| Message ID | 20231214173614.2820929-1-gnstark@salutedevices.com |
|---|---|
| Headers | show |
| Series | devm_led_classdev_register() usage problem | expand |
On Thu, 14 Dec 2023, George Stark wrote: > This patch series fixes the problem of devm_led_classdev_register misusing. > > The basic problem is described in [1]. Shortly when devm_led_classdev_register() > is used then led_classdev_unregister() called after driver's remove() callback. > led_classdev_unregister() calls driver's brightness_set callback and that callback > may use resources which were destroyed already in driver's remove(). > > After discussion with maintainers [2] [3] we decided: > 1) don't touch led subsytem core code and don't remove led_set_brightness() from it > but fix drivers > 2) don't use devm_led_classdev_unregister > > So the solution is to use devm wrappers for all resources > driver's brightness_set() depends on. And introduce dedicated devm wrapper > for mutex as it's often used resource. > > [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ > [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f > [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 > > Changelog: > v1->v2: > revise patch series completely > > v2->v3: > locking: add define if mutex_destroy() is not an empty function > new patch, discussed here [8] > > devm-helpers: introduce devm_mutex_init > previous version [4] > - revise code based on mutex_destroy define > - update commit message > - update devm_mutex_init()'s description > > leds: aw2013: unlock mutex before destroying it > previous version [5] > - make this patch first in the series > - add tags Fixes and RvB by Andy > > leds: aw2013: use devm API to cleanup module's resources > previous version [6] > - make aw2013_chip_disable_action()'s body oneline > - don't shadow devm_mutex_init() return code > > leds: aw200xx: use devm API to cleanup module's resources > previous version [7] > - make aw200xx_*_action()'s bodies oneline > - don't shadow devm_mutex_init() return code > > leds: lm3532: use devm API to cleanup module's resources > leds: nic78bx: use devm API to cleanup module's resources > leds: mlxreg: use devm_mutex_init for mutex initializtion > leds: an30259a: use devm_mutext_init for mutext initialization > leds: powernv: add LED_RETAIN_AT_SHUTDOWN flag for leds > - those pathes were planned but not sent in the series #2 due to mail server > problem on my side. I revised them according to the comments. > > v3->v4: > locking: introduce devm_mutex_init > new patch > - move devm_mutex_init implementation completely from devm-helpers.h to mutex.h > > locking: add define if mutex_destroy() is not an empty function > drop the patch [9] > > devm-helpers: introduce devm_mutex_init > drop the patch [10] > > leds: aw2013: use devm API to cleanup module's resources > - add tag Tested-by: Nikita Travkin <nikita@trvn.ru> > > [4] https://lore.kernel.org/lkml/20231204180603.470421-1-gnstark@salutedevices.com/T/#mf500af0eda2a9ffc95594607dbe4cb64f2e3c9a8 > [5] https://lore.kernel.org/lkml/20231204180603.470421-1-gnstark@salutedevices.com/T/#mc92df4fb4f7d4187fb01cc1144acfa5fb5230dd2 > [6] https://lore.kernel.org/lkml/20231204180603.470421-1-gnstark@salutedevices.com/T/#m300df89710c43cc2ab598baa16c68dd0a0d7d681 > [7] https://lore.kernel.org/lkml/20231204180603.470421-1-gnstark@salutedevices.com/T/#m8e5c65e0c6b137c91fa00bb9320ad581164d1d0b > [8] https://lore.kernel.org/lkml/377e4437-7051-4d88-ae68-1460bcd692e1@redhat.com/T/#m5f84a4a2f387d49678783e652b9e658e02c27450 > [9] https://lore.kernel.org/lkml/20231213223020.2713164-1-gnstark@salutedevices.com/T/#m19ad1fc04c560012c1e27418e3156d0c9306dd84 > [10] https://lore.kernel.org/lkml/20231213223020.2713164-1-gnstark@salutedevices.com/T/#m63126025f5d1bdcef69bcad50f2e58274d42e2d7 > > George Stark (10): > leds: aw2013: unlock mutex before destroying it > locking: introduce devm_mutex_init > leds: aw2013: use devm API to cleanup module's resources > leds: aw200xx: use devm API to cleanup module's resources > leds: lp3952: use devm API to cleanup module's resources > leds: lm3532: use devm API to cleanup module's resources > leds: nic78bx: use devm API to cleanup module's resources > leds: mlxreg: use devm_mutex_init for mutex initializtion > leds: an30259a: use devm_mutext_init for mutext initialization > leds: powernv: use LED_RETAIN_AT_SHUTDOWN flag for leds > > drivers/leds/leds-an30259a.c | 15 +++++---------- > drivers/leds/leds-aw200xx.c | 33 ++++++++++++++++++++++----------- > drivers/leds/leds-aw2013.c | 27 +++++++++++++++------------ > drivers/leds/leds-lm3532.c | 30 ++++++++++++++++++------------ > drivers/leds/leds-lp3952.c | 21 +++++++++++---------- > drivers/leds/leds-mlxreg.c | 17 ++++++----------- > drivers/leds/leds-nic78bx.c | 25 +++++++++++++------------ > drivers/leds/leds-powernv.c | 23 ++++++++--------------- FYI: I'll conduct my review once the locking side is settled. > include/linux/mutex.h | 23 +++++++++++++++++++++++ > kernel/locking/mutex-debug.c | 22 ++++++++++++++++++++++ > 10 files changed, 143 insertions(+), 93 deletions(-) > > -- > 2.25.1 > >
On Thu, Dec 21, 2023 at 03:11:11PM +0000, Lee Jones wrote: > On Thu, 14 Dec 2023, George Stark wrote: > > > This patch series fixes the problem of devm_led_classdev_register misusing. > > > > The basic problem is described in [1]. Shortly when devm_led_classdev_register() > > is used then led_classdev_unregister() called after driver's remove() callback. > > led_classdev_unregister() calls driver's brightness_set callback and that callback > > may use resources which were destroyed already in driver's remove(). > > > > After discussion with maintainers [2] [3] we decided: > > 1) don't touch led subsytem core code and don't remove led_set_brightness() from it > > but fix drivers > > 2) don't use devm_led_classdev_unregister > > > > So the solution is to use devm wrappers for all resources > > driver's brightness_set() depends on. And introduce dedicated devm wrapper > > for mutex as it's often used resource. > > > > [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ > > [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f > > [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 ... > FYI: I'll conduct my review once the locking side is settled. To reduce burden can you apply the first one? It's a fix.
On Thu, Dec 14, 2023 at 08:36:04PM +0300, George Stark wrote: > This patch series fixes the problem of devm_led_classdev_register misusing. > > The basic problem is described in [1]. Shortly when devm_led_classdev_register() > is used then led_classdev_unregister() called after driver's remove() callback. > led_classdev_unregister() calls driver's brightness_set callback and that callback > may use resources which were destroyed already in driver's remove(). > > After discussion with maintainers [2] [3] we decided: > 1) don't touch led subsytem core code and don't remove led_set_brightness() from it > but fix drivers > 2) don't use devm_led_classdev_unregister > > So the solution is to use devm wrappers for all resources > driver's brightness_set() depends on. And introduce dedicated devm wrapper > for mutex as it's often used resource. > > [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ > [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f > [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 Are you going to send an updated version with the amended second patch?
