| [eclipse-pmc] On using WebKitGTK |
It's been clear for a while now that using XULRunner as the basis for the SWT Browser control on Linux has caused significant problems for our consumers. In current Linux distros, there is a compelling alternative, WebKitGTK, which is licensed under LGPL. The Eclipse Board has recently approved a new policy to allow the use of LGPL APIs in certain circumstances. According to this policy, to approve the request to use the APIs, the board must determine that:
a) The library or application invoked by the APIs is a “pre-req,” as defined in the Eclipse Foundation’s Guidelines for the Review of Third Party Dependencies.
b) There is no alternative open source software that provides the same or similar function to the WebKit APIs that can be obtained under a more permissive license.
c) The functionality, usability or consumability of the Eclipse project would be severely restricted or reduced without the APIs.
In determining these things the board may rely on information provided by the PMC, and to help in this disregard, the PMC (with the help of our browser guru, Grant Gayed) investigated and determined the following w.r.t. the above points:
a) SWT's Browser control has remained stable and reliable on Windows and OS X since the Eclipse 3.0 release, and there are many Eclipse-based applications that have a hard dependency on having a working Browser control at runtime. Eclipse/SWT strives to provide a cross-platform experience that is as consistent as possible for its applications, so reliably providing a working Browser control on a heavily-consumed platform like Linux is essential.
b) The goal of the Browser control is to embed a natively-available web browser without requiring the user to specially install it (i.e. it should just be there from the OS installation). WebKitGTK is the only embeddable web browser that provides a frozen API that is rich enough to meet the Browser's needs and can be counted on to be pre-installed in modern Linux distributions.
c) Without the use of WebKitGTK the instability that currently plagues the Browser on Linux will continue. Given the breadth of Eclipse applications that depend on having a working Browser control and that wish to target Linux as a supported platform, this instability inevitably reduces the consumability of Eclipse as an application platform.
Every major release of XULRunner on Linux has caused SWT's Browser control to become less-/non-functional, and we have been forced to adapt it to the native-level XULRunner changes in our subsequent Eclipse/SWT releases (see http://www.eclipse.org/swt/faq.php#browserlinux). This results in Eclipse-based applications not being able to rely on their Browser controls to work well on Linux distributions with newer native browser versions than existed when the application's shipped Browser control was released.
Consequently, Eclipse-based applications deployed to the field may become broken at a later date, and maintainers of these applications must constantly monitor when such breakages occur and issue updates based on corresponding SWT updates. Alternatively, consumers must specify strict compatibility requirements to ensure such breakages don't occur, which may force users to move to older browsers, which may not always be possible. As an example, given that Ubuntu 9.10 ships with XULRunner 1.9.1.3, an application must be based on Eclipse 3.5.x or newer in order for its Browser control to work on this Linux distribution.
A sample of the Eclipse/SWT problems that have been caused by the need to rely on non-frozen interfaces appear below [1]. Other similar but unrelated problems caused by interface changes have been fixed without accompanying bug reports [2]. Additionally, significant native behavioural changes have caused other Browser functions to become broken [3]. A general question pertaining to these breakages was asked last year (http://markmail.org/message/uapflk6lgkaicsqz#query:+page:1+mid:vdesauobtmoa2n55+state:results ), but the suggestions that were received in response did not adequately meet the Browser's needs [4]. Being able to rely on OS-provided frozen native API is what enables other facets of SWT releases to continue to just work for years without major change, and with the use of WebKitGTK, the Browser control should be able to provide a similar level of stability on Linux.
------------------
[1] Example Eclipse/SWT problems caused by the need to rely on non-frozen interfaces:
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=213194 : setText crash with xulrunner 1.9 stream (+13 duplicates)
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=214802 : xulrunner 1.9 uses new interface for authentication
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=242644 : eclipse crashes when downloading a file in the internal browser
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=268651 : Make sure Eclipse works with xulrunner 1.9.1beta
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=296284 : xulrunner 1.9.2 changes
[2] Other similar problems:
- Mozilla 1.7: nsIDOMSerializer
- XULRunner 1.8: nsIFilePicker
- XULRunner 1.8: nsIHelperAppLauncher
- XULRunner 1.8: nsIProgressDialog
- XULRunner 1.9: nsICookieService
- XULRunner 1.9.1: nsIScriptSecurityManager
[3] Changes in native behaviour:
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=214807 : xulrunner 1.9 has changed invalid certificate behaviour
- https://bugs.eclipse.org/bugs/show_bug.cgi?id=259687 : _javascript_ executes asynchronously with xulrunner 1.9
[4] Suggested workarounds for XULRunner, and why they did not meet our requirements:
i. Suggestion: Only use frozen interfaces.
- The Browser cannot do this without providing significantly less functionality than is needed to fulfill its responsibilities.
ii. Suggestion: Don't use nsIDocShell anymore.
- This change was made in Eclipse/SWT 3.5. However, this does not help with the various other unfrozen interfaces currently being used.
iii. Suggestion: Specify the maxversion more tightly when auto-detecting a native browser to use.
- This could avoid some crashes, but would result in no native browser being found for use on newer Linux distros. Since having a working Browser control on Linux is essential, this is not an adequate solution.
iv. Suggestion: Ship XULRunner, "treating Mozilla as a system library is pain and hurt"
- This has been reported to have limited success since the back-end API needs of XULRunner are diverse, and unsatisfied shared library dependencies can still occur. A native browser that is properly integrated with the target OS is needed.
v.Suggestion: Make XPCOM calls via _javascript_
- This could alleviate some problems related to member positioning within a class, but would not have helped other breakages, such as:
- nsIHelperAppLauncherDialog.PromptForSaveToFile() (the Browser implements this interface):
- grew an argument in Mozilla 1.5
- changed an argument type in XULRunner 1.8
- grew another argument in XULRunner 1.9
- nsIDOMSerializer.SerializeToString() changed an argument type in Mozilla 1.7
- nsIFilePicker.Init() and nsIFilePicker.SetDefaultString() changed their argument types in XULRunner 1.8 (the Browser implements this interface)
- nsIDownload changed significantly (>10 method signatures) in XULRunner 1.8 (the Browser implements this interface)
- the contractID for registering a download factory changed in XULRunner 1.8
- the XULRunner 1.9 behavioral breakages in [3]