Re: [RFC PATCH v1 01/10] s390/uaccess: Add storage key checked access to user memory

[Heiko Carstens]

On Fri, Jan 21, 2022 at 08:32:25AM +0100, Christian Borntraeger wrote:
> So in essence adding something like this and then providing raw_copy_from/to_user_key?
> (whitespace damaged, just pasted in)
> 
> diff --git a/include/linux/uaccess.h b/include/linux/uaccess.h
> index ac0394087f7d..3b6e78ee211c 100644
> --- a/include/linux/uaccess.h
> +++ b/include/linux/uaccess.h
> @@ -201,6 +201,59 @@ copy_to_user(void __user *to, const void *from, unsigned long n)
>         return n;
>  }
> +
> +#if defined(__s390x__) && defined(CONFIG_KVM)
> +/*
> + * Variants that pass along an access key. Uses by KVM on s390x to implement
> + * key checks for guests that use storage keys Must be kept in sync with the
> + * non-key variants from above. The only difference is the _key suffix when
> + * calling raw_copy_from/to_user_key.
> + */

This is too architecture specific, I wouldn't like to see __s390__ or
KVM dependencies. This should be a bit more generic, so other
architectures _might_ also make use of this interface that is:

> +static inline __must_check unsigned long
> +_copy_from_user_key(void *to, const void __user *from, unsigned long n, u8 key)

Make key unsigned long, add support for INLINE_COPY_TO_USER, and maybe
add a wrapper, so this works on all architectures, e.g. if
raw_copy_to_user_key() is not defined, then fall back to
raw_copy_to_user() and ignore the key parameter.

Just some ideas. The _only_ concern about the previous implementation
I have, is that we could run out-of-sync with common code wrt
instrumentation and other options. Given how sensitive uaccess ops
are, I think we should try to avoid that by all cost, if possible.

Does that make sense?