net/sched/act_police.c

   1 /*
   2  * net/sched/police.c   Input police filter.
   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:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *              J Hadi Salim (action changes)
  11  */
  12
  13 #include <asm/uaccess.h>
  14 #include <asm/system.h>
  15 #include <linux/bitops.h>
  16 #include <linux/module.h>
  17 #include <linux/types.h>
  18 #include <linux/kernel.h>
  19 #include <linux/string.h>
  20 #include <linux/mm.h>
  21 #include <linux/socket.h>
  22 #include <linux/sockios.h>
  23 #include <linux/in.h>
  24 #include <linux/errno.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/netdevice.h>
  27 #include <linux/skbuff.h>
  28 #include <linux/module.h>
  29 #include <linux/rtnetlink.h>
  30 #include <linux/init.h>
  31 #include <net/sock.h>
  32 #include <net/act_api.h>
  33 #include <net/netlink.h>
  34
  35 #define L2T(p,L)   ((p)->tcfp_R_tab->data[(L)>>(p)->tcfp_R_tab->rate.cell_log])
  36 #define L2T_P(p,L) ((p)->tcfp_P_tab->data[(L)>>(p)->tcfp_P_tab->rate.cell_log])
  37
  38 #define POL_TAB_MASK     15
  39 static struct tcf_common *tcf_police_ht[POL_TAB_MASK + 1];
  40 static u32 police_idx_gen;
  41 static DEFINE_RWLOCK(police_lock);
  42
  43 static struct tcf_hashinfo police_hash_info = {
  44         .htab   =       tcf_police_ht,
  45         .hmask  =       POL_TAB_MASK,
  46         .lock   =       &police_lock,
  47 };
  48
  49 /* old policer structure from before tc actions */
  50 struct tc_police_compat
  51 {
  52         u32                     index;
  53         int                     action;
  54         u32                     limit;
  55         u32                     burst;
  56         u32                     mtu;
  57         struct tc_ratespec      rate;
  58         struct tc_ratespec      peakrate;
  59 };
  60
  61 /* Each policer is serialized by its individual spinlock */
  62
  63 #ifdef CONFIG_NET_CLS_ACT
  64 static int tcf_act_police_walker(struct sk_buff *skb, struct netlink_callback *cb,
  65                               int type, struct tc_action *a)
  66 {
  67         struct tcf_common *p;
  68         int err = 0, index = -1, i = 0, s_i = 0, n_i = 0;
  69         struct rtattr *r;
  70
  71         read_lock(&police_lock);
  72
  73         s_i = cb->args[0];
  74
  75         for (i = 0; i < (POL_TAB_MASK + 1); i++) {
  76                 p = tcf_police_ht[tcf_hash(i, POL_TAB_MASK)];
  77
  78                 for (; p; p = p->tcfc_next) {
  79                         index++;
  80                         if (index < s_i)
  81                                 continue;
  82                         a->priv = p;
  83                         a->order = index;
  84                         r = (struct rtattr *)skb_tail_pointer(skb);
  85                         RTA_PUT(skb, a->order, 0, NULL);
  86                         if (type == RTM_DELACTION)
  87                                 err = tcf_action_dump_1(skb, a, 0, 1);
  88                         else
  89                                 err = tcf_action_dump_1(skb, a, 0, 0);
  90                         if (err < 0) {
  91                                 index--;
  92                                 nlmsg_trim(skb, r);
  93                                 goto done;
  94                         }
  95                         r->rta_len = skb_tail_pointer(skb) - (u8 *)r;
  96                         n_i++;
  97                 }
  98         }
  99 done:
 100         read_unlock(&police_lock);
 101         if (n_i)
 102                 cb->args[0] += n_i;
 103         return n_i;
 104
 105 rtattr_failure:
 106         nlmsg_trim(skb, r);
 107         goto done;
 108 }
 109 #endif
 110
 111 void tcf_police_destroy(struct tcf_police *p)
 112 {
 113         unsigned int h = tcf_hash(p->tcf_index, POL_TAB_MASK);
 114         struct tcf_common **p1p;
 115
 116         for (p1p = &tcf_police_ht[h]; *p1p; p1p = &(*p1p)->tcfc_next) {
