This patch set depends on: - https://lore.kernel.org/linux-integrity/20210723085304.1760138-1-roberto.sassu@xxxxxxxxxx/ - https://lore.kernel.org/linux-integrity/20210705115650.3373599-1-roberto.sassu@xxxxxxxxxx/ I still kept pointer math to optimize the size of the digest_list_item_ref structure. Replacing offsets with pointers would cause the size of the structure to double. I could do this in the next version of the patch set if the size change is acceptable. Digest Lists Integrity Module (DIGLIM) is a new component added to the integrity subsystem in the kernel, primarily aiming to aid Integrity Measurement Architecture (IMA) in the process of checking the integrity of file content and metadata. It accomplishes this task by storing reference values coming from software vendors and by reporting whether or not the digest of file content or metadata calculated by IMA (or EVM) is found among those values. In this way, IMA can decide, depending on the result of a query, if a measurement should be taken or access to the file should be granted. The Security Assumptions section explains more in detail why this component has been placed in the kernel. The main benefits of using IMA in conjunction with DIGLIM are the ability to implement advanced remote attestation schemes based on the usage of a TPM key for establishing a TLS secure channel [1][2], and to reduce the burden on Linux distribution vendors to extend secure boot at OS level to applications. DIGLIM does not have the complexity of feature-rich databases. In fact, its main functionality comes from the hash table primitives already in the kernel. It does not have an ad-hoc storage module, it just indexes data in a fixed format (digest lists, a set of concatenated digests preceded by a header), copied to kernel memory as they are. Lastly, it does not support database-oriented languages such as SQL, but only accepts a digest and its algorithm as a query. The only digest list format supported by DIGLIM is called compact. However, Linux distribution vendors don't have to generate new digest lists in this format for the packages they release, as already available information, such as RPM headers and DEB package metadata, can be already used as a source for reference values (they already include file digests), with a user space parser taking care of the conversion to the compact format. Although one might perceive that storing file or metadata digests for a Linux distribution would significantly increase the memory usage, this does not seem to be the case. As an anticipation of the evaluation done in the Preliminary Performance Evaluation section, protecting binaries and shared libraries of a minimal Fedora 33 installation requires 208K of memory for the digest lists plus 556K for indexing. In exchange for a slightly increased memory usage, DIGLIM improves the performance of the integrity subsystem. In the considered scenario, IMA measurement and appraisal with digest lists requires respectively less than one quarter and less than half the time, compared to the current solution. DIGLIM also keeps track of whether digest lists have been processed in some way (e.g. measured or appraised by IMA). This is important for example for remote attestation, so that remote verifiers understand what has been uploaded to the kernel. DIGLIM behaves like a transactional database, i.e. it has the ability to roll back to the beginning of the transaction if an error occurred during the addition of a digest list (the deletion operation always succeeds). This capability has been tested with an ad-hoc fault injection mechanism capable of simulating failures during the operations. Finally, DIGLIM exposes to user space, through securityfs, the digest lists currently loaded, the number of digests added, a query interface and an interface to set digest list labels. [1] LSS EU 2019 - slides: https://static.sched.com/hosted_files/lsseu2019/bd/secure_attested_communication_channels_lss_eu_2019.pdf - video: https://youtu.be/mffdQgkvDNY [2] FutureTPM EU project, final review meeting demo - slides: https://futuretpm.eu/images/07-3-FutureTPM-Final-Review-Slides-WP6-Device-Management-Use-Case-HWDU.pdf - video: https://vimeo.com/528251864/4c1d55abcd Binary Integrity Integrity is a fundamental security property in information systems. Integrity could be described as the condition in which a generic component is just after it has been released by the entity that created it. One way to check whether a component is in this condition (called binary integrity) is to