[RFC 00/20] ns: Introduce Time Namespace

[Thomas Gleixner]

Eric,

On Fri, 28 Sep 2018, Eric W. Biederman wrote:
> Thomas Gleixner <tglx at linutronix.de> writes:
> > On Wed, 26 Sep 2018, Eric W. Biederman wrote:
> >> At the same time using the techniques from the nohz work and a little
> >> smarts I expect we could get the code to scale.
> >
> > You'd need to invoke the update when the namespace is switched in and
> > hasn't been updated since the last tick happened. That might be doable, but
> > you also need to take the wraparound constraints of the underlying
> > clocksources into account, which again can cause walking all name spaces
> > when they are all idle long enough.
> 
> The wrap around constraints being how long before the time sources wrap
> around so you have to read them once per wrap around?  I have not dug
> deeply enough into the code to see that yet.

It's done by limiting the NOHZ idle time when all CPUs are going into deep
sleep for a long time, i.e. we make sure that at least one CPU comes back
sufficiently _before_ the wraparound happens and invokes the update
function.

It's not so much a problem for TSC, but not every clocksource the kernel
supports has wraparound times in the range of hundreds of years.

But yes, your idea of keeping track of wraparounds might work. Tricky, but
looks feasible on first sight, but we should be aware of the dragons.

> Please pardon me for thinking out load.
> 
> There are one or more time sources that we use to compute the time
> and for each time source we have a conversion from ticks of the
> time source to nanoseconds.
> 
> Each time source needs to be sampled at least once per wrap-around
> and something incremented so that we don't loose time when looking
> at that time source.
> 
> There are several clocks presented to userspace and they all share the
> same length of second and are all fundamentally offsets from
> CLOCK_MONOTONIC.

Yes. That's the readout side. This one is doable. But now look at timers.

If you arm the timer from a name space, then it needs to be converted to
host time in order to sort it into the hrtimer queue and at some point arm
the clockevent device for it. This works as long as host and name space
time have a constant offset and the same skew.

Once the name space time has a different skew this falls apart because the
armed timer will either expire late or early.

Late might be acceptable, early violates the spec. You could do an extra
check for rescheduling it, if it's early, but that requires to store the
name space time accessor in the hrtimer itself because not every timer
expiry happens so that it can be checked in the name space context (think
signal based timers). We need to add this extra magic right into
__hrtimer_run_queues() which is called from the hard and soft interrupt. We
really don't want to touch all relevant callbacks or syscalls. The latter
is not sufficient anyway for signal based timer delivery.

That's going to be interesting in terms of synchronization and might also
cause substantial overhead at least for the timers which belong to name
spaces.

But that also means that anything which is early can and probably will
cause rearming of the timer hardware possibly for a very short delta. We
need to think about whether this can be abused to create interrupt storms.

Now if you accept a bit late, which I'm not really happy about, then you
surely won't accept very late, i.e. hours, days. But that can happen when
settimeofday() comes into play. Right now with a single time domain, this
is easy. When settimeofday() or adjtimex() makes time jump, we just go and
reprogramm the hardware timers accordingly, which might also result in
immediate expiry of timers.

But this does not help for time jumps in name spaces because the timer is
enqueued on the host time base.

And no, we should not think about creating per name space hrtimer queues
and then have to walk through all of them for finding the first expiring
timer in order to arm the hardware. That cannot scale.

Walking all hrtimer bases on all CPUs and check all queued timers whether
they belong to the affected name space does not scale either.

So we'd need to keep track of queued timers belonging to a name space and
then just handle them. Interesting locking problem and also a scalability
issue because this might need to be done on all online CPUs. Haven't
thought it through, but it makes me shudder.

> I see two fundamental driving cases for a time namespace.

<SNIP>

I completely understand the problem you are trying to solve and yes, the
read out of time should be a solvable problem.

> For timers my inclination would be to assume no adjustments to the
> current time parameters and set the timer to go off then.   If the time
> on the appropriate clock has been changed since the timer was set and
> the timer is going off early reschedule so the timer fires at the
> appropriate time.

See above.

> Not that I think a final implementation would necessary look like what I
> have described. I just think it is possible with extreme care to evolve
> the current code base into something that can efficiently handle
> multiple time domains with slightly different lenghts of second.

Yes, it really needs some serious thoughts and timekeeping is a really
complex place especially with NTP/PTP in play. We had quite some quality
time to make it work correctly and reliably, now you come along and want to
transform it into a multidimensional puzzle. :)

> Thomas does it sound like I am completely out of touch with reality?

Which reality are you talking about? :)

> It does though sound like it is going to take some serious digging
> through the code to understand how what everything does and how and why
> everthing works the way it does.  Not something grafted on top with just
> a cursory understanding of how the code works.

I fully agree and I'm happy to help with explanations and ideas and being
the one who shoots holes into yours.

Thanks,

	tglx