net/sched/act_ctinfo.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /* net/sched/act_ctinfo.c  netfilter ctinfo connmark actions
   3  *
   4  * Copyright (c) 2019 Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
   5  */
   6
   7 #include <linux/module.h>
   8 #include <linux/init.h>
   9 #include <linux/kernel.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/rtnetlink.h>
  12 #include <linux/pkt_cls.h>
  13 #include <linux/ip.h>
  14 #include <linux/ipv6.h>
  15 #include <net/netlink.h>
  16 #include <net/pkt_sched.h>
  17 #include <net/act_api.h>
  18 #include <net/pkt_cls.h>
  19 #include <uapi/linux/tc_act/tc_ctinfo.h>
  20 #include <net/tc_act/tc_ctinfo.h>
  21
  22 #include <net/netfilter/nf_conntrack.h>
  23 #include <net/netfilter/nf_conntrack_core.h>
  24 #include <net/netfilter/nf_conntrack_ecache.h>
  25 #include <net/netfilter/nf_conntrack_zones.h>
  26
  27 static struct tc_action_ops act_ctinfo_ops;
  28 static unsigned int ctinfo_net_id;
  29
  30 static void tcf_ctinfo_dscp_set(struct nf_conn *ct, struct tcf_ctinfo *ca,
  31                                 struct tcf_ctinfo_params *cp,
  32                                 struct sk_buff *skb, int wlen, int proto)
  33 {
  34         u8 dscp, newdscp;
  35
  36         newdscp = (((ct->mark & cp->dscpmask) >> cp->dscpmaskshift) << 2) &
  37                      ~INET_ECN_MASK;
  38
  39         switch (proto) {
  40         case NFPROTO_IPV4:
  41                 dscp = ipv4_get_dsfield(ip_hdr(skb)) & ~INET_ECN_MASK;
  42                 if (dscp != newdscp) {
  43                         if (likely(!skb_try_make_writable(skb, wlen))) {
  44                                 ipv4_change_dsfield(ip_hdr(skb),
  45                                                     INET_ECN_MASK,
  46                                                     newdscp);
  47                                 ca->stats_dscp_set++;
  48                         } else {
  49                                 ca->stats_dscp_error++;
  50                         }
  51                 }
  52                 break;
  53         case NFPROTO_IPV6:
  54                 dscp = ipv6_get_dsfield(ipv6_hdr(skb)) & ~INET_ECN_MASK;
  55                 if (dscp != newdscp) {
  56                         if (likely(!skb_try_make_writable(skb, wlen))) {
  57                                 ipv6_change_dsfield(ipv6_hdr(skb),
  58                                                     INET_ECN_MASK,
  59                                                     newdscp);
  60                                 ca->stats_dscp_set++;
  61                         } else {
  62                                 ca->stats_dscp_error++;
  63                         }
  64                 }
  65                 break;
  66         default:
  67                 break;
  68         }
  69 }
  70
  71 static void tcf_ctinfo_cpmark_set(struct nf_conn *ct, struct tcf_ctinfo *ca,
  72                                   struct tcf_ctinfo_params *cp,
  73                                   struct sk_buff *skb)
  74 {
  75         ca->stats_cpmark_set++;
  76         skb->mark = ct->mark & cp->cpmarkmask;
  77 }
  78
  79 static int tcf_ctinfo_act(struct sk_buff *skb, const struct tc_action *a,
  80                           struct tcf_result *res)
  81 {
  82         const struct nf_conntrack_tuple_hash *thash = NULL;
  83         struct tcf_ctinfo *ca = to_ctinfo(a);
  84         struct nf_conntrack_tuple tuple;
  85         struct nf_conntrack_zone zone;
  86         enum ip_conntrack_info ctinfo;
  87         struct tcf_ctinfo_params *cp;
  88         struct nf_conn *ct;
  89         int proto, wlen;
  90         int action;
  91
  92         cp = rcu_dereference_bh(ca->params);
  93
  94         tcf_lastuse_update(&ca->tcf_tm);
  95         bstats_update(&ca->tcf_bstats, skb);
  96         action = READ_ONCE(ca->tcf_action);
  97
  98         wlen = skb_network_offset(skb);
  99         if (tc_skb_protocol(skb) == htons(ETH_P_IP)) {
 100                 wlen += sizeof(struct iphdr);
 101                 if (!pskb_may_pull(skb, wlen))
 102                         goto out;
 103
 104                 proto = NFPROTO_IPV4;