 117                 if (*p1p == &p->common) {
 118                         write_lock_bh(&police_lock);
 119                         *p1p = p->tcf_next;
 120                         write_unlock_bh(&police_lock);
 121                         gen_kill_estimator(&p->tcf_bstats,
 122                                            &p->tcf_rate_est);
 123                         if (p->tcfp_R_tab)
 124                                 qdisc_put_rtab(p->tcfp_R_tab);
 125                         if (p->tcfp_P_tab)
 126                                 qdisc_put_rtab(p->tcfp_P_tab);
 127                         kfree(p);
 128                         return;
 129                 }
 130         }
 131         BUG_TRAP(0);
 132 }
 133
 134 #ifdef CONFIG_NET_CLS_ACT
 135 static int tcf_act_police_locate(struct rtattr *rta, struct rtattr *est,
 136                                  struct tc_action *a, int ovr, int bind)
 137 {
 138         unsigned h;
 139         int ret = 0, err;
 140         struct rtattr *tb[TCA_POLICE_MAX];
 141         struct tc_police *parm;
 142         struct tcf_police *police;
 143         struct qdisc_rate_table *R_tab = NULL, *P_tab = NULL;
 144         int size;
 145
 146         if (rta == NULL || rtattr_parse_nested(tb, TCA_POLICE_MAX, rta) < 0)
 147                 return -EINVAL;
 148
 149         if (tb[TCA_POLICE_TBF-1] == NULL)
 150                 return -EINVAL;
 151         size = RTA_PAYLOAD(tb[TCA_POLICE_TBF-1]);
 152         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 153                 return -EINVAL;
 154         parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
 155
 156         if (tb[TCA_POLICE_RESULT-1] != NULL &&
 157             RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
 158                 return -EINVAL;
 159         if (tb[TCA_POLICE_RESULT-1] != NULL &&
 160             RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
 161                 return -EINVAL;
 162
 163         if (parm->index) {
 164                 struct tcf_common *pc;
 165
 166                 pc = tcf_hash_lookup(parm->index, &police_hash_info);
 167                 if (pc != NULL) {
 168                         a->priv = pc;
 169                         police = to_police(pc);
 170                         if (bind) {
 171                                 police->tcf_bindcnt += 1;
 172                                 police->tcf_refcnt += 1;
 173                         }
 174                         if (ovr)
 175                                 goto override;
 176                         return ret;
 177                 }
 178         }
 179
 180         police = kzalloc(sizeof(*police), GFP_KERNEL);
 181         if (police == NULL)
 182                 return -ENOMEM;
 183         ret = ACT_P_CREATED;
 184         police->tcf_refcnt = 1;
 185         spin_lock_init(&police->tcf_lock);
 186         if (bind)
 187                 police->tcf_bindcnt = 1;
 188 override:
 189         if (parm->rate.rate) {
 190                 err = -ENOMEM;
 191                 R_tab = qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1]);
 192                 if (R_tab == NULL)
 193                         goto failure;
 194                 if (parm->peakrate.rate) {
 195                         P_tab = qdisc_get_rtab(&parm->peakrate,
 196                                                tb[TCA_POLICE_PEAKRATE-1]);
 197                         if (P_tab == NULL) {
 198                                 qdisc_put_rtab(R_tab);
 199                                 goto failure;
 200                         }
 201                 }
 202         }
 203         /* No failure allowed after this point */
 204         spin_lock_bh(&police->tcf_lock);
 205         if (R_tab != NULL) {
 206                 qdisc_put_rtab(police->tcfp_R_tab);
 207                 police->tcfp_R_tab = R_tab;
 208         }
