1Linux and the Device Tree
   3The Linux usage model for device tree data
   5Author: Grant Likely <>
   7This article describes how Linux uses the device tree.  An overview of
   8the device tree data format can be found on the device tree usage page
  13The "Open Firmware Device Tree", or simply Device Tree (DT), is a data
  14structure and language for describing hardware.  More specifically, it
  15is a description of hardware that is readable by an operating system
  16so that the operating system doesn't need to hard code details of the
  19Structurally, the DT is a tree, or acyclic graph with named nodes, and
  20nodes may have an arbitrary number of named properties encapsulating
  21arbitrary data.  A mechanism also exists to create arbitrary
  22links from one node to another outside of the natural tree structure.
  24Conceptually, a common set of usage conventions, called 'bindings',
  25is defined for how data should appear in the tree to describe typical
  26hardware characteristics including data busses, interrupt lines, GPIO
  27connections, and peripheral devices.
  29As much as possible, hardware is described using existing bindings to
  30maximize use of existing support code, but since property and node
  31names are simply text strings, it is easy to extend existing bindings
  32or create new ones by defining new nodes and properties.  Be wary,
  33however, of creating a new binding without first doing some homework
  34about what already exists.  There are currently two different,
  35incompatible, bindings for i2c busses that came about because the new
  36binding was created without first investigating how i2c devices were
  37already being enumerated in existing systems.
  391. History
  41The DT was originally created by Open Firmware as part of the
  42communication method for passing data from Open Firmware to a client
  43program (like to an operating system).  An operating system used the
  44Device Tree to discover the topology of the hardware at runtime, and
  45thereby support a majority of available hardware without hard coded
  46information (assuming drivers were available for all devices).
  48Since Open Firmware is commonly used on PowerPC and SPARC platforms,
  49the Linux support for those architectures has for a long time used the
  50Device Tree.
  52In 2005, when PowerPC Linux began a major cleanup and to merge 32-bit
  53and 64-bit support, the decision was made to require DT support on all
  54powerpc platforms, regardless of whether or not they used Open
  55Firmware.  To do this, a DT representation called the Flattened Device
  56Tree (FDT) was created which could be passed to the kernel as a binary
  57blob without requiring a real Open Firmware implementation.  U-Boot,
  58kexec, and other bootloaders were modified to support both passing a
  59Device Tree Binary (dtb) and to modify a dtb at boot time.  DT was
  60also added to the PowerPC boot wrapper (arch/powerpc/boot/*) so that
  61a dtb could be wrapped up with the kernel image to support booting
  62existing non-DT aware firmware.
  64Some time later, FDT infrastructure was generalized to be usable by
  65all architectures.  At the time of this writing, 6 mainlined
  66architectures (arm, microblaze, mips, powerpc, sparc, and x86) and 1
  67out of mainline (nios) have some level of DT support.
  692. Data Model
  71If you haven't already read the Device Tree Usage[1] page,
  72then go read it now.  It's okay, I'll wait....
  742.1 High Level View
  76The most important thing to understand is that the DT is simply a data
  77structure that describes the hardware.  There is nothing magical about
  78it, and it doesn't magically make all hardware configuration problems
  79go away.  What it does do is provide a language for decoupling the
  80hardware configuration from the board and device driver support in the
  81Linux kernel (or any other operating system for that matter).  Using
  82it allows board and device support to become data driven; to make
  83setup decisions based on data passed into the kernel instead of on
  84per-machine hard coded selections.
  86Ideally, data driven platform setup should result in less code
  87duplication and make it easier to support a wide range of hardware
  88with a single kernel image.
  90Linux uses DT data for three major purposes:
  911) platform identification,
  922) runtime configuration, and
  933) device population.
  952.2 Platform Identification
  97First and foremost, the kernel will use data in the DT to identify the
  98specific machine.  In a perfect world, the specific platform shouldn't
  99matter to the kernel because all platform details would be described
 100perfectly by the device tree in a consistent and reliable manner.
 101Hardware is not perfect though, and so the kernel must identify the
 102machine during early boot so that it has the opportunity to run
 103machine-specific fixups.
 105In the majority of cases, the machine identity is irrelevant, and the
 106kernel will instead select setup code based on the machine's core
 107CPU or SoC.  On ARM for example, setup_arch() in
 108arch/arm/kernel/setup.c will call setup_machine_fdt() in
 109arch/arm/kernel/devicetree.c which searches through the machine_desc
 110table and selects the machine_desc which best matches the device tree
 111data.  It determines the best match by looking at the 'compatible'
 112property in the root device tree node, and comparing it with the
 113dt_compat list in struct machine_desc.
