[PATCH v3] watchdog: sunxi_wdt: stop the watchdog across system suspend
From: Yang Siyuan
Date: Tue Nov 17 2026 - 18:19:16 EST
The H616 watchdog has a maximum timeout of 16 seconds. systemd keeps
/dev/watchdog0 open with RuntimeWatchdogSec=16s, but during system
suspend all userspace is frozen and the watchdog core does not stop the
hardware. A suspend longer than the timeout therefore resets the board
before it can resume.
Add system sleep PM ops to stop the watchdog before suspend and restart
it after resume if it was running. The driver did not store its private
data in the platform device, so add platform_set_drvdata() for the PM
callbacks.
Use DEFINE_SIMPLE_DEV_PM_OPS() and pm_sleep_ptr(), and use
watchdog_hw_running() instead of a private flag to decide whether the
hardware needs to be stopped and restarted.
Tested on an Anbernic RG35XX-SP: a 30-second s2idle suspend resumes
without the 16-second watchdog reset, and /dev/watchdog0 keeps working
afterwards.
Signed-off-by: Yang Siyuan <hsyysy@xxxxxxx>
---
Changes in v3:
- Mark the PM callbacks __maybe_unused. With CONFIG_PM_SLEEP=n the
pm_sleep_ptr() references expand to NULL, so both callbacks were unused
and a W=1 build warned about them. Reported by the kernel test robot:
https://lore.kernel.org/oe-kbuild-all/202611172344.RbQl66xu-lkp@xxxxxxxxx/
- Reword the v2 changelog entry: pm_sleep_ptr() alone does not keep the
callbacks referenced when CONFIG_PM_SLEEP is disabled.
Changes in v2:
- Use DEFINE_SIMPLE_DEV_PM_OPS()/pm_sleep_ptr() to avoid unused-function
warnings with CONFIG_PM_SLEEP=n.
- Use watchdog_hw_running() instead of tracking running_at_suspend.
- Drop the inaccurate ping-worker reference from the comment.
drivers/watchdog/sunxi_wdt.c | 32 ++++++++++++++++++++++++++++++++
1 file changed, 32 insertions(+)
diff --git a/drivers/watchdog/sunxi_wdt.c b/drivers/watchdog/sunxi_wdt.c
index a6500eb..3fdc7bf 100644
--- a/drivers/watchdog/sunxi_wdt.c
+++ b/drivers/watchdog/sunxi_wdt.c
@@ -225,6 +225,35 @@ static int sunxi_wdt_start(struct watchdog_device *wdt_dev)
return 0;
}
+/*
+ * The hardware watchdog has a maximum timeout of 16 seconds. A system
+ * suspend easily lasts longer than that, and the watchdog core does not
+ * stop the hardware during system sleep. Stop the watchdog explicitly
+ * while suspended and re-arm it on resume if it was running.
+ */
+static int __maybe_unused sunxi_wdt_suspend(struct device *dev)
+{
+ struct sunxi_wdt_dev *sunxi_wdt = dev_get_drvdata(dev);
+
+ if (watchdog_hw_running(&sunxi_wdt->wdt_dev))
+ sunxi_wdt_stop(&sunxi_wdt->wdt_dev);
+
+ return 0;
+}
+
+static int __maybe_unused sunxi_wdt_resume(struct device *dev)
+{
+ struct sunxi_wdt_dev *sunxi_wdt = dev_get_drvdata(dev);
+
+ if (watchdog_hw_running(&sunxi_wdt->wdt_dev))
+ sunxi_wdt_start(&sunxi_wdt->wdt_dev);
+
+ return 0;
+}
+
+static DEFINE_SIMPLE_DEV_PM_OPS(sunxi_wdt_pm_ops,
+ sunxi_wdt_suspend, sunxi_wdt_resume);
+
static const struct watchdog_info sunxi_wdt_info = {
.identity = DRV_NAME,
.options = WDIOF_SETTIMEOUT |
@@ -307,6 +336,8 @@ static int sunxi_wdt_probe(struct platform_device *pdev)
if (IS_ERR(sunxi_wdt->wdt_base))
return PTR_ERR(sunxi_wdt->wdt_base);
+ platform_set_drvdata(pdev, sunxi_wdt);
+
sunxi_wdt->wdt_dev.info = &sunxi_wdt_info;
sunxi_wdt->wdt_dev.ops = &sunxi_wdt_ops;
sunxi_wdt->wdt_dev.timeout = WDT_MAX_TIMEOUT;
@@ -348,6 +379,7 @@ static struct platform_driver sunxi_wdt_driver = {
.driver = {
.name = DRV_NAME,
.of_match_table = sunxi_wdt_dt_ids,
+ .pm = pm_sleep_ptr(&sunxi_wdt_pm_ops),
},
};
base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3
--
2.56.1