bug#35575: Some graphical programs borked with Guix on Arch

bug#35575: Some graphical programs borked with Guix on Arch

[Brendan Tildesley]

With Guix on Arch, Running Guix's Godot and cool-retro-term programs
results in a completely garbled GUI. Godot ran with OpenGL ES2 instead
of ES3 looks a bit better except the fonts fail to render correctly.
Cool-retro-term complains about how using a variable width font may
cause display alignment issues, although I'm not sure that is related.
Perhaps it indicates it is a font issue? The Arch versions work just
fine. I haven't installed Guix System  on this machine yet so I'm not
sure if they would work that way.

I have seen this kind of graphical artefact before with  KDE when I've
suspended my machine and it's come back all broken.. It may be related
to the fact that I have an AMD Radeon Fury X graphics card. (I haven't
explicitly installed any proprietary drivers; I think AMDGPU is free).

Any one have any clues on how to debug such a thing?

https://brendan.scot/p/borked-godot.png

bug#35575: Some graphical programs borked with Guix on Arch

[Boris A. Dekshteyn]

Hi,

Brendan Tildesley <mail@brendan.scot> writes:

> Any one have any clues on how to debug such a thing?
>
> https://brendan.scot/p/borked-godot.png

check dmesg for something like:

--8<---------------cut here---------------start------------->8---
[  337.066640] amdgpu 0000:01:00.0: GPU fault detected: 146 0x0000480c for process Xorg pid 845 thread Xorg:cs0 pid 846
[  337.068114] amdgpu 0000:01:00.0:   VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x00000000
[  337.069354] amdgpu 0000:01:00.0:   VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x0404800C
[  337.070674] amdgpu 0000:01:00.0: VM fault (0x0c, vmid 2, pasid 32768) at page 0, read from 'TC0' (0x54433000) (72)
--8<---------------cut here---------------end--------------->8---

if you have something similar, try doing:

--8<---------------cut here---------------start------------->8---
rm -rf $HOME/.cache/mesa_shader_cache/
--8<---------------cut here---------------end--------------->8---

-- 
WBR, Boris Dekshteyn

bug#35575: Some graphical programs borked with Guix on Arch

[Brendan Tildesley]

Wow you are a genius, that fixed it. How did you know? Sorry I got no
notification of your email, maybe because you emailed bug-guix instead
of 35575@debbugs.gnu.org.

So that's fixed my issue but I wonder how it can be fixed in general so
others don't have this issue.

bug#35575: Some graphical programs borked with Guix on Arch

[Boris A. Dekshteyn]

Brendan Tildesley <mail@brendan.scot> writes:

> Wow you are a genius, that fixed it. How did you know? Sorry I got no
> notification of your email, maybe because you emailed bug-guix instead
> of 35575@debbugs.gnu.org.
>
> So that's fixed my issue but I wonder how it can be fixed in general so
> others don't have this issue.

I suspect that the problem is caused by the transition from llvm6 to
llvm7 in the mesa package.

So, i think, the right decision would be if mesa do check cache version
and invalidate it in case of mismatch. 

On the other side — Guix itself, like any other distribution, should not
change the contents of user's home directories. May be, proper solution
will be indicate in the ~guix pull --news~, like it's implemented in debian
or gentoo.

-- 
WBR, Boris Dekshteyn

bug#35575: logo,Some graphical programs borked with Guix on Arch

[Brendan Tildesley]

To follow up on this old bug, I believe the issue may come from here: 
https://gitlab.freedesktop.org/mesa/mesa/-/blob/master/src/compiler/glsl/shader_cache.cpp#L144

Mesa calculates a sha1 based on some things they reason affect the 
output, but likely it is not truely a function of every parameter than 
can make a difference to the shader output. When we updated from llvm6 
to lvm7 I'm guessing it changed the shaders somehow, and the llvm 
version is not included in the hash. Since I have zero understanding 
mesa, I'm not capable of determining the best solution. One thought is 
that if we included the mesa /gnu/store path in the calculation, this 
would make the hash's truely unique for a given mesa version, but also 
cached shaders that /would/ work would be routinely discarded after an 
update (i assume?). Would this be sensible or completely break something 
else? Should we just add the llvm version, or just start a mesa bug 
report asking for input?

