Re: vmstat: On demand vmstat workers V4

[Paul E. McKenney]

On Sat, May 10, 2014 at 12:57:15AM +0200, Thomas Gleixner wrote:
> On Fri, 9 May 2014, Christoph Lameter wrote:
> > On Fri, 9 May 2014, Thomas Gleixner wrote:
> > > I understand why you want to get this done by a housekeeper, I just
> > > did not understand why we need this whole move it around business is
> > > required.
> > 
> > This came about because of another objection against having it simply
> > fixed to a processor. After all that processor may be disabled etc etc.
> 
> I really regret that I did not pay more attention (though my cycle
> constraints simply do not allow it).

As far as I can see, the NO_HZ_FULL timekeeping CPU is always zero.  If it
can change in NO_HZ_FULL kernels, RCU will do some very strange things!

One possible issue here is that Christoph's patch is unconditional.
It takes effect for both NO_HZ_FULL and !NO_HZ_FULL.  If I recall
correctly, the timekeeping CPU -can- change in !NO_HZ_FULL kernels,
which might be what Christoph was trying to take into account.

> This is the typical overengineering failure: 
> 
>   Before we even have a working proof that we can solve the massive
>   complex basic problem with the price of a dedicated housekeeper, we
>   try to make the housekeeper itself a moving target with the price of
>   making the problem exponential(unknown) instead of simply unknown.
> 
> I really cannot figure out why a moving housekeeper would be a
> brilliant idea at all, but I'm sure there is some magic use case in
> some other disjunct universe.
> 
> Whoever complained and came up with the NOT SO brilliant idea to make
> the housekeeper a moving target, come please forth and explain:
> 
> - How this can be done without having a working solution with a
>   dedicated housekeeper in the first place
> 
> - How this can be done without knowing what implication it has w/o
>   seing the complexity of a dedicated housekeeper upfront.
> 
> Keep it simple has always been and still is the best engineering
> principle.

If someone decides to make tick_do_timer_cpu non-constant in NO_HZ_FULL
CPUs, they will break unless/until I make RCU deal with that sort
of thing, at least for NO_HZ_FULL_SYSIDLE kernels.  ;-)

> We all know that we can do large scale overhauls in a very controlled
> way if the need arises. But going for the most complex solution while
> not knowing whether the least complex solution is feasible at all is
> outright stupid or beyond.
> 
> Unless someone comes up with a reasonable explantion for all of this I
> put a general NAK on patches which are directed to kernel/time/*
> 
> Correction:
> 
> I'm taking patches right away which undo any damage which has been
> applied w/o me noticing because I trusted the responsible developers /
> maintainers.
> 
> Preferrably those patches arrive before my return from LinuxCon Japan.

I could easily have missed something, but as far as I know, there is
nothing in the current kernel that allows tick_do_timer_cpu to move in
NO_HZ_FULL kernels.

Hmmm...  Well, I -do- have a gratuitous ACCESS_ONCE() around one fetch
of tick_do_timer_cpu that happens only in NO_HZ_FULL kernels.  I will
remove it.

							Thanx, Paul

> Thanks,
> 
> 	tglx
> 

--
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/