Cobalt Evergreen is an end-to-end framework for cloud-based deployment of Cobalt updates without the need for supplemental Cobalt integration work on device platforms.
For a bit of background context, as the number of Cobalt devices in the field increases there is a growing proliferation of version fragmentation. Many of these devices are unable to take advantage of the benefits of Cobalt performance, security, and functional improvements as it is costly to integrate and port new versions of Cobalt. We recognized this issue, listened to feedback from the Cobalt community and as a result developed Cobalt Evergreen as a way to make updating Cobalt a much simpler process for everyone involved.
This relies on separating the Starboard(platform) and Cobalt(core) components of a Cobalt implementation into the following discrete components:
Google-built (on Google toolchain)
Partner-built (on Partner toolchain)
With this new Cobalt platform architecture, less engineering effort is necessary for a full Cobalt integration/deployment. The idea here is you should only need to implement Starboard one time and any Cobalt-level updates should only require platform testing. NOTE that certain new Cobalt features may require Starboard changes, so if you want to take advantage of some of these new features, Starboard changes may be necessary.
crashpad_handlercomponents required to be built on platform toolchains
There are minimal differences in switching to Evergreen as the Cobalt team has already done a majority of the work building the necessary components to support the Evergreen architecture. You will still be responsible for building the Starboard and platform-specific components as usual. Thereafter, switching to Evergreen is as simple as building a different configuration. Please see the Raspberry Pi 2 Evergreen reference port (Instructions) for an example.
Cobalt Evergreen requires that there are two separate build(
gyp) configurations used due to the separation of the Cobalt core(
libcobalt.so) and the platform-specific Starboard layer(
loader_app). As a result, you will have to initiate a separate gyp process for each. This is required since the Cobalt core binary is built with the Google toolchain settings and the platform-specific Starboard layer is built with partner toolchain configurations.
Cobalt Evergreen is built by a separate gyp platform using the Google toolchain:
$ cobalt/build/gyp_cobalt evergreen-arm-softfp-sbversion-12 $ ninja -C out/evergreen-arm-softfp-sbversion-12_qa cobalt
Which produces a shared library
libcobalt.so targeted for specific architecture, ABI and Starboard version.
The gyp variable
sb_evergreen is set to 1 when building
The partner port of Starboard is built with the partner’s toolchain and is linked into the **
loader_app which knows how to dynamically load
libcobalt.so, and the
crashpad_handler which handles crashes.
cobalt/build/gyp_cobalt <partner_port_name> ninja -C out/<partner_port_name>_qa loader_app crashpad_handler
Partners should set
sb_evergreen_compatible to 1 in their gyp platform config. DO NOT set the
sb_evergreen to 1 in your platform-specific configuration as it is used only by Cobalt when building with the Google toolchain.
Additionally, partners should install crash handlers as instructed in the Installing Crash Handlers for Cobalt guide.
The following additional Starboard interfaces are necessary to implement for Evergreen:
#define SB_CAN_MAP_EXECUTABLE_MEMORY 1
Only if necessary, create a customized SABI configuration for your architecture. Note, we do not anticipate that you will need to make a new configuration for your platform unless it is not one of our supported architectures:
If your target architecture falls outside the support list above, please reach out to us for guidance.
Some common versions of Cobalt in the field may show a bug in the implementation of the CSS which can cause layout behavior to cause components to overlap and give users a poor user experience. A fix for this is identified and pushed to Cobalt open source ready for integration and deployment on devices.
Though a fix for this was made available in the latest Cobalt open source, affected devices in the field are not getting updated (e.g. due to engineering resources, timing, device end-of-life), users continue to have a poor experience and have negative sentiment against a device. In parallel, the web app team determines a workaround for this particular situation, but the workaround is obtuse and causes app bloat and technical debt from on-going maintenance of workarounds for various Cobalt versions.
The Cobalt team can work with you to guide validation and deployment of a shared Cobalt library to all affected devices much more quickly without all the engineering effort required to deploy a new Cobalt build. With this simpler updating capability, device behavior will be more consistent and there is less technical debt from workarounds on the web application side. Additionally, users can benefit from the latest performance, security, and functional fixes.
Cobalt Evergreen currently supports the following
Additional reserved storage (96MB) is required for Evergreen binaries. We expect Evergreen implementations to have an initial Cobalt preloaded on the device and an additional reserved space for additional Cobalt update storage.
