xref: /linux/Documentation/security/ipe.rst (revision ac6731870ed943c7c6a8d4114b3ccaddfbdf7d58)
1.. SPDX-License-Identifier: GPL-2.0
2
3Integrity Policy Enforcement (IPE) - Kernel Documentation
4=========================================================
5
6.. NOTE::
7
8   This is documentation targeted at developers, instead of administrators.
9   If you're looking for documentation on the usage of IPE, please see
10   :doc:`IPE admin guide </admin-guide/LSM/ipe>`.
11
12Historical Motivation
13---------------------
14
15The original issue that prompted IPE's implementation was the creation
16of a locked-down system. This system would be born-secure, and have
17strong integrity guarantees over both the executable code, and specific
18*data files* on the system, that were critical to its function. These
19specific data files would not be readable unless they passed integrity
20policy. A mandatory access control system would be present, and
21as a result, xattrs would have to be protected. This lead to a selection
22of what would provide the integrity claims. At the time, there were two
23main mechanisms considered that could guarantee integrity for the system
24with these requirements:
25
26  1. IMA + EVM Signatures
27  2. DM-Verity
28
29Both options were carefully considered, however the choice to use DM-Verity
30over IMA+EVM as the *integrity mechanism* in the original use case of IPE
31was due to three main reasons:
32
33  1. Protection of additional attack vectors:
34
35    * With IMA+EVM, without an encryption solution, the system is vulnerable
36      to offline attack against the aforementioned specific data files.
37
38      Unlike executables, read operations (like those on the protected data
39      files), cannot be enforced to be globally integrity verified. This means
40      there must be some form of selector to determine whether a read should
41      enforce the integrity policy, or it should not.
42
43      At the time, this was done with mandatory access control labels. An IMA
44      policy would indicate what labels required integrity verification, which
45      presented an issue: EVM would protect the label, but if an attacker could
46      modify filesystem offline, the attacker could wipe all the xattrs -
47      including the SELinux labels that would be used to determine whether the
48      file should be subject to integrity policy.
49
50      With DM-Verity, as the xattrs are saved as part of the Merkel tree, if
51      offline mount occurs against the filesystem protected by dm-verity, the
52      checksum no longer matches and the file fails to be read.
53
54    * As userspace binaries are paged in Linux, dm-verity also offers the
55      additional protection against a hostile block device. In such an attack,
56      the block device reports the appropriate content for the IMA hash
57      initially, passing the required integrity check. Then, on the page fault
58      that accesses the real data, will report the attacker's payload. Since
59      dm-verity will check the data when the page fault occurs (and the disk
60      access), this attack is mitigated.
61
62  2. Performance:
63
64    * dm-verity provides integrity verification on demand as blocks are
65      read versus requiring the entire file being read into memory for
66      validation.
67
68  3. Simplicity of signing:
69
70    * No need for two signatures (IMA, then EVM): one signature covers
71      an entire block device.
72    * Signatures can be stored externally to the filesystem metadata.
73    * The signature supports an x.509-based signing infrastructure.
74
75The next step was to choose a *policy* to enforce the integrity mechanism.
76The minimum requirements for the policy were:
77
78  1. The policy itself must be integrity verified (preventing trivial
79     attack against it).
80  2. The policy itself must be resistant to rollback attacks.
81  3. The policy enforcement must have a permissive-like mode.
82  4. The policy must be able to be updated, in its entirety, without
83     a reboot.
84  5. Policy updates must be atomic.
85  6. The policy must support *revocations* of previously authored
86     components.
87  7. The policy must be auditable, at any point-of-time.
88
89IMA, as the only integrity policy mechanism at the time, was
90considered against these list of requirements, and did not fulfill
91all of the minimum requirements. Extending IMA to cover these
92requirements was considered, but ultimately discarded for a
93two reasons:
94
95  1. Regression risk; many of these changes would result in
96     dramatic code changes to IMA, which is already present in the
97     kernel, and therefore might impact users.
98
99  2. IMA was used in the system for measurement and attestation;
100     separation of measurement policy from local integrity policy
101     enforcement was considered favorable.
102
103Due to these reasons, it was decided that a new LSM should be created,
104whose responsibility would be only the local integrity policy enforcement.
105
106Role and Scope
107--------------
108
109IPE, as its name implies, is fundamentally an integrity policy enforcement
110solution; IPE does not mandate how integrity is provided, but instead
111leaves that decision to the system administrator to set the security bar,
112via the mechanisms that they select that suit their individual needs.
113There are several different integrity solutions that provide a different
114level of security guarantees; and IPE allows sysadmins to express policy for
115theoretically all of them.
