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/percpu.h>
  40 #include <linux/rtnetlink.h>
  41 #include <linux/skbuff.h>
  42 #include <linux/bitmap.h>
  43 #include <net/netlink.h>
  44 #include <net/act_api.h>
  45 #include <net/pkt_cls.h>
  46
  47 struct tc_u_knode {
  48         struct tc_u_knode __rcu *next;
  49         u32                     handle;
  50         struct tc_u_hnode __rcu *ht_up;
  51         struct tcf_exts         exts;
  52 #ifdef CONFIG_NET_CLS_IND
  53         int                     ifindex;
  54 #endif
  55         u8                      fshift;
  56         struct tcf_result       res;
  57         struct tc_u_hnode __rcu *ht_down;
  58 #ifdef CONFIG_CLS_U32_PERF
  59         struct tc_u32_pcnt __percpu *pf;
  60 #endif
  61 #ifdef CONFIG_CLS_U32_MARK
  62         u32                     val;
  63         u32                     mask;
  64         u32 __percpu            *pcpu_success;
  65 #endif
  66         struct tcf_proto        *tp;
  67         struct tc_u32_sel       sel;
  68         struct rcu_head         rcu;
  69 };
  70
  71 struct tc_u_hnode {
  72         struct tc_u_hnode __rcu *next;
  73         u32                     handle;
  74         u32                     prio;
  75         struct tc_u_common      *tp_c;
  76         int                     refcnt;
  77         unsigned int            divisor;
  78         struct tc_u_knode __rcu *ht[1];
  79         struct rcu_head         rcu;
  80 };
  81
  82 struct tc_u_common {
  83         struct tc_u_hnode __rcu *hlist;
  84         struct Qdisc            *q;
  85         int                     refcnt;
  86         u32                     hgenerator;
  87         struct rcu_head         rcu;
  88 };
  89
  90 static inline unsigned int u32_hash_fold(__be32 key,
  91                                          const struct tc_u32_sel *sel,
  92                                          u8 fshift)
  93 {
  94         unsigned int h = ntohl(key & sel->hmask) >> fshift;
  95
  96         return h;
  97 }
  98
  99 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
 100 {
 101         struct {
 102                 struct tc_u_knode *knode;
 103                 unsigned int      off;
 104         } stack[TC_U32_MAXDEPTH];
 105
 106         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 107         unsigned int off = skb_network_offset(skb);
 108         struct tc_u_knode *n;
 109         int sdepth = 0;
 110         int off2 = 0;
 111         int sel = 0;
 112 #ifdef CONFIG_CLS_U32_PERF
 113         int j;
 114 #endif
 115         int i, r;
 116
 117 next_ht:
 118         n = rcu_dereference_bh(ht->ht[sel]);
 119
 120 next_knode:
 121         if (n) {
 122                 struct tc_u32_key *key = n->sel.keys;
 123
 124 #ifdef CONFIG_CLS_U32_PERF
 125                 __this_cpu_inc(n->pf->rcnt);
 126                 j = 0;
 127 #endif
 128
 129 #ifdef CONFIG_CLS_U32_MARK
 130                 if ((skb->mark & n->mask) != n->val) {
 131                         n = rcu_dereference_bh(n->next);
 132                         goto next_knode;
 133                 } else {
 134                         __this_cpu_inc(*n->pcpu_success);
 135                 }
 136 #endif
 137
 138                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 139                         int toff = off + key->off + (off2 & key->offmask);
 140                         __be32 *data, hdata;
 141
 142                         if (skb_headroom(skb) + toff > INT_MAX)
 143                                 goto out;
 144
 145                         data = skb_header_pointer(skb, toff, 4, &hdata);
 146                         if (!data)
 147                                 goto out;
 148                         if ((*data ^ key->val) & key->mask) {
 149                                 n = rcu_dereference_bh(n->next);
 150                                 goto next_knode;
 151                         }
 152 #ifdef CONFIG_CLS_U32_PERF
 153                         __this_cpu_inc(n->pf->kcnts[j]);
 154                         j++;
 155 #endif
 156                 }
 157
 158                 ht = rcu_dereference_bh(n->ht_down);
 159                 if (!ht) {
 160 check_terminal:
 161                         if (n->sel.flags & TC_U32_TERMINAL) {
 162
 163                                 *res = n->res;
 164 #ifdef CONFIG_NET_CLS_IND
 165                                 if (!tcf_match_indev(skb, n->ifindex)) {
 166                                         n = rcu_dereference_bh(n->next);
 167                                         goto next_knode;
 168                                 }
 169 #endif
 170 #ifdef CONFIG_CLS_U32_PERF
 171                                 __this_cpu_inc(n->pf->rhit);
 172 #endif
 173                                 r = tcf_exts_exec(skb, &n->exts, res);
 174                                 if (r < 0) {
 175                                         n = rcu_dereference_bh(n->next);
 176                                         goto next_knode;
 177                                 }
 178
 179                                 return r;
 180                         }
 181                         n = rcu_dereference_bh(n->next);
 182                         goto next_knode;
