auxdisplay: add devm_linedisp_register() and fix work initialization races
cover auxdisplay: add devm_linedisp_register() and fix work initialization races
From: Surendra Singh Chouhan <[email protected]>
Per maintainer review feedback from Andy Shevchenko and Chris Packham, this patch series introduces a devm-managed devm_linedisp_register() helper for the line-display core to avoid use-after-free teardown races across auxdisplay drivers.
Patch 1 adds devm_linedisp_register() to line-display core.
Patch 2 fixes delayed work initialization races in seg-led-gpio.c, moves INIT_DELAYED_WORK() to probe(), uses devm_add_action_or_reset() for work cancellation, and converts to devm_linedisp_register(), removing seg_led_remove().
Patch 3 fixes delayed work initialization races in max6959.c, moves INIT_DELAYED_WORK() to probe(), uses devm_add_action_or_reset() for work cancellation, and converts to devm_linedisp_register(), removing max6959_i2c_remove().
Surendra Singh Chouhan (3):
auxdisplay: line-display: add devm_linedisp_register()
auxdisplay: seg-led-gpio: fix work initialization race and convert to
devm_linedisp_register()
auxdisplay: max6959: fix work initialization race and convert to
devm_linedisp_register()
drivers/auxdisplay/line-display.c | 32 +++++++++++++++++++++++++++++++ drivers/auxdisplay/line-display.h | 2 ++ drivers/auxdisplay/max6959.c | 26 ++++++++++++------------- drivers/auxdisplay/seg-led-gpio.c | 27 +++++++++++++------------- 4 files changed, 61 insertions(+), 26 deletions(-)
Signature · 1 line
2.55.0
PATCH 1/3 auxdisplay: line-display: add devm_linedisp_register()
From: Surendra Singh Chouhan <[email protected]>
Add devm_linedisp_register() to manage character line display registration via devres. This simplifies driver cleanup and prevents use-after-free bugs when unregistering line displays on driver detach.
--- drivers/auxdisplay/line-display.c | 32 +++++++++++++++++++++++++++++++ drivers/auxdisplay/line-display.h | 2 ++ 2 files changed, 34 insertions(+)
drivers/auxdisplay/line-display.c +32 -0
@@ -592,5 +592,37 @@ void linedisp_unregister(struct linedisp *linedisp) }EXPORT_SYMBOL_NS_GPL (linedisp_unregister,"LINEDISP" ); +static void devm_linedisp_unregister (void *data) +{ +struct linedisp *linedisp = data; + +linedisp_unregister (linedisp); +} + +/** +* devm_linedisp_register - register a character line display +* @dev: device being registered +* @linedisp: pointer to character line display structure +* @num_chars: the number of characters that can be displayed +* @ops: character line display operations +* +* Managed linedisp_register(). Line display registered with this function will +* automatically be unregistered on driver detach. +* +* Return: zero on success, else a negative error code. +*/ +int devm_linedisp_register (struct device *dev,struct linedisp *linedisp, +unsigned int num_chars,const struct linedisp_ops *ops) +{ +int err; + + err =linedisp_register (linedisp, dev, num_chars, ops); +if (err) +return err; + +return devm_add_action_or_reset (dev, devm_linedisp_unregister, linedisp); +} +EXPORT_SYMBOL_NS_GPL (devm_linedisp_register,"LINEDISP" ); +MODULE_DESCRIPTION ("Character line display core support" );MODULE_LICENSE ("GPL" );
drivers/auxdisplay/line-display.h +2 -0
@@ -88,5 +88,7 @@ void linedisp_detach(struct device *dev);int linedisp_register (struct linedisp *linedisp,struct device *parent,unsigned int num_chars,const struct linedisp_ops *ops);void linedisp_unregister (struct linedisp *linedisp); +int devm_linedisp_register (struct device *dev,struct linedisp *linedisp, +unsigned int num_chars,const struct linedisp_ops *ops);#endif /* LINEDISP_H */
Signature · 1 line
2.55.0
Re: [PATCH 0/3 1/3] auxdisplay: line-display: add devm_linedisp_register()
Quoted 3 lines · On Wed, Aug 19, 2026 at 08:15:54AM +0530, [email protected] wrote:
Add
devm_linedisp_register()to manage character line display registration via devres. This simplifies driver cleanup and prevents use-after-free bugs when unregistering line displays on driver detach.
