net/sched/act_police.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * net/sched/act_police.c       Input police filter
   4  *
   5  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   6  *              J Hadi Salim (action changes)
   7  */
   8
   9 #include <linux/module.h>
  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/init.h>
  17 #include <linux/slab.h>
  18 #include <net/act_api.h>
  19 #include <net/netlink.h>
  20 #include <net/pkt_cls.h>
  21 #include <net/tc_act/tc_police.h>
  22
  23 /* Each policer is serialized by its individual spinlock */
  24
  25 static struct tc_action_ops act_police_ops;
  26
  27 static const struct nla_policy police_policy[TCA_POLICE_MAX + 1] = {
  28         [TCA_POLICE_RATE]       = { .len = TC_RTAB_SIZE },
  29         [TCA_POLICE_PEAKRATE]   = { .len = TC_RTAB_SIZE },
  30         [TCA_POLICE_AVRATE]     = { .type = NLA_U32 },
  31         [TCA_POLICE_RESULT]     = { .type = NLA_U32 },
  32         [TCA_POLICE_RATE64]     = { .type = NLA_U64 },
  33         [TCA_POLICE_PEAKRATE64] = { .type = NLA_U64 },
  34         [TCA_POLICE_PKTRATE64]  = { .type = NLA_U64, .min = 1 },
  35         [TCA_POLICE_PKTBURST64] = { .type = NLA_U64, .min = 1 },
  36 };
  37
  38 static int tcf_police_init(struct net *net, struct nlattr *nla,
  39                                struct nlattr *est, struct tc_action **a,
  40                                struct tcf_proto *tp, u32 flags,
  41                                struct netlink_ext_ack *extack)
  42 {
  43         int ret = 0, tcfp_result = TC_ACT_OK, err, size;
  44         bool bind = flags & TCA_ACT_FLAGS_BIND;
  45         struct nlattr *tb[TCA_POLICE_MAX + 1];
  46         struct tcf_chain *goto_ch = NULL;
  47         struct tc_police *parm;
  48         struct tcf_police *police;
  49         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
  50         struct tc_action_net *tn = net_generic(net, act_police_ops.net_id);
  51         struct tcf_police_params *new;
  52         bool exists = false;
  53         u32 index;
  54         u64 rate64, prate64;
  55         u64 pps, ppsburst;
  56
  57         if (nla == NULL)
  58                 return -EINVAL;
  59
  60         err = nla_parse_nested_deprecated(tb, TCA_POLICE_MAX, nla,
  61                                           police_policy, NULL);
  62         if (err < 0)
  63                 return err;
  64
  65         if (tb[TCA_POLICE_TBF] == NULL)
  66                 return -EINVAL;
  67         size = nla_len(tb[TCA_POLICE_TBF]);
  68         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
  69                 return -EINVAL;
  70
  71         parm = nla_data(tb[TCA_POLICE_TBF]);
  72         index = parm->index;
  73         err = tcf_idr_check_alloc(tn, &index, a, bind);
  74         if (err < 0)
  75                 return err;
  76         exists = err;
  77         if (exists && bind)
  78                 return 0;
  79
  80         if (!exists) {
  81                 ret = tcf_idr_create(tn, index, NULL, a,
  82                                      &act_police_ops, bind, true, flags);
  83                 if (ret) {
  84                         tcf_idr_cleanup(tn, index);
  85                         return ret;
  86                 }
  87                 ret = ACT_P_CREATED;
  88                 spin_lock_init(&(to_police(*a)->tcfp_lock));
  89         } else if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
  90                 tcf_idr_release(*a, bind);
  91                 return -EEXIST;
  92         }
  93         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