 209         if (P_tab != NULL) {
 210                 qdisc_put_rtab(police->tcfp_P_tab);
 211                 police->tcfp_P_tab = P_tab;
 212         }
 213
 214         if (tb[TCA_POLICE_RESULT-1])
 215                 police->tcfp_result = *(u32*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
 216         police->tcfp_toks = police->tcfp_burst = parm->burst;
 217         police->tcfp_mtu = parm->mtu;
 218         if (police->tcfp_mtu == 0) {
 219                 police->tcfp_mtu = ~0;
 220                 if (police->tcfp_R_tab)
 221                         police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
 222         }
 223         if (police->tcfp_P_tab)
 224                 police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
 225         police->tcf_action = parm->action;
 226
 227         if (tb[TCA_POLICE_AVRATE-1])
 228                 police->tcfp_ewma_rate =
 229                         *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
 230         if (est)
 231                 gen_replace_estimator(&police->tcf_bstats,
 232                                       &police->tcf_rate_est,
 233                                       &police->tcf_lock, est);
 234
 235         spin_unlock_bh(&police->tcf_lock);
 236         if (ret != ACT_P_CREATED)
 237                 return ret;
 238
 239         police->tcfp_t_c = psched_get_time();
 240         police->tcf_index = parm->index ? parm->index :
 241                 tcf_hash_new_index(&police_idx_gen, &police_hash_info);
 242         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 243         write_lock_bh(&police_lock);
 244         police->tcf_next = tcf_police_ht[h];
 245         tcf_police_ht[h] = &police->common;
 246         write_unlock_bh(&police_lock);
 247
 248         a->priv = police;
 249         return ret;
 250
 251 failure:
 252         if (ret == ACT_P_CREATED)
 253                 kfree(police);
 254         return err;
 255 }
 256
 257 static int tcf_act_police_cleanup(struct tc_action *a, int bind)
 258 {
 259         struct tcf_police *p = a->priv;
 260
 261         if (p != NULL)
 262                 return tcf_police_release(p, bind);
 263         return 0;
 264 }
 265
 266 static int tcf_act_police(struct sk_buff *skb, struct tc_action *a,
 267                           struct tcf_result *res)
 268 {
 269         struct tcf_police *police = a->priv;
 270         psched_time_t now;
 271         long toks;
 272         long ptoks = 0;
 273
 274         spin_lock(&police->tcf_lock);
 275
 276         police->tcf_bstats.bytes += skb->len;
 277         police->tcf_bstats.packets++;
 278
 279         if (police->tcfp_ewma_rate &&
 280             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 281                 police->tcf_qstats.overlimits++;
 282                 spin_unlock(&police->tcf_lock);
 283                 return police->tcf_action;
 284         }
 285
 286         if (skb->len <= police->tcfp_mtu) {
 287                 if (police->tcfp_R_tab == NULL) {
 288                         spin_unlock(&police->tcf_lock);
 289                         return police->tcfp_result;
 290                 }
 291
 292                 now = psched_get_time();
 293                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 294                                             police->tcfp_burst);
 295                 if (police->tcfp_P_tab) {
 296                         ptoks = toks + police->tcfp_ptoks;
 297                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 298                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 299                         ptoks -= L2T_P(police, skb->len);
 300                 }
 301                 toks += police->tcfp_toks;
 302                 if (toks > (long)police->tcfp_burst)
 303                         toks = police->tcfp_burst;
 304                 toks -= L2T(police, skb->len);