 183                 }
 184
 185                 /* PUSH */
 186                 if (sdepth >= TC_U32_MAXDEPTH)
 187                         goto deadloop;
 188                 stack[sdepth].knode = n;
 189                 stack[sdepth].off = off;
 190                 sdepth++;
 191
 192                 ht = rcu_dereference_bh(n->ht_down);
 193                 sel = 0;
 194                 if (ht->divisor) {
 195                         __be32 *data, hdata;
 196
 197                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 198                                                   &hdata);
 199                         if (!data)
 200                                 goto out;
 201                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 202                                                           n->fshift);
 203                 }
 204                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 205                         goto next_ht;
 206
 207                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 208                         off2 = n->sel.off + 3;
 209                         if (n->sel.flags & TC_U32_VAROFFSET) {
 210                                 __be16 *data, hdata;
 211
 212                                 data = skb_header_pointer(skb,
 213                                                           off + n->sel.offoff,
 214                                                           2, &hdata);
 215                                 if (!data)
 216                                         goto out;
 217                                 off2 += ntohs(n->sel.offmask & *data) >>
 218                                         n->sel.offshift;
 219                         }
 220                         off2 &= ~3;
 221                 }
 222                 if (n->sel.flags & TC_U32_EAT) {
 223                         off += off2;
 224                         off2 = 0;
 225                 }
 226
 227                 if (off < skb->len)
 228                         goto next_ht;
 229         }
 230
 231         /* POP */
 232         if (sdepth--) {
 233                 n = stack[sdepth].knode;
 234                 ht = rcu_dereference_bh(n->ht_up);
 235                 off = stack[sdepth].off;
 236                 goto check_terminal;
 237         }
 238 out:
 239         return -1;
 240
 241 deadloop:
 242         net_warn_ratelimited("cls_u32: dead loop\n");
 243         return -1;
 244 }
 245
 246 static struct tc_u_hnode *
 247 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 248 {
 249         struct tc_u_hnode *ht;
 250
 251         for (ht = rtnl_dereference(tp_c->hlist);
 252              ht;
 253              ht = rtnl_dereference(ht->next))
 254                 if (ht->handle == handle)
 255                         break;
 256
 257         return ht;
 258 }
 259
 260 static struct tc_u_knode *
 261 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 262 {
 263         unsigned int sel;
 264         struct tc_u_knode *n = NULL;
 265
 266         sel = TC_U32_HASH(handle);
 267         if (sel > ht->divisor)
 268                 goto out;
 269
 270         for (n = rtnl_dereference(ht->ht[sel]);
 271              n;
 272              n = rtnl_dereference(n->next))
 273                 if (n->handle == handle)
 274                         break;
 275 out:
 276         return n;
 277 }
 278
 279
 280 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 281 {
 282         struct tc_u_hnode *ht;
 283         struct tc_u_common *tp_c = tp->data;
 284
 285         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 286                 ht = rtnl_dereference(tp->root);
 287         else
 288                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 289
 290         if (!ht)
 291                 return 0;
 292
 293         if (TC_U32_KEY(handle) == 0)
 294                 return (unsigned long)ht;
 295
 296         return (unsigned long)u32_lookup_key(ht, handle);
 297 }
 298
 299 static void u32_put(struct tcf_proto *tp, unsigned long f)
 300 {
 301 }
 302
 303 static u32 gen_new_htid(struct tc_u_common *tp_c)
 304 {
 305         int i = 0x800;
 306
 307         /* hgenerator only used inside rtnl lock it is safe to increment
 308          * without read _copy_ update semantics
 309          */
 310         do {
 311                 if (++tp_c->hgenerator == 0x7FF)
 312                         tp_c->hgenerator = 1;
 313         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 314
 315         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 316 }
 317
 318 static int u32_init(struct tcf_proto *tp)
 319 {
 320         struct tc_u_hnode *root_ht;
 321         struct tc_u_common *tp_c;
 322
 323         tp_c = tp->q->u32_node;
 324
 325         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 326         if (root_ht == NULL)
 327                 return -ENOBUFS;
 328
 329         root_ht->divisor = 0;
 330         root_ht->refcnt++;
 331         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 332         root_ht->prio = tp->prio;
 333
 334         if (tp_c == NULL) {
 335                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 336                 if (tp_c == NULL) {
