Skip to contentSkip to conversation

rtc: zynqmp: ensure correct RTC calibration

PATCH In review Lothar Rubusch 2 participants 6 messages Aug 4, 2025 – Aug 8, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover rtc: zynqmp: ensure correct RTC calibration

We've observed intermittent failures with the current implementation on our Xilinx hardware. These issues were not present prior to the commit referenced by the Fixes tag, using the same hardware setup. The problem appears to be identical to the one described here:
https://adaptivesupport.amd.com/s/article/000036886?language=en_US

Instead of explicitly writing 0x7FFF to the calibration register using devmem, the patch now resets the register.

Ivan Vera (1): rtc: zynqmp: ensure correct RTC calibration

 drivers/rtc/rtc-zynqmp.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

base-commit: 260f6f4fda93c8485c8037865c941b42b9cba5d2

Signature · 1 line
2.39.5
#1 lore raw

PATCH 1/1 rtc: zynqmp: ensure correct RTC calibration

From: Ivan Vera <[email protected]>

In the event of an uninitialized calibration register, ensure the register is reset and properly programmed during the probe sequence.

At present, only the calibration register is evaluated. If it holds invalid values after a power cycle, there's no longer a way to reset it, for instance, via a devicetree entry to 0x7FFF. This issue is documented here:
https://adaptivesupport.amd.com/s/article/000036886?language=en_US

The fix prioritizes an optional calibration value provided via the devicetree over the value in the register.

Fixes 07dcc6f9c76275d6679f28a69e042a2f9dc8f128 ("rtc: zynqmp: Add calibration set and get support")
---
 drivers/rtc/rtc-zynqmp.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/rtc/rtc-zynqmp.c +3 -3
index f39102b66eac..0c063c3fae52 100644
@@ -331,9 +331,9 @@ static int xlnx_rtc_probe(struct platform_device *pdev)
 		if (ret)
 			xrtcdev->freq = RTC_CALIB_DEF;
 	}
-	ret = readl(xrtcdev->reg_base + RTC_CALIB_RD);
-	if (!ret)
-		writel(xrtcdev->freq, (xrtcdev->reg_base + RTC_CALIB_WR));
+
+	/* Enable unconditional re-calibration to RTC_CALIB_DEF or DTB entry. */
+	writel(xrtcdev->freq, xrtcdev->reg_base + RTC_CALIB_WR);
 
 	xlnx_init_rtc(xrtcdev);
 
Signature · 1 line
2.39.5
#2 lore raw
Quoted 25 lines · Lothar Rubusch

From: Ivan Vera <[email protected]>

In the event of an uninitialized calibration register, ensure the register is reset and properly programmed during the probe sequence.

At present, only the calibration register is evaluated. If it holds invalid values after a power cycle, there's no longer a way to reset it, for instance, via a devicetree entry to 0x7FFF. This issue is documented here:
https://adaptivesupport.amd.com/s/article/000036886?language=en_US

The fix prioritizes an optional calibration value provided via the devicetree over the value in the register.

Fixes 07dcc6f9c76275d6679f28a69e042a2f9dc8f128 ("rtc: zynqmp: Add calibration set and get support")
---
 drivers/rtc/rtc-zynqmp.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/rtc/rtc-zynqmp.c +0 -0
index f39102b66eac..0c063c3fae52 100644

drivers/rtc/rtc-zynqmp.c static int xlnx_rtc_probe(struct platform_device *pdev) · line 336

Quoted 9 lines
drivers/rtc/rtc-zynqmp.c +3 -3
@@ -331,9 +331,9 @@ static int xlnx_rtc_probe(struct platform_device *pdev)
 		if (ret)
 			xrtcdev->freq = RTC_CALIB_DEF;
 	}
-	ret = readl(xrtcdev->reg_base + RTC_CALIB_RD);
-	if (!ret)
-		writel(xrtcdev->freq, (xrtcdev->reg_base + RTC_CALIB_WR));
+
+	/* Enable unconditional re-calibration to RTC_CALIB_DEF or DTB entry. */
+	writel(xrtcdev->freq, xrtcdev->reg_base + RTC_CALIB_WR);

Doesn't this forcefully overwrite the proper value that has been set from userspace and so trashes the time at each reboot?

Isn't the proper way to reset it to simply set the offset from userspace again?

Signature · 3 lines
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#3 lore raw
Quoted 3 lines · Alexandre Belloni
Quoted 1 line · Lothar Rubusch

From: Ivan Vera <[email protected]>

(...)

Quoted 9 lines
Quoted 9 lines
drivers/rtc/rtc-zynqmp.c +0 -1
index f39102b66eac..0c063c3fae52 100644
@@ -331,9 +331,9 @@ static int xlnx_rtc_probe(struct platform_device *pdev)
              if (ret)
                      xrtcdev->freq = RTC_CALIB_DEF;
      }
-     ret = readl(xrtcdev->reg_base + RTC_CALIB_RD);
Quoted 9 lines

drivers/rtc/rtc-zynqmp.c static int xlnx_rtc_probe(struct platform_device *pdev) · line 336

Quoted 5 lines
drivers/rtc/rtc-zynqmp.c +3 -2
-     if (!ret)
-             writel(xrtcdev->freq, (xrtcdev->reg_base + RTC_CALIB_WR));
+
+     /* Enable unconditional re-calibration to RTC_CALIB_DEF or DTB entry. */
+     writel(xrtcdev->freq, xrtcdev->reg_base + RTC_CALIB_WR);

Doesn't this forcefully overwrite the proper value that has been set from userspace and so trashes the time at each reboot?

