
>>>> +}
>>>> +
>>>> +static inline void cdns3_role_stop(struct cdns3 *cdns)
>>>> +{
>>>> +     enum cdns3_roles role = cdns->role;
>>>> +
>>>> +     if (role == CDNS3_ROLE_END)
>>>
>>> WARN_ON(role >= CNDS3_ROLE_END) ?
>>>
>>>> +             return;
>>>> +
>>>> +     mutex_lock(&cdns->mutex);
>>>> +     cdns->roles[role]->stop(cdns);
>>>> +     cdns->role = CDNS3_ROLE_END;
>>>
>>> Why change the role here? You are just stopping the role not changing it.
>>> I think cdns->role should remain unchanged, so we can call cdns3_role_start()
>>> if required without error.
>>>
>>
>> The current version of this IP has some issues to detect vbus status correctly,
>> we have to force vbus status accordingly, so we need a status to indicate
>> vbus disconnection, and add some code to let controller know vbus
>> removal, in that case, the controller's state machine can be correct.
>> So, we increase one role 'CDNS3_ROLE_END' to for this purpose.
>>
>> CDNS3_ROLE_GADGET: gadget mode and VBUS on
>> CDNS3_ROLE_HOST: host mode and VBUS on
>> CDNS3_ROLE_END: VBUS off, eg either host or device cable on the port.
>>
>> So, we may start role from CDNS3_ROLE_END at probe when nothing is connected,
>> and need to set role as CDNS3_ROLE_END at ->stop for further handling at
>> role switch routine.
>
>OK. but still this (changing to ROLE_END) must be moved to the role switch routine
>and the explanation you just mentioned must be added there.
>
>>
>>>> +     mutex_unlock(&cdns->mutex);
>>>> +}
>>>> +
>>>> +static enum cdns3_roles cdns3_get_role(struct cdns3 *cdns)
>>>> +{
>>>> +     if (cdns->roles[CDNS3_ROLE_HOST] && cdns->roles[CDNS3_ROLE_GADGET]) {
>>>> +             //TODO: implements selecting device/host mode
>>>> +             return CDNS3_ROLE_HOST;
>>>> +     }
>>>> +     return cdns->roles[CDNS3_ROLE_HOST]
>>>> +             ? CDNS3_ROLE_HOST
>>>> +             : CDNS3_ROLE_GADGET;
>>>
>>> Why not just
>>>         return cdns->role;
>>>
>>> I'm wondering if we really need this function.
>>
>> cdns->role gets from cdns3_get_role, and this API tells role at the runtime.
>> If both roles are supported, the role is decided by external
>> conditions, eg, vbus/id
>> or external connector. If only single role is supported, only one role structure
>> is allocated, cdns->roles[CDNS3_ROLE_HOST] or cdns->roles[CDNS3_ROLE_GADGET]
>>
>
>How about adding this description in function documentation.

Ok, I will do it. 
>
>>>> +}
>>>
>>>> +
>>>> +/**
>>>> + * cdns3_core_init_role - initialize role of operation
>>>> + * @cdns: Pointer to cdns3 structure
>>>> + *
>>>> + * Returns 0 on success otherwise negative errno
>>>> + */
>>>> +static int cdns3_core_init_role(struct cdns3 *cdns)
>>>> +{
>>>> +     struct device *dev = cdns->dev;
>>>> +     enum usb_dr_mode dr_mode;
>>>> +
>>>> +     dr_mode = usb_get_dr_mode(dev);
>>>> +     cdns->role = CDNS3_ROLE_END;
>>>> +
>>>> +     /*
>>>> +      * If driver can't read mode by means of usb_get_dr_mdoe function then
>>>> +      * chooses mode according with Kernel configuration. This setting
>>>> +      * can be restricted later depending on strap pin configuration.
