Fix crash after reseting device.

Fix crash after reseting device.

After calling Reset in testDeviceLost to acknowledge a display mode change, D3D sometimes crashes on the next Present call on an existing swap chain. I couldn't get it to happen with nVidia. The stack looks like this with the debug runtime:

 	ntdll.dll!_ZwRaiseException@12()  + 0x12 bytes	
 	ntdll.dll!_ZwRaiseException@12()  + 0x12 bytes	
 	d3d9d.dll!__DDAssert()  + 0x1ca bytes	
 	d3d9d.dll!CSwapChain::PresentMain()  + 0xf52 bytes	
 	d3d9d.dll!CSwapChain::Present()  + 0x24 bytes	
 	libEGL.dll!egl::Surface::swapRect(int x, int y, int width, int height)  Line 458 + 0x2a bytes	C++
>	libEGL.dll!egl::Surface::swap()  Line 605	C++
 	libEGL.dll!eglSwapBuffers(void * dpy, void * surface)  Line 1063 + 0x8 bytes	C++
 	gl_wrapper.dll!gfx::EGLApiBase::eglSwapBuffersFn(void * dpy, void * surface)  Line 813 + 0x18 bytes	C++
 	gl_wrapper.dll!gfx::NativeViewGLSurfaceEGL::SwapBuffers()  Line 363 + 0x35 bytes	C++
 	gl_wrapper.dll!gfx::GLSurfaceAdapter::SwapBuffers()  Line 201 + 0x1d bytes	C++
 	content.dll!content::PassThroughImageTransportSurface::SwapBuffers()  Line 216 + 0xb bytes	C++

The error is "BitBlt or StretchBlt failed in Present".

With the release runtime either Present continually fails with E_INVALIDARG or, less commonly, crashes.

It seems that by recreating all the swap chains after reset, it recovers. I didn't call resetSwapChain because there is no need to recreate all the other surfaces.

As far as I know, this should not be necessary with WDDM.
Committed: https://code.google.com/p/angleproject/source/detail?r=1993

[apatrick1, 2013-03-08 00:21:11]

This is a bit of a hack. Let me know if you have better ideas.

Thanks. Al.

[apatrick1, 2013-03-11 19:54:00]

Is this going to be okay? The GPU process crash rate has increased about 4x
because of this.

[apatrick1, 2013-03-13 18:55:16]

This issue has made it to Chrome Dev channel so I've got to fix it pronto.
Handing over to jbauman for review but I'm happy to change it later if you have
input.

[John Bauman, 2013-03-13 19:07:59]

lgtm

This is probably ok, but I'd be worried about getting into a state where the
backbuffers have all been thrown away and the application doesn't realize it.

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Display.cpp
File src/libEGL/Display.cpp (right):

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Display.cpp#newco...
src/libEGL/Display.cpp:923: // Existing swap chains sometimes crash on the next
present after a reset. Seems to be ATI specific.
Unfortunately, I guess this will cause us to throw away the surface even on
WDDM, and the application won't know that anything went wrong because it won't
get a device lost.

Do we know if there's a way to narrow down when this is needed or not?

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Surface.cpp
File src/libEGL/Surface.cpp (right):

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Surface.cpp#newco...
src/libEGL/Surface.cpp:373: result = mSwapChain->GetBackBuffer(0,
D3DBACKBUFFER_TYPE_MONO, &mBackBuffer);
I think you need to release mBackBuffer as well.

[nicolas, 2013-03-13 20:21:07]

Sorry for the delay. Other than what John already mentioned it looks ok to me
too. If it could only be done for ATI cards that would be even better, since
other hardware shouldn't pay for what appears to be an ATI driver bug, in my
humble opinion.

[apatrick1, 2013-03-13 20:38:10]

Since I wrote this patch I've started seeing reports of a similar crash that I
can't reproduce on Optimus. They will identify themselves as Intel vendor.

Thread 0 *CRASHED* ( EXCEPTION_ACCESS_VIOLATION_EXEC @ 0x00000000 )

0x74558d50	 [nvumdshim.dll]	 + 0x00018d50]	
0x6ff5b30c	 [d3d9.dll]	 + 0x0011b30c]	CBatchFilterI::LHBatchOpenResource(void
*,_D3DDDIARG_OPENRESOURCE *)
0x6fe69b5f	 [d3d9.dll]	 + 0x00029b5f]	OpenResourceLH
0x6fe69cdd	 [d3d9.dll]	 + 0x00029cdd]	OpenResourceLH
0x6fe47640	 [d3d9.dll]	 + 0x00007640]	CSwapChain::PresentMain(tagRECT const
*,tagRECT const *,HWND__ *,_RGNDATA const *,unsigned long)
0x6fe47842	 [d3d9.dll]	 + 0x00007842]	CSwapChain::Present(tagRECT const
*,tagRECT const *,HWND__ *,_RGNDATA const *,unsigned long)
0x743d19fc	 [libegl.dll]	 -
surface.cpp:427]	egl::Surface::swapRect(int,int,int,int)
0x743d1b3d	 [libegl.dll]	 - surface.cpp:573]	egl::Surface::swap()
0x743d8754	 [libegl.dll]	 - libegl.cpp:1063]	eglSwapBuffers
0x657e367d	 [chrome.dll]	 -
gl_surface_egl.cc:371]	gfx::NativeViewGLSurfaceEGL::SwapBuffers()

Hopefully this will fix both.

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Display.cpp
File src/libEGL/Display.cpp (right):

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Display.cpp#newco...
src/libEGL/Display.cpp:923: // Existing swap chains sometimes crash on the next
present after a reset. Seems to be ATI specific.
This path is always WDDM. I'll update the comment though. From the crash
reports, it is not ATI specific.

I think what might be happening is the call to Reset resets the, dummy in our
case, default device swap chain but additional swap chains are left in some kind
of broken state.

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Surface.cpp
File src/libEGL/Surface.cpp (right):

https://codereview.appspot.com/7567045/diff/5001/src/libEGL/Surface.cpp#newco...
src/libEGL/Surface.cpp:373: result = mSwapChain->GetBackBuffer(0,
D3DBACKBUFFER_TYPE_MONO, &mBackBuffer);
On 2013/03/13 19:07:59, John Bauman wrote:
> I think you need to release mBackBuffer as well.

Right.