===========================================
Seccomp BPF (SECure COMPuting with filters)
===========================================

Introduction
============

A large number of system calls are exposed to every userland process
with many of them going unused for the entire lifetime of the process.
As system calls change and mature, bugs are found and eradicated.  A
certain subset of userland applications benefit by having a reduced set
of available system calls.  The resulting set reduces the total kernel
surface exposed to the application.  System call filtering is meant for
use with those applications.

Seccomp filtering provides a means for a process to specify a filter for
incoming system calls.  The filter is expressed as a Berkeley Packet
Filter (BPF) program, as with socket filters, except that the data
operated on is related to the system call being made: system call
number and the system call arguments.  This allows for expressive
filtering of system calls using a filter program language with a long
history of being exposed to userland and a straightforward data set.

Additionally, BPF makes it impossible for users of seccomp to fall prey
to time-of-check-time-of-use (TOCTOU) attacks that are common in system
call interposition frameworks.  BPF programs may not dereference
pointers which constrains all filters to solely evaluating the system
call arguments directly.

What it isn't
=============

System call filtering isn't a sandbox.  It provides a clearly defined
mechanism for minimizing the exposed kernel surface.  It is meant to be
a tool for sandbox developers to use.  Beyond that, policy for logical
behavior and information flow should be managed with a combination of
other system hardening techniques and, potentially, an LSM of your
choosing.  Expressive, dynamic filters provide further options down this
path (avoiding pathological sizes or selecting which of the multiplexed
system calls in socketcall() is allowed, for instance) which could be
construed, incorrectly, as a more complete sandboxing solution.

Usage
=====

An additional seccomp mode is added and is enabled using the same
prctl(2) call as the strict seccomp.  If the architecture has
``CONFIG_HAVE_ARCH_SECCOMP_FILTER``, then filters may be added as below:

``PR_SET_SECCOMP``:
	Now takes an additional argument which specifies a new filter
	using a BPF program.
	The BPF program will be executed over struct seccomp_data
	reflecting the system call number, arguments, and other
	metadata.  The BPF program must then return one of the
	acceptable values to inform the kernel which action should be
	taken.

	Usage::

		prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, prog);

	The 'prog' argument is a pointer to a struct sock_fprog which
	will contain the filter program.  If the program is invalid, the
	call will return -1 and set errno to ``EINVAL``.

	If ``fork``/``clone`` and ``execve`` are allowed by @prog, any child
	processes will be constrained to the same filters and system
	call ABI as the parent.

	Prior to use, the task must call ``prctl(PR_SET_NO_NEW_PRIVS, 1)`` or
	run with ``CAP_SYS_ADMIN`` privileges in its namespace.  If these are not
	true, ``-EACCES`` will be returned.  This requirement ensures that filter
	programs cannot be applied to child processes with greater privileges
	than the task that installed them.

	Additionally, if ``prctl(2)`` is allowed by the attached filter,
	additional filters may be layered on which will increase evaluation
	time, but allow for further decreasing the attack surface during
	execution of a process.

The above call returns 0 on success and non-zero on error.

Return values
=============

A seccomp filter may return any of the following values. If multiple
filters exist, the return value for the evaluation of a given system
call will always use the highest precedent value. (For example,
``SECCOMP_RET_KILL_PROCESS`` will always take precedence.)

In precedence order, they are:

``SECCOMP_RET_KILL_PROCESS``:
	Results in the entire process exiting immediately without executing
	the system call.  The exit status of the task (``status & 0x7f``)
	will be ``SIGSYS``, not ``SIGKILL``.

``SECCOMP_RET_KILL_THREAD``:
	Results in the task exiting immediately without executing the
	system call.  The exit status of the task (``status & 0x7f``) will
	be ``SIGSYS``, not ``SIGKILL``.

``SECCOMP_RET_TRAP``:
	Results in the kernel sending a ``SIGSYS`` signal to the triggering
	task without executing the system call. ``siginfo->si_call_addr``
	will show the address of the system call instruction, and
	``siginfo->si_syscall`` and ``siginfo->si_arch`` will indicate which
	syscall was attempted.  The program counter will be as though
	the syscall happened (i.e. it will not point to the syscall
	instruction).  The return value register will contain an arch-
	dependent value -- if resuming execution, set it to something
	sensible.  (The architecture dependency is because replacing
	it with ``-ENOSYS`` could overwrite some useful information.)

	The ``SECCOMP_RET_DATA`` portion of the return value will be passed
	as ``si_errno``.

	``SIGSYS`` triggered by seccomp will have a si_code of ``SYS_SECCOMP``.

``SECCOMP_RET_ERRNO``:
	Results in the lower 16-bits of the return value being passed
	to userland as the errno without executing the system call.

``SECCOMP_RET_USER_NOTIF``:
	Results in a ``struct seccomp_notif`` message sent on the userspace
	notification fd, if it is attached, or ``-ENOSYS`` if it is not. See
	below on discussion of how to handle user notifications.

``SECCOMP_RET_TRACE``:
	When returned, this value will cause the kernel to attempt to
	notify a ``ptrace()``-based tracer prior to executing the system
	call.  If there is no tracer present, ``-ENOSYS`` is returned to
	userland and the system call is not executed.

