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,
 358                            struct tc_u_knode *n,
 359                            bool free_pf)
 360 {
 361         tcf_exts_destroy(&n->exts);
 362         if (n->ht_down)
 363                 n->ht_down->refcnt--;
 364 #ifdef CONFIG_CLS_U32_PERF
 365         if (free_pf)
 366                 free_percpu(n->pf);
 367 #endif
 368 #ifdef CONFIG_CLS_U32_MARK
 369         if (free_pf)
 370                 free_percpu(n->pcpu_success);
 371 #endif
 372         kfree(n);
 373         return 0;
 374 }
 375
 376 /* u32_delete_key_rcu should be called when free'ing a copied
 377  * version of a tc_u_knode obtained from u32_init_knode(). When
 378  * copies are obtained from u32_init_knode() the statistics are
 379  * shared between the old and new copies to allow readers to
 380  * continue to update the statistics during the copy. To support
 381  * this the u32_delete_key_rcu variant does not free the percpu
 382  * statistics.
 383  */
 384 static void u32_delete_key_rcu(struct rcu_head *rcu)
 385 {
 386         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 387
 388         u32_destroy_key(key->tp, key, false);
 389 }
 390
 391 /* u32_delete_key_freepf_rcu is the rcu callback variant
 392  * that free's the entire structure including the statistics
 393  * percpu variables. Only use this if the key is not a copy
 394  * returned by u32_init_knode(). See u32_delete_key_rcu()
 395  * for the variant that should be used with keys return from
 396  * u32_init_knode()
 397  */
 398 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 399 {
 400         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 401
 402         u32_destroy_key(key->tp, key, true);
 403 }
 404
 405 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 406 {
 407         struct tc_u_knode __rcu **kp;
 408         struct tc_u_knode *pkp;
 409         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 410
 411         if (ht) {
 412                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 413                 for (pkp = rtnl_dereference(*kp); pkp;
 414                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 415                         if (pkp == key) {
 416                                 RCU_INIT_POINTER(*kp, key->next);
 417
 418                                 tcf_unbind_filter(tp, &key->res);
 419                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 420                                 return 0;
 421                         }
 422                 }
 423         }
 424         WARN_ON(1);
 425         return 0;
 426 }
 427
 428 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 429 {
 430         struct tc_u_knode *n;
 431         unsigned int h;
 432
 433         for (h = 0; h <= ht->divisor; h++) {
 434                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 435                         RCU_INIT_POINTER(ht->ht[h],
 436                                          rtnl_dereference(n->next));
 437                         tcf_unbind_filter(tp, &n->res);
 438                         call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 439                 }
 440         }
 441 }
 442
 443 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 444 {
 445         struct tc_u_common *tp_c = tp->data;
 446         struct tc_u_hnode __rcu **hn;
 447         struct tc_u_hnode *phn;
 448
 449         WARN_ON(ht->refcnt);
 450
 451         u32_clear_hnode(tp, ht);
 452
 453         hn = &tp_c->hlist;
 454         for (phn = rtnl_dereference(*hn);
 455              phn;
 456              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 457                 if (phn == ht) {
 458                         RCU_INIT_POINTER(*hn, ht->next);
 459                         kfree_rcu(ht, rcu);
 460                         return 0;
 461                 }
 462         }
 463
 464         return -ENOENT;
 465 }
 466
 467 static void u32_destroy(struct tcf_proto *tp)
 468 {
 469         struct tc_u_common *tp_c = tp->data;
 470         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 471
 472         WARN_ON(root_ht == NULL);
 473
 474         if (root_ht && --root_ht->refcnt == 0)
 475                 u32_destroy_hnode(tp, root_ht);
 476
 477         if (--tp_c->refcnt == 0) {
 478                 struct tc_u_hnode *ht;
 479
 480                 tp->q->u32_node = NULL;
 481
 482                 for (ht = rtnl_dereference(tp_c->hlist);
 483                      ht;
 484                      ht = rtnl_dereference(ht->next)) {
 485                         ht->refcnt--;
 486                         u32_clear_hnode(tp, ht);
 487                 }
 488
 489                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 490                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 491                         kfree_rcu(ht, rcu);
 492                 }
 493
 494                 kfree(tp_c);
 495         }
 496
 497         tp->data = NULL;
 498 }
 499
 500 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 501 {
 502         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 503         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 504
 505         if (ht == NULL)
 506                 return 0;
 507
 508         if (TC_U32_KEY(ht->handle))
 509                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 510
 511         if (root_ht == ht)
 512                 return -EINVAL;
 513
 514         if (ht->refcnt == 1) {
 515                 ht->refcnt--;
 516                 u32_destroy_hnode(tp, ht);
 517         } else {
 518                 return -EBUSY;
 519         }
