net/sched/cls_u32.c

   1 /*
   2  * net/sched/cls_u32.c  Ugly (or Universal) 32bit key Packet Classifier.
   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  *
  11  *      The filters are packed to hash tables of key nodes
  12  *      with a set of 32bit key/mask pairs at every node.
  13  *      Nodes reference next level hash tables etc.
  14  *
  15  *      This scheme is the best universal classifier I managed to
  16  *      invent; it is not super-fast, but it is not slow (provided you
  17  *      program it correctly), and general enough.  And its relative
  18  *      speed grows as the number of rules becomes larger.
  19  *
  20  *      It seems that it represents the best middle point between
  21  *      speed and manageability both by human and by machine.
  22  *
  23  *      It is especially useful for link sharing combined with QoS;
  24  *      pure RSVP doesn't need such a general approach and can use
  25  *      much simpler (and faster) schemes, sort of cls_rsvp.c.
  26  *
  27  *      JHS: We should remove the CONFIG_NET_CLS_IND from here
  28  *      eventually when the meta match extension is made available
  29  *
  30  *      nfmark match added by Catalin(ux aka Dino) BOIE <catab at umbrella.ro>
  31  */
  32
  33 #include <linux/module.h>
  34 #include <linux/slab.h>
  35 #include <linux/types.h>
  36 #include <linux/kernel.h>
  37 #include <linux/string.h>
  38 #include <linux/errno.h>
  39 #include <linux/rtnetlink.h>
  40 #include <linux/skbuff.h>
  41 #include <linux/bitmap.h>
  42 #include <net/netlink.h>
  43 #include <net/act_api.h>
  44 #include <net/pkt_cls.h>
  45
  46 struct tc_u_knode {
  47         struct tc_u_knode       *next;
  48         u32                     handle;
  49         struct tc_u_hnode       *ht_up;
  50         struct tcf_exts         exts;
  51 #ifdef CONFIG_NET_CLS_IND
  52         int                     ifindex;
  53 #endif
  54         u8                      fshift;
  55         struct tcf_result       res;
  56         struct tc_u_hnode       *ht_down;
  57 #ifdef CONFIG_CLS_U32_PERF
  58         struct tc_u32_pcnt __percpu *pf;
  59 #endif
  60 #ifdef CONFIG_CLS_U32_MARK
  61         u32                     val;
  62         u32                     mask;
  63         u32 __percpu            *pcpu_success;
  64 #endif
  65         struct tc_u32_sel       sel;
  66 };
  67
  68 struct tc_u_hnode {
  69         struct tc_u_hnode       *next;
  70         u32                     handle;
  71         u32                     prio;
  72         struct tc_u_common      *tp_c;
  73         int                     refcnt;
  74         unsigned int            divisor;
  75         struct tc_u_knode       *ht[1];
  76 };
  77
  78 struct tc_u_common {
  79         struct tc_u_hnode       *hlist;
  80         struct Qdisc            *q;
  81         int                     refcnt;
  82         u32                     hgenerator;
  83 };
  84
  85 static inline unsigned int u32_hash_fold(__be32 key,
  86                                          const struct tc_u32_sel *sel,
  87                                          u8 fshift)
  88 {
  89         unsigned int h = ntohl(key & sel->hmask) >> fshift;
  90
  91         return h;
  92 }
  93
  94 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
  95 {
  96         struct {
  97                 struct tc_u_knode *knode;
  98                 unsigned int      off;
  99         } stack[TC_U32_MAXDEPTH];
 100
 101         struct tc_u_hnode *ht = tp->root;
 102         unsigned int off = skb_network_offset(skb);
 103         struct tc_u_knode *n;
 104         int sdepth = 0;
 105         int off2 = 0;
 106         int sel = 0;
 107 #ifdef CONFIG_CLS_U32_PERF
 108         int j;
 109 #endif
 110         int i, r;
 111
 112 next_ht:
 113         n = ht->ht[sel];
 114
 115 next_knode:
 116         if (n) {
 117                 struct tc_u32_key *key = n->sel.keys;
 118
 119 #ifdef CONFIG_CLS_U32_PERF
 120                 __this_cpu_inc(n->pf->rcnt);
 121                 j = 0;
 122 #endif
 123
 124 #ifdef CONFIG_CLS_U32_MARK
 125                 if ((skb->mark & n->mask) != n->val) {
 126                         n = n->next;
 127                         goto next_knode;
 128                 } else {
 129                         __this_cpu_inc(*n->pcpu_success);
 130                 }
 131 #endif
 132
 133                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 134                         int toff = off + key->off + (off2 & key->offmask);
 135                         __be32 *data, hdata;
 136
 137                         if (skb_headroom(skb) + toff > INT_MAX)
 138                                 goto out;
 139
 140                         data = skb_header_pointer(skb, toff, 4, &hdata);
 141                         if (!data)
 142                                 goto out;
 143                         if ((*data ^ key->val) & key->mask) {
 144                                 n = n->next;
 145                                 goto next_knode;
 146                         }
 147 #ifdef CONFIG_CLS_U32_PERF