  94         if (err < 0)
  95                 goto release_idr;
  96
  97         police = to_police(*a);
  98         if (parm->rate.rate) {
  99                 err = -ENOMEM;
 100                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE], NULL);
 101                 if (R_tab == NULL)
 102                         goto failure;
 103
 104                 if (parm->peakrate.rate) {
 105                         P_tab = qdisc_get_rtab(&parm->peakrate,
 106                                                tb[TCA_POLICE_PEAKRATE], NULL);
 107                         if (P_tab == NULL)
 108                                 goto failure;
 109                 }
 110         }
 111
 112         if (est) {
 113                 err = gen_replace_estimator(&police->tcf_bstats,
 114                                             police->common.cpu_bstats,
 115                                             &police->tcf_rate_est,
 116                                             &police->tcf_lock,
 117                                             false, est);
 118                 if (err)
 119                         goto failure;
 120         } else if (tb[TCA_POLICE_AVRATE] &&
 121                    (ret == ACT_P_CREATED ||
 122                     !gen_estimator_active(&police->tcf_rate_est))) {
 123                 err = -EINVAL;
 124                 goto failure;
 125         }
 126
 127         if (tb[TCA_POLICE_RESULT]) {
 128                 tcfp_result = nla_get_u32(tb[TCA_POLICE_RESULT]);
 129                 if (TC_ACT_EXT_CMP(tcfp_result, TC_ACT_GOTO_CHAIN)) {
 130                         NL_SET_ERR_MSG(extack,
 131                                        "goto chain not allowed on fallback");
 132                         err = -EINVAL;
 133                         goto failure;
 134                 }
 135         }
 136
 137         if ((tb[TCA_POLICE_PKTRATE64] && !tb[TCA_POLICE_PKTBURST64]) ||
 138             (!tb[TCA_POLICE_PKTRATE64] && tb[TCA_POLICE_PKTBURST64])) {
 139                 NL_SET_ERR_MSG(extack,
 140                                "Both or neither packet-per-second burst and rate must be provided");
 141                 err = -EINVAL;
 142                 goto failure;
 143         }
 144
 145         if (tb[TCA_POLICE_PKTRATE64] && R_tab) {
 146                 NL_SET_ERR_MSG(extack,
 147                                "packet-per-second and byte-per-second rate limits not allowed in same action");
 148                 err = -EINVAL;
 149                 goto failure;
 150         }
 151
 152         new = kzalloc(sizeof(*new), GFP_KERNEL);
 153         if (unlikely(!new)) {
 154                 err = -ENOMEM;
 155                 goto failure;
 156         }
 157
 158         /* No failure allowed after this point */
 159         new->tcfp_result = tcfp_result;
 160         new->tcfp_mtu = parm->mtu;
 161         if (!new->tcfp_mtu) {
 162                 new->tcfp_mtu = ~0;
 163                 if (R_tab)
 164                         new->tcfp_mtu = 255 << R_tab->rate.cell_log;
 165         }
 166         if (R_tab) {
 167                 new->rate_present = true;
 168                 rate64 = tb[TCA_POLICE_RATE64] ?
 169                          nla_get_u64(tb[TCA_POLICE_RATE64]) : 0;
 170                 psched_ratecfg_precompute(&new->rate, &R_tab->rate, rate64);
 171                 qdisc_put_rtab(R_tab);
 172         } else {
 173                 new->rate_present = false;
 174         }
 175         if (P_tab) {
 176                 new->peak_present = true;
 177                 prate64 = tb[TCA_POLICE_PEAKRATE64] ?
