thermal/core: Fix refcount bugs in __thermal_cooling_device_register()
PATCH 1/1 thermal/core: Fix refcount bugs in __thermal_cooling_device_register()
For each new reference of 'device_node', we should increase its refcount. Otherwise, there will be premature free.
For example, in drivers\thermal\tegra\soctherm.c, the function soctherm_init_hw_throt_cdev() will use for_each_child_of_node() to iterate its child device_node which will be then passed into
__thermal_cooling_device_register(). As for_each_xxx OF APIs will
automatically increase and decrease the refcount of 'device_node', we should use additional of_node_get() to record the new refernece.
NOTE, we should also call the corresponding of_node_put() in fail path or when the *_unregister() function is called.
---
I cannot confirm, in *_unregister, we should put of_node_put() in or
out of the *_lock/*_unlock functions. Please check it carefully.
drivers/thermal/thermal_core.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/thermal/thermal_core.c +3 -1
@@ -919,7 +919,7 @@ __thermal_cooling_device_register(struct device_node *np,mutex_init (&cdev->lock);INIT_LIST_HEAD (&cdev->thermal_instances); - cdev->np = np; + cdev->np =of_node_get (np); cdev->ops = ops; cdev->updated = false; cdev->device.class = &thermal_class;
@@ -947,6 +947,7 @@ __thermal_cooling_device_register(struct device_node *np,return cdev; out_kfree_type: +of_node_put (cdev->np);thermal_cooling_device_destroy_sysfs (cdev);kfree (cdev->type);put_device (&cdev->device);
@@ -1111,6 +1112,7 @@ void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)mutex_unlock (&thermal_list_lock); +of_node_put (cdev->np);ida_simple_remove (&thermal_cdev_ida, cdev->id);device_del (&cdev->device);thermal_cooling_device_destroy_sysfs (cdev);
Signature · 1 line
2.25.1
Rafael J. Wysocki
maintainerQuoted 10 lines · Liang He
For each new reference of '
device_node', we should increase its refcount. Otherwise, there will be premature free.For example, in drivers\thermal\tegra\soctherm.c, the function
soctherm_init_hw_throt_cdev()will usefor_each_child_of_node()to iterate its childdevice_nodewhich will be then passed into
__thermal_cooling_device_register(). Asfor_each_xxxOF APIs will
automatically increase and decrease the refcount of 'device_node', we should use additionalof_node_get()to record the new refernece.
reference
Quoted 3 lines
NOTE, we should also call the corresponding
of_node_put()in fail path or when the*_unregister()function is called.
The NOTE in capitals above is somewhat confusing. I would just say "Accordingly, the corresponding of_node_put() needs to be run in the error code path and on cooling device unregistration."
Quoted 6 lines
Fixes a116b5d44f14 ("thermal: core: introduce thermal_of_cooling_device_register")Signed-off-by Liang He <[email protected]>---
I cannot confirm, in*_unregister, we should putof_node_put()in or
out of the *_lock/*_unlock functions. Please check it carefully.
This doesn't matter too much AFAICS.
Please note that the of_node_put() can still "leak" into the critical section through the "unlock" operation, because the latter is not a full memory barrier.
Moreover, dropping the reference means that the object in question won't be used any more by the holder of that reference and it is no reason I can see why it would be necessary to hold the lock while doing that.
Quoted 25 lines
drivers/thermal/thermal_core.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)drivers/thermal/thermal_core.c +2
-1@@ -919,7 +919,7 @@ __thermal_cooling_device_register(struct device_node *np,mutex_init (&cdev->lock);INIT_LIST_HEAD (&cdev->thermal_instances); - cdev->np = np; + cdev->np =of_node_get (np); cdev->ops = ops; cdev->updated = false; cdev->device.class = &thermal_class;@@ -947,6 +947,7 @@ __thermal_cooling_device_register(struct device_node *np,return cdev; out_kfree_type: +of_node_put (cdev->np);thermal_cooling_device_destroy_sysfs (cdev);kfree (cdev->type);put_device (&cdev->device);
drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev) · line 1115
Quoted 4 lines
drivers/thermal/thermal_core.c +1
-0@@ -1111,6 +1112,7 @@ void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)mutex_unlock (&thermal_list_lock); +of_node_put (cdev->np);
Could this be done right before the thermal_cooling_device_destroy_sysfs() below? Then the sequence would be completely analogous to the error code path above.
drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)
Quoted 4 lines
drivers/thermal/thermal_core.c +0
-1ida_simple_remove (&thermal_cdev_ida, cdev->id);device_del (&cdev->device);thermal_cooling_device_destroy_sysfs (cdev); --
Overall, this looks like a genuine fix to me.
