libhybris: eglplatform_wayland.so unresolved symbol issue

The problem

Attempt to start plasma mobile session using kwin_hwcomposer backend results in kwin_wayland running, but plasmashell not present, and logs:

klte:~$ cat /tmp/plasmashell_logs 
Error: could not determine $DISPLAY.
org.kde.kwindowsystem: Loaded plugin "/usr/lib/qt5/plugins/kf5/org.kde.kwindowsystem.platforms/KF5WindowSystemKWaylandPlugin.so" for platform "wayland"
Using Wayland-EGL
ERROR: /usr/lib/libhybris/eglplatform_wayland.so
        Error relocating /usr/lib/libhybris/eglplatform_wayland.so: hybris_egl_display_get_mapping: symbol not found
Assertion failed: 0 (ws.c: _init_ws: 49)
KCrash: Attempting to start /usr/bin/plasmashell from kdeinit
Warning: connect() failed: : Connection refused
KCrash: Attempting to start /usr/bin/plasmashell directly
KCrash: Application 'plasmashell' crashing...
org.kde.kwindowsystem: Loaded plugin "/usr/lib/qt5/plugins/kf5/org.kde.kwindowsystem.platforms/KF5WindowSystemKWaylandPlugin.so" for platform "wayland"
Using Wayland-EGL
ERROR: /usr/lib/libhybris/eglplatform_wayland.so
        Error relocating /usr/lib/libhybris/eglplatform_wayland.so: hybris_egl_display_get_mapping: symbol not found
Assertion failed: 0 (ws.c: _init_ws: 49)

"Famous issue"

Jan 23 17:01:21 <NotKit> Using Wayland-EGL … ERROR: /usr/lib/libhybris/eglplatform_wayland.so …
    Error relocating /usr/lib/libhybris/eglplatform_wayland.so: hybris_egl_display_get_mapping: symbol not found
Jan 23 17:01:25 <NotKit> oh, that famous issue
Jan 23 17:01:53 <bhushanshah> huh?
Jan 23 17:02:18 <NotKit> I think that also happened on Arch
Jan 23 17:02:34 <bhushanshah> do you know solution?
Jan 23 17:04:52 <NotKit> I remember that it required linker arguments change during libhybris build
Jan 23 17:05:24 <NotKit> used export LD_PRELOAD=/usr/lib/libEGL.so.1.0.0 as quick hack

Turns out that issue is known. Since 2017!

krnlyng    bshah, try adding -lEGL to LDFLAGS of hybris/egl/platforms/wayland/Makefile.am

relevant chat logs link: http://www.merproject.org/logs/%23sailfishos-porters/%23sailfishos-porters.2017-10-01.log.html

Workaround:

Edit plaMo startup script /usr/bin/plasma-phone like that, by adding a line with LD_PRELOAD definition before the line that starts plsamashell:

...
export LD_PRELOAD=/usr/lib/libhybris-egl/libEGL.so.1.0.0
exec /usr/bin/plasmashell -p org.kde.plasma.phone 2>/tmp/plasmashell_logs

Explanation

The issue is that hybris's eglplatform_wayland.so plugin is using a function hybris_egl_display_get_mapping() that is implemented inside a hybris's own libEGL.so (probably some kind of private hybris API). And eglplatform_wayland.so is not linked to libEGL.so, because, I assume: in libhybris's buildsystem plugins are built first, then libEGL.so is built after plugins. (libhybris provides its own libEGL to be used instead of mesa's).

No platform plugin in libhybris tree has -lEGL link flag and the package itself of course cannot depend on mesa (as provider of libEGL). Probably platform plugin shouldn't link to libEGL at all (because it loads android libs at runtime) and this is some bug in libhybris, or misconfiguration/misplacement of hybris_egl_display_get_mapping() function. Hard to tell.

The reason this bug is unknown for other distros is probably because something else loads libEGL (provided by libhybris) in process before eglplatform_wayland is loaded, as proved by LD_PRELOAD hack working.

Links

• function definition is here: https://github.com/libhybris/libhybris/blob/master/hybris/egl/egl.c#L171
• usage is here: https://github.com/libhybris/libhybris/blob/master/hybris/egl/platforms/wayland/eglplatform_wayland.cpp#L214

Possible solutions:

1. LD_PRELOAD hack mentioned above is only good as temporary solution, or to prove the issue

2. We can force eglplatform_wayland to link to libEGL, by just adding -lEGL to linker flags. But because libEGL does not exist at the time of eglplatform_wayland linking we would have to add mesa-dev to deps to have libEGL available and avoid linking error. Another problem is that --as-needed is passed to linker flags, and libraries are linked only if they are really used. So we have to actually use some public libEGL API. I tried this option in https://gitlab.com/minlexx/pmaports/commit/26c5346bcf45fb51368a343b222d0ff6c65b1247 (I used eglQueryString() as harmless call) but decided that this it too much hacks already and wrong overall.

3. Apparently platform plugin is not supposed to link to to libEGL. So, the correct approach would be to just load libEGL at runtime using dlopen() and resolve hybris_egl_display_get_mapping() with dlsym().

I'm going to fix that.