| Message ID | 20210904000543.2019010-2-saravanak@google.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Geert Uytterhoeven |
| Headers | show |
| Series | Fix simple-bus issues with fw_devlink | expand |
Hi Saravana, Thanks for your patch! CC linux-pm, Lee (mfd) On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > fw_devlink could end up creating device links for bus only devices. > However, bus only devices don't get probed and can block probe() or > sync_state() [1] call backs of other devices. To avoid this, probe these > devices using the simple-pm-bus driver. > > However, there are instances of devices that are not simple buses (they > get probed by their specific drivers) that also list the "simple-bus" > (or other bus only compatible strings) in their compatible property to > automatically populate their child devices. We still want these devices > to get probed by their specific drivers. So, we make sure this driver > only probes devices that are only buses. Note that this can also be the case for buses declaring compatibility with "simple-pm-bus". However, at the moment, none of such device nodes in upstream DTS files have device-specific drivers. > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > Signed-off-by: Saravana Kannan <saravanak@google.com> > Tested-by: Saravana Kannan <saravanak@google.com> > --- a/drivers/bus/simple-pm-bus.c > +++ b/drivers/bus/simple-pm-bus.c > @@ -13,11 +13,26 @@ > #include <linux/platform_device.h> > #include <linux/pm_runtime.h> > > - > static int simple_pm_bus_probe(struct platform_device *pdev) > { > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > - struct device_node *np = pdev->dev.of_node; > + const struct device *dev = &pdev->dev; > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > + struct device_node *np = dev->of_node; > + const struct of_device_id *match; > + > + match = of_match_device(dev->driver->of_match_table, dev); > + > + /* > + * These are transparent bus devices (not simple-pm-bus matches) that > + * have their child nodes populated automatically. So, don't need to > + * do anything more. > + */ > + if (match && match->data) { > + if (of_property_match_string(np, "compatible", match->compatible) == 0) Does this work as expected? Having multiple compatible values in a device node does not guarantee there exist a separate driver for any of the device-specific compatible values. > + return 0; > + else > + return -ENODEV; So if we get here, as both branches use "return", we skip the pm_runtime_enable() and of_platform_populate() below: - of_platform_populate() is handled for these buses by of_platform_default_populate(), so that's OK, - I'm wondering if any of the simple-mfd sub-devices use Runtime PM, but currently fail to save power because pm_runtime_enable() is never called for the MFD container, just like with simple-bus... > + } > > dev_dbg(&pdev->dev, "%s\n", __func__); > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > static int simple_pm_bus_remove(struct platform_device *pdev) > { > + const void *data = of_device_get_match_data(&pdev->dev); > + > + if (data) > + return 0; > + > dev_dbg(&pdev->dev, "%s\n", __func__); > > pm_runtime_disable(&pdev->dev); > return 0; > } > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > + > static const struct of_device_id simple_pm_bus_of_match[] = { > { .compatible = "simple-pm-bus", }, > + { .compatible = "simple-bus", .data = ONLY_BUS }, > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > + { .compatible = "isa", .data = ONLY_BUS }, #ifdef CONFIG_ARM_AMBA ? > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > { /* sentinel */ } This is now (almost[*]) the same as of_default_bus_match_table[] in drivers/of/platform.c. Perhaps it can be shared? [*] Especially if "simple-pm-bus" and "simple-bus" would be treated the same. > }; > MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match); Gr{oetje,eeting}s, Geert
On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > Hi Saravana, > > Thanks for your patch! > > CC linux-pm, Lee (mfd) > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > fw_devlink could end up creating device links for bus only devices. > > However, bus only devices don't get probed and can block probe() or > > sync_state() [1] call backs of other devices. To avoid this, probe these > > devices using the simple-pm-bus driver. > > > > However, there are instances of devices that are not simple buses (they > > get probed by their specific drivers) that also list the "simple-bus" > > (or other bus only compatible strings) in their compatible property to > > automatically populate their child devices. We still want these devices > > to get probed by their specific drivers. So, we make sure this driver > > only probes devices that are only buses. > > Note that this can also be the case for buses declaring compatibility > with "simple-pm-bus". However, at the moment, none of such device > nodes in upstream DTS files have device-specific drivers. Not sure about mfd, but I want to make sure we don't confuse busses (which are typically added to a class) with these "simple bus" devices that are added to platform_bus. Also if these other buses are actually causing an issue, then then should implement their own stub driver or use try patch[2] if they are added to classes (devices on classes don't probe) [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > Tested-by: Saravana Kannan <saravanak@google.com> > > > --- a/drivers/bus/simple-pm-bus.c > > +++ b/drivers/bus/simple-pm-bus.c > > @@ -13,11 +13,26 @@ > > #include <linux/platform_device.h> > > #include <linux/pm_runtime.h> > > > > - > > static int simple_pm_bus_probe(struct platform_device *pdev) > > { > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > - struct device_node *np = pdev->dev.of_node; > > + const struct device *dev = &pdev->dev; > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > + struct device_node *np = dev->of_node; > > + const struct of_device_id *match; > > + > > + match = of_match_device(dev->driver->of_match_table, dev); > > + > > + /* > > + * These are transparent bus devices (not simple-pm-bus matches) that > > + * have their child nodes populated automatically. So, don't need to > > + * do anything more. > > + */ > > + if (match && match->data) { > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > Does this work as expected? Having multiple compatible values in a > device node does not guarantee there exist a separate driver for any > of the device-specific compatible values. Right, and if they are platform devices that are equivalent to simple-bus (meaning, they don't do anything in Linux and just have their devices populated) we can add those to this list too. > > > + return 0; > > + else > > + return -ENODEV; > > So if we get here, as both branches use "return", we skip the > pm_runtime_enable() and of_platform_populate() below: > - of_platform_populate() is handled for these buses by > of_platform_default_populate(), so that's OK, > - I'm wondering if any of the simple-mfd sub-devices use Runtime > PM, but currently fail to save power because pm_runtime_enable() > is never called for the MFD container, just like with simple-bus... But this doesn't affect simple-mfd though. > > > + } > > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > > > static int simple_pm_bus_remove(struct platform_device *pdev) > > { > > + const void *data = of_device_get_match_data(&pdev->dev); > > + > > + if (data) > > + return 0; > > + > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > pm_runtime_disable(&pdev->dev); > > return 0; > > } > > > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > > + > > static const struct of_device_id simple_pm_bus_of_match[] = { > > { .compatible = "simple-pm-bus", }, > > + { .compatible = "simple-bus", .data = ONLY_BUS }, > > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > > + { .compatible = "isa", .data = ONLY_BUS }, > > #ifdef CONFIG_ARM_AMBA ? Not needed? If CONFIG_ARM_AMBA isn't enabled, the device wouldn't be created in the first place. > > > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > > { /* sentinel */ } > > This is now (almost[*]) the same as of_default_bus_match_table[] > in drivers/of/platform.c. Perhaps it can be shared? I wanted this table to be expandable as needed. That's why I'm intentionally not using of_default_bus_match_table[]. > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > the same. They are not treated the same way. -Saravana > > > }; > > MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match); > > Gr{oetje,eeting}s, > > Geert > > -- > Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org > > In personal conversations with technical people, I call myself a hacker. But > when I'm talking to journalists I just say "programmer" or something like that. > -- Linus Torvalds
