net/sched/act_mirred.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/act_mirred.c       packet mirroring and redirect actions
   4  *
   5  * Authors:     Jamal Hadi Salim (2002-4)
   6  *
   7  * TODO: Add ingress support (and socket redirect support)
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/kernel.h>
  12 #include <linux/string.h>
  13 #include <linux/errno.h>
  14 #include <linux/skbuff.h>
  15 #include <linux/rtnetlink.h>
  16 #include <linux/module.h>
  17 #include <linux/init.h>
  18 #include <linux/gfp.h>
  19 #include <linux/if_arp.h>
  20 #include <net/net_namespace.h>
  21 #include <net/netlink.h>
  22 #include <net/pkt_sched.h>
  23 #include <net/pkt_cls.h>
  24 #include <linux/tc_act/tc_mirred.h>
  25 #include <net/tc_act/tc_mirred.h>
  26
  27 static LIST_HEAD(mirred_list);
  28 static DEFINE_SPINLOCK(mirred_list_lock);
  29
  30 #define MIRRED_RECURSION_LIMIT    4
  31 static DEFINE_PER_CPU(unsigned int, mirred_rec_level);
  32
  33 static bool tcf_mirred_is_act_redirect(int action)
  34 {
  35         return action == TCA_EGRESS_REDIR || action == TCA_INGRESS_REDIR;
  36 }
  37
  38 static bool tcf_mirred_act_wants_ingress(int action)
  39 {
  40         switch (action) {
  41         case TCA_EGRESS_REDIR:
  42         case TCA_EGRESS_MIRROR:
  43                 return false;
  44         case TCA_INGRESS_REDIR:
  45         case TCA_INGRESS_MIRROR:
  46                 return true;
  47         default:
  48                 BUG();
  49         }
  50 }
  51
  52 static bool tcf_mirred_can_reinsert(int action)
  53 {
  54         switch (action) {
  55         case TC_ACT_SHOT:
  56         case TC_ACT_STOLEN:
  57         case TC_ACT_QUEUED:
  58         case TC_ACT_TRAP:
  59                 return true;
  60         }
  61         return false;
  62 }
  63
  64 static struct net_device *tcf_mirred_dev_dereference(struct tcf_mirred *m)
  65 {
  66         return rcu_dereference_protected(m->tcfm_dev,
  67                                          lockdep_is_held(&m->tcf_lock));
  68 }
  69
  70 static void tcf_mirred_release(struct tc_action *a)
  71 {
  72         struct tcf_mirred *m = to_mirred(a);
  73         struct net_device *dev;
  74
  75         spin_lock(&mirred_list_lock);
  76         list_del(&m->tcfm_list);
  77         spin_unlock(&mirred_list_lock);
  78
  79         /* last reference to action, no need to lock */
  80         dev = rcu_dereference_protected(m->tcfm_dev, 1);
  81         if (dev)
  82                 dev_put(dev);
  83 }
  84
  85 static const struct nla_policy mirred_policy[TCA_MIRRED_MAX + 1] = {
  86         [TCA_MIRRED_PARMS]      = { .len = sizeof(struct tc_mirred) },
  87 };
  88
  89 static unsigned int mirred_net_id;
  90 static struct tc_action_ops act_mirred_ops;
  91
  92 static int tcf_mirred_init(struct net *net, struct nlattr *nla,
  93                            struct nlattr *est, struct tc_action **a,
  94                            int ovr, int bind, bool rtnl_held,
  95                            struct tcf_proto *tp,
  96                            u32 flags, struct netlink_ext_ack *extack)
  97 {
  98         struct tc_action_net *tn = net_generic(net, mirred_net_id);
  99         struct nlattr *tb[TCA_MIRRED_MAX + 1];
 100         struct tcf_chain *goto_ch = NULL;
 101         bool mac_header_xmit = false;
 102         struct tc_mirred *parm;
 103         struct tcf_mirred *m;
 104         struct net_device *dev;
 105         bool exists = false;
 106         int ret, err;
 107         u32 index;
 108
 109         if (!nla) {
 110                 NL_SET_ERR_MSG_MOD(extack, "Mirred requires attributes to be passed");
 111                 return -EINVAL;
 112         }
 113         ret = nla_parse_nested_deprecated(tb, TCA_MIRRED_MAX, nla,