calculate its digest and to compare it with a reference value (i.e. the digest calculated in controlled conditions, when the component is released). IMA, a software part of the integrity subsystem, can perform such evaluation and execute different actions: - store the digest in an integrity-protected measurement list, so that it can be sent to a remote verifier for analysis; - compare the calculated digest with a reference value (usually protected with a signature) and deny operations if the file is found corrupted; - store the digest in the system log. Contribution DIGLIM further enhances the capabilities offered by IMA-based solutions and, at the same time, makes them more practical to adopt by reusing existing sources as reference values for integrity decisions. Possible sources for digest lists are: - RPM headers; - Debian repository metadata. Benefits for IMA Measurement One of the issues that arises when files are measured by the OS is that, due to parallel execution, the order in which file accesses happen cannot be predicted. Since the TPM Platform Configuration Register (PCR) extend operation, executed after each file measurement, cryptographically binds the current measurement to the previous ones, the PCR value at the end of a workload cannot be predicted too. Thus, even if the usage of a TPM key, bound to a PCR value, should be allowed when only good files were accessed, the TPM could unexpectedly deny an operation on that key if files accesses did not happen as stated by the key policy (which allows only one of the possible sequences). DIGLIM solves this issue by making the PCR value stable over the time and not dependent on file accesses. The following figure depicts the current and the new approaches: IMA measurement list (current) entry# 1st boot 2nd boot 3rd boot +----+---------------+ +----+---------------+ +----+---------------+ 1: | 10 | file1 measur. | | 10 | file3 measur. | | 10 | file2 measur. | +----+---------------+ +----+---------------+ +----+---------------+ 2: | 10 | file2 measur. | | 10 | file2 measur. | | 10 | file3 measur. | +----+---------------+ +----+---------------+ +----+---------------+ 3: | 10 | file3 measur. | | 10 | file1 measur. | | 10 | file4 measur. | +----+---------------+ +----+---------------+ +----+---------------+ PCR: Extend != Extend != Extend file1, file2, file3 file3, file2, file1 file2, file3, file4 PCR Extend definition: PCR(new value) = Hash(Hash(meas. entry), PCR(previous value)) A new entry in the measurement list is created by IMA for each file access. Assuming that file1, file2 and file3 are files provided by the software vendor, file4 is an unknown file, the first two PCR values above represent a good system state, the third a bad system state. The PCR values are the result of the PCR extend operation performed for each measurement entry with the digest of the measurement entry as an input. IMA measurement list (with DIGLIM) dlist +--------------+ | header | +--------------+ | file1 digest | | file2 digest | | file3 digest | +--------------+ dlist is a digest list containing the digest of file1, file2 and file3. In the intended scenario, it is generated by a software vendor at the end of the building process, and retrieved by the administrator of the system where the digest list is loaded. entry# 1st boot 2nd boot 3rd boot +----+---------------+ +----+---------------+ +----+---------------+ 0: | 11 | dlist measur. | | 11 | dlist measur. | | 11 | dlist measur. | +----+---------------+ +----+---------------+ +----+---------------+ 1: < file1 measur. skip > < file3 measur. skip > < file2 measur. skip > 2: < file2 measur. skip > < file2 measur. skip > < file3 measur. skip > +----+---------------+ 3: < file3 measur. skip > < file1 measur. skip > | 11 | file4 measur. | +----+---------------+ PCR: Extend = Extend != Extend dlist dlist dlist, file4 The first entry in the measurement list contains the digest of the digest list uploaded to the kernel at kernel initialization time. When a file is accessed, IMA queries DIGLIM with the calculated file digest and, if it is found, IMA skips the measurement. Thus, the only information sent to remote verifiers are: the list of files that could possibly be accessed (from the digest list), but not if they were accessed and when; the measurement of unknown files. Despite providing less information, this solution has the advantage that the good system state (i.e. when only file1, file2 and file3 are accessed) now can be represented with a deterministic PCR value (the PCR is extended only