Re: [PATCH] FS: timerfd: Fix unexpected return value of timerfd_read function.

[Arul Jeniston]

Hi tglx,

>Changing the return value to 1 would be just a cosmetic workaround.

Agreed. Returning 1 is incorrect as It causes the next read() to
return before the interval time passed.

>So I rather change the documentation (this applies only to CLOCK_REALTIME
>and CLOCK_REALTIME_ALARM) and explain the rationale.

When timerfd_read() returns 0, hrtimer_forward() doesn't change expiry
time, So, instead of modifying the man page, can we call
timerfd_read() functionality once again from kernel.

For example:-
timerfd_read_wrapper()
{
   do {
     ret = timerfd_read(...);
   } while (ret == 0);
}

Let us know whether you see any problem in handling this race in kernel.

Regards,
Arul


On Thu, Sep 5, 2019 at 9:04 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>
> Arul,
>
> On Thu, 5 Sep 2019, Arul Jeniston wrote:
> > When we adjust the date setting using date command we observed
> > 'timerfd_read()' on CLOCK_REALTIME (TFD_TIMER_ABSTIME flag is set)
> > returns 0.
> > we don't see any hardware influence here and we are able to recreate
> > it consistently. Is it expected? if yes, isn't it something to be
> > documented in timerfd read() man page?
>
> It's expected, yes. Simply because it hits the following condition:
>
>      armtimer(T1)
>
>      settime(T1 + X)  --> causes timer to fire
>
>                                  wakeup reader
>      settime(T0)
>
>                                  read number of intervals: 0
>
>                                  i.e. timer did not expire
>
> Changing the return value to 1 would be just a cosmetic workaround. We
> could also jump back and wait again. But that's all not consistent because
>
>      armtimer(T1)
>
>      settime(T1 + X)  --> causes timer to fire
>
>                                  wakeup reader
>
>                                  read number of intervals: 1
>      settime(T0)
>
>                                  user space reads time and figures that
>                                  the returned tick is bogus.
>
> So I rather change the documentation (this applies only to CLOCK_REALTIME
> and CLOCK_REALTIME_ALARM) and explain the rationale.
>
> For applications which care about notifications when the time was set,
> timerfd_settime() provides TFD_TIMER_CANCEL_ON_SET which causes the timer
> to be canceled when time is set and returns -ECANCELED from the
> read. That's unambiguous.
>
> Thanks,
>
>         tglx