Re: Add pselect, ppoll system calls.

[Valdis.Kletnieks@vt.edu]

On Tue, 14 Jun 2005 16:07:54 +0200, =?ISO-8859-1?Q?Mattias Engdeg=E5rd?= said:
> >Monotonic clocks are guaranteed to not go backward. A sudden warp 35
> >seconds into the future when you have timers set for 15 and 20
> >seconds into the future is still ugly....
> 
> I don't have the POSIX specs handy, but I see no reason we could not let
> it use a warpless monotonic clock.

Right - I just mentioned it because "warpless" is an additional constraint
on the clock implementation.

The last time I looked at the NTP code, there was a definite upper limit to
how much warpless slewing of the clock was easily achievable - if you had a
large initial adjustment to make, you still have to either take one large
step, or be prepared to spend literally hours/days slewing the clock.

And then there's always the sysadmin issuing a 'date' command to manually
set the clock.  The kernel *still* needs to be able to handle warps correctly,
because I believe 'date' *is* in Posix? ;)

(The other possible way to handle this is the old IBM S/370 way - where setting
the system clock by warping it was only practical during system boot (as it was
done before any timers got launched), and required an operator intervention to
hit the 'TOD Clock Set Enable' button.  So the operator would look at their
watch, enter a time 15-20 seconds into the future, and whack the button at the
specified time).

Attachment: pgpaXM7wc5h0p.pgp
Description: PGP signature