 115The 'compatible' property contains a sorted list of strings starting
 116with the exact name of the machine, followed by an optional list of
 117boards it is compatible with sorted from most compatible to least.  For
 118example, the root compatible properties for the TI BeagleBoard and its
 119successor, the BeagleBoard xM board might look like:
 121        compatible = "ti,omap3-beagleboard", "ti,omap3450", "ti,omap3";
 122        compatible = "ti,omap3-beagleboard-xm", "ti,omap3450", "ti,omap3";
 124Where "ti,omap3-beagleboard-xm" specifies the exact model, it also
 125claims that it compatible with the OMAP 3450 SoC, and the omap3 family
 126of SoCs in general.  You'll notice that the list is sorted from most
 127specific (exact board) to least specific (SoC family).
 129Astute readers might point out that the Beagle xM could also claim
 130compatibility with the original Beagle board.  However, one should be
 131cautioned about doing so at the board level since there is typically a
 132high level of change from one board to another, even within the same
 133product line, and it is hard to nail down exactly what is meant when one
 134board claims to be compatible with another.  For the top level, it is
 135better to err on the side of caution and not claim one board is
 136compatible with another.  The notable exception would be when one
 137board is a carrier for another, such as a CPU module attached to a
 138carrier board.
 140One more note on compatible values.  Any string used in a compatible
 141property must be documented as to what it indicates.  Add
 142documentation for compatible strings in Documentation/devicetree/bindings.
 144Again on ARM, for each machine_desc, the kernel looks to see if
 145any of the dt_compat list entries appear in the compatible property.
 146If one does, then that machine_desc is a candidate for driving the
 147machine.  After searching the entire table of machine_descs,
 148setup_machine_fdt() returns the 'most compatible' machine_desc based
 149on which entry in the compatible property each machine_desc matches
 150against.  If no matching machine_desc is found, then it returns NULL.
 152The reasoning behind this scheme is the observation that in the majority
 153of cases, a single machine_desc can support a large number of boards
 154if they all use the same SoC, or same family of SoCs.  However,
 155invariably there will be some exceptions where a specific board will
 156require special setup code that is not useful in the generic case.
 157Special cases could be handled by explicitly checking for the
 158troublesome board(s) in generic setup code, but doing so very quickly
 159becomes ugly and/or unmaintainable if it is more than just a couple of
 162Instead, the compatible list allows a generic machine_desc to provide
 163support for a wide common set of boards by specifying "less
 164compatible" value in the dt_compat list.  In the example above,
 165generic board support can claim compatibility with "ti,omap3" or
 166"ti,omap3450".  If a bug was discovered on the original beagleboard
 167that required special workaround code during early boot, then a new
 168machine_desc could be added which implements the workarounds and only
 169matches on "ti,omap3-beagleboard".
 171PowerPC uses a slightly different scheme where it calls the .probe()
 172hook from each machine_desc, and the first one returning TRUE is used.
 173However, this approach does not take into account the priority of the
 174compatible list, and probably should be avoided for new architecture
 1772.3 Runtime configuration
 179In most cases, a DT will be the sole method of communicating data from
 180firmware to the kernel, so also gets used to pass in runtime and
 181configuration data like the kernel parameters string and the location
 182of an initrd image.
 184Most of this data is contained in the /chosen node, and when booting
 185Linux it will look something like this:
 187        chosen {
 188                bootargs = "console=ttyS0,115200 loglevel=8";
 189                initrd-start = <0xc8000000>;
 190                initrd-end = <0xc8200000>;
 191        };
 193The bootargs property contains the kernel arguments, and the initrd-*
 194properties define the address and size of an initrd blob.  Note that
 195initrd-end is the first address after the initrd image, so this doesn't
 196match the usual semantic of struct resource.  The chosen node may also
 197optionally contain an arbitrary number of additional properties for
 198platform-specific configuration data.
 200During early boot, the architecture setup code calls of_scan_flat_dt()
 201several times with different helper callbacks to parse device tree
 202data before paging is setup.  The of_scan_flat_dt() code scans through
 203the device tree and uses the helpers to extract information required
 204during early boot.  Typically the early_init_dt_scan_chosen() helper
 205is used to parse the chosen node including kernel parameters,
 206early_init_dt_scan_root() to initialize the DT address space model,
 207and early_init_dt_scan_memory() to determine the size and
 208location of usable RAM.
