We were getting the last_jiffy per-cpu variable set ahead of the current
timebase in smp_space_timers on SMP machines.  This caused the loop in
timer_interrupt to loop virtually forever, since tb_ticks_since assumes
that it will never be called with the timebase behind the last_jiffy
value.
Signed-off-by: Paul Mackerras <paulus@samba.org>
        unsigned long offset = tb_ticks_per_jiffy / max_cpus;
        unsigned long previous_tb = per_cpu(last_jiffy, boot_cpuid);
 
+       /* make sure tb > per_cpu(last_jiffy, cpu) for all cpus always */
+       previous_tb -= tb_ticks_per_jiffy;
        for_each_cpu(i) {
                if (i != boot_cpuid) {
                        previous_tb += offset;