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