1406230 - Linux: GPU Process crash @ missing FinishObjectAPPLE

Status: RESOLVED FIXED

(Core :: Graphics: WebRender, defect, P3)

Description

[Darkspirit]

Nightly 58 x64 20171005100211 de_DE @ Debian Testing (KDE / Radeon RX480).
layers.gpu-process.enabled;true
layers.gpu-process.force-enabled;true
layers.acceleration.force-enabled;true
( gfx.webrender.enabled;true )

Other remote webext panels (at their regular position on the right top, like uBlock Origin) are working fine.

I would like to use remote webextensions just for testing with webrender etc.

As I only get this gpu process crash signature when remote webextensions are enabled, I think crash signature and this text belong together.

Example addon: https://addons.mozilla.org/firefox/addon/ipvfoo-pmarks/

Comment 1

[Darkspirit]

Linux x64 debug try build from bug 1406008 comment 18:
> [GPU 18711, Compositor] WARNING: [GLX] XGetWindowAttributes() failed: file /builds/worker/workspace/build/src/gfx/gl/GLContextProviderGLX.cpp, line 932
> [GPU 18711, Compositor] WARNING: Failed to create CompositorOGL context: file /builds/worker/workspace/build/src/gfx/layers/opengl/CompositorOGL.cpp, line 148
> Crash Annotation GraphicsCriticalError: |[G0][GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT (t=14.019) |[G1][GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT (t=59.3221) [GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT

Comment 2

[Darkspirit]

If I click on the crash signature above I don't get any results. But when I go into a crash report and click on "More Reports" I get this (don't forget to adjust dates):
https://crash-stats.mozilla.com/signature/?product=Firefox&signature=gfxPlatform%3A%3AInit%20%20%7C%20mozilla%3A%3Awidget%3A%3AWindowSurfaceX11%3A%3AGetVisualFormat&date=%3E%3D2017-09-06T15%3A38%3A02.000Z&date=%3C2017-10-06T15%3A38%3A02.000Z&_columns=date&_columns=product&_columns=version&_columns=build_id&_columns=platform&_columns=reason&_columns=address&_columns=install_time&_sort=-date&page=1#reports

Currently nobody else than me tries to use OOP Webextensions (for testing) + GPU Process (for WebRender stability) on Linux, but it would be helpful for testing (as this is the expected future).

> Process Type: gpu (web)
> MOZ_CRASH Reason: MOZ_RELEASE_ASSERT(!XRE_IsGPUProcess()) (GFX: Not allowed in GPU process.)

Comment 3

[Darkspirit]

GPU Process + OOP Webextensions prevent WebRender from functioning at all:

This is what I get after enabling OOP Webextensions and restarting Nightly, when those "Panels of remote webextension icons inside locationbar are invisible" afterwards:

about:support
> Error	Compositors might be mixed (5,2)
> GP+[GFX1-]: Failed GL context creation for WebRender: 0
> GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT
> GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT

This can be fixed by just disabling the GPU Process, but this would very bad for WebRender stability (bug 1402067/regressions for example) as "full browser crashes" would be too often then.

---
GPU Process + OOP Webextensions + without Webrender
behaves the same and would only show this on about:support:
> (#0) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT
> (#1) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT

Comment 5

[Darkspirit]

(In reply to Jan Andre Ikenmeyer [:darkspirit] from comment #2)
> If I click on the crash signature above I don't get any results.

Socorro showed [@ gfxPlatform::Init | mozilla::widget::WindowSurfaceX11::GetVisualFormat ],
but it had [@ gfxPlatform::Init  | mozilla::widget::WindowSurfaceX11::GetVisualFormat ]
as html source code and two spaces (%20) in the link.
Filed bug 1424659 to let Socorro do something like "Init  |".

Comment 6

[guigs [:guigs] PST (please needinfo me!)]

