v2: drop the user namespace uuid and instead add an IMA specific label (residing in an opaque IMA structure pointed to by the user_ns) that is settable once before use, but if not set reverts to a randomly generated uuid. Over the past five years there have been several attempts to namespace IMA [1,2,3]. All of them were eventually fairly huge patch series, which try to solve every issue and potential issue all at once, making them incredibly difficult to review and highly dependent on an array of non-IMA features which causes huge upporting difficulty as the patch progresses. Given this, I thought we'd try a different way: introduce a minimal namespacing of IMA and try to build on it in subsequent patches. This first patch set namespaces IMA by tying it to the user namespace. We're still discussing whether this is a good idea, so I'll pass on the justification and note that the only addition is a ima_ns_info pointer which points to a structure that has a lifetime longer than the namespace, so the whole machinery for managing this could be transferred to a different namespace. Within this pointer is a label for the IMA namespace, which has a set API (but no exposure in the current patch se) and if the label isn't set before the namespace causes an IMA log entry, a uuid is placed into the label. All this patch set does is add a new template 'ima-ns' which includes the namespace label (added by the first patch) in the IMA log. Using uuids gives us probabalistically unique identifiers for all namespaces without having to check them for uniqueness. Once we have the container being logged, it stands to reason that the ima inode cache needs to record one event per namespace per inode instead of the one global event per inode, so if I enter the ima namespace and execute the same measured command, it will log again with the new namespace uuid even if the hash is the same: > ls > grep ls /sys/kernel/security/integrity/ima/ascii_runtime_measurements 10 c70c7b851f83c8c71ee7b508c8468383c0d2c154 ima-ns sha256:1f7f27ef1052e33731c9ff56a36ac3af4437e3f95ad55f6813c320b087b5d356 /usr/bin/ls 6582e360-1354-42b9-a6ef-ee1993d982da > unshare --user -r # ls # exit > grep ls /sys/kernel/security/integrity/ima/ascii_runtime_measurements 10 c70c7b851f83c8c71ee7b508c8468383c0d2c154 ima-ns sha256:1f7f27ef1052e33731c9ff56a36ac3af4437e3f95ad55f6813c320b087b5d356 /usr/bin/ls 6582e360-1354-42b9-a6ef-ee1993d982da 10 144a73d85e9cf999c4abbc99f3c41e9422c8016e ima-ns sha256:1f7f27ef1052e33731c9ff56a36ac3af4437e3f95ad55f6813c320b087b5d356 /usr/bin/ls e496e384-4133-4d57-b93a-1812b83badf2 Note that this namespacing if the iint cache is in the third patch and could be dropped if there's huge opposition. Some things to note are that the IMA securityfs entries aren't virtualized. This is planned for a follow up patch (so currently the admin can't even view the ima log in the container). Everything that's logged goes through the main IMA log and the physical TPM. This means that the admin of the physical system can attest to the log, but the containers would have to trust the admins attestation of their log pieces. The initial IMA policy is also inherited from the physical system and can't currently be changed. The rough plan of action for follow up patches is 1. Namespace securityfs so container admin can read the IMA files like log which would only show entries related to the container (so only entries generated by the current and all child namespaces) and policy. 2. Add per namespace policies by writing to the policy file in the container. Obviously implementation of this would have to preserve the security of the system, so the new namespace couldn't stop logging something the physical host required to be logged, but it could add additional classes of things to log. 3. Add the ima keyrings and the ability to appraise inside the container. There could be other phases beyond this, including possibly optionally attaching a vtpm to the container to provide local quotes but this should be need driven. Some non problems of this approach are: * The continuous growth of the IMA log. This is already a problem with non-namespaced IMA. One can argue that the above implementation makes the problem worse, but it is unarguable that if the problem were solved generally it would no logner be an issue for containers. * attesting to the in-container IMA log. Given it's being logged through the physical TPM, the physical system owner will have to publish a mechanism for attesting to particular container entries of the log. [1] https://lore.kernel.org/all/20200818152037.11869-1-krzysztof.struczynski@xxxxxxxxxx [2] https://lore.kernel.org/all/20180511144230.75384-1-stefanb@xxxxxxxxxxxxxxxxxx [3] https://lore.kernel.org/all/1494511203-8397-1-git-send-email-guilherme.magalhaes@xxxxxxx James --- James Bottomley (3): userns: add ima_ns_info field containing a settable namespace label ima: show the namespace label in the ima-ns template ima: make the integrity inode cache per namespace include/linux/ima.h | 15 +- include/linux/user_namespace.h | 7 + kernel/user.c | 1 + kernel/user_namespace.c | 6 + security/integrity/iint.c | 4 +- security/integrity/ima/Kconfig | 6 +- security/integrity/ima/Makefile | 2 +- security/integrity/ima/ima.h | 26 +++- security/integrity/ima/ima_api.c | 7 +- security/integrity/ima/ima_main.c | 21 +-- security/integrity/ima/ima_ns.c | 169 ++++++++++++++++++++++ security/integrity/ima/ima_policy.c | 2 +- security/integrity/ima/ima_template.c | 6 +- security/integrity/ima/ima_template_lib.c | 17 +++ security/integrity/ima/ima_template_lib.h | 4 + security/integrity/integrity.h | 11 +- 16 files changed, 284 insertions(+), 20 deletions(-) create mode 100644 security/integrity/ima/ima_ns.c -- 2.33.0
