kernel/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 rcu_head         rcu;
  82         /* The 'ht' field MUST be the last field in structure to allow for
  83          * more entries allocated at end of structure.
  84          */
  85         struct tc_u_knode __rcu *ht[1];
  86 };
  87
  88 struct tc_u_common {
  89         struct tc_u_hnode __rcu *hlist;
  90         struct Qdisc            *q;
  91         int                     refcnt;
  92         u32                     hgenerator;
  93         struct rcu_head         rcu;
  94 };
  95
  96 static inline unsigned int u32_hash_fold(__be32 key,
  97                                          const struct tc_u32_sel *sel,
  98                                          u8 fshift)
  99 {
 100         unsigned int h = ntohl(key & sel->hmask) >> fshift;
 101
 102         return h;
 103 }
 104
 105 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
 106 {
 107         struct {
 108                 struct tc_u_knode *knode;
 109                 unsigned int      off;
 110         } stack[TC_U32_MAXDEPTH];
 111
 112         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 113         unsigned int off = skb_network_offset(skb);
 114         struct tc_u_knode *n;
 115         int sdepth = 0;
 116         int off2 = 0;
 117         int sel = 0;
 118 #ifdef CONFIG_CLS_U32_PERF
 119         int j;
 120 #endif
 121         int i, r;
 122
 123 next_ht:
 124         n = rcu_dereference_bh(ht->ht[sel]);
 125
 126 next_knode:
 127         if (n) {
 128                 struct tc_u32_key *key = n->sel.keys;
 129
 130 #ifdef CONFIG_CLS_U32_PERF
 131                 __this_cpu_inc(n->pf->rcnt);
 132                 j = 0;
 133 #endif
 134
 135 #ifdef CONFIG_CLS_U32_MARK
 136                 if ((skb->mark & n->mask) != n->val) {
 137                         n = rcu_dereference_bh(n->next);
 138                         goto next_knode;
 139                 } else {
 140                         __this_cpu_inc(*n->pcpu_success);
 141                 }
 142 #endif
 143
 144                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 145                         int toff = off + key->off + (off2 & key->offmask);
 146                         __be32 *data, hdata;
 147
 148                         if (skb_headroom(skb) + toff > INT_MAX)
 149                                 goto out;
 150
 151                         data = skb_header_pointer(skb, toff, 4, &hdata);
 152                         if (!data)
 153                                 goto out;
 154                         if ((*data ^ key->val) & key->mask) {
 155                                 n = rcu_dereference_bh(n->next);
 156                                 goto next_knode;
 157                         }
 158 #ifdef CONFIG_CLS_U32_PERF
 159                         __this_cpu_inc(n->pf->kcnts[j]);
 160                         j++;
 161 #endif
 162                 }
 163
 164                 ht = rcu_dereference_bh(n->ht_down);
 165                 if (!ht) {
 166 check_terminal:
 167                         if (n->sel.flags & TC_U32_TERMINAL) {
 168
 169                                 *res = n->res;
 170 #ifdef CONFIG_NET_CLS_IND
 171                                 if (!tcf_match_indev(skb, n->ifindex)) {
 172                                         n = rcu_dereference_bh(n->next);
 173                                         goto next_knode;
 174                                 }
 175 #endif
 176 #ifdef CONFIG_CLS_U32_PERF
 177                                 __this_cpu_inc(n->pf->rhit);
 178 #endif
 179                                 r = tcf_exts_exec(skb, &n->exts, res);
 180                                 if (r < 0) {
 181                                         n = rcu_dereference_bh(n->next);
 182                                         goto next_knode;
 183                                 }
 184
 185                                 return r;
 186                         }
 187                         n = rcu_dereference_bh(n->next);
 188                         goto next_knode;
 189                 }
 190
 191                 /* PUSH */
 192                 if (sdepth >= TC_U32_MAXDEPTH)
 193                         goto deadloop;
 194                 stack[sdepth].knode = n;
 195                 stack[sdepth].off = off;
 196                 sdepth++;
 197
 198                 ht = rcu_dereference_bh(n->ht_down);
 199                 sel = 0;
 200                 if (ht->divisor) {
 201                         __be32 *data, hdata;
 202
 203                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 204                                                   &hdata);
 205                         if (!data)
 206                                 goto out;
 207                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 208                                                           n->fshift);
 209                 }
 210                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 211                         goto next_ht;
 212
 213                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 214                         off2 = n->sel.off + 3;
 215                         if (n->sel.flags & TC_U32_VAROFFSET) {
 216                                 __be16 *data, hdata;
 217
 218                                 data = skb_header_pointer(skb,