 520
 521         return 0;
 522 }
 523
 524 #define NR_U32_NODE (1<<12)
 525 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 526 {
 527         struct tc_u_knode *n;
 528         unsigned long i;
 529         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 530                                         GFP_KERNEL);
 531         if (!bitmap)
 532                 return handle | 0xFFF;
 533
 534         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 535              n;
 536              n = rtnl_dereference(n->next))
 537                 set_bit(TC_U32_NODE(n->handle), bitmap);
 538
 539         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 540         if (i >= NR_U32_NODE)
 541                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 542
 543         kfree(bitmap);
 544         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 545 }
 546
 547 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 548         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 549         [TCA_U32_HASH]          = { .type = NLA_U32 },
 550         [TCA_U32_LINK]          = { .type = NLA_U32 },
 551         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 552         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 553         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 554         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 555 };
 556
 557 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 558                          unsigned long base, struct tc_u_hnode *ht,
 559                          struct tc_u_knode *n, struct nlattr **tb,
 560                          struct nlattr *est, bool ovr)
 561 {
 562         int err;
 563         struct tcf_exts e;
 564
 565         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 566         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 567         if (err < 0)
 568                 return err;
 569
 570         err = -EINVAL;
 571         if (tb[TCA_U32_LINK]) {
 572                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 573                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 574
 575                 if (TC_U32_KEY(handle))
 576                         goto errout;
 577
 578                 if (handle) {
 579                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 580
 581                         if (ht_down == NULL)
 582                                 goto errout;
 583                         ht_down->refcnt++;
 584                 }
 585
 586                 ht_old = rtnl_dereference(n->ht_down);
 587                 rcu_assign_pointer(n->ht_down, ht_down);
 588
 589                 if (ht_old)
 590                         ht_old->refcnt--;
 591         }
 592         if (tb[TCA_U32_CLASSID]) {
 593                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 594                 tcf_bind_filter(tp, &n->res, base);
 595         }
 596
 597 #ifdef CONFIG_NET_CLS_IND
 598         if (tb[TCA_U32_INDEV]) {
 599                 int ret;
 600                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 601                 if (ret < 0)
 602                         goto errout;
 603                 n->ifindex = ret;
 604         }
 605 #endif
 606         tcf_exts_change(tp, &n->exts, &e);
 607
 608         return 0;
 609 errout:
 610         tcf_exts_destroy(&e);
 611         return err;
 612 }
 613
 614 static void u32_replace_knode(struct tcf_proto *tp,
 615                               struct tc_u_common *tp_c,
 616                               struct tc_u_knode *n)
 617 {
 618         struct tc_u_knode __rcu **ins;
 619         struct tc_u_knode *pins;
 620         struct tc_u_hnode *ht;
 621
 622         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 623                 ht = rtnl_dereference(tp->root);
 624         else
 625                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 626
 627         ins = &ht->ht[TC_U32_HASH(n->handle)];
 628
 629         /* The node must always exist for it to be replaced if this is not the
 630          * case then something went very wrong elsewhere.
 631          */
 632         for (pins = rtnl_dereference(*ins); ;
 633              ins = &pins->next, pins = rtnl_dereference(*ins))
 634                 if (pins->handle == n->handle)
 635                         break;
 636
 637         RCU_INIT_POINTER(n->next, pins->next);
 638         rcu_assign_pointer(*ins, n);
 639 }
 640
 641 static struct tc_u_knode *u32_init_knode(struct tcf_proto *tp,
 642                                          struct tc_u_knode *n)
 643 {
 644         struct tc_u_knode *new;
 645         struct tc_u32_sel *s = &n->sel;
 646
 647         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 648                       GFP_KERNEL);
 649
 650         if (!new)
 651                 return NULL;
 652
 653         RCU_INIT_POINTER(new->next, n->next);
 654         new->handle = n->handle;
 655         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 656
 657 #ifdef CONFIG_NET_CLS_IND
 658         new->ifindex = n->ifindex;
 659 #endif
 660         new->fshift = n->fshift;
 661         new->res = n->res;
 662         RCU_INIT_POINTER(new->ht_down, n->ht_down);
 663
 664         /* bump reference count as long as we hold pointer to structure */
 665         if (new->ht_down)
 666                 new->ht_down->refcnt++;
 667
 668 #ifdef CONFIG_CLS_U32_PERF
 669         /* Statistics may be incremented by readers during update
 670          * so we must keep them in tact. When the node is later destroyed
 671          * a special destroy call must be made to not free the pf memory.
