Skip to content

Force GCC linker to see that we need pthread to work around nvidia driver bug.

Robert Maynard requested to merge robertmaynard/vtk-m:correct_transfer_to_opengl_test into master

  1. Some Linux distributions default linker implicitly enables the as-needed linking flag. This means that your shared library or executable will only link to libraries from which they use symbols. So if you explicitly link to pthread but don't use any symbols you wont have a 'DT_NEEDED' entry for pthread.

  2. NVidia libGL (driver version 352 ) uses pthread but doesn't have a DT_NEEDED entry for the library. When you run ldd or readelf on the library you won't detect any reference to the pthread library. Aside this is odd since the mesa version does explicitly link to pthread. But if you run the following command: "strings /usr/lib/nvidia-352/libGL.so.1 | grep pthread | less" You will see the following: { pthread_create pthread_self pthread_equal pthread_key_crea ... libpthread.so.0 libpthread.so pthread_create } This is very strong evidence that this library is using pthread.

  3. So what does this all mean? It means that on system that use the linking flag 'as-needed', are using the nvidia driver, and don't use pthread will generate binaries that crash on launch. The only way to work around this issue is to do either: A: Specify 'no-as-needed' to the linker potentially causing over-linking and a slow down in link time B: Use a method from pthread, making the linker realize that pthread is needed.

We went with method B.

Merge request reports