lkml.org 
[lkml]   [2020]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Date
SubjectRe: [PATCH bpf-next v7 8/8] bpf: lsm: Add Documentation
Thanks for the reviews!

On 26-Mär 12:31, Andrii Nakryiko wrote:
> On Thu, Mar 26, 2020 at 7:29 AM KP Singh <kpsingh@chromium.org> wrote:
> >
> > From: KP Singh <kpsingh@google.com>
> >
> > Document how eBPF programs (BPF_PROG_TYPE_LSM) can be loaded and
> > attached (BPF_LSM_MAC) to the LSM hooks.
> >
> > Signed-off-by: KP Singh <kpsingh@google.com>
> > Reviewed-by: Brendan Jackman <jackmanb@google.com>
> > Reviewed-by: Florent Revest <revest@google.com>
> > Reviewed-by: Thomas Garnier <thgarnie@google.com>
> > ---
>
> This needs another pass and re-reading, has a bunch of outdated info :)

Indeed :)

>
> > Documentation/bpf/bpf_lsm.rst | 150 ++++++++++++++++++++++++++++++++++
> > Documentation/bpf/index.rst | 1 +
> > 2 files changed, 151 insertions(+)
> > create mode 100644 Documentation/bpf/bpf_lsm.rst
> >
> > diff --git a/Documentation/bpf/bpf_lsm.rst b/Documentation/bpf/bpf_lsm.rst
> > new file mode 100644
> > index 000000000000..2a2c3b4a74d4
> > --- /dev/null
> > +++ b/Documentation/bpf/bpf_lsm.rst
> > @@ -0,0 +1,150 @@
> > +.. SPDX-License-Identifier: GPL-2.0+
> > +.. Copyright (C) 2020 Google LLC.
> > +
> > +================
> > +LSM BPF Programs
> > +================
> > +
> > +These BPF programs allow runtime instrumentation of the LSM hooks by privileged
> > +users to implement system-wide MAC (Mandatory Access Control) and Audit
> > +policies using eBPF. Since these program end up modifying the MAC policies of
> > +the system, they require both ``CAP_MAC_ADMIN`` and also require
> > +``CAP_SYS_ADMIN`` for the loading of BPF programs.
> > +
> > +Structure
> > +---------
> > +
> > +The example shows an eBPF program that can be attached to the ``file_mprotect``
> > +LSM hook:
> > +
> > +.. c:function:: int file_mprotect(struct vm_area_struct *vma, unsigned long reqprot, unsigned long prot);
> > +
> > +Other LSM hooks which can be instrumented can be found in
> > +``include/linux/lsm_hooks.h``.
> > +
> > +eBPF programs that use :doc:`/bpf/btf` do not need to include kernel headers
> > +for accessing information from the attached eBPF program's context. They can
> > +simply declare the structures in the eBPF program and only specify the fields
> > +that need to be accessed.
> > +
> > +.. code-block:: c
> > +
> > + struct mm_struct {
> > + unsigned long start_brk, brk, start_stack;
> > + } __attribute__((preserve_access_index));
> > +
> > + struct vm_area_struct {
> > + unsigned long start_brk, brk, start_stack;
> > + unsigned long vm_start, vm_end;
> > + struct mm_struct *vm_mm;
> > + } __attribute__((preserve_access_index));
> > +
> > +
> > +.. note:: Only the size and the names of the fields must match the type in the
> > + kernel and the order of the fields is irrelevant.
>
> type should match/be compatible as well?

I changed it to simply be:

.. note:: The order of the fields is irrelevant.

>
> > +
> > +This can be further simplified (if one has access to the BTF information at
> > +build time) by generating the ``vmlinux.h`` with:
> > +
> > +.. code-block:: console
> > +
> > + # bpftool dump file <path-to-btf-vmlinux> format c > vmlinux.h
> > +
>
> bpftool btf *dump* file

Done.

>
> > +.. note:: ``path-to-btf-vmlinux`` can be ``/sys/kernel/btf/vmlinux`` if the
> > + build environment matches the environment the BPF programs are
> > + deployed in.
> > +
> > +The ``vmlinux.h`` can then simply be included in the BPF programs without
> > +requiring the definition of the types.
> > +
> > +The eBPF programs can be declared using the``BPF_PROG``
> > +macros defined in `tools/lib/bpf/bpf_tracing.h`_. In this
> > +example:
> > +
> > + * ``"lsm/file_mprotect"`` indicates the LSM hook that the program must
> > + be attached to
> > + * ``mprotect_audit`` is the name of the eBPF program
> > +
> > +.. code-block:: c
> > +
> > + SEC("lsm/file_mprotect")
> > + int BPF_PROG(mprotect_audit, struct vm_area_struct *vma,
> > + unsigned long reqprot, unsigned long prot, int ret)
> > + {
> > + /* Ret is the return value from the previous BPF program
> > + * or 0 if it's the first hook.
> > + */
> > + if (ret != 0)
> > + return ret;
> > +
> > + int is_heap;
> > +
> > + is_heap = (vma->vm_start >= vma->vm_mm->start_brk &&
> > + vma->vm_end <= vma->vm_mm->brk);
> > +
> > + /* Return an -EPERM or write information to the perf events buffer
> > + * for auditing
> > + */
>
> return missing?