 210On ARM, the function setup_machine_fdt() is responsible for early
 211scanning of the device tree after selecting the correct machine_desc
 212that supports the board.
 2142.4 Device population
 216After the board has been identified, and after the early configuration data
 217has been parsed, then kernel initialization can proceed in the normal
 218way.  At some point in this process, unflatten_device_tree() is called
 219to convert the data into a more efficient runtime representation.
 220This is also when machine-specific setup hooks will get called, like
 221the machine_desc .init_early(), .init_irq() and .init_machine() hooks
 222on ARM.  The remainder of this section uses examples from the ARM
 223implementation, but all architectures will do pretty much the same
 224thing when using a DT.
 226As can be guessed by the names, .init_early() is used for any machine-
 227specific setup that needs to be executed early in the boot process,
 228and .init_irq() is used to set up interrupt handling.  Using a DT
 229doesn't materially change the behaviour of either of these functions.
 230If a DT is provided, then both .init_early() and .init_irq() are able
 231to call any of the DT query functions (of_* in include/linux/of*.h) to
 232get additional data about the platform.
 234The most interesting hook in the DT context is .init_machine() which
 235is primarily responsible for populating the Linux device model with
 236data about the platform.  Historically this has been implemented on
 237embedded platforms by defining a set of static clock structures,
 238platform_devices, and other data in the board support .c file, and
 239registering it en-masse in .init_machine().  When DT is used, then
 240instead of hard coding static devices for each platform, the list of
 241devices can be obtained by parsing the DT, and allocating device
 242structures dynamically.
 244The simplest case is when .init_machine() is only responsible for
 245registering a block of platform_devices.  A platform_device is a concept
 246used by Linux for memory or I/O mapped devices which cannot be detected
 247by hardware, and for 'composite' or 'virtual' devices (more on those
 248later).  While there is no 'platform device' terminology for the DT,
 249platform devices roughly correspond to device nodes at the root of the
 250tree and children of simple memory mapped bus nodes.
 252About now is a good time to lay out an example.  Here is part of the
 253device tree for the NVIDIA Tegra board.
 256        compatible = "nvidia,harmony", "nvidia,tegra20";
 257        #address-cells = <1>;
 258        #size-cells = <1>;
 259        interrupt-parent = <&intc>;
 261        chosen { };
 262        aliases { };
 264        memory {
 265                device_type = "memory";
 266                reg = <0x00000000 0x40000000>;
 267        };
 269        soc {
 270                compatible = "nvidia,tegra20-soc", "simple-bus";
 271                #address-cells = <1>;
 272                #size-cells = <1>;
 273                ranges;
 275                intc: interrupt-controller@50041000 {
 276                        compatible = "nvidia,tegra20-gic";
 277                        interrupt-controller;
 278                        #interrupt-cells = <1>;
 279                        reg = <0x50041000 0x1000>, < 0x50040100 0x0100 >;
 280                };
 282                serial@70006300 {
 283                        compatible = "nvidia,tegra20-uart";
 284                        reg = <0x70006300 0x100>;
 285                        interrupts = <122>;
 286                };
 288                i2s1: i2s@70002800 {
 289                        compatible = "nvidia,tegra20-i2s";
 290                        reg = <0x70002800 0x100>;
 291                        interrupts = <77>;
 292                        codec = <&wm8903>;
 293                };
 295                i2c@7000c000 {
 296                        compatible = "nvidia,tegra20-i2c";
 297                        #address-cells = <1>;
 298                        #size-cells = <0>;
 299                        reg = <0x7000c000 0x100>;
 300                        interrupts = <70>;
 302                        wm8903: codec@1a {
 303                                compatible = "wlf,wm8903";
 304                                reg = <0x1a>;
 305                                interrupts = <347>;
 306                        };
 307                };
 308        };
 310        sound {
 311                compatible = "nvidia,harmony-sound";
 312                i2s-controller = <&i2s1>;
 313                i2s-codec = <&wm8903>;
 314        };
 317At .init_machine() time, Tegra board support code will need to look at
 318this DT and decide which nodes to create platform_devices for.
