platform/x86: dell-ddv: Fix taking the psy->extensions_sem lock twice
Subsystems (2)
Hans de Goede
maintainerPATCH 1/1 platform/x86: dell-ddv: Fix taking the psy->extensions_sem lock twice
dell_wmi_ddv_get_property() gets called with psy->extensions_sem read-locked, it calls dell_wmi_ddv_battery_translate() which calls power_supply_get_property() on the same psy which again read-locks psy->extensions_sem.
Lockdep rightfully complains about this:
============================================
WARNING: possible recursive locking detected
...
kworker/16:3/1230 is trying to acquire lock:
ffff8c3143417658 (&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
but task is already holding lock:
ffff8c3143417658 (&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
...
Possible unsafe locking scenario:
CPU0
---- lock(&psy->extensions_sem);
lock(&psy->extensions_sem);
*** DEADLOCK ***
...
Call Trace: <TASK> ... down_read+0x3e/0x180 ? power_supply_get_property.part.0+0x23/0x160 power_supply_get_property.part.0+0x23/0x160 dell_wmi_ddv_battery_translate+0x68/0x1d0 [dell_wmi_ddv] ? lock_acquire+0xd9/0x2c0 dell_wmi_ddv_get_property+0x25/0x240 [dell_wmi_ddv] power_supply_get_property.part.0+0x87/0x160 power_supply_format_property+0xc4/0x3d0 add_prop_uevent+0x26/0x90 power_supply_uevent+0xb9/0xf0
This usually works fine, because read-locking can be done multiple times but if someone tries to write-lock between the 2 read-lock calls then the second read-lock will block on the write-lock and the write-lock will be blocked on the first read-lock leading to a deadlock.
The serial is part of the main psy device, not of an extension. Directly call psy->desc->get_property() in dell_wmi_ddv_battery_translate() to fix the double-lock issue.
Note this also influences eppid_show() which is called directly rather then through power_supply_get_property(). This is ok since the ACPI battery is fully ready to be used when the battery hook's add_battery callback is called.
--- drivers/platform/x86/dell/dell-wmi-ddv.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/platform/x86/dell/dell-wmi-ddv.c +5 -3
@@ -689,9 +689,11 @@ static int dell_wmi_ddv_battery_translate(struct dell_wmi_ddv_data *data,dev_dbg (&data->wdev->dev,"Translation cache miss\n" ); -/* Perform a translation between a ACPI battery and a battery index */ - - ret =power_supply_get_property (battery,POWER_SUPPLY_PROP_SERIAL_NUMBER , &val); +/* +* Perform a translation between a ACPI battery and a battery index. Directly call +* desc->get_property() to avoid locking battery->extensions_sem a second time. +*/ + ret = battery->desc->get_property (battery,POWER_SUPPLY_PROP_SERIAL_NUMBER , &val);if (ret <0 )return ret;
Signature · 1 line
2.49.0
Armin Wolf
maintainerAm 20.06.25 um 19:58 schrieb Hans de Goede:
Quoted 45 lines
dell_wmi_ddv_get_property()gets called withpsy->extensions_semread-locked, it callsdell_wmi_ddv_battery_translate()which callspower_supply_get_property()on the same psy which again read-lockspsy->extensions_sem.Lockdep rightfully complains about this:
============================================
WARNING: possible recursive locking detected
...
kworker/16:3/1230 is trying to acquire lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
but task is already holding lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
...Possible unsafe locking scenario:CPU0 ---- lock(&psy->extensions_sem); lock(&psy->extensions_sem);*** DEADLOCK *** ... Call Trace: <TASK> ... down_read+0x3e/0x180 ? power_supply_get_property.part.0+0x23/0x160 power_supply_get_property.part.0+0x23/0x160 dell_wmi_ddv_battery_translate+0x68/0x1d0 [dell_wmi_ddv] ? lock_acquire+0xd9/0x2c0 dell_wmi_ddv_get_property+0x25/0x240 [dell_wmi_ddv] power_supply_get_property.part.0+0x87/0x160 power_supply_format_property+0xc4/0x3d0 add_prop_uevent+0x26/0x90 power_supply_uevent+0xb9/0xf0This usually works fine, because read-locking can be done multiple times but if someone tries to write-lock between the 2 read-lock calls then the second read-lock will block on the write-lock and the write-lock will be blocked on the first read-lock leading to a deadlock.
