Re: [Intel-gfx] [PATCH v9 1/8] overflow: Move and add few utility macros into overflow

[Kees Cook <keescook@chromium.org>]

On Wed, Aug 24, 2022 at 05:45:07PM +0900, Gwan-gyeong Mun wrote:
> It moves overflows_type utility macro into overflow header from i915_utils
> header. The overflows_type can be used to catch the truncaion (overflow)
> between different data types. And it adds check_assign() macro which
> performs an assigning source value into destination ptr along with an
> overflow check. overflow_type macro has been improved to handle the signbit
> by gcc's built-in overflow check function. And it adds overflows_ptr()
> helper macro for checking the overflows between a value and a pointer
> type/value.
> 
> v3: Add is_type_unsigned() macro (Mauro)
>     Modify overflows_type() macro to consider signed data types (Mauro)
>     Fix the problem that safe_conversion() macro always returns true
> v4: Fix kernel-doc markups
> v6: Move macro addition location so that it can be used by other than drm
>     subsystem (Jani, Mauro, Andi)
>     Change is_type_unsigned to is_unsigned_type to have the same name form
>     as is_signed_type macro
> v8: Add check_assign() and remove safe_conversion() (Kees)
>     Fix overflows_type() to use gcc's built-in overflow function (Andrzej)
>     Add overflows_ptr() to allow overflow checking when assigning a value
>     into a pointer variable (G.G.)
> v9: Fix overflows_type() to use __builtin_add_overflow() instead of
>     __builtin_add_overflow_p() (Andrzej)
>     Fix overflows_ptr() to use overflows_type() with the unsigned long type
>     (Andrzej)
> 
> Signed-off-by: Gwan-gyeong Mun <gwan-gyeong.mun@intel.com>
> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Cc: Matthew Auld <matthew.auld@intel.com>
> Cc: Nirmoy Das <nirmoy.das@intel.com>
> Cc: Jani Nikula <jani.nikula@intel.com>
> Cc: Andi Shyti <andi.shyti@linux.intel.com>
> Cc: Andrzej Hajda <andrzej.hajda@intel.com>
> Cc: Mauro Carvalho Chehab <mauro.chehab@linux.intel.com>
> Cc: Kees Cook <keescook@chromium.org>
> Reviewed-by: Mauro Carvalho Chehab <mchehab@kernel.org> (v5)
> ---
>  drivers/gpu/drm/i915/i915_user_extensions.c |  3 +-
>  drivers/gpu/drm/i915/i915_utils.h           |  5 +-
>  include/linux/overflow.h                    | 62 +++++++++++++++++++++
>  3 files changed, 64 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_user_extensions.c b/drivers/gpu/drm/i915/i915_user_extensions.c
> index c822d0aafd2d..6f6b5b910968 100644
> --- a/drivers/gpu/drm/i915/i915_user_extensions.c
> +++ b/drivers/gpu/drm/i915/i915_user_extensions.c
> @@ -50,8 +50,7 @@ int i915_user_extensions(struct i915_user_extension __user *ext,
>  		if (err)
>  			return err;
>  
> -		if (get_user(next, &ext->next_extension) ||
> -		    overflows_type(next, ext))
> +		if (get_user(next, &ext->next_extension) || overflows_ptr(next))
>  			return -EFAULT;
>  
>  		ext = u64_to_user_ptr(next);

I continue to dislike the layers of macros and specialization here.
This is just a fancy version of check_assign():

	if (get_user(next, &ext->next_extension) || check_assign(next, &ext))
		return -EFAULT;

However, the __builtin_*_overflow() family only wants to work on
integral types, so this needs to be slightly expanded:

	uintptr_t kptr;
	...
	if (get_user(next, &ext->next_extension) || check_assign(next, &kptr))
		return -EFAULT;

	ext = (void * __user)kptr;

But, it does seem like the actual problem here is that u64_to_user_ptr()
is not performing the checking? It's used heavily in the drm code.

Is a check_assign_user_ptr() needed?

