Re: nanosecond debiting cv_timedwait

[Joerg Sonnenberger <joerg%britannica.bec.de@localhost>]

On Mon, Mar 23, 2015 at 07:36:52AM +0000, Taylor R Campbell wrote:
>    Date: Sat, 21 Mar 2015 16:20:32 +0100
>    From: Joerg Sonnenberger <joerg%britannica.bec.de@localhost>
> 
>    My consideration is for the future of when the callout handling itself
>    uses precise times. In that cacse, bintime does simplify quite a few
>    computation, since i.e. for differences, the overflow/underflow of the
>    sub-second field needs only binary ops to handle, no conditions. I'd
>    expect most timeouts in drivers to be constants, so the conversion of
>    ms to bintime can be done at compile time.
> 
> Is it easier for high-resolution timer hardware drivers to deal in
> nanoseconds or bintime?  I'm not familiar enough with the hardware to
> say.

The timer hardware is pretty much irrelevant, it seldomly runs at a nice
frequency as you said. Difference is for the kernel, where all arithmetic
is just plain shift, add, and (for overflow).

> In that case, we'd better also add ns2bintime, us2bintime, ms2bintime:
> 
> 	struct bintime timeout = us2bintime(100);
> 	int error;
> 
> 	while (!condition) {
> 		error = cv_bintimedwait(&sc->sc_cv, &sc->sc_lock, &timeout);
> 		if (error)
> 			goto fail;
> 	}
> 
> I'm OK with this, or with cv_timedwaitbt, if nobody else objects.

We have timespec and timeval conversions, direct ms/us/ns conversion
woudl be easy to add.

Joerg