[PATCH v3] watchdog: mtk_wdt: Stop a running watchdog over system sleep even if not opened

[PATCH v3] watchdog: mtk_wdt: Stop a running watchdog over system sleep even if not opened

[zoan37]

mtk_wdt_suspend() and mtk_wdt_resume() only stop and restart the
watchdog when it is active, that is, when userspace has opened it.

But a watchdog that the bootloader left enabled stays enabled at probe,
marked WDOG_HW_RUNNING, and the watchdog core pings it until userspace
takes over; that has been the case since
commit bbece05c0d3a ("watchdog: mtk_wdt: Remove mtk_wdt_stop() in probe()
to prevent the system freeze and it doesn't reboot by watchdog problem").
If the system suspends before userspace opens the device, or userspace
never does, nothing stops the watchdog and nothing pings it while the
system sleeps, so it resets the system one timeout (31 s by default)
into the sleep.

Also stop it on suspend and restart it on resume when it is running in
hardware, like sp805_wdt does.

The watchdog core keeps pinging such a watchdog until
mtk_wdt_suspend() stops it, so it stays fed while tasks are frozen
before the devices suspend. It may also ping once after resume, before
mtk_wdt_resume() has started the watchdog again; that only reloads the
counter of a stopped watchdog, and the watchdog stays powered over
system sleep (it is what resets the system from s2idle without this
patch). watchdog_stop_ping_on_suspend() is deliberately not used: it
stops the pings from PM_SUSPEND_PREPARE on, which would leave a running
watchdog unfed while tasks are frozen and could reset the system there
with a short timeout.

Fixes: bbece05c0d3a ("watchdog: mtk_wdt: Remove mtk_wdt_stop() in probe() to prevent the system freeze and it doesn't reboot by watchdog problem")
Assisted-by: LLM
Signed-off-by: zoan37 <agentzoan@gmail.com>
---
Changes in v3:
- Drop watchdog_stop_ping_on_suspend() again. It stops the core's pings
  at PM_SUSPEND_PREPARE, before tasks are frozen, while the watchdog
  keeps running until mtk_wdt_suspend(); with a short timeout that could
  reset the system during the freeze (Sashiko review of v2). The ping
  after resume it was meant to avoid (Sashiko review of v1) is harmless,
  see the commit message.
- The code is v1 again; the commit message explains the choice.
- v2: https://lore.kernel.org/all/20261011011536.1139162-1-agentzoan@gmail.com/

Changes in v2:
- Call watchdog_stop_ping_on_suspend() in probe.
- v1: https://lore.kernel.org/all/20261011004322.1118027-1-agentzoan@gmail.com/

Testing: on an MT8189 Chromebook (next-20261008 plus MT8189 support)
whose firmware leaves the watchdog off; a local module parameter
(mtk_wdt.start_timeout=31) starts it in probe the same way a
bootloader-enabled watchdog is picked up (WDOG_HW_RUNNING set, fed by
the watchdog core, nobody opening /dev/watchdog). Without this patch
the board reset about 30 s into s2idle. With this code (= v1): a
10-minute s2idle sleep and 40+ shorter s2idle cycles came back without
a reset. Not tested with a watchdog actually left enabled by a
bootloader. Built with W=1, checkpatch --strict clean apart from the
sign-off.

 drivers/watchdog/mtk_wdt.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/watchdog/mtk_wdt.c b/drivers/watchdog/mtk_wdt.c
index 1630ab65d593..ff72e0fb1154 100644
--- a/drivers/watchdog/mtk_wdt.c
+++ b/drivers/watchdog/mtk_wdt.c
@@ -572,7 +572,8 @@ static int mtk_wdt_suspend(struct device *dev)
 {
 	struct mtk_wdt_dev *mtk_wdt = dev_get_drvdata(dev);
 
-	if (watchdog_active(&mtk_wdt->wdt_dev))
+	if (watchdog_active(&mtk_wdt->wdt_dev) ||
+	    watchdog_hw_running(&mtk_wdt->wdt_dev))
 		mtk_wdt_stop(&mtk_wdt->wdt_dev);
 
 	return 0;
@@ -582,7 +583,8 @@ static int mtk_wdt_resume(struct device *dev)
 {
 	struct mtk_wdt_dev *mtk_wdt = dev_get_drvdata(dev);
 
-	if (watchdog_active(&mtk_wdt->wdt_dev)) {
+	if (watchdog_active(&mtk_wdt->wdt_dev) ||
+	    watchdog_hw_running(&mtk_wdt->wdt_dev)) {
 		mtk_wdt_start(&mtk_wdt->wdt_dev);
 		mtk_wdt_ping(&mtk_wdt->wdt_dev);
 	}

base-commit: aac26bee2287c88af5be5a5ff96d783b19a28790
-- 
2.43.0