pci: bus: only call of_platform_populate() if CONFIG_OF is defined

[PATCH] pci: bus: only call of_platform_populate() if CONFIG_OF is defined

Posted by Bert Karwatzki 1 year, 5 months ago

If of_platform_populate() is called when CONFIG_OF is not defined this
leads to spurious error messages of the following type:
 pci 0000:00:01.1: failed to populate child OF nodes (-19)
 pci 0000:00:02.1: failed to populate child OF nodes (-19)

Fixes: 8fb18619d910 ("PCI/pwrctl: Create platform devices for child OF nodes of the port node")
Signed-off-by: Bert Karwatzki <spasswolf@web.de>
---
 drivers/pci/bus.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
index e4735428814d..b363010664cd 100644
--- a/drivers/pci/bus.c
+++ b/drivers/pci/bus.c
@@ -350,6 +350,7 @@ void pci_bus_add_device(struct pci_dev *dev)

 	pci_dev_assign_added(dev, true);

+#ifdef CONFIG_OF
 	if (pci_is_bridge(dev)) {
 		retval = of_platform_populate(dev->dev.of_node, NULL, NULL,
 					      &dev->dev);
@@ -357,6 +358,7 @@ void pci_bus_add_device(struct pci_dev *dev)
 			pci_err(dev, "failed to populate child OF nodes (%d)\n",
 				retval);
 	}
+#endif
 }
 EXPORT_SYMBOL_GPL(pci_bus_add_device);

--
2.45.2

The mentioned error message occur on systems without CONFIG_OF, i.e.
x86_64. The call to of_platform_depopulate() in drivers/pci/remove.c
does not need #ifdef CONFIG_OF as the return value is not checked (it
will most likely be optimized away on platforms witout OF where the
of_platform_{de,}populate() functions just return -ENODEV)

Please CC me when replying, I'm not subscribed to the lists.

Bert Karwatzki

Re: [PATCH] pci: bus: only call of_platform_populate() if CONFIG_OF is defined

Posted by Bartosz Golaszewski 1 year, 5 months ago

On Wed, Jul 3, 2024 at 11:15 AM Bert Karwatzki <spasswolf@web.de> wrote:
>
> If of_platform_populate() is called when CONFIG_OF is not defined this
> leads to spurious error messages of the following type:
>  pci 0000:00:01.1: failed to populate child OF nodes (-19)
>  pci 0000:00:02.1: failed to populate child OF nodes (-19)
>
> Fixes: 8fb18619d910 ("PCI/pwrctl: Create platform devices for child OF nodes of the port node")
> Signed-off-by: Bert Karwatzki <spasswolf@web.de>
> ---
>  drivers/pci/bus.c | 2 ++
>  1 file changed, 2 insertions(+)
>
> diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
> index e4735428814d..b363010664cd 100644
> --- a/drivers/pci/bus.c
> +++ b/drivers/pci/bus.c
> @@ -350,6 +350,7 @@ void pci_bus_add_device(struct pci_dev *dev)
>
>         pci_dev_assign_added(dev, true);
>
> +#ifdef CONFIG_OF
>         if (pci_is_bridge(dev)) {
>                 retval = of_platform_populate(dev->dev.of_node, NULL, NULL,
>                                               &dev->dev);
> @@ -357,6 +358,7 @@ void pci_bus_add_device(struct pci_dev *dev)
>                         pci_err(dev, "failed to populate child OF nodes (%d)\n",
>                                 retval);
>         }
> +#endif
>  }
>  EXPORT_SYMBOL_GPL(pci_bus_add_device);
>
> --
> 2.45.2
>
> The mentioned error message occur on systems without CONFIG_OF, i.e.
> x86_64. The call to of_platform_depopulate() in drivers/pci/remove.c
> does not need #ifdef CONFIG_OF as the return value is not checked (it
> will most likely be optimized away on platforms witout OF where the
> of_platform_{de,}populate() functions just return -ENODEV)
>
> Please CC me when replying, I'm not subscribed to the lists.
>
> Bert Karwatzki

There's a better (less ifdeffery) fix on the list that I'll pick up
later today[1].

Bart

[1] https://lore.kernel.org/lkml/20240702180839.71491-1-superm1@kernel.org/T/

Re: [PATCH] pci: bus: only call of_platform_populate() if CONFIG_OF is defined

Posted by Lukas Wunner 1 year, 5 months ago

On Wed, Jul 03, 2024 at 11:32:05AM +0200, Bartosz Golaszewski wrote:
> On Wed, Jul 3, 2024 at 11:15 AM Bert Karwatzki <spasswolf@web.de> wrote:
> > If of_platform_populate() is called when CONFIG_OF is not defined this
> > leads to spurious error messages of the following type:
> >  pci 0000:00:01.1: failed to populate child OF nodes (-19)
> >  pci 0000:00:02.1: failed to populate child OF nodes (-19)
> >
> > Fixes: 8fb18619d910 ("PCI/pwrctl: Create platform devices for child OF nodes of the port node")
> > Signed-off-by: Bert Karwatzki <spasswolf@web.de>
> > ---
> >  drivers/pci/bus.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
> > index e4735428814d..b363010664cd 100644
> > --- a/drivers/pci/bus.c
> > +++ b/drivers/pci/bus.c
> > @@ -350,6 +350,7 @@ void pci_bus_add_device(struct pci_dev *dev)
> >
> >         pci_dev_assign_added(dev, true);
> >
> > +#ifdef CONFIG_OF
> >         if (pci_is_bridge(dev)) {

Per section 21 of Documentation/process/coding-style.rst,
IS_ENABLED() is strongly preferred to #ifdef.