Hello Andy. I haven't lose hope for the devm_mutex thing and keep pinging those guys from time to time. Sure I can single out the fix-only patch I'll do it tomorrow. On 2/9/24 20:11, Andy Shevchenko wrote: > On Thu, Dec 21, 2023 at 03:11:11PM +0000, Lee Jones wrote: >> On Thu, 14 Dec 2023, George Stark wrote: >> >>> This patch series fixes the problem of devm_led_classdev_register misusing. >>> >>> The basic problem is described in [1]. Shortly when devm_led_classdev_register() >>> is used then led_classdev_unregister() called after driver's remove() callback. >>> led_classdev_unregister() calls driver's brightness_set callback and that callback >>> may use resources which were destroyed already in driver's remove(). >>> >>> After discussion with maintainers [2] [3] we decided: >>> 1) don't touch led subsytem core code and don't remove led_set_brightness() from it >>> but fix drivers >>> 2) don't use devm_led_classdev_unregister >>> >>> So the solution is to use devm wrappers for all resources >>> driver's brightness_set() depends on. And introduce dedicated devm wrapper >>> for mutex as it's often used resource. >>> >>> [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ >>> [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f >>> [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 > > ... > >> FYI: I'll conduct my review once the locking side is settled. > > To reduce burden can you apply the first one? It's a fix. >
On Mon, Feb 12, 2024 at 1:52 AM George Stark <gnstark@salutedevices.com> wrote: > I haven't lose hope for the devm_mutex thing and keep pinging those guys > from time to time. I don't understand. According to v4 thread Christophe proposed on how the patch should look like. What you need is to incorporate an updated version into your series. Am I wrong? > Sure I can single out the fix-only patch I'll do it tomorrow. I believe it can be handled without issuing it separately. `b4` tool is capable of selective choices. It was rather Q to Lee if he can/want to apply it right away. > On 2/9/24 20:11, Andy Shevchenko wrote: > > On Thu, Dec 21, 2023 at 03:11:11PM +0000, Lee Jones wrote: > >> On Thu, 14 Dec 2023, George Stark wrote: > >> > >>> This patch series fixes the problem of devm_led_classdev_register misusing. > >>> > >>> The basic problem is described in [1]. Shortly when devm_led_classdev_register() > >>> is used then led_classdev_unregister() called after driver's remove() callback. > >>> led_classdev_unregister() calls driver's brightness_set callback and that callback > >>> may use resources which were destroyed already in driver's remove(). > >>> > >>> After discussion with maintainers [2] [3] we decided: > >>> 1) don't touch led subsytem core code and don't remove led_set_brightness() from it > >>> but fix drivers > >>> 2) don't use devm_led_classdev_unregister > >>> > >>> So the solution is to use devm wrappers for all resources > >>> driver's brightness_set() depends on. And introduce dedicated devm wrapper > >>> for mutex as it's often used resource. > >>> > >>> [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ > >>> [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f > >>> [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 > > > > ... > > > >> FYI: I'll conduct my review once the locking side is settled. > > > > To reduce burden can you apply the first one? It's a fix.
Hello Andy On 2/12/24 12:53, Andy Shevchenko wrote: > On Mon, Feb 12, 2024 at 1:52 AM George Stark <gnstark@salutedevices.com> wrote: >> I haven't lose hope for the devm_mutex thing and keep pinging those guys >> from time to time. > > I don't understand. According to v4 thread Christophe proposed on how > the patch should look like. What you need is to incorporate an updated > version into your series. Am I wrong? We agreed that the effective way of implementing devm_mutex_init() is in mutex.h using forward declaration of struct device. The only inconvenient thing is that in the mutex.h mutex_init() declared after mutex_destroy() so we'll have to use condition #ifdef CONFIG_DEBUG_MUTEXES twice. Waiman Long proposed great cleanup patch [1] that eliminates the need of doubling #ifdef. That patch was reviewed a bit but it's still unapplied (near 2 months). I'm still trying to contact mutex.h guys but there're no any feedback yet. [1] https://lore.kernel.org/lkml/20231216013656.1382213-2-longman@redhat.com/T/#m795b230d662c1debb28463ad721ddba5b384340a > >> Sure I can single out the fix-only patch I'll do it tomorrow. > > I believe it can be handled without issuing it separately. `b4` tool > is capable of selective choices. It was rather Q to Lee if he can/want > to apply it right away. Oh ok, that would be great. > >> On 2/9/24 20:11, Andy Shevchenko wrote: >>> On Thu, Dec 21, 2023 at 03:11:11PM +0000, Lee Jones wrote: >>>> On Thu, 14 Dec 2023, George Stark wrote: >>>> >>>>> This patch series fixes the problem of devm_led_classdev_register misusing. >>>>> >>>>> The basic problem is described