This is not enough per se. Copied'n'pasted the reply I made to Rajat who reported the issue:
-->8----8<--
Quoted 20 lines
Quoted 20 lines
6. PROPOSED FIX
Replace
devm_kzallocwith plain kzalloc and tie the container lifetime to the embedded device's refcount:drivers/auxdisplay/line-display.c +0
-0@@ (
linedisp_releasecallback)static void
linedisp_release(struct device *dev){ struct linedisp *linedisp = to_linedisp(dev);drivers/auxdisplay/line-display.c +2
-0+ struct container *priv = container_of(linedisp, ...);kfree (linedisp->map);kfree (linedisp->message);kfree (linedisp->buf); +kfree (priv);/* free container when refcount reaches 0 */ }
Quoted 9 lines
Quoted 4 lines
Each affected driver (img-ascii-lcd, max6959, seg-led-gpio) must change
devm_kzallocto kzalloc for the container struct, and ensure the release function frees it. This aligns the container lifetime with the embedded device refcount.That won't scale as the device drivers are free to call
devm_kzalloc()and similar for their private data structures. What we should do is to prevent a device from unbinding when one or more files are open (via sysfs). TL;DR: downgradingdevm_kzalloc()is not an option.
So, the fix as I see it is much more intrusive. The linedisp_register() should be split to _alloc() and _register() APIs, and then at least the first one being also wrapped with devm_*() for users that want this. See how devm_iio_device_alloc() and devm_iio_device_registers() are implemented.
Signature · 2 lines
With Best Regards, Andy Shevchenko
PATCH 2/3 auxdisplay: seg-led-gpio: fix work initialization race and convert to devm_linedisp_register()
In reply to [email protected]From: Surendra Singh Chouhan <[email protected]>
INIT_DELAYED_WORK(&priv->work, seg_led_update) was previously called inside seg_led_linedisp_get_map_type(), which is invoked during/after linedisp_register(). Initializing a delayed_work structure inside a map query callback can re-initialize an active work item or race with seg_led_linedisp_update().
In addition, seg_led_remove() called cancel_delayed_work_sync(&priv->work) before linedisp_unregister(&priv->linedisp), allowing sysfs updates to reschedule work after cancel_delayed_work_sync() completed.
Fix these by moving INIT_DELAYED_WORK() to probe(), using devm_add_action_or_reset() for work cancellation, and converting to devm_linedisp_register(). Registering devm_linedisp_register() after work cancellation action ensures proper LIFO teardown order (sysfs interface unregistered first, followed by work cancellation), allowing seg_led_remove() to be removed entirely.
--- drivers/auxdisplay/seg-led-gpio.c | 27 ++++++++++++++------------- 1 file changed, 14 insertions(+), 13 deletions(-)
drivers/auxdisplay/seg-led-gpio.c +14 -13
@@ -38,11 +38,15 @@ static void seg_led_update(struct work_struct *work)gpiod_multi_set_value_cansleep (priv->segment_gpios, values); } -static int seg_led_linedisp_get_map_type (struct linedisp *linedisp) +static void seg_led_cancel_work (void *data) { - struct seg_led_priv *priv = container_of(linedisp, struct seg_led_priv, linedisp); +struct delayed_work *work = data; -INIT_DELAYED_WORK (&priv->work, seg_led_update); +cancel_delayed_work_sync (work); +} + +static int seg_led_linedisp_get_map_type (struct linedisp *linedisp) +{return LINEDISP_MAP_SEG7 ; }
@@ -62,11 +66,17 @@ static int seg_led_probe(struct platform_device *pdev) {struct seg_led_priv *priv;struct device *dev = &pdev->dev; +int ret; priv =devm_kzalloc (dev,sizeof (*priv),GFP_KERNEL );if (!priv)return -ENOMEM ; +INIT_DELAYED_WORK (&priv->work, seg_led_update); + ret =devm_add_action_or_reset (dev, seg_led_cancel_work, &priv->work); +if (ret) +return ret; +platform_set_drvdata (pdev, priv); priv->segment_gpios =devm_gpiod_get_array (dev,"segment" ,GPIOD_OUT_LOW );
@@ -76,15 +86,7 @@ static int seg_led_probe(struct platform_device *pdev)if (priv->segment_gpios->ndescs <7 || priv->segment_gpios->ndescs >8 )return -EINVAL ; -return linedisp_register (&priv->linedisp, dev,1 , &seg_led_linedisp_ops); -} - -static void seg_led_remove (struct platform_device *pdev) -{ -struct seg_led_priv *priv =platform_get_drvdata (pdev); - -cancel_delayed_work_sync (&priv->work); -linedisp_unregister (&priv->linedisp); +return devm_linedisp_register (dev, &priv->linedisp,1 , &seg_led_linedisp_ops); }static const struct of_device_id seg_led_of_match[] = {
@@ -95,7 +97,6 @@ MODULE_DEVICE_TABLE(of, seg_led_of_match);static struct platform_driver seg_led_driver = { .probe = seg_led_probe, - .remove = seg_led_remove, .driver = { .name ="seg-led-gpio" , .of_match_table = seg_led_of_match,
Signature · 1 line
2.55.0
PATCH 3/3 auxdisplay: max6959: fix work initialization race and convert to devm_linedisp_register()
In reply to [email protected]From: Surendra Singh Chouhan <[email protected]>
INIT_DELAYED_WORK(&priv->work, max6959_disp_update) was previously called inside max6959_linedisp_get_map_type(), which is invoked during/after linedisp_register(). Initializing a delayed_work structure inside a map query callback can re-initialize an active work item or race with max6959_linedisp_update().
