Skip to contentSkip to conversation

pwm: lp3943: validate DT output indices in lp3943_pwm_parse_dt()

PATCH In review Manush Prajwal 3 participants 3 messages Sep 6 – Nov 13

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

PATCH v2 1/1 pwm: lp3943: validate DT output indices in lp3943_pwm_parse_dt()

The ti,pwm0/ti,pwm1 devicetree properties are raw u32 arrays naming which of the 16 physical LP3943 outputs (0-15, per the documented binding and the 16-entry enum lp3943_pwm_output) each PWM channel drives. lp3943_pwm_parse_dt() reads these values straight from the devicetree with of_property_read_u32_array() and stores them in pwm_map->output[] without checking they are actually in range.

Both consumers of pwm_map->output[] then use an out-of-range value directly as an index/bit-offset with no bounds check of their own:

- lp3943_pwm_request_map() does

       test_and_set_bit(offset, &lp3943->pin_used)

with offset taken straight from output[i], corrupting memory beyond the single 'unsigned long pin_used' field for offset >= BITS_PER_LONG.

- lp3943_pwm_set_mode() does
mux[index].reg / .mask / .shift
with index taken straight from output[i], reading past the 16-entry lp3943_mux_cfg[] array and then issuing a regmap write built from whatever garbage register/mask/shift values were read out of bounds.

A devicetree (malformed board file or a loaded overlay) with an out-of-range ti,pwm0/ti,pwm1 value is enough to reach either path; nothing upstream of these two functions validates the value.

Add a range check in lp3943_pwm_parse_dt() right after the values are read, rejecting the devicetree with -EINVAL before any out-of-range value can reach pwm_map->output[]. Check explicitly for negative values in addition to the upper bound: enum lp3943_pwm_output's underlying integer representation is implementation-defined per the C standard, so do not rely on the compiler having chosen an unsigned type for it even though it always does in practice for GCC (verified empirically, with and without -fshort-enums, since none of the enumerators are negative). Add the LP3943_NUM_OUTPUTS(16) constant next to the existing enum it bounds, alongside the existing LP3943_NUM_PWMS(2) constant.

---
Changes since v1:
- Sashiko AI review flagged that the bounds check could theoretically be bypassed by a devicetree value with the MSB set, if the compiler chose a signed underlying type for enum lp3943_pwm_output. I checked this empirically with gcc 13.2.0, both with and without -fshort-enums: since none of the enum's values are negative, gcc always picks an unsigned underlying type, so output[j] >= 16 already rejects such values today. Added the explicit "< 0" check anyway, since the C standard leaves the enum's underlying type implementation-defined and the check costs nothing.
- The same review also flagged a missing NULL check on pdata->pwms[hwpwm] in lp3943_pwm_request_map() as a pre-existing, separate bug (not introduced by this patch, confirmed by inspection) - sending that as its own follow-up patch rather than folding an unrelated fix in here.

 drivers/pwm/pwm-lp3943.c   | 7 ++++++-
 include/linux/mfd/lp3943.h | 2 ++
 2 files changed, 8 insertions(+), 1 deletion(-)
drivers/pwm/pwm-lp3943.c +6 -1
index 10537e74b..e8bb8dbb9 100644
@@ -218,7 +218,7 @@ static int lp3943_pwm_parse_dt(struct device *dev,
 	struct lp3943_platform_data *pdata;
 	struct lp3943_pwm_map *pwm_map;
 	enum lp3943_pwm_output *output;
-	int i, err, num_outputs, count = 0;
+	int i, j, err, num_outputs, count = 0;
 
 	if (!node)
 		return -EINVAL;
@@ -247,6 +247,11 @@ static int lp3943_pwm_parse_dt(struct device *dev,
 		if (err)
 			return err;
 
