Skip to contentSkip to conversation

hwmon: (ina2xx) Fix overflow issues

PATCH Superseded Guenter Roeck 2 participants 2 messages Jun 10 Hardware monitoring

v2 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Guenter Roeck

author maintainer
#0 lore raw

PATCH v2 1/1 hwmon: (ina2xx) Fix overflow issues

Sashiko reports the following overflow problems:

In ina2xx_get_value(), the INA2XX_POWER calculation is:

    val = regval * data->power_lsb_uW;

The result is returned as a signed 32-bit int. For the INA232 with a common 2mOhm shunt, power_lsb_uW becomes 40,000. When the 16-bit regval exceeds 53,687, the product exceeds INT_MAX. This overflows the 32-bit signed math, wrapping to a negative integer and reporting erroneous negative power readings to userspace.

For INA2XX_POWER, the upper bound is clamped:

    val = clamp_val(val, 0, UINT_MAX - data->power_lsb_uW);

clamp_val() implicitly casts the upper bound to a 32-bit signed long on 32-bit platforms. This results in a negative number, meaning any valid power limit is clamped to a negative value and ultimately programmed as 0. Similarly, for INA2XX_SHUNT_VOLTAGE, the initial clamp uses SHRT_MAX * shunt_div instead of division. The subsequent multiplication:

    val *= data->config->shunt_div;

overflows LONG_MAX on 32-bit platforms for high inputs, wrapping to a negative value and also programming the hardware limit to 0.

For INA2XX_BUS_VOLTAGE on parts with bus_voltage_shift > 0, the calculation (val * 1000) << shift can exceed LONG_MAX for limits over ~134V, wrapping to negative and setting the limit to 0.

For INA2XX_CURRENT:
On 32-bit systems, long is 32-bit signed. For configurations with small shunts, current_lsb_uA can be very large (e.g., 40,000,000). When multiplied by a large regval, the product can reach 1.3 trillion, massively exceeding LONG_MAX (2.14 billion).
This will silently overflow the 32-bit signed math, wrapping to a negative value.

If a large limit is provided to effectively disable the alert (e.g., INT_MAX / 1000), the intermediate value after DIV_ROUND_CLOSEST can be large (e.g., 200,000,000). When left-shifted by current_shift (e.g., 4 for INA234), it becomes 3.2 billion, exceeding LONG_MAX and wrapping to a negative number.

In sy24655_average_power_read(), the accumulator quotient multiplied by power_lsb_uW can overflow the 32-bit signed math before the assignment.

Have ina2xx_get_value() return a long variable to improve the supported value range on 64-bit systems and to match the type of values returned to the hwmon core.

Clamp the result of 'regval * data->power_lsb_uW' to LONG_MAX to fix the INA2XX_POWER calculation overflow.

Change the initial clamp for INA2XX_SHUNT_VOLTAGE to SHRT_MAX / shunt_div to fix the shunt limit overflow.

For INA2XX_BUS_VOLTAGE, limit the initial clamp to 130V instead of 200V to avoid the overflow.

For INA2XX_CURRENT, improve clamping to avoid the overflow.

To address the INA2XX_CURRENT problem in ina226_alert_to_reg(), take current_shift into acount for the initial clamp to avoid the overflow.

In sy24655_average_power_read, use a temporary 64-bit variable to store the multiplication result and clamp the result against LONG_MAX.

Cc Loic Poulain <[email protected]>
Fixes ab7fbee452be ("hwmon: (ina2xx) Fix various overflow issues")

---
v2: Fixed several additional overflow conditions reported by Sashiko
after v1

 drivers/hwmon/ina2xx.c | 29 ++++++++++++++++++-----------
 1 file changed, 18 insertions(+), 11 deletions(-)
drivers/hwmon/ina2xx.c +18 -11
index c4742e84b999..1ba46c9c3e20 100644
@@ -16,6 +16,7 @@
 #include <linux/i2c.h>
 #include <linux/init.h>
 #include <linux/kernel.h>
+#include <linux/limits.h>
 #include <linux/module.h>
 #include <linux/property.h>
 #include <linux/regmap.h>
@@ -266,10 +267,11 @@ static u16 ina226_interval_to_reg(long interval)
 	return FIELD_PREP(INA226_AVG_RD_MASK, avg_bits);
 }
 
