[PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Guenter Roeck]

Building the driver on xtensa fails with

tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
	in function `tegra186_timer_remove':
timer-tegra186.c:(.text+0x350):
	undefined reference to `__udivdi3'

Avoid the problem by rearranging the offending code to avoid the 64-bit
divide operation.

Fixes: 28c842c8b0f5 ("clocksource/drivers/timer-tegra186: Add WDIOC_GETTIMELEFT support")
Cc: Pohsun Su <pohsuns@nvidia.com>
Cc: Robert Lin <robelin@nvidia.com>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
---
 drivers/clocksource/timer-tegra186.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/clocksource/timer-tegra186.c b/drivers/clocksource/timer-tegra186.c
index e5394f98a02e..7b506de65438 100644
--- a/drivers/clocksource/timer-tegra186.c
+++ b/drivers/clocksource/timer-tegra186.c
@@ -267,7 +267,7 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
 	 * counter value to the time of the counter expirations that
 	 * remain.
 	 */
-	timeleft += (((u64)wdt->base.timeout * USEC_PER_SEC) / 5) * (4 - expiration);
+	timeleft += ((u64)wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
 
 	/*
 	 * Convert the current counter value to seconds,
-- 
2.45.2

[PATCH 2/2] clocksource/drivers/timer-tegra186: Simplify calculating timeleft

[Guenter Roeck]

It is not necessary to use 64-bit operations to calculate the
remaining watchdog timeout. Simplify to use 32-bit operations,
and add comments explaining why there will be no overflow.

Cc: Pohsun Su <pohsuns@nvidia.com>
Cc: Robert Lin <robelin@nvidia.com>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
---
 drivers/clocksource/timer-tegra186.c | 25 +++++++++++++++----------
 1 file changed, 15 insertions(+), 10 deletions(-)

diff --git a/drivers/clocksource/timer-tegra186.c b/drivers/clocksource/timer-tegra186.c
index 7b506de65438..6ed319bb4e06 100644
--- a/drivers/clocksource/timer-tegra186.c
+++ b/drivers/clocksource/timer-tegra186.c
@@ -231,7 +231,7 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
 {
 	struct tegra186_wdt *wdt = to_tegra186_wdt(wdd);
 	u32 expiration, val;
-	u64 timeleft;
+	u32 timeleft;
 
 	if (!watchdog_active(&wdt->base)) {
 		/* return zero if the watchdog timer is not activated. */
@@ -266,21 +266,26 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
 	 * Calculate the time remaining by adding the time for the
 	 * counter value to the time of the counter expirations that
 	 * remain.
+	 * Note: Since wdt->base.timeout is bound to 255, the maximum
+	 * value added to timeleft is
+	 *   255 * (1,000,000 / 5) * 4
+	 * = 255 * 200,000 * 4
+	 * = 204,000,000
+	 * TMRSR_PCV is a 29-bit field.
+	 * Its maximum value is 0x1fffffff = 536,870,911.
+	 * 204,000,000 + 536,870,911 = 740,870,911 = 0x2C28CAFF.
+	 * timeleft can therefore not overflow, and 64-bit calculations
+	 * are not necessary.
 	 */
-	timeleft += ((u64)wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
+	timeleft += (wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
 
 	/*
 	 * Convert the current counter value to seconds,
-	 * rounding up to the nearest second. Cast u64 to
-	 * u32 under the assumption that no overflow happens
-	 * when coverting to seconds.
+	 * rounding to the nearest second.
 	 */
-	timeleft = DIV_ROUND_CLOSEST_ULL(timeleft, USEC_PER_SEC);
+	timeleft = DIV_ROUND_CLOSEST(timeleft, USEC_PER_SEC);
 
-	if (WARN_ON_ONCE(timeleft > U32_MAX))
-		return U32_MAX;
-
-	return lower_32_bits(timeleft);
+	return timeleft;
 }
 
 static const struct watchdog_ops tegra186_wdt_ops = {
-- 
2.45.2

Re: [PATCH 2/2] clocksource/drivers/timer-tegra186: Simplify calculating timeleft

[Jon Hunter]

