[PATCH v8 3/6] stack: Optionally randomize kernel stack offset each syscall

[Kees Cook]

On Thu, Apr 01, 2021 at 11:15:43AM +0000, David Laight wrote:
> From: Will Deacon
> > Sent: 01 April 2021 09:31
> ...
> > > +/*
> > > + * These macros must be used during syscall entry when interrupts and
> > > + * preempt are disabled, and after user registers have been stored to
> > > + * the stack.
> > > + */
> > > +#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(KSTACK_OFFSET_MAX(offset));	\
> > > +		asm volatile("" : "=m"(*ptr) :: "memory");		\
> > 
> > Using the "m" constraint here is dangerous if you don't actually evaluate it
> > inside the asm. For example, if the compiler decides to generate an
> > addressing mode relative to the stack but with writeback (autodecrement), then
> > the stack pointer will be off by 8 bytes. Can you use "o" instead?

I see other examples of empty asm, but it's true, none are using "=m" read
constraints. But, yes, using "o" appears to work happily.

> Is it allowed to use such a mode?
> It would have to know that the "m" was substituted exactly once.
> I think there are quite a few examples with 'strange' uses of memory
> asm arguments.
> 
> However, in this case, isn't it enough to ensure the address is 'saved'?
> So:
> 	asm volatile("" : "=r"(ptr) );
> should be enough.

It isn't, it seems.

Here's a comparison:

https://godbolt.org/z/xYGn9GfGY

So, I'll resend with "o", and with raw_cpu_*().

Thanks!

-- 
Kees Cook