I have a user experiencing crashes similar to this with the Radeon 480 x graphics - in Firefox 61 
https://crash-stats.mozilla.com/report/index/b1f131e2-20cc-45cf-b8fb-b9fe80180529

Is this the correct bug?
https://support.mozilla.org/en-US/questions/1219637

Comment 7

[Darkspirit]

(In reply to guigs [guigs] PST (please needinfo me!) from comment #6)
> I have a user experiencing crashes similar to this with the Radeon 480 x graphics - in Firefox 61 
> https://crash-stats.mozilla.com/report/index/b1f131e2-20cc-45cf-b8fb-b9fe80180529

bug 1458331: It looks like a problem with webm.

> Is this the correct bug?
> https://support.mozilla.org/en-US/questions/1219637

This bug is about WebRender on Linux with OOP webextensions which are not enabled by default yet. None of those reports have WebRender enabled and they're all on Windows.

Comment 8

[Darkspirit]

Behavior has changed a bit over time:
I haven't seen [@ gfxPlatform::Init | mozilla::widget::WindowSurfaceX11::GetVisualFormat ] for a long time.
It still doesn't work, but it doesn't seem to crash with that signature. Altough something is horribly failing in the background.

https://mozilla.github.io/mozregression/install.html

This is a day before bug 1406533, but behavior seems to be the same today:

=== The underlying problem: GPU Process, OOP Webextensions, OpenGL ===

> mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true layers.acceleration.force-enabled:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ipvfoo-pmarks/'
1. Install IPvFoo.
2. Reload the AMO page. On the right side of the locationbar you see an "?" icon. Click on it. Its panel is invisible/broken.
3. Open about:support
> Compositing	OpenGL
> [...]
> Failure Log
> (#0) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT


> mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true layers.acceleration.force-enabled:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'
1. Install uBlock Origin.
2. Click on its icon. The browser was white for a moment.
3. Open about:support
> Compositing	OpenGL
> [...]
> Failure Log
> (#0) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT


=== GPU Process, OOP Webextensions, WebRender ===

> mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true gfx.webrender.all:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ipvfoo-pmarks/'
1. Install IPvFoo.
2. Reload the AMO page. On the right side of the locationbar you see an "?" icon. Click on it. Its panel is invisible/broken. The browser was white for a moment.
3. Open about:support
> Compositing	OpenGL
> [...]
> Failure Log
> (#0) Error	Compositors might be mixed (5,2)
> (#1) 	GP+[GFX1-]: Failed GL context creation for WebRender: 0
> (#2) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT



> mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true gfx.webrender.all:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'
1. Install uBlock Origin.
2. Click on its icon. The browser was white for a moment.
3. Open about:support
> Compositing	OpenGL
> [...]
> Failure Log
> (#0) Error	Compositors might be mixed (5,2)
> (#1) 	GP+[GFX1-]: Failed GL context creation for WebRender: 0
> (#2) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT

Comment 9

[Martin Stránský [:stransky] (ni? me)]

There's a backtrace I get on Mesa DRI Intel(R) HD Graphics 530 (Skylake GT2):

