net/sched/act_mirred.c

   1 /*
   2  * net/sched/act_mirred.c       packet mirroring and redirect actions
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Jamal Hadi Salim (2002-4)
  10  *
  11  * TODO: Add ingress support (and socket redirect support)
  12  *
  13  */
  14
  15 #include <linux/types.h>
  16 #include <linux/kernel.h>
  17 #include <linux/string.h>
  18 #include <linux/errno.h>
  19 #include <linux/skbuff.h>
  20 #include <linux/rtnetlink.h>
  21 #include <linux/module.h>
  22 #include <linux/init.h>
  23 #include <linux/gfp.h>
  24 #include <linux/if_arp.h>
  25 #include <net/net_namespace.h>
  26 #include <net/netlink.h>
  27 #include <net/pkt_sched.h>
  28 #include <net/pkt_cls.h>
  29 #include <linux/tc_act/tc_mirred.h>
  30 #include <net/tc_act/tc_mirred.h>
  31
  32 static LIST_HEAD(mirred_list);
  33 static DEFINE_SPINLOCK(mirred_list_lock);
  34
  35 static bool tcf_mirred_is_act_redirect(int action)
  36 {
  37         return action == TCA_EGRESS_REDIR || action == TCA_INGRESS_REDIR;
  38 }
  39
  40 static bool tcf_mirred_act_wants_ingress(int action)
  41 {
  42         switch (action) {
  43         case TCA_EGRESS_REDIR:
  44         case TCA_EGRESS_MIRROR:
  45                 return false;
  46         case TCA_INGRESS_REDIR:
  47         case TCA_INGRESS_MIRROR:
  48                 return true;
  49         default:
  50                 BUG();
  51         }
  52 }
  53
  54 static bool tcf_mirred_can_reinsert(int action)
  55 {
  56         switch (action) {
  57         case TC_ACT_SHOT:
  58         case TC_ACT_STOLEN:
  59         case TC_ACT_QUEUED:
  60         case TC_ACT_TRAP:
  61                 return true;
  62         }
  63         return false;
  64 }
  65
  66 static struct net_device *tcf_mirred_dev_dereference(struct tcf_mirred *m)
  67 {
  68         return rcu_dereference_protected(m->tcfm_dev,
  69                                          lockdep_is_held(&m->tcf_lock));
  70 }
  71
  72 static void tcf_mirred_release(struct tc_action *a)
  73 {
  74         struct tcf_mirred *m = to_mirred(a);
  75         struct net_device *dev;
  76
  77         spin_lock(&mirred_list_lock);
  78         list_del(&m->tcfm_list);
  79         spin_unlock(&mirred_list_lock);
  80
  81         /* last reference to action, no need to lock */
  82         dev = rcu_dereference_protected(m->tcfm_dev, 1);
  83         if (dev)
  84                 dev_put(dev);
  85 }
  86
  87 static const struct nla_policy mirred_policy[TCA_MIRRED_MAX + 1] = {
  88         [TCA_MIRRED_PARMS]      = { .len = sizeof(struct tc_mirred) },
  89 };
  90
  91 static unsigned int mirred_net_id;
  92 static struct tc_action_ops act_mirred_ops;
  93
  94 static int tcf_mirred_init(struct net *net, struct nlattr *nla,
  95                            struct nlattr *est, struct tc_action **a,
  96                            int ovr, int bind, bool rtnl_held,
  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         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(tb, TCA_MIRRED_MAX, nla, 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(tn, index, est, a,
 151                                      &act_mirred_ops, bind, true);
 152                 if (ret) {
 153                         tcf_idr_cleanup(tn, index);
 154                         return ret;
 155                 }
 156                 ret = ACT_P_CREATED;
 157         } else if (!ovr) {