 148                         __this_cpu_inc(n->pf->kcnts[j]);
 149                         j++;
 150 #endif
 151                 }
 152                 if (n->ht_down == NULL) {
 153 check_terminal:
 154                         if (n->sel.flags & TC_U32_TERMINAL) {
 155
 156                                 *res = n->res;
 157 #ifdef CONFIG_NET_CLS_IND
 158                                 if (!tcf_match_indev(skb, n->ifindex)) {
 159                                         n = n->next;
 160                                         goto next_knode;
 161                                 }
 162 #endif
 163 #ifdef CONFIG_CLS_U32_PERF
 164                                 __this_cpu_inc(n->pf->rhit);
 165 #endif
 166                                 r = tcf_exts_exec(skb, &n->exts, res);
 167                                 if (r < 0) {
 168                                         n = n->next;
 169                                         goto next_knode;
 170                                 }
 171
 172                                 return r;
 173                         }
 174                         n = n->next;
 175                         goto next_knode;
 176                 }
 177
 178                 /* PUSH */
 179                 if (sdepth >= TC_U32_MAXDEPTH)
 180                         goto deadloop;
 181                 stack[sdepth].knode = n;
 182                 stack[sdepth].off = off;
 183                 sdepth++;
 184
 185                 ht = n->ht_down;
 186                 sel = 0;
 187                 if (ht->divisor) {
 188                         __be32 *data, hdata;
 189
 190                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 191                                                   &hdata);
 192                         if (!data)
 193                                 goto out;
 194                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 195                                                           n->fshift);
 196                 }
 197                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 198                         goto next_ht;
 199
 200                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 201                         off2 = n->sel.off + 3;
 202                         if (n->sel.flags & TC_U32_VAROFFSET) {
 203                                 __be16 *data, hdata;
 204
 205                                 data = skb_header_pointer(skb,
 206                                                           off + n->sel.offoff,
 207                                                           2, &hdata);
 208                                 if (!data)
 209                                         goto out;
 210                                 off2 += ntohs(n->sel.offmask & *data) >>
 211                                         n->sel.offshift;
 212                         }
 213                         off2 &= ~3;
 214                 }
 215                 if (n->sel.flags & TC_U32_EAT) {
 216                         off += off2;
 217                         off2 = 0;
 218                 }
 219
 220                 if (off < skb->len)
 221                         goto next_ht;
 222         }
 223
 224         /* POP */
 225         if (sdepth--) {
 226                 n = stack[sdepth].knode;
 227                 ht = n->ht_up;
 228                 off = stack[sdepth].off;
 229                 goto check_terminal;
 230         }
 231 out:
 232         return -1;
 233
 234 deadloop:
 235         net_warn_ratelimited("cls_u32: dead loop\n");
 236         return -1;
 237 }
 238
 239 static struct tc_u_hnode *
 240 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 241 {
 242         struct tc_u_hnode *ht;
 243
 244         for (ht = tp_c->hlist; ht; ht = ht->next)
 245                 if (ht->handle == handle)
 246                         break;
 247
 248         return ht;
 249 }
 250
 251 static struct tc_u_knode *
 252 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 253 {
 254         unsigned int sel;
 255         struct tc_u_knode *n = NULL;
 256
 257         sel = TC_U32_HASH(handle);
 258         if (sel > ht->divisor)
 259                 goto out;
 260
 261         for (n = ht->ht[sel]; n; n = n->next)
 262                 if (n->handle == handle)
 263                         break;
 264 out:
 265         return n;
 266 }
 267
 268
 269 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 270 {
 271         struct tc_u_hnode *ht;
 272         struct tc_u_common *tp_c = tp->data;
 273
 274         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 275                 ht = tp->root;
 276         else
 277                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 278
 279         if (!ht)
 280                 return 0;
 281
 282         if (TC_U32_KEY(handle) == 0)
 283                 return (unsigned long)ht;
 284
 285         return (unsigned long)u32_lookup_key(ht, handle);
 286 }
 287
 288 static void u32_put(struct tcf_proto *tp, unsigned long f)
 289 {
 290 }
 291
 292 static u32 gen_new_htid(struct tc_u_common *tp_c)
 293 {
 294         int i = 0x800;
 295
 296         do {
