Re: alarm timer/timerfd expiration does not abort suspend operation

[Rafael J. Wysocki]

On Fri, Feb 10, 2017 at 8:23 PM, Gabriel Beddingfield <gabe@xxxxxxxxxxxx> wrote:
> Hi John,
>
> Re-sending because VGER rejected my hipster HTML mail... sorry!
>
> On Fri, Feb 10, 2017 at 10:49 AM, John Stultz <john.stultz@xxxxxxxxxx> wrote:
>>> I see a few ways to fix it:
>>>
>>> 1. Create a wakeup_source for each timerfd, and if it's an alarm timer
>>> call
>>> __pm_stay_awake() in timerfd_triggered() and __pm_relax() in
>>> timerfd_read().
>>> 2. call pm_system_wakeup() in alarmtimer_fired()
>>> 3. call `if (isalarm(ctx)) pm_system_wakeup();' in timerfd_triggered()
>>> 4. call __pm_wakeup_event(ws, 2 * MSECS_PER_SEC) in alarmtimer_fired()
>>> 5. call `if (isalarm(ctc)) __pm_wakeup_event(ws, 2 * MSECS_PER_SEC);'
>>> in
>>> timerfd_triggered() (using a static struct wakeup_source).
> [snip]
>>> * #4 Matches the current behavior of the "happy case" if and only if
>>> userspace is using the 'wakeup' system, otherwise doesn't change any
>>> behavior. But, I wonder how many people think the current behavior is a
>>> bug.
> [snip]
>> The approach you took in your patch looks basically ok to me, though I
>>
>> think the __pm_wakeup_event() method in #4 sounds safer, just to avoid
>> the problematic issue if no one is waiting on the fd.
>>
>> Though I worry I'm not quite understanding the con for that case
>> properly, so maybe you can clarify what concerns you there?
>
> The concern is born of my personal experience: I was ignorant of the
> "wakeup_count" protocol, and so I wasn't using it. Because of this
> __pm_wakeup_event() would not block a suspend because I never wrote to
> wakeup_count. On the other hand, method #2 will work unconditionally.

Right.

So if you want the timerfd behavior to not depend on whether or not
wakeup_count is used by user space, #2 is the way to go.

That said, creating a wakeup source for each timerfd would be nicer
from the diagnostics perspective, so maybe you can combine #1 and #2
such that if user space doesn't use wakeup_count, alarmtimer_fired()
will still abort suspends in a hard way?

Thanks,
Rafael