| Message ID | 20220929144618.1086985-2-marten.lindahl@axis.com |
|---|---|
| State | New |
| Headers | show |
| Series | [1/2] PM: runtime: Synchronize PM runtime enable state with parent | expand |
On Sun, Nov 6, 2022 at 4:33 PM Jonathan Cameron <jic23@kernel.org> wrote: > > On Mon, 31 Oct 2022 17:48:39 +0100 > Marten Lindahl <martenli@axis.com> wrote: > > > On Tue, Oct 25, 2022 at 06:20:10PM +0200, Rafael J. Wysocki wrote: > > > On Thu, Sep 29, 2022 at 4:46 PM Mårten Lindahl <marten.lindahl@axis.com> wrote: > > > > Hi! Thanks for your feedback! > > > > > > > > > > A device that creates a child character device with cdev_device_add by > > > > default create a PM sysfs group with power attributes for userspace > > > > control. This means that the power attributes monitors the child device > > > > only and thus does not reflect the parent device PM runtime behavior. > > > > > > It looks like device_set_pm_not_required() should be used on the child then. > > > > > > > But as the PM runtime framework (rpm_suspend/rpm_resume) operates not > > > > only on a single device that has been enabled for runtime PM, but also > > > > on its parent, it should be possible to synchronize the child and the > > > > parent so that the power attribute monitoring reflects the child and the > > > > parent as one. > > > > > > > > As an example, if an i2c_client device registers an iio_device, the > > > > iio_device will create sysfs power/attribute nodes for userspace > > > > control. But if the dev_pm_ops with resume/suspend callbacks is attached > > > > to the struct i2c_driver.driver.pm, the PM runtime needs to be enabled > > > > for the i2c_client device and not for the child iio_device. > > > > > > > > In this case PM runtime can be enabled for the i2c_client device and > > > > suspend/resume callbacks will be triggered, but the child sysfs power > > > > attributes will be visible but marked as 'unsupported' and can not be > > > > used for control or monitoring. This can be confusing as the sysfs > > > > device node presents the i2c_client and the iio_device as one device. > > > > > > I don't quite understand the last sentence. > > > > > > They are separate struct device objects and so they each have a > > > directory in sysfs, right? > > > > > > > Yes, they do have separate directories and if using device_set_pm_not_required > > on the child it will make it clearer which device is PM runtime regulated, so > > I guess that is what should be done. > > > > I think it all depends on where in sysfs the user accesses the device from. My > > point with these patches is that the iio_device may be perceived to be an > > iio device that should be possible to manually power control, as the power > > directory is visble. If looking at it from here: > > > > ~# ls /sys/bus/iio/devices/iio:device0/ > > in_illuminance_raw in_proximity_raw power > > in_illuminance_scale name subsystem > > in_proximity_nearlevel of_node uevent > > > > my idea is to let this power directory inherity the parent power control. But > > as you say, it is probably better to not create it at all, as the actual manual > > power control can be done here: > > > > ~# ls /sys/devices/platform/soc/.../i2c-2/2-0060/ > > driver modalias of_node subsystem > > iio:device1 name power uevent > > > > where it is more clear which device (the i2c parent) that can be power > > controlled. > > > > > > Add a function to synchronize the runtime PM enable state of a device > > > > with its parent. As there already exists a link from the child to its > > > > parent and both are enabled, all sysfs control/monitoring can reflect > > > > both devices, which from a userspace perspective makes more sense. > > > > > > Except that user space will be able to change "control" to "on" for > > > the parent alone AFAICS which still will be confusing. > > > > Yes, that is true. > > > > > > For devices that are pure software constructs it only makes sense to > > > expose the PM-runtime interface for them if the plan is to indirectly > > > control the parent's runtime PM through them. > > > > I will abandon this patchset and send a single patch for the iio device. > > I entirely agree with the statement that these are pointless and should not > be exposed. However I don't want to see a per device tweak. If we get > rid of these, we get rid of them for all of the iio:device0/ > devices (and the various other types of device IIO uses). > > The risk here is that, although pointless, some userspace is relying on the > ABI in sysfs. Do people thing it's worth the gamble of getting rid > of this non functioning interface for the whole of IIO? I think so. It is better to avoid exposing garbage to user space even if the scope of it is limited IMV. > So far I think this is only been done for a few similar cases > and for new subsystems.
