August 2021 - webkit-dev - lists.webkit.org

Smart Pointer Analysis Tool for WebKit
by Ryosuke Niwa 01 Mar '23

01 Mar '23

Hi all, I’ve been working with Geoff (ggaren) and Jan Korous from Apple's compiler team to create a static analyzer which detects dangerous use of ref counted objects, and we’re looking for folks who are interested in trying out this tool and helping us refine the tool. Please let me know if you’re interested in using this tool & try applying code changes to WebKit. Our goal is to eventually integrate this tool into buildbot and EWS so that we can catch dangerous code before they get committed. *What is Dangerous?* So *what is* a *dangerous* use of ref counted objects you may ask? It’s any use of which we can’t trivially conclude that it doesn’t lead to a use-after-free. For now, we don’t detect dangerous use of non-ref counted objects including ones that can vend WeakPtr. It’s on us, humans, to decide which objects need to be ref counted or need to be CanMakeWeakPtr. Consider the following example. This code most certainly will lead to a use-after-free of “parent” in the third line because the code doesn’t keep the parent alive. Because isContentEditable updates style in some cases, it could lead to arbitrary script execution which could remove the parent from the document. Node* parent = element->parentElement(); if (parent->isContentEditable()) parent->scrollIntoView(); In general, relying on a complex data structure such as DOM tree to keep RefCounted objects alive while we call a non-trivial function is not safe. All it takes for the code to have a use-after-free is for someone to start updating style, layout, etc… inside the function either directly or indirectly. And we don’t want to make WebKit development really hard by forcing anyone who modifies a function to check every caller of the function and their callers, etc… to make sure it’s safe to do so. For this reason, it’s *dangerous* to store a raw pointer or a reference to a ref counted object as a local variable and use it after calling a non-trivial function. We did a similar analysis of a number of other patterns and usage of ref counted objects in WebKit and came up with the following basic rules for using ref counted objects in a safe manner. We’re hoping that these rules would be eventually incorporated into our coding style guideline: https://webkit.org/code-style-guidelines/ *Rules for Using Ref Counted Objects* 1. Every data member to a ref counted object must use either Ref, RefPtr, or WeakPtr. webkit.NoUncountedMemberChecker <https://github.com/llvm/llvm-project/blob/master/clang/docs/analyzer/checke…> 2. Every ref counted base class, if it has a derived class, must define a virtual destructor. webkit.RefCntblBaseVirtualDtor <https://github.com/llvm/llvm-project/blob/master/clang/docs/analyzer/checke…> 3. Every ref counted object passed to a non-trivial function as an argument (including "this" pointer) should be stored as a Ref or RefPtr in the caller’s local scope unless it's an argument to the caller itself by transitive property [1]. alpha.webkit.UncountedCallArgsChecker <https://github.com/llvm/llvm-project/blob/master/clang/docs/analyzer/checke…> 4. Every ref counted object must be captured using Ref or RefPtr for lambda. webkit.UncountedLambdaCapturesChecker <https://github.com/llvm/llvm-project/blob/master/clang/docs/analyzer/checke…> 5. Local variables - we’re still working on this ( https://reviews.llvm.org/D83259) Below, I've dissected each one of these rules with the real warning emitted by the analysis tool in development. Please let me know if you have any comments / concerns. ------------------------------ (1) is pretty trivial. Every ref counted data member should be stored using Ref, RefPtr, or WeakPtr since it would be not trivially obvious that life cycles of two or more objects are correctly tied or managed together. ------------------------------ (2) is also pretty easy to understand. In the following example, if someone destroys an instance of B using Ref<A>, then it would result in an undefined behavior. struct A : public RefCounted<A> { Vector<int> someData; }; struct B : public A { Vector<int> otherData; }; ------------------------------ For (3), passing a ref counted object as a raw pointer or reference to a function as an argument, the tool may generate a warning like this: Source/WebCore/html/FormAssociatedElement.cpp:181:13: warning: [WebKit] call argument is a raw pointers to a ref-countable type [-Wfunction-arg-ptr-to-refcntbl] setForm(findAssociatedForm(&asHTMLElement(), originalForm.get())); ^ This warns that void setForm(HTMLFormElement*) is called with the result of findAssociatedForm, which returns HTMLFormElement* without storing it anywhere. If setForm can somehow cause HTMLFormElement to be deleted, then this can result in the use-after-free in setForm. void FormAssociatedElement::resetFormOwner() { RefPtr<HTMLFormElement> originalForm = m_form.get(); setForm(findAssociatedForm(&asHTMLElement(), originalForm.get())); HTMLElement& element = asHTMLElement(); auto* newForm = m_form.get(); if (newForm && newForm != originalForm && newForm->isConnected()) element.document().didAssociateFormControl(element); } Why, you may ask, we don't put HTMLFormElement* in a Ref or RefPtr in setForm instead? This is because a lot of code has calls to a function with a local variable. Because the callee doesn’t know whether the caller has already stored each argument in Ref / RefPtr or not, we need to be safe and store them again in Ref / RefPtr, resulting in an unnecessary ref-churn. Additionally, if we took the approach of the callee being responsible for keeping every argument alive, then every non-trivial member function of a ref counted object must have "protectedThis" at the beginning of the function, which would be particularly wasteful if those member functions are sometimes called by other member functions. Once this rule is applied everywhere, for example, we can get rid of “protectedThis” from our codebase because it would be redundant. Furthermore, this rule transitively allows a raw pointer or a reference passed in as an argument to be used as arguments to call another function without storing it in Ref or RefPtr. For example, the following function satisfies this rule because both oldDocument and newDocument are raw references passed to this function as arguments, and therefore the caller of this function must have already stored as Ref / RefPtr as local variables. void HTMLMediaElement::didMoveToNewDocument(Document& oldDocument, Document& newDocument) { ASSERT_WITH_SECURITY_IMPLICATION(&document() == &newDocument); if (m_shouldDelayLoadEvent) { oldDocument.decrementLoadEventDelayCount(); newDocument.incrementLoadEventDelayCount(); } unregisterWithDocument(oldDocument); registerWithDocument(newDocument); HTMLElement::didMoveToNewDocument(oldDocument, newDocument); updateShouldAutoplay(); } ------------------------------ For (4), capturing a ref counted object as a raw pointer or a reference, the tool may generate a warning like this: Source/WebKit/UIProcess/API/APIHTTPCookieStore.cpp:115:74: warning: [WebKit] captured a raw pointer to a ref-countable type [-Wlambda-capture-ptr-to-refcntbl] cookieManager->deleteCookie(m_owningDataStore->sessionID(), cookie, [pool = WTFMove(pool), completionHandler = WTFMove(completionHandler)]() mutable { ^ This is warning about webPage object getting captured using a raw reference in the following function: void HTTPCookieStore::deleteCookie(const WebCore::Cookie& cookie, CompletionHandler<void()>&& completionHandler) { auto* pool = m_owningDataStore->processPoolForCookieStorageOperations(); if (!pool) { if (m_owningDataStore->sessionID() == PAL::SessionID::defaultSessionID() && !cookie.session) deleteCookieFromDefaultUIProcessCookieStore(cookie); else m_owningDataStore->removePendingCookie(cookie); RunLoop::main().dispatch([completionHandler = WTFMove(completionHandler)] () mutable { completionHandler(); }); return; } auto* cookieManager = pool->supplement<WebKit::WebCookieManagerProxy>(); cookieManager->deleteCookie(m_owningDataStore->sessionID(), cookie, [pool = WTFMove(pool), completionHandler = WTFMove(completionHandler)]() mutable { completionHandler(); }); } Here, we should have stored WebProcessPool in Ref instead. - R. Niwa