with the measurement of the digest list). Also, the bad system state can still be distinguished from the good state (the PCR is extended also with the measurement of file4). If a TPM key is bound to the good PCR value, the TPM would allow the key to be used if file1, file2 or file3 are accessed, regardless of the sequence in which they are accessed (the PCR value does not change), and would revoke the permission when the unknown file4 is accessed (the PCR value changes). If a system is able to establish a TLS connection with a peer, this implicitly means that the system was in a good state (i.e. file4 was not accessed, otherwise the TPM would have denied the usage of the TPM key due to the key policy). Benefits for IMA Appraisal Extending secure boot to applications means being able to verify the provenance of files accessed. IMA does it by verifying file signatures with a key that it trusts, which requires Linux distribution vendors to additionally include in the package header a signature for each file that must be verified (there is the dedicated RPMTAG_FILESIGNATURES section in the RPM header). The proposed approach would be instead to verify data provenance from already available metadata (file digests) in existing packages. IMA would verify the signature of package metadata and search file digests extracted from package metadata and added to the hash table in the kernel. For RPMs, file digests can be found in the RPMTAG_FILEDIGESTS section of RPMTAG_IMMUTABLE, whose signature is in RPMTAG_RSAHEADER. For DEBs, file digests (unsafe to use due to a weak digest algorithm) can be found in the md5sum file, which can be indirectly verified from Release.gpg. The following figure highlights the differences between the current and the proposed approach. IMA appraisal (current solution, with file signatures): appraise +-----------+ V | +-------------------------+-----+ +-------+-----+ | | RPM header | | ima rpm | file1 | sig | | | ... | | plugin +-------+-----+ +-----+ | file1 sig [to be added] | sig |--------> ... | IMA | | ... | | +-------+-----+ +-----+ | fileN sig [to be added] | | | fileN | sig | +-------------------------+-----+ +-------+-----+ In this case, file signatures must be added to the RPM header, so that the ima rpm plugin can extract them together with the file content. The RPM header signature is not used. IMA appraisal (with DIGLIM): kernel hash table with RPM header content +---+ +--------------+ | |--->| file1 digest | +---+ +--------------+ ... +---+ appraise (file1) | | <--------------+ +----------------+-----+ +---+ | | RPM header | | ^ | | ... | | digest_list | | | file1 digest | sig | rpm plugin | +-------+ +-----+ | ... | |-------------+--->| file1 | | IMA | | fileN digest | | +-------+ +-----+ +----------------+-----+ | ^ | +------------------------------------+ appraise (RPM header) In this case, the RPM header is used as it is, and its signature is used for IMA appraisal. Then, the digest_list rpm plugin executes the user space parser to parse the RPM header and add the extracted digests to an hash table in the kernel. IMA appraisal of the files in the RPM package consists in searching their digest in the hash table. Other than reusing available information as digest list, another advantage is the lower computational overhead compared to the solution with file signatures (only one signature verification for many files and digest lookup, instead of per file signature verification, see Preliminary Performance Evaluation for more details). Lifecycle The lifecycle of DIGLIM is represented in the following figure: Vendor premises (release process with modifications): +------------+ +-----------------------+ +------------------------+ | 1. build a | | 2. generate and sign | | 3. publish the package | | package |-->| a digest list from |-->| and digest list in | | | | packaged files | | a repository | +------------+ +-----------------------+ +------------------------+ | | User premises: | V +---------------------+ +------------------------+ +-----------------+ | 6. use digest lists | | 5. download the digest | | 4. download and | | for measurement |<--| list and upload to |<--| install the | | and/or appraisal | | the kernel | | package | +---------------------+ +------------------------+ +-----------------+ The figure above represents all the steps when a digest list is generated separately. However, as mentioned in Contribution, in most cases existing packages can be already used as a source for digest lists, limiting the effort for software vendors. If, for