Hi Kishon, On Wed, Nov 27, 2019 at 5:45 AM Kishon Vijay Abraham I <kishon@xxxxxx> wrote: > > Hi, > > On 13/11/19 9:38 AM, Rob Herring wrote: > > On Thu, Nov 07, 2019 at 09:08:35PM +0100, Geert Uytterhoeven wrote: > >> Hi Prabhakar, > >> > >> On Thu, Nov 7, 2019 at 10:26 AM Lad, Prabhakar > >> <prabhakar.csengg@xxxxxxxxx> wrote: > >>> On Thu, Nov 7, 2019 at 8:44 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote: > >>>> On Wed, Nov 6, 2019 at 8:36 PM Lad Prabhakar <prabhakar.csengg@xxxxxxxxx> wrote: > >>>>> From: "Lad, Prabhakar" <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx> > >>>>> > >>>>> This patch adds the bindings for the R-Car PCIe endpoint driver. > >>>>> > >>>>> Signed-off-by: Lad, Prabhakar <prabhakar.mahadev-lad.rj@xxxxxxxxxxxxxx> > >>>> > >>>> Thanks for your patch! > >>>> > >>>>> --- /dev/null > >>>>> +++ b/Documentation/devicetree/bindings/pci/rcar-pci-ep.txt > >>>>> @@ -0,0 +1,43 @@ > >>>>> +* Renesas R-Car PCIe Endpoint Controller DT description > >>>>> + > >>>>> +Required properties: > >>>>> + "renesas,pcie-ep-r8a774c0" for the R8A774C0 SoC; > >>>>> + "renesas,pcie-ep-rcar-gen3" for a generic R-Car Gen3 or > >>>>> + RZ/G2 compatible device. > >>>> > >>>> Unless I'm missing something, this is for the exact same hardware block as > >>>> Documentation/devicetree/bindings/pci/rcar-pci.txt? > >>>> So shouldn't you amend those bindings, instead of adding new compatible > >>>> values? > >>>> Please remember that DT describes hardware, not software policy. > >>>> So IMHO choosing between host and endpoint is purely a configuration > >>>> issue, and could be indicated by the presence or lack of some DT properties. > >>>> E.g. host mode requires both "bus-range" and "device_type" properties, > >>>> so their absence could indicate endpoint mode. > >>>> > >>> yes its the same hardware block as described in the rcar-pci.txt, I > >>> did think about amending it > >>> but it might turn out to be bit messy, > >>> > >>> required properties host ======required properties Endpoint > >>> ====================||================== > >>> 1: reg || reg > >>> 2:bus-range || reg names > >>> 3: device_type || resets > >>> 4: ranges || clocks > >>> 5: dma-ranges || clock-names > >>> 6: interrupts || > >>> 7: interrupt-cells || > >>> 8: interrupt-map-mask || > >>> 9: clocks || > >>> 10: clock-names || > >> > >> We have a similar situation with SPI, where a controller can operate in > >> master or slave mode, based on the absence or presence of the > >> "spi-slave" DT property. > >> > >>> and if I go ahead with the same compatible string that would mean to > >>> add support for endpoint > >>> mode in the host driver itself. I did follow the examples of > >> > >> You can still have two separate drivers, binding against the same > >> compatible value. Just let the .probe() function return -ENODEV if it > >> discovers (by looking at DT properties) if the node is configured for > >> the other mode. > >> Which brings us to my next questions: is there any code that could be > >> shared between the drivers for the two modes? > >> > >>> rockchip/cadence/designware where > >>> its the same hardware block but has two different binding files one > >>> for host mode and other for > >>> endpoint mode. > >> > >> Having two separate DT binding documents sounds fine to me, if unifying > >> them makes things too complex. > >> However, I think they should use the same compatible value, because the > >> hardware block is the same, but just used in a different mode. > >> > >> Rob/Mark: Any input from the DT maintainers? > > > > Separate files makes sense because different modes will want to > > include different common schemas. We've generally been doing different > > compatibles too which makes validating the node has the right set of > > properties easier. > > > >>>>> +- reg: Five register ranges as listed in the reg-names property > >>>>> +- reg-names: Must include the following names > >>>>> + - "apb-base" > >>>>> + - "memory0" > >>>>> + - "memory1" > >>>>> + - "memory2" > >>>>> + - "memory3" > >>>> > >>>> What is the purpose of the last 4 regions? > >>>> Can they be chosen by the driver, at runtime? > >>>> > >>> no the driver cannot choose them at runtime, as these are the only > >>> PCIE memory(0/1/2/3) ranges > >>> in the AXI address space where host memory can be mapped. > >> > >> Are they fixed by the PCIe hardware, i.e. could they be looked up by the > >> driver based on the compatible value? > > > > That would be strange for a memory range. > > > > Sounds like like 'ranges' though I'm not sure if 'ranges' for an EP > > makes sense or what that should look like. > > These are similar to "memory node" with multiple address, size pairs. I'm > thinking if these should be added as a subnode within PCIe EP controller device > tree node? > +1 something similar like below ? pcie_ep: pcie_ep@fe000000 { compatible = "renesas,pcie-r8a7791", "renesas,pcie-rcar-gen2"; reg = <0 0xfe000000 0 0x80000>; reg-names = "apb-base"; clocks = <&cpg CPG_MOD 319>; clock-names = "pcie"; power-domains = <&sysc R8A774C0_PD_ALWAYS_ON>; resets = <&cpg 319>; mem-region { base = <0x0 0xfe100000 0 0x100000>, <0x0 0xfe200000 0 0x200000>, <0x0 0x30000000 0 0x8000000>, <0x0 0x38000000 0 0x8000000>; }; }; Cheers, --Prabhakar