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 <linux/netdevice.h>
  44 #include <linux/hash.h>
  45 #include <net/netlink.h>
  46 #include <net/act_api.h>
  47 #include <net/pkt_cls.h>
  48 #include <linux/netdevice.h>
  49
  50 struct tc_u_knode {
  51         struct tc_u_knode __rcu *next;
  52         u32                     handle;
  53         struct tc_u_hnode __rcu *ht_up;
  54         struct tcf_exts         exts;
  55 #ifdef CONFIG_NET_CLS_IND
  56         int                     ifindex;
  57 #endif
  58         u8                      fshift;
  59         struct tcf_result       res;
  60         struct tc_u_hnode __rcu *ht_down;
  61 #ifdef CONFIG_CLS_U32_PERF
  62         struct tc_u32_pcnt __percpu *pf;
  63 #endif
  64         u32                     flags;
  65 #ifdef CONFIG_CLS_U32_MARK
  66         u32                     val;
  67         u32                     mask;
  68         u32 __percpu            *pcpu_success;
  69 #endif
  70         struct tcf_proto        *tp;
  71         union {
  72                 struct work_struct      work;
  73                 struct rcu_head         rcu;
  74         };
  75         /* The 'sel' field MUST be the last field in structure to allow for
  76          * tc_u32_keys allocated at end of structure.
  77          */
  78         struct tc_u32_sel       sel;
  79 };
  80
  81 struct tc_u_hnode {
  82         struct tc_u_hnode __rcu *next;
  83         u32                     handle;
  84         u32                     prio;
  85         struct tc_u_common      *tp_c;
  86         int                     refcnt;
  87         unsigned int            divisor;
  88         struct rcu_head         rcu;
  89         /* The 'ht' field MUST be the last field in structure to allow for
  90          * more entries allocated at end of structure.
  91          */
  92         struct tc_u_knode __rcu *ht[1];
  93 };
  94
  95 struct tc_u_common {
  96         struct tc_u_hnode __rcu *hlist;
  97         struct Qdisc            *q;
  98         int                     refcnt;
  99         u32                     hgenerator;
 100         struct hlist_node       hnode;
 101         struct rcu_head         rcu;
 102 };
 103
 104 static inline unsigned int u32_hash_fold(__be32 key,
 105                                          const struct tc_u32_sel *sel,
 106                                          u8 fshift)
 107 {
 108         unsigned int h = ntohl(key & sel->hmask) >> fshift;
 109
 110         return h;
 111 }
 112
 113 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 114                         struct tcf_result *res)
 115 {
 116         struct {
 117                 struct tc_u_knode *knode;
 118                 unsigned int      off;
 119         } stack[TC_U32_MAXDEPTH];
 120
 121         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 122         unsigned int off = skb_network_offset(skb);
 123         struct tc_u_knode *n;
 124         int sdepth = 0;
 125         int off2 = 0;
 126         int sel = 0;
 127 #ifdef CONFIG_CLS_U32_PERF
 128         int j;
 129 #endif
 130         int i, r;
 131
 132 next_ht:
 133         n = rcu_dereference_bh(ht->ht[sel]);
 134
 135 next_knode:
 136         if (n) {
 137                 struct tc_u32_key *key = n->sel.keys;
 138
 139 #ifdef CONFIG_CLS_U32_PERF
 140                 __this_cpu_inc(n->pf->rcnt);
 141                 j = 0;
 142 #endif
 143
 144                 if (tc_skip_sw(n->flags)) {
 145                         n = rcu_dereference_bh(n->next);
 146                         goto next_knode;
 147                 }
 148
 149 #ifdef CONFIG_CLS_U32_MARK
 150                 if ((skb->mark & n->mask) != n->val) {
 151                         n = rcu_dereference_bh(n->next);
 152                         goto next_knode;
 153                 } else {
 154                         __this_cpu_inc(*n->pcpu_success);
 155                 }
 156 #endif
 157
 158                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 159                         int toff = off + key->off + (off2 & key->offmask);
 160                         __be32 *data, hdata;
 161
 162                         if (skb_headroom(skb) + toff > INT_MAX)
 163                                 goto out;
 164
 165                         data = skb_header_pointer(skb, toff, 4, &hdata);
 166                         if (!data)
 167                                 goto out;
 168                         if ((*data ^ key->val) & key->mask) {
 169                                 n = rcu_dereference_bh(n->next);
 170                                 goto next_knode;
 171                         }
 172 #ifdef CONFIG_CLS_U32_PERF
 173                         __this_cpu_inc(n->pf->kcnts[j]);
 174                         j++;
 175 #endif
 176                 }
 177
 178                 ht = rcu_dereference_bh(n->ht_down);
 179                 if (!ht) {
 180 check_terminal:
 181                         if (n->sel.flags & TC_U32_TERMINAL) {
 182
 183                                 *res = n->res;
 184 #ifdef CONFIG_NET_CLS_IND
 185                                 if (!tcf_match_indev(skb, n->ifindex)) {
 186                                         n = rcu_dereference_bh(n->next);
 187                                         goto next_knode;
 188                                 }
 189 #endif
 190 #ifdef CONFIG_CLS_U32_PERF
 191                                 __this_cpu_inc(n->pf->rhit);
 192 #endif
 193                                 r = tcf_exts_exec(skb, &n->exts, res);
 194                                 if (r < 0) {
