[webkit-dev] Networking proxy on iOS

Fri May 20 08:58:13 PDT 2016

> On May 20, 2016, at 2:30 AM, Daniel Olegovich Lazarenko <danielo at opera.com> wrote:
> 
> Thank you for such a fast reply. That is amazing! :)
> Back to questions...
> 
> > Are you primarily focused on a custom networking layer (e.g. your own HTTP implementation?),
> > or with custom protocol handling for non-http protocols?
> 
> I'm primarily concerned about HTTP and friends (HTTPS, SPDY, HTTP2), but if there are any other widely used protocols that's also interesting. FTP support is not interesting for me. Do you have any other specific things in mind?
> 
> If there's a custom proprietary protocol that people handle in the app with their own code, for example, acme://acme.com:1234 <http://acme.com:1234/>, then proxying this thing is not very interesting to me, because it's very easy to proxy my own protocol handled by my own code. There's a case when "acme" is provided by some 3rd party, and the app author doesn't have the processing code. In such case it might be interesting to proxy it as well, but then again, I'm asking for a concrete example of such protocol (in WebKit context).

In a WebKit1 app (WebView on Mac, UIWebView on iOS), app authors were able to use NSURLProtocol to override any scheme they wished.

While some such as yourself might’ve used it to override http from the default handling, *many more* used it to implement custom protocols. e.g. “acme://my-app-specific-url”

We are exploring ways to restore that full functionality - https://bugs.webkit.org/show_bug.cgi?id=138169

> > You seem to dismiss the Networking process’ crash recovery aspect.
> > "because in practice we know that most of the crashes happen in the WebProcess parts”.
> > I’m curious what data you’re using to make that claim?
> 
> Well, I'm not dismissing it, it's definitely a trade off that an app author will make by choosing to enable this option.
> The data comes from our web browser apps. We certainly see networking faults, but in total it was usually a minor part of all the WebKit crashes. To not sound subjective, I've looked through our current app version, which already has enough data to judge, and in the top WebKit crashes there are none in the network code. Most are crashes in JavaScriptCore, DOM and graphics subsystems. This is the experience we have over many versions and years of service. I might be able to show you the data in private if you want, although I'm sure that you have your own crash analysis system with much more data.

Without getting in to specifics, the NetworkingProcess does crash.

And while the WebContent process does crash way more, it usually only effects that one web page.

If networking code was back inside the UI Process, and it crashed, that takes down the whole browser.

Doing so would be reverting towards the single process architecture of yesteryear, not progressing away from it.

> Let's discuss the sandboxing a little bit. First of all, and correct me if I'm wrong: I thought that there's no 3rd party code execution in the networking process,

Currently, there is no 3rd party code execution in the networking process.

> and all the JS execution happens inside the WebProcess.

That’s correct, but I’m not necessarily talking about JS.

> The code that we want to run is our own code that we control that we implement in a safe and secure manner.

The 1st party is “the Modern WebKit framework.”
The 3rd party is “the application’s code, as well as stuff downloaded from the Internet.

> In this sense it's no less secure.

A single process app that does whatever it wants with the networking stack is less secure to the user than a multi-process app that has well defined behavior inside certain sandboxed parts of the app.

> In the end we are always protected by the iOS/Mac sandbox.

For a WKWebView app on iOS, while the app itself is sandboxed, the Networking process is *much more* sandboxed.

> Maybe I'm wrong about JS here, or do you have some other use case in mind?

I think you misunderstood that I was talking about the native app code that is a 3rd party client to the WebKit framework.

But, there is an important JS aspect to this; Since there is messaging between the WebContent process and the Networking process, a vulnerability in the WebContent process can turn into a “remote exploit” of the Networking process.
Therefore it is beneficial to have the Networking process tightened down as much as possible.

---

There’s another very important reason why offering a “networking-in-UI-process mode” is not appealing. From a project maintenance standpoint, the more configurations we have the harder it is to test and develop within the project.

Adding this “optional” networking mode would double certain testing matrices and make development in this area (which there has been a lot of lately) more difficult.

> > Debugging the multi process architecture of WebKit2 
> > has not gotten any harder in years, active developers have all adapted, 
> > and new developers tend to pick it up pretty quickly. This is not a useful point.
> 
> I'm sorry that you are rejecting this. Of course you can adapt to that, but it inevitably has a steeper learning curve, and takes longer time. Many app developers come from a single-process background and find multi-process debugging much harder. Often it's a real challenge to understand what's going on. I'm sure that you in your team have multiple stories that show how non-trivial it is, and tricks about dealing with it. Nevertheless, I agree, it's not a decisive point.

