Re: [PATCH nf-next 3/5] nft_set_pipapo: Prepare for vectorised implementation: alignment

[Florian Westphal <fw@strlen.de>]

Stefano Brivio <sbrivio@redhat.com> wrote:
>  struct nft_pipapo_field {
> @@ -439,6 +456,9 @@ struct nft_pipapo_field {
>  	unsigned long rules;
>  	size_t bsize;
>  	int bb;
> +#ifdef NFT_PIPAPO_ALIGN
> +	unsigned long *lt_aligned;
> +#endif
>  	unsigned long *lt;
>  	union nft_pipapo_map_bucket *mt;
>  };

I wonder if these structs can be compressed.
AFAICS bsize is in sizes of longs, so when this number is
large then we also need to kvmalloc a large blob of memory.

I think u32 would be enough?

nft_pipapo_field is probably the most relevant one wrt. to size.

>  struct nft_pipapo_match {
>  	int field_count;
> +#ifdef NFT_PIPAPO_ALIGN
> +	unsigned long * __percpu *scratch_aligned;
> +#endif
>  	unsigned long * __percpu *scratch;
>  	size_t bsize_max;

Same here (bsize_max -- could fit with hole after field_count)?

Also, since you know the size of nft_pipapo_match (including the
dynamically allocated array at the end), you could store the
original memory (*scratch) and the rcu_head at the end, since
they are not needed at lookup time and a little overhead to calculate
their storage offset is fine.

Not sure its worth it, just an idea.