The rollback process is rollable as well. What you do when applying PSU patch is update binaries on each cluster node one by one, the shared components like OCR are not changing. That’s why rollable operation is possible. When all nodes are patched, you apply sql script only once, to update data dictionary. The word of caution here, the apply script may interfere with your application, so always test on test environment before production. You may want to check PSUSUPPLEMENTAL README note 1494646.1 for more details how to perform manual apply/rollabck. In this note you can find that rollback auto is bound to local node.
We never use opatch auto, because we like to be in control of what we are doing. For manual rollback, make sure you specify -local to do operation only on local node.
opatch rollback -local -id <GI_components_number> -oh <GI_HOME>
The best backout strategy for PSU patching is to use minimal downtime using cloning methodology. When rollback is necessary its just a matter of switching back to old home. Check 1136544.1 for more info.
Again, always test your strategy on test environment before production.