Re: [RFC][PATCH 4/4] time: Do leapsecond adjustment in gettime fastpaths

[John Stultz]

On Sat, Jun 6, 2015 at 2:44 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Fri, 5 Jun 2015, Peter Zijlstra wrote:
>
>> On Fri, 2015-06-05 at 11:04 +0200, Richard Cochran wrote:
>> > On Fri, Jun 05, 2015 at 09:29:13AM +0200, Peter Zijlstra wrote:
>> > > That leaves the question; for who is this exact second edge important?
>> >
>> > Distributed applications using the UTC time scale.
>> >
>> > Many control applications are done with a 1 millisecond period.
>> > Having the time wrong by a second for 10 or 100 loops is bad news.
>>
>> Firstly I would strongly suggest such applications not use UTC because
>> of this, I think TAI was invented for just this reason.
>>
>> Secondly, how would John's patches help with this? Usespace loops
>> reading time would be using the VDSO and would still not get the right
>> time, and timers would be subject to the same IRQ latency that a hrtimer
>> based leap second insert would, and would still very much not be in-sync
>> across the cluster.
>
> So the only thing which is fixed are in kernel users and therefor
> hrtimers.

Well, for vdso systems, hrtimers and adjtimex (which is the only
interface that provides enough information to understand where in a
leapsecond you actually are).

And again, vdsos are fixable, but I hesitated due to my concerns about
the extra performance overhead, the smaller benefit it provides
relative to not having timers expiring early.

> That means the whole leap mess added into the gettime fast path is
> just constant overhead for that corner case.
>
> We can be smarter than that and just block hrtimer delivery for clock
> realtime timers across the leap edge. There should be ways to do that
> non intrusive if we think hard enough about it.

This approach seems like it would add more work to the timer-add
function (to check leapstate and adjust the expiration), which might
be a heavier use case (we adjust for each timer) then the similar
logic done in the update_base_offsets_now() at hrtimer_interrupt time
(adjust for each hrtimer_interrupt).

Now, It could be possible to do a lighter weight version of my patch,
which just does the adjustment only for the hrtimer_interrupt code
(leaving the rest of the read paths alone).  If that is something
you'd prefer.  I'll have to think a bit to ensure the internal
inconsistency isn't otherwise problematic.

Though I suspect fixing adjtimex is worth it as well, since its really
the only interface that can provide a sane view of the leapsecond, and
isn't normally considered a hot path.

thanks
-john
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/