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/idr.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         unsigned int            in_hw_count;
  66 #ifdef CONFIG_CLS_U32_MARK
  67         u32                     val;
  68         u32                     mask;
  69         u32 __percpu            *pcpu_success;
  70 #endif
  71         struct rcu_work         rwork;
  72         /* The 'sel' field MUST be the last field in structure to allow for
  73          * tc_u32_keys allocated at end of structure.
  74          */
  75         struct tc_u32_sel       sel;
  76 };
  77
  78 struct tc_u_hnode {
  79         struct tc_u_hnode __rcu *next;
  80         u32                     handle;
  81         u32                     prio;
  82         int                     refcnt;
  83         unsigned int            divisor;
  84         struct idr              handle_idr;
  85         bool                    is_root;
  86         struct rcu_head         rcu;
  87         u32                     flags;
  88         /* The 'ht' field MUST be the last field in structure to allow for
  89          * more entries allocated at end of structure.
  90          */
  91         struct tc_u_knode __rcu *ht[1];
  92 };
  93
  94 struct tc_u_common {
  95         struct tc_u_hnode __rcu *hlist;
  96         void                    *ptr;
  97         int                     refcnt;
  98         struct idr              handle_idr;
  99         struct hlist_node       hnode;
 100         long                    knodes;
 101 };
 102
 103 static inline unsigned int u32_hash_fold(__be32 key,
 104                                          const struct tc_u32_sel *sel,
 105                                          u8 fshift)
 106 {
 107         unsigned int h = ntohl(key & sel->hmask) >> fshift;
 108
 109         return h;
 110 }
 111
 112 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 113                         struct tcf_result *res)
 114 {
 115         struct {
 116                 struct tc_u_knode *knode;
 117                 unsigned int      off;
 118         } stack[TC_U32_MAXDEPTH];
 119
 120         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 121         unsigned int off = skb_network_offset(skb);
 122         struct tc_u_knode *n;
 123         int sdepth = 0;
 124         int off2 = 0;
 125         int sel = 0;
 126 #ifdef CONFIG_CLS_U32_PERF
 127         int j;
 128 #endif
 129         int i, r;
 130
 131 next_ht:
 132         n = rcu_dereference_bh(ht->ht[sel]);
 133
 134 next_knode:
 135         if (n) {
 136                 struct tc_u32_key *key = n->sel.keys;
 137
 138 #ifdef CONFIG_CLS_U32_PERF
 139                 __this_cpu_inc(n->pf->rcnt);
 140                 j = 0;
 141 #endif
 142
 143                 if (tc_skip_sw(n->flags)) {
 144                         n = rcu_dereference_bh(n->next);
 145                         goto next_knode;
 146                 }
 147
 148 #ifdef CONFIG_CLS_U32_MARK
 149                 if ((skb->mark & n->mask) != n->val) {
 150                         n = rcu_dereference_bh(n->next);
 151                         goto next_knode;
 152                 } else {
 153                         __this_cpu_inc(*n->pcpu_success);
 154                 }
 155 #endif
 156
 157                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 158                         int toff = off + key->off + (off2 & key->offmask);
 159                         __be32 *data, hdata;
 160
 161                         if (skb_headroom(skb) + toff > INT_MAX)
 162                                 goto out;
 163
 164                         data = skb_header_pointer(skb, toff, 4, &hdata);
 165                         if (!data)
 166                                 goto out;
 167                         if ((*data ^ key->val) & key->mask) {
 168                                 n = rcu_dereference_bh(n->next);
 169                                 goto next_knode;
 170                         }
 171 #ifdef CONFIG_CLS_U32_PERF
 172                         __this_cpu_inc(n->pf->kcnts[j]);
 173                         j++;
 174 #endif
 175                 }
 176
 177                 ht = rcu_dereference_bh(n->ht_down);
 178                 if (!ht) {
 179 check_terminal:
 180                         if (n->sel.flags & TC_U32_TERMINAL) {
 181
 182                                 *res = n->res;
 183 #ifdef CONFIG_NET_CLS_IND
 184                                 if (!tcf_match_indev(skb, n->ifindex)) {
 185                                         n = rcu_dereference_bh(n->next);
 186                                         goto next_knode;
 187                                 }
 188 #endif
 189 #ifdef CONFIG_CLS_U32_PERF
 190                                 __this_cpu_inc(n->pf->rhit);
 191 #endif
 192                                 r = tcf_exts_exec(skb, &n->exts, res);
 193                                 if (r < 0) {
 194                                         n = rcu_dereference_bh(n->next);
 195                                         goto next_knode;