On Sat, 4 Sept 2021 at 02:05, Saravana Kannan <saravanak@google.com> wrote: > > fw_devlink could end up creating device links for bus only devices. > However, bus only devices don't get probed and can block probe() or > sync_state() [1] call backs of other devices. To avoid this, probe these > devices using the simple-pm-bus driver. > > However, there are instances of devices that are not simple buses (they > get probed by their specific drivers) that also list the "simple-bus" > (or other bus only compatible strings) in their compatible property to > automatically populate their child devices. We still want these devices > to get probed by their specific drivers. So, we make sure this driver > only probes devices that are only buses. > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > Signed-off-by: Saravana Kannan <saravanak@google.com> > Tested-by: Saravana Kannan <saravanak@google.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> I will run some tests as soon as I can and let you know the results. Kind regards Uffe > --- > drivers/bus/simple-pm-bus.c | 32 +++++++++++++++++++++++++++++--- > 1 file changed, 29 insertions(+), 3 deletions(-) > > diff --git a/drivers/bus/simple-pm-bus.c b/drivers/bus/simple-pm-bus.c > index 01a3d0cd08ed..3e086a9f34cb 100644 > --- a/drivers/bus/simple-pm-bus.c > +++ b/drivers/bus/simple-pm-bus.c > @@ -13,11 +13,26 @@ > #include <linux/platform_device.h> > #include <linux/pm_runtime.h> > > - > static int simple_pm_bus_probe(struct platform_device *pdev) > { > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > - struct device_node *np = pdev->dev.of_node; > + const struct device *dev = &pdev->dev; > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > + struct device_node *np = dev->of_node; > + const struct of_device_id *match; > + > + match = of_match_device(dev->driver->of_match_table, dev); > + > + /* > + * These are transparent bus devices (not simple-pm-bus matches) that > + * have their child nodes populated automatically. So, don't need to > + * do anything more. > + */ > + if (match && match->data) { > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > + return 0; > + else > + return -ENODEV; > + } > > dev_dbg(&pdev->dev, "%s\n", __func__); > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > static int simple_pm_bus_remove(struct platform_device *pdev) > { > + const void *data = of_device_get_match_data(&pdev->dev); > + > + if (data) > + return 0; > + > dev_dbg(&pdev->dev, "%s\n", __func__); > > pm_runtime_disable(&pdev->dev); > return 0; > } > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > + > static const struct of_device_id simple_pm_bus_of_match[] = { > { .compatible = "simple-pm-bus", }, > + { .compatible = "simple-bus", .data = ONLY_BUS }, > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > + { .compatible = "isa", .data = ONLY_BUS }, > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > { /* sentinel */ } > }; > MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match); > -- > 2.33.0.153.gba50c8fa24-goog >
On Tue, Sep 7, 2021 at 3:41 AM Ulf Hansson <ulf.hansson@linaro.org> wrote: > > On Sat, 4 Sept 2021 at 02:05, Saravana Kannan <saravanak@google.com> wrote: > > > > fw_devlink could end up creating device links for bus only devices. > > However, bus only devices don't get probed and can block probe() or > > sync_state() [1] call backs of other devices. To avoid this, probe these > > devices using the simple-pm-bus driver. > > > > However, there are instances of devices that are not simple buses (they > > get probed by their specific drivers) that also list the "simple-bus" > > (or other bus only compatible strings) in their compatible property to > > automatically populate their child devices. We still want these devices > > to get probed by their specific drivers. So, we make sure this driver > > only probes devices that are only buses. > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > Tested-by: Saravana Kannan <saravanak@google.com> > > Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> > > I will run some tests as soon as I can and let you know the results. Thanks. -Saravana > > Kind regards > Uffe > > > --- > > drivers/bus/simple-pm-bus.c | 32 +++++++++++++++++++++++++++++--- > > 1 file changed, 29 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/bus/simple-pm-bus.c b/drivers/bus/simple-pm-bus.c > > index 01a3d0cd08ed..3e086a9f34cb 100644 > > --- a/drivers/bus/simple-pm-bus.c > > +++ b/drivers/bus/simple-pm-bus.c > > @@ -13,11 +13,26 @@ > > #include <linux/platform_device.h> > > #include <linux/pm_runtime.h> > > > > - > > static int simple_pm_bus_probe(struct platform_device *pdev) > > { > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > - struct device_node *np = pdev->dev.of_node; > > + const struct device *dev = &pdev->dev; > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > + struct device_node *np = dev->of_node; > > + const struct of_device_id *match; > > + > > + match = of_match_device(dev->driver->of_match_table, dev); > > + > > + /* > > + * These are transparent bus devices (not simple-pm-bus matches) that > > + * have their child nodes populated automatically. So, don't need to > > + * do anything more. > > + */ > > + if (match && match->data) { > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > + return 0; > > + else > > + return -ENODEV; > > + } > > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > > > static int simple_pm_bus_remove(struct platform_device *pdev) > > { > > + const void *data = of_device_get_match_data(&pdev->dev); > > + > > + if (data) > > + return 0; > > + > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > pm_runtime_disable(&pdev->dev); > > return 0; > > } > > > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > > + > > static const struct of_device_id simple_pm_bus_of_match[] = { > > { .compatible = "simple-pm-bus", }, > > + { .compatible = "simple-bus", .data = ONLY_BUS }, > > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > > + { .compatible = "isa", .data = ONLY_BUS }, > > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > > { /* sentinel */ } > > }; > > MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match); > > -- > > 2.33.0.153.gba50c8fa24-goog > >
On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak@google.com> wrote: > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > > > Hi Saravana, > > > > Thanks for your patch! > > > > CC linux-pm, Lee (mfd) > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > > fw_devlink could end up creating device links for bus only devices. > > > However, bus only devices don't get probed and can block probe() or > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > devices using the simple-pm-bus driver. > > > > > > However, there are instances of devices that are not simple buses (they > > > get probed by their specific drivers) that also list the "simple-bus" > > > (or other bus only compatible strings) in their compatible property to > > > automatically populate their child devices. We still want these devices > > > to get probed by their specific drivers. So, we make sure this driver > > > only probes devices that are only buses. > > > > Note that this can also be the case for buses declaring compatibility > > with "simple-pm-bus". However, at the moment, none of such device > > nodes in upstream DTS files have device-specific drivers. > > Not sure about mfd, but I want to make sure we don't confuse busses > (which are typically added to a class) with these "simple bus" devices > that are added to platform_bus. Also if these other buses are actually > causing an issue, then then should implement their own stub driver or > use try patch[2] if they are added to classes (devices on classes > don't probe) > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > > --- a/drivers/bus/simple-pm-bus.c > > > +++ b/drivers/bus/simple-pm-bus.c > > > @@ -13,11 +13,26 @@ > > > #include <linux/platform_device.h> > > > #include <linux/pm_runtime.h> > > > > > > - > > > static