Re: [PATCH] netfilter: nfnetlink_queue: optimize verdict lookup with hash table

[Florian Westphal]

Scott Mitchell <scott.k.mitch1@xxxxxxxxx> wrote:
>  static inline u_int8_t instance_hashfn(u_int16_t queue_num)
>  {
>  	return ((queue_num >> 8) ^ queue_num) % INSTANCE_BUCKETS;
> @@ -114,13 +153,63 @@ instance_lookup(struct nfnl_queue_net *q, u_int16_t queue_num)
>  	return NULL;
>  }
>  
> +static int
> +nfqnl_hash_resize(struct nfqnl_instance *inst, u32 hash_size)
> +{
> +	struct hlist_head *new_hash, *old_hash;
> +	struct nf_queue_entry *entry;
> +	unsigned int h, hash_mask;
> +
> +	/* lock scope includes kcalloc/kfree to bound memory if concurrent resizes.
> +	 * lock scope could be reduced to exclude the  kcalloc/kfree at the cost
> +	 * of increased code complexity (re-check of hash_size) and relaxed memory
> +	 * bounds (concurrent resize may each do allocations). since resize is
> +	 * expected to be rare, the broader lock scope is simpler and preferred.
> +	 */

I'm all for simplicity. but I don't see how concurrent resizes are
possible.  NFQNL_MSG_CONFIG runs under nfnetlink subsystem mutex.

Or did I miss something?

> +	new_hash = kcalloc(hash_size, sizeof(*new_hash), GFP_ATOMIC);

Since the hash table could be large I would prefer if this could
be GFP_KERNEL_ACCOUNT + kvcalloc to permit vmalloc fallback.

> +	if (nfqa[NFQA_CFG_HASH_SIZE]) {
> +		hash_size = ntohl(nla_get_be32(nfqa[NFQA_CFG_HASH_SIZE]));
> +	}

Nit, no { } here.