Yes, it overwrites the calibration, i.e. counting 1sec in about 1sec. No, the time/date is not actually "trashed" (I double-checked that with timesyncd disabled, having and not having register content and over several reboots keeping a bogus date/time - it psersistet in the same time space. The current patch always overwrites the calib register content. So, a manual userspace setting will be lost after reboot. That's true.

Would it rather make sense to extend it, say, instead of merely checking whether the calibration register contains any data - which could potentially be incorrect - also check for the presence of a calibration property in the devicetree (or a similar property, since 'calibration' may be deprecated)? If such a property exists, perform a re-calibration based on the devicetree at every reboot. Otherwise, retain the current behavior of checking whether the register is empty?

Quoted 3 lines

Isn't the proper way to reset it to simply set the offset from userspace again?

Hm.. I'm unsure if I understood you correctly. You mean the way as described in AMD's link to perform the reset by executing the devmem from Linux manually? If so, why is it preferable to adjust the RTC calibration manually every time this happens, rather than to simply put a correction value into the devicetree properties for problematic setups? Or do I miss something, is there a config file for RTC calibration for doing this persistently from Linux, that I'm not aware of?

Before, the devicetree properties seemed to have generally priority over userspace settings. Now, after the calibration rework, this priorization seems to have changed and a devicetree calib correction for such problematic cases will generally be ignored, with a recommendation by Xilinx/AMD (see link in cover letter) to execute a devmem command from off Linux (...). I mean, can't this be elaborated a bit more to allow for a persistent correction method?

Just, let me know what you think about. Thank you for the feedback. Best,
L

Quoted 4 lines

--
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com

#4 lore raw
Quoted 24 lines · Lothar Rubusch
Quoted 3 lines · Alexandre Belloni
Quoted 1 line · Lothar Rubusch

From: Ivan Vera <[email protected]>

(...)

Quoted 9 lines
Quoted 9 lines
drivers/rtc/rtc-zynqmp.c +0 -1
index f39102b66eac..0c063c3fae52 100644
@@ -331,9 +331,9 @@ static int xlnx_rtc_probe(struct platform_device *pdev)
              if (ret)
                      xrtcdev->freq = RTC_CALIB_DEF;
      }
-     ret = readl(xrtcdev->reg_base + RTC_CALIB_RD);
Quoted 9 lines

drivers/rtc/rtc-zynqmp.c static int xlnx_rtc_probe(struct platform_device *pdev) · line 336

Quoted 5 lines
drivers/rtc/rtc-zynqmp.c +3 -2
-     if (!ret)
-             writel(xrtcdev->freq, (xrtcdev->reg_base + RTC_CALIB_WR));
+
+     /* Enable unconditional re-calibration to RTC_CALIB_DEF or DTB entry. */
+     writel(xrtcdev->freq, xrtcdev->reg_base + RTC_CALIB_WR);

Doesn't this forcefully overwrite the proper value that has been set from userspace and so trashes the time at each reboot?

Quoted 7 lines

Yes, it overwrites the calibration, i.e. counting 1sec in about 1sec. No, the time/date is not actually "trashed" (I double-checked that with timesyncd disabled, having and not having register content and over several reboots keeping a bogus date/time - it psersistet in the same time space. The current patch always overwrites the calib register content. So, a manual userspace setting will be lost after reboot. That's true.

It is about 1sec on your platform because it didn't deviate too much from the expected value but what if another platform needs a way different value? Then you are introducing the same issue as the one you are trying to fix but it will have it at each reboot.

Quoted 21 lines

Would it rather make sense to extend it, say, instead of merely checking whether the calibration register contains any data - which could potentially be incorrect - also check for the presence of a calibration property in the devicetree (or a similar property, since 'calibration' may be deprecated)? If such a property exists, perform a re-calibration based on the devicetree at every reboot. Otherwise, retain the current behavior of checking whether the register is empty?

Quoted 3 lines

Isn't the proper way to reset it to simply set the offset from userspace again?

Hm.. I'm unsure if I understood you correctly. You mean the way as described in AMD's link to perform the reset by executing the devmem from Linux manually? If so, why is it preferable to adjust the RTC calibration manually every time this happens, rather than to simply put a correction value into the devicetree properties for problematic setups? Or do I miss something, is there a config file for RTC calibration for doing this persistently from Linux, that I'm not aware of?

Quoted 7 lines

Before, the devicetree properties seemed to have generally priority over userspace settings. Now, after the calibration rework, this priorization seems to have changed and a devicetree calib correction for such problematic cases will generally be ignored, with a recommendation by Xilinx/AMD (see link in cover letter) to execute a devmem command from off Linux (...). I mean, can't this be elaborated a bit more to allow for a persistent correction method?

