selftests/bpf: Reject graph kptr use after RCU unlockgrafted

Add a sleepable verifier test that loads a graph-node local kptr in an
explicit RCU read-side critical section, then passes its node to
bpf_rbtree_remove() after the section ends.

Before the verifier fix, the stale NON_OWN_REF flag makes the node look like
a live borrowed reference and the program is accepted. After the fix, the
pointer is demoted without NON_OWN_REF and the graph kfunc argument is
rejected.

Also exercise a graph kptr loaded while a spin lock provides implicit RCU
protection. The pointer must be invalidated when the lock is released, which
guards the required ordering between non-owning-reference invalidation and
RCU demotion.

Update the existing fault-protected load test state description. The
post-unlock pointer no longer carries NON_OWN_REF, but remains readable
because the load is rewritten to use BPF_PROBE_MEM.

Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20260904084325.52250-7-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

1 files changed, 118 insertions, 0 deletions

diff --git a/Documentation/admin-guide/LSM/SafeSetID.rst b/Documentation/admin-guide/LSM/SafeSetID.rst
new file mode 100644
index 000000000..6d439c987
--- /dev/null
+++ b/Documentation/admin-guide/LSM/SafeSetID.rst
@@ -0,0 +1,118 @@
+=========
+SafeSetID
+=========
+SafeSetID is an LSM module that gates the setid family of syscalls to restrict
+UID/GID transitions from a given UID/GID to only those approved by a
+system-wide allowlist. These restrictions also prohibit the given UIDs/GIDs
+from obtaining auxiliary privileges associated with CAP_SET{U/G}ID, such as
+allowing a user to set up user namespace UID/GID mappings.
+
+
+Background
+==========
+In absence of file capabilities, processes spawned on a Linux system that need
+to switch to a different user must be spawned with CAP_SETUID privileges.
+CAP_SETUID is granted to programs running as root or those running as a non-root
+user that have been explicitly given the CAP_SETUID runtime capability. It is
+often preferable to use Linux runtime capabilities rather than file
+capabilities, since using file capabilities to run a program with elevated
+privileges opens up possible security holes since any user with access to the
+file can exec() that program to gain the elevated privileges.
+
+While it is possible to implement a tree of processes by giving full
+CAP_SET{U/G}ID capabilities, this is often at odds with the goals of running a
+tree of processes under non-root user(s) in the first place. Specifically,
+since CAP_SETUID allows changing to any user on the system, including the root
+user, it is an overpowered capability for what is needed in this scenario,
+especially since programs often only call setuid() to drop privileges to a
+lesser-privileged user -- not elevate privileges. Unfortunately, there is no
+generally feasible way in Linux to restrict the potential UIDs that a user can
+switch to through setuid() beyond allowing a switch to any user on the system.
+This SafeSetID LSM seeks to provide a solution for restricting setid
+capabilities in such a way.
+
+The main use case for this LSM is to allow a non-root program to transition to
+other untrusted uids without full blown CAP_SETUID capabilities. The non-root
+program would still need CAP_SETUID to do any kind of transition, but the
+additional restrictions imposed by this LSM would mean it is a "safer" version
+of CAP_SETUID since the non-root program cannot take advantage of CAP_SETUID to
+do any unapproved actions (e.g. setuid to uid 0 or create/enter new user
+namespace). The higher level goal is to allow for uid-based sandboxing of system
+services without having to give out CAP_SETUID all over the place just so that
+non-root programs can drop to even-lesser-privileged uids. This is especially
+relevant when one non-root daemon on the system should be allowed to spawn other
+processes as different uids, but it's undesirable to give the daemon a
+basically-root-equivalent CAP_SETUID.
+
+
+Other Approaches Considered
+===========================
+
+Solve this problem in userspace
+-------------------------------
+For candidate applications that would like to have restricted setid capabilities
+as implemented in this LSM, an alternative option would be to simply take away
+setid capabilities from the application completely and refactor the process
+spawning semantics in the application (e.g. by using a privileged helper program
+to do process spawning and UID/GID transitions). Unfortunately, there are a
+number of semantics around process spawning that would be affected by this, such
+as fork() calls where the program doesn't immediately call exec() after the
+fork(), parent processes specifying custom environment variables or command line
+args for spawned child processes, or inheritance of file handles across a
+fork()/exec(). Because of this, as solution that uses a privileged helper in
+userspace would likely be less appealing to incorporate into existing projects
+that rely on certain process-spawning semantics in Linux.
+
+Use user namespaces
+-------------------
+Another possible approach would be to run a given process tree in its own user
+namespace and give programs in the tree setid capabilities. In this way,
+programs in the tree could change to any desired UID/GID in the context of their
+own user namespace, and only approved UIDs/GIDs could be mapped back to the
+initial system user namespace, affectively preventing privilege escalation.
+Unfortunately, it is not generally feasible to use user namespaces in isolation,
+without pairing them with other namespace types, which is not always an option.
+Linux checks for capabilities based off of the user namespace that "owns" some
+entity. For example, Linux has the notion that network namespaces are owned by
+the user namespace in which they were created. A consequence of this is that
+capability checks for access to a given network namespace are done by checking
+whether a task has the given capability in the context of the user namespace
+that owns the network namespace -- not necessarily the user namespace under
+which the given task runs. Therefore spawning a process in a new user namespace
+effectively prevents it from accessing the network namespace owned by the
+initial namespace. This is a deal-breaker for any application that expects to
+retain the CAP_NET_ADMIN capability for the purpose of adjusting network
+configurations. Using user namespaces in isolation causes problems regarding
+other system interactions, including use of pid namespaces and device creation.
+
+Use an existing LSM
+-------------------
+None of the other in-tree LSMs have the capability to gate setid transitions, or
+even employ the security_task_fix_setuid hook at all. SELinux says of that hook:
+"Since setuid only affects the current process, and since the SELinux controls
+are not based on the Linux identity attributes, SELinux does not need to control
+this operation."
+
+
+Directions for use
+==================
+This LSM hooks the setid syscalls to make sure transitions are allowed if an
+applicable restriction policy is in place. Policies are configured through
+securityfs by writing to the safesetid/uid_allowlist_policy and
+safesetid/gid_allowlist_policy files at the location where securityfs is
+mounted. The format for adding a policy is '<UID>:<UID>' or '<GID>:<GID>',
+using literal numbers, and ending with a newline character such as '123:456\n'.
+Writing an empty string "" will flush the policy. Again, configuring a policy
+for a UID/GID will prevent that UID/GID from obtaining auxiliary setid
+privileges, such as allowing a user to set up user namespace UID/GID mappings.
+
+Note on GID policies and setgroups()
+====================================
+In v5.9 we are adding support for limiting CAP_SETGID privileges as was done
+previously for CAP_SETUID. However, for compatibility with common sandboxing
+related code conventions in userspace, we currently allow arbitrary
+setgroups() calls for processes with CAP_SETGID restrictions. Until we add
+support in a future release for restricting setgroups() calls, these GID
+policies add no meaningful security. setgroups() restrictions will be enforced
+once we have the policy checking code in place, which will rely on GID policy
+configuration code added in v5.9.