Re: [PATCH v2 02/10] irqchip: mtk-sysirq: extend intpol base to arbitrary number

[Mars Cheng <mars.cheng-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org>]

On Thu, 2017-02-09 at 09:03 +0000, Marc Zyngier wrote:
> On 06/02/17 12:15, Mars Cheng wrote:
> > Originally driver only supports one base. However, MT6797 has
> > more than one bases to configure interrupt polarity. To support
> > possible design change, here comes a solution to use arbitrary
> > number of bases.
> > 
> > Signed-off-by: Mars Cheng <mars.cheng-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org>
> > ---
> >  drivers/irqchip/irq-mtk-sysirq.c |   71 +++++++++++++++++++++++++++-----------
> >  1 file changed, 50 insertions(+), 21 deletions(-)
> > 
> > diff --git a/drivers/irqchip/irq-mtk-sysirq.c b/drivers/irqchip/irq-mtk-sysirq.c
> > index 63ac73b..2645706 100644
> > --- a/drivers/irqchip/irq-mtk-sysirq.c
> > +++ b/drivers/irqchip/irq-mtk-sysirq.c
> > @@ -24,7 +24,9 @@
> >  
> >  struct mtk_sysirq_chip_data {
> >  	spinlock_t lock;
> > -	void __iomem *intpol_base;
> > +	u32 nr_intpol_bases;
> > +	void __iomem **intpol_bases;
> > +	u32 *intpol_words;
> >  };
> >  
> >  static int mtk_sysirq_set_type(struct irq_data *data, unsigned int type)
> > @@ -33,13 +35,15 @@ static int mtk_sysirq_set_type(struct irq_data *data, unsigned int type)
> >  	struct mtk_sysirq_chip_data *chip_data = data->chip_data;
> >  	u32 offset, reg_index, value;
> >  	unsigned long flags;
> > -	int ret;
> > +	int ret, i;
> >  
> >  	offset = hwirq & 0x1f;
> >  	reg_index = hwirq >> 5;
> > +	for (i = 0; reg_index >= chip_data->intpol_words[i]; i++)
> > +		reg_index -= chip_data->intpol_words[i];
> 
> Two questions:
> - What guarantees that two successive regions deal with consecutive
> interrupts? For example, if I have region A that deals with interrupts
> 0-31, what guarantees that region B covers 32-63?

It is guaranteed by mediatek's chip design team. For those chips with
multiple bases, they all have the consecutive interrupts in the next
region.

> - Given that there is a static relation between a region and a hwirq,
> can't you compute this relation at init time, and let set_type be a fast
> path?
> 

sure, I can do this to optimize set_type. will do it in v3

> >  
> >  	spin_lock_irqsave(&chip_data->lock, flags);
> > -	value = readl_relaxed(chip_data->intpol_base + reg_index * 4);
> > +	value = readl_relaxed(chip_data->intpol_bases[i] + reg_index * 4);
> >  	if (type == IRQ_TYPE_LEVEL_LOW || type == IRQ_TYPE_EDGE_FALLING) {
> >  		if (type == IRQ_TYPE_LEVEL_LOW)
> >  			type = IRQ_TYPE_LEVEL_HIGH;
> > @@ -49,7 +53,8 @@ static int mtk_sysirq_set_type(struct irq_data *data, unsigned int type)
> >  	} else {
> >  		value &= ~(1 << offset);
> >  	}
> > -	writel(value, chip_data->intpol_base + reg_index * 4);
> > +
> > +	writel(value, chip_data->intpol_bases[i] + reg_index * 4);
> 
> Why do you have a writel here, while you're using relaxed accessors
> otherwise? Is there anything else that needs to be made visible to the
> irqchip?
> 

before we call spin_unlock_irqrestore() in the end of set_type, polarity
should take effect so we use writel() here. This patch did not change
the usage.

