Skip to contentSkip to conversation

i2c: designware: Fix corrupted memory seen in the ISR

PATCH Applied Jan Bottorff 6 participants 10 messages Nov 9, 2023 – Nov 16, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 i2c: designware: Fix corrupted memory seen in the ISR

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

---
ChangeLog
v3->v4: add missing changelog
v2->v3: regmap accessors use non-relaxed form instead of wmb barrier v1->v2: Commit message improvements
v1: insert wmb barrier before enabling interrupts

 drivers/i2c/busses/i2c-designware-common.c | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)
drivers/i2c/busses/i2c-designware-common.c +8 -8
index affcfb243f0f..35f762872b8a 100644
@@ -63,7 +63,7 @@ static int dw_reg_read(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = readl_relaxed(dev->base + reg);
+	*val = readl(dev->base + reg);
 
 	return 0;
 }
@@ -72,7 +72,7 @@ static int dw_reg_write(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writel_relaxed(val, dev->base + reg);
+	writel(val, dev->base + reg);
 
 	return 0;
 }
@@ -81,7 +81,7 @@ static int dw_reg_read_swab(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = swab32(readl_relaxed(dev->base + reg));
+	*val = swab32(readl(dev->base + reg));
 
 	return 0;
 }
@@ -90,7 +90,7 @@ static int dw_reg_write_swab(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writel_relaxed(swab32(val), dev->base + reg);
+	writel(swab32(val), dev->base + reg);
 
 	return 0;
 }
@@ -99,8 +99,8 @@ static int dw_reg_read_word(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = readw_relaxed(dev->base + reg) |
-		(readw_relaxed(dev->base + reg + 2) << 16);
+	*val = readw(dev->base + reg) |
+		(readw(dev->base + reg + 2) << 16);
 
 	return 0;
 }
@@ -109,8 +109,8 @@ static int dw_reg_write_word(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writew_relaxed(val, dev->base + reg);
-	writew_relaxed(val >> 16, dev->base + reg + 2);
+	writew(val, dev->base + reg);
+	writew(val >> 16, dev->base + reg + 2);
 
 	return 0;
 }
Signature · 1 line
2.34.1
#1 lore raw
Quoted 23 lines · Jan Bottorff

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

...

Quoted 1 line

v3->v4: add missing changelog

Side note: Usually it's enough to just reply to the patch with the changelog.

...

drivers/i2c/busses/i2c-designware-common.c static int dw_reg_read_swab(void *context, unsigned int reg, unsigned int *val)

Quoted 5 lines
drivers/i2c/busses/i2c-designware-common.c +2 -2
-	*val = swab32(readl_relaxed(dev->base + reg));
+	*val = swab32(readl(dev->base + reg));
 
-	writel_relaxed(swab32(val), dev->base + reg);
+	writel(swab32(val), dev->base + reg);

I'm wondering why ioread32be() / iowrite32be() can't be used here...

Probably it would require to switch entire IO to use ioreadXX() / iowriteXX() APIs and since we touch all of them (?) may be it makes sense convert to use them at the same time. Dunno.

Signature · 2 lines
With Best Regards,
Andy Shevchenko
Acked-by
#2 lore raw
Quoted 36 lines · Andy Shevchenko
Quoted 23 lines · Jan Bottorff

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

...

Quoted 1 line

v3->v4: add missing changelog

Side note: Usually it's enough to just reply to the patch with the changelog.

...

Quoted 2 lines
drivers/i2c/busses/i2c-designware-common.c +1 -1
-	*val = swab32(readl_relaxed(dev->base + reg));
+	*val = swab32(readl(dev->base + reg));
Quoted 9 lines

drivers/i2c/busses/i2c-designware-common.c static int dw_reg_write_swab(void *context, unsigned int reg, unsigned int val) · line 93

Quoted 2 lines
drivers/i2c/busses/i2c-designware-common.c +1 -1
-	writel_relaxed(swab32(val), dev->base + reg);
+	writel(swab32(val), dev->base + reg);

I'm wondering why ioread32be() / iowrite32be() can't be used here...

Probably it would require to switch entire IO to use ioreadXX() / iowriteXX() APIs and since we touch all of them (?) may be it makes sense convert to use them at the same time. Dunno.

I would say cosmetic conversions are better to go to another patch.

For this patch:

Applied
#3 lore raw
In reply to Jan Bottorff
Quoted 24 lines · Jan Bottorff

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

Quoted 1 line

Applied to for-current, thanks!

Tested-by
#4 lore raw
Quoted 25 lines · Wolfram Sang
Quoted 24 lines · Jan Bottorff

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

Quoted 3 lines
Quoted 1 line

Applied to for-current, thanks!

