[Devel] Re: [NETFILTER] early_drop() imrovement (v4)

[Vasily Averin]

Patrick McHardy wrote:
> Vasily Averin wrote:
>> it is incorrect,
>> We should count the number of checked _conntracks_, but you count the number of
>> hash buckets. I.e "i" should be incremented/checked inside the nested loop.
> 
> 
> I misunderstood your patch then. This one should be better.
> +static int early_drop(unsigned int hash)
>  {
>  	/* Use oldest entry, which is roughly LRU */
>  	struct nf_conntrack_tuple_hash *h;
>  	struct nf_conn *ct = NULL, *tmp;
>  	struct hlist_node *n;
> -	int dropped = 0;
> +	unsigned int i;
> +	int dropped = 0, cnt = NF_CT_EVICTION_RANGE;
>  
>  	read_lock_bh(&nf_conntrack_lock);
> -	hlist_for_each_entry(h, n, chain, hnode) {
> -		tmp = nf_ct_tuplehash_to_ctrack(h);
> -		if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
> -			ct = tmp;
> +	for (i = 0; i < nf_conntrack_htable_size; i++) {
> +		hlist_for_each_entry(h, n, &nf_conntrack_hash[hash], hnode) {
> +			tmp = nf_ct_tuplehash_to_ctrack(h);
> +			if (!test_bit(IPS_ASSURED_BIT, &tmp->status))
> +				ct = tmp;
> +			if (--cnt <= 0)
> +				break;
> +		}
> +		hash = (hash + 1) % nf_conntrack_htable_size;
>  	}

it is incorrect again: when cnt=0 you should break both cycles.

thank you,
	Vasily Averin