Re: [PATCH 00/21] add and use devm_irq_of_parse_and_map()

[Andreas Larsson <andreas@gaisler.com>]

On 2014-06-12 21:02, Nikita Yushchenko wrote:
>>> Currently many device tree aware drivers use irq_of_parse_and_map() to
>>> get
>>> IRQ number and then devm_request_irq() to set up IRQ handler.
>>>
>>> This causes a problem for exit path and for error paths: undo action for
>>> irq_of_parse_and_map() is irq_dispose_mapping() that must not be called
>>> while IRQ handler is defined, but devres moves free_irq() out of driver
>>> so driver does not have safe place to call irq_dispose_mapping().
>>>
>>> So many drivers do not call irq_dispose_mapping() at all, others call it
>>> while IRQ handler is defined (which is incorrect and causes error
>>> messages
>>> in logs), even others refuse to use devres for managing IRQ handlers.
>>>
>>> This patchset adds devres version of irq_of_parse_and_map(), and updates
>>> in-tree users of both devres and irq_of_parse_and_map() to use
>>> devm_irq_of_parse_and_map() instead.
>>
>>
>> SPARC does not use OF_IRQ and has a different implementation of
>> irq_of_parse_and_map than the one in drivers/of/irq.c. All code
>> converted from irq_of_parse_and_map to devm_irq_of_parse_and_map in this
>> patch set will be unlinkable for SPARC. This includes SPI in general and
>> many drivers that are used for SPARC (of which several are currently
>> only used on SPARC platforms).
>
> Can this be fixed by adding a copy of devm_irq_of_parse_and_map() to
> arch/sparc/kernel/of_device_common.c ?

Not a copy of the version in irq.c no. On SPARC, IRQ_DOMAIN is not 
selected in general.

However, there is no technical problem that I can see with having a 
SPARC version of devm_irq_of_parse_and_map that just calls 
irq_of_parse_and_map as there are no mappings that needs to be disposed 
of. (The empty dummy for irq_dispose_mapping is used if any 
irq_dispose_mapping calls are made from drivers).

Best regards,
Andreas Larsson