 196                                 }
 197
 198                                 return r;
 199                         }
 200                         n = rcu_dereference_bh(n->next);
 201                         goto next_knode;
 202                 }
 203
 204                 /* PUSH */
 205                 if (sdepth >= TC_U32_MAXDEPTH)
 206                         goto deadloop;
 207                 stack[sdepth].knode = n;
 208                 stack[sdepth].off = off;
 209                 sdepth++;
 210
 211                 ht = rcu_dereference_bh(n->ht_down);
 212                 sel = 0;
 213                 if (ht->divisor) {
 214                         __be32 *data, hdata;
 215
 216                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 217                                                   &hdata);
 218                         if (!data)
 219                                 goto out;
 220                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 221                                                           n->fshift);
 222                 }
 223                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 224                         goto next_ht;
 225
 226                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 227                         off2 = n->sel.off + 3;
 228                         if (n->sel.flags & TC_U32_VAROFFSET) {
 229                                 __be16 *data, hdata;
 230
 231                                 data = skb_header_pointer(skb,
 232                                                           off + n->sel.offoff,
 233                                                           2, &hdata);
 234                                 if (!data)
 235                                         goto out;
 236                                 off2 += ntohs(n->sel.offmask & *data) >>
 237                                         n->sel.offshift;
 238                         }
 239                         off2 &= ~3;
 240                 }
 241                 if (n->sel.flags & TC_U32_EAT) {
 242                         off += off2;
 243                         off2 = 0;
 244                 }
 245
 246                 if (off < skb->len)
 247                         goto next_ht;
 248         }
 249
 250         /* POP */
 251         if (sdepth--) {
 252                 n = stack[sdepth].knode;
 253                 ht = rcu_dereference_bh(n->ht_up);
 254                 off = stack[sdepth].off;
 255                 goto check_terminal;
 256         }
 257 out:
 258         return -1;
 259
 260 deadloop:
 261         net_warn_ratelimited("cls_u32: dead loop\n");
 262         return -1;
 263 }
 264
 265 static struct tc_u_hnode *u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 266 {
 267         struct tc_u_hnode *ht;
 268
 269         for (ht = rtnl_dereference(tp_c->hlist);
 270              ht;
 271              ht = rtnl_dereference(ht->next))
 272                 if (ht->handle == handle)
 273                         break;
 274
 275         return ht;
 276 }
 277
 278 static struct tc_u_knode *u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 279 {
 280         unsigned int sel;
 281         struct tc_u_knode *n = NULL;
 282
 283         sel = TC_U32_HASH(handle);
 284         if (sel > ht->divisor)
 285                 goto out;
 286
 287         for (n = rtnl_dereference(ht->ht[sel]);
 288              n;
 289              n = rtnl_dereference(n->next))
 290                 if (n->handle == handle)
 291                         break;
 292 out:
 293         return n;
 294 }
 295
 296
 297 static void *u32_get(struct tcf_proto *tp, u32 handle)
 298 {
 299         struct tc_u_hnode *ht;
 300         struct tc_u_common *tp_c = tp->data;
 301
 302         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 303                 ht = rtnl_dereference(tp->root);
 304         else
 305                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 306
 307         if (!ht)
 308                 return NULL;
 309
 310         if (TC_U32_KEY(handle) == 0)
 311                 return ht;
 312
 313         return u32_lookup_key(ht, handle);
 314 }
 315
 316 /* Protected by rtnl lock */
 317 static u32 gen_new_htid(struct tc_u_common *tp_c, struct tc_u_hnode *ptr)
 318 {
 319         int id = idr_alloc_cyclic(&tp_c->handle_idr, ptr, 1, 0x7FF, GFP_KERNEL);
 320         if (id < 0)
 321                 return 0;
 322         return (id | 0x800U) << 20;
 323 }
 324
 325 static struct hlist_head *tc_u_common_hash;
 326
 327 #define U32_HASH_SHIFT 10
 328 #define U32_HASH_SIZE (1 << U32_HASH_SHIFT)
 329
 330 static void *tc_u_common_ptr(const struct tcf_proto *tp)
 331 {
 332         struct tcf_block *block = tp->chain->block;
 333
 334         /* The block sharing is currently supported only
 335          * for classless qdiscs. In that case we use block
 336          * for tc_u_common identification. In case the
 337          * block is not shared, block->q is a valid pointer
 338          * and we can use that. That works for classful qdiscs.
