[PATCH 2/2] clocksource: timer-econet-en751221: fix IRQ ack storm on interrupt handler

[PATCH 2/2] clocksource: timer-econet-en751221: fix IRQ ack storm on interrupt handler

[kpursoty]

The interrupt handler acked by writing the current count back to compare:

  iowrite32(ioread32(reg_count(cpu)), reg_compare(cpu));

Since the interrupt fires when count >= compare, and the counter advances
past the written value before the MMIO write reaches the peripheral, the
condition re-asserts on the same bus cycle. This causes a tight IRQ storm
on CPU0, stalling boot at T=62s (measured: boot never progresses past
clockevents registration).

The EcoNet GPL kernel (arch/mips/econet/time2.c, cputmr_timer_ack()) acks
by writing the *next* deadline, which puts count < compare and clears the
pending condition. Write U32_MAX (unreachable before set_next_event()
programs the real deadline), consistent with cevt_dev_init() which also
writes U32_MAX to park the compare register at init.

Fixes: 3b4c33ac87d0 ("clocksource/drivers: Add EcoNet Timer HPT driver")
Signed-off-by: Kervin Pursoty <kpursoty@proton.me>
---
 drivers/clocksource/timer-econet-en751221.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/clocksource/timer-econet-en751221.c b/drivers/clocksource/timer-econet-en751221.c
--- a/drivers/clocksource/timer-econet-en751221.c
+++ b/drivers/clocksource/timer-econet-en751221.c
@@ -68,7 +68,9 @@ static irqreturn_t cevt_interrupt(int irq, void *dev_id)
 	if (!cevt_is_pending(cpu))
 		return IRQ_NONE;

-	iowrite32(ioread32(reg_count(cpu)), reg_compare(cpu));
+	/* Ack: write a future value so count < compare, clearing the pending
+	 * condition. set_next_event() will program the real deadline. */
+	iowrite32(U32_MAX, reg_compare(cpu));
 	dev->event_handler(dev);
 	return IRQ_HANDLED;
 }

Re: [PATCH 2/2] clocksource: timer-econet-en751221: fix IRQ ack storm on interrupt handler

[Sergey Shtylyov]

On 4/13/26 1:12 PM, kpursoty@proton.me wrote:

> The interrupt handler acked by writing the current count back to compare:
> 
>   iowrite32(ioread32(reg_count(cpu)), reg_compare(cpu));
> 
> Since the interrupt fires when count >= compare, and the counter advances
> past the written value before the MMIO write reaches the peripheral, the
> condition re-asserts on the same bus cycle. This causes a tight IRQ storm
> on CPU0, stalling boot at T=62s (measured: boot never progresses past
> clockevents registration).
> 
> The EcoNet GPL kernel (arch/mips/econet/time2.c, cputmr_timer_ack()) acks
> by writing the *next* deadline, which puts count < compare and clears the
> pending condition. Write U32_MAX (unreachable before set_next_event()
> programs the real deadline), consistent with cevt_dev_init() which also
> writes U32_MAX to park the compare register at init.
> 
> Fixes: 3b4c33ac87d0 ("clocksource/drivers: Add EcoNet Timer HPT driver")
> Signed-off-by: Kervin Pursoty <kpursoty@proton.me>
> ---
>  drivers/clocksource/timer-econet-en751221.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/clocksource/timer-econet-en751221.c b/drivers/clocksource/timer-econet-en751221.c
> --- a/drivers/clocksource/timer-econet-en751221.c
> +++ b/drivers/clocksource/timer-econet-en751221.c
> @@ -68,7 +68,9 @@ static irqreturn_t cevt_interrupt(int irq, void *dev_id)
>  	if (!cevt_is_pending(cpu))
>  		return IRQ_NONE;
> 
> -	iowrite32(ioread32(reg_count(cpu)), reg_compare(cpu));
> +	/* Ack: write a future value so count < compare, clearing the pending
> +	 * condition. set_next_event() will program the real deadline. */

   The multiline comments in the kernel (with a little exception of the networking)
should be formatter like this:

	/*
	 * blah blah blah
	 */

[...]

MBR, Sergey