Package Instance

Apache Hadoop argument injection vulnerability
Apache Hadoop's `FileUtil.unTar(File, File)` API does not escape the input file name before being passed to the shell. An attacker can inject arbitrary commands. This is only used in Hadoop 3.3 InMemoryAliasMap.completeBootstrapTransfer, which is only ever run by a local user. It has been used in Hadoop 2.x for yarn localization, which does enable remote code execution. It is used in Apache Spark, from the SQL command ADD ARCHIVE. As the ADD ARCHIVE command adds new binaries to the classpath, being able to execute shell scripts does not confer new permissions to the caller. SPARK-38305. "Check existence of file before untarring/zipping", which is included in 3.3.0, 3.1.4, 3.2.2, prevents shell commands being executed, regardless of which version of the hadoop libraries are in use. Users should upgrade to Apache Hadoop 2.10.2, 3.2.4, 3.3.3 or upper (including HADOOP-18136).

Untrusted Search Path
Relative library resolution in linux container-executor binary in Apache Hadoop 3.3.1-3.3.4 on Linux allows local user to gain root privileges. If the YARN cluster is accepting work from remote (authenticated) users, this MAY permit remote users to gain root privileges.

Hadoop 3.3.0 updated the " YARN Secure Containers https://hadoop.apache.org/docs/stable/hadoop-yarn/hadoop-yarn-site/SecureContainer.html " to add a feature for executing user-submitted applications in isolated linux containers.

The native binary HADOOP_HOME/bin/container-executor is used to launch these containers; it must be owned by root and have the suid bit set in order for the YARN processes to run the containers as the specific users submitting the jobs.

The patch " YARN-10495 https://issues.apache.org/jira/browse/YARN-10495 . make the rpath of container-executor configurable" modified the library loading path for loading .so files from "$ORIGIN/" to ""$ORIGIN/:../lib/native/". This is the a path through which libcrypto.so is located. Thus it is is possible for a user with reduced privileges to install a malicious libcrypto library into a path to which they have write access, invoke the container-executor command, and have their modified library executed as root.
If the YARN cluster is accepting work from remote (authenticated) users, and these users' submitted job are executed in the physical host, rather than a container, then the CVE permits remote users to gain root privileges.

The fix for the vulnerability is to revert the change, which is done in YARN-11441 https://issues.apache.org/jira/browse/YARN-11441, "Revert YARN-10495". This patch is in hadoop-3.3.5.

To determine whether a version of container-executor is vulnerable, use the readelf command. If the RUNPATH or RPATH value contains the relative path "./lib/native/" then it is at risk

$ readelf -d container-executor|grep 'RUNPATH\|RPATH' 
0x000000000000001d (RUNPATH) Library runpath: [$ORIGIN/:../lib/native/]

If it does not, then it is safe:

$ readelf -d container-executor|grep 'RUNPATH\|RPATH' 
0x000000000000001d (RUNPATH) Library runpath: [$ORIGIN/]

For an at-risk version of container-executor to enable privilege escalation, the owner must be root and the suid bit must be set

$ ls -laF /opt/hadoop/bin/container-executor
---Sr-s---. 1 root hadoop 802968 May 9 20:21 /opt/hadoop/bin/container-executor

A safe installation lacks the suid bit; ideally is also not owned by root.

$ ls -laF /opt/hadoop/bin/container-executor
-rwxr-xr-x. 1 yarn hadoop 802968 May 9 20:21 /opt/hadoop/bin/container-executor

This configuration does not support Yarn Secure Containers, but all other hadoop services, including YARN job execution outside secure containers continue to work.