	A tracer will be notified if it requests ``PTRACE_O_TRACESECCOMP``
	using ``ptrace(PTRACE_SETOPTIONS)``.  The tracer will be notified
	of a ``PTRACE_EVENT_SECCOMP`` and the ``SECCOMP_RET_DATA`` portion of
	the BPF program return value will be available to the tracer
	via ``PTRACE_GETEVENTMSG``.

	The tracer can skip the system call by changing the syscall number
	to -1.  Alternatively, the tracer can change the system call
	requested by changing the system call to a valid syscall number.  If
	the tracer asks to skip the system call, then the system call will
	appear to return the value that the tracer puts in the return value
	register.

	The seccomp check will not be run again after the tracer is
	notified.  (This means that seccomp-based sandboxes MUST NOT
	allow use of ptrace, even of other sandboxed processes, without
	extreme care; ptracers can use this mechanism to escape.)

``SECCOMP_RET_LOG``:
	Results in the system call being executed after it is logged. This
	should be used by application developers to learn which syscalls their
	application needs without having to iterate through multiple test and
	development cycles to build the list.

	This action will only be logged if "log" is present in the
	actions_logged sysctl string.

``SECCOMP_RET_ALLOW``:
	Results in the system call being executed.

If multiple filters exist, the return value for the evaluation of a
given system call will always use the highest precedent value.

Precedence is only determined using the ``SECCOMP_RET_ACTION`` mask.  When
multiple filters return values of the same precedence, only the
``SECCOMP_RET_DATA`` from the most recently installed filter will be
returned.

Pitfalls
========

The biggest pitfall to avoid during use is filtering on system call
number without checking the architecture value.  Why?  On any
architecture that supports multiple system call invocation conventions,
the system call numbers may vary based on the specific invocation.  If
the numbers in the different calling conventions overlap, then checks in
the filters may be abused.  Always check the arch value!

Example
=======

The ``samples/seccomp/`` directory contains both an x86-specific example
and a more generic example of a higher level macro interface for BPF
program generation.

Userspace Notification
======================

The ``SECCOMP_RET_USER_NOTIF`` return code lets seccomp filters pass a
particular syscall to userspace to be handled. This may be useful for
applications like container managers, which wish to intercept particular
syscalls (``mount()``, ``finit_module()``, etc.) and change their behavior.

To acquire a notification FD, use the ``SECCOMP_FILTER_FLAG_NEW_LISTENER``
argument to the ``seccomp()`` syscall:

.. code-block:: c

    fd = seccomp(SECCOMP_SET_MODE_FILTER, SECCOMP_FILTER_FLAG_NEW_LISTENER, &prog);

which (on success) will return a listener fd for the filter, which can then be
passed around via ``SCM_RIGHTS`` or similar. Note that filter fds correspond to
a particular filter, and not a particular task. So if this task then forks,
notifications from both tasks will appear on the same filter fd. Reads and
writes to/from a filter fd are also synchronized, so a filter fd can safely
have many readers.

The interface for a seccomp notification fd consists of two structures:

.. code-block:: c

    struct seccomp_notif_sizes {
        __u16 seccomp_notif;
        __u16 seccomp_notif_resp;
        __u16 seccomp_data;
    };

    struct seccomp_notif {
        __u64 id;
        __u32 pid;
        __u32 flags;
        struct seccomp_data data;
    };

    struct seccomp_notif_resp {
        __u64 id;
        __s64 val;
        __s32 error;
        __u32 flags;
    };

The ``struct seccomp_notif_sizes`` structure can be used to determine the size
of the various structures used in seccomp notifications. The size of ``struct
seccomp_data`` may change in the future, so code should use:

.. code-block:: c

    struct seccomp_notif_sizes sizes;
    seccomp(SECCOMP_GET_NOTIF_SIZES, 0, &sizes);

to determine the size of the various structures to allocate. See
samples/seccomp/user-trap.c for an example.

Users can read via ``ioctl(SECCOMP_IOCTL_NOTIF_RECV)``  (or ``poll()``) on a
seccomp notification fd to receive a ``struct seccomp_notif``, which contains
five members: the input length of the structure, a unique-per-filter ``id``,
the ``pid`` of the task which triggered this request (which may be 0 if the
task is in a pid ns not visible from the listener's pid namespace). The
notification also contains the ``data`` passed to seccomp, and a filters flag.
The structure should be zeroed out prior to calling the ioctl.

Userspace can then make a decision based on this information about what to do,
and ``ioctl(SECCOMP_IOCTL_NOTIF_SEND)`` a response, indicating what should be
returned to userspace. The ``id`` member of ``struct seccomp_notif_resp`` should
be the same ``id`` as in ``struct seccomp_notif``.