The value depends on each manufactured machine/board as it is supposed to correct for imprecision on the input clock which is either a crystal or derived from a crystal. This crystal may be more or less accurate and its accuracy will change over time notably because of temperature changes or simply because it is aging. Having the value in the device tree is as good as having it hardcoded in the driver which is not far from what your are doing here. It makes the feature useless.

What I was suggesting is simply to do the right thing, compute the inaccuracy and correct it from userspace, using the proper interface that is sysfs or the RTC_PARAM_SET ioctl for RTC_PARAM_CORRECTION This has to be done regularly anyway so I guess it would catch and correct any corrupted value in the register.

Signature · 3 lines
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
#5 lore raw
Quoted 34 lines · Alexandre Belloni
Quoted 24 lines · Lothar Rubusch
Quoted 3 lines · Alexandre Belloni
Quoted 1 line · Lothar Rubusch

From: Ivan Vera <[email protected]>

(...)

Quoted 9 lines
Quoted 9 lines
drivers/rtc/rtc-zynqmp.c +0 -1
index f39102b66eac..0c063c3fae52 100644
@@ -331,9 +331,9 @@ static int xlnx_rtc_probe(struct platform_device *pdev)
              if (ret)
                      xrtcdev->freq = RTC_CALIB_DEF;
      }
-     ret = readl(xrtcdev->reg_base + RTC_CALIB_RD);
Quoted 9 lines

drivers/rtc/rtc-zynqmp.c static int xlnx_rtc_probe(struct platform_device *pdev) · line 336

Quoted 5 lines
drivers/rtc/rtc-zynqmp.c +3 -2
-     if (!ret)
-             writel(xrtcdev->freq, (xrtcdev->reg_base + RTC_CALIB_WR));
+
+     /* Enable unconditional re-calibration to RTC_CALIB_DEF or DTB entry. */
+     writel(xrtcdev->freq, xrtcdev->reg_base + RTC_CALIB_WR);

Doesn't this forcefully overwrite the proper value that has been set from userspace and so trashes the time at each reboot?

Quoted 7 lines

Yes, it overwrites the calibration, i.e. counting 1sec in about 1sec. No, the time/date is not actually "trashed" (I double-checked that with timesyncd disabled, having and not having register content and over several reboots keeping a bogus date/time - it psersistet in the same time space. The current patch always overwrites the calib register content. So, a manual userspace setting will be lost after reboot. That's true.

Quoted 5 lines

It is about 1sec on your platform because it didn't deviate too much from the expected value but what if another platform needs a way different value? Then you are introducing the same issue as the one you are trying to fix but it will have it at each reboot.

I guess you missunderstood me here a bit. I understand that every scenario will need individual calibration especially over time.

Quoted 29 lines
Quoted 21 lines

Would it rather make sense to extend it, say, instead of merely checking whether the calibration register contains any data - which could potentially be incorrect - also check for the presence of a calibration property in the devicetree (or a similar property, since 'calibration' may be deprecated)? If such a property exists, perform a re-calibration based on the devicetree at every reboot. Otherwise, retain the current behavior of checking whether the register is empty?

Quoted 3 lines

Isn't the proper way to reset it to simply set the offset from userspace again?

Hm.. I'm unsure if I understood you correctly. You mean the way as described in AMD's link to perform the reset by executing the devmem from Linux manually? If so, why is it preferable to adjust the RTC calibration manually every time this happens, rather than to simply put a correction value into the devicetree properties for problematic setups? Or do I miss something, is there a config file for RTC calibration for doing this persistently from Linux, that I'm not aware of?

Quoted 7 lines

Before, the devicetree properties seemed to have generally priority over userspace settings. Now, after the calibration rework, this priorization seems to have changed and a devicetree calib correction for such problematic cases will generally be ignored, with a recommendation by Xilinx/AMD (see link in cover letter) to execute a devmem command from off Linux (...). I mean, can't this be elaborated a bit more to allow for a persistent correction method?

Quoted 8 lines

The value depends on each manufactured machine/board as it is supposed to correct for imprecision on the input clock which is either a crystal or derived from a crystal. This crystal may be more or less accurate and its accuracy will change over time notably because of temperature changes or simply because it is aging. Having the value in the device tree is as good as having it hardcoded in the driver which is not far from what your are doing here. It makes the feature useless.

Yes, I see your point.

Quoted 6 lines

What I was suggesting is simply to do the right thing, compute the inaccuracy and correct it from userspace, using the proper interface that is sysfs or the RTC_PARAM_SET ioctl for RTC_PARAM_CORRECTION This has to be done regularly anyway so I guess it would catch and correct any corrupted value in the register.

The degradation over time and or temperature does not match the static approach we took of our v1 patch. I modified it still a bit to keep userspace configurations, and use the optional devicetree property for a correction. But this does not cancel out your argumentation, since the approach is still static. So, I have to agree, thanks.

Best,
L

Quoted 4 lines

--
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com