 195                                         n = rcu_dereference_bh(n->next);
 196                                         goto next_knode;
 197                                 }
 198
 199                                 return r;
 200                         }
 201                         n = rcu_dereference_bh(n->next);
 202                         goto next_knode;
 203                 }
 204
 205                 /* PUSH */
 206                 if (sdepth >= TC_U32_MAXDEPTH)
 207                         goto deadloop;
 208                 stack[sdepth].knode = n;
 209                 stack[sdepth].off = off;
 210                 sdepth++;
 211
 212                 ht = rcu_dereference_bh(n->ht_down);
 213                 sel = 0;
 214                 if (ht->divisor) {
 215                         __be32 *data, hdata;
 216
 217                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 218                                                   &hdata);
 219                         if (!data)
 220                                 goto out;
 221                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 222                                                           n->fshift);
 223                 }
 224                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 225                         goto next_ht;
 226
 227                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 228                         off2 = n->sel.off + 3;
 229                         if (n->sel.flags & TC_U32_VAROFFSET) {
 230                                 __be16 *data, hdata;
 231
 232                                 data = skb_header_pointer(skb,
 233                                                           off + n->sel.offoff,
 234                                                           2, &hdata);
 235                                 if (!data)
 236                                         goto out;
 237                                 off2 += ntohs(n->sel.offmask & *data) >>
 238                                         n->sel.offshift;
 239                         }
 240                         off2 &= ~3;
 241                 }
 242                 if (n->sel.flags & TC_U32_EAT) {
 243                         off += off2;
 244                         off2 = 0;
 245                 }
 246
 247                 if (off < skb->len)
 248                         goto next_ht;
 249         }
 250
 251         /* POP */
 252         if (sdepth--) {
 253                 n = stack[sdepth].knode;
 254                 ht = rcu_dereference_bh(n->ht_up);
 255                 off = stack[sdepth].off;
 256                 goto check_terminal;
 257         }
 258 out:
 259         return -1;
 260
 261 deadloop:
 262         net_warn_ratelimited("cls_u32: dead loop\n");
 263         return -1;
 264 }
 265
 266 static struct tc_u_hnode *u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 267 {
 268         struct tc_u_hnode *ht;
 269
 270         for (ht = rtnl_dereference(tp_c->hlist);
 271              ht;
 272              ht = rtnl_dereference(ht->next))
 273                 if (ht->handle == handle)
 274                         break;
 275
 276         return ht;
 277 }
 278
 279 static struct tc_u_knode *u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 280 {
 281         unsigned int sel;
 282         struct tc_u_knode *n = NULL;
 283
 284         sel = TC_U32_HASH(handle);
 285         if (sel > ht->divisor)
 286                 goto out;
 287
 288         for (n = rtnl_dereference(ht->ht[sel]);
 289              n;
 290              n = rtnl_dereference(n->next))
 291                 if (n->handle == handle)
 292                         break;
 293 out:
 294         return n;
 295 }
 296
 297
 298 static void *u32_get(struct tcf_proto *tp, u32 handle)
 299 {
 300         struct tc_u_hnode *ht;
 301         struct tc_u_common *tp_c = tp->data;
 302
 303         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 304                 ht = rtnl_dereference(tp->root);
 305         else
 306                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 307
 308         if (!ht)
 309                 return NULL;
 310
 311         if (TC_U32_KEY(handle) == 0)
 312                 return ht;
 313
 314         return u32_lookup_key(ht, handle);
 315 }
 316
 317 static u32 gen_new_htid(struct tc_u_common *tp_c)
 318 {
 319         int i = 0x800;
 320
 321         /* hgenerator only used inside rtnl lock it is safe to increment
 322          * without read _copy_ update semantics
 323          */
 324         do {
 325                 if (++tp_c->hgenerator == 0x7FF)
 326                         tp_c->hgenerator = 1;
 327         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 328
 329         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 330 }
 331
 332 static struct hlist_head *tc_u_common_hash;
 333
 334 #define U32_HASH_SHIFT 10
 335 #define U32_HASH_SIZE (1 << U32_HASH_SHIFT)
 336
 337 static unsigned int tc_u_hash(const struct tcf_proto *tp)
 338 {
 339         struct net_device *dev = tp->q->dev_queue->dev;
 340         u32 qhandle = tp->q->handle;
 341         int ifindex = dev->ifindex;
 342
 343         return hash_64((u64)ifindex << 32 | qhandle, U32_HASH_SHIFT);
 344 }
 345
 346 static struct tc_u_common *tc_u_common_find(const struct tcf_proto *tp)
 347 {
 348         struct tc_u_common *tc;
 349         unsigned int h;
 350
 351         h = tc_u_hash(tp);
 352         hlist_for_each_entry(tc, &tc_u_common_hash[h], hnode) {
 353                 if (tc->q == tp->q)