 105         } else if (tc_skb_protocol(skb) == htons(ETH_P_IPV6)) {
 106                 wlen += sizeof(struct ipv6hdr);
 107                 if (!pskb_may_pull(skb, wlen))
 108                         goto out;
 109
 110                 proto = NFPROTO_IPV6;
 111         } else {
 112                 goto out;
 113         }
 114
 115         ct = nf_ct_get(skb, &ctinfo);
 116         if (!ct) { /* look harder, usually ingress */
 117                 if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
 118                                        proto, cp->net, &tuple))
 119                         goto out;
 120                 zone.id = cp->zone;
 121                 zone.dir = NF_CT_DEFAULT_ZONE_DIR;
 122
 123                 thash = nf_conntrack_find_get(cp->net, &zone, &tuple);
 124                 if (!thash)
 125                         goto out;
 126
 127                 ct = nf_ct_tuplehash_to_ctrack(thash);
 128         }
 129
 130         if (cp->mode & CTINFO_MODE_DSCP)
 131                 if (!cp->dscpstatemask || (ct->mark & cp->dscpstatemask))
 132                         tcf_ctinfo_dscp_set(ct, ca, cp, skb, wlen, proto);
 133
 134         if (cp->mode & CTINFO_MODE_CPMARK)
 135                 tcf_ctinfo_cpmark_set(ct, ca, cp, skb);
 136
 137         if (thash)
 138                 nf_ct_put(ct);
 139 out:
 140         return action;
 141 }
 142
 143 static const struct nla_policy ctinfo_policy[TCA_CTINFO_MAX + 1] = {
 144         [TCA_CTINFO_ACT]                  = { .type = NLA_EXACT_LEN,
 145                                               .len = sizeof(struct
 146                                                             tc_ctinfo) },
 147         [TCA_CTINFO_ZONE]                 = { .type = NLA_U16 },
 148         [TCA_CTINFO_PARMS_DSCP_MASK]      = { .type = NLA_U32 },
 149         [TCA_CTINFO_PARMS_DSCP_STATEMASK] = { .type = NLA_U32 },
 150         [TCA_CTINFO_PARMS_CPMARK_MASK]    = { .type = NLA_U32 },
 151 };
 152
 153 static int tcf_ctinfo_init(struct net *net, struct nlattr *nla,
 154                            struct nlattr *est, struct tc_action **a,
 155                            int ovr, int bind, bool rtnl_held,
 156                            struct tcf_proto *tp,
 157                            struct netlink_ext_ack *extack)
 158 {
 159         struct tc_action_net *tn = net_generic(net, ctinfo_net_id);
 160         u32 dscpmask = 0, dscpstatemask, index;
 161         struct nlattr *tb[TCA_CTINFO_MAX + 1];
 162         struct tcf_ctinfo_params *cp_new;
 163         struct tcf_chain *goto_ch = NULL;
 164         struct tc_ctinfo *actparm;
 165         struct tcf_ctinfo *ci;
 166         u8 dscpmaskshift;
 167         int ret = 0, err;
 168
 169         if (!nla) {
 170                 NL_SET_ERR_MSG_MOD(extack, "ctinfo requires attributes to be passed");
 171                 return -EINVAL;
 172         }
 173
 174         err = nla_parse_nested(tb, TCA_CTINFO_MAX, nla, ctinfo_policy, extack);
 175         if (err < 0)
 176                 return err;
 177
 178         if (!tb[TCA_CTINFO_ACT]) {
 179                 NL_SET_ERR_MSG_MOD(extack,
 180                                    "Missing required TCA_CTINFO_ACT attribute");
 181                 return -EINVAL;
 182         }
 183         actparm = nla_data(tb[TCA_CTINFO_ACT]);
 184
 185         /* do some basic validation here before dynamically allocating things */
 186         /* that we would otherwise have to clean up.                          */
 187         if (tb[TCA_CTINFO_PARMS_DSCP_MASK]) {
 188                 dscpmask = nla_get_u32(tb[TCA_CTINFO_PARMS_DSCP_MASK]);
 189                 /* need contiguous 6 bit mask */
 190                 dscpmaskshift = dscpmask ? __ffs(dscpmask) : 0;
 191                 if ((~0 & (dscpmask >> dscpmaskshift)) != 0x3f) {
 192                         NL_SET_ERR_MSG_ATTR(extack,
 193                                             tb[TCA_CTINFO_PARMS_DSCP_MASK],
 194                                             "dscp mask must be 6 contiguous bits");
 195                         return -EINVAL;
 196                 }
 197                 dscpstatemask = tb[TCA_CTINFO_PARMS_DSCP_STATEMASK] ?
