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 rcu_head         rcu;
  68         /* The 'sel' field MUST be the last field in structure to allow for
  69          * tc_u32_keys allocated at end of structure.
  70          */
  71         struct tc_u32_sel       sel;
  72 };
  73
  74 struct tc_u_hnode {
  75         struct tc_u_hnode __rcu *next;
  76         u32                     handle;
  77         u32                     prio;
  78         struct tc_u_common      *tp_c;
  79         int                     refcnt;
  80         unsigned int            divisor;
  81         struct tc_u_knode __rcu *ht[1];
  82         struct rcu_head         rcu;
  83 };
  84
  85 struct tc_u_common {
  86         struct tc_u_hnode __rcu *hlist;
  87         struct Qdisc            *q;
  88         int                     refcnt;
  89         u32                     hgenerator;
  90         struct rcu_head         rcu;
  91 };
  92
  93 static inline unsigned int u32_hash_fold(__be32 key,
  94                                          const struct tc_u32_sel *sel,
  95                                          u8 fshift)
  96 {
  97         unsigned int h = ntohl(key & sel->hmask) >> fshift;
  98
  99         return h;
 100 }
 101
 102 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
 103 {
 104         struct {
 105                 struct tc_u_knode *knode;
 106                 unsigned int      off;
 107         } stack[TC_U32_MAXDEPTH];
 108
 109         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 110         unsigned int off = skb_network_offset(skb);
 111         struct tc_u_knode *n;
 112         int sdepth = 0;
 113         int off2 = 0;
 114         int sel = 0;
 115 #ifdef CONFIG_CLS_U32_PERF
 116         int j;
 117 #endif
 118         int i, r;
 119
 120 next_ht:
 121         n = rcu_dereference_bh(ht->ht[sel]);
 122
 123 next_knode:
 124         if (n) {
 125                 struct tc_u32_key *key = n->sel.keys;
 126
 127 #ifdef CONFIG_CLS_U32_PERF
 128                 __this_cpu_inc(n->pf->rcnt);
 129                 j = 0;
 130 #endif
 131
 132 #ifdef CONFIG_CLS_U32_MARK
 133                 if ((skb->mark & n->mask) != n->val) {
 134                         n = rcu_dereference_bh(n->next);
 135                         goto next_knode;
 136                 } else {
 137                         __this_cpu_inc(*n->pcpu_success);
 138                 }
 139 #endif
 140
 141                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 142                         int toff = off + key->off + (off2 & key->offmask);
 143                         __be32 *data, hdata;
 144
 145                         if (skb_headroom(skb) + toff > INT_MAX)
 146                                 goto out;
 147
 148                         data = skb_header_pointer(skb, toff, 4, &hdata);
 149                         if (!data)
 150                                 goto out;
 151                         if ((*data ^ key->val) & key->mask) {
 152                                 n = rcu_dereference_bh(n->next);
 153                                 goto next_knode;
 154                         }
 155 #ifdef CONFIG_CLS_U32_PERF
 156                         __this_cpu_inc(n->pf->kcnts[j]);
 157                         j++;
 158 #endif
 159                 }
 160
 161                 ht = rcu_dereference_bh(n->ht_down);
 162                 if (!ht) {
 163 check_terminal:
 164                         if (n->sel.flags & TC_U32_TERMINAL) {
 165
 166                                 *res = n->res;
 167 #ifdef CONFIG_NET_CLS_IND
 168                                 if (!tcf_match_indev(skb, n->ifindex)) {
 169                                         n = rcu_dereference_bh(n->next);
 170                                         goto next_knode;
 171                                 }
 172 #endif
 173 #ifdef CONFIG_CLS_U32_PERF
 174                                 __this_cpu_inc(n->pf->rhit);
 175 #endif
 176                                 r = tcf_exts_exec(skb, &n->exts, res);
 177                                 if (r < 0) {
 178                                         n = rcu_dereference_bh(n->next);
 179                                         goto next_knode;
 180                                 }
 181
 182                                 return r;
 183                         }
 184                         n = rcu_dereference_bh(n->next);
 185                         goto next_knode;
 186                 }
 187
 188                 /* PUSH */
 189                 if (sdepth >= TC_U32_MAXDEPTH)
 190                         goto deadloop;
 191                 stack[sdepth].knode = n;
 192                 stack[sdepth].off = off;
 193                 sdepth++;
 194
 195                 ht = rcu_dereference_bh(n->ht_down);
 196                 sel = 0;
 197                 if (ht->divisor) {
 198                         __be32 *data, hdata;
 199
 200                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 201                                                   &hdata);
 202                         if (!data)
 203                                 goto out;
