Re: [PATCH v8] ACPI: button: Report wakeup key only for power button wakeups

From: Rafael J. Wysocki (Intel)

Date: Fri Nov 20 2026 - 09:01:32 EST


On Tue, Nov 17, 2026 at 8:17 AM Baorui.Liu <baorliu@xxxxxxx> wrote:
>
> The ACPI button driver reports KEY_WAKEUP from the Power Button input
> device to let userspace know that the system was resumed by a power
> button wakeup.
>
> However, reporting KEY_WAKEUP from generic system resume paths can make
> userspace observe a Power Button wakeup even when the system was resumed
> by a different wake source.
>
> This is reproducible on an AMD Android 15 Xen guest. With a kernel
> without this fix, a non-power-button S3 resume:
>
> echo mem > /sys/power/state
> xl trigger android s3resume
>
> makes the Power Button input device report KEY_WAKEUP. The same test on
> a kernel with this fix no longer reports KEY_WAKEUP from the Power Button
> input device.
>
> Track whether a power button event/notify is observed while the ACPI
> button device is suspended, and report KEY_WAKEUP on resume only in that
> case.
>
> When the fixed power-button status bit is set during early S3 resume,
> clear it to avoid a deferred KEY_POWER delivery to userspace, and set a
> sleep.c-owned flag so button resume can still report KEY_WAKEUP. Do not
> call into the button driver from that IRQ-disabled path. Also clear
> wakeup_pending before setting suspended, with a compiler barrier, and
> drain ACPI notify work before clearing suspended on resume. Report
> KEY_WAKEUP before clearing suspended so a late notify cannot drop both
> KEY_WAKEUP and KEY_POWER.
>
> Reported-by: kernel test robot <lkp@xxxxxxxxx>
> Closes: https://lore.kernel.org/r/202609290454.LJrhMCb3-lkp@xxxxxxxxx/
> Reported-by: kernel test robot <lkp@xxxxxxxxx>
> Closes: https://lore.kernel.org/oe-kbuild-all/202611030112.cVsBZFFE-lkp@xxxxxxxxx/
> Reported-by: kernel test robot <lkp@xxxxxxxxx>
> Closes: https://lore.kernel.org/oe-kbuild-all/202611030545.3st9p1ha-lkp@xxxxxxxxx/
> Reported-by: kernel test robot <lkp@xxxxxxxxx>
> Closes: https://lore.kernel.org/oe-kbuild-all/202611100542.1xljQv20-lkp@xxxxxxxxx/
> Signed-off-by: Baorui.Liu <baorliu@xxxxxxx>
>
> Changes since v7:
> - Guard acpi_sleep_pwr_btn_wakeup_pending() with CONFIG_ACPI_SLEEP,
> matching its definition in sleep.c, so button resume does not link
> against a missing symbol when CONFIG_PM_SLEEP=y and CONFIG_ACPI_SLEEP
> is unset (0day/LKP ERROR on x86_64 randconfig).
>
> Changes since v6:
> - Include <linux/types.h> in include/acpi/button.h so the bool
> declaration/stub of acpi_sleep_pwr_btn_wakeup_pending() builds when
> the header is included without prior bool typedefs (0day/LKP ERROR
> via nouveau_connector.c on parisc64-defconfig and arm64-randconfig).

Sashiko has feedback on this:

https://sashiko.dev/#/patchset/20261117071709.428-1-baorliu%40amd.com

Please let me know what you think about it.

Also, if you decide to send a new version, please do so in a new thread.

Thanks!

