Add tests for the hostfs filesystems to make sure it has a consistent
inode management, which is required for Landlock's file hierarchy
identification. This adds 5 new tests for layout3_fs with the hostfs
variant.
Add hostfs to the new (architecture-specific) config.um file.
The hostfs filesystem, only available for an User-Mode Linux kernel, is
special because we cannot explicitly mount it. The layout3_fs.hostfs
variant tests are skipped if the current test directory is not backed by
this filesystem.
The layout3_fs.hostfs.tag_inode_dir_child and
layout3_fs.hostfs.tag_inode_file tests pass thanks to a previous commit
fixing hostfs inode management. Without this fix, the deny-by-default
policy would apply and all access requests would be denied.
Link: https://lore.kernel.org/r/20230612191430.339153-7-mic@digikod.net
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Add generic and read-only tests for 6 pseudo filesystems to make sure
they have a consistent inode management, which is required for
Landlock's file hierarchy identification:
- tmpfs
- ramfs
- cgroup2
- proc
- sysfs
Update related kernel configuration to support these new filesystems,
remove useless CONFIG_SECURITY_PATH, and sort all entries. If these
filesystems are not supported by the kernel running tests, the related
tests are skipped.
Expanding variants, this adds 25 new tests for layout3_fs:
- tag_inode_dir_parent
- tag_inode_dir_mnt
- tag_inode_dir_child
- tag_inode_dir_file
- release_inodes
Test coverage for security/landlock with kernel debug code:
- 94.7% of 835 lines according to gcc/gcov-12
- 93.0% of 852 lines according to gcc/gcov-13
Test coverage for security/landlock without kernel debug code:
- 95.5% of 624 lines according to gcc/gcov-12
- 93.1% of 641 lines according to gcc/gcov-13
Link: https://lore.kernel.org/r/20230612191430.339153-6-mic@digikod.net
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Add and use a layout0 test fixture to not populate the tmpfs filesystem
if it is not required for tests: unknown_access_rights, proc_nsfs,
unpriv and max_layers.
This doesn't change these tests but it speeds up their setup and makes
them less prone to error. This prepare the ground for a next commit.
Link: https://lore.kernel.org/r/20230612191430.339153-3-mic@digikod.net
Signed-off-by: Mickaël Salaün <mic@digikod.net>
All file descriptors that are truncatable need to have the Landlock
access rights set correctly on the file's Landlock security blob. This
is also the case for files that are opened by other means than
open(2).
Test coverage for security/landlock is 94.7% of 838 lines according to
gcc/gcov-11.
Signed-off-by: Günther Noack <gnoack3000@gmail.com>
Link: https://lore.kernel.org/r/20221018182216.301684-10-gnoack3000@gmail.com
[mic: Add test coverage in commit message]
Signed-off-by: Mickaël Salaün <mic@digikod.net>
A file descriptor created in a restricted process carries Landlock
restrictions with it which will apply even if the same opened file is
used from an unrestricted process.
This change extracts suitable FD-passing helpers from base_test.c and
moves them to common.h. We use the fixture variants from the ftruncate
fixture to exercise the same scenarios as in the open_and_ftruncate
test, but doing the Landlock restriction and open() in a different
process than the ftruncate() call.
Signed-off-by: Günther Noack <gnoack3000@gmail.com>
Link: https://lore.kernel.org/r/20221018182216.301684-9-gnoack3000@gmail.com
Signed-off-by: Mickaël Salaün <mic@digikod.net>
These tests exercise the following truncation operations:
* truncate() (truncate by path)
* ftruncate() (truncate by file descriptor)
* open with the O_TRUNC flag
* special case: creat(), which is open with O_CREAT|O_WRONLY|O_TRUNC.
in the following scenarios:
* Files with read, write and truncate rights.
* Files with read and truncate rights.
* Files with the truncate right.
* Files without the truncate right.
In particular, the following scenarios are enforced with the test:
* open() with O_TRUNC requires the truncate right, if it truncates a file.
open() already checks security_path_truncate() in this case,
and it required no additional check in the Landlock LSM's file_open hook.
* creat() requires the truncate right
when called with an existing filename.
* creat() does *not* require the truncate right
when it's creating a new file.