As Cobalt Evergreen is intended to be updated from Google Cloud architecture without the need for device FW updates, it is important that this can be done easily and securely on the target platform. There are a set of general minimum requirements to do so:
PROT_EXEC) for loading in-memory and performing relocations for Cobalt Evergreen binaries
elf_loader_sandbox binary can be used to run tests in Evergreen mode. This is much more lightweight than the
loader_app, and does not have any knowledge about installations or downloading updates.
elf_loader_sandbox is run using two command line switches:
--evergreen_content. These switches are the path to the shared library to be run and the path to that shared library's content. These paths should be relative to the content of the elf_loader_sandbox.
For example, if we wanted to run the NPLB set of tests and had the following directory tree,
.../elf_loader_sandbox .../content/app/nplb/lib/libnplb.so .../content/app/nplb/content
we would use the following command to run NPLB:
.../elf_loader_sandbox --evergreen_library=app/nplb/lib/libnplb.so --evergreen_content=app/nplb/content
Building tests is identical to how they are already built except that a different platform configuration must be used. The platform configuration should be an Evergreen platform configuration, and have a Starboard ABI file that matches the file used by the platform configuration used to build the
For example, building these targets for the Raspberry Pi 2 would use the
evergreen-arm-hardfp platform configurations.
In order to verify the platform requirements you should run the ‘nplb_evergreen_compat_tests’. These tests ensure that the platform is configured appropriately for Evergreen.
To enable the test, set the
sb_evergreen_compatible gyp variable to 1 in the
gyp_configuration.gypi. For more details please take a look at the Raspberry Pi 2 gyp files.
There is a reference implementation available for Raspberry Pi 2 with instructions available here.
The above diagram is a high-level overview of the components in the Cobalt Evergreen architecture.
Partner-built represents components the Partner is responsible for implementing and building.
Cobalt-built represents components the Cobalt team is responsible for implementing and building.
This is a new component in the Cobalt Shared Library component that is built on top of the Starboard API. The purpose of this module is to check the update servers if there is a new version of the Cobalt Shared Library available for the target device. If a new version is available, the Cobalt updater will download, verify, and install the new package on the target platform. The new package can be used the next time Cobalt is started or it can be forced to update immediately and restart Cobalt as soon as the new package is available and verified on the platform. This behavior will take into account the suspend/resume logic on the target platform.
Functionally, the Cobalt Updater itself runs as a separate thread within the Cobalt process when Cobalt is running. This behavior depends on what the target platform allows.
For more detailed information on Cobalt Updater, please take a look here.
We use Google Update as the infrastructure to manage the Cobalt Evergreen package install and update process. This has been heavily used across Google for quite some time and has the level of reliability required for Cobalt Evergreen. There are also other important features such as:
For more detailed information on Google Update for Cobalt Evergreen, please take a look here.
We use Google Downloads to manage the downloads available for Cobalt Evergreen. The Cobalt Updater will use Google Downloads in order to download available packages onto the target platform. We selected Google Downloads for this purpose due to its ability to scale across billions of devices as well as the flexibility to control download behavior for reliability.
For more detailed information on Google Downloads (Download Server) for Cobalt Evergreen, please take a look here.
The Starboard ABI was introduced to provide a single, consistent method for specifying the Starboard API version and the ABI. This specification ensures that any two binaries, built with the same Starboard ABI and with arbitrary toolchains, are compatible.
Note that Cobalt already provides default SABI files for the following architectures:
You should not need to make a new SABI file for your target platform unless it is not a currently supported architecture. We recommend that you do not make any SABI file changes. If you believe it is necessary to create a new SABI file for your target platform, please reach out to the Cobalt team to advise.
For more detailed information on the Starboard ABI for Cobalt Evergreen, please take a look here.
Cobalt Evergreen provides support for maintaining multiple, separate versions of the Cobalt binary on a platform. These versions are stored in installation slots(i.e. known locations on disk), and are used to significantly improve the resilience and reliability of Cobalt updates.
All slot configurations assume the following:
The number of installation slots available will be determined by the platform owner. 3 slots is the default configuration for Evergreen. There can be
N installation slots configured with the only limitation being available storage.
NOTE: 3-slots is the DEFAULT configuration.
The number of installation slots is directly controlled using
kMaxNumInstallations, defined in loader_app.cc.
It is worth noting that all slot configurations specify that the first installation slot (
SLOT_0) will always be the read-only factory system image. This is permanently installed on the platform and is used as a fail-safe option. This is stored in the directory specified by
kSbSystemPathContentDirectory under the
All of the other installation slots are located within the storage directory specified by
kSbSystemPathStorageDirectory. This will vary depending on the platform.
For example, on the Raspberry Pi the
kSbSystemPathStorageDirectory directory is
/home/pi/.cobalt_storage, and the paths to all existing installation slots will be as follows:
/home/pi/<kSbSystemPathContentDirectory>/app/cobalt (system image installation SLOT_0) (read-only) /home/pi/.cobalt_storage/installation_1 (SLOT_1) /home/pi/.cobalt_storage/installation_2 (SLOT_2) ... /home/pi/.cobalt_storage/installation_N (SLOT_N)
Where the most recent update is stored will alternate between the available writable slots. In the above example, this would be
Slots are used to manage Cobalt Evergreen binaries with associated app metadata to select the appropriate Cobalt Evergreen binaries.
See the below structures for an example 3-slot configuration.
kSbSystemPathContentDirectory used for the read-only System Image required for all slot configurations:
. ├── content <--(kSbSystemPathContentDirectory) │ └── fonts <--(kSbSystemPathFontDirectory, `standard` or `limit` configuration, to be explained below) │ └── app │ └── cobalt <--(SLOT_0) │ ├── content <--(relative path defined in kSystemImageContentPath) │ │ ├── fonts <--(`minimal` configuration) │ │ ├── (icu) <--(only present when it needs to be updated by Cobalt Update) │ │ ├── licenses │ │ ├── ssl │ ├── lib │ │ └── libcobalt.so <--(System image version of libcobalt.so) │ └── manifest.json └── loader_app <--(Cobalt launcher binary) └── crashpad_handler <--(Cobalt crash handler)
kSbSystemPathStorageDirectory used for future Cobalt Evergreen updates in an example 3-slot configuration:
├── .cobalt_storage <--(kSbSystemPathStorageDirectory) ├── cobalt_updater │ └── prefs_<APP_KEY>.json ├── installation_1 <--(SLOT_1 - currently unused) ├── installation_2 <--(SLOT_2 - contains new Cobalt version) │ ├── content │ │ ├── fonts <--(`minimal` configuration) │ │ ├── (icu) <--(only present when it needs to be updated by Cobalt Update) │ │ ├── licenses │ │ ├── ssl │ ├── lib │ │ └── libcobalt.so <--(SLOT_2 version of libcobalt.so) │ ├── manifest.fingerprint │ └── manifest.json <-- (Evergreen version information of libcobalt.so under SLOT_2) ├── installation_store_<APP_KEY>.pb └── icu (default location shared by installation slots, to be explained below)
Each Cobalt Evergreen application has a set of unique metadata to track slot selection. The following set of files are unique per application via a differentiating <APP_KEY> identifier, which is a Base64 hash appended to the filename.
You should NOT change any of these files and they are highlighted here just for reference.
The system font directory
kSbSystemPathFontDirectory should be configured to point to the
standard (23MB) or the
limited (3.1MB) cobalt font packages. An easy way to do that is to use the
kSbSystemPathContentDirectory to contain the system font directory and setting the
limited in your port.
Cobalt Evergreen (built by Google), will by default use the
minimal font package which is around 16KB to minimize storage requirements. A separate
cobalt_font_package variable is set to
minimal in the Evergreen platform.
On Raspberry Pi this is:
minimal set of fonts under:
limited set of fonts under:
The ICU table should be deployed under the
kSbSystemPathStorageDirectory. This way all Cobalt Evergreen installations would be able to share the same tables. The current storage size for the ICU tables is 7MB.
On Raspberry Pi this is:
The Cobalt Evergreen package will not carry ICU tables by default but may add them in the future if needed. When the package has ICU tables they would be stored under the content location for the installation:
Pending updates will be picked up on the next application start, which means that on platforms that support suspending the platform should check
loader_app::IsPendingRestart and call
SbSystemRequestStop instead of suspending if there is a pending restart.
suspend_signals.cc for an example.
Evergreen can support multiple apps that share a Cobalt binary. This is a very common way to save space and keep all your Cobalt apps using the latest version of Cobalt. We understand that there are situations where updates are only needed for certain apps, so we have provided a way where Cobalt Updater behavior can be easily configured on a per-app basis with simple loader_app command-line flags.
Currently, the only configurable option for Cobalt Updater configuration is:
Each app’s Cobalt Updater will perform an independent, regular check for new Cobalt Evergreen updates. Note that all apps will share the same set of slots, but each app will maintain metadata about which slots are “good” (working) or “bad” (error detected) and use the appropriate slot. Sharing slots allows Evergreen to download Cobalt updates a single time and be able to use it across all Evergreen-enabled apps.