 339          */
 340         if (tcf_block_shared(block))
 341                 return block;
 342         else
 343                 return block->q;
 344 }
 345
 346 static struct hlist_head *tc_u_hash(void *key)
 347 {
 348         return tc_u_common_hash + hash_ptr(key, U32_HASH_SHIFT);
 349 }
 350
 351 static struct tc_u_common *tc_u_common_find(void *key)
 352 {
 353         struct tc_u_common *tc;
 354         hlist_for_each_entry(tc, tc_u_hash(key), hnode) {
 355                 if (tc->ptr == key)
 356                         return tc;
 357         }
 358         return NULL;
 359 }
 360
 361 static int u32_init(struct tcf_proto *tp)
 362 {
 363         struct tc_u_hnode *root_ht;
 364         void *key = tc_u_common_ptr(tp);
 365         struct tc_u_common *tp_c = tc_u_common_find(key);
 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, root_ht) : 0x80000000;
 373         root_ht->prio = tp->prio;
 374         root_ht->is_root = true;
 375         idr_init(&root_ht->handle_idr);
 376
 377         if (tp_c == NULL) {
 378                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 379                 if (tp_c == NULL) {
 380                         kfree(root_ht);
 381                         return -ENOBUFS;
 382                 }
 383                 tp_c->ptr = key;
 384                 INIT_HLIST_NODE(&tp_c->hnode);
 385                 idr_init(&tp_c->handle_idr);
 386
 387                 hlist_add_head(&tp_c->hnode, tc_u_hash(key));
 388         }
 389
 390         tp_c->refcnt++;
 391         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 392         rcu_assign_pointer(tp_c->hlist, root_ht);
 393
 394         root_ht->refcnt++;
 395         rcu_assign_pointer(tp->root, root_ht);
 396         tp->data = tp_c;
 397         return 0;
 398 }
 399
 400 static int u32_destroy_key(struct tc_u_knode *n, bool free_pf)
 401 {
 402         struct tc_u_hnode *ht = rtnl_dereference(n->ht_down);
 403
 404         tcf_exts_destroy(&n->exts);
 405         tcf_exts_put_net(&n->exts);
 406         if (ht && --ht->refcnt == 0)
 407                 kfree(ht);
 408 #ifdef CONFIG_CLS_U32_PERF
 409         if (free_pf)
 410                 free_percpu(n->pf);
 411 #endif
 412 #ifdef CONFIG_CLS_U32_MARK
 413         if (free_pf)
 414                 free_percpu(n->pcpu_success);
 415 #endif
 416         kfree(n);
 417         return 0;
 418 }
 419
 420 /* u32_delete_key_rcu should be called when free'ing a copied
 421  * version of a tc_u_knode obtained from u32_init_knode(). When
 422  * copies are obtained from u32_init_knode() the statistics are
 423  * shared between the old and new copies to allow readers to
 424  * continue to update the statistics during the copy. To support
 425  * this the u32_delete_key_rcu variant does not free the percpu
 426  * statistics.
