Filter already tracked directories from ldcache update

[elezar]

This change ensures that directories that are already included in the ld.conf are not added to the NVCR-specific config files. The intent is to not inadventently promote existing directories to a higher priority.

See #123 where this was already an issue in the libnvidia-container-based implementation.

Testing

build the following docker image

docker build -t libordering \
            - <<EOF
FROM ubuntu
ENV NVIDIA_VISIBLE_DEVICES=all
RUN mkdir -p /extra/lib
RUN cp /usr/lib/$(uname -m)-linux-gnu/libc.so.? /extra/lib/
RUN echo "/extra/lib" > /etc/ld.so.conf.d/00-xxx.conf
RUN ldconfig
EOF

Without the nvidia runtime:

$ docker run --rm -ti --runtime=runc libordering bash -c "ldconfig -p | grep libc.so."
        libc.so.6 (libc6,AArch64) => /extra/lib/libc.so.6
        libc.so.6 (libc6,AArch64) => /lib/aarch64-linux-gnu/libc.so.6

Without this change:

$ docker run --rm -ti --runtime=nvidia libordering bash -c "ldconfig -p | grep libc.so."
        libc.so.6 (libc6,AArch64) => /usr/lib/aarch64-linux-gnu/libc.so.6
        libc.so.6 (libc6,AArch64) => /extra/lib/libc.so.6

(Note that the libc.so.6 in /usr/lib/aarch64-linux-gnu takes precedence).

With this change:

$ docker run --rm -ti --runtime=nvidia libordering bash -c "ldconfig -p | grep libc.so."
        libc.so.6 (libc6,AArch64) => /extra/lib/libc.so.6
        libc.so.6 (libc6,AArch64) => /lib/aarch64-linux-gnu/libc.so.6

[elezar]

/cherry-pick release-1.18

[ArangoGutierrez]

LGTM
2 non Blocking comments

[ArangoGutierrez]

Question: If the function doesn't need to access or modify the state of an Ldconfig object, why not to be a regular standalone function?

[elezar]

I was originally accessing a member, but I can remove it.

[jasonzlai]

(Jason from Uber AI Infra here - we filed the ticket) What will be passed in as directories?

[elezar]

directories contains the container paths of parent directories for any CUDA user-mode driver libraries that are mounted from the host. Note that at present the contianer paths of the libraries are the same as the host path.

On an ubuntu based system the list is typically:

/usr/lib/${ARCH_SPECIFIC_LIB_DIR}/
/usr/lib/${ARCH_SPECIFIC_LIB_DIR}/vdpau

which results in the following CDI hook:

        - hookName: createContainer
          path: /usr/bin/nvidia-cdi-hook
          args:
            - nvidia-cdi-hook
            - update-ldcache
            - --folder
            - /lib/x86_64-linux-gnu
            - --folder
            - /lib/x86_64-linux-gnu/vdpau
          env:
            - NVIDIA_CTK_DEBUG=false

If you wanted to know what the list is on your system, you could run the nvidia-ctk cdi generate command and check the output for the generated update-ldcache hook.

[elezar]

@jasonzlai do you have a simple reproducer Dockerfile that we could add to our test suite to address this?

Additional changes.

[ArangoGutierrez]

lgtm

[github-actions]

🤖 Backport PR created for release-1.18: #1449 ✅