#4  0x00007fdf2358f070 in <signal handler called> () at /lib64/libc.so.6
#5  0x00007fdf0e024611 in mozilla::gl::GLContext::fFinishObjectAPPLE(unsigned int, int) (this=0x7fdef3c46000, object=5890, name=4) at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/GLContext.h:3311
#6  0x00007fdf0e03361e in mozilla::layers::DirectMapTextureSource::Sync(bool) (this=0x7fdef312d900, aBlocking=true)
    at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/opengl/TextureHostOGL.cpp:346
#7  0x00007fdf0e1b33b7 in mozilla::layers::BufferTextureHost::ReadUnlock() (this=0x7fdef312b3d0)
    at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TextureHost.cpp:903
#8  0x00007fdf0e1b3466 in mozilla::layers::BufferTextureHost::UnbindTextureSource() (this=0x7fdef312b3d0)
    at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TextureHost.cpp:936
#9  0x00007fdf0e04174b in mozilla::layers::TextureHost::ReleaseCompositableRef() (this=0x7fdef312b3d0)
    at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/mozilla/layers/TextureHost.h:615
#10 0x00007fdf0e0663e4 in mozilla::layers::CompositableTextureRef<mozilla::layers::TextureHost>::~CompositableTextureRef() (this=0x7fdef40f0408, __in_chrg=<optimized out>)
    at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/mozilla/layers/TextureHost.h:247
#11 0x00007fdf0e1a8f42 in mozilla::layers::TileHost::~TileHost() (this=0x7fdef40f0408, __in_chrg=<optimized out>)
    at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TiledContentHost.h:44
#12 0x00007fdf0e1c5d50 in nsTArrayElementTraits<mozilla::layers::TileHost>::Destruct(mozilla::layers::TileHost*) (aE=0x7fdef40f0408) at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/nsTArray.h:544
#13 0x00007fdf0e1c43a4 in nsTArray_Impl<mozilla::layers::TileHost, nsTArrayInfallibleAllocator>::DestructRange(unsigned long, unsigned long) (this=0x7fdf09ffda00, aStart=0, aCount=1)
    at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/nsTArray.h:2150
#14 0x00007fdf0e1c1b01 in nsTArray_Impl<mozilla::layers::TileHost, nsTArrayInfallibleAllocator>::ClearAndRetainStorage() (this=0x7fdf09ffda00) at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/nsTArray.h:1367
#15 0x00007fdf0e1bf1d1 in nsTArray_Impl<mozilla::layers::TileHost, nsTArrayInfallibleAllocator>::~nsTArray_Impl() (this=0x7fdf09ffda00, __in_chrg=<optimized out>) at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/nsTArray.h:924
#16 0x00007fdf0e1b5478 in nsTArray<mozilla::layers::TileHost>::~nsTArray() (this=0x7fdf09ffda00, __in_chrg=<optimized out>)
    at /home/komat/tmp676-trunk-gtk3/src2/objdir/dist/include/nsTArray.h:2399
#17 0x00007fdf0e1b5d68 in mozilla::layers::TextureSourceRecycler::~TextureSourceRecycler() (this=0x7fdf09ffda00, __in_chrg=<optimized out>) at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TiledContentHost.cpp:207
#18 0x00007fdf0e1b6726 in mozilla::layers::TiledLayerBufferComposite::UseTiles(mozilla::layers::SurfaceDescriptorTiles const&, mozilla::layers::HostLayerManager*, mozilla::layers::ISurfaceAllocator*) (this=0x7fdef312b508, aTiles=..., aLayerManager=0x7fdef4126940, aAllocator=0x7fdef42fd448) at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TiledContentHost.cpp:301
#19 0x00007fdf0e1b5835 in mozilla::layers::TiledContentHost::UseTiledLayerBuffer(mozilla::layers::ISurfaceAllocator*, mozilla::layers::SurfaceDescriptorTiles const&) (this=0x7fdef312b4c0, aAllocator=0x7fdef42fd448, aTiledDescriptor=...)
    at /home/komat/tmp676-trunk-gtk3/src2/gfx/layers/composite/TiledContentHost.cpp:176

GL_APPLE_fence is not present on my extension list so it crashes as it's missing FinishObjectAPPLE routine.

Comment 10

[Martin Stránský [:stransky] (ni? me)]

Jan, can you check if the GL_APPLE_fence gl extension is present on your box where you see this crash? (by glxinfo | grep APPLE for instance). Thanks.

Comment 11

[Martin Stránský [:stransky] (ni? me)]

This seems to be rated to Bug 1265824.

Comment 12

[Martin Stránský [:stransky] (ni? me)]

Attachment: Bug 1406230 - Check for APPLE_fence/APPLE_texture_range/APPLE_client_storage extensions,

Review commit: https://reviewboard.mozilla.org/r/259900/diff/#index_header
See other reviews: https://reviewboard.mozilla.org/r/259900/

Comment 13

[Darkspirit]

(In reply to Martin Stránský [:stransky] from comment #10)
> Jan, can you check if the GL_APPLE_fence gl extension is present on your box where you see this crash? (by glxinfo | grep APPLE for instance). Thanks.