* ftruncate() requires that the file was opened by a thread that had
the truncate right for the file at the time of open(). (The rights
are carried along with the opened file.)
Signed-off-by: Günther Noack <gnoack3000@gmail.com>
Link: https://lore.kernel.org/r/20221018182216.301684-6-gnoack3000@gmail.com
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Introduce the LANDLOCK_ACCESS_FS_TRUNCATE flag for file truncation.
This flag hooks into the path_truncate, file_truncate and
file_alloc_security LSM hooks and covers file truncation using
truncate(2), ftruncate(2), open(2) with O_TRUNC, as well as creat().
This change also increments the Landlock ABI version, updates
corresponding selftests, and updates code documentation to document
the flag.
In security/security.c, allocate security blobs at pointer-aligned
offsets. This fixes the problem where one LSM's security blob can
shift another LSM's security blob to an unaligned address (reported
by Nathan Chancellor).
The following operations are restricted:
open(2): requires the LANDLOCK_ACCESS_FS_TRUNCATE right if a file gets
implicitly truncated as part of the open() (e.g. using O_TRUNC).
Notable special cases:
* open(..., O_RDONLY|O_TRUNC) can truncate files as well in Linux
* open() with O_TRUNC does *not* need the TRUNCATE right when it
creates a new file.
truncate(2) (on a path): requires the LANDLOCK_ACCESS_FS_TRUNCATE
right.
ftruncate(2) (on a file): requires that the file had the TRUNCATE
right when it was previously opened. File descriptors acquired by
other means than open(2) (e.g. memfd_create(2)) continue to support
truncation with ftruncate(2).
Cc: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Günther Noack <gnoack3000@gmail.com>
Acked-by: Paul Moore <paul@paul-moore.com> (LSM)
Link: https://lore.kernel.org/r/20221018182216.301684-5-gnoack3000@gmail.com
Signed-off-by: Mickaël Salaün <mic@digikod.net>
This change fixes a mis-handling of the LANDLOCK_ACCESS_FS_REFER right
when multiple rulesets/domains are stacked. The expected behaviour was
that an additional ruleset can only restrict the set of permitted
operations, but in this particular case, it was potentially possible to
re-gain the LANDLOCK_ACCESS_FS_REFER right.
With the introduction of LANDLOCK_ACCESS_FS_REFER, we added the first
globally denied-by-default access right. Indeed, this lifted an initial
Landlock limitation to rename and link files, which was initially always
denied when the source or the destination were different directories.
This led to an inconsistent backward compatibility behavior which was
only taken into account if no domain layer were using the new
LANDLOCK_ACCESS_FS_REFER right. However, when restricting a thread with
a new ruleset handling LANDLOCK_ACCESS_FS_REFER, all inherited parent
rulesets/layers not explicitly handling LANDLOCK_ACCESS_FS_REFER would
behave as if they were handling this access right and with all their
rules allowing it. This means that renaming and linking files could
became allowed by these parent layers, but all the other required
accesses must also be granted: all layers must allow file removal or
creation, and renaming and linking operations cannot lead to privilege
escalation according to the Landlock policy. See detailed explanation
in commit b91c3e4ea7 ("landlock: Add support for file reparenting with
LANDLOCK_ACCESS_FS_REFER").
To say it another way, this bug may lift the renaming and linking
limitations of the initial Landlock version, and a same ruleset can
enforce different restrictions depending on previous or next enforced
ruleset (i.e. inconsistent behavior). The LANDLOCK_ACCESS_FS_REFER right
cannot give access to data not already allowed, but this doesn't follow
the contract of the first Landlock ABI. This fix puts back the
limitation for sandboxes that didn't opt-in for this additional right.
For instance, if a first ruleset allows LANDLOCK_ACCESS_FS_MAKE_REG on
/dst and LANDLOCK_ACCESS_FS_REMOVE_FILE on /src, renaming /src/file to
/dst/file is denied. However, without this fix, stacking a new ruleset
which allows LANDLOCK_ACCESS_FS_REFER on / would now permit the
sandboxed thread to rename /src/file to /dst/file .
This change fixes the (absolute) rule access rights, which now always
forbid LANDLOCK_ACCESS_FS_REFER except when it is explicitly allowed
when creating a rule.
