Re: [PATCH] tick/nohz_full: don't abuse smp_call_function_single() in tick_setup_device()

[Frederic Weisbecker <frederic@kernel.org>]

Le Tue, May 28, 2024 at 02:20:19PM +0200, Oleg Nesterov a écrit :
> After the recent commit 5097cbcb38e6 ("sched/isolation: Prevent boot
> crash when the boot CPU is nohz_full") the kernel no longer crashes, but
> there is another problem.
> 
> In this case tick_setup_device() calls tick_take_do_timer_from_boot() to
> update tick_do_timer_cpu and this triggers the WARN_ON_ONCE(irqs_disabled)
> in smp_call_function_single().
> 
> Kill tick_take_do_timer_from_boot() and just use WRITE_ONCE(), the new
> comment tries to explain why this is safe (thanks Thomas!).
> 
> Fixes: 08ae95f4fd3b ("nohz_full: Allow the boot CPU to be nohz_full")
> Link: https://lore.kernel.org/all/20240522151742.GA10400@redhat.com
> Signed-off-by: Oleg Nesterov <oleg@redhat.com>
> ---
>  kernel/time/tick-common.c | 39 +++++++++++++--------------------------
>  1 file changed, 13 insertions(+), 26 deletions(-)
> 
> diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
> index d88b13076b79..27d0018c8b05 100644
> --- a/kernel/time/tick-common.c
> +++ b/kernel/time/tick-common.c
> @@ -178,26 +178,6 @@ void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
>  	}
>  }
>  
> -#ifdef CONFIG_NO_HZ_FULL
> -static void giveup_do_timer(void *info)
> -{
> -	int cpu = *(unsigned int *)info;
> -
> -	WARN_ON(tick_do_timer_cpu != smp_processor_id());
> -
> -	tick_do_timer_cpu = cpu;
> -}
> -
> -static void tick_take_do_timer_from_boot(void)
> -{
> -	int cpu = smp_processor_id();
> -	int from = tick_do_timer_boot_cpu;
> -
> -	if (from >= 0 && from != cpu)
> -		smp_call_function_single(from, giveup_do_timer, &cpu, 1);
> -}
> -#endif
> -
>  /*
>   * Setup the tick device
>   */
> @@ -221,19 +201,26 @@ static void tick_setup_device(struct tick_device *td,
>  			tick_next_period = ktime_get();
>  #ifdef CONFIG_NO_HZ_FULL
>  			/*
> -			 * The boot CPU may be nohz_full, in which case set
> -			 * tick_do_timer_boot_cpu so the first housekeeping
> -			 * secondary that comes up will take do_timer from
> -			 * us.
> +			 * The boot CPU may be nohz_full, in which case the
> +			 * first housekeeping secondary will take do_timer()
> +			 * from us.
>  			 */
>  			if (tick_nohz_full_cpu(cpu))
>  				tick_do_timer_boot_cpu = cpu;
>  
>  		} else if (tick_do_timer_boot_cpu != -1 &&
>  						!tick_nohz_full_cpu(cpu)) {
> -			tick_take_do_timer_from_boot();
>  			tick_do_timer_boot_cpu = -1;
> -			WARN_ON(READ_ONCE(tick_do_timer_cpu) != cpu);
> +			/*
> +			 * The boot CPU will stay in periodic (NOHZ disabled)
> +			 * mode until clocksource_done_booting() called after
> +			 * smp_init() selects a high resolution clocksource and
> +			 * timekeeping_notify() kicks the NOHZ stuff alive.
> +			 *
> +			 * So this WRITE_ONCE can only race with the READ_ONCE
> +			 * check in tick_periodic() but this race is harmless.
> +			 */
> +			WRITE_ONCE(tick_do_timer_cpu, cpu);

Looks good, but can we have a WARN_ONCE(tick_do_timer_cpu != tick_do_timer_boot_cpu)
right before that, just to make sure our assumptions above are right forever and
the boot CPU hasn't stopped the tick up to that point?