example, RPMs are used as a source for digest lists, the figure above becomes: Vendor premises (release process without modifications): +------------+ +------------------------+ | 1. build a | | 2. publish the package | | package |-->| in a repository |---------------------+ | | | | | +------------+ +------------------------+ | | | User premises: | V +---------------------+ +------------------------+ +-----------------+ | 5. use digest lists | | 4. extract digest list | | 3. download and | | for measurement |<--| from the package |<--| install the | | and/or appraisal | | and upload to the | | package | | | | kernel | | | +---------------------+ +------------------------+ +-----------------+ Step 4 can be performed with the digest_list rpm plugin and the user space parser, without changes to rpm itself. Security Assumptions As mentioned in the Introduction, DIGLIM will be primarily used in conjunction with IMA to enforce a mandatory policy on all user space processes, including those owned by root. Even root, in a system with a locked-down kernel, cannot affect the enforcement of the mandatory policy or, if changes are permitted, it cannot do so without being detected. Given that the target of the enforcement are user space processes, DIGLIM cannot be placed in the target, as a Mandatory Access Control (MAC) design is required to have the components responsible to enforce the mandatory policy separated from the target. While locking-down a system and limiting actions with a mandatory policy is generally perceived by users as an obstacle, it has noteworthy benefits for the users themselves. First, it would timely block attempts by malicious software to steal or misuse user assets. Although users could query the package managers to detect them, detection would happen after the fact, or it wouldn't happen at all if the malicious software tampered with package managers. With a mandatory policy enforced by the kernel, users would still be able to decide which software they want to be executed except that, unlike package managers, the kernel is not affected by user space processes or root. Second, it might make systems more easily verifiable from outside, due to the limited actions the system allows. When users connect to a server, not only they would be able to verify the server identity, which is already possible with communication protocols like TLS, but also if the software running on that server can be trusted to handle their sensitive data. Adoption A former version of DIGLIM is used in the following OSes: - openEuler 20.09 https://github.com/openeuler-mirror/kernel/tree/openEuler-20.09 - openEuler 21.03 https://github.com/openeuler-mirror/kernel/tree/openEuler-21.03 Originally, DIGLIM was part of IMA (known as IMA Digest Lists). In this version, it has been redesigned as a standalone module with an API that makes its functionality accessible by IMA and, eventually, other subsystems. User Space Support Digest lists can be generated and managed with digest-list-tools: https://github.com/openeuler-mirror/digest-list-tools It includes two main applications: - gen_digest_lists: generates digest lists from files in the filesystem or from the RPM database (more digest list sources can be supported); - manage_digest_lists: converts and uploads digest lists to the kernel. Integration with rpm is done with the digest_list plugin: https://gitee.com/src-openeuler/rpm/blob/master/Add-digest-list-plugin.patch This plugin writes the RPM header and its signature to a file, so that the file is ready to be appraised by IMA, and calls the user space parser to convert and upload the digest list to the kernel. Simple Usage Example (Tested with Fedora 33) 1. Digest list generation (RPM headers and their signature are copied to the specified directory): # mkdir /etc/digest_lists # gen_digest_lists -t file -f rpm+db -d /etc/digest_lists -o add 2. Digest list upload with the user space parser: # manage_digest_lists -p add-digest -d /etc/digest_lists 3. First digest list query: # echo sha256-$(sha256sum /bin/cat) > /sys/kernel/security/integrity/diglim/digest_query # cat /sys/kernel/security/integrity/diglim/digest_query sha256-[...]-0-file_list-rpm-coreutils-8.32-18.fc33.x86_64 (actions: 0): version: 1, algo: sha256, type: 2, modifiers: 1, count: 106, datalen: 3392 4. Second digest list query: # echo sha256-$(sha256sum /bin/zip) > /sys/kernel/security/integrity/diglim/digest_query # cat /sys/kernel/security/integrity/diglim/digest_query sha256-[...]