 427  */
 428 static void u32_delete_key_work(struct work_struct *work)
 429 {
 430         struct tc_u_knode *key = container_of(to_rcu_work(work),
 431                                               struct tc_u_knode,
 432                                               rwork);
 433         rtnl_lock();
 434         u32_destroy_key(key, false);
 435         rtnl_unlock();
 436 }
 437
 438 /* u32_delete_key_freepf_rcu is the rcu callback variant
 439  * that free's the entire structure including the statistics
 440  * percpu variables. Only use this if the key is not a copy
 441  * returned by u32_init_knode(). See u32_delete_key_rcu()
 442  * for the variant that should be used with keys return from
 443  * u32_init_knode()
 444  */
 445 static void u32_delete_key_freepf_work(struct work_struct *work)
 446 {
 447         struct tc_u_knode *key = container_of(to_rcu_work(work),
 448                                               struct tc_u_knode,
 449                                               rwork);
 450         rtnl_lock();
 451         u32_destroy_key(key, true);
 452         rtnl_unlock();
 453 }
 454
 455 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 456 {
 457         struct tc_u_common *tp_c = tp->data;
 458         struct tc_u_knode __rcu **kp;
 459         struct tc_u_knode *pkp;
 460         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 461
 462         if (ht) {
 463                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 464                 for (pkp = rtnl_dereference(*kp); pkp;
 465                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 466                         if (pkp == key) {
 467                                 RCU_INIT_POINTER(*kp, key->next);
 468                                 tp_c->knodes--;
 469
 470                                 tcf_unbind_filter(tp, &key->res);
 471                                 idr_remove(&ht->handle_idr, key->handle);
 472                                 tcf_exts_get_net(&key->exts);
 473                                 tcf_queue_work(&key->rwork, u32_delete_key_freepf_work);
 474                                 return 0;
 475                         }
 476                 }
 477         }
 478         WARN_ON(1);
 479         return 0;
 480 }
 481
 482 static void u32_clear_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h,
 483                                struct netlink_ext_ack *extack)
 484 {
 485         struct tcf_block *block = tp->chain->block;
 486         struct tc_cls_u32_offload cls_u32 = {};
 487
 488         tc_cls_common_offload_init(&cls_u32.common, tp, h->flags, extack);
 489         cls_u32.command = TC_CLSU32_DELETE_HNODE;
 490         cls_u32.hnode.divisor = h->divisor;
 491         cls_u32.hnode.handle = h->handle;
 492         cls_u32.hnode.prio = h->prio;
 493
 494         tc_setup_cb_call(block, TC_SETUP_CLSU32, &cls_u32, false);
 495 }
 496
 497 static int u32_replace_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h,
 498                                 u32 flags, struct netlink_ext_ack *extack)
 499 {
 500         struct tcf_block *block = tp->chain->block;
 501         struct tc_cls_u32_offload cls_u32 = {};
 502         bool skip_sw = tc_skip_sw(flags);
 503         bool offloaded = false;
 504         int err;
 505
 506         tc_cls_common_offload_init(&cls_u32.common, tp, flags, extack);
 507         cls_u32.command = TC_CLSU32_NEW_HNODE;
 508         cls_u32.hnode.divisor = h->divisor;
 509         cls_u32.hnode.handle = h->handle;
 510         cls_u32.hnode.prio = h->prio;
 511
 512         err = tc_setup_cb_call(block, TC_SETUP_CLSU32, &cls_u32, skip_sw);
 513         if (err < 0) {
 514                 u32_clear_hw_hnode(tp, h, NULL);
 515                 return err;
 516         } else if (err > 0) {
 517                 offloaded = true;
 518         }
 519
 520         if (skip_sw && !offloaded)
 521                 return -EINVAL;
 522
 523         return 0;
 524 }
 525
 526 static void u32_remove_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 527                                 struct netlink_ext_ack *extack)
 528 {
 529         struct tcf_block *block = tp->chain->block;
 530         struct tc_cls_u32_offload cls_u32 = {};
 531
 532         tc_cls_common_offload_init(&cls_u32.common, tp, n->flags, extack);
 533         cls_u32.command = TC_CLSU32_DELETE_KNODE;
 534         cls_u32.knode.handle = n->handle;
 535
 536         tc_setup_cb_call(block, TC_SETUP_CLSU32, &cls_u32, false);
 537         tcf_block_offload_dec(block, &n->flags);
 538 }
 539
 540 static int u32_replace_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 541                                 u32 flags, struct netlink_ext_ack *extack)
 542 {
 543         struct tc_u_hnode *ht = rtnl_dereference(n->ht_down);
 544         struct tcf_block *block = tp->chain->block;
