[Fixed] Issue with mesa-git (Manjaro ARM Stable Update 2020-09-22))

[carlosqueso]

Hi,

Updating to mesa-git-20.3.0_devel.128278.20a4235c4c9-1 caused a kwin_x11 crash, preventing it to use OpenGL as the compositor engine which resulted in a sluggish device (even after rebooting it).

After rebooting, the crash didn’t happen again, but it appeared that I’d lost transparency (during animations like moving windows, using the drop-down terminal, moving the mouse cursor, even the shutdown menu background was completely black) and the system felt sluggish when it used to display transparency before the update (like it was dropping frames and it kinda felt like it was using the CPU instead of the GPU to render the desktop).

To fix this, downgrade mesa-git to the previous version (mesa-git-20.3.0_devel.127750.ac6b8e42ce2-1), then go to System Settings > Hardware > Display and Monitor > Compositor. There should be a message on top saying that OpenGL detection was turned off because it caused crashes in the past and asking to re-enable it. Re-enabling it and rebooting fixed the issue for me.

Hint to everyone here: This issue isn't going to automatically fix itself.

I haven't been able to reproduce this myself, so someone here who has the problem will have to help find the problem by bisecting.

You'll have to use git bisect on Mesa, with a "good" commit of ac6b8e42ce2 and a "bad" commit of 20a4235c4c9 (man git-bisect for docs).

Configure Mesa from the build directory like this:

Code:

meson -Dgallium-drivers=panfrost,kmsro -Dvulkan-drivers= -Db_ndebug=false --buildtype=plain

Compile and install with `ninja; sudo ninja install`

Install ccache beforehand to speed up the bisect process significantly.



Once an issue is narrowed down to a specific commit, it's usually pretty easy to find what the problem is and fix it.

I'm happy to put some time in helping with this, but while I'm fully capable of running commands and determining if something is broken, I'm not a developer.  I tried to run the meson command and got asked for a directory.  Do I need to navigate somewhere, or should a specify a directory?

[icecream95]

Hint to everyone here: This issue isn't going to automatically fix itself.

I haven't been able to reproduce this myself, so someone here who has the problem will have to help find the problem by bisecting.

You'll have to use git bisect on Mesa, with a "good" commit of ac6b8e42ce2 and a "bad" commit of 20a4235c4c9 (man git-bisect for docs).

Configure Mesa from the build directory like this:

Code:

meson -Dgallium-drivers=panfrost,kmsro -Dvulkan-drivers= -Db_ndebug=false --buildtype=plain

Compile and install with `ninja; sudo ninja install`

Install ccache beforehand to speed up the bisect process significantly.



Once an issue is narrowed down to a specific commit, it's usually pretty easy to find what the problem is and fix it.

I'm happy to put some time in helping with this, but while I'm fully capable of running commands and determining if something is broken, I'm not a developer.  I tried to run the meson command and got asked for a directory.  Do I need to navigate somewhere, or should a specify a directory?

With a new enough meson, the command should work as is if you're in a 'build' subdirectory of the root Mesa directory.

You should add  -Dlibunwind=false  to the meson command, though, as on ARM libunwind is broken for many distros.

Note that `git bisect` needs to be run from the root directory of the repository, but you can pass `-C build` to ninja to avoid having to go back into the build directory. You shouldn't have to re-run meson at all while bisecting.

[MeLikeyTacos]

I think I will just stick with the plain vanilla mesa. The only reason I installed mesa-git was it was mentioned as being beneficial in the wiki, but the normal mesa seems fine.

[icecream95]

I'm happy to put some time in helping with this, but while I'm fully capable of running commands and determining if something is broken, I'm not a developer.  I tried to run the meson command and got asked for a directory.  Do I need to navigate somewhere, or should a specify a directory?

Have you made any progress on this?

Mesa 20.3 will be released in just a few weeks, so it would be good to have a fix by then.

[JasonG-FR]

> Hint to everyone here: This issue isn't going to automatically fix itself.

Obviously. I don't expect it to fix itself automatically, and that's exactly why I'm trying to help.

> Have you made any progress on this?

No. I probably did something wrong though because I could only guess the commands to run from your previous posts and 'man git-bisect'.

