218100 2020-10-22 14:47:34 -0700 Rapidly resizing a WebGL-rendered canvas leaks memory 2021-08-30 06:32:52 -0700 1 1 1 Unclassified WebKit WebGL WebKit Nightly Build Unspecified Unspecified RESOLVED DUPLICATE 217212 https://bugs.webkit.org/show_bug.cgi?id=219780 https://bugs.webkit.org/show_bug.cgi?id=229606 P2 Major --- 217212 1 kbr kbr dino jdarpinian kkinnunen oldest_to_newest 1700847 0 kbr 2020-10-22 14:47:34 -0700 Sebastien Vandenberghe from Microsoft's Babylon.js team reports that rapidly resizing a WebGL-rendered canvas leaks significant amounts of memory: https://jsfiddle.net/kr0atxuf/2/ Safari reloads the tab after about 30 seconds because it's consuming too much memory. This happens repeatedly until the browser says it failed too many times. This is reproducible in Safari Technology Preview 114 on macOS 10.15.7, so isn't fixed by the ANGLE backend. Looking into Source/WebCore/platform/graphics/cocoa/WebGLLayer.mm, while the old IOSurfaces should be being cleaned up because they're held onto via std::unique_ptr, it looks like perhaps other associated graphics resources aren't being cleaned up properly. Filing as Major severity because this will significantly impact WebGL applications' reliability. 1700879 1 kbr 2020-10-22 16:28:35 -0700 After preparing a patch fixing this against somewhat old WebKit sources and rebasing to top-of-tree, I found Kimmo's fix for Bug 209139 might have already fixed this. Rebuilding to see. 1700880 2 kbr 2020-10-22 16:29:02 -0700 Sorry, that should have been Bug 217212. 1700895 3 kbr 2020-10-22 17:15:16 -0700 Yes, Kimmo's fix for Bug 217212 fixed this memory leak. I'll ask about getting that merged into the Safari 14 release branch. *** This bug has been marked as a duplicate of bug 217212 *** 1700929 4 kkinnunen 2020-10-22 23:13:01 -0700 I'll dupe it to bug 217581 which is open. Although I tried to be a bit more meticulous with the underlying gpu object references when refactoring, I don't think the original code had that significant leaks (that the new code does not, at least). *** This bug has been marked as a duplicate of bug 217581 *** 1701183 5 kbr 2020-10-23 18:30:02 -0700 The original code leaked EGL pbuffers that were associated with the IOSurfaces, and it looked to me like those were somehow retaining the IOSurfaces. Fixing the leak in the old [WebGLLayer allocateIOSurfaceBackingStoreWithSize:usingAlpha:] method definitely addressed the previous leak, which also doesn't exist after the fix for Bug 217212. 1701202 6 kkinnunen 2020-10-23 21:04:05 -0700 > The original code leaked EGL pbuffers that were associated with the IOSurfaces, and it looked to me like those were somehow retaining the IOSurfaces EGL pbuffers do retain the iosurfaces and unretain only in destructors. I'll take your word on the original code leaking pbuffers and dupe this back. *** This bug has been marked as a duplicate of bug 217212 ***