Re: [PATCH v2 02/16] slab: Introduce kmalloc_size_roundup()

[Hyeonggon Yoo]

On Fri, Sep 23, 2022 at 01:28:08PM -0700, Kees Cook wrote:
> In the effort to help the compiler reason about buffer sizes, the
> __alloc_size attribute was added to allocators. This improves the scope
> of the compiler's ability to apply CONFIG_UBSAN_BOUNDS and (in the near
> future) CONFIG_FORTIFY_SOURCE. For most allocations, this works well,
> as the vast majority of callers are not expecting to use more memory
> than what they asked for.
> 
> There is, however, one common exception to this: anticipatory resizing
> of kmalloc allocations. These cases all use ksize() to determine the
> actual bucket size of a given allocation (e.g. 128 when 126 was asked
> for). This comes in two styles in the kernel:
> 
> 1) An allocation has been determined to be too small, and needs to be
>    resized. Instead of the caller choosing its own next best size, it
>    wants to minimize the number of calls to krealloc(), so it just uses
>    ksize() plus some additional bytes, forcing the realloc into the next
>    bucket size, from which it can learn how large it is now. For example:
> 
> 	data = krealloc(data, ksize(data) + 1, gfp);
> 	data_len = ksize(data);
> 
> 2) The minimum size of an allocation is calculated, but since it may
>    grow in the future, just use all the space available in the chosen
>    bucket immediately, to avoid needing to reallocate later. A good
>    example of this is skbuff's allocators:
> 
> 	data = kmalloc_reserve(size, gfp_mask, node, &pfmemalloc);
> 	...
> 	/* kmalloc(size) might give us more room than requested.
> 	 * Put skb_shared_info exactly at the end of allocated zone,
> 	 * to allow max possible filling before reallocation.
> 	 */
> 	osize = ksize(data);
>         size = SKB_WITH_OVERHEAD(osize);
> 
> In both cases, the "how much was actually allocated?" question is answered
> _after_ the allocation, where the compiler hinting is not in an easy place
> to make the association any more. This mismatch between the compiler's
> view of the buffer length and the code's intention about how much it is
> going to actually use has already caused problems[1]. It is possible to
> fix this by reordering the use of the "actual size" information.
> 
> We can serve the needs of users of ksize() and still have accurate buffer
> length hinting for the compiler by doing the bucket size calculation
> _before_ the allocation. Code can instead ask "how large an allocation
> would I get for a given size?".
> 
> Introduce kmalloc_size_roundup(), to serve this function so we can start
> replacing the "anticipatory resizing" uses of ksize().
> 
> [1] https://github.com/ClangBuiltLinux/linux/issues/1599
>     https://github.com/KSPP/linux/issues/183
> 
> Cc: Vlastimil Babka <vbabka@xxxxxxx>
> Cc: Christoph Lameter <cl@xxxxxxxxx>
> Cc: Pekka Enberg <penberg@xxxxxxxxxx>
> Cc: David Rientjes <rientjes@xxxxxxxxxx>
> Cc: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: linux-mm@xxxxxxxxx
> Signed-off-by: Kees Cook <keescook@xxxxxxxxxxxx>
> ---
>  include/linux/slab.h | 31 +++++++++++++++++++++++++++++++
>  mm/slab.c            |  9 ++++++---
>  mm/slab_common.c     | 20 ++++++++++++++++++++
>  3 files changed, 57 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/slab.h b/include/linux/slab.h
> index 41bd036e7551..727640173568 100644
> --- a/include/linux/slab.h
> +++ b/include/linux/slab.h
> @@ -188,7 +188,21 @@ void * __must_check krealloc(const void *objp, size_t new_size, gfp_t flags) __r
>  void kfree(const void *objp);
>  void kfree_sensitive(const void *objp);
>  size_t __ksize(const void *objp);
> +
> +/**
> + * ksize - Report actual allocation size of associated object
> + *
> + * @objp: Pointer returned from a prior kmalloc()-family allocation.
> + *
> + * This should not be used for writing beyond the originally requested
> + * allocation size. Either use krealloc() or round up the allocation size
> + * with kmalloc_size_roundup() prior to allocation. If this is used to
> + * access beyond the originally requested allocation size, UBSAN_BOUNDS
> + * and/or FORTIFY_SOURCE may trip, since they only know about the
> + * originally allocated size via the __alloc_size attribute.
> + */
>  size_t ksize(const void *objp);
> +

With this now we have two conflicting kernel-doc comments
about ksize in mm/slab_common.c and include/linux/slab.h.