 114                                           mirred_policy, extack);
 115         if (ret < 0)
 116                 return ret;
 117         if (!tb[TCA_MIRRED_PARMS]) {
 118                 NL_SET_ERR_MSG_MOD(extack, "Missing required mirred parameters");
 119                 return -EINVAL;
 120         }
 121         parm = nla_data(tb[TCA_MIRRED_PARMS]);
 122         index = parm->index;
 123         err = tcf_idr_check_alloc(tn, &index, a, bind);
 124         if (err < 0)
 125                 return err;
 126         exists = err;
 127         if (exists && bind)
 128                 return 0;
 129
 130         switch (parm->eaction) {
 131         case TCA_EGRESS_MIRROR:
 132         case TCA_EGRESS_REDIR:
 133         case TCA_INGRESS_REDIR:
 134         case TCA_INGRESS_MIRROR:
 135                 break;
 136         default:
 137                 if (exists)
 138                         tcf_idr_release(*a, bind);
 139                 else
 140                         tcf_idr_cleanup(tn, index);
 141                 NL_SET_ERR_MSG_MOD(extack, "Unknown mirred option");
 142                 return -EINVAL;
 143         }
 144
 145         if (!exists) {
 146                 if (!parm->ifindex) {
 147                         tcf_idr_cleanup(tn, index);
 148                         NL_SET_ERR_MSG_MOD(extack, "Specified device does not exist");
 149                         return -EINVAL;
 150                 }
 151                 ret = tcf_idr_create_from_flags(tn, index, est, a,
 152                                                 &act_mirred_ops, bind, flags);
 153                 if (ret) {
 154                         tcf_idr_cleanup(tn, index);
 155                         return ret;
 156                 }
 157                 ret = ACT_P_CREATED;
 158         } else if (!ovr) {
 159                 tcf_idr_release(*a, bind);
 160                 return -EEXIST;
 161         }
 162
 163         m = to_mirred(*a);
 164         if (ret == ACT_P_CREATED)
 165                 INIT_LIST_HEAD(&m->tcfm_list);
 166
 167         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 168         if (err < 0)
 169                 goto release_idr;
 170
 171         spin_lock_bh(&m->tcf_lock);
 172
 173         if (parm->ifindex) {
 174                 dev = dev_get_by_index(net, parm->ifindex);
 175                 if (!dev) {
 176                         spin_unlock_bh(&m->tcf_lock);
 177                         err = -ENODEV;
 178                         goto put_chain;
 179                 }
 180                 mac_header_xmit = dev_is_mac_header_xmit(dev);
 181                 dev = rcu_replace_pointer(m->tcfm_dev, dev,
 182                                           lockdep_is_held(&m->tcf_lock));
 183                 if (dev)
 184                         dev_put(dev);
 185                 m->tcfm_mac_header_xmit = mac_header_xmit;
 186         }
 187         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 188         m->tcfm_eaction = parm->eaction;
 189         spin_unlock_bh(&m->tcf_lock);
 190         if (goto_ch)
 191                 tcf_chain_put_by_act(goto_ch);
 192
 193         if (ret == ACT_P_CREATED) {
 194                 spin_lock(&mirred_list_lock);
 195                 list_add(&m->tcfm_list, &mirred_list);
 196                 spin_unlock(&mirred_list_lock);
 197         }
 198
 199         return ret;
 200 put_chain:
 201         if (goto_ch)
 202                 tcf_chain_put_by_act(goto_ch);
 203 release_idr:
 204         tcf_idr_release(*a, bind);
 205         return err;
 206 }
 207
 208 static int tcf_mirred_forward(bool want_ingress, struct sk_buff *skb)
 209 {
 210         int err;
 211
 212         if (!want_ingress)
 213                 err = tcf_dev_queue_xmit(skb, dev_queue_xmit);
 214         else