 672          */
 673         new->pf = n->pf;
 674 #endif
 675
 676 #ifdef CONFIG_CLS_U32_MARK
 677         new->val = n->val;
 678         new->mask = n->mask;
 679         /* Similarly success statistics must be moved as pointers */
 680         new->pcpu_success = n->pcpu_success;
 681 #endif
 682         new->tp = tp;
 683         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 684
 685         tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE);
 686
 687         return new;
 688 }
 689
 690 static int u32_change(struct net *net, struct sk_buff *in_skb,
 691                       struct tcf_proto *tp, unsigned long base, u32 handle,
 692                       struct nlattr **tca,
 693                       unsigned long *arg, bool ovr)
 694 {
 695         struct tc_u_common *tp_c = tp->data;
 696         struct tc_u_hnode *ht;
 697         struct tc_u_knode *n;
 698         struct tc_u32_sel *s;
 699         struct nlattr *opt = tca[TCA_OPTIONS];
 700         struct nlattr *tb[TCA_U32_MAX + 1];
 701         u32 htid;
 702         int err;
 703 #ifdef CONFIG_CLS_U32_PERF
 704         size_t size;
 705 #endif
 706
 707         if (opt == NULL)
 708                 return handle ? -EINVAL : 0;
 709
 710         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 711         if (err < 0)
 712                 return err;
 713
 714         n = (struct tc_u_knode *)*arg;
 715         if (n) {
 716                 struct tc_u_knode *new;
 717
 718                 if (TC_U32_KEY(n->handle) == 0)
 719                         return -EINVAL;
 720
 721                 new = u32_init_knode(tp, n);
 722                 if (!new)
 723                         return -ENOMEM;
 724
 725                 err = u32_set_parms(net, tp, base,
 726                                     rtnl_dereference(n->ht_up), new, tb,
 727                                     tca[TCA_RATE], ovr);
 728
 729                 if (err) {
 730                         u32_destroy_key(tp, new, false);
 731                         return err;
 732                 }
 733
 734                 u32_replace_knode(tp, tp_c, new);
 735                 tcf_unbind_filter(tp, &n->res);
 736                 call_rcu(&n->rcu, u32_delete_key_rcu);
 737                 return 0;
 738         }
 739
 740         if (tb[TCA_U32_DIVISOR]) {
 741                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 742
 743                 if (--divisor > 0x100)
 744                         return -EINVAL;
 745                 if (TC_U32_KEY(handle))
 746                         return -EINVAL;
 747                 if (handle == 0) {
 748                         handle = gen_new_htid(tp->data);
 749                         if (handle == 0)
 750                                 return -ENOMEM;
 751                 }
 752                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 753                 if (ht == NULL)
 754                         return -ENOBUFS;
 755                 ht->tp_c = tp_c;
 756                 ht->refcnt = 1;
 757                 ht->divisor = divisor;
 758                 ht->handle = handle;
 759                 ht->prio = tp->prio;
 760                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 761                 rcu_assign_pointer(tp_c->hlist, ht);
 762                 *arg = (unsigned long)ht;
 763                 return 0;
 764         }
 765
 766         if (tb[TCA_U32_HASH]) {
 767                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 768                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 769                         ht = rtnl_dereference(tp->root);
 770                         htid = ht->handle;
 771                 } else {
 772                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 773                         if (ht == NULL)
 774                                 return -EINVAL;
 775                 }
 776         } else {
 777                 ht = rtnl_dereference(tp->root);
 778                 htid = ht->handle;
 779         }
 780
 781         if (ht->divisor < TC_U32_HASH(htid))
 782                 return -EINVAL;
 783
 784         if (handle) {
 785                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 786                         return -EINVAL;
 787                 handle = htid | TC_U32_NODE(handle);
 788         } else
 789                 handle = gen_new_kid(ht, htid);
 790
 791         if (tb[TCA_U32_SEL] == NULL)
 792                 return -EINVAL;
 793
 794         s = nla_data(tb[TCA_U32_SEL]);
 795
 796         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 797         if (n == NULL)
 798                 return -ENOBUFS;
 799
 800 #ifdef CONFIG_CLS_U32_PERF