Quoted 8 lines
The serial is part of the main psy device, not of an extension. Directly call
psy->desc->get_property()indell_wmi_ddv_battery_translate()to fix the double-lock issue.Note this also influences
eppid_show()which is called directly rather then throughpower_supply_get_property(). This is ok since the ACPI battery is fully ready to be used when the battery hook'sadd_batterycallback is called.
Thank you very much for finding this issue, but i think that simply calling battery->desc->get_property() is not the right solution for this:
1. We should still call psy_desc_has_property() to determine if the power supply actually support POWER_SUPPLY_PROP_SERIAL_NUMBER.
2. At least another power supply extension user (the uniwill-laptop driver currently being under review) suffers from a similar problem, so a more generic solution is needed.
Maybe we could introduce a new function for reading power supply properties that ignores any power supply extensions? This way future extension could use this function too.
I envision something like this:
int power_supply_get_property_direct(struct power_supply *psy,
enum power_supply_property psp, union power_supply_propval *val)
{
if (atomic_read(&psy->use_cnt) <= 0) {
if (!psy->initialized)
return -EAGAIN;
return -ENODEV;
}
if (psy_desc_has_property(psy->desc, psp))
return psy->desc->get_property(psy, psp, val);
else if (power_supply_battery_info_has_prop(psy->battery_info, psp))
return power_supply_battery_info_get_prop(psy->battery_info, psp, val);
else
return -EINVAL;
}EXPORT_SYMBOL_GPL(power_supply_get_property_direct);
It basically is power_supply_get_property() without the extension logic. I can also write some documentation on how to implement power supply extensions in general.
What do you thing?
Thanks,
Armin Wolf
Quoted 25 lines
Fixes 058de163a376 ("platform/x86: dell-ddv: Implement the battery matching algorithm")Signed-off-by Hans de Goede <[email protected]>--- drivers/platform/x86/dell/dell-wmi-ddv.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/platform/x86/dell/dell-wmi-ddv.c +5
-3@@ -689,9 +689,11 @@ static int dell_wmi_ddv_battery_translate(struct dell_wmi_ddv_data *data,dev_dbg (&data->wdev->dev,"Translation cache miss\n" ); -/* Perform a translation between a ACPI battery and a battery index */ - - ret =power_supply_get_property (battery,POWER_SUPPLY_PROP_SERIAL_NUMBER , &val); +/* +* Perform a translation between a ACPI battery and a battery index. Directly call +* desc->get_property() to avoid locking battery->extensions_sem a second time. +*/ + ret = battery->desc->get_property (battery,POWER_SUPPLY_PROP_SERIAL_NUMBER , &val);if (ret <0 )return ret;
Hans de Goede
maintainer+ Cc sre, whom I should have Cc-ed from the beginning.
Hi Armin,
Quoted 56 lines · Armin Wolf
Am 20.06.25 um 19:58 schrieb Hans de Goede:
Quoted 45 lines
dell_wmi_ddv_get_property()gets called withpsy->extensions_semread-locked, it callsdell_wmi_ddv_battery_translate()which callspower_supply_get_property()on the same psy which again read-lockspsy->extensions_sem.Lockdep rightfully complains about this:
============================================
WARNING: possible recursive locking detected
...