 305                 if ((toks|ptoks) >= 0) {
 306                         police->tcfp_t_c = now;
 307                         police->tcfp_toks = toks;
 308                         police->tcfp_ptoks = ptoks;
 309                         spin_unlock(&police->tcf_lock);
 310                         return police->tcfp_result;
 311                 }
 312         }
 313
 314         police->tcf_qstats.overlimits++;
 315         spin_unlock(&police->tcf_lock);
 316         return police->tcf_action;
 317 }
 318
 319 static int
 320 tcf_act_police_dump(struct sk_buff *skb, struct tc_action *a, int bind, int ref)
 321 {
 322         unsigned char *b = skb_tail_pointer(skb);
 323         struct tcf_police *police = a->priv;
 324         struct tc_police opt;
 325
 326         opt.index = police->tcf_index;
 327         opt.action = police->tcf_action;
 328         opt.mtu = police->tcfp_mtu;
 329         opt.burst = police->tcfp_burst;
 330         opt.refcnt = police->tcf_refcnt - ref;
 331         opt.bindcnt = police->tcf_bindcnt - bind;
 332         if (police->tcfp_R_tab)
 333                 opt.rate = police->tcfp_R_tab->rate;
 334         else
 335                 memset(&opt.rate, 0, sizeof(opt.rate));
 336         if (police->tcfp_P_tab)
 337                 opt.peakrate = police->tcfp_P_tab->rate;
 338         else
 339                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 340         RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 341         if (police->tcfp_result)
 342                 RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int),
 343                         &police->tcfp_result);
 344         if (police->tcfp_ewma_rate)
 345                 RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &police->tcfp_ewma_rate);
 346         return skb->len;
 347
 348 rtattr_failure:
 349         nlmsg_trim(skb, b);
 350         return -1;
 351 }
 352
 353 MODULE_AUTHOR("Alexey Kuznetsov");
 354 MODULE_DESCRIPTION("Policing actions");
 355 MODULE_LICENSE("GPL");
 356
 357 static struct tc_action_ops act_police_ops = {
 358         .kind           =       "police",
 359         .hinfo          =       &police_hash_info,
 360         .type           =       TCA_ID_POLICE,
 361         .capab          =       TCA_CAP_NONE,
 362         .owner          =       THIS_MODULE,
 363         .act            =       tcf_act_police,
 364         .dump           =       tcf_act_police_dump,
 365         .cleanup        =       tcf_act_police_cleanup,
 366         .lookup         =       tcf_hash_search,
 367         .init           =       tcf_act_police_locate,
 368         .walk           =       tcf_act_police_walker
 369 };
 370
 371 static int __init
 372 police_init_module(void)
 373 {
 374         return tcf_register_action(&act_police_ops);
 375 }
 376
 377 static void __exit
 378 police_cleanup_module(void)
 379 {
 380         tcf_unregister_action(&act_police_ops);
 381 }
 382
 383 module_init(police_init_module);
 384 module_exit(police_cleanup_module);
 385
 386 #else /* CONFIG_NET_CLS_ACT */
 387
 388 static struct tcf_common *tcf_police_lookup(u32 index)
 389 {
 390         struct tcf_hashinfo *hinfo = &police_hash_info;
 391         struct tcf_common *p;
 392
 393         read_lock(hinfo->lock);
 394         for (p = hinfo->htab[tcf_hash(index, hinfo->hmask)]; p;
 395              p = p->tcfc_next) {
 396                 if (p->tcfc_index == index)
 397                         break;
 398         }
 399         read_unlock(hinfo->lock);
 400
 401         return p;
 402 }
 403
 404 static u32 tcf_police_new_index(void)
 405 {
 406         u32 *idx_gen = &police_idx_gen;
 407         u32 val = *idx_gen;
 408
 409         do {
 410                 if (++val == 0)
 411                         val = 1;
 412         } while (tcf_police_lookup(val));
 413
 414         return (*idx_gen = val);
 415 }
 416
 417 struct tcf_police *tcf_police_locate(struct rtattr *rta, struct rtattr *est)
