A review of 'arch/parisc/kernel/time.c' and in particular
'timer_interrupt' and 'gettimeoffset'  show that this code has a
couple of design flaws.

1. cr16 is treated as signed.
2. No attempt is made to handle cr16 wrapping.
3. jiffies vs. wall_jiffies adjustment is incorrect.

I have rewritten 'timer_interrupt'.

A. cr16 is treated as an unsigned long.
B. The following 3 scenarios exist.

1. The timer interrupt fires, and 'now' comes after 'next_tick'
2. The timer interrupt fires, and 'now' has wrapped and is before 'next_tick'
3. The timer interrupt fires, and 'now' is *just* before 'next_tick'

In theory 99% of the time we should be in scenario 1. On occasion we
may miss a timer interrupt, end up close to wrapping, and we get
scenario 2. Scenario 3 is just plain wrong and for now I will BUG() if
the timer fires before we intended.

I have adjusted 'gettimeoffset' in the following fashion:

A. cr16 is treated as an unsigned long.
B. Never ever ever return a negative adjustment.
C. (jiffies - wall_jiffies) adjustment is always positive and added to usec.
D. The 3 timer scenarios (same as above), exist.

I assert that 'gettimeoffset' should never return a negative value. It
represents the postive adjustment accounting for the fact that we are
*part* of the way through a tick.

The patch is attached. Please review. I booted this on my a500 without
any problems. I don't really know how to look for problems so I ran
the glibc testsuite and didn't get any extra regressions.

The real test will be to run this on a 715/50 and look for Guy's
reported problems. Guy, would you mind patching your kernel with this
patch and testing again?

Cheers,
Carlos.