| Message ID | 20250215130849.227812-1-claudiu.beznea.uj@bp.renesas.com |
|---|---|
| State | New |
| Headers | show |
| Series | driver core: platform: Use devres group to free driver probe resources | expand |
Hi, Rafael, On 06.03.2025 08:11, Dmitry Torokhov wrote: > On Wed, Mar 05, 2025 at 02:03:09PM +0000, Jonathan Cameron wrote: >> On Wed, 19 Feb 2025 14:45:07 +0200 >> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >> >>> Hi, Daniel, Jonathan, >>> >>> On 15.02.2025 15:51, Claudiu Beznea wrote: >>>> Hi, Greg, >>>> >>>> On 15.02.2025 15:25, Greg KH wrote: >>>>> On Sat, Feb 15, 2025 at 03:08:49PM +0200, Claudiu wrote: >>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>> >>>>>> On the Renesas RZ/G3S (and other Renesas SoCs, e.g., RZ/G2{L, LC, UL}), >>>>>> clocks are managed through PM domains. These PM domains, registered on >>>>>> behalf of the clock controller driver, are configured with >>>>>> GENPD_FLAG_PM_CLK. In most of the Renesas drivers used by RZ SoCs, the >>>>>> clocks are enabled/disabled using runtime PM APIs. The power domains may >>>>>> also have power_on/power_off support implemented. After the device PM >>>>>> domain is powered off any CPU accesses to these domains leads to system >>>>>> aborts. >>>>>> >>>>>> During probe, devices are attached to the PM domain controlling their >>>>>> clocks and power. Similarly, during removal, devices are detached from the >>>>>> PM domain. >>>>>> >>>>>> The detachment call stack is as follows: >>>>>> >>>>>> device_driver_detach() -> >>>>>> device_release_driver_internal() -> >>>>>> __device_release_driver() -> >>>>>> device_remove() -> >>>>>> platform_remove() -> >>>>>> dev_pm_domain_detach() >>>>>> >>>>>> During driver unbind, after the device is detached from its PM domain, >>>>>> the device_unbind_cleanup() function is called, which subsequently invokes >>>>>> devres_release_all(). This function handles devres resource cleanup. >>>>>> >>>>>> If runtime PM is enabled in driver probe via devm_pm_runtime_enable(), the >>>>>> cleanup process triggers the action or reset function for disabling runtime >>>>>> PM. This function is pm_runtime_disable_action(), which leads to the >>>>>> following call stack of interest when called: >>>>>> >>>>>> pm_runtime_disable_action() -> >>>>>> pm_runtime_dont_use_autosuspend() -> >>>>>> __pm_runtime_use_autosuspend() -> >>>>>> update_autosuspend() -> >>>>>> rpm_idle() >>>>>> >>>>>> The rpm_idle() function attempts to resume the device at runtime. However, >>>>>> at the point it is called, the device is no longer part of a PM domain >>>>>> (which manages clocks and power states). If the driver implements its own >>>>>> runtime PM APIs for specific functionalities - such as the rzg2l_adc >>>>>> driver - while also relying on the power domain subsystem for power >>>>>> management, rpm_idle() will invoke the driver's runtime PM API. However, >>>>>> since the device is no longer part of a PM domain at this point, the PM >>>>>> domain's runtime PM APIs will not be called. This leads to system aborts on >>>>>> Renesas SoCs. >>>>>> >>>>>> Another identified case is when a subsystem performs various cleanups >>>>>> using device_unbind_cleanup(), calling driver-specific APIs in the process. >>>>>> A known example is the thermal subsystem, which may call driver-specific >>>>>> APIs to disable the thermal device. The relevant call stack in this case >>>>>> is: >>>>>> >>>>>> device_driver_detach() -> >>>>>> device_release_driver_internal() -> >>>>>> device_unbind_cleanup() -> >>>>>> devres_release_all() -> >>>>>> devm_thermal_of_zone_release() -> >>>>>> thermal_zone_device_disable() -> >>>>>> thermal_zone_device_set_mode() -> >>>>>> struct thermal_zone_device_ops::change_mode() >>>>>> >>>>>> At the moment the driver-specific change_mode() API is called, the device >>>>>> is no longer part of its PM domain. Accessing its registers without proper >>>>>> power management leads to system aborts. >>>>>> >>>>>> Open a devres group before calling the driver probe, and close it >>>>>> immediately after the driver remove function is called and before >>>>>> dev_pm_domain_detach(). This ensures that driver-specific devm actions or >>>>>> reset functions are executed immediately after the driver remove function >>>>>> completes. Additionally, it prevents driver-specific runtime PM APIs from >>>>>> being called when the device is no longer part of its power domain. >>>>>> >>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>> --- >>>>>> >>>>>> Hi, >> >> Hi Claudiu, Greg, >> >> Sorry, I missed this thread whilst travelling and only saw it because >> of reference from the in driver solution. >> >>>>>> >>>>>> Although Ulf gave its green light for the approaches on both IIO [1], >>>>>> [2] and thermal subsystems [3], Jonathan considered unacceptable the >>>>>> approaches in [1], [2] as he considered it may lead to dificult to >>>>>> maintain code and code opened to subtle bugs (due to the potential of >>>>>> mixing devres and non-devres calls). He pointed out a similar approach >>>>>> that was done for the I2C bus [4], [5]. >>>>>> >>>>>> As the discussions in [1], [2] stopped w/o a clear conclusion, this >>>>>> patch tries to revive it by proposing a similar approach that was done >>>>>> for the I2C bus. >>>>>> >>>>>> Please let me know you input. >>>>> >>>>> I'm with Jonathan here, the devres stuff is getting crazy here and you >>>>> have drivers mixing them and side affects happening and lots of >>>>> confusion. Your change here is only going to make it even more >>>>> confusing, and shouldn't actually solve it for other busses (i.e. what >>>>> about iio devices NOT on the platform bus?) >> >> In some cases they are already carrying the support as per the link >> above covering all i2c drivers. I'd like to see a generic solution and >> I suspect pushing it to the device drivers rather than the bus code >> will explode badly and leave us with subtle bugs where people don't >> realise it is necessary. >> >> https://lore.kernel.org/all/20250224120608.1769039-1-claudiu.beznea.uj@bp.renesas.com/ >> is a lot nastier looking than what we have here. I'll review that in a minute >> to show that it need not be that bad, but none the less not pleasant. >> >> +CC linux-iio to join up threads and Dmitry wrt to i2c case (and HID that does >> similar) > > We should not expect individual drivers handle this, because this is a > layering violation: they need to know implementation details of the bus > code to know if the bus is using non-devres managed resources, and > adjust their behavior. Moving this into driver core is also not > feasible, as not all buses need it. So IMO this should belong to > individual bus code. > > Instead of using devres group a bus may opt to use > devm_add_action_or_reset() and other devm APIs to make sure bus' > resource unwinding is carried in the correct order relative to freeing > driver-owned resources. Can you please let us know your input on the approach proposed in this patch? Or if you would prefer devm_add_action_or_reset() as suggested by Dmitry? Or if you consider another approach would fit better? Currently there were issues identified with the rzg2l-adc driver (driver based solution proposed in [1]) and with the rzg3s thermal driver (solved by function rzg3s_thermal_probe() from [2]). As expressed previously by Jonathan and Dimitry this is a common problem and as the issue is due to a call in the bus driver, would be better and simpler to handle it in the bus driver. Otherwise, individual drivers would have to be adjusted in a similar way. Thank you, Claudiu [1] https://lore.kernel.org/all/20250324122627.32336-2-claudiu.beznea.uj@bp.renesas.com/ [2] https://lore.kernel.org/all/20250324135701.179827-3-claudiu.beznea.uj@bp.renesas.com/ > >> >>>> >>>> You're right, other busses will still have this problem. >>>> >>>>> >>>>> Why can't your individual driver handle this instead? >> >> In my mind because it's the bus code that is doing the unexpected part by >> making calls in the remove path that are effectively not in the same order >> as probe because they occur between driver remove and related devres cleanup >> for stuff registered in probe. >> >>>> >>>> Initially I tried it at the driver level by using non-devres PM runtime >>>> enable API but wasn't considered OK by all parties. >>>> >>>> I haven't thought about having devres_open_group()/devres_close_group() in >>>> the driver itself but it should work. >>> >>> Are you OK with having the devres_open_group()/devres_close_group() in the >>> currently known affected drivers (drivers/iio/adc/rzg2l_adc.c and the >>> proposed drivers/thermal/renesas/rzg3s_thermal.c [1]) ? >> >> I guess it may be the best of a bunch of not particularly nasty solutions... > > We need to update _ALL_ platform drivers using devm then, and this is > clearly not scalable. > > Thanks. >
On Thu, 27 Mar 2025 18:47:53 +0200 Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > Hi, Rafael, > > On 06.03.2025 08:11, Dmitry Torokhov wrote: > > On Wed, Mar 05, 2025 at 02:03:09PM +0000, Jonathan Cameron wrote: > >> On Wed, 19 Feb 2025 14:45:07 +0200 > >> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > >> > >>> Hi, Daniel, Jonathan, > >>> > >>> On 15.02.2025 15:51, Claudiu Beznea wrote: > >>>> Hi, Greg, > >>>> > >>>> On 15.02.2025 15:25, Greg KH wrote: > >>>>> On Sat, Feb 15, 2025 at 03:08:49PM +0200, Claudiu wrote: > >>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> > >>>>>> > >>>>>> On the Renesas RZ/G3S (and other Renesas SoCs, e.g., RZ/G2{L, LC, UL}), > >>>>>> clocks are managed through PM domains. These PM domains, registered on > >>>>>> behalf of the clock controller driver, are configured with > >>>>>> GENPD_FLAG_PM_CLK. In most of the Renesas drivers used by RZ SoCs, the > >>>>>> clocks are enabled/disabled using runtime PM APIs. The power domains may > >>>>>> also have power_on/power_off support implemented. After the device PM > >>>>>> domain is powered off any CPU accesses to these domains leads to system > >>>>>> aborts. > >>>>>> > >>>>>> During probe, devices are attached to the PM domain controlling their > >>>>>> clocks and power. Similarly, during removal, devices are detached from the > >>>>>> PM domain. > >>>>>> > >>>>>> The detachment call stack is as follows: > >>>>>> > >>>>>> device_driver_detach() -> > >>>>>> device_release_driver_internal() -> > >>>>>> __device_release_driver() -> > >>>>>> device_remove() -> > >>>>>> platform_remove() -> > >>>>>> dev_pm_domain_detach() > >>>>>> > >>>>>> During driver unbind, after the device is detached from its PM domain, > >>>>>> the device_unbind_cleanup() function is called, which subsequently invokes > >>>>>> devres_release_all(). This function handles devres resource cleanup. > >>>>>> > >>>>>> If runtime PM is enabled in driver probe via devm_pm_runtime_enable(), the > >>>>>> cleanup process triggers the action or reset function for disabling runtime > >>>>>> PM. This function is pm_runtime_disable_action(), which leads to the > >>>>>> following call stack of interest when called: > >>>>>> > >>>>>> pm_runtime_disable_action() -> > >>>>>> pm_runtime_dont_use_autosuspend() -> > >>>>>> __pm_runtime_use_autosuspend() -> > >>>>>> update_autosuspend() -> > >>>>>> rpm_idle() > >>>>>> > >>>>>> The rpm_idle() function attempts to resume the device at runtime. However, > >>>>>> at the point it is called, the device is no longer part of a PM domain > >>>>>> (which manages clocks and power states). If the driver implements its own > >>>>>> runtime PM APIs for specific functionalities - such as the rzg2l_adc > >>>>>> driver - while also relying on the power domain subsystem for power > >>>>>> management, rpm_idle() will invoke the driver's runtime PM API. However, > >>>>>> since the device is no longer part of a PM domain at this point, the PM > >>>>>> domain's runtime PM APIs will not be called. This leads to system aborts on > >>>>>> Renesas SoCs. > >>>>>> > >>>>>> Another identified case is when a subsystem performs various cleanups > >>>>>> using device_unbind_cleanup(), calling driver-specific APIs in the process. > >>>>>> A known example is the thermal subsystem, which may call driver-specific > >>>>>> APIs to disable the thermal device. The relevant call stack in this case > >>>>>> is: > >>>>>> > >>>>>> device_driver_detach() -> > >>>>>> device_release_driver_internal() -> > >>>>>> device_unbind_cleanup() -> > >>>>>> devres_release_all() -> > >>>>>> devm_thermal_of_zone_release() -> > >>>>>> thermal_zone_device_disable() -> > >>>>>> thermal_zone_device_set_mode() -> > >>>>>> struct thermal_zone_device_ops::change_mode() > >>>>>> > >>>>>> At the moment the driver-specific change_mode() API is called, the device > >>>>>> is no longer part of its PM domain. Accessing its registers without proper > >>>>>> power management leads to system aborts. > >>>>>> > >>>>>> Open a devres group before calling the driver probe, and close it > >>>>>> immediately after the driver remove function is called and before > >>>>>> dev_pm_domain_detach(). This ensures that driver-specific devm actions or > >>>>>> reset functions are executed immediately after the driver remove function > >>>>>> completes. Additionally, it prevents driver-specific runtime PM APIs from > >>>>>> being called when the device is no longer part of its power domain. > >>>>>> > >>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> > >>>>>> --- > >>>>>> > >>>>>> Hi, > >> > >> Hi Claudiu, Greg, > >> > >> Sorry, I missed this thread whilst travelling and only saw it because > >> of reference from the in driver solution. > >> > >>>>>> > >>>>>> Although Ulf gave its green light for the approaches on both IIO [1], > >>>>>> [2] and thermal subsystems [3], Jonathan considered unacceptable the > >>>>>> approaches in [1], [2] as he considered it may lead to dificult to > >>>>>> maintain code and code opened to subtle bugs (due to the potential of > >>>>>> mixing devres and non-devres calls). He pointed out a similar approach > >>>>>> that was done for the I2C bus [4], [5]. > >>>>>> > >>>>>> As the discussions in [1], [2] stopped w/o a clear conclusion, this > >>>>>> patch tries to revive it by proposing a similar approach that was done > >>>>>> for the I2C bus. > >>>>>> > >>>>>> Please let me know you input. > >>>>> > >>>>> I'm with Jonathan here, the devres stuff is getting crazy here and you > >>>>> have drivers mixing them and side affects happening and lots of > >>>>> confusion. Your change here is only going to make it even more > >>>>> confusing, and shouldn't actually solve it for other busses (i.e. what > >>>>> about iio devices NOT on the platform bus?) > >> > >> In some cases they are already carrying the support as per the link > >> above covering all i2c drivers. I'd like to see a generic solution and > >> I suspect pushing it to the device drivers rather than the bus code > >> will explode badly and leave us with subtle bugs where people don't > >> realise it is necessary. > >> > >> https://lore.kernel.org/all/20250224120608.1769039-1-claudiu.beznea.uj@bp.renesas.com/ > >> is a lot nastier looking than what we have here. I'll review that in a minute > >> to show that it need not be that bad, but none the less not pleasant. > >> > >> +CC linux-iio to join up threads and Dmitry wrt to i2c case (and HID that does > >> similar) > > > > We should not expect individual drivers handle this, because this is a > > layering violation: they need to know implementation details of the bus > > code to know if the bus is using non-devres managed resources, and > > adjust their behavior. Moving this into driver core is also not > > feasible, as not all buses need it. So IMO this should belong to > > individual bus code. > > > > Instead of using devres group a bus may opt to use > > devm_add_action_or_reset() and other devm APIs to make sure bus' > > resource unwinding is carried in the correct order relative to freeing > > driver-owned resources. > > Can you please let us know your input on the approach proposed in this > patch? Or if you would prefer devm_add_action_or_reset() as suggested by > Dmitry? Or if you consider another approach would fit better? > > Currently there were issues identified with the rzg2l-adc driver (driver > based solution proposed in [1]) and with the rzg3s thermal driver (solved > by function rzg3s_thermal_probe() from [2]). > > As expressed previously by Jonathan and Dimitry this is a common problem > and as the issue is due to a call in the bus driver, would be better and > simpler to handle it in the bus driver. Otherwise, individual drivers would > have to be adjusted in a similar way. > Rafael, Greg suggested we ask for your input on the right option: https://lore.kernel.org/all/2025032703-genre-excitable-9473@gregkh/ (that thread has the other option). Jonathan > Thank you, > Claudiu > > [1] > https://lore.kernel.org/all/20250324122627.32336-2-claudiu.beznea.uj@bp.renesas.com/ > [2] > https://lore.kernel.org/all/20250324135701.179827-3-claudiu.beznea.uj@bp.renesas.com/ > > > > >> > >>>> > >>>> You're right, other busses will still have this problem. > >>>> > >>>>> > >>>>> Why can't your individual driver handle this instead? > >> > >> In my mind because it's the bus code that is doing the unexpected part by > >> making calls in the remove path that are effectively not in the same order > >> as probe because they occur between driver remove and related devres cleanup > >> for stuff registered in probe. > >> > >>>> > >>>> Initially I tried it at the driver level by using non-devres PM runtime > >>>> enable API but wasn't considered OK by all parties. > >>>> > >>>> I haven't thought about having devres_open_group()/devres_close_group() in > >>>> the driver itself but it should work. > >>> > >>> Are you OK with having the devres_open_group()/devres_close_group() in the > >>> currently known affected drivers (drivers/iio/adc/rzg2l_adc.c and the > >>> proposed drivers/thermal/renesas/rzg3s_thermal.c [1]) ? > >> > >> I guess it may be the best of a bunch of not particularly nasty solutions... > > > > We need to update _ALL_ platform drivers using devm then, and this is > > clearly not scalable. > > > > Thanks. > > >
Hi, Rafael, On 30.03.2025 18:31, Jonathan Cameron wrote: > On Thu, 27 Mar 2025 18:47:53 +0200 > Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > >> Hi, Rafael, >> >> On 06.03.2025 08:11, Dmitry Torokhov wrote: >>> On Wed, Mar 05, 2025 at 02:03:09PM +0000, Jonathan Cameron wrote: >>>> On Wed, 19 Feb 2025 14:45:07 +0200 >>>> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >>>> >>>>> Hi, Daniel, Jonathan, >>>>> >>>>> On 15.02.2025 15:51, Claudiu Beznea wrote: >>>>>> Hi, Greg, >>>>>> >>>>>> On 15.02.2025 15:25, Greg KH wrote: >>>>>>> On Sat, Feb 15, 2025 at 03:08:49PM +0200, Claudiu wrote: >>>>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>>>> >>>>>>>> On the Renesas RZ/G3S (and other Renesas SoCs, e.g., RZ/G2{L, LC, UL}), >>>>>>>> clocks are managed through PM domains. These PM domains, registered on >>>>>>>> behalf of the clock controller driver, are configured with >>>>>>>> GENPD_FLAG_PM_CLK. In most of the Renesas drivers used by RZ SoCs, the >>>>>>>> clocks are enabled/disabled using runtime PM APIs. The power domains may >>>>>>>> also have power_on/power_off support implemented. After the device PM >>>>>>>> domain is powered off any CPU accesses to these domains leads to system >>>>>>>> aborts. >>>>>>>> >>>>>>>> During probe, devices are attached to the PM domain controlling their >>>>>>>> clocks and power. Similarly, during removal, devices are detached from the >>>>>>>> PM domain. >>>>>>>> >>>>>>>> The detachment call stack is as follows: >>>>>>>> >>>>>>>> device_driver_detach() -> >>>>>>>> device_release_driver_internal() -> >>>>>>>> __device_release_driver() -> >>>>>>>> device_remove() -> >>>>>>>> platform_remove() -> >>>>>>>> dev_pm_domain_detach() >>>>>>>> >>>>>>>> During driver unbind, after the device is detached from its PM domain, >>>>>>>> the device_unbind_cleanup() function is called, which subsequently invokes >>>>>>>> devres_release_all(). This function handles devres resource cleanup. >>>>>>>> >>>>>>>> If runtime PM is enabled in driver probe via devm_pm_runtime_enable(), the >>>>>>>> cleanup process triggers the action or reset function for disabling runtime >>>>>>>> PM. This function is pm_runtime_disable_action(), which leads to the >>>>>>>> following call stack of interest when called: >>>>>>>> >>>>>>>> pm_runtime_disable_action() -> >>>>>>>> pm_runtime_dont_use_autosuspend() -> >>>>>>>> __pm_runtime_use_autosuspend() -> >>>>>>>> update_autosuspend() -> >>>>>>>> rpm_idle() >>>>>>>> >>>>>>>> The rpm_idle() function attempts to resume the device at runtime. However, >>>>>>>> at the point it is called, the device is no longer part of a PM domain >>>>>>>> (which manages clocks and power states). If the driver implements its own >>>>>>>> runtime PM APIs for specific functionalities - such as the rzg2l_adc >>>>>>>> driver - while also relying on the power domain subsystem for power >>>>>>>> management, rpm_idle() will invoke the driver's runtime PM API. However, >>>>>>>> since the device is no longer part of a PM domain at this point, the PM >>>>>>>> domain's runtime PM APIs will not be called. This leads to system aborts on >>>>>>>> Renesas SoCs. >>>>>>>> >>>>>>>> Another identified case is when a subsystem performs various cleanups >>>>>>>> using device_unbind_cleanup(), calling driver-specific APIs in the process. >>>>>>>> A known example is the thermal subsystem, which may call driver-specific >>>>>>>> APIs to disable the thermal device. The relevant call stack in this case >>>>>>>> is: >>>>>>>> >>>>>>>> device_driver_detach() -> >>>>>>>> device_release_driver_internal() -> >>>>>>>> device_unbind_cleanup() -> >>>>>>>> devres_release_all() -> >>>>>>>> devm_thermal_of_zone_release() -> >>>>>>>> thermal_zone_device_disable() -> >>>>>>>> thermal_zone_device_set_mode() -> >>>>>>>> struct thermal_zone_device_ops::change_mode() >>>>>>>> >>>>>>>> At the moment the driver-specific change_mode() API is called, the device >>>>>>>> is no longer part of its PM domain. Accessing its registers without proper >>>>>>>> power management leads to system aborts. >>>>>>>> >>>>>>>> Open a devres group before calling the driver probe, and close it >>>>>>>> immediately after the driver remove function is called and before >>>>>>>> dev_pm_domain_detach(). This ensures that driver-specific devm actions or >>>>>>>> reset functions are executed immediately after the driver remove function >>>>>>>> completes. Additionally, it prevents driver-specific runtime PM APIs from >>>>>>>> being called when the device is no longer part of its power domain. >>>>>>>> >>>>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>>>> --- >>>>>>>> >>>>>>>> Hi, >>>> >>>> Hi Claudiu, Greg, >>>> >>>> Sorry, I missed this thread whilst travelling and only saw it because >>>> of reference from the in driver solution. >>>> >>>>>>>> >>>>>>>> Although Ulf gave its green light for the approaches on both IIO [1], >>>>>>>> [2] and thermal subsystems [3], Jonathan considered unacceptable the >>>>>>>> approaches in [1], [2] as he considered