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