Re: [PATCH 08/12] kasan, mm: optimize krealloc poisoning

[Marco Elver <elver@google.com>]

On Mon, Feb 01, 2021 at 08:43PM +0100, Andrey Konovalov wrote:
> Currently, krealloc() always calls ksize(), which unpoisons the whole
> object including the redzone. This is inefficient, as kasan_krealloc()
> repoisons the redzone for objects that fit into the same buffer.
> 
> This patch changes krealloc() instrumentation to use uninstrumented
> __ksize() that doesn't unpoison the memory. Instead, kasan_kreallos()
> is changed to unpoison the memory excluding the redzone.
> 
> For objects that don't fit into the old allocation, this patch disables
> KASAN accessibility checks when copying memory into a new object instead
> of unpoisoning it.
> 
> Signed-off-by: Andrey Konovalov <andreyknvl@google.com>
> ---
>  mm/kasan/common.c | 12 ++++++++++--
>  mm/slab_common.c  | 20 ++++++++++++++------
>  2 files changed, 24 insertions(+), 8 deletions(-)
> 
> diff --git a/mm/kasan/common.c b/mm/kasan/common.c
> index 9c64a00bbf9c..a51d6ea580b0 100644
> --- a/mm/kasan/common.c
> +++ b/mm/kasan/common.c
> @@ -476,7 +476,7 @@ static void *____kasan_kmalloc(struct kmem_cache *cache, const void *object,
>  
>  	/*
>  	 * The object has already been unpoisoned by kasan_slab_alloc() for
> -	 * kmalloc() or by ksize() for krealloc().
> +	 * kmalloc() or by kasan_krealloc() for krealloc().
>  	 */
>  
>  	/*
> @@ -526,7 +526,7 @@ void * __must_check __kasan_kmalloc_large(const void *ptr, size_t size,
>  
>  	/*
>  	 * The object has already been unpoisoned by kasan_alloc_pages() for
> -	 * alloc_pages() or by ksize() for krealloc().
> +	 * alloc_pages() or by kasan_krealloc() for krealloc().
>  	 */
>  
>  	/*
> @@ -554,8 +554,16 @@ void * __must_check __kasan_krealloc(const void *object, size_t size, gfp_t flag
>  	if (unlikely(object == ZERO_SIZE_PTR))
>  		return (void *)object;
>  
> +	/*
> +	 * Unpoison the object's data.
> +	 * Part of it might already have been unpoisoned, but it's unknown
> +	 * how big that part is.
> +	 */
> +	kasan_unpoison(object, size);
> +
>  	page = virt_to_head_page(object);
>  
> +	/* Piggy-back on kmalloc() instrumentation to poison the redzone. */
>  	if (unlikely(!PageSlab(page)))
>  		return __kasan_kmalloc_large(object, size, flags);
>  	else
> diff --git a/mm/slab_common.c b/mm/slab_common.c
> index dad70239b54c..821f657d38b5 100644
> --- a/mm/slab_common.c
> +++ b/mm/slab_common.c
> @@ -1140,19 +1140,27 @@ static __always_inline void *__do_krealloc(const void *p, size_t new_size,
>  	void *ret;
>  	size_t ks;
>  
> -	if (likely(!ZERO_OR_NULL_PTR(p)) && !kasan_check_byte(p))
> -		return NULL;
> -
> -	ks = ksize(p);
> +	/* Don't use instrumented ksize to allow precise KASAN poisoning. */
> +	if (likely(!ZERO_OR_NULL_PTR(p))) {
> +		if (!kasan_check_byte(p))
> +			return NULL;
> +		ks = __ksize(p);
> +	} else
> +		ks = 0;
>  

This unfortunately broke KFENCE:
https://syzkaller.appspot.com/bug?extid=e444e1006d07feef0ef3 + various
other false positives.

We need to use ksize() here, as __ksize() is unaware of KFENCE. Or
somehow add the same check here that ksize() uses to get the real object
size.

> +	/* If the object still fits, repoison it precisely. */
>  	if (ks >= new_size) {
>  		p = kasan_krealloc((void *)p, new_size, flags);
>  		return (void *)p;
>  	}
>  
>  	ret = kmalloc_track_caller(new_size, flags);
> -	if (ret && p)
> -		memcpy(ret, p, ks);
> +	if (ret && p) {
> +		/* Disable KASAN checks as the object's redzone is accessed. */
> +		kasan_disable_current();
> +		memcpy(ret, kasan_reset_tag(p), ks);
> +		kasan_enable_current();
> +	}
>  
>  	return ret;
>  }
> -- 
> 2.30.0.365.g02bc693789-goog
>