> There's a better (less ifdeffery) fix on the list that I'll pick up
> later today[1].
> 
> [1] https://lore.kernel.org/lkml/20240702180839.71491-1-superm1@kernel.org/T/

That other fix doesn't feel very robust as it depends on
of_platform_populate() never returning -ENODEV in the
CONFIG_OF=y case.

Thanks,

Lukas

Re: [PATCH] pci: bus: only call of_platform_populate() if CONFIG_OF is defined

Posted by Bartosz Golaszewski 1 year, 5 months ago

On Sat, Jul 6, 2024 at 10:43 AM Lukas Wunner <lukas@wunner.de> wrote:
>
> On Wed, Jul 03, 2024 at 11:32:05AM +0200, Bartosz Golaszewski wrote:
> > On Wed, Jul 3, 2024 at 11:15 AM Bert Karwatzki <spasswolf@web.de> wrote:
> > > If of_platform_populate() is called when CONFIG_OF is not defined this
> > > leads to spurious error messages of the following type:
> > >  pci 0000:00:01.1: failed to populate child OF nodes (-19)
> > >  pci 0000:00:02.1: failed to populate child OF nodes (-19)
> > >
> > > Fixes: 8fb18619d910 ("PCI/pwrctl: Create platform devices for child OF nodes of the port node")
> > > Signed-off-by: Bert Karwatzki <spasswolf@web.de>
> > > ---
> > >  drivers/pci/bus.c | 2 ++
> > >  1 file changed, 2 insertions(+)
> > >
> > > diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
> > > index e4735428814d..b363010664cd 100644
> > > --- a/drivers/pci/bus.c
> > > +++ b/drivers/pci/bus.c
> > > @@ -350,6 +350,7 @@ void pci_bus_add_device(struct pci_dev *dev)
> > >
> > >         pci_dev_assign_added(dev, true);
> > >
> > > +#ifdef CONFIG_OF
> > >         if (pci_is_bridge(dev)) {
>
> Per section 21 of Documentation/process/coding-style.rst,
> IS_ENABLED() is strongly preferred to #ifdef.
>
>
> > There's a better (less ifdeffery) fix on the list that I'll pick up
> > later today[1].
> >
> > [1] https://lore.kernel.org/lkml/20240702180839.71491-1-superm1@kernel.org/T/
>
> That other fix doesn't feel very robust as it depends on
> of_platform_populate() never returning -ENODEV in the
> CONFIG_OF=y case.
>

If we even have to play these ifdef games then the stubs for
of_platform_populate() are broken and should probably return 0 with
!OF as otherwise the stubs themselves are useless.

Bart

Re: [PATCH] pci: bus: only call of_platform_populate() if CONFIG_OF is defined

Posted by Lukas Wunner 1 year, 5 months ago

On Sat, Jul 06, 2024 at 07:12:48PM +0200, Bartosz Golaszewski wrote:
> On Sat, Jul 6, 2024 at 10:43???AM Lukas Wunner <lukas@wunner.de> wrote:
> > On Wed, Jul 03, 2024 at 11:32:05AM +0200, Bartosz Golaszewski wrote:
> > > On Wed, Jul 3, 2024 at 11:15 AM Bert Karwatzki <spasswolf@web.de> wrote:
> > > > If of_platform_populate() is called when CONFIG_OF is not defined this
> > > > leads to spurious error messages of the following type:
> > > >  pci 0000:00:01.1: failed to populate child OF nodes (-19)
> > > >  pci 0000:00:02.1: failed to populate child OF nodes (-19)
[...]
> > > > --- a/drivers/pci/bus.c
> > > > +++ b/drivers/pci/bus.c
> > > > @@ -350,6 +350,7 @@ void pci_bus_add_device(struct pci_dev *dev)
> > > >
> > > >         pci_dev_assign_added(dev, true);
> > > >
> > > > +#ifdef CONFIG_OF
> > > >         if (pci_is_bridge(dev)) {
> > > 
> > > There's a better (less ifdeffery) fix on the list that I'll pick up
> > > later today[1].
> > >
> > > [1] https://lore.kernel.org/lkml/20240702180839.71491-1-superm1@kernel.org/T/
> >
> > That other fix doesn't feel very robust as it depends on
> > of_platform_populate() never returning -ENODEV in the
> > CONFIG_OF=y case.
> 
> If we even have to play these ifdef games then the stubs for
> of_platform_populate() are broken and should probably return 0 with
> !OF as otherwise the stubs themselves are useless.

The stub was introduced by commit 964dba283439 ("devicetree: Add empty
of_platform_populate() for !CONFIG_OF_ADDRESS (sparc)") some twelve
years ago.

From a brief look I'm under the impression that the commit used -ENODEV
on purpose (instead of 0).

There are 100 files in the kernel source tree referencing
of_platform_populate().  The majority is probably not compiled
in the CONFIG_OF=n case, nevertheless changing the stub risks
regressing some of those 100 callers unless great care is
observed.

Just making this conditional on IS_ENABLED(CONFIG_OF) is the
least risky option.

Thanks,

Lukas