116
117IPE does not have an inherent mechanism to ensure integrity on its own.
118Instead, there are more effective layers available for building systems that
119can guarantee integrity. It's important to note that the mechanism for proving
120integrity is independent of the policy for enforcing that integrity claim.
121
122Therefore, IPE was designed around:
123
124  1. Easy integrations with integrity providers.
125  2. Ease of use for platform administrators/sysadmins.
126
127Design Rationale:
128-----------------
129
130IPE was designed after evaluating existing integrity policy solutions
131in other operating systems and environments. In this survey of other
132implementations, there were a few pitfalls identified:
133
134  1. Policies were not readable by humans, usually requiring a binary
135     intermediary format.
136  2. A single, non-customizable action was implicitly taken as a default.
137  3. Debugging the policy required manual steps to determine what rule was violated.
138  4. Authoring a policy required an in-depth knowledge of the larger system,
139     or operating system.
140
141IPE attempts to avoid all of these pitfalls.
142
143Policy
144~~~~~~
145
146Plain Text
147^^^^^^^^^^
148
149IPE's policy is plain-text. This introduces slightly larger policy files than
150other LSMs, but solves two major problems that occurs with some integrity policy
151solutions on other platforms.
152
153The first issue is one of code maintenance and duplication. To author policies,
154the policy has to be some form of string representation (be it structured,
155through XML, JSON, YAML, etcetera), to allow the policy author to understand
156what is being written. In a hypothetical binary policy design, a serializer
157is necessary to write the policy from the human readable form, to the binary
158form, and a deserializer is needed to interpret the binary form into a data
159structure in the kernel.
160
161Eventually, another deserializer will be needed to transform the binary from
162back into the human-readable form with as much information preserved. This is because a
163user of this access control system will have to keep a lookup table of a checksum
164and the original file itself to try to understand what policies have been deployed
165on this system and what policies have not. For a single user, this may be alright,
166as old policies can be discarded almost immediately after the update takes hold.
167For users that manage computer fleets in the thousands, if not hundreds of thousands,
168with multiple different operating systems, and multiple different operational needs,
169this quickly becomes an issue, as stale policies from years ago may be present,
170quickly resulting in the need to recover the policy or fund extensive infrastructure
171to track what each policy contains.
172
173With now three separate serializer/deserializers, maintenance becomes costly. If the
174policy avoids the binary format, there is only one required serializer: from the
175human-readable form to the data structure in kernel, saving on code maintenance,
176and retaining operability.
177
178The second issue with a binary format is one of transparency. As IPE controls
179access based on the trust of the system's resources, it's policy must also be
180trusted to be changed. This is done through signatures, resulting in needing
181signing as a process. Signing, as a process, is typically done with a
182high security bar, as anything signed can be used to attack integrity
183enforcement systems. It is also important that, when signing something, that
184the signer is aware of what they are signing. A binary policy can cause
185obfuscation of that fact; what signers see is an opaque binary blob. A
186plain-text policy, on the other hand, the signers see the actual policy
187submitted for signing.
188
189Boot Policy
190~~~~~~~~~~~
191
192IPE, if configured appropriately, is able to enforce a policy as soon as a
193kernel is booted and usermode starts. That implies some level of storage
194of the policy to apply the minute usermode starts. Generally, that storage
195can be handled in one of three ways:
196
197  1. The policy file(s) live on disk and the kernel loads the policy prior
198     to an code path that would result in an enforcement decision.
199  2. The policy file(s) are passed by the bootloader to the kernel, who
200     parses the policy.
201  3. There is a policy file that is compiled into the kernel that is
202     parsed and enforced on initialization.
203
204The first option has problems: the kernel reading files from userspace
205is typically discouraged and very uncommon in the kernel.
206
207The second option also has problems: Linux supports a variety of bootloaders
208across its entire ecosystem - every bootloader would have to support this
209new methodology or there must be an independent source. It would likely
210result in more drastic changes to the kernel startup than necessary.
211
212The third option is the best but it's important to be aware that the policy
213will take disk space against the kernel it's compiled in. It's important to
214keep this policy generalized enough that userspace can load a new, more
215complicated policy, but restrictive enough that it will not overauthorize
216and cause security issues.
217
218The initramfs provides a way that this bootup path can be established. The
219kernel starts with a minimal policy, that trusts the initramfs only. Inside
220the initramfs, when the real rootfs is mounted, but not yet transferred to,
221it deploys and activates a policy that trusts the new root filesystem.
222This prevents overauthorization at any step, and keeps the kernel policy
223to a minimal size.