To illustrate, a simple example:
[APP_1] (currently using SLOT_1, using Cobalt v4) [APP_2] (currently using SLOT_0, using Cobalt v3) [APP_3] (currently using SLOT_0, using Cobalt v3)
Now remember, apps could share the same Cobalt binary. Let’s say
APP_1 has detected an update available and downloads the latest update (Cobalt v5) into SLOT_2. The next time
APP_2 runs, it may detect Cobalt v5 as well. It would then simply do a
request_roll_forward operation to switch to SLOT_2 and does not have to download a new update since the latest is already available in an existing slot. In this case,
APP_2 are now using the same Cobalt binaries in SLOT_2.
APP_3 has not been launched, not run through a regular Cobalt Updater check, or launched with the
--disable_updates flag, it stays with its current configuration.
[APP_1] (currently using SLOT_2, using Cobalt v5) [APP_2] (currently using SLOT_2, using Cobalt v5) [APP_3] (currently using SLOT_0, using Cobalt v3)
Now that we have gone through an example scenario, we can cover some examples of how to configure Cobalt Updater behavior and
Some example configurations include:
# All Cobalt-based apps get Evergreen Updates [APP_1] (Cobalt Updater ENABLED) [APP_2] (Cobalt Updater ENABLED) [APP_3] (Cobalt Updater ENABLED) loader_app --url="<YOUR_APP_1_URL>" loader_app --url="<YOUR_APP_2_URL>" loader_app --url="<YOUR_APP_3_URL>" # Only APP_1 gets Evergreen Updates, APP_2 wants to use an alternate splash screen [APP_1] (Cobalt Updater ENABLED) [APP_2] (Cobalt Updater DISABLED) [APP_3] (Cobalt Updater DISABLED) loader_app --url="<YOUR_APP_1_URL>" loader_app --url="<YOUR_APP_2_URL>" --disable_updates \ --fallback_splash_screen_url="/<PATH_TO_APP_2>/app_2_splash_screen.html" loader_app --url="<YOUR_APP_3_URL>" --disable_updates # APP_3 is a local app, wants Cobalt Updater disabled, and uses an alternate content directory # (This configuration is common for System UI apps. APP_3 in this example.) [APP_1] (Cobalt Updater ENABLED) [APP_2] (Cobalt Updater ENABLED) [APP_3] (Cobalt Updater DISABLED) loader_app --url="<YOUR_APP_1_URL>" loader_app --url="<YOUR_APP_2_URL>" loader_app --csp_mode=disable --allow_http --url="file:///<PATH_TO_APP_3>/index.html" --content="/<PATH_TO_APP_3>/content" --disable_updates
loader_app_switches.cc for full list of available command-line flags.
As Cobalt binary packages (CRX format) are downloaded from the Google Downloads server, the verification of the Cobalt update package is critical to the reliability of Cobalt Evergreen. There are mechanisms in place to ensure that the binary is verified and a chain of trust is formed. The Cobalt Updater is responsible for downloading the available Cobalt update package and verifies that the package is authored by Cobalt(and not an imposter), before trying to install the downloaded package.
In the above diagram, the Cobalt Updater downloads the update package if available, and parses the CRX header of the package for verification, before unpacking the whole package. A copy of the Cobalt public key is contained in the CRX header, so the updater retrieves the key and generates the hash of the key coming from the header of the package, say Key hash1.
At the same time, the updater has the hash of the Cobalt public key hard-coded locally, say Key hash2.
The updater compares Key hash1 with Key hash2. If they match, verification succeeds.
Not at this time. All Cobalt updates will be deployed through Google infrastructure. We believe Google hosting the Cobalt core binaries allows us to ensure a high-level of reliability and monitoring in case issues arise.
We expect performance to be similar to a standard non-Evergreen Cobalt port.
Google will work closely with device partners to ensure that the appropriate testing is in place to prevent regressions.
Yes, there are tests available to help validate the implementation:
The Cobalt team is focusing a large amount of effort to identify and integrate various methods to reduce the size of the Cobalt binary such as compression and using less fonts.
Yes, this is one of the benefits of Evergreen. We can initiate an update from the server side that addresses problems that were undetected during full testing. There are a formal set of guidelines to verify an updated binary deployed to the device to ensure that it will work properly with no regressions that partners should follow to ensure that there are no regressions. In addition, it is also critical to do your own testing to exercise platform-specific behavior.
Much of the optimization work remains in the Starboard layer and configuration so you should still expect good performance using Cobalt Evergreen. That being said, the Cobalt Evergreen configuration allows you to customize Cobalt features and settings as before.