Re: [PATCH v12 05/13] PCI: dwc: Add dw_pcie_parent_bus_offset()

[Frank Li <Frank.li@nxp.com>]

On Tue, Mar 25, 2025 at 11:59:14PM +0530, Manivannan Sadhasivam wrote:
> On Mon, Mar 24, 2025 at 01:28:27PM -0500, Bjorn Helgaas wrote:
> > On Mon, Mar 24, 2025 at 10:48:23PM +0530, Manivannan Sadhasivam wrote:
> > > On Sat, Mar 15, 2025 at 03:15:40PM -0500, Bjorn Helgaas wrote:
> > > > From: Frank Li <Frank.Li@nxp.com>
> > > >
> > > > Return the offset from CPU physical address to the parent bus address of
> > > > the specified element of the devicetree 'reg' property.
> >
> > > > +resource_size_t dw_pcie_parent_bus_offset(struct dw_pcie *pci,
> > > > +					  const char *reg_name,
> > > > +					  resource_size_t cpu_phy_addr)
> > > > +{
> > >
> > > s/cpu_phy_addr/cpu_phys_addr/g
> >
> > Fixed, thanks!
> >
> > > > +	struct device *dev = pci->dev;
> > > > +	struct device_node *np = dev->of_node;
> > > > +	int index;
> > > > +	u64 reg_addr;
> > > > +
> > > > +	/* Look up reg_name address on parent bus */
> > >
> > > 'parent bus' is not accurate as the below code checks for the 'reg_name' in
> > > current PCI controller node.
> >
> > We want the address of "reg_name" on the node's primary side.  We've
> > been calling that the "parent bus address", I guess because it's the
> > address on the "parent bus" of the node.
> >
>
> Yeah, 'parent bus address' sounds bogus to me. 'ranges' property is described
> as:
>
> 	ranges = <child_addr parent_addr child_size>
>
> Here, child_addr refers to the PCIe host controller's view of the address space
> and parent_addr refers to the CPU's view of the address space.
>
> So the register address described in the PCIe controller node is not a 'parent
> bus address'.


All should be parent bus address. See Rob's comments

https://lore.kernel.org/imx/20240927221831.GA135061-robh@kernel.org/


bus{
	ranges = <child_addr parent_addr child_size>
	pcie {

		All address here, will be translated by bus's ranges, which
use 1:1 map if out of ranges by default.

		from pcie node (children node of bus) view, the 'child_addr'
		is parent node (bus)'s output address.

	}
}


bus may not only one layer to CPU.

bus1 {
	ranges = <...>

	bus2 {
		ranges = <...>

		bus3 {
			ranges = <...>

			All address here is parent's node (bus3)'s bus address
			So, 'parent bus address' means the parent node's
			output bus address.
		};
	};
};

Frank

>
> > I'm not sure what the best term is for this.  Do you have a
> > suggestion?
> >
>
> We are just extracting the offset between translated (cpu_phy_addr) and
> untranslated (reg_addr) addresses of a specific register. Maybe the function
> should just return the 'untranslated address' and the caller should compute the
> offset to make it simple?
>
> > If "parent bus address" is the wrong term, maybe we need to rename
> > dw_pcie_parent_bus_offset() itself?
> >
>
> Yes!
>
> > Currently we pass in cpu_phys_addr, but this function doesn't need it
> > except for the debug code added later.  I would really rather have
> > something like this in the callers:
> >
> >   pci->parent_bus_offset = pp->cfg0_base -
> >       dw_pcie_parent_bus_addr(pci, "config");
> >
>
> I agree. This should become, dw_pcie_get_untranslated_addr().
>
> > because then the offset is computed sort of at the same level where
> > it's used, and a grep for "cfg0_base" would find both the set and the
> > use and they would be easy to match up.
> >
> > > > +	index = of_property_match_string(np, "reg-names", reg_name);
> > > > +
> > > > +	if (index < 0) {
> > > > +		dev_err(dev, "No %s in devicetree \"reg\" property\n", reg_name);
> > >
> > > Both of these callers are checking for the existence of the
> > > 'reg_name' property before calling this API. So this check seems to
> > > be redundant (for now).
> >
> > True, but I don't see a way to enforce the caller checks.  I don't
> > like the idea of calling of_property_read_reg(np, index, ...) where we
> > have to look the caller to verify that "index" is valid.
> >
>
> Ok.
>
> - Mani
>
> --
> மணிவண்ணன் சதாசிவம்