IMA protects files by storing a hash (or a signature thereof) of their content in the security.ima xattr. While the security.ima xattr itself is protected by EVM with either a HMAC or a digital signature, no mechanism is currently in place to ensure that the security.ima xattr was generated with a strong digest algorithm, and any hash defined in the kernel will be accepted, even obsolete formats like MD4 and MD5. The kernel itself will only write to this xattr with the 'ima_hash' parameter, fixed at init, but it will also happily accept userland writes for said xattr, and those writes may use arbitrary hash algorithms as long as the kernel have support for it. One important point is safeguarding users from mislabelling their files when using userland utilities to update their files, as this is the kind of behavior one can observe with evmctl (`evmctl ima_hash` defaults to sha1). Another group that may be interested is those that have deployed IMA years ago, possibly using algorithms that was then deemed sufficiently collision-resistant, but that proved to be weak with the passage of time (note that this could also happen in the future with algorithms considered safe today). This patch provides a migration path of sorts for these users. This patch series gives users the ability to restrict the algorithms accepted by their system, both when writing/updating xattrs, and when appraising files, while retaining a permissive behavior by default to preserve backward compatibility. To provide these features, alter the behavior of setxattr to only accept hashes built in the kernel, instead of any hash listed in the kernel (complete list crypto/hash_info.c). In addition, the user can define in his IMA policy the list of digest algorithms allowed for writing to the security.ima xattr. In that case, only algorithms present in that list are accepted for writing. In addition, users may opt-in to allowlist hash algorithms for ppraising thanks to the new 'appraise_algos' IMA policy option. By default IMA will keep accepting any hash algorithm, but specifying that option will make appraisal of files hashed with another algorithm fail. Even when using this option to restrict accepted hashes, a migration to a new algorithm is still possible. Suppose your policy states you must migrate from 'old_algo' (e.g. sha1) to 'new_algo' (e.g. one of sha256/384/512). You can upgrade without relaxing the hash requirements: alter your policy rules from 'appraise_algos=old_algo' to 'appraise_algos=old_algo,new_algo', load a new SETXATTR_CHECK policy rule that accept writes using 'new_algo', reboot, relabel all your files with 'new_algo', alter your policy rules from 'appraise_algos=old_algo,new_algo' to 'appraise_algos=new_algo', and you're done. While this represent a significant amount of work, it is important to showcase that this patchset is flexible enough to let users upgrade if needed. Here is also a short description of the new audit messages, but I can send it in a followup mail if that is not the proper place: When writing the xattr with an algorithm not built in the kernel (here the kernel was built with CONFIG_CRYPTO_MD5 unset), e.g. with "evmctl ima_hash -a md5 /usr/bin/strace": audit(1628066120.418:121): pid=1344 uid=0 auid=0 ses=1 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 op=set_data cause=unavailable-hash-algorithm comm="evmctl" name="/usr/bin/strace" dev="dm-0" ino=2632657 res=0 errno=0 With the same command and the policy rule "appraise func=SETXATTR_CHECK appraise_algos=sha256,sha384,sha512", we get: audit(1628066210.141:127): pid=1362 uid=0 auid=0 ses=1 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 op=set_data cause=denied-hash-algorithm comm="evmctl" name="/usr/bin/strace" dev="dm-0" ino=2632657 res=0 errno=0 Note that the cause is now 'denied-hash-algorithm' instead of 'unavailable-hash-algorithm'. We get that audit message for any algorithm outside of sha256/384/512 (including algorithms not compiled in the kernel like MD5). In a sense, 'denied-hash-algorithm' takes predecence over 'unavailable-hash-algorithm'. When appraising files, e.g. trying to execute a file whose xattr was hashed with sha1 while the policy rule "appraise func=BPRM_CHECK fowner=0 appraise_algos=sha256" is enabled: audit(1628066349.230:130): pid=1369 uid=0 auid=0 ses=1 subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 op=collect_data cause=denied-hash-algorithm comm="bash" name="/usr/bin/strace" dev="dm-0" ino=2632657 res=0 errno=0 This series is based on the following repo/branch: repo: https://git.kernel.org/pub/scm/linux/kernel/git/zohar/linux-integrity.git branch: next-integrity-testing commit e37be5343ae2b9419aea1442b07e5d2428b437b4 ("Merge branch 'ima-buffer-measurement-changes-v4' into next-integrity") Changelog since v7: - Add a new (tiny) patch on top to restrict the hash algorithms in 'appraise_algos' to algorithms enabled in the kernel configuration - Fix the Signed-off-by header because my git configuration didn't match how my corporate email provider was rewriting the 'FROM' header Changelog