Re: [RFC][PATCH 1/3] time/tick: Pipe tick count down through cputime accounting

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, Jan 27 2025 at 22:32, John Stultz wrote:
> In working up the dynHZ patch, I found that the skipping of
> ticks would result in large latencies for itimers.
>
> As I dug into it, I realized there is still some logic that
> assumes we don't miss ticks, resulting in late expiration of
> cputime timers.
>
> So this patch pipes the actual number of ticks passed down
> through cputime accounting.

This word salad does not qualify as a proper technical changelog. See
Documentation/

>  /*
>   * Must be called with interrupts disabled !
>   */
> -static void tick_do_update_jiffies64(ktime_t now)
> +static unsigned long tick_do_update_jiffies64(ktime_t now)
>  {
>  	unsigned long ticks = 1;
>  	ktime_t delta, nextp;
> @@ -70,7 +70,7 @@ static void tick_do_update_jiffies64(ktime_t now)
>  	 */
>  	if (IS_ENABLED(CONFIG_64BIT)) {
>  		if (ktime_before(now, smp_load_acquire(&tick_next_period)))
> -			return;
> +			return 0;

So if the CPU's tick handler does not update jiffies, then this returns
zero ticks....

> -static void tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
> +static unsigned long tick_sched_do_timer(struct tick_sched *ts, ktime_t now)
>  {
>  	int tick_cpu, cpu = smp_processor_id();
> -
> +	unsigned long ticks = 0;

And you have also zero ticks, when the CPU is not the tick_cpu:

>  	/* Check if jiffies need an update */
>  	if (tick_cpu == cpu)
> -		tick_do_update_jiffies64(now);
> +		ticks = tick_do_update_jiffies64(now);

...

> +	return ticks;
>  }
>  
> -static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
> +static void tick_sched_handle(struct tick_sched *ts, unsigned long ticks, struct pt_regs *regs)
>  {
>  	/*
>  	 * When we are idle and the tick is stopped, we have to touch
> @@ -264,7 +266,7 @@ static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
>  	    tick_sched_flag_test(ts, TS_FLAG_STOPPED)) {
>  		touch_softlockup_watchdog_sched();
>  		if (is_idle_task(current))
> -			ts->idle_jiffies++;
> +			ts->idle_jiffies += ticks;
>  		/*
>  		 * In case the current tick fired too early past its expected
>  		 * expiration, make sure we don't bypass the next clock reprogramming
> @@ -273,7 +275,7 @@ static void tick_sched_handle(struct tick_sched *ts, struct pt_regs *regs)
>  		ts->next_tick = 0;
>  	}
>  
> -	update_process_times(user_mode(regs));
> +	update_process_times(ticks, user_mode(regs));

Which is then fed to update_process_times() and subsequently to
account_process_ticks().

IOW, any CPUs tick handler which does not actually advance jiffies is
going to account ZERO ticks...

Seriously?

Thanks,

        tglx