Good Morning. "glxinfo | grep APPLE" has no output, while glxinfo itself has a long output.
You can see Nvidia GLX stuff here: attachment 8995221 [details]

Comment 14

[Darkspirit]

(In reply to Martin Stránský [:stransky] from comment #12)
> Created attachment 8995458 [details]
> Bug 1406230 - Check for APPLE_fence/APPLE_texture_range/APPLE_client_storage extensions,

This patch might need to be adjusted to the reviewed changes in bug 1478704 comment 5.

Comment 15

[Alex Thayer [:alexical] (she/her)]

I believe bug 1478704 is separate. I'll have a patch up for this one by end of day. It should be as simple as turning this section:

    return mGLContext->IsExtensionSupported(gl::GLContext::APPLE_client_storage) &&
           mGLContext->IsExtensionSupported(gl::GLContext::APPLE_texture_range);

Into this:

    return mGLContext->IsExtensionSupported(gl::GLContext::APPLE_client_storage) &&
           mGLContext->IsExtensionSupported(gl::GLContext::APPLE_texture_range) &&
           mGLContext->IsExtensionSupported(gl::GLContext::APPLE_fence);

Missing the APPLE_fence requirement was just a silly oversight :( - sorry about that.

Comment 16

[Alex Thayer [:alexical] (she/her)]

Just noticed the patch - Martin, you probably just want to adjust here:

https://searchfox.org/mozilla-central/rev/d47c829065767b6f36d29303d650bffb7c4f94a3/gfx/layers/opengl/CompositorOGL.cpp#2024

Comment 17

[Matt Woodrow (:mattwoodrow)]

Comment on attachment 8995458 [details]
Bug 1406230 - Check for APPLE_fence/APPLE_texture_range/APPLE_client_storage extensions,

https://reviewboard.mozilla.org/r/259900/#review267152

I think we want to check before using DirectMapTextureSource, like Doug suggested in comment 12.

Comment 18

[Alex Thayer [:alexical] (she/her)]

Looking a little closer, I think CompositorOGL::CreateDataTextureSourceAround[1] and CompositorOGL::CreateDataTextureSourceAroundYCbCr need to fall back to the equivalent BasicCompositor methods if SupportsTextureDirectMapping() is false, which might be more what we're running into here.

[1] https://searchfox.org/mozilla-central/rev/033d45ca70ff32acf04286244644d19308c359d5/gfx/layers/opengl/CompositorOGL.cpp#1987

Comment 19

[Martin Stránský [:stransky] (ni? me)]

Thanks, I'll update the patch.

Comment 20

[Darkspirit]

For completeness:
"glxinfo | grep APPLE" is empty with proprietary Nvidia driver (known problems with and without gpu process),
but with Nouveau/Mesa (no problems since bug 1401634; it seems comment 15 wasn't needed for Nouveau) I get exactly this:
>    GL_APPLE_packed_pixels, GL_ARB_ES2_compatibility, 
>    GL_APPLE_texture_max_level, GL_EXT_base_instance,

Comment 21

[Darkspirit]

(In reply to Jan Andre Ikenmeyer [:darkspirit] from comment #8)
> === The underlying problem: GPU Process, OOP Webextensions, OpenGL ===

> > mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true layers.acceleration.force-enabled:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'
> 1. Install uBlock Origin.
> 2. Click on its icon. The browser was white for a moment.
> 3. Open about:support
> > Compositing	OpenGL
> > [...]
> > Failure Log
> > (#0) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT


Proprietary Nvidia now:

mozregression --launch 20180730221422 --pref layers.gpu-process.enabled:true layers.acceleration.force-enabled:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'

This is fine and probably also fixed by bug 1401634.
Your crash from comment 9 seems to be a different than the original bug.