 337                         kfree(root_ht);
 338                         return -ENOBUFS;
 339                 }
 340                 tp_c->q = tp->q;
 341                 tp->q->u32_node = tp_c;
 342         }
 343
 344         tp_c->refcnt++;
 345         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 346         rcu_assign_pointer(tp_c->hlist, root_ht);
 347         root_ht->tp_c = tp_c;
 348
 349         rcu_assign_pointer(tp->root, root_ht);
 350         tp->data = tp_c;
 351         return 0;
 352 }
 353
 354 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
 355 {
 356         tcf_unbind_filter(tp, &n->res);
 357         tcf_exts_destroy(tp, &n->exts);
 358         if (n->ht_down)
 359                 n->ht_down->refcnt--;
 360 #ifdef CONFIG_CLS_U32_PERF
 361         free_percpu(n->pf);
 362 #endif
 363         kfree(n);
 364         return 0;
 365 }
 366
 367 static void u32_delete_key_rcu(struct rcu_head *rcu)
 368 {
 369         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 370
 371         u32_destroy_key(key->tp, key);
 372 }
 373
 374 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 375 {
 376         struct tc_u_knode __rcu **kp;
 377         struct tc_u_knode *pkp;
 378         struct tc_u_hnode *ht = key->ht_up;
 379
 380         if (ht) {
 381                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 382                 for (pkp = rtnl_dereference(*kp); pkp;
 383                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 384                         if (pkp == key) {
 385                                 RCU_INIT_POINTER(*kp, key->next);
 386
 387                                 call_rcu(&key->rcu, u32_delete_key_rcu);
 388                                 return 0;
 389                         }
 390                 }
 391         }
 392         WARN_ON(1);
 393         return 0;
 394 }
 395
 396 static void u32_clear_hnode(struct tc_u_hnode *ht)
 397 {
 398         struct tc_u_knode *n;
 399         unsigned int h;
 400
 401         for (h = 0; h <= ht->divisor; h++) {
 402                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 403                         RCU_INIT_POINTER(ht->ht[h],
 404                                          rtnl_dereference(n->next));
 405                         call_rcu(&n->rcu, u32_delete_key_rcu);
 406                 }
 407         }
 408 }
 409
 410 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 411 {
 412         struct tc_u_common *tp_c = tp->data;
 413         struct tc_u_hnode __rcu **hn;
 414         struct tc_u_hnode *phn;
 415
 416         WARN_ON(ht->refcnt);
 417
 418         u32_clear_hnode(ht);
 419
 420         hn = &tp_c->hlist;
 421         for (phn = rtnl_dereference(*hn);
 422              phn;
 423              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 424                 if (phn == ht) {
 425                         RCU_INIT_POINTER(*hn, ht->next);
 426                         kfree_rcu(ht, rcu);
 427                         return 0;
 428                 }
 429         }
 430
 431         return -ENOENT;
 432 }
 433
 434 static void u32_destroy(struct tcf_proto *tp)
 435 {
 436         struct tc_u_common *tp_c = tp->data;
 437         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 438
 439         WARN_ON(root_ht == NULL);
 440
 441         if (root_ht && --root_ht->refcnt == 0)
 442                 u32_destroy_hnode(tp, root_ht);
 443
 444         if (--tp_c->refcnt == 0) {
 445                 struct tc_u_hnode *ht;
 446
 447                 tp->q->u32_node = NULL;
 448
 449                 for (ht = rtnl_dereference(tp_c->hlist);
 450                      ht;
 451                      ht = rtnl_dereference(ht->next)) {
 452                         ht->refcnt--;
 453                         u32_clear_hnode(ht);
 454                 }
 455
 456                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 457                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 458                         kfree_rcu(ht, rcu);
 459                 }
 460
 461                 kfree(tp_c);
 462         }
 463
 464         tp->data = NULL;
 465 }
 466
 467 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 468 {
 469         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 470         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 471
 472         if (ht == NULL)
 473                 return 0;
 474
 475         if (TC_U32_KEY(ht->handle))
 476                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 477
 478         if (root_ht == ht)
 479                 return -EINVAL;
 480
 481         if (ht->refcnt == 1) {
 482                 ht->refcnt--;
 483                 u32_destroy_hnode(tp, ht);
 484         } else {
 485                 return -EBUSY;
 486         }
 487
 488         return 0;
 489 }
 490
 491 #define NR_U32_NODE (1<<12)
 492 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 493 {
 494         struct tc_u_knode *n;
 495         unsigned long i;
 496         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 497                                         GFP_KERNEL);
 498         if (!bitmap)
 499                 return handle | 0xFFF;
 500
