Re: [PATCH 5/6] posix-cpu-timers: Force next expiration recalc after early timer firing

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Jun 04, 2021 at 01:31:58PM +0200, Frederic Weisbecker wrote:
> diff --git a/kernel/time/posix-cpu-timers.c b/kernel/time/posix-cpu-timers.c
> index 0b5715c8db04..d8325a906314 100644
> --- a/kernel/time/posix-cpu-timers.c
> +++ b/kernel/time/posix-cpu-timers.c
> @@ -405,6 +405,21 @@ static int posix_cpu_timer_create(struct k_itimer *new_timer)
>  	return 0;
>  }
>  
> +static void __disarm_timer(struct k_itimer *timer, struct task_struct *p,
> +			   u64 old_expires)
> +{
> +	int clkidx = CPUCLOCK_WHICH(timer->it_clock);
> +	struct posix_cputimer_base *base;
> +
> +	if (CPUCLOCK_PERTHREAD(timer->it_clock))
> +		base = p->posix_cputimers.bases + clkidx;
> +	else
> +		base = p->signal->posix_cputimers.bases + clkidx;
> +
> +	if (old_expires == base->nextevt)
> +		base->nextevt = 0;
> +}
> +
>  /*
>   * Dequeue the timer and reset the base if it was its earliest expiration.
>   * It makes sure the next tick recalculates the base next expiration so we
> @@ -415,24 +430,14 @@ static void disarm_timer(struct k_itimer *timer, struct task_struct *p)
>  {
>  	struct cpu_timer *ctmr = &timer->it.cpu;
>  	u64 old_expires = cpu_timer_getexpires(ctmr);
> -	struct posix_cputimer_base *base;
>  	bool queued;
> -	int clkidx;
>  
>  	queued = cpu_timer_dequeue(ctmr);
>  	cpu_timer_setexpires(ctmr, 0);
>  	if (!queued)
>  		return;
>  
> -	clkidx = CPUCLOCK_WHICH(timer->it_clock);
> -
> -	if (CPUCLOCK_PERTHREAD(timer->it_clock))
> -		base = p->posix_cputimers.bases + clkidx;
> -	else
> -		base = p->signal->posix_cputimers.bases + clkidx;
> -
> -	if (old_expires == base->nextevt)
> -		base->nextevt = 0;
> +	__disarm_timer(timer, p, old_expires);
>  }
>  
>  
> @@ -686,8 +691,7 @@ static int posix_cpu_timer_set(struct k_itimer *timer, int timer_flags,
>  			u64 exp = bump_cpu_timer(timer, val);
>  
>  			if (val < exp) {
> -				old_expires = exp - val;
> -				old->it_value = ns_to_timespec64(old_expires);
> +				old->it_value = ns_to_timespec64(exp - val);
>  			} else {
>  				old->it_value.tv_nsec = 1;
>  				old->it_value.tv_sec = 0;
> @@ -748,9 +752,28 @@ static int posix_cpu_timer_set(struct k_itimer *timer, int timer_flags,
>  		 * accumulate more time on this clock.
>  		 */
>  		cpu_timer_fire(timer);
> +
> +		sighand = lock_task_sighand(p, &flags);
> +		if (sighand == NULL)
> +			goto out;
> +		if (!cpu_timer_queued(&timer->it.cpu)) {
> +			/*
> +			 * Disarm the previous timer to deactivate the tick
> +			 * dependency and process wide cputime counter if
> +			 * necessary.
> +			 */
> +			__disarm_timer(timer, p, old_expires);
> +			/*
> +			 * If the previous timer was deactivated, we might have
> +			 * just started the process wide cputime counter. Make
> +			 * sure we poke the tick to deactivate it then.
> +			 */
> +			if (!old_expires && !CPUCLOCK_PERTHREAD(timer->it_clock))
> +				p->signal->posix_cputimers.bases[clkid].nextevt = 0;
> +		}
> +		unlock_task_sighand(p, &flags);
>  	}

I'm thinking this is a better fix than patch #2. AFAICT you can now go
back to unconditionally doing start, and then if we fire it early, we'll
disarm the thing.

That would avoid the disconnect between the start condition and the fire
condition.

Hmm?