On 14/06/2025 18:55, Guenter Roeck wrote:
> It is not necessary to use 64-bit operations to calculate the
> remaining watchdog timeout. Simplify to use 32-bit operations,
> and add comments explaining why there will be no overflow.
> 
> Cc: Pohsun Su <pohsuns@nvidia.com>
> Cc: Robert Lin <robelin@nvidia.com>
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---
>   drivers/clocksource/timer-tegra186.c | 25 +++++++++++++++----------
>   1 file changed, 15 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/clocksource/timer-tegra186.c b/drivers/clocksource/timer-tegra186.c
> index 7b506de65438..6ed319bb4e06 100644
> --- a/drivers/clocksource/timer-tegra186.c
> +++ b/drivers/clocksource/timer-tegra186.c
> @@ -231,7 +231,7 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
>   {
>   	struct tegra186_wdt *wdt = to_tegra186_wdt(wdd);
>   	u32 expiration, val;
> -	u64 timeleft;
> +	u32 timeleft;
>   
>   	if (!watchdog_active(&wdt->base)) {
>   		/* return zero if the watchdog timer is not activated. */
> @@ -266,21 +266,26 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
>   	 * Calculate the time remaining by adding the time for the
>   	 * counter value to the time of the counter expirations that
>   	 * remain.
> +	 * Note: Since wdt->base.timeout is bound to 255, the maximum
> +	 * value added to timeleft is
> +	 *   255 * (1,000,000 / 5) * 4
> +	 * = 255 * 200,000 * 4
> +	 * = 204,000,000
> +	 * TMRSR_PCV is a 29-bit field.
> +	 * Its maximum value is 0x1fffffff = 536,870,911.
> +	 * 204,000,000 + 536,870,911 = 740,870,911 = 0x2C28CAFF.
> +	 * timeleft can therefore not overflow, and 64-bit calculations
> +	 * are not necessary.

Nit, I may have worded this the other way around and said that 'timeleft 
  cannot overflow a 32-bit type and so 32-bit variables are sufficient 
for this calculation'. Simply because there are no longer any 64-bit 
variables used now. Anyway from the history of the change it will be 
clear where this came from.

>   	 */
> -	timeleft += ((u64)wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
> +	timeleft += (wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
>   
>   	/*
>   	 * Convert the current counter value to seconds,
> -	 * rounding up to the nearest second. Cast u64 to
> -	 * u32 under the assumption that no overflow happens
> -	 * when coverting to seconds.
> +	 * rounding to the nearest second.
>   	 */
> -	timeleft = DIV_ROUND_CLOSEST_ULL(timeleft, USEC_PER_SEC);
> +	timeleft = DIV_ROUND_CLOSEST(timeleft, USEC_PER_SEC);
>   
> -	if (WARN_ON_ONCE(timeleft > U32_MAX))
> -		return U32_MAX;
> -
> -	return lower_32_bits(timeleft);
> +	return timeleft;
>   }
>   
>   static const struct watchdog_ops tegra186_wdt_ops = {

Otherwise ...

Reviewed-by: Jon Hunter <jonathanh@nvidia.com>

Thanks!
Jon

-- 
nvpublic

Re: [PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Jon Hunter]

On 14/06/2025 18:55, Guenter Roeck wrote:
> Building the driver on xtensa fails with
> 
> tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
> 	in function `tegra186_timer_remove':
> timer-tegra186.c:(.text+0x350):
> 	undefined reference to `__udivdi3'
> 
> Avoid the problem by rearranging the offending code to avoid the 64-bit
> divide operation.
> 
> Fixes: 28c842c8b0f5 ("clocksource/drivers/timer-tegra186: Add WDIOC_GETTIMELEFT support")
> Cc: Pohsun Su <pohsuns@nvidia.com>
> Cc: Robert Lin <robelin@nvidia.com>
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---
>   drivers/clocksource/timer-tegra186.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/clocksource/timer-tegra186.c b/drivers/clocksource/timer-tegra186.c
> index e5394f98a02e..7b506de65438 100644
> --- a/drivers/clocksource/timer-tegra186.c
> +++ b/drivers/clocksource/timer-tegra186.c
> @@ -267,7 +267,7 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
>   	 * counter value to the time of the counter expirations that
>   	 * remain.
>   	 */
> -	timeleft += (((u64)wdt->base.timeout * USEC_PER_SEC) / 5) * (4 - expiration);
> +	timeleft += ((u64)wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
>   
>   	/*
>   	 * Convert the current counter value to seconds,


Reviewed-by: Jon Hunter <jonathanh@nvidia.com>

Thanks!
Jon

-- 
nvpublic

Re: [PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Daniel Lezcano]

On Sat, Jun 14, 2025 at 10:55:55AM -0700, Guenter Roeck wrote:
> Building the driver on xtensa fails with
> 
> tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
> 	in function `tegra186_timer_remove':
> timer-tegra186.c:(.text+0x350):
> 	undefined reference to `__udivdi3'
> 
> Avoid the problem by rearranging the offending code to avoid the 64-bit
> divide operation.
> 