 501         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 502              n;
 503              n = rtnl_dereference(n->next))
 504                 set_bit(TC_U32_NODE(n->handle), bitmap);
 505
 506         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 507         if (i >= NR_U32_NODE)
 508                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 509
 510         kfree(bitmap);
 511         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 512 }
 513
 514 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 515         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 516         [TCA_U32_HASH]          = { .type = NLA_U32 },
 517         [TCA_U32_LINK]          = { .type = NLA_U32 },
 518         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 519         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 520         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 521         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 522 };
 523
 524 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 525                          unsigned long base, struct tc_u_hnode *ht,
 526                          struct tc_u_knode *n, struct nlattr **tb,
 527                          struct nlattr *est, bool ovr)
 528 {
 529         int err;
 530         struct tcf_exts e;
 531
 532         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 533         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 534         if (err < 0)
 535                 return err;
 536
 537         err = -EINVAL;
 538         if (tb[TCA_U32_LINK]) {
 539                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 540                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 541
 542                 if (TC_U32_KEY(handle))
 543                         goto errout;
 544
 545                 if (handle) {
 546                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 547
 548                         if (ht_down == NULL)
 549                                 goto errout;
 550                         ht_down->refcnt++;
 551                 }
 552
 553                 ht_old = rtnl_dereference(n->ht_down);
 554                 rcu_assign_pointer(n->ht_down, ht_down);
 555
 556                 if (ht_old)
 557                         ht_old->refcnt--;
 558         }
 559         if (tb[TCA_U32_CLASSID]) {
 560                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 561                 tcf_bind_filter(tp, &n->res, base);
 562         }
 563
 564 #ifdef CONFIG_NET_CLS_IND
 565         if (tb[TCA_U32_INDEV]) {
 566                 int ret;
 567                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 568                 if (ret < 0)
 569                         goto errout;
 570                 n->ifindex = ret;
 571         }
 572 #endif
 573         tcf_exts_change(tp, &n->exts, &e);
 574
 575         return 0;
 576 errout:
 577         tcf_exts_destroy(tp, &e);
 578         return err;
 579 }
 580
 581 static int u32_change(struct net *net, struct sk_buff *in_skb,
 582                       struct tcf_proto *tp, unsigned long base, u32 handle,
 583                       struct nlattr **tca,
 584                       unsigned long *arg, bool ovr)
 585 {
 586         struct tc_u_common *tp_c = tp->data;
 587         struct tc_u_hnode *ht;
 588         struct tc_u_knode *n;
 589         struct tc_u32_sel *s;
 590         struct nlattr *opt = tca[TCA_OPTIONS];
 591         struct nlattr *tb[TCA_U32_MAX + 1];
 592         u32 htid;
 593         int err;
 594 #ifdef CONFIG_CLS_U32_PERF
 595         size_t size;
 596 #endif
 597
 598         if (opt == NULL)
 599                 return handle ? -EINVAL : 0;
 600
 601         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 602         if (err < 0)
 603                 return err;
 604
 605         n = (struct tc_u_knode *)*arg;
 606         if (n) {
 607                 if (TC_U32_KEY(n->handle) == 0)
 608                         return -EINVAL;
 609
 610                 return u32_set_parms(net, tp, base, n->ht_up, n, tb,
 611                                      tca[TCA_RATE], ovr);
 612         }
 613
 614         if (tb[TCA_U32_DIVISOR]) {
 615                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 616
 617                 if (--divisor > 0x100)
 618                         return -EINVAL;
 619                 if (TC_U32_KEY(handle))
 620                         return -EINVAL;
 621                 if (handle == 0) {
 622                         handle = gen_new_htid(tp->data);
 623                         if (handle == 0)
 624                                 return -ENOMEM;
 625                 }
 626                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 627                 if (ht == NULL)
 628                         return -ENOBUFS;
 629                 ht->tp_c = tp_c;
 630                 ht->refcnt = 1;
 631                 ht->divisor = divisor;
 632                 ht->handle = handle;
 633                 ht->prio = tp->prio;
 634                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 635                 rcu_assign_pointer(tp_c->hlist, ht);
 636                 *arg = (unsigned long)ht;