it may lead to dificult to >>>>>>>> maintain code and code opened to subtle bugs (due to the potential of >>>>>>>> mixing devres and non-devres calls). He pointed out a similar approach >>>>>>>> that was done for the I2C bus [4], [5]. >>>>>>>> >>>>>>>> As the discussions in [1], [2] stopped w/o a clear conclusion, this >>>>>>>> patch tries to revive it by proposing a similar approach that was done >>>>>>>> for the I2C bus. >>>>>>>> >>>>>>>> Please let me know you input. >>>>>>> >>>>>>> I'm with Jonathan here, the devres stuff is getting crazy here and you >>>>>>> have drivers mixing them and side affects happening and lots of >>>>>>> confusion. Your change here is only going to make it even more >>>>>>> confusing, and shouldn't actually solve it for other busses (i.e. what >>>>>>> about iio devices NOT on the platform bus?) >>>> >>>> In some cases they are already carrying the support as per the link >>>> above covering all i2c drivers. I'd like to see a generic solution and >>>> I suspect pushing it to the device drivers rather than the bus code >>>> will explode badly and leave us with subtle bugs where people don't >>>> realise it is necessary. >>>> >>>> https://lore.kernel.org/all/20250224120608.1769039-1-claudiu.beznea.uj@bp.renesas.com/ >>>> is a lot nastier looking than what we have here. I'll review that in a minute >>>> to show that it need not be that bad, but none the less not pleasant. >>>> >>>> +CC linux-iio to join up threads and Dmitry wrt to i2c case (and HID that does >>>> similar) >>> >>> We should not expect individual drivers handle this, because this is a >>> layering violation: they need to know implementation details of the bus >>> code to know if the bus is using non-devres managed resources, and >>> adjust their behavior. Moving this into driver core is also not >>> feasible, as not all buses need it. So IMO this should belong to >>> individual bus code. >>> >>> Instead of using devres group a bus may opt to use >>> devm_add_action_or_reset() and other devm APIs to make sure bus' >>> resource unwinding is carried in the correct order relative to freeing >>> driver-owned resources. >> >> Can you please let us know your input on the approach proposed in this >> patch? Or if you would prefer devm_add_action_or_reset() as suggested by >> Dmitry? Or if you consider another approach would fit better? >> >> Currently there were issues identified with the rzg2l-adc driver (driver >> based solution proposed in [1]) and with the rzg3s thermal driver (solved >> by function rzg3s_thermal_probe() from [2]). >> >> As expressed previously by Jonathan and Dimitry this is a common problem >> and as the issue is due to a call in the bus driver, would be better and >> simpler to handle it in the bus driver. Otherwise, individual drivers would >> have to be adjusted in a similar way. >> > > Rafael, > > Greg suggested we ask for your input on the right option: > > https://lore.kernel.org/all/2025032703-genre-excitable-9473@gregkh/ > (that thread has the other option). Can you please let us know your opinion on this? Thank you, Claudiu > > Jonathan > >> Thank you, >> Claudiu >> >> [1] >> https://lore.kernel.org/all/20250324122627.32336-2-claudiu.beznea.uj@bp.renesas.com/ >> [2] >> https://lore.kernel.org/all/20250324135701.179827-3-claudiu.beznea.uj@bp.renesas.com/ >> >>> >>>> >>>>>> >>>>>> You're right, other busses will still have this problem. >>>>>> >>>>>>> >>>>>>> Why can't your individual driver handle this instead? >>>> >>>> In my mind because it's the bus code that is doing the unexpected part by >>>> making calls in the remove path that are effectively not in the same order >>>> as probe because they occur between driver remove and related devres cleanup >>>> for stuff registered in probe. >>>> >>>>>> >>>>>> Initially I tried it at the driver level by using non-devres PM runtime >>>>>> enable API but wasn't considered OK by all parties. >>>>>> >>>>>> I haven't thought about having devres_open_group()/devres_close_group() in >>>>>> the driver itself but it should work. >>>>> >>>>> Are you OK with having the devres_open_group()/devres_close_group() in the >>>>> currently known affected drivers (drivers/iio/adc/rzg2l_adc.c and the >>>>> proposed drivers/thermal/renesas/rzg3s_thermal.c [1]) ? >>>> >>>> I guess it may be the best of a bunch of not particularly nasty solutions... >>> >>> We need to update _ALL_ platform drivers using devm then, and this is >>> clearly not scalable. >>> >>> Thanks. >>> >> >
Hi, Ulf, Thank you for your input! On 09.05.2025 16:07, Ulf Hansson wrote: > On Fri, 9 May 2025 at 13:51, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >> >> Hi, Rafael, Ulf, PM list, >> >> >> On 09.04.2025 19:12, Claudiu Beznea wrote: >>> Hi, Rafael, >>> >>> On 30.03.2025 18:31, Jonathan Cameron wrote: >>>> On Thu, 27 Mar 2025 18:47:53 +0200 >>>> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >>>> >>>>> Hi, Rafael, >>>>> >>>>> On 06.03.2025 08:11, Dmitry Torokhov wrote: >>>>>> On Wed, Mar 05, 2025 at 02:03:09PM +0000, Jonathan Cameron wrote: >>>>>>> On Wed, 19 Feb 2025 14:45:07 +0200 >>>>>>> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >>>>>>> >>>>>>>> Hi, Daniel, Jonathan, >>>>>>>> >>>>>>>> On 15.02.2025 15:51, Claudiu Beznea wrote: >>>>>>>>> Hi, Greg, >>>>>>>>> >>>>>>>>> On 15.02.2025 15:25, Greg KH wrote: >>>>>>>>>> On Sat, Feb 15, 2025 at 03:08:49PM +0200, Claudiu wrote: >>>>>>>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>>>>>>> >>>>>>>>>>> On the Renesas RZ/G3S (and other Renesas SoCs, e.g., RZ/G2{L, LC, UL}), >>>>>>>>>>> clocks are managed through PM domains. These PM domains, registered on >>>>>>>>>>> behalf of the clock controller driver, are configured with >>>>>>>>>>> GENPD_FLAG_PM_CLK. In most of the Renesas drivers used by RZ SoCs, the >>>>>>>>>>> clocks are enabled/disabled using runtime PM APIs. The power domains may >>>>>>>>>>> also have power_on/power_off support implemented. After the device PM >>>>>>>>>>> domain is powered off any CPU accesses to these domains leads to system >>>>>>>>>>> aborts. >>>>>>>>>>> >>>>>>>>>>> During probe, devices are attached to the PM domain controlling their >>>>>>>>>>> clocks and power. Similarly, during removal, devices are detached from the >>>>>>>>>>> PM domain. >>>>>>>>>>> >>>>>>>>>>> The detachment call stack is as follows: >>>>>>>>>>> >>>>>>>>>>> device_driver_detach() -> >>>>>>>>>>> device_release_driver_internal() -> >>>>>>>>>>> __device_release_driver() -> >>>>>>>>>>> device_remove() -> >>>>>>>>>>> platform_remove() -> >>>>>>>>>>> dev_pm_domain_detach() >>>>>>>>>>> >>>>>>>>>>> During driver unbind, after the device is detached from its PM domain, >>>>>>>>>>> the device_unbind_cleanup() function is called, which subsequently invokes >>>>>>>>>>> devres_release_all(). This function handles devres resource cleanup. >>>>>>>>>>> >>>>>>>>>>> If runtime PM is enabled in driver probe via devm_pm_runtime_enable(), the >>>>>>>>>>> cleanup process triggers the action or reset function for disabling runtime >>>>>>>>>>> PM. This function is pm_runtime_disable_action(), which leads to the >>>>>>>>>>> following call stack of interest when called: >>>>>>>>>>> >>>>>>>>>>> pm_runtime_disable_action() -> >>>>>>>>>>> pm_runtime_dont_use_autosuspend() -> >>>>>>>>>>> __pm_runtime_use_autosuspend() -> >>>>>>>>>>> update_autosuspend() -> >>>>>>>>>>> rpm_idle() >>>>>>>>>>> >>>>>>>>>>> The rpm_idle() function attempts to resume the device at runtime. However, >>>>>>>>>>> at the point it is called, the device is no longer part of a PM domain >>>>>>>>>>> (which manages clocks and power states). If the driver implements its own >>>>>>>>>>> runtime PM APIs for specific functionalities - such as the rzg2l_adc >>>>>>>>>>> driver - while also relying on the power domain subsystem for power >>>>>>>>>>> management, rpm_idle() will invoke the driver's runtime PM API. However, >>>>>>>>>>> since the device is no longer part of a PM domain at this point, the PM >>>>>>>>>>> domain's runtime PM APIs will not be called. This leads to system aborts on >>>>>>>>>>> Renesas SoCs. >>>>>>>>>>> >>>>>>>>>>> Another identified case is when a subsystem performs various cleanups >>>>>>>>>>> using device_unbind_cleanup(), calling driver-specific APIs in the process. >>>>>>>>>>> A known example is the thermal subsystem, which may call driver-specific >>>>>>>>>>> APIs to disable the thermal device. The relevant call stack in this case >>>>>>>>>>> is: >>>>>>>>>>> >>>>>>>>>>> device_driver_detach() -> >>>>>>>>>>> device_release_driver_internal() -> >>>>>>>>>>> device_unbind_cleanup() -> >>>>>>>>>>> devres_release_all() -> >>>>>>>>>>> devm_thermal_of_zone_release() -> >>>>>>>>>>> thermal_zone_device_disable() -> >>>>>>>>>>> thermal_zone_device_set_mode() -> >>>>>>>>>>> struct thermal_zone_device_ops::change_mode() >>>>>>>>>>> >>>>>>>>>>> At the moment the driver-specific change_mode() API is called, the device >>>>>>>>>>> is no longer part of its PM domain. Accessing its registers without proper >>>>>>>>>>> power management leads to system aborts. >>>>>>>>>>> >>>>>>>>>>> Open a devres group before calling the driver probe, and close it >>>>>>>>>>> immediately after the driver remove function is called and before >>>>>>>>>>> dev_pm_domain_detach(). This ensures that driver-specific devm actions or >>>>>>>>>>> reset functions are executed immediately after the driver remove function >>>>>>>>>>> completes. Additionally, it prevents driver-specific runtime PM APIs from >>>>>>>>>>> being called when the device is no longer part of its power domain. >>>>>>>>>>> >>>>>>>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> >>>>>>>>>>> --- >>>>>>>>>>> >>>>>>>>>>> Hi, >>>>>>> >>>>>>> Hi Claudiu, Greg, >>>>>>> >>>>>>> Sorry, I missed this thread whilst travelling and only saw it because >>>>>>> of reference from the in driver solution. >>>>>>> >>>>>>>>>>> >>>>>>>>>>> Although Ulf gave its green light for the approaches on both IIO [1], >>>>>>>>>>> [2] and thermal subsystems [3], Jonathan considered unacceptable the >>>>>>>>>>> approaches in [1], [2] as he considered it may lead to dificult to >>>>>>>>>>> maintain code and code opened to subtle bugs (due to the potential of >>>>>>>>>>> mixing devres and non-devres calls). He pointed out a similar approach >>>>>>>>>>> that was done for the I2C bus [4], [5]. >>>>>>>>>>> >>>>>>>>>>> As the discussions in [1], [2] stopped w/o a clear conclusion, this >>>>>>>>>>> patch tries to revive it by proposing a similar approach that was done >>>>>>>>>>> for the I2C bus. >>>>>>>>>>> >>>>>>>>>>> Please let me know you input. >>>>>>>>>> >>>>>>>>>> I'm with Jonathan here, the devres stuff is getting crazy here and you >>>>>>>>>> have drivers mixing them and side affects happening and lots of >>>>>>>>>> confusion. Your change here is only going to make it even more >>>>>>>>>> confusing, and shouldn't actually solve it for other busses (i.e. what >>>>>>>>>> about iio devices NOT on the platform bus?) >>>>>>> >>>>>>> In some cases they are already carrying the support as per the link >>>>>>> above covering all i2c drivers. I'd like to see a generic solution and >>>>>>> I suspect pushing it to the device drivers rather than the bus code >>>>>>> will explode badly and leave us with subtle bugs where people don't >>>>>>> realise it is necessary. >>>>>>> >>>>>>> https://lore.kernel.org/all/20250224120608.1769039-1-claudiu.beznea.uj@bp.renesas.com/ >>>>>>> is a lot nastier looking than what we have here. I'll review that in a minute >>>>>>> to show that it need not be that bad, but none the less not pleasant. >>>>>>> >>>>>>> +CC linux-iio to