[Flurry] screensaver crashes in __strtod_internal

[diver]

hrev50849 x86_64.

Selecting Flurry screensaver crashes ScreenSaver preflet.

Not reproducible in x86_gcc2.

[diver]

It looks like it crashes here ​https://cgit.freedesktop.org/mesa/glu/tree/src/libutil/mipmap.c#n3414

[Premislaus]

Another crash report.

[Andrew Aldridge]

This is because the FlurryView::AttachedToWindow() logic looks like the following:

void FlurryView::AttachedToWindow() {
  LockGL();
  BGLView::AttachedToWindow(); // Calls UnlockGL()
  // Do stuff assuming a context is active
}

When glGetString(GL_VERSION) is called, it returns a null string. The GLife screensaver has the same problem, but instead of crashing it just renders incorrectly.

For both Flurry and GLife, moving the LockGL call *after* the call to BGLView::AttachedToWindow fixes the problem.

I'll try to create a patch tomorrow.

[Andrew Aldridge]

(Maybe BGLView::AttachedToWindow() shouldn't unconditionally unlock the GL context; I can persue that route instead if it's preferable)

[waddlesplash]

What did BeOS do re. locking? We should try to mimic that.

[Andrew Aldridge]

The BeBook seems to indicate that the existing behavior of the screensavers should be correct, meaning my patch should be rejected. The current BGLView::AttachedToWindow() locking logic was added to Haiku in fafc6e4b776ba7c9eff20696976d9fdb85113682 (2010) without explanation.

There are two options I can see:

• Revert that part of the change, making Haiku match BeOS more closely based on the BeBook's code samples, but breaking applications written to depend on Haiku's current behavior.
• Only call UnlockGL() iff the initial state of BGLView::fDisplayLocker::IsLocked() is false. That is, restore the original state right after setting up the glViewport.

I'll pursue the second option as a pull request to the mesa package in haikuports, if that's the right procedure?

[waddlesplash]

CC'ing kallisti5 as he's in charge of Mesa-related stuff.

Yes, that's probably correct. Note that there are 2 versions of Mesa though (Mesa 7 for GCC2 and current Mesa for GCC5+) and both will need to be patched I think.

[PulkoMandy]

Aren't LockGL() and UnlockGL() recursive locks? (sorry, I'm not familiar with the API)

If they are, then there should be no problem here - AttachedToWindow could safely call LockGL then UnlockGL which would restore the initial state of the lock.

A patch was submitted here: ​https://github.com/haikuports/haikuports/pull/2070/files Submitting directly to github.com/haiku/mesa (for gcc2) and upstream mesa (for gcc5) would be better than having the patch in haikuports.

[Andrew Aldridge]

LockGL() and UnlockGL() currently don't behave recursively. I think they probably should?

I appreciate everybody taking the time to help me muddle my way through what a proper fix for this looks like :D

[waddlesplash]

The Be Book does not specify whether or not they are recursive. Guess someone needs to fire up BeOS and find out...

[PulkoMandy]

Let's dig out that BeOS machine. Maybe someone has code I could easily test?

[cocobean]

Rebuilt Flurry code and tested on hrev51880 x86_64. I can confirm this points to the patchset for Mesa 17.1.7 as the culprit. The only other warning is a "_start" symbol resolving from Flurry.

Confirmed the odd default misrendering with GLife (although GLTeapots rendering looks fine). Mesa 7.9 x86 renders those screensavers fine on Haiku x86. Didn't test Mesa 17.1 on Haiku x86 nor Mesa 7.9 on x86_64.

I can assist work on a 'clean room' Mesa 17.1.10 re-implementation for Haiku - if we all come to the same conclusions.

[Alexander von Gluck]

The core problem is people keep bumping the version of Mesa in our Haikuports repository without testing if it actually works. Mesa 12-15 were stable, but without the bandwith to work on it bugs creep in.

I've been trying to keep up, but real life and all the other Haiku work I have get in the way.

The SCons build system we use in Mesa is slowly being decommissioned. I've added Haiku support to their new Meson build system... but haven't had time to track down all the bugs.

Help wanted!

And *please* don't start talking of forking Mesa. The project moves fast... lets just not bump versions until we know that the version we bump to is stable + functional :-)

[PulkoMandy]

We need a release branch at haikuports and not let random people push/merge changes to it; at least for packages required for Haiku to run.

[cocobean]

Issue (Haiku hrev51877, Mesa 17.1.7-7): GLinfo - Segmentation fault on exit with Mesa 17.1.7, GLife - Misrendering (cubes larger than normal), Flurry - fails to start properly.

Note: haiku hrev51880 x86 - Mesa 7.9.2 works and renders fine (i.e. good enough).

This Mesa 17.0.7 setup renders fine with Flurry on Linux:

OpenGL vendor string: X.Org
OpenGL renderer string: Gallium 0.4 on AMD RV710 (DRM 2.49.0 / 4.10.0-38-generic, LLVM 4.0.0)
OpenGL core profile version string: 3.3 (Core Profile) Mesa 17.0.7
OpenGL core profile shading language version string: 3.30
OpenGL core profile context flags: (none)
OpenGL core profile profile mask: core profile
OpenGL core profile extensions:
OpenGL version string: 3.0 Mesa 17.0.7
OpenGL shading language version string: 1.30
OpenGL context flags: (none)
OpenGL extensions:
OpenGL ES profile version string: OpenGL ES 3.0 Mesa 17.0.7
OpenGL ES profile shading language version string: OpenGL ES GLSL ES 3.00
OpenGL ES profile extensions:

On Haiku hrev51877 x86_64 with radeon_hd driver and Mesa 17.1.7-7:

Gallium 0.4 on llvmpipe (LLVM 5.0, 128 bits)
VMware Inc.
3.0 Mesa 17.1.7
GLU 1.3, GLU 1.5

[cocobean]

GLInfo crash on exit

[cocobean]

Flurry crash on hrev51877 x86_64

[cocobean]

Syslog after crashes hrev51877 x86_64

[cocobean]

See attachment: https://dev.haiku-os.org/attachment/ticket/13191/haiku_flurry_x64.jpeg

Confirming that GLife, GL Info, and Flurry all work, exit, and render properly on Haiku hrev49856 x86_64:

Software Rasterizer
Mesa Project
2.1 Mesa 10.5.2
GLU 1.3, GLU 1.5

Note: A user also confirmed that Haiku x86_64 hrev51173 with mesa_swpipe 17.0.2-2 works.

IMPORTANT: The current nightly ISO x86_64 image builds have the issue.

[cocobean]

Haiku hrev49856 x86_64 w/Mesa 10.5.2 working properly with GLife, GL Info, and Flurry.

[cocobean]

Mesa-swrast review on Haiku hrev51904 x86_64

[cocobean]

The issue is within mesa_swpipe. I swapped out mesa_swpipe with mesa_swrast and it fixed all my 3D rendering issues and crash issues. So, that is the quick workaround for now. Need to review GLView locking/unlocking in apps per dev specs. Fixing mesa_swpipe is long term fix. Short term fix is to provide mesa_swrast in Haiku Depot/nightly images.