Re: usercopy_warn in __copy_to_user

[Kees Cook]

On Wed, Dec 19, 2018 at 11:51 AM <rishabhb@xxxxxxxxxxxxxx> wrote:
>
>
> In the 4.19 kernel, we are seeing a USERCOPY_WARN in __copy_to_user
> during bootup.
> The code-flow is something like this:
>
> (arch/arm64/kernel/signal.c)
> struct sigset_t *set;
> __copy_to_user(&sf->uc.uc_sigmask, set, sizeof(*set))
>
> (include/linux/uaccess.h)
> __copy_to_user(void __user *to, const void *from, unsigned long n)
> {
>                  might_fault();
>                  kasan_check_read(from, n);
>                  check_object_size(from, n, true);
>                  return raw_copy_to_user(to, from, n);
> }
>
> (include/linux/thread_info.h)
> static __always_inline void check_object_size(const void *ptr,
>                                   unsigned long n, bool to_user)
> {
>                  if (!__builtin_constant_p(n))
>                                  __check_object_size(ptr, n, to_user);
> }
>
> Since sizeof(*set) is constant, __builtin_constant_p(n) should return
> true.
> But we are seeing that its returning the value as false. Because of
> which
> the code goes on to __check_object_size and generates a USERCOPY_WARN
> ("usercopy: WARN() on slab cache usercopy
> region violations").
>
> We are using LLVM clang version 6.0 to compile the kernel and not gcc.

This is fixed in the latest LLVM and Clang:

https://github.com/ClangBuiltLinux/linux/issues/7

-Kees

> In clang, __builtin_constant_p is evaluated immediately, before inlining
> or other optimizations run, gcc evaluates it later.
> We believe that maybe causing __builtin_constant_p(n) to return false.
> Thereâs upstream work to change LLVM, so __builtin_constant_p works more
> like gcc when optimization is enabled, but its still in progress.
>
>
> For this scenario is there a way to avoid the warning? Should the code
> be
> written in a different to avoid dependency on compiler?
>
> Thanks,
> Rishabh



-- 
Kees Cook