Re: [PATCH] usercopy: delete __noreturn from usercopy_abort

[Kees Cook]

On Tue, Mar 05, 2024 at 11:31:06AM +0800, Jiangfeng Xiao wrote:
> 
> 
> On 2024/3/5 1:40, Kees Cook wrote:
> > On Mon, Mar 04, 2024 at 04:15:07PM +0100, Jann Horn wrote:
> >> On Mon, Mar 4, 2024 at 3:02 AM Jiangfeng Xiao <xiaojiangfeng@xxxxxxxxxx> wrote:
> >>> When the last instruction of a noreturn function is a call
> >>> to another function, the return address falls outside
> >>> of the function boundary. This seems to cause kernel
> >>> to interrupt the backtrace.
> > 
> > FWIW, all email from huawei.com continues to get eaten by anti-spam
> > checking. I've reported this a few times -- it'd be really nice if the
> > domain configuration could get fixed.
> > 
> >> [...]
> >>> Delete __noreturn from usercopy_abort,
> >>
> >> This sounds like the actual bug is in the backtracing logic? I don't
> >> think removing __noreturn annotations from an individual function is a
> >> good fix, since the same thing can happen with other __noreturn
> >> functions depending on what choices the compiler makes.
> > 
> > Yeah, NAK. usercopy_abort() doesn't return. It ends with BUG().
> > 
> When the user directly or indirectly calls usercopy_abort,
> the final call stack is incorrect, and the
> code where the problem occurs cannot be located.
> In this case, the user will be frustrated.

Can you please give an example of this?

> For the usercopy_abort function, whether '__noreturn' is added
> does not affect the internal behavior of the usercopy_abort function.
> Therefore, it is recommended that '__noreturn' be deleted
> so that backtrace can work properly.

This isn't acceptable. Removing __noreturn this will break
objtool's processing of execution flow for livepatching, IBT, and
KCFI instrumentation. These all depend on an accurate control flow
descriptions, and usercopy_abort is correctly marked __noreturn.

-- 
Kees Cook