 354                         return tc;
 355         }
 356         return NULL;
 357 }
 358
 359 static int u32_init(struct tcf_proto *tp)
 360 {
 361         struct tc_u_hnode *root_ht;
 362         struct tc_u_common *tp_c;
 363         unsigned int h;
 364
 365         tp_c = tc_u_common_find(tp);
 366
 367         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 368         if (root_ht == NULL)
 369                 return -ENOBUFS;
 370
 371         root_ht->refcnt++;
 372         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 373         root_ht->prio = tp->prio;
 374
 375         if (tp_c == NULL) {
 376                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 377                 if (tp_c == NULL) {
 378                         kfree(root_ht);
 379                         return -ENOBUFS;
 380                 }
 381                 tp_c->q = tp->q;
 382                 INIT_HLIST_NODE(&tp_c->hnode);
 383
 384                 h = tc_u_hash(tp);
 385                 hlist_add_head(&tp_c->hnode, &tc_u_common_hash[h]);
 386         }
 387
 388         tp_c->refcnt++;
 389         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 390         rcu_assign_pointer(tp_c->hlist, root_ht);
 391         root_ht->tp_c = tp_c;
 392
 393         rcu_assign_pointer(tp->root, root_ht);
 394         tp->data = tp_c;
 395         return 0;
 396 }
 397
 398 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n,
 399                            bool free_pf)
 400 {
 401         tcf_exts_destroy(&n->exts);
 402         tcf_exts_put_net(&n->exts);
 403         if (n->ht_down)
 404                 n->ht_down->refcnt--;
 405 #ifdef CONFIG_CLS_U32_PERF
 406         if (free_pf)
 407                 free_percpu(n->pf);
 408 #endif
 409 #ifdef CONFIG_CLS_U32_MARK
 410         if (free_pf)
 411                 free_percpu(n->pcpu_success);
 412 #endif
 413         kfree(n);
 414         return 0;
 415 }
 416
 417 /* u32_delete_key_rcu should be called when free'ing a copied
 418  * version of a tc_u_knode obtained from u32_init_knode(). When
 419  * copies are obtained from u32_init_knode() the statistics are
 420  * shared between the old and new copies to allow readers to
 421  * continue to update the statistics during the copy. To support
 422  * this the u32_delete_key_rcu variant does not free the percpu
 423  * statistics.
 424  */
 425 static void u32_delete_key_work(struct work_struct *work)
 426 {
 427         struct tc_u_knode *key = container_of(work, struct tc_u_knode, work);
 428
 429         rtnl_lock();
 430         u32_destroy_key(key->tp, key, false);
 431         rtnl_unlock();
 432 }
 433
 434 static void u32_delete_key_rcu(struct rcu_head *rcu)
 435 {
 436         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 437
 438         INIT_WORK(&key->work, u32_delete_key_work);
 439         tcf_queue_work(&key->work);
 440 }
 441
 442 /* u32_delete_key_freepf_rcu is the rcu callback variant
 443  * that free's the entire structure including the statistics
 444  * percpu variables. Only use this if the key is not a copy
 445  * returned by u32_init_knode(). See u32_delete_key_rcu()
 446  * for the variant that should be used with keys return from
 447  * u32_init_knode()
 448  */
 449 static void u32_delete_key_freepf_work(struct work_struct *work)
 450 {
 451         struct tc_u_knode *key = container_of(work, struct tc_u_knode, work);
 452
 453         rtnl_lock();
 454         u32_destroy_key(key->tp, key, true);
 455         rtnl_unlock();
 456 }
 457
 458 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 459 {
 460         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 461
 462         INIT_WORK(&key->work, u32_delete_key_freepf_work);
 463         tcf_queue_work(&key->work);
 464 }
 465
 466 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 467 {
 468         struct tc_u_knode __rcu **kp;
 469         struct tc_u_knode *pkp;
 470         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 471
 472         if (ht) {
 473                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 474                 for (pkp = rtnl_dereference(*kp); pkp;
 475                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 476                         if (pkp == key) {
 477                                 RCU_INIT_POINTER(*kp, key->next);
 478
 479                                 tcf_unbind_filter(tp, &key->res);
 480                                 tcf_exts_get_net(&key->exts);
 481                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 482                                 return 0;
 483                         }
 484                 }
 485         }
 486         WARN_ON(1);
 487         return 0;
 488 }
 489
 490 static void u32_remove_hw_knode(struct tcf_proto *tp, u32 handle)
 491 {
 492         struct net_device *dev = tp->q->dev_queue->dev;
 493         struct tc_cls_u32_offload cls_u32 = {};
 494
 495         if (!tc_should_offload(dev, 0))
 496                 return;
 497
 498         tc_cls_common_offload_init(&cls_u32.common, tp);
 499         cls_u32.command = TC_CLSU32_DELETE_KNODE;
 500         cls_u32.knode.handle = handle;
 501
 502         dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, &cls_u32);
 503 }
 504
 505 static int u32_replace_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h,
 506                                 u32 flags)
 507 {
 508         struct net_device *dev = tp->q->dev_queue->dev;
 509         struct tc_cls_u32_offload cls_u32 = {};
 510         int err;
 511
 512         if (!tc_should_offload(dev, flags))