 158                 tcf_idr_release(*a, bind);
 159                 return -EEXIST;
 160         }
 161         m = to_mirred(*a);
 162
 163         if (ret == ACT_P_CREATED)
 164                 INIT_LIST_HEAD(&m->tcfm_list);
 165
 166         spin_lock_bh(&m->tcf_lock);
 167         m->tcf_action = parm->action;
 168         m->tcfm_eaction = parm->eaction;
 169
 170         if (parm->ifindex) {
 171                 dev = dev_get_by_index(net, parm->ifindex);
 172                 if (!dev) {
 173                         spin_unlock_bh(&m->tcf_lock);
 174                         tcf_idr_release(*a, bind);
 175                         return -ENODEV;
 176                 }
 177                 mac_header_xmit = dev_is_mac_header_xmit(dev);
 178                 rcu_swap_protected(m->tcfm_dev, dev,
 179                                    lockdep_is_held(&m->tcf_lock));
 180                 if (dev)
 181                         dev_put(dev);
 182                 m->tcfm_mac_header_xmit = mac_header_xmit;
 183         }
 184         spin_unlock_bh(&m->tcf_lock);
 185
 186         if (ret == ACT_P_CREATED) {
 187                 spin_lock(&mirred_list_lock);
 188                 list_add(&m->tcfm_list, &mirred_list);
 189                 spin_unlock(&mirred_list_lock);
 190
 191                 tcf_idr_insert(tn, *a);
 192         }
 193
 194         return ret;
 195 }
 196
 197 static int tcf_mirred_act(struct sk_buff *skb, const struct tc_action *a,
 198                           struct tcf_result *res)
 199 {
 200         struct tcf_mirred *m = to_mirred(a);
 201         struct sk_buff *skb2 = skb;
 202         bool m_mac_header_xmit;
 203         struct net_device *dev;
 204         int retval, err = 0;
 205         bool use_reinsert;
 206         bool want_ingress;
 207         bool is_redirect;
 208         int m_eaction;
 209         int mac_len;
 210
 211         tcf_lastuse_update(&m->tcf_tm);
 212         bstats_cpu_update(this_cpu_ptr(m->common.cpu_bstats), skb);
 213
 214         m_mac_header_xmit = READ_ONCE(m->tcfm_mac_header_xmit);
 215         m_eaction = READ_ONCE(m->tcfm_eaction);
 216         retval = READ_ONCE(m->tcf_action);
 217         dev = rcu_dereference_bh(m->tcfm_dev);
 218         if (unlikely(!dev)) {
 219                 pr_notice_once("tc mirred: target device is gone\n");
 220                 goto out;
 221         }
 222
 223         if (unlikely(!(dev->flags & IFF_UP)) || !netif_carrier_ok(dev)) {
 224                 net_notice_ratelimited("tc mirred to Houston: device %s is down\n",
 225                                        dev->name);
 226                 goto out;
 227         }
 228
 229         /* we could easily avoid the clone only if called by ingress and clsact;
 230          * since we can't easily detect the clsact caller, skip clone only for
 231          * ingress - that covers the TC S/W datapath.
 232          */
 233         is_redirect = tcf_mirred_is_act_redirect(m_eaction);
 234         use_reinsert = skb_at_tc_ingress(skb) && is_redirect &&
 235                        tcf_mirred_can_reinsert(retval);
 236         if (!use_reinsert) {
 237                 skb2 = skb_clone(skb, GFP_ATOMIC);
 238                 if (!skb2)
 239                         goto out;
 240         }
 241
 242         /* If action's target direction differs than filter's direction,
 243          * and devices expect a mac header on xmit, then mac push/pull is
 244          * needed.
