Re: [PATCH 08/11] PCI: mvebu: Use child_ops API

["Pali Rohár" <pali@kernel.org>]

On Wednesday 05 January 2022 09:41:51 Rob Herring wrote:
> > @@ -347,6 +375,11 @@ static int mvebu_pcie_hw_wr_conf(struct mvebu_pcie_port *port,
> >         return PCIBIOS_SUCCESSFUL;
> >  }
> >
> > +static struct pci_ops mvebu_pcie_child_ops = {
> > +       .read = mvebu_pcie_child_rd_conf,
> > +       .write = mvebu_pcie_child_wr_conf,
> > +};
> > +
> >  /*
> >   * Remove windows, starting from the largest ones to the smallest
> >   * ones.
> > @@ -862,25 +895,12 @@ static int mvebu_pcie_wr_conf(struct pci_bus *bus, u32 devfn,
> >  {
> >         struct mvebu_pcie *pcie = bus->sysdata;
> >         struct mvebu_pcie_port *port;
> > -       int ret;
> >
> >         port = mvebu_pcie_find_port(pcie, bus, devfn);
> >         if (!port)
> >                 return PCIBIOS_DEVICE_NOT_FOUND;
> 
> It would be nice to go from 'bus' to 'bridge' ptr directly, but I
> still had this in my version. I guess a standard RP struct as part of
> decoupling host bridges from RPs would solve this issue.

Hello!

The problem is somewhere else. This driver is misusing bus 0 for doing
non-transparent bus-bridging between multiple PCI domains by registering
roots ports across all domains into domain 0, bus 0. All details are
in this my patch which documents this strange driver behavior:
https://lore.kernel.org/linux-pci/20211125124605.25915-12-pali@kernel.org/

So the correct solution is is split these multidomain mixing and then
every domain would have exactly one root port (as it is designed in HW).