Making all domain handle LANDLOCK_ACCESS_FS_REFER was an initial
approach but there is two downsides:
* it makes the code more complex because we still want to check that a
rule allowing LANDLOCK_ACCESS_FS_REFER is legitimate according to the
ruleset's handled access rights (i.e. ABI v1 != ABI v2);
* it would not allow to identify if the user created a ruleset
explicitly handling LANDLOCK_ACCESS_FS_REFER or not, which will be an
issue to audit Landlock.
Instead, this change adds an ACCESS_INITIALLY_DENIED list of
denied-by-default rights, which (only) contains
LANDLOCK_ACCESS_FS_REFER. All domains are treated as if they are also
handling this list, but without modifying their fs_access_masks field.
A side effect is that the errno code returned by rename(2) or link(2)
*may* be changed from EXDEV to EACCES according to the enforced
restrictions. Indeed, we now have the mechanic to identify if an access
is denied because of a required right (e.g. LANDLOCK_ACCESS_FS_MAKE_REG,
LANDLOCK_ACCESS_FS_REMOVE_FILE) or if it is denied because of missing
LANDLOCK_ACCESS_FS_REFER rights. This may result in different errno
codes than for the initial Landlock version, but this approach is more
consistent and better for rename/link compatibility reasons, and it
wasn't possible before (hence no backport to ABI v1). The
layout1.rename_file test reflects this change.
Add 4 layout1.refer_denied_by_default* test suites to check that the
behavior of a ruleset not handling LANDLOCK_ACCESS_FS_REFER (ABI v1) is
unchanged even if another layer handles LANDLOCK_ACCESS_FS_REFER (i.e.
ABI v1 precedence). Make sure rule's absolute access rights are correct
by testing with and without a matching path. Add test_rename() and
test_exchange() helpers.
Extend layout1.inval tests to check that a denied-by-default access
right is not necessarily part of a domain's handled access rights.
Test coverage for security/landlock is 95.3% of 599 lines according to
gcc/gcov-11.
Fixes: b91c3e4ea7 ("landlock: Add support for file reparenting with LANDLOCK_ACCESS_FS_REFER")
Reviewed-by: Paul Moore <paul@paul-moore.com>
Reviewed-by: Günther Noack <gnoack3000@gmail.com>
Link: https://lore.kernel.org/r/20220831203840.1370732-1-mic@digikod.net
Cc: stable@vger.kernel.org
[mic: Constify and slightly simplify test helpers]
Signed-off-by: Mickaël Salaün <mic@digikod.net>
These test suites try to check all edge cases for directory and file
renaming or linking involving a new parent directory, with and without
LANDLOCK_ACCESS_FS_REFER and other access rights.
layout1:
* reparent_refer: Tests simple FS_REFER usage.
* reparent_link: Tests a mix of FS_MAKE_REG and FS_REFER with links.
* reparent_rename: Tests a mix of FS_MAKE_REG and FS_REFER with renames
and RENAME_EXCHANGE.
* reparent_exdev_layers_rename1/2: Tests renames with two layers.
* reparent_exdev_layers_exchange1/2/3: Tests exchanges with two layers.
* reparent_remove: Tests file and directory removal with rename.
* reparent_dom_superset: Tests access partial ordering.
layout1_bind:
* reparent_cross_mount: Tests FS_REFER propagation across mount points.
Test coverage for security/landlock is 95.4% of 604 lines according to
gcc/gcov-11.
Cc: Paul Moore <paul@paul-moore.com>
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Link: https://lore.kernel.org/r/20220506161102.525323-9-mic@digikod.net
Add a new LANDLOCK_ACCESS_FS_REFER access right to enable policy writers
to allow sandboxed processes to link and rename files from and to a
specific set of file hierarchies. This access right should be composed
with LANDLOCK_ACCESS_FS_MAKE_* for the destination of a link or rename,
and with LANDLOCK_ACCESS_FS_REMOVE_* for a source of a rename. This
lift a Landlock limitation that always denied changing the parent of an
inode.
Renaming or linking to the same directory is still always allowed,
whatever LANDLOCK_ACCESS_FS_REFER is used or not, because it is not
considered a threat to user data.
However, creating multiple links or renaming to a different parent
directory may lead to privilege escalations if not handled properly.