Here is what I did:

Code:

git clone https://github.com/mesa3d/mesa.git
cd mesa
mkdir build
git bisect start
git bisect good ac6b8e42ce2
git bisect bad 20a4235c4c9
cd build
meson -Dgallium-drivers=panfrost,kmsro -Dvulkan-drivers= -Db_ndebug=false --buildtype=plain -Dlibunwind=false
cd ..
ninja -C build
sudo ninja install -C build/
reboot

For each bisect I did this (I never experienced the crash):

Code:

git bisect good
ninja -C build
sudo ninja install -C build/
reboot

In the end the git bisect result says that the last commit (20a4235c4c9) is the bad one.

Because I found this very unlikely, I compiled and installed this supposedly bad commit by following the same procedure (except I did 'git checkout 20a4235c4c9' instead of 'git bisect good') and I didn't have any issue after rebooting the PBP.

In fact, I doubt that 'sudo ninja install -C build/' worked because 'glxinfo' always returned 'mesa 20.2.1' instead of 'mesa 20.3.x'.

Am I doing something wrong?

[NathanB]

Code:

meson -Dgallium-drivers=panfrost,kmsro -Dvulkan-drivers= -Db_ndebug=false --buildtype=plain -Dlibunwind=false

In fact, I doubt that 'sudo ninja install -C build/' worked because 'glxinfo' always returned 'mesa 20.2.1' instead of 'mesa 20.3.x'.

Am I doing something wrong?

Hi JasonG-FR,

You need to add --prefix=/usr to meson command otherwise the files will install to /usr/local.

Also, I couldn't find anything wrong with any commit in the mesa git repo. This is an issue with the mesa-git package in the pacman repository. Even 20a4235c4c9 from the git repo doesn't cause a kwin crash. If I pull the mesa-git package down with Pacman, it causes the crashes.

[carlosqueso]

I'm happy to put some time in helping with this, but while I'm fully capable of running commands and determining if something is broken, I'm not a developer.  I tried to run the meson command and got asked for a directory.  Do I need to navigate somewhere, or should a specify a directory?

Have you made any progress on this?

Mesa 20.3 will be released in just a few weeks, so it would be good to have a fix by then.

Unfortunately, the day after I started on this, all hell broke loose.  This is my first time on this computer since.  It looks like JasonG-FR beat me to it anyway. Sorry.

[icecream95]

Code:

meson -Dgallium-drivers=panfrost,kmsro -Dvulkan-drivers= -Db_ndebug=false --buildtype=plain -Dlibunwind=false

In fact, I doubt that 'sudo ninja install -C build/' worked because 'glxinfo' always returned 'mesa 20.2.1' instead of 'mesa 20.3.x'.

Am I doing something wrong?

Hi JasonG-FR,

You need to add --prefix=/usr to meson command otherwise the files will install to /usr/local.

Also, I couldn't find anything wrong with any commit in the mesa git repo. This is an issue with the mesa-git package in the pacman repository. Even 20a4235c4c9 from the git repo doesn't cause a kwin crash. If I pull the mesa-git package down with Pacman, it causes the crashes.

Can you try building Mesa using the exact PKGBUILD used for mesa-git, to see if compiling it locally still causes the crash?

If it does crash, try removing the line '-D b_lto=true' from the PKGBUILD and rebuilding (makepkg -ef).

If it works then, it is likely a compiler bug. Otherwise, you can try bisecting using the PKGBUILD (still with the lto line removed to keep down build times), again using `makepkg -ef` to rebuild the package.

[as400]

@icecream95 can you take a look at this thread ? ---> https://forum.pine64.org/showthread.php?...4#pid82324

[JasonG-FR]

Can you try building Mesa using the exact PKGBUILD used for mesa-git, to see if compiling it locally still causes the crash?

If it does crash, try removing the line '-D b_lto=true' from the PKGBUILD and rebuilding (makepkg -ef).

If it works then, it is likely a compiler bug. Otherwise, you can try bisecting using the PKGBUILD (still with the lto line removed to keep down build times), again using `makepkg -ef` to rebuild the package.

I built mesa-git using the PKGBUILD and it still causes the crash, even by removing the '-D b_lto=true' line.

The first compilation took close to one hour to complete on the PBP, but rebuilding using 'makepkg -ef' only took a few minutes, is that the expected behavior or am I doing something wrong?

I'll try to do the bisecting on the mesa-git repo to see if I can find something.