BPF token support was introduced to allow a privileged process to delegate
limited BPF functionality—such as map creation and program loading—to
an unprivileged process:
https://lore.kernel.org/linux-security-module/20231130185229.2688956-1-andrii@kernel.org/
This patch adds SELinux support for controlling BPF token access. With
this change, SELinux policies can now enforce constraints on BPF token
usage based on both the delegating (privileged) process and the recipient
(unprivileged) process.
Supported operations currently include:
- map_create
- prog_load
High-level workflow:
1. An unprivileged process creates a VFS context via `fsopen()` and
obtains a file descriptor.
2. This descriptor is passed to a privileged process, which configures
BPF token delegation options and mounts a BPF filesystem.
3. SELinux records the `creator_sid` of the privileged process during
mount setup.
4. The unprivileged process then uses this BPF fs mount to create a
token and attach it to subsequent BPF syscalls.
5. During verification of `map_create` and `prog_load`, SELinux uses
`creator_sid` and the current SID to check policy permissions via:
avc_has_perm(creator_sid, current_sid, SECCLASS_BPF,
BPF__MAP_CREATE, NULL);
The implementation introduces two new permissions:
- map_create_as
- prog_load_as
At token creation time, SELinux verifies that the current process has the
appropriate `*_as` permission (depending on the `allowed_cmds` value in
the bpf_token) to act on behalf of the `creator_sid`.
Example SELinux policy:
allow test_bpf_t self:bpf {
map_create map_read map_write prog_load prog_run
map_create_as prog_load_as
};
Additionally, a new policy capability bpf_token_perms is added to ensure
backward compatibility. If disabled, previous behavior ((checks based on
current process SID)) is preserved.
Signed-off-by: Eric Suen <ericsu@linux.microsoft.com>
Tested-by: Daniel Durning <danieldurning.work@gmail.com>
Reviewed-by: Daniel Durning <danieldurning.work@gmail.com>
[PM: merge fuzz, subject tweaks, whitespace tweaks, line length tweaks]
Signed-off-by: Paul Moore <paul@paul-moore.com>
Prior to this change, no security hooks were called at the creation of a
memfd file. It means that, for SELinux as an example, it will receive
the default type of the filesystem that backs the in-memory inode. In
most cases, that would be tmpfs, but if MFD_HUGETLB is passed, it will
be hugetlbfs. Both can be considered implementation details of memfd.
It also means that it is not possible to differentiate between a file
coming from memfd_create and a file coming from a standard tmpfs mount
point.
Additionally, no permission is validated at creation, which differs from
the similar memfd_secret syscall.
Call security_inode_init_security_anon during creation. This ensures
that the file is setup similarly to other anonymous inodes. On SELinux,
it means that the file will receive the security context of its task.
The ability to limit fexecve on memfd has been of interest to avoid
potential pitfalls where /proc/self/exe or similar would be executed
[1][2]. Reuse the "execute_no_trans" and "entrypoint" access vectors,
similarly to the file class. These access vectors may not make sense for
the existing "anon_inode" class. Therefore, define and assign a new
class "memfd_file" to support such access vectors.
Guard these changes behind a new policy capability named "memfd_class".
[1] https://crbug.com/1305267
[2] https://lore.kernel.org/lkml/20221215001205.51969-1-jeffxu@google.com/
Signed-off-by: Thiébaud Weksteen <tweek@google.com>
Reviewed-by: Stephen Smalley <stephen.smalley.work@gmail.com>
Tested-by: Stephen Smalley <stephen.smalley.work@gmail.com>
Acked-by: Hugh Dickins <hughd@google.com>
[PM: subj tweak]
Signed-off-by: Paul Moore <paul@paul-moore.com>
This patch adds support for genfscon per-file labeling of functionfs
files as well as support for userspace to apply labels after new
functionfs endpoints are created.
This allows for separate labels and therefore access control on a
per-endpoint basis. An example use case would be for the default
endpoint EP0 used as a restricted control endpoint, and additional
usb endpoints to be used by other more permissive domains.
