Intent to Remove: SwiftShader Fallback

[David Adrian]

Contact emails

dad...@google.com, geof...@chromium.org

Summary

Allowing automatic fallback to WebGL backed by SwiftShader is deprecated and will be removed. This has been noted in DevTools since Chrome 130. 

WebGL context creation will fail instead of falling back to SwiftShader. This is for two primary reasons:

1. SwiftShader is a high security risk due to JIT-ed code running in Chromium's GPU process.

2. Users have a poor experience when falling back from a high-performance GPU-backed WebGL to a CPU-backed implementation. Users have no control over this behavior and it is difficult to describe in bug reports.

SwiftShader is a useful tool for web developers to test their sites on systems that are headless or do not have a supported GPU. This use case will still be supported by opting in but is not intended for running untrusted content.

To opt-in to lower security guarantees and allow SwiftShader for WebGL, run the chrome executable with the --enable-unsafe-swiftshader command-line switch.

During the deprecation period, a warning will appear in the javascript console when a WebGL context is created and backed with SwiftShader. Passing --enable-unsafe-swiftshader will remove this warning message. This deprecation period began in Chrome 130.

Chromium and other browsers do not guarantee WebGL availability. Please test and handle WebGL context creation failure and fall back to other web APIs such as Canvas2D or an appropriate message to the user.

SwiftShader is an internal implementation detail of Chromium, not a public web standard, therefore buy-in from other browsers is not required. The devices covered by SwiftShader (primarily older Windows devices) are likely already incompatible with WebGL in other browsers.

SwiftShader is not used on mobile; this only applies to Desktop platforms.

Blink component

Blink>WebGL

Motivation

https://chromium.googlesource.com/chromium/src/+/main/docs/gpu/swiftshader.md#automatic-swiftshader-webgl-fallback-is-deprecated

Risks

SwiftShader is used by devices without hardware acceleration for WebGL. This is approximately 2.7% of WebGL contexts. However, WebGL is considered fallible and in many cases, these draws are not performant and provide an effectively unusable experience for users. Many applications, such as Google Maps, prefer to fail out rather than use SwiftShader.

Debuggability

None

Flag name on about://flags

--enable-unsafe-swiftshader command-line switch.

Finch feature name

AllowSwiftShaderFallback

Tracking bug

https://issues.chromium.org/40277080

Launch bug

https://launch.corp.google.com/launch/4351104

Estimated milestones

Shipping on Desktop 137

Link to entry on the Chrome Platform Status

https://chromestatus.com/feature/5166674414927872?gate=5188866041184256

This intent message was generated by Chrome Platform Status.

[Ashley Gullen]

I wrote about this previously but I'm still concerned this is a major breaking change for existing published WebGL content on the web. If the figure of 2.7% comes from my previous citing of Web3DSurvey (https://web3dsurvey.com/) then this should be seen as very much an underestimate, because that site uses a relatively small sample size that is more likely to be focused on high-end devices (samples are taken from developer-focused sites like the three.js website, WebGPU fundamentals etc). I would not be surprised if the real worldwide average was more like 4-5%. Then if that's a worldwide average, there will probably be some specific countries or markets where the figure could be more like 10%.

Suppose this change rolls out and we get reports that say our WebGL content no longer works for 10% of users in a South American market. Then what? There is nothing feasible we can do about it. These customers were previously getting by with SwiftShader, but now they get an error message. So I fear this risks disaster for web games in some markets.

Does Google have their own internal data about the usage of SwiftShader? Can more data about this be shared? I respect the work done by Web3DSurvey but unfortunately for the reasons I mentioned I don't think it should be used as evidence to make such a big change as this. Maybe in some places it will affect 25% or 50% of users - who knows? How can we be sure?

Can there not be some other fallback implemented? V8 does JIT with untrusted JavaScript code and that is generally considered reasonably secure, is there any particular technical reason SwiftShader is not considered as secure?

I'd also point out that any website that has a poor experience with SwiftShader can already opt-out using the failIfMajorPerformanceCaveat context flag. If there is some other mode that can be used instead, or just showing an error message is acceptable, then websites can already implement that. In our case with Construct we specifically attempt to obtain hardware-accelerated WebGPU, WebGL 2, or WebGL 1; only failing that do we resort to using SwiftShader on the basis that showing the content with potentially poor performance is better than not showing it at all.

--
You received this message because you are subscribed to the Google Groups "blink-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blink-dev+...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAGkh42KV4DrSSyEgJaF4DnFOXAye-wRLrfD-LKGNkWhyWzshLA%40mail.gmail.com.