 245          */
 246         want_ingress = tcf_mirred_act_wants_ingress(m_eaction);
 247         if (skb_at_tc_ingress(skb) != want_ingress && m_mac_header_xmit) {
 248                 if (!skb_at_tc_ingress(skb)) {
 249                         /* caught at egress, act ingress: pull mac */
 250                         mac_len = skb_network_header(skb) - skb_mac_header(skb);
 251                         skb_pull_rcsum(skb2, mac_len);
 252                 } else {
 253                         /* caught at ingress, act egress: push mac */
 254                         skb_push_rcsum(skb2, skb->mac_len);
 255                 }
 256         }
 257
 258         skb2->skb_iif = skb->dev->ifindex;
 259         skb2->dev = dev;
 260
 261         /* mirror is always swallowed */
 262         if (is_redirect) {
 263                 skb2->tc_redirected = 1;
 264                 skb2->tc_from_ingress = skb2->tc_at_ingress;
 265
 266                 /* let's the caller reinsert the packet, if possible */
 267                 if (use_reinsert) {
 268                         res->ingress = want_ingress;
 269                         res->qstats = this_cpu_ptr(m->common.cpu_qstats);
 270                         return TC_ACT_REINSERT;
 271                 }
 272         }
 273
 274         if (!want_ingress)
 275                 err = dev_queue_xmit(skb2);
 276         else
 277                 err = netif_receive_skb(skb2);
 278
 279         if (err) {
 280 out:
 281                 qstats_overlimit_inc(this_cpu_ptr(m->common.cpu_qstats));
 282                 if (tcf_mirred_is_act_redirect(m_eaction))
 283                         retval = TC_ACT_SHOT;
 284         }
 285
 286         return retval;
 287 }
 288
 289 static void tcf_stats_update(struct tc_action *a, u64 bytes, u32 packets,
 290                              u64 lastuse)
 291 {
 292         struct tcf_mirred *m = to_mirred(a);
 293         struct tcf_t *tm = &m->tcf_tm;
 294
 295         _bstats_cpu_update(this_cpu_ptr(a->cpu_bstats), bytes, packets);
 296         tm->lastuse = max_t(u64, tm->lastuse, lastuse);
 297 }
 298
 299 static int tcf_mirred_dump(struct sk_buff *skb, struct tc_action *a, int bind,
 300                            int ref)
 301 {
 302         unsigned char *b = skb_tail_pointer(skb);
 303         struct tcf_mirred *m = to_mirred(a);
 304         struct tc_mirred opt = {
 305                 .index   = m->tcf_index,
 306                 .refcnt  = refcount_read(&m->tcf_refcnt) - ref,
 307                 .bindcnt = atomic_read(&m->tcf_bindcnt) - bind,
 308         };
 309         struct net_device *dev;
 310         struct tcf_t t;
 311
 312         spin_lock_bh(&m->tcf_lock);
 313         opt.action = m->tcf_action;
 314         opt.eaction = m->tcfm_eaction;
 315         dev = tcf_mirred_dev_dereference(m);
 316         if (dev)
 317                 opt.ifindex = dev->ifindex;
 318
 319         if (nla_put(skb, TCA_MIRRED_PARMS, sizeof(opt), &opt))
 320                 goto nla_put_failure;
 321
 322         tcf_tm_dump(&t, &m->tcf_tm);
 323         if (nla_put_64bit(skb, TCA_MIRRED_TM, sizeof(t), &t, TCA_MIRRED_PAD))
 324                 goto nla_put_failure;
 325         spin_unlock_bh(&m->tcf_lock);
 326
 327         return skb->len;
 328
 329 nla_put_failure:
 330         spin_unlock_bh(&m->tcf_lock);
 331         nlmsg_trim(skb, b);
 332         return -1;
 333 }
 334
 335 static int tcf_mirred_walker(struct net *net, struct sk_buff *skb,
 336                              struct netlink_callback *cb, int type,
 337                              const struct tc_action_ops *ops,
 338                              struct netlink_ext_ack *extack)
 339 {
 340         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 341
 342         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 343 }
 344
 345 static int tcf_mirred_search(struct net *net, struct tc_action **a, u32 index,
 346                              struct netlink_ext_ack *extack)
 347 {
 348         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 349
 350         return tcf_idr_search(tn, a, index);
 351 }
 352
 353 static int mirred_device_event(struct notifier_block *unused,
 354                                unsigned long event, void *ptr)
 355 {
 356         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 357         struct tcf_mirred *m;
 358
 359         ASSERT_RTNL();
 360         if (event == NETDEV_UNREGISTER) {
 361                 spin_lock(&mirred_list_lock);
 362                 list_for_each_entry(m, &mirred_list, tcfm_list) {
 363                         spin_lock_bh(&m->tcf_lock);
 364                         if (tcf_mirred_dev_dereference(m) == dev) {
 365                                 dev_put(dev);
 366                                 /* Note : no rcu grace period necessary, as
 367                                  * net_device are already rcu protected.