The code:



ralloc_asprintf_append(&buf, "tf: %d ", prog->TransformFeedback.BufferMode);
    for (unsigned int i = 0; i < prog->TransformFeedback.NumVarying; i++) {
       ralloc_asprintf_append(&buf, "%s ",
prog->TransformFeedback.VaryingNames[i]);
    }

    /* SSO has an effect on the linked program so include this when 
generating
     * the sha also.
     */
    ralloc_asprintf_append(&buf, "sso: %s\n",
                           prog->SeparateShader ? "T" : "F");

    /* A shader might end up producing different output depending on the 
glsl
     * version supported by the compiler. For example a different path 
might be
     * taken by the preprocessor, so add the version to the hash input.
     */
    ralloc_asprintf_append(&buf, "api: %d glsl: %d fglsl: %d\n",
                           ctx->API, ctx->Const.GLSLVersion,
                           ctx->Const.ForceGLSLVersion);

    /* We run the preprocessor on shaders after hashing them, so we need to
     * add any extension override vars to the hash. If we don't do this the
     * preprocessor could result in different output and we could load the
     * wrong shader.
     */
    char *ext_override = getenv("MESA_EXTENSION_OVERRIDE");
    if (ext_override) {
       ralloc_asprintf_append(&buf, "ext:%s", ext_override);
    }

    /* DRI config options may also change the output from the compiler so
     * include them as an input to sha1 creation.
     */
    char sha1buf[41];
    _mesa_sha1_format(sha1buf, ctx->Const.dri_config_options_sha1);
    ralloc_strcat(&buf, sha1buf);

    for (unsigned i = 0; i < prog->NumShaders; i++) {
       struct gl_shader *sh = prog->Shaders[i];
       _mesa_sha1_format(sha1buf, sh->sha1);
       ralloc_asprintf_append(&buf, "%s: %s\n",
_mesa_shader_stage_to_abbrev(sh->Stage), sha1buf);

bug#35575: logo,Some graphical programs borked with Guix on Arch

[Marius Bakke]

[-- Attachment #1: Type: text/plain, Size: 1311 bytes --]

Brendan Tildesley <mail@brendan.scot> writes:

> To follow up on this old bug, I believe the issue may come from here: 
> https://gitlab.freedesktop.org/mesa/mesa/-/blob/master/src/compiler/glsl/shader_cache.cpp#L144
>
> Mesa calculates a sha1 based on some things they reason affect the 
> output, but likely it is not truely a function of every parameter than 
> can make a difference to the shader output. When we updated from llvm6 
> to lvm7 I'm guessing it changed the shaders somehow, and the llvm 
> version is not included in the hash. Since I have zero understanding 
> mesa, I'm not capable of determining the best solution. One thought is 
> that if we included the mesa /gnu/store path in the calculation, this 
> would make the hash's truely unique for a given mesa version, but also 
> cached shaders that /would/ work would be routinely discarded after an 
> update (i assume?). Would this be sensible or completely break something 
> else? Should we just add the llvm version, or just start a mesa bug 
> report asking for input?

Is this still relevant?  I haven't heard reports about this in a long
time, nor experienced it (anymore) on my super-experimental systems that
switch LLVM and Mesa versions all the time.  So I think the issue might
have been fixed upstream?

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 487 bytes --]

bug#35575: closing. bug has not reoccurred since

[Brendan Tildesley]

On 1/4/20 12:53 am, Marius Bakke wrote:
> Is this still relevant?  I haven't heard reports about this in a long
> time, nor experienced it (anymore) on my super-experimental systems that
> switch LLVM and Mesa versions all the time.  So I think the issue might
> have been fixed upstream?


Closing because the issue seems to have only occurred once  a long time 
ago and may have been fixed up stream.