> === GPU Process, OOP Webextensions, WebRender ===

> > mozregression --launch 2018-07-23 --pref layers.gpu-process.enabled:true gfx.webrender.all:true extensions.webextensions.remote:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'
> 1. Install uBlock Origin.
> 2. Click on its icon. The browser was white for a moment.
> 3. Open about:support
> > Compositing	OpenGL
> > [...]
> > Failure Log
> > (#0) Error	Compositors might be mixed (5,2)
> > (#1) 	GP+[GFX1-]: Failed GL context creation for WebRender: 0
> > (#2) 	GP+[GFX1-]: [OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT

Proprietary Nvidia now:
mozregression --launch 20180730221422 --pref layers.gpu-process.enabled:true gfx.webrender.all:true -a 'https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/'
> Compositing	OpenGL
> [...]
> Failure Log
> (#0) Error	Compositors might be mixed (5,2)
> (#1) 	GP+[GFX1-]: Failed GL context creation for WebRender: 0

The line "[OPENGL] Failed to init compositor with reason: FEATURE_FAILURE_OPENGL_CREATE_CONTEXT" is gone.
So it looks like only bug 1478454 is left for proprietary Nvidia both with and without gpu process.

Comment 22

[Martin Stránský [:stransky] (ni? me)]

(In reply to Jan Andre Ikenmeyer [:darkspirit] from comment #21)
> This is fine and probably also fixed by bug 1401634.
> Your crash from comment 9 seems to be a different than the original bug.

IMHO you're wrong here. From some reason (I'm still investigation it) the latest trunk does not use OGL compositor for WebRender but only a basic one, which apparently does not use DirectMapTextureSource by default (BasicCompositor::CreateDataTextureSourceAround is called here).

Comment 23

[Martin Stránský [:stransky] (ni? me)]

Dough, looks like the APPLE_client_storage/APPLE_texture_range checks are enough, no need to add APPLE_fence there as it's just a subset of them (IMHO). 

The issue it that at 428469:7ba07ef0e453 revision the DirectMapTextureSource is used without the CompositorOGL::SupportsTextureDirectMapping() call.

OTOH the latest trunk seems to be "fixed" as it ignores OGL compositor completely and creates only a BasicCompositor in GPU process (CompositorBridgeParent::AllocPLayerTransactionParent() has only BASIC hint). Are you aware of that change?

Comment 24

[Darkspirit]

bug 1478454 seems to be fixed with patch from bug 1479181 comment 18 as I noted there.

(Proprietary Nvidia)
Now I ran my local build like this:
./mach run --enable-crash-reporter --temp-profile --setpref gfx.webrender.all=true --setpref layers.gpu-process.enabled=true --setpref gfx.webrender.debug.profiler=true --setpref gfx.webrender.debug.compact-profiler=true https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/

uBlock's panel has a black background instead of transparency around it and it uses WR because there's a WR profiler in it. 
The black background is the same as with Nouveau and current m-c. Looking on about:support there is no fallback to OpenGL.

Comment 25

[Darkspirit]

(In reply to Martin Stránský [:stransky] from comment #9)
> There's a backtrace I get on Mesa DRI Intel(R) HD Graphics 530 (Skylake GT2):

Debian Testing, Mesa 18.1.4, Mesa DRI Intel(R) HD Graphics 630 (Kaby Lake GT2) 
(Integrated Graphics of this Desktop computer)

latest m-c:
mozregression --launch af6a7edf0069 --pref layers.gpu-process.force-enabled:true gfx.webrender.all:true gfx.webrender.debug.profiler:true gfx.webrender.debug.compact-profiler:true -a https://addons.mozilla.org/en-US/firefox/addon/ublock-origin/
* black background around OOP webext panels, everything WebRender, no fallback, no failure log on about:support

$ glxinfo | grep APPLE
    GL_APPLE_object_purgeable, GL_ARB_ES2_compatibility, 
    GL_APPLE_object_purgeable, GL_APPLE_packed_pixels, 
    GL_ANGLE_texture_compression_dxt5, GL_APPLE_texture_max_level,

Comment 26

[Martin Stránský [:stransky] (ni? me)]

I don't know why but the crash is fixed now.