From: George Cherian george.cherian@marvell.com
[ Upstream commit 000987a38b53c172f435142a4026dd71378ca464 ]
Make sure to honour the max_hw_heartbeat_ms while programming the timeout value to WOR. Clamp the timeout passed to sbsa_gwdt_set_timeout() to make sure the programmed value is within the permissible range.
Fixes: abd3ac7902fb ("watchdog: sbsa: Support architecture version 1")
Signed-off-by: George Cherian george.cherian@marvell.com Reviewed-by: Guenter Roeck linux@roeck-us.net Link: https://lore.kernel.org/r/20230209021117.1512097-1-george.cherian@marvell.co... Signed-off-by: Guenter Roeck linux@roeck-us.net Signed-off-by: Wim Van Sebroeck wim@linux-watchdog.org Signed-off-by: Tyler Hicks (Microsoft) code@tyhicks.com ---
The Fixes line in the original commit is incorrect. This commit fixes a bug that goes all the way back to v4.6 commit 57d2caaabfc7 ("Watchdog: introduce ARM SBSA watchdog driver") when only 32-bit Watchdog Offset Registers (WOR) were supported.
Without this fix, there's a truncation on the first argument, of u32 type, passed to writel() in the following situation situation:
Generic Watchdog architecture version is 1 (WOR is 32-bit) action is 1 timeout is 240s CNTFRQ_EL0 is 25000050 Hz wdd.max_hw_heartbeat_ms is 171s
25000050 * 240 = 6000012000 <--- requires 33 bits to store 6000012000 & 0xFFFFFFFF = 1705044704 <--- truncated value written to WOR 1705044704 / 25000050 = 68.2s <--- timeout incorrectly set to 68.2s
The timeout from userspace is greater than wdd.max_hw_heartbeat_ms so the watchdog core pings at 69s (240 - 171) which results in intermittent and unexpected panics (action=1).
With this patch applied, the timeout passed to writel() never exceeds 32-bits and the watchdog core + systemd keeps the watchdog happy.
I've validated this fix on real hardware running a linux-5.10.y stable kernel. Please apply this patch to 5.10 through 4.14. Thanks!
Tyler
drivers/watchdog/sbsa_gwdt.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/watchdog/sbsa_gwdt.c b/drivers/watchdog/sbsa_gwdt.c index f0f1e3b2e463..4cbe6ba52754 100644 --- a/drivers/watchdog/sbsa_gwdt.c +++ b/drivers/watchdog/sbsa_gwdt.c @@ -121,6 +121,7 @@ static int sbsa_gwdt_set_timeout(struct watchdog_device *wdd, struct sbsa_gwdt *gwdt = watchdog_get_drvdata(wdd);
wdd->timeout = timeout; + timeout = clamp_t(unsigned int, timeout, 1, wdd->max_hw_heartbeat_ms / 1000);
if (action) writel(gwdt->clk * timeout,
On 4/13/23 13:48, Tyler Hicks (Microsoft) wrote:
From: George Cherian george.cherian@marvell.com
[ Upstream commit 000987a38b53c172f435142a4026dd71378ca464 ]
Make sure to honour the max_hw_heartbeat_ms while programming the timeout value to WOR. Clamp the timeout passed to sbsa_gwdt_set_timeout() to make sure the programmed value is within the permissible range.
Fixes: abd3ac7902fb ("watchdog: sbsa: Support architecture version 1")
Signed-off-by: George Cherian george.cherian@marvell.com Reviewed-by: Guenter Roeck linux@roeck-us.net Link: https://lore.kernel.org/r/20230209021117.1512097-1-george.cherian@marvell.co... Signed-off-by: Guenter Roeck linux@roeck-us.net Signed-off-by: Wim Van Sebroeck wim@linux-watchdog.org Signed-off-by: Tyler Hicks (Microsoft) code@tyhicks.com
Reviewed-by: Guenter Roeck linux@roeck-us.net
The Fixes line in the original commit is incorrect. This commit fixes a bug that goes all the way back to v4.6 commit 57d2caaabfc7 ("Watchdog: introduce ARM SBSA watchdog driver") when only 32-bit Watchdog Offset Registers (WOR) were supported.
Without this fix, there's a truncation on the first argument, of u32 type, passed to writel() in the following situation situation:
Generic Watchdog architecture version is 1 (WOR is 32-bit) action is 1 timeout is 240s CNTFRQ_EL0 is 25000050 Hz wdd.max_hw_heartbeat_ms is 171s
25000050 * 240 = 6000012000 <--- requires 33 bits to store 6000012000 & 0xFFFFFFFF = 1705044704 <--- truncated value written to WOR 1705044704 / 25000050 = 68.2s <--- timeout incorrectly set to 68.2s
The timeout from userspace is greater than wdd.max_hw_heartbeat_ms so the watchdog core pings at 69s (240 - 171) which results in intermittent and unexpected panics (action=1).
With this patch applied, the timeout passed to writel() never exceeds 32-bits and the watchdog core + systemd keeps the watchdog happy.
I've validated this fix on real hardware running a linux-5.10.y stable kernel. Please apply this patch to 5.10 through 4.14. Thanks!
Tyler
drivers/watchdog/sbsa_gwdt.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/watchdog/sbsa_gwdt.c b/drivers/watchdog/sbsa_gwdt.c index f0f1e3b2e463..4cbe6ba52754 100644 --- a/drivers/watchdog/sbsa_gwdt.c +++ b/drivers/watchdog/sbsa_gwdt.c @@ -121,6 +121,7 @@ static int sbsa_gwdt_set_timeout(struct watchdog_device *wdd, struct sbsa_gwdt *gwdt = watchdog_get_drvdata(wdd); wdd->timeout = timeout;
- timeout = clamp_t(unsigned int, timeout, 1, wdd->max_hw_heartbeat_ms / 1000);
if (action) writel(gwdt->clk * timeout,
linux-stable-mirror@lists.linaro.org