> ---
> drivers/acpi/button.c | 63 ++++++++++++++++++++++++++++++++++++-------
> drivers/acpi/sleep.c | 47 +++++++++++++++++++++-----------
> include/acpi/button.h | 19 +++++++++++++
> 3 files changed, 104 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/acpi/button.c b/drivers/acpi/button.c
> index c41fba556030..9a80a27e4297 100644
> --- a/drivers/acpi/button.c
> +++ b/drivers/acpi/button.c
> @@ -203,6 +203,7 @@ struct acpi_button {
> bool last_state;
> ktime_t last_time;
> bool suspended;
> + bool wakeup_pending;
> bool lid_state_initialized;
> bool gpe_enabled;
> };
> @@ -488,7 +489,11 @@ static void acpi_button_notify(acpi_handle handle, u32 event, void *data)
>
> acpi_pm_wakeup_event(button->dev);
>
> - if (button->suspended || event == ACPI_BUTTON_NOTIFY_WAKE)
> + if (button->type == ACPI_BUTTON_TYPE_POWER &&
> + READ_ONCE(button->suspended))
> + WRITE_ONCE(button->wakeup_pending, true);
> +
> + if (READ_ONCE(button->suspended) || event == ACPI_BUTTON_NOTIFY_WAKE)
> return;
>
> input = button->input;
> @@ -510,6 +515,12 @@ static void acpi_button_notify_run(void *data)
>
> static u32 acpi_button_event(void *data)
> {
> + struct acpi_button *button = data;
> +
> + if (button->type == ACPI_BUTTON_TYPE_POWER &&
> + READ_ONCE(button->suspended))
> + WRITE_ONCE(button->wakeup_pending, true);
> +
> acpi_os_execute(OSL_NOTIFY_HANDLER, acpi_button_notify_run, data);
> return ACPI_INTERRUPT_HANDLED;
> }
> @@ -519,7 +530,15 @@ static int acpi_button_suspend(struct device *dev)
> {
> struct acpi_button *button = dev_get_drvdata(dev);
>
> - button->suspended = true;
> + /*
> + * Clear wakeup_pending before marking suspended. Otherwise a power
> + * button SCI between the two stores could set wakeup_pending and then
> + * be overwritten here, losing a real wakeup. barrier() prevents the
> + * compiler from reordering the stores relative to that IRQ handler.
> + */
> + WRITE_ONCE(button->wakeup_pending, false);
> + barrier();
> + WRITE_ONCE(button->suspended, true);
> return 0;
> }
>
> @@ -527,21 +546,45 @@ static int acpi_button_resume(struct device *dev)
> {
> struct acpi_button *button = dev_get_drvdata(dev);
> struct input_dev *input;
> + bool report_wakeup = false;
> +
> + /*
> + * Drain notify work queued while suspended so a late power-button
> + * notify still observes suspended == true, sets wakeup_pending, and
> + * does not synthesize KEY_POWER for userspace.
> + */
> + acpi_os_wait_events_complete();
> +
> + /*
> + * Sample and report KEY_WAKEUP before clearing suspended. Otherwise a
> + * notify that arrives after sampling but before suspended is cleared
> + * can set wakeup_pending and then be dropped without KEY_WAKEUP or
> + * KEY_POWER. Clear suspended before wakeup_pending so a new SCI sees
> + * suspended == false and takes the normal KEY_POWER path; barrier()
> + * keeps that store order.
> + */
> + if (button->type == ACPI_BUTTON_TYPE_POWER) {
> + report_wakeup = READ_ONCE(button->wakeup_pending) ||
> + acpi_sleep_pwr_btn_wakeup_pending();
> + if (report_wakeup) {
> + input = button->input;
> + input_report_key(input, KEY_WAKEUP, 1);
> + input_sync(input);
> + input_report_key(input, KEY_WAKEUP, 0);
> + input_sync(input);
> + }
> + }
> +
> + WRITE_ONCE(button->suspended, false);
> + barrier();
> + WRITE_ONCE(button->wakeup_pending, false);
>
> - button->suspended = false;
> if (button->type == ACPI_BUTTON_TYPE_LID) {
> button->last_state = !!acpi_lid_evaluate_state(ACPI_HANDLE(dev));
> button->last_time = ktime_get();
> acpi_lid_initialize_state(button);
> }
>
> - if (button->type == ACPI_BUTTON_TYPE_POWER) {
> - input = button->input;
> - input_report_key(input, KEY_WAKEUP, 1);
> - input_sync(input);
> - input_report_key(input, KEY_WAKEUP, 0);
> - input_sync(input);
> - }
> return 0;
> }
> #endif
> diff --git a/drivers/acpi/sleep.c b/drivers/acpi/sleep.c
> index 7c1fb5b844be..a17318a7a3d6 100644
> --- a/drivers/acpi/sleep.c
> +++ b/drivers/acpi/sleep.c
> @@ -22,6 +22,7 @@
> #include <linux/syscore_ops.h>
> #include <asm/io.h>
> #include <trace/events/power.h>
> +#include <acpi/button.h>
>
> #include "internal.h"
> #include "sleep.h"
> @@ -105,6 +106,23 @@ u32 acpi_target_system_state(void)
