Re: [Linux-mips-kernel]Re: PATCH: pci_auto bridge support

[Jun Sun <jsun@mvista.com>]

"Bradley D. LaRonde" wrote:
> 
> ----- Original Message -----
> From: "Jun Sun" <jsun@mvista.com>
> To: "Bradley D. LaRonde" <brad@ltc.com>
> Cc: <linux-mips@oss.sgi.com>; <linux-mips-kernel@lists.sourceforge.net>
> Sent: Monday, October 29, 2001 5:38 PM
> Subject: [Linux-mips-kernel]Re: PATCH: pci_auto bridge support
> 
> > "Bradley D. LaRonde" wrote:
> > >
> > > I considered that, but since only this small chuck of run-once surrogate
> > > bios autoconfig code needs to know, I figured better keep it separate.
> > >
> >
> > I would vote to put it inside the hose structure:
> >
> > . It makes a workaround look like a real fix. :-)
> >
> > . In other implementations of pci_auto, hose is the private sys data of a
> pci
> > dev. Having a bus number inside is very useful (e.g., pci_ops can tell
> whether
> > it is type0 of type1 configuration based on the bus number rather than a
> shaky
> > NULL parent bus pointer).  In the future, all pci_auto should be combined
> into
> > the pci driver.  So that is probably the right direction to go.
> >
> > I think hose may evolve to be the data structure that represents the
> topology
> > of PCI buses.  It should have more uses in the future (e.g., the standard
> IRQ
> > routing across PCI-PCI bridges).
> 
> Isn't the bus topology already adequately represented in the pci_dev and
> pci_channel structures?
> 

Not really.  For example, we don't know which bus is the sub-bus of which,
directly, and how their address space translate into each other.  Those data
structures are needed when we start to support dynamically mapped and/or
arbitrarily mapped PCI memory spaces.

> I look at the pci autoconfig stuff as a bios replacement.  The fact that we
> can use some of the same structures and functions to help us implement it is
> a bonus, but not a mandate to mess with the existing model.
>

That is a right point.  I might be too far ahead of myself. :-)
 
> Isn't Linux already handling PCI-PCI bridges and multiple PCI channles fine
> already, or has our autoconfig code exposed some existing non-arch-specific
> weakness?
>

I think on PCs, P2P still largely depends on BIOS.  (Correct me if I am
wrong).

There are some post-scan_bus() mechanism to setup P2P bridges.  I have not
looked into it closely. 

Jun