Re: linux-next: manual merge of the tiny tree with the tip tree

[Josh Triplett <josh@joshtriplett.org>]

On Mon, Nov 24, 2014 at 10:30:10PM -0800, John Stultz wrote:
> On Mon, Nov 24, 2014 at 10:16 PM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > * Stephen Rothwell <sfr@canb.auug.org.au> wrote:
> >
> >> Hi Josh,
> >>
> >> Today's linux-next merge of the tiny tree got a conflict in
> >> kernel/time/Makefile between commit fd866e2b116b ("time: Rename
> >> udelay_test.c to test_udelay.c") from the tip tree and commit
> >> d1f6d68d03ea ("kernel: time: Compile out NTP support") from the tiny
> >> tree.
> >
> > So I think a timer subsystem commit d1f6d68d03ea with this
> > magnitude of linecount increase:
> >
> >  Signed-off-by: Catalina Mocanu <catalina.mocanu@gmail.com>
> >  [josh: Handle CONFIG_COMPAT=y.]
> >  Reviewed-by: Josh Triplett <josh@joshtriplett.org>
> >  Signed-off-by: Josh Triplett <josh@joshtriplett.org>
> >  ---
> >   drivers/pps/Kconfig        |  2 +-
> >   include/linux/timex.h      | 15 +++++++++++++--
> >   init/Kconfig               | 10 ++++++++++
> >   kernel/compat.c            |  8 ++++++--
> >   kernel/sys_ni.c            |  4 ++++
> >   kernel/time/Makefile       |  3 ++-
> >   kernel/time/ntp_internal.h | 28 ++++++++++++++++++++++++++++
> >   kernel/time/posix-timers.c |  2 ++
> >   kernel/time/time.c         |  2 ++
> >   kernel/time/timekeeping.c  |  2 ++
> >   10 files changed, 70 insertions(+), 6 deletions(-)
> >
> > at minimum needs the ack of timer folks, before it can be
> > committed to Git. Or is the tiny tree plan to submit all
> > patches to the appropriate subsystem or gather acks, before
> > sending it upstream?
> 
> Yea.  From first glance d1f6d68d03ea looks fairly broken.
> 
> Returning 0 for ntp_tick_length() (which should be the current tick
> length in NTP_SCALE_SHIFT shifted ns), seems like it would cause major
> timekeeping problems.

Ouch, yeah; I'm impressed the kernel successfully booted that way (which
I did test).

Computing the tick_length to return seems to require a div_u64; is it
safe to initialize a static const with the result of calling div_u64, or
does the intializer need manual constant-folding to make the expression
compile-time computable?

Going by the logic in ntp_update_frequency, it looks like the stub
ntp_tick_length needs to return:

tick_length_base = 0;
tick_usec = TICK_USEC;
second_length = (u64)(tick_usec * NSEC_PER_USEC * USER_HZ) << NTP_SCALE_SHIFT;
new_base = div_u64(second_length, NTP_INTERVAL_FREQ);
tick_length += new_base - tick_length_base;

(tick_length starts out 0, gets new_base - 0 added initially, and every
subsequent time gets 0 added since tick_length_base won't change.)

Substituting and simplifying:

tick_length = new_base
            = div_u64(second_length, NTP_INTERVAL_FREQ)
            = div_u64((TICK_USEC * NSEC_PER_USEC * USER_HZ) << NTP_SCALE_SHIFT, NTP_INTERVAL_FREQ)

The numerator there could potentially be simplified, but I don't see an
obvious way around the division by NTP_INTERVAL_FREQ (defined as HZ).

- Josh Triplett