leds: pwm: Annotate leds[] with __counted_by()
PATCH v3 1/1 leds: pwm: Annotate leds[] with __counted_by()
Add the __counted_by() attribute to the flexible array member leds[] in struct led_pwm_priv so the compiler and runtime (e.g. FORTIFY_SOURCE, UBSAN_BOUNDS) can bounds-check accesses against num_leds.
For the annotation to be correct, num_leds must equal the number of allocated elements before leds[] is accessed. Even taking the address &priv->leds[i] is bounds-checked against num_leds under UBSAN_BOUNDS. The driver allocates device_get_child_node_count() elements up front, so set num_leds to that count right after allocation.
While here, pass the led_pwm_data element into led_pwm_add() instead of the whole led_pwm_priv, so the helper no longer needs to index the array; the caller walks a led_pwm_data pointer over priv->leds.
No functional change intended.
---
Changes in v3:
- v2 followed the review suggestion to pass the element and increment num_leds in the caller. That is exactly what Sashiko caught: with __counted_by(num_leds), &priv->leds[num_leds] is bounds-checked, so incrementing num_leds from 0 as the loop runs trips UBSAN_BOUNDS on probe. num_leds is now set to the full count up front and the caller walks a led_pwm_data pointer, which keeps the element-passing and the dropped index variable, just without the running counter.
Changes in v2 (per Lee Jones review):
- Pass the led_pwm_data element into led_pwm_add() so it drops the priv and index arguments.
Build-tested only (no PWM-LED hardware). The __counted_by bounds behaviour that motivated v3 was verified with a standalone UBSAN reproducer on gcc and clang.
drivers/leds/leds-pwm.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-)
drivers/leds/leds-pwm.c +6 -5
@@ -36,7 +36,7 @@ struct led_pwm_data {struct led_pwm_priv {int num_leds; -struct led_pwm_data leds[]; + struct led_pwm_data leds[] __counted_by(num_leds); }; static int led_pwm_set(struct led_classdev *led_cdev,
@@ -81,10 +81,9 @@ static int led_pwm_default_brightness_get(struct fwnode_handle *fwnode, } __attribute__((nonnull)) -static int led_pwm_add (struct device *dev,struct led_pwm_priv *priv, +static int led_pwm_add (struct device *dev,struct led_pwm_data *led_data,struct led_pwm *led,struct fwnode_handle *fwnode) { -struct led_pwm_data *led_data = &priv->leds[priv->num_leds];struct led_init_data init_data = { .fwnode = fwnode };int ret;
@@ -167,12 +166,12 @@ static int led_pwm_add(struct device *dev, struct led_pwm_priv *priv, } } - priv->num_leds++;return 0 ; }static int led_pwm_create_fwnode (struct device *dev,struct led_pwm_priv *priv) { +struct led_pwm_data *led_data = priv->leds;struct led_pwm led;int ret;
@@ -193,7 +192,7 @@ static int led_pwm_create_fwnode(struct device *dev, struct led_pwm_priv *priv) led.default_state =led_init_default_state_get (fwnode); - ret =led_pwm_add (dev, priv, &led, fwnode); + ret =led_pwm_add (dev, led_data++, &led, fwnode);if (ret)return ret; }
@@ -217,6 +216,8 @@ static int led_pwm_probe(struct platform_device *pdev)if (!priv)return -ENOMEM ; + priv->num_leds = count; + ret =led_pwm_create_fwnode (&pdev->dev, priv);if (ret)
Signature · 1 line
2.55.0
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] The refactoring to use pointer arithmetic (led_data++) bypasses the newly added __counted_by()…
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] The refactoring to use pointer arithmetic (led_data++) bypasses the newly added __counted_by() bounds checking. --
--- Patch [1]: [PATCH v3] leds: pwm: Annotate leds[] with __counted_by() --- commit c418c024dbd5055ff67d14fc9a244c4d09795169
Author: Mert Seftali <[email protected]>
leds: pwm: Annotate leds[] with __counted_by()
Adds the __counted_by() attribute to the flexible array member leds[] in struct led_pwm_priv for bounds checking. To avoid UBSAN triggers during initialization, the code is refactored to pass a led_pwm_data pointer into led_pwm_add() instead of indexing the array directly.
Quoted 12 lines
drivers/leds/leds-pwm.c +1
-2@@ -81,10 +81,9 @@ static int led_pwm_default_brightness_get(struct fwnode_handle *fwnode, } __attribute__((nonnull)) -static int led_pwm_add (struct device *dev,struct led_pwm_priv *priv, +static int led_pwm_add (struct device *dev,struct led_pwm_data *led_data,struct led_pwm *led,struct fwnode_handle *fwnode) { -struct led_pwm_data *led_data = &priv->leds[priv->num_leds];
[ ... ]
drivers/leds/leds-pwm.c static int led_pwm_create_fwnode(struct device *dev, struct led_pwm_priv *priv) · line 195
Quoted 12 lines
static int
led_pwm_create_fwnode(struct device *dev, struct led_pwm_priv *priv){drivers/leds/leds-pwm.c +2
-1+struct led_pwm_data *led_data = priv->leds;struct led_pwm led;int ret;@@ -193,7 +192,7 @@ static int led_pwm_create_fwnode(struct device *dev, struct led_pwm_priv *priv) led.default_state =led_init_default_state_get (fwnode); - ret =led_pwm_add (dev, priv, &led, fwnode); + ret =led_pwm_add (dev, led_data++, &led, fwnode);
[Severity: Medium]
By decaying the flexible array to a pointer and using pointer arithmetic (led_data++), doesn't the compiler lose the context needed to enforce the __counted_by bounds check?
Compilers and sanitizers typically enforce __counted_by mitigations during array subscripting operations like priv->leds[i]. Because priv->leds is decayed into a standalone pointer and passed to led_pwm_add(), all bounds information relative to priv->num_leds appears to be lost.
Will the primary objective of adding __counted_by be neutralized by passing the element as a bare pointer here rather than using an index?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Hi Lee,
Sashiko is right and I tested it too now to confirm. Under UBSAN_BOUNDS &priv->leds[i] traps on an out-of-range index, but the led_data++ walk in v3 does not: decaying the array to a plain led_pwm_data * drops the __counted_by association, so num_leds is only enforced on subscript accesses and v3 has none.
Getting the check back means &priv->leds[i] subscripting. I can still pass that element into led_pwm_add() as you suggested, it just brings a loop index back. So it is a tradeoff: the cleaner pointer form, or the annotation actually enforced.
Since the element-passing was your suggestion, which would you prefer:
keep the pointer walk (annotation stays documentary), or respin with subscripting so num_leds is enforced? Happy to do either.
Thanks,
Mert