Bounce Tracking Protection

Bounce Tracking Protection (BTP) is an anti-tracking feature in Gecko which detects bounce trackers (navigational tracking) based on a set of heuristics. As opposed to the cookie purging feature it does not rely on a list of trackers which makes it more webcompat friendly while also covering unknown bounce trackers.

Standardization

The protection is a work item of the PrivacyCG. The implementation in Gecko closely follows the Bounce Tracking Mitigations spec draft.

Mozilla also has a specification position on Bounce Tracking Mitigations.

Behavior

BTP detects bounce trackers by looking at navigation timing. It establishes the concept of an extended navigation which can encompass a chain of short-lived redirects. These short-lived redirects are commonly used by bounce trackers. Any site that appears as an intermediary in such a short-lived redirect gets added to a classification list. Classified bounce trackers have their cookies, site data and cache purged periodically. In order to avoid false positives and purging data that may be important for users, sites which the user directly interacted with in the last 45 days are exempt from being classified or purged.

See Bounce Tracking Mitigations Explainer for a more detailed (albeit chromium-oriented) description of the feature and how trackers are classified.

Deviations from the spec

Initial host of an extended navigation

The spec’s process navigation start algorithm derives the extended navigation’s initial host from the navigation’s source document. Because the initial host is exempt from classification, that lets a cross-site iframe which navigates the top level nominate itself as the initial host and escape being classified for its own bounce (Bug 2060310).

Gecko instead uses the site the extended navigation is leaving: the site of the most recent non-initial document committed in the context being navigated. A context which has never committed a document of its own is one the navigation opened (window.open, target="_blank"), so the opener’s top level document is used instead. Both signals are read at commit time, so a navigation which starts before the current document has finished loading still attributes that document.

Reading the opener rather than the navigation’s initiator is load bearing. A context opened by a cross-site frame holds an initial about:blank whose principal is inherited from that frame, and the initiator can resolve back into the opened context itself, so either would let the frame name itself again through an empty popup.

Because the derivation does not depend on the navigation having an initiator at all, a navigation started by the browser rather than by content — address bar, bookmark, session restore — also exempts the site it is leaving. This matches how the spec attributes user activation, which already uses the top level traversable’s active document.

Comparison with Chromium

Chromium’s implementation starts a redirect chain the same way. In btm_bounce_detector.cc, BtmBounceDetector::DidStartNavigation takes the chain start from the tab’s last committed URL and only consults the navigation’s initiator when the tab has nothing committed. BtmServiceImpl::HandleRedirects in btm_service_impl.cc then skips a redirector whose site equals the chain’s initial or final site, which is what RecordStatefulBounces does with the initial and final host.

Gecko is stricter in one case. Chromium’s fallback is the initiator’s origin as-is, so for a context opened by a cross-site frame it resolves to the frame’s own site and the frame can still name itself. Gecko resolves the opener to its top level document instead.

Gecko Implementation

Work for the Gecko implementation in tracked under following meta-bug: Bug 1839915 - [meta] Bounce Tracking Protection.

A simplified UML diagram of the BTP implementation in Gecko. Note that some classes and attributes have been omitted for readability. You can use the diagram feature in Searchfox to view the full diagram (example).

        classDiagram
    class BounceTrackingProtection {
        - mBounceTrackingPurgeTimer: nsITimer
        - mStorage: BounceTrackingProtectionStorage
    }

    note for BounceTrackingProtection "Singleton class to manage the feature."

    class BounceTrackingAllowList {
        <!-- [...] -->
    }

    class nsIBTPExceptionList {
        <!-- Bug 1930704: Empty classes lead to build errors. Adding a comment inside the class fixes the issue -->
        <!-- [...] -->
    }

    class BounceTrackingStateGlobal {
        - mUserActivation: nsTHashMap&lt;nsCStringHashKey, PRTime&gt;
        - mBounceTrackers: nsTHashMap&lt;nsCStringHashKey, PRTime&gt;
        - mOriginAttributes: OriginAttributes
    }

    note for BounceTrackingStateGlobal "Manages the global maps for bounce tracker candidates
                                        and user activation for a specific OriginAttributes dict."

    class BounceTrackingProtectionStorage {
        - mStateGlobal : nsTHashMap&lt;OriginAttributesHashKey, RefPtr&lt;BounceTrackingStateGlobal&gt;&gt;
        - mDatabaseFile : nsCOMPtr&lt;nsIFile&gt;
    }


    class BounceTrackingRecord {
        - mInitialHost: nsAutoCString
        - mBounceHosts: nsTHashSet&lt;nsCStringHashKey&gt;
        - mFinalHost: nsAutoCString
    }

    note for BounceTrackingRecord "Encapsulates the per-tab navigation state
                                   during an extended navigation."

