diff --git a/Bindings/bus/nvidia,tegra20-gmi.txt b/Bindings/bus/nvidia,tegra20-gmi.txt
new file mode 100644
index 0000000..c1e7062
--- /dev/null
+++ b/Bindings/bus/nvidia,tegra20-gmi.txt
@@ -0,0 +1,128 @@
+Device tree bindings for NVIDIA Tegra Generic Memory Interface bus
+
+The Generic Memory Interface bus enables memory transfers between internal and
+external memory. Can be used to attach various high speed devices such as
+synchronous/asynchronous NOR, FPGA, UARTS and more.
+
+The actual devices are instantiated from the child nodes of a GMI node.
+
+Required properties:
+ - compatible : Should contain one of the following:
+        For Tegra20 must contain "nvidia,tegra20-gmi".
+        For Tegra30 must contain "nvidia,tegra30-gmi".
+ - reg: Should contain GMI controller registers location and length.
+ - clocks: Must contain an entry for each entry in clock-names.
+ - clock-names: Must include the following entries: "gmi"
+ - resets : Must contain an entry for each entry in reset-names.
+ - reset-names : Must include the following entries: "gmi"
+ - #address-cells: The number of cells used to represent physical base
+   addresses in the GMI address space. Should be 2.
+ - #size-cells: The number of cells used to represent the size of an address
+   range in the GMI address space. Should be 1.
+ - ranges: Must be set up to reflect the memory layout with three integer values
+   for each chip-select line in use (only one entry is supported, see below
+   comments):
+   <cs-number> <offset> <physical address of mapping> <size>
+
+Note that the GMI controller does not have any internal chip-select address
+decoding, because of that chip-selects either need to be managed via software
+or by employing external chip-select decoding logic.
+
+If external chip-select logic is used to support multiple devices it is assumed
+that the devices use the same timing and so are probably the same type. It also
+assumes that they can fit in the 256MB address range. In this case only one
+child device is supported which represents the active chip-select line, see
+examples for more insight.
+
+The chip-select number is decoded from the child nodes second address cell of
+'ranges' property, if 'ranges' property is not present or empty chip-select will
+then be decoded from the first cell of the 'reg' property.
+
+Optional child cs node properties:
+
+ - nvidia,snor-data-width-32bit: Use 32bit data-bus, default is 16bit.
+ - nvidia,snor-mux-mode: Enable address/data MUX mode.
+ - nvidia,snor-rdy-active-before-data: Assert RDY signal one cycle before data.
+   If omitted it will be asserted with data.
+ - nvidia,snor-rdy-active-high: RDY signal is active high
+ - nvidia,snor-adv-active-high: ADV signal is active high
+ - nvidia,snor-oe-active-high: WE/OE signal is active high
+ - nvidia,snor-cs-active-high: CS signal is active high
+
+  Note that there is some special handling for the timing values.
+  From Tegra TRM:
+  Programming 0 means 1 clock cycle: actual cycle = programmed cycle + 1
+
+ - nvidia,snor-muxed-width: Number of cycles MUX address/data asserted on the
+   bus. Valid values are 0-15, default is 1
+ - nvidia,snor-hold-width: Number of cycles CE stays asserted after the
+   de-assertion of WR_N (in case of SLAVE/MASTER Request) or OE_N
+   (in case of MASTER Request). Valid values are 0-15, default is 1
+ - nvidia,snor-adv-width: Number of cycles during which ADV stays asserted.
+   Valid values are 0-15, default is 1.
+ - nvidia,snor-ce-width: Number of cycles before CE is asserted.
+   Valid values are 0-15, default is 4
+ - nvidia,snor-we-width: Number of cycles during which WE stays asserted.
+   Valid values are 0-15, default is 1
+ - nvidia,snor-oe-width: Number of cycles during which OE stays asserted.
+   Valid values are 0-255, default is 1
+ - nvidia,snor-wait-width: Number of cycles before READY is asserted.
+   Valid values are 0-255, default is 3
+
+Example with two SJA1000 CAN controllers connected to the GMI bus. We wrap the
+controllers with a simple-bus node since they are all connected to the same
+chip-select (CS4), in this example external address decoding is provided:
+
+gmi@70009000 {
+	compatible = "nvidia,tegra20-gmi";
+	reg = <0x70009000 0x1000>;
+	#address-cells = <2>;
+	#size-cells = <1>;
+	clocks = <&tegra_car TEGRA20_CLK_NOR>;
+	clock-names = "gmi";
+	resets = <&tegra_car 42>;
+	reset-names = "gmi";
+	ranges = <4 0 0xd0000000 0xfffffff>;
+
+	bus@4,0 {
+		compatible = "simple-bus";
+		#address-cells = <1>;
+		#size-cells = <1>;
+		ranges = <0 4 0 0x40100>;
+
+		nvidia,snor-mux-mode;
+		nvidia,snor-adv-active-high;
+
+		can@0 {
+			reg = <0 0x100>;
+			...
+		};
+
+		can@40000 {
+			reg = <0x40000 0x100>;
+			...
+		};
+	};
+};
+
+Example with one SJA1000 CAN controller connected to the GMI bus
+on CS4:
+
+gmi@70009000 {
+	compatible = "nvidia,tegra20-gmi";
+	reg = <0x70009000 0x1000>;
+	#address-cells = <2>;
+	#size-cells = <1>;
+	clocks = <&tegra_car TEGRA20_CLK_NOR>;
+	clock-names = "gmi";
+	resets = <&tegra_car 42>;
+	reset-names = "gmi";
+	ranges = <4 0 0xd0000000 0xfffffff>;
+
+	can@4,0 {
+		reg = <4 0 0x100>;
+		nvidia,snor-mux-mode;
+		nvidia,snor-adv-active-high;
+		...
+	};
+};