net/netfilter/xt_cluster.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * (C) 2008-2009 Pablo Neira Ayuso <pablo@netfilter.org>
0004  */
0005 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0006 #include <linux/module.h>
0007 #include <linux/skbuff.h>
0008 #include <linux/jhash.h>
0009 #include <linux/ip.h>
0010 #include <net/ipv6.h>
0011
0012 #include <linux/netfilter/x_tables.h>
0013 #include <net/netfilter/nf_conntrack.h>
0014 #include <linux/netfilter/xt_cluster.h>
0015
0016 static inline u32 nf_ct_orig_ipv4_src(const struct nf_conn *ct)
0017 {
0018     return (__force u32)ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
0019 }
0020
0021 static inline const u32 *nf_ct_orig_ipv6_src(const struct nf_conn *ct)
0022 {
0023     return (__force u32 *)ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6;
0024 }
0025
0026 static inline u_int32_t
0027 xt_cluster_hash_ipv4(u_int32_t ip, const struct xt_cluster_match_info *info)
0028 {
0029     return jhash_1word(ip, info->hash_seed);
0030 }
0031
0032 static inline u_int32_t
0033 xt_cluster_hash_ipv6(const void *ip, const struct xt_cluster_match_info *info)
0034 {
0035     return jhash2(ip, NF_CT_TUPLE_L3SIZE / sizeof(__u32), info->hash_seed);
0036 }
0037
0038 static inline u_int32_t
0039 xt_cluster_hash(const struct nf_conn *ct,
0040         const struct xt_cluster_match_info *info)
0041 {
0042     u_int32_t hash = 0;
0043
0044     switch(nf_ct_l3num(ct)) {
0045     case AF_INET:
0046         hash = xt_cluster_hash_ipv4(nf_ct_orig_ipv4_src(ct), info);
0047         break;
0048     case AF_INET6:
0049         hash = xt_cluster_hash_ipv6(nf_ct_orig_ipv6_src(ct), info);
0050         break;
0051     default:
0052         WARN_ON(1);
0053         break;
0054     }
0055
0056     return reciprocal_scale(hash, info->total_nodes);
0057 }
0058
0059 static inline bool
0060 xt_cluster_is_multicast_addr(const struct sk_buff *skb, u_int8_t family)
0061 {
0062     bool is_multicast = false;
0063
0064     switch(family) {
0065     case NFPROTO_IPV4:
0066         is_multicast = ipv4_is_multicast(ip_hdr(skb)->daddr);
0067         break;
0068     case NFPROTO_IPV6:
0069         is_multicast = ipv6_addr_is_multicast(&ipv6_hdr(skb)->daddr);
0070         break;
0071     default:
0072         WARN_ON(1);
0073         break;
0074     }
0075     return is_multicast;
0076 }
0077
0078 static bool
0079 xt_cluster_mt(const struct sk_buff *skb, struct xt_action_param *par)
0080 {
0081     struct sk_buff *pskb = (struct sk_buff *)skb;
0082     const struct xt_cluster_match_info *info = par->matchinfo;
0083     const struct nf_conn *ct;
0084     enum ip_conntrack_info ctinfo;
0085     unsigned long hash;
0086
0087     /* This match assumes that all nodes see the same packets. This can be
0088      * achieved if the switch that connects the cluster nodes support some
0089      * sort of 'port mirroring'. However, if your switch does not support
0090      * this, your cluster nodes can reply ARP request using a multicast MAC
0091      * address. Thus, your switch will flood the same packets to the
0092      * cluster nodes with the same multicast MAC address. Using a multicast
0093      * link address is a RFC 1812 (section 3.3.2) violation, but this works
0094      * fine in practise.
0095      *
0096      * Unfortunately, if you use the multicast MAC address, the link layer
0097      * sets skbuff's pkt_type to PACKET_MULTICAST, which is not accepted
0098      * by TCP and others for packets coming to this node. For that reason,
0099      * this match mangles skbuff's pkt_type if it detects a packet
0100      * addressed to a unicast address but using PACKET_MULTICAST. Yes, I
0101      * know, matches should not alter packets, but we are doing this here
0102      * because we would need to add a PKTTYPE target for this sole purpose.
0103      */
0104     if (!xt_cluster_is_multicast_addr(skb, xt_family(par)) &&
0105         skb->pkt_type == PACKET_MULTICAST) {
0106             pskb->pkt_type = PACKET_HOST;
0107     }
0108
0109     ct = nf_ct_get(skb, &ctinfo);
0110     if (ct == NULL)
0111         return false;
0112
0113     if (ct->master)
0114         hash = xt_cluster_hash(ct->master, info);
0115     else
0116         hash = xt_cluster_hash(ct, info);
0117
0118     return !!((1 << hash) & info->node_mask) ^
0119            !!(info->flags & XT_CLUSTER_F_INV);
0120 }
0121
0122 static int xt_cluster_mt_checkentry(const struct xt_mtchk_param *par)
0123 {
0124     struct xt_cluster_match_info *info = par->matchinfo;
0125     int ret;
0126
0127     if (info->total_nodes > XT_CLUSTER_NODES_MAX) {
0128         pr_info_ratelimited("you have exceeded the maximum number of cluster nodes (%u > %u)\n",
0129                     info->total_nodes, XT_CLUSTER_NODES_MAX);
0130         return -EINVAL;
0131     }
0132     if (info->node_mask >= (1ULL << info->total_nodes)) {
0133         pr_info_ratelimited("node mask cannot exceed total number of nodes\n");
0134         return -EDOM;
0135     }
0136
0137     ret = nf_ct_netns_get(par->net, par->family);
0138     if (ret < 0)
0139         pr_info_ratelimited("cannot load conntrack support for proto=%u\n",
0140                     par->family);
0141     return ret;
0142 }
0143
0144 static void xt_cluster_mt_destroy(const struct xt_mtdtor_param *par)
0145 {
0146     nf_ct_netns_put(par->net, par->family);
0147 }
0148
0149 static struct xt_match xt_cluster_match __read_mostly = {
0150     .name       = "cluster",
0151     .family     = NFPROTO_UNSPEC,
0152     .match      = xt_cluster_mt,
0153     .checkentry = xt_cluster_mt_checkentry,
0154     .matchsize  = sizeof(struct xt_cluster_match_info),
0155     .destroy    = xt_cluster_mt_destroy,
0156     .me     = THIS_MODULE,
0157 };
0158
0159 static int __init xt_cluster_mt_init(void)
0160 {
0161     return xt_register_match(&xt_cluster_match);
0162 }
0163
0164 static void __exit xt_cluster_mt_fini(void)
0165 {
0166     xt_unregister_match(&xt_cluster_match);
0167 }
0168
0169 MODULE_AUTHOR("Pablo Neira Ayuso <pablo@netfilter.org>");
0170 MODULE_LICENSE("GPL");
0171 MODULE_DESCRIPTION("Xtables: hash-based cluster match");
0172 MODULE_ALIAS("ipt_cluster");
0173 MODULE_ALIAS("ip6t_cluster");
0174 module_init(xt_cluster_mt_init);
0175 module_exit(xt_cluster_mt_fini);