It should be noted that if there are multiple functionfs mounts on a
system, genfs file labels will apply to all mounts, and therefore will not
likely be as useful as the userspace relabeling portion of this patch -
the addition to selinux_is_genfs_special_handling().
This patch introduces the functionfs_seclabel policycap to maintain
existing functionfs genfscon behavior unless explicitly enabled.
Signed-off-by: Neill Kapron <nkapron@google.com>
Acked-by: Stephen Smalley <stephen.smalley.work@gmail.com>
[PM: trim changelog, apply boolean logic fixup]
Signed-off-by: Paul Moore <paul@paul-moore.com>
Currently, genfscon only supports string prefix match to label files.
Thus, labeling numerous dynamic sysfs entries requires many specific
path rules. For example, labeling device paths such as
`/sys/devices/pci0000:00/0000:00:03.1/<...>/0000:04:00.1/wakeup`
requires listing all specific PCI paths, which is challenging to
maintain. While user-space restorecon can handle these paths with
regular expression rules, relabeling thousands of paths under sysfs
after it is mounted is inefficient compared to using genfscon.
This commit adds wildcard matching to genfscon to make rules more
efficient and expressive. This new behavior is enabled by
genfs_seclabel_wildcard capability. With this capability, genfscon does
wildcard matching instead of prefix matching. When multiple wildcard
rules match against a path, then the longest rule (determined by the
length of the rule string) will be applied. If multiple rules of the
same length match, the first matching rule encountered in the given
genfscon policy will be applied. Users are encouraged to write longer,
more explicit path rules to avoid relying on this behavior.
This change resulted in nice real-world performance improvements. For
example, boot times on test Android devices were reduced by 15%. This
improvement is due to the elimination of the "restorecon -R /sys" step
during boot, which takes more than two seconds in the worst case.
Signed-off-by: Takaya Saeki <takayas@chromium.org>
Signed-off-by: Paul Moore <paul@paul-moore.com>
Add support for wildcard matching of network interface names. This is
useful for auto-generated interfaces, for example podman creates network
interfaces for containers with the naming scheme podman0, podman1,
podman2, ...
To maintain backward compatibility guard this feature with a new policy
capability 'netif_wildcard'.
Netifcon definitions are compared against in the order given by the
policy, so userspace tools should sort them in a reasonable order.
Signed-off-by: Christian Göttsche <cgzones@googlemail.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>
Reuse the existing extended permissions infrastructure to support
policies based on the netlink message types.
A new policy capability "netlink_xperm" is introduced. When disabled,
the previous behaviour is preserved. That is, netlink_send will rely on
the permission mappings defined in nlmsgtab.c (e.g, nlmsg_read for
RTM_GETADDR on NETLINK_ROUTE). When enabled, the mappings are ignored
and the generic "nlmsg" permission is used instead.
The new "nlmsg" permission is an extended permission. The 16 bits of the
extended permission are mapped to the nlmsg_type field.
Example policy on Android, preventing regular apps from accessing the
device's MAC address and ARP table, but allowing this access to
privileged apps, looks as follows:
allow netdomain self:netlink_route_socket {
create read getattr write setattr lock append connect getopt
setopt shutdown nlmsg
};
allowxperm netdomain self:netlink_route_socket nlmsg ~{
RTM_GETLINK RTM_GETNEIGH RTM_GETNEIGHTBL
};
allowxperm priv_app self:netlink_route_socket nlmsg {
RTM_GETLINK RTM_GETNEIGH RTM_GETNEIGHTBL
};
The constants in the example above (e.g., RTM_GETLINK) are explicitly
defined in the policy.
