Re: [PATCH] alarmtimer: return EINVAL instead of ENOTSUPP if rtcdev doesn't exist

[KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>]

On 10/18/2013 6:39 PM, John Stultz wrote:
> On 10/17/2013 06:12 PM, KOSAKI Motohiro wrote:
>> (10/17/13 1:05 PM), John Stultz wrote:
>>> On 10/14/2013 02:33 PM, kosaki.motohiro@gmail.com wrote:
>>>> From: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
>>>>
>>>> Fedora Ruby maintainer reported latest Ruby doesn't work on Fedora
>>>> Rawhide
>>>> on ARM. (http://bugs.ruby-lang.org/issues/9008)
>>>>
>>>> Because of, commit 1c6b39ad3f (alarmtimers: Return -ENOTSUPP if no
>>>> RTC device is present) intruduced to return ENOTSUPP when
>>>> clock_get{time,res} can't find a RTC device. However it is incorrect.
>>>>
>>>> Posix and Linux man pages agree that clock_gettime and clock_getres
>>>> should return EINVAL if clk_id argument is invalid. This is significant
>>>> different from timer_create API.
>>>>
>>>> This patch fixes it.
>>>
>>> Hrm... So I feel like there is a difference here. The clockid for
>>> CLOCK_BOOTTIME_ALARM and CLOCK_REALTIME_ALARM are both valid.
>>>
>>> Its just that they're not supported on this specific hardware because it
>>> apparently lacks a RTC that has told the system it can be used as a
>>> wakeup device (Its actually quite likely on the hardware that the RTC
>>> can be a wakeup device, but that the driver is probably setting the
>>> wakeup flag after the RTC registered - so there is probably a driver bug
>>> here too).
>>>
>>> So I feel like in this case EINVAL isn't quite right.  I'll admit it is
>>> somewhat new behavior, because we haven't had any clockids before that
>>> were dependent on the particular hardware, they either existed in a
>>> kernel verison or didn't.
>>>
>>> Would updating the manpage be a better route?
>>
>> Nope.
>>
>> ENOTSUPP is not exported to userland. ENOTSUP (single P) and EOPNOTSUP is
>> valid errno (and they are same on linux), but ENOTSUPP is a kernel
>> internal specific.
>>
>> Moreover, I completely disagree your position. Both
>> CLOCK_REALTIME_ALARM unsupported
>> kernel and ARM which doesn't support RTC should use the same error
>> because application
>> need the same fallback.
> 
> Ok. You're right. The technicality that the clockid is valid but
> unsupported isn't really useful to the applications, since the effect is
> the same.
> 
> What is the urgency on this? As the issue has been around since 3.0, is
> it ok if it gets queued for 3.13 and marked for stable, or does it need
> to land in 3.12?

3.13 is OK to me.