[PATCH 4/4] arm64: ftrace: Add return address protection

[Mark Rutland]

On Thu, Dec 01, 2022 at 04:05:35PM +0100, Ard Biesheuvel wrote:
> On Thu, 1 Dec 2022 at 15:40, Mark Rutland <mark.rutland at arm.com> wrote:
> >
> > On Thu, Dec 01, 2022 at 02:09:41PM +0100, Ard Biesheuvel wrote:
> > > On Wed, 30 Nov 2022 at 18:45, Mark Rutland <mark.rutland at arm.com> wrote:
> > > >
> > > > On Wed, Nov 30, 2022 at 03:26:19PM +0100, Ard Biesheuvel wrote:
> > > > > On Wed, 30 Nov 2022 at 15:04, Mark Rutland <mark.rutland at arm.com> wrote:
> > > > > > On Tue, Nov 29, 2022 at 03:18:03PM +0100, Ard Biesheuvel wrote:
> > > > > > > Use the newly added asm macros to protect and restore the return address
> > > > > > > in the ftrace call wrappers, based on whichever method is active (PAC
> > > > > > > and/or shadow call stack).
> > > > > > >
> > > > > > > If the graph tracer is in use, this covers both the return address *to*
> > > > > > > the ftrace call site as well as the return address *at* the call site,
> > > > > > > and the latter will either be restored in return_to_handler(), or before
> > > > > > > returning to the call site.
> > > > > > >
> > > > > > > Signed-off-by: Ard Biesheuvel <ardb at kernel.org>
> > > > > > > ---
> > > > > > >  arch/arm64/kernel/entry-ftrace.S | 28 +++++++++++++++++++-
> > > > > > >  1 file changed, 27 insertions(+), 1 deletion(-)
> > > > > >
> > > > > > As a heads-up, this code has recently changed quite significantly, and this
> > > > > > won't apply to the version queued in arm64's for-next/{ftrace,core} branches.
> > > > > >
> > > > > > I had a direction of travel in mind with some changes for better stacktracing,
> > > > > > which won't work with the approach here, so I'd prefer we do this a bit
> > > > > > differently; more on that below.
> > > > > >
> > > > > > > diff --git a/arch/arm64/kernel/entry-ftrace.S b/arch/arm64/kernel/entry-ftrace.S
> > > > > > > index 795344ab4ec45889..c744e4dd8c90a352 100644
> > > > > > > --- a/arch/arm64/kernel/entry-ftrace.S
> > > > > > > +++ b/arch/arm64/kernel/entry-ftrace.S
> > > > > > > @@ -35,6 +35,11 @@
> > > > > > >   * is missing from the LR and existing chain of frame records.
> > > > > > >   */
> > > > > > >       .macro  ftrace_regs_entry, allregs=0
> > > > > > > +#ifdef CONFIG_FUNCTION_GRAPH_TRACER
> > > > > > > +     protect_return_address x9
> > > > > > > +#endif
> > > > > > > +     protect_return_address x30
> > > > > >
> > > > > > I think if we're going to protect the callsite's original LR (x9 here), we
> > > > > > should do that regardless of CONFIG_FUNCTION_GRAPH_TRACER; what matters is
> > > > > > whether that's vulnerable rather than whether we intend to modify it, so I
> > > > > > don't think it makes sene to protect it conditionally based on
> > > > > > CONFIG_FUNCTION_GRAPH_TRACER.
> > > > >
> > > > > My reasoning was that if we are not going to return from it (in
> > > > > return_to_handler()), we can rely on the interrupted function to
> > > > > sign/authenticate it as usual. So the only reason for signing it here
> > > > > is so that we can authenticate it in return_to_handler() if that
> > > > > exists on the call path, removing a potentially vulnerable sequence
> > > > > from that function.
> > > >
> > > > What I was trying to point out is that there is a window where this is spilled
> > > > to the stack (and hence is potentially vulnerable) between
> > > > ftrace_{caller,regs_caller}() and the end of ftrace_common().
> > > >
> > > > So if we don't protect this when CONFIG_FUNCTION_GRAPH_TRACER=n, it could be
> > > > clobbered during that window (e.g. while function tracers are invoked),
> > > > *before* we return back into the instrumented function and sign the
> > > > (potentially already clobbered) value.
> > >
> > > Agreed.
> > >
> > > But to clarify, the intent of this series is not to add protection to
> > > ftrace, the intent is to get rid of the gadgets from the ftrace code
> > > that can be abused even if you don't use ftrace at all.
> >
> > Ok; sorry for missing that; I'll need to think a little harder.
> >
> 
> You said it :-)
> 
> > > > Hence, my thinking is that we should sign this regardless of
> > > > CONFIG_FUNCTION_GRAPH_TRACER to mitigate that case. I agree that we also want
> > > > it to be signed while it's in the graph return stack (i.e. until the
> > > > instrumented function returns back to return_to_handler()). In general, we
> > > > should sign the value if it's going to be spilled to the stack.
> > >
> > > Sure, but it solves a different problem.
> >
> > Fair enough!
> >
> > I think we're agreed that something which solves both issues makes sense, even
> > if that's not necessary for the gadgetisation issue specifically?
> 
> Of course.

Great -- just wanted to check there wasn't an inverse problem I'd missed!