 297                 if (++tp_c->hgenerator == 0x7FF)
 298                         tp_c->hgenerator = 1;
 299         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 300
 301         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 302 }
 303
 304 static int u32_init(struct tcf_proto *tp)
 305 {
 306         struct tc_u_hnode *root_ht;
 307         struct tc_u_common *tp_c;
 308
 309         tp_c = tp->q->u32_node;
 310
 311         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 312         if (root_ht == NULL)
 313                 return -ENOBUFS;
 314
 315         root_ht->divisor = 0;
 316         root_ht->refcnt++;
 317         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 318         root_ht->prio = tp->prio;
 319
 320         if (tp_c == NULL) {
 321                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 322                 if (tp_c == NULL) {
 323                         kfree(root_ht);
 324                         return -ENOBUFS;
 325                 }
 326                 tp_c->q = tp->q;
 327                 tp->q->u32_node = tp_c;
 328         }
 329
 330         tp_c->refcnt++;
 331         root_ht->next = tp_c->hlist;
 332         tp_c->hlist = root_ht;
 333         root_ht->tp_c = tp_c;
 334
 335         tp->root = root_ht;
 336         tp->data = tp_c;
 337         return 0;
 338 }
 339
 340 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
 341 {
 342         tcf_unbind_filter(tp, &n->res);
 343         tcf_exts_destroy(tp, &n->exts);
 344         if (n->ht_down)
 345                 n->ht_down->refcnt--;
 346 #ifdef CONFIG_CLS_U32_PERF
 347         free_percpu(n->pf);
 348 #endif
 349         kfree(n);
 350         return 0;
 351 }
 352
 353 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 354 {
 355         struct tc_u_knode **kp;
 356         struct tc_u_hnode *ht = key->ht_up;
 357
 358         if (ht) {
 359                 for (kp = &ht->ht[TC_U32_HASH(key->handle)]; *kp; kp = &(*kp)->next) {
 360                         if (*kp == key) {
 361                                 tcf_tree_lock(tp);
 362                                 *kp = key->next;
 363                                 tcf_tree_unlock(tp);
 364
 365                                 u32_destroy_key(tp, key);
 366                                 return 0;
 367                         }
 368                 }
 369         }
 370         WARN_ON(1);
 371         return 0;
 372 }
 373
 374 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 375 {
 376         struct tc_u_knode *n;
 377         unsigned int h;
 378
 379         for (h = 0; h <= ht->divisor; h++) {
 380                 while ((n = ht->ht[h]) != NULL) {
 381                         ht->ht[h] = n->next;
 382
 383                         u32_destroy_key(tp, n);
 384                 }
 385         }
 386 }
 387
 388 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 389 {
 390         struct tc_u_common *tp_c = tp->data;
 391         struct tc_u_hnode **hn;
 392
 393         WARN_ON(ht->refcnt);
 394
 395         u32_clear_hnode(tp, ht);
 396
 397         for (hn = &tp_c->hlist; *hn; hn = &(*hn)->next) {
 398                 if (*hn == ht) {
 399                         *hn = ht->next;
 400                         kfree(ht);
 401                         return 0;
 402                 }
 403         }
 404
 405         WARN_ON(1);
 406         return -ENOENT;
 407 }
 408
 409 static void u32_destroy(struct tcf_proto *tp)
 410 {
 411         struct tc_u_common *tp_c = tp->data;
 412         struct tc_u_hnode *root_ht = tp->root;
 413
 414         WARN_ON(root_ht == NULL);
 415
 416         if (root_ht && --root_ht->refcnt == 0)
 417                 u32_destroy_hnode(tp, root_ht);
 418
 419         if (--tp_c->refcnt == 0) {
 420                 struct tc_u_hnode *ht;
 421
 422                 tp->q->u32_node = NULL;
 423
 424                 for (ht = tp_c->hlist; ht; ht = ht->next) {
 425                         ht->refcnt--;
 426                         u32_clear_hnode(tp, ht);
 427                 }
 428
 429                 while ((ht = tp_c->hlist) != NULL) {
 430                         tp_c->hlist = ht->next;
 431
 432                         WARN_ON(ht->refcnt != 0);
 433
 434                         kfree(ht);
 435                 }
 436
 437                 kfree(tp_c);
 438         }
 439
 440         tp->data = NULL;
 441 }
 442
 443 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 444 {
 445         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 446
 447         if (ht == NULL)
 448                 return 0;
 449
 450         if (TC_U32_KEY(ht->handle))
 451                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 452
 453         if (tp->root == ht)
 454                 return -EINVAL;
 455
 456         if (ht->refcnt == 1) {
 457                 ht->refcnt--;
 458                 u32_destroy_hnode(tp, ht);
 459         } else {
 460                 return -EBUSY;
 461         }
 462
 463         return 0;
 464 }
 465
 466 #define NR_U32_NODE (1<<12)
 467 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 468 {