-0-file_list-rpm-zip-3.0-27.fc33.x86_64 (actions: 0): version: 1, algo: sha256, type: 2, modifiers: 1, count: 4, datalen: 128 Preliminary Performance Evaluation This section provides an initial estimation of the overhead introduced by DIGLIM. The estimation has been performed on a Fedora 33 virtual machine with 1447 packages installed. The virtual machine has 16 vCPU (host CPU: AMD Ryzen Threadripper PRO 3955WX 16-Cores) and 2G of RAM (host memory: 64G). The virtual machine also has a vTPM with libtpms and swtpm as backend. After writing the RPM headers to files, the size of the directory containing them is 36M. After converting the RPM headers to the compact digest list, the size of the data being uploaded to the kernel is 3.6M. The time to load the entire RPM database is 0.628s. After loading the digest lists to the kernel, the slab usage due to indexing is (obtained with slab_nomerge in the kernel command line): OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME 118144 118144 100% 0,03K 923 128 3692K digest_list_item_ref_cache 102400 102400 100% 0,03K 800 128 3200K digest_item_cache 2646 2646 100% 0,09K 63 42 252K digest_list_item_cache The stats, obtained from the digests_count interface, introduced later, are: Parser digests: 0 File digests: 99100 Metadata digests: 0 Digest list digests: 1423 On this installation, this would be the worst case in which all files are measured and/or appraised, which is currently not recommended without enforcing an integrity policy protecting mutable files. Infoflow LSM is a component to accomplish this task: https://patchwork.kernel.org/project/linux-integrity/cover/20190818235745.1417-1-roberto.sassu@xxxxxxxxxx/ The first manageable goal of IMA with DIGLIM is to use an execution policy, with measurement and/or appraisal of files executed or mapped in memory as executable (in addition to kernel modules and firmware). In this case, the digest list contains the digest only for those files. The numbers above change as follows. After converting the RPM headers to the compact digest list, the size of the data being uploaded to the kernel is 208K. The time to load the digest of binaries and shared libraries is 0.062s. After loading the digest lists to the kernel, the slab usage due to indexing is: OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME 7168 7168 100% 0,03K 56 128 224K digest_list_item_ref_cache 7168 7168 100% 0,03K 56 128 224K digest_item_cache 1134 1134 100% 0,09K 27 42 108K digest_list_item_cache The stats, obtained from the digests_count interface, are: Parser digests: 0 File digests: 5986 Metadata digests: 0 Digest list digests: 1104 Comparison with IMA This section compares the performance between the current solution for IMA measurement and appraisal, and IMA with DIGLIM. Workload A (without DIGLIM): 1. cat file[0-5985] > /dev/null Workload B (with DIGLIM): 1. echo $PWD/0-file_list-compact-file[0-1103] > <securityfs>/integrity/diglim/digest_list_add 2. cat file[0-5985] > /dev/null Workload A execution time without IMA policy: real 0m0,155s user 0m0,008s sys 0m0,066s Measurement IMA policy: measure fowner=2000 func=FILE_CHECK mask=MAY_READ use_diglim=allow pcr=11 ima_template=ima-sig use_diglim is a policy keyword not yet supported by IMA. Workload A execution time with IMA and 5986 files with signature measured: real 0m8,273s user 0m0,008s sys 0m2,537s Workload B execution time with IMA, 1104 digest lists with signature measured and uploaded to the kernel, and 5986 files with signature accessed but not measured (due to the file digest being found in the hash table): real 0m1,837s user 0m0,036s sys 0m0,583s Appraisal IMA policy: appraise fowner=2000 func=FILE_CHECK mask=MAY_READ use_diglim=allow use_diglim is a policy keyword not yet supported by IMA. Workload A execution time with IMA and 5986 files with file signature appraised: real 0m2,197s user 0m0,011s sys 0m2,022s Workload B execution time with IMA, 1104 digest lists with signature appraised and uploaded to the kernel, and with 5986 files with signature not verified (due to the file digest being found in the hash table): real 0m0,982s user 0m0,020s sys 0m0,865s Changelog v1: - remove 'ima: Add digest, algo, measured parameters to ima_measure_critical_data()', replaced by: https://lore.kernel.org/linux-integrity/20210705090922.3321178-1-roberto.sassu@xxxxxxxxxx/ - add 'Lifecycle' subsection to better clarify how digest lists are generated and used (suggested by Greg KH) - remove 'Possible Usages' subsection and add 'Benefits for IMA Measurement' and 'Benefits for IMA Appraisal' subsubsections - add 'Preliminary