>>>> +      */
>>>> +     if (dr_mode == USB_DR_MODE_UNKNOWN) {
>>>> +             if (IS_ENABLED(CONFIG_USB_CDNS3_HOST) &&
>>>> +                 IS_ENABLED(CONFIG_USB_CDNS3_GADGET))
>>>> +                     dr_mode = USB_DR_MODE_OTG;
>>>> +             else if (IS_ENABLED(CONFIG_USB_CDNS3_HOST))
>>>> +                     dr_mode = USB_DR_MODE_HOST;
>>>> +             else if (IS_ENABLED(CONFIG_USB_CDNS3_GADGET))
>>>> +                     dr_mode = USB_DR_MODE_PERIPHERAL;
>>>> +     }
>>>> +
>>>> +     if (dr_mode == USB_DR_MODE_OTG || dr_mode == USB_DR_MODE_HOST) {
>>>> +             //TODO: implements host initialization
>>>
>>>                 /* TODO: Add host role */ ?
>>>
>>>> +     }
>>>> +
>>>> +     if (dr_mode == USB_DR_MODE_OTG || dr_mode == USB_DR_MODE_PERIPHERAL) {
>>>> +             //TODO: implements device initialization
>>>
>>>                 /* TODO: Add device role */ ?
>>>
>>
>> Yes, it needs to allocate cdns->roles[CDNS3_ROLE_HOST] and
>> cdns->roles[CDNS3_ROLE_GADGET].
>>
>>>> +     }
>>>> +
>>>> +     if (!cdns->roles[CDNS3_ROLE_HOST] && !cdns->roles[CDNS3_ROLE_GADGET]) {
>>>> +             dev_err(dev, "no supported roles\n");
>>>> +             return -ENODEV;
>>>> +     }
>>>> +
>>>> +     cdns->dr_mode = dr_mode;
>>
>> Pawel, why dr_mode needs to be introduced?
>>
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +/**
>>>> + * cdns3_irq - interrupt handler for cdns3 core device
>>>> + *
>>>> + * @irq: irq number for cdns3 core device
>>>> + * @data: structure of cdns3
>>>> + *
>>>> + * Returns IRQ_HANDLED or IRQ_NONE
>>>> + */
>>>> +static irqreturn_t cdns3_irq(int irq, void *data)
>>>> +{
>>>> +     struct cdns3 *cdns = data;
>>>> +     irqreturn_t ret = IRQ_NONE;
>>>> +
>>>> +     /* Handle device/host interrupt */
>>>> +     if (cdns->role != CDNS3_ROLE_END)
>>>
>>> Is it because of this that you need to set role to END at role_stop?
>>> I think it is better to add a state variable to struct cdns3_role_driver, so we can
>>> check if it is active or stopped.
>>>
>>> e.g.
>>>         if (cdns3_get_current_role_driver(cdns)->state == CDNS3_ROLE_STATE_ACTIVE)
>>>
>>>> +             ret = cdns3_get_current_role_driver(cdns)->irq(cdns);
>>>> +
>>>> +     return ret;
>>>> +}
>>>> +
>>
>>  CDNS3_ROLE_END is introduced from above comments, we don't
>> need another flag for it.
>> If cdns->role == CDNS3_ROLE_END, it handles VBUS and ID interrupt.
>>
>>>> +static void cdns3_remove_roles(struct cdns3 *cdns)
>>>
>>> Should this be called cdns3_exit_roles() to be opposite of cdns3_init_roles()?
>>>
>>
>> It is planed to called when at ->remove.
>>>> +{
>>>> +     //TODO: implements this function
>>>> +}
>>>
>>>> +
>>>> +static int cdns3_do_role_switch(struct cdns3 *cdns, enum cdns3_roles role)
>>>> +{
>>>> +     enum cdns3_roles current_role;
>>>> +     int ret = 0;
>>>> +
>>>> +     current_role = cdns->role;
>>>> +
>>>> +     if (role == CDNS3_ROLE_END)
>>>> +             return 0;
>>>
>>> role == END looks like error state. and it should never happen.
>>> WARN here?
>>>
>>
>> See my comments above.
>>
>>>> +
>>>> +     dev_dbg(cdns->dev, "Switching role");
>>>> +
>>>
>>> Don't you have to stop the previous role before starting the new role?