int simple_pm_bus_probe(struct platform_device *pdev) > > > { > > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > > - struct device_node *np = pdev->dev.of_node; > > > + const struct device *dev = &pdev->dev; > > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > > + struct device_node *np = dev->of_node; > > > + const struct of_device_id *match; > > > + > > > + match = of_match_device(dev->driver->of_match_table, dev); > > > + > > > + /* > > > + * These are transparent bus devices (not simple-pm-bus matches) that > > > + * have their child nodes populated automatically. So, don't need to > > > + * do anything more. > > > + */ > > > + if (match && match->data) { > > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > > > Does this work as expected? Having multiple compatible values in a > > device node does not guarantee there exist a separate driver for any > > of the device-specific compatible values. > > Right, and if they are platform devices that are equivalent to > simple-bus (meaning, they don't do anything in Linux and just have > their devices populated) we can add those to this list too. I think this needs to be a list of compatibles we have drivers for instead. A more specific compatible that the OS doesn't understand shouldn't cause a change in behavior and adding one would. I expect it to be a short list. We are guaranteed that of_match_device() returns the best match in the match list, so we really just need 1 list here with a boolean to bail or not. > > > + return 0; > > > + else > > > + return -ENODEV; > > > > So if we get here, as both branches use "return", we skip the > > pm_runtime_enable() and of_platform_populate() below: > > - of_platform_populate() is handled for these buses by > > of_platform_default_populate(), so that's OK, > > - I'm wondering if any of the simple-mfd sub-devices use Runtime > > PM, but currently fail to save power because pm_runtime_enable() > > is never called for the MFD container, just like with simple-bus... > > But this doesn't affect simple-mfd though. > > > > > > + } > > > > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > > > > > static int simple_pm_bus_remove(struct platform_device *pdev) > > > { > > > + const void *data = of_device_get_match_data(&pdev->dev); > > > + > > > + if (data) > > > + return 0; > > > + > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > > > pm_runtime_disable(&pdev->dev); > > > return 0; > > > } > > > > > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > > > + > > > static const struct of_device_id simple_pm_bus_of_match[] = { > > > { .compatible = "simple-pm-bus", }, > > > + { .compatible = "simple-bus", .data = ONLY_BUS }, > > > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > > > + { .compatible = "isa", .data = ONLY_BUS }, > > > > #ifdef CONFIG_ARM_AMBA ? > > Not needed? If CONFIG_ARM_AMBA isn't enabled, the device wouldn't be > created in the first place. > > > > > > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > > > { /* sentinel */ } > > > > This is now (almost[*]) the same as of_default_bus_match_table[] > > in drivers/of/platform.c. Perhaps it can be shared? > > I wanted this table to be expandable as needed. That's why I'm > intentionally not using of_default_bus_match_table[]. > > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > > the same. > > They are not treated the same way. I think it would be better if they were. IOW, the core code stops descending into simple-bus, etc. nodes and they are populated here. Then we just get rid of of_default_bus_match_table. That could cause some issues with init ordering. As I recall the at91 gpio and pinctrl drivers are sensitive to this. The default call to of_platform_populate doesn't work on those systems because the devices get created later than when their machine specific call happens. It may have been a case of a parent probe assuming a child probe completed after of_platform_populate returns (also a problem for Qcom with DWC3). There's a fix for at91 somewhere in the git history after I broke it. I started trying to untangle things with at91, but never finished that. Rob
On Wed, Sep 8, 2021 at 5:16 PM Rob Herring <robh+dt@kernel.org> wrote: > > On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak@google.com> wrote: > > > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > > > > > Hi Saravana, > > > > > > Thanks for your patch! > > > > > > CC linux-pm, Lee (mfd) > > > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > > > fw_devlink could end up creating device links for bus only devices. > > > > However, bus only devices don't get probed and can block probe() or > > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > > devices using the simple-pm-bus driver. > > > > > > > > However, there are instances of devices that are not simple buses (they > > > > get probed by their specific drivers) that also list the "simple-bus" > > > > (or other bus only compatible strings) in their compatible property to > > > > automatically populate their child devices. We still want these devices > > > > to get probed by their specific drivers. So, we make sure this driver > > > > only probes devices that are only buses. > > > > > > Note that this can also be the case for buses declaring compatibility > > > with "simple-pm-bus". However, at the moment, none of such device > > > nodes in upstream DTS files have device-specific drivers. > > > > Not sure about mfd, but I want to make sure we don't confuse busses > > (which are typically added to a class) with these "simple bus" devices > > that are added to platform_bus. Also if these other buses are actually > > causing an issue, then then should implement their own stub driver or > > use try patch[2] if they are added to classes (devices on classes > > don't probe) > > > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > > > > --- a/drivers/bus/simple-pm-bus.c > > > > +++ b/drivers/bus/simple-pm-bus.c > > > > @@ -13,11 +13,26 @@ > > > > #include <linux/platform_device.h> > > > > #include <linux/pm_runtime.h> > > > > > > > > - > > > > static int simple_pm_bus_probe(struct platform_device *pdev) > > > > { > > > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > > > - struct device_node *np = pdev->dev.of_node; > > > > + const struct device *dev = &pdev->dev; > > > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > > > + struct device_node *np = dev->of_node; > > > > + const struct of_device_id *match; > > > > + > > > > + match = of_match_device(dev->driver->of_match_table, dev); > > > > + > > > > + /* > > > > + * These are transparent bus devices (not simple-pm-bus matches) that > > > > + * have their child nodes populated automatically. So, don't need to > > > > + * do anything more. > > > > + */ > > > > + if (match && match->data) { > > > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > > > > > Does this work as expected? Having multiple compatible values in a > > > device node does not guarantee there exist a separate driver for any > > > of the device-specific compatible values. > > > > Right, and if they are platform devices that are equivalent to > > simple-bus (meaning, they don't do anything in Linux and just have > > their devices populated) we can add those to this list too. > > I think this needs to be a list of compatibles we have drivers for > instead. I don't think a "denylist" (devices we shouldn't probe with this driver) would be a short list. As of today, literally any device that has children could add a "simple-bus" to the compatible property and get its child devices populated for free. If we use a denylist, we'll have to update it every time someone adds "simple-bus" as a general match to a DT node (new or otherwise) that isn't in the denylist. The list would blow up and be a maintenance headache. Also, a denylist won't capture any DT that isn't part of the kernel repo but depends on "simple-bus" to populate the device's child nodes. Keep in mind this could be true even for completely upstream drivers today. And on top of that, this will also break for downstream drivers and platforms in the development stage. The allowlist is much smaller and manageable. > A more specific compatible that the OS doesn't understand > shouldn't cause a change in behavior and adding one would. I think the amount of specific compatible strings that'll be added, but won't have drivers added to Linux AND would want to boot with Linux is much less