 178                           nla_get_u64(tb[TCA_POLICE_PEAKRATE64]) : 0;
 179                 psched_ratecfg_precompute(&new->peak, &P_tab->rate, prate64);
 180                 qdisc_put_rtab(P_tab);
 181         } else {
 182                 new->peak_present = false;
 183         }
 184
 185         new->tcfp_burst = PSCHED_TICKS2NS(parm->burst);
 186         if (new->peak_present)
 187                 new->tcfp_mtu_ptoks = (s64)psched_l2t_ns(&new->peak,
 188                                                          new->tcfp_mtu);
 189
 190         if (tb[TCA_POLICE_AVRATE])
 191                 new->tcfp_ewma_rate = nla_get_u32(tb[TCA_POLICE_AVRATE]);
 192
 193         if (tb[TCA_POLICE_PKTRATE64]) {
 194                 pps = nla_get_u64(tb[TCA_POLICE_PKTRATE64]);
 195                 ppsburst = nla_get_u64(tb[TCA_POLICE_PKTBURST64]);
 196                 new->pps_present = true;
 197                 new->tcfp_pkt_burst = PSCHED_TICKS2NS(ppsburst);
 198                 psched_ppscfg_precompute(&new->ppsrate, pps);
 199         }
 200
 201         spin_lock_bh(&police->tcf_lock);
 202         spin_lock_bh(&police->tcfp_lock);
 203         police->tcfp_t_c = ktime_get_ns();
 204         police->tcfp_toks = new->tcfp_burst;
 205         if (new->peak_present)
 206                 police->tcfp_ptoks = new->tcfp_mtu_ptoks;
 207         spin_unlock_bh(&police->tcfp_lock);
 208         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 209         new = rcu_replace_pointer(police->params,
 210                                   new,
 211                                   lockdep_is_held(&police->tcf_lock));
 212         spin_unlock_bh(&police->tcf_lock);
 213
 214         if (goto_ch)
 215                 tcf_chain_put_by_act(goto_ch);
 216         if (new)
 217                 kfree_rcu(new, rcu);
 218
 219         return ret;
 220
 221 failure:
 222         qdisc_put_rtab(P_tab);
 223         qdisc_put_rtab(R_tab);
 224         if (goto_ch)
 225                 tcf_chain_put_by_act(goto_ch);
 226 release_idr:
 227         tcf_idr_release(*a, bind);
 228         return err;
 229 }
 230
 231 static bool tcf_police_mtu_check(struct sk_buff *skb, u32 limit)
 232 {
 233         u32 len;
 234
 235         if (skb_is_gso(skb))
 236                 return skb_gso_validate_mac_len(skb, limit);
 237
 238         len = qdisc_pkt_len(skb);
 239         if (skb_at_tc_ingress(skb))
 240                 len += skb->mac_len;
 241
 242         return len <= limit;
 243 }
 244
 245 static int tcf_police_act(struct sk_buff *skb, const struct tc_action *a,
 246                           struct tcf_result *res)
 247 {
 248         struct tcf_police *police = to_police(a);
 249         s64 now, toks, ppstoks = 0, ptoks = 0;
 250         struct tcf_police_params *p;
 251         int ret;
 252
 253         tcf_lastuse_update(&police->tcf_tm);
 254         bstats_update(this_cpu_ptr(police->common.cpu_bstats), skb);
 255
 256         ret = READ_ONCE(police->tcf_action);
 257         p = rcu_dereference_bh(police->params);
 258
 259         if (p->tcfp_ewma_rate) {
 260                 struct gnet_stats_rate_est64 sample;
 261
 262                 if (!gen_estimator_read(&police->tcf_rate_est, &sample) ||
 263                     sample.bps >= p->tcfp_ewma_rate)
 264                         goto inc_overlimits;
 265         }
 266
 267         if (tcf_police_mtu_check(skb, p->tcfp_mtu)) {
 268                 if (!p->rate_present && !p->pps_present) {
 269                         ret = p->tcfp_result;
 270                         goto end;
 271                 }
 272
 273                 now = ktime_get_ns();
 274                 spin_lock_bh(&police->tcfp_lock);
 275                 toks = min_t(s64, now - police->tcfp_t_c, p->tcfp_burst);
 276                 if (p->peak_present) {
