Re: [PATCH bpf-next v1 00/13] MAC and Audit policy using eBPF (KRSI)

[Andrii Nakryiko]

On Mon, Dec 30, 2019 at 6:59 AM KP Singh <kpsingh@xxxxxxxxxxxx> wrote:
>
> On 21-Dez 17:27, Alexei Starovoitov wrote:
> > On Fri, Dec 20, 2019 at 04:41:55PM +0100, KP Singh wrote:
> > > // Declare the eBPF program mprotect_audit which attaches to
> > > // to the file_mprotect LSM hook and accepts three arguments.
> > > BPF_TRACE_3("lsm/file_mprotect", mprotect_audit,
> > >         struct vm_area_struct *, vma,
> > >         unsigned long, reqprot, unsigned long, prot
> > > {
> > >     unsigned long vm_start = _(vma->vm_start);
> > >     return 0;
> > > }
> >
>
> Hi Alexei,
>
> Thanks for the feedback. This is really helpful!
>
> > I think the only sore point of the patchset is:
> > security/bpf/include/hooks.h   | 1015 ++++++++++++++++++++++++++++++++
> > With bpf trampoline this type of 'kernel types -> bpf types' converters
> > are no longer necessary. Please take a look at tcp-congestion-control patchset:
> > https://patchwork.ozlabs.org/cover/1214417/
> > Instead of doing similar thing (like your patch 1 plus patch 6) it's using
> > trampoline to provide bpf based congestion control callbacks into tcp stack.
> > The same trampoline-based mechanism can be reused by bpf_lsm.
> > Then all manual work of doing BPF_LSM_HOOK(...) for every hook won't be
> > necessary. It will also prove the point that attaching BPF to raw LSM hooks
> > doesn't freeze them into stable abi.
>
> Really cool!
>
> I looked into how BPF trampolines are being used in tracing and the
> new STRUCT_OPS patchset and was able protoype
> (https://github.com/sinkap/linux-krsi/tree/patch/v1/trampoline_prototype,
> not ready for review yet) which:
>
> * Gets rid of security/bpf/include/hooks.h and all of the static
>   macro magic essentially making the LSM ~truly instrumentable~ at
>   runtime.
> * Gets rid of the generation of any new types as we already have
>   all the BTF information in the kernel in the following two types:
>
> struct security_hook_heads {
>         .
>         .
>         struct hlist_head file_mprotect;   <- Append the callback at this offset
>         .
>         .
> };
>
> and
>
> union security_list_options {
>         int (*file_mprotect)(struct vm_area_struct *vma, unsigned long reqprot,
>                                 unsigned long prot);
> };
>
> Which is the same type as the typedef that's currently being generated
> , i.e. lsm_btf_file_mprotect
>
> In the current prototype, libbpf converts the name of the hook into an
> offset into the security_hook_heads and the verifier does the
> following when a program is loaded:
>
> * Verifies the offset and the type at the offset (struct hlist_head).
> * Resolves the func_proto (by looking up the type in
>   security_list_options) and updates prog->aux with the name and
>   func_proto which are then verified similar to raw_tp programs with
>   btf_ctx_access.
>
> On attachment:
>
> * A trampoline is created and appended to the security_hook_heads
>   for the BPF LSM.
> * An anonymous FD is returned and the attachment is conditional on the
>   references to FD (as suggested and similar to fentry/fexit tracing
>   programs).
>
> This implies that the BPF programs are "the LSM hook" as opposed to
> being executed inside a statically defined hook body which requires
> mutable LSM hooks for which I was able to re-use some of ideas in
> Sargun's patch:
>
> https://lore.kernel.org/lkml/20180408065916.GA2832@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>
> to maintain a separate security_hook_heads struct for dynamically
> added LSM hooks by the BPF LSM which are executed after all the
> statically defined hooks.
>
> > Longer program names are supplied via btf's func_info.
> > It feels that:
> > cat /sys/kernel/security/bpf/process_execution
> > env_dumper__v2
> > is reinventing the wheel. bpftool is the main introspection tool.
> > It can print progs attached to perf, cgroup, networking. I think it's better to
> > stay consistent and do the same with bpf-lsm.
>
> I agree, based on the new feedback, I don't think we need securityFS
> attachment points anymore. I was able to get rid of it completely.
>
> >
> > Another issue is in proposed attaching method:
> > hook_fd = open("/sys/kernel/security/bpf/process_execution");
> > sys_bpf(attach, prog_fd, hook_fd);
> > With bpf tracing we moved to FD-based attaching, because permanent attaching is
> > problematic in production. We're going to provide FD-based api to attach to
> > networking as well, because xdp/tc/cgroup prog attaching suffers from the same
> > production issues. Mainly with permanent attaching there is no ownership of
> > attachment. Everything is global and permanent. It's not clear what
> > process/script suppose to detach/cleanup. I suggest bpf-lsm use FD-based
> > attaching from the beginning. Take a look at raw_tp/tp_btf/fentry/fexit style
> > of attaching. All of them return FD which represents what libbpf calls
> > 'bpf_link' concept. Once refcnt of that FD goes to zero that link (attachment)
> > is destroyed and program is detached _by the kernel_. To make such links
> > permanent the application can pin them in bpffs. The pinning patches haven't
> > landed yet, but the concept of the link is quite powerful and much more
> > production friendly than permanent attaching.
>
> I like this. This also means we don't immediately need the handling of
> duplicate names so I dropped that bit of the patch as well and updated
> the attachment to use this mechanism.
>
> > bpf-lsm will still be able to attach multiple progs to the same hook and
> > see what is attached via bpftool.
> >
> > The rest looks good. Thank you for working on it.
>
> There are some choices we need to make here from an API perspective:
>
> * Should we "repurpose" attr->attach_btf_id and use it as an offset
>   into security_hook_heads or add a new attribute
>   (e.g lsm_hook_offset) for the offset or use name of the LSM hook
>   (e.g. lsm_hook_name).

I think setting this to member index inside union
security_list_options will be better? Or member index inside struct
security_hook_heads. Seems like kernel will have to "join" those two
anyways, right (one for type info, another for trampoline)? Offset is
less convenient either way.

> * Since we don't have the files in securityFS, the attachment does not
>   have a target_fd. Should we add a new type of BPF command?
>   e.g. LSM_HOOK_OPEN?

Semantics of LSM program seems closer to fentry/fexit/raw_tp, so maybe
instead use BPF_RAW_TRACEPOINT_OPEN command? On libbpf side it's all
going to be abstracted behind bpf_program__attach() anyways.

>
> I will clean up the prototype, incorporate some of the other feedback
> received, and send a v2.
>
> Wishing everyone a very Happy New Year!

Thanks, you too!

>
> - KP
>