 219                                                           off + n->sel.offoff,
 220                                                           2, &hdata);
 221                                 if (!data)
 222                                         goto out;
 223                                 off2 += ntohs(n->sel.offmask & *data) >>
 224                                         n->sel.offshift;
 225                         }
 226                         off2 &= ~3;
 227                 }
 228                 if (n->sel.flags & TC_U32_EAT) {
 229                         off += off2;
 230                         off2 = 0;
 231                 }
 232
 233                 if (off < skb->len)
 234                         goto next_ht;
 235         }
 236
 237         /* POP */
 238         if (sdepth--) {
 239                 n = stack[sdepth].knode;
 240                 ht = rcu_dereference_bh(n->ht_up);
 241                 off = stack[sdepth].off;
 242                 goto check_terminal;
 243         }
 244 out:
 245         return -1;
 246
 247 deadloop:
 248         net_warn_ratelimited("cls_u32: dead loop\n");
 249         return -1;
 250 }
 251
 252 static struct tc_u_hnode *
 253 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 254 {
 255         struct tc_u_hnode *ht;
 256
 257         for (ht = rtnl_dereference(tp_c->hlist);
 258              ht;
 259              ht = rtnl_dereference(ht->next))
 260                 if (ht->handle == handle)
 261                         break;
 262
 263         return ht;
 264 }
 265
 266 static struct tc_u_knode *
 267 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 268 {
 269         unsigned int sel;
 270         struct tc_u_knode *n = NULL;
 271
 272         sel = TC_U32_HASH(handle);
 273         if (sel > ht->divisor)
 274                 goto out;
 275
 276         for (n = rtnl_dereference(ht->ht[sel]);
 277              n;
 278              n = rtnl_dereference(n->next))
 279                 if (n->handle == handle)
 280                         break;
 281 out:
 282         return n;
 283 }
 284
 285
 286 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 287 {
 288         struct tc_u_hnode *ht;
 289         struct tc_u_common *tp_c = tp->data;
 290
 291         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 292                 ht = rtnl_dereference(tp->root);
 293         else
 294                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 295
 296         if (!ht)
 297                 return 0;
 298
 299         if (TC_U32_KEY(handle) == 0)
 300                 return (unsigned long)ht;
 301
 302         return (unsigned long)u32_lookup_key(ht, handle);
 303 }
 304
 305 static u32 gen_new_htid(struct tc_u_common *tp_c)
 306 {
 307         int i = 0x800;
 308
 309         /* hgenerator only used inside rtnl lock it is safe to increment
 310          * without read _copy_ update semantics
 311          */
 312         do {
 313                 if (++tp_c->hgenerator == 0x7FF)
 314                         tp_c->hgenerator = 1;
 315         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 316
 317         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 318 }
 319
 320 static int u32_init(struct tcf_proto *tp)
 321 {
 322         struct tc_u_hnode *root_ht;
 323         struct tc_u_common *tp_c;
 324
 325         tp_c = tp->q->u32_node;
 326
 327         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 328         if (root_ht == NULL)
 329                 return -ENOBUFS;
 330
 331         root_ht->divisor = 0;
 332         root_ht->refcnt++;
 333         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 334         root_ht->prio = tp->prio;
 335
 336         if (tp_c == NULL) {
 337                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 338                 if (tp_c == NULL) {
 339                         kfree(root_ht);
 340                         return -ENOBUFS;
 341                 }
 342                 tp_c->q = tp->q;
 343                 tp->q->u32_node = tp_c;
 344         }
 345
 346         tp_c->refcnt++;
 347         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 348         rcu_assign_pointer(tp_c->hlist, root_ht);
 349         root_ht->tp_c = tp_c;
 350
 351         rcu_assign_pointer(tp->root, root_ht);
 352         tp->data = tp_c;
 353         return 0;
 354 }
 355
 356 static int u32_destroy_key(struct tcf_proto *tp,
 357                            struct tc_u_knode *n,
 358                            bool free_pf)
 359 {
 360         tcf_exts_destroy(&n->exts);
 361         if (n->ht_down)
 362                 n->ht_down->refcnt--;
 363 #ifdef CONFIG_CLS_U32_PERF
 364         if (free_pf)
 365                 free_percpu(n->pf);
 366 #endif
 367 #ifdef CONFIG_CLS_U32_MARK
 368         if (free_pf)
 369                 free_percpu(n->pcpu_success);
 370 #endif
 371         kfree(n);
 372         return 0;
 373 }
 374
 375 /* u32_delete_key_rcu should be called when free'ing a copied
 376  * version of a tc_u_knode obtained from u32_init_knode(). When
 377  * copies are obtained from u32_init_knode() the statistics are
 378  * shared between the old and new copies to allow readers to
 379  * continue to update the statistics during the copy. To support
 380  * this the u32_delete_key_rcu variant does not free the percpu
 381  * statistics.
