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
 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(tb, TCA_MIRRED_MAX, nla, mirred_policy, extack);
 113         if (ret < 0)
 114                 return ret;
 115         if (!tb[TCA_MIRRED_PARMS]) {
 116                 NL_SET_ERR_MSG_MOD(extack, "Missing required mirred parameters");
 117                 return -EINVAL;
 118         }
 119         parm = nla_data(tb[TCA_MIRRED_PARMS]);
 120
 121         err = tcf_idr_check_alloc(tn, &parm->index, a, bind);
 122         if (err < 0)
 123                 return err;
 124         exists = err;
 125         if (exists && bind)
 126                 return 0;
 127
 128         switch (parm->eaction) {
 129         case TCA_EGRESS_MIRROR:
 130         case TCA_EGRESS_REDIR:
 131         case TCA_INGRESS_REDIR:
 132         case TCA_INGRESS_MIRROR:
 133                 break;
 134         default:
 135                 if (exists)
 136                         tcf_idr_release(*a, bind);
 137                 else
 138                         tcf_idr_cleanup(tn, parm->index);
 139                 NL_SET_ERR_MSG_MOD(extack, "Unknown mirred option");
 140                 return -EINVAL;
 141         }
 142
 143         if (!exists) {
 144                 if (!parm->ifindex) {
 145                         tcf_idr_cleanup(tn, parm->index);
 146                         NL_SET_ERR_MSG_MOD(extack, "Specified device does not exist");
 147                         return -EINVAL;
 148                 }
 149                 ret = tcf_idr_create(tn, parm->index, est, a,
 150                                      &act_mirred_ops, bind, true);
 151                 if (ret) {
 152                         tcf_idr_cleanup(tn, parm->index);
 153                         return ret;
 154                 }
 155                 ret = ACT_P_CREATED;
 156         } else if (!ovr) {
 157                 tcf_idr_release(*a, bind);
 158                 return -EEXIST;
 159         }
 160         m = to_mirred(*a);
 161
 162         spin_lock_bh(&m->tcf_lock);
 163         m->tcf_action = parm->action;
 164         m->tcfm_eaction = parm->eaction;
 165
 166         if (parm->ifindex) {
 167                 dev = dev_get_by_index(net, parm->ifindex);
 168                 if (!dev) {
 169                         spin_unlock_bh(&m->tcf_lock);
 170                         tcf_idr_release(*a, bind);
 171                         return -ENODEV;
 172                 }
 173                 mac_header_xmit = dev_is_mac_header_xmit(dev);
 174                 rcu_swap_protected(m->tcfm_dev, dev,
 175                                    lockdep_is_held(&m->tcf_lock));
 176                 if (dev)
 177                         dev_put(dev);
 178                 m->tcfm_mac_header_xmit = mac_header_xmit;
 179         }
 180         spin_unlock_bh(&m->tcf_lock);
 181
 182         if (ret == ACT_P_CREATED) {
 183                 spin_lock(&mirred_list_lock);
 184                 list_add(&m->tcfm_list, &mirred_list);
 185                 spin_unlock(&mirred_list_lock);
 186
 187                 tcf_idr_insert(tn, *a);
 188         }
 189
 190         return ret;
 191 }
 192
 193 static int tcf_mirred_act(struct sk_buff *skb, const struct tc_action *a,
 194                           struct tcf_result *res)
 195 {
 196         struct tcf_mirred *m = to_mirred(a);
 197         struct sk_buff *skb2 = skb;
 198         bool m_mac_header_xmit;
 199         struct net_device *dev;
 200         int retval, err = 0;
 201         bool use_reinsert;
 202         bool want_ingress;
 203         bool is_redirect;
 204         int m_eaction;
 205         int mac_len;
 206
 207         tcf_lastuse_update(&m->tcf_tm);
 208         bstats_cpu_update(this_cpu_ptr(m->common.cpu_bstats), skb);
 209
 210         m_mac_header_xmit = READ_ONCE(m->tcfm_mac_header_xmit);
 211         m_eaction = READ_ONCE(m->tcfm_eaction);
 212         retval = READ_ONCE(m->tcf_action);
 213         dev = rcu_dereference_bh(m->tcfm_dev);
 214         if (unlikely(!dev)) {
 215                 pr_notice_once("tc mirred: target device is gone\n");
 216                 goto out;
 217         }
 218
 219         if (unlikely(!(dev->flags & IFF_UP))) {
 220                 net_notice_ratelimited("tc mirred to Houston: device %s is down\n",
 221                                        dev->name);
 222                 goto out;
 223         }
 224
 225         /* we could easily avoid the clone only if called by ingress and clsact;
 226          * since we can't easily detect the clsact caller, skip clone only for
 227          * ingress - that covers the TC S/W datapath.