 215                 err = netif_receive_skb(skb);
 216
 217         return err;
 218 }
 219
 220 static int tcf_mirred_act(struct sk_buff *skb, const struct tc_action *a,
 221                           struct tcf_result *res)
 222 {
 223         struct tcf_mirred *m = to_mirred(a);
 224         struct sk_buff *skb2 = skb;
 225         bool m_mac_header_xmit;
 226         struct net_device *dev;
 227         unsigned int rec_level;
 228         int retval, err = 0;
 229         bool use_reinsert;
 230         bool want_ingress;
 231         bool is_redirect;
 232         bool expects_nh;
 233         int m_eaction;
 234         int mac_len;
 235         bool at_nh;
 236
 237         rec_level = __this_cpu_inc_return(mirred_rec_level);
 238         if (unlikely(rec_level > MIRRED_RECURSION_LIMIT)) {
 239                 net_warn_ratelimited("Packet exceeded mirred recursion limit on dev %s\n",
 240                                      netdev_name(skb->dev));
 241                 __this_cpu_dec(mirred_rec_level);
 242                 return TC_ACT_SHOT;
 243         }
 244
 245         tcf_lastuse_update(&m->tcf_tm);
 246         tcf_action_update_bstats(&m->common, skb);
 247
 248         m_mac_header_xmit = READ_ONCE(m->tcfm_mac_header_xmit);
 249         m_eaction = READ_ONCE(m->tcfm_eaction);
 250         retval = READ_ONCE(m->tcf_action);
 251         dev = rcu_dereference_bh(m->tcfm_dev);
 252         if (unlikely(!dev)) {
 253                 pr_notice_once("tc mirred: target device is gone\n");
 254                 goto out;
 255         }
 256
 257         if (unlikely(!(dev->flags & IFF_UP))) {
 258                 net_notice_ratelimited("tc mirred to Houston: device %s is down\n",
 259                                        dev->name);
 260                 goto out;
 261         }
 262
 263         /* we could easily avoid the clone only if called by ingress and clsact;
 264          * since we can't easily detect the clsact caller, skip clone only for
 265          * ingress - that covers the TC S/W datapath.
 266          */
 267         is_redirect = tcf_mirred_is_act_redirect(m_eaction);
 268         use_reinsert = skb_at_tc_ingress(skb) && is_redirect &&
 269                        tcf_mirred_can_reinsert(retval);
 270         if (!use_reinsert) {
 271                 skb2 = skb_clone(skb, GFP_ATOMIC);
 272                 if (!skb2)
 273                         goto out;
 274         }
 275
 276         /* All mirred/redirected skbs should clear previous ct info */
 277         nf_reset_ct(skb2);
 278
 279         want_ingress = tcf_mirred_act_wants_ingress(m_eaction);
 280
 281         expects_nh = want_ingress || !m_mac_header_xmit;
 282         at_nh = skb->data == skb_network_header(skb);
 283         if (at_nh != expects_nh) {
 284                 mac_len = skb_at_tc_ingress(skb) ? skb->mac_len :
 285                           skb_network_header(skb) - skb_mac_header(skb);
 286                 if (expects_nh) {
 287                         /* target device/action expect data at nh */
 288                         skb_pull_rcsum(skb2, mac_len);
 289                 } else {
 290                         /* target device/action expect data at mac */
 291                         skb_push_rcsum(skb2, mac_len);
 292                 }
 293         }
 294
 295         skb2->skb_iif = skb->dev->ifindex;
 296         skb2->dev = dev;
 297
 298         /* mirror is always swallowed */
 299         if (is_redirect) {
 300                 skb_set_redirected(skb2, skb2->tc_at_ingress);
 301
 302                 /* let's the caller reinsert the packet, if possible */
 303                 if (use_reinsert) {
 304                         res->ingress = want_ingress;
 305                         err = tcf_mirred_forward(res->ingress, skb);