 418 {
 419         unsigned int h;
 420         struct tcf_police *police;
 421         struct rtattr *tb[TCA_POLICE_MAX];
 422         struct tc_police *parm;
 423         int size;
 424
 425         if (rtattr_parse_nested(tb, TCA_POLICE_MAX, rta) < 0)
 426                 return NULL;
 427
 428         if (tb[TCA_POLICE_TBF-1] == NULL)
 429                 return NULL;
 430         size = RTA_PAYLOAD(tb[TCA_POLICE_TBF-1]);
 431         if (size != sizeof(*parm) && size != sizeof(struct tc_police_compat))
 432                 return NULL;
 433
 434         parm = RTA_DATA(tb[TCA_POLICE_TBF-1]);
 435
 436         if (parm->index) {
 437                 struct tcf_common *pc;
 438
 439                 pc = tcf_police_lookup(parm->index);
 440                 if (pc) {
 441                         police = to_police(pc);
 442                         police->tcf_refcnt++;
 443                         return police;
 444                 }
 445         }
 446         police = kzalloc(sizeof(*police), GFP_KERNEL);
 447         if (unlikely(!police))
 448                 return NULL;
 449
 450         police->tcf_refcnt = 1;
 451         spin_lock_init(&police->tcf_lock);
 452         if (parm->rate.rate) {
 453                 police->tcfp_R_tab =
 454                         qdisc_get_rtab(&parm->rate, tb[TCA_POLICE_RATE-1]);
 455                 if (police->tcfp_R_tab == NULL)
 456                         goto failure;
 457                 if (parm->peakrate.rate) {
 458                         police->tcfp_P_tab =
 459                                 qdisc_get_rtab(&parm->peakrate,
 460                                                tb[TCA_POLICE_PEAKRATE-1]);
 461                         if (police->tcfp_P_tab == NULL)
 462                                 goto failure;
 463                 }
 464         }
 465         if (tb[TCA_POLICE_RESULT-1]) {
 466                 if (RTA_PAYLOAD(tb[TCA_POLICE_RESULT-1]) != sizeof(u32))
 467                         goto failure;
 468                 police->tcfp_result = *(u32*)RTA_DATA(tb[TCA_POLICE_RESULT-1]);
 469         }
 470         if (tb[TCA_POLICE_AVRATE-1]) {
 471                 if (RTA_PAYLOAD(tb[TCA_POLICE_AVRATE-1]) != sizeof(u32))
 472                         goto failure;
 473                 police->tcfp_ewma_rate =
 474                         *(u32*)RTA_DATA(tb[TCA_POLICE_AVRATE-1]);
 475         }
 476         police->tcfp_toks = police->tcfp_burst = parm->burst;
 477         police->tcfp_mtu = parm->mtu;
 478         if (police->tcfp_mtu == 0) {
 479                 police->tcfp_mtu = ~0;
 480                 if (police->tcfp_R_tab)
 481                         police->tcfp_mtu = 255<<police->tcfp_R_tab->rate.cell_log;
 482         }
 483         if (police->tcfp_P_tab)
 484                 police->tcfp_ptoks = L2T_P(police, police->tcfp_mtu);
 485         police->tcfp_t_c = psched_get_time();
 486         police->tcf_index = parm->index ? parm->index :
 487                 tcf_police_new_index();
 488         police->tcf_action = parm->action;
 489         if (est)
 490                 gen_new_estimator(&police->tcf_bstats, &police->tcf_rate_est,
 491                                   &police->tcf_lock, est);
 492         h = tcf_hash(police->tcf_index, POL_TAB_MASK);
 493         write_lock_bh(&police_lock);
 494         police->tcf_next = tcf_police_ht[h];
 495         tcf_police_ht[h] = &police->common;
 496         write_unlock_bh(&police_lock);
 497         return police;
 498
 499 failure:
 500         if (police->tcfp_R_tab)
 501                 qdisc_put_rtab(police->tcfp_R_tab);
 502         kfree(police);
 503         return NULL;
 504 }
 505
 506 int tcf_police(struct sk_buff *skb, struct tcf_police *police)
 507 {
 508         psched_time_t now;
 509         long toks;
 510         long ptoks = 0;
 511
 512         spin_lock(&police->tcf_lock);
 513
