@@ -561,6 +561,20 @@ static struct pci_ops u4_pcie_pci_ops =
.write = u4_pcie_write_config,
};
+static void __devinit pmac_pci_fixup_u4_of_node(struct pci_dev *dev)
+{
+ /* Apple's device-tree "hides" the root complex virtual P2P bridge
+ * on U4. However, Linux sees it, causing the PCI <-> OF matching
+ * code to fail to properly match devices below it. This works around
+ * it by setting the node of the bridge to point to the PHB node,
+ * which is not entirely correct but fixes the matching code and
+ * doesn't break anything else. It's also the simplest possible fix.
+ */
+ if (dev->dev.of_node == NULL)
+ dev->dev.of_node = pcibios_get_phb_of_node(dev->bus);
+}
+DECLARE_PCI_FIXUP_HEADER(PCI_VENDOR_ID_APPLE, 0x5b, pmac_pci_fixup_u4_of_node);
+
#endif /* CONFIG_PPC64 */
#ifdef CONFIG_PPC32
Apple Quad G5 has some oddity in it's device-tree which causes the new generic matching code to fail to relate nodes for PCI-E devices below U4 with their respective struct pci_dev. This breaks graphics on those machines among others. This fixes it using a quirk which copies the node pointer from the host bridge for the root complex, which makes the generic code work for the children afterward. Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org> --- Linus, please apply, this is a 3.1 regression. Thanks !