Skip to contentSkip to conversation

i2c: qcom-cci: fix device_node refcount leak in cci_probe()/cci_remove()

PATCH Superseded Liu Zhenlong 4 participants 9 messages Aug 15 – Sep 16

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 i2c: qcom-cci: fix device_node refcount leak in cci_probe()/cci_remove()

cci_probe() calls of_node_get() to take an extra reference on the child device_node when assigning it to the adapter device. The matching of_node_put() calls exist in both the error cleanup path and cci_remove(), but they are placed after i2c_del_adapter().

i2c_del_adapter() clears adap->dev with memset() at the end (commit bd4bc3dbded9 ("i2c: Clear i2c_adapter.dev on adapter removal")), which zeroes adap->dev.of_node before of_node_put() runs, turning it into a no-op. The reference taken by of_node_get() is never released, leaking the device_node on every cleanup of already-registered adapters and every adapter removal.

The commit that added of_node_get() and the matching of_node_put() calls placed the puts after i2c_del_adapter(), so the bug has been present since the fix was introduced.

Cache the pointer before calling i2c_del_adapter(), the same approach used in i2c-mux (i2c_mux_del_adapters) and mtd (commit 56570bdad5e3 ("mtd: core: Fix refcount error in del_mtd_device()")).

The of_node_put() in the i2c_add_adapter() failure path (before any i2c_del_adapter() runs) is correct and left unchanged.

Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5
---
 drivers/i2c/busses/i2c-qcom-cci.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/i2c/busses/i2c-qcom-cci.c +6 -2
index bdeda3979c48..61c535b8196f 100644
@@ -618,8 +618,10 @@ static int cci_probe(struct platform_device *pdev)
 
 	for (--i ; i >= 0; i--) {
 		if (cci->master[i].cci) {
+			struct device_node *node = cci->master[i].adap.dev.of_node;
+
 			i2c_del_adapter(&cci->master[i].adap);
-			of_node_put(cci->master[i].adap.dev.of_node);
+			of_node_put(node);
 		}
 	}
 disable_clocks:
@@ -635,8 +637,10 @@ static void cci_remove(struct platform_device *pdev)
 
 	for (i = 0; i < cci->data->num_masters; i++) {
 		if (cci->master[i].cci) {
+			struct device_node *node = cci->master[i].adap.dev.of_node;
+
 			i2c_del_adapter(&cci->master[i].adap);
-			of_node_put(cci->master[i].adap.dev.of_node);
+			of_node_put(node);
 			cci_halt(cci, i);
 		}
 	}
Signature · 1 line
2.55.0
#1 lore raw
Quoted 36 lines · Liu Zhenlong

cci_probe() calls of_node_get() to take an extra reference on the child device_node when assigning it to the adapter device. The matching of_node_put() calls exist in both the error cleanup path and cci_remove(), but they are placed after i2c_del_adapter().

i2c_del_adapter() clears adap->dev with memset() at the end (commit bd4bc3dbded9 ("i2c: Clear i2c_adapter.dev on adapter removal")), which zeroes adap->dev.of_node before of_node_put() runs, turning it into a no-op. The reference taken by of_node_get() is never released, leaking the device_node on every cleanup of already-registered adapters and every adapter removal.

The commit that added of_node_get() and the matching of_node_put() calls placed the puts after i2c_del_adapter(), so the bug has been present since the fix was introduced.

Cache the pointer before calling i2c_del_adapter(), the same approach used in i2c-mux (i2c_mux_del_adapters) and mtd (commit 56570bdad5e3 ("mtd: core: Fix refcount error in del_mtd_device()")).

The of_node_put() in the i2c_add_adapter() failure path (before any i2c_del_adapter() runs) is correct and left unchanged.

Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5
---
 drivers/i2c/busses/i2c-qcom-cci.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/i2c/busses/i2c-qcom-cci.c +0 -0
index bdeda3979c48..61c535b8196f 100644

drivers/i2c/busses/i2c-qcom-cci.c static int cci_probe(struct platform_device *pdev) · line 624

Quoted 8 lines
drivers/i2c/busses/i2c-qcom-cci.c +3 -1
@@ -618,8 +618,10 @@ static int cci_probe(struct platform_device *pdev)
 
 	for (--i ; i >= 0; i--) {
 		if (cci->master[i].cci) {
+			struct device_node *node = cci->master[i].adap.dev.of_node;
+
 			i2c_del_adapter(&cci->master[i].adap);
-			of_node_put(cci->master[i].adap.dev.of_node);
+			of_node_put(node);

The fix seems correct, but the way it's done is still fragile - someone/a bot will surely come around in a couple weeks with a ""simplification"" undoing your change. Would a devm action here be a better choice?

Konrad

#2 lore raw

PATCH v2 1/1 i2c: qcom-cci: fix device_node refcount leak in cci_probe()/cci_remove()

In reply to Liu Zhenlong

The of_node_put() matching of_node_get() runs after i2c_del_adapter(), whose trailing memset() zeroes adap->dev and thus adap->dev.of_node, making the put a no-op and leaking the node on every adapter removal and error cleanup.

Use a devm action: the pointer is captured at registration, out of reach of that memset(), and devres runs the put once on probe failure and detach, replacing the three manual of_node_put() calls. The setup loop uses the scoped iterator form so the child node is released automatically if devm_add_action_or_reset() fails mid-loop.

Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5

---
Changes in v2:
- Rework the fix to use a devm action (cci_put_of_node) instead of caching the pointer before i2c_del_adapter(), per Konrad Dybcio. The pointer is captured at registration, out of reach of the memset() in i2c_del_adapter(); the three manual of_node_put() calls are removed.
- Use for_each_available_child_of_node_scoped() so the child reference is released if devm_add_action_or_reset() fails mid-loop.

 drivers/i2c/busses/i2c-qcom-cci.c | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)
drivers/i2c/busses/i2c-qcom-cci.c +11 -9
index bdeda3979c48..d3528c7d15bd 100644
@@ -497,10 +497,14 @@ static const struct dev_pm_ops qcom_cci_pm = {
 	SET_RUNTIME_PM_OPS(cci_suspend_runtime, cci_resume_runtime, NULL)
 };
 
+static void cci_put_of_node(void *data)
+{
+	of_node_put(data);
+}
+
 static int cci_probe(struct platform_device *pdev)
 {
 	struct device *dev = &pdev->dev;
-	struct device_node *child;
 	struct resource *r;
 	struct cci *cci;
 	int ret, i;
@@ -516,7 +520,7 @@ static int cci_probe(struct platform_device *pdev)
 	if (!cci->data)
 		return -ENOENT;
 
-	for_each_available_child_of_node(dev->of_node, child) {
+	for_each_available_child_of_node_scoped(dev->of_node, child) {
 		struct cci_master *master;
 		u32 idx;
 
@@ -537,6 +541,9 @@ static int cci_probe(struct platform_device *pdev)
 		master->adap.algo = &cci_algo;
 		master->adap.dev.parent = dev;
 		master->adap.dev.of_node = of_node_get(child);
+		ret = devm_add_action_or_reset(dev, cci_put_of_node, child);
+		if (ret)
+			return ret;
 		master->master = idx;
 		master->cci = cci;
 
@@ -604,10 +611,8 @@ static int cci_probe(struct platform_device *pdev)
 			continue;
 
 		ret = i2c_add_adapter(&cci->master[i].adap);
-		if (ret < 0) {
-			of_node_put(cci->master[i].adap.dev.of_node);
+		if (ret < 0)
 			goto error_i2c;
-		}
 	}
 
 	return 0;
@@ -617,10 +622,8 @@ static int cci_probe(struct platform_device *pdev)
 	pm_runtime_dont_use_autosuspend(dev);
 
 	for (--i ; i >= 0; i--) {
-		if (cci->master[i].cci) {
+		if (cci->master[i].cci)
 			i2c_del_adapter(&cci->master[i].adap);
-			of_node_put(cci->master[i].adap.dev.of_node);
-		}
 	}
 disable_clocks:
 	cci_disable_clocks(cci);
@@ -636,7 +639,6 @@ static void cci_remove(struct platform_device *pdev)
 	for (i = 0; i < cci->data->num_masters; i++) {
 		if (cci->master[i].cci) {
 			i2c_del_adapter(&cci->master[i].adap);
-			of_node_put(cci->master[i].adap.dev.of_node);
 			cci_halt(cci, i);
 		}
 	}
Signature · 1 line
2.55.0
Reviewed-by on v2 1/1
#3 lore raw

Hi Liu.