9 17

Request for Position: MSE-in-Workers
by Matt Wolenetz 27 Jul '22

27 Jul '22

Hi webkit-dev, This is a request for Webkit's position on the MSE-in-Workers feature: Explainer: https://github.com/wicg/media-source/blob/mse-in-workers-using-handle/mse-i… Main MSE spec issue tracking this feature: https://github.com/w3c/media-source/issues/175 ChromeStatus entry: https://chromestatus.com/features/5177263249162240 Draft specification: https://github.com/w3c/media-source/pull/282 TAG review: https://github.com/w3ctag/design-reviews/issues/656 Demo page (Chromium has experimental implementation): https://wolenetz.github.io/mse-in-workers-demo/mse-in-workers-demo.html Please let us know if you have any feedback! Thanks, Matt

1 1

Request for Position: User Preference Media Features Client Hints Headers
by Thomas Steiner 29 Jun '22

29 Jun '22

Hi WebKittens, We are interested in hearing your thoughts on User Preference Media Features Client Hints Headers, a set of HTTP Client Hints headers around user preference media features as defined by Media Queries Level 5. Specification Title: User Preference Media Features Client Hints Headers Specification or proposal URL: https://tomayac.github.io/user-preference-media-features-headers/ Explainer URL: https://github.com/tomayac/user-preference-media-features-headers/blob/main… TAG review: https://github.com/w3ctag/design-reviews/issues/632 Mozilla Standards Position: https://github.com/mozilla/standards-positions/issues/526 Thank you very much for your time and consideration! Cheers, Tom

1 5

Request for feedback on CompressionStream and DecompressionStream
by Adam Rice 02 May '22

02 May '22

I am trying to gauge feedback on compression streams with a view to shipping them in Chromium. Very briefly, they are a way to do gzip and gunzip in the browser. Less briefly, the explainer https://github.com/WICG/compression/blob/master/explainer.md goes into some detail of the how and why. The specification https://wicg.github.io/compression/ gives verbose detail of how. You may also find the W3C TAG review https://github.com/w3ctag/design-reviews/issues/410 interesting. Any and all feedback is welcome! FAQ: Q. Will this support brotli, zstd, etc.? A. Not yet. It's on the radar. Thanks, Adam Rice, Chromium

3 3

Request for Position: Multi-Screen Window Placement API
by Ben Morss 12 Apr '22

12 Apr '22

Hi all, We'd love to get WebKit's position on the new Multi-Screen Window Placement API. The proposal is to make incremental improvements to existing screen information and window placement APIs, allowing web applications to be intentional about the experiences they offer to users of multi-screen devices. We'd like to make it simple to: - detect when screen properties changes - discover the presence of extended screen areas - request extended screen information - place a window on a specific screen - show a window fullscreen on a specific screen This is useful for applications like the following: - a slideshow app shows slides full-screen on a projector, while displaying speaker notes on a laptop - a creativity app pops up secondary windows, like palettes, on a separate screen - a videoconferencing app shows video on a large screen, and controls on a smaller screen - a medical app shows images on multiple displays - a financial app shows dashboards on multiple monitors Future work could involve support for foldable devices, showing a window that spans multiple screens, etc. For more details, please see these links: - Explainer: https://github.com/webscreens/window-placement/blob/main/EXPLAINER.md - Draft Community Group Report: https://webscreens.github.io/window-placement/ - Additional ideas: https://github.com/webscreens/window-placement/blob/main/additional_explora… Thanks for your feedback on this! All best, Ben

1 1

Re: [webkit-dev] Request For Position on CSS containment
by cathiechen 24 Dec '21

24 Dec '21

Hi webkit-dev, We made a lot of progress regarding CSS containment [1]. Rob and I have finished the layout containment and size containment [2].\o/ And the patches of paint containment and style containment are ready for review now [3]. So we think now it's time to move on to content-visibility: (https://www.w3.org/TR/css-contain-2/#content-visibility) Since content-visibility depends on paint and style containment, we will do some specification research first, then prototype it based on Rob's patches. Thank you for your feedback! [1] https://bugs.webkit.org/show_bug.cgi?id=172026 [2] https://bugs.webkit.org/show_bug.cgi?id=223569, https://bugs.webkit.org/show_bug.cgi?id=223570 [3] https://bugs.webkit.org/show_bug.cgi?id=224742, https://bugs.webkit.org/show_bug.cgi?id=226458 Best, Rob and Cathie

3 5

Request for position: CSS highlight inheritance (css-pseudo #highlight-cascade)
by Delan Azabani 28 Oct '21

