Skip to contentSkip to conversation

A few fixes to sprd watchdog driver

PATCH Superseded Chunyan Zhang 2 participants 10 messages Oct 26, 2020 – Oct 27, 2020

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Chunyan Zhang

author reviewer
#0 lore raw

cover A few fixes to sprd watchdog driver

From: Chunyan Zhang <[email protected]>

A few issues about sprd watchdog driver were found recently, this patchset would fix them.

Lingling Xu (3):
watchdog: sprd: should not disable watchdog in resume
watchdog: sprd: change timeout value from 1000 to 2000
watchdog: sprd: check busy bit before kick watchdog

 drivers/watchdog/sprd_wdt.c | 38 +++++++++++++++++--------------------
 1 file changed, 17 insertions(+), 21 deletions(-)
Signature · 1 line
2.20.1
author reviewer
#1 lore raw

PATCH 1/3 watchdog: sprd: should not disable watchdog in resume

From: Lingling Xu <[email protected]>

Don't disable watchdog in resume process, otherwise system would crash once kick watchdog.

Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/watchdog/sprd_wdt.c +2 -7
index 65cb55f3916f..f3c90b4afead 100644
@@ -345,15 +345,10 @@ static int __maybe_unused sprd_wdt_pm_resume(struct device *dev)
 	if (ret)
 		return ret;
 
-	if (watchdog_active(&wdt->wdd)) {
+	if (watchdog_active(&wdt->wdd))
 		ret = sprd_wdt_start(&wdt->wdd);
-		if (ret) {
-			sprd_wdt_disable(wdt);
-			return ret;
-		}
-	}
 
-	return 0;
+	return ret;
 }
 
 static const struct dev_pm_ops sprd_wdt_pm_ops = {
Signature · 1 line
2.20.1
maintainer Reviewed-by on v1 1/3
#2 lore raw

Re: [PATCH 1/3] watchdog: sprd: should not disable watchdog in resume

Quoted 5 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

Don't disable watchdog in resume process, otherwise system would crash once kick watchdog.

This is a bit misleading: It is only disabled if the attempt to start it has failed. Was this observed in practice ? If so, it might make sense to identify and fix the underlying problem instead of trying to work around it (or is this addressed with the second patch of the series ?)

Anyway, the patch itself is fine.

Thanks,
Guenter

Quoted 30 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/watchdog/sprd_wdt.c +2 -7
index 65cb55f3916f..f3c90b4afead 100644
@@ -345,15 +345,10 @@ static int __maybe_unused sprd_wdt_pm_resume(struct device *dev)
 	if (ret)
 		return ret;
 
-	if (watchdog_active(&wdt->wdd)) {
+	if (watchdog_active(&wdt->wdd))
 		ret = sprd_wdt_start(&wdt->wdd);
-		if (ret) {
-			sprd_wdt_disable(wdt);
-			return ret;
-		}
-	}
 
-	return 0;
+	return ret;
 }
 
 static const struct dev_pm_ops sprd_wdt_pm_ops = {
author reviewer
#3 lore raw

Re: [PATCH 1/3] watchdog: sprd: should not disable watchdog in resume

Quoted 4 lines · Guenter Roeck
Quoted 2 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

Quoted 8 lines
Quoted 3 lines

Don't disable watchdog in resume process, otherwise system would crash once kick watchdog.

This is a bit misleading: It is only disabled if the attempt to start it has failed. Was this observed in practice ? If so, it might make sense to identify and fix the underlying problem instead of trying to work around it (or is this addressed with the second patch of the series ?)

Yes, I think the root cause of this problem was like what I explained in the 3rd patch in this series.
Lingling found there was something wrong in sprd_wdt_pm_resume() when debugging that issue, then we had this patch.

Quoted 4 lines

Anyway, the patch itself is fine.