likely than the amount of times we'd have to update a denylist. Also, if we do hit the cases you mention and we want those devices to get probed anyway, with my current allowlist approach, we could use "driver_override" to force this driver to match them. If you use a denylist like you said, there's no way you can get the simple-pm-bus to unbind and let the more specific driver to bind. > I expect it to be a short list. > > We are guaranteed that of_match_device() returns the best match in the > match list, so we really just need 1 list here with a boolean to bail > or not. > > > > > + return 0; > > > > + else > > > > + return -ENODEV; > > > > > > So if we get here, as both branches use "return", we skip the > > > pm_runtime_enable() and of_platform_populate() below: > > > - of_platform_populate() is handled for these buses by > > > of_platform_default_populate(), so that's OK, > > > - I'm wondering if any of the simple-mfd sub-devices use Runtime > > > PM, but currently fail to save power because pm_runtime_enable() > > > is never called for the MFD container, just like with simple-bus... > > > > But this doesn't affect simple-mfd though. > > > > > > > > > + } > > > > > > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > > > > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > > > > > > > static int simple_pm_bus_remove(struct platform_device *pdev) > > > > { > > > > + const void *data = of_device_get_match_data(&pdev->dev); > > > > + > > > > + if (data) > > > > + return 0; > > > > + > > > > dev_dbg(&pdev->dev, "%s\n", __func__); > > > > > > > > pm_runtime_disable(&pdev->dev); > > > > return 0; > > > > } > > > > > > > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > > > > + > > > > static const struct of_device_id simple_pm_bus_of_match[] = { > > > > { .compatible = "simple-pm-bus", }, > > > > + { .compatible = "simple-bus", .data = ONLY_BUS }, > > > > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > > > > + { .compatible = "isa", .data = ONLY_BUS }, > > > > > > #ifdef CONFIG_ARM_AMBA ? > > > > Not needed? If CONFIG_ARM_AMBA isn't enabled, the device wouldn't be > > created in the first place. > > > > > > > > > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > > > > { /* sentinel */ } > > > > > > This is now (almost[*]) the same as of_default_bus_match_table[] > > > in drivers/of/platform.c. Perhaps it can be shared? > > > > I wanted this table to be expandable as needed. That's why I'm > > intentionally not using of_default_bus_match_table[]. > > > > > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > > > the same. > > > > They are not treated the same way. > > I think it would be better if they were. IOW, the core code stops > descending into simple-bus, etc. nodes and they are populated here. > Then we just get rid of of_default_bus_match_table. Right, I would if we could. But we can't simply stop descending the simple-bus nodes in the core code because all the specific drivers that used to have their child devices populated automatically would stop working and would need to be updated to populate their child devices. And I'm sure there are a ton more downstream kernels and downstream DTs (that use upstream kernels) that we would break. If you really want to do that go for it, but I'd rather not do all this as part of trying to fix the issue Ulf reported that needs simple-bus only devices probed. > That could cause some issues with init ordering. As I recall the at91 > gpio and pinctrl drivers are sensitive to this. The default call to > of_platform_populate doesn't work on those systems because the devices > get created later than when their machine specific call happens. It > may have been a case of a parent probe assuming a child probe > completed after of_platform_populate returns (also a problem for Qcom > with DWC3). There's a fix for at91 somewhere in the git history after > I broke it. I started trying to untangle things with at91, but never > finished that. I think it'll cause a lot of issues if we stop descending simple-bus nodes in the core code. We're just scratching the surface here. -Saravana
On Sat, 4 Sept 2021 at 02:05, Saravana Kannan <saravanak@google.com> wrote: > > fw_devlink could end up creating device links for bus only devices. > However, bus only devices don't get probed and can block probe() or > sync_state() [1] call backs of other devices. To avoid this, probe these > devices using the simple-pm-bus driver. > > However, there are instances of devices that are not simple buses (they > get probed by their specific drivers) that also list the "simple-bus" > (or other bus only compatible strings) in their compatible property to > automatically populate their child devices. We still want these devices > to get probed by their specific drivers. So, we make sure this driver > only probes devices that are only buses. > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > Signed-off-by: Saravana Kannan <saravanak@google.com> > Tested-by: Saravana Kannan <saravanak@google.com> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Kind regards Uffe > --- > drivers/bus/simple-pm-bus.c | 32 +++++++++++++++++++++++++++++--- > 1 file changed, 29 insertions(+), 3 deletions(-) > > diff --git a/drivers/bus/simple-pm-bus.c b/drivers/bus/simple-pm-bus.c > index 01a3d0cd08ed..3e086a9f34cb 100644 > --- a/drivers/bus/simple-pm-bus.c > +++ b/drivers/bus/simple-pm-bus.c > @@ -13,11 +13,26 @@ > #include <linux/platform_device.h> > #include <linux/pm_runtime.h> > > - > static int simple_pm_bus_probe(struct platform_device *pdev) > { > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > - struct device_node *np = pdev->dev.of_node; > + const struct device *dev = &pdev->dev; > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > + struct device_node *np = dev->of_node; > + const struct of_device_id *match; > + > + match = of_match_device(dev->driver->of_match_table, dev); > + > + /* > + * These are transparent bus devices (not simple-pm-bus matches) that > + * have their child nodes populated automatically. So, don't need to > + * do anything more. > + */ > + if (match && match->data) { > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > + return 0; > + else > + return -ENODEV; > + } > > dev_dbg(&pdev->dev, "%s\n", __func__); > > @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) > > static int simple_pm_bus_remove(struct platform_device *pdev) > { > + const void *data = of_device_get_match_data(&pdev->dev); > + > + if (data) > + return 0; > + > dev_dbg(&pdev->dev, "%s\n", __func__); > > pm_runtime_disable(&pdev->dev); > return 0; > } > > +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ > + > static const struct of_device_id simple_pm_bus_of_match[] = { > { .compatible = "simple-pm-bus", }, > + { .compatible = "simple-bus", .data = ONLY_BUS }, > + { .compatible = "simple-mfd", .data = ONLY_BUS }, > + { .compatible = "isa", .data = ONLY_BUS }, > + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, > { /* sentinel */ } > }; > MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match); > -- > 2.33.0.153.gba50c8fa24-goog >
On Wed, Sep 8, 2021 at 7:58 PM Saravana Kannan <saravanak@google.com> wrote: > > On Wed, Sep 8, 2021 at 5:16 PM Rob Herring <robh+dt@kernel.org> wrote: > > > > On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > > > > > > > Hi Saravana, > > > > > > > > Thanks for your patch! > > > > > > > > CC linux-pm, Lee (mfd) > > > > > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > fw_devlink could end up creating device links for bus only devices. > > > > > However, bus only devices don't get probed and can block probe() or > > > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > > > devices using the simple-pm-bus driver. > > > > > > > > > > However, there are instances of devices that are not simple buses (they > > > > > get probed by their specific drivers) that also list the "simple-bus" > > > > > (or other bus only compatible strings) in their compatible property to > > > > > automatically populate their child devices. We still want these devices > > > > > to get probed by their specific drivers. So, we make sure this driver > > > > > only probes devices that are only buses. > > > > > > > > Note that this can also be the case for buses declaring compatibility > > > > with "simple-pm-bus". However, at the moment, none of such device > > > > nodes in upstream DTS files have device-specific drivers. > > > > > > Not