Re: [PATCH] simplify update_times (avoid jiffies/jiffies_64 aliasing problem)

[Atsushi Nemoto <anemo@mba.ocn.ne.jp>]

>>>>> On Sat, 4 Mar 2006 23:33:04 +1100, Paul Mackerras <paulus@samba.org> said:
>> > > Also I assume Atsushi-san did the patch because he saw a real problem?
>> > 
>> > Yes, one which I also saw on PPC.  The compiler (gcc-4) emits loads
>> > for jiffies, jiffies64 and wall_jiffies before storing the incremented
>> > jiffies64 value back.
>> > 
>> 
>> What was the effect of that?

paulus> The effect is that the first call to do_timer doesn't increment xtime.
paulus> This explains why the code I have to detect disagreements between
paulus> xtime and the time of day as computed from the timebase register was
paulus> finding a disagreement on the first tick, which I was scratching my
paulus> head over.

paulus> There may be other effects on architectures which use wall_jiffies to
paulus> detect lost timer ticks.  We don't have that problem on PPC and we
paulus> don't use wall_jiffies in computing time of day.

Well, I found this problem during investigating an another MIPS
do_gettimeofday() bug which was fixed recently.  It was:

	lost = jiffies - wall_jiffies;
	...
	} else if (unlikely(lost))
		usecs += lost * tick_usec;

The tick_usec should not be used here since it is based on USER_HZ.
It is 'usecs += lost * (USECS_PER_SEC /HZ)' now.

I wondered why this bug really affects gettimeofday() always though
the 'lost' should usually be 0.  And finally I noticed the
jiffies/jiffies_64 optimization issue.

---
Atsushi Nemoto