On Fri, Jun 13, 2025 at 03:47:58PM +0200, Herve Codina wrote:
> PCI drivers can use a device-tree overlay to describe the hardware
> available on the PCI board. This is the case, for instance, of the
> LAN966x PCI device driver.
>
> Adding some more nodes in the device-tree overlay adds some more
> consumer/supplier relationship between devices instantiated from this
> overlay.
>
> Those fw_node consumer/supplier relationships are handled by fw_devlink
> and are created based on the device-tree parsing done by the
> of_fwnode_add_links() function.
>
> Those consumer/supplier links are needed in order to ensure a correct PM
> runtime management and a correct removal order between devices.
>
> For instance, without those links a supplier can be removed before its
> consumers is removed leading to all kind of issue if this consumer still
> want the use the already removed supplier.
>
> The support for the usage of an overlay from a PCI driver has been added
> on x86 systems in commit 1f340724419ed ("PCI: of: Create device tree PCI
> host bridge node").
>
> In the past, support for fw_devlink on x86 had been tried but this
> support has been removed in commit 4a48b66b3f52 ("of: property: Disable
> fw_devlink DT support for X86"). Indeed, this support was breaking some
> x86 systems such as OLPC system and the regression was reported in [0].
>
> Instead of disabling this support for all x86 system, a first approach
> would be to use a finer grain and disable this support only for the
> possible problematic subset of x86 systems (at least OLPC and CE4100).
>
> This first approach could still leads to issues. Indeed, the list of
> possible problematic system and the way to identify them using Kconfig
> symbols is not well defined and so some system can be missed leading to
> kernel regressions on those missing systems.
>
> Use an other way and enable the support on x86 system only when this
> support is needed by some specific feature. The usage of a device-tree
> overlay by a PCI driver and thus the creation of PCI device-tree nodes
> is a feature that needs it.
>
> Signed-off-by: Herve Codina <herve.codina@xxxxxxxxxxx>
> Link: https://lore.kernel.org/lkml/3c1f2473-92ad-bfc4-258e-a5a08ad73dd0@xxxxxx/ [0]
> Reviewed-by: Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>
> ---
> drivers/of/property.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/of/property.c b/drivers/of/property.c
> index c1feb631e383..8b5cfee696e2 100644
> --- a/drivers/of/property.c
> +++ b/drivers/of/property.c
> @@ -1605,7 +1605,7 @@ static int of_fwnode_add_links(struct fwnode_handle *fwnode)
> const struct property *p;
> struct device_node *con_np = to_of_node(fwnode);
>
> - if (IS_ENABLED(CONFIG_X86))
> + if (IS_ENABLED(CONFIG_X86) && !IS_ENABLED(CONFIG_PCI_DYNAMIC_OF_NODES))
I really want CONFIG_PCI_DYNAMIC_OF_NODES to go away at some point, not
add more users.
I think this should instead check for specific platforms not with
kconfig symbols but DT properties. For ce4100, you can just check the
root compatible string. For OLPC, there isn't a root compatible (in the
DT I have). You could check for /architecture == OLPC instead. There's
some virtualization guests using DT now too. I would think their DT's
are simple enough to avoid any fw_devlink issues.
Alternatively, we could perhaps make x86 fw_devlink default off and then
enable it only when you create nodes. Maybe it has to be restricted a
sub tree of the DT to avoid any later interactions if devices are
unbound and rebound. Not a fully fleshed out idea...
Rob