sure about mfd, but I want to make sure we don't confuse busses > > > (which are typically added to a class) with these "simple bus" devices > > > that are added to platform_bus. Also if these other buses are actually > > > causing an issue, then then should implement their own stub driver or > > > use try patch[2] if they are added to classes (devices on classes > > > don't probe) > > > > > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > > > > > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > > > > > > --- a/drivers/bus/simple-pm-bus.c > > > > > +++ b/drivers/bus/simple-pm-bus.c > > > > > @@ -13,11 +13,26 @@ > > > > > #include <linux/platform_device.h> > > > > > #include <linux/pm_runtime.h> > > > > > > > > > > - > > > > > static int simple_pm_bus_probe(struct platform_device *pdev) > > > > > { > > > > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > > > > - struct device_node *np = pdev->dev.of_node; > > > > > + const struct device *dev = &pdev->dev; > > > > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > > > > + struct device_node *np = dev->of_node; > > > > > + const struct of_device_id *match; > > > > > + > > > > > + match = of_match_device(dev->driver->of_match_table, dev); > > > > > + > > > > > + /* > > > > > + * These are transparent bus devices (not simple-pm-bus matches) that > > > > > + * have their child nodes populated automatically. So, don't need to > > > > > + * do anything more. > > > > > + */ > > > > > + if (match && match->data) { > > > > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > > > > > > > Does this work as expected? Having multiple compatible values in a > > > > device node does not guarantee there exist a separate driver for any > > > > of the device-specific compatible values. > > > > > > Right, and if they are platform devices that are equivalent to > > > simple-bus (meaning, they don't do anything in Linux and just have > > > their devices populated) we can add those to this list too. > > > > I think this needs to be a list of compatibles we have drivers for > > instead. > > I don't think a "denylist" (devices we shouldn't probe with this > driver) would be a short list. As of today, literally any device that > has children could add a "simple-bus" to the compatible property and > get its child devices populated for free. If we use a denylist, we'll > have to update it every time someone adds "simple-bus" as a general > match to a DT node (new or otherwise) that isn't in the denylist. The > list would blow up and be a maintenance headache. > > Also, a denylist won't capture any DT that isn't part of the kernel > repo but depends on "simple-bus" to populate the device's child nodes. > Keep in mind this could be true even for completely upstream drivers > today. And on top of that, this will also break for downstream drivers > and platforms in the development stage. You've got this all backwards. The list would be compatibles for which they have their own driver. If they have their own driver, we know that because there is a driver in the kernel. If you have an out of tree bus driver, then I guess you shouldn't be claiming compatibility with 'simple-bus'. > The allowlist is much smaller and manageable. I count 140 cases of 'simple-bus' with another compatible. I find roughly 24 of those under drivers/ that have a driver (and look, there's at91 pinctrl/gpio). There's some more under arch/, but I'm not sure if they are drivers. This is what I ran: git grep -ho '".*", "simple-bus"' -- arch/ | cut -d' ' -f1 | grep -oE '".+"' | grep -v '"syscon"' | sort -u > buses.txt git grep -f buses.txt -- drivers/ I haven't looked at 'simple-mfd', but that was supposed to always mean 'no driver'. To put it another way, let's look at 3 possibilities: 'simple-bus' 'foo,has-no-driver', 'simple-bus' 'foo,has-a-driver', 'simple-bus' The first case is easy. The last 2 cases are not. We have no way to handle them differently without a list. > > A more specific compatible that the OS doesn't understand > > shouldn't cause a change in behavior and adding one would. Again, if a dt is modified from case 1 to case 2, there should not be a change in behavior. Presumably if Ulf changed his test in this way, it would again fail, right? > I think the amount of specific compatible strings that'll be added, > but won't have drivers added to Linux AND would want to boot with > Linux is much less likely than the amount of times we'd have to update > a denylist. > > Also, if we do hit the cases you mention and we want those devices to > get probed anyway, with my current allowlist approach, we could use > "driver_override" to force this driver to match them. If you use a > denylist like you said, there's no way you can get the simple-pm-bus > to unbind and let the more specific driver to bind. Where would we set "driver_override"? Aren't we going to end up with a driver_override list? [...] > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > > > > the same. > > > > > > They are not treated the same way. > > > > I think it would be better if they were. IOW, the core code stops > > descending into simple-bus, etc. nodes and they are populated here. > > Then we just get rid of of_default_bus_match_table. > > Right, I would if we could. But we can't simply stop descending the > simple-bus nodes in the core code because all the specific drivers > that used to have their child devices populated automatically would > stop working and would need to be updated to populate their child > devices. And I'm sure there are a ton more downstream kernels and > downstream DTs (that use upstream kernels) that we would break. > > If you really want to do that go for it, but I'd rather not do all > this as part of trying to fix the issue Ulf reported that needs > simple-bus only devices probed. Agreed. > > That could cause some issues with init ordering. As I recall the at91 > > gpio and pinctrl drivers are sensitive to this. The default call to > > of_platform_populate doesn't work on those systems because the devices > > get created later than when their machine specific call happens. It > > may have been a case of a parent probe assuming a child probe > > completed after of_platform_populate returns (also a problem for Qcom > > with DWC3). There's a fix for at91 somewhere in the git history after > > I broke it. I started trying to untangle things with at91, but never > > finished that. > > I think it'll cause a lot of issues if we stop descending simple-bus > nodes in the core code. We're just scratching the surface here. Perhaps, but if there's cases which are that fragile, we should fix them. Rob
On Thu, Sep 9, 2021 at 7:02 AM Rob Herring <robh+dt@kernel.org> wrote: > > On Wed, Sep 8, 2021 at 7:58 PM Saravana Kannan <saravanak@google.com> wrote: > > > > On Wed, Sep 8, 2021 at 5:16 PM Rob Herring <robh+dt@kernel.org> wrote: > > > > > > On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > > > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > > > > > > > > > Hi Saravana, > > > > > > > > > > Thanks for your patch! > > > > > > > > > > CC linux-pm, Lee (mfd) > > > > > > > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > > fw_devlink could end up creating device links for bus only devices. > > > > > > However, bus only devices don't get probed and can block probe() or > > > > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > > > > devices using the simple-pm-bus driver. > > > > > > > > > > > > However, there are instances of devices that are not simple buses (they > > > > > > get probed by their specific drivers) that also list the "simple-bus" > > > > > > (or other bus only compatible strings) in their compatible property to > > > > > > automatically populate their child devices. We still want these devices > > > > > > to get probed by their specific drivers. So, we make sure this driver > > > > > > only probes devices that are only buses. > > > > > > > > > > Note that this can also be the case for buses declaring compatibility > > > > > with "simple-pm-bus". However, at the moment, none of such device > > > > > nodes in upstream DTS files have device-specific drivers. > > > > > > > > Not sure about mfd, but I want to make sure we don't confuse busses > > > > (which are typically added to a class) with these "simple bus" devices > > > > that are added to platform_bus. Also if these other buses are actually > > > > causing an issue, then then should implement