 204                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 205                                                           n->fshift);
 206                 }
 207                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 208                         goto next_ht;
 209
 210                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 211                         off2 = n->sel.off + 3;
 212                         if (n->sel.flags & TC_U32_VAROFFSET) {
 213                                 __be16 *data, hdata;
 214
 215                                 data = skb_header_pointer(skb,
 216                                                           off + n->sel.offoff,
 217                                                           2, &hdata);
 218                                 if (!data)
 219                                         goto out;
 220                                 off2 += ntohs(n->sel.offmask & *data) >>
 221                                         n->sel.offshift;
 222                         }
 223                         off2 &= ~3;
 224                 }
 225                 if (n->sel.flags & TC_U32_EAT) {
 226                         off += off2;
 227                         off2 = 0;
 228                 }
 229
 230                 if (off < skb->len)
 231                         goto next_ht;
 232         }
 233
 234         /* POP */
 235         if (sdepth--) {
 236                 n = stack[sdepth].knode;
 237                 ht = rcu_dereference_bh(n->ht_up);
 238                 off = stack[sdepth].off;
 239                 goto check_terminal;
 240         }
 241 out:
 242         return -1;
 243
 244 deadloop:
 245         net_warn_ratelimited("cls_u32: dead loop\n");
 246         return -1;
 247 }
 248
 249 static struct tc_u_hnode *
 250 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 251 {
 252         struct tc_u_hnode *ht;
 253
 254         for (ht = rtnl_dereference(tp_c->hlist);
 255              ht;
 256              ht = rtnl_dereference(ht->next))
 257                 if (ht->handle == handle)
 258                         break;
 259
 260         return ht;
 261 }
 262
 263 static struct tc_u_knode *
 264 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 265 {
 266         unsigned int sel;
 267         struct tc_u_knode *n = NULL;
 268
 269         sel = TC_U32_HASH(handle);
 270         if (sel > ht->divisor)
 271                 goto out;
 272
 273         for (n = rtnl_dereference(ht->ht[sel]);
 274              n;
 275              n = rtnl_dereference(n->next))
 276                 if (n->handle == handle)
 277                         break;
 278 out:
 279         return n;
 280 }
 281
 282
 283 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 284 {
 285         struct tc_u_hnode *ht;
 286         struct tc_u_common *tp_c = tp->data;
 287
 288         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 289                 ht = rtnl_dereference(tp->root);
 290         else
 291                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 292
 293         if (!ht)
 294                 return 0;
 295
 296         if (TC_U32_KEY(handle) == 0)
 297                 return (unsigned long)ht;
 298
 299         return (unsigned long)u32_lookup_key(ht, handle);
 300 }
 301
 302 static void u32_put(struct tcf_proto *tp, unsigned long f)
 303 {
 304 }
 305
 306 static u32 gen_new_htid(struct tc_u_common *tp_c)
 307 {
 308         int i = 0x800;
 309
 310         /* hgenerator only used inside rtnl lock it is safe to increment
 311          * without read _copy_ update semantics
 312          */
 313         do {
 314                 if (++tp_c->hgenerator == 0x7FF)
 315                         tp_c->hgenerator = 1;
 316         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 317
 318         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 319 }
 320
 321 static int u32_init(struct tcf_proto *tp)
 322 {
 323         struct tc_u_hnode *root_ht;
 324         struct tc_u_common *tp_c;
 325
 326         tp_c = tp->q->u32_node;
 327
 328         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 329         if (root_ht == NULL)
 330                 return -ENOBUFS;
 331
 332         root_ht->divisor = 0;
 333         root_ht->refcnt++;
 334         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 335         root_ht->prio = tp->prio;
 336
 337         if (tp_c == NULL) {
 338                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 339                 if (tp_c == NULL) {
 340                         kfree(root_ht);
 341                         return -ENOBUFS;
 342                 }
 343                 tp_c->q = tp->q;
 344                 tp->q->u32_node = tp_c;
 345         }
 346
 347         tp_c->refcnt++;
 348         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 349         rcu_assign_pointer(tp_c->hlist, root_ht);
 350         root_ht->tp_c = tp_c;
 351
 352         rcu_assign_pointer(tp->root, root_ht);
 353         tp->data = tp_c;
 354         return 0;
 355 }
 356
 357 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