kworker/16:3/1230 is trying to acquire lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160 but task is already holding lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
...Possible unsafe locking scenario:CPU0
----
lock(&psy->extensions_sem);
lock(&psy->extensions_sem);*** DEADLOCK ***
...Call Trace: <TASK> ... down_read+0x3e/0x180 ? power_supply_get_property.part.0+0x23/0x160 power_supply_get_property.part.0+0x23/0x160 dell_wmi_ddv_battery_translate+0x68/0x1d0 [dell_wmi_ddv] ? lock_acquire+0xd9/0x2c0 dell_wmi_ddv_get_property+0x25/0x240 [dell_wmi_ddv] power_supply_get_property.part.0+0x87/0x160 power_supply_format_property+0xc4/0x3d0 add_prop_uevent+0x26/0x90 power_supply_uevent+0xb9/0xf0This usually works fine, because read-locking can be done multiple times but if someone tries to write-lock between the 2 read-lock calls then the second read-lock will block on the write-lock and the write-lock will be blocked on the first read-lock leading to a deadlock.
Quoted 8 lines
The serial is part of the main psy device, not of an extension. Directly call
psy->desc->get_property()indell_wmi_ddv_battery_translate()to fix the double-lock issue.Note this also influences
eppid_show()which is called directly rather then throughpower_supply_get_property(). This is ok since the ACPI battery is fully ready to be used when the battery hook'sadd_batterycallback is called.
Quoted 5 lines
Thank you very much for finding this issue, but i think that simply calling
battery->desc->get_property()is not the right solution for this:1. We should still call
psy_desc_has_property()to determine if the power supply actually supportPOWER_SUPPLY_PROP_SERIAL_NUMBER.
Although it is currently not enforced in power_supply_core.c it seems reasonable to assume that any powersupply must have a get_property callback in their desc (the core also unconditionally calls this).
And most (all?) psy drivers I've seen have a default which returns -EINVAL in their get_property() implementation which works just as well as calling has_prop ...
My bigger worry is the lack of the:
Quoted 5 lines
if (atomic_read(&psy->use_cnt) <= 0) {
if (!psy->initialized)
return -EAGAIN;
return -ENODEV;
}Check TBH.
Quoted 27 lines
2. At least another power supply extension user (the uniwill-laptop driver currently being under review) suffers from a similar problem, so a more generic solution is needed.
Maybe we could introduce a new function for reading power supply properties that ignores any power supply extensions? This way future extension could use this function too.
I envision something like this:
intpower_supply_get_property_direct(struct power_supply *psy, enum power_supply_property psp, union power_supply_propval *val)
{
if (atomic_read(&psy->use_cnt)<= 0) {
if (!psy->initialized)
return-EAGAIN;
return-ENODEV;
}if (
psy_desc_has_property(psy->desc, psp))
returnpsy->desc->get_property(psy, psp, val);
else if (power_supply_battery_info_has_prop(psy->battery_info, psp)) returnpower_supply_battery_info_get_prop(psy->battery_info, psp, val); else
return-EINVAL;
}EXPORT_SYMBOL_GPL(power_supply_get_property_direct);
Quoted 1 line
It basically is
power_supply_get_property()without the extension logic.
While working on this fix I was thinking that something like this would be useful, so +1 for this.
Maybe first do a prep patch where the extension handling in power_supply_get_property() is moved last, then power_supply_get_property() can just wrap this new helprr and on -EINVAL check the extensions.
Actually if you move the extensions check to last then the whole doublelock issue goes away because the serial-number will be found before checking extensions.
Or if you want to keep checking the extensions first change the current power_supply_get_property() into a new __power_supply_get_property() with a "bool check_extensions" argument and make power_supply_get_property() wrap it pasing true for check_extensions. Or some such, whatever you do try to avoid code duplication but you already know this ...
Quoted 2 lines
I can also write some
documentation on how to implement power supply extensions in general.
That would also be good to have.