 469         struct tc_u_knode *n;
 470         unsigned long i;
 471         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 472                                         GFP_KERNEL);
 473         if (!bitmap)
 474                 return handle | 0xFFF;
 475
 476         for (n = ht->ht[TC_U32_HASH(handle)]; n; n = n->next)
 477                 set_bit(TC_U32_NODE(n->handle), bitmap);
 478
 479         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 480         if (i >= NR_U32_NODE)
 481                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 482
 483         kfree(bitmap);
 484         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 485 }
 486
 487 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 488         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 489         [TCA_U32_HASH]          = { .type = NLA_U32 },
 490         [TCA_U32_LINK]          = { .type = NLA_U32 },
 491         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 492         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 493         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 494         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 495 };
 496
 497 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 498                          unsigned long base, struct tc_u_hnode *ht,
 499                          struct tc_u_knode *n, struct nlattr **tb,
 500                          struct nlattr *est, bool ovr)
 501 {
 502         int err;
 503         struct tcf_exts e;
 504
 505         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 506         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 507         if (err < 0)
 508                 return err;
 509
 510         err = -EINVAL;
 511         if (tb[TCA_U32_LINK]) {
 512                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 513                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 514
 515                 if (TC_U32_KEY(handle))
 516                         goto errout;
 517
 518                 if (handle) {
 519                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 520
 521                         if (ht_down == NULL)
 522                                 goto errout;
 523                         ht_down->refcnt++;
 524                 }
 525
 526                 tcf_tree_lock(tp);
 527                 ht_old = n->ht_down;
 528                 n->ht_down = ht_down;
 529                 tcf_tree_unlock(tp);
 530
 531                 if (ht_old)
 532                         ht_old->refcnt--;
 533         }
 534         if (tb[TCA_U32_CLASSID]) {
 535                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 536                 tcf_bind_filter(tp, &n->res, base);
 537         }
 538
 539 #ifdef CONFIG_NET_CLS_IND
 540         if (tb[TCA_U32_INDEV]) {
 541                 int ret;
 542                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 543                 if (ret < 0)
 544                         goto errout;
 545                 n->ifindex = ret;
 546         }
 547 #endif
 548         tcf_exts_change(tp, &n->exts, &e);
 549
 550         return 0;
 551 errout:
 552         tcf_exts_destroy(tp, &e);
 553         return err;
 554 }
 555
 556 static int u32_change(struct net *net, struct sk_buff *in_skb,
 557                       struct tcf_proto *tp, unsigned long base, u32 handle,
 558                       struct nlattr **tca,
 559                       unsigned long *arg, bool ovr)
 560 {
 561         struct tc_u_common *tp_c = tp->data;
 562         struct tc_u_hnode *ht;
 563         struct tc_u_knode *n;
 564         struct tc_u32_sel *s;
 565         struct nlattr *opt = tca[TCA_OPTIONS];
 566         struct nlattr *tb[TCA_U32_MAX + 1];
 567         u32 htid;
 568         int err;
 569 #ifdef CONFIG_CLS_U32_PERF
 570         size_t size;
 571 #endif
 572
 573         if (opt == NULL)
 574                 return handle ? -EINVAL : 0;
 575
 576         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 577         if (err < 0)
 578                 return err;
 579
 580         n = (struct tc_u_knode *)*arg;
 581         if (n) {
 582                 if (TC_U32_KEY(n->handle) == 0)
 583                         return -EINVAL;
 584
 585                 return u32_set_parms(net, tp, base, n->ht_up, n, tb,
 586                                      tca[TCA_RATE], ovr);
 587         }
 588
 589         if (tb[TCA_U32_DIVISOR]) {
 590                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 591
 592                 if (--divisor > 0x100)
 593                         return -EINVAL;
 594                 if (TC_U32_KEY(handle))
 595                         return -EINVAL;
 596                 if (handle == 0) {
 597                         handle = gen_new_htid(tp->data);
 598                         if (handle == 0)
 599                                 return -ENOMEM;
 600                 }
 601                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 602                 if (ht == NULL)
 603                         return -ENOBUFS;
 604                 ht->tp_c = tp_c;