 801         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 802         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 803         if (!n->pf) {
 804                 kfree(n);
 805                 return -ENOBUFS;
 806         }
 807 #endif
 808
 809         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 810         RCU_INIT_POINTER(n->ht_up, ht);
 811         n->handle = handle;
 812         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 813         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 814         n->tp = tp;
 815
 816 #ifdef CONFIG_CLS_U32_MARK
 817         n->pcpu_success = alloc_percpu(u32);
 818         if (!n->pcpu_success) {
 819                 err = -ENOMEM;
 820                 goto errout;
 821         }
 822
 823         if (tb[TCA_U32_MARK]) {
 824                 struct tc_u32_mark *mark;
 825
 826                 mark = nla_data(tb[TCA_U32_MARK]);
 827                 n->val = mark->val;
 828                 n->mask = mark->mask;
 829         }
 830 #endif
 831
 832         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
 833         if (err == 0) {
 834                 struct tc_u_knode __rcu **ins;
 835                 struct tc_u_knode *pins;
 836
 837                 ins = &ht->ht[TC_U32_HASH(handle)];
 838                 for (pins = rtnl_dereference(*ins); pins;
 839                      ins = &pins->next, pins = rtnl_dereference(*ins))
 840                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
 841                                 break;
 842
 843                 RCU_INIT_POINTER(n->next, pins);
 844                 rcu_assign_pointer(*ins, n);
 845
 846                 *arg = (unsigned long)n;
 847                 return 0;
 848         }
 849
 850 #ifdef CONFIG_CLS_U32_MARK
 851         free_percpu(n->pcpu_success);
 852 errout:
 853 #endif
 854
 855 #ifdef CONFIG_CLS_U32_PERF
 856         free_percpu(n->pf);
 857 #endif
 858         kfree(n);
 859         return err;
 860 }
 861
 862 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 863 {
 864         struct tc_u_common *tp_c = tp->data;
 865         struct tc_u_hnode *ht;
 866         struct tc_u_knode *n;
 867         unsigned int h;
 868
 869         if (arg->stop)
 870                 return;
 871
 872         for (ht = rtnl_dereference(tp_c->hlist);
 873              ht;
 874              ht = rtnl_dereference(ht->next)) {
 875                 if (ht->prio != tp->prio)
 876                         continue;
 877                 if (arg->count >= arg->skip) {
 878                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 879                                 arg->stop = 1;
 880                                 return;
 881                         }
 882                 }
 883                 arg->count++;
 884                 for (h = 0; h <= ht->divisor; h++) {
 885                         for (n = rtnl_dereference(ht->ht[h]);
 886                              n;
 887                              n = rtnl_dereference(n->next)) {
 888                                 if (arg->count < arg->skip) {
 889                                         arg->count++;
 890                                         continue;
 891                                 }
 892                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 893                                         arg->stop = 1;
 894                                         return;
 895                                 }
 896                                 arg->count++;
 897                         }
 898                 }
 899         }
 900 }
 901
 902 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 903                      struct sk_buff *skb, struct tcmsg *t)
 904 {
 905         struct tc_u_knode *n = (struct tc_u_knode *)fh;
 906         struct tc_u_hnode *ht_up, *ht_down;
 907         struct nlattr *nest;
 908
 909         if (n == NULL)
 910                 return skb->len;
 911
 912         t->tcm_handle = n->handle;
 913
 914         nest = nla_nest_start(skb, TCA_OPTIONS);
 915         if (nest == NULL)
 916                 goto nla_put_failure;
 917
 918         if (TC_U32_KEY(n->handle) == 0) {
 919                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
 920                 u32 divisor = ht->divisor + 1;
 921
 922                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
 923                         goto nla_put_failure;
 924         } else {
 925 #ifdef CONFIG_CLS_U32_PERF
 926                 struct tc_u32_pcnt *gpf;
 927                 int cpu;
 928 #endif
 929
 930                 if (nla_put(skb, TCA_U32_SEL,
 931                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 932                             &n->sel))