>>>
>>
>> Yes, it is needed. Pawel may simply some flows to suit his platform.
>>
>>>> +     ret = cdns3_role_start(cdns, role);
>>>> +     if (ret) {
>>>> +             /* Back to current role */
>>>> +             dev_err(cdns->dev, "set %d has failed, back to %d\n",
>>>> +                     role, current_role);
>>>> +             ret = cdns3_role_start(cdns, current_role);
>>>> +     }
>>>> +
>>>> +     return ret;
>>>> +}
>>>> +
>>>> +/**
>>>> + * cdns3_role_switch - work queue handler for role switch
>>>> + *
>>>> + * @work: work queue item structure
>>>> + *
>>>> + * Handles below events:
>>>> + * - Role switch for dual-role devices
>>>> + * - CDNS3_ROLE_GADGET <--> CDNS3_ROLE_END for peripheral-only devices
>>>> + */
>>>> +static void cdns3_role_switch(struct work_struct *work)
>>>> +{
>>>> +     enum cdns3_roles role = CDNS3_ROLE_END;
>>>> +     struct cdns3 *cdns;
>>>> +     bool device, host;
>>>> +
>>>> +     cdns = container_of(work, struct cdns3, role_switch_wq);
>>>> +
>>>> +     //TODO: implements this functions.
>>>> +     //host = cdns3_is_host(cdns);
>>>> +     //device = cdns3_is_device(cdns);
>>>> +     host = 1;
>>>> +     device = 0;
>>>> +
>>>> +     if (host)
>>>> +             role = CDNS3_ROLE_HOST;
>>>> +     else if (device)
>>>> +             role = CDNS3_ROLE_GADGET;
>>>> +
>>>> +     if (cdns->desired_dr_mode == cdns->current_dr_mode &&
>>>> +         cdns->role == role)
>>>> +             return;
>>>> +
>>>
>>> I think all the below code can be moved to cdns3_do_role_switch().
>>>
>>>> +     pm_runtime_get_sync(cdns->dev);
>>>> +     cdns3_role_stop(cdns);
>>>> +
>>>> +     if (host) {
>>>> +             if (cdns->roles[CDNS3_ROLE_HOST])
>>>> +                     cdns3_do_role_switch(cdns, CDNS3_ROLE_HOST);
>>>> +             pm_runtime_put_sync(cdns->dev);
>>>> +             return;
>>>> +     }
>>>> +
>>>> +     if (device)
>>>> +             cdns3_do_role_switch(cdns, CDNS3_ROLE_GADGET);
>>>> +     else
>>>> +             cdns3_do_role_switch(cdns, CDNS3_ROLE_END);
>>>> +
>>>> +     pm_runtime_put_sync(cdns->dev);
>>>> +}
>>>> +
>>>> +/**
>>>> + * cdns3_probe - probe for cdns3 core device
>>>> + * @pdev: Pointer to cdns3 core platform device
>>>> + *
>>>> + * Returns 0 on success otherwise negative errno
>>>> + */
>>>> +static int cdns3_probe(struct platform_device *pdev)
>>>> +{
>>>> +     struct device *dev = &pdev->dev;
>>>> +     struct resource *res;
>>>> +     struct cdns3 *cdns;
>>>> +     void __iomem *regs;
>>>> +     int ret;
>>>> +
>>>> +     cdns = devm_kzalloc(dev, sizeof(*cdns), GFP_KERNEL);
>>>> +     if (!cdns)
>>>> +             return -ENOMEM;
>>>> +
>>>> +     cdns->dev = dev;
>>>> +
>>>> +     platform_set_drvdata(pdev, cdns);
>>>> +
>>>> +     res = platform_get_resource(pdev, IORESOURCE_IRQ, 0);
>>>> +     if (!res) {
>>>> +             dev_err(dev, "missing IRQ\n");
>>>> +             return -ENODEV;
>>>> +     }
>>>> +     cdns->irq = res->start;
>>>> +
>>>> +     /*
>>>> +      * Request memory region
>>>> +      * region-0: xHCI
>>>> +      * region-1: Peripheral
>>>> +      * region-2: OTG registers
>>>> +      */
>>>
>>> The memory region order is different from the dt-binding.
