Time bug on 2.2.* kernels

[Henry Cejtin]

The bug I mentioned before, about the race when the counter wraps, is easy to
produce on 2.4.2 kernels with SMP machines.   I  tried  to  see  if  I  could
produce the problem with 2.2 kernels.  I assumed that I could not, since they
had much much less multi-threading, but I seem to recall that  we  have  seen
this problem on such machines.

I  ran  the  C  grinder on a 2.2.14 machine with 2 CPUs, and although it took
much longer, I got the bug there, but slightly differently.  On that machine,
when the wrap happens you get back a tv_usec field which contains 10^6.  This
is actually better, since you could work around it, but I don't  see  at  all
how this is happening since the kernel code, after expanding the macros, is
    res.ru_stime.tv_sec = <utime-counter> / 100;
    res.ru_stime.tv_use = (<utime-counter> % 100) * 10000;
so it should have the same bug, and should never ever return a tv_usec larger
than 990,000.