 319However, looking at the tree, it is not immediately obvious what kind
 320of device each node represents, or even if a node represents a device
 321at all.  The /chosen, /aliases, and /memory nodes are informational
 322nodes that don't describe devices (although arguably memory could be
 323considered a device).  The children of the /soc node are memory mapped
 324devices, but the codec@1a is an i2c device, and the sound node
 325represents not a device, but rather how other devices are connected
 326together to create the audio subsystem.  I know what each device is
 327because I'm familiar with the board design, but how does the kernel
 328know what to do with each node?
 330The trick is that the kernel starts at the root of the tree and looks
 331for nodes that have a 'compatible' property.  First, it is generally
 332assumed that any node with a 'compatible' property represents a device
 333of some kind, and second, it can be assumed that any node at the root
 334of the tree is either directly attached to the processor bus, or is a
 335miscellaneous system device that cannot be described any other way.
 336For each of these nodes, Linux allocates and registers a
 337platform_device, which in turn may get bound to a platform_driver.
 339Why is using a platform_device for these nodes a safe assumption?
 340Well, for the way that Linux models devices, just about all bus_types
 341assume that its devices are children of a bus controller.  For
 342example, each i2c_client is a child of an i2c_master.  Each spi_device
 343is a child of an SPI bus.  Similarly for USB, PCI, MDIO, etc.  The
 344same hierarchy is also found in the DT, where I2C device nodes only
 345ever appear as children of an I2C bus node.  Ditto for SPI, MDIO, USB,
 346etc.  The only devices which do not require a specific type of parent
 347device are platform_devices (and amba_devices, but more on that
 348later), which will happily live at the base of the Linux /sys/devices
 349tree.  Therefore, if a DT node is at the root of the tree, then it
 350really probably is best registered as a platform_device.
 352Linux board support code calls of_platform_populate(NULL, NULL, NULL, NULL)
 353to kick off discovery of devices at the root of the tree.  The
 354parameters are all NULL because when starting from the root of the
 355tree, there is no need to provide a starting node (the first NULL), a
 356parent struct device (the last NULL), and we're not using a match
 357table (yet).  For a board that only needs to register devices,
 358.init_machine() can be completely empty except for the
 359of_platform_populate() call.
 361In the Tegra example, this accounts for the /soc and /sound nodes, but
 362what about the children of the SoC node?  Shouldn't they be registered
 363as platform devices too?  For Linux DT support, the generic behaviour
 364is for child devices to be registered by the parent's device driver at
 365driver .probe() time.  So, an i2c bus device driver will register a
 366i2c_client for each child node, an SPI bus driver will register
 367its spi_device children, and similarly for other bus_types.
 368According to that model, a driver could be written that binds to the
 369SoC node and simply registers platform_devices for each of its
 370children.  The board support code would allocate and register an SoC
 371device, a (theoretical) SoC device driver could bind to the SoC device,
 372and register platform_devices for /soc/interrupt-controller, /soc/serial,
 373/soc/i2s, and /soc/i2c in its .probe() hook.  Easy, right?
 375Actually, it turns out that registering children of some
 376platform_devices as more platform_devices is a common pattern, and the
 377device tree support code reflects that and makes the above example
 378simpler.  The second argument to of_platform_populate() is an
 379of_device_id table, and any node that matches an entry in that table
 380will also get its child nodes registered.  In the tegra case, the code
 381can look something like this:
 383static void __init harmony_init_machine(void)
 385        /* ... */
 386        of_platform_populate(NULL, of_default_bus_match_table, NULL, NULL);
 389"simple-bus" is defined in the ePAPR 1.0 specification as a property
 390meaning a simple memory mapped bus, so the of_platform_populate() code
 391could be written to just assume simple-bus compatible nodes will
 392always be traversed.  However, we pass it in as an argument so that
 393board support code can always override the default behaviour.
 395[Need to add discussion of adding i2c/spi/etc child devices]
 397Appendix A: AMBA devices
 400ARM Primecells are a certain kind of device attached to the ARM AMBA
 401bus which include some support for hardware detection and power
 402management.  In Linux, struct amba_device and the amba_bus_type is
 403used to represent Primecell devices.  However, the fiddly bit is that
 404not all devices on an AMBA bus are Primecells, and for Linux it is
 405typical for both amba_device and platform_device instances to be
 406siblings of the same bus segment.
 408When using the DT, this creates problems for of_platform_populate()
 409because it must decide whether to register each node as either a
 410platform_device or an amba_device.  This unfortunately complicates the
 411device creation model a little bit, but the solution turns out not to
 412be too invasive.  If a node is compatible with "arm,amba-primecell", then
 413of_platform_populate() will register it as an amba_device instead of a