 513                 return tc_skip_sw(flags) ? -EINVAL : 0;
 514
 515         tc_cls_common_offload_init(&cls_u32.common, tp);
 516         cls_u32.command = TC_CLSU32_NEW_HNODE;
 517         cls_u32.hnode.divisor = h->divisor;
 518         cls_u32.hnode.handle = h->handle;
 519         cls_u32.hnode.prio = h->prio;
 520
 521         err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, &cls_u32);
 522         if (tc_skip_sw(flags))
 523                 return err;
 524
 525         return 0;
 526 }
 527
 528 static void u32_clear_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h)
 529 {
 530         struct net_device *dev = tp->q->dev_queue->dev;
 531         struct tc_cls_u32_offload cls_u32 = {};
 532
 533         if (!tc_should_offload(dev, 0))
 534                 return;
 535
 536         tc_cls_common_offload_init(&cls_u32.common, tp);
 537         cls_u32.command = TC_CLSU32_DELETE_HNODE;
 538         cls_u32.hnode.divisor = h->divisor;
 539         cls_u32.hnode.handle = h->handle;
 540         cls_u32.hnode.prio = h->prio;
 541
 542         dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, &cls_u32);
 543 }
 544
 545 static int u32_replace_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 546                                 u32 flags)
 547 {
 548         struct net_device *dev = tp->q->dev_queue->dev;
 549         struct tc_cls_u32_offload cls_u32 = {};
 550         int err;
 551
 552         if (!tc_should_offload(dev, flags))
 553                 return tc_skip_sw(flags) ? -EINVAL : 0;
 554
 555         tc_cls_common_offload_init(&cls_u32.common, tp);
 556         cls_u32.command = TC_CLSU32_REPLACE_KNODE;
 557         cls_u32.knode.handle = n->handle;
 558         cls_u32.knode.fshift = n->fshift;
 559 #ifdef CONFIG_CLS_U32_MARK
 560         cls_u32.knode.val = n->val;
 561         cls_u32.knode.mask = n->mask;
 562 #else
 563         cls_u32.knode.val = 0;
 564         cls_u32.knode.mask = 0;
 565 #endif
 566         cls_u32.knode.sel = &n->sel;
 567         cls_u32.knode.exts = &n->exts;
 568         if (n->ht_down)
 569                 cls_u32.knode.link_handle = n->ht_down->handle;
 570
 571         err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, &cls_u32);
 572
 573         if (!err)
 574                 n->flags |= TCA_CLS_FLAGS_IN_HW;
 575
 576         if (tc_skip_sw(flags))
 577                 return err;
 578
 579         return 0;
 580 }
 581
 582 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 583 {
 584         struct tc_u_knode *n;
 585         unsigned int h;
 586
 587         for (h = 0; h <= ht->divisor; h++) {
 588                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 589                         RCU_INIT_POINTER(ht->ht[h],
 590                                          rtnl_dereference(n->next));
 591                         tcf_unbind_filter(tp, &n->res);
 592                         u32_remove_hw_knode(tp, n->handle);
 593                         if (tcf_exts_get_net(&n->exts))
 594                                 call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 595                         else
 596                                 u32_destroy_key(n->tp, n, true);
 597                 }
 598         }
 599 }
 600
 601 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 602 {
 603         struct tc_u_common *tp_c = tp->data;
 604         struct tc_u_hnode __rcu **hn;
 605         struct tc_u_hnode *phn;
 606
 607         WARN_ON(ht->refcnt);
 608
 609         u32_clear_hnode(tp, ht);
 610
 611         hn = &tp_c->hlist;
 612         for (phn = rtnl_dereference(*hn);
 613              phn;
 614              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 615                 if (phn == ht) {
 616                         u32_clear_hw_hnode(tp, ht);
 617                         RCU_INIT_POINTER(*hn, ht->next);
 618                         kfree_rcu(ht, rcu);
 619                         return 0;
 620                 }
 621         }
 622
 623         return -ENOENT;
 624 }
 625
 626 static bool ht_empty(struct tc_u_hnode *ht)
 627 {
 628         unsigned int h;
 629
 630         for (h = 0; h <= ht->divisor; h++)
 631                 if (rcu_access_pointer(ht->ht[h]))
 632                         return false;
 633
 634         return true;
 635 }
 636
 637 static void u32_destroy(struct tcf_proto *tp)
 638 {
 639         struct tc_u_common *tp_c = tp->data;
 640         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 641
 642         WARN_ON(root_ht == NULL);
 643
 644         if (root_ht && --root_ht->refcnt == 0)
 645                 u32_destroy_hnode(tp, root_ht);
 646
 647         if (--tp_c->refcnt == 0) {
 648                 struct tc_u_hnode *ht;
 649
 650                 hlist_del(&tp_c->hnode);
 651
 652                 for (ht = rtnl_dereference(tp_c->hlist);
 653                      ht;
 654                      ht = rtnl_dereference(ht->next)) {
 655                         ht->refcnt--;
 656                         u32_clear_hnode(tp, ht);
 657                 }
 658
 659                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 660                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 661                         kfree_rcu(ht, rcu);
 662                 }
 663
 664                 kfree(tp_c);
 665         }
 666
 667         tp->data = NULL;
 668 }
 669
 670 static int u32_delete(struct tcf_proto *tp, void *arg, bool *last)
 671 {
 672         struct tc_u_hnode *ht = arg;
 673         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 674         struct tc_u_common *tp_c = tp->data;