 358 {
 359         tcf_unbind_filter(tp, &n->res);
 360         tcf_exts_destroy(tp, &n->exts);
 361         if (n->ht_down)
 362                 n->ht_down->refcnt--;
 363 #ifdef CONFIG_CLS_U32_PERF
 364         free_percpu(n->pf);
 365 #endif
 366         kfree(n);
 367         return 0;
 368 }
 369
 370 static void u32_delete_key_rcu(struct rcu_head *rcu)
 371 {
 372         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 373
 374         u32_destroy_key(key->tp, key);
 375 }
 376
 377 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 378 {
 379         struct tc_u_knode __rcu **kp;
 380         struct tc_u_knode *pkp;
 381         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 382
 383         if (ht) {
 384                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 385                 for (pkp = rtnl_dereference(*kp); pkp;
 386                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 387                         if (pkp == key) {
 388                                 RCU_INIT_POINTER(*kp, key->next);
 389
 390                                 call_rcu(&key->rcu, u32_delete_key_rcu);
 391                                 return 0;
 392                         }
 393                 }
 394         }
 395         WARN_ON(1);
 396         return 0;
 397 }
 398
 399 static void u32_clear_hnode(struct tc_u_hnode *ht)
 400 {
 401         struct tc_u_knode *n;
 402         unsigned int h;
 403
 404         for (h = 0; h <= ht->divisor; h++) {
 405                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 406                         RCU_INIT_POINTER(ht->ht[h],
 407                                          rtnl_dereference(n->next));
 408                         call_rcu(&n->rcu, u32_delete_key_rcu);
 409                 }
 410         }
 411 }
 412
 413 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 414 {
 415         struct tc_u_common *tp_c = tp->data;
 416         struct tc_u_hnode __rcu **hn;
 417         struct tc_u_hnode *phn;
 418
 419         WARN_ON(ht->refcnt);
 420
 421         u32_clear_hnode(ht);
 422
 423         hn = &tp_c->hlist;
 424         for (phn = rtnl_dereference(*hn);
 425              phn;
 426              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 427                 if (phn == ht) {
 428                         RCU_INIT_POINTER(*hn, ht->next);
 429                         kfree_rcu(ht, rcu);
 430                         return 0;
 431                 }
 432         }
 433
 434         return -ENOENT;
 435 }
 436
 437 static void u32_destroy(struct tcf_proto *tp)
 438 {
 439         struct tc_u_common *tp_c = tp->data;
 440         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 441
 442         WARN_ON(root_ht == NULL);
 443
 444         if (root_ht && --root_ht->refcnt == 0)
 445                 u32_destroy_hnode(tp, root_ht);
 446
 447         if (--tp_c->refcnt == 0) {
 448                 struct tc_u_hnode *ht;
 449
 450                 tp->q->u32_node = NULL;
 451
 452                 for (ht = rtnl_dereference(tp_c->hlist);
 453                      ht;
 454                      ht = rtnl_dereference(ht->next)) {
 455                         ht->refcnt--;
 456                         u32_clear_hnode(ht);
 457                 }
 458
 459                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 460                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 461                         kfree_rcu(ht, rcu);
 462                 }
 463
 464                 kfree(tp_c);
 465         }
 466
 467         tp->data = NULL;
 468 }
 469
 470 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 471 {
 472         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 473         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 474
 475         if (ht == NULL)
 476                 return 0;
 477
 478         if (TC_U32_KEY(ht->handle))
 479                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 480
 481         if (root_ht == ht)
 482                 return -EINVAL;
 483
 484         if (ht->refcnt == 1) {
 485                 ht->refcnt--;
 486                 u32_destroy_hnode(tp, ht);
 487         } else {
 488                 return -EBUSY;
 489         }
 490
 491         return 0;
 492 }
 493
 494 #define NR_U32_NODE (1<<12)
 495 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 496 {
 497         struct tc_u_knode *n;
 498         unsigned long i;
 499         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 500                                         GFP_KERNEL);
 501         if (!bitmap)
 502                 return handle | 0xFFF;
 503
 504         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 505              n;
 506              n = rtnl_dereference(n->next))
 507                 set_bit(TC_U32_NODE(n->handle), bitmap);
 508
 509         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 510         if (i >= NR_U32_NODE)
 511                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 512
 513         kfree(bitmap);
 514         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 515 }