 637                 return 0;
 638         }
 639
 640         if (tb[TCA_U32_HASH]) {
 641                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 642                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 643                         ht = rtnl_dereference(tp->root);
 644                         htid = ht->handle;
 645                 } else {
 646                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 647                         if (ht == NULL)
 648                                 return -EINVAL;
 649                 }
 650         } else {
 651                 ht = rtnl_dereference(tp->root);
 652                 htid = ht->handle;
 653         }
 654
 655         if (ht->divisor < TC_U32_HASH(htid))
 656                 return -EINVAL;
 657
 658         if (handle) {
 659                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 660                         return -EINVAL;
 661                 handle = htid | TC_U32_NODE(handle);
 662         } else
 663                 handle = gen_new_kid(ht, htid);
 664
 665         if (tb[TCA_U32_SEL] == NULL)
 666                 return -EINVAL;
 667
 668         s = nla_data(tb[TCA_U32_SEL]);
 669
 670         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 671         if (n == NULL)
 672                 return -ENOBUFS;
 673
 674 #ifdef CONFIG_CLS_U32_PERF
 675         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 676         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 677         if (!n->pf) {
 678                 kfree(n);
 679                 return -ENOBUFS;
 680         }
 681 #endif
 682
 683         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 684         n->ht_up = ht;
 685         n->handle = handle;
 686         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 687         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 688         n->tp = tp;
 689
 690 #ifdef CONFIG_CLS_U32_MARK
 691         n->pcpu_success = alloc_percpu(u32);
 692
 693         if (tb[TCA_U32_MARK]) {
 694                 struct tc_u32_mark *mark;
 695
 696                 mark = nla_data(tb[TCA_U32_MARK]);
 697                 n->val = mark->val;
 698                 n->mask = mark->mask;
 699         }
 700 #endif
 701
 702         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
 703         if (err == 0) {
 704                 struct tc_u_knode __rcu **ins;
 705                 struct tc_u_knode *pins;
 706
 707                 ins = &ht->ht[TC_U32_HASH(handle)];
 708                 for (pins = rtnl_dereference(*ins); pins;
 709                      ins = &pins->next, pins = rtnl_dereference(*ins))
 710                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
 711                                 break;
 712
 713                 RCU_INIT_POINTER(n->next, pins);
 714                 rcu_assign_pointer(*ins, n);
 715
 716                 *arg = (unsigned long)n;
 717                 return 0;
 718         }
 719 #ifdef CONFIG_CLS_U32_PERF
 720         free_percpu(n->pf);
 721 #endif
 722         kfree(n);
 723         return err;
 724 }
 725
 726 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 727 {
 728         struct tc_u_common *tp_c = tp->data;
 729         struct tc_u_hnode *ht;
 730         struct tc_u_knode *n;
 731         unsigned int h;
 732
 733         if (arg->stop)
 734                 return;
 735
 736         for (ht = rtnl_dereference(tp_c->hlist);
 737              ht;
 738              ht = rtnl_dereference(ht->next)) {
 739                 if (ht->prio != tp->prio)
 740                         continue;
 741                 if (arg->count >= arg->skip) {
 742                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 743                                 arg->stop = 1;
 744                                 return;
 745                         }
 746                 }
 747                 arg->count++;
 748                 for (h = 0; h <= ht->divisor; h++) {
 749                         for (n = rtnl_dereference(ht->ht[h]);
 750                              n;
 751                              n = rtnl_dereference(n->next)) {
 752                                 if (arg->count < arg->skip) {
 753                                         arg->count++;
 754                                         continue;
 755                                 }
 756                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 757                                         arg->stop = 1;
 758                                         return;
 759                                 }
 760                                 arg->count++;
 761                         }
 762                 }
 763         }
 764 }
 765
 766 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 767                      struct sk_buff *skb, struct tcmsg *t)
 768 {
 769         struct tc_u_knode *n = (struct tc_u_knode *)fh;
 770         struct tc_u_hnode *ht_up, *ht_down;
 771         struct nlattr *nest;
 772
 773         if (n == NULL)
 774                 return skb->len;
 775
 776         t->tcm_handle = n->handle;
 777
 778         nest = nla_nest_start(skb, TCA_OPTIONS);
 779         if (nest == NULL)
 780                 goto nla_put_failure;