Daniel, what do you think?
Quoted 5 lines · Rafael J. Wysocki
Quoted 4 lines · Liang He
For each new reference of '
device_node', we should increase its refcount. Otherwise, there will be premature free.
Quoted 8 lines
Quoted 6 lines
For example, in drivers\thermal\tegra\soctherm.c, the function
soctherm_init_hw_throt_cdev()will usefor_each_child_of_node()to iterate its childdevice_nodewhich will be then passed into
__thermal_cooling_device_register(). Asfor_each_xxxOF APIs will
automatically increase and decrease the refcount of 'device_node', we should use additionalof_node_get()to record the new refernece.reference
Thanks!
Quoted 9 lines
Quoted 4 lines
NOTE, we should also call the corresponding
of_node_put()in fail path or when the*_unregister()function is called.The NOTE in capitals above is somewhat confusing. I would just say "Accordingly, the corresponding
of_node_put()needs to be run in the error code path and on cooling device unregistration."
Thanks, I will change that in new version.
Quoted 44 lines
Quoted 6 lines
Fixes a116b5d44f14 ("thermal: core: introduce thermal_of_cooling_device_register")Signed-off-by Liang He <[email protected]>---
I cannot confirm, in*_unregister, we should putof_node_put()in or
out of the *_lock/*_unlock functions. Please check it carefully.This doesn't matter too much AFAICS.
Please note that the
of_node_put()can still "leak" into the critical section through the "unlock" operation, because the latter is not a full memory barrier.Moreover, dropping the reference means that the object in question won't be used any more by the holder of that reference and it is no reason I can see why it would be necessary to hold the lock while doing that.
Quoted 26 lines
drivers/thermal/thermal_core.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)drivers/thermal/thermal_core.c +2
-1@@ -919,7 +919,7 @@ __thermal_cooling_device_register(struct device_node *np,mutex_init (&cdev->lock);INIT_LIST_HEAD (&cdev->thermal_instances); - cdev->np = np; + cdev->np =of_node_get (np); cdev->ops = ops; cdev->updated = false; cdev->device.class = &thermal_class;@@ -947,6 +947,7 @@ __thermal_cooling_device_register(struct device_node *np,return cdev; out_kfree_type: +of_node_put (cdev->np);thermal_cooling_device_destroy_sysfs (cdev);kfree (cdev->type);put_device (&cdev->device);@@ -1111,6 +1112,7 @@ void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)
Quoted 8 lines
drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev) · line 1115
Quoted 3 lines
drivers/thermal/thermal_core.c +1
-0mutex_unlock (&thermal_list_lock); +of_node_put (cdev->np);Could this be done right before the
thermal_cooling_device_destroy_sysfs()below? Then the sequence would be completely analogous to the error code path above.
That souds good, I will change it in new version.
Quoted 8 lines
drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)
Quoted 4 lines
drivers/thermal/thermal_core.c +0
-1ida_simple_remove (&thermal_cdev_ida, cdev->id);device_del (&cdev->device);thermal_cooling_device_destroy_sysfs (cdev); --Overall, this looks like a genuine fix to me.
Daniel, what do you think?
I will also wait Daniel's reponse before I send new version.
Thanks again,
Liang
Quoted 1 line · Rafael J. Wysocki
On Thu, Jul 7, 2022 at 8:21 AM Liang He <[email protected]> wrote:
[ ... ]
Quoted 23 lines
Quoted 14 lines
drivers/thermal/thermal_core.c +2
-1- cdev->np = np; + cdev->np =of_node_get (np); cdev->ops = ops; cdev->updated = false; cdev->device.class = &thermal_class;@@ -947,6 +947,7 @@ __thermal_cooling_device_register(struct device_node *np,return cdev; out_kfree_type: +of_node_put (cdev->np);thermal_cooling_device_destroy_sysfs (cdev);kfree (cdev->type);put_device (&cdev->device);drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev) · line 1115
Quoted 4 lines
drivers/thermal/thermal_core.c +1
-0@@ -1111,6 +1112,7 @@ void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)mutex_unlock (&thermal_list_lock); +of_node_put (cdev->np);Could this be done right before the
thermal_cooling_device_destroy_sysfs()below? Then the sequence would be completely analogous to the error code path above.
