Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Message-ID: <92a767c4-09e1-8783-2581-9848bb72890d@linux.alibaba.com>
Date:   Mon, 7 Mar 2022 07:16:36 -0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.6.1
Subject: Re: [PATCH v3 2/2] lkdtm: Add Shadow Call Stack tests
Content-Language: en-US
To:     Kees Cook <keescook@chromium.org>
Cc:     akpm@linux-foundation.org, arnd@arndb.de, catalin.marinas@arm.com,
        gregkh@linuxfoundation.org, linux@roeck-us.net,
        luc.vanoostenryck@gmail.com, elver@google.com,
        mark.rutland@arm.com, masahiroy@kernel.org, ojeda@kernel.org,
        nathan@kernel.org, npiggin@gmail.com, ndesaulniers@google.com,
        samitolvanen@google.com, shuah@kernel.org, tglx@linutronix.de,
        will@kernel.org, linux-arm-kernel@lists.infradead.org,
        linux-kernel@vger.kernel.org, linux-kselftest@vger.kernel.org,
        llvm@lists.linux.dev, linux-hardening@vger.kernel.org
References: <20220303073340.86008-1-ashimida@linux.alibaba.com>
 <20220303074339.86337-1-ashimida@linux.alibaba.com>
 <202203031010.0A492D114@keescook> <202203031105.A1B4CAE6@keescook>
From:   Dan Li <ashimida@linux.alibaba.com>
In-Reply-To: <202203031105.A1B4CAE6@keescook>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 3/3/22 11:09, Kees Cook wrote:
> On Thu, Mar 03, 2022 at 10:42:45AM -0800, Kees Cook wrote:
>> And we should, actually, be able to make the "set_lr" functions be
>> arch-specific, leaving the test itself arch-agnostic....
> 
> Yeah, as a tested example, this works for x86_64, and based on what you
> had, I'd expect it to work on arm64 too:
> 
> #include <stdio.h>
> 
> static __attribute__((noinline))
> void set_return_addr(unsigned long *expected, unsigned long *addr)
> {
>      /* Use of volatile is to make sure final write isn't seen as a dead store. */
>      unsigned long * volatile *ret_addr = (unsigned long **)__builtin_frame_address(0) + 1;
> 
>      /* Make sure we've found the right place on the stack before writing it. */
>      if (*ret_addr == expected)
>          *ret_addr = addr;
> }
> 
> volatile int force_label;
> int main(void)
> {
>      do {
>          /* Keep labels in scope. */
>          if (force_label)
>              goto normal;
>          if (force_label)
>              goto redirected;
> 
>          set_return_addr(&&normal, &&redirected);
> normal:
>          printf("I should be skipped\n");
>          break;
> redirected:
>          printf("Redirected\n");
>      } while (0);
> 
>      return 0;
> }
> 
> 
> It does _not_ work under Clang, though, which I'm still looking at.
> 

The following code seems to work fine under clang/gcc, x86_64/aarch64
(also tested in lkdtm_CFI_BACKWARD_SHADOW):

#include <stdio.h>

static __attribute__((noinline))
void set_return_addr(unsigned long *expected, unsigned long *addr)
{
     /* Use of volatile is to make sure final write isn't seen as a dead store. */
     unsigned long * volatile *ret_addr = (unsigned long **)__builtin_frame_address(0) + 1;

     /* Make sure we've found the right place on the stack before writing it. */
     if(*ret_addr == expected)
         *ret_addr = (addr);
}

static volatile int force_label;

int main(void)
{
     void *array[] = {0, &&normal, &&redirected};

     if (force_label) {
         /* Call it with a NULL to avoid parameters being treated as constants in -02. */
         set_return_addr(NULL, NULL);
         goto * array[force_label];
     }

     do {

         set_return_addr(&&normal, &&redirected);

normal:
         printf("I should be skipped\n");
         break;

redirected:
         printf("Redirected\n");

     } while (0);

     return 0;
}

But currently it still crashes when I try to enable
"-mbranch-protection=pac-ret+leaf+bti".

Because the address of "&&redirected" is not encrypted under pac,
the autiasp check will fail when set_return_addr returns, and
eventually cause the function to crash when it returns to "&&redirected"
("&&redirected" as a reserved label always seems to start with a bti j
insn).

For lkdtm, if we're going to handle both cases in one function, maybe
it would be better to turn off the -mbranch-protection=pac-ret+leaf+bti
and maybe also turn off -O2 options for the function :)

Thanks,
Dan.