Re: [PATCH 4/6] Export ns irqtimes from IRQ_TIME_ACCOUNTING through /proc/stat

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2010-10-21 at 12:25 -0700, Venkatesh Pallipadi wrote:
> > I'd do:
> >
> >  - hardirq
> >  - softirq
> >  - user
> >  - system
> >     - guest
> >     - really system
> >  - idle
> >
> > Since otherwise tiny slices of softirq would need to wait for a system
> > tick to happen before you fold them.
> >
> > Also, it is possible that in a single tick multiple counters overflow
> > the jiffy boundary, so something like:
> >
> >  if (irqtime_account_hi_update())
> >        cpustat->irq = ...
> >
> >  if (irqtime_account_si_update())
> >        cpustate->softirq = ...
> >
> >  if (user_tick) {
> >  } else if (...) {
> >
> >  } else ...
> >
> > would seem like the better approach.
> >
> 
> I am not sure about checking for both si and hi. That would result in
> double accounting a tick and have some side-effects.

Depends on how you look at it I guess, in order for this to occur a
previous tick would have to be not reported, eg. consider the case where
during two consecutive ticks the time is 50% for both sirq and hirq.

Then, after the first tick, nothing will have progressed because they're
both at 50% of a tick, after the second tick both will have reached a
full jiffy's worth of time and need to roll over.

In total two ticks happened, two ticks got accounted, {0,2}, your
approach would make it look like {0,1,1} two ticks worth of work
happened, two ticks got accounted, but it takes 3 ticks for that to
happen.

> Regarding moving si above user: Yes. That seems good.
> idle after system, That may not make so much of a difference, as there
> is no special way to check for system time, other than !idle. 

Right, so about user and system... we have a bit of a problem there.
There is overlap between si/hi and system. ksoftirqd time would be
accounted as system and si.

Then there is the whole issue of per-task accounting not actually using
the system/user ticks. They use the ticks as a ratio for
se.sum_exec_runtime.