 605                 ht->refcnt = 1;
 606                 ht->divisor = divisor;
 607                 ht->handle = handle;
 608                 ht->prio = tp->prio;
 609                 ht->next = tp_c->hlist;
 610                 tp_c->hlist = ht;
 611                 *arg = (unsigned long)ht;
 612                 return 0;
 613         }
 614
 615         if (tb[TCA_U32_HASH]) {
 616                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 617                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 618                         ht = tp->root;
 619                         htid = ht->handle;
 620                 } else {
 621                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 622                         if (ht == NULL)
 623                                 return -EINVAL;
 624                 }
 625         } else {
 626                 ht = tp->root;
 627                 htid = ht->handle;
 628         }
 629
 630         if (ht->divisor < TC_U32_HASH(htid))
 631                 return -EINVAL;
 632
 633         if (handle) {
 634                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 635                         return -EINVAL;
 636                 handle = htid | TC_U32_NODE(handle);
 637         } else
 638                 handle = gen_new_kid(ht, htid);
 639
 640         if (tb[TCA_U32_SEL] == NULL)
 641                 return -EINVAL;
 642
 643         s = nla_data(tb[TCA_U32_SEL]);
 644
 645         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 646         if (n == NULL)
 647                 return -ENOBUFS;
 648
 649 #ifdef CONFIG_CLS_U32_PERF
 650         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 651         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 652         if (!n->pf) {
 653                 kfree(n);
 654                 return -ENOBUFS;
 655         }
 656 #endif
 657
 658         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 659         n->ht_up = ht;
 660         n->handle = handle;
 661         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 662         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 663
 664 #ifdef CONFIG_CLS_U32_MARK
 665         n->pcpu_success = alloc_percpu(u32);
 666
 667         if (tb[TCA_U32_MARK]) {
 668                 struct tc_u32_mark *mark;
 669
 670                 mark = nla_data(tb[TCA_U32_MARK]);
 671                 n->val = mark->val;
 672                 n->mask = mark->mask;
 673         }
 674 #endif
 675
 676         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
 677         if (err == 0) {
 678                 struct tc_u_knode **ins;
 679                 for (ins = &ht->ht[TC_U32_HASH(handle)]; *ins; ins = &(*ins)->next)
 680                         if (TC_U32_NODE(handle) < TC_U32_NODE((*ins)->handle))
 681                                 break;
 682
 683                 n->next = *ins;
 684                 tcf_tree_lock(tp);
 685                 *ins = n;
 686                 tcf_tree_unlock(tp);
 687
 688                 *arg = (unsigned long)n;
 689                 return 0;
 690         }
 691 #ifdef CONFIG_CLS_U32_PERF
 692         kfree(n->pf);
 693 #endif
 694         kfree(n);
 695         return err;
 696 }
 697
 698 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 699 {
 700         struct tc_u_common *tp_c = tp->data;
 701         struct tc_u_hnode *ht;
 702         struct tc_u_knode *n;
 703         unsigned int h;
 704
 705         if (arg->stop)
 706                 return;
 707
 708         for (ht = tp_c->hlist; ht; ht = ht->next) {
 709                 if (ht->prio != tp->prio)
 710                         continue;
 711                 if (arg->count >= arg->skip) {
 712                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 713                                 arg->stop = 1;
 714                                 return;
 715                         }
 716                 }
 717                 arg->count++;
 718                 for (h = 0; h <= ht->divisor; h++) {
 719                         for (n = ht->ht[h]; n; n = n->next) {
 720                                 if (arg->count < arg->skip) {
 721                                         arg->count++;
 722                                         continue;
 723                                 }
 724                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 725                                         arg->stop = 1;
 726                                         return;
 727                                 }
 728                                 arg->count++;
 729                         }
 730                 }
 731         }
 732 }
 733
 734 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 735                      struct sk_buff *skb, struct tcmsg *t)
 736 {
 737         struct tc_u_knode *n = (struct tc_u_knode *)fh;
 738         struct nlattr *nest;
 739
 740         if (n == NULL)
 741                 return skb->len;
 742
 743         t->tcm_handle = n->handle;
 744
 745         nest = nla_nest_start(skb, TCA_OPTIONS);
 746         if (nest == NULL)
 747                 goto nla_put_failure;