224
225Startup
226^^^^^^^
227
228Not every system, however starts with an initramfs, so the startup policy
229compiled into the kernel will need some flexibility to express how trust
230is established for the next phase of the bootup. To this end, if we just
231make the compiled-in policy a full IPE policy, it allows system builders
232to express the first stage bootup requirements appropriately.
233
234Updatable, Rebootless Policy
235~~~~~~~~~~~~~~~~~~~~~~~~~~~~
236
237As requirements change over time (vulnerabilities are found in previously
238trusted applications, keys roll, etcetera). Updating a kernel to change the
239meet those security goals is not always a suitable option, as updates are not
240always risk-free, and blocking a security update leaves systems vulnerable.
241This means IPE requires a policy that can be completely updated (allowing
242revocations of existing policy) from a source external to the kernel (allowing
243policies to be updated without updating the kernel).
244
245Additionally, since the kernel is stateless between invocations, and reading
246policy files off the disk from kernel space is a bad idea(tm), then the
247policy updates have to be done rebootlessly.
248
249To allow an update from an external source, it could be potentially malicious,
250so this policy needs to have a way to be identified as trusted. This is
251done via a signature chained to a trust source in the kernel. Arbitrarily,
252this is  the ``SYSTEM_TRUSTED_KEYRING``, a keyring that is initially
253populated at kernel compile-time, as this matches the expectation that the
254author of the compiled-in policy described above is the same entity that can
255deploy policy updates.
256
257Anti-Rollback / Anti-Replay
258~~~~~~~~~~~~~~~~~~~~~~~~~~~
259
260Over time, vulnerabilities are found and trusted resources may not be
261trusted anymore. IPE's policy has no exception to this. There can be
262instances where a mistaken policy author deploys an insecure policy,
263before correcting it with a secure policy.
264
265Assuming that as soon as the insecure policy is signed, and an attacker
266acquires the insecure policy, IPE needs a way to prevent rollback
267from the secure policy update to the insecure policy update.
268
269Initially, IPE's policy can have a policy_version that states the
270minimum required version across all policies that can be active on
271the system. This will prevent rollback while the system is live.
272
273.. WARNING::
274
275  However, since the kernel is stateless across boots, this policy
276  version will be reset to 0.0.0 on the next boot. System builders
277  need to be aware of this, and ensure the new secure policies are
278  deployed ASAP after a boot to ensure that the window of
279  opportunity is minimal for an attacker to deploy the insecure policy.
280
281Implicit Actions:
282~~~~~~~~~~~~~~~~~
283
284The issue of implicit actions only becomes visible when you consider
285a mixed level of security bars across multiple operations in a system.
286For example, consider a system that has strong integrity guarantees
287over both the executable code, and specific *data files* on the system,
288that were critical to its function. In this system, three types of policies
289are possible:
290
291  1. A policy in which failure to match any rules in the policy results
292     in the action being denied.
293  2. A policy in which failure to match any rules in the policy results
294     in the action being allowed.
295  3. A policy in which the action taken when no rules are matched is
296     specified by the policy author.
297
298The first option could make a policy like this::
299
300  op=EXECUTE integrity_verified=YES action=ALLOW
301
302In the example system, this works well for the executables, as all
303executables should have integrity guarantees, without exception. The
304issue becomes with the second requirement about specific data files.
305This would result in a policy like this (assuming each line is
306evaluated in order)::
307
308  op=EXECUTE integrity_verified=YES action=ALLOW
309
310  op=READ integrity_verified=NO label=critical_t action=DENY
311  op=READ action=ALLOW
312
313This is somewhat clear if you read the docs, understand the policy
314is executed in order and that the default is a denial; however, the
315last line effectively changes that default to an ALLOW. This is
316required, because in a realistic system, there are some unverified
317reads (imagine appending to a log file).
318
319The second option, matching no rules results in an allow, is clearer
320for the specific data files::
321
322  op=READ integrity_verified=NO label=critical_t action=DENY
323
324And, like the first option, falls short with the execution scenario,
325effectively needing to override the default::
326
327  op=EXECUTE integrity_verified=YES action=ALLOW
328  op=EXECUTE action=DENY
329
330  op=READ integrity_verified=NO label=critical_t action=DENY
331
332This leaves the third option. Instead of making users be clever
333and override the default with an empty rule, force the end-user
334to consider what the appropriate default should be for their
335scenario and explicitly state it::
336
337  DEFAULT op=EXECUTE action=DENY
338  op=EXECUTE integrity_verified=YES action=ALLOW
339
340  DEFAULT op=READ action=ALLOW
341  op=READ integrity_verified=NO label=critical_t action=DENY
342
343Policy Debugging:
344~~~~~~~~~~~~~~~~~
345
346When developing a policy, it is useful to know what line of the policy
347is being violated to reduce debugging costs; narrowing the scope of the
348investigation to the exact line that resulted in the action. Some integrity
349policy systems do not provide this information, instead providing the
350information that was used in the evaluation. This then requires a correlation
351with the policy to evaluate what went wrong.