> EXPORT_SYMBOL_GPL(acpi_target_system_state);
>
> static bool pwr_btn_event_pending;
> +static bool pwr_btn_wakeup_pending;
> +
> +/**
> + * acpi_sleep_pwr_btn_wakeup_pending - Consume fixed power-button wake hint.
> + *
> + * Returns true once if acpi_suspend_enter() observed and cleared the fixed
> + * power-button status bit during early S3 resume. Used by the button driver
> + * to report KEY_WAKEUP without sleeping in that IRQ-disabled path.
> + */
> +bool acpi_sleep_pwr_btn_wakeup_pending(void)
> +{
> + bool pending = READ_ONCE(pwr_btn_wakeup_pending);
> +
> + WRITE_ONCE(pwr_btn_wakeup_pending, false);
> + return pending;
> +}
> +EXPORT_SYMBOL_GPL(acpi_sleep_pwr_btn_wakeup_pending);
>
> /*
> * The ACPI specification wants us to save NVS memory regions during hibernation
> @@ -504,12 +522,12 @@ static void acpi_pm_finish(void)
>
> acpi_resume_power_resources();
>
> - /* If we were woken with the fixed power button, provide a small
> - * hint to userspace in the form of a wakeup event on the fixed power
> - * button device (if it can be found).
> - *
> - * We delay the event generation til now, as the PM layer requires
> - * timekeeping to be running before we generate events. */
> + /*
> + * If woken by the fixed power button, provide a wakeup event on that
> + * device. KEY_WAKEUP input reporting is handled by the button driver
> + * via acpi_sleep_pwr_btn_wakeup_pending() / wakeup_pending; do not
> + * synthesize KEY_POWER here.
> + */
> if (!pwr_btn_event_pending)
> return;
>
> @@ -625,14 +643,13 @@ static int acpi_suspend_enter(suspend_state_t pm_state)
> /* Reprogram control registers */
> acpi_leave_sleep_state_prep(acpi_state);
>
> - /* ACPI 3.0 specs (P62) says that it's the responsibility
> - * of the OSPM to clear the status bit [ implying that the
> - * POWER_BUTTON event should not reach userspace ]
> - *
> - * However, we do generate a small hint for userspace in the form of
> - * a wakeup event. We flag this condition for now and generate the
> - * event later, as we're currently too early in resume to be able to
> - * generate wakeup events.
> + /*
> + * ACPI 3.0 (P62): OSPM should clear the fixed power-button status bit
> + * so the event does not reach userspace as KEY_POWER. Only set plain
> + * flags here: this path runs with IRQs disabled, so do not call into
> + * the button driver (it may take sleeping locks). Button resume later
> + * consumes pwr_btn_wakeup_pending for KEY_WAKEUP; acpi_pm_finish()
> + * uses pwr_btn_event_pending for the PM wakeup event.
> */
> if (ACPI_SUCCESS(status) && (acpi_state == ACPI_STATE_S3)) {
> acpi_event_status pwr_btn_status = ACPI_EVENT_FLAG_DISABLED;
> @@ -641,8 +658,8 @@ static int acpi_suspend_enter(suspend_state_t pm_state)
>
> if (pwr_btn_status & ACPI_EVENT_FLAG_STATUS_SET) {
> acpi_clear_event(ACPI_EVENT_POWER_BUTTON);
> - /* Flag for later */
> pwr_btn_event_pending = true;
> + WRITE_ONCE(pwr_btn_wakeup_pending, true);
> }
> }
>
> diff --git a/include/acpi/button.h b/include/acpi/button.h
> index af2fce5d2ee3..83f5dd373f39 100644
> --- a/include/acpi/button.h
> +++ b/include/acpi/button.h
> @@ -2,6 +2,8 @@
> #ifndef ACPI_BUTTON_H
> #define ACPI_BUTTON_H
>
> +#include <linux/types.h> /* for bool */
> +
> #define ACPI_BUTTON_HID_POWER "PNP0C0C"
> #define ACPI_BUTTON_HID_LID "PNP0C0D"
> #define ACPI_BUTTON_HID_SLEEP "PNP0C0E"
> @@ -15,4 +17,21 @@ static inline int acpi_lid_open(void)
> }
> #endif /* IS_ENABLED(CONFIG_ACPI_BUTTON) */
>
> +/*
> + * Set in acpi_suspend_enter() when the fixed power-button status bit is
> + * observed (IRQ-disabled context). Consumed by the button driver on resume
> + * so KEY_WAKEUP can be reported without calling into the button driver from
> + * that early path. Defined in sleep.c under CONFIG_ACPI_SLEEP. CONFIG_PM_SLEEP
> + * can be enabled without CONFIG_ACPI_SLEEP, so the stub keeps button.o
> + * linkable in that case. Safe when CONFIG_ACPI_BUTTON=m.
> + */
> +#if IS_ENABLED(CONFIG_ACPI_SLEEP)
> +bool acpi_sleep_pwr_btn_wakeup_pending(void);
> +#else
> +static inline bool acpi_sleep_pwr_btn_wakeup_pending(void)
> +{
> + return false;
> +}
> +#endif
> +
> #endif /* ACPI_BUTTON_H */
>
> base-commit: 0069320cef67ba6ca745a05868c40e4c46b61284
> --
> 2.34.1
>