Re: gpio-pch: BUG - driver does not honour IRQF_ONESHOTn

[Jean-Francois Dagenais]

On Apr 25, 2012, at 19:03, Thomas Gleixner wrote:

> On Wed, 25 Apr 2012, Jean-Francois Dagenais wrote:
> 
>> Isn't anyone concerned with this? I am replying to myself to provoke a response.
> 
> Here you go.
> 
>> Configuring a gpio pin with the gpio-pch driver with
>> "IRQF_TRIGGER_LOW | IRQF_ONESHOT" generates an interrupt
>> storm for threaded ISR until the ISR thread actually gets to physically clear
>> the interrupt on the triggering chip!! The immediate observable symptom
>> is the high CPU usage for my ISR thread task and the interrupt count in
>> /proc/interrupts incrementing radically.
>> See below for more details.
>> 
> 
>>> Unfortunately, since all of the handling of the adp5588 is done in
>>> a thread function, the interrupt stays low between the moment the
>>> hard handler is run and the 5588 function is run. Since
>>> pch_gpio_handler clears the interrupt status right away, I get an
>>> interrupt storm for the pch_gpio_handler function.
> 
>>> the line that does "iowrite32(BIT(i), &chip->reg->iclr);" right
>>> before calling generic_handle_irq should be executed only after
>>> the corresponding nested ISR has run it's thread function.
> 
> Sigh, that driver is broken in several ways.
> 
> Does the following untested patch solve your problem?
it seems to be working ok, I have an ad714x on it and it requested
a threaded ISR with "IRQF_SHARED | IRQF_TRIGGER_LOW | IRQF_ONESHOT".

I will continue to use it and update you if I find anything wrong. If you format a
proper commit patch, I can ack it.
> 
> It's not complete and I neither have access to that hardware nor did I
> bother to look for docs.
> 
> Fact is, that using handle_simple_irq is preventing the usage of
> oneshot threaded handlers for the demultiplexed interrupts, because
> the simple handler does not deal with masking/unmasking. So yes you
> run into an irq storm.
> 
> The "ack" of the sub interrupt before calling the handler is crap as
> well. Though it kinda works as long as you are not using threaded
> interrupts... 
> 
> The other nonsense is that the set_type function unconditionally
> enables the interrupt. It's supposed to set the type and nothing
> else. The unmasking is done by the core code. If that hardware
> requires to do the type changes in masked state, then it should
> indicate it to the core by adding the IRQCHIP_SET_TYPE_MASKED flag to
> the irq chip. But I can't see it masking it. I just see the
> unconditional unmask and enable. Looks like one of those, it works but
> we don't know why thingies....
> 
> That said, the patch is just done from my core code POV and the
> limited information I was able to pull out of the driver code.
> 
> Thanks,

