To expose your feature via the origin trials framework, there are a few code changes required.
[TOC]
NOTE: You can land these code changes before requesting to run an origin trial. These code changes make it possible to control a feature via an origin trial, but don't require an origin trial to be approved. For more on the process, see Running an Origin Trial.
First, you’ll need to configure runtime_enabled_features.json5. This is
explained in the file, but you use origin_trial_feature_name
to associate your
runtime feature flag with a name for your origin trial. The name can be the
same as your runtime feature flag, or different. Eventually, this configured
name will be used in the origin trials developer console. You can have both
status: experimental
and origin_trial_feature_name
if you want your feature
to be enabled either by using the --enable-experimental-web-platform-features
flag or the origin trial.
You may have a feature that is not available on all platforms, or need to limit
the trial to specific platforms. Use origin_trial_os: [list]
to specify which
platforms will allow the trial to be enabled. The list values are case-
insensitive, but must match one of the defined OS_<platform>
macros (see
build_config.h).
Flag name and trial name are the same:
{
name: "MyFeature",
origin_trial_feature_name: "MyFeature",
status: "experimental",
},
Flag name and trial name are different:
{
name: "MyFeature",
origin_trial_feature_name: "MyFeatureTrial",
status: "experimental",
},
Trial limited to specific platform:
{
name: "MyFeature",
origin_trial_feature_name: "MyFeature",
origin_trial_os: ["android"],
status: "experimental",
},
You can also run experiment for new CSS properties with origin trial. After you
have configured your feature in runtime_enabled_features.json5 as above, head
to css_properties.json5. As explained in the file, you use runtime_flag
to associate
the CSS property with the feature you just defined. This will automatically link the CSS
property to the origin trial defined in the runtime feature. It will be available
in both JavaScript (Element.style
) and CSS (including @supports
) when the trial
is enabled.
Example: origin-trial-test-property defines a test css property controlled via
runtime feature OriginTrialsSampleAPI
and subsequently an origin trial named Frobulate
.
Once configured, there are two mechanisms to gate access to your feature behind an origin trial. You can use either mechanism, or both, as appropriate to your feature implementation.
- A native C++ method that you can call in Blink code at runtime to expose your
feature:
bool RuntimeEnabledFeatures::MyFeatureEnabled(ExecutionContext*)
- An IDL attribute [RuntimeEnabled] that you can use to automatically generate code to expose and hide JavaScript methods/attributes/objects.
[RuntimeEnabled=MyFeature]
partial interface Navigator {
readonly attribute MyFeatureManager myFeature;
}
NOTE: Your feature implementation must not persist the result of the enabled
check. Your code should simply call
RuntimeEnabledFeatures::MyFeatureEnabled(ExecutionContext*)
as often as
necessary to gate access to your feature.
NOTE: For CSS properties, you do not need to edit the IDL files, as the exposure on the CSSStyleDeclaration is handled at runtime.
ISSUE: In the rare cases where the origin trial token is added via script after
the css style declaration, the css property will be enabled and is fully functional,
however it will not appear on the CSSStyleDeclaration interface, i.e. not accessible
in Element.style
. This issue is tracked in crbug/1041993.
Once the feature is created, in order to run the origin trial you need to track how often users use your feature. You can do it in two ways.
- Add your feature counter to end of web_feature.mojom:
enum WebFeature {
// ...
kLastFeatureBeforeYours = 1235,
// Here, increment the last feature count before yours by 1.
kMyFeature = 1236,
kNumberOfFeatures, // This enum value must be last.
};
-
Run update_use_counter_feature_enum.py to update the UMA mapping.
-
Increment your feature counter in c++ code.
#include "third_party/blink/renderer/platform/instrumentation/use_counter.h"
// ...
if (RuntimeEnabledFeatures::MyFeatureEnabled(context)) {
UseCounter::Count(context, WebFeature::kMyFeature);
}
- Add [Measure] IDL attribute
partial interface Navigator {
[RuntimeEnabled=MyFeature, Measure]
readonly attribute MyFeatureManager myFeature;
- The code to increment your feature counter will be generated in V8 automatically. But it requires you to follow [Measure] IDL attribute naming convention when you will add your feature counter to web_feature.mojom.
enum WebFeature {
// ...
kLastFeatureBeforeYours = 1235,
// Here, increment the last feature count before yours by 1.
kV8Navigator_MyFeature_AttributeGetter = 1236,
kNumberOfFeatures, // This enum value must be last.
};
What you can't do, because of the nature of these origin trials, is know at
either browser or renderer startup time whether your feature is going to be used
in the current page/context. This means that if you require lots of expensive
processing to begin (say you index the user's hard drive, or scan an entire city
for interesting weather patterns,) that you will have to either do it on browser
startup for all users, just in case it's used, or do it on first access. (If
you go with first access, then only people trying the experiment will notice the
delay, and hopefully only the first time they use it.). We are investigating
providing a method like OriginTrials::myFeatureShouldInitialize()
that will
hint if you should do startup initialization. For example, this could include
checks for trials that have been revoked (or throttled) due to usage, if the
entire origin trials framework has been disabled, etc. The method would be
conservative and assume initialization is required, but it could avoid expensive
startup in some known scenarios.
Similarly, if you need to know in the browser process whether a feature should be enabled, then you will have to either have the renderer inform it at runtime, or else just assume that it's always enabled, and gate access to the feature from the renderer.
To test an origin trial feature during development, follow these steps:
-
Use generate_token.py to generate a token signed with the test private key. You can generate signed tokens for any origin that you need to help you test, including localhost or 127.0.0.1. Example:
tools/origin_trials/generate_token.py http://localhost:8000 MyFeature
There are additional flags to generate third-party tokens, set the expiry date, and control other options. See the command help for details (
--help
). For example, to generate a third-party token, with user subset exclusion:tools/origin_trials/generate_token.py --is-third-party --usage-restriction=subset http://localhost:8000 MyFeature
-
Copy the token from the end of the output and use it in a
<meta>
tag or anOrigin-Trial
header as described in the Developer Guide. -
Run Chrome with the test public key by passing:
--origin-trial-public-key=dRCs+TocuKkocNKa0AtZ4awrt9XKH2SQCI6o4FY6BNA=
You can also run Chrome with both the test public key and the default public key along side by passing:
--origin-trial-public-key=dRCs+TocuKkocNKa0AtZ4awrt9XKH2SQCI6o4FY6BNA=,fMS4mpO6buLQ/QMd+zJmxzty/VQ6B1EUZqoCU04zoRU=
The --origin-trial-public-key
switch is not needed with content_shell
, as it
uses the test public key by default.
The test private key is stored in the repo at tools/origin_trials/eftest.key
.
It's also used by Origin Trials unit tests and web tests.
If you cannot set command-line switches (e.g., on Chrome OS), you can also directly modify chrome_origin_trial_policy.cc.
To see additional information about origin trial token parsing (including reasons for failures, or token names for successful tokens), you can add these switches:
--vmodule=trial_token=2,origin_trial_context=1
If you are building with is_debug=false
, then you will also need to add
dcheck_always_on=true
to your build options, and add this to the command line:
--enable-logging=stderr
When using the [RuntimeEnabled] IDL attribute, you should add web tests to verify that the V8 bindings code is working as expected. Depending on how your feature is exposed, you'll want tests for the exposed interfaces, as well as tests for script-added tokens. For examples, refer to the existing tests in origin_trials/webexposed.