On Sun, 6 Nov 2022 18:16:10 +0100 "Rafael J. Wysocki" <rafael@kernel.org> wrote: > On Sun, Nov 6, 2022 at 4:33 PM Jonathan Cameron <jic23@kernel.org> wrote: > > > > On Mon, 31 Oct 2022 17:48:39 +0100 > > Marten Lindahl <martenli@axis.com> wrote: > > > > > On Tue, Oct 25, 2022 at 06:20:10PM +0200, Rafael J. Wysocki wrote: > > > > On Thu, Sep 29, 2022 at 4:46 PM Mårten Lindahl <marten.lindahl@axis.com> wrote: > > > > > > Hi! Thanks for your feedback! > > > > > > > > > > > > > A device that creates a child character device with cdev_device_add by > > > > > default create a PM sysfs group with power attributes for userspace > > > > > control. This means that the power attributes monitors the child device > > > > > only and thus does not reflect the parent device PM runtime behavior. > > > > > > > > It looks like device_set_pm_not_required() should be used on the child then. > > > > > > > > > But as the PM runtime framework (rpm_suspend/rpm_resume) operates not > > > > > only on a single device that has been enabled for runtime PM, but also > > > > > on its parent, it should be possible to synchronize the child and the > > > > > parent so that the power attribute monitoring reflects the child and the > > > > > parent as one. > > > > > > > > > > As an example, if an i2c_client device registers an iio_device, the > > > > > iio_device will create sysfs power/attribute nodes for userspace > > > > > control. But if the dev_pm_ops with resume/suspend callbacks is attached > > > > > to the struct i2c_driver.driver.pm, the PM runtime needs to be enabled > > > > > for the i2c_client device and not for the child iio_device. > > > > > > > > > > In this case PM runtime can be enabled for the i2c_client device and > > > > > suspend/resume callbacks will be triggered, but the child sysfs power > > > > > attributes will be visible but marked as 'unsupported' and can not be > > > > > used for control or monitoring. This can be confusing as the sysfs > > > > > device node presents the i2c_client and the iio_device as one device. > > > > > > > > I don't quite understand the last sentence. > > > > > > > > They are separate struct device objects and so they each have a > > > > directory in sysfs, right? > > > > > > > > > > Yes, they do have separate directories and if using device_set_pm_not_required > > > on the child it will make it clearer which device is PM runtime regulated, so > > > I guess that is what should be done. > > > > > > I think it all depends on where in sysfs the user accesses the device from. My > > > point with these patches is that the iio_device may be perceived to be an > > > iio device that should be possible to manually power control, as the power > > > directory is visble. If looking at it from here: > > > > > > ~# ls /sys/bus/iio/devices/iio:device0/ > > > in_illuminance_raw in_proximity_raw power > > > in_illuminance_scale name subsystem > > > in_proximity_nearlevel of_node uevent > > > > > > my idea is to let this power directory inherity the parent power control. But > > > as you say, it is probably better to not create it at all, as the actual manual > > > power control can be done here: > > > > > > ~# ls /sys/devices/platform/soc/.../i2c-2/2-0060/ > > > driver modalias of_node subsystem > > > iio:device1 name power uevent > > > > > > where it is more clear which device (the i2c parent) that can be power > > > controlled. > > > > > > > > Add a function to synchronize the runtime PM enable state of a device > > > > > with its parent. As there already exists a link from the child to its > > > > > parent and both are enabled, all sysfs control/monitoring can reflect > > > > > both devices, which from a userspace perspective makes more sense. > > > > > > > > Except that user space will be able to change "control" to "on" for > > > > the parent alone AFAICS which still will be confusing. > > > > > > Yes, that is true. > > > > > > > > For devices that are pure software constructs it only makes sense to > > > > expose the PM-runtime interface for them if the plan is to indirectly > > > > control the parent's runtime PM through them. > > > > > > I will abandon this patchset and send a single patch for the iio device. > > > > I entirely agree with the statement that these are pointless and should not > > be exposed. However I don't want to see a per device tweak. If we get > > rid of these, we get rid of them for all of the iio:device0/ > > devices (and the various other types of device IIO uses). > > > > The risk here is that, although pointless, some userspace is relying on the > > ABI in sysfs. Do people thing it's worth the gamble of getting rid > > of this non functioning interface for the whole of IIO? > > I think so. > > It is better to avoid exposing garbage to user space even if the scope > of it is limited IMV. Marten, do you want to spin a patch doing this in the iio core? If not I'll do so sometime soon. Given where we are in the cycle, and the need to keep such a patch up for review for at least a few weeks, we can look to get it into next early next cycle and hopefully any issues will become apparent. Jonathan > > > So far I think this is only been done for a few similar cases > > and for new subsystems.