 516
 517 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 518         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 519         [TCA_U32_HASH]          = { .type = NLA_U32 },
 520         [TCA_U32_LINK]          = { .type = NLA_U32 },
 521         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 522         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 523         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 524         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 525 };
 526
 527 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 528                          unsigned long base, struct tc_u_hnode *ht,
 529                          struct tc_u_knode *n, struct nlattr **tb,
 530                          struct nlattr *est, bool ovr)
 531 {
 532         int err;
 533         struct tcf_exts e;
 534
 535         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 536         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 537         if (err < 0)
 538                 return err;
 539
 540         err = -EINVAL;
 541         if (tb[TCA_U32_LINK]) {
 542                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 543                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 544
 545                 if (TC_U32_KEY(handle))
 546                         goto errout;
 547
 548                 if (handle) {
 549                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 550
 551                         if (ht_down == NULL)
 552                                 goto errout;
 553                         ht_down->refcnt++;
 554                 }
 555
 556                 ht_old = rtnl_dereference(n->ht_down);
 557                 rcu_assign_pointer(n->ht_down, ht_down);
 558
 559                 if (ht_old)
 560                         ht_old->refcnt--;
 561         }
 562         if (tb[TCA_U32_CLASSID]) {
 563                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 564                 tcf_bind_filter(tp, &n->res, base);
 565         }
 566
 567 #ifdef CONFIG_NET_CLS_IND
 568         if (tb[TCA_U32_INDEV]) {
 569                 int ret;
 570                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 571                 if (ret < 0)
 572                         goto errout;
 573                 n->ifindex = ret;
 574         }
 575 #endif
 576         tcf_exts_change(tp, &n->exts, &e);
 577
 578         return 0;
 579 errout:
 580         tcf_exts_destroy(tp, &e);
 581         return err;
 582 }
 583
 584 static int u32_change(struct net *net, struct sk_buff *in_skb,
 585                       struct tcf_proto *tp, unsigned long base, u32 handle,
 586                       struct nlattr **tca,
 587                       unsigned long *arg, bool ovr)
 588 {
 589         struct tc_u_common *tp_c = tp->data;
 590         struct tc_u_hnode *ht;
 591         struct tc_u_knode *n;
 592         struct tc_u32_sel *s;
 593         struct nlattr *opt = tca[TCA_OPTIONS];
 594         struct nlattr *tb[TCA_U32_MAX + 1];
 595         u32 htid;
 596         int err;
 597 #ifdef CONFIG_CLS_U32_PERF
 598         size_t size;
 599 #endif
 600
 601         if (opt == NULL)
 602                 return handle ? -EINVAL : 0;
 603
 604         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 605         if (err < 0)
 606                 return err;
 607
 608         n = (struct tc_u_knode *)*arg;
 609         if (n) {
 610                 if (TC_U32_KEY(n->handle) == 0)
 611                         return -EINVAL;
 612
 613                 return u32_set_parms(net, tp, base,
 614                                      rtnl_dereference(n->ht_up), n, tb,
 615                                      tca[TCA_RATE], ovr);
 616         }
 617
 618         if (tb[TCA_U32_DIVISOR]) {
 619                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 620
 621                 if (--divisor > 0x100)
 622                         return -EINVAL;
 623                 if (TC_U32_KEY(handle))
 624                         return -EINVAL;
 625                 if (handle == 0) {
 626                         handle = gen_new_htid(tp->data);
 627                         if (handle == 0)
 628                                 return -ENOMEM;
 629                 }
 630                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 631                 if (ht == NULL)
 632                         return -ENOBUFS;
 633                 ht->tp_c = tp_c;
 634                 ht->refcnt = 1;
 635                 ht->divisor = divisor;
 636                 ht->handle = handle;
 637                 ht->prio = tp->prio;
 638                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 639                 rcu_assign_pointer(tp_c->hlist, ht);
 640                 *arg = (unsigned long)ht;
 641                 return 0;
 642         }
 643
 644         if (tb[TCA_U32_HASH]) {
 645                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 646                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 647                         ht = rtnl_dereference(tp->root);