Regards,
Hans
Quoted 23 lines
Quoted 23 lines
Fixes 058de163a376 ("platform/x86: dell-ddv: Implement the battery matching algorithm")Signed-off-by Hans de Goede <[email protected]>---
drivers/platform/x86/dell/dell-wmi-ddv.c| 8 +++++---1 file changed, 5 insertions(+), 3 deletions(-)drivers/platform/x86/dell/dell-wmi-ddv.c +0
-0@@ -689,9 +689,11 @@ static int dell_wmi_ddv_battery_translate(struct dell_wmi_ddv_data *data, dev_dbg(&data->wdev->dev, "Translation cache miss\n");- /* Perform a translation between a ACPI battery and a battery index */ -
- ret =power_supply_get_property(battery, POWER_SUPPLY_PROP_SERIAL_NUMBER, &val); + /*
+ * Perform a translation between a ACPI battery and a battery index. Directly call + *desc->get_property()to avoid lockingbattery->extensions_sema second time. + */
+ ret =battery->desc->get_property(battery, POWER_SUPPLY_PROP_SERIAL_NUMBER, &val); if (ret < 0)
return ret;
Armin Wolf
maintainerAm 22.06.25 um 22:59 schrieb Hans de Goede:
Quoted 127 lines
+ Cc sre, whom I should have Cc-ed from the beginning.
Hi Armin,
Quoted 55 lines · Armin Wolf
Am 20.06.25 um 19:58 schrieb Hans de Goede:
Quoted 45 lines
dell_wmi_ddv_get_property()gets called withpsy->extensions_semread-locked, it callsdell_wmi_ddv_battery_translate()which callspower_supply_get_property()on the same psy which again read-lockspsy->extensions_sem.Lockdep rightfully complains about this:
============================================
WARNING: possible recursive locking detected
...
kworker/16:3/1230 is trying to acquire lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160 but task is already holding lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
...Possible unsafe locking scenario:CPU0
----
lock(&psy->extensions_sem);
lock(&psy->extensions_sem);*** DEADLOCK ***
...Call Trace: <TASK> ... down_read+0x3e/0x180 ? power_supply_get_property.part.0+0x23/0x160 power_supply_get_property.part.0+0x23/0x160 dell_wmi_ddv_battery_translate+0x68/0x1d0 [dell_wmi_ddv] ? lock_acquire+0xd9/0x2c0 dell_wmi_ddv_get_property+0x25/0x240 [dell_wmi_ddv] power_supply_get_property.part.0+0x87/0x160 power_supply_format_property+0xc4/0x3d0 add_prop_uevent+0x26/0x90 power_supply_uevent+0xb9/0xf0This usually works fine, because read-locking can be done multiple times but if someone tries to write-lock between the 2 read-lock calls then the second read-lock will block on the write-lock and the write-lock will be blocked on the first read-lock leading to a deadlock.
Quoted 8 lines
The serial is part of the main psy device, not of an extension. Directly call
psy->desc->get_property()indell_wmi_ddv_battery_translate()to fix the double-lock issue.Note this also influences
eppid_show()which is called directly rather then throughpower_supply_get_property(). This is ok since the ACPI battery is fully ready to be used when the battery hook'sadd_batterycallback is called.Quoted 5 lines
Thank you very much for finding this issue, but i think that simply calling
battery->desc->get_property()is not the right solution for this:1. We should still call
psy_desc_has_property()to determine if the power supply actually supportPOWER_SUPPLY_PROP_SERIAL_NUMBER.Although it is currently not enforced in
power_supply_core.cit seems reasonable to assume that any powersupply must have aget_propertycallback in their desc (the core also unconditionally calls this).And most (all?) psy drivers I've seen have a default which returns
-EINVALin theirget_property()implementation which works just as well as callinghas_prop...My bigger worry is the lack of the:
Quoted 5 lines
if (atomic_read(&psy->use_cnt) <= 0) { if (!psy->initialized) return -EAGAIN; return -ENODEV; }Check TBH.
Quoted 27 lines
2. At least another power supply extension user (the uniwill-laptop driver currently being under review) suffers from a similar problem, so a more generic solution is needed.
