Ludovic Courtès <ludovic.courtes@inria.fr> skribis:
Toggle quote (13 lines)
> The audit library created by ‘guix pack -RR’, called ‘pack-audit.so’, is
> dynamically-linked. Thus, it cannot be loaded (unless there are working
> libgcc.so and libc.so in the search path?), leading to failures like
> this:
>
> $ GUIX_EXECUTION_ENGINE=performance ./bin/sinfo --version
> ERROR: ld.so: object '/…/lcourtes/tmp/t/gnu/store/0n6nnvzgxyisg0bszb5zqxp2gzdwh7h3-pack-audit.so' cannot be loaded as audit interface: cannot open shared object file; ignored.
> /…/lcourtes/tmp/t/gnu/store/3dhy2f3djmm1h5ix5aa84lrskxzrl6d0-slurm-19.05.3-2/bin//sinfo: error while loading shared libraries: libslurmfull.so: cannot open shared object file: No such file or directory
>
> The root cause is that ‘pack-audit.so’ NEEDs libgcc_s.so, which on this
> machine could not be loaded, and thus resolving the other dependencies
> of the wrapped executable, like ‘libslurmfull.so’ here, fails.
Fixed by commit c6c0d5a22c2ee3d7164dab0129b2e4852a4ae76c.
How could this happen? I think there are two key factors:
1. Most likely I tested v1 of the patch series on “real” non-Guix
machines, but when iterating on v2 (which removed ‘--library-path’
from the ld.so invocation¹) I probably relied on
‘tests/guix-pack-relocatable.sh’, which emulates a machine where
/gnu is empty by mounting a tmpfs on it.
2. There was another bug in ‘exec_with_loader’ whereby it would
symlink its store to that writable /gnu. Thus, the tests would run
as if the store was already available under /gnu/store, thereby
hiding the actual issue. This second bug is fixed by
c088aa2988ef82289c87ebfd6d07d8f1464dd8f0.
AFAICS we’re all fine now.
You’re very welcome to give it a spin on a Guix-less machine! Just do:
guix pack -RR your favorite packages -S /bin=bin
Send the resulting tarball to that machine, and then:
mkdir test
cd test
tar xf /path/to/pack.tar.gz
GUIX_EXECUTION_ENGINE=fakechroot ./bin/your-favorite-command
Ludo’.
