Re: [PATCH 1/6] delayacct: Use sched_clock()

[Johannes Weiner]

On Thu, May 06, 2021 at 04:17:33PM +0200, Peter Zijlstra wrote:
> On Thu, May 06, 2021 at 09:59:11AM -0400, Johannes Weiner wrote:
> > On Wed, May 05, 2021 at 12:59:41PM +0200, Peter Zijlstra wrote:
> > > @@ -42,10 +42,9 @@ void __delayacct_tsk_init(struct task_st
> > >   * Finish delay accounting for a statistic using its timestamps (@start),
> > >   * accumalator (@total) and @count
> > >   */
> > > -static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total,
> > > -			  u32 *count)
> > > +static void delayacct_end(raw_spinlock_t *lock, u64 *start, u64 *total, u32 *count)
> > >  {
> > > -	s64 ns = ktime_get_ns() - *start;
> > > +	s64 ns = local_clock() - *start;
> > 
> > I don't think this is safe. These time sections that have preemption
> > and migration enabled and so might span multiple CPUs. local_clock()
> > could end up behind *start, AFAICS.
> 
> Only if you have really crummy hardware, and in that case the drift is
> bounded by around 1 tick. Also, this function actually checks: ns > 0.

Oh, I didn't realize it was that close. I just went off the dramatic
warnings on cpu_clock() :-) But yeah, that seems plenty accurate for
this purpose.

Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>