Re: [PATCH 3/3] pinctrl: qcom: Clear possible pending irq when remuxing GPIOs

[Marc Zyngier <maz@kernel.org>]

On 2020-11-24 12:43, Maulik Shah wrote:
> Hi Marc,
> 
> On 11/24/2020 4:45 PM, Marc Zyngier wrote:
>> On 2020-11-24 10:37, Maulik Shah wrote:
>> 
>> [...]
>> 
>>>>   static int msm_pinmux_set_mux(struct pinctrl_dev *pctldev,
>>>>                     unsigned function,
>>>>                     unsigned group)
>>>>   {
>>>>       struct msm_pinctrl *pctrl = pinctrl_dev_get_drvdata(pctldev);
>>>> +    struct gpio_chip *gc = &pctrl->chip;
>>>> +    unsigned int irq = irq_find_mapping(gc->irq.domain, group);
>>>>       const struct msm_pingroup *g;
>>>>       unsigned long flags;
>>>>       u32 val, mask;
>>>> +    u32 oldval;
>>>> +    u32 old_i;
>>>>       int i;
>>>>         g = &pctrl->soc->groups[group];
>>>> @@ -187,15 +215,26 @@ static int msm_pinmux_set_mux(struct 
>>>> pinctrl_dev *pctldev,
>>>>       if (WARN_ON(i == g->nfuncs))
>>>>           return -EINVAL;
>>>>   -    raw_spin_lock_irqsave(&pctrl->lock, flags);
>>>> +    disable_irq(irq);
>>>>   -    val = msm_readl_ctl(pctrl, g);
>>>> +    raw_spin_lock_irqsave(&pctrl->lock, flags);
>>>> +    oldval = val = msm_readl_ctl(pctrl, g);
>>>>       val &= ~mask;
>>>>       val |= i << g->mux_bit;
>>>>       msm_writel_ctl(val, pctrl, g);
>>>> -
>>>>       raw_spin_unlock_irqrestore(&pctrl->lock, flags);
>>>>   +    /*
>>>> +     * Clear IRQs if switching to/from GPIO mode since muxing 
>>>> to/from
>>>> +     * the GPIO path can cause phantom edges.
>>>> +     */
>>>> +    old_i = (oldval & mask) >> g->mux_bit;
>>>> +    if (old_i != i &&
>>>> +        (i == pctrl->soc->gpio_func || old_i == 
>>>> pctrl->soc->gpio_func))
>>>> +        msm_pinctrl_clear_pending_irq(pctrl, group, irq);
>>> 
>>> disable_irq() and enable_irq() should be moved inside this if loop. 
>>> as
>>> only use for this is to mask the IRQ when switching back to gpio IRQ
>>> mode?
>>> 
>>> i also don't think we should leave IRQ enabled at the end of this
>>> function by default, probably need to check if IRQ was already
>>> unmasked before disabling it, then only call enable_irq().
>> 
>> Why? It looks to me that this reproduces the behaviour of 
>> IRQCHIP_SET_TYPE_MASKED, which is highly desirable. What
>> problem are you trying to address with this?
> 
> Correct, here trying to reproduce the behaviour of
> IRQCHIP_SET_TYPE_MASKED which i guess is ok once its moved inside if
> loop as this is the place its switching to IRQ mode.
> 
> but there is a problem to leave it enabled at the end of set_direction
> callbacks, see below.
> 
>> 
>>> 
>>>> +
>>>> +    enable_irq(irq);
>>>> +
>>>>       return 0;
>>>>   }
>>>>   @@ -456,32 +495,45 @@ static const struct pinconf_ops 
>>>> msm_pinconf_ops = {
>>>>   static int msm_gpio_direction_input(struct gpio_chip *chip, 
>>>> unsigned offset)
>>>>   {
>>>>       const struct msm_pingroup *g;
>>>> +    unsigned int irq = irq_find_mapping(chip->irq.domain, offset);
>>>>       struct msm_pinctrl *pctrl = gpiochip_get_data(chip);
>>>>       unsigned long flags;
>>>> +    u32 oldval;
>>>>       u32 val;
>>>>         g = &pctrl->soc->groups[offset];
>>>>   +    disable_irq(irq);
>>>> +
>>>>       raw_spin_lock_irqsave(&pctrl->lock, flags);
>>>>   -    val = msm_readl_ctl(pctrl, g);
>>>> +    oldval = val = msm_readl_ctl(pctrl, g);
>>>>       val &= ~BIT(g->oe_bit);
>>>>       msm_writel_ctl(val, pctrl, g);
>>>>         raw_spin_unlock_irqrestore(&pctrl->lock, flags);
>>>>   +    if (oldval != val)
>>>> +        msm_pinctrl_clear_pending_irq(pctrl, offset, irq);
>>>> +
>>>> +    enable_irq(irq);
>>> 
>>> i do not think we need disable_irq() and enable_irq() here, changing
>>> direction to input does not mean its being used for interrupt only, 
>>> it
>>> may be set to use something like Rx mode in UART.
>>> 
>>> the client driver should enable IRQ when needed.
>> 
>> And the kernel doesn't expect random interrupts to fire. Again, what
>> are you trying to fix by removing these?
> 
> I see leaving IRQ enabled here can cause problems. For example in
> qcom_geni_serial.c driver before requesting IRQ, it sets the
> IRQ_NOAUTOEN flag to not keep it enabled.
> 
> see the below snippet
>         irq_set_status_flags(uport->irq, IRQ_NOAUTOEN);
>         ret = devm_request_irq(uport->dev, uport->irq, 
> qcom_geni_serial_isr,
>                         IRQF_TRIGGER_HIGH, port->name, uport);
> 
> later when this devm_request_irq() invokes .irq_request_resources
> callback it will reach msm_gpio_irq_reqres() from
> where msm_gpio_direction_input() is called which leaves the irq
> enabled at the end with enable_irq() which was not expected by driver.

No it doesn't. disable_irq()/enable_irq() are designed to nest.
If the interrupt line was disabled before the disable/enable
sequence, it will still be disabled after.

> It will cause is IRQ storm since the UART geni driver uses GPIO in Rx
> mode when out of suspend. The IRQ mode in GPIO is enabled
> with suspend entry only. During resume the IRQ will again be disabled
> and GPIO will be switched to Rx mode.

I don't see how this contradicts what is above. If the interrupt was
disabled before hitting this sequence, it will still be disabled after.
Am I missing something? Have you actually seen the problem on HW?

         M.
-- 
Jazz is not dead. It just smells funny...