Thanks,
Chunyan

Quoted 35 lines

Thanks,
Guenter

Quoted 30 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/watchdog/sprd_wdt.c +2 -7
index 65cb55f3916f..f3c90b4afead 100644
@@ -345,15 +345,10 @@ static int __maybe_unused sprd_wdt_pm_resume(struct device *dev)
      if (ret)
              return ret;
 
-     if (watchdog_active(&wdt->wdd)) {
+     if (watchdog_active(&wdt->wdd))
              ret = sprd_wdt_start(&wdt->wdd);
-             if (ret) {
-                     sprd_wdt_disable(wdt);
-                     return ret;
-             }
-     }
 
-     return 0;
+     return ret;
 }
 
 static const struct dev_pm_ops sprd_wdt_pm_ops = {
author reviewer
#4 lore raw

PATCH 2/3 watchdog: sprd: change timeout value from 1000 to 2000

In reply to Chunyan Zhang

From: Lingling Xu <[email protected]>

Because cpu_relax() takes different time on different SoCs, for some rare cases, it would take more than 1000 cycles for waitting load operation finished. The result of many times testing verified that changing the timeout value to 2000 can solve the issue.

Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/watchdog/sprd_wdt.c +1 -1
index f3c90b4afead..4f2a8c6d6485 100644
@@ -53,7 +53,7 @@
 
 #define SPRD_WDT_CNT_HIGH_SHIFT		16
 #define SPRD_WDT_LOW_VALUE_MASK		GENMASK(15, 0)
-#define SPRD_WDT_LOAD_TIMEOUT		1000
+#define SPRD_WDT_LOAD_TIMEOUT		2000
 
 struct sprd_wdt {
 	void __iomem *base;
Signature · 1 line
2.20.1
maintainer
#5 lore raw

Re: [PATCH 2/3] watchdog: sprd: change timeout value from 1000 to 2000

Quoted 4 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

Because cpu_relax() takes different time on different SoCs, for some rare cases, it would take more than 1000 cycles for waitting load operation

waiting

Quoted 3 lines

finished. The result of many times testing verified that changing the timeout value to 2000 can solve the issue.

This is just a kludge that doesn't address the underlying problem. As the wait loop states, "Waiting the load value operation done, it needs two or three RTC clock cycles". This means the loop should wait for a maximum number of clock cycles, and not run as hot loop. If we assume that clk_get_rate() returns the clock frequency, that frequency can be used to determine how long this needs to be retried. It might also make sense - depending on how long this actually takes - to use usleep_range() instead of cpu_relax() to avoid the hot loop.

Guenter

Quoted 21 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/watchdog/sprd_wdt.c +1 -1
index f3c90b4afead..4f2a8c6d6485 100644
@@ -53,7 +53,7 @@
 
 #define SPRD_WDT_CNT_HIGH_SHIFT		16
 #define SPRD_WDT_LOW_VALUE_MASK		GENMASK(15, 0)
-#define SPRD_WDT_LOAD_TIMEOUT		1000
+#define SPRD_WDT_LOAD_TIMEOUT		2000
 
 struct sprd_wdt {
 	void __iomem *base;
author reviewer
#6 lore raw

Re: [PATCH 2/3] watchdog: sprd: change timeout value from 1000 to 2000

Quoted 8 lines · Guenter Roeck
Quoted 4 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

Because cpu_relax() takes different time on different SoCs, for some rare cases, it would take more than 1000 cycles for waitting load operation

waiting

Ok.

Quoted 5 lines
Quoted 4 lines

finished. The result of many times testing verified that changing the timeout value to 2000 can solve the issue.

Quoted 9 lines

This is just a kludge that doesn't address the underlying problem. As the wait loop states, "Waiting the load value operation done, it needs two or three RTC clock cycles". This means the loop should wait for a maximum number of clock cycles, and not run as hot loop. If we assume that clk_get_rate() returns the clock frequency, that frequency can be used to determine how long this needs to be retried. It might also make sense - depending on how long this actually takes - to use usleep_range() instead of cpu_relax() to avoid the hot loop.