their own stub driver or > > > > use try patch[2] if they are added to classes (devices on classes > > > > don't probe) > > > > > > > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > > > > > > > > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > > > > > > > > --- a/drivers/bus/simple-pm-bus.c > > > > > > +++ b/drivers/bus/simple-pm-bus.c > > > > > > @@ -13,11 +13,26 @@ > > > > > > #include <linux/platform_device.h> > > > > > > #include <linux/pm_runtime.h> > > > > > > > > > > > > - > > > > > > static int simple_pm_bus_probe(struct platform_device *pdev) > > > > > > { > > > > > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > > > > > - struct device_node *np = pdev->dev.of_node; > > > > > > + const struct device *dev = &pdev->dev; > > > > > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > > > > > + struct device_node *np = dev->of_node; > > > > > > + const struct of_device_id *match; > > > > > > + > > > > > > + match = of_match_device(dev->driver->of_match_table, dev); > > > > > > + > > > > > > + /* > > > > > > + * These are transparent bus devices (not simple-pm-bus matches) that > > > > > > + * have their child nodes populated automatically. So, don't need to > > > > > > + * do anything more. > > > > > > + */ > > > > > > + if (match && match->data) { > > > > > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > > > > > > > > > Does this work as expected? Having multiple compatible values in a > > > > > device node does not guarantee there exist a separate driver for any > > > > > of the device-specific compatible values. > > > > > > > > Right, and if they are platform devices that are equivalent to > > > > simple-bus (meaning, they don't do anything in Linux and just have > > > > their devices populated) we can add those to this list too. > > > > > > I think this needs to be a list of compatibles we have drivers for > > > instead. > > > > I don't think a "denylist" (devices we shouldn't probe with this > > driver) would be a short list. As of today, literally any device that > > has children could add a "simple-bus" to the compatible property and > > get its child devices populated for free. If we use a denylist, we'll > > have to update it every time someone adds "simple-bus" as a general > > match to a DT node (new or otherwise) that isn't in the denylist. The > > list would blow up and be a maintenance headache. > > > > Also, a denylist won't capture any DT that isn't part of the kernel > > repo but depends on "simple-bus" to populate the device's child nodes. > > Keep in mind this could be true even for completely upstream drivers > > today. And on top of that, this will also break for downstream drivers > > and platforms in the development stage. > > You've got this all backwards. I don't think so. I got what you are saying. > The list would be compatibles for which > they have their own driver. If they have their own driver, we know > that because there is a driver in the kernel. If you have an out of > tree bus driver, then I guess you shouldn't be claiming compatibility > with 'simple-bus'. But it isn't just out of tree drivers though. Take for example (a random compatible string with a driver in upstream) "qcom,sdm845-gpucc" that has DT nodes in a couple of DTS files. It's perfectly valid today today to add child nodes under it and have the child nodes get populated by making the compatible property: "qcom,sdm845-gpucc", "simple-bus" The only thing not upstreamed in that case is the DT file. And I'm fairly certain our position is not "if your DT isn't in the kernel repo we won't support it". In this situation we have no way of knowing we need to add "qcom,sdm845-gpucc" to the "it has a driver, don't probe it" list. And even if they upstream it, this list is going to blow up. And we are breaking backward compatibility that can't be fixed without a kernel update (no runtime fixes possible). > > The allowlist is much smaller and manageable. > > I count 140 cases of 'simple-bus' with another compatible. I find > roughly 24 of those under drivers/ that have a driver (and look, > there's at91 pinctrl/gpio). There's some more under arch/, but I'm not > sure if they are drivers. This is what I ran: > > git grep -ho '".*", "simple-bus"' -- arch/ | cut -d' ' -f1 | grep -oE > '".+"' | grep -v '"syscon"' | sort -u > buses.txt > git grep -f buses.txt -- drivers/ > > I haven't looked at 'simple-mfd', but that was supposed to always mean > 'no driver'. > > To put it another way, let's look at 3 possibilities: > > 'simple-bus' > 'foo,has-no-driver', 'simple-bus' > 'foo,has-a-driver', 'simple-bus' > > The first case is easy. The last 2 cases are not. We have no way to > handle them differently without a list. > > > > A more specific compatible that the OS doesn't understand > > > shouldn't cause a change in behavior and adding one would. > > Again, if a dt is modified from case 1 to case 2, there should not be > a change in behavior. My point is that if DT is modified from 1 to 2, we'll need to add "foo,has-no-driver" to the list we'll maintain. That preserves backward compatibility with existing DTs and if someone modifies the DT it's more likely that they can modify the kernel too. See more below. > Presumably if Ulf changed his test in this way, > it would again fail, right? > > > > I think the amount of specific compatible strings that'll be added, > > but won't have drivers added to Linux AND would want to boot with > > Linux is much less likely than the amount of times we'd have to update > > a denylist. > > > > Also, if we do hit the cases you mention and we want those devices to > > get probed anyway, with my current allowlist approach, we could use > > "driver_override" to force this driver to match them. If you use a > > denylist like you said, there's no way you can get the simple-pm-bus > > to unbind and let the more specific driver to bind. > > Where would we set "driver_override"? Aren't we going to end up with a > driver_override list? I'm not saying we maintain a list in the kernel. I'm saying if someone changed their DT to add 'foo,has-no-driver', but hit the unlikely case where they can update the DT but not the kernel, they can still get the 'foo,has-no-driver' to probe by using the driver_override file for that device in sysfs. TLDR: they can still get it to work without having to modify the kernel. > > [...] > > > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > > > > > the same. > > > > > > > > They are not treated the same way. > > > > > > I think it would be better if they were. IOW, the core code stops > > > descending into simple-bus, etc. nodes and they are populated here. > > > Then we just get rid of of_default_bus_match_table. > > > > Right, I would if we could. But we can't simply stop descending the > > simple-bus nodes in the core code because all the specific drivers > > that used to have their child devices populated automatically would > > stop working and would need to be updated to populate their child > > devices. And I'm sure there are a ton more downstream kernels and > > downstream DTs (that use upstream kernels) that we would break. > > > > If you really want to do that go for it, but I'd rather not do all > > this as part of trying to fix the issue Ulf reported that needs > > simple-bus only devices probed. > > Agreed. I'm confused. So you are okay with this patch to merge now? > > > That could cause some issues with init ordering. As I recall the at91 > > > gpio and pinctrl drivers are sensitive to this. The default call to > > > of_platform_populate doesn't work on those systems because the devices > > > get created later than when their machine specific call happens. It > > > may have been a case of a parent probe assuming a child probe > > > completed after of_platform_populate returns (also a problem for Qcom > > > with DWC3). There's a fix for at91 somewhere in the git history after > > > I broke it. I started trying to untangle things with at91, but never > > > finished that. > > > > I think it'll cause a lot of issues if we stop descending simple-bus > > nodes in the core code. We're just scratching the surface here. > > Perhaps, but if there's cases which are that fragile, we should fix them. Sure, but I'm not sure I want to open that can of works, -Saravana
