Changeset 203172 in webkit

Timestamp:

Jul 13, 2016 12:17:46 PM (8 years ago)

Author:

Antti Koivisto

Message:

v2: WebContent crash due to RELEASE_ASSERT(!m_inLoadPendingImages) in StyleResolver::~StyleResolver()
​https://bugs.webkit.org/show_bug.cgi?id=159722

Reviewed by Andreas Kling.

We have crashes where a StyleResolver is deleted underneath pseudoStyleForElement (key parts of the stack):

0 WebCore::StyleResolver::~StyleResolver
3 WebCore::AuthorStyleSheets::updateActiveStyleSheets
4 WebCore::Document::styleResolverChanged
5 WebKit::WebPage::viewportConfigurationChanged()
6 WebKit::WebPage::mainFrameDidLayout()
9 WebCore::FrameLoader::checkCompleted
13 WebCore::ResourceLoader::cancel
19 WebKit::WebLoaderStrategy::loadResource
24 WebCore::Style::loadPendingImage
27 WebCore::StyleResolver::pseudoStyleForElement
29 WebCore::RenderTreeUpdater::updateBeforeOrAfterPseudoElement
33 WebCore::Document::recalcStyle

This appears to be happening when a content blocker blocks a resource load for an image referenced from a stylesheet
and triggers synchronous cancellation of the load. With engine in suitable state this can clear style resolver.

No test, don't know how to make one. This is very timing and engine state dependent.

• dom/AuthorStyleSheets.cpp:

(WebCore::AuthorStyleSheets::updateActiveStyleSheets):

We have an existing check here that prevents destruction of the style resolver when we are in the middle of
a style resolution. However the old inStyleRecalc() bit no longer covers the render tree update phase. Pseudo
elements are resolved during render tree update.

Fix by adding a check for inRenderTreeUpdate() bit too.

This just fixes a regression. A proper fix would be to gather all resources during style resolution
and trigger the loads afterwards.

Location:

trunk/Source/WebCore

Files:

• ChangeLog (modified) (1 diff)
• dom/AuthorStyleSheets.cpp (modified) (1 diff)

trunk/Source/WebCore/ChangeLog

@@ +1 @@
+2016-07-13  Antti Koivisto  <antti@apple.com>
+
+        v2: WebContent crash due to RELEASE_ASSERT(!m_inLoadPendingImages) in StyleResolver::~StyleResolver()
+        https://bugs.webkit.org/show_bug.cgi?id=159722
+
+        Reviewed by Andreas Kling.
+
+        We have crashes where a StyleResolver is deleted underneath pseudoStyleForElement (key parts of the stack):
+
+        0   WebCore::StyleResolver::~StyleResolver
+        3   WebCore::AuthorStyleSheets::updateActiveStyleSheets
+        4   WebCore::Document::styleResolverChanged
+        5   WebKit::WebPage::viewportConfigurationChanged()
+        6   WebKit::WebPage::mainFrameDidLayout()
+        9   WebCore::FrameLoader::checkCompleted
+        13  WebCore::ResourceLoader::cancel
+        19  WebKit::WebLoaderStrategy::loadResource
+        24  WebCore::Style::loadPendingImage
+        27  WebCore::StyleResolver::pseudoStyleForElement
+        29  WebCore::RenderTreeUpdater::updateBeforeOrAfterPseudoElement
+        33  WebCore::Document::recalcStyle
+
+        This appears to be happening when a content blocker blocks a resource load for an image referenced from a stylesheet
+        and triggers synchronous cancellation of the load. With engine in suitable state this can clear style resolver.
+
+        No test, don't know how to make one. This is very timing and engine state dependent.
+
+        * dom/AuthorStyleSheets.cpp:
+        (WebCore::AuthorStyleSheets::updateActiveStyleSheets):
+
+        We have an existing check here that prevents destruction of the style resolver when we are in the middle of
+        a style resolution. However the old inStyleRecalc() bit no longer covers the render tree update phase. Pseudo
+        elements are resolved during render tree update.
+
+        Fix by adding a check for inRenderTreeUpdate() bit too.
+
+        This just fixes a regression. A proper fix would be to gather all resources during style resolution
+        and trigger the loads afterwards.
+
 2016-07-13  Frederic Wang  <fred.wang@free.fr>
 

trunk/Source/WebCore/dom/AuthorStyleSheets.cpp

@@ -292 +292 @@
 bool AuthorStyleSheets::updateActiveStyleSheets(UpdateFlag updateFlag)
 {
-    if (m_document.inStyleRecalc()) {
-        // SVG <use> element may manage to invalidate style selector in the middle of a style recalc.
-        // https://bugs.webkit.org/show_bug.cgi?id=54344
-        // FIXME: This should be fixed in SVG and the call site replaced by ASSERT(!m_inStyleRecalc).
+    if (m_document.inStyleRecalc() || m_document.inRenderTreeUpdate()) {
+        // Protect against deleting style resolver in the middle of a style resolution.
+        // Crash stacks indicate we can get here when z resource load fails synchronously (for example due to content blocking).
+        // FIXME: These kind of cases should be eliminated and this path replaced by an assert.
         m_pendingUpdateType = FullUpdate;
         m_document.scheduleForcedStyleRecalc();