 675         int ret = 0;
 676
 677         if (ht == NULL)
 678                 goto out;
 679
 680         if (TC_U32_KEY(ht->handle)) {
 681                 u32_remove_hw_knode(tp, ht->handle);
 682                 ret = u32_delete_key(tp, (struct tc_u_knode *)ht);
 683                 goto out;
 684         }
 685
 686         if (root_ht == ht)
 687                 return -EINVAL;
 688
 689         if (ht->refcnt == 1) {
 690                 ht->refcnt--;
 691                 u32_destroy_hnode(tp, ht);
 692         } else {
 693                 return -EBUSY;
 694         }
 695
 696 out:
 697         *last = true;
 698         if (root_ht) {
 699                 if (root_ht->refcnt > 1) {
 700                         *last = false;
 701                         goto ret;
 702                 }
 703                 if (root_ht->refcnt == 1) {
 704                         if (!ht_empty(root_ht)) {
 705                                 *last = false;
 706                                 goto ret;
 707                         }
 708                 }
 709         }
 710
 711         if (tp_c->refcnt > 1) {
 712                 *last = false;
 713                 goto ret;
 714         }
 715
 716         if (tp_c->refcnt == 1) {
 717                 struct tc_u_hnode *ht;
 718
 719                 for (ht = rtnl_dereference(tp_c->hlist);
 720                      ht;
 721                      ht = rtnl_dereference(ht->next))
 722                         if (!ht_empty(ht)) {
 723                                 *last = false;
 724                                 break;
 725                         }
 726         }
 727
 728 ret:
 729         return ret;
 730 }
 731
 732 #define NR_U32_NODE (1<<12)
 733 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 734 {
 735         struct tc_u_knode *n;
 736         unsigned long i;
 737         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 738                                         GFP_KERNEL);
 739         if (!bitmap)
 740                 return handle | 0xFFF;
 741
 742         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 743              n;
 744              n = rtnl_dereference(n->next))
 745                 set_bit(TC_U32_NODE(n->handle), bitmap);
 746
 747         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 748         if (i >= NR_U32_NODE)
 749                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 750
 751         kfree(bitmap);
 752         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 753 }
 754
 755 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 756         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 757         [TCA_U32_HASH]          = { .type = NLA_U32 },
 758         [TCA_U32_LINK]          = { .type = NLA_U32 },
 759         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 760         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 761         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 762         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 763         [TCA_U32_FLAGS]         = { .type = NLA_U32 },
 764 };
 765
 766 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 767                          unsigned long base, struct tc_u_hnode *ht,
 768                          struct tc_u_knode *n, struct nlattr **tb,
 769                          struct nlattr *est, bool ovr)
 770 {
 771         int err;
 772
 773         err = tcf_exts_validate(net, tp, tb, est, &n->exts, ovr);
 774         if (err < 0)
 775                 return err;
 776
 777         if (tb[TCA_U32_LINK]) {
 778                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 779                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 780
 781                 if (TC_U32_KEY(handle))
 782                         return -EINVAL;
 783
 784                 if (handle) {
 785                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 786
 787                         if (ht_down == NULL)
 788                                 return -EINVAL;
 789                         ht_down->refcnt++;
 790                 }
 791
 792                 ht_old = rtnl_dereference(n->ht_down);
 793                 rcu_assign_pointer(n->ht_down, ht_down);
 794
 795                 if (ht_old)
 796                         ht_old->refcnt--;
 797         }
 798         if (tb[TCA_U32_CLASSID]) {
 799                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 800                 tcf_bind_filter(tp, &n->res, base);
 801         }
 802
 803 #ifdef CONFIG_NET_CLS_IND
 804         if (tb[TCA_U32_INDEV]) {
 805                 int ret;
 806                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 807                 if (ret < 0)
 808                         return -EINVAL;
 809                 n->ifindex = ret;
 810         }
 811 #endif
 812         return 0;
 813 }
 814
 815 static void u32_replace_knode(struct tcf_proto *tp, struct tc_u_common *tp_c,
 816                               struct tc_u_knode *n)
 817 {
 818         struct tc_u_knode __rcu **ins;
 819         struct tc_u_knode *pins;
 820         struct tc_u_hnode *ht;
 821
 822         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 823                 ht = rtnl_dereference(tp->root);
 824         else
 825                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 826
 827         ins = &ht->ht[TC_U32_HASH(n->handle)];
 828
 829         /* The node must always exist for it to be replaced if this is not the
 830          * case then something went very wrong elsewhere.