Indeed, we must be sure that the source doesn't gain more privileges by
being accessible from the destination. This is handled by making sure
that the source hierarchy (including the referenced file or directory
itself) restricts at least as much the destination hierarchy. If it is
not the case, an EXDEV error is returned, making it potentially possible
for user space to copy the file hierarchy instead of moving or linking
it.
Instead of creating different access rights for the source and the
destination, we choose to make it simple and consistent for users.
Indeed, considering the previous constraint, it would be weird to
require such destination access right to be also granted to the source
(to make it a superset). Moreover, RENAME_EXCHANGE would also add to
the confusion because of paths being both a source and a destination.
See the provided documentation for additional details.
New tests are provided with a following commit.
Reviewed-by: Paul Moore <paul@paul-moore.com>
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Link: https://lore.kernel.org/r/20220506161102.525323-8-mic@digikod.net
The original behavior was to check if the full set of requested accesses
was allowed by at least a rule of every relevant layer. This didn't
take into account requests for multiple accesses and same-layer rules
allowing the union of these accesses in a complementary way. As a
result, multiple accesses requested on a file hierarchy matching rules
that, together, allowed these accesses, but without a unique rule
allowing all of them, was illegitimately denied. This case should be
rare in practice and it can only be triggered by the path_rename or
file_open hook implementations.
For instance, if, for the same layer, a rule allows execution
beneath /a/b and another rule allows read beneath /a, requesting access
to read and execute at the same time for /a/b should be allowed for this
layer.
This was an inconsistency because the union of same-layer rule accesses
was already allowed if requested once at a time anyway.
This fix changes the way allowed accesses are gathered over a path walk.
To take into account all these rule accesses, we store in a matrix all
layer granting the set of requested accesses, according to the handled
accesses. To avoid heap allocation, we use an array on the stack which
is 2*13 bytes. A following commit bringing the LANDLOCK_ACCESS_FS_REFER
access right will increase this size to reach 112 bytes (2*14*4) in case
of link or rename actions.
Add a new layout1.layer_rule_unions test to check that accesses from
different rules pertaining to the same layer are ORed in a file
hierarchy. Also test that it is not the case for rules from different
layers.
Reviewed-by: Paul Moore <paul@paul-moore.com>
Link: https://lore.kernel.org/r/20220506161102.525323-5-mic@digikod.net
Cc: stable@vger.kernel.org
Signed-off-by: Mickaël Salaün <mic@digikod.net>
The maximum number of nested Landlock domains is currently 64. Because
of the following fix and to help reduce the stack size, let's reduce it
to 16. This seems large enough for a lot of use cases (e.g. sandboxed
init service, spawning a sandboxed SSH service, in nested sandboxed
containers). Reducing the number of nested domains may also help to
discover misuse of Landlock (e.g. creating a domain per rule).
Add and use a dedicated layer_mask_t typedef to fit with the number of
layers. This might be useful when changing it and to keep it consistent
with the maximum number of layers.
Reviewed-by: Paul Moore <paul@paul-moore.com>
Link: https://lore.kernel.org/r/20220506161102.525323-3-mic@digikod.net
Cc: stable@vger.kernel.org
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Replace SYS_<syscall> with __NR_<syscall>. Using the __NR_<syscall>
notation, provided by UAPI, is useful to build tests on systems without
the SYS_<syscall> definitions.
Replace SYS_pivot_root with __NR_pivot_root, and SYS_move_mount with
__NR_move_mount.
Define renameat2() and RENAME_EXCHANGE if they are unknown to old build
systems.
Cc: Shuah Khan <shuah@kernel.org>
Link: https://lore.kernel.org/r/20220506160820.524344-3-mic@digikod.net
Cc: stable@vger.kernel.org
Signed-off-by: Mickaël Salaün <mic@digikod.net>
In preparation to a following commit, add clang-format on and
clang-format off stanzas around constant definitions and the TEST_F_FORK
macro. This enables to keep aligned values, which is much more readable
than packed definitions.
Add other clang-format exceptions for FIXTURE() and
FIXTURE_VARIANT_ADD() declarations to force space before open brace,
which is reported by checkpatch.pl .
Link: https://lore.kernel.org/r/20220506160513.523257-4-mic@digikod.net
Cc: stable@vger.kernel.org
Signed-off-by: Mickaël Salaün <mic@digikod.net>