 306                         if (err)
 307                                 tcf_action_inc_overlimit_qstats(&m->common);
 308                         __this_cpu_dec(mirred_rec_level);
 309                         return TC_ACT_CONSUMED;
 310                 }
 311         }
 312
 313         err = tcf_mirred_forward(want_ingress, skb2);
 314         if (err) {
 315 out:
 316                 tcf_action_inc_overlimit_qstats(&m->common);
 317                 if (tcf_mirred_is_act_redirect(m_eaction))
 318                         retval = TC_ACT_SHOT;
 319         }
 320         __this_cpu_dec(mirred_rec_level);
 321
 322         return retval;
 323 }
 324
 325 static void tcf_stats_update(struct tc_action *a, u64 bytes, u64 packets,
 326                              u64 drops, u64 lastuse, bool hw)
 327 {
 328         struct tcf_mirred *m = to_mirred(a);
 329         struct tcf_t *tm = &m->tcf_tm;
 330
 331         tcf_action_update_stats(a, bytes, packets, drops, hw);
 332         tm->lastuse = max_t(u64, tm->lastuse, lastuse);
 333 }
 334
 335 static int tcf_mirred_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 336                            int ref)
 337 {
 338         unsigned char *b = skb_tail_pointer(skb);
 339         struct tcf_mirred *m = to_mirred(a);
 340         struct tc_mirred opt = {
 341                 .index   = m->tcf_index,
 342                 .refcnt  = refcount_read(&m->tcf_refcnt) - ref,
 343                 .bindcnt = atomic_read(&m->tcf_bindcnt) - bind,
 344         };
 345         struct net_device *dev;
 346         struct tcf_t t;
 347
 348         spin_lock_bh(&m->tcf_lock);
 349         opt.action = m->tcf_action;
 350         opt.eaction = m->tcfm_eaction;
 351         dev = tcf_mirred_dev_dereference(m);
 352         if (dev)
 353                 opt.ifindex = dev->ifindex;
 354
 355         if (nla_put(skb, TCA_MIRRED_PARMS, sizeof(opt), &opt))
 356                 goto nla_put_failure;
 357
 358         tcf_tm_dump(&t, &m->tcf_tm);
 359         if (nla_put_64bit(skb, TCA_MIRRED_TM, sizeof(t), &t, TCA_MIRRED_PAD))
 360                 goto nla_put_failure;
 361         spin_unlock_bh(&m->tcf_lock);
 362
 363         return skb->len;
 364
 365 nla_put_failure:
 366         spin_unlock_bh(&m->tcf_lock);
 367         nlmsg_trim(skb, b);
 368         return -1;
 369 }
 370
 371 static int tcf_mirred_walker(struct net *net, struct sk_buff *skb,
 372                              struct netlink_callback *cb, int type,
 373                              const struct tc_action_ops *ops,
 374                              struct netlink_ext_ack *extack)
 375 {
 376         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 377
 378         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 379 }
 380
 381 static int tcf_mirred_search(struct net *net, struct tc_action **a, u32 index)
 382 {
 383         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 384
 385         return tcf_idr_search(tn, a, index);
 386 }
 387
 388 static int mirred_device_event(struct notifier_block *unused,
 389                                unsigned long event, void *ptr)
 390 {
 391         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 392         struct tcf_mirred *m;
 393
 394         ASSERT_RTNL();
 395         if (event == NETDEV_UNREGISTER) {
 396                 spin_lock(&mirred_list_lock);
 397                 list_for_each_entry(m, &mirred_list, tcfm_list) {
 398                         spin_lock_bh(&m->tcf_lock);
 399                         if (tcf_mirred_dev_dereference(m) == dev) {
 400                                 dev_put(dev);
 401                                 /* Note : no rcu grace period necessary, as
 402                                  * net_device are already rcu protected.