On Thu, Sep 9, 2021 at 1:11 PM Saravana Kannan <saravanak@google.com> wrote: > TLDR: Read the end first... > On Thu, Sep 9, 2021 at 7:02 AM Rob Herring <robh+dt@kernel.org> wrote: > > > > On Wed, Sep 8, 2021 at 7:58 PM Saravana Kannan <saravanak@google.com> wrote: > > > > > > On Wed, Sep 8, 2021 at 5:16 PM Rob Herring <robh+dt@kernel.org> wrote: > > > > > > > > On Tue, Sep 7, 2021 at 2:01 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > > > > > > On Mon, Sep 6, 2021 at 12:54 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote: > > > > > > > > > > > > Hi Saravana, > > > > > > > > > > > > Thanks for your patch! > > > > > > > > > > > > CC linux-pm, Lee (mfd) > > > > > > > > > > > > On Sat, Sep 4, 2021 at 2:05 AM Saravana Kannan <saravanak@google.com> wrote: > > > > > > > fw_devlink could end up creating device links for bus only devices. > > > > > > > However, bus only devices don't get probed and can block probe() or > > > > > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > > > > > devices using the simple-pm-bus driver. > > > > > > > > > > > > > > However, there are instances of devices that are not simple buses (they > > > > > > > get probed by their specific drivers) that also list the "simple-bus" > > > > > > > (or other bus only compatible strings) in their compatible property to > > > > > > > automatically populate their child devices. We still want these devices > > > > > > > to get probed by their specific drivers. So, we make sure this driver > > > > > > > only probes devices that are only buses. > > > > > > > > > > > > Note that this can also be the case for buses declaring compatibility > > > > > > with "simple-pm-bus". However, at the moment, none of such device > > > > > > nodes in upstream DTS files have device-specific drivers. > > > > > > > > > > Not sure about mfd, but I want to make sure we don't confuse busses > > > > > (which are typically added to a class) with these "simple bus" devices > > > > > that are added to platform_bus. Also if these other buses are actually > > > > > causing an issue, then then should implement their own stub driver or > > > > > use try patch[2] if they are added to classes (devices on classes > > > > > don't probe) > > > > > > > > > > [2] - https://lore.kernel.org/lkml/20210831224510.703253-1-saravanak@google.com/ > > > > > > > > > > > > > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > > > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > > > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > > > > > > > > > > --- a/drivers/bus/simple-pm-bus.c > > > > > > > +++ b/drivers/bus/simple-pm-bus.c > > > > > > > @@ -13,11 +13,26 @@ > > > > > > > #include <linux/platform_device.h> > > > > > > > #include <linux/pm_runtime.h> > > > > > > > > > > > > > > - > > > > > > > static int simple_pm_bus_probe(struct platform_device *pdev) > > > > > > > { > > > > > > > - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); > > > > > > > - struct device_node *np = pdev->dev.of_node; > > > > > > > + const struct device *dev = &pdev->dev; > > > > > > > + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); > > > > > > > + struct device_node *np = dev->of_node; > > > > > > > + const struct of_device_id *match; > > > > > > > + > > > > > > > + match = of_match_device(dev->driver->of_match_table, dev); > > > > > > > + > > > > > > > + /* > > > > > > > + * These are transparent bus devices (not simple-pm-bus matches) that > > > > > > > + * have their child nodes populated automatically. So, don't need to > > > > > > > + * do anything more. > > > > > > > + */ > > > > > > > + if (match && match->data) { > > > > > > > + if (of_property_match_string(np, "compatible", match->compatible) == 0) > > > > > > > > > > > > Does this work as expected? Having multiple compatible values in a > > > > > > device node does not guarantee there exist a separate driver for any > > > > > > of the device-specific compatible values. > > > > > > > > > > Right, and if they are platform devices that are equivalent to > > > > > simple-bus (meaning, they don't do anything in Linux and just have > > > > > their devices populated) we can add those to this list too. > > > > > > > > I think this needs to be a list of compatibles we have drivers for > > > > instead. > > > > > > I don't think a "denylist" (devices we shouldn't probe with this > > > driver) would be a short list. As of today, literally any device that > > > has children could add a "simple-bus" to the compatible property and > > > get its child devices populated for free. If we use a denylist, we'll > > > have to update it every time someone adds "simple-bus" as a general > > > match to a DT node (new or otherwise) that isn't in the denylist. The > > > list would blow up and be a maintenance headache. > > > > > > Also, a denylist won't capture any DT that isn't part of the kernel > > > repo but depends on "simple-bus" to populate the device's child nodes. > > > Keep in mind this could be true even for completely upstream drivers > > > today. And on top of that, this will also break for downstream drivers > > > and platforms in the development stage. > > > > You've got this all backwards. > > I don't think so. I got what you are saying. > > > The list would be compatibles for which > > they have their own driver. If they have their own driver, we know > > that because there is a driver in the kernel. If you have an out of > > tree bus driver, then I guess you shouldn't be claiming compatibility > > with 'simple-bus'. > > But it isn't just out of tree drivers though. Take for example (a > random compatible string with a driver in upstream) > "qcom,sdm845-gpucc" that has DT nodes in a couple of DTS files. It's > perfectly valid today today to add child nodes under it and have the > child nodes get populated by making the compatible property: > "qcom,sdm845-gpucc", "simple-bus" IMO, that is not valid and I'd reject adding 'simple-bus' to the schema (so we would know). Also, if we had only "qcom,sdm845-gpucc" to start with, then we probably have a driver already. > The only thing not upstreamed in that case is the DT file. And I'm > fairly certain our position is not "if your DT isn't in the kernel > repo we won't support it". In this situation we have no way of knowing > we need to add "qcom,sdm845-gpucc" to the "it has a driver, don't > probe it" list. And even if they upstream it, this list is going to > blow up. And we are breaking backward compatibility that can't be > fixed without a kernel update (no runtime fixes possible). > > > > The allowlist is much smaller and manageable. > > > > I count 140 cases of 'simple-bus' with another compatible. I find > > roughly 24 of those under drivers/ that have a driver (and look, > > there's at91 pinctrl/gpio). There's some more under arch/, but I'm not > > sure if they are drivers. This is what I ran: > > > > git grep -ho '".*", "simple-bus"' -- arch/ | cut -d' ' -f1 | grep -oE > > '".+"' | grep -v '"syscon"' | sort -u > buses.txt > > git grep -f buses.txt -- drivers/ > > > > I haven't looked at 'simple-mfd', but that was supposed to always mean > > 'no driver'. > > > > To put it another way, let's look at 3 possibilities: > > > > 'simple-bus' > > 'foo,has-no-driver', 'simple-bus' > > 'foo,has-a-driver', 'simple-bus' > > > > The first case is easy. The last 2 cases are not. We have no way to > > handle them differently without a list. > > > > > > A more specific compatible that the OS doesn't understand > > > > shouldn't cause a change in behavior and adding one would. > > > > Again, if a dt is modified from case 1 to case 2, there should not be > > a change in behavior. > > My point is that if DT is modified from 1 to 2, we'll need to add > "foo,has-no-driver" to the list we'll maintain. That preserves > backward compatibility with existing DTs and if someone modifies the > DT it's more likely that they can modify the kernel too. See more > below. I assume this list is just cases of I need to bind the simple-bus driver for devlinks, right? Being able to update the kernel is not a valid assumption. While the kernel sources may be updated, it is perfectly valid to update a dtb and not the kernel. This is what SUSE does (or was doing?). DTBs from the latest kernel (because that's the most complete h/w description, right?) and an LTS kernel. Think about updating your PC's OS and firmware. Do you want updating either of those independently to work? Probably so. It's the same thing with kernel and dtb(in firmware) updates. While specific platforms or environments (Android) may not care, in general we have to. However, we