Agree, using usleep_range() instead makes more sense, I will look into that.

Thanks for your review.

Chunyan

Quoted 25 lines

Guenter

Quoted 21 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/watchdog/sprd_wdt.c +1 -1
index f3c90b4afead..4f2a8c6d6485 100644
@@ -53,7 +53,7 @@
 
 #define SPRD_WDT_CNT_HIGH_SHIFT              16
 #define SPRD_WDT_LOW_VALUE_MASK              GENMASK(15, 0)
-#define SPRD_WDT_LOAD_TIMEOUT                1000
+#define SPRD_WDT_LOAD_TIMEOUT                2000
 
 struct sprd_wdt {
      void __iomem *base;
author reviewer
#7 lore raw

PATCH 3/3 watchdog: sprd: check busy bit before kick watchdog

In reply to Chunyan Zhang

From: Lingling Xu <[email protected]>

As the specification described, checking busy bit must be done before kick watchdog.

Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)
drivers/watchdog/sprd_wdt.c +14 -13
index 4f2a8c6d6485..14071c66ff49 100644
@@ -108,20 +108,8 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
 	u32 tmr_step = timeout * SPRD_WDT_CNT_STEP;
 	u32 prtmr_step = pretimeout * SPRD_WDT_CNT_STEP;
 
-	sprd_wdt_unlock(wdt->base);
-	writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-		      SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
-	writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
-		       wdt->base + SPRD_WDT_LOAD_LOW);
-	writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-			SPRD_WDT_LOW_VALUE_MASK,
-		       wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
-	writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
-		       wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
-	sprd_wdt_lock(wdt->base);
-
 	/*
-	 * Waiting the load value operation done,
+	 * Waiting the last load value operation done,
 	 * it needs two or three RTC clock cycles.
 	 */
 	do {
@@ -134,6 +122,19 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
 
 	if (delay_cnt >= SPRD_WDT_LOAD_TIMEOUT)
 		return -EBUSY;
+
+	sprd_wdt_unlock(wdt->base);
+	writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+		      SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
+	writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
+		       wdt->base + SPRD_WDT_LOAD_LOW);
+	writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+			SPRD_WDT_LOW_VALUE_MASK,
+		       wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
+	writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
+		       wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
+	sprd_wdt_lock(wdt->base);
+
 	return 0;
 }
 
Signature · 1 line
2.20.1
maintainer
#8 lore raw

Re: [PATCH 3/3] watchdog: sprd: check busy bit before kick watchdog

Quoted 5 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

As the specification described, checking busy bit must be done before kick watchdog.

That is a key functional change: So far the code checked if a value was accepted after loading it. That is no longer the case. Effectively, with this change, the _next_ operation will now check if the previous operation was accepted. Is this intentional ?

Also, does this really solve a problem, or is it just an optimization ? By checking for busy prior to an operation instead of after it the only real difference is that the busy check will most likely succeed immediately because enough time has passed since the last write.

Ultimately it is your call how you want to handle this, but I think the impact should be spelled out.

Guenter

Quoted 54 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)
drivers/watchdog/sprd_wdt.c +14 -13
index 4f2a8c6d6485..14071c66ff49 100644
@@ -108,20 +108,8 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
 	u32 tmr_step = timeout * SPRD_WDT_CNT_STEP;
 	u32 prtmr_step = pretimeout * SPRD_WDT_CNT_STEP;
 