 368                                  */
 369                                 RCU_INIT_POINTER(m->tcfm_dev, NULL);
 370                         }
 371                         spin_unlock_bh(&m->tcf_lock);
 372                 }
 373                 spin_unlock(&mirred_list_lock);
 374         }
 375
 376         return NOTIFY_DONE;
 377 }
 378
 379 static struct notifier_block mirred_device_notifier = {
 380         .notifier_call = mirred_device_event,
 381 };
 382
 383 static struct net_device *tcf_mirred_get_dev(const struct tc_action *a)
 384 {
 385         struct tcf_mirred *m = to_mirred(a);
 386         struct net_device *dev;
 387
 388         rcu_read_lock();
 389         dev = rcu_dereference(m->tcfm_dev);
 390         if (dev)
 391                 dev_hold(dev);
 392         rcu_read_unlock();
 393
 394         return dev;
 395 }
 396
 397 static void tcf_mirred_put_dev(struct net_device *dev)
 398 {
 399         dev_put(dev);
 400 }
 401
 402 static struct tc_action_ops act_mirred_ops = {
 403         .kind           =       "mirred",
 404         .type           =       TCA_ACT_MIRRED,
 405         .owner          =       THIS_MODULE,
 406         .act            =       tcf_mirred_act,
 407         .stats_update   =       tcf_stats_update,
 408         .dump           =       tcf_mirred_dump,
 409         .cleanup        =       tcf_mirred_release,
 410         .init           =       tcf_mirred_init,
 411         .walk           =       tcf_mirred_walker,
 412         .lookup         =       tcf_mirred_search,
 413         .size           =       sizeof(struct tcf_mirred),
 414         .get_dev        =       tcf_mirred_get_dev,
 415         .put_dev        =       tcf_mirred_put_dev,
 416 };
 417
 418 static __net_init int mirred_init_net(struct net *net)
 419 {
 420         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 421
 422         return tc_action_net_init(net, tn, &act_mirred_ops);
 423 }
 424
 425 static void __net_exit mirred_exit_net(struct list_head *net_list)
 426 {
 427         tc_action_net_exit(net_list, mirred_net_id);
 428 }
 429
 430 static struct pernet_operations mirred_net_ops = {
 431         .init = mirred_init_net,
 432         .exit_batch = mirred_exit_net,
 433         .id   = &mirred_net_id,
 434         .size = sizeof(struct tc_action_net),
 435 };
 436
 437 MODULE_AUTHOR("Jamal Hadi Salim(2002)");
 438 MODULE_DESCRIPTION("Device Mirror/redirect actions");
 439 MODULE_LICENSE("GPL");
 440
 441 static int __init mirred_init_module(void)
 442 {
 443         int err = register_netdevice_notifier(&mirred_device_notifier);
 444         if (err)
 445                 return err;
 446
 447         pr_info("Mirror/redirect action on\n");
 448         err = tcf_register_action(&act_mirred_ops, &mirred_net_ops);
 449         if (err)
 450                 unregister_netdevice_notifier(&mirred_device_notifier);
 451
 452         return err;
 453 }
 454
 455 static void __exit mirred_cleanup_module(void)
 456 {
 457         tcf_unregister_action(&act_mirred_ops, &mirred_net_ops);
 458         unregister_netdevice_notifier(&mirred_device_notifier);
 459 }
 460
 461 module_init(mirred_init_module);
 462 module_exit(mirred_cleanup_module);