Re: [PATCH v2 1/3] process cputimer is moving faster than its corresponding clock

[Olivier Langlois <olivier@trillion01.com>]

On Mon, 2013-04-29 at 13:10 -0400, Olivier Langlois wrote:
> On Mon, 2013-04-29 at 01:06 -0400, KOSAKI Motohiro wrote:
> > (4/27/13 12:40 AM), Olivier Langlois wrote:
> > > 
> > > 
> > > Forbids the cputimer to drift ahead of its process clock by
> > > blocking its update when a tick occurs while a autoreaping task
> > > is currently in do_exit() between the call to release_task() and
> > > its final call to schedule().
> > > 
> > > Any task stats update after having called release_task() will
> > > be lost because they are added to the global process stats located
> > > in the signal struct from release_task().
> > > 
> > > Ideally, you should postpone the release_task() call after the
> > > final context switch to get all the stats added but this is
> > > more complex to achieve.
> > > 
> > > In other words, this is slowing down the cputimer so it keep the same
> > > pace than the process clock but in fact, what should be done is to
> > > speed up the process clock by adding the missing stats to it.
> > > 
> > > Signed-off-by: Olivier Langlois <olivier@trillion01.com>
> > > ---
> > >  kernel/sched/fair.c | 10 +++++++++-
> > >  1 file changed, 9 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 7a33e59..52d7b10 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -708,7 +708,15 @@ static void update_curr(struct cfs_rq *cfs_rq)
> > >  
> > >                 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
> > >                 cpuacct_charge(curtask, delta_exec);
> > > -               account_group_exec_runtime(curtask, delta_exec);
> > > +               /*
> > > +                * Do not update the cputimer if the task is already released by
> > > +                * release_task().
> > > +                *
> > > +                * it would preferable to defer the autoreap release_task
> > > +                * after the last context switch but harder to do.
> > > +                */
> > > +               if (likely(curtask->sighand))
> > > +                       account_group_exec_runtime(curtask, delta_exec);
> > >         }
> > 
> > I'm confused. glibc's rt/tst-cputimer1 doesn't have thread exiting code. I have
> > no seen any issue in this accounting.
> 
> glibc launch a helper thread to receive timer signal and will also
> create a new thread upon signal reception when a timer is created with
> sigev_notify = SIGEV_THREAD;
> 
> please see:
> 
> glibc-2.17/nptl/sysdeps/unix/sysv/linux/timer_create.c
> glibc-2.17/nptl/sysdeps/unix/sysv/linux/timer_routines.c
> 
One very easy way to see the problem is to add a printk statement inside
update_gt_cputime()

        if (b->sum_exec_runtime > a->sum_exec_runtime)
                a->sum_exec_runtime = b->sum_exec_runtime;
	else
		printk( KERN_DEBUG "cputimer %llu, process clock %llu, diff %llu\n",
a->sum_exec_runtime, b->sum_exec_runtime,
a->sum_exec_runtime-b->sum_exec_runtime);

Check the output with and without the fair.c modif when executing
tst-cputimer1. As an extra bonus, this trace will show the spurious
start/stop cputimer problem that I was trying to explain to Frederic.