diff --git a/drivers/base/power/runtime.c b/drivers/base/power/runtime.c index 997be3ac20a7..b202d513684a 100644 --- a/drivers/base/power/runtime.c +++ b/drivers/base/power/runtime.c @@ -1499,6 +1499,24 @@ void pm_runtime_enable(struct device *dev) } EXPORT_SYMBOL_GPL(pm_runtime_enable); +/** + * pm_runtime_sync_parent_enable - Synchronize enable state. + * @dev: Device to handle. + * + * Synchronize the device enable state with its parent. + * NOTE: This function should only be used if the parent will never disable + * PM runtime (i.e. calling __pm_runtime_disable()) without telling its child. + */ +void pm_runtime_sync_parent_enable(struct device *dev) +{ + struct device *parent = dev->parent; + + if (parent && !parent->power.ignore_children + && pm_runtime_enabled(parent)) + pm_runtime_enable(dev); +} +EXPORT_SYMBOL_GPL(pm_runtime_sync_parent_enable); + static void pm_runtime_disable_action(void *data) { pm_runtime_dont_use_autosuspend(data); diff --git a/include/linux/pm_runtime.h b/include/linux/pm_runtime.h index 0a41b2dcccad..5054427adbd1 100644 --- a/include/linux/pm_runtime.h +++ b/include/linux/pm_runtime.h @@ -89,6 +89,7 @@ extern void pm_runtime_put_suppliers(struct device *dev); extern void pm_runtime_new_link(struct device *dev); extern void pm_runtime_drop_link(struct device_link *link); extern void pm_runtime_release_supplier(struct device_link *link); +extern void pm_runtime_sync_parent_enable(struct device *dev); extern int devm_pm_runtime_enable(struct device *dev); @@ -315,6 +316,7 @@ static inline void pm_runtime_put_suppliers(struct device *dev) {} static inline void pm_runtime_new_link(struct device *dev) {} static inline void pm_runtime_drop_link(struct device_link *link) {} static inline void pm_runtime_release_supplier(struct device_link *link) {} +static inline void pm_runtime_sync_parent_enable(struct device *dev) {} #endif /* !CONFIG_PM */
A device that creates a child character device with cdev_device_add by default create a PM sysfs group with power attributes for userspace control. This means that the power attributes monitors the child device only and thus does not reflect the parent device PM runtime behavior. But as the PM runtime framework (rpm_suspend/rpm_resume) operates not only on a single device that has been enabled for runtime PM, but also on its parent, it should be possible to synchronize the child and the parent so that the power attribute monitoring reflects the child and the parent as one. As an example, if an i2c_client device registers an iio_device, the iio_device will create sysfs power/attribute nodes for userspace control. But if the dev_pm_ops with resume/suspend callbacks is attached to the struct i2c_driver.driver.pm, the PM runtime needs to be enabled for the i2c_client device and not for the child iio_device. In this case PM runtime can be enabled for the i2c_client device and suspend/resume callbacks will be triggered, but the child sysfs power attributes will be visible but marked as 'unsupported' and can not be used for control or monitoring. This can be confusing as the sysfs device node presents the i2c_client and the iio_device as one device. Add a function to synchronize the runtime PM enable state of a device with its parent. As there already exists a link from the child to its parent and both are enabled, all sysfs control/monitoring can reflect both devices, which from a userspace perspective makes more sense. Signed-off-by: Mårten Lindahl <marten.lindahl@axis.com> --- drivers/base/power/runtime.c | 18 ++++++++++++++++++ include/linux/pm_runtime.h | 2 ++ 2 files changed, 20 insertions(+)