612123 - [D2D] Firefox 4.0b7 crash [@ cairo_d2d_surface_create_for_texture ] when trying to turn off HW acceleration

Status: RESOLVED FIXED

Description

[Scoobidiver (away)]

It is a new crash signature that exists in 4.0b7, b8pre.
It is #133 top crasher in 4.0b7 for the last week.

Here are some useful comments:
"tried to turn off hardware acceleration"
"Firefox crashes every single time I try and turn off hardware acceleration in the settings. I am doing this because I am running a GPU client of Folding@home constantly in the background and Firefox doesn't get on well with this when hardware acceleration is on."
"It crashed when I tried to disable the option "Use Hardware Acceleration When Available". Please improve the rendering of the fonts."

Signature	cairo_d2d_surface_create_for_texture
UUID	c30d9924-9aa2-4abc-b9a3-a51702101112
Time 	2010-11-12 10:49:14.581543
Uptime	51041
Last Crash	51046 seconds (14.2 hours) before submission
Install Age	51041 seconds (14.2 hours) since version was first installed.
Product	Firefox
Version	4.0b7
Build ID	20101104142426
Branch	2.0
OS	Windows NT
OS Version	6.1.7601 Service Pack 1, v.721
CPU	x86
CPU Info	GenuineIntel family 6 model 30 stepping 5
Crash Reason	EXCEPTION_ACCESS_VIOLATION_WRITE
Crash Address	0x4
App Notes 	AdapterVendorID: 1002, AdapterDeviceID: 68a0

Frame 	Module 	Signature [Expand] 	Source
0 	xul.dll 	cairo_d2d_surface_create_for_texture 	gfx/cairo/cairo/src/cairo-d2d-surface.cpp:3880
1 	xul.dll 	gfxD2DSurface::gfxD2DSurface 	gfx/thebes/gfxD2DSurface.cpp:62
2 	xul.dll 	mozilla::layers::LayerManagerD3D10::CreateOptimalSurface 	gfx/layers/d3d10/LayerManagerD3D10.cpp:316
3 	xul.dll 	nsCanvasRenderingContext2D::SetDimensions 	content/canvas/src/nsCanvasRenderingContext2D.cpp:1086
4 	xul.dll 	nsHTMLCanvasElement::UpdateContext 	content/html/content/src/nsHTMLCanvasElement.cpp:514
5 	xul.dll 	nsHTMLCanvasElement::GetContext 	content/html/content/src/nsHTMLCanvasElement.cpp:447
6 	xul.dll 	nsIDOMHTMLCanvasElement_GetContext 	obj-firefox/js/src/xpconnect/src/dom_quickstubs.cpp:20968
7 	mozjs.dll 	js::Interpret 	js/src/jsinterp.cpp:4744
8 	mozjs.dll 	js::RunScript 	js/src/jsinterp.cpp:665
9 	mozjs.dll 	js::Invoke 	js/src/jsinterp.cpp:768
10 	mozjs.dll 	js::ExternalInvoke 	js/src/jsinterp.cpp:881
11 	mozjs.dll 	JS_CallFunctionValue 	js/src/jsapi.cpp:4898
12 	xul.dll 	nsXPCWrappedJSClass::CallMethod 	js/src/xpconnect/src/xpcwrappedjsclass.cpp:1694
13 	xul.dll 	nsXPCWrappedJS::CallMethod 	js/src/xpconnect/src/xpcwrappedjs.cpp:577
14 	xul.dll 	PrepareAndDispatch 	xpcom/reflect/xptcall/src/md/win32/xptcstubs.cpp:114
15 	xul.dll 	SharedStub 	xpcom/reflect/xptcall/src/md/win32/xptcstubs.cpp:141
16 	xul.dll 	mozilla::widget::TaskbarPreview::DrawBitmap 	widget/src/windows/TaskbarPreview.cpp:406

More reports at:
http://crash-stats.mozilla.com/report/list?range_value=4&range_unit=weeks&signature=cairo_d2d_surface_create_for_texture

Comment 1

[Bas Schouten (:bas.schouten)]

Hrm, I think I understand what happens here. When we update the RenderMode the LayerManagers should be cancelled for all Widgets. Not just the one that triggered the render mode update.

Reliable STR are as follows:

- Start firefox with HW accel on
- Ctrl+N to open new window
- Go into options in new window
- Disable hardware acceleration
- Mouse-over taskbar icon to draw preview for the other window

Comment 2

[Bas Schouten (:bas.schouten)]

Created attachment 491313 [details] [diff] [review]
Clear all D3D10 layer managers

This patch clears all D3D10 layer managers when the D2D device changes/becomes unavailable. It's not particularly pretty though.

Comment 3

[Bas Schouten (:bas.schouten)]

Created attachment 491328 [details] [diff] [review]
Check D3D10 device upon GetLayerManager

As per IRC discussion this will check the D3D10 device when the layer manager is retrieved.

Comment 4

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Comment on attachment 491328 [details] [diff] [review]
Check D3D10 device upon GetLayerManager

Looks good, but I think we need an additional patch to prevent the crash if a D3D10 device goes dead after the GetLayerManager() has succeeded. Right?

Comment 5

[Bas Schouten (:bas.schouten)]

(In reply to comment #4)
> Comment on attachment 491328 [details] [diff] [review]
> Check D3D10 device upon GetLayerManager
> 
> Looks good, but I think we need an additional patch to prevent the crash if a
> D3D10 device goes dead after the GetLayerManager() has succeeded. Right?

A D3D10 device never goes dead. It continues working, it might not display though, this is why UpdateRenderMode is called each paint to check if it's still 'working' if it's not, it will change the render mode, and this patch should automatically adapt all the layer managers. The crash in this case was UpdateRenderMode NULLs out the D2D device causing D2D surface creation functions to stop working.

Comment 6

[Scoobidiver (away)]

Is this patch going to fix bug 612515? (same STR)

Comment 7

[Bas Schouten (:bas.schouten)]

(In reply to comment #6)
> Is this patch going to fix bug 612515? (same STR)

I believe so.

Comment 8

[Bas Schouten (:bas.schouten)]

http://hg.mozilla.org/mozilla-central/rev/3ae93548e27a