>  #ifdef CONFIG_PRINTK
>  bool kmem_valid_obj(void *object);
>  void kmem_dump_obj(void *object);
> @@ -779,6 +793,23 @@ extern void kvfree(const void *addr);
>  extern void kvfree_sensitive(const void *addr, size_t len);
>  
>  unsigned int kmem_cache_size(struct kmem_cache *s);
> +
> +/**
> + * kmalloc_size_roundup - Report allocation bucket size for the given size
> + *
> + * @size: Number of bytes to round up from.
> + *
> + * This returns the number of bytes that would be available in a kmalloc()
> + * allocation of @size bytes. For example, a 126 byte request would be
> + * rounded up to the next sized kmalloc bucket, 128 bytes. (This is strictly
> + * for the general-purpose kmalloc()-based allocations, and is not for the
> + * pre-sized kmem_cache_alloc()-based allocations.)
> + *
> + * Use this to kmalloc() the full bucket size ahead of time instead of using
> + * ksize() to query the size after an allocation.
> + */
> +size_t kmalloc_size_roundup(size_t size);
> +
>  void __init kmem_cache_init_late(void);
>  
>  #if defined(CONFIG_SMP) && defined(CONFIG_SLAB)
> diff --git a/mm/slab.c b/mm/slab.c
> index 10e96137b44f..2da862bf6226 100644
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -4192,11 +4192,14 @@ void __check_heap_object(const void *ptr, unsigned long n,
>  #endif /* CONFIG_HARDENED_USERCOPY */
>  
>  /**
> - * __ksize -- Uninstrumented ksize.
> + * __ksize -- Report full size of underlying allocation
>   * @objp: pointer to the object
>   *
> - * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
> - * safety checks as ksize() with KASAN instrumentation enabled.
> + * This should only be used internally to query the true size of allocations.
> + * It is not meant to be a way to discover the usable size of an allocation
> + * after the fact. Instead, use kmalloc_size_roundup(). Using memory beyond
> + * the originally requested allocation size may trigger KASAN, UBSAN_BOUNDS,
> + * and/or FORTIFY_SOURCE.
>   *
>   * Return: size of the actual memory used by @objp in bytes
>   */
> diff --git a/mm/slab_common.c b/mm/slab_common.c
> index 457671ace7eb..d7420cf649f8 100644
> --- a/mm/slab_common.c
> +++ b/mm/slab_common.c
> @@ -721,6 +721,26 @@ struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
>  	return kmalloc_caches[kmalloc_type(flags)][index];
>  }
>  
> +size_t kmalloc_size_roundup(size_t size)
> +{
> +	struct kmem_cache *c;
> +
> +	/* Short-circuit the 0 size case. */
> +	if (unlikely(size == 0))
> +		return 0;
> +	/* Short-circuit saturated "too-large" case. */
> +	if (unlikely(size == SIZE_MAX))
> +		return SIZE_MAX;
> +	/* Above the smaller buckets, size is a multiple of page size. */
> +	if (size > KMALLOC_MAX_CACHE_SIZE)
> +		return PAGE_SIZE << get_order(size);
> +
> +	/* The flags don't matter since size_index is common to all. */
> +	c = kmalloc_slab(size, GFP_KERNEL);
> +	return c ? c->object_size : 0;
> +}
> +EXPORT_SYMBOL(kmalloc_size_roundup);
> +
>  #ifdef CONFIG_ZONE_DMA
>  #define KMALLOC_DMA_NAME(sz)	.name[KMALLOC_DMA] = "dma-kmalloc-" #sz,
>  #else
> -- 
> 2.34.1

Otherwise looks good!

-- 
Thanks,
Hyeonggon