in [1]. Shortly when devm_led_classdev_register() >>>>> is used then led_classdev_unregister() called after driver's remove() callback. >>>>> led_classdev_unregister() calls driver's brightness_set callback and that callback >>>>> may use resources which were destroyed already in driver's remove(). >>>>> >>>>> After discussion with maintainers [2] [3] we decided: >>>>> 1) don't touch led subsytem core code and don't remove led_set_brightness() from it >>>>> but fix drivers >>>>> 2) don't use devm_led_classdev_unregister >>>>> >>>>> So the solution is to use devm wrappers for all resources >>>>> driver's brightness_set() depends on. And introduce dedicated devm wrapper >>>>> for mutex as it's often used resource. >>>>> >>>>> [1] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/ >>>>> [2] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mc132b9b350fa51931b4fcfe14705d9f06e91421f >>>>> [3] https://lore.kernel.org/lkml/8704539b-ed3b-44e6-aa82-586e2f895e2b@salutedevices.com/T/#mdbf572a85c33f869a553caf986b6228bb65c8383 >>> >>> ... >>> >>>> FYI: I'll conduct my review once the locking side is settled. >>> >>> To reduce burden can you apply the first one? It's a fix. >
On Tue, Feb 13, 2024 at 2:14 AM George Stark <gnstark@salutedevices.com> wrote: > > Hello Andy > > On 2/12/24 12:53, Andy Shevchenko wrote: > > On Mon, Feb 12, 2024 at 1:52 AM George Stark <gnstark@salutedevices.com> wrote: > >> I haven't lose hope for the devm_mutex thing and keep pinging those guys > >> from time to time. > > > > I don't understand. According to v4 thread Christophe proposed on how > > the patch should look like. What you need is to incorporate an updated > > version into your series. Am I wrong? > > We agreed that the effective way of implementing devm_mutex_init() is in > mutex.h using forward declaration of struct device. > The only inconvenient thing is that in the mutex.h mutex_init() declared > after mutex_destroy() so we'll have to use condition #ifdef > CONFIG_DEBUG_MUTEXES twice. Waiman Long proposed great cleanup patch [1] > that eliminates the need of doubling #ifdef. That patch was reviewed a > bit but it's still unapplied (near 2 months). I'm still trying to > contact mutex.h guys but there're no any feedback yet. So the roadmap (as I see it) is: - convince Lee to take the first patch while waiting for the others - incorporate the above mentioned patch into your series - make an ultimatum in case there is no reaction to get it applied on deadline, let's say within next cycle (if Lee is okay with a such, but this is normal practice when some maintainers are non-responsive) P.S. In case Lee doesn't want to take the first patch separately (let's say this week), send a new version with amended patches included. > [1] > https://lore.kernel.org/lkml/20231216013656.1382213-2-longman@redhat.com/T/#m795b230d662c1debb28463ad721ddba5b384340a
Hello Andy On 2/13/24 13:55, Andy Shevchenko wrote: > On Tue, Feb 13, 2024 at 2:14 AM George Stark <gnstark@salutedevices.com> wrote: >> >> Hello Andy >> >> On 2/12/24 12:53, Andy Shevchenko wrote: >>> On Mon, Feb 12, 2024 at 1:52 AM George Stark <gnstark@salutedevices.com> wrote: >>>> I haven't lose hope for the devm_mutex thing and keep pinging those guys >>>> from time to time. >>> >>> I don't understand. According to v4 thread Christophe proposed on how >>> the patch should look like. What you need is to incorporate an updated >>> version into your series. Am I wrong? >> >> We agreed that the effective way of implementing devm_mutex_init() is in >> mutex.h using forward declaration of struct device. >> The only inconvenient thing is that in the mutex.h mutex_init() declared >> after mutex_destroy() so we'll have to use condition #ifdef >> CONFIG_DEBUG_MUTEXES twice. Waiman Long proposed great cleanup patch [1] >> that eliminates the need of doubling #ifdef. That patch was reviewed a >> bit but it's still unapplied (near 2 months). I'm still trying to >> contact mutex.h guys but there're no any feedback yet. > > So the roadmap (as I see it) is: > - convince Lee to take the first patch while waiting for the others > - incorporate the above mentioned patch into your series > - make an ultimatum in case there is no reaction to get it applied on > deadline, let's say within next cycle (if Lee is okay with a such, but > this is normal practice when some maintainers are non-responsive) Well, it was interesting to know that there is such a practice. Waiman Long has just updated his patches with mutex.h cleanup [1] so I think we can wait for that series to be merged than I'll prepare may patchseries with or w\o the first patch. [1] https://lore.kernel.org/all/20240213031656.1375951-4-longman@redhat.com/T/ > > P.S. In case Lee doesn't want to take the first patch separately > (let's say this week), send a new version with amended patches > included. Ok > >> [1] >> https://lore.kernel.org/lkml/20231216013656.1382213-2-longman@redhat.com/T/#m795b230d662c1debb28463ad721ddba5b384340a > >