 831          */
 832         for (pins = rtnl_dereference(*ins); ;
 833              ins = &pins->next, pins = rtnl_dereference(*ins))
 834                 if (pins->handle == n->handle)
 835                         break;
 836
 837         RCU_INIT_POINTER(n->next, pins->next);
 838         rcu_assign_pointer(*ins, n);
 839 }
 840
 841 static struct tc_u_knode *u32_init_knode(struct tcf_proto *tp,
 842                                          struct tc_u_knode *n)
 843 {
 844         struct tc_u_knode *new;
 845         struct tc_u32_sel *s = &n->sel;
 846
 847         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 848                       GFP_KERNEL);
 849
 850         if (!new)
 851                 return NULL;
 852
 853         RCU_INIT_POINTER(new->next, n->next);
 854         new->handle = n->handle;
 855         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 856
 857 #ifdef CONFIG_NET_CLS_IND
 858         new->ifindex = n->ifindex;
 859 #endif
 860         new->fshift = n->fshift;
 861         new->res = n->res;
 862         new->flags = n->flags;
 863         RCU_INIT_POINTER(new->ht_down, n->ht_down);
 864
 865         /* bump reference count as long as we hold pointer to structure */
 866         if (new->ht_down)
 867                 new->ht_down->refcnt++;
 868
 869 #ifdef CONFIG_CLS_U32_PERF
 870         /* Statistics may be incremented by readers during update
 871          * so we must keep them in tact. When the node is later destroyed
 872          * a special destroy call must be made to not free the pf memory.
 873          */
 874         new->pf = n->pf;
 875 #endif
 876
 877 #ifdef CONFIG_CLS_U32_MARK
 878         new->val = n->val;
 879         new->mask = n->mask;
 880         /* Similarly success statistics must be moved as pointers */
 881         new->pcpu_success = n->pcpu_success;
 882 #endif
 883         new->tp = tp;
 884         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 885
 886         if (tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE)) {
 887                 kfree(new);
 888                 return NULL;
 889         }
 890
 891         return new;
 892 }
 893
 894 static int u32_change(struct net *net, struct sk_buff *in_skb,
 895                       struct tcf_proto *tp, unsigned long base, u32 handle,
 896                       struct nlattr **tca, void **arg, bool ovr)
 897 {
 898         struct tc_u_common *tp_c = tp->data;
 899         struct tc_u_hnode *ht;
 900         struct tc_u_knode *n;
 901         struct tc_u32_sel *s;
 902         struct nlattr *opt = tca[TCA_OPTIONS];
 903         struct nlattr *tb[TCA_U32_MAX + 1];
 904         u32 htid, flags = 0;
 905         int err;
 906 #ifdef CONFIG_CLS_U32_PERF
 907         size_t size;
 908 #endif
 909
 910         if (opt == NULL)
 911                 return handle ? -EINVAL : 0;
 912
 913         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy, NULL);
 914         if (err < 0)
 915                 return err;
 916
 917         if (tb[TCA_U32_FLAGS]) {
 918                 flags = nla_get_u32(tb[TCA_U32_FLAGS]);
 919                 if (!tc_flags_valid(flags))
 920                         return -EINVAL;
 921         }
 922
 923         n = *arg;
 924         if (n) {
 925                 struct tc_u_knode *new;
 926
 927                 if (TC_U32_KEY(n->handle) == 0)
 928                         return -EINVAL;
 929
 930                 if (n->flags != flags)
 931                         return -EINVAL;
 932
 933                 new = u32_init_knode(tp, n);
 934                 if (!new)
 935                         return -ENOMEM;
 936
 937                 err = u32_set_parms(net, tp, base,
 938                                     rtnl_dereference(n->ht_up), new, tb,
 939                                     tca[TCA_RATE], ovr);
 940
 941                 if (err) {
 942                         u32_destroy_key(tp, new, false);
 943                         return err;
 944                 }
 945
 946                 err = u32_replace_hw_knode(tp, new, flags);
 947                 if (err) {
 948                         u32_destroy_key(tp, new, false);
 949                         return err;
 950                 }
 951
 952                 if (!tc_in_hw(new->flags))
 953                         new->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
 954
 955                 u32_replace_knode(tp, tp_c, new);
 956                 tcf_unbind_filter(tp, &n->res);
 957                 tcf_exts_get_net(&n->exts);
 958                 call_rcu(&n->rcu, u32_delete_key_rcu);
 959                 return 0;
 960         }
 961
 962         if (tb[TCA_U32_DIVISOR]) {
 963                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 964
 965                 if (--divisor > 0x100)
 966                         return -EINVAL;
 967                 if (TC_U32_KEY(handle))
 968                         return -EINVAL;
 969                 if (handle == 0) {
 970                         handle = gen_new_htid(tp->data);
 971                         if (handle == 0)
 972                                 return -ENOMEM;
 973                 }
 974                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 975                 if (ht == NULL)
 976                         return -ENOBUFS;
 977                 ht->tp_c = tp_c;
 978                 ht->refcnt = 1;
 979                 ht->divisor = divisor;
 980                 ht->handle = handle;
 981                 ht->prio = tp->prio;
 982
 983                 err = u32_replace_hw_hnode(tp, ht, flags);
 984                 if (err) {
 985                         kfree(ht);
 986                         return err;
 987                 }
 988
 989                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 990                 rcu_assign_pointer(tp_c->hlist, ht);
 991                 *arg = ht;
 992
 993                 return 0;
 994         }
 995
 996         if (tb[TCA_U32_HASH]) {