Performance Evaluation' subsection - declare digest_offset and hdr_offset in the digest_list_item_ref structure as u32 (sufficient for digest lists of 4G) to make room for a list_head structure (digest_list_item_ref size: 32) - implement digest list reference management with a linked list instead of an array - reorder structure members for better alignment (suggested by Mauro) - rename digest_lookup() to __digest_lookup() (suggested by Mauro) - introduce an object cache for each defined structure - replace atomic_long_t with unsigned long in h_table structure definition (suggested by Greg KH) - remove GPL2 license text and file names (suggested by Greg KH) - ensure that the _reserved field of compact_list_hdr is equal to zero (suggested by Greg KH) - dynamically allocate the buffer in digest_lists_show_htable_len() to avoid frame size warning (reported by kernel test robot, dynamic allocation suggested by Mauro) - split documentation in multiple files and reference the source code (suggested by Mauro) - use #ifdef in include/linux/diglim.h - improve generation of event name for IMA measurements - add new patch to introduce the 'Remote Attestation' section in the documentation - fix assignment of actions variable in digest_list_read() and digest_list_write() - always release dentry reference when digest_list_get_secfs_files() is called - rewrite add/del and query interfaces to take advantage of m->private - prevent deletion of a digest list only if there are actions done at addition time that are not currently being performed - fix doc warnings (replace Returns with Return:) - perform queries of digest list digests in the existing tests - add new tests: digest_list_add_del_test_file_upload_measured, digest_list_check_measurement_list_test_file_upload and digest_list_check_measurement_list_test_buffer_upload - don't return a value from digest_del(), digest_list_ref_del, and digest_list_del() - improve Makefile for tests Roberto Sassu (12): diglim: Overview diglim: Basic definitions diglim: Objects diglim: Methods diglim: Parser diglim: Interfaces - digest_list_add, digest_list_del diglim: Interfaces - digest_lists_loaded diglim: Interfaces - digest_label diglim: Interfaces - digest_query diglim: Interfaces - digests_count diglim: Remote Attestation diglim: Tests .../security/diglim/architecture.rst | 45 + .../security/diglim/implementation.rst | 255 +++ Documentation/security/diglim/index.rst | 14 + .../security/diglim/introduction.rst | 631 ++++++++ .../security/diglim/remote_attestation.rst | 87 ++ Documentation/security/diglim/tests.rst | 66 + Documentation/security/index.rst | 1 + MAINTAINERS | 19 + include/linux/diglim.h | 28 + include/linux/kernel_read_file.h | 1 + include/uapi/linux/diglim.h | 51 + security/integrity/Kconfig | 1 + security/integrity/Makefile | 1 + security/integrity/diglim/Kconfig | 11 + security/integrity/diglim/Makefile | 8 + security/integrity/diglim/diglim.h | 157 ++ security/integrity/diglim/fs.c | 782 ++++++++++ security/integrity/diglim/methods.c | 499 ++++++ security/integrity/diglim/parser.c | 274 ++++ security/integrity/integrity.h | 4 + tools/testing/selftests/Makefile | 1 + tools/testing/selftests/diglim/Makefile | 19 + tools/testing/selftests/diglim/common.c | 115 ++ tools/testing/selftests/diglim/common.h | 31 + tools/testing/selftests/diglim/config | 3 + tools/testing/selftests/diglim/selftest.c | 1382 +++++++++++++++++ 26 files changed, 4486 insertions(+) create mode 100644 Documentation/security/diglim/architecture.rst create mode 100644 Documentation/security/diglim/implementation.rst create mode 100644 Documentation/security/diglim/index.rst create mode 100644 Documentation/security/diglim/introduction.rst create mode 100644 Documentation/security/diglim/remote_attestation.rst create mode 100644 Documentation/security/diglim/tests.rst create mode 100644 include/linux/diglim.h create mode 100644 include/uapi/linux/diglim.h create mode 100644 security/integrity/diglim/Kconfig create mode 100644 security/integrity/diglim/Makefile create mode 100644 security/integrity/diglim/diglim.h create mode 100644 security/integrity/diglim/fs.c create mode 100644 security/integrity/diglim/methods.c create mode 100644 security/integrity/diglim/parser.c create mode 100644 tools/testing/selftests/diglim/Makefile create mode 100644 tools/testing/selftests/diglim/common.c create mode 100644 tools/testing/selftests/diglim/common.h create mode 100644 tools/testing/selftests/diglim/config create mode 100644 tools/testing/selftests/diglim/selftest.c -- 2.25.1