 198                         nla_get_u32(tb[TCA_CTINFO_PARMS_DSCP_STATEMASK]) : 0;
 199                 /* mask & statemask must not overlap */
 200                 if (dscpmask & dscpstatemask) {
 201                         NL_SET_ERR_MSG_ATTR(extack,
 202                                             tb[TCA_CTINFO_PARMS_DSCP_STATEMASK],
 203                                             "dscp statemask must not overlap dscp mask");
 204                         return -EINVAL;
 205                 }
 206         }
 207
 208         /* done the validation:now to the actual action allocation */
 209         index = actparm->index;
 210         err = tcf_idr_check_alloc(tn, &index, a, bind);
 211         if (!err) {
 212                 ret = tcf_idr_create(tn, index, est, a,
 213                                      &act_ctinfo_ops, bind, false);
 214                 if (ret) {
 215                         tcf_idr_cleanup(tn, index);
 216                         return ret;
 217                 }
 218                 ret = ACT_P_CREATED;
 219         } else if (err > 0) {
 220                 if (bind) /* don't override defaults */
 221                         return 0;
 222                 if (!ovr) {
 223                         tcf_idr_release(*a, bind);
 224                         return -EEXIST;
 225                 }
 226         } else {
 227                 return err;
 228         }
 229
 230         err = tcf_action_check_ctrlact(actparm->action, tp, &goto_ch, extack);
 231         if (err < 0)
 232                 goto release_idr;
 233
 234         ci = to_ctinfo(*a);
 235
 236         cp_new = kzalloc(sizeof(*cp_new), GFP_KERNEL);
 237         if (unlikely(!cp_new)) {
 238                 err = -ENOMEM;
 239                 goto put_chain;
 240         }
 241
 242         cp_new->net = net;
 243         cp_new->zone = tb[TCA_CTINFO_ZONE] ?
 244                         nla_get_u16(tb[TCA_CTINFO_ZONE]) : 0;
 245         if (dscpmask) {
 246                 cp_new->dscpmask = dscpmask;
 247                 cp_new->dscpmaskshift = dscpmaskshift;
 248                 cp_new->dscpstatemask = dscpstatemask;
 249                 cp_new->mode |= CTINFO_MODE_DSCP;
 250         }
 251
 252         if (tb[TCA_CTINFO_PARMS_CPMARK_MASK]) {
 253                 cp_new->cpmarkmask =
 254                                 nla_get_u32(tb[TCA_CTINFO_PARMS_CPMARK_MASK]);
 255                 cp_new->mode |= CTINFO_MODE_CPMARK;
 256         }
 257
 258         spin_lock_bh(&ci->tcf_lock);
 259         goto_ch = tcf_action_set_ctrlact(*a, actparm->action, goto_ch);
 260         rcu_swap_protected(ci->params, cp_new,
 261                            lockdep_is_held(&ci->tcf_lock));
 262         spin_unlock_bh(&ci->tcf_lock);
 263
 264         if (goto_ch)
 265                 tcf_chain_put_by_act(goto_ch);
 266         if (cp_new)
 267                 kfree_rcu(cp_new, rcu);
 268
 269         if (ret == ACT_P_CREATED)
 270                 tcf_idr_insert(tn, *a);
 271
 272         return ret;
 273
 274 put_chain:
 275         if (goto_ch)
 276                 tcf_chain_put_by_act(goto_ch);
 277 release_idr:
 278         tcf_idr_release(*a, bind);
 279         return err;
 280 }
 281
 282 static int tcf_ctinfo_dump(struct sk_buff *skb, struct tc_action *a,
 283                            int bind, int ref)
 284 {
 285         struct tcf_ctinfo *ci = to_ctinfo(a);
 286         struct tc_ctinfo opt = {
 287                 .index   = ci->tcf_index,
 288                 .refcnt  = refcount_read(&ci->tcf_refcnt) - ref,
 289                 .bindcnt = atomic_read(&ci->tcf_bindcnt) - bind,
 290         };
 291         unsigned char *b = skb_tail_pointer(skb);
 292         struct tcf_ctinfo_params *cp;
 293         struct tcf_t t;
 294
 295         spin_lock_bh(&ci->tcf_lock);
 296         cp = rcu_dereference_protected(ci->params,
 297                                        lockdep_is_held(&ci->tcf_lock));
 298
 299         tcf_tm_dump(&t, &ci->tcf_tm);
 300         if (nla_put_64bit(skb, TCA_CTINFO_TM, sizeof(t), &t, TCA_CTINFO_PAD))