A bit late but FYI:

Signature · 1 line
Yann
Tested-by Reviewed-by on v4 1/1
#5 lore raw
In reply to Jan Bottorff
Quoted 24 lines · Jan Bottorff

When running on a many core ARM64 server, errors were happening in the ISR that looked like corrupted memory. These corruptions would fix themselves if small delays were inserted in the ISR. Errors reported by the driver included "i2c_designware APMC0D0F:00: i2c_dw_xfer_msg: invalid target address" and "i2c_designware APMC0D0F:00:controller timed out" during in-band IPMI SSIF stress tests.

The problem was determined to be memory writes in the driver were not becoming visible to all cores when execution rapidly shifted between cores, like when a register write immediately triggers an ISR. Processors with weak memory ordering, like ARM64, make no guarantees about the order normal memory writes become globally visible, unless barrier instructions are used to control ordering.

To solve this, regmap accessor functions configured by this driver were changed to use non-relaxed forms of the low-level register access functions, which include a barrier on platforms that require it. This assures memory writes before a controller register access are visible to all cores. The community concluded defaulting to correct operation outweighed defaulting to the small performance gains from using relaxed access functions. Being a low speed device added weight to this choice of default register access behavior.

Quoted 1 line

The patch has already been merged in, but in anyway:

For the record. Next time don't forget to add all the reviewers to the Cc-list on the patch(set) respin, otherwise the new version might get to be missed amongst the other messages in their inbox. That in its turn won't let them test it out and finish the review on time.

-Serge(y)

Quoted 75 lines

---
ChangeLog
v3->v4: add missing changelog
v2->v3: regmap accessors use non-relaxed form instead of wmb barrier v1->v2: Commit message improvements
v1: insert wmb barrier before enabling interrupts

 drivers/i2c/busses/i2c-designware-common.c | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)
drivers/i2c/busses/i2c-designware-common.c +8 -8
index affcfb243f0f..35f762872b8a 100644
@@ -63,7 +63,7 @@ static int dw_reg_read(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = readl_relaxed(dev->base + reg);
+	*val = readl(dev->base + reg);
 
 	return 0;
 }
@@ -72,7 +72,7 @@ static int dw_reg_write(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writel_relaxed(val, dev->base + reg);
+	writel(val, dev->base + reg);
 
 	return 0;
 }
@@ -81,7 +81,7 @@ static int dw_reg_read_swab(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = swab32(readl_relaxed(dev->base + reg));
+	*val = swab32(readl(dev->base + reg));
 
 	return 0;
 }
@@ -90,7 +90,7 @@ static int dw_reg_write_swab(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writel_relaxed(swab32(val), dev->base + reg);
+	writel(swab32(val), dev->base + reg);
 
 	return 0;
 }
@@ -99,8 +99,8 @@ static int dw_reg_read_word(void *context, unsigned int reg, unsigned int *val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	*val = readw_relaxed(dev->base + reg) |
-		(readw_relaxed(dev->base + reg + 2) << 16);
+	*val = readw(dev->base + reg) |
+		(readw(dev->base + reg + 2) << 16);
 
 	return 0;
 }
@@ -109,8 +109,8 @@ static int dw_reg_write_word(void *context, unsigned int reg, unsigned int val)
 {
 	struct dw_i2c_dev *dev = context;
 
-	writew_relaxed(val, dev->base + reg);
-	writew_relaxed(val >> 16, dev->base + reg + 2);
+	writew(val, dev->base + reg);
+	writew(val >> 16, dev->base + reg + 2);
 
 	return 0;
 }
Signature · 1 line
2.34.1
#8 lore raw
Quoted 4 lines · Wolfram Sang
Quoted 3 lines

Thanks to a restrictive hotel network, I haven't pushed out yet, and could still add your tags. Thanks!

Hi Wolfram,

Any chance we could get the "Cc: [email protected]" tag added to this patch? More than one large cloud company would like to see this in the stable kernel, as it significantly improves the reliability of IPMI transactions on platforms that use i2c for this communication.

Sorry for not including this tag in the submission.

Jan

#9 lore raw
Quoted 3 lines

Any chance we could get the "Cc: [email protected]" tag added to this patch? More than one large cloud company would like to see this in the

Well, no, it is already pushed out...

Quoted 2 lines

stable kernel, as it significantly improves the reliability of IPMI transactions on platforms that use i2c for this communication.

... but with that commit desc it will surely be backported nonetheless. If all fails, it can be manually requested to go to stable, but I am sure this will not be needed.

Quoted 1 line

Sorry for not including this tag in the submission.

No worries.