 228          */
 229         is_redirect = tcf_mirred_is_act_redirect(m_eaction);
 230         use_reinsert = skb_at_tc_ingress(skb) && is_redirect &&
 231                        tcf_mirred_can_reinsert(retval);
 232         if (!use_reinsert) {
 233                 skb2 = skb_clone(skb, GFP_ATOMIC);
 234                 if (!skb2)
 235                         goto out;
 236         }
 237
 238         /* If action's target direction differs than filter's direction,
 239          * and devices expect a mac header on xmit, then mac push/pull is
 240          * needed.
 241          */
 242         want_ingress = tcf_mirred_act_wants_ingress(m_eaction);
 243         if (skb_at_tc_ingress(skb) != want_ingress && m_mac_header_xmit) {
 244                 if (!skb_at_tc_ingress(skb)) {
 245                         /* caught at egress, act ingress: pull mac */
 246                         mac_len = skb_network_header(skb) - skb_mac_header(skb);
 247                         skb_pull_rcsum(skb2, mac_len);
 248                 } else {
 249                         /* caught at ingress, act egress: push mac */
 250                         skb_push_rcsum(skb2, skb->mac_len);
 251                 }
 252         }
 253
 254         skb2->skb_iif = skb->dev->ifindex;
 255         skb2->dev = dev;
 256
 257         /* mirror is always swallowed */
 258         if (is_redirect) {
 259                 skb2->tc_redirected = 1;
 260                 skb2->tc_from_ingress = skb2->tc_at_ingress;
 261                 if (skb2->tc_from_ingress)
 262                         skb2->tstamp = 0;
 263                 /* let's the caller reinsert the packet, if possible */
 264                 if (use_reinsert) {
 265                         res->ingress = want_ingress;
 266                         res->qstats = this_cpu_ptr(m->common.cpu_qstats);
 267                         return TC_ACT_REINSERT;
 268                 }
 269         }
 270
 271         if (!want_ingress)
 272                 err = dev_queue_xmit(skb2);
 273         else
 274                 err = netif_receive_skb(skb2);
 275
 276         if (err) {
 277 out:
 278                 qstats_overlimit_inc(this_cpu_ptr(m->common.cpu_qstats));
 279                 if (tcf_mirred_is_act_redirect(m_eaction))
 280                         retval = TC_ACT_SHOT;
 281         }
 282
 283         return retval;
 284 }
 285
 286 static void tcf_stats_update(struct tc_action *a, u64 bytes, u32 packets,
 287                              u64 lastuse, bool hw)
 288 {
 289         struct tcf_mirred *m = to_mirred(a);
 290         struct tcf_t *tm = &m->tcf_tm;
 291
 292         _bstats_cpu_update(this_cpu_ptr(a->cpu_bstats), bytes, packets);
 293         if (hw)
 294                 _bstats_cpu_update(this_cpu_ptr(a->cpu_bstats_hw),
 295                                    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 {
 347         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 348
 349         return tcf_idr_search(tn, a, index);
 350 }
 351
 352 static int mirred_device_event(struct notifier_block *unused,
 353                                unsigned long event, void *ptr)
 354 {
 355         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 356         struct tcf_mirred *m;
 357
 358         ASSERT_RTNL();
 359         if (event == NETDEV_UNREGISTER) {
 360                 spin_lock(&mirred_list_lock);
 361                 list_for_each_entry(m, &mirred_list, tcfm_list) {
 362                         spin_lock_bh(&m->tcf_lock);
 363                         if (tcf_mirred_dev_dereference(m) == dev) {
 364                                 dev_put(dev);
 365                                 /* Note : no rcu grace period necessary, as
 366                                  * net_device are already rcu protected.