And after all, pushing a bit further your subsequent patch, can we get rid of
tick_do_timer_boot_cpu and ifdefery altogether? Such as:

diff --git a/kernel/time/tick-common.c b/kernel/time/tick-common.c
index fb0fdec8719a..63a7bd405de7 100644
--- a/kernel/time/tick-common.c
+++ b/kernel/time/tick-common.c
@@ -48,14 +48,6 @@ ktime_t tick_next_period;
  *    procedure also covers cpu hotplug.
  */
 int tick_do_timer_cpu __read_mostly = TICK_DO_TIMER_BOOT;
-#ifdef CONFIG_NO_HZ_FULL
-/*
- * tick_do_timer_boot_cpu indicates the boot CPU temporarily owns
- * tick_do_timer_cpu and it should be taken over by an eligible secondary
- * when one comes online.
- */
-static int tick_do_timer_boot_cpu __read_mostly = -1;
-#endif
 
 /*
  * Debugging: see timer_list.c
@@ -177,26 +169,6 @@ void tick_setup_periodic(struct clock_event_device *dev, int broadcast)
 	}
 }
 
-#ifdef CONFIG_NO_HZ_FULL
-static void giveup_do_timer(void *info)
-{
-	int cpu = *(unsigned int *)info;
-
-	WARN_ON(tick_do_timer_cpu != smp_processor_id());
-
-	tick_do_timer_cpu = cpu;
-}
-
-static void tick_take_do_timer_from_boot(void)
-{
-	int cpu = smp_processor_id();
-	int from = tick_do_timer_boot_cpu;
-
-	if (from >= 0 && from != cpu)
-		smp_call_function_single(from, giveup_do_timer, &cpu, 1);
-}
-#endif
-
 /*
  * Setup the tick device
  */
@@ -211,29 +183,28 @@ static void tick_setup_device(struct tick_device *td,
 	 * First device setup ?
 	 */
 	if (!td->evtdev) {
+		int timekeeper = READ_ONCE(tick_do_timer_cpu);
 		/*
 		 * If no cpu took the do_timer update, assign it to
 		 * this cpu:
 		 */
-		if (tick_do_timer_cpu == TICK_DO_TIMER_BOOT) {
+		if (timekeeper == TICK_DO_TIMER_BOOT) {
 			tick_do_timer_cpu = cpu;
 			tick_next_period = ktime_get();
-#ifdef CONFIG_NO_HZ_FULL
+		} else if (timekeeper == TICK_DO_TIMER_NONE) {
+			if (WARN_ON_ONCE(tick_nohz_full_enabled()))
+				WRITE_ONCE(tick_do_timer_cpu, cpu);
+		} else if (tick_nohz_full_cpu(timekeeper) && !tick_nohz_full_cpu(cpu)) {
 			/*
-			 * The boot CPU may be nohz_full, in which case set
-			 * tick_do_timer_boot_cpu so the first housekeeping
-			 * secondary that comes up will take do_timer from
-			 * us.
+			 * The boot CPU will stay in periodic (NOHZ disabled)
+			 * mode until clocksource_done_booting() called after
+			 * smp_init() selects a high resolution clocksource and
+			 * timekeeping_notify() kicks the NOHZ stuff alive.
+			 *
+			 * So this WRITE_ONCE can only race with the READ_ONCE
+			 * check in tick_periodic() but this race is harmless.
 			 */
-			if (tick_nohz_full_cpu(cpu))
-				tick_do_timer_boot_cpu = cpu;
-
-		} else if (tick_do_timer_boot_cpu != -1 &&
-						!tick_nohz_full_cpu(cpu)) {
-			tick_take_do_timer_from_boot();
-			tick_do_timer_boot_cpu = -1;
-			WARN_ON(tick_do_timer_cpu != cpu);
-#endif
+			WRITE_ONCE(tick_do_timer_cpu, cpu);
 		}
 
 		/*