Re: tick/sched: Make jiffies update quick check more robust

[Peter Zijlstra]

On Mon, Dec 07, 2020 at 03:41:47PM +0100, Thomas Gleixner wrote:
> On Mon, Dec 07 2020 at 10:59, Peter Zijlstra wrote:
> >> +	if (IS_ENABLED(CONFIG_64BIT)) {
> >> +		if (ktime_before(now, smp_load_acquire(&tick_next_period)))
> >> +			return;
> >
> > Explicit ACQUIRE
> >
> >> +	} else {
> >> +		unsigned int seq;
> >> +
> >> +		/*
> >> +		 * Avoid contention on jiffies_lock and protect the quick
> >> +		 * check with the sequence count.
> >> +		 */
> >> +		do {
> >> +			seq = read_seqcount_begin(&jiffies_seq);
> >> +			nextp = tick_next_period;
> >> +		} while (read_seqcount_retry(&jiffies_seq, seq));
> >> +
> >> +		if (ktime_before(now, nextp))
> >> +			return;
> >
> > Actually has an implicit ACQUIRE:
> >
> > 	read_seqcount_retry() implies smp_rmb(), which ensures
> > 	LOAD->LOAD order, IOW any later load must happen after our
> > 	@tick_next_period load.
> >
> > 	Then it has a control dependency on ktime_before(,nextp), which
> > 	ensures LOAD->STORE order.
> >
> > 	Combined we have a LOAD->{LOAD,STORE} order on the
> > 	@tick_next_period load, IOW ACQUIRE.

It's actually the whole of:

+               } while (read_seqcount_retry(&jiffies_seq, seq));

That implies the ACQUIRE, don't need the rest.

> >> +	}
> >>  
> >> +	/* Quick check failed, i.e. update is required. */
> >>  	raw_spin_lock(&jiffies_lock);
> >
> > Another ACQUIRE, which means the above ACQUIRE only ensures we load the
> > lock value after?
> >
> > Or are we trying to guarantee the caller is sure to observe the new
> > jiffies value if we return?
> 
> The guarantee we need on 64bit for the check w/o seqcount is:
> 
> CPU0                                         CPU1
> 
>  if (ktime_before(now, tick_next_period))
>  	return;
> 
>  raw_spin_lock(&jiffies_lock);
>  ....
>  jiffies_64 += ticks;                           
>  
>  tick_next_period = next;                   if (ktime_before(now, tick_next_period))
>   	                                           return;
> 
> When CPU1 returns because it observes the new value in tick_next_period
> then it has to be guaranteed that jiffies_64 is observable as well.

Right, it does that. Good.