--- /dev/null
+==========================================
+ARM CPUs capacity bindings
+==========================================
+
+==========================================
+1 - Introduction
+==========================================
+
+ARM systems may be configured to have cpus with different power/performance
+characteristics within the same chip. In this case, additional information has
+to be made available to the kernel for it to be aware of such differences and
+take decisions accordingly.
+
+==========================================
+2 - CPU capacity definition
+==========================================
+
+CPU capacity is a number that provides the scheduler information about CPUs
+heterogeneity. Such heterogeneity can come from micro-architectural differences
+(e.g., ARM big.LITTLE systems) or maximum frequency at which CPUs can run
+(e.g., SMP systems with multiple frequency domains). Heterogeneity in this
+context is about differing performance characteristics; this binding tries to
+capture a first-order approximation of the relative performance of CPUs.
+
+CPU capacities are obtained by running a suitable benchmark. This binding makes
+no guarantees on the validity or suitability of any particular benchmark, the
+final capacity should, however, be:
+
+* A "single-threaded" or CPU affine benchmark
+* Divided by the running frequency of the CPU executing the benchmark
+* Not subject to dynamic frequency scaling of the CPU
+
+For the time being we however advise usage of the Dhrystone benchmark. What
+above thus becomes:
+
+CPU capacities are obtained by running the Dhrystone benchmark on each CPU at
+max frequency (with caches enabled). The obtained DMIPS score is then divided
+by the frequency (in MHz) at which the benchmark has been run, so that
+DMIPS/MHz are obtained.  Such values are then normalized w.r.t. the highest
+score obtained in the system.
+
+==========================================
+3 - capacity-dmips-mhz
+==========================================
+
+capacity-dmips-mhz is an optional cpu node [1] property: u32 value
+representing CPU capacity expressed in normalized DMIPS/MHz. At boot time, the
+maximum frequency available to the cpu is then used to calculate the capacity
+value internally used by the kernel.
+
+capacity-dmips-mhz property is all-or-nothing: if it is specified for a cpu
+node, it has to be specified for every other cpu nodes, or the system will
+fall back to the default capacity value for every CPU. If cpufreq is not
+available, final capacities are calculated by directly using capacity-dmips-
+mhz values (normalized w.r.t. the highest value found while parsing the DT).
+
+===========================================
+4 - Examples
+===========================================
+
+Example 1 (ARM 64-bit, 6-cpu system, two clusters):
+capacities-dmips-mhz are scaled w.r.t. 1024 (cpu@0 and cpu@1)
+supposing cluster0@max-freq=1100 and custer1@max-freq=850,
+final capacities are 1024 for cluster0 and 446 for cluster1
+
+cpus {
+       #address-cells = <2>;
+       #size-cells = <0>;
+
+       cpu-map {
+               cluster0 {
+                       core0 {
+                               cpu = <&A57_0>;
+                       };
+                       core1 {
+                               cpu = <&A57_1>;
+                       };
+               };
+
+               cluster1 {
+                       core0 {
+                               cpu = <&A53_0>;
+                       };
+                       core1 {
+                               cpu = <&A53_1>;
+                       };
+                       core2 {
+                               cpu = <&A53_2>;
+                       };
+                       core3 {
+                               cpu = <&A53_3>;
+                       };
+               };
+       };
+
+       idle-states {
+               entry-method = "arm,psci";
+
+               CPU_SLEEP_0: cpu-sleep-0 {
+                       compatible = "arm,idle-state";
+                       arm,psci-suspend-param = <0x0010000>;
+                       local-timer-stop;
+                       entry-latency-us = <100>;
+                       exit-latency-us = <250>;
+                       min-residency-us = <150>;
+               };
+
+               CLUSTER_SLEEP_0: cluster-sleep-0 {
+                       compatible = "arm,idle-state";
+                       arm,psci-suspend-param = <0x1010000>;
+                       local-timer-stop;
+                       entry-latency-us = <800>;
+                       exit-latency-us = <700>;
+                       min-residency-us = <2500>;
+               };
+       };
+
+       A57_0: cpu@0 {
+               compatible = "arm,cortex-a57","arm,armv8";
+               reg = <0x0 0x0>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A57_L2>;
+               clocks = <&scpi_dvfs 0>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <1024>;
+       };
+
+       A57_1: cpu@1 {
+               compatible = "arm,cortex-a57","arm,armv8";
+               reg = <0x0 0x1>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A57_L2>;
+               clocks = <&scpi_dvfs 0>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <1024>;
+       };
+
+       A53_0: cpu@100 {
+               compatible = "arm,cortex-a53","arm,armv8";
+               reg = <0x0 0x100>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A53_L2>;
+               clocks = <&scpi_dvfs 1>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <578>;
+       };
+
+       A53_1: cpu@101 {
+               compatible = "arm,cortex-a53","arm,armv8";
+               reg = <0x0 0x101>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A53_L2>;
+               clocks = <&scpi_dvfs 1>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <578>;
+       };
+
+       A53_2: cpu@102 {
+               compatible = "arm,cortex-a53","arm,armv8";
+               reg = <0x0 0x102>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A53_L2>;
+               clocks = <&scpi_dvfs 1>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <578>;
+       };
+
+       A53_3: cpu@103 {
+               compatible = "arm,cortex-a53","arm,armv8";
+               reg = <0x0 0x103>;
+               device_type = "cpu";
+               enable-method = "psci";
+               next-level-cache = <&A53_L2>;
+               clocks = <&scpi_dvfs 1>;
+               cpu-idle-states = <&CPU_SLEEP_0 &CLUSTER_SLEEP_0>;
+               capacity-dmips-mhz = <578>;
+       };
+
+       A57_L2: l2-cache0 {
+               compatible = "cache";
+       };
+
+       A53_L2: l2-cache1 {
+               compatible = "cache";
+       };
+};
+
+Example 2 (ARM 32-bit, 4-cpu system, two clusters,
+          cpus 0,1@1GHz, cpus 2,3@500MHz):
+capacities-dmips-mhz are scaled w.r.t. 2 (cpu@0 and cpu@1), this means that first
+cpu@0 and cpu@1 are twice fast than cpu@2 and cpu@3 (at the same frequency)
+
+cpus {
+       #address-cells = <1>;
+       #size-cells = <0>;
+
+       cpu0: cpu@0 {
+               device_type = "cpu";
+               compatible = "arm,cortex-a15";
+               reg = <0>;
+               capacity-dmips-mhz = <2>;
+       };
+
+       cpu1: cpu@1 {
+               device_type = "cpu";
+               compatible = "arm,cortex-a15";
+               reg = <1>;
+               capacity-dmips-mhz = <2>;
+       };
+
+       cpu2: cpu@2 {
+               device_type = "cpu";
+               compatible = "arm,cortex-a15";
+               reg = <0x100>;
+               capacity-dmips-mhz = <1>;
+       };
+
+       cpu3: cpu@3 {
+               device_type = "cpu";
+               compatible = "arm,cortex-a15";
+               reg = <0x101>;
+               capacity-dmips-mhz = <1>;
+       };
+};
+
+===========================================
+5 - References
+===========================================
+
+[1] ARM Linux Kernel documentation - CPUs bindings
+    Documentation/devicetree/bindings/arm/cpus.txt