join up threads and Dmitry wrt to i2c case (and HID that does >>>>>>> similar) >>>>>> >>>>>> We should not expect individual drivers handle this, because this is a >>>>>> layering violation: they need to know implementation details of the bus >>>>>> code to know if the bus is using non-devres managed resources, and >>>>>> adjust their behavior. Moving this into driver core is also not >>>>>> feasible, as not all buses need it. So IMO this should belong to >>>>>> individual bus code. >>>>>> >>>>>> Instead of using devres group a bus may opt to use >>>>>> devm_add_action_or_reset() and other devm APIs to make sure bus' >>>>>> resource unwinding is carried in the correct order relative to freeing >>>>>> driver-owned resources. >>>>> >>>>> Can you please let us know your input on the approach proposed in this >>>>> patch? Or if you would prefer devm_add_action_or_reset() as suggested by >>>>> Dmitry? Or if you consider another approach would fit better? >>>>> >>>>> Currently there were issues identified with the rzg2l-adc driver (driver >>>>> based solution proposed in [1]) and with the rzg3s thermal driver (solved >>>>> by function rzg3s_thermal_probe() from [2]). >>>>> >>>>> As expressed previously by Jonathan and Dimitry this is a common problem >>>>> and as the issue is due to a call in the bus driver, would be better and >>>>> simpler to handle it in the bus driver. Otherwise, individual drivers would >>>>> have to be adjusted in a similar way. >>>>> >>>> >>>> Rafael, >>>> >>>> Greg suggested we ask for your input on the right option: >>>> >>>> https://lore.kernel.org/all/2025032703-genre-excitable-9473@gregkh/ >>>> (that thread has the other option). >>> >>> Can you please let us know your opinion on this? >> Can you please let us know if you have any suggestions for this? > > It's been a while since I looked at this. Although as I understand it, > the main issue comes from using devm_pm_runtime_enable(). Yes, it comes from the usage of devm_pm_runtime_enable() in drivers and the dev_pm_domain_detach() call in platform_remove() right after calling driver's remove function. On the platform I experienced issues with, the dev_pm_domain_detach() drops the clocks from the device power domain and any subsequent PM runtime resume calls (that may happen in the devres cleanup phase) have no effect on enabling the clocks. If driver has functions registered (e.g. through devm_add_action_or_reset()), or driver specific runtime PM functions that access directly registers in the devres cleanup phase this leads to system aborts. > > As I have tried to argue before, I think devm_pm_runtime_enable() > should *not* be used. Not here, not at all. Runtime PM isn't like any > other resources that we fetch/release. Instead, it's a behaviour that > you turn on and off, which needs to be managed more carefully, rather > than relying on fetch/release ordering from devres. > > That said, I would convert the driver to use pm_runtime_enable() and > pm_runtime_disable() instead. I've tried this approach previously but it resulted in more complicated code and thus, Jonathan wasn't happy with it [1]. Another approach I've tried was to have devres group opened/closed in the driver itself [2], [3] but it was postponed as this approach may have a chance. At the moment I have 2 drivers waiting for a resolution on this [2], [3] and I recently posted a new one [4] that uses driver specific local devres group to avoid this issue. Thank you, Claudiu [1] https://lore.kernel.org/all/20250224120608.1769039-2-claudiu.beznea.uj@bp.renesas.com [2] https://lore.kernel.org/all/20250324122627.32336-2-claudiu.beznea.uj@bp.renesas.com [3] https://lore.kernel.org/all/20250324135701.179827-3-claudiu.beznea.uj@bp.renesas.com/ [4] https://lore.kernel.org/all/20250430103236.3511989-6-claudiu.beznea.uj@bp.renesas.com
Hi Ulf, On Mon, May 19, 2025 at 05:02:05PM +0200, Ulf Hansson wrote: > On Fri, 9 May 2025 at 16:12, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > > > Hi, Ulf, > > > > Thank you for your input! > > > > On 09.05.2025 16:07, Ulf Hansson wrote: > > > On Fri, 9 May 2025 at 13:51, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > >> > > >> Hi, Rafael, Ulf, PM list, > > >> > > >> > > >> On 09.04.2025 19:12, Claudiu Beznea wrote: > > >>> Hi, Rafael, > > >>> > > >>> On 30.03.2025 18:31, Jonathan Cameron wrote: > > >>>> On Thu, 27 Mar 2025 18:47:53 +0200 > > >>>> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > >>>> > > >>>>> Hi, Rafael, > > >>>>> > > >>>>> On 06.03.2025 08:11, Dmitry Torokhov wrote: > > >>>>>> On Wed, Mar 05, 2025 at 02:03:09PM +0000, Jonathan Cameron wrote: > > >>>>>>> On Wed, 19 Feb 2025 14:45:07 +0200 > > >>>>>>> Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > >>>>>>> > > >>>>>>>> Hi, Daniel, Jonathan, > > >>>>>>>> > > >>>>>>>> On 15.02.2025 15:51, Claudiu Beznea wrote: > > >>>>>>>>> Hi, Greg, > > >>>>>>>>> > > >>>>>>>>> On 15.02.2025 15:25, Greg KH wrote: > > >>>>>>>>>> On Sat, Feb 15, 2025 at 03:08:49PM +0200, Claudiu wrote: > > >>>>>>>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> > > >>>>>>>>>>> > > >>>>>>>>>>> On the Renesas RZ/G3S (and other Renesas SoCs, e.g., RZ/G2{L, LC, UL}), > > >>>>>>>>>>> clocks are managed through PM domains. These PM domains, registered on > > >>>>>>>>>>> behalf of the clock controller driver, are configured with > > >>>>>>>>>>> GENPD_FLAG_PM_CLK. In most of the Renesas drivers used by RZ SoCs, the > > >>>>>>>>>>> clocks are enabled/disabled using runtime PM APIs. The power domains may > > >>>>>>>>>>> also have power_on/power_off support implemented. After the device PM > > >>>>>>>>>>> domain is powered off any CPU accesses to these domains leads to system > > >>>>>>>>>>> aborts. > > >>>>>>>>>>> > > >>>>>>>>>>> During probe, devices are attached to the PM domain controlling their > > >>>>>>>>>>> clocks and power. Similarly, during removal, devices are detached from the > > >>>>>>>>>>> PM domain. > > >>>>>>>>>>> > > >>>>>>>>>>> The detachment call stack is as follows: > > >>>>>>>>>>> > > >>>>>>>>>>> device_driver_detach() -> > > >>>>>>>>>>> device_release_driver_internal() -> > > >>>>>>>>>>> __device_release_driver() -> > > >>>>>>>>>>> device_remove() -> > > >>>>>>>>>>> platform_remove() -> > > >>>>>>>>>>> dev_pm_domain_detach() > > >>>>>>>>>>> > > >>>>>>>>>>> During driver unbind, after the device is detached from its PM domain, > > >>>>>>>>>>> the device_unbind_cleanup() function is called, which subsequently invokes > > >>>>>>>>>>> devres_release_all(). This function handles devres resource cleanup. > > >>>>>>>>>>> > > >>>>>>>>>>> If runtime PM is enabled in driver probe via devm_pm_runtime_enable(), the > > >>>>>>>>>>> cleanup process triggers the action or reset function for disabling runtime > > >>>>>>>>>>> PM. This function is pm_runtime_disable_action(), which leads to the > > >>>>>>>>>>> following call stack of interest when called: > > >>>>>>>>>>> > > >>>>>>>>>>> pm_runtime_disable_action() -> > > >>>>>>>>>>> pm_runtime_dont_use_autosuspend() -> > > >>>>>>>>>>> __pm_runtime_use_autosuspend() -> > > >>>>>>>>>>> update_autosuspend() -> > > >>>>>>>>>>> rpm_idle() > > >>>>>>>>>>> > > >>>>>>>>>>> The rpm_idle() function attempts to resume the device at runtime. However, > > >>>>>>>>>>> at the point it is called, the device is no longer part of a PM domain > > >>>>>>>>>>> (which manages clocks and power states). If the driver implements its own > > >>>>>>>>>>> runtime PM APIs for specific functionalities - such as the rzg2l_adc > > >>>>>>>>>>> driver - while also relying on the power domain subsystem for power > > >>>>>>>>>>> management, rpm_idle() will invoke the driver's runtime PM API. However, > > >>>>>>>>>>> since the device is no longer part of a PM domain at this point, the PM > > >>>>>>>>>>> domain's runtime PM APIs will not be called. This leads to system aborts on > > >>>>>>>>>>> Renesas SoCs. > > >>>>>>>>>>> > > >>>>>>>>>>> Another identified case is when a subsystem performs various cleanups > > >>>>>>>>>>> using device_unbind_cleanup(), calling driver-specific APIs in the process. > > >>>>>>>>>>> A known example is the thermal subsystem, which may call driver-specific > > >>>>>>>>>>> APIs to disable the thermal device. The relevant call stack in this case > > >>>>>>>>>>> is: > > >>>>>>>>>>> > > >>>>>>>>>>> device_driver_detach() -> > > >>>>>>>>>>> device_release_driver_internal() -> > > >>>>>>>>>>> device_unbind_cleanup() -> > > >>>>>>>>>>> devres_release_all() -> > > >>>>>>>>>>> devm_thermal_of_zone_release() -> > > >>>>>>>>>>> thermal_zone_device_disable() -> > > >>>>>>>>>>> thermal_zone_device_set_mode() -> > > >>>>>>>>>>> struct thermal_zone_device_ops::change_mode() > > >>>>>>>>>>> > > >>>>>>>>>>> At the moment the driver-specific change_mode() API is called, the device > > >>>>>>>>>>> is no longer part of its PM domain. Accessing its registers without proper > > >>>>>>>>>>> power management leads to system aborts. > > >>>>>>>>>>> > > >>>>>>>>>>> Open a devres group before calling the driver probe, and close it > > >>>>>>>>>>> immediately after the driver remove function is called and before > > >>>>>>>>>>> dev_pm_domain_detach(). This ensures that driver-specific devm actions or > > >>>>>>>>>>> reset functions are executed immediately after the driver remove function > > >>>>>>>>>>> completes. Additionally, it prevents driver-specific runtime PM APIs from > > >>>>>>>>>>> being called when the device is no longer part of its power domain. > > >>>>>>>>>>> > > >>>>>>>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com> > > >>>>>>>>>>> --- > > >>>>>>>>>>> > > >>>>>>>>>>> Hi, > > >>>>>>> > > >>>>>>> Hi Claudiu, Greg, > > >>>>>>> > > >>>>>>> Sorry, I missed this thread whilst travelling and only saw it because > > >>>>>>> of reference from the in driver solution. > > >>>>>>> > > >>>>>>>>>>> > > >>>>>>>>>>> Although Ulf gave its green light for the approaches on both IIO [1], > > >>>>>>>>>>> [2] and thermal subsystems [3], Jonathan considered unacceptable the > > >>>>>>>>>>> approaches in [1], [2] as he considered it may lead to dificult to > > >>>>>>>>>>> maintain code and code opened to subtle bugs (due to the potential of > > >>>>>>>>>>> mixing devres and non-devres calls). He pointed out a similar approach > > >>>>>>>>>>> that was done for the I2C bus [4], [5]. > > >>>>>>>>>>> > > >>>>>>>>>>> As the discussions in [1], [2] stopped w/o a clear conclusion, this > > >>>>>>>>>>> patch tries to revive it by proposing a similar approach that was done > > >>>>>>>>>>> for the I2C bus. > > >>>>>>>>>>> > > >>>>>>>>>>> Please let me know you input. > > >>>>>>>>>> > > >>>>>>>>>> I'm with Jonathan here, the devres stuff is getting crazy here and you > > >>>>>>>>>> have drivers mixing them and side affects happening and lots of > > >>>>>>>>>> confusion. Your change here is only going to make it even more > > >>>>>>>>>> confusing, and shouldn't actually solve it for other busses (i.e. what > > >>>>>>>>>> about iio devices NOT on the platform bus?) > > >>>>>>> > > >>>>>>> In some cases they are already carrying the support as per the link > > >>>>>>> above covering all i2c drivers. I'd like to see a generic solution and > > >>>>>>> I suspect pushing it to the device drivers rather than the bus code > > >>>>>>> will explode badly and leave us with subtle bugs where people don't > > >>>>>>> realise it is necessary. > > >>>>>>> > > >>>>>>> https://lore.kernel.org/all/20250224120608.1769039-1-claudiu.beznea.uj@bp.renesas.com/ > > >>>>>>> is a lot nastier looking than what we have here. I'll review that in a minute > > >>>>>>> to show that it need not be