Re: [PATCH v4] kprobes: unpoison stack in jprobe_return() for KASAN

[Mark Rutland]

On Fri, Oct 14, 2016 at 01:54:30PM +0200, Dmitry Vyukov wrote:
> KASAN stack instrumentation poisons stack redzones on function entry
> and unpoisons them on function exit. If a function exits abnormally
> (e.g. with a longjmp like jprobe_return()), stack redzones are left
> poisoned. Later this leads to random KASAN false reports.
> 
> Unpoison stack redzones in the frames we are going to jump over
> before doing actual longjmp in jprobe_return().
> 
> Signed-off-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> Reviewed-by: Mark Rutland <mark.rutland@xxxxxxx>

... judging by the kbuild test robot I spoke too soon, and should have
been more thorough. :/

> +/*
> + * Clear all poison for the region between the current SP and a provided
> + * watermark value, as is sometimes required prior to hand-crafted asm function
> + * returns in the middle of functions.
> + */
> +void kasan_unpoison_stack_above_sp_to(const void *watermark)
> +{
> +	const void *sp = (void *)current_stack_pointer();

Aargh; it seems current_stack_pointer() is only function-like on some
arches, and not on others (arm64 included). I should have known better;
sorry for the bad suggestion.

I'm not overjoyed about taking the address of a stack variable to
implement this ourselves. Can we use __builtin_frame_address(0) instead?
Or are there cases where that won't work on x86?

> +	size_t size = watermark - sp;
> +
> +	if (WARN_ON(sp > watermark))
> +		return;

... not a new problem, but we should also include <linux/bug.h> for
WARN_ON().

Thanks,
Mark.