Fixed.

>
> > + }
> > +
> > +The ``__attribute__((preserve_access_index))`` is a clang feature that allows
> > +the BPF verifier to update the offsets for the access at runtime using the
> > +:doc:`/bpf/btf` information. Since the BPF verifier is aware of the types, it
> > +also validates all the accesses made to the various types in the eBPF program.
> > +
> > +Loading
> > +-------
> > +
> > +eBPF programs can be loaded with the :manpage:`bpf(2)` syscall's
> > +``BPF_PROG_LOAD`` operation or more simply by using the the libbpf helper
> > +``bpf_prog_load_xattr``:
> > +
> > +
> > +.. code-block:: c
> > +
> > + struct bpf_prog_load_attr attr = {
> > + .file = "./prog.o",
> > + };
> > + struct bpf_object *prog_obj;
> > + struct bpf_program *prog;
> > + int prog_fd;
> > +
> > + bpf_prog_load_xattr(&attr, &prog_obj, &prog_fd);
>
> Can you please update this to not use deprecated/legacy APIs. Please
> suggest bpf_object__open/bpf_object__load and/or BPF skeleton as an
> example.


Simplified and modernized this section as:


Loading
-------

eBPF programs can be loaded with the :manpage:`bpf(2)` syscall's
``BPF_PROG_LOAD`` operation:

.. code-block:: c

struct bpf_object *obj;

obj = bpf_object__open("./my_prog.o");
bpf_object__load(obj);

This can be simplified by using a skeleton header generated by ``bpftool``:

.. code-block:: console

# bpftool gen skeleton my_prog.o > my_prog.skel.h

and the program can be loaded by including ``my_prog.skel.h`` and using
the generated helper, ``my_prog__open_and_load``.

Attachment to LSM Hooks
-----------------------

The LSM allows attachment of eBPF programs as LSM hooks using :manpage:`bpf(2)`
syscall's ``BPF_RAW_TRACEPOINT_OPEN`` operation or more simply by
using the libbpf helper ``bpf_program__attach_lsm``.

The program can be detached from the LSM hook by *destroying* the ``link``
link returned by ``bpf_program__attach_lsm`` using ``bpf_link__destroy``.

One can also use the helpers generated in ``my_prog.skel.h`` i.e.
``my_prog__attach`` for attachment and ``my_prog__destroy`` for cleaning up.

</end>

If this looks okay, I will send a v8 with this updated and other
fixes.

- KP

>
> > +
> > +Attachment to LSM Hooks
> > +-----------------------
> > +
> > +The LSM allows attachment of eBPF programs as LSM hooks using :manpage:`bpf(2)`
> > +syscall's ``BPF_PROG_ATTACH`` operation or more simply by
>
> BPF_PROG_ATTACH is incorrect, it's RAW_TRACEPOINT_OPEN, isn't it?

Correct, updated. Thanks!

>
> > +using the libbpf helper ``bpf_program__attach_lsm``. In the code shown below
> > +``prog`` is the eBPF program loaded using ``BPF_PROG_LOAD``:
> > +
> > +.. code-block:: c
> > +
> > + struct bpf_link *link;
> > +
> > + link = bpf_program__attach_lsm(prog);
> > +
> > +The program can be detached from the LSM hook by *destroying* the ``link``
> > +link returned by ``bpf_program__attach_lsm``:
> > +
> > +.. code-block:: c
> > +
> > + link->destroy();
>
> that's not how it works in C ;)

Oops, I incorrectly picked it up from link->destroy(link); and wrote
something stupid.

>
> bpf_link__destroy(link);

Updated in the snippet posted above.

- KP

>
> > +
> > +Examples
> > +--------
> > +
> > +An example eBPF programs can be found in
> > +`tools/testing/selftests/bpf/progs/lsm.c`_ and the corresponding
> > +userspace code in `tools/testing/selftests/bpf/prog_tests/test_lsm.c`_
> > +
> > +.. Links
> > +.. _tools/lib/bpf/bpf_tracing.h:
> > + https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/tools/lib/bpf/bpf_tracing.h
> > +.. _tools/testing/selftests/bpf/progs/lsm.c:
> > + https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/tools/testing/selftests/bpf/progs/lsm.c
> > +.. _tools/testing/selftests/bpf/progs/lsm_void_hook.c:
> > + https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/tools/testing/selftests/bpf/progs/lsm_void_hook.c
> > +.. _tools/testing/selftests/bpf/prog_tests/test_lsm.c:
> > + https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/tree/tools/testing/selftests/bpf/prog_tests/test_lsm.c
> > diff --git a/Documentation/bpf/index.rst b/Documentation/bpf/index.rst
> > index 7be43c5f2dcf..f99677f3572f 100644
> > --- a/Documentation/bpf/index.rst
> > +++ b/Documentation/bpf/index.rst
> > @@ -45,6 +45,7 @@ Program types
> > prog_cgroup_sockopt
> > prog_cgroup_sysctl
> > prog_flow_dissector
> > + bpf_lsm
> >
> >
> > Testing and debugging BPF
> > --
> > 2.20.1
> >

\
 
 \ /
  Last update: 2020-03-26 21:56    [W:0.070 / U:23.272 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site