that bad, but none the less not pleasant. > > >>>>>>> > > >>>>>>> +CC linux-iio to join up threads and Dmitry wrt to i2c case (and HID that does > > >>>>>>> similar) > > >>>>>> > > >>>>>> We should not expect individual drivers handle this, because this is a > > >>>>>> layering violation: they need to know implementation details of the bus > > >>>>>> code to know if the bus is using non-devres managed resources, and > > >>>>>> adjust their behavior. Moving this into driver core is also not > > >>>>>> feasible, as not all buses need it. So IMO this should belong to > > >>>>>> individual bus code. > > >>>>>> > > >>>>>> Instead of using devres group a bus may opt to use > > >>>>>> devm_add_action_or_reset() and other devm APIs to make sure bus' > > >>>>>> resource unwinding is carried in the correct order relative to freeing > > >>>>>> driver-owned resources. > > >>>>> > > >>>>> Can you please let us know your input on the approach proposed in this > > >>>>> patch? Or if you would prefer devm_add_action_or_reset() as suggested by > > >>>>> Dmitry? Or if you consider another approach would fit better? > > >>>>> > > >>>>> Currently there were issues identified with the rzg2l-adc driver (driver > > >>>>> based solution proposed in [1]) and with the rzg3s thermal driver (solved > > >>>>> by function rzg3s_thermal_probe() from [2]). > > >>>>> > > >>>>> As expressed previously by Jonathan and Dimitry this is a common problem > > >>>>> and as the issue is due to a call in the bus driver, would be better and > > >>>>> simpler to handle it in the bus driver. Otherwise, individual drivers would > > >>>>> have to be adjusted in a similar way. > > >>>>> > > >>>> > > >>>> Rafael, > > >>>> > > >>>> Greg suggested we ask for your input on the right option: > > >>>> > > >>>> https://lore.kernel.org/all/2025032703-genre-excitable-9473@gregkh/ > > >>>> (that thread has the other option). > > >>> > > >>> Can you please let us know your opinion on this? > > >> Can you please let us know if you have any suggestions for this? > > > > > > It's been a while since I looked at this. Although as I understand it, > > > the main issue comes from using devm_pm_runtime_enable(). > > > > Yes, it comes from the usage of devm_pm_runtime_enable() in drivers and the > > dev_pm_domain_detach() call in platform_remove() right after calling > > driver's remove function. > > Okay. This is not the root of the problem though. There is nothing really special about power domain and runtime power management. The root of the problem is that current code violates the order of releasing resources by mixing devm- and normal resource management together. Usually it is individual driver's fault, but in this case it is bus code that uses the manual release (dev_om_domain_detach) that violates the "release in opposite order to acquisition" rule. > > > > > On the platform I experienced issues with, the dev_pm_domain_detach() drops > > the clocks from the device power domain and any subsequent PM runtime > > resume calls (that may happen in the devres cleanup phase) have no effect > > on enabling the clocks. If driver has functions registered (e.g. through > > devm_add_action_or_reset()), or driver specific runtime PM functions that > > access directly registers in the devres cleanup phase this leads to system > > aborts. > > So if you move away from using devm_pm_runtime_enable() things would > be easier to manage and there is no additional new devres-management > needed. How exactly will it improve the situation? You still need to make sure that you are not disabling things out of the order. You simply moving the complexity to the driver, essentially forbidding it (and any other driver on platform bus) from using any devm APIs. > > > > > > > > > > > As I have tried to argue before, I think devm_pm_runtime_enable() > > > should *not* be used. Not here, not at all. Runtime PM isn't like any > > > other resources that we fetch/release. Instead, it's a behaviour that > > > you turn on and off, which needs to be managed more carefully, rather > > > than relying on fetch/release ordering from devres. I disagree. It is a resource that you turn on and off, same as clocks, regulators, interrupts, etc. We manage those during lifetime of the device, disable them when going into low power mode/suspend, reenable them upon resume, may disable and reenable them for other reasons. PM is not any more special here. As long as you keep the proper order of operations it works as well. > > > > > > That said, I would convert the driver to use pm_runtime_enable() and > > > pm_runtime_disable() instead. > > > > I've tried this approach previously but it resulted in more complicated > > code and thus, Jonathan wasn't happy with it [1]. > > I understand that you have been trying to move forward to address > people's opinions. It's not always easy to keep everybody happy. :-) > > That said, I still think this is the most viable option as it's how > the vast majority of drivers do it today. A few lines of additional > code shouldn't really be a big problem in my opinion. Have you tried making such change? Again, you will need to abandon use of most other devm APIs so that you keep the order of releasing resources. The only devm that you can still use is devm_k*alloc(), the rest has to be converted into unmanaged. Thanks.
Hi, Ulf, On 20.05.2025 15:09, Ulf Hansson wrote: > For example, even if the order is made correctly, suppose a driver's > ->remove() callback completes by turning off the resources for its > device and leaves runtime PM enabled, as it relies on devres to do it > some point later. Beyond this point, nothing would prevent userspace > for runtime resuming/suspending the device via sysfs. If I'm not wrong, that can't happen? The driver_sysfs_remove() is called before device_remove() (which calls the driver remove) is called, this being the call path: device_driver_detach() -> device_release_driver_internal() -> __device_release_driver() -> driver_sysfs_remove() // ... device_remove() And the driver_sysfs_remove() calls in the end __kernfs_remove() which looks to me like the place that actually drops the entries from sysfs, this being a call path for it: driver_sysfs_remove() -> sysfs_remove_link() -> kernfs_remove_by_name() -> kernfs_remove_by_name_ns() -> __kernfs_remove() -> activating the following line in __kernfs_remove(): pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); leads to the following prints when unbinding the watchdog device from its watchdog driver (attached to platform bus) on my board: https://p.fr33tux.org/935252 Thank you, Claudiu
On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > Hi, Ulf, > > On 20.05.2025 15:09, Ulf Hansson wrote: > > For example, even if the order is made correctly, suppose a driver's > > ->remove() callback completes by turning off the resources for its > > device and leaves runtime PM enabled, as it relies on devres to do it > > some point later. Beyond this point, nothing would prevent userspace > > for runtime resuming/suspending the device via sysfs. > > If I'm not wrong, that can't happen? The driver_sysfs_remove() is called > before device_remove() (which calls the driver remove) is called, this > being the call path: > > device_driver_detach() -> > device_release_driver_internal() -> > __device_release_driver() -> > driver_sysfs_remove() > // ... > device_remove() > > And the driver_sysfs_remove() calls in the end __kernfs_remove() which > looks to me like the place that actually drops the entries from sysfs, this > being a call path for it: > > driver_sysfs_remove() -> > sysfs_remove_link() -> > kernfs_remove_by_name() -> > kernfs_remove_by_name_ns() -> > __kernfs_remove() -> > > activating the following line in __kernfs_remove(): > > pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); > > leads to the following prints when unbinding the watchdog device from its > watchdog driver (attached to platform bus) on my board: > https://p.fr33tux.org/935252 Indeed this is a very good point you make! I completely overlooked this fact, thanks a lot for clarifying this! However, my main point still stands. In the end, there is nothing preventing rpm_suspend|resume|idle() in drivers/base/power/runtime.c from running (don't forget runtime PM is asynchronous too) for the device in question. This could lead to that a ->runtime_suspend|resume|idle() callback becomes executed at any point in time, as long as we haven't called pm_runtime_disable() for the device. That's why the devm_pm_runtime_enable() should be avoided as it simply introduces a race-condition. Drivers need to be more careful and use pm_runtime_enable|disable() explicitly to control the behaviour. Kind regards Uffe
On Wed, May 21, 2025 at 02:37:08PM +0200, Ulf Hansson wrote: > On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > > > Hi, Ulf, > > > > On 20.05.2025 15:09, Ulf Hansson wrote: > > > For example, even if the order is made correctly, suppose a driver's > > > ->remove() callback completes by turning off the resources for its > > > device and leaves runtime PM enabled, as it relies on devres to do it > > > some point later. Beyond this point, nothing would prevent userspace > > > for runtime resuming/suspending the device via sysfs. > > > > If I'm not wrong, that can't happen? The driver_sysfs_remove() is called > > before device_remove() (which calls the driver remove) is called, this > > being the call path: > > > > device_driver_detach() -> > > device_release_driver_internal() -> > > __device_release_driver() -> > > driver_sysfs_remove() > > // ... > > device_remove() > > > > And the driver_sysfs_remove() calls in the end __kernfs_remove() which > > looks to me like the place that actually drops the entries from sysfs, this > > being a call path for it: > > > > driver_sysfs_remove() -> > > sysfs_remove_link() -> > > kernfs_remove_by_name() -> > > kernfs_remove_by_name_ns() -> > > __kernfs_remove() -> > > > > activating the following line in __kernfs_remove(): > > > > pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); > > > > leads to the following prints when unbinding the watchdog device from its > > watchdog driver (attached to platform bus) on my board: > > https://p.fr33tux.org/935252 > > Indeed this is a very good point you make! I completely overlooked > this fact, thanks a lot for clarifying this! > > However, my main point still stands. > > In the end, there is nothing preventing rpm_suspend|resume|idle() in > drivers/base/power/runtime.c from running (don't forget runtime PM is > asynchronous too) for the device in question. This could lead to that > a ->runtime_suspend|resume|idle() callback becomes executed at any > point in time, as long as we haven't called pm_runtime_disable() for > the device. So exactly the same may happen if you enter driver->remove() and something calls runtime API before pm_runtime_disable() is called. The driver has (as they should be doing currently) be prepared for this. > > That's why the devm_pm_runtime_enable() should be avoided as it simply > introduces a race-condition. Drivers need to be more careful and use > pm_runtime_enable|disable() explicitly to control the behaviour. You make it sound like we are dealing with some non-deterministic process, like garbage collector, where runtime disable done by devm happens at some unspecified point in the future. However we are dealing with very well defined order of operations, all happening within __device_release_driver() call. It is the same scope as when using manual pm_runtime_disable(). Just the order is wrong, that is it. Thanks.