> >  
> >  	data = data->parent_data;
> >  	ret = data->chip->irq_set_type(data, type);
> > @@ -124,8 +129,7 @@ static int __init mtk_sysirq_of_init(struct device_node *node,
> >  {
> >  	struct irq_domain *domain, *domain_parent;
> >  	struct mtk_sysirq_chip_data *chip_data;
> > -	int ret, size, intpol_num;
> > -	struct resource res;
> > +	int ret, size, intpol_num = 0, nr_intpol_bases, i;
> >  
> >  	domain_parent = irq_find_host(parent);
> >  	if (!domain_parent) {
> > @@ -133,36 +137,61 @@ static int __init mtk_sysirq_of_init(struct device_node *node,
> >  		return -EINVAL;
> >  	}
> >  
> > -	ret = of_address_to_resource(node, 0, &res);
> > -	if (ret)
> > -		return ret;
> > -
> >  	chip_data = kzalloc(sizeof(*chip_data), GFP_KERNEL);
> >  	if (!chip_data)
> >  		return -ENOMEM;
> >  
> > -	size = resource_size(&res);
> > -	intpol_num = size * 8;
> > -	chip_data->intpol_base = ioremap(res.start, size);
> > -	if (!chip_data->intpol_base) {
> > -		pr_err("mtk_sysirq: unable to map sysirq register\n");
> > -		ret = -ENXIO;
> > -		goto out_free;
> > +	if (of_property_read_u32(node, "#intpol-bases", &nr_intpol_bases))
> > +		nr_intpol_bases = 1;
> > +
> > +	chip_data->intpol_words =
> > +		kcalloc(nr_intpol_bases, sizeof(u32), GFP_KERNEL);
> 
> Please keep the assignment on a single line.
> 

Got it, but another warning (prefer 75 char in single line) would pop
up. Would you still like me to keep it on a single line?

> > +	if (!chip_data->intpol_words) {
> > +		ret = -ENOMEM;
> > +		goto out_free_chip;
> > +	}
> > +
> > +	chip_data->intpol_bases =
> > +		kcalloc(nr_intpol_bases, sizeof(void __iomem *), GFP_KERNEL);
> 
> Same here.
> 
> > +	if (!chip_data->intpol_bases) {
> > +		ret = -ENOMEM;
> > +		goto out_free_intpol_words;
> > +	}
> > +
> > +	for (i = 0; i < nr_intpol_bases; i++) {
> > +		struct resource res;
> > +
> > +		ret = of_address_to_resource(node, i, &res);
> > +		size = resource_size(&res);
> > +		intpol_num += size * 8;
> > +		chip_data->intpol_words[i] = size / 4;
> > +		chip_data->intpol_bases[i] = of_iomap(node, i);
> > +		if (ret || !chip_data->intpol_bases[i]) {
> > +			pr_err("%s: couldn't map region %d\n",
> > +			       node->full_name, i);
> > +			ret = -ENODEV;
> > +			goto out_free_intpol;
> > +		}
> >  	}
> >  
> >  	domain = irq_domain_add_hierarchy(domain_parent, 0, intpol_num, node,
> >  					  &sysirq_domain_ops, chip_data);
> >  	if (!domain) {
> >  		ret = -ENOMEM;
> > -		goto out_unmap;
> > +		goto out_free_intpol;
> >  	}
> >  	spin_lock_init(&chip_data->lock);
> >  
> >  	return 0;
> >  
> > -out_unmap:
> > -	iounmap(chip_data->intpol_base);
> > -out_free:
> > +out_free_intpol:
> > +	for (i = 0; i < nr_intpol_bases; i++)
> > +		if (chip_data->intpol_bases[i])
> > +			iounmap(chip_data->intpol_bases[i]);
> > +	kfree(chip_data->intpol_bases);
> > +out_free_intpol_words:
> > +	kfree(chip_data->intpol_words);
> > +out_free_chip:
> >  	kfree(chip_data);
> >  	return ret;
> >  }
> > 
> 
> Thanks,
> 
> 	M.

Thanks,

          Mars Cheng