Re: [PATCH v4 3/5] stack: Optionally randomize kernel stack offset each syscall

[Kees Cook]

On Mon, Jun 22, 2020 at 10:07:37PM +0200, Jann Horn wrote:
> On Mon, Jun 22, 2020 at 9:31 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> > This provides the ability for architectures to enable kernel stack base
> > address offset randomization. This feature is controlled by the boot
> > param "randomize_kstack_offset=on/off", with its default value set by
> > CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT.
> [...]
> > +#define add_random_kstack_offset() do {                                        \
> > +       if (static_branch_maybe(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT, \
> > +                               &randomize_kstack_offset)) {            \
> > +               u32 offset = this_cpu_read(kstack_offset);              \
> > +               u8 *ptr = __builtin_alloca(offset & 0x3FF);             \
> > +               asm volatile("" : "=m"(*ptr));                          \
> > +       }                                                               \
> > +} while (0)
> 
> clang generates better code here if the mask is stack-aligned -
> otherwise it needs to round the stack pointer / the offset:

Interesting. I was hoping to avoid needing to know the architecture
stack alignment (leaving it up to the compiler).

> 
> $ cat alloca_align.c
> #include <alloca.h>
> void callee(void);
> 
> void alloca_blah(unsigned long rand) {
>   asm volatile(""::"r"(alloca(rand & MASK)));
>   callee();
> }
> $ clang -O3 -c -o alloca_align.o alloca_align.c -DMASK=0x3ff
> $ objdump -d alloca_align.o
> [...]
>    0: 55                    push   %rbp
>    1: 48 89 e5              mov    %rsp,%rbp
>    4: 81 e7 ff 03 00 00    and    $0x3ff,%edi
>    a: 83 c7 0f              add    $0xf,%edi
>    d: 83 e7 f0              and    $0xfffffff0,%edi
>   10: 48 89 e0              mov    %rsp,%rax
>   13: 48 29 f8              sub    %rdi,%rax
>   16: 48 89 c4              mov    %rax,%rsp
>   19: e8 00 00 00 00        callq  1e <alloca_blah+0x1e>
>   1e: 48 89 ec              mov    %rbp,%rsp
>   21: 5d                    pop    %rbp
>   22: c3                    retq
> $ clang -O3 -c -o alloca_align.o alloca_align.c -DMASK=0x3f0
> $ objdump -d alloca_align.o
> [...]
>    0: 55                    push   %rbp
>    1: 48 89 e5              mov    %rsp,%rbp
>    4: 48 89 e0              mov    %rsp,%rax
>    7: 81 e7 f0 03 00 00    and    $0x3f0,%edi
>    d: 48 29 f8              sub    %rdi,%rax
>   10: 48 89 c4              mov    %rax,%rsp
>   13: e8 00 00 00 00        callq  18 <alloca_blah+0x18>
>   18: 48 89 ec              mov    %rbp,%rsp
>   1b: 5d                    pop    %rbp
>   1c: c3                    retq
> $
> 
> (From a glance at the assembly, gcc seems to always assume that the
> length may be misaligned.)

Right -- this is why I didn't bother with it, since it didn't seem to
notice what I'd already done to the alloca() argument. (But from what I
could measure on cycle counts, the additional ALU didn't seem to really
make much difference ... it _would_ be nice to avoid it, of course.)

> Maybe this should be something along the lines of
> __builtin_alloca(offset & (0x3ff & ARCH_STACK_ALIGN_MASK)) (with
> appropriate definitions of the stack alignment mask depending on the
> architecture's choice of stack alignment for kernel code).

Is that explicitly selected anywhere in the kernel? I thought the
alignment was left up to the compiler (as in I've seen bugs fixed where
the kernel had to deal with the alignment choices the compiler was
making...)

-- 
Kees Cook