Re: [PATCH] [RFC] Potential fix for leapsecond caused futex related load spikes

[Sytse Wielinga <sytse@swielinga.nl>]

On Tue, Jul 03, 2012 at 11:23:25AM +0200, Richard Cochran wrote:
> I think the established practice of announcing the event by network is
> the only sane way of handling this issue. The list of TAI-UTC offsets
> belongs to what David Mills has called our "institutional memory", and
> this is a user space issue. The kernel's job is to just live in the
> moment and provide the right time for *now*.

I do suppose hardware clock and file system times will have to be UTC (or
UTC-based local time) though?  Or do you think the 35 seconds difference
simply will be so small of a problem that it's not worth fussing over?

Doing this translation in libc and keeping fs times in TAI+tz offset would
seem to necessitate at least modifications to every single program accessing
file system data directly, and would still cause minor problems with
multiboot; doing it in the kernel would mean adding a step to the boot
sequence (before mounting root r/w) for loading the current TAI-UTC difference
into the kernel; also, it'd mean splitting 'kernel time' into multiple times.
Which solution did you have in mind?

> > > There have been a number of clock/timer/leap bugs over the last
> > > years. Some of these might have been avoided by using a continuous
> > > scale, since no special timer actions would be needed during a leap
> > > second.
> > > 
> > > The run time cost is low, just one additional test and addition when
> > > reading the time. It might be worth it for the peace of mind when
> > > the next leap second rolls around.
> > 
> > I don't know if reworking the system that's been in place for ages is a good
> > way to give us 'peace of mind'.  Then again, I love to be enlightened :-)
> 
> There have been lockups and other kernel issues due to leap second
> bugs. That is a fact. Does that give you peace of mind?
> 
> My own computers were off for the last leap second. But some people
> cannot afford to do this. I suggest that changing the code so that no
> special actions occur at a leap second would be more reliable than
> having rarely tested code paths just for leap second handling.

I suppose you're right; the new code might be buggy, but at least it'd get
year-round testing instead of just once every few years or so.

Then again, you and John have come up with a good regression test.

Sytse