Quoted 59 lines · Liu Zhenlong

The of_node_put() matching of_node_get() runs after i2c_del_adapter(), whose trailing memset() zeroes adap->dev and thus adap->dev.of_node, making the put a no-op and leaking the node on every adapter removal and error cleanup.

Use a devm action: the pointer is captured at registration, out of reach of that memset(), and devres runs the put once on probe failure and detach, replacing the three manual of_node_put() calls. The setup loop uses the scoped iterator form so the child node is released automatically if devm_add_action_or_reset() fails mid-loop.

Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5

---
Changes in v2:
- Rework the fix to use a devm action (cci_put_of_node) instead of caching the pointer before i2c_del_adapter(), per Konrad Dybcio. The pointer is captured at registration, out of reach of the memset() in i2c_del_adapter(); the three manual of_node_put() calls are removed.
- Use for_each_available_child_of_node_scoped() so the child reference is released if devm_add_action_or_reset() fails mid-loop.

  drivers/i2c/busses/i2c-qcom-cci.c | 20 +++++++++++---------
  1 file changed, 11 insertions(+), 9 deletions(-)
drivers/i2c/busses/i2c-qcom-cci.c +6 -2
index bdeda3979c48..d3528c7d15bd 100644
@@ -497,10 +497,14 @@ static const struct dev_pm_ops qcom_cci_pm = {
  	SET_RUNTIME_PM_OPS(cci_suspend_runtime, cci_resume_runtime, NULL)
  };
 
+static void cci_put_of_node(void *data)
+{
+	of_node_put(data);
+}
+
  static int cci_probe(struct platform_device *pdev)
  {
  	struct device *dev = &pdev->dev;
-	struct device_node *child;
  	struct resource *r;
  	struct cci *cci;
  	int ret, i;
@@ -516,7 +520,7 @@ static int cci_probe(struct platform_device *pdev)
  	if (!cci->data)
  		return -ENOENT;
 
-	for_each_available_child_of_node(dev->of_node, child) {
+	for_each_available_child_of_node_scoped(dev->of_node, child) {
  		struct cci_master *master;
  		u32 idx;
 

drivers/i2c/busses/i2c-qcom-cci.c static int cci_probe(struct platform_device *pdev) · line 544

Quoted 4 lines
drivers/i2c/busses/i2c-qcom-cci.c +1 -0
@@ -537,6 +541,9 @@ static int cci_probe(struct platform_device *pdev)
  		master->adap.algo = &cci_algo;
  		master->adap.dev.parent = dev;
  		master->adap.dev.of_node = of_node_get(child);
+		ret = devm_add_action_or_reset(dev, cci_put_of_node, child);

I believe the new cci_put_of_node() and the original of_node_put() functions are type compatible, therefore a function type cast could be sufficient here:

(void (*)(void *))of_node_put

In any case the change seems to correct, thank you for the fix!

Signature · 2 lines
Best wishes,
Vladimir
#4 lore raw

Hi Vladimir,

Thanks for the review and the Reviewed-by!

On the cast: I considered (void (*)(void *))of_node_put, but kept the small wrapper. Casting to an incompatible function pointer type and calling through it is technically undefined, and tegra_dc_of_node_put() in drivers/gpu/drm/tegra/rgb.c uses the same one-line wrapper around of_node_put(), so I followed that existing pattern. Happy to switch to the cast if you'd still prefer it.

