Re: [PATCH 2/2] PM: sleep: Do not arm wakeup IRQs for hibernation freeze and quiesce

From: Ulf Hansson

Date: Thu Nov 12 2026 - 08:52:42 EST


On Mon, Nov 2, 2026 at 3:47 PM Abel Vesa <abel.vesa@xxxxxxxxxxxxxxxx> wrote:
>
> Freezing devices to take a hibernation image and quiescing them before
> jumping into one go through the regular noirq suspend path, which arms
> the wakeup IRQs of all devices that may wake up the system. A wakeup
> IRQ firing then aborts the hibernation or the restore:
>
> PM: Triggering wakeup from IRQ 180
> Wakeup pending. Abort CPU freeze
> PM: hibernation: Failed to load image, recovering.
>
> On a Qualcomm Glymur CRD this happens on every hibernation and every
> restore. The WWAN modem keeps asserting WAKE# once its root port goes
> down, after its driver has shut it down for the freeze, and the WAKE#
> GPIO of a PCIe port is a dedicated wakeup IRQ with wakeup enabled by
> default.
>
> Neither transition is one the system can be woken up from, as nothing
> goes to sleep, and Documentation/driver-api/pm/devices.rst already says
> that in the freeze_late and freeze_noirq phases a device "should not be
> allowed to generate wakeup events". Do not arm the wakeup IRQs for
> them, and only disarm them in the noirq resume path if they have been
> armed. Wakeup IRQs are still armed for the final transition to sleep,
> PM_EVENT_HIBERNATE, and for suspend.
>
> Signed-off-by: Abel Vesa <abel.vesa@xxxxxxxxxxxxxxxx>

This seems reasonable to me, but I can't judge from a reasonably brief
review if this is the proper place to fix this. Rafael, what do you
think?

Nevertheless, please add:

Reviewed-by: Ulf Hansson <ulfh@xxxxxxxxxx>

Kind regards
Uffe

> ---
> drivers/base/power/main.c | 21 +++++++++++++++++++--
> 1 file changed, 19 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/base/power/main.c b/drivers/base/power/main.c
> index 16a1f017051e..34ba0e5941a3 100644
> --- a/drivers/base/power/main.c
> +++ b/drivers/base/power/main.c
> @@ -949,6 +949,9 @@ static void dpm_noirq_resume_devices(pm_message_t state)
> trace_suspend_resume(TPS("dpm_resume_noirq"), state.event, false);
> }
>
> +/* Whether dpm_suspend_noirq() has armed the wakeup IRQs of devices */
> +static bool dpm_wake_irqs_armed;
> +
> /**
> * dpm_resume_noirq - Execute "noirq resume" callbacks for all devices.
> * @state: PM transition of the system being carried out.
> @@ -961,7 +964,11 @@ void dpm_resume_noirq(pm_message_t state)
> dpm_noirq_resume_devices(state);
>
> resume_device_irqs();
> - device_wakeup_disarm_wake_irqs();
> +
> + if (dpm_wake_irqs_armed) {
> + device_wakeup_disarm_wake_irqs();
> + dpm_wake_irqs_armed = false;
> + }
> }
>
> static void async_resume_early(void *data, async_cookie_t cookie);
> @@ -1685,7 +1692,17 @@ int dpm_suspend_noirq(pm_message_t state)
> {
> int ret;
>
> - device_wakeup_arm_wake_irqs();
> + /*
> + * Freezing devices to take a hibernation image, and quiescing them
> + * before jumping into one, are not transitions the system can be woken
> + * up from, as nothing goes to sleep. A wakeup IRQ of a device that
> + * signals for attention once its link goes down, like a PCIe endpoint
> + * asserting WAKE#, would only abort the hibernation or the restore.
> + */
> + if (state.event != PM_EVENT_FREEZE && state.event != PM_EVENT_QUIESCE) {
> + device_wakeup_arm_wake_irqs();
> + dpm_wake_irqs_armed = true;
> + }
> suspend_device_irqs();
>
> ret = dpm_noirq_suspend_devices(state);
>
> --
> 2.54.0
>