Re: [PATCH 09/10] s390/cputime: delayed accounting of system time
From: Frederic Weisbecker
Date: Tue Dec 13 2016 - 21:38:53 EST
On Tue, Dec 13, 2016 at 02:21:21PM +0100, Martin Schwidefsky wrote:
> On Tue, 13 Dec 2016 12:13:22 +0100
> Martin Schwidefsky <schwidefsky@xxxxxxxxxx> wrote:
> > On Mon, 12 Dec 2016 16:02:30 +0100
> > Frederic Weisbecker <fweisbec@xxxxxxxxx> wrote:
> > > On Mon, Dec 12, 2016 at 11:27:54AM +0100, Martin Schwidefsky wrote:
> > > > 3) The call to vtime_flush in account_process_tick is done in irq context from
> > > > update_process_times. hardirq_offset==1 is also correct.
> > >
> > > Let's see this for example:
> > >
> > > + if ((tsk->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
> > > + S390_lowcore.guest_timer += timer;
> > >
> > > If the tick is interrupting guest, we have accounted the guest time on tick IRQ entry.
> > > Now we are in the middle of the tick interrupt and since hardirq_offset is 1, we
> > > are taking the above path by accounting half of the tick-IRQ time as guest, which is wrong,
> > > it's actually IRQ time.
> > Hmm, you got me there. The system time from irq_enter until account_process_tick
> > is reached is indeed IRQ time. It is not much but it is incorrect. The best fix
> > would be to rip out the accounting of the system time from account_process_tick
> > as irq_enter / irq_exit will do system time accounting anyway. To do that
> > do_account_vtime needs to be split, because for the task switch we need to
> > account the system time of the previous task.
> New patch for the delayed cputime account. I can not just rip out system time
> accounting from account_process_tick after all, I need a sync point for the
> steal time calculation. It basically is the same patch as before but with a new
> helper update_tsk_timer, the removal of hardirq_offset and a simplification
> for do_account_vtime: the last accounting delta is either hardirq time for
> the tick or system time for the task switch.
> Keeping my finger crossed..
The patch looks good. But you might want to remove the hardirq_offset in a
separate patch to queue for this merge window (I'm not sure if it needs a
stable tag, the argument may be there since the beginning).
Because the rest depends on the series that is unlikely to be queued in this
merge window at this stage.