 997                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 998                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 999                         ht = rtnl_dereference(tp->root);
1000                         htid = ht->handle;
1001                 } else {
1002                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
1003                         if (ht == NULL)
1004                                 return -EINVAL;
1005                 }
1006         } else {
1007                 ht = rtnl_dereference(tp->root);
1008                 htid = ht->handle;
1009         }
1010
1011         if (ht->divisor < TC_U32_HASH(htid))
1012                 return -EINVAL;
1013
1014         if (handle) {
1015                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
1016                         return -EINVAL;
1017                 handle = htid | TC_U32_NODE(handle);
1018         } else
1019                 handle = gen_new_kid(ht, htid);
1020
1021         if (tb[TCA_U32_SEL] == NULL)
1022                 return -EINVAL;
1023
1024         s = nla_data(tb[TCA_U32_SEL]);
1025
1026         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
1027         if (n == NULL)
1028                 return -ENOBUFS;
1029
1030 #ifdef CONFIG_CLS_U32_PERF
1031         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
1032         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
1033         if (!n->pf) {
1034                 kfree(n);
1035                 return -ENOBUFS;
1036         }
1037 #endif
1038
1039         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
1040         RCU_INIT_POINTER(n->ht_up, ht);
1041         n->handle = handle;
1042         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
1043         n->flags = flags;
1044         n->tp = tp;
1045
1046         err = tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
1047         if (err < 0)
1048                 goto errout;
1049
1050 #ifdef CONFIG_CLS_U32_MARK
1051         n->pcpu_success = alloc_percpu(u32);
1052         if (!n->pcpu_success) {
1053                 err = -ENOMEM;
1054                 goto errout;
1055         }
1056
1057         if (tb[TCA_U32_MARK]) {
1058                 struct tc_u32_mark *mark;
1059
1060                 mark = nla_data(tb[TCA_U32_MARK]);
1061                 n->val = mark->val;
1062                 n->mask = mark->mask;
1063         }
1064 #endif
1065
1066         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
1067         if (err == 0) {
1068                 struct tc_u_knode __rcu **ins;
1069                 struct tc_u_knode *pins;
1070
1071                 err = u32_replace_hw_knode(tp, n, flags);
1072                 if (err)
1073                         goto errhw;
1074
1075                 if (!tc_in_hw(n->flags))
1076                         n->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1077
1078                 ins = &ht->ht[TC_U32_HASH(handle)];
1079                 for (pins = rtnl_dereference(*ins); pins;
1080                      ins = &pins->next, pins = rtnl_dereference(*ins))
1081                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
1082                                 break;
1083
1084                 RCU_INIT_POINTER(n->next, pins);
1085                 rcu_assign_pointer(*ins, n);
1086                 *arg = n;
1087                 return 0;
1088         }
1089
1090 errhw:
1091 #ifdef CONFIG_CLS_U32_MARK
1092         free_percpu(n->pcpu_success);
1093 #endif
1094
1095 errout:
1096         tcf_exts_destroy(&n->exts);
1097 #ifdef CONFIG_CLS_U32_PERF
1098         free_percpu(n->pf);
1099 #endif
1100         kfree(n);
1101         return err;
1102 }
1103
1104 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1105 {
1106         struct tc_u_common *tp_c = tp->data;
1107         struct tc_u_hnode *ht;
1108         struct tc_u_knode *n;
1109         unsigned int h;
1110
1111         if (arg->stop)
1112                 return;
1113
1114         for (ht = rtnl_dereference(tp_c->hlist);
1115              ht;
1116              ht = rtnl_dereference(ht->next)) {
1117                 if (ht->prio != tp->prio)
1118                         continue;
1119                 if (arg->count >= arg->skip) {
1120                         if (arg->fn(tp, ht, arg) < 0) {
1121                                 arg->stop = 1;
1122                                 return;
1123                         }
1124                 }
1125                 arg->count++;
1126                 for (h = 0; h <= ht->divisor; h++) {
1127                         for (n = rtnl_dereference(ht->ht[h]);
1128                              n;
1129                              n = rtnl_dereference(n->next)) {
1130                                 if (arg->count < arg->skip) {
1131                                         arg->count++;
1132                                         continue;
1133                                 }
1134                                 if (arg->fn(tp, n, arg) < 0) {
1135                                         arg->stop = 1;
1136                                         return;
1137                                 }
1138                                 arg->count++;
1139                         }
1140                 }
1141         }
1142 }
1143
1144 static void u32_bind_class(void *fh, u32 classid, unsigned long cl)
1145 {
1146         struct tc_u_knode *n = fh;
1147
1148         if (n && n->res.classid == classid)
1149                 n->res.class = cl;
1150 }
1151
1152 static int u32_dump(struct net *net, struct tcf_proto *tp, void *fh,
1153                     struct sk_buff *skb, struct tcmsg *t)
1154 {
1155         struct tc_u_knode *n = fh;
1156         struct tc_u_hnode *ht_up, *ht_down;
1157         struct nlattr *nest;
1158
1159         if (n == NULL)
1160                 return skb->len;
1161
1162         t->tcm_handle = n->handle;
1163
1164         nest = nla_nest_start(skb, TCA_OPTIONS);
1165         if (nest == NULL)
1166                 goto nla_put_failure;