Hi, Ulf, On 21.05.2025 17:57, Dmitry Torokhov wrote: > On Wed, May 21, 2025 at 02:37:08PM +0200, Ulf Hansson wrote: >> On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >>> >>> Hi, Ulf, >>> >>> On 20.05.2025 15:09, Ulf Hansson wrote: >>>> For example, even if the order is made correctly, suppose a driver's >>>> ->remove() callback completes by turning off the resources for its >>>> device and leaves runtime PM enabled, as it relies on devres to do it >>>> some point later. Beyond this point, nothing would prevent userspace >>>> for runtime resuming/suspending the device via sysfs. >>> >>> If I'm not wrong, that can't happen? The driver_sysfs_remove() is called >>> before device_remove() (which calls the driver remove) is called, this >>> being the call path: >>> >>> device_driver_detach() -> >>> device_release_driver_internal() -> >>> __device_release_driver() -> >>> driver_sysfs_remove() >>> // ... >>> device_remove() >>> >>> And the driver_sysfs_remove() calls in the end __kernfs_remove() which >>> looks to me like the place that actually drops the entries from sysfs, this >>> being a call path for it: >>> >>> driver_sysfs_remove() -> >>> sysfs_remove_link() -> >>> kernfs_remove_by_name() -> >>> kernfs_remove_by_name_ns() -> >>> __kernfs_remove() -> >>> >>> activating the following line in __kernfs_remove(): >>> >>> pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); >>> >>> leads to the following prints when unbinding the watchdog device from its >>> watchdog driver (attached to platform bus) on my board: >>> https://p.fr33tux.org/935252 >> >> Indeed this is a very good point you make! I completely overlooked >> this fact, thanks a lot for clarifying this! >> >> However, my main point still stands. >> >> In the end, there is nothing preventing rpm_suspend|resume|idle() in >> drivers/base/power/runtime.c from running (don't forget runtime PM is >> asynchronous too) for the device in question. This could lead to that >> a ->runtime_suspend|resume|idle() callback becomes executed at any >> point in time, as long as we haven't called pm_runtime_disable() for >> the device. > > So exactly the same may happen if you enter driver->remove() and > something calls runtime API before pm_runtime_disable() is called. > The driver has (as they should be doing currently) be prepared for this. I took the time and tried to do a comparison of the current solutions (describing the bad and good things I see), trying to understand your concerns with regards to RPM suspend|resume|idle while unbinding a device from its driver. I see the following cases: Case 1/ the current approach when devm_pm_runtime_enable() is used in driver's ->probe() with the current code base: - right after driver ->remove() finish its execution clocks are detached from the PM domain, through dev_pm_domain_detach() call in platform_remove() - any subsequent RPM resume|suspend|idle will lead to failure if the driver specific RPM APIs access directly registers and counts on PM domain to enable/disable the clocks - at this point, if the IRQs are shared (but not only) and devm requested the driver's IRQ handler can still be called asynchronously; driver should be prepared for such events and should be written to work for such scenarios; but as the clocks are not in the PM domain anymore and RPM is still enabled at this point, if the driver don't run runtime suspend on probe (and runtime resume/suspend on runtime), I think (because I haven't investigated this yet) it can't rely on pm_runtime_active()/ pm_runtime_suspended() checks in interrupt handlers and can't decide if it can interrogate registers or not; interrogating should lead to failure at this stage as the clocks are disabled; drivers should work in such scenario and the CONFIG_DEBUG_SHIRQ is a way to check they can; I previously debugged a similar issue on drivers/net/ethernet/ renesas/ravb driver where using devm_pm_runtime_enable() in probe and pm_runtime_suspended() checks in IRQ handlers was the way to make this scenario happy; at that time I wasn't able to find that dev_pm_domain_detach() have the impact discussed in this thread Case 2/ What is proposed in this patch: devm_pm_runtime_enable() used + open devres group after dev_pm_domain_attach() (in probe) and close the devres group before dev_pm_domain_attach() (in remove): - right after the driver ->remove() is executed only the driver allocated devres resources are freed; this happens before dev_pm_domain_deattach() is called, though the proposed devres_release_group() call in this patch - while doing this, driver can still get async RPM suspend|resume|idle requests; is like the execution is in the driver ->remove() but the pm_runtime_disable() hasn't been called yet - as the runtime PM is enabled in driver's ->probe() mostly after the HW is prepared to take requests and all the other devm resources are allocated, the RPM disable is going to be among the first things to be called by the devres_release_group() - then, after RPM disable, all the devres resources allocated only in the driver's ->probe() are cleaned up in reverse order, just like device_unbind_cleanup() -> devres_release_all() call in __device_release_driver() is doing, but limited only to the resources allocated by the driver itself; I personally see this like manually allocating and freeing resources in the driver itself w/o relying on devres - then it comes the turn of dev_pm_domain_detach() call in platform_remove(): at the time dev_pm_domain_detach() is executed the runtime PM is disabled and all the devres resources allocated by driver are freed as well - after the dev_pm_domain_detach() is executed all the driver resources are cleaned up, the driver can't get IRQs as it's handler was already unregistered, no other user can execute rpm suspend|resume|idle as the RPM is disabled at this time Case 3/ devm_pm_runtime_enabled() dropped and replaced by manual cleanup: - the driver code is going be complicated, difficult to maintain and error prone I may have missed considering things when describing the case 2 (which is what is proposed by this patch) as I don't have the full picture behind the dev_pm_domain_detach() call in platform bus remove. If so, please correct me. Thank you, Claudiu > >> >> That's why the devm_pm_runtime_enable() should be avoided as it simply >> introduces a race-condition. Drivers need to be more careful and use >> pm_runtime_enable|disable() explicitly to control the behaviour. > > You make it sound like we are dealing with some non-deterministic > process, like garbage collector, where runtime disable done by devm > happens at some unspecified point in the future. However we are dealing > with very well defined order of operations, all happening within > __device_release_driver() call. It is the same scope as when using > manual pm_runtime_disable(). Just the order is wrong, that is it. > > Thanks. >
On Thu, 22 May 2025 at 11:48, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > Hi, Ulf, > > On 21.05.2025 17:57, Dmitry Torokhov wrote: > > On Wed, May 21, 2025 at 02:37:08PM +0200, Ulf Hansson wrote: > >> On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > >>> > >>> Hi, Ulf, > >>> > >>> On 20.05.2025 15:09, Ulf Hansson wrote: > >>>> For example, even if the order is made correctly, suppose a driver's > >>>> ->remove() callback completes by turning off the resources for its > >>>> device and leaves runtime PM enabled, as it relies on devres to do it > >>>> some point later. Beyond this point, nothing would prevent userspace > >>>> for runtime resuming/suspending the device via sysfs. > >>> > >>> If I'm not wrong, that can't happen? The driver_sysfs_remove() is called > >>> before device_remove() (which calls the driver remove) is called, this > >>> being the call path: > >>> > >>> device_driver_detach() -> > >>> device_release_driver_internal() -> > >>> __device_release_driver() -> > >>> driver_sysfs_remove() > >>> // ... > >>> device_remove() > >>> > >>> And the driver_sysfs_remove() calls in the end __kernfs_remove() which > >>> looks to me like the place that actually drops the entries from sysfs, this > >>> being a call path for it: > >>> > >>> driver_sysfs_remove() -> > >>> sysfs_remove_link() -> > >>> kernfs_remove_by_name() -> > >>> kernfs_remove_by_name_ns() -> > >>> __kernfs_remove() -> > >>> > >>> activating the following line in __kernfs_remove(): > >>> > >>> pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); > >>> > >>> leads to the following prints when unbinding the watchdog device from its > >>> watchdog driver (attached to platform bus) on my board: > >>> https://p.fr33tux.org/935252 > >> > >> Indeed this is a very good point you make! I completely overlooked > >> this fact, thanks a lot for clarifying this! > >> > >> However, my main point still stands. > >> > >> In the end, there is nothing preventing rpm_suspend|resume|idle() in > >> drivers/base/power/runtime.c from running (don't forget runtime PM is > >> asynchronous too) for the device in question. This could lead to that > >> a ->runtime_suspend|resume|idle() callback becomes executed at any > >> point in time, as long as we haven't called pm_runtime_disable() for > >> the device. > > > > So exactly the same may happen if you enter driver->remove() and > > something calls runtime API before pm_runtime_disable() is called. > > The driver has (as they should be doing currently) be prepared for this. > > I took the time and tried to do a comparison of the current solutions > (describing the bad and good things I see), trying to understand your > concerns with regards to RPM suspend|resume|idle while unbinding a device > from its driver. > > I see the following cases: > > Case 1/ the current approach when devm_pm_runtime_enable() is used in > driver's ->probe() with the current code base: > > - right after driver ->remove() finish its execution clocks are detached > from the PM domain, through dev_pm_domain_detach() call in > platform_remove() > > - any subsequent RPM resume|suspend|idle will lead to failure if the driver > specific RPM APIs access directly registers and counts on PM domain to > enable/disable the clocks > > - at this point, if the IRQs are shared (but not only) and devm requested > the driver's IRQ handler can still be called asynchronously; driver > should be prepared for such events and should be written to work for such > scenarios; but as the clocks are not in the PM domain anymore and RPM is > still enabled at this point, if the driver don't run runtime suspend on > probe (and runtime resume/suspend on runtime), I think (because I haven't > investigated this yet) it can't rely on pm_runtime_active()/ > pm_runtime_suspended() checks in interrupt handlers > and can't decide if it can interrogate registers or not; interrogating > should lead to failure at this stage as the clocks are disabled; drivers > should work in such scenario and the CONFIG_DEBUG_SHIRQ is a way to check > they can; I previously debugged a similar issue on drivers/net/ethernet/ > renesas/ravb driver where using devm_pm_runtime_enable() in probe and > pm_runtime_suspended() checks in IRQ handlers was the way to make this > scenario happy; at that time I wasn't able to find that > dev_pm_domain_detach() have the impact discussed in this thread > > Case 2/ What is proposed in this patch: devm_pm_runtime_enable() used + > open devres group after dev_pm_domain_attach() (in probe) and close the > devres group before dev_pm_domain_attach() (in remove): > > - right after the driver ->remove() is executed only the driver allocated > devres resources are freed; this happens before dev_pm_domain_deattach() > is called, though the proposed devres_release_group() call in this patch > > - while doing this, driver can still get async RPM suspend|resume|idle > requests; is like the execution is in the driver ->remove() > but the pm_runtime_disable() hasn't been called yet > > - as the runtime PM is enabled in driver's ->probe() mostly after the HW is > prepared to take requests and all the other devm resources are allocated, > the RPM disable is going to be among the first things to be called by the > devres_release_group() > > - then, after RPM disable, all the devres resources allocated only in the > driver's ->probe() are cleaned up in reverse order, just like > device_unbind_cleanup() -> devres_release_all() call in > __device_release_driver() is doing, but limited only to the resources > allocated by the driver itself; I personally see this like manually > allocating and freeing resources in the driver itself w/o relying on > devres > > - then it comes the turn of dev_pm_domain_detach() call in > platform_remove(): at the time dev_pm_domain_detach() is executed the > runtime PM is disabled and all the devres resources allocated by driver > are freed as well > > - after the dev_pm_domain_detach() is executed all the driver resources > are cleaned up, the driver can't get IRQs as it's handler was already > unregistered, no other user can execute rpm suspend|resume|idle > as the RPM is disabled at this time > > Case 3/ devm_pm_runtime_enabled() dropped and replaced by manual cleanup: > - the driver code is