[PATCH v6 10/30] PCI: Introduce pci_host_bridge_list to manage host bridges

[bhelgaas@google.com (Bjorn Helgaas)]

On Thu, Mar 12, 2015 at 09:03:12PM +0800, Yijing Wang wrote:
> On 2015/3/12 10:55, Bjorn Helgaas wrote:
> > On Mon, Mar 09, 2015 at 10:34:07AM +0800, Yijing Wang wrote:
> >> Introduce pci_host_bridge_list to manage pci host
> >> bridges in system, so we could detect whether
> >> the host in domain:bus is alreay registered.
> >> Then we could remove bus alreay exist test in
> >> __pci_create_root_bus().
> > 
> > It's a nice idea to move this test into the core.  While you're at it, why
> > don't you check for any overlap with the bus ranges of existing host
> > bridges?  For example, if we're trying to create a new host bridge to
> > [bus 40-7f], it should conflict with existing bridges to [bus 00-7f]
> > as well as to [bus 40-ff].  I think your current patch will detect the
> > latter conflict but not the former.
> 
> Now pci host bridge may only know its start bus number, like acpi _BBN provided,
> but does not limit the end bus number, Eg. two pci roots report _BBN 0x0 and 0x80,
> so we have two bus number resource (0, 0xff) and (0x80, 0xff), if we check it strictly,
> some pci scan would fail which currently scan success.

_BBN is not the correct source for the bridge's bus number range.  There's
a comment in acpi_pci_root_add() that explains why:

  * We need both the start and end of the downstream bus range
  * to interpret _CBA (MMCONFIG base address), so it really is
  * supposed to be in _CRS.  If we don't find it there, all we
  * can do is assume [_BBN-0xFF] or [0-0xFF].

A platform SHOULD know the start and and end bus number.  If it doesn't I
think it's the platform's responsibility to carve up the bus number range.
Maybe this can be done by trimming the range of the [bus 00-ff] bridge when
we discover another bridge that leads to bus 80.