The "mull" instruction in __const_udelay() cuts off the lower 32 bits --
so, it is "rounding down".  This is both an issue for small ndelay()s for
_all_ values for loops_per_jiffy and for certain {n,u}delay()s for many
loops_per_jiffy values.

Assuming

LPJ = 1501115

udelay(87)

results in

130597 loops to be spent.

However, 1000 * 130597 / 1501115 is 86.999997 us, so we're actually
_rounding down_.  1000 * 130598 / 1501115 is 87.000662841, which would be
the technically correct thing to do.  Of course, for the TSC case this
won't matter as the maths take some time, so the actual delay is

1000 * __udelay(x) / lpj + __OVERHEAD(x)

Anybody worried about both the additional overhead and the fact that the
overhead takes some time to run should add a check

        if (unlikely(xloops < OVERHEAD))
                return;
        xloops -= OVERHEAD;

to the delay() routines in arch/i386/kernel/timers/*.c and determine
what the OVERHEAD is.
Signed-off-by: Dominik Brodowski <linux@brodo.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>