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