going be complicated, difficult to maintain and error > prone Yes, the driver's code would become slightly more complicated, but more importantly it would be correct. To me it sounds like the driver's ->remove() callback could do this: pm_runtime_get_sync() pm_runtime_disable() pm_runtime_put_noidle() In this way, the driver will runtime resume its device, allowing devres to drop/turn-off resources in the order we want. Except for the clocks, as those would be turned off via dev_pm_domain_detach() before the IRQ handler is freed (via devres), right? To avoid getting the IRQ handler to be called when it can't access registers, we could do one of the below: *) Look for a condition in the IRQ handler and bail-out when we know we should not manage IRQs. Is using pm_runtime_enabled() sufficient, you think? Otherwise we need a driver specific flag, which should be set in ->remove(). *) Don't use devm* when registering the IRQ handler. Yes, both options further contribute to making the driver code slightly more complicated, but if you want to solve the problem sooner than later, I think this is what you need to do. Yet, I think there is another option too, see below. > > I may have missed considering things when describing the case 2 (which is > what is proposed by this patch) as I don't have the full picture behind the > dev_pm_domain_detach() call in platform bus remove. If so, please correct me. The dev_pm_domain_attach|detach() calls in bus level code (probe/remove) were added there a long time ago, way before devres was being used like today. Currently we also have devm_pm_domain_attach_list(), which is used when devices have multiple PM domains to attach too. This is *not* called by bus-level code, but by the driver themselves. For these cases, we would not encounter the problems you have been facing with clocks/IRQ-handler, I think - because the devres order is maintained for PM domains too. That said, I think adding a devm_pm_domain_attach() interface would make perfect sense. Then we can try to replace dev_pm_domain_attach|detach() in bus level code, with just a call to devm_pm_domain_attach(). In this way, we should preserve the expectation for drivers around devres for PM domains. Even if it would change the behaviour for some drivers, it still sounds like the correct thing to do in my opinion. [...] Kind regards Uffe
Hi, Ulf, On 22.05.2025 14:53, Ulf Hansson wrote: > On Thu, 22 May 2025 at 11:48, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >> >> Hi, Ulf, >> >> On 21.05.2025 17:57, Dmitry Torokhov wrote: >>> On Wed, May 21, 2025 at 02:37:08PM +0200, Ulf Hansson wrote: >>>> On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: >>>>> >>>>> Hi, Ulf, >>>>> >>>>> On 20.05.2025 15:09, Ulf Hansson wrote: >>>>>> For example, even if the order is made correctly, suppose a driver's >>>>>> ->remove() callback completes by turning off the resources for its >>>>>> device and leaves runtime PM enabled, as it relies on devres to do it >>>>>> some point later. Beyond this point, nothing would prevent userspace >>>>>> for runtime resuming/suspending the device via sysfs. >>>>> >>>>> If I'm not wrong, that can't happen? The driver_sysfs_remove() is called >>>>> before device_remove() (which calls the driver remove) is called, this >>>>> being the call path: >>>>> >>>>> device_driver_detach() -> >>>>> device_release_driver_internal() -> >>>>> __device_release_driver() -> >>>>> driver_sysfs_remove() >>>>> // ... >>>>> device_remove() >>>>> >>>>> And the driver_sysfs_remove() calls in the end __kernfs_remove() which >>>>> looks to me like the place that actually drops the entries from sysfs, this >>>>> being a call path for it: >>>>> >>>>> driver_sysfs_remove() -> >>>>> sysfs_remove_link() -> >>>>> kernfs_remove_by_name() -> >>>>> kernfs_remove_by_name_ns() -> >>>>> __kernfs_remove() -> >>>>> >>>>> activating the following line in __kernfs_remove(): >>>>> >>>>> pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); >>>>> >>>>> leads to the following prints when unbinding the watchdog device from its >>>>> watchdog driver (attached to platform bus) on my board: >>>>> https://p.fr33tux.org/935252 >>>> >>>> Indeed this is a very good point you make! I completely overlooked >>>> this fact, thanks a lot for clarifying this! >>>> >>>> However, my main point still stands. >>>> >>>> In the end, there is nothing preventing rpm_suspend|resume|idle() in >>>> drivers/base/power/runtime.c from running (don't forget runtime PM is >>>> asynchronous too) for the device in question. This could lead to that >>>> a ->runtime_suspend|resume|idle() callback becomes executed at any >>>> point in time, as long as we haven't called pm_runtime_disable() for >>>> the device. >>> >>> So exactly the same may happen if you enter driver->remove() and >>> something calls runtime API before pm_runtime_disable() is called. >>> The driver has (as they should be doing currently) be prepared for this. >> >> I took the time and tried to do a comparison of the current solutions >> (describing the bad and good things I see), trying to understand your >> concerns with regards to RPM suspend|resume|idle while unbinding a device >> from its driver. >> >> I see the following cases: >> >> Case 1/ the current approach when devm_pm_runtime_enable() is used in >> driver's ->probe() with the current code base: >> >> - right after driver ->remove() finish its execution clocks are detached >> from the PM domain, through dev_pm_domain_detach() call in >> platform_remove() >> >> - any subsequent RPM resume|suspend|idle will lead to failure if the driver >> specific RPM APIs access directly registers and counts on PM domain to >> enable/disable the clocks >> >> - at this point, if the IRQs are shared (but not only) and devm requested >> the driver's IRQ handler can still be called asynchronously; driver >> should be prepared for such events and should be written to work for such >> scenarios; but as the clocks are not in the PM domain anymore and RPM is >> still enabled at this point, if the driver don't run runtime suspend on >> probe (and runtime resume/suspend on runtime), I think (because I haven't >> investigated this yet) it can't rely on pm_runtime_active()/ >> pm_runtime_suspended() checks in interrupt handlers >> and can't decide if it can interrogate registers or not; interrogating >> should lead to failure at this stage as the clocks are disabled; drivers >> should work in such scenario and the CONFIG_DEBUG_SHIRQ is a way to check >> they can; I previously debugged a similar issue on drivers/net/ethernet/ >> renesas/ravb driver where using devm_pm_runtime_enable() in probe and >> pm_runtime_suspended() checks in IRQ handlers was the way to make this >> scenario happy; at that time I wasn't able to find that >> dev_pm_domain_detach() have the impact discussed in this thread >> >> Case 2/ What is proposed in this patch: devm_pm_runtime_enable() used + >> open devres group after dev_pm_domain_attach() (in probe) and close the >> devres group before dev_pm_domain_attach() (in remove): >> >> - right after the driver ->remove() is executed only the driver allocated >> devres resources are freed; this happens before dev_pm_domain_deattach() >> is called, though the proposed devres_release_group() call in this patch >> >> - while doing this, driver can still get async RPM suspend|resume|idle >> requests; is like the execution is in the driver ->remove() >> but the pm_runtime_disable() hasn't been called yet >> >> - as the runtime PM is enabled in driver's ->probe() mostly after the HW is >> prepared to take requests and all the other devm resources are allocated, >> the RPM disable is going to be among the first things to be called by the >> devres_release_group() >> >> - then, after RPM disable, all the devres resources allocated only in the >> driver's ->probe() are cleaned up in reverse order, just like >> device_unbind_cleanup() -> devres_release_all() call in >> __device_release_driver() is doing, but limited only to the resources >> allocated by the driver itself; I personally see this like manually >> allocating and freeing resources in the driver itself w/o relying on >> devres >> >> - then it comes the turn of dev_pm_domain_detach() call in >> platform_remove(): at the time dev_pm_domain_detach() is executed the >> runtime PM is disabled and all the devres resources allocated by driver >> are freed as well >> >> - after the dev_pm_domain_detach() is executed all the driver resources >> are cleaned up, the driver can't get IRQs as it's handler was already >> unregistered, no other user can execute rpm suspend|resume|idle >> as the RPM is disabled at this time >> >> Case 3/ devm_pm_runtime_enabled() dropped and replaced by manual cleanup: >> - the driver code is going be complicated, difficult to maintain and error >> prone > > Yes, the driver's code would become slightly more complicated, but > more importantly it would be correct. > > To me it sounds like the driver's ->remove() callback could do this: > > pm_runtime_get_sync() > pm_runtime_disable() > pm_runtime_put_noidle() In my case it was just pm_runtime_disable() at the end of driver ->remove() and pm_runtime_active() checks in IRQ handlers which didn't worked after driver ->remove() finished execution due to disable_depth being incremented in pm_runtime_disable(). The IRQs were devm requested. The solution found at the that time was to use devm_pm_runtime_enable() in probe and pm_runtime_suspended() calls in IRQ handlers. > > In this way, the driver will runtime resume its device, allowing > devres to drop/turn-off resources in the order we want. > Except for the > clocks, as those would be turned off via dev_pm_domain_detach() before > the IRQ handler is freed (via devres), right? > > To avoid getting the IRQ handler to be called when it can't access > registers, we could do one of the below: > *) Look for a condition in the IRQ handler and bail-out when we know > we should not manage IRQs. Is using pm_runtime_enabled() sufficient, > you think? Otherwise we need a driver specific flag, which should be > set in ->remove(). > *) Don't use devm* when registering the IRQ handler. That's true. > > Yes, both options further contribute to making the driver code > slightly more complicated, but if you want to solve the problem sooner > than later, I think this is what you need to do. Yet, I think there is > another option too, see below. > >> >> I may have missed considering things when describing the case 2 (which is >> what is proposed by this patch) as I don't have the full picture behind the >> dev_pm_domain_detach() call in platform bus remove. If so, please correct me. > > The dev_pm_domain_attach|detach() calls in bus level code > (probe/remove) were added there a long time ago, way before devres was > being used like today. > > Currently we also have devm_pm_domain_attach_list(), which is used > when devices have multiple PM domains to attach too. This is *not* > called by bus-level code, but by the driver themselves. For these > cases, we would not encounter the problems you have been facing with > clocks/IRQ-handler, I think - because the devres order is maintained > for PM domains too. > > That said, I think adding a devm_pm_domain_attach() interface would > make perfect sense. Then we can try to replace > dev_pm_domain_attach|detach() in bus level code, with just a call to > devm_pm_domain_attach(). In this way, we should preserve the > expectation for drivers around devres for PM domains. Even if it would > change the behaviour for some drivers, it still sounds like the > correct thing to do in my opinion. This looks good to me, as well. I did prototype it on my side and tested on all my failure cases and it works. Thank you, Claudiu