 648                         htid = ht->handle;
 649                 } else {
 650                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 651                         if (ht == NULL)
 652                                 return -EINVAL;
 653                 }
 654         } else {
 655                 ht = rtnl_dereference(tp->root);
 656                 htid = ht->handle;
 657         }
 658
 659         if (ht->divisor < TC_U32_HASH(htid))
 660                 return -EINVAL;
 661
 662         if (handle) {
 663                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 664                         return -EINVAL;
 665                 handle = htid | TC_U32_NODE(handle);
 666         } else
 667                 handle = gen_new_kid(ht, htid);
 668
 669         if (tb[TCA_U32_SEL] == NULL)
 670                 return -EINVAL;
 671
 672         s = nla_data(tb[TCA_U32_SEL]);
 673
 674         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 675         if (n == NULL)
 676                 return -ENOBUFS;
 677
 678 #ifdef CONFIG_CLS_U32_PERF
 679         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 680         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 681         if (!n->pf) {
 682                 kfree(n);
 683                 return -ENOBUFS;
 684         }
 685 #endif
 686
 687         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 688         RCU_INIT_POINTER(n->ht_up, ht);
 689         n->handle = handle;
 690         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 691         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 692         n->tp = tp;
 693
 694 #ifdef CONFIG_CLS_U32_MARK
 695         n->pcpu_success = alloc_percpu(u32);
 696
 697         if (tb[TCA_U32_MARK]) {
 698                 struct tc_u32_mark *mark;
 699
 700                 mark = nla_data(tb[TCA_U32_MARK]);
 701                 n->val = mark->val;
 702                 n->mask = mark->mask;
 703         }
 704 #endif
 705
 706         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
 707         if (err == 0) {
 708                 struct tc_u_knode __rcu **ins;
 709                 struct tc_u_knode *pins;
 710
 711                 ins = &ht->ht[TC_U32_HASH(handle)];
 712                 for (pins = rtnl_dereference(*ins); pins;
 713                      ins = &pins->next, pins = rtnl_dereference(*ins))
 714                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
 715                                 break;
 716
 717                 RCU_INIT_POINTER(n->next, pins);
 718                 rcu_assign_pointer(*ins, n);
 719
 720                 *arg = (unsigned long)n;
 721                 return 0;
 722         }
 723 #ifdef CONFIG_CLS_U32_PERF
 724         free_percpu(n->pf);
 725 #endif
 726         kfree(n);
 727         return err;
 728 }
 729
 730 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 731 {
 732         struct tc_u_common *tp_c = tp->data;
 733         struct tc_u_hnode *ht;
 734         struct tc_u_knode *n;
 735         unsigned int h;
 736
 737         if (arg->stop)
 738                 return;
 739
 740         for (ht = rtnl_dereference(tp_c->hlist);
 741              ht;
 742              ht = rtnl_dereference(ht->next)) {
 743                 if (ht->prio != tp->prio)
 744                         continue;
 745                 if (arg->count >= arg->skip) {
 746                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 747                                 arg->stop = 1;
 748                                 return;
 749                         }
 750                 }
 751                 arg->count++;
 752                 for (h = 0; h <= ht->divisor; h++) {
 753                         for (n = rtnl_dereference(ht->ht[h]);
 754                              n;
 755                              n = rtnl_dereference(n->next)) {
 756                                 if (arg->count < arg->skip) {
 757                                         arg->count++;
 758                                         continue;
 759                                 }
 760                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 761                                         arg->stop = 1;
 762                                         return;
 763                                 }
 764                                 arg->count++;
 765                         }
 766                 }
 767         }
 768 }
 769
 770 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 771                      struct sk_buff *skb, struct tcmsg *t)
 772 {
 773         struct tc_u_knode *n = (struct tc_u_knode *)fh;
 774         struct tc_u_hnode *ht_up, *ht_down;
 775         struct nlattr *nest;
 776
 777         if (n == NULL)
 778                 return skb->len;
 779
 780         t->tcm_handle = n->handle;
 781
 782         nest = nla_nest_start(skb, TCA_OPTIONS);
 783         if (nest == NULL)
 784                 goto nla_put_failure;
 785
 786         if (TC_U32_KEY(n->handle) == 0) {
 787                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
