Re: [RFC 3/4] watchdog: ni9x3x_wdt: Add timeout_action sysfs attribute

[Guenter Roeck <linux@roeck-us.net>]

On 02/05/2016 01:13 PM, Kyle Roeschley wrote:
> On Fri, Feb 05, 2016 at 02:08:03PM -0600, Josh Cartwright wrote:
>> On Thu, Feb 04, 2016 at 07:28:02PM -0600, Kyle Roeschley wrote:
>> [..]
>>> diff --git a/drivers/watchdog/ni9x3x_wdt.c b/drivers/watchdog/ni9x3x_wdt.c
>>> index 13f5c10..1fce4d0 100644
>>> --- a/drivers/watchdog/ni9x3x_wdt.c
>>> +++ b/drivers/watchdog/ni9x3x_wdt.c
>>> @@ -15,6 +15,7 @@
>>>   #include <linux/acpi.h>
>>>   #include <linux/interrupt.h>
>>>   #include <linux/module.h>
>>> +#include <linux/sysfs.h>
>>>   #include <linux/watchdog.h>
>>>
>>>   #define NIWD_CONTROL	0x01
>>> @@ -28,6 +29,7 @@
>>>   #define NIWD_IO_SIZE	0x08
>>>
>>>   #define NIWD_CONTROL_MODE		0x80
>>> +#define NIWD_CONTROL_PROC_INTERRUPT	0x40
>>>   #define NIWD_CONTROL_PROC_RESET		0x20
>>>   #define NIWD_CONTROL_PET		0x10
>>>   #define NIWD_CONTROL_RUNNING		0x08
>>> @@ -48,6 +50,7 @@ struct ni9x3x_wdt {
>>>   	struct acpi_device *acpi_device;
>>>   	u16 io_base;
>>>   	u16 io_size;
>>> +	u32 irq;
>>
>> Interrupts have type 'int'.
>>
>>>   	spinlock_t lock;
>>>   	struct watchdog_device wdog;
>>>   };
>>> @@ -101,6 +104,34 @@ static unsigned int ni9x3x_wdt_wdd_get_timeleft(struct watchdog_device *wdd)
>>>   	return (unsigned int)((counter * (u64)NIWD_PERIOD_NS) / 1000000000);
>>>   }
>>>
>>> +static irqreturn_t ni9x3x_wdt_irq_handler(int irq, void *data)
>>> +{
>>> +	struct ni9x3x_wdt *wdt = data;
>>> +	irqreturn_t ret = IRQ_NONE;
>>> +	u8 control;
>>> +	unsigned long flags;
>>> +
>>> +	spin_lock_irqsave(&wdt->lock, flags);
>>
>> Interrupts are disabled already, so the _irqsave()/_irqrestore() is
>> unnecessary.
>>
>>> +
>>> +	control = inb(wdt->io_base + NIWD_CONTROL);
>>> +
>>> +	if (!(NIWD_CONTROL_ALARM & control)) {
>>> +		dev_err(&wdt->acpi_device->dev,
>>> +			"Spurious watchdog interrupt, 0x%02X\n", control);
>>
>> No need to print anything here, just return IRQ_NONE.  The irq core will
>> print a much better (and rate-limited) error when/if this occurs.
>>
>>> +		goto out_unlock;
>>> +	}
>>> +
>>> +	/* Acknowledge the interrupt. */
>>> +	outb(control | NIWD_CONTROL_RESET, wdt->io_base + NIWD_CONTROL);
>>> +
>>> +	ret = IRQ_HANDLED;
>>> +
>>> +out_unlock:
>>> +	spin_unlock_irqrestore(&wdt->lock, flags);
>>> +
>>> +	return ret;
>>> +}
>>
>> This is...strange.  You're allowing a user to enable an interrupt reset
>> action, but...there is no way that a user can actually _do_ anything
>> when that action occurs.
>>
>> I'm reminded of one of these:
>>
>>    https://www.youtube.com/watch?v=aqAUmgE3WyM
>>
>> (The only reason I can think of where this would be a legitimate thing
>> to do was if we were relying on the watchdog interrupt to bring the CPU
>> out of a low power state...but AFAIK that's not something we do).
>>
>>    Josh
>
> This is another reason (that I forgot to mention) for this being an RFC.
> Ideally we'd poll or select the /dev/watchdogN file to wait on interrupt, but
> this isn't supported by the watchdog core. My next instinct would be to have an
> sysfs attribute named interrupt or event which a user could read/poll/select
> to look for a 1 value indicating that an interrupt has occurred. Else, maybe
> polling support could be added to the core, but I don't expect it would be very
> useful.
>

I _really_ get the impression that you are trying to use the watchdog
subsystem for something it isn't supposed to support. poll/select on the
watchdog device is pretty much the opposite of what should happen.
the watchdog device node is intended to be written to to trigger a
heartbeat, not a device to use to wait for an interrupt. There is
no way we'll ever add support for this into the watchdog subsystem.

I'll really have to spend much more time on this, probably much more than
I have. I would suggest to either implement a traditional, standard,
watchdog driver, or consider implementing a non-watchdog driver if
your use case is that much different.

Thanks,
Guenter