On Tue, Mar 04, 2025 at 02:31:46PM -0600, Bjorn Helgaas wrote:
> On Tue, Mar 04, 2025 at 03:12:11PM -0500, Frank Li wrote:
> > On Tue, Mar 04, 2025 at 01:08:16PM -0600, Bjorn Helgaas wrote:
> > > On Tue, Mar 04, 2025 at 12:49:36PM -0500, Frank Li wrote:
> > > > Remove artpec6_pcie_cpu_addr_fixup() as the DT bus fabric should provide correct
> > > > address translation. Set use_parent_dt_ranges to allow the DWC core driver to
> > > > fetch address translation from the device tree.
> > >
> > > Shouldn't we be able to detect platforms where DT doesn't describe the
> > > translation correctly? E.g., by running .cpu_addr_fixup() on a
> > > res.start value and comparing the result to the parent_bus_addr()?
> > > Then we could complain about it if they don't match.
> >
> > Can't detect because:
> >
> > There are case, driver have not provide .cpu_addr_fixup, but dts still be
> > wrong. such as
> >
> > bus@10000000
> > {
> > ranges = <0xdeaddead 0x1000000 size>;
> > pci@90000000 {
> >
> > reg = <...>, <0xdeaddead>;
> > reg-names = <...>, <config>;
> > }
> >
> > };
> >
> > above dts can work with current driver, but parent bus address 0xdeaddead
> > is totally fake address. We can't detect this case because no
> > .cpu_addr_fixup() at all.
>
> If there's no .cpu_addr_fixup(), primary-side ATU addresses must be
> identical to CPU addresses. If the DT parent bus address is
> different, can't we assume the DT is broken?
I think so.
Frank