 277                         ptoks = toks + police->tcfp_ptoks;
 278                         if (ptoks > p->tcfp_mtu_ptoks)
 279                                 ptoks = p->tcfp_mtu_ptoks;
 280                         ptoks -= (s64)psched_l2t_ns(&p->peak,
 281                                                     qdisc_pkt_len(skb));
 282                 }
 283                 if (p->rate_present) {
 284                         toks += police->tcfp_toks;
 285                         if (toks > p->tcfp_burst)
 286                                 toks = p->tcfp_burst;
 287                         toks -= (s64)psched_l2t_ns(&p->rate, qdisc_pkt_len(skb));
 288                 } else if (p->pps_present) {
 289                         ppstoks = min_t(s64, now - police->tcfp_t_c, p->tcfp_pkt_burst);
 290                         ppstoks += police->tcfp_pkttoks;
 291                         if (ppstoks > p->tcfp_pkt_burst)
 292                                 ppstoks = p->tcfp_pkt_burst;
 293                         ppstoks -= (s64)psched_pkt2t_ns(&p->ppsrate, 1);
 294                 }
 295                 if ((toks | ptoks | ppstoks) >= 0) {
 296                         police->tcfp_t_c = now;
 297                         police->tcfp_toks = toks;
 298                         police->tcfp_ptoks = ptoks;
 299                         police->tcfp_pkttoks = ppstoks;
 300                         spin_unlock_bh(&police->tcfp_lock);
 301                         ret = p->tcfp_result;
 302                         goto inc_drops;
 303                 }
 304                 spin_unlock_bh(&police->tcfp_lock);
 305         }
 306
 307 inc_overlimits:
 308         qstats_overlimit_inc(this_cpu_ptr(police->common.cpu_qstats));
 309 inc_drops:
 310         if (ret == TC_ACT_SHOT)
 311                 qstats_drop_inc(this_cpu_ptr(police->common.cpu_qstats));
 312 end:
 313         return ret;
 314 }
 315
 316 static void tcf_police_cleanup(struct tc_action *a)
 317 {
 318         struct tcf_police *police = to_police(a);
 319         struct tcf_police_params *p;
 320
 321         p = rcu_dereference_protected(police->params, 1);
 322         if (p)
 323                 kfree_rcu(p, rcu);
 324 }
 325
 326 static void tcf_police_stats_update(struct tc_action *a,
 327                                     u64 bytes, u64 packets, u64 drops,
 328                                     u64 lastuse, bool hw)
 329 {
 330         struct tcf_police *police = to_police(a);
 331         struct tcf_t *tm = &police->tcf_tm;
 332
 333         tcf_action_update_stats(a, bytes, packets, drops, hw);
 334         tm->lastuse = max_t(u64, tm->lastuse, lastuse);
 335 }
 336
 337 static int tcf_police_dump(struct sk_buff *skb, struct tc_action *a,
 338                                int bind, int ref)
 339 {
 340         unsigned char *b = skb_tail_pointer(skb);
 341         struct tcf_police *police = to_police(a);
 342         struct tcf_police_params *p;
 343         struct tc_police opt = {
 344                 .index = police->tcf_index,
 345                 .refcnt = refcount_read(&police->tcf_refcnt) - ref,
 346                 .bindcnt = atomic_read(&police->tcf_bindcnt) - bind,
 347         };
 348         struct tcf_t t;
 349
 350         spin_lock_bh(&police->tcf_lock);
 351         opt.action = police->tcf_action;
 352         p = rcu_dereference_protected(police->params,
 353                                       lockdep_is_held(&police->tcf_lock));
 354         opt.mtu = p->tcfp_mtu;
 355         opt.burst = PSCHED_NS2TICKS(p->tcfp_burst);
 356         if (p->rate_present) {
 357                 psched_ratecfg_getrate(&opt.rate, &p->rate);
 358                 if ((police->params->rate.rate_bytes_ps >= (1ULL << 32)) &&
 359                     nla_put_u64_64bit(skb, TCA_POLICE_RATE64,