1167
1168         if (TC_U32_KEY(n->handle) == 0) {
1169                 struct tc_u_hnode *ht = fh;
1170                 u32 divisor = ht->divisor + 1;
1171
1172                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
1173                         goto nla_put_failure;
1174         } else {
1175 #ifdef CONFIG_CLS_U32_PERF
1176                 struct tc_u32_pcnt *gpf;
1177                 int cpu;
1178 #endif
1179
1180                 if (nla_put(skb, TCA_U32_SEL,
1181                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
1182                             &n->sel))
1183                         goto nla_put_failure;
1184
1185                 ht_up = rtnl_dereference(n->ht_up);
1186                 if (ht_up) {
1187                         u32 htid = n->handle & 0xFFFFF000;
1188                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
1189                                 goto nla_put_failure;
1190                 }
1191                 if (n->res.classid &&
1192                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
1193                         goto nla_put_failure;
1194
1195                 ht_down = rtnl_dereference(n->ht_down);
1196                 if (ht_down &&
1197                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
1198                         goto nla_put_failure;
1199
1200                 if (n->flags && nla_put_u32(skb, TCA_U32_FLAGS, n->flags))
1201                         goto nla_put_failure;
1202
1203 #ifdef CONFIG_CLS_U32_MARK
1204                 if ((n->val || n->mask)) {
1205                         struct tc_u32_mark mark = {.val = n->val,
1206                                                    .mask = n->mask,
1207                                                    .success = 0};
1208                         int cpum;
1209
1210                         for_each_possible_cpu(cpum) {
1211                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
1212
1213                                 mark.success += cnt;
1214                         }
1215
1216                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
1217                                 goto nla_put_failure;
1218                 }
1219 #endif
1220
1221                 if (tcf_exts_dump(skb, &n->exts) < 0)
1222                         goto nla_put_failure;
1223
1224 #ifdef CONFIG_NET_CLS_IND
1225                 if (n->ifindex) {
1226                         struct net_device *dev;
1227                         dev = __dev_get_by_index(net, n->ifindex);
1228                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
1229                                 goto nla_put_failure;
1230                 }
1231 #endif
1232 #ifdef CONFIG_CLS_U32_PERF
1233                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1234                               n->sel.nkeys * sizeof(u64),
1235                               GFP_KERNEL);
1236                 if (!gpf)
1237                         goto nla_put_failure;
1238
1239                 for_each_possible_cpu(cpu) {
1240                         int i;
1241                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1242
1243                         gpf->rcnt += pf->rcnt;
1244                         gpf->rhit += pf->rhit;
1245                         for (i = 0; i < n->sel.nkeys; i++)
1246                                 gpf->kcnts[i] += pf->kcnts[i];
1247                 }
1248
1249                 if (nla_put_64bit(skb, TCA_U32_PCNT,
1250                                   sizeof(struct tc_u32_pcnt) +
1251                                   n->sel.nkeys * sizeof(u64),
1252                                   gpf, TCA_U32_PAD)) {
1253                         kfree(gpf);
1254                         goto nla_put_failure;
1255                 }
1256                 kfree(gpf);
1257 #endif
1258         }
1259
1260         nla_nest_end(skb, nest);
1261
1262         if (TC_U32_KEY(n->handle))
1263                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1264                         goto nla_put_failure;
1265         return skb->len;
1266
1267 nla_put_failure:
1268         nla_nest_cancel(skb, nest);
1269         return -1;
1270 }
1271
1272 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1273         .kind           =       "u32",
1274         .classify       =       u32_classify,
1275         .init           =       u32_init,
1276         .destroy        =       u32_destroy,
1277         .get            =       u32_get,
1278         .change         =       u32_change,
1279         .delete         =       u32_delete,
1280         .walk           =       u32_walk,
1281         .dump           =       u32_dump,
1282         .bind_class     =       u32_bind_class,
1283         .owner          =       THIS_MODULE,
1284 };
1285
1286 static int __init init_u32(void)
1287 {
1288         int i, ret;
1289
1290         pr_info("u32 classifier\n");
1291 #ifdef CONFIG_CLS_U32_PERF
1292         pr_info("    Performance counters on\n");
1293 #endif
1294 #ifdef CONFIG_NET_CLS_IND
1295         pr_info("    input device check on\n");
1296 #endif
1297 #ifdef CONFIG_NET_CLS_ACT
1298         pr_info("    Actions configured\n");
1299 #endif
1300         tc_u_common_hash = kvmalloc_array(U32_HASH_SIZE,
1301                                           sizeof(struct hlist_head),
1302                                           GFP_KERNEL);
1303         if (!tc_u_common_hash)
1304                 return -ENOMEM;
1305
1306         for (i = 0; i < U32_HASH_SIZE; i++)
1307                 INIT_HLIST_HEAD(&tc_u_common_hash[i]);
1308
1309         ret = register_tcf_proto_ops(&cls_u32_ops);
1310         if (ret)
1311                 kvfree(tc_u_common_hash);
1312         return ret;
1313 }
1314
1315 static void __exit exit_u32(void)
1316 {
1317         unregister_tcf_proto_ops(&cls_u32_ops);
1318         kvfree(tc_u_common_hash);
1319 }
1320
1321 module_init(init_u32)
1322 module_exit(exit_u32)
1323 MODULE_LICENSE("GPL");