 301                 goto nla_put_failure;
 302
 303         opt.action = ci->tcf_action;
 304         if (nla_put(skb, TCA_CTINFO_ACT, sizeof(opt), &opt))
 305                 goto nla_put_failure;
 306
 307         if (nla_put_u16(skb, TCA_CTINFO_ZONE, cp->zone))
 308                 goto nla_put_failure;
 309
 310         if (cp->mode & CTINFO_MODE_DSCP) {
 311                 if (nla_put_u32(skb, TCA_CTINFO_PARMS_DSCP_MASK,
 312                                 cp->dscpmask))
 313                         goto nla_put_failure;
 314                 if (nla_put_u32(skb, TCA_CTINFO_PARMS_DSCP_STATEMASK,
 315                                 cp->dscpstatemask))
 316                         goto nla_put_failure;
 317         }
 318
 319         if (cp->mode & CTINFO_MODE_CPMARK) {
 320                 if (nla_put_u32(skb, TCA_CTINFO_PARMS_CPMARK_MASK,
 321                                 cp->cpmarkmask))
 322                         goto nla_put_failure;
 323         }
 324
 325         if (nla_put_u64_64bit(skb, TCA_CTINFO_STATS_DSCP_SET,
 326                               ci->stats_dscp_set, TCA_CTINFO_PAD))
 327                 goto nla_put_failure;
 328
 329         if (nla_put_u64_64bit(skb, TCA_CTINFO_STATS_DSCP_ERROR,
 330                               ci->stats_dscp_error, TCA_CTINFO_PAD))
 331                 goto nla_put_failure;
 332
 333         if (nla_put_u64_64bit(skb, TCA_CTINFO_STATS_CPMARK_SET,
 334                               ci->stats_cpmark_set, TCA_CTINFO_PAD))
 335                 goto nla_put_failure;
 336
 337         spin_unlock_bh(&ci->tcf_lock);
 338         return skb->len;
 339
 340 nla_put_failure:
 341         spin_unlock_bh(&ci->tcf_lock);
 342         nlmsg_trim(skb, b);
 343         return -1;
 344 }
 345
 346 static int tcf_ctinfo_walker(struct net *net, struct sk_buff *skb,
 347                              struct netlink_callback *cb, int type,
 348                              const struct tc_action_ops *ops,
 349                              struct netlink_ext_ack *extack)
 350 {
 351         struct tc_action_net *tn = net_generic(net, ctinfo_net_id);
 352
 353         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 354 }
 355
 356 static int tcf_ctinfo_search(struct net *net, struct tc_action **a, u32 index)
 357 {
 358         struct tc_action_net *tn = net_generic(net, ctinfo_net_id);
 359
 360         return tcf_idr_search(tn, a, index);
 361 }
 362
 363 static struct tc_action_ops act_ctinfo_ops = {
 364         .kind   = "ctinfo",
 365         .id     = TCA_ID_CTINFO,
 366         .owner  = THIS_MODULE,
 367         .act    = tcf_ctinfo_act,
 368         .dump   = tcf_ctinfo_dump,
 369         .init   = tcf_ctinfo_init,
 370         .walk   = tcf_ctinfo_walker,
 371         .lookup = tcf_ctinfo_search,
 372         .size   = sizeof(struct tcf_ctinfo),
 373 };
 374
 375 static __net_init int ctinfo_init_net(struct net *net)
 376 {
 377         struct tc_action_net *tn = net_generic(net, ctinfo_net_id);
 378
 379         return tc_action_net_init(tn, &act_ctinfo_ops);
 380 }
 381
 382 static void __net_exit ctinfo_exit_net(struct list_head *net_list)
 383 {
 384         tc_action_net_exit(net_list, ctinfo_net_id);
 385 }
 386
 387 static struct pernet_operations ctinfo_net_ops = {
 388         .init           = ctinfo_init_net,
 389         .exit_batch     = ctinfo_exit_net,
 390         .id             = &ctinfo_net_id,
 391         .size           = sizeof(struct tc_action_net),
 392 };
 393
 394 static int __init ctinfo_init_module(void)
 395 {
 396         return tcf_register_action(&act_ctinfo_ops, &ctinfo_net_ops);
 397 }
 398
 399 static void __exit ctinfo_cleanup_module(void)
 400 {
 401         tcf_unregister_action(&act_ctinfo_ops, &ctinfo_net_ops);
 402 }
 403
 404 module_init(ctinfo_init_module);
 405 module_exit(ctinfo_cleanup_module);
 406 MODULE_AUTHOR("Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>");
 407 MODULE_DESCRIPTION("Connection tracking mark actions");
 408 MODULE_LICENSE("GPL");