can cheat a little bit if the kernel update goes to stable kernels, but that should be the exception and I don't think we should rely on that. Would you want your PC to stop booting due to a firmware update without a stable kernel update? > > Presumably if Ulf changed his test in this way, > > it would again fail, right? > > > > > > > I think the amount of specific compatible strings that'll be added, > > > but won't have drivers added to Linux AND would want to boot with > > > Linux is much less likely than the amount of times we'd have to update > > > a denylist. > > > > > > Also, if we do hit the cases you mention and we want those devices to > > > get probed anyway, with my current allowlist approach, we could use > > > "driver_override" to force this driver to match them. If you use a > > > denylist like you said, there's no way you can get the simple-pm-bus > > > to unbind and let the more specific driver to bind. > > > > Where would we set "driver_override"? Aren't we going to end up with a > > driver_override list? > > I'm not saying we maintain a list in the kernel. I'm saying if someone > changed their DT to add 'foo,has-no-driver', but hit the unlikely case > where they can update the DT but not the kernel, they can still get > the 'foo,has-no-driver' to probe by using the driver_override file for > that device in sysfs. TLDR: they can still get it to work without > having to modify the kernel. That doesn't seem like a great experience. How's that going to work if you didn't boot up enough to get to sysfs? > > [...] > > > > > > > > [*] Especially if "simple-pm-bus" and "simple-bus" would be treated > > > > > > the same. > > > > > > > > > > They are not treated the same way. > > > > > > > > I think it would be better if they were. IOW, the core code stops > > > > descending into simple-bus, etc. nodes and they are populated here. > > > > Then we just get rid of of_default_bus_match_table. > > > > > > Right, I would if we could. But we can't simply stop descending the > > > simple-bus nodes in the core code because all the specific drivers > > > that used to have their child devices populated automatically would > > > stop working and would need to be updated to populate their child > > > devices. And I'm sure there are a ton more downstream kernels and > > > downstream DTs (that use upstream kernels) that we would break. > > > > > > If you really want to do that go for it, but I'd rather not do all > > > this as part of trying to fix the issue Ulf reported that needs > > > simple-bus only devices probed. > > > > Agreed. > > I'm confused. So you are okay with this patch to merge now? No, I'm okay with the separate step of removing descending 'simple-bus', etc. nodes out of of_platform_populate() to do it in the simple-bus driver. Backing up... The problem here is letting what 'simple-bus' (and simple-mfd) means change with the presence of another compatible and we can't determine if it changes or not. Has a driver or not doesn't really work. I think the only thing that works is 'simple-bus' can only mean child nodes have no dependency on the simple-bus node. If there is a dependency, then it is not a simple-bus. We might have some cases that violated this definition and they've gotten lucky so far. I think we just have to fix these cases as they arise. That's at least bounded by cases where we already have drivers and shouldn't have new cases introduced. I still like the idea that 'simple-bus' always binds to a driver, but I think for that to work we've got to support rebinding to a new driver when a better matching driver is found. We're never going to know for sure if/when there is a better match. There's a couple of cases where we need this. Rob
On Thu, 9 Sept 2021 at 13:01, Ulf Hansson <ulf.hansson@linaro.org> wrote: > > On Sat, 4 Sept 2021 at 02:05, Saravana Kannan <saravanak@google.com> wrote: > > > > fw_devlink could end up creating device links for bus only devices. > > However, bus only devices don't get probed and can block probe() or > > sync_state() [1] call backs of other devices. To avoid this, probe these > > devices using the simple-pm-bus driver. > > > > However, there are instances of devices that are not simple buses (they > > get probed by their specific drivers) that also list the "simple-bus" > > (or other bus only compatible strings) in their compatible property to > > automatically populate their child devices. We still want these devices > > to get probed by their specific drivers. So, we make sure this driver > > only probes devices that are only buses. > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > Tested-by: Saravana Kannan <saravanak@google.com> > > Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Saravana, Rob, I have been following your latest discussion in this thread - and it looks like you guys are moving towards a consensus. Although, if there is anything I can do to help to complete this, just tell me and I will jump in immediately. [...] Kind regards Uffe
On Fri, Sep 24, 2021 at 4:50 AM Ulf Hansson <ulf.hansson@linaro.org> wrote: > > On Thu, 9 Sept 2021 at 13:01, Ulf Hansson <ulf.hansson@linaro.org> wrote: > > > > On Sat, 4 Sept 2021 at 02:05, Saravana Kannan <saravanak@google.com> wrote: > > > > > > fw_devlink could end up creating device links for bus only devices. > > > However, bus only devices don't get probed and can block probe() or > > > sync_state() [1] call backs of other devices. To avoid this, probe these > > > devices using the simple-pm-bus driver. > > > > > > However, there are instances of devices that are not simple buses (they > > > get probed by their specific drivers) that also list the "simple-bus" > > > (or other bus only compatible strings) in their compatible property to > > > automatically populate their child devices. We still want these devices > > > to get probed by their specific drivers. So, we make sure this driver > > > only probes devices that are only buses. > > > > > > [1] - https://lore.kernel.org/lkml/CAPDyKFo9Bxremkb1dDrr4OcXSpE0keVze94Cm=zrkOVxHHxBmQ@mail.gmail.com/ > > > Signed-off-by: Saravana Kannan <saravanak@google.com> > > > Tested-by: Saravana Kannan <saravanak@google.com> > > > > Tested-by: Ulf Hansson <ulf.hansson@linaro.org> > > Saravana, Rob, > > I have been following your latest discussion in this thread - and it > looks like you guys are moving towards a consensus. > > Although, if there is anything I can do to help to complete this, just > tell me and I will jump in immediately. Looks like we settled on the allow list approach during the BoF in LPC. I'll send out a v4 with some tweaks. -Saravana
diff --git a/drivers/bus/simple-pm-bus.c b/drivers/bus/simple-pm-bus.c index 01a3d0cd08ed..3e086a9f34cb 100644 --- a/drivers/bus/simple-pm-bus.c +++ b/drivers/bus/simple-pm-bus.c @@ -13,11 +13,26 @@ #include <linux/platform_device.h> #include <linux/pm_runtime.h> - static int simple_pm_bus_probe(struct platform_device *pdev) { - const struct of_dev_auxdata *lookup = dev_get_platdata(&pdev->dev); - struct device_node *np = pdev->dev.of_node; + const struct device *dev = &pdev->dev; + const struct of_dev_auxdata *lookup = dev_get_platdata(dev); + struct device_node *np = dev->of_node; + const struct of_device_id *match; + + match = of_match_device(dev->driver->of_match_table, dev); + + /* + * These are transparent bus devices (not simple-pm-bus matches) that + * have their child nodes populated automatically. So, don't need to + * do anything more. + */ + if (match && match->data) { + if (of_property_match_string(np, "compatible", match->compatible) == 0) + return 0; + else + return -ENODEV; + } dev_dbg(&pdev->dev, "%s\n", __func__); @@ -31,14 +46,25 @@ static int simple_pm_bus_probe(struct platform_device *pdev) static int simple_pm_bus_remove(struct platform_device *pdev) { + const void *data = of_device_get_match_data(&pdev->dev); + + if (data) + return 0; + dev_dbg(&pdev->dev, "%s\n", __func__); pm_runtime_disable(&pdev->dev); return 0; } +#define ONLY_BUS ((void *) 1) /* Match if the device is only a bus. */ + static const struct of_device_id simple_pm_bus_of_match[] = { { .compatible = "simple-pm-bus", }, + { .compatible = "simple-bus", .data = ONLY_BUS }, + { .compatible = "simple-mfd", .data = ONLY_BUS }, + { .compatible = "isa", .data = ONLY_BUS }, + { .compatible = "arm,amba-bus", .data = ONLY_BUS }, { /* sentinel */ } }; MODULE_DEVICE_TABLE(of, simple_pm_bus_of_match);