> So the issue we are talking about here is the fact that you might be
> able to attack the ftrace infrastructure while it is being used so
> that the function return from ftrace_common() is made to point
> somewhere else.

Yup. I suspect the risk is must lower due to the smaller amount of code there,
but given things like fprobe and BPF hooks, there might be code that gets
injected there which isn't as careful as we'd like, so it would be nice to
protect.

> I agree that this is something we might want to
> harden, and I also wonder whether we should perhaps insert three NOPs
> instead of two, or teach the compiler to put its PACIASP right after
> so that we can use BR instead of RET to perform the return.

I think that approach is a mixed bag :/

I was hoping that we could reduce the set of BTI-compatible instructions we
have, and I'd like to get to a point where we can set SCTLR_ELx.BT1=1 so that
PACIASP isn't an implicit BTI in the kernel. That way we'd be in a similar boat
to x86 after redundant ENDBRs are removed, with forward-edge protection being
strengthened and EXPORT-control being strengthened. That needs new compiler
help to output separate BTI and PACIASP instructions, but otherwise that's
relatively simple, and could significantly reduce the set of gadgetizable
functions regardless of whether ftrace is in use.

Given that, I'm not keen on adding an extra BTI-compatible instruction into
function prologues.

> But again, this is ground that I am currently not attempting to cover.
> 
> > > > > > I'm a bit worried this might confuse some ftrace code manipulating the return
> > > > > > address (e.g. manipulation of the ftrace graph return stack), as I don't think
> > > > > > that's all PAC-clean, and might need some modification.
> > > > >
> > > > > This is the reason for the xpaci instruction below.
> > > >
> > > > Unfortunately, that alone isn't sufficient.
> > > >
> > > > What I was alluding to is that this change means the ftrace graph return stack
> > > > contains signed addresses, and other code doesn't expect that. For example,
> > > > arm64's stacktrace code currently depends on the graph return stack containing
> > > > plain pointers, and so that gets broken as of this patch when function graph
> > > > tracing is enabled:
> > > >
> > > > | # uname -a
> > > > | # Linux buildroot 6.1.0-rc7-00003-g44a67f0b8ac7 #1 SMP PREEMPT Wed Nov 30 17:19:38 GMT 2022 aarch64 GNU/Linux
> > > > | # cat /proc/self/stack
> > > > | [<0>] proc_pid_stack+0xc0/0x130
> > > > | [<0>] proc_single_show+0x68/0x120
> > > > | [<0>] seq_read_iter+0x16c/0x45c
> > > > | [<0>] seq_read+0x98/0xd0
> > > > | [<0>] vfs_read+0xc8/0x2c0
> > > > | [<0>] ksys_read+0x78/0x110
> > > > | [<0>] __arm64_sys_read+0x24/0x30
> > > > | [<0>] invoke_syscall+0x50/0x120
> > > > | [<0>] el0_svc_common.constprop.0+0xd4/0xf4
> > > > | [<0>] do_el0_svc+0x34/0xd0
> > > > | [<0>] el0_svc+0x2c/0x84
> > > > | [<0>] el0t_64_sync_handler+0xf4/0x120
> > > > | [<0>] el0t_64_sync+0x18c/0x190
> > > > | # echo function_graph > /sys/kernel/debug/tracing/current_tracer
> > > > | # cat /proc/self/stack
> > > > | [<0>] 0xf5f98000083dff40
> > > > | [<0>] 0xd6b88000083e0f68
> > > > | [<0>] 0x21ac800008381ad0
> > > > | [<0>] 0xd0bc800008381e58
> > > > | [<0>] 0x22b280000834bc28
> > > > | [<0>] 0xf0ca80000834c5c8
> > > > | [<0>] 0x299080000834c684
> > > > | [<0>] 0xb1a1800008029cf0
> > > > | [<0>] 0x9bd0800008029e94
> > > > | [<0>] 0x1788800008029ee8
> > > > | [<0>] 0xa08680000916dd5c
> > > > | [<0>] el0t_64_sync_handler+0xf4/0x120
> > > > | [<0>] el0t_64_sync+0x18c/0x190
> > > >
> > > > That's unfortunate (and would break RELIABLE_STACKTRACE, which we're slowly
> > > > getting towards being able to implement), but it's simple enough to account for
> > > > in the stacktrace code.
> > > >
> > >
> > > Indeed. Those functions should just strip the PAC bits, no?
> >
> > For that case, yup. That was roughly what I meant about it being simple to deal
> > with in the stacktrace code. :)
> 
> Right. So given that this is an issue for PAC but not for shadow call
> stack, we might consider a shorter term fix where we push/pop these
> addresses to the shadow call stack, and address the PAC clearing more
> comprehensively once we get around to it.

I'm not necessarily opposed to that, and TBH we might not need the address in
the graph return stack to be signed, since the graph return stack itself is a
shadow stack.

I think we can restructure things such that the values on the graph return
stack would remain unsigned, but we'd still always protect spills to the
regular stack *AND* the assembly would be structured to ensure to remove the
return gadgets.

As before, I'll have a go at that and try to get it out shortly.

[...]

> > With that in mind, I think we should also fix up
> > qcom_link_stack_sanitisation(), since that ends up creating a gadget of the form:
> >
> >         MOV     X30, Xn
> >         RET

> Yeah, I'm sure that's the last one :-)

:)

Mark.