NVIDIA / nvidia-container-toolkit / 34668084772

44%

main: 44%

push

github

Fix intermittent EBADF in create-symlinks hook

createSymlinkInRoot extracted the raw fd of the link's parent directory
with (*os.File).Fd() and then only used the integer in Symlinkat and
Renameat. Nothing kept the *os.File reachable after that point, so the
garbage collector could run its finalizer and close the fd while the
syscalls were still using it. The hook then failed with

  failed to create link [...]: failed to create symlink: bad file descriptor

and crun aborted container creation with
"error executing hook /usr/bin/nvidia-cdi-hook (exit code: 1)". The
failure is timing dependent and rare, roughly 1 in 1000-2000 container
starts on busy hosts, which is enough to break large multi-node jobs
where every container has to start.

v1.19.1 had `defer linkDirInRoot.Close()` here; the line was dropped in
v1.20.0. Restore it so the file stays alive for the whole function and
the descriptor is closed deterministically.

Signed-off-by: Daniel Ohayon <dohayon@meta.com>

1 of 1 new or added line in 1 file covered. (100.0%)

5952 of 13518 relevant lines covered (44.03%)

0.44 hits per line