 933                         goto nla_put_failure;
 934
 935                 ht_up = rtnl_dereference(n->ht_up);
 936                 if (ht_up) {
 937                         u32 htid = n->handle & 0xFFFFF000;
 938                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
 939                                 goto nla_put_failure;
 940                 }
 941                 if (n->res.classid &&
 942                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
 943                         goto nla_put_failure;
 944
 945                 ht_down = rtnl_dereference(n->ht_down);
 946                 if (ht_down &&
 947                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
 948                         goto nla_put_failure;
 949
 950 #ifdef CONFIG_CLS_U32_MARK
 951                 if ((n->val || n->mask)) {
 952                         struct tc_u32_mark mark = {.val = n->val,
 953                                                    .mask = n->mask,
 954                                                    .success = 0};
 955                         int cpum;
 956
 957                         for_each_possible_cpu(cpum) {
 958                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
 959
 960                                 mark.success += cnt;
 961                         }
 962
 963                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
 964                                 goto nla_put_failure;
 965                 }
 966 #endif
 967
 968                 if (tcf_exts_dump(skb, &n->exts) < 0)
 969                         goto nla_put_failure;
 970
 971 #ifdef CONFIG_NET_CLS_IND
 972                 if (n->ifindex) {
 973                         struct net_device *dev;
 974                         dev = __dev_get_by_index(net, n->ifindex);
 975                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
 976                                 goto nla_put_failure;
 977                 }
 978 #endif
 979 #ifdef CONFIG_CLS_U32_PERF
 980                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
 981                               n->sel.nkeys * sizeof(u64),
 982                               GFP_KERNEL);
 983                 if (!gpf)
 984                         goto nla_put_failure;
 985
 986                 for_each_possible_cpu(cpu) {
 987                         int i;
 988                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
 989
 990                         gpf->rcnt += pf->rcnt;
 991                         gpf->rhit += pf->rhit;
 992                         for (i = 0; i < n->sel.nkeys; i++)
 993                                 gpf->kcnts[i] += pf->kcnts[i];
 994                 }
 995
 996                 if (nla_put(skb, TCA_U32_PCNT,
 997                             sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(u64),
 998                             gpf)) {
 999                         kfree(gpf);
1000                         goto nla_put_failure;
1001                 }
1002                 kfree(gpf);
1003 #endif
1004         }
1005
1006         nla_nest_end(skb, nest);
1007
1008         if (TC_U32_KEY(n->handle))
1009                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1010                         goto nla_put_failure;
1011         return skb->len;
1012
1013 nla_put_failure:
1014         nla_nest_cancel(skb, nest);
1015         return -1;
1016 }
1017
1018 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1019         .kind           =       "u32",
1020         .classify       =       u32_classify,
1021         .init           =       u32_init,
1022         .destroy        =       u32_destroy,
1023         .get            =       u32_get,
1024         .put            =       u32_put,
1025         .change         =       u32_change,
1026         .delete         =       u32_delete,
1027         .walk           =       u32_walk,
1028         .dump           =       u32_dump,
1029         .owner          =       THIS_MODULE,
1030 };
1031
1032 static int __init init_u32(void)
1033 {
1034         pr_info("u32 classifier\n");
1035 #ifdef CONFIG_CLS_U32_PERF
1036         pr_info("    Performance counters on\n");
1037 #endif
1038 #ifdef CONFIG_NET_CLS_IND
1039         pr_info("    input device check on\n");
1040 #endif
1041 #ifdef CONFIG_NET_CLS_ACT
1042         pr_info("    Actions configured\n");
1043 #endif
1044         return register_tcf_proto_ops(&cls_u32_ops);
1045 }
1046
1047 static void __exit exit_u32(void)
1048 {
1049         unregister_tcf_proto_ops(&cls_u32_ops);
1050 }
1051
1052 module_init(init_u32)
1053 module_exit(exit_u32)
1054 MODULE_LICENSE("GPL");