 545         struct tc_cls_u32_offload cls_u32 = {};
 546         bool skip_sw = tc_skip_sw(flags);
 547         int err;
 548
 549         tc_cls_common_offload_init(&cls_u32.common, tp, flags, extack);
 550         cls_u32.command = TC_CLSU32_REPLACE_KNODE;
 551         cls_u32.knode.handle = n->handle;
 552         cls_u32.knode.fshift = n->fshift;
 553 #ifdef CONFIG_CLS_U32_MARK
 554         cls_u32.knode.val = n->val;
 555         cls_u32.knode.mask = n->mask;
 556 #else
 557         cls_u32.knode.val = 0;
 558         cls_u32.knode.mask = 0;
 559 #endif
 560         cls_u32.knode.sel = &n->sel;
 561         cls_u32.knode.res = &n->res;
 562         cls_u32.knode.exts = &n->exts;
 563         if (n->ht_down)
 564                 cls_u32.knode.link_handle = ht->handle;
 565
 566         err = tc_setup_cb_call(block, TC_SETUP_CLSU32, &cls_u32, skip_sw);
 567         if (err < 0) {
 568                 u32_remove_hw_knode(tp, n, NULL);
 569                 return err;
 570         } else if (err > 0) {
 571                 n->in_hw_count = err;
 572                 tcf_block_offload_inc(block, &n->flags);
 573         }
 574
 575         if (skip_sw && !(n->flags & TCA_CLS_FLAGS_IN_HW))
 576                 return -EINVAL;
 577
 578         return 0;
 579 }
 580
 581 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht,
 582                             struct netlink_ext_ack *extack)
 583 {
 584         struct tc_u_common *tp_c = tp->data;
 585         struct tc_u_knode *n;
 586         unsigned int h;
 587
 588         for (h = 0; h <= ht->divisor; h++) {
 589                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 590                         RCU_INIT_POINTER(ht->ht[h],
 591                                          rtnl_dereference(n->next));
 592                         tp_c->knodes--;
 593                         tcf_unbind_filter(tp, &n->res);
 594                         u32_remove_hw_knode(tp, n, extack);
 595                         idr_remove(&ht->handle_idr, n->handle);
 596                         if (tcf_exts_get_net(&n->exts))
 597                                 tcf_queue_work(&n->rwork, u32_delete_key_freepf_work);
 598                         else
 599                                 u32_destroy_key(n, true);
 600                 }
 601         }
 602 }
 603
 604 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht,
 605                              struct netlink_ext_ack *extack)
 606 {
 607         struct tc_u_common *tp_c = tp->data;
 608         struct tc_u_hnode __rcu **hn;
 609         struct tc_u_hnode *phn;
 610
 611         WARN_ON(--ht->refcnt);
 612
 613         u32_clear_hnode(tp, ht, extack);
 614
 615         hn = &tp_c->hlist;
 616         for (phn = rtnl_dereference(*hn);
 617              phn;
 618              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 619                 if (phn == ht) {
 620                         u32_clear_hw_hnode(tp, ht, extack);
 621                         idr_destroy(&ht->handle_idr);
 622                         idr_remove(&tp_c->handle_idr, ht->handle);
 623                         RCU_INIT_POINTER(*hn, ht->next);
 624                         kfree_rcu(ht, rcu);
 625                         return 0;
 626                 }
 627         }
 628
 629         return -ENOENT;
 630 }
 631
 632 static void u32_destroy(struct tcf_proto *tp, bool rtnl_held,
 633                         struct netlink_ext_ack *extack)
 634 {
 635         struct tc_u_common *tp_c = tp->data;
 636         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 637
 638         WARN_ON(root_ht == NULL);
 639
 640         if (root_ht && --root_ht->refcnt == 1)
 641                 u32_destroy_hnode(tp, root_ht, extack);
 642
 643         if (--tp_c->refcnt == 0) {
 644                 struct tc_u_hnode *ht;
 645
 646                 hlist_del(&tp_c->hnode);
 647
 648                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 649                         u32_clear_hnode(tp, ht, extack);
 650                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 651
 652                         /* u32_destroy_key() will later free ht for us, if it's
 653                          * still referenced by some knode
 654                          */
 655                         if (--ht->refcnt == 0)
 656                                 kfree_rcu(ht, rcu);
 657                 }
 658
 659                 idr_destroy(&tp_c->handle_idr);
 660                 kfree(tp_c);
 661         }
 662
 663         tp->data = NULL;
 664 }
 665
 666 static int u32_delete(struct tcf_proto *tp, void *arg, bool *last,
 667                       bool rtnl_held, struct netlink_ext_ack *extack)
 668 {
 669         struct tc_u_hnode *ht = arg;
 670         struct tc_u_common *tp_c = tp->data;
 671         int ret = 0;