 382  */
 383 static void u32_delete_key_rcu(struct rcu_head *rcu)
 384 {
 385         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 386
 387         u32_destroy_key(key->tp, key, false);
 388 }
 389
 390 /* u32_delete_key_freepf_rcu is the rcu callback variant
 391  * that free's the entire structure including the statistics
 392  * percpu variables. Only use this if the key is not a copy
 393  * returned by u32_init_knode(). See u32_delete_key_rcu()
 394  * for the variant that should be used with keys return from
 395  * u32_init_knode()
 396  */
 397 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 398 {
 399         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 400
 401         u32_destroy_key(key->tp, key, true);
 402 }
 403
 404 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 405 {
 406         struct tc_u_knode __rcu **kp;
 407         struct tc_u_knode *pkp;
 408         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 409
 410         if (ht) {
 411                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 412                 for (pkp = rtnl_dereference(*kp); pkp;
 413                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 414                         if (pkp == key) {
 415                                 RCU_INIT_POINTER(*kp, key->next);
 416
 417                                 tcf_unbind_filter(tp, &key->res);
 418                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 419                                 return 0;
 420                         }
 421                 }
 422         }
 423         WARN_ON(1);
 424         return 0;
 425 }
 426
 427 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 428 {
 429         struct tc_u_knode *n;
 430         unsigned int h;
 431
 432         for (h = 0; h <= ht->divisor; h++) {
 433                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 434                         RCU_INIT_POINTER(ht->ht[h],
 435                                          rtnl_dereference(n->next));
 436                         tcf_unbind_filter(tp, &n->res);
 437                         call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 438                 }
 439         }
 440 }
 441
 442 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 443 {
 444         struct tc_u_common *tp_c = tp->data;
 445         struct tc_u_hnode __rcu **hn;
 446         struct tc_u_hnode *phn;
 447
 448         WARN_ON(ht->refcnt);
 449
 450         u32_clear_hnode(tp, ht);
 451
 452         hn = &tp_c->hlist;
 453         for (phn = rtnl_dereference(*hn);
 454              phn;
 455              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 456                 if (phn == ht) {
 457                         RCU_INIT_POINTER(*hn, ht->next);
 458                         kfree_rcu(ht, rcu);
 459                         return 0;
 460                 }
 461         }
 462
 463         return -ENOENT;
 464 }
 465
 466 static bool ht_empty(struct tc_u_hnode *ht)
 467 {
 468         unsigned int h;
 469
 470         for (h = 0; h <= ht->divisor; h++)
 471                 if (rcu_access_pointer(ht->ht[h]))
 472                         return false;
 473
 474         return true;
 475 }
 476
 477 static bool u32_destroy(struct tcf_proto *tp, bool force)
 478 {
 479         struct tc_u_common *tp_c = tp->data;
 480         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 481
 482         WARN_ON(root_ht == NULL);
 483
 484         if (!force) {
 485                 if (root_ht) {
 486                         if (root_ht->refcnt > 1)
 487                                 return false;
 488                         if (root_ht->refcnt == 1) {
 489                                 if (!ht_empty(root_ht))
 490                                         return false;
 491                         }
 492                 }
 493         }
 494
 495         if (root_ht && --root_ht->refcnt == 0)
 496                 u32_destroy_hnode(tp, root_ht);
 497
 498         if (--tp_c->refcnt == 0) {
 499                 struct tc_u_hnode *ht;
 500
 501                 tp->q->u32_node = NULL;
 502
 503                 for (ht = rtnl_dereference(tp_c->hlist);
 504                      ht;
 505                      ht = rtnl_dereference(ht->next)) {
 506                         ht->refcnt--;
 507                         u32_clear_hnode(tp, ht);
 508                 }
 509
 510                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 511                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 512                         kfree_rcu(ht, rcu);
 513                 }
 514
 515                 kfree(tp_c);
 516         }
 517
 518         tp->data = NULL;
 519         return true;
 520 }
 521
