Re: [Y2038] [PATCH v3] Staging: media: Replace timeval with ktime_t

[Mauro Carvalho Chehab <mchehab@osg.samsung.com>]

Em Wed, 13 May 2015 14:45:38 -0700
John Stultz <john.stultz@linaro.org> escreveu:

> On Wed, May 13, 2015 at 2:10 PM, Mauro Carvalho Chehab
> <mchehab@osg.samsung.com> wrote:
> > Em Wed, 13 May 2015 21:53:07 +0200
> > Arnd Bergmann <arnd@arndb.de> escreveu:
> >
> >> On Wednesday 13 May 2015 10:04:48 John Stultz wrote:
> >> > On Wed, May 13, 2015 at 9:57 AM, Ksenija Stanojevic
> >> > <ksenija.stanojevic@gmail.com> wrote:
> >> > > 'struct timeval last_tv' is used to get the time of last signal change
> >> > > and 'struct timeval last_intr_tv' is used to get the time of last UART
> >> > > interrupt.
> >> > > 32-bit systems using 'struct timeval' will break in the year 2038, so we
> >> > > have to replace that code with more appropriate types.
> >> > > Here struct timeval is replaced with ktime_t.
> >> > >
> >> > > Signed-off-by: Ksenija Stanojevic <ksenija.stanojevic@gmail.com>
> >>
> >> An additional comment: as drivers/staging/media refers to a whole subsystem
> >> with mutually independent drivers, the subject line should mention 'lirc',
> >> either in addition to, or instead of 'media'.
> >>
> >> > > -static long delta(struct timeval *tv1, struct timeval *tv2)
> >> > > +static inline long delta(ktime_t t1, ktime_t t2)
> >> > >  {
> >> > > -       unsigned long deltv;
> >> > > -
> >> > > -       deltv = tv2->tv_sec - tv1->tv_sec;
> >> > > -       if (deltv > 15)
> >> > > -               deltv = 0xFFFFFF;
> >> > > -       else
> >> > > -               deltv = deltv*1000000 +
> >> > > -                       tv2->tv_usec -
> >> > > -                       tv1->tv_usec;
> >> > > -       return deltv;
> >> > > +       /* return the delta in 32bit usecs, but cap to UINTMAX in case the
> >> > > +        * delta is greater then 32bits */
> >> > > +       return (long) min((unsigned int) ktime_us_delta(t1, t2), UINT_MAX);
> >> > >  }
> >> >
> >> > This probably needs some close review from the media folks. Thinking
> >> > about it more, I'm really not certain the 15sec cap was to avoid a
> >> > 32bit overflow or if there's some other subtle undocumented reason.
> >>
> >> The new code is clearly wrong, as the cast to 'unsigned int' already truncates
> >> the value to at most UINT_MAX, and the min() does not have any effect.
> >>
> >> The correct way to write what was intended here is
> >>
> >>       return min_t(long long, ktime_us_delta(t1, t2), UINT_MAX);
> >>
> >> which will truncate delta to an unsigned integer. The return type of the
> >> delta() function would need to be changed to 'unsigned long' as well to
> >> make this work.
> >>
> >> However, I think you are right that we should probably not change the
> >> behavior, unless someone who understands the purpose better can say
> >> what it really should be.
> >
> > Inside the remote controller code, we have measurements for pulse/space
> > encodings on a IR transmission. The duration of a pulse or space is
> > generally in the other of microseconds. On the standard protocols, the
> > maximum duration is on NEC protocol, where a pulse of 9 ms is sent at
> > the beginning:
> >         http://www.sbprojects.com/knowledge/ir/nec.php
> >
> > It should be noticed that bigger time intervals can be used to indicate
> > key repeat. Again, in the NEC protocol, the space between key repeats
> > are 110 ms.
> >
> > So, everything above 110 ms is actually an infinite time.
> >
> > As the Kernel implementation was built to be generic enough, we consider
> > (u32)-1 (e. g. about 4 seconds) as the maximum possible time.
> 
> So that's ~4 seconds of nanoseconds, but since we're talking usecs,
> UINT_MAX is something like 70-some minutes.

Well this is a staging driver. The ones outside staging use nanoseconds
internally.

Btw, on almost all non-staging drivers, the timer is provided by the
hardware, and not measured using do_gettimeofday (or similar).

The drivers at staging are actually for the older legacy remote
controllers. Those drivers are there basically because nobody
cared enough to convert them to use the Remote Controller subsystem.

I wouldn't mind if we decide to remove those drivers due to Y2038
changes ;)

> > This is due to the fact that some IR protocols use u32 for the pulse/space
> > time shifts. So, any duration bigger than that could actually be
> > rounded to (u32)-1.
> 
> Sure. The part that confused me is that the delta function is checking
> if the second delta is larger then 15 seconds, and I couldn't quite
> understand the significance of that check. If it really is just to
> make sure the timeval -> usec conversion doesn't overflow, then that's
> easy enough to solve. But if the 15 second check has some other
> meaning, we'd want to understand before changing this.

It is likely some maximum time. As I said, we're using ~4 second
checks internally at the Kernel.

I can't find any usage for a remote controller that would have a 15
seconds of timeout... actually, even 4 seconds is a looooong time ;)

> > That's said, I really don't see the need of "fixing" it on the y2038
> > patchset. All that it is needed is to warrant that the time difference
> > will be positive.
> >
> > I would, instead, remove the delta function, and replace:
> >
> >         do_gettimeofday(&curr_tv);
> >         deltv = delta(&last_tv, &curr_tv);
> >
> > (and other equivalent parts)
> >
> > By an equivalent logic that would be reading the timestamp from a
> > high precision clock.
> >
> > That's said, I suspect that this driver is broken, as I doubt that
> > do_gettimeofday() gets enough precision needed for IR decoding. Also,
> > as this returns a non-monotonic timestamp, it will break if one adjusts
> > the clock while IR keys are being pressed.
> 
> Yep. Moving to ktime_get should help since it uses the monotonic clock.

Yep.

Regards,
Mauro