28 Oct '21

I’m seeking WebKit’s position on the “Cascading and Per-Element Highlight Styles” section of css-pseudo-4 [0]. Most impls (with the sole exception of Presto) have historically not propagated highlight styles through the element tree. This has meant that only the global ::selection selector (*::selection) worked intuitively, and made it awkward to style the highlights on some subset of elements. Authors had to repeat each selector with a “*” descendant, and ensure that “overriding” rules win the cascade. For example, to highlight content under p.foo in yellow, but content under <em> anywhere in the document in cyan, one might: p.foo::selection, p.foo *::selection { background-color: yellow; } em:not(#specificity)::selection, em:not(#specificity) *::selection { background-color: cyan; } The spec gives highlights an inheritance-based processing model, where all properties (inherited or otherwise) default to the value on the parent’s highlight. Note that the spec’s processing model is no longer cascade-based, as of w3c/csswg-drafts#2474 [1]. We believe the compat risk is minimal, because the new model won’t break existing content with global ::selection rules, nor will it break rules that use the cascade to build an ersatz “propagation” like the example above. As far as we can tell, only content that relies on descendants of a highlight-styled element being reset to initial styles should be affected. [0] https://drafts.csswg.org/css-pseudo-4/#highlight-cascade [1] https://github.com/w3c/csswg-drafts/issues/2474 Cheers, Delan Azabani Igalia // web platform

1 1

New fast commit queue mode: fast-cq
by Aakash Jain 25 Oct '21

25 Oct '21

Hi Everyone, I am happy to inform you that I have added a fast-commit-queue mode to the Commit-Queue. In this mode Commit-Queue would land patches quickly (typically taking ~1 minute, e.g.: https://ews-build.webkit.org/#/builders/28/builds/10876) This mode skips building and testing. It basically does sanity check (like validating ChangeLog, Reviewer, Committer etc.) and commits the patch. This is intended to be used for scenarios where the patch is a build fix, urgent fix, quick typo/follow-up fix etc. To use this mode, you can use the --fast-cq parameter while uploading the patch using webkit-patch command. e.g.: 'webkit-patch upload --fast-cq'. For patches which are already uploaded to Bugzilla, you can rename the patch (by clicking on 'Details' button for the patch), and prefix patch name with [fast-cq], e.g.: '[fast-cq] Patch for landing'. Pro-tip: you can also rename the patch to include [fast-cq] prefix while marking the patch cq+, from patch 'Details' page, and this mode would work. Please let me know if you have any questions/feedback. Thanks Aakash

2 4

Request for position on Reporting API (now with structured headers!)
by Ian Clelland 02 Sep '21

02 Sep '21

Hi WebKit! In Chrome-land, we've been working on updating our implementation of the Reporting API to align with the spec changes made over the last year or so, which resolve a number of privacy issues, separating reports generated within the context of a single document from those which relate to overall network conditions, and isolating reports from unrelated documents. This also includes a new header, Reporting-Endpoints, whose value is not persisted between document loads. We're hoping to ship these updates sometime early in the new year, and would love to get the thoughts / opinions / feedback from WebKit developers. Thanks! Ian Some relevant links: * Explainer: https://github.com/w3c/reporting * Spec: https://w3c.github.io/reporting * Recently-opened TAG review: https://github.com/w3ctag/design-reviews/issues/585

1 1

Request for Position on URLPattern
by Ben Kelly 31 Aug '21

31 Aug '21

Hello, I would like to request an official WebKit position on URLPattern [0]. This is a web API for matching against URLs which is often done in javascript routing systems. We also intend to use this same matching primitive in future changes to service worker scopes. This larger set of use cases is described in the explainer. [1] This particular request, however, is just for the URLPattern API and does not yet touch service workers. URLPattern was discussed in a couple of past TPAC meetings: * TPAC 2019 [2] * TPAC 2020 Video Call [3] In case it's helpful, we also have a summary of web developer feedback we've received so far. [4] Thank you. Ben [0]: https://wicg.github.io/urlpattern [1]: https://github.com/WICG/urlpattern/blob/main/explainer.md [2]: https://docs.google.com/document/d/1q090ovJ4gd8wSfVtvuoZLMZ51YkiFDsEZ0Jiqi4… [3]: https://github.com/w3c/ServiceWorker/issues/1535 [4]: https://github.com/WICG/urlpattern/blob/main/dev-trials-feedack.md

1 1