[David Adrian]

> I wrote about this previously but I'm still concerned this is a major breaking change for existing published WebGL content on the web. If the figure of 2.7% comes from my previous citing of Web3DSurvey

It does, not it comes from Chrome's metrics system.

> Does Google have their own internal data about the usage of SwiftShader?

It is the 2.7% number.

> Suppose this change rolls out and we get reports that say our WebGL content no longer works for 10% of users in a South American market. Then what? There is nothing feasible we can do about it. These customers were previously getting by with SwiftShader, but now they get an error message. So I fear this risks disaster for web games in some markets.

> I mentioned I don't think it should be used as evidence to make such a big change as this. Maybe in some places it will affect 25% or 50% of users - who knows? How can we be sure?

Can you share the number for Construct about what percentage of your users are using the SwiftShader fallback? Again, our numbers indicate that these are primarily older Windows workstations. Notably, SwiftShader is not used at all on mobile.

> V8 does JIT with untrusted JavaScript code and that is generally considered reasonably secure, is there any particular technical reason SwiftShader is not considered as secure?

Yes. The GPU process is shared between all sites, whereas the V8 JIT is per-site. This means compromising the GPU process can be enough to bypass site isolation protections with a single bug. Additionally, V8 bugs can be reliably patched in the browser, whereas SwiftShader "bugs" can be user-mode graphics driver bugs that are simply more exposed via SwiftShader than they would be otherwise. In this case, the browser can't patch the bug because it's in the driver.

[Rick Byers]

Sorry for the delay from API owners, as discussed on chat the chromestatus entry wasn't set up properly to request API owner review (now fixed).

This is a tricky one indeed (thanks for your input Ashley!). It looks like WebGL is used on about 20% of page loads, so 2.7% of that is ~0.5% of page loads which is very high risk according to our rules of thumb. 

Of course that's an upper-bound, how many will have a fallback? One option would be to collect some UKM data for SwiftShader usage and review a random ~50 sites to observe the user experience in practice. That could give us a better sense of what the real user impact would likely be. Or Maybe Ashley can give us some examples of some web games just to confirm they indeed go from being playable to unplayable without swiftshader on some specific devices? David, do you have a device yourself you can test with that doesn't support GPU WebGL?

Regardless, unless sites have been really good about almost always falling back somehow, I suspect we'll find that there's enough end-user impact to make this a high-risk change (but I could be convinced otherwise such as via a thorough UKM analysis). In which case then we could start working through our playbook for a phased plan for risky breaking changes. Not unlike what we did for flash removal, or WebSQL (both big security benefit but big web compat risk). For example:

• Explore whether we can build swiftshader into a wasm module that sites can use as a (probably even slower) fallback themselves. This turned out to be the key to making WebSQL deprecation tractable. In general our policy is that we don't take functionality away that developers can't replace with some other substitute except in pretty extreme circumstances. 
• Prompt the user on whether or not to enable it per-origin (like a permission)
• Put 3p usage behind a permission policy so the top-level site has to opt-in to allow 3p iframes to use swiftshader
• Rely on some heuristics, (perhaps crowd-sourced signals) to try to find a sweet spot in the safety vs. functionality tradeoff space. We did this for flash initially with things like blocking it for very small canvases. 
• Anything we can do to make WebGL work on a larger set of devices?
• Probably lots of other ideas that aren't occurring to me right now, more examples in bit.ly/blink.compat.

On the other side of the equation, API owners can be convinced to accept more compat risk the more significant the security benefits are. Are there more details you can share? Such as:

• Are we confident that an attacker can only trigger swiftshader on somewhere around 3% of users (vs. some knob which can force it to be used on a larger fraction)? To what extent do we have reason to believe that the vulnerable population size is large enough to be a plausible target for attackers? Is there anything we can do to make the vulnerable user base more reliably contained?
• How does swiftshader compare to other areas in terms of the number of vulnerabilities we've found in practice? Are there any reports of ITW exploits of it? It looks like since 2020 SwiftShader has been about 8% of Chrome's VRP spend - that seems quite significant to me, but probably not in the top 5 areas of concern. This was obviously key to the immense cost and pain of Flash removal - we kept having severe security incidents in practice.

So assuming Ashley and I are right that this isn't likely to be easy, that means it's likely quite a lot of work to attempt to phase-out SwiftShader in a responsible fashion. But with luck maybe we can find a first step that is a good cost-benefit tradeoff (like putting 3P usage behind a permission prompt)? Or maybe it's just a better cost-benefit tradeoff to invest in other areas which pose a threat to a greater number users (hardening ANGLE perhaps)? But of course I will defer to the judgement of security and GPU experts like yourself on that question, I'm happy to consult and support if you want to invest in a plan that API owners can approve.