 672
 673         if (TC_U32_KEY(ht->handle)) {
 674                 u32_remove_hw_knode(tp, (struct tc_u_knode *)ht, extack);
 675                 ret = u32_delete_key(tp, (struct tc_u_knode *)ht);
 676                 goto out;
 677         }
 678
 679         if (ht->is_root) {
 680                 NL_SET_ERR_MSG_MOD(extack, "Not allowed to delete root node");
 681                 return -EINVAL;
 682         }
 683
 684         if (ht->refcnt == 1) {
 685                 u32_destroy_hnode(tp, ht, extack);
 686         } else {
 687                 NL_SET_ERR_MSG_MOD(extack, "Can not delete in-use filter");
 688                 return -EBUSY;
 689         }
 690
 691 out:
 692         *last = tp_c->refcnt == 1 && tp_c->knodes == 0;
 693         return ret;
 694 }
 695
 696 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 htid)
 697 {
 698         u32 index = htid | 0x800;
 699         u32 max = htid | 0xFFF;
 700
 701         if (idr_alloc_u32(&ht->handle_idr, NULL, &index, max, GFP_KERNEL)) {
 702                 index = htid + 1;
 703                 if (idr_alloc_u32(&ht->handle_idr, NULL, &index, max,
 704                                  GFP_KERNEL))
 705                         index = max;
 706         }
 707
 708         return index;
 709 }
 710
 711 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 712         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 713         [TCA_U32_HASH]          = { .type = NLA_U32 },
 714         [TCA_U32_LINK]          = { .type = NLA_U32 },
 715         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 716         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 717         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 718         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 719         [TCA_U32_FLAGS]         = { .type = NLA_U32 },
 720 };
 721
 722 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 723                          unsigned long base,
 724                          struct tc_u_knode *n, struct nlattr **tb,
 725                          struct nlattr *est, bool ovr,
 726                          struct netlink_ext_ack *extack)
 727 {
 728         int err;
 729
 730         err = tcf_exts_validate(net, tp, tb, est, &n->exts, ovr, true, extack);
 731         if (err < 0)
 732                 return err;
 733
 734         if (tb[TCA_U32_LINK]) {
 735                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 736                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 737
 738                 if (TC_U32_KEY(handle)) {
 739                         NL_SET_ERR_MSG_MOD(extack, "u32 Link handle must be a hash table");
 740                         return -EINVAL;
 741                 }
 742
 743                 if (handle) {
 744                         ht_down = u32_lookup_ht(tp->data, handle);
 745
 746                         if (!ht_down) {
 747                                 NL_SET_ERR_MSG_MOD(extack, "Link hash table not found");
 748                                 return -EINVAL;
 749                         }
 750                         if (ht_down->is_root) {
 751                                 NL_SET_ERR_MSG_MOD(extack, "Not linking to root node");
 752                                 return -EINVAL;
 753                         }
 754                         ht_down->refcnt++;
 755                 }
 756
 757                 ht_old = rtnl_dereference(n->ht_down);
 758                 rcu_assign_pointer(n->ht_down, ht_down);
 759
 760                 if (ht_old)
 761                         ht_old->refcnt--;
 762         }
 763         if (tb[TCA_U32_CLASSID]) {
 764                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 765                 tcf_bind_filter(tp, &n->res, base);
 766         }
 767
 768 #ifdef CONFIG_NET_CLS_IND
 769         if (tb[TCA_U32_INDEV]) {
 770                 int ret;
 771                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV], extack);
 772                 if (ret < 0)
 773                         return -EINVAL;
 774                 n->ifindex = ret;
 775         }
 776 #endif
 777         return 0;
 778 }
 779
 780 static void u32_replace_knode(struct tcf_proto *tp, struct tc_u_common *tp_c,
 781                               struct tc_u_knode *n)
 782 {
 783         struct tc_u_knode __rcu **ins;
 784         struct tc_u_knode *pins;
 785         struct tc_u_hnode *ht;
 786
 787         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 788                 ht = rtnl_dereference(tp->root);
 789         else
 790                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 791
 792         ins = &ht->ht[TC_U32_HASH(n->handle)];
 793
 794         /* The node must always exist for it to be replaced if this is not the
 795          * case then something went very wrong elsewhere.