 360                                       police->params->rate.rate_bytes_ps,
 361                                       TCA_POLICE_PAD))
 362                         goto nla_put_failure;
 363         }
 364         if (p->peak_present) {
 365                 psched_ratecfg_getrate(&opt.peakrate, &p->peak);
 366                 if ((police->params->peak.rate_bytes_ps >= (1ULL << 32)) &&
 367                     nla_put_u64_64bit(skb, TCA_POLICE_PEAKRATE64,
 368                                       police->params->peak.rate_bytes_ps,
 369                                       TCA_POLICE_PAD))
 370                         goto nla_put_failure;
 371         }
 372         if (p->pps_present) {
 373                 if (nla_put_u64_64bit(skb, TCA_POLICE_PKTRATE64,
 374                                       police->params->ppsrate.rate_pkts_ps,
 375                                       TCA_POLICE_PAD))
 376                         goto nla_put_failure;
 377                 if (nla_put_u64_64bit(skb, TCA_POLICE_PKTBURST64,
 378                                       PSCHED_NS2TICKS(p->tcfp_pkt_burst),
 379                                       TCA_POLICE_PAD))
 380                         goto nla_put_failure;
 381         }
 382         if (nla_put(skb, TCA_POLICE_TBF, sizeof(opt), &opt))
 383                 goto nla_put_failure;
 384         if (p->tcfp_result &&
 385             nla_put_u32(skb, TCA_POLICE_RESULT, p->tcfp_result))
 386                 goto nla_put_failure;
 387         if (p->tcfp_ewma_rate &&
 388             nla_put_u32(skb, TCA_POLICE_AVRATE, p->tcfp_ewma_rate))
 389                 goto nla_put_failure;
 390
 391         tcf_tm_dump(&t, &police->tcf_tm);
 392         if (nla_put_64bit(skb, TCA_POLICE_TM, sizeof(t), &t, TCA_POLICE_PAD))
 393                 goto nla_put_failure;
 394         spin_unlock_bh(&police->tcf_lock);
 395
 396         return skb->len;
 397
 398 nla_put_failure:
 399         spin_unlock_bh(&police->tcf_lock);
 400         nlmsg_trim(skb, b);
 401         return -1;
 402 }
 403
 404 static int tcf_police_act_to_flow_act(int tc_act, u32 *extval,
 405                                       struct netlink_ext_ack *extack)
 406 {
 407         int act_id = -EOPNOTSUPP;
 408
 409         if (!TC_ACT_EXT_OPCODE(tc_act)) {
 410                 if (tc_act == TC_ACT_OK)
 411                         act_id = FLOW_ACTION_ACCEPT;
 412                 else if (tc_act ==  TC_ACT_SHOT)
 413                         act_id = FLOW_ACTION_DROP;
 414                 else if (tc_act == TC_ACT_PIPE)
 415                         act_id = FLOW_ACTION_PIPE;
 416                 else if (tc_act == TC_ACT_RECLASSIFY)
 417                         NL_SET_ERR_MSG_MOD(extack, "Offload not supported when conform/exceed action is \"reclassify\"");
 418                 else
 419                         NL_SET_ERR_MSG_MOD(extack, "Unsupported conform/exceed action offload");
 420         } else if (TC_ACT_EXT_CMP(tc_act, TC_ACT_GOTO_CHAIN)) {
 421                 act_id = FLOW_ACTION_GOTO;
 422                 *extval = tc_act & TC_ACT_EXT_VAL_MASK;
 423         } else if (TC_ACT_EXT_CMP(tc_act, TC_ACT_JUMP)) {
 424                 act_id = FLOW_ACTION_JUMP;
 425                 *extval = tc_act & TC_ACT_EXT_VAL_MASK;
 426         } else if (tc_act == TC_ACT_UNSPEC) {
 427                 act_id = FLOW_ACTION_CONTINUE;
 428         } else {
 429                 NL_SET_ERR_MSG_MOD(extack, "Unsupported conform/exceed action offload");
 430         }
 431
 432         return act_id;
 433 }
 434
 435 static int tcf_police_offload_act_setup(struct tc_action *act, void *entry_data,
 436                                         u32 *index_inc, bool bind,
 437                                         struct netlink_ext_ack *extack)
 438 {
 439         if (bind) {