-	sprd_wdt_unlock(wdt->base);
-	writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-		      SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
-	writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
-		       wdt->base + SPRD_WDT_LOAD_LOW);
-	writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-			SPRD_WDT_LOW_VALUE_MASK,
-		       wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
-	writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
-		       wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
-	sprd_wdt_lock(wdt->base);
-
 	/*
-	 * Waiting the load value operation done,
+	 * Waiting the last load value operation done,
 	 * it needs two or three RTC clock cycles.
 	 */
 	do {
@@ -134,6 +122,19 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
 
 	if (delay_cnt >= SPRD_WDT_LOAD_TIMEOUT)
 		return -EBUSY;
+
+	sprd_wdt_unlock(wdt->base);
+	writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+		      SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
+	writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
+		       wdt->base + SPRD_WDT_LOAD_LOW);
+	writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+			SPRD_WDT_LOW_VALUE_MASK,
+		       wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
+	writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
+		       wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
+	sprd_wdt_lock(wdt->base);
+
 	return 0;
 }
 
author reviewer
#9 lore raw

Re: [PATCH 3/3] watchdog: sprd: check busy bit before kick watchdog

Quoted 4 lines · Guenter Roeck
Quoted 2 lines · Chunyan Zhang

From: Lingling Xu <[email protected]>

Quoted 8 lines
Quoted 3 lines

As the specification described, checking busy bit must be done before kick watchdog.

That is a key functional change: So far the code checked if a value was accepted after loading it. That is no longer the case. Effectively, with this change, the _next_ operation will now check if the previous operation was accepted. Is this intentional ?

Yes, the busy bit indicates whether the previous operation is done, so we have to make sure the last loading completed (the busy bit is not set) before new loading.

The spec says that this bit is set after a new loading, and would last 2 or 3 RTC clock cycles.

Quoted 8 lines

Also, does this really solve a problem, or is it just an optimization ? By checking for busy prior to an operation instead of after it the only real difference is that the busy check will most likely succeed immediately because enough time has passed since the last write.

Ultimately it is your call how you want to handle this, but I think the impact should be spelled out.

Ok, I will add more details in the commit message.

Many thanks for the review!

Chunyan

Quoted 58 lines

Guenter

Quoted 54 lines
Fixes 477603467009 ("watchdog: Add Spreadtrum watchdog driver")
---
 drivers/watchdog/sprd_wdt.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)
drivers/watchdog/sprd_wdt.c +14 -13
index 4f2a8c6d6485..14071c66ff49 100644
@@ -108,20 +108,8 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
      u32 tmr_step = timeout * SPRD_WDT_CNT_STEP;
      u32 prtmr_step = pretimeout * SPRD_WDT_CNT_STEP;
 
-     sprd_wdt_unlock(wdt->base);
-     writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-                   SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
-     writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
-                    wdt->base + SPRD_WDT_LOAD_LOW);
-     writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
-                     SPRD_WDT_LOW_VALUE_MASK,
-                    wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
-     writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
-                    wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
-     sprd_wdt_lock(wdt->base);
-
      /*
-      * Waiting the load value operation done,
+      * Waiting the last load value operation done,
       * it needs two or three RTC clock cycles.
       */
      do {
@@ -134,6 +122,19 @@ static int sprd_wdt_load_value(struct sprd_wdt *wdt, u32 timeout,
 
      if (delay_cnt >= SPRD_WDT_LOAD_TIMEOUT)
              return -EBUSY;
+
+     sprd_wdt_unlock(wdt->base);
+     writel_relaxed((tmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+                   SPRD_WDT_LOW_VALUE_MASK, wdt->base + SPRD_WDT_LOAD_HIGH);
+     writel_relaxed((tmr_step & SPRD_WDT_LOW_VALUE_MASK),
+                    wdt->base + SPRD_WDT_LOAD_LOW);
+     writel_relaxed((prtmr_step >> SPRD_WDT_CNT_HIGH_SHIFT) &
+                     SPRD_WDT_LOW_VALUE_MASK,
+                    wdt->base + SPRD_WDT_IRQ_LOAD_HIGH);
+     writel_relaxed(prtmr_step & SPRD_WDT_LOW_VALUE_MASK,
+                    wdt->base + SPRD_WDT_IRQ_LOAD_LOW);
+     sprd_wdt_lock(wdt->base);
+
      return 0;
 }