In addition, max6959_i2c_remove() called cancel_delayed_work_sync() before linedisp_unregister(&priv->linedisp), allowing sysfs updates to reschedule work after cancel_delayed_work_sync() completed.
Fix these by moving INIT_DELAYED_WORK() to probe(), using devm_add_action_or_reset() for work cancellation, and converting to devm_linedisp_register(). Registering devm_linedisp_register() after work cancellation action ensures proper LIFO teardown order, allowing max6959_i2c_remove() to be removed entirely.
--- drivers/auxdisplay/max6959.c | 26 +++++++++++++------------- 1 file changed, 13 insertions(+), 13 deletions(-)
drivers/auxdisplay/max6959.c +13 -13
@@ -63,11 +63,15 @@ static void max6959_disp_update(struct work_struct *work)regmap_bulk_write (priv->regmap,REG_DIGIT (0 ), buf,ARRAY_SIZE (buf)); } -static int max6959_linedisp_get_map_type (struct linedisp *linedisp) +static void max6959_cancel_work (void *data) { - struct max6959_priv *priv = container_of(linedisp, struct max6959_priv, linedisp); +struct delayed_work *work = data; -INIT_DELAYED_WORK (&priv->work, max6959_disp_update); +cancel_delayed_work_sync (work); +} + +static int max6959_linedisp_get_map_type (struct linedisp *linedisp) +{return LINEDISP_MAP_SEG7 ; }
@@ -123,6 +127,11 @@ static int max6959_i2c_probe(struct i2c_client *client)if (!priv)return -ENOMEM ; +INIT_DELAYED_WORK (&priv->work, max6959_disp_update); + ret =devm_add_action_or_reset (dev, max6959_cancel_work, &priv->work); +if (ret) +return ret; + priv->regmap =devm_regmap_init_i2c (client, &max6959_regmap_config);if (IS_ERR (priv->regmap))return PTR_ERR (priv->regmap);
@@ -131,7 +140,7 @@ static int max6959_i2c_probe(struct i2c_client *client)if (ret)return ret; - ret =linedisp_register (&priv->linedisp, dev,4 , &max6959_linedisp_ops); + ret =devm_linedisp_register (dev, &priv->linedisp,4 , &max6959_linedisp_ops);if (ret)return ret;
@@ -140,14 +149,6 @@ static int max6959_i2c_probe(struct i2c_client *client)return 0 ; } -static void max6959_i2c_remove (struct i2c_client *client) -{ -struct max6959_priv *priv =i2c_get_clientdata (client); - -cancel_delayed_work_sync (&priv->work); -linedisp_unregister (&priv->linedisp); -} -static int max6959_suspend (struct device *dev) {return max6959_enable (dev_get_drvdata (dev), false);
@@ -179,7 +180,6 @@ static struct i2c_driver max6959_i2c_driver = { .of_match_table = max6959_of_table, }, .probe = max6959_i2c_probe, - .remove = max6959_i2c_remove, .id_table = max6959_i2c_id, };module_i2c_driver (max6959_i2c_driver);
Signature · 1 line
2.55.0
Re: [PATCH 0/3 3/3] auxdisplay: max6959: fix work initialization race and convert to devm_linedisp_register()
Quoted 10 lines · On Wed, Aug 19, 2026 at 08:15:56AM +0530, [email protected] wrote:
INIT_DELAYED_WORK(&priv->work, max6959_disp_update)was previously called insidemax6959_linedisp_get_map_type(), which is invoked during/afterlinedisp_register(). Initializing adelayed_workstructure inside a map query callback can re-initialize an active work item or race withmax6959_linedisp_update().In addition,
max6959_i2c_remove()calledcancel_delayed_work_sync()beforelinedisp_unregister(&priv->linedisp), allowing sysfs updates to reschedule work aftercancel_delayed_work_sync()completed.
Quoted 5 lines
Fix these by moving
INIT_DELAYED_WORK()toprobe(), usingdevm_add_action_or_reset()for work cancellation, and converting todevm_linedisp_register(). Registeringdevm_linedisp_register()after work cancellation action ensures proper LIFO teardown order, allowingmax6959_i2c_remove()to be removed entirely.
Use devm_delayed_work_autocancel() from devm-helpers.h.
Also try to squeeze this most likely AI-assisted commit message to the point.
...
Same comment to the other fix-patch.
Signature · 2 lines
With Best Regards, Andy Shevchenko