Is it feasible to have SwiftShader (or WARP) run in its own process with a stronger sandbox?
On Fri, 28 Feb 2025 at 15:25, Geoff Lang <geoffl...@google.com> wrote: > 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 > > On Thu, Feb 27, 2025 at 3:28 PM Erik Anderson <erik.ander...@microsoft.com> > wrote: > >> 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! >> >> >> >> *From:* 'Ashley Gullen' via blink-dev <blink-dev@chromium.org> >> *Sent:* Thursday, February 27, 2025 4:14 AM >> *To:* Rick Byers <rby...@chromium.org> >> *Cc:* David Adrian <dadr...@google.com>; blink-dev < >> blink-dev@chromium.org>; geof...@chromium.org <geoffl...@chromium.org> >> *Subject:* [EXTERNAL] Re: [blink-dev] Intent to Remove: SwiftShader >> Fallback >> >> >> >> 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. >> >> >> >> >> >> On Tue, 25 Feb 2025 at 22:31, Rick Byers <rby...@chromium.org> wrote: >> >> 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 <https://chromestatus.com/metrics/feature/timeline/popularity/4026> >> 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 >> <https://docs.google.com/document/d/1RC-pBBvsazYfCNNUSkPqAVpSpNJ96U8trhNkfV0v9fk/edit?tab=t.0#heading=h.mqfkui78vo5z> >> . >> >> >> >> 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 >> <https://www.chromium.org/flash-roadmap/>, or WebSQL >> <https://developer.chrome.com/blog/deprecating-web-sql> (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 >> >> <https://docs.google.com/document/d/1RC-pBBvsazYfCNNUSkPqAVpSpNJ96U8trhNkfV0v9fk/edit?tab=t.0#heading=h.x5bhg5grhfeo> >> 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 >> >> <https://chrome-commit-tracker.arthursonzogni.com/cve/reward_per_components?start=2019-12-27&end=2025-02-25> >> 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 >> >> >> >> >> >> >> >> On Wed, Feb 19, 2025 at 2:48 PM 'David Adrian' via blink-dev < >> blink-dev@chromium.org> wrote: >> >> > 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. >> >> >> >> On Thursday, February 13, 2025 at 12:12:07 PM UTC-5 ash...@scirra.com >> wrote: >> >> 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. >> >> >> >> >> >> On Thu, 13 Feb 2025 at 15:46, 'David Adrian' via blink-dev < >> blin...@chromium.org> wrote: >> >> 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 >> <https://issues.chromium.org/issues?q=customfield1222907:%22Blink%3EWebGL%22> >> >> >> 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 >> <https://chromestatus.com/>. >> >> >> >> -- >> 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 >> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAGkh42KV4DrSSyEgJaF4DnFOXAye-wRLrfD-LKGNkWhyWzshLA%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> >> -- >> 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+unsubscr...@chromium.org. >> To view this discussion visit >> https://groups.google.com/a/chromium.org/d/msgid/blink-dev/c5131675-dff4-4aa0-8e84-4cdc373e3035n%40chromium.org >> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/c5131675-dff4-4aa0-8e84-4cdc373e3035n%40chromium.org?utm_medium=email&utm_source=footer> >> . >> >> -- >> 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+unsubscr...@chromium.org. >> To view this discussion visit >> https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAABs73jWBkuxvj%3DDDXmEQNwLfCa_uV5OZZ5nZJRj9ZMgP9yk7Q%40mail.gmail.com >> <https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAABs73jWBkuxvj%3DDDXmEQNwLfCa_uV5OZZ5nZJRj9ZMgP9yk7Q%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> > -- 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+unsubscr...@chromium.org. To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/blink-dev/CAABs73hYs9O-2hdmfn37fQb-U-8m_-08i3Qg9dkUhKNQQvNLSg%40mail.gmail.com.
