[PATCH 1/6] watchdog: watchdog_dev: WATCHDOG_KEEP_ON feature

[Janusz Użycki]

W dniu 2014-09-30 06:37, Guenter Roeck pisze:
> On 09/29/2014 09:25 AM, Janusz Użycki wrote:
>>
>>> This patch set is trying to solve four problems at once:
>>>
>>> 1) Auto-start watchdog when its driver registers
>>> 2) Keep watchdog running when its driver registers until userspace 
>>> opens it
>>> 3) Handle watchdogs which can not be stopped after being started
>>> 4) Keep watchdog running with kernel timer after it has been closed,
>>>    even if it can be stopped.
>>>
>>> The next time adds 'boot time protection', which is really another term
>>> for an initial timeout, and case 5).
>>>
>>> That is a bit too much for a single patch and, even more so, a 
>>> single flag.
>>
>> OK, but I think [PATCH 3/6] could be applied.
>> Do you agree? Should I resent it separately?
>
> Yes, it looks ok.

Can you apply it?

>
>> I omited in the comment
>> "The patch adds suspend/resume PM support to stmp3xxx_rtc_wdt
>> watchdog driver" because the subject is almost the same.
>>
>>> Let's look at one case after another.
>>>
>>> Auto-start watchdog when its driver registers - this makes sense as a
>>> feature just by itself. A good name for its flag might be something 
>>> like
>>> WDT_AUTOSTART. A matching module parameter might also make sense.
>>>
>>> autostart:
>>>     Set to 0 to disable, -1 to enable with unlimited timeout,
>>>     or <n> for an initial timeout of <n> seconds.
>>
>> Current start(1) + keep-on(2,3,4) + boottime(5) combined. It looks OK.
>>
> Maybe for you. For me they are different cases.

They are different. I missed some words in the first sentence :)
However autostart automatically combine them together again :)
The problem is common timer so the patches are dependent.
There is no reason to use more timers.

>
>>>
>>> This could be accompanied by a variable in watchdog_device:
>>>     int init_timeout;    /* initial timeout in seconds */
>>
>> As the module parameter, instead of "boottime" in watchdog_core?
>>
>>>
>>> An API function such as watchdog_set_autostart() with the initial 
>>> timeout
>>> as parameter would also be helpful. This function could then be used to
>>> implement 2).
>>>
>>>     if (autostart || (keep_running && this_watchdog_is_running())
>>>         watchdog_set_autostart(&wdd, autostart ? : keep_running);
>>>
>> I don't understand the difference exactly and why to check the watchdog
>> is running? This means watchdog is active or something new?
>>
>>> keep_running could then be a another module parameter with the same 
>>> meaning
>>> as autostart.
>> But autostart and keep_running aren't in conflict.
>> So I don't understand also "autostart ? : keep_running".
>>
>
> The functionality is distinctively different.
>
> autostart: start watchdog on module load
> keep_running: If the watchdog is already running, keep it running. 
> Otherwise do nothing.
>
> Both have different use cases which should not be combined.

according to autostart value:
-1: current "keep-on" feature
 > 0: current "boottime" feature

Does watchdog_set_autostart() start/activate watchog?
Does keep_running differ from autostart=-1 "this_watchdog_is_running()" 
only?
How/where is this_watchdog_is_running() implemented?

What type is keep_running?
* bool
* 0/1
* -1/0 for watchdog_set_autostart()?

>
>>>
>>> Together this would also solve problem 5) while at the same time 
>>> keeping
>>> the use cases separate.
>>
>> It is solved by current code.
>>
>>>
>>> For 3) we really need another flag. Actually, it might be sufficient 
>>> to have
>>> watchdog drivers with this condition simply not provide a 'stop' 
>>> function.
>> or use "NOT SUPPORTED" error code in stop,
>> Stop could be called on register and new flag is set.
>>
>
> Seems to add complexity for no real benefit. Please explain why you think
> it is a good idea to have multiple drivers implement the same function
> just to return an error and do nothing else,

Specific driver knows if the watchdog is stoppable.
wddev->ops->stop() is called from watchdog_dev and if the stop() returns 
an error
watchdog_dev prints warning. It assumes wddev->ops->stop() is always 
implemented
even if not supported, ie. there is no condition. So either ENOSUP can 
be used
or the condition for optional (not mandatory) wddev->ops->stop().
Today a lot of drivers returns 0 only instead of ENOSUP.
So stop() could be called before autostart() to set the flag (code 
complexity)
or as you wrote just use the flag directly. The last one is indeed better.

>
>>> If we use a flag, something like WDOG_HW_NO_WAY_OUT with matching
>>> WATCHDOG_HW_NOWAYOUT might make sense. Its functionality is slightly
>>
>> What is difference between WDOG_HW_NO_WAY_OUT
>> and WATCHDOG_HW_NOWAYOUT?
>>
> Similar to other flags
>
> #define WDOG_HW_NO_WAY_OUT    5
> #define WATCHDOG_HW_NOWAYOUT    (1 << WDOG_HW_NO_WAY_OUT)

oh, right

>
>>> different to the other conditions: It would not auto-start a watchdog,
>>> but keep it running with the internal timer when the watchdog file 
>>> is closed.
>>>
>>> As for 4), I don't really know if it makes sense to have this 
>>> functionality.
>>
>> Yes, it is rootfs specific need. Script based code runs watchdog before
>> critical function and after exit the watchdog using magic char.
>> Critical section has timeout equal watchdog timeout value.
>> The feature allow to avoid userland application for watchdog
>> and does not cost much in the kernel.
>>
> Sorry, I can not follow your logic here. A basic userland implementation
> doesn't cost much either, is much safer, and even init systems such as
> systemd implement it nowadays.

True but not always for embedded systems where systemd is overweight today.
Let's notice that 4) works only if magic char is sent on exit.
Thanks to 4) magic char works also for non-stoppable watchdogs.

Janusz