Hi Guenter,

Arnd posted a version which seems more complete.

https://lore.kernel.org/all/20250620111939.3395525-1-arnd@kernel.org/

> Fixes: 28c842c8b0f5 ("clocksource/drivers/timer-tegra186: Add WDIOC_GETTIMELEFT support")
> Cc: Pohsun Su <pohsuns@nvidia.com>
> Cc: Robert Lin <robelin@nvidia.com>
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---
>  drivers/clocksource/timer-tegra186.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/clocksource/timer-tegra186.c b/drivers/clocksource/timer-tegra186.c
> index e5394f98a02e..7b506de65438 100644
> --- a/drivers/clocksource/timer-tegra186.c
> +++ b/drivers/clocksource/timer-tegra186.c
> @@ -267,7 +267,7 @@ static unsigned int tegra186_wdt_get_timeleft(struct watchdog_device *wdd)
>  	 * counter value to the time of the counter expirations that
>  	 * remain.
>  	 */
> -	timeleft += (((u64)wdt->base.timeout * USEC_PER_SEC) / 5) * (4 - expiration);
> +	timeleft += ((u64)wdt->base.timeout * (USEC_PER_SEC / 5)) * (4 - expiration);
>  
>  	/*
>  	 * Convert the current counter value to seconds,
> -- 
> 2.45.2
> 

-- 

 <http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog

Re: [PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Jon Hunter]

Hi Daniel,

On 24/06/2025 15:32, Daniel Lezcano wrote:
> On Sat, Jun 14, 2025 at 10:55:55AM -0700, Guenter Roeck wrote:
>> Building the driver on xtensa fails with
>>
>> tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
>> 	in function `tegra186_timer_remove':
>> timer-tegra186.c:(.text+0x350):
>> 	undefined reference to `__udivdi3'
>>
>> Avoid the problem by rearranging the offending code to avoid the 64-bit
>> divide operation.
>>
> 
> Hi Guenter,
> 
> Arnd posted a version which seems more complete.
> 
> https://lore.kernel.org/all/20250620111939.3395525-1-arnd@kernel.org/

Actually Arnd responded here saying this version looks better :-)

https://lore.kernel.org/all/5a571236-7de0-42bc-bfb2-52a57e210852@app.fastmail.com/

OK, to pick up this series now?

Thanks
Jon

-- 
nvpublic

Re: [PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Daniel Lezcano]

On 03/07/2025 14:34, Jon Hunter wrote:
> Hi Daniel,
> 
> On 24/06/2025 15:32, Daniel Lezcano wrote:
>> On Sat, Jun 14, 2025 at 10:55:55AM -0700, Guenter Roeck wrote:
>>> Building the driver on xtensa fails with
>>>
>>> tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
>>>     in function `tegra186_timer_remove':
>>> timer-tegra186.c:(.text+0x350):
>>>     undefined reference to `__udivdi3'
>>>
>>> Avoid the problem by rearranging the offending code to avoid the 64-bit
>>> divide operation.
>>>
>>
>> Hi Guenter,
>>
>> Arnd posted a version which seems more complete.
>>
>> https://lore.kernel.org/all/20250620111939.3395525-1-arnd@kernel.org/
> 
> Actually Arnd responded here saying this version looks better :-)
> 
> https://lore.kernel.org/all/5a571236-7de0-42bc- 
> bfb2-52a57e210852@app.fastmail.com/
> 
> OK, to pick up this series now?

Yes, I'm ok.

May be Thomas can pick them up for the next -rc ?

-- 
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog

Re: [PATCH 1/2] clocksource/drivers/timer-tegra186: Avoid 64-bit divide operation

[Daniel Lezcano]

On Sat, Jun 14, 2025 at 10:55:55AM -0700, Guenter Roeck wrote:
> Building the driver on xtensa fails with
> 
> tensa-linux-ld: drivers/clocksource/timer-tegra186.o:
> 	in function `tegra186_timer_remove':
> timer-tegra186.c:(.text+0x350):
> 	undefined reference to `__udivdi3'
> 
> Avoid the problem by rearranging the offending code to avoid the 64-bit
> divide operation.
> 
> Fixes: 28c842c8b0f5 ("clocksource/drivers/timer-tegra186: Add WDIOC_GETTIMELEFT support")
> Cc: Pohsun Su <pohsuns@nvidia.com>
> Cc: Robert Lin <robelin@nvidia.com>
> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
> ---

Applied, thanks

-- 

 <http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog