Layered Components: Design
Chrome for iOS cannot use the content layer as-is due to the restrictions that Apple places on third-party browsers on iOS. Given this fact, the goal of this project is to enable iOS’s usage of chrome- and component-level features that:
- Minimizes pain on Chromium developers, both for iOS and for other platforms
- Enables Chromium developers developing multi-platform features to target iOS as well
To that end, we will be performing refactorings to introduce layered components: components that consist of shared code that is content-free together with a driver based on the content layer and an alternative driver for iOS that uses an iOS-specific API.
- Cannot use the existing content/ implementation, which is based on V8/Blink and is multiprocess.
- Could provide at best a partial alternative implementation of the content API due to the black-box nature of UIWebView.
As a result, iOS usage of chrome- and component-level features currently requires a sea of ifdef’s, code modifications, and code additions. Upstreaming this surgery would introduce maintainability and readability challenges.
The fact that most Chromium developers have no visibility into iOS’s restrictions and requirements additionally causes a raised level of difficulty for developers looking to target feature work to iOS in addition to other platforms.
Our approach to the challenges outlined above is to gradually eliminate dependence on the content API in code that is used on iOS (see the below background for discussion of alternative approaches and how we converged on this approach). In determining how to structure this approach within the codebase, we took the position that DEPS restrictions are easiest for developers to understand and introduce the least amount of maintainability pain when they are put within a recognizable structure. We will introduce layered components. These components will live in one of a small number of known locations within the source tree (e.g., src/components). Each such component will consist of the following:
- A core/ subdirectory consisting of code shared by all platforms. This code cannot depend on content/, and will contain e.g. core model logic of the component.
- An ios/ subdirectory that contains code that depends on an iOS-specific API and drives the shared code for iOS.
- A content/ subdirectory that contains code that can depend on content/ and drives the shared code for all other platforms. For example, this subdirectory might contain a WebContentsObserver that invokes behavior of the shared code when a WebContents is destroyed.
- Won’t creating this structure introduce extra layering in the affected features? Inevitably it will. However, we have explicit goals of (a) minimizing added layering to reduce maintainability pain and (b) making decouplings as logical as possible (e.g., decoupling model code from code interacting with the renderer via IPC). We plan to work together with feature owners to develop agreeable designs.
- What is this “iOS-specific API”? It will be an API that will have a similar purpose as the content API (e.g., allow an object to know when a webpage has finished loading) but will be much smaller. In a sense, it will be a “content API around UIWebView.”
- Will there be a wrapper API around the content API and the iOS API? Unknown at this time. The goal is to avoid such an API; however, if it turns out that having a thin wrapper API would greatly simplify the structure of multiple components, such an API might be introduced.
- Without a wrapper API, how will you possibly handle WebContents/NavigationController/BrowserContext/etc.? These questions will have to be examined on a case-by-case basis. Starting from the position that we would like to avoid a wrapper API will enable isolating the cases (if any) where the lack of such a wrapper API causes significant pain.
- What if the shared code of a layered component needs to talk to its driver? We expect that in general communication between the shared code of a layered component and its driver will be bi-directional. Communication from the shared code to a driver will be done via classes that the shared code declares and drivers define (the exact fashion in which this will be done will likely have to be approached on a case-by-case basis with the goal of minimizing boilerplate/code duplication).
- Will this work make the Chromium codebase less maintainable as a
whole? While the primary motivation of this project is to enable
upstream code/features to be shared by (and developed for) iOS in
the most maintainable way, we see several other potential benefits
that are independent of iOS:
- Mitigate the monolithic nature of chrome/browser
- Enable simpler unit testing of the code that will be shared
- Perform refactorings that increase understandability of the impacted features
- How large is the scope of this project? Of the ~80 directories under chrome/browser, iOS uses ~30.
- I think that approach <X> would be better! We welcome feedback and discussion; please read the background section below as context.
We have examined several approaches to iOS’s usage of content:
- iOS uses all the parts of the content API that it needs/can support, providing alternative partial implementations of interfaces for which it cannot share the existing full-fledged implementation (e.g., WebContents).
- iOS uses only the parts of the content API for which it can share the implementations, refactoring the Chromium codebase to remove and disallow usage of problematic parts of the content API (e.g., WebContents) in shared code.
- iOS does not use the content API (the approach that we ultimately decided on).
The conclusion re: Approach 1 was that the fact that iOS would only be able to provide a partial implementation of the content API would make it more challenging for developers to develop against the API: they would have to always be aware of (a) whether the given piece of code that they were working on was shared on iOS, and (b) which subsets of the content API were OK to use in code shared on iOS. The conclusion re: approach 2 was similar to that re: approach 1, with the added concern that performing somewhat arbitrary-looking refactorings throughout the codebase would introduce maintainability and readability problems (initial CLs targeting this approach bore out these fears).
Stuart Morgan (stuartmorgan@) and Colin Blundell (blundell@) are the leads on this project. Darin Fisher (darin@) and John Abd-El-Malek (jam@) are providing guidance, direction and approval.