Re: [PATCH] xtensa: fix {get,put}_user() for 64bit values

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Oct 09, 2019 at 06:38:12PM -0700, Max Filippov wrote:

> There's also the following code in the callers of this macro, e.g. in
> __get_user_nocheck:
> 
>         long __gu_err, __gu_val;                                \
>         __get_user_size(__gu_val, (ptr), (size), __gu_err);     \
>         (x) = (__force __typeof__(*(ptr)))__gu_val;             \
> 
> the last line is important for sizes 1..4, because it takes care of
> sign extension of the value loaded by the assembly.
> At the same time the first line doesn't make sense for the size 8
> as it will result in value truncation.

Right you are...

> +       long __gu_err;                                          \
> +       __typeof__(*(ptr) + 0) __gu_val;                        \

What would __u64 __gu_val; end up with for smaller sizes?  I don't have
xtensa cross-toolchain at the moment, so I can't check it easily;
what does =r constraint generate in such case?

Another thing is, you want to zero it on failure, to avoid an uninitialized
value ending up someplace interesting....

> @@ -198,7 +200,7 @@ do {
>                          \
>         case 1: __get_user_asm(x, ptr, retval, 1, "l8ui", __cb);  break;\
>         case 2: __get_user_asm(x, ptr, retval, 2, "l16ui", __cb); break;\
>         case 4: __get_user_asm(x, ptr, retval, 4, "l32i", __cb);  break;\
> -       case 8: retval = __copy_from_user(&x, ptr, 8);    break;        \
> +       case 8: retval = __copy_from_user(&x, ptr, 8) ? -EFAULT : 0;
>  break;  \
>         default: (x) = __get_user_bad();                                \
>         }                                                               \
>  } while (0)
> 
> Here __typeof__(*(ptr) + 0) makes enough room for all cases
> in the __get_user_size and the "+0" part takes care of pointers
> to const data.