 522 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 523 {
 524         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 525         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 526
 527         if (ht == NULL)
 528                 return 0;
 529
 530         if (TC_U32_KEY(ht->handle))
 531                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 532
 533         if (root_ht == ht)
 534                 return -EINVAL;
 535
 536         if (ht->refcnt == 1) {
 537                 ht->refcnt--;
 538                 u32_destroy_hnode(tp, ht);
 539         } else {
 540                 return -EBUSY;
 541         }
 542
 543         return 0;
 544 }
 545
 546 #define NR_U32_NODE (1<<12)
 547 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 548 {
 549         struct tc_u_knode *n;
 550         unsigned long i;
 551         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 552                                         GFP_KERNEL);
 553         if (!bitmap)
 554                 return handle | 0xFFF;
 555
 556         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 557              n;
 558              n = rtnl_dereference(n->next))
 559                 set_bit(TC_U32_NODE(n->handle), bitmap);
 560
 561         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 562         if (i >= NR_U32_NODE)
 563                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 564
 565         kfree(bitmap);
 566         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 567 }
 568
 569 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 570         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 571         [TCA_U32_HASH]          = { .type = NLA_U32 },
 572         [TCA_U32_LINK]          = { .type = NLA_U32 },
 573         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 574         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 575         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 576         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 577 };
 578
 579 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 580                          unsigned long base, struct tc_u_hnode *ht,
 581                          struct tc_u_knode *n, struct nlattr **tb,
 582                          struct nlattr *est, bool ovr)
 583 {
 584         int err;
 585         struct tcf_exts e;
 586
 587         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 588         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 589         if (err < 0)
 590                 return err;
 591
 592         err = -EINVAL;
 593         if (tb[TCA_U32_LINK]) {
 594                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 595                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 596
 597                 if (TC_U32_KEY(handle))
 598                         goto errout;
 599
 600                 if (handle) {
 601                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 602
 603                         if (ht_down == NULL)
 604                                 goto errout;
 605                         ht_down->refcnt++;
 606                 }
 607
 608                 ht_old = rtnl_dereference(n->ht_down);
 609                 rcu_assign_pointer(n->ht_down, ht_down);
 610
 611                 if (ht_old)
 612                         ht_old->refcnt--;
 613         }
 614         if (tb[TCA_U32_CLASSID]) {
 615                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 616                 tcf_bind_filter(tp, &n->res, base);
 617         }
 618
 619 #ifdef CONFIG_NET_CLS_IND
 620         if (tb[TCA_U32_INDEV]) {
 621                 int ret;
 622                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 623                 if (ret < 0)
 624                         goto errout;
 625                 n->ifindex = ret;
 626         }
 627 #endif
 628         tcf_exts_change(tp, &n->exts, &e);
 629
 630         return 0;
 631 errout:
 632         tcf_exts_destroy(&e);
 633         return err;
 634 }
 635
 636 static void u32_replace_knode(struct tcf_proto *tp,
 637                               struct tc_u_common *tp_c,
 638                               struct tc_u_knode *n)
 639 {
 640         struct tc_u_knode __rcu **ins;
 641         struct tc_u_knode *pins;
 642         struct tc_u_hnode *ht;
 643
 644         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 645                 ht = rtnl_dereference(tp->root);
 646         else
 647                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 648
 649         ins = &ht->ht[TC_U32_HASH(n->handle)];
 650
 651         /* The node must always exist for it to be replaced if this is not the
 652          * case then something went very wrong elsewhere.