+		for (j = 0; j < num_outputs; j++) {
+			if (output[j] < 0 || output[j] >= LP3943_NUM_OUTPUTS)
+				return -EINVAL;
+		}
+
 		pwm_map = devm_kzalloc(dev, sizeof(*pwm_map), GFP_KERNEL);
 		if (!pwm_map)
 			return -ENOMEM;
include/linux/mfd/lp3943.h +2 -0
index 5d2d172d3..844938ca7 100644
@@ -53,6 +53,8 @@ enum lp3943_pwm_output {
 	LP3943_PWM_OUT15,
 };
 
+#define LP3943_NUM_OUTPUTS		16
+
 /*
  * struct lp3943_pwm_map
  * @output: Output pins which are mapped to each PWM channel
Signature · 1 line
2.46.2.windows.1
#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:

Pre-existing issues:
- [Critical] Passing an enum array pointer to of_property_read_u32_array causes a heap buffer overflow and array stride mismatch if the compiler uses a smaller underlying type (e.g., with -fshort-enums).
- [High] Missing NULL check on pdata->pwms[hwpwm] in lp3943_pwm_request_map() causes a kernel crash if an unconfigured PWM channel is requested.
--

commit 2bff859954d5825f1b63378dcea622868d103bda
Author: Manush Prajwal <[email protected]>

pwm: lp3943: validate DT output indices in lp3943_pwm_parse_dt()

This commit adds bounds checking for the ti,pwm0 and ti,pwm1 devicetree properties in lp3943_pwm_parse_dt(). It verifies the raw u32 array values are within the valid range of 0-15 before storing them, preventing memory corruption and out-of-bounds reads in downstream consumers.

Quoted 7 lines

Both consumers of pwm_map->output[] then use an out-of-range value directly as an index/bit-offset with no bounds check of their own:

- lp3943_pwm_request_map() does
test_and_set_bit(offset, &lp3943->pin_used)
with offset taken straight from output[i], corrupting memory beyond the single 'unsigned long pin_used' field for offset >= BITS_PER_LONG.

[Severity: High]
This isn't a bug introduced by this patch, but does lp3943_pwm_request_map() lack a NULL pointer check for an unconfigured hardware PWM channel?

If a board DT does not define all PWM channels (so pdata->pwms[1] is NULL), and a consumer requests it, won't we dereference a NULL pointer here:

drivers/pwm/pwm-lp3943.c:lp3943_pwm_request_map() {
    ...
    pwm_map->output = pdata->pwms[hwpwm]->output;
    ...
}

[ ... ]

Quoted 4 lines
drivers/pwm/pwm-lp3943.c +0 -0
index 10537e74be288..e8bb8dbb9b67b 100644

[ ... ]

drivers/pwm/pwm-lp3943.c static int lp3943_pwm_parse_dt(struct device *dev, · line 250

Quoted 5 lines
drivers/pwm/pwm-lp3943.c +1 -0
@@ -247,6 +247,11 @@ static int lp3943_pwm_parse_dt(struct device *dev,
 		if (err)
 			return err;
 
+		for (j = 0; j < num_outputs; j++) {

[Severity: Critical]
This is a pre-existing issue, but can passing an enum array pointer to of_property_read_u32_array() cause a heap buffer overflow?

If the compiler uses a smaller underlying type for enum lp3943_pwm_output (e.g., with -fshort-enums), devm_kcalloc() will allocate less than 4 bytes per element.

Since of_property_read_u32_array() expects a u32 * buffer and writes 4 bytes per element, won't it overwrite adjacent heap memory? Also, wouldn't array indexing with output[j] then read with a mismatched stride?

Quoted 7 lines
drivers/pwm/pwm-lp3943.c +4 -0
+			if (output[j] < 0 || output[j] >= LP3943_NUM_OUTPUTS)
+				return -EINVAL;
+		}
+
 		pwm_map = devm_kzalloc(dev, sizeof(*pwm_map), GFP_KERNEL);
 		if (!pwm_map)
 			return -ENOMEM;
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Manush Prajwal
Quoted 73 lines · Manush Prajwal

The ti,pwm0/ti,pwm1 devicetree properties are raw u32 arrays naming which of the 16 physical LP3943 outputs (0-15, per the documented binding and the 16-entry enum lp3943_pwm_output) each PWM channel drives. lp3943_pwm_parse_dt() reads these values straight from the devicetree with of_property_read_u32_array() and stores them in pwm_map->output[] without checking they are actually in range.

