[PATCH] irq: Consider a negative return value of irq_startup() as an error

[Jean-Jacques Hiblot]

2014-02-26 23:15 GMT+01:00 Thomas Gleixner <tglx at linutronix.de>:
> On Tue, 25 Feb 2014, Jean-Jacques Hiblot wrote:
>
>> The irq_startup() function returns the return value of the
>> irq_startup callback of the underlying irq_chip. Currently this
>> value only tells if the interrupt is pending, but we can make it
>> also return an error code when it fails.
>>
>> Signed-off-by: Jean-Jacques Hiblot <jjhiblot at traphandler.com>
>> ---
>>
>> Hi Thomas,
>>
>> This patch updates the semantic of the return value of the irq_startup()
>> callback of an irq_chip : a negative value indicates an error. The purpose is
>> to make __setup_irq() fail if the irq_chip can't startup the interrupt.
>
> How does that work?
>
>     unsigned int    (*irq_startup)(struct irq_data *data);
well .. no excuse sir
>
>> I checked in the kernel for drivers impacted by this. It turns out that most
>> of the implementation of irq_startup() return 0. The only drivers that return
>> non-zero values are:
>>  * arch/x86/kernel/apic/io_apic.c: returns 1 when an interrupt is pending
>>  * drivers/pinctrl/pinctrl-adi2.c: may return -ENODEV.
>
>    which is wrong to begin with.
>
>> GPIO-based irq_chip could use this feature because an output gpio
>> can't be used as an interrupt, and thus a call to irq_startup() is
>> likely to fail. You can check out
>> http://www.spinics.net/lists/arm-kernel/msg302713.html for a
>> reference.
>
> And that tells me that the design of the gpio stuff is just wrong.
>
> The whole drviers/gpio directory is full of this gpio_lock_as_irq()
> called from the guts of irq_chip callbacks braindamage.
>
> The comment above gpiod_lock_as_irq() is so wrong it's not even funny
> anymore.
>
> "....or in the .irq_enable() from its irq_chip implementation ..."
>
>         void            (*irq_enable)(struct irq_data *data);
>
> So how does that help, if the gpio is not available as irq?
>
> And no, you are not going to cure that design brainfart by adding a
> negative return value to a function returning unsigned int. And I'm
> not going to accept anything which is just a duct tape based
> workaround for a complete braindamaged design.
>
> Lets look at the usage sites:
>
> The following irqchip callback implementations printk some blurb and
> proceed:
>
>  sirfsoc_gpio_irq_startup
>  nmk_gpio_irq_startup
>  msm_gpio_irq_startup
>  u300_gpio_irq_enable
>  byt_irq_startup
>  mcp23s08_irq_startup
>  lp_irq_startup
>  intel_mid_irq_startup
>  em_gio_irq_startup
>  bcm_kona_gpio_irq_startup
>  adnp_irq_startup
>
> The following functions return an error code from the set_type
> callback:
>
>  tegra_gpio_irq_set_type
>  gpio_irq_type
>
> The whole idiocy starts with commit d468bf9e, which tells people to
> call this from irq_startup/enable. But it fails to tell, that this is
> pointless because it wont prevent the interrupt from starting up.
>
>From a GPIO subsystem standpoint, the IRQ lock is sensible because
using a gpio as an interrupt source, adds some restrictions to its
usage. Ouput configuration and IRQ configuration are exclusive for
example.
That  request_irq() succeeds if a GPIO can't be used as interrupt, is
however a problem. It can be fixed.


> So 11 SoC lemmings followed and added this completely pointless
> nonsense to their drivers.
>
> Two lemmings added it to a function which can actually fail. Failure
> as well, because there is no guarantee that this function gets invoked
> before request_irq(). What's worse is that if request_irq() succeeds
> then the following code which tries to set the type fails for no
> obvious reason.
>
> Now at least Jean-Jacques tried to make it actually fail, but that
> turns out to be a failure on its own.
just to clarify: is the failure the problem with the unsigned or the
principle of making request_irq fail ?

>
> Dammit, I told you folks often enough that workarounds in some
> subsystem do not actually cure a shortcoming in the core code. When
> the core code was written, the GPIO case which might actually fail was
> definitly not thought of. But that's not a reason for adding
> tasteless and useless workarounds to the GPIO subsystem.
>
> Of course you encouraged people to copy that nonsense all over the
> place. All startup functions do the same thing:
>
>     if (gpio_lock_as_irq() < 0)
>        dev_err("BLA");
>     unmask_irq();
>
> Plus the shutdown functions having the gpio_unlock_as_irq() +
> mask_irq() sequence. Sigh.
>
> So the proper solution to the problem if we want to use irq_startup()
> is:
>
>   Change the return value of the irq_startup() callback to int and
>   fixup all existing implementations. This wants to be done with a
>   coccinelle script. If you hurry up, I might squeeze it into 3.14 as
>   there is no risk at all. Otherwise this is going to happen after the
>   3.15 merge window closes.
I'll post a patch doing just this in the next email.

Jean-Jacques
>
>   After that's done, remove all the copies of startup/shutdown
>   nonsense in drivers/gpio and force all gpio chips to call
>   gpio_set_irq_chip*() instead of irq_set_irq_chip*() and add generic
>   irq_startup/shutdown callbacks which are implemented in the gpiolib
>   core code.
>
> There is a less intrusive alternative solution:
>
>   Mark the interrupt as GPIO based in the core and let the core call
>   conditonally the gpio_[un]lock_as_irq functions.
>
> And no, I'm not going to provide you the proper solution on the silver
> tablet this time. Go and figure it out yourself and if you're
> confident enough that it might be an acceptable solution, post it and
> we'll see.
>
> Thanks,
>
>         tglx