Good job, Thank YOU!
> 
> 	tglx
> 
> diff --git a/drivers/gpio/gpio-pch.c b/drivers/gpio/gpio-pch.c
> index e8729cc..8b92fec 100644
> --- a/drivers/gpio/gpio-pch.c
> +++ b/drivers/gpio/gpio-pch.c
> @@ -230,16 +230,12 @@ static void pch_gpio_setup(struct pch_gpio *chip)
> 
> static int pch_irq_type(struct irq_data *d, unsigned int type)
> {
> -	u32 im;
> -	u32 __iomem *im_reg;
> -	u32 ien;
> -	u32 im_pos;
> -	int ch;
> -	unsigned long flags;
> -	u32 val;
> -	int irq = d->irq;
> 	struct irq_chip_generic *gc = irq_data_get_irq_chip_data(d);
> 	struct pch_gpio *chip = gc->private;
> +	u32 im, im_pos, val;
> +	u32 __iomem *im_reg;
> +	unsigned long flags;
> +	int ch, irq = d->irq;
> 
> 	ch = irq - chip->irq_base;
> 	if (irq <= chip->irq_base + 7) {
> @@ -270,30 +266,22 @@ static int pch_irq_type(struct irq_data *d, unsigned int type)
> 	case IRQ_TYPE_LEVEL_LOW:
> 		val = PCH_LEVEL_L;
> 		break;
> -	case IRQ_TYPE_PROBE:
> -		goto end;
> 	default:
> -		dev_warn(chip->dev, "%s: unknown type(%dd)",
> -			__func__, type);
> -		goto end;
> +		goto unlock;
> 	}
> 
> 	/* Set interrupt mode */
> 	im = ioread32(im_reg) & ~(PCH_IM_MASK << (im_pos * 4));
> 	iowrite32(im | (val << (im_pos * 4)), im_reg);
> 
> -	/* iclr */
> -	iowrite32(BIT(ch), &chip->reg->iclr);
> +	/* And the handler */
> +	if (type & (IRQ_TYPE_LEVEL_LOW | IRQ_TYPE_LEVEL_HIGH))
> +		__irq_set_handler_locked(d->irq, handle_level_irq);
> +	else if (type & (IRQ_TYPE_EDGE_FALLING | IRQ_TYPE_EDGE_RISING))
> +		__irq_set_handler_locked(d->irq, handle_edge_irq);
> 
> -	/* IMASKCLR */
> -	iowrite32(BIT(ch), &chip->reg->imaskclr);
> -
> -	/* Enable interrupt */
> -	ien = ioread32(&chip->reg->ien);
> -	iowrite32(ien | BIT(ch), &chip->reg->ien);
> -end:
> +unlock:
> 	spin_unlock_irqrestore(&chip->spinlock, flags);
> -
> 	return 0;
> }
> 
> @@ -313,18 +301,24 @@ static void pch_irq_mask(struct irq_data *d)
> 	iowrite32(1 << (d->irq - chip->irq_base), &chip->reg->imask);
> }
> 
> +static void pch_irq_ack(struct irq_data *d)
> +{
> +	struct irq_chip_generic *gc = irq_data_get_irq_chip_data(d);
> +	struct pch_gpio *chip = gc->private;
> +
> +	iowrite32(1 << (d->irq - chip->irq_base), &chip->reg->iclr);
> +}
> +
> static irqreturn_t pch_gpio_handler(int irq, void *dev_id)
> {
> 	struct pch_gpio *chip = dev_id;
> 	u32 reg_val = ioread32(&chip->reg->istatus);
> -	int i;
> -	int ret = IRQ_NONE;
> +	int i, ret = IRQ_NONE;
> 
> 	for (i = 0; i < gpio_pins[chip->ioh]; i++) {
> 		if (reg_val & BIT(i)) {
> 			dev_dbg(chip->dev, "%s:[%d]:irq=%d  status=0x%x\n",
> 				__func__, i, irq, reg_val);
> -			iowrite32(BIT(i), &chip->reg->iclr);
> 			generic_handle_irq(chip->irq_base + i);
> 			ret = IRQ_HANDLED;
> 		}
> @@ -343,6 +337,7 @@ static __devinit void pch_gpio_alloc_generic_chip(struct pch_gpio *chip,
> 	gc->private = chip;
> 	ct = gc->chip_types;
> 
> +	ct->chip.irq_ack = pch_irq_ack;
> 	ct->chip.irq_mask = pch_irq_mask;
> 	ct->chip.irq_unmask = pch_irq_unmask;
> 	ct->chip.irq_set_type = pch_irq_type;
> @@ -357,6 +352,7 @@ static int __devinit pch_gpio_probe(struct pci_dev *pdev,
> 	s32 ret;
> 	struct pch_gpio *chip;
> 	int irq_base;
> +	u32 msk;
> 
> 	chip = kzalloc(sizeof(*chip), GFP_KERNEL);
> 	if (chip == NULL)
> @@ -418,8 +414,10 @@ static int __devinit pch_gpio_probe(struct pci_dev *pdev,
> 
> 	pch_gpio_alloc_generic_chip(chip, irq_base, gpio_pins[chip->ioh]);
> 
> -	/* Initialize interrupt ien register */
> -	iowrite32(0, &chip->reg->ien);
> +	/* Mask all interrupts, but enable them */
> +	msk = (1 << gpio_pins[chip->ioh]) - 1;
> +	iowrite32(msk, &chip->reg->imask);
> +	iowrite32(msk, &chip->reg->ien);
> end:
> 	return 0;
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/