Re: [PATCH 2/2] mm: Add kvmalloc_ab_c and kvzalloc_struct

From: Kees Cook
Date: Wed Feb 14 2018 - 14:22:47 EST


On Wed, Feb 14, 2018 at 10:26 AM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> From: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
>
> We have kvmalloc_array in order to safely allocate an array with a
> number of elements specified by userspace (avoiding arithmetic overflow
> leading to a buffer overrun). But it's fairly common to have a header
> in front of that array (eg specifying the length of the array), so we
> need a helper function for that situation.
>
> kvmalloc_ab_c() is the workhorse that does the calculation, but in spite
> of our best efforts to name the arguments, it's really hard to remember
> which order to put the arguments in. kvzalloc_struct() eliminates that
> effort; you tell it about the struct you're allocating, and it puts the
> arguments in the right order for you (and checks that the arguments
> you've given are at least plausible).
>
> For comparison between the three schemes:
>
> sev = kvzalloc(sizeof(*sev) + sizeof(struct v4l2_kevent) * elems,
> GFP_KERNEL);
> sev = kvzalloc_ab_c(elems, sizeof(struct v4l2_kevent), sizeof(*sev),
> GFP_KERNEL);
> sev = kvzalloc_struct(sev, events, elems, GFP_KERNEL);
>
> Signed-off-by: Matthew Wilcox <mawilcox@xxxxxxxxxxxxx>
> ---
> include/linux/mm.h | 51 +++++++++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 51 insertions(+)
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 81bd7f0be286..ddf929c5aaee 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -557,6 +557,57 @@ static inline void *kvmalloc_array(size_t n, size_t size, gfp_t flags)
> return kvmalloc(n * size, flags);
> }
>
> +/**
> + * kvmalloc_ab_c() - Allocate memory.

Longer description, maybe? "Allocate a *b + c bytes of memory"?

> + * @n: Number of elements.
> + * @size: Size of each element (should be constant).
> + * @c: Size of header (should be constant).

If these should be constant, should we mark them as "const"? Or WARN
if __builtin_constant_p() isn't true?

> + * @gfp: Memory allocation flags.
> + *
> + * Use this function to allocate @n * @size + @c bytes of memory. This
> + * function is safe to use when @n is controlled from userspace; it will
> + * return %NULL if the required amount of memory cannot be allocated.
> + * Use kvfree() to free the allocated memory.
> + *
> + * The kvzalloc_hdr_arr() function is easier to use as it has typechecking

renaming typo? Should this be "kvzalloc_struct()"?

> + * and you do not need to remember which of the arguments should be constants.
> + *
> + * Context: Process context. May sleep; the @gfp flags should be based on
> + * %GFP_KERNEL.
> + * Return: A pointer to the allocated memory or %NULL.
> + */
> +static inline __must_check
> +void *kvmalloc_ab_c(size_t n, size_t size, size_t c, gfp_t gfp)
> +{
> + if (size != 0 && n > (SIZE_MAX - c) / size)
> + return NULL;
> +
> + return kvmalloc(n * size + c, gfp);
> +}
> +#define kvzalloc_ab_c(a, b, c, gfp) kvmalloc_ab_c(a, b, c, gfp | __GFP_ZERO)

Nit: "(gfp) | __GFP_ZERO" just in case of insane usage.

> +
> +/**
> + * kvzalloc_struct() - Allocate and zero-fill a structure containing a
> + * variable length array.
> + * @p: Pointer to the structure.
> + * @member: Name of the array member.
> + * @n: Number of elements in the array.
> + * @gfp: Memory allocation flags.
> + *
> + * Allocate (and zero-fill) enough memory for a structure with an array
> + * of @n elements. This function is safe to use when @n is specified by
> + * userspace as the arithmetic will not overflow.
> + * Use kvfree() to free the allocated memory.
> + *
> + * Context: Process context. May sleep; the @gfp flags should be based on
> + * %GFP_KERNEL.
> + * Return: Zero-filled memory or a NULL pointer.
> + */
> +#define kvzalloc_struct(p, member, n, gfp) \
> + (typeof(p))kvzalloc_ab_c(n, \
> + sizeof(*(p)->member) + __must_be_array((p)->member), \
> + offsetof(typeof(*(p)), member), gfp)
> +
> extern void kvfree(const void *addr);
>
> static inline atomic_t *compound_mapcount_ptr(struct page *page)

It might be nice to include another patch that replaces some of the
existing/common uses of a*b+c with the new function...

Otherwise, yes, please. We could build a coccinelle rule for
additional replacements...

-Kees

--
Kees Cook
Pixel Security