Re: [PATCH 02/25] sched/vtime: Protect idle accounting under vtime seqcount

[Peter Zijlstra]

On Wed, Nov 14, 2018 at 03:45:46AM +0100, Frederic Weisbecker wrote:
> Locking the seqcount on idle vtime accounting wasn't thought to be
> necessary because the readers of idle cputime don't use vtime (yet).
> 
> Now updating vtime expect the related seqcount to be locked so do it
> for locking coherency purposes.
> 
> Also idle cputime updates use vtime, but idle cputime readers use the
> traditional ad-hoc nohz time delta. We may want to consider moving
> readers to use vtime to consolidate the overall accounting scheme. The
> seqcount will be a functional requirement for it.
> 
> Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> Cc: Yauheni Kaliuta <yauheni.kaliuta@xxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: Rik van Riel <riel@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Wanpeng Li <wanpengli@xxxxxxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> ---
>  kernel/sched/cputime.c | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
> index 54eb945..6e74ec2 100644
> --- a/kernel/sched/cputime.c
> +++ b/kernel/sched/cputime.c
> @@ -800,7 +800,11 @@ EXPORT_SYMBOL_GPL(vtime_guest_exit);
>  
>  void vtime_account_idle(struct task_struct *tsk)
>  {
> +	struct vtime *vtime = &tsk->vtime;
> +
> +	write_seqcount_begin(&vtime->seqcount);
>  	account_idle_time(get_vtime_delta(&tsk->vtime));
> +	write_seqcount_end(&vtime->seqcount);
>  }

So this makes switching away from idle more expensive ? Also,
vtime_account_system() has this fast-path check in there before taking
that lock, should we not do the same? Or should it be removed from
vtime_account_system() ?