Re: [PATCH v3 3/9] PCI: of_property: Restore the arguments of the next level parent

On Thu, Aug 21, 2025 at 10:40:40AM GMT, Claudiu Beznea wrote:

Hi, Manivannan,

On 20.08.2025 20:47, Manivannan Sadhasivam wrote:

quoted

On Fri, Jul 04, 2025 at 07:14:03PM GMT, Claudiu wrote:

quoted

From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>

of_pci_make_dev_node() creates a device tree node for the PCIe bridge it
detects. The node name follows the format: pci_type@pci_slot,pci_func. If
such a node already exists in the current device tree, a new one is not
created.

When the node is created, its contents are populated with information from
the parent node. In the case of root complex nodes described in the device
tree, the created node duplicates the interrupt-map property. However, the
duplicated interrupt-map property does not correctly point to the next
interrupt controller.

For example, in the case of the Renesas RZ/G3S SoC, the resulting device
tree node is as follows (only relevant DT properties are shown):

pcie@11e40000 {

    // ...

    interrupt-map = <0x00 0x00 0x00 0x01 0x1f 0x00 0x00 0x00 0x00
                     0x00 0x00 0x00 0x02 0x1f 0x00 0x00 0x00 0x01
                     0x00 0x00 0x00 0x03 0x1f 0x00 0x00 0x00 0x02
                     0x00 0x00 0x00 0x04 0x1f 0x00 0x00 0x00 0x03>;
    interrupt-map-mask = <0x00 0x00 0x00 0x07>;
    interrupt-controller;
    #interrupt-cells = <0x01>;

    #address-cells = <0x03>;
    #size-cells = <0x02>;

    phandle = <0x1f>;

    // ...

    pci@0,0 {
        reg = <0x00 0x00 0x00 0x00 0x00>;
        interrupt-map = <0x10000 0x00 0x00 0x01 0x1f 0x00 0x11e40000 0x00 0x00
                         0x10000 0x00 0x00 0x02 0x1f 0x00 0x11e40000 0x00 0x01
                         0x10000 0x00 0x00 0x03 0x1f 0x00 0x11e40000 0x00 0x02
                         0x10000 0x00 0x00 0x04 0x1f 0x00 0x11e40000 0x00 0x03>;
        interrupt-map-mask = <0xffff00 0x00 0x00 0x07>;
        #interrupt-cells = <0x01>;

        #address-cells = <0x03>;
        #size-cells = <0x02>;

        // ...
    };
};

With this pci@0,0 node, the interrupt-map parsing code behaves as follows:

When a PCIe endpoint is enumerated and it requests to map a legacy
interrupt, of_irq_parse_raw() is called requesting the interrupt from
pci@0,0. If INTA is requested, of_irq_parse_raw() first matches:

interrupt-map = <0x10000 0x00 0x00 0x01 0x1f 0x00 0x11e40000 0x00 0x00>

from the pci@0,0 node. It then follows the phandle 0x1f to the interrupt
parent, looking for a mapping for interrupt ID 0x00
(0x00 0x11e40000 0x00 0x00). However, the root complex node does not
provide this mapping in its interrupt-map property, causing the interrupt
request to fail.

Are you trying to say that the generated bridge node incorrectly uses Root
Complex node as the interrupt parent?

I'm trying to say that the generated bridge node is wrong because it copies
the interrupt-map from the root complex mapping int 0x1 to 0x0 in the
bridge node, while it should have map the int 0x1 to something valid for
root complex mapping.

E.g. when some device requests INT with id 0x1 from bridge the bridge
mapping returns 0x0 then the returned 0x0 is used to find a new mapping on
the root complex based on what is provided for in with interrupt-map DT
property.

TBH, I don't know why it generates the 'interrupt-map' property for the bridge
node first place. It just creates two level lookup and in the absence, the IRQ
code will traverse up the node to find the interrupt parent anyway.

Maybe the intention was to avoid doing the traversal.

quoted

I'm getting confused since your example above shows '0x1f' as the interrupt
parent phandle for both Root Complex and bridge nodes. But I don't know to which
node this phandle corresponds to.

Root complex node from this patch description has:

phandle = <0x1f>;

Oops. I failed to spot it.

- Mani

-- 
மணிவண்ணன் சதாசிவம்