 367                                  */
 368                                 RCU_INIT_POINTER(m->tcfm_dev, NULL);
 369                         }
 370                         spin_unlock_bh(&m->tcf_lock);
 371                 }
 372                 spin_unlock(&mirred_list_lock);
 373         }
 374
 375         return NOTIFY_DONE;
 376 }
 377
 378 static struct notifier_block mirred_device_notifier = {
 379         .notifier_call = mirred_device_event,
 380 };
 381
 382 static struct net_device *tcf_mirred_get_dev(const struct tc_action *a)
 383 {
 384         struct tcf_mirred *m = to_mirred(a);
 385         struct net_device *dev;
 386
 387         rcu_read_lock();
 388         dev = rcu_dereference(m->tcfm_dev);
 389         if (dev)
 390                 dev_hold(dev);
 391         rcu_read_unlock();
 392
 393         return dev;
 394 }
 395
 396 static void tcf_mirred_put_dev(struct net_device *dev)
 397 {
 398         dev_put(dev);
 399 }
 400
 401 static struct tc_action_ops act_mirred_ops = {
 402         .kind           =       "mirred",
 403         .id             =       TCA_ID_MIRRED,
 404         .owner          =       THIS_MODULE,
 405         .act            =       tcf_mirred_act,
 406         .stats_update   =       tcf_stats_update,
 407         .dump           =       tcf_mirred_dump,
 408         .cleanup        =       tcf_mirred_release,
 409         .init           =       tcf_mirred_init,
 410         .walk           =       tcf_mirred_walker,
 411         .lookup         =       tcf_mirred_search,
 412         .size           =       sizeof(struct tcf_mirred),
 413         .get_dev        =       tcf_mirred_get_dev,
 414         .put_dev        =       tcf_mirred_put_dev,
 415 };
 416
 417 static __net_init int mirred_init_net(struct net *net)
 418 {
 419         struct tc_action_net *tn = net_generic(net, mirred_net_id);
 420
 421         return tc_action_net_init(tn, &act_mirred_ops);
 422 }
 423
 424 static void __net_exit mirred_exit_net(struct list_head *net_list)
 425 {
 426         tc_action_net_exit(net_list, mirred_net_id);
 427 }
 428
 429 static struct pernet_operations mirred_net_ops = {
 430         .init = mirred_init_net,
 431         .exit_batch = mirred_exit_net,
 432         .id   = &mirred_net_id,
 433         .size = sizeof(struct tc_action_net),
 434 };
 435
 436 MODULE_AUTHOR("Jamal Hadi Salim(2002)");
 437 MODULE_DESCRIPTION("Device Mirror/redirect actions");
 438 MODULE_LICENSE("GPL");
 439
 440 static int __init mirred_init_module(void)
 441 {
 442         int err = register_netdevice_notifier(&mirred_device_notifier);
 443         if (err)
 444                 return err;
 445
 446         pr_info("Mirror/redirect action on\n");
 447         return tcf_register_action(&act_mirred_ops, &mirred_net_ops);
 448 }
 449
 450 static void __exit mirred_cleanup_module(void)
 451 {
 452         tcf_unregister_action(&act_mirred_ops, &mirred_net_ops);
 453         unregister_netdevice_notifier(&mirred_device_notifier);
 454 }
 455
 456 module_init(mirred_init_module);
 457 module_exit(mirred_cleanup_module);