 403                                  */
 404                                 RCU_INIT_POINTER(m->tcfm_dev, NULL);
 405                         }
 406                         spin_unlock_bh(&m->tcf_lock);
 407                 }
 408                 spin_unlock(&mirred_list_lock);
 409         }
 410
 411         return NOTIFY_DONE;
 412 }
 413
 414 static struct notifier_block mirred_device_notifier = {
 415         .notifier_call = mirred_device_event,
 416 };
 417
 418 static void tcf_mirred_dev_put(void *priv)
 419 {
 420         struct net_device *dev = priv;
 421
 422         dev_put(dev);
 423 }
 424
 425 static struct net_device *
 426 tcf_mirred_get_dev(const struct tc_action *a,
 427                    tc_action_priv_destructor *destructor)
 428 {
 429         struct tcf_mirred *m = to_mirred(a);
 430         struct net_device *dev;
 431
 432         rcu_read_lock();
 433         dev = rcu_dereference(m->tcfm_dev);
 434         if (dev) {
 435                 dev_hold(dev);
 436                 *destructor = tcf_mirred_dev_put;
 437         }
 438         rcu_read_unlock();
 439
 440         return dev;
 441 }
 442
 443 static size_t tcf_mirred_get_fill_size(const struct tc_action *act)
 444 {
 445         return nla_total_size(sizeof(struct tc_mirred));
 446 }
 447
 448 static struct tc_action_ops act_mirred_ops = {
 449         .kind           =       "mirred",
 450         .id             =       TCA_ID_MIRRED,
 451         .owner          =       THIS_MODULE,
 452         .act            =       tcf_mirred_act,
 453         .stats_update   =       tcf_stats_update,
 454         .dump           =       tcf_mirred_dump,
 455         .cleanup        =       tcf_mirred_release,
 456         .init           =       tcf_mirred_init,
 457         .walk           =       tcf_mirred_walker,
 458         .lookup         =       tcf_mirred_search,
 459         .get_fill_size  =       tcf_mirred_get_fill_size,
 460         .size           =       sizeof(struct tcf_mirred),
 461         .get_dev        =       tcf_mirred_get_dev,
 462 };
 463
 464 static __net_init int mirred_init_net(struct net *net)
 465 {
 466         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 467
 468         return tc_action_net_init(net, tn, &act_mirred_ops);
 469 }
 470
 471 static void __net_exit mirred_exit_net(struct list_head *net_list)
 472 {
 473         tc_action_net_exit(net_list, mirred_net_id);
 474 }
 475
 476 static struct pernet_operations mirred_net_ops = {
 477         .init = mirred_init_net,
 478         .exit_batch = mirred_exit_net,
 479         .id   = &mirred_net_id,
 480         .size = sizeof(struct tc_action_net),
 481 };
 482
 483 MODULE_AUTHOR("Jamal Hadi Salim(2002)");
 484 MODULE_DESCRIPTION("Device Mirror/redirect actions");
 485 MODULE_LICENSE("GPL");
 486
 487 static int __init mirred_init_module(void)
 488 {
 489         int err = register_netdevice_notifier(&mirred_device_notifier);
 490         if (err)
 491                 return err;
 492
 493         pr_info("Mirror/redirect action on\n");
 494         err = tcf_register_action(&act_mirred_ops, &mirred_net_ops);
 495         if (err)
 496                 unregister_netdevice_notifier(&mirred_device_notifier);
 497
 498         return err;
 499 }
 500
 501 static void __exit mirred_cleanup_module(void)
 502 {
 503         tcf_unregister_action(&act_mirred_ops, &mirred_net_ops);
 504         unregister_netdevice_notifier(&mirred_device_notifier);
 505 }
 506
 507 module_init(mirred_init_module);
 508 module_exit(mirred_cleanup_module);