It is possible to generate policies to support kernels that may or
may not have the capability enabled by generating a rule for each
scenario. For instance:
allow domain self:netlink_audit_socket nlmsg_read;
allow domain self:netlink_audit_socket nlmsg;
allowxperm domain self:netlink_audit_socket nlmsg { AUDIT_GET };
The approach of defining a new permission ("nlmsg") instead of relying
on the existing permissions (e.g., "nlmsg_read", "nlmsg_readpriv" or
"nlmsg_tty_audit") has been preferred because:
1. This is similar to the other extended permission ("ioctl");
2. With the new extended permission, the coarse-grained mapping is not
necessary anymore. It could eventually be removed, which would be
impossible if the extended permission was defined below these.
3. Having a single extra extended permission considerably simplifies
the implementation here and in libselinux.
Signed-off-by: Thiébaud Weksteen <tweek@google.com>
Signed-off-by: Bram Bonné <brambonne@google.com>
[PM: manual merge fixes for sock_skip_has_perm()]
Signed-off-by: Paul Moore <paul@paul-moore.com>
As part of on ongoing effort to perform more automated testing and
provide more tools for individual developers to validate their
patches before submitting, we are trying to make our code
"clang-format clean". My hope is that once we have fixed all of our
style "quirks", developers will be able to run clang-format on their
patches to help avoid silly formatting problems and ensure their
changes fit in well with the rest of the SELinux kernel code.
Signed-off-by: Paul Moore <paul@paul-moore.com>
Currently, SELinux doesn't allow distinguishing between kernel threads
and userspace processes that are started before the policy is first
loaded - both get the label corresponding to the kernel SID. The only
way a process that persists from early boot can get a meaningful label
is by doing a voluntary dyntransition or re-executing itself.
Reusing the kernel label for userspace processes is problematic for
several reasons:
1. The kernel is considered to be a privileged domain and generally
needs to have a wide range of permissions allowed to work correctly,
which prevents the policy writer from effectively hardening against
early boot processes that might remain running unintentionally after
the policy is loaded (they represent a potential extra attack surface
that should be mitigated).
2. Despite the kernel being treated as a privileged domain, the policy
writer may want to impose certain special limitations on kernel
threads that may conflict with the requirements of intentional early
boot processes. For example, it is a good hardening practice to limit
what executables the kernel can execute as usermode helpers and to
confine the resulting usermode helper processes. However, a
(legitimate) process surviving from early boot may need to execute a
different set of executables.
3. As currently implemented, overlayfs remembers the security context of
the process that created an overlayfs mount and uses it to bound
subsequent operations on files using this context. If an overlayfs
mount is created before the SELinux policy is loaded, these "mounter"
checks are made against the kernel context, which may clash with
restrictions on the kernel domain (see 2.).
To resolve this, introduce a new initial SID (reusing the slot of the
former "init" initial SID) that will be assigned to any userspace
process started before the policy is first loaded. This is easy to do,
as we can simply label any process that goes through the
bprm_creds_for_exec LSM hook with the new init-SID instead of
propagating the kernel SID from the parent.
To provide backwards compatibility for existing policies that are
unaware of this new semantic of the "init" initial SID, introduce a new
policy capability "userspace_initial_context" and set the "init" SID to
the same context as the "kernel" SID unless this capability is set by
the policy.
Another small backwards compatibility measure is needed in
security_sid_to_context_core() for before the initial SELinux policy
load - see the code comment for explanation.
Signed-off-by: Ondrej Mosnacek <omosnace@redhat.com>
Reviewed-by: Stephen Smalley <stephen.smalley.work@gmail.com>
[PM: edited comments based on feedback/discussion]
Signed-off-by: Paul Moore <paul@paul-moore.com>
This commit reverts 5b0eea835d ("selinux: introduce an initial SID
for early boot processes") as it was found to cause problems on
distros with old SELinux userspace tools/libraries, specifically
Ubuntu 16.04.
Hopefully we will be able to re-add this functionality at a later
date, but let's revert this for now to help ensure a stable and
backwards compatible SELinux tree.
Link: https://lore.kernel.org/selinux/87edkseqf8.fsf@mail.lhotse
Acked-by: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>
Currently, SELinux doesn't allow distinguishing between kernel threads
and userspace processes that are started before the policy is first
loaded - both get the label corresponding to the kernel SID. The only
way a process that persists from early boot can get a meaningful label
is by doing a voluntary dyntransition or re-executing itself.