    class BounceTrackingState {
        - mBounceTrackingProtection: BounceTrackingProtection
        - mBounceTrackingRecord: BounceTrackingRecord
        - mClientBounceDetectionTimeout: nsReadOnlyTimer
        - mOriginAttributes: OriginAttributes
    }


    class BrowsingContextWebProgress {
        - mBounceTrackingState: BounceTrackingState
    }


    class DocumentLoadListener{
        <!-- [...] -->
    }

    note for BrowsingContextWebProgress "Every tab has a web progress
                                         and therefore a BounceTrackingState"

    BounceTrackingProtection *-- BounceTrackingProtectionStorage
    BounceTrackingState --o BounceTrackingProtection
    BounceTrackingState --> BounceTrackingProtection : RecordStatefulBounces() on extended nav end
    BounceTrackingState *-- BounceTrackingRecord
    BrowsingContextWebProgress *-- BounceTrackingState
    BounceTrackingStateGlobal --* BounceTrackingProtectionStorage
    BounceTrackingStateGlobal --> BounceTrackingProtectionStorage : Persists state changes in storage.
    BounceTrackingProtection --> BounceTrackingStateGlobal : Manages global state
    nsIBTPExceptionList --* BounceTrackingProtection
    nsIBTPExceptionList --> BounceTrackingProtection : Site host purge exceptions from RemoteSettings.
    BounceTrackingAllowList --* BounceTrackingProtection
    BounceTrackingAllowList --> BounceTrackingProtection :  Site host purge exceptions from PermissionManager.
    BrowsingContextWebProgress --> BounceTrackingState : Navigation signals
    DocumentLoadListener --> BounceTrackingState : Navigation signals

Preferences

The feature can be enabled and it’s behavior can be adjusted using the privacy.bounceTrackingProtection.* prefs. See StaticPrefList.yaml for a list of prefs with descriptions.

The main feature pref is privacy.bounceTrackingProtection.mode where 0 is fully disabled and 1 is fully enabled. See nsIBounceTrackingProtection.idl for a full list of options.

Nimbus Integration

A subset of the BTP prefs can be controlled via Nimbus. See definition here: FeatureManifest.yaml.

Logging

BTP has a logger which can be enabled by starting Firefox with the MOZ_LOG environment variable. Use MOZ_LOG=BounceTrackingProtection:5 for verbose logging for every navigation and MOZ_LOG=BounceTrackingProtection:3 for more concise logging focused on classification and purging.

Console Messages

You can check the developer tools console for warning messages which will be logged when a site gets classified. Example:

“bounce-tracking-demo-tracker-server.glitch.me” has been classified as a bounce tracker. If it does not receive user activation within the next 3,600 seconds it will have its state purged.

When a site has recently been purged (since last restart), upon next visit, Firefox will also log a warning to the website console:

The state of “bounce-tracking-demo-tracker-server.glitch.me” was recently purged because it was detected as a bounce tracker.

Testing

When testing sites to ensure they don’t get purged for bounce tracking behavior you can use both logging (as described above) to observe classification and direct calls to the feature via the Browser Toolbox to trigger the purging early.

The snippets in the following section need to be executed in the Browser Toolbox. Note that while the toolbox looks like the regular devtools it’s a special console used to debug Firefox itself rather than websites.

Trigger a purge of all classified trackers:

Before navigating to the site set privacy.bounceTrackingProtection.bounceTrackingGracePeriodSec to 0 or a low number. This controls how fast after (classified) bounce a site may be purged. If you don’t update this pref you need to wait up to 1h for a site to be purged.

Purges normally run every hour. To trigger a purge manually you can use the following snippet:

await Cc[
  "@mozilla.org/bounce-tracking-protection;1"
].getService(Ci.nsIBounceTrackingProtection).testRunPurgeBounceTrackers();

The return value will be an array of sites that have been purged. Note that purging applies for the entire domain (eTLD+1).

List recently purged sites

You can obtain a list of recently purged sites (since the last restart) by calling:

await Cc[
  "@mozilla.org/bounce-tracking-protection;1"
].getService(Ci.nsIBounceTrackingProtection).testGetRecentlyPurgedTrackers({});

This only shows sites which have been purged in normal browsing. If you want data from private browsing or containers you need to pass in a non-default OriginAttributes object, e.g. { privateBrowsingId: 1 }.

There is also hasRecentlyPurgedSite which can be used to check if a specific site has been recently purged (across all OriginAttributes contexts).

Test Page

https://bounce-tracking-demo.glitch.me/ is a demo page with two links that exhibit bounce tracking behaviour. You can use it combined with the methods above to verify that the mechanism is running.