since v6: - removed the IS_ENABLED(CONFIG_CRYPTO_MD5) as IMA fails gracefully and already defaults to sha256 in that case (with a more helpful error message than "invalid hash algorithm for template") - Update the commit message of patch 2/5 to fix a typo (reported by Mimi Zohar and Lakshmi Ramasubramanian) - Update the kernel-doc (I decided to keep it as it is acceptable to document static fucntions according to 'Documentation/doc-guide/kernel-doc.rst') (reported by Mimi Zohar) - Prevent file re-validation when the hash algorithm is invalid on setxattr() (reported by Mimi Zohar) - Update the operation in the audit log from 'collect_data' to 'set_data' when performing setxatttr()s. (reported by Mimi Zohar) - Rename the user options from 'appraise_hash' to 'appraise_algos' (reported by Mimi Zohar) - Reject the policy update when the 'appraise_algos' parameter cannot be parsed (reported by Lakshmi Ramasubramanian) - Kept the (single) whitespace->tab change in 'Documentation/ABI/testing/ima_policy' for consistency Changelog since v5: - Adapt a few places where lines were longer than 80 characters (suggested by Mimi Zohar) - Rebase on top on next-integrity-testing (suggested by Mimi Zohar) - Replace "whitelist" with "allowlist" (suggested by both Mimi Zohar and Lakshmi Ramasubramanian) - Fix a broken kernel-doc (suggested by Mimi Zohar) - Prune the feature that updated "func=SETXATTR_CHECK" (suggested by Mimi Zohar) - Update the commit messages with suggestions from Lakshmi Ramasubramanian - Short-circuit a bit the evaluation in ima_appraise.c/validate_hash_algo to return earlier in case of memory exhaustion (suggested by Lakshmi Ramasubramanian) - Renamed "forbidden-hash-algorithm" in ima_main.c to reduce the number of Changelog since v4: - Deleting the ability to remove SETXATTR_CHECK rules, as it added a lot of concurrency troubles while creating a special case for SETXATTR_CHECK rules only, which is rarely a good idea (suggested by Mimi Zohar) - Change from #ifdef CONFIG_... to IS_ENABLED (suggested by Mimi Zohar) - Various fixes (code style, english grammar errors, double initialization to zero) reported by Mimi Zohar - Fixed a logic inversion error introduced in v4 where checks where performed when no SETXATTR_CHECK rule was enabled. - Do not log partial audit messages under memory pressure (suggested by Mimi Zohar) Changelog since v3: - fixed an issue where the first write to the policy would ignore the SETXATTR_CHECK attribute - fixed potential concurrency issues (I would greatly like external opinions on this, because I clearly don't know much about RCU. Beside maybe it's better to completely ignore the duplicates SETXATTR_CHECK issue and not update the IMA policy in any case) - remove the CONFIG_CRYPTO_MD5 requirement for IMA (suggested by Mimi Zohar) - updated commit messages to follow more closely the kernel style guide (suggested by Mimi Zohar) - moved the hash verification code on appraisal a bit later, to prevent issues when using the code with IMA in a disable/auditing mode (suggested by Mimi Zohar) - limit the 'appraise_hash' parameter to the 'appraise' action (suggested by Mimi Zohar) Changelog since v2: - remove the SecureBoot-specific behavior (suggested by Mimi Zohar) - users can now tweak through policy both the algorithms for appraising files (a feature already present in v2) and for writing with the new SETXATTR_CHECK value for the 'func' ima policy flag - updating 'forbidden-hash-algorithm' to 'denied-hash-algorithm' and 'unsupported-hash-algorithm' to disambiguate cases when the user asked for an algorithm not present in the kernel and when the system vendor explicitly opted in to a restricted list of accepted algorithms (suggested by Mimi Zohar) - change the order of the patches to be bisect-safe while retaining the guarantee that a policy cannot be accepted but not enforced (suggested by Mimi Zohar) Changelog since v1: - Remove the two boot parameters (suggested by Mimi Zohar) - filter out hash algorithms not compiled in the kernel on xattr writes (suggested by Mimi Zohar) - add a special case when secure boot is enabled: only the ima_hash algorithm is accepted on userland writes - add a policy option to opt-in to restricting digest algorithms at a per-rule granularity (suggested by Mimi Zohar) THOBY Simon (6): IMA: remove the dependency on CRYPTO_MD5 IMA: block writes of the security.ima xattr with unsupported algorithms IMA: add support to restrict the hash algorithms used for file appraisal IMA: add a policy option to restrict xattr hash algorithms on appraisal IMA: introduce a new policy option func=SETXATTR_CHECK IMA: prevent SETXATTR_CHECK policy rules with unavailable algorithms Documentation/ABI/testing/ima_policy | 15 ++- security/integrity/ima/Kconfig | 1 - security/integrity/ima/ima.h | 14 ++- security/integrity/ima/ima_api.c | 6 +- security/integrity/ima/ima_appraise.c | 73 ++++++++++- security/integrity/ima/ima_main.c | 20 ++- security/integrity/ima/ima_policy.c | 174 +++++++++++++++++++++++--- 7 files changed, 268 insertions(+), 35 deletions(-) -- 2.31.1