 653          */
 654         for (pins = rtnl_dereference(*ins); ;
 655              ins = &pins->next, pins = rtnl_dereference(*ins))
 656                 if (pins->handle == n->handle)
 657                         break;
 658
 659         RCU_INIT_POINTER(n->next, pins->next);
 660         rcu_assign_pointer(*ins, n);
 661 }
 662
 663 static struct tc_u_knode *u32_init_knode(struct tcf_proto *tp,
 664                                          struct tc_u_knode *n)
 665 {
 666         struct tc_u_knode *new;
 667         struct tc_u32_sel *s = &n->sel;
 668
 669         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 670                       GFP_KERNEL);
 671
 672         if (!new)
 673                 return NULL;
 674
 675         RCU_INIT_POINTER(new->next, n->next);
 676         new->handle = n->handle;
 677         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 678
 679 #ifdef CONFIG_NET_CLS_IND
 680         new->ifindex = n->ifindex;
 681 #endif
 682         new->fshift = n->fshift;
 683         new->res = n->res;
 684         RCU_INIT_POINTER(new->ht_down, n->ht_down);
 685
 686         /* bump reference count as long as we hold pointer to structure */
 687         if (new->ht_down)
 688                 new->ht_down->refcnt++;
 689
 690 #ifdef CONFIG_CLS_U32_PERF
 691         /* Statistics may be incremented by readers during update
 692          * so we must keep them in tact. When the node is later destroyed
 693          * a special destroy call must be made to not free the pf memory.
 694          */
 695         new->pf = n->pf;
 696 #endif
 697
 698 #ifdef CONFIG_CLS_U32_MARK
 699         new->val = n->val;
 700         new->mask = n->mask;
 701         /* Similarly success statistics must be moved as pointers */
 702         new->pcpu_success = n->pcpu_success;
 703 #endif
 704         new->tp = tp;
 705         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 706
 707         tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE);
 708
 709         return new;
 710 }
 711
 712 static int u32_change(struct net *net, struct sk_buff *in_skb,
 713                       struct tcf_proto *tp, unsigned long base, u32 handle,
 714                       struct nlattr **tca,
 715                       unsigned long *arg, bool ovr)
 716 {
 717         struct tc_u_common *tp_c = tp->data;
 718         struct tc_u_hnode *ht;
 719         struct tc_u_knode *n;
 720         struct tc_u32_sel *s;
 721         struct nlattr *opt = tca[TCA_OPTIONS];
 722         struct nlattr *tb[TCA_U32_MAX + 1];
 723         u32 htid;
 724         int err;
 725 #ifdef CONFIG_CLS_U32_PERF
 726         size_t size;
 727 #endif
 728
 729         if (opt == NULL)
 730                 return handle ? -EINVAL : 0;
 731
 732         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 733         if (err < 0)
 734                 return err;
 735
 736         n = (struct tc_u_knode *)*arg;
 737         if (n) {
 738                 struct tc_u_knode *new;
 739
 740                 if (TC_U32_KEY(n->handle) == 0)
 741                         return -EINVAL;
 742
 743                 new = u32_init_knode(tp, n);
 744                 if (!new)
 745                         return -ENOMEM;
 746
 747                 err = u32_set_parms(net, tp, base,
 748                                     rtnl_dereference(n->ht_up), new, tb,
 749                                     tca[TCA_RATE], ovr);
 750
 751                 if (err) {
 752                         u32_destroy_key(tp, new, false);
 753                         return err;
 754                 }
 755
 756                 u32_replace_knode(tp, tp_c, new);
 757                 tcf_unbind_filter(tp, &n->res);
 758                 call_rcu(&n->rcu, u32_delete_key_rcu);
 759                 return 0;
 760         }
 761
 762         if (tb[TCA_U32_DIVISOR]) {
 763                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 764
 765                 if (--divisor > 0x100)
 766                         return -EINVAL;
 767                 if (TC_U32_KEY(handle))
 768                         return -EINVAL;
 769                 if (handle == 0) {
 770                         handle = gen_new_htid(tp->data);
 771                         if (handle == 0)
 772                                 return -ENOMEM;
 773                 }
 774                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 775                 if (ht == NULL)
 776                         return -ENOBUFS;
 777                 ht->tp_c = tp_c;
 778                 ht->refcnt = 1;
 779                 ht->divisor = divisor;
 780                 ht->handle = handle;
 781                 ht->prio = tp->prio;
 782                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 783                 rcu_assign_pointer(tp_c->hlist, ht);
 784                 *arg = (unsigned long)ht;
 785                 return 0;
 786         }
 787
 788         if (tb[TCA_U32_HASH]) {
 789                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 790                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 791                         ht = rtnl_dereference(tp->root);
 792                         htid = ht->handle;