Maybe we could introduce a new function for reading power supply properties that ignores any power supply extensions? This way future extension could use this function too.
I envision something like this:
int power_supply_get_property_direct(struct
power_supply*psy,
enumpower_supply_propertypsp, unionpower_supply_propval*val)
{
if (atomic_read(&psy->use_cnt)<= 0) {
if (!psy->initialized)
return-EAGAIN;
return-ENODEV;
}if (
psy_desc_has_property(psy->desc, psp))
returnpsy->desc->get_property(psy, psp, val);
else if (power_supply_battery_info_has_prop(psy->battery_info, psp)) returnpower_supply_battery_info_get_prop(psy->battery_info, psp, val); else
return-EINVAL;
}EXPORT_SYMBOL_GPL(power_supply_get_property_direct);Quoted 1 line
It basically is
power_supply_get_property()without the extension logic.While working on this fix I was thinking that something like this would be useful, so +1 for this.
Maybe first do a prep patch where the extension handling in
power_supply_get_property()is moved last, thenpower_supply_get_property()can just wrap this new helprr and on-EINVALcheck the extensions.Actually if you move the extensions check to last then the whole doublelock issue goes away because the serial-number will be found before checking extensions.
Or if you want to keep checking the extensions first change the current
power_supply_get_property()into a new__power_supply_get_property()with a "boolcheck_extensions" argument and makepower_supply_get_property()wrap it pasing true forcheck_extensions. Or some such, whatever you do try to avoid code duplication but you already know this ...
Quoted 7 lines
Quoted 2 lines
I can also write some
documentation on how to implement power supply extensions in general.That would also be good to have.
Regards,
Hans
Alright, i will send the necessary patches soon.
Thanks,
Armin Wolf
Quoted 24 lines
Quoted 23 lines
Quoted 23 lines
Fixes 058de163a376 ("platform/x86: dell-ddv: Implement the battery matching algorithm")Signed-off-by Hans de Goede <[email protected]>--- drivers/platform/x86/dell/dell-wmi-ddv.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/platform/x86/dell/dell-wmi-ddv.c +5
-2@@ -689,9 +689,11 @@ static int dell_wmi_ddv_battery_translate(struct dell_wmi_ddv_data *data, dev_dbg(&data->wdev->dev, "Translation cache miss\n"); - /* Perform a translation between a ACPI battery and a battery index */ - - ret = power_supply_get_property(battery, POWER_SUPPLY_PROP_SERIAL_NUMBER, &val); + /* + * Perform a translation between a ACPI battery and a battery index. Directly call + * desc->get_property() to avoid locking battery->extensions_sem a second time. + */ + ret = battery->desc->get_property(battery, POWER_SUPPLY_PROP_SERIAL_NUMBER, &val); if (ret < 0) return ret;
Hi,
Quoted 130 lines · Armin Wolf
Am 22.06.25 um 22:59 schrieb Hans de Goede:
Quoted 128 lines
+ Cc sre, whom I should have Cc-ed from the beginning.
Hi Armin,
Quoted 55 lines · Armin Wolf
Am 20.06.25 um 19:58 schrieb Hans de Goede:
Quoted 45 lines
dell_wmi_ddv_get_property()gets called withpsy->extensions_semread-locked, it callsdell_wmi_ddv_battery_translate()which callspower_supply_get_property()on the same psy which again read-lockspsy->extensions_sem.Lockdep rightfully complains about this:
============================================
WARNING: possible recursive locking detected
...