>>> There it is OTG, host(xhci), device (peripheral).
>>>
>>>> +     res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
>>>> +     regs = devm_ioremap_resource(dev, res);
>>>> +
>>>> +     if (IS_ERR(regs))
>>>> +             return PTR_ERR(regs);
>>>> +     cdns->xhci_regs = regs;
>>>> +     cdns->xhci_res = res;
>>>> +
>>>> +     res = platform_get_resource(pdev, IORESOURCE_MEM, 1);
>>>> +     regs = devm_ioremap_resource(dev, res);
>>>> +     if (IS_ERR(regs))
>>>> +             return PTR_ERR(regs);
>>>> +     cdns->dev_regs  = regs;
>>>> +
>>>> +     res = platform_get_resource(pdev, IORESOURCE_MEM, 2);
>>>> +     regs = devm_ioremap_resource(dev, res);
>>>> +     if (IS_ERR(regs))
>>>> +             return PTR_ERR(regs);
>>>> +     cdns->otg_regs = regs;
>>>> +
>>>> +     mutex_init(&cdns->mutex);
>>>> +
>>>> +     cdns->phy = devm_phy_get(dev, "cdns3,usbphy");
>>>
>>> "cdns3,usbphy" is not documented in dt-binding.
>>>
>>>> +     if (IS_ERR(cdns->phy)) {
>>>> +             dev_info(dev, "no generic phy found\n");
>>>> +             cdns->phy = NULL;
>>>> +             /*
>>>> +              * fall through here!
>>>> +              * if no generic phy found, phy init
>>>> +              * should be done under boot!
>>>> +              */
>>>
>>> No you shouldn't fall through always if it is an error condition.
>>> Something like this should work better.
>>>
>>>         if (IS_ERR(cnds->phy)) {
>>>                 ret = PTR_ERR(cdns->phy);
>>>                 if (ret == -ENOSYS || ret == -ENODEV) {
>>>                         cdns->phy = NULL;
>>>                 } else if (ret == -EPROBE_DEFER) {
>>>                         return ret;
>>>                 } else {
>>>                         dev_err(dev, "no phy found\n");
>>>                         goto err0;
>>>                 }
>>>         }
>>>
>>> So if PHY was provided in DT, and PHY support/drivers is present
>>> and error condition means something is wrong and we have to error out.
>>>
>>>> +     } else {
>>>> +             phy_init(cdns->phy);
>>>> +     }
>>>
>>> You can do phy_init() outside the else.
>>>
>>>> +
>>>> +     ret = cdns3_core_init_role(cdns);
>>>> +     if (ret)
>>>> +             goto err1;
>>>> +
>>>> +     INIT_WORK(&cdns->role_switch_wq, cdns3_role_switch);
>>>> +     if (ret)
>>>> +             goto err2;
>>>> +
>>>> +     if (ret)
>>>> +             goto err2;
>>>> +
>>>> +     cdns->role = cdns3_get_role(cdns);
>>>
>>> I think this should move to cdns3_core_init_role().
>>>
>>
>> I agree.
>>
>>>> +
>>>> +     ret = devm_request_irq(dev, cdns->irq, cdns3_irq, IRQF_SHARED,
>>>> +                            dev_name(dev), cdns);
>>>> +
>>>> +     if (ret)
>>>> +             goto err2;
>>>
>>> How about moving request_irq to before cdsn3_core_init_role()?
>>>
>>> Then you can move cdns3_role_start() as well to core_init_role().
>>>
>>
>> Usually, we request irq after hardware initialization has finished, if not,
>> there may unexpected interrupt.
>
>Doesn't kernel warn if interrupt happens and there is no handler?
>To avoid that I was suggesting to request_irq first.
>
Cheers
Pawel