AFAIK, we haven’t had a potential contributor to the WebKit Open Source project decide to not contribute because of the multi-process architecture.

But, regardless, the MP architecture is primarily how well WebKit works as a browser engine for the user, and not about how easy it is for single-process-only developers to contribute.

Such developers can still actively contribute to the cross platform code of the project (WebCore) in single process mode using MiniBrowser or DumpRenderTree.

Thanks,
~Brady
> 
> 
> 
> On Thu, May 19, 2016 at 6:43 PM, Brady Eidson <beidson at apple.com <mailto:beidson at apple.com>> wrote:
> 
>> On May 19, 2016, at 8:41 AM, Daniel Olegovich Lazarenko <danielo at opera.com <mailto:danielo at opera.com>> wrote:
>> 
>> I'd like to ask your for advice about implementation of a custom networking layer
> 
> Are you primarily focused on a custom networking layer (e.g. your own HTTP implementation?), or with custom protocol handling for non-http protocols?
> 
>> ...with WKWebView on iOS.
> 
> WKWebView is an API that ships on both OS X and iOS. When a design aspect of it affects both platforms (such as the networking behavior), we must consider both platforms.
> 
>> Our current solution is based on NSURLProtocol, and the issues we had with it in 2014 are unresolved:
>> https://bugs.webkit.org/show_bug.cgi?id=137302 <https://bugs.webkit.org/show_bug.cgi?id=137302>
>> https://bugs.webkit.org/show_bug.cgi?id=138131 <https://bugs.webkit.org/show_bug.cgi?id=138131>
>> 
>> It was kind of a shoehorn hack, and so it was rejected by Benjamin Poulain and Alexey Proskuryakov among other reviewers.
> 
> I’m not sure it’s useful for WebKit to spend energy testing and maintaining a mechanism that *only* allows for HTTP-handling replacement and doesn’t also allow for the oft-requested feature of custom protocol handling.
> 
>> Now I'm again looking for a better solution.
>> I'd really like to discuss it with somebody responsible,
> 
> There is no single person responsible; the project works largely on consensus. When dealing with platform specific concerns such as this, it works on consensus of the platform owners.
> 
> That said, I have been the primary caretaker of the Networking process since it’s inception, as well one of the primary caretakers of Mac/iOS networking in general for many years, so I’ll share my thoughts below.
> 
>> There's currently 2 solutions I'm weighting:
>> Pass and use NetworkProcessCreationParameters.httpProxy to NSURLSessionConfiguration (in NetworkSession and maybe other places). The httpProxy solution is easy to implement and would look clean design-wise. It would let us spawn an HTTP proxy on localhost and filter the traffic there. There might be some complications, because it's not fully transparent to the client side. For example HTTPS will have issues. All in all this could be a fine short-term solution.
> While ToT WebKit contains an NSURLSession-based networking implementation for Mac/iOS, it also still contains an NSURLConnection implementation, which is unaffected by NSURLSession considerations.
> 
> That a solution doesn’t work on all supported platforms is not a deal breaker, but it certainly makes it less interesting than one that does.
> 
> HTTPS losing reliability is likely an unacceptable red flag. 
> 
> I’m not sure it’s useful for WebKit to spend energy testing and maintaining a mechanism that *only* allows for HTTP-handling replacement and doesn’t also allow for the oft-requested feature of custom protocol handling.
>> Add a new mode to the NetworkProcess, which would do all networking in UIProcess (instead of spawning a new process). A mode would be optional and controlled with some configuration setting (or NSUserDefaults).
>> The UIProcess solution is harder to implement, and it will affect more code. It is somewhat controversial. One of the reasons of splitting out a NetworkProcess was to have it respawn after crashes. Nevertheless we can take this risk, because in practice we know that most of the crashes happen in the WebProcess parts.
> 
> You seem to dismiss the Networking process’ crash recovery aspect. "because in practice we know that most of the crashes happen in the WebProcess parts”.  I’m curious what data you’re using to make that claim?
> 
>>  I don't see any other significant downsides of having the UIProcess handling networking.
> 
> The Networking process provides significant benefit unrelated to crash recovery that should not be abandoned for convenience sake. e.g. Sandboxing.
> 
> Especially when moving the networking to the UI process would also end up moving 3rd party code execution into the UI process, this seems like an unacceptable regression.
> 
>> In addition it can simplify the NetworkProcess debugging.
> 
> Debugging the multi process architecture of WebKit2 has not gotten any harder in years, active developers have all adapted, and new developers tend to pick it up pretty quickly. This is not a useful point.
> 
> Thanks,
> ~Brady
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.webkit.org/pipermail/webkit-dev/attachments/20160520/8fcfa78c/attachment.html>