 514         police->tcf_bstats.bytes += skb->len;
 515         police->tcf_bstats.packets++;
 516
 517         if (police->tcfp_ewma_rate &&
 518             police->tcf_rate_est.bps >= police->tcfp_ewma_rate) {
 519                 police->tcf_qstats.overlimits++;
 520                 spin_unlock(&police->tcf_lock);
 521                 return police->tcf_action;
 522         }
 523         if (skb->len <= police->tcfp_mtu) {
 524                 if (police->tcfp_R_tab == NULL) {
 525                         spin_unlock(&police->tcf_lock);
 526                         return police->tcfp_result;
 527                 }
 528
 529                 now = psched_get_time();
 530                 toks = psched_tdiff_bounded(now, police->tcfp_t_c,
 531                                             police->tcfp_burst);
 532                 if (police->tcfp_P_tab) {
 533                         ptoks = toks + police->tcfp_ptoks;
 534                         if (ptoks > (long)L2T_P(police, police->tcfp_mtu))
 535                                 ptoks = (long)L2T_P(police, police->tcfp_mtu);
 536                         ptoks -= L2T_P(police, skb->len);
 537                 }
 538                 toks += police->tcfp_toks;
 539                 if (toks > (long)police->tcfp_burst)
 540                         toks = police->tcfp_burst;
 541                 toks -= L2T(police, skb->len);
 542                 if ((toks|ptoks) >= 0) {
 543                         police->tcfp_t_c = now;
 544                         police->tcfp_toks = toks;
 545                         police->tcfp_ptoks = ptoks;
 546                         spin_unlock(&police->tcf_lock);
 547                         return police->tcfp_result;
 548                 }
 549         }
 550
 551         police->tcf_qstats.overlimits++;
 552         spin_unlock(&police->tcf_lock);
 553         return police->tcf_action;
 554 }
 555 EXPORT_SYMBOL(tcf_police);
 556
 557 int tcf_police_dump(struct sk_buff *skb, struct tcf_police *police)
 558 {
 559         unsigned char *b = skb_tail_pointer(skb);
 560         struct tc_police opt;
 561
 562         opt.index = police->tcf_index;
 563         opt.action = police->tcf_action;
 564         opt.mtu = police->tcfp_mtu;
 565         opt.burst = police->tcfp_burst;
 566         if (police->tcfp_R_tab)
 567                 opt.rate = police->tcfp_R_tab->rate;
 568         else
 569                 memset(&opt.rate, 0, sizeof(opt.rate));
 570         if (police->tcfp_P_tab)
 571                 opt.peakrate = police->tcfp_P_tab->rate;
 572         else
 573                 memset(&opt.peakrate, 0, sizeof(opt.peakrate));
 574         RTA_PUT(skb, TCA_POLICE_TBF, sizeof(opt), &opt);
 575         if (police->tcfp_result)
 576                 RTA_PUT(skb, TCA_POLICE_RESULT, sizeof(int),
 577                         &police->tcfp_result);
 578         if (police->tcfp_ewma_rate)
 579                 RTA_PUT(skb, TCA_POLICE_AVRATE, 4, &police->tcfp_ewma_rate);
 580         return skb->len;
 581
 582 rtattr_failure:
 583         nlmsg_trim(skb, b);
 584         return -1;
 585 }
 586
 587 int tcf_police_dump_stats(struct sk_buff *skb, struct tcf_police *police)
 588 {
 589         struct gnet_dump d;
 590
 591         if (gnet_stats_start_copy_compat(skb, TCA_STATS2, TCA_STATS,
 592                                          TCA_XSTATS, &police->tcf_lock,
 593                                          &d) < 0)
 594                 goto errout;
 595
 596         if (gnet_stats_copy_basic(&d, &police->tcf_bstats) < 0 ||
 597             gnet_stats_copy_rate_est(&d, &police->tcf_rate_est) < 0 ||
 598             gnet_stats_copy_queue(&d, &police->tcf_qstats) < 0)
 599                 goto errout;
 600
 601         if (gnet_stats_finish_copy(&d) < 0)
 602                 goto errout;
 603
 604         return 0;
 605
 606 errout:
 607         return -1;
 608 }
 609
 610 #endif /* CONFIG_NET_CLS_ACT */