-static int ina2xx_get_value(struct ina2xx_data *data, u8 reg,
-			    unsigned int regval)
+static long ina2xx_get_value(struct ina2xx_data *data, u8 reg,
+			     unsigned int regval)
 {
-	int val;
+	s64 val64;
+	long val;
 
 	switch (reg) {
 	case INA2XX_SHUNT_VOLTAGE:
@@ -283,13 +285,13 @@ static int ina2xx_get_value(struct ina2xx_data *data, u8 reg,
 		val = DIV_ROUND_CLOSEST(val, 1000);
 		break;
 	case INA2XX_POWER:
-		val = regval * data->power_lsb_uW;
+		val = clamp_val((u64)regval * data->power_lsb_uW, 0, LONG_MAX);
 		break;
 	case INA2XX_CURRENT:
 		/* signed register, result in mA */
-		val = ((s16)regval >> data->config->current_shift) *
+		val64 = (s64)((s16)regval >> data->config->current_shift) *
 		  data->current_lsb_uA;
-		val = DIV_ROUND_CLOSEST(val, 1000);
+		val = clamp_val(DIV_ROUND_CLOSEST(val64, 1000), LONG_MIN, LONG_MAX);
 		break;
 	case INA2XX_CALIBRATION:
 		val = regval;
@@ -378,23 +380,26 @@ static int ina2xx_read_init(struct device *dev, int reg, long *val)
  */
 static u16 ina226_alert_to_reg(struct ina2xx_data *data, int reg, long val)
 {
+	long limit;
+
 	switch (reg) {
 	case INA2XX_SHUNT_VOLTAGE:
-		val = clamp_val(val, 0, SHRT_MAX * data->config->shunt_div);
+		val = clamp_val(val, 0, DIV_ROUND_CLOSEST(SHRT_MAX, data->config->shunt_div));
 		val *= data->config->shunt_div;
 		val <<= data->config->shunt_voltage_shift;
 		return clamp_val(val, 0, SHRT_MAX);
 	case INA2XX_BUS_VOLTAGE:
-		val = clamp_val(val, 0, 200000);
+		val = clamp_val(val, 0, 130000);
 		val = (val * 1000) << data->config->bus_voltage_shift;
 		val = DIV_ROUND_CLOSEST(val, data->config->bus_voltage_lsb);
 		return clamp_val(val, 0, USHRT_MAX);
 	case INA2XX_POWER:
-		val = clamp_val(val, 0, UINT_MAX - data->power_lsb_uW);
+		val = clamp_val(val, 0, LONG_MAX - data->power_lsb_uW);
 		val = DIV_ROUND_CLOSEST(val, data->power_lsb_uW);
 		return clamp_val(val, 0, USHRT_MAX);
 	case INA2XX_CURRENT:
-		val = clamp_val(val, INT_MIN / 1000, INT_MAX / 1000);
+		limit = (INT_MAX / 1000) >> data->config->current_shift;
+		val = clamp_val(val, -limit, limit);
 		/* signed register, result in mA */
 		val = DIV_ROUND_CLOSEST(val * 1000, data->current_lsb_uA);
 		val <<= data->config->current_shift;
@@ -537,6 +542,7 @@ static int sy24655_average_power_read(struct ina2xx_data *data, u8 reg, long *va
 	u8 template[6];
 	int ret;
 	long accumulator_24, sample_count;
+	u64 val64;
 
 	/* 48-bit register read */
 	ret = i2c_smbus_read_i2c_block_data(data->client, reg, 6, template);
@@ -555,7 +561,8 @@ static int sy24655_average_power_read(struct ina2xx_data *data, u8 reg, long *va
 		return 0;
 	}
 
-	*val = DIV_ROUND_CLOSEST(accumulator_24, sample_count) * data->power_lsb_uW;
+	val64 = (u64)DIV_ROUND_CLOSEST(accumulator_24, sample_count) * data->power_lsb_uW;
+	*val = clamp_val(val64, 0, LONG_MAX);
 
 	return 0;
 }
Signature · 1 line
2.45.2
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [Low] Build breakage: Using DIV_ROUND_CLOSEST() with an s64 dividend causes a linker error (undefined reference to '__divdi3') on 32-bit platforms.