> diff --git a/drivers/gpu/drm/i915/i915_utils.h b/drivers/gpu/drm/i915/i915_utils.h
> index c10d68cdc3ca..eb0ded23fa9c 100644
> --- a/drivers/gpu/drm/i915/i915_utils.h
> +++ b/drivers/gpu/drm/i915/i915_utils.h
> @@ -32,6 +32,7 @@
>  #include <linux/types.h>
>  #include <linux/workqueue.h>
>  #include <linux/sched/clock.h>
> +#include <linux/overflow.h>
>  
>  #ifdef CONFIG_X86
>  #include <asm/hypervisor.h>
> @@ -111,10 +112,6 @@ bool i915_error_injected(void);
>  #define range_overflows_end_t(type, start, size, max) \
>  	range_overflows_end((type)(start), (type)(size), (type)(max))
>  
> -/* Note we don't consider signbits :| */
> -#define overflows_type(x, T) \
> -	(sizeof(x) > sizeof(T) && (x) >> BITS_PER_TYPE(T))
> -
>  #define ptr_mask_bits(ptr, n) ({					\
>  	unsigned long __v = (unsigned long)(ptr);			\
>  	(typeof(ptr))(__v & -BIT(n));					\
> diff --git a/include/linux/overflow.h b/include/linux/overflow.h
> index f1221d11f8e5..6af9df1d67a1 100644
> --- a/include/linux/overflow.h
> +++ b/include/linux/overflow.h
> @@ -52,6 +52,68 @@ static inline bool __must_check __must_check_overflow(bool overflow)
>  	return unlikely(overflow);
>  }
>  
> +/**
> + * overflows_type - helper for checking the overflows between data types or
> + *                  values
> + *
> + * @x: Source value or data type for overflow check
> + * @T: Destination value or data type for overflow check
> + *
> + * It compares the values or data type between the first and second argument to
> + * check whether overflow can occur when assigning the first argument to the
> + * variable of the second argument. Source and Destination can be singned or
> + * unsigned data types. Source and Destination can be different data types.
> + *
> + * Returns:
> + * True if overflow can occur, false otherwise.
> + */
> +#define overflows_type(x, T) __must_check_overflow(({	\
> +	typeof(T) v = 0;				\
> +	__builtin_add_overflow((x), v, &v);		\
> +}))

I'd like to avoid "externalizing" these kinds of checks when the better
path is to catch the issue at operation type (add, sub, mul, assign).
Looking at existing users, I see stuff like:

                if (overflows_type(item.query_id - 1, unsigned long))
                        return -EINVAL;

                func_idx = item.query_id - 1;

This requires too much open-coded checking, IMO. It's better as:

		if (check_assign(item.query_id - 1, &func_idx))
			return -EINVAL;

or other similar:

        if (overflows_type(user->slice_mask, context->slice_mask) ||
	...
        context->slice_mask = user->slice_mask;

and some that don't make sense to me. Why check argument types? Is this
trying to avoid implicit type conversions?

So, if it's _really_ needed, I can live with adding overflows_type().

> +
> +/**
> + * overflows_ptr - helper for checking the occurrence of overflows when a value
> + *                 assigns to  the pointer data type
> + *
> + * @x: Source value for overflow check
> + *
> + * gcc's built-in overflow check functions don't support checking between the
> + * pointer type and non-pointer type. And ILP32 and LP64 have the same bit size
> + * between long and pointer. Therefore it internally compares the source value
> + * and unsigned long data type for checking overflow.
> + *
> + * Returns:
> + * True if overflow can occur, false otherwise.
> + */
> +#define overflows_ptr(x) __must_check_overflow(overflows_type(x, unsigned long))

I'd rather not have this -- it's just a specialized use of
overflows_type(), and only used in 1 place.

> +
> +/**
> + * check_assign - perform an assigning source value into destination ptr along
> + *                with an overflow check.
> + *
> + * @value: Source value
> + * @ptr: Destination pointer address, If the pointer type is not used,
> + *       a warning message is output during build.
> + *
> + * It checks internally the ptr is a pointer type. And it uses gcc's built-in
> + * overflow check function.
> + * It does not use the check_*() wrapper functions, but directly uses gcc's
> + * built-in overflow check function so that it can be used even when
> + * the type of value and the type pointed to by ptr are different without build
> + * warning messages.

This is a good point: the check_{add,sub,mul}_overflow() helpers
currently require all the params be the same type, which rather limits
their usage. Perhaps this can be weakened now that we're not using[1]
the fallback logic any more? (Separate patch.)

> + *
> + * Returns:
> + * If the value would overflow the destination, it returns true. If not return
> + * false. When overflow does not occur, the assigning into ptr from value
> + * succeeds. It follows the return policy as other check_*_overflow() functions
> + * return non-zero as a failure.
> + */
> +#define check_assign(value, ptr) __must_check_overflow(({	\
> +	typecheck_pointer(ptr); 		\
> +	__builtin_add_overflow(0, value, ptr);	\
> +}))

But yes, this looks correct. I really like it. :)

> +
>  /*
>   * For simplicity and code hygiene, the fallback code below insists on
>   * a, b and *d having the same type (similar to the min() and max()
> -- 
> 2.37.1
> 

-Kees

[1] 4eb6bd55cfb2 ("compiler.h: drop fallback overflow checkers")

-- 
Kees Cook