 793                 } else {
 794                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 795                         if (ht == NULL)
 796                                 return -EINVAL;
 797                 }
 798         } else {
 799                 ht = rtnl_dereference(tp->root);
 800                 htid = ht->handle;
 801         }
 802
 803         if (ht->divisor < TC_U32_HASH(htid))
 804                 return -EINVAL;
 805
 806         if (handle) {
 807                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 808                         return -EINVAL;
 809                 handle = htid | TC_U32_NODE(handle);
 810         } else
 811                 handle = gen_new_kid(ht, htid);
 812
 813         if (tb[TCA_U32_SEL] == NULL)
 814                 return -EINVAL;
 815
 816         s = nla_data(tb[TCA_U32_SEL]);
 817
 818         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 819         if (n == NULL)
 820                 return -ENOBUFS;
 821
 822 #ifdef CONFIG_CLS_U32_PERF
 823         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 824         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 825         if (!n->pf) {
 826                 kfree(n);
 827                 return -ENOBUFS;
 828         }
 829 #endif
 830
 831         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 832         RCU_INIT_POINTER(n->ht_up, ht);
 833         n->handle = handle;
 834         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 835         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 836         n->tp = tp;
 837
 838 #ifdef CONFIG_CLS_U32_MARK
 839         n->pcpu_success = alloc_percpu(u32);
 840         if (!n->pcpu_success) {
 841                 err = -ENOMEM;
 842                 goto errout;
 843         }
 844
 845         if (tb[TCA_U32_MARK]) {
 846                 struct tc_u32_mark *mark;
 847
 848                 mark = nla_data(tb[TCA_U32_MARK]);
 849                 n->val = mark->val;
 850                 n->mask = mark->mask;
 851         }
 852 #endif
 853
 854         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
 855         if (err == 0) {
 856                 struct tc_u_knode __rcu **ins;
 857                 struct tc_u_knode *pins;
 858
 859                 ins = &ht->ht[TC_U32_HASH(handle)];
 860                 for (pins = rtnl_dereference(*ins); pins;
 861                      ins = &pins->next, pins = rtnl_dereference(*ins))
 862                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
 863                                 break;
 864
 865                 RCU_INIT_POINTER(n->next, pins);
 866                 rcu_assign_pointer(*ins, n);
 867
 868                 *arg = (unsigned long)n;
 869                 return 0;
 870         }
 871
 872 #ifdef CONFIG_CLS_U32_MARK
 873         free_percpu(n->pcpu_success);
 874 errout:
 875 #endif
 876
 877 #ifdef CONFIG_CLS_U32_PERF
 878         free_percpu(n->pf);
 879 #endif
 880         kfree(n);
 881         return err;
 882 }
 883
 884 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 885 {
 886         struct tc_u_common *tp_c = tp->data;
 887         struct tc_u_hnode *ht;
 888         struct tc_u_knode *n;
 889         unsigned int h;
 890
 891         if (arg->stop)
 892                 return;
 893
 894         for (ht = rtnl_dereference(tp_c->hlist);
 895              ht;
 896              ht = rtnl_dereference(ht->next)) {
 897                 if (ht->prio != tp->prio)
 898                         continue;
 899                 if (arg->count >= arg->skip) {
 900                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 901                                 arg->stop = 1;
 902                                 return;
 903                         }
 904                 }
 905                 arg->count++;
 906                 for (h = 0; h <= ht->divisor; h++) {
 907                         for (n = rtnl_dereference(ht->ht[h]);
 908                              n;
 909                              n = rtnl_dereference(n->next)) {
 910                                 if (arg->count < arg->skip) {
 911                                         arg->count++;
 912                                         continue;
 913                                 }
 914                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 915                                         arg->stop = 1;
 916                                         return;
 917                                 }
 918                                 arg->count++;
 919                         }
 920                 }
 921         }
 922 }
 923
 924 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 925                      struct sk_buff *skb, struct tcmsg *t)
 926 {
 927         struct tc_u_knode *n = (struct tc_u_knode *)fh;
 928         struct tc_u_hnode *ht_up, *ht_down;
 929         struct nlattr *nest;
 930
 931         if (n == NULL)
 932                 return skb->len;
 933
 934         t->tcm_handle = n->handle;
 935
 936         nest = nla_nest_start(skb, TCA_OPTIONS);
 937         if (nest == NULL)
 938                 goto nla_put_failure;
 939
