Re: [PATCH v2] lkdtm: Add CFI_BACKWARD to test ROP mitigations

[Kees Cook <keescook@chromium.org>]

On Tue, Dec 06, 2022 at 06:28:53PM -0600, Daniel Díaz wrote:
> Hello!
> 
> On Sat, 16 Apr 2022 at 00:30, Kees Cook <keescook@chromium.org> wrote:
> > In order to test various backward-edge control flow integrity methods,
> > add a test that manipulates the return address on the stack. Currently
> > only arm64 Pointer Authentication and Shadow Call Stack is supported.
> >
> >  $ echo CFI_BACKWARD | cat >/sys/kernel/debug/provoke-crash/DIRECT
> >
> > Under SCS, successful test of the mitigation is reported as:
> >
> >  lkdtm: Performing direct entry CFI_BACKWARD
> >  lkdtm: Attempting unchecked stack return address redirection ...
> >  lkdtm: ok: redirected stack return address.
> >  lkdtm: Attempting checked stack return address redirection ...
> >  lkdtm: ok: control flow unchanged.
> >
> > Under PAC, successful test of the mitigation is reported by the PAC
> > exception handler:
> >
> >  lkdtm: Performing direct entry CFI_BACKWARD
> >  lkdtm: Attempting unchecked stack return address redirection ...
> >  lkdtm: ok: redirected stack return address.
> >  lkdtm: Attempting checked stack return address redirection ...
> >  Unable to handle kernel paging request at virtual address bfffffc0088d0514
> >  Mem abort info:
> >    ESR = 0x86000004
> >    EC = 0x21: IABT (current EL), IL = 32 bits
> >    SET = 0, FnV = 0
> >    EA = 0, S1PTW = 0
> >    FSC = 0x04: level 0 translation fault
> >  [bfffffc0088d0514] address between user and kernel address ranges
> >  ...
> >
> > If the CONFIGs are missing (or the mitigation isn't working), failure
> > is reported as:
> >
> >  lkdtm: Performing direct entry CFI_BACKWARD
> >  lkdtm: Attempting unchecked stack return address redirection ...
> >  lkdtm: ok: redirected stack return address.
> >  lkdtm: Attempting checked stack return address redirection ...
> >  lkdtm: FAIL: stack return address was redirected!
> >  lkdtm: This is probably expected, since this kernel was built *without* CONFIG_ARM64_PTR_AUTH_KERNEL=y nor CONFIG_SHADOW_CALL_STACK=y
> >
> > Co-developed-by: Dan Li <ashimida@linux.alibaba.com>
> > Signed-off-by: Dan Li <ashimida@linux.alibaba.com>
> > Cc: Arnd Bergmann <arnd@arndb.de>
> > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> > Signed-off-by: Kees Cook <keescook@chromium.org>
> > ---
> > v1: https://lore.kernel.org/lkml/20220413213917.711770-1-keescook@chromium.org
> > v2:
> >  - add PAGE_OFFSET setting for PAC bits (Dan Li)
> > ---
> >  drivers/misc/lkdtm/cfi.c                | 134 ++++++++++++++++++++++++
> >  tools/testing/selftests/lkdtm/tests.txt |   1 +
> >  2 files changed, 135 insertions(+)
> >
> > diff --git a/drivers/misc/lkdtm/cfi.c b/drivers/misc/lkdtm/cfi.c
> > index e88f778be0d5..804965a480b7 100644
> > --- a/drivers/misc/lkdtm/cfi.c
> > +++ b/drivers/misc/lkdtm/cfi.c
> > @@ -3,6 +3,7 @@
> >   * This is for all the tests relating directly to Control Flow Integrity.
> >   */
> >  #include "lkdtm.h"
> > +#include <asm/page.h>
> >
> >  static int called_count;
> >
> > @@ -42,8 +43,141 @@ static void lkdtm_CFI_FORWARD_PROTO(void)
> >         pr_expected_config(CONFIG_CFI_CLANG);
> >  }
> >
> > +/*
> > + * This can stay local to LKDTM, as there should not be a production reason
> > + * to disable PAC && SCS.
> > + */
> > +#ifdef CONFIG_ARM64_PTR_AUTH_KERNEL
> > +# ifdef CONFIG_ARM64_BTI_KERNEL
> > +#  define __no_pac             "branch-protection=bti"
> > +# else
> > +#  define __no_pac             "branch-protection=none"
> > +# endif
> > +# define __no_ret_protection   __noscs __attribute__((__target__(__no_pac)))
> > +#else
> > +# define __no_ret_protection   __noscs
> > +#endif
> 
> We're seeing this problem with allmodconfig on arm64 and GCC 8 (this
> one observed on 6.0.12-rc3):
> 
> -----8<----------8<----------8<-----
> make --silent --keep-going --jobs=8
> O=/home/tuxbuild/.cache/tuxmake/builds/2/build
> CROSS_COMPILE_COMPAT=arm-linux-gnueabihf- ARCH=arm64
> CROSS_COMPILE=aarch64-linux-gnu- 'CC=sccache aarch64-linux-gnu-gcc'
> 'HOSTCC=sccache gcc'
> /builds/linux/drivers/misc/lkdtm/cfi.c:67:1: error: pragma or
> attribute 'target("branch-protection=none")' is not valid
>  {
>  ^

Uuuh... how is CONFIG_ARM64_PTR_AUTH_KERNEL getting set if the compiler
can't support the 'target("branch-protection=none")' attribute?

-- 
Kees Cook