January 2022 - 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

Using C++20 in WebKit
by Alex Christensen 10 Mar '22

10 Mar '22

In April 2019 in https://bugs.webkit.org/show_bug.cgi?id=197131 <https://bugs.webkit.org/show_bug.cgi?id=197131> I increased WebKit’s minimum C++ language requirement from C++14 to C++17. In 2022 I’m planning to increase WebKit’s minimum C++ requirement from C++17 to C++20. Would April 2022 be a good time to do that?

4 14

Request for position: Cookie Expires/Max-Age attribute upper limit
by Ari Chivukula 23 Feb '22

23 Feb '22

I'd like to get WebKit's position on: (1) Having an explicit upper limit for Cookie Expires/Max-Age attributes (2) Having an explicit upper limit for Cookie Expires/Max-Age attributes that's less than or equal to 400 days https://httpwg.org/http-extensions/draft-ietf-httpbis-rfc6265bis.html#name-… https://github.com/httpwg/http-extensions/pull/1732 https://github.com/mozilla/standards-positions/issues/592 https://bugs.chromium.org/p/chromium/issues/detail?id=1264458 The draft of rfc6265bis now contains an upper limit for Cookie Expires/Max-Age attributes. As written: `The user agent MUST limit the maximum value of the [Max-Age/Expiration] attribute. The limit MUST NOT be greater than 400 days (34560000 seconds) in duration. The RECOMMENDED limit is 400 days in duration, but the user agent MAY adjust the limit to be less. [Max-Age/Expiration] attributes that are greater than the limit MUST be reduced to the limit.` 400 days was chosen as a round number close to 13 months in duration. 13 months was chosen to ensure that sites one visits roughly once a year (e.g., picking health insurance benefits) will continue to work. Safari is already partially compliant (has an upper age limit of 7 days when cookies are set client side), while Firefox and Chrome both support cookies with expiration dates orders of magnitude longer than a millenia in the future. According to measurements in Chrome of all cookies set about 20% have an Expires/Max-Age further than 400 days in the future. Of that 20%: half target 2 years, a quarter target 10 years or more, and the remainder are spread over the rest of the range. ~ Ari Chivukula (Their/There/They're)

3 3

Request for position: blocking="render" attribute on <link>, <script> and <style>
by Xiaocheng Hu 14 Jan '22

14 Jan '22

Hi webkit-dev, This is a request for WebKit's position on the blocking="render" attribute on <link>, <script> and <style> elements. This is a new feature currently being proposed to the HTML spec: - Full proposal: https://gist.github.com/xiaochengh/9404abbecdc3b45c0e9f3d5e99fbc65d - Discussion at WHATWG: https://github.com/whatwg/html/issues/7131 All current browsers treat stylesheets as render-blocking resources, and do not start rendering the page before they are loaded. This feature extends this mechanism by allowing authors to explicitly mark an external resource also as render-blocking to ensure better page load performance in more cases. Regards, Xiaocheng

1 0

Request for Position: Priority Hints
by Patrick Meenan 10 Jan '22

10 Jan '22

Hi webkit-dev, This is a request for Webkit's position on Priority Hints. Draft Spec: https://wicg.github.io/priority-hints/ Explainer: https://github.com/WICG/priority-hints/blob/main/EXPLAINER.md Priority Hints was explored in 2018 as part of the w3c web performance working group but went stale. The work has re-started now that there are cleaner use cases and the scope has been reduced to only reflect the importance of the fetch for the initial resource referenced by an element (rather than cascading to all fetches within a frame or triggered by a given script). Chrome has been running a field trial that started in 96 and early feedback is very positive and the plan is to push forward for a launch (any feedback on keeping "importance" or finding something more fetch-specific on the linked whatwg issue above would be appreciated). The prioritization changes under the cover as a result of developer hinting can be applied both to internal browser fetch ordering as well as plumbed through to protocol-level prioritization for multiplexed protocols that support prioritization (HTTP/2 and HTTP/3 currently). Use Cases There are 3 concrete use-cases that have emerged that are current developer pain points that Priority Hints provides a relatively clean solution for with minimal effort on the developer's part. Signaling Important Images Identifying which images are important in markup gives a parse-time signal to the browser for ordering the image fetches. Chromium already has logic that boosts the priority of visible images in the viewport but that information is only available after styles have been applied and layout has been performed which is relatively late in the lifecycle of fetching resources for the page. We have seen early gains in the 20-40% range for Largest Contentful Paint for sites in the origin trial where they can benefit from signaling of important images (Content pages with a main hero image and product detail pages for e-commerce sites are easy wins). Differentiating API fetches Script using the fetch API to both react to user input (like with autocomplete suggestions based on typing) and performing background bulk fetches benefits from being able to differentiate the priorities of the fetches. In this case, it works best if the transport protocol and origin all support prioritization but it provides application-level control to make the API calls more responsive than is otherwise possible. Signaling the priority of async scripts Async (and defer) scripts don't block tree-building so Chromium usually assigns them a lower priority. For some sites, the core framework and application logic are all loaded through async scripts and the lower priority can harm performance. There is a chromium-specific hack that relies on preloading the same script to get it to load with a higher priority. It would be a lot cleaner to signal it directly. Thanks, -Pat

1 0