Re: [RFT PATCH v2] gpiolib: allow gpio irqchip to map irqs dynamically

[Grygorii Strashko]

On 10/07/2017 07:27 PM, Linus Walleij wrote:
> On Thu, Sep 28, 2017 at 10:33 AM, Mika Westerberg
> <mika.westerberg@xxxxxxxxxxxxxxx> wrote:
>> On Fri, Jul 21, 2017 at 11:49:00AM -0500, Grygorii Strashko wrote:
>>> Now IRQ mappings are always created for all (allowed) GPIOs in gpiochip in
>>> gpiochip_irqchip_add_key() which goes against the idea of SPARSE_IRQ and,
>>> as result, leads to:
>>>   - increasing of memory consumption for IRQ descriptors most of which will
>>> never ever be used (espessially on platform with a high number of GPIOs).
>>> (sizeof(struct irq_desc) == 256 on my tested platforms)
>>>   - imposibility to use GPIO irqchip APIs by gpio drivers when HW implements
>>> GPIO IRQ functionality as IRQ crossbar/router which has only limited
>>> number of IRQ outputs (example from [1], all GPIOs can be mapped on only 8
>>> IRQs).
>>>
>>> Hence, remove static IRQ mapping code from gpiochip_irqchip_add_key() and
>>> instead replace irq_find_mapping() with irq_create_mapping() in
>>> gpiochip_to_irq(). Also add additional gpiochip_irqchip_irq_valid() calls
>>> in gpiochip_to_irq() and gpiochip_irq_map().
>>>
>>> After this change gpio2irq mapping will happen the following way when GPIO
>>> irqchip APIs are used by gpio driver:
>>>   - IRQ mappings will be created statically if driver passes first_irq>0
>>> vlaue in gpiochip_irqchip_add_key().
>>>   - IRQ mappings will be created dynamically from gpio_to_irq() or
>>> of_irq_get().
>>>
>>> Tested on am335x-evm and dra72-evm-revc.
>>> - dra72-evm-revc: number of created irq mappings decreased from 402 -> 135
>>>    Mem savings 267*256 = 68352 (66kB)
>>> - am335x-evm: number of created irq mappings decreased from 188 -> 63
>>>    Mem savings 125*256 = 32000 (31kB)
>>>
>>> [1] https://lkml.org/lkml/2017/6/15/428
>>> Signed-off-by: Grygorii Strashko <grygorii.strashko@xxxxxx>
>>
>> Hi Grygorii,
>>
>> It looks like dc749a09ea5e41 ("gpiolib: allow gpio irqchip to map irqs
>> dynamically") broke the quirk we added for some Intel Cherryview
>> systems [1]. Basically after this keyboard of those systems stopped
>> working again.
> 
> It seems like we need to revert this change.
> 
> I have also noted the following:
> 
> -       /*
> -        * Prepare the mapping since the irqchip shall be orthogonal to
> -        * any gpiochip calls. If the first_irq was zero, this is
> -        * necessary to allocate descriptors for all IRQs.
> -        */
> -       for (offset = 0; offset < gpiochip->ngpio; offset++) {
> -               if (!gpiochip_irqchip_irq_valid(gpiochip, offset))
> -                       continue;
> -               irq_base = irq_create_mapping(gpiochip->irqdomain, offset);
> -               if (!irq_base_set) {
> -                       /*
> -                        * Store the base into the gpiochip to be used when
> -                        * unmapping the irqs.
> -                        */
> -                       gpiochip->irq_base = irq_base;
> -                       irq_base_set = true;
> -               }
> 
> Note assignment of gpiochip->irq_base.
> 
> A whole slew of drivers use this.
> 
> We cannot merge a change like this without also fixing all the users
> of .irq_base, as it will cause regressions on them.

At the moment it was merged there were no user of irq_base except gpio-mockup.c.
And actually there are shouldn't as calling irq_create_mapping() in cycle
will not guarantee sequential Linux IRQ numbers allocation.



-- 
regards,
-grygorii