Both consumers of pwm_map->output[] then use an out-of-range value directly as an index/bit-offset with no bounds check of their own:

- lp3943_pwm_request_map() does

       test_and_set_bit(offset, &lp3943->pin_used)

with offset taken straight from output[i], corrupting memory beyond the single 'unsigned long pin_used' field for offset >= BITS_PER_LONG.

- lp3943_pwm_set_mode() does
mux[index].reg / .mask / .shift
with index taken straight from output[i], reading past the 16-entry lp3943_mux_cfg[] array and then issuing a regmap write built from whatever garbage register/mask/shift values were read out of bounds.

A devicetree (malformed board file or a loaded overlay) with an out-of-range ti,pwm0/ti,pwm1 value is enough to reach either path; nothing upstream of these two functions validates the value.

Add a range check in lp3943_pwm_parse_dt() right after the values are read, rejecting the devicetree with -EINVAL before any out-of-range value can reach pwm_map->output[]. Check explicitly for negative values in addition to the upper bound: enum lp3943_pwm_output's underlying integer representation is implementation-defined per the C standard, so do not rely on the compiler having chosen an unsigned type for it even though it always does in practice for GCC (verified empirically, with and without -fshort-enums, since none of the enumerators are negative). Add the LP3943_NUM_OUTPUTS(16) constant next to the existing enum it bounds, alongside the existing LP3943_NUM_PWMS(2) constant.

---
Changes since v1:
- Sashiko AI review flagged that the bounds check could theoretically be bypassed by a devicetree value with the MSB set, if the compiler chose a signed underlying type for enum lp3943_pwm_output. I checked this empirically with gcc 13.2.0, both with and without -fshort-enums: since none of the enum's values are negative, gcc always picks an unsigned underlying type, so output[j] >= 16 already rejects such values today. Added the explicit "< 0" check anyway, since the C standard leaves the enum's underlying type implementation-defined and the check costs nothing.
- The same review also flagged a missing NULL check on pdata->pwms[hwpwm] in lp3943_pwm_request_map() as a pre-existing, separate bug (not introduced by this patch, confirmed by inspection) - sending that as its own follow-up patch rather than folding an unrelated fix in here.

 drivers/pwm/pwm-lp3943.c   | 7 ++++++-
 include/linux/mfd/lp3943.h | 2 ++
 2 files changed, 8 insertions(+), 1 deletion(-)
drivers/pwm/pwm-lp3943.c +1 -1
index 10537e74b..e8bb8dbb9 100644
@@ -218,7 +218,7 @@ static int lp3943_pwm_parse_dt(struct device *dev,
 	struct lp3943_platform_data *pdata;
 	struct lp3943_pwm_map *pwm_map;
 	enum lp3943_pwm_output *output;
-	int i, err, num_outputs, count = 0;
+	int i, j, err, num_outputs, count = 0;
 
 	if (!node)
 		return -EINVAL;

drivers/pwm/pwm-lp3943.c static int lp3943_pwm_parse_dt(struct device *dev, · line 254

Quoted 8 lines
drivers/pwm/pwm-lp3943.c +5 -0
@@ -247,6 +247,11 @@ static int lp3943_pwm_parse_dt(struct device *dev,
 		if (err)
 			return err;
 
+		for (j = 0; j < num_outputs; j++) {
+			if (output[j] < 0 || output[j] >= LP3943_NUM_OUTPUTS)
+				return -EINVAL;
+		}
+

A few lines above we have:

                num_outputs = of_property_count_u32_elems(node, name[i]);
                if (num_outputs <= 0)
                        continue;

I wonder if it makes sense to clamp num_outputs also using LP3943_NUM_OUTPUTS?!

Best regards
Uwe