Userspace can also add file descriptors to the notifying process via
``ioctl(SECCOMP_IOCTL_NOTIF_ADDFD)``. The ``id`` member of
``struct seccomp_notif_addfd`` should be the same ``id`` as in
``struct seccomp_notif``. The ``newfd_flags`` flag may be used to set flags
like O_CLOEXEC on the file descriptor in the notifying process. If the supervisor
wants to inject the file descriptor with a specific number, the
``SECCOMP_ADDFD_FLAG_SETFD`` flag can be used, and set the ``newfd`` member to
the specific number to use. If that file descriptor is already open in the
notifying process it will be replaced. The supervisor can also add an FD, and
respond atomically by using the ``SECCOMP_ADDFD_FLAG_SEND`` flag and the return
value will be the injected file descriptor number.

The notifying process can be preempted, resulting in the notification being
aborted. This can be problematic when trying to take actions on behalf of the
notifying process that are long-running and typically retryable (mounting a
filesytem). Alternatively, at filter installation time, the
``SECCOMP_FILTER_FLAG_WAIT_KILLABLE_RECV`` flag can be set. This flag makes it
such that when a user notification is received by the supervisor, the notifying
process will ignore non-fatal signals until the response is sent. Signals that
are sent prior to the notification being received by userspace are handled
normally.

It is worth noting that ``struct seccomp_data`` contains the values of register
arguments to the syscall, but does not contain pointers to memory. The task's
memory is accessible to suitably privileged traces via ``ptrace()`` or
``/proc/pid/mem``. However, care should be taken to avoid the TOCTOU mentioned
above in this document: all arguments being read from the tracee's memory
should be read into the tracer's memory before any policy decisions are made.
This allows for an atomic decision on syscall arguments.

Sysctls
=======

Seccomp's sysctl files can be found in the ``/proc/sys/kernel/seccomp/``
directory. Here's a description of each file in that directory:

``actions_avail``:
	A read-only ordered list of seccomp return values (refer to the
	``SECCOMP_RET_*`` macros above) in string form. The ordering, from
	left-to-right, is the least permissive return value to the most
	permissive return value.

	The list represents the set of seccomp return values supported
	by the kernel. A userspace program may use this list to
	determine if the actions found in the ``seccomp.h``, when the
	program was built, differs from the set of actions actually
	supported in the current running kernel.

``actions_logged``:
	A read-write ordered list of seccomp return values (refer to the
	``SECCOMP_RET_*`` macros above) that are allowed to be logged. Writes
	to the file do not need to be in ordered form but reads from the file
	will be ordered in the same way as the actions_avail sysctl.

	The ``allow`` string is not accepted in the ``actions_logged`` sysctl
	as it is not possible to log ``SECCOMP_RET_ALLOW`` actions. Attempting
	to write ``allow`` to the sysctl will result in an EINVAL being
	returned.

Adding architecture support
===========================

See ``arch/Kconfig`` for the authoritative requirements.  In general, if an
architecture supports both ptrace_event and seccomp, it will be able to
support seccomp filter with minor fixup: ``SIGSYS`` support and seccomp return
value checking.  Then it must just add ``CONFIG_HAVE_ARCH_SECCOMP_FILTER``
to its arch-specific Kconfig.



Caveats
=======

The vDSO can cause some system calls to run entirely in userspace,
leading to surprises when you run programs on different machines that
fall back to real syscalls.  To minimize these surprises on x86, make
sure you test with
``/sys/devices/system/clocksource/clocksource0/current_clocksource`` set to
something like ``acpi_pm``.

On x86-64, vsyscall emulation is enabled by default.  (vsyscalls are
legacy variants on vDSO calls.)  Currently, emulated vsyscalls will
honor seccomp, with a few oddities:

- A return value of ``SECCOMP_RET_TRAP`` will set a ``si_call_addr`` pointing to
  the vsyscall entry for the given call and not the address after the
  'syscall' instruction.  Any code which wants to restart the call
  should be aware that (a) a ret instruction has been emulated and (b)
  trying to resume the syscall will again trigger the standard vsyscall
  emulation security checks, making resuming the syscall mostly
  pointless.

- A return value of ``SECCOMP_RET_TRACE`` will signal the tracer as usual,
  but the syscall may not be changed to another system call using the
  orig_rax register. It may only be changed to -1 order to skip the
  currently emulated call. Any other change MAY terminate the process.
  The rip value seen by the tracer will be the syscall entry address;
  this is different from normal behavior.  The tracer MUST NOT modify
  rip or rsp.  (Do not rely on other changes terminating the process.
  They might work.  For example, on some kernels, choosing a syscall
  that only exists in future kernels will be correctly emulated (by
  returning ``-ENOSYS``).

To detect this quirky behavior, check for ``addr & ~0x0C00 ==
0xFFFFFFFFFF600000``.  (For ``SECCOMP_RET_TRACE``, use rip.  For
``SECCOMP_RET_TRAP``, use ``siginfo->si_call_addr``.)  Do not check any other
condition: future kernels may improve vsyscall emulation and current
kernels in vsyscall=native mode will behave differently, but the
instructions at ``0xF...F600{0,4,8,C}00`` will not be system calls in these
cases.

Note that modern systems are unlikely to use vsyscalls at all -- they
are a legacy feature and they are considerably slower than standard
syscalls.  New code will use the vDSO, and vDSO-issued system calls
are indistinguishable from normal system calls.