[PATCH v1] Arm64: ASLR: fix text randomization

[Arun Chandran]

Hi Mark,

On Tue, Oct 7, 2014 at 7:13 PM, Mark Rutland <mark.rutland at arm.com> wrote:
>
> On Tue, Oct 07, 2014 at 01:40:28PM +0100, Arun Chandran wrote:
> > This is due to incorrect definition of ELF_ET_DYN_BASE. It
> > introduces randomization for text even if user does a "echo 0 >
> > /proc/sys/kernel/randomize_va_space"
>
> Interesting.
>
> It looks like this was a copy of what powerpc and s390 do (authors
> Cc'd), and the generic support came later. powerpc gained support in
> 501cb16d3cfdcca9 (powerpc: Randomise PIEs), but the generic support was
> enabled later in e39f560239984c30 (fs: binfmt_elf: create Kconfig
> variable for PIE randomization).
>

I did not understand why they need a special architecture randomize_et_dyn()
function to handle the situation.

I have tested PIE on arm and x86 (which don't have a randomize_et_dyn()) and
it works as expected.

>
> The policy of disabling PIE randomization was added in a3defbe5c337dbc6
> (binfmt_elf: fix PIE execution with randomization disabled), after the
> powerpc implementation, but before the x86 implementation was made
> generic.

Thought about extending the policy(a3defbe5c337dbc6) to arm64 by doing

#############
diff --git a/arch/arm64/include/asm/elf.h b/arch/arm64/include/asm/elf.h
index 01d3aab..401b1e8 100644
--- a/arch/arm64/include/asm/elf.h
+++ b/arch/arm64/include/asm/elf.h
@@ -127,6 +127,7 @@ typedef struct user_fpsimd_state elf_fpregset_t;
  */
 extern unsigned long randomize_et_dyn(unsigned long base);
 #define ELF_ET_DYN_BASE        (randomize_et_dyn(2 * TASK_SIZE_64 / 3))
+#define ARM64_ELF_ET_CONST_BASE                (2 * TASK_SIZE_64 / 3)

 /*
  * When the program starts, a1 contains a pointer to a function to be
diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c
index 29d4869..5115f80 100644
--- a/arch/arm64/kernel/process.c
+++ b/arch/arm64/kernel/process.c
@@ -406,5 +406,8 @@ unsigned long arch_randomize_brk(struct mm_struct *mm)

 unsigned long randomize_et_dyn(unsigned long base)
 {
-       return randomize_base(base);
+       if (current->flags & PF_RANDOMIZE)
+               return randomize_base(base);
+       else
+               return ARM64_ELF_ET_CONST_BASE;
 }
##############

then discarded it after seeing the same thing works on x86 and arm.
In arm64(and in ppc and s390) why we need a special randomize_et_dyn()?

>
>
> I wasn't able to spot where the randomness came from in the
> ARCH_BINFMT_ELF_RANDOMIZE_PIE case, so it's not clear to me if the
> generic implementation behaves identically other than disabling
> randomization when told to via proc.

I also don't know from where it is coming; but it works on arm and x86 :)
>
>
> Assuming it behaves similarly enough, it looks like arm64, powerpc, and
> s390 should all be moved over.
>
> >
> > Signed-off-by: Arun Chandran <achandran at mvista.com>
> > ---
> > This can be tested using the code below
> >
> > #include <stdio.h>
> >
> > int main(int argc, char *argv)
> > {
> >     printf("main = %p\n", main);
> >     return 0;
> > }
> >
> > * compile it possition independently
> >   aarch64-linux-gnu-gcc -fPIE -pie aslr.c -o aslr
> >
> > * run it on the target
> >
> >       # ./aslr
> >       main = 0x7f87138950
> >       # ./aslr
> >       main = 0x7f94a10950
> >       # ./aslr
> >       main = 0x7f94fee950
> >       # ./aslr text
> >       main = 0x7f8cb72950
> >
> >       # echo 0 > /proc/sys/kernel/randomize_va_space
> >       # ./aslr text
> >       main = 0x5555555950
> >       # ./aslr
> >       main = 0x5555555950
> >       # ./aslr
> >       main = 0x5555555950
> >       # ./aslr
> >       main = 0x5555555950
>
> It would be worth pointing out that this is after your patch is applied.
> Before your patch I get randomized VAs even after writing 0 to
> randomize_va_spave.

Ok.

--Arun