Reusing the kernel label for userspace processes is problematic for
several reasons:
1. The kernel is considered to be a privileged domain and generally
needs to have a wide range of permissions allowed to work correctly,
which prevents the policy writer from effectively hardening against
early boot processes that might remain running unintentionally after
the policy is loaded (they represent a potential extra attack surface
that should be mitigated).
2. Despite the kernel being treated as a privileged domain, the policy
writer may want to impose certain special limitations on kernel
threads that may conflict with the requirements of intentional early
boot processes. For example, it is a good hardening practice to limit
what executables the kernel can execute as usermode helpers and to
confine the resulting usermode helper processes. However, a
(legitimate) process surviving from early boot may need to execute a
different set of executables.
3. As currently implemented, overlayfs remembers the security context of
the process that created an overlayfs mount and uses it to bound
subsequent operations on files using this context. If an overlayfs
mount is created before the SELinux policy is loaded, these "mounter"
checks are made against the kernel context, which may clash with
restrictions on the kernel domain (see 2.).
To resolve this, introduce a new initial SID (reusing the slot of the
former "init" initial SID) that will be assigned to any userspace
process started before the policy is first loaded. This is easy to do,
as we can simply label any process that goes through the
bprm_creds_for_exec LSM hook with the new init-SID instead of
propagating the kernel SID from the parent.
To provide backwards compatibility for existing policies that are
unaware of this new semantic of the "init" initial SID, introduce a new
policy capability "userspace_initial_context" and set the "init" SID to
the same context as the "kernel" SID unless this capability is set by
the policy.
Signed-off-by: Ondrej Mosnacek <omosnace@redhat.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>
The arrays for the policy capability names, the initial sid identifiers
and the class and permission names are not changed at runtime. Declare
them const to avoid accidental modification.
Do not override the classmap and the initial sid list in the build time
script genheaders.
Check flose(3) is successful in genheaders.c, otherwise the written data
might be corrupted or incomplete.
Signed-off-by: Christian Göttsche <cgzones@googlemail.com>
[PM: manual merge due to fuzz, minor style tweaks]
Signed-off-by: Paul Moore <paul@paul-moore.com>
The SELinux policy capability enum names are rather long and follow
the "POLICYDB_CAPABILITY_XXX format". While the "POLICYDB_" prefix
is helpful in tying the enums to other SELinux policy constants,
macros, etc. there is no reason why we need to spell out
"CAPABILITY" completely. Shorten "CAPABILITY" to "CAP" in order to
make things a bit shorter and cleaner.
Moving forward, the SELinux policy capability enum names should
follow the "POLICYDB_CAP_XXX" format.
Signed-off-by: Paul Moore <paul@paul-moore.com>
These ioctls are equivalent to fcntl(fd, F_SETFD, flags), which SELinux
always allows too. Furthermore, a failed FIOCLEX could result in a file
descriptor being leaked to a process that should not have access to it.
As this patch removes access controls, a policy capability needs to be
enabled in policy to always allow these ioctls.
Based-on-patch-by: Demi Marie Obenour <demiobenour@gmail.com>
Signed-off-by: Richard Haines <richard_c_haines@btinternet.com>
[PM: subject line tweak]
Signed-off-by: Paul Moore <paul@paul-moore.com>
Presently mdp does not enable any SELinux policy capabilities
in the dummy policy it generates. Thus, policies derived from
it will by default lack various features commonly used in modern
policies such as open permission, extended socket classes, network
peer controls, etc. Split the policy capability definitions out into
their own headers so that we can include them into mdp without pulling in
other kernel headers and extend mdp generate policycap statements for the
policy capabilities known to the kernel. Policy authors may wish to
selectively remove some of these from the generated policy.
Signed-off-by: Stephen Smalley <stephen.smalley.work@gmail.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>
