Re: [PATCH 2/4] alarmtimer: Make alarmtimer platform device child of RTC device

[Stephen Boyd]

Quoting Doug Anderson (2020-01-15 11:22:26)
> Hi,
> 
> On Wed, Jan 15, 2020 at 2:07 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > Doug Anderson <dianders@xxxxxxxxxxxx> writes:
> > > On Thu, Jan 9, 2020 at 7:59 AM Stephen Boyd <swboyd@xxxxxxxxxxxx> wrote:
> > >> diff --git a/kernel/time/alarmtimer.c b/kernel/time/alarmtimer.c
> > >> index 4b11f0309eee..ccb6aea4f1d4 100644
> > >> --- a/kernel/time/alarmtimer.c
> > >> +++ b/kernel/time/alarmtimer.c
> > >> @@ -88,6 +88,7 @@ static int alarmtimer_rtc_add_device(struct device *dev,
> > >>         unsigned long flags;
> > >>         struct rtc_device *rtc = to_rtc_device(dev);
> > >>         struct wakeup_source *__ws;
> > >> +       struct platform_device *pdev;
> > >>         int ret = 0;
> > >>
> > >>         if (rtcdev)
> > >> @@ -99,6 +100,7 @@ static int alarmtimer_rtc_add_device(struct device *dev,
> > >>                 return -1;
> > >>
> > >>         __ws = wakeup_source_register(dev, "alarmtimer");
> > >> +       pdev = platform_device_register_data(dev, "alarmtimer", -1, NULL, 0);
> > >
> > > Don't you need to check for an error here?  If pdev is an error you'll
> > > continue on your merry way.  Before your patch if you got an error
> > > registering the device it would have caused probe to fail.
> >
> > Yes, that return value should be checked
> >
> > > I guess you'd only want it to be an error if "rtcdev" is NULL?
> >
> > If rtcdev is not NULL then this code is not reached. See the begin of
> > this function :)
> 
> Wow, not sure how I missed that.  I guess the one at the top of the
> function is an optimization, though?  It's being accessed without the
> spinlock which means that it's not necessarily reliable, right?  I
> guess once the rtcdev has been set then it is never unset, but it does
> seem like if two threads could call alarmtimer_rtc_add_device() at the
> same time then it's possible that we could end up calling
> wakeup_source_register() for both of them.  Did I understand that
> correctly?  If I did then maybe it deserves a comment?

It also looks like we call wakeup_source_register() and get lucky that
they're both called alarmtimer but they don't conflict with each other
in sysfs. Once we try to add a device named "alarmtimer" to the platform
bus it is the only one that can be added. I'll have to make this
autogenerate some number for the device instead of using -1 as the id so
that we don't see device name conflicts on the same bus.