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