 940         if (TC_U32_KEY(n->handle) == 0) {
 941                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
 942                 u32 divisor = ht->divisor + 1;
 943
 944                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
 945                         goto nla_put_failure;
 946         } else {
 947 #ifdef CONFIG_CLS_U32_PERF
 948                 struct tc_u32_pcnt *gpf;
 949                 int cpu;
 950 #endif
 951
 952                 if (nla_put(skb, TCA_U32_SEL,
 953                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 954                             &n->sel))
 955                         goto nla_put_failure;
 956
 957                 ht_up = rtnl_dereference(n->ht_up);
 958                 if (ht_up) {
 959                         u32 htid = n->handle & 0xFFFFF000;
 960                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
 961                                 goto nla_put_failure;
 962                 }
 963                 if (n->res.classid &&
 964                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
 965                         goto nla_put_failure;
 966
 967                 ht_down = rtnl_dereference(n->ht_down);
 968                 if (ht_down &&
 969                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
 970                         goto nla_put_failure;
 971
 972 #ifdef CONFIG_CLS_U32_MARK
 973                 if ((n->val || n->mask)) {
 974                         struct tc_u32_mark mark = {.val = n->val,
 975                                                    .mask = n->mask,
 976                                                    .success = 0};
 977                         int cpum;
 978
 979                         for_each_possible_cpu(cpum) {
 980                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
 981
 982                                 mark.success += cnt;
 983                         }
 984
 985                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
 986                                 goto nla_put_failure;
 987                 }
 988 #endif
 989
 990                 if (tcf_exts_dump(skb, &n->exts) < 0)
 991                         goto nla_put_failure;
 992
 993 #ifdef CONFIG_NET_CLS_IND
 994                 if (n->ifindex) {
 995                         struct net_device *dev;
 996                         dev = __dev_get_by_index(net, n->ifindex);
 997                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
 998                                 goto nla_put_failure;
 999                 }
1000 #endif
1001 #ifdef CONFIG_CLS_U32_PERF
1002                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1003                               n->sel.nkeys * sizeof(u64),
1004                               GFP_KERNEL);
1005                 if (!gpf)
1006                         goto nla_put_failure;
1007
1008                 for_each_possible_cpu(cpu) {
1009                         int i;
1010                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1011
1012                         gpf->rcnt += pf->rcnt;
1013                         gpf->rhit += pf->rhit;
1014                         for (i = 0; i < n->sel.nkeys; i++)
1015                                 gpf->kcnts[i] += pf->kcnts[i];
1016                 }
1017
1018                 if (nla_put(skb, TCA_U32_PCNT,
1019                             sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
1020                             gpf)) {
1021                         kfree(gpf);
1022                         goto nla_put_failure;
1023                 }
1024                 kfree(gpf);
1025 #endif
1026         }
1027
1028         nla_nest_end(skb, nest);
1029
1030         if (TC_U32_KEY(n->handle))
1031                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1032                         goto nla_put_failure;
1033         return skb->len;
1034
1035 nla_put_failure:
1036         nla_nest_cancel(skb, nest);
1037         return -1;
1038 }
1039
1040 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1041         .kind           =       "u32",
1042         .classify       =       u32_classify,
1043         .init           =       u32_init,
1044         .destroy        =       u32_destroy,
1045         .get            =       u32_get,
1046         .change         =       u32_change,
1047         .delete         =       u32_delete,
1048         .walk           =       u32_walk,
1049         .dump           =       u32_dump,
1050         .owner          =       THIS_MODULE,
1051 };
1052
1053 static int __init init_u32(void)
1054 {
1055         pr_info("u32 classifier\n");
1056 #ifdef CONFIG_CLS_U32_PERF
1057         pr_info("    Performance counters on\n");
1058 #endif
1059 #ifdef CONFIG_NET_CLS_IND
1060         pr_info("    input device check on\n");
1061 #endif
1062 #ifdef CONFIG_NET_CLS_ACT
1063         pr_info("    Actions configured\n");
1064 #endif
1065         return register_tcf_proto_ops(&cls_u32_ops);
1066 }
1067
1068 static void __exit exit_u32(void)
1069 {
1070         unregister_tcf_proto_ops(&cls_u32_ops);
1071 }
1072
1073 module_init(init_u32)
1074 module_exit(exit_u32)
1075 MODULE_LICENSE("GPL");