 796          */
 797         for (pins = rtnl_dereference(*ins); ;
 798              ins = &pins->next, pins = rtnl_dereference(*ins))
 799                 if (pins->handle == n->handle)
 800                         break;
 801
 802         idr_replace(&ht->handle_idr, n, n->handle);
 803         RCU_INIT_POINTER(n->next, pins->next);
 804         rcu_assign_pointer(*ins, n);
 805 }
 806
 807 static struct tc_u_knode *u32_init_knode(struct net *net, struct tcf_proto *tp,
 808                                          struct tc_u_knode *n)
 809 {
 810         struct tc_u_hnode *ht = rtnl_dereference(n->ht_down);
 811         struct tc_u32_sel *s = &n->sel;
 812         struct tc_u_knode *new;
 813
 814         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 815                       GFP_KERNEL);
 816
 817         if (!new)
 818                 return NULL;
 819
 820         RCU_INIT_POINTER(new->next, n->next);
 821         new->handle = n->handle;
 822         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 823
 824 #ifdef CONFIG_NET_CLS_IND
 825         new->ifindex = n->ifindex;
 826 #endif
 827         new->fshift = n->fshift;
 828         new->res = n->res;
 829         new->flags = n->flags;
 830         RCU_INIT_POINTER(new->ht_down, ht);
 831
 832         /* bump reference count as long as we hold pointer to structure */
 833         if (ht)
 834                 ht->refcnt++;
 835
 836 #ifdef CONFIG_CLS_U32_PERF
 837         /* Statistics may be incremented by readers during update
 838          * so we must keep them in tact. When the node is later destroyed
 839          * a special destroy call must be made to not free the pf memory.
 840          */
 841         new->pf = n->pf;
 842 #endif
 843
 844 #ifdef CONFIG_CLS_U32_MARK
 845         new->val = n->val;
 846         new->mask = n->mask;
 847         /* Similarly success statistics must be moved as pointers */
 848         new->pcpu_success = n->pcpu_success;
 849 #endif
 850         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 851
 852         if (tcf_exts_init(&new->exts, net, TCA_U32_ACT, TCA_U32_POLICE)) {
 853                 kfree(new);
 854                 return NULL;
 855         }
 856
 857         return new;
 858 }
 859
 860 static int u32_change(struct net *net, struct sk_buff *in_skb,
 861                       struct tcf_proto *tp, unsigned long base, u32 handle,
 862                       struct nlattr **tca, void **arg, bool ovr, bool rtnl_held,
 863                       struct netlink_ext_ack *extack)
 864 {
 865         struct tc_u_common *tp_c = tp->data;
 866         struct tc_u_hnode *ht;
 867         struct tc_u_knode *n;
 868         struct tc_u32_sel *s;
 869         struct nlattr *opt = tca[TCA_OPTIONS];
 870         struct nlattr *tb[TCA_U32_MAX + 1];
 871         u32 htid, flags = 0;
 872         size_t sel_size;
 873         int err;
 874 #ifdef CONFIG_CLS_U32_PERF
 875         size_t size;
 876 #endif
 877
 878         if (!opt) {
 879                 if (handle) {
 880                         NL_SET_ERR_MSG_MOD(extack, "Filter handle requires options");
 881                         return -EINVAL;
 882                 } else {
 883                         return 0;
 884                 }
 885         }
 886
 887         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy, extack);
 888         if (err < 0)
 889                 return err;
 890
 891         if (tb[TCA_U32_FLAGS]) {
 892                 flags = nla_get_u32(tb[TCA_U32_FLAGS]);
 893                 if (!tc_flags_valid(flags)) {
 894                         NL_SET_ERR_MSG_MOD(extack, "Invalid filter flags");
 895                         return -EINVAL;
 896                 }
 897         }
 898
 899         n = *arg;
 900         if (n) {
 901                 struct tc_u_knode *new;
 902
 903                 if (TC_U32_KEY(n->handle) == 0) {
 904                         NL_SET_ERR_MSG_MOD(extack, "Key node id cannot be zero");
 905                         return -EINVAL;
 906                 }
 907
 908                 if ((n->flags ^ flags) &
 909                     ~(TCA_CLS_FLAGS_IN_HW | TCA_CLS_FLAGS_NOT_IN_HW)) {
 910                         NL_SET_ERR_MSG_MOD(extack, "Key node flags do not match passed flags");
 911                         return -EINVAL;
 912                 }
 913
 914                 new = u32_init_knode(net, tp, n);
 915                 if (!new)
 916                         return -ENOMEM;
 917
 918                 err = u32_set_parms(net, tp, base, new, tb,
 919                                     tca[TCA_RATE], ovr, extack);
 920
 921                 if (err) {
 922                         u32_destroy_key(new, false);
 923                         return err;
