Re: Patch: 2.5.45 PCI Fixups for PCI HotPlug

[Ivan Kokshaysky <ink@jurassic.park.msu.ru>]

On Mon, Nov 04, 2002 at 12:29:37PM -0800, Adam J. Richter wrote:
> 	What are you advocating?  Do you want all quirk routines
> marked __devinit?  Wouldn't you agree at least that if we can
> establish that device *cannot* be hot plugged, that its quirk routine
> can be __init?

You cannot mark individual quirk routines differently as long as they
belong in the same quirk list. If the list is __devinitdata and some
of routines in it are __init, you'll have an oops in the hotplug path.

What we need is an additional quirk list, say, "hotplug_pci_fixups"
and a global flag "init_gone" (probably free_initmem() should set it).
Then we'll have

void pci_fixup_device(int pass, struct pci_dev *dev)
{
-	pci_do_fixups(dev, pass, pcibios_fixups);
-	pci_do_fixups(dev, pass, pci_fixups);
+	if (!init_gone) {
+		pci_do_fixups(dev, pass, pcibios_fixups);
+		pci_do_fixups(dev, pass, pci_fixups);
+	}
+	pci_do_fixups(dev, pass, hotplug_pci_fixups);
}

Perhaps arch specific "hotplug_pcibios_fixups" is also needed, as archs
may work around the same pci bug differently.

What should go into the new list is another story. Obviously the fixups for
host bridges must be __init, as well as some south bridges that look like
pci devices but actually aren't.

Ivan.