 440                 struct flow_action_entry *entry = entry_data;
 441                 struct tcf_police *police = to_police(act);
 442                 struct tcf_police_params *p;
 443                 int act_id;
 444
 445                 p = rcu_dereference_protected(police->params,
 446                                               lockdep_is_held(&police->tcf_lock));
 447
 448                 entry->id = FLOW_ACTION_POLICE;
 449                 entry->police.burst = tcf_police_burst(act);
 450                 entry->police.rate_bytes_ps =
 451                         tcf_police_rate_bytes_ps(act);
 452                 entry->police.peakrate_bytes_ps = tcf_police_peakrate_bytes_ps(act);
 453                 entry->police.avrate = tcf_police_tcfp_ewma_rate(act);
 454                 entry->police.overhead = tcf_police_rate_overhead(act);
 455                 entry->police.burst_pkt = tcf_police_burst_pkt(act);
 456                 entry->police.rate_pkt_ps =
 457                         tcf_police_rate_pkt_ps(act);
 458                 entry->police.mtu = tcf_police_tcfp_mtu(act);
 459
 460                 act_id = tcf_police_act_to_flow_act(police->tcf_action,
 461                                                     &entry->police.exceed.extval,
 462                                                     extack);
 463                 if (act_id < 0)
 464                         return act_id;
 465
 466                 entry->police.exceed.act_id = act_id;
 467
 468                 act_id = tcf_police_act_to_flow_act(p->tcfp_result,
 469                                                     &entry->police.notexceed.extval,
 470                                                     extack);
 471                 if (act_id < 0)
 472                         return act_id;
 473
 474                 entry->police.notexceed.act_id = act_id;
 475
 476                 *index_inc = 1;
 477         } else {
 478                 struct flow_offload_action *fl_action = entry_data;
 479
 480                 fl_action->id = FLOW_ACTION_POLICE;
 481         }
 482
 483         return 0;
 484 }
 485
 486 MODULE_AUTHOR("Alexey Kuznetsov");
 487 MODULE_DESCRIPTION("Policing actions");
 488 MODULE_LICENSE("GPL");
 489
 490 static struct tc_action_ops act_police_ops = {
 491         .kind           =       "police",
 492         .id             =       TCA_ID_POLICE,
 493         .owner          =       THIS_MODULE,
 494         .stats_update   =       tcf_police_stats_update,
 495         .act            =       tcf_police_act,
 496         .dump           =       tcf_police_dump,
 497         .init           =       tcf_police_init,
 498         .cleanup        =       tcf_police_cleanup,
 499         .offload_act_setup =    tcf_police_offload_act_setup,
 500         .size           =       sizeof(struct tcf_police),
 501 };
 502
 503 static __net_init int police_init_net(struct net *net)
 504 {
 505         struct tc_action_net *tn = net_generic(net, act_police_ops.net_id);
 506
 507         return tc_action_net_init(net, tn, &act_police_ops);
 508 }
 509
 510 static void __net_exit police_exit_net(struct list_head *net_list)
 511 {
 512         tc_action_net_exit(net_list, act_police_ops.net_id);
 513 }
 514
 515 static struct pernet_operations police_net_ops = {
 516         .init = police_init_net,
 517         .exit_batch = police_exit_net,
 518         .id   = &act_police_ops.net_id,
 519         .size = sizeof(struct tc_action_net),
 520 };
 521
 522 static int __init police_init_module(void)
 523 {
 524         return tcf_register_action(&act_police_ops, &police_net_ops);
 525 }
 526
 527 static void __exit police_cleanup_module(void)
 528 {
 529         tcf_unregister_action(&act_police_ops, &police_net_ops);
 530 }
 531
 532 module_init(police_init_module);
 533 module_exit(police_cleanup_module);