 781
 782         if (TC_U32_KEY(n->handle) == 0) {
 783                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
 784                 u32 divisor = ht->divisor + 1;
 785
 786                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
 787                         goto nla_put_failure;
 788         } else {
 789 #ifdef CONFIG_CLS_U32_PERF
 790                 struct tc_u32_pcnt *gpf;
 791 #endif
 792                 int cpu;
 793
 794                 if (nla_put(skb, TCA_U32_SEL,
 795                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 796                             &n->sel))
 797                         goto nla_put_failure;
 798
 799                 ht_up = rtnl_dereference(n->ht_up);
 800                 if (ht_up) {
 801                         u32 htid = n->handle & 0xFFFFF000;
 802                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
 803                                 goto nla_put_failure;
 804                 }
 805                 if (n->res.classid &&
 806                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
 807                         goto nla_put_failure;
 808
 809                 ht_down = rtnl_dereference(n->ht_down);
 810                 if (ht_down &&
 811                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
 812                         goto nla_put_failure;
 813
 814 #ifdef CONFIG_CLS_U32_MARK
 815                 if ((n->val || n->mask)) {
 816                         struct tc_u32_mark mark = {.val = n->val,
 817                                                    .mask = n->mask,
 818                                                    .success = 0};
 819
 820                         for_each_possible_cpu(cpu) {
 821                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpu);
 822
 823                                 mark.success += cnt;
 824                         }
 825
 826                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
 827                                 goto nla_put_failure;
 828                 }
 829 #endif
 830
 831                 if (tcf_exts_dump(skb, &n->exts) < 0)
 832                         goto nla_put_failure;
 833
 834 #ifdef CONFIG_NET_CLS_IND
 835                 if (n->ifindex) {
 836                         struct net_device *dev;
 837                         dev = __dev_get_by_index(net, n->ifindex);
 838                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
 839                                 goto nla_put_failure;
 840                 }
 841 #endif
 842 #ifdef CONFIG_CLS_U32_PERF
 843                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
 844                               n->sel.nkeys * sizeof(u64),
 845                               GFP_KERNEL);
 846                 if (!gpf)
 847                         goto nla_put_failure;
 848
 849                 for_each_possible_cpu(cpu) {
 850                         int i;
 851                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
 852
 853                         gpf->rcnt += pf->rcnt;
 854                         gpf->rhit += pf->rhit;
 855                         for (i = 0; i < n->sel.nkeys; i++)
 856                                 gpf->kcnts[i] += pf->kcnts[i];
 857                 }
 858
 859                 if (nla_put(skb, TCA_U32_PCNT,
 860                             sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
 861                             gpf)) {
 862                         kfree(gpf);
 863                         goto nla_put_failure;
 864                 }
 865                 kfree(gpf);
 866 #endif
 867         }
 868
 869         nla_nest_end(skb, nest);
 870
 871         if (TC_U32_KEY(n->handle))
 872                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
 873                         goto nla_put_failure;
 874         return skb->len;
 875
 876 nla_put_failure:
 877         nla_nest_cancel(skb, nest);
 878         return -1;
 879 }
 880
 881 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
 882         .kind           =       "u32",
 883         .classify       =       u32_classify,
 884         .init           =       u32_init,
 885         .destroy        =       u32_destroy,
 886         .get            =       u32_get,
 887         .put            =       u32_put,
 888         .change         =       u32_change,
 889         .delete         =       u32_delete,
 890         .walk           =       u32_walk,
 891         .dump           =       u32_dump,
 892         .owner          =       THIS_MODULE,
 893 };
 894
 895 static int __init init_u32(void)
 896 {
 897         pr_info("u32 classifier\n");
 898 #ifdef CONFIG_CLS_U32_PERF
 899         pr_info("    Performance counters on\n");
 900 #endif
 901 #ifdef CONFIG_NET_CLS_IND
 902         pr_info("    input device check on\n");
 903 #endif
 904 #ifdef CONFIG_NET_CLS_ACT
 905         pr_info("    Actions configured\n");
 906 #endif
 907         return register_tcf_proto_ops(&cls_u32_ops);
 908 }
 909
 910 static void __exit exit_u32(void)
 911 {
 912         unregister_tcf_proto_ops(&cls_u32_ops);
 913 }
 914
 915 module_init(init_u32)
 916 module_exit(exit_u32)
 917 MODULE_LICENSE("GPL");