 788                 u32 divisor = ht->divisor + 1;
 789
 790                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
 791                         goto nla_put_failure;
 792         } else {
 793 #ifdef CONFIG_CLS_U32_PERF
 794                 struct tc_u32_pcnt *gpf;
 795                 int cpu;
 796 #endif
 797
 798                 if (nla_put(skb, TCA_U32_SEL,
 799                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 800                             &n->sel))
 801                         goto nla_put_failure;
 802
 803                 ht_up = rtnl_dereference(n->ht_up);
 804                 if (ht_up) {
 805                         u32 htid = n->handle & 0xFFFFF000;
 806                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
 807                                 goto nla_put_failure;
 808                 }
 809                 if (n->res.classid &&
 810                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
 811                         goto nla_put_failure;
 812
 813                 ht_down = rtnl_dereference(n->ht_down);
 814                 if (ht_down &&
 815                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
 816                         goto nla_put_failure;
 817
 818 #ifdef CONFIG_CLS_U32_MARK
 819                 if ((n->val || n->mask)) {
 820                         struct tc_u32_mark mark = {.val = n->val,
 821                                                    .mask = n->mask,
 822                                                    .success = 0};
 823                         int cpum;
 824
 825                         for_each_possible_cpu(cpum) {
 826                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
 827
 828                                 mark.success += cnt;
 829                         }
 830
 831                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
 832                                 goto nla_put_failure;
 833                 }
 834 #endif
 835
 836                 if (tcf_exts_dump(skb, &n->exts) < 0)
 837                         goto nla_put_failure;
 838
 839 #ifdef CONFIG_NET_CLS_IND
 840                 if (n->ifindex) {
 841                         struct net_device *dev;
 842                         dev = __dev_get_by_index(net, n->ifindex);
 843                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
 844                                 goto nla_put_failure;
 845                 }
 846 #endif
 847 #ifdef CONFIG_CLS_U32_PERF
 848                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
 849                               n->sel.nkeys * sizeof(u64),
 850                               GFP_KERNEL);
 851                 if (!gpf)
 852                         goto nla_put_failure;
 853
 854                 for_each_possible_cpu(cpu) {
 855                         int i;
 856                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
 857
 858                         gpf->rcnt += pf->rcnt;
 859                         gpf->rhit += pf->rhit;
 860                         for (i = 0; i < n->sel.nkeys; i++)
 861                                 gpf->kcnts[i] += pf->kcnts[i];
 862                 }
 863
 864                 if (nla_put(skb, TCA_U32_PCNT,
 865                             sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
 866                             gpf)) {
 867                         kfree(gpf);
 868                         goto nla_put_failure;
 869                 }
 870                 kfree(gpf);
 871 #endif
 872         }
 873
 874         nla_nest_end(skb, nest);
 875
 876         if (TC_U32_KEY(n->handle))
 877                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
 878                         goto nla_put_failure;
 879         return skb->len;
 880
 881 nla_put_failure:
 882         nla_nest_cancel(skb, nest);
 883         return -1;
 884 }
 885
 886 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
 887         .kind           =       "u32",
 888         .classify       =       u32_classify,
 889         .init           =       u32_init,
 890         .destroy        =       u32_destroy,
 891         .get            =       u32_get,
 892         .put            =       u32_put,
 893         .change         =       u32_change,
 894         .delete         =       u32_delete,
 895         .walk           =       u32_walk,
 896         .dump           =       u32_dump,
 897         .owner          =       THIS_MODULE,
 898 };
 899
 900 static int __init init_u32(void)
 901 {
 902         pr_info("u32 classifier\n");
 903 #ifdef CONFIG_CLS_U32_PERF
 904         pr_info("    Performance counters on\n");
 905 #endif
 906 #ifdef CONFIG_NET_CLS_IND
 907         pr_info("    input device check on\n");
 908 #endif
 909 #ifdef CONFIG_NET_CLS_ACT
 910         pr_info("    Actions configured\n");
 911 #endif
 912         return register_tcf_proto_ops(&cls_u32_ops);
 913 }
 914
 915 static void __exit exit_u32(void)
 916 {
 917         unregister_tcf_proto_ops(&cls_u32_ops);
 918 }
 919
 920 module_init(init_u32)
 921 module_exit(exit_u32)
 922 MODULE_LICENSE("GPL");