My newyear's resolution is...

[Paul Herman]

So just to wrap up / sum up this thread, tvtohz() was fixed and
Matt came up with a good idea to add a sysctl kern.sleep_hardloop
which will force nanosleep() to try to be as accurate as possible at
the expense of a few CPU cycles.

Additionally, Matt added a check in i386/isa/clock.c to slow down
the HZ timer to compensate for the fact that the i8254 generally
doesn't have a high enough resolution to count out *exactly* 1/HZ
seconds worth of tick.  That seemed to keep the aliasing/drift
errors at bay.

Now DragonFly has a HZ timer that ticks as fast (and only as fast)
as it should.  Way to go Matt.

-Paul.