Re: [PATCH] Consolidate multiple implementations of jiffies-msecsconversions.

From: Sridhar Samudrala
Date: Thu Mar 25 2004 - 18:54:47 EST

On Thu, 25 Mar 2004, Bill Rugolsky Jr. wrote:

> On Thu, Mar 25, 2004 at 12:17:41PM -0800, Sridhar Samudrala wrote:
> > The following patch to 2.6.5-rc2 consolidates 6 different implementations
> > of msecs to jiffies and 3 different implementation of jiffies to msecs.
> > All of them now use the generic msecs_to_jiffies() and jiffies_to_msecs()
> > that are added to include/linux/time.h
> I was inplementing precisely the same diff while cleaning up the
> select/poll timeout logic, when I came across this lkml post by George
> Anziger,
> the relevant part of which is:
> As to small drifts of ~170 PPM, they are caused by code (ps I would
> guess)
> that assumes that jiffies is exactly 1/HZ whereas it is NOT in the 2.6.*
> kernel. The size of the jiffie that the kernel uses is returned by:
> struct timespec tv;
> :
> :
> clock_res(CLOCK_REALTIME, &tv);
> I inferred from the above that a generic msec_to_jiffies()/jiffies_to_msec()
> ought to use TICK_NSEC, as with the other routines in time.h. For 32-bit
> platforms the scaled arithmetic is simple; one has to be more careful
> when BITS_PER_LONG == 64.
> After e-mailing George about it, he replied:
> You might want to look at the code in time.h that does the jiffies to
> timespec conversion. We did a lot of work to get that right for both 32 and
> 64 bit platforms. I don't think you really need more precision than we get
> on the 32 bit platforms (if I recall correctly it is in the 100 PPB range,
> yeah that is parts per BILLION).
> Unless this conversion is done a lot, I would just start with the timespec
> conversions and convert from / to using simple math. As it is all power of
> 10 stuff it should not have a precision problem.
> He also supplied me with his test harness.
> I haven't gotten around to doing this properly yet. It seems that the
> only place where precision is actually important (due to the possibility
> of very long timeouts) is in poll/epoll, so perhaps it is best
> to just code up a special version for them, as the simple version
> becomes a no-op everywhere else for the default HZ==1000.

Do poll/epoll also use similar routines to do msecs/jiffies conversions?
If so, i seem to have missed them.
But i think for all the other users which i have consolidated, the simpler
version should be good enough. In fact, some of the existing implementations
do overflow for large values.

Once this patch is integrated, I guess you can either update the generic
version to do a more precise calculation or create a special version for

> Regards,
> Bill Rugolsky
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at