Rick

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/c5131675-dff4-4aa0-8e84-4cdc373e3035n%40chromium.org.

[Ashley Gullen]

Thanks for the response Rick, I agree with much of what you've said and I think your views and suggested workarounds are all generally reasonable.

I just realised I previously responded to this thread but only replied to David - for transparency I've copied my previous response below.

I can confirm all content made with Construct since about 2018 requires WebGL to work and will show an error message if WebGL is unavailable. I've included a screenshot of the message Construct content published to the web will display when WebGL is not supported, saying "Software update needed", since that has usually been the best advice in that situation in the past. As my previous message says we long ago removed any other fallback and are now likely too dependent on WebGL to feasibly reimplement a canvas2d fallback.

Some other thoughts about workarounds/mitigations:

• A swiftshader WASM module would at least give us a workaround, but if that was something like a ~10 MB+ module it would be a very high download overhead which we'd be obligated to include in every Construct export for compatibility
• Swiftshader could be removed from insecure origins with little impact to us, and using a permission policy for cross-site iframes should be straightforward to work with
• If it helps reduce the attack surface, we could live with SwiftShader support for WebGL 1 only (no WebGL 2) with minimum capabilities (no extensions).
• A permission prompt to the user is not ideal but better than nothing, and I imagine it would be tricky to explain to a normal web user though the prompt message (and makes obtaining a WebGL context async...)
• Regarding getting WebGL to work on more devices, as I mentioned in my previous message, reviewing the GPU blocklist to re-enable WebGL for older devices if drivers have been updated or workarounds for issues can be found would help reduce the number of devices subject to SwiftShader. Being able to enable at least WebGL 1 will still help with Construct content.
• If a software fallback can be securely implemented for WebGPU, Construct has a WebGPU renderer too now so that would give us a workaround (and potentially for any other WebGL content - AFAIK many widely used libraries like three.js now either support WebGPU or are working on it)

Thanks for the consideration all.

Copy of my previous message:

-----

OK, thanks for the information. I just want to point out that even stopping WebGL content for only 2.7% of users is still potentially very disruptive. Consider a web game on Poki that requires WebGL and gets a million players. With this change, now 27,000 users will see a "WebGL not supported" error message. That's then potentially a huge number of new support requests to deal with.

> Can you share the number for Construct about what percentage of your users are using the SwiftShader fallback? Again, our numbers indicate that these are primarily older Windows workstations.

For the Construct editor itself, it is around 3%, so that seems in line. But the key point here is that Construct is middleware: it is a tool our users develop web games in and then publish independently of us. It is much more important that WebGL works for players of those games than it does for Construct itself.

Lots of people use older Windows workstations. We've had issues before where for example a graphics driver bug affecting WebGL 1 caused a great deal of trouble in a South American market, even though I suspect it only affected a small percentage of devices - see https://issues.chromium.org/issues/40941645 which was never resolved. There are probably places in the world where there are large numbers of people using older Windows workstations. I fear that pulling WebGL support from those devices may result in much higher numbers of unsupported users, and many more support requests, in the specific markets where such devices are common.

Is there anything that can be done to mitigate this change? Given SwiftShader allowed WebGL to be considered ubiquitous for many years, engines like Construct long ago removed any fallback for systems that do not support WebGL; we moved forward assuming we could rely on WebGL, and so now it's probably infeasible to bring back any fallback as we have too many key features that fundamentally require WebGL. Could SwiftShader be adapted to not use JIT? Could some other fallback be found? Could the GPU blocklist be revised to enable WebGL on as many older devices as possible?

I think the number of affected users should be <1% to minimise the impact from such a change. At web scale 2.7% is still a lot. Perhaps with revising the GPU blocklist and adding more workarounds this is feasible. I fear if this goes ahead without any mitigation, it will cause a great deal of trouble and is exactly the kind of thing sceptics of the web will bring up to say that web technology sucks, browsers can't be trusted, and people should just develop desktop games instead.

[Erik Anderson]

Hi David,

The initial message states that SwiftShader primarily covers older Windows devices. Beyond those, there are a non-trivial set of enterprise users that use thin clients to connect to a remote Windows device which is often running in a VM without access to a physical GPU.

For example, this applies to the Microsoft Dev Box offering (https://azure.microsoft.com/en-us/products/dev-box/).

 

Unfortunately, enterprise clients often turn off telemetry. So, I would assume any UMA-derived metrics to be undercounting the population.

It’s likely there are certain line-of-business and/or consumer-oriented sites that have a hard dependency on WebGL to be fully functional.

