Quoting Rob Herring (2022-06-29 10:58:52) > On Wed, Jun 29, 2022 at 9:01 AM Pin-yen Lin <treapking@xxxxxxxxxxxx> wrote: > > > > > > Yes it6505 is just a protocol converter. But in our use case, the output DP > > > lines are connected to the Type-C ports and the chip has to know which > > > port has DP Alt mode enabled. Does this justify a child node here? > > > > > > Does it make more sense if we we eliminate the usb-switch node here > > > and list the ports in the top level? > > In the it6505 node? No, the it6505 h/w knows nothing about Type-C > switching so neither should its binding. > > What device controls the switching in this case? Again, block diagrams > please if you want advice on what the binding should look like. My understanding is there are 4 DP lanes on it6505 and two lanes are connected to one usb-c-connector and the other two lanes are connected to a different usb-c-connector. The IT6505 driver will send DP out on the associated two DP lanes depending on which usb-c-connector has DP pins assigned by the typec manager. +-------+ | | +--------+ /----+ usb-c | | IT6505 | / /---+ | | +------------ lane 0 ------/ / | | | +------------ lane 1 -------/ +-------+ DPI -----+ | | | +-------+ | | | | | +------------ lane 2 -------------+ usb-c | | +------------ lane 3 -------------+ | | | | | +--------+ +-------+ The bridge is a mux that steers DP to one or the other usb-c-connector based on what the typec manager decides. I would expect this to be described with the existing port binding in the it6505 node. The binding would need to be extended to describe the output side. bridge@5c { compatible = "ite,it6505"; ... ports { #address-cells = <1>; #size-cells = <0>; port@0 { reg = <0>; it6505_in: endpoint { remote-endpoint = <&dpi_out>; }; }; port@1 { #address-cells = <1>; #size-cells = <0>; reg = <1>; it6505_out_lanes_01: endpoint@0 { reg = <0> data-lanes = <0 1>; remote-endpoint = <&typec0>; }; it6505_out_lanes_23: endpoint@1 { reg = <1> data-lanes = <2 3>; remote-endpoint = <&typec1>; }; }; }; }; usb-c-connector { compatible = "usb-c-connector"; .... ports { #address-cells = <1>; #size-cells = <0>; port@1 { reg = <1>; typec0: endpoint { remote-endpoint = <&it6505_out_lanes_01>; }; }; }; }; Note: port@1 in usb-c-connector above is for superspeed lines, which technically DP reuses. But we can also shove USB3 superspeed lines over the other two superspeed pins in the usb-c-connector pinout. Probably port@1 should have two endpoints so we can connect usb superspeed lines there in addition to DP lines. Another use case would be to have the IT6505 send 4 lanes of DP to a dp-connector. Or send one lane of DP to 4 dp-connectors? Sounds awful but possible if this bridge can drive one lane DP out on any DP output lane. bridge@5c { compatible = "ite,it6505"; ... ports { #address-cells = <1>; #size-cells = <0>; port@0 { reg = <0>; it6505_in: endpoint { remote-endpoint = <&dpi_out>; }; }; port@1 { #address-cells = <1>; #size-cells = <0>; reg = <1>; it6505_out_lane_0: endpoint@0 { reg = <0> data-lanes = <0>; remote-endpoint = <&dp0>; }; it6505_out_lane_1: endpoint@1 { reg = <1> data-lanes = <1>; remote-endpoint = <&dp1>; }; it6505_out_lane_2: endpoint@2 { reg = <2> data-lanes = <2>; remote-endpoint = <&dp1>; }; it6505_out_lane_3: endpoint@3 { reg = <3> data-lanes = <3>; remote-endpoint = <&dp1>; }; }; }; }; Or we could send 4 lanes of DP to a typec redriver that's controlled by firmware outside of the kernel where the redriver takes in 4 lanes DP and USB3 and muxes USB or USB+DP or just DP. bridge@5c { compatible = "ite,it6505"; ... ports { #address-cells = <1>; #size-cells = <0>; port@0 { reg = <0>; it6505_in: endpoint { remote-endpoint = <&dpi_out>; }; }; port@1 { #address-cells = <1>; #size-cells = <0>; reg = <1>; it6505_out: endpoint@0 { reg = <0> data-lanes = <0 1 2 3>; remote-endpoint = <&cros_ec_typec_dp_in>; }; }; }; }; cros_ec@0 { compatible = "google,cros-ec-spi"; ... cros_ec_typec: typec { compatible = "google,cros-ec-typec"; ports { #address-cells = <1>; #size-cells = <0>; port@0 { cros_ec_typec_dp_in: endpoint { remote-endpoint = <&it6505_out>; }; }; }; }; }; In this case we would want to have cros_ec_typec describe some sort of input graph binding to accept DP. I imagine the EC as a dp-bridge in this scenario, where it takes in DP and muxes it along with USB onto a usb-c-connector. If the EC is managing multiple usb-c-connectors then the typec framework may need help determining which port DP is going to, especially in the case that two DP bridges are used and they're both sent to the EC so that the EC can mux them onto a usb-c-connector along with USB. (I can draw more block diagrams if you prefer) > > > > > > > If the child node represents what the output is connected to (like a > > > > > > bus), then yes that is a pattern we have used. > > > > > > > > > > For the anx7625 case, the child node does represent what the output is connected > > > > > to (the usb-c-connector via the switch). Does that not qualify? Or do you mean > > > > > the child node should be a usb-c-connector itself? > > > > > > > > > > > For example, a panel > > > > > > represented as child node of a display controller. However, that only > > > > > > works for simple cases, and is a pattern we have gotten away from in > > > > > > favor of using the graph binding. > > > > > > > > > > The child node will still use a OF graph binding to connect to the > > > > > usb-c-connector. > > > > > Is that insufficient to consider a child node usage here? > > > > > By "using the graph binding", do you mean "only use the top-level ports" ? > > > > > I'm trying to clarify this, so that it will inform future versions and patches. > > > > > > > > What I want to see is block diagrams of possible h/w with different > > > > scenarios and then what the binding looks like in those cases. The > > > > switching/muxing could be in the SoC, a bridge chip, a Type C > > > > controller, a standalone mux chip, or ????. If you want a somewhat > > > > genericish binding, then you need to consider all of these. > > > > Then, is it suitable to put the switch binding into the drivers own bindings > > (i.e., the bindings for it6505 and anx7625), and introduce some helper > > functions to eliminate the duplication in the code? > > Only for h/w devices that have switch h/w. > > > Though we will have two similar bindings in two drivers with this approach. > > While the anx7625 driver having Type-C awareness makes sense given it > has a switch and a type-C controller, that doesn't make sense for > it6505 (and every other bridge driver that's just providing DP > output). > I don't see the benefit to making a genericish binding for typec switches, even if the hardware has typec awareness like anx7625. It looks like the graph binding can already handle what we need. By putting it in the top-level ports node we have one way to describe the input/output of the device instead of describing it in the top-level in the display connector case and the child typec switch node in the usb c connector case. I think the difficulty comes from the combinatorial explosion of possible configurations. As evidenced here, hardware engineers can take a DP bridge and use it as a DP mux as long as the bridge has lane control. Or they can take a device like anx7625 and ignore the USB aspect and use the internal crosspoint switch as a DP mux. The anx7625 part could be a MIPI-to-DP display bridge plus mux that is connected to two dp-connectors, in which case typec isn't even involved, but we could mux between two dp connectors. Also, the typec framework would like to simply walk the graph from the usb-c-connector looking for nodes that have 'mode-switch' or 'orientation-switch' properties and treat those devices as the typec switches for the connector. This means that we have to add these typec properties like 'mode-switch' to something like the IT6505 bridge binding, which is a little awkward. I wonder if those properties aren't really required. Would it be sufficient if the framework could walk the graph and look for registered typec switches in the kernel that have a matching of_node?