Pre-existing issues:
- [High] Arithmetic overflow in ina226_alert_to_reg() during DIV_ROUND_CLOSEST intermediate calculation on 32-bit platforms. --

commit 2efec75618467bc75c51301a14034dc4018a9013
Author: Guenter Roeck <[email protected]>

hwmon: (ina2xx) Fix overflow issues

This patch aims to fix multiple overflow issues in the ina2xx driver, specifically when calculating power, voltage, and current values on 32-bit platforms. It introduces 64-bit variables and clamped limits to prevent numbers from wrapping around.

Quoted 4 lines
drivers/hwmon/ina2xx.c +0 -0
index 613ffb622b7c4..b2f543b89f9e8 100644

[ ... ]

Quoted 4 lines
drivers/hwmon/ina2xx.c +0 -0
@@ -283,13 +285,13 @@ static int ina2xx_get_value(struct ina2xx_data *data, u8 reg,
 		val = DIV_ROUND_CLOSEST(val, 1000);
 		break;
 	case INA2XX_POWER:

drivers/hwmon/ina2xx.c static int ina2xx_get_value(struct ina2xx_data *data, u8 reg, · line 294

Quoted 10 lines
drivers/hwmon/ina2xx.c +3 -3
-		val = regval * data->power_lsb_uW;
+		val = clamp_val((u64)regval * data->power_lsb_uW, 0, LONG_MAX);
 		break;
 	case INA2XX_CURRENT:
 		/* signed register, result in mA */
-		val = ((s16)regval >> data->config->current_shift) *
+		val64 = (s64)((s16)regval >> data->config->current_shift) *
 		  data->current_lsb_uA;
-		val = DIV_ROUND_CLOSEST(val, 1000);
+		val = clamp_val(DIV_ROUND_CLOSEST(val64, 1000), LONG_MIN, LONG_MAX);

[Severity: Low]
Does this cause a build failure on 32-bit platforms?

The variable val64 is an s64, and DIV_ROUND_CLOSEST relies on the native division operator. Doing a 64-bit signed division by a non-power-of-2 constant like 1000 typically causes a linker error (undefined reference to __divdi3) because the kernel does not link against libgcc.

Should this use div_s64() instead?

[ ... ]

drivers/hwmon/ina2xx.c static u16 ina226_alert_to_reg(struct ina2xx_data *data, int reg, long val) · line 388

Quoted 12 lines
drivers/hwmon/ina2xx.c +3 -2
@@ -378,23 +380,26 @@ static u16 ina226_alert_to_reg(struct ina2xx_data *data, int reg, long val)
 	case INA2XX_POWER:
-		val = clamp_val(val, 0, UINT_MAX - data->power_lsb_uW);
+		val = clamp_val(val, 0, LONG_MAX - data->power_lsb_uW);
 		val = DIV_ROUND_CLOSEST(val, data->power_lsb_uW);
 		return clamp_val(val, 0, USHRT_MAX);
 	case INA2XX_CURRENT:
-		val = clamp_val(val, INT_MIN / 1000, INT_MAX / 1000);
+		limit = (INT_MAX / 1000) >> data->config->current_shift;
+		val = clamp_val(val, -limit, limit);
 		/* signed register, result in mA */
 		val = DIV_ROUND_CLOSEST(val * 1000, data->current_lsb_uA);

[Severity: High]
This is a pre-existing issue, but can the intermediate calculation inside DIV_ROUND_CLOSEST overflow the 32-bit signed integer?

The macro internally performs:
(val * 1000) + (data->current_lsb_uA / 2)

On 32-bit systems where long is 32-bit, val * 1000 can be up to 2,147,483,000. If data->current_lsb_uA is large (e.g., 40,000,000 as noted in the commit message), adding current_lsb_uA / 2 (20,000,000) yields 2,167,483,000.

This exceeds LONG_MAX (2,147,483,647) and silently wraps to a negative value. Will this result in a large positive limit requested by the user being incorrectly programmed as a negative limit in hardware?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1