Re: [PATCH] skbuff: Reallocate to ksize() in __build_skb_around()

[Jakub Kicinski]

On Tue,  6 Dec 2022 15:17:14 -0800 Kees Cook wrote:
> -	unsigned int size = frag_size ? : ksize(data);
> +	unsigned int size = frag_size;
> +
> +	/* When frag_size == 0, the buffer came from kmalloc, so we
> +	 * must find its true allocation size (and grow it to match).
> +	 */
> +	if (unlikely(size == 0)) {
> +		void *resized;
> +
> +		size = ksize(data);
> +		/* krealloc() will immediate return "data" when
> +		 * "ksize(data)" is requested: it is the existing upper
> +		 * bounds. As a result, GFP_ATOMIC will be ignored.
> +		 */
> +		resized = krealloc(data, size, GFP_ATOMIC);
> +		if (WARN_ON(resized != data))
> +			data = resized;
> +	}
>  

Aammgh. build_skb(0) is plain silly, AFAIK. The performance hit of
using kmalloc()'ed heads is large because GRO can't free the metadata.
So we end up carrying per-MTU skbs across to the application and then
freeing them one by one. With pages we just aggregate up to 64k of data
in a single skb.

I can only grep out 3 cases of build_skb(.. 0), could we instead
convert them into a new build_skb_slab(), and handle all the silliness
in such a new helper? That'd be a win both for the memory safety and one
fewer branch for the fast path.

I think it's worth doing, so LMK if you're okay to do this extra work,
otherwise I can help (unless e.g. Eric tells me I'm wrong..).