 748
 749         if (TC_U32_KEY(n->handle) == 0) {
 750                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
 751                 u32 divisor = ht->divisor + 1;
 752
 753                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
 754                         goto nla_put_failure;
 755         } else {
 756 #ifdef CONFIG_CLS_U32_PERF
 757                 struct tc_u32_pcnt *gpf;
 758 #endif
 759                 int cpu;
 760
 761                 if (nla_put(skb, TCA_U32_SEL,
 762                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 763                             &n->sel))
 764                         goto nla_put_failure;
 765                 if (n->ht_up) {
 766                         u32 htid = n->handle & 0xFFFFF000;
 767                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
 768                                 goto nla_put_failure;
 769                 }
 770                 if (n->res.classid &&
 771                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
 772                         goto nla_put_failure;
 773                 if (n->ht_down &&
 774                     nla_put_u32(skb, TCA_U32_LINK, n->ht_down->handle))
 775                         goto nla_put_failure;
 776
 777 #ifdef CONFIG_CLS_U32_MARK
 778                 if ((n->val || n->mask)) {
 779                         struct tc_u32_mark mark = {.val = n->val,
 780                                                    .mask = n->mask,
 781                                                    .success = 0};
 782
 783                         for_each_possible_cpu(cpu) {
 784                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpu);
 785
 786                                 mark.success += cnt;
 787                         }
 788
 789                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
 790                                 goto nla_put_failure;
 791                 }
 792 #endif
 793
 794                 if (tcf_exts_dump(skb, &n->exts) < 0)
 795                         goto nla_put_failure;
 796
 797 #ifdef CONFIG_NET_CLS_IND
 798                 if (n->ifindex) {
 799                         struct net_device *dev;
 800                         dev = __dev_get_by_index(net, n->ifindex);
 801                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
 802                                 goto nla_put_failure;
 803                 }
 804 #endif
 805 #ifdef CONFIG_CLS_U32_PERF
 806                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
 807                               n->sel.nkeys * sizeof(u64),
 808                               GFP_KERNEL);
 809                 if (!gpf)
 810                         goto nla_put_failure;
 811
 812                 for_each_possible_cpu(cpu) {
 813                         int i;
 814                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
 815
 816                         gpf->rcnt += pf->rcnt;
 817                         gpf->rhit += pf->rhit;
 818                         for (i = 0; i < n->sel.nkeys; i++)
 819                                 gpf->kcnts[i] += pf->kcnts[i];
 820                 }
 821
 822                 if (nla_put(skb, TCA_U32_PCNT,
 823                             sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
 824                             gpf)) {
 825                         kfree(gpf);
 826                         goto nla_put_failure;
 827                 }
 828                 kfree(gpf);
 829 #endif
 830         }
 831
 832         nla_nest_end(skb, nest);
 833
 834         if (TC_U32_KEY(n->handle))
 835                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
 836                         goto nla_put_failure;
 837         return skb->len;
 838
 839 nla_put_failure:
 840         nla_nest_cancel(skb, nest);
 841         return -1;
 842 }
 843
 844 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
 845         .kind           =       "u32",
 846         .classify       =       u32_classify,
 847         .init           =       u32_init,
 848         .destroy        =       u32_destroy,
 849         .get            =       u32_get,
 850         .put            =       u32_put,
 851         .change         =       u32_change,
 852         .delete         =       u32_delete,
 853         .walk           =       u32_walk,
 854         .dump           =       u32_dump,
 855         .owner          =       THIS_MODULE,
 856 };
 857
 858 static int __init init_u32(void)
 859 {
 860         pr_info("u32 classifier\n");
 861 #ifdef CONFIG_CLS_U32_PERF
 862         pr_info("    Performance counters on\n");
 863 #endif
 864 #ifdef CONFIG_NET_CLS_IND
 865         pr_info("    input device check on\n");
 866 #endif
 867 #ifdef CONFIG_NET_CLS_ACT
 868         pr_info("    Actions configured\n");
 869 #endif
 870         return register_tcf_proto_ops(&cls_u32_ops);
 871 }
 872
 873 static void __exit exit_u32(void)
 874 {
 875         unregister_tcf_proto_ops(&cls_u32_ops);
 876 }
 877
 878 module_init(init_u32)
 879 module_exit(exit_u32)
 880 MODULE_LICENSE("GPL");