kworker/16:3/1230 is trying to acquire lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160 but task is already holding lock:
ffff8c3143417658(&psy->extensions_sem){++++}-{4:4},
at: power_supply_get_property.part.0+0x23/0x160
...Possible unsafe locking scenario:CPU0
----
lock(&psy->extensions_sem);
lock(&psy->extensions_sem);*** DEADLOCK ***
...Call Trace: <TASK> ... down_read+0x3e/0x180 ? power_supply_get_property.part.0+0x23/0x160 power_supply_get_property.part.0+0x23/0x160 dell_wmi_ddv_battery_translate+0x68/0x1d0 [dell_wmi_ddv] ? lock_acquire+0xd9/0x2c0 dell_wmi_ddv_get_property+0x25/0x240 [dell_wmi_ddv] power_supply_get_property.part.0+0x87/0x160 power_supply_format_property+0xc4/0x3d0 add_prop_uevent+0x26/0x90 power_supply_uevent+0xb9/0xf0This usually works fine, because read-locking can be done multiple times but if someone tries to write-lock between the 2 read-lock calls then the second read-lock will block on the write-lock and the write-lock will be blocked on the first read-lock leading to a deadlock.
Quoted 8 lines
The serial is part of the main psy device, not of an extension. Directly call
psy->desc->get_property()indell_wmi_ddv_battery_translate()to fix the double-lock issue.Note this also influences
eppid_show()which is called directly rather then throughpower_supply_get_property(). This is ok since the ACPI battery is fully ready to be used when the battery hook'sadd_batterycallback is called.Quoted 5 lines
Thank you very much for finding this issue, but i think that simply calling
battery->desc->get_property()is not the right solution for this:1. We should still call
psy_desc_has_property()to determine if the power supply actually supportPOWER_SUPPLY_PROP_SERIAL_NUMBER.Although it is currently not enforced in
power_supply_core.cit seems reasonable to assume that any powersupply must have aget_propertycallback in their desc (the core also unconditionally calls this).And most (all?) psy drivers I've seen have a default which returns
-EINVALin theirget_property()implementation which works just as well as callinghas_prop...My bigger worry is the lack of the:
Quoted 5 lines
if (atomic_read(&psy->use_cnt) <= 0) { if (!psy->initialized) return -EAGAIN; return -ENODEV; }Check TBH.
Quoted 26 lines
2. At least another power supply extension user (the uniwill-laptop driver currently being under review) suffers from a similar problem, so a more generic solution is needed.
Maybe we could introduce a new function for reading power supply properties that ignores any power supply extensions? This way future extension could use this function too.
I envision something like this:
int power_supply_get_property_direct(structpower_supply*psy,
enumpower_supply_propertypsp, unionpower_supply_propval*val)
{
if (atomic_read(&psy->use_cnt)<= 0) {
if (!psy->initialized)
return-EAGAIN;
return-ENODEV;
}if (
psy_desc_has_property(psy->desc, psp))
returnpsy->desc->get_property(psy, psp, val);
else if (power_supply_battery_info_has_prop(psy->battery_info, psp)) returnpower_supply_battery_info_get_prop(psy->battery_info, psp, val); else
return-EINVAL;
}EXPORT_SYMBOL_GPL(power_supply_get_property_direct);Quoted 1 line
It basically is
power_supply_get_property()without the extension logic.While working on this fix I was thinking that something like this would be useful, so +1 for this.
Maybe first do a prep patch where the extension handling in
power_supply_get_property()is moved last, thenpower_supply_get_property()can just wrap this new helprr and on-EINVALcheck the extensions.Actually if you move the extensions check to last then the whole doublelock issue goes away because the serial-number will be found before checking extensions.
Or if you want to keep checking the extensions first change the current
power_supply_get_property()into a new__power_supply_get_property()with a "boolcheck_extensions" argument and makepower_supply_get_property()wrap it pasing true forcheck_extensions. Or some such, whatever you do try to avoid code duplication but you already know this ...Quoted 2 lines
I can also write some
documentation on how to implement power supply extensions in general.
Quoted 7 lines
Quoted 5 lines
That would also be good to have.
Regards,
Hans
Alright, i will send the necessary patches soon.
Sounds good to me. Please also Cc Thomas Weißschuh, who implemented all the extension code :)
Greetings,
-- Sebastian