352
353Instead, IPE just emits the rule that was matched. This limits the scope
354of the investigation to the exact policy line (in the case of a specific
355rule), or the section (in the case of a DEFAULT). This decreases iteration
356and investigation times when policy failures are observed while evaluating
357policies.
358
359IPE's policy engine is also designed in a way that it makes it obvious to
360a human of how to investigate a policy failure. Each line is evaluated in
361the sequence that is written, so the algorithm is very simple to follow
362for humans to recreate the steps and could have caused the failure. In other
363surveyed systems, optimizations occur (sorting rules, for instance) when loading
364the policy. In those systems, it requires multiple steps to debug, and the
365algorithm may not always be clear to the end-user without reading the code first.
366
367Simplified Policy:
368~~~~~~~~~~~~~~~~~~
369
370Finally, IPE's policy is designed for sysadmins, not kernel developers. Instead
371of covering individual LSM hooks (or syscalls), IPE covers operations. This means
372instead of sysadmins needing to know that the syscalls ``mmap``, ``mprotect``,
373``execve``, and ``uselib`` must have rules protecting them, they must simple know
374that they want to restrict code execution. This limits the amount of bypasses that
375could occur due to a lack of knowledge of the underlying system; whereas the
376maintainers of IPE, being kernel developers can make the correct choice to determine
377whether something maps to these operations, and under what conditions.
378
379Implementation Notes
380--------------------
381
382Anonymous Memory
383~~~~~~~~~~~~~~~~
384
385Anonymous memory isn't treated any differently from any other access in IPE.
386When anonymous memory is mapped with ``+X``, it still comes into the ``file_mmap``
387or ``file_mprotect`` hook, but with a ``NULL`` file object. This is submitted to
388the evaluation, like any other file. However, all current trust properties will
389evaluate to false, as they are all file-based and the operation is not
390associated with a file.
391
392.. WARNING::
393
394  This also occurs with the ``kernel_load_data`` hook, when the kernel is
395  loading data from a userspace buffer that is not backed by a file. In this
396  scenario all current trust properties will also evaluate to false.
397
398Securityfs Interface
399~~~~~~~~~~~~~~~~~~~~
400
401The per-policy securityfs tree is somewhat unique. For example, for
402a standard securityfs policy tree::
403
404  MyPolicy
405    |- active
406    |- delete
407    |- name
408    |- pkcs7
409    |- policy
410    |- update
411    |- version
412
413The policy is stored in the ``->i_private`` data of the MyPolicy inode.
414
415Tests
416-----
417
418IPE has KUnit Tests for the policy parser. Recommended kunitconfig::
419
420  CONFIG_KUNIT=y
421  CONFIG_SECURITY=y
422  CONFIG_SECURITYFS=y
423  CONFIG_PKCS7_MESSAGE_PARSER=y
424  CONFIG_SYSTEM_DATA_VERIFICATION=y
425  CONFIG_FS_VERITY=y
426  CONFIG_FS_VERITY_BUILTIN_SIGNATURES=y
427  CONFIG_BLOCK=y
428  CONFIG_MD=y
429  CONFIG_BLK_DEV_DM=y
430  CONFIG_DM_VERITY=y
431  CONFIG_DM_VERITY_VERIFY_ROOTHASH_SIG=y
432  CONFIG_NET=y
433  CONFIG_AUDIT=y
434  CONFIG_AUDITSYSCALL=y
435  CONFIG_BLK_DEV_INITRD=y
436
437  CONFIG_SECURITY_IPE=y
438  CONFIG_IPE_PROP_DM_VERITY=y
439  CONFIG_IPE_PROP_DM_VERITY_SIGNATURE=y
440  CONFIG_IPE_PROP_FS_VERITY=y
441  CONFIG_IPE_PROP_FS_VERITY_BUILTIN_SIG=y
442  CONFIG_SECURITY_IPE_KUNIT_TEST=y
443
444In addition, IPE has a python based integration
445`test suite <https://github.com/microsoft/ipe/tree/test-suite>`_ that
446can test both user interfaces and enforcement functionalities.
447