Missing ACPI PCI devices in acpi_pcidev_scan

[Grégoire Sutre <gregoire.sutre%labri.fr@localhost>]

Hi,

I'm trying to refactor an old ACPI display patch that I sent [1] in dec. 
2008, and that I've been using since (on netbsd-5).  The patch 
introduced helper functions to look for a given pci device in the ACPI 
tree.  But, as noted in [2], and I gladly admit it, the implementation 
was not efficient (it used too many ACPI walks without even caching 
results).  A new implementation now exists in functions acpi_pcidev_scan 
and acpi_pcidev_find of acpi_pci.c [3], and I would like to use them.
However, acpi_pcidev_scan does not detect all my PCI devices, in fact it 
only detects one: PCI0.  This is due to the fact that the function 
acpi_pcidev_scan discards devices that do not have the "_BBN" child 
integer value (PCI bus number).  IMHO, PCI devices are not required to 
have this "_BBN" child, except for PCI host bridges.  This is how I 
interpret the spec, and also what I observe on two PCs (with iasl).
It was previously argued in the list that there should be a unique ACPI 
namespace walk to build an in-kernel representation of the ACPI tree 
from which drivers can get all the information they need.  I like this 
idea, but I don't see how acpi_pcidev_scan can find the PCI bus number 
of PCI devices without the parent-child information of the ACPI tree, 
and AFAICS this information is currently missing in the in-kernel 
representation (struct acpi_devnode).  As AcpiWalkNamespace performs a 
DFS (with both pre and post order visitor callbacks), the parent-child 
information could be stored during the ACPI walk done by acpi_build_tree 
(in acpi.c), provided that struct acpi_devnode is modified to contain 
that information.  What do you think?
Thanks for your help,

Grégoire

[1] http://mail-index.netbsd.org/tech-kern/2008/12/11/msg003849.html
[2] http://mail-index.netbsd.org/tech-kern/2009/11/18/msg006507.html
[3] http://mail-index.netbsd.org/tech-kern/2009/11/28/msg006552.html