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