Quoted 8 lines
drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)
Quoted 4 lines
drivers/thermal/thermal_core.c +0
-1ida_simple_remove (&thermal_cdev_ida, cdev->id);device_del (&cdev->device);thermal_cooling_device_destroy_sysfs (cdev); --Overall, this looks like a genuine fix to me.
Daniel, what do you think?
Yes, the of_node_put() is often missing when there is the for_each_xxx OF API. But here the cdev->np is only used to compare pointers so used as an identifier, not de-referenced just comparing the addresses.
It is used to bind a thermal zone with a cooling device:
thermal_of.c: if (tcbp->cooling_device == cdev->np) {
thermal_of.c: if (tcbp->cooling_device == cdev->np) {That is probably why no refcount was taken on this device node.
Moreover, this will go away with the series reworking the thermal-of I sent last week
Signature · 5 lines
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook | <http://twitter.com/#!/linaroorg> Twitter | <http://www.linaro.org/linaro-blog/> Blog
Quoted 33 lines · Daniel Lezcano
Quoted 1 line · Rafael J. Wysocki
On Thu, Jul 7, 2022 at 8:21 AM Liang He <[email protected]> wrote:
[ ... ]
Quoted 28 lines
Quoted 14 lines
drivers/thermal/thermal_core.c +2
-1- cdev->np = np; + cdev->np =of_node_get (np); cdev->ops = ops; cdev->updated = false; cdev->device.class = &thermal_class;@@ -947,6 +947,7 @@ __thermal_cooling_device_register(struct device_node *np,return cdev; out_kfree_type: +of_node_put (cdev->np);thermal_cooling_device_destroy_sysfs (cdev);kfree (cdev->type);put_device (&cdev->device);drivers/thermal/thermal_core.c void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev) · line 1115
Quoted 4 lines
drivers/thermal/thermal_core.c +1
-0@@ -1111,6 +1112,7 @@ void thermal_cooling_device_unregister(struct thermal_cooling_device *cdev)mutex_unlock (&thermal_list_lock); +of_node_put (cdev->np);Could this be done right before the
thermal_cooling_device_destroy_sysfs()below? Then the sequence would be completely analogous to the error code path above.Quoted 4 lines
drivers/thermal/thermal_core.c +0
-1ida_simple_remove (&thermal_cdev_ida, cdev->id);device_del (&cdev->device);thermal_cooling_device_destroy_sysfs (cdev); --
Quoted 9 lines
Quoted 3 lines
Overall, this looks like a genuine fix to me.
Daniel, what do you think?
Yes, the
of_node_put()is often missing when there is thefor_each_xxxOF API. But here thecdev->npis only used to compare pointers so used as an identifier, not de-referenced just comparing the addresses.
Thanks, this is a good lesson that explains when there is no need to refcount new reference.
So I think there is also no need to patch anything, right?
Thanks again,
Liang
Quoted 18 lines
It is used to bind a thermal zone with a cooling device:
thermal_of.c: if (tcbp->cooling_device == cdev->np) { thermal_of.c: if (tcbp->cooling_device == cdev->np) {That is probably why no refcount was taken on this device node.
Moreover, this will go away with the series reworking the thermal-of I sent last week
Quoted 1 line
Signature · 5 lines
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook | <http://twitter.com/#!/linaroorg> Twitter | <http://www.linaro.org/linaro-blog/> Blog
On 17/07/2022 04:57, Liang He wrote:
[ ... ]
Quoted 9 lines
Quoted 5 lines
Yes, the
of_node_put()is often missing when there is thefor_each_xxxOF API. But here thecdev->npis only used to compare pointers so used as an identifier, not de-referenced just comparing the addresses.Thanks, this is a good lesson that explains when there is no need to refcount new reference.
So I think there is also no need to patch anything, right?
Right, no need a change for this.
Thanks anyway for tracking down the refcount in the code
[ ... ]
Signature · 5 lines
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook | <http://twitter.com/#!/linaroorg> Twitter | <http://www.linaro.org/linaro-blog/> Blog