Starboard is Cobalt's porting layer and OS abstraction. It attempts to encompass all the platform-specific functionality that Cobalt actually uses, and nothing that it does not.
Cobalt and Starboard have been migrated from the GYP build system to the GN build system. This readme only contains instructions for GN, as GYP is no longer supported.
starboard/doc for more detailed documentation.
All source locations are specified relative to
starboard/ (this directory).
. - This is the root directory for the Starboard project, and contains all the public headers that Starboard defines.
examples/ - Example code demonstrating various aspects of Starboard API usage.
stub/ - The home of the Stub Starboard implementation. This contains a
BUILD.gn file that defines a library with all the source files needed to provide a complete linkable Starboard implementation.
nplb/ - “No Platform Left Behind,” Starboard's platform verification test suite.
shared/ - The home of all code that can be shared between Starboard implementations. Subdirectories delimit code that can be shared between platforms that share some facet of their OS API.
Before starting a Cobalt/Starboard port, first you will need to define the canonical names for your set of platform configurations. These will be used when organizing the code for your platforms.
What determines what goes into one platform configuration versus another? A platform configuration has a one-to-one mapping to a production binary. So, if you will need to produce a new binary, you are going to need a new platform configuration for that.
The recommended naming convention for a
<family-name> is a name specific to the family of products you are porting to Starboard and
<binary-variant> is one or more tokens that uniquely describe the specifics of the binary you want that configuration to produce.
For example, let's say your company is named BobCo. BobCo employs multiple different device architectures so it will need to define multiple platform configurations.
All the BobCo devices are called BobBox, so it's a reasonable choice as a product
<family-name>. But they have both big- and little-endian ARM chips. So they might define two platform configurations:
bobbox-armeb - For big-endian ARM devices.
bobbox-armel - For little-endian ARM devices.
To be perfectly compatible with the Cobalt source tree layout, any code that is written by a party that isn't the Cobalt team should be in the
third_party/ directory. The choice is up to you, but we recommend that you follow this practice, even if, as we expect to be common, you do not plan on sharing your Starboard implementation with anyone.
Primarily, following this convention ensures that no future changes to Cobalt or Starboard will conflict with your source code additions. Starboard is intended to be a junction where new Cobalt versions or Starboard implementations can be replaced without significant (and hopefully, any) code changes.
We recommend that you place your code here in the source tree:
shared/ - For code shared between architectures within a product family.
<binary-variant>/ - For any code that is specific to a specific binary variant. Each one of these must at least have
In the BobCo's BobBox example, we would see something like:
And so on.
You can start off by copying files from a reference port to your port's location. Currently these reference ports include:
BUILD.gn contains absolute paths, so the paths will still be valid if you copy it to a new directory. You can then incrementally replace files with new implementations as necessary.
The cleanest, simplest starting point is from the Stub reference implementation. Nothing will work, but you should be able to compile and link it with your toolchain. You can then replace stub implementations with implementations from
starboard/shared or your own custom implementations module-by-module, until you have gone through all modules.
You may also choose to copy either the Desktop Linux or Raspberry Pi ports and work backwards fixing things that don't compile or work on your platform.
For example, for
bobbox-armel, you might do:
mkdir -p third_party/starboard/bobbox cp -R starboard/stub third_party/starboard/bobbox/armel
Modify the files in
<binary-variant>/ as appropriate (you will probably be coming back to these files a lot).
<binary-variant>/BUILD.gn to point at all the source files that you want to build as your new Starboard implementation. The
// expression in GN refers to the top-level directory of your source tree. Otherwise, files are assumed to be relative to the directory the
.gni file is in. The
BUILD.gn file contains absolute paths, so the paths will still be valid if you copy it to a new directory. You can then incrementally replace files with new implementations as necessary.
In order to use a new platform configuration in a build, you need to ensure that you have a
platform_configuration/BUILD.gn in their own directory for each binary variant, plus the header files
atomic_public.h. You must add your platform name to
starboard/build/platforms.py along with the path to the port to be able to build it.
third_party/starboard/<family-name>/<binary-variant>. You may also consider copying from another reference platform, like
gl_type to the appropriate value if it is not the default
false. This enables or disables the DIAL server that runs inside Cobalt, only when Cobalt is running. You do not want in-app DIAL if you already have system-wide DIAL support.
clang_toolchain template and pass the base path to your toolchain, or use the
gcc_toolchain and pass the full path to each tool you use.
is_clang = true in
gcc_toolchain if the toolchain uses clang.
configuration_public.h and adjust all the configuration values as appropriate for your platform.
BUILD.gn to point at all the source files you want to build as part of your new Starboard implementation (as mentioned above).
atomic_public.h as necessary to point at the appropriate shared or custom implementations.
If you want to use
cobalt/build/gn.py, you‘ll also need a
third_party/starboard/<family-name>/<binary-variant>/args.gn file. This should contain the gn arguments necessary to build your platform. Pay particular attention to
is_clang. The defaults for each can be found in
starboard/build/config/BUILDCONFIG.gn. If you don’t care about using
gn.py, all of these arguments can be passed using
gn args or
gn gen with
--args. For example, the first command below might instead look like
gn gen out/bobbox-armeb_debug --args='target_platform="bobbox-armeb" build_type="debug"'.
You should now be able to run GN with your new port. From your the top level directory:
$ python cobalt/build/gn.py -c debug -p bobbox-armeb $ ninja -C out/bobbox-armeb_debug nplb
This will attempt to build the “No Platform Left Behind” test suite with your new Starboard implementation, and you are ready to start porting!
When bringing up a new Starboard platform, it is suggested that you try to get the NPLB tests passing module-by-module. Because of dependencies between modules, you will find it easier to get some modules passing sooner than other modules.
Here's a recommended module implementation order in which to get things going (still significantly subject to change based on feedback):