Implement LANDLOCK_ADD_RULE_QUIET

Hi,

This is the v11 of the "quiet flag" series, implementing the feature as proposed in [1].

v10: https://lore.kernel.org/all/[email protected]/
v9: https://lore.kernel.org/all/[email protected]/
v8: https://lore.kernel.org/all/[email protected]/
v7: https://lore.kernel.org/all/[email protected]/
v6: https://lore.kernel.org/all/[email protected]/
v5: https://lore.kernel.org/all/[email protected]/
v4: https://lore.kernel.org/all/[email protected]/
v3: https://lore.kernel.org/all/[email protected]/
v2: https://lore.kernel.org/all/[email protected]/
v1: https://lore.kernel.org/all/[email protected]/

v10..v11:
- doc and style fixes
- s/audit log/log/
- u32 flags arguments
- stop using bitfields for quiet_optional_accesses and fown_layer
- fixed bitfield using different types
- use u8 instead of bool in struct landlock_layer
- sandboxer: merge LL_{FS,NET,SCOPED}_QUIET_ACCESS into a single LL_QUIET_ACCESS with more descriptive values.
- selftests: also test the quiet_access_net and quiet_scoped fields.

(Kept ABI version at 10)

All text following this line is unchanged except for the demo
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

v9..v10:
- clang-format on .h files
- doc changes
- remove stray __attribute__((fallthrough));

v8..v9:
- Refactor to store the collected rule flags in layer_masks instead (renamed from layer_access_masks). Got rid of layer_mask_t again.
- Rebase sandboxer and net_tests on top of UDP support, resolving conflicts
- Additional small changes, noted in each patch

v7..v8:

- Rebase to mic/next
- Re-introduced layer_mask_t due to need in first patch
- Plumb through rule flags in hook_unix_find()
- Some selftests patches were not properly clang-format'd, fixed now.
- Minor env var handling change in sandboxer
- Fix selftests use of audit_count_records() without EXPECT_EQ

v6..v7:

- Remove "landlock: Fix wrong type usage" (merged)
- Revert back to taking rule_flags separately from landlock_request until
we call landlock_log_denial (https://lore.kernel.org/all/[email protected]/)
- Rebase to mic/next

v5..v6 rebases on top of the new simpler disconnected directory handling, change some bools into u32, and fix some typo and style.

v4..v5 addresses review feedbacks, most significantly:
- reduces code changes by pushing rule_flags into landlock_request.
- adding test cases for two layers handling different access bits.

v3..v4 is a one-character formatting change, plus more tests.

We now have 5 patches for the selftest - I'm happy to squash it into one depending on preference (and happy for Mickaël to do the squash if no other feedback):
- selftests/landlock: Replace hard-coded 16 with a constant
- selftests/landlock: add tests for quiet flag with fs rules
- selftests/landlock: add tests for quiet flag with net rules
- selftests/landlock: Add tests for quiet flag with scope
- selftests/landlock: Add tests for invalid use of quiet flag

v2..v3:
Not much has changed in the actual functionality except various comment, typing, asserts and general style fixes based on feedback. The major new thing here is tests (a bit of KUnit squashed into the optional access commit, a lot of selftests especially in fs_tests.c).

The added fs_tests should exercise code path for optional and non-optional access, renames, and mountpoint and disconnected directory handling. I will add the above missing bits to v4.

Removed:
- "Implement quiet for optional accesses"
(squashed into "landlock: Suppress logging when quiet flag is present")

Old feature summary below:

The quiet flag allows a sandboxer to suppress audit logs for uninteresting denials. The flag can be set on objects and inherits downward in the filesystem hierarchy. On a denial, the youngest denying layer's quiet flag setting decides whether to audit. The motivation for this feature is to reduce audit noise, and also prepare for a future supervisor feature which will use this bit to suppress supervisor notifications.

This patch introduces a new quiet access mask in the ruleset_attr, which gets eventually stored in the hierarchy. This allows the user to specify which access should be affected by quiet bits. One can then, for example, make it such that read accesses to certain files are not audited (but still denied), but all writes are still audited, regardless of location.

The sandboxer is extended to show example usage of this feature, supporting quieting filesystem, network and scope accesses.

Demo:

    /# LL_FS_RO=/usr LL_FS_RW= LL_FORCE_LOG=1 LL_FS_QUIET=/dev:/tmp:/etc LL_QUIET_ACCESS=read ./sandboxer bash
    ...
    audit: type=1423 audit(1759680175.562:195): domain=15bb25f6b blockers=fs.write_file,fs.read_file path="/dev/tty" dev="devtmpfs" ino=11
    ^^^^^^^^
    # note: because write is not quieted, we see the above line. blockers
    # contains read as well since that's the originally requested access.
    audit: type=1424 audit(1759680175.562:195): domain=15bb25f6b status=allocated mode=enforcing pid=616 uid=0 exe="/sandboxer" comm="sandboxer"
    audit: type=1300 audit(1759680175.562:195): arch=c000003e syscall=257 success=no exit=-13 a0=ffffffffffffff9c a1=5565c86113d1 a2=802 a3=0 items=0 ppid=605 pid=616 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295 comm="bash" exe="/usr/bin/bash" key=(null)
    audit: type=1327 audit(1759680175.562:195): proctitle="bash"
    bash: cannot set terminal process group (605): Inappropriate ioctl for device
    bash: no job control in this shell
    bash: /etc/bash.bashrc: Permission denied
    audit: type=1423 audit(1759680175.570:196): domain=15bb25f6b blockers=fs.read_file path="/.bash_history" dev="virtiofs" ino=36963
    ^^^^^^^^
    # read outside /dev:/tmp:/etc - not quieted
    audit: type=1300 audit(1759680175.570:196): arch=c000003e syscall=257 success=no exit=-13 a0=ffffffffffffff9c a1=5565c868e400 a2=0 a3=0 items=0 ppid=605 pid=616 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295 comm="bash" exe="/usr/bin/bash" key=(null)
    audit: type=1327 audit(1759680175.570:196): proctitle="bash"
    audit: type=1423 audit(1759680175.570:197): domain=15bb25f6b blockers=fs.read_file path="/.bash_history" dev="virtiofs" ino=36963
    audit: type=1300 audit(1759680175.570:197): arch=c000003e syscall=257 success=no exit=-13 a0=ffffffffffffff9c a1=5565c868e400 a2=0 a3=0 items=0 ppid=605 pid=616 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295 comm="bash" exe="/usr/bin/bash" key=(null)
    audit: type=1327 audit(1759680175.570:197): proctitle="bash"

    bash-5.2# head /etc/passwd
    head: cannot open '/etc/passwd' for reading: Permission denied
    ^^^^^^^^
    # reads to /etc are quieted

    bash-5.2# echo evil >> /etc/passwd
    bash: /etc/passwd: Permission denied
    audit: type=1423 audit(1759680227.030:198): domain=15bb25f6b blockers=fs.write_file path="/etc/passwd" dev="virtiofs" ino=790
    ^^^^^^^^
    # writes are not quieted
    audit: type=1300 audit(1759680227.030:198): arch=c000003e syscall=257 success=no exit=-13 a0=ffffffffffffff9c a1=5565c86ab030 a2=441 a3=1b6 items=0 ppid=605 pid=616 auid=4294967295 uid=0 gid=0 euid=0 suid=0 fsuid=0 egid=0 sgid=0 fsgid=0 tty=(none) ses=4294967295 comm="bash" exe="/usr/bin/bash" key=(null)
    audit: type=1327 audit(1759680227.030:198): proctitle="bash"

Design:

- The user can set the quiet flag for a layer on any part of the fs hierarchy (whether it allows any access on it or not), and the flag inherits down (no support for "cancelling" the inheritance of the flag in specific subdirectories).

- The youngest layer that denies a request gets to decide whether the denial is audited or not. This means that a compromised binary, for example, cannot "turn off" Landlock auditing when it tries to access files, unless it denies access to the files itself. There is some debate to be had on whether, if a parent layer sets the quiet flag, but the request is denied by a deeper layer, whether Landlock should still audit anyway (since the rule author of the child layer likely did not expect the denial, so it would be good diagnostic). The current approach is to ignore the quiet on the parent layer and audit anyway.

[1]: https://github.com/landlock-lsm/linux/issues/44#issuecomment-2876500918

Kind regards,
Tingmao

Tingmao Wang (9):
landlock: Add a place for flags to layer rules
landlock: Add API support and docs for the quiet flags
landlock: Suppress logging when quiet flag is present samples/landlock: Add quiet flag support to sandboxer selftests/landlock: Replace hard-coded 16 with a constant selftests/landlock: add tests for quiet flag with fs rules selftests/landlock: add tests for quiet flag with net rules selftests/landlock: Add tests for quiet flag with scope selftests/landlock: Add tests for invalid use of quiet flag

 Documentation/admin-guide/LSM/landlock.rst    |    9 +-
 Documentation/userspace-api/landlock.rst      |   14 +
 include/uapi/linux/landlock.h                 |   61 +
 samples/landlock/sandboxer.c                  |  138 +-
 security/landlock/access.h                    |   44 +-
 security/landlock/audit.c                     |  288 +-
 security/landlock/audit.h                     |    3 +-
 security/landlock/domain.c                    |   57 +-
 security/landlock/domain.h                    |   11 +-
 security/landlock/fs.c                        |  157 +-
 security/landlock/fs.h                        |   21 +-
 security/landlock/limits.h                    |    3 +
 security/landlock/net.c                       |   22 +-
 security/landlock/net.h                       |    5 +-
 security/landlock/ruleset.c                   |   45 +-
 security/landlock/ruleset.h                   |   29 +-
 security/landlock/syscalls.c                  |   71 +-
 tools/testing/selftests/landlock/audit_test.c |   27 +-
 tools/testing/selftests/landlock/base_test.c  |  118 +-
 tools/testing/selftests/landlock/common.h     |    2 +
 tools/testing/selftests/landlock/fs_test.c    | 2450 ++++++++++++++++-
 tools/testing/selftests/landlock/net_test.c   |  138 +-
 .../landlock/scoped_abstract_unix_test.c      |   77 +-
 23 files changed, 3576 insertions(+), 214 deletions(-)

base-commit: a6f0a6f5377fae42a8028f63c89d544c68f24b60

2.54.0