Regards,
Liu Zhenlong

#5 lore raw
Quoted 7 lines · Liu Zhenlong

Hi Vladimir,

Thanks for the review and the Reviewed-by!

On the cast: I considered (void (*)(void *))of_node_put, but kept the small wrapper. Casting to an incompatible function pointer type and calling through it is technically undefined,

Exactly, according to my reading of C99 section 6.7.5.3, paragraph 15, I believe it is a cast to a compatible function, as I've said earlier.

Quoted 5 lines

and tegra_dc_of_node_put()
in drivers/gpu/drm/tegra/rgb.c uses the same one-line wrapper around of_node_put(), so I followed that existing pattern. Happy to switch to the cast if you'd still prefer it.

I would not insist on the suggested change, as for me both versions have incomparable advantages.

Signature · 2 lines
Best wishes,
Vladimr
Reviewed-by
#6 lore raw
Quoted 12 lines · Vladimir Zapolskiy
Quoted 7 lines · Liu Zhenlong

Hi Vladimir,

Thanks for the review and the Reviewed-by!

On the cast: I considered (void (*)(void *))of_node_put, but kept the small wrapper.  Casting to an incompatible function pointer type and calling through it is technically undefined,

Exactly, according to my reading of C99 section 6.7.5.3, paragraph 15, I believe it is a cast to a compatible function, as I've said earlier.

Quoted 8 lines
Quoted 5 lines

and tegra_dc_of_node_put()
in drivers/gpu/drm/tegra/rgb.c uses the same one-line wrapper around of_node_put(), so I followed that existing pattern.  Happy to switch to the cast if you'd still prefer it.

I would not insist on the suggested change, as for me both versions have incomparable advantages.

Just seeing (void (*)(void*))foo scares me

So for the current iteration

Konrad

maintainer Applied
#7 lore raw
In reply to Liu Zhenlong

Hi Liu,

Quoted 11 lines · Liu Zhenlong

The of_node_put() matching of_node_get() runs after i2c_del_adapter(), whose trailing memset() zeroes adap->dev and thus adap->dev.of_node, making the put a no-op and leaking the node on every adapter removal and error cleanup.

Use a devm action: the pointer is captured at registration, out of reach of that memset(), and devres runs the put once on probe failure and detach, replacing the three manual of_node_put() calls. The setup loop uses the scoped iterator form so the child node is released automatically if devm_add_action_or_reset() fails mid-loop.

Quoted 5 lines
Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5

pushed to i2c/i2c-fixes.

Please, next time don't send patches as --in-reply-to.

We should train LLM's not to take much freedom as to send the next versions in the same threads.

Thanks,
Andi

#8 lore raw
Quoted 20 lines · Andi Shyti

Hi Liu,

Quoted 11 lines · Liu Zhenlong

The of_node_put() matching of_node_get() runs after i2c_del_adapter(), whose trailing memset() zeroes adap->dev and thus adap->dev.of_node, making the put a no-op and leaking the node on every adapter removal and error cleanup.

Use a devm action: the pointer is captured at registration, out of reach of that memset(), and devres runs the put once on probe failure and detach, replacing the three manual of_node_put() calls. The setup loop uses the scoped iterator form so the child node is released automatically if devm_add_action_or_reset() fails mid-loop.

Quoted 5 lines
Fixes 02a4a69667a2 ("i2c: qcom-cci: don't put a device tree node before i2c_add_adapter()")
Assisted-by Claude:claude-opus-5
Quoted 6 lines

pushed to i2c/i2c-fixes.

Please, next time don't send patches as --in-reply-to.

We should train LLM's not to take much freedom as to send the next versions in the same threads.

The process is well documented, but the LLMs don't really seem to care about the current chain of:

README.md:
find "Who are you / * AI Coding Assistant"
go to the "AI Coding Assistant" paragraph
read the link to "Documentation/process/coding-assistants.rst"
coding-assistants.rst:
read the links to process docs
read the whole file

Might be because we have no AGENTS.md/CLAUDE.md which the tools look at by default. But it's not guaranteed that it'd be a panaceum either

Konrad