Have you considered, on Windows, targeting WARP (https://learn.microsoft.com/en-us/windows/win32/direct3darticles/directx-warp) instead? I don’t know if there are other viable alternatives on other OSes, but if the primary impacted clients are Windows perhaps that would be a sufficient mitigation.

To help enterprise customers reason about how much this is going to impact them, it would be helpful to have an enterprise policy to control this. This is a common pattern for potentially high-impact changes.

In its initial phase, the policy would enable motivated enterprises to forcibly disable SwiftShader as a scream test. And after you switch over the default, it could enable enterprises caught unaware to have some additional window of time to plan mitigations (by re-enabling it via policy) before you proceed with fully deprecating support and remove the policy.

Can you comment on if you plan to add such a policy or, if not, why not?

Thanks!

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAABs73jWBkuxvj%3DDDXmEQNwLfCa_uV5OZZ5nZJRj9ZMgP9yk7Q%40mail.gmail.com.

[Geoff Lang]

Hey Erik, Ashley, Rick,

I want to be clear that I think having high WebGL availability is a good thing. I don't think that users with software WebGL have a great experience but it's likely better than no availability, at least for drawing static things. What pushes this over the line and warrants this discussion is that JITing code in the GPU process is a huge vulnerability and is a rapidly increasing attack target. 

We're investigating WARP as an alternative on Windows. You are right that a large portion of the SwiftShader fallback is on machines with no GPUs (headless or VMs). There are just many unknowns about the quality and security of WARP, it will take a while to be confident in such a change and it still does not resolve the issue of JITing code in the weakly sandboxed GPU process.

Regarding corporate policy, I'd much rather have these users fall back in the same way as everyone else and work towards lowering the number of users in this position.  It would mean supporting and testing a feature only used by enterprise users when we have no visibility into crashes, bugs or vulnerabilities that they face.

We're also disabling software fallback due to a crashes in the GPU driver (as opposed to blocklisted GPU). Right now any user can fairly easily trigger a GPU crash and fall back to software WebGL which opens up vulnerabilities to all users instead of the 2.7%.

Geoff

[Ashley Gullen]

Is it feasible to have SwiftShader (or WARP) run in its own process with a stronger sandbox?

[Ken Russell]

It's feasible, but a significant amount of engineering work that our (Chrome Graphics) team would not be able to prioritize versus other current work that would impact a larger user base.

-Ken

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAABs73hYs9O-2hdmfn37fQb-U-8m_-08i3Qg9dkUhKNQQvNLSg%40mail.gmail.com.

[Erik Anderson]

Hi Geoff,

 

My suggestion re: a policy was not to have one that is supported indefinitely.

 

Many high-risk deprecations have had a policy lasting for, I believe, as little as 3 major version releases. Having such a thing helps mitigate the concern that the risk analysis was way off (which could then mean needing to do a stable respin if your risk analysis was off). If a policy is available, impacted enterprises can quickly self-remediate, report what broke once you flip over the default, and have a little bit more of a window to plan mitigations tied to the removal of the policy (since they’d now be aware of what broke and why).

 

Thanks,

Erik

[Rick Byers]

+1 that providing temporary enterprise policy exceptions is standard practice for breaking changes that we predict may have enterprise impact. 

Rick

[David Adrian]

To cover the testing use case, we have provided a CLI flag to enable SwiftShader. This has been in the release notes since November. If this is insufficient, we could add an enterprise policy. 

However, rather than attempt a straight removal, we are going to take a multi-pronged approach to attempt to simultaneously reduce the situations where SwiftShader is available, while maintaining compatibility with devices that require it due to the GPU blocklist.

• SwiftShader is already unused on many Mac clients, since it does not support ARM. We will run an experiment where we fully remove it on Mac, where usage is much smaller. We expect this will be ~fine.
  
• Similarly, we will try the same on Linux, although this may not go as well, as there are not a large number of ARM Linux clients.
  
• We will experiment with removing the automatic fallback to SwiftShader after 3 OOMs, limiting it to just the devices without a GPU or on the GPU blocklist. This should reduce the attack surface across the board, as attackers would be unable to arbitrarily cause SwiftShader to be used. In conjunction with this, we'll see if we can leverage Warp on Windows.

If we can get to a state where SwiftShader is off by default on Mac and Linux, and replaced with Warp on Windows aside from the CLI flag, and the fallback is not triggerable by an attacker on systems with a "normal" GPU, we'll be in much better shape from a security standpoint.

We will update this thread with the progress and results of these experiments as they roll out.

[Ashley Gullen]

Thanks for the update David - that sounds like a much less disruptive approach than removing it completely.