On Thu, 22 May 2025 at 16:08, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > Hi, Ulf, > > On 22.05.2025 14:53, Ulf Hansson wrote: > > On Thu, 22 May 2025 at 11:48, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > >> > >> Hi, Ulf, > >> > >> On 21.05.2025 17:57, Dmitry Torokhov wrote: > >>> On Wed, May 21, 2025 at 02:37:08PM +0200, Ulf Hansson wrote: > >>>> On Wed, 21 May 2025 at 07:41, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > >>>>> > >>>>> Hi, Ulf, > >>>>> > >>>>> On 20.05.2025 15:09, Ulf Hansson wrote: > >>>>>> For example, even if the order is made correctly, suppose a driver's > >>>>>> ->remove() callback completes by turning off the resources for its > >>>>>> device and leaves runtime PM enabled, as it relies on devres to do it > >>>>>> some point later. Beyond this point, nothing would prevent userspace > >>>>>> for runtime resuming/suspending the device via sysfs. > >>>>> > >>>>> If I'm not wrong, that can't happen? The driver_sysfs_remove() is called > >>>>> before device_remove() (which calls the driver remove) is called, this > >>>>> being the call path: > >>>>> > >>>>> device_driver_detach() -> > >>>>> device_release_driver_internal() -> > >>>>> __device_release_driver() -> > >>>>> driver_sysfs_remove() > >>>>> // ... > >>>>> device_remove() > >>>>> > >>>>> And the driver_sysfs_remove() calls in the end __kernfs_remove() which > >>>>> looks to me like the place that actually drops the entries from sysfs, this > >>>>> being a call path for it: > >>>>> > >>>>> driver_sysfs_remove() -> > >>>>> sysfs_remove_link() -> > >>>>> kernfs_remove_by_name() -> > >>>>> kernfs_remove_by_name_ns() -> > >>>>> __kernfs_remove() -> > >>>>> > >>>>> activating the following line in __kernfs_remove(): > >>>>> > >>>>> pr_debug("kernfs %s: removing\n", kernfs_rcu_name(kn)); > >>>>> > >>>>> leads to the following prints when unbinding the watchdog device from its > >>>>> watchdog driver (attached to platform bus) on my board: > >>>>> https://p.fr33tux.org/935252 > >>>> > >>>> Indeed this is a very good point you make! I completely overlooked > >>>> this fact, thanks a lot for clarifying this! > >>>> > >>>> However, my main point still stands. > >>>> > >>>> In the end, there is nothing preventing rpm_suspend|resume|idle() in > >>>> drivers/base/power/runtime.c from running (don't forget runtime PM is > >>>> asynchronous too) for the device in question. This could lead to that > >>>> a ->runtime_suspend|resume|idle() callback becomes executed at any > >>>> point in time, as long as we haven't called pm_runtime_disable() for > >>>> the device. > >>> > >>> So exactly the same may happen if you enter driver->remove() and > >>> something calls runtime API before pm_runtime_disable() is called. > >>> The driver has (as they should be doing currently) be prepared for this. > >> > >> I took the time and tried to do a comparison of the current solutions > >> (describing the bad and good things I see), trying to understand your > >> concerns with regards to RPM suspend|resume|idle while unbinding a device > >> from its driver. > >> > >> I see the following cases: > >> > >> Case 1/ the current approach when devm_pm_runtime_enable() is used in > >> driver's ->probe() with the current code base: > >> > >> - right after driver ->remove() finish its execution clocks are detached > >> from the PM domain, through dev_pm_domain_detach() call in > >> platform_remove() > >> > >> - any subsequent RPM resume|suspend|idle will lead to failure if the driver > >> specific RPM APIs access directly registers and counts on PM domain to > >> enable/disable the clocks > >> > >> - at this point, if the IRQs are shared (but not only) and devm requested > >> the driver's IRQ handler can still be called asynchronously; driver > >> should be prepared for such events and should be written to work for such > >> scenarios; but as the clocks are not in the PM domain anymore and RPM is > >> still enabled at this point, if the driver don't run runtime suspend on > >> probe (and runtime resume/suspend on runtime), I think (because I haven't > >> investigated this yet) it can't rely on pm_runtime_active()/ > >> pm_runtime_suspended() checks in interrupt handlers > >> and can't decide if it can interrogate registers or not; interrogating > >> should lead to failure at this stage as the clocks are disabled; drivers > >> should work in such scenario and the CONFIG_DEBUG_SHIRQ is a way to check > >> they can; I previously debugged a similar issue on drivers/net/ethernet/ > >> renesas/ravb driver where using devm_pm_runtime_enable() in probe and > >> pm_runtime_suspended() checks in IRQ handlers was the way to make this > >> scenario happy; at that time I wasn't able to find that > >> dev_pm_domain_detach() have the impact discussed in this thread > >> > >> Case 2/ What is proposed in this patch: devm_pm_runtime_enable() used + > >> open devres group after dev_pm_domain_attach() (in probe) and close the > >> devres group before dev_pm_domain_attach() (in remove): > >> > >> - right after the driver ->remove() is executed only the driver allocated > >> devres resources are freed; this happens before dev_pm_domain_deattach() > >> is called, though the proposed devres_release_group() call in this patch > >> > >> - while doing this, driver can still get async RPM suspend|resume|idle > >> requests; is like the execution is in the driver ->remove() > >> but the pm_runtime_disable() hasn't been called yet > >> > >> - as the runtime PM is enabled in driver's ->probe() mostly after the HW is > >> prepared to take requests and all the other devm resources are allocated, > >> the RPM disable is going to be among the first things to be called by the > >> devres_release_group() > >> > >> - then, after RPM disable, all the devres resources allocated only in the > >> driver's ->probe() are cleaned up in reverse order, just like > >> device_unbind_cleanup() -> devres_release_all() call in > >> __device_release_driver() is doing, but limited only to the resources > >> allocated by the driver itself; I personally see this like manually > >> allocating and freeing resources in the driver itself w/o relying on > >> devres > >> > >> - then it comes the turn of dev_pm_domain_detach() call in > >> platform_remove(): at the time dev_pm_domain_detach() is executed the > >> runtime PM is disabled and all the devres resources allocated by driver > >> are freed as well > >> > >> - after the dev_pm_domain_detach() is executed all the driver resources > >> are cleaned up, the driver can't get IRQs as it's handler was already > >> unregistered, no other user can execute rpm suspend|resume|idle > >> as the RPM is disabled at this time > >> > >> Case 3/ devm_pm_runtime_enabled() dropped and replaced by manual cleanup: > >> - the driver code is going be complicated, difficult to maintain and error > >> prone > > > > Yes, the driver's code would become slightly more complicated, but > > more importantly it would be correct. > > > > To me it sounds like the driver's ->remove() callback could do this: > > > > pm_runtime_get_sync() > > pm_runtime_disable() > > pm_runtime_put_noidle() > > In my case it was just pm_runtime_disable() at the end of driver ->remove() > and pm_runtime_active() checks in IRQ handlers which didn't worked after > driver ->remove() finished execution due to disable_depth being incremented > in pm_runtime_disable(). The IRQs were devm requested. > > The solution found at the that time was to use devm_pm_runtime_enable() in > probe and pm_runtime_suspended() calls in IRQ handlers. > > > > > In this way, the driver will runtime resume its device, allowing > > devres to drop/turn-off resources in the order we want. > > Except for the > > clocks, as those would be turned off via dev_pm_domain_detach() before > > the IRQ handler is freed (via devres), right? > > > > To avoid getting the IRQ handler to be called when it can't access > > registers, we could do one of the below: > > *) Look for a condition in the IRQ handler and bail-out when we know > > we should not manage IRQs. Is using pm_runtime_enabled() sufficient, > > you think? Otherwise we need a driver specific flag, which should be > > set in ->remove(). > > *) Don't use devm* when registering the IRQ handler. > > That's true. > > > > > Yes, both options further contribute to making the driver code > > slightly more complicated, but if you want to solve the problem sooner > > than later, I think this is what you need to do. Yet, I think there is > > another option too, see below. > > > >> > >> I may have missed considering things when describing the case 2 (which is > >> what is proposed by this patch) as I don't have the full picture behind the > >> dev_pm_domain_detach() call in platform bus remove. If so, please correct me. > > > > The dev_pm_domain_attach|detach() calls in bus level code > > (probe/remove) were added there a long time ago, way before devres was > > being used like today. > > > > Currently we also have devm_pm_domain_attach_list(), which is used > > when devices have multiple PM domains to attach too. This is *not* > > called by bus-level code, but by the driver themselves. For these > > cases, we would not encounter the problems you have been facing with > > clocks/IRQ-handler, I think - because the devres order is maintained > > for PM domains too. > > > > That said, I think adding a devm_pm_domain_attach() interface would > > make perfect sense. Then we can try to replace > > dev_pm_domain_attach|detach() in bus level code, with just a call to > > devm_pm_domain_attach(). In this way, we should preserve the > > expectation for drivers around devres for PM domains. Even if it would > > change the behaviour for some drivers, it still sounds like the > > correct thing to do in my opinion. > > This looks good to me, as well. I did prototype it on my side and tested on > all my failure cases and it works. That's great! I am happy to help review, if/when you decide to post it. Kind regards Uffe
On Thu, 22 May 2025 at 20:47, Dmitry Torokhov <dmitry.torokhov@gmail.com> wrote: > > On Thu, May 22, 2025 at 06:28:44PM +0200, Ulf Hansson wrote: > > On Thu, 22 May 2025 at 16:08, Claudiu Beznea <claudiu.beznea@tuxon.dev> wrote: > > > > > > Hi, Ulf, > > > > > > On 22.05.2025 14:53, Ulf Hansson wrote: > > > > > > > > That said, I think adding a devm_pm_domain_attach() interface would > > > > make perfect sense. Then we can try to replace > > > > dev_pm_domain_attach|detach() in bus level code, with just a call to > > > > devm_pm_domain_attach(). In this way, we should preserve the > > > > expectation for drivers around devres for PM domains. Even if it would > > > > change the behaviour for some drivers, it still sounds like the > > > > correct thing to do in my opinion. > > > > > > This looks good to me, as well. I did prototype it on my side and tested on > > > all my failure cases and it works. > > > > That's great! I am happy to help review, if/when you decide to post it. > > So you are saying you'd be OK with essentially the following (with > devm_pm_domain_attach() actually being elsewhere in a real patch and not > necessarily mimicked by devm_add_action_or_reset()): Correct! > > diff --git a/drivers/base/platform.c b/drivers/base/platform.c > index cfccf3ff36e7..1e017bfa5caf 100644 > --- a/drivers/base/platform.c > +++ b/drivers/base/platform.c > @@ -1376,6 +1376,27 @@ static int platform_uevent(const struct device *dev, struct kobj_uevent_env *env > return 0; > } > > + > +static void platform_pm_domain_detach(void *d) > +{ > + dev_pm_domain_detach(d, true); > +} Well, I would not limit this to the platform bus, even if that is the most widely used. Let's add the new generic interface along with dev_pm_domain_attach|detach* and friends instead. Then we can convert bus level code (and others), such as the platform bus to use it, in a step-by-step approach. > + > +static int devm_pm_domain_attach(struct device *dev) > +{ > + int error; > + > + error = dev_pm_domain_attach(dev, true); > + if (error) > + return error; > + > + error = devm_add_action_or_reset(dev, platform_pm_domain_detach, dev); > + if (error) > + return error; > + > + return 0; > +} > + > static int platform_probe(struct device *_dev) > { > struct platform_driver *drv = to_platform_driver(_dev->driver); > @@ -1396,15 +1417,12 @@ static int platform_probe(struct device *_dev) > if (ret < 0) > return ret; > > - ret = dev_pm_domain_attach(_dev, true); > + ret = devm_pm_domain_attach(_dev); > if (ret) > goto out; > > - if (drv->probe) { > + if (drv->probe) > ret = drv->probe(dev); > - if (ret) > - dev_pm_domain_detach(_dev, true); > - } > > out: > if (drv->prevent_deferred_probe && ret == -EPROBE_DEFER) { > @@ -1422,7 +1440,6 @@ static void platform_remove(struct device *_dev) > > if (drv->remove) > drv->remove(dev); > - dev_pm_domain_detach(_dev, true); > } > > static void platform_shutdown(struct device *_dev) > > > If so, then OK, it will work for me as well. This achieves the > same behavior as with using devres group. The only difference is that if > we ever need to extend the platform bus to acquire/release more > resources they will also have to use devm API and not the regular one. Sounds reasonable to me! Thanks for a nice discussion! When it comes to the devm_pm_runtime_enable() API, I think we seriously should consider removing it. Let me have a closer look at that. Kind regards Uffe
diff --git a/drivers/base/platform.c b/drivers/base/platform.c index 6f2a33722c52..1b64c4a44263 100644 --- a/drivers/base/platform.c +++ b/drivers/base/platform.c @@ -1401,9 +1401,15 @@ static int platform_probe(struct device *_dev) goto out; if (drv->probe) { + dev->devres_group_id = devres_open_group(&dev->dev, NULL, GFP_KERNEL); + if (!dev->devres_group_id) + return -ENOMEM; + ret = drv->probe(dev); - if (ret) + if (ret) { + devres_close_group(&dev->dev, dev->devres_group_id); dev_pm_domain_detach(_dev, true); + } } out: @@ -1422,6 +1428,8 @@ static void platform_remove(struct device *_dev) if (drv->remove) drv->remove(dev); + if (dev->devres_group_id) + devres_release_group(&dev->dev, dev->devres_group_id); dev_pm_domain_detach(_dev, true); } diff --git a/include/linux/platform_device.h b/include/linux/platform_device.h index 074754c23d33..e842ad243bef 100644 --- a/include/linux/platform_device.h +++ b/include/linux/platform_device.h @@ -40,6 +40,9 @@ struct platform_device { /* MFD cell pointer */ struct mfd_cell *mfd_cell; + /* ID of the probe devres group. */ + void *devres_group_id; + /* arch specific additions */ struct pdev_archdata archdata; };