 924                 }
 925
 926                 err = u32_replace_hw_knode(tp, new, flags, extack);
 927                 if (err) {
 928                         u32_destroy_key(new, false);
 929                         return err;
 930                 }
 931
 932                 if (!tc_in_hw(new->flags))
 933                         new->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
 934
 935                 u32_replace_knode(tp, tp_c, new);
 936                 tcf_unbind_filter(tp, &n->res);
 937                 tcf_exts_get_net(&n->exts);
 938                 tcf_queue_work(&n->rwork, u32_delete_key_work);
 939                 return 0;
 940         }
 941
 942         if (tb[TCA_U32_DIVISOR]) {
 943                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 944
 945                 if (!is_power_of_2(divisor)) {
 946                         NL_SET_ERR_MSG_MOD(extack, "Divisor is not a power of 2");
 947                         return -EINVAL;
 948                 }
 949                 if (divisor-- > 0x100) {
 950                         NL_SET_ERR_MSG_MOD(extack, "Exceeded maximum 256 hash buckets");
 951                         return -EINVAL;
 952                 }
 953                 if (TC_U32_KEY(handle)) {
 954                         NL_SET_ERR_MSG_MOD(extack, "Divisor can only be used on a hash table");
 955                         return -EINVAL;
 956                 }
 957                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 958                 if (ht == NULL)
 959                         return -ENOBUFS;
 960                 if (handle == 0) {
 961                         handle = gen_new_htid(tp->data, ht);
 962                         if (handle == 0) {
 963                                 kfree(ht);
 964                                 return -ENOMEM;
 965                         }
 966                 } else {
 967                         err = idr_alloc_u32(&tp_c->handle_idr, ht, &handle,
 968                                             handle, GFP_KERNEL);
 969                         if (err) {
 970                                 kfree(ht);
 971                                 return err;
 972                         }
 973                 }
 974                 ht->refcnt = 1;
 975                 ht->divisor = divisor;
 976                 ht->handle = handle;
 977                 ht->prio = tp->prio;
 978                 idr_init(&ht->handle_idr);
 979                 ht->flags = flags;
 980
 981                 err = u32_replace_hw_hnode(tp, ht, flags, extack);
 982                 if (err) {
 983                         idr_remove(&tp_c->handle_idr, handle);
 984                         kfree(ht);
 985                         return err;
 986                 }
 987
 988                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 989                 rcu_assign_pointer(tp_c->hlist, ht);
 990                 *arg = ht;
 991
 992                 return 0;
 993         }
 994
 995         if (tb[TCA_U32_HASH]) {
 996                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 997                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 998                         ht = rtnl_dereference(tp->root);
 999                         htid = ht->handle;
1000                 } else {
1001                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
1002                         if (!ht) {
1003                                 NL_SET_ERR_MSG_MOD(extack, "Specified hash table not found");
1004                                 return -EINVAL;
1005                         }
1006                 }
1007         } else {
1008                 ht = rtnl_dereference(tp->root);
1009                 htid = ht->handle;
1010         }
1011
1012         if (ht->divisor < TC_U32_HASH(htid)) {
1013                 NL_SET_ERR_MSG_MOD(extack, "Specified hash table buckets exceed configured value");
1014                 return -EINVAL;
1015         }
1016
1017         if (handle) {
1018                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle ^ htid)) {
1019                         NL_SET_ERR_MSG_MOD(extack, "Handle specified hash table address mismatch");
1020                         return -EINVAL;
1021                 }
1022                 handle = htid | TC_U32_NODE(handle);
1023                 err = idr_alloc_u32(&ht->handle_idr, NULL, &handle, handle,
1024                                     GFP_KERNEL);
1025                 if (err)
1026                         return err;
1027         } else
1028                 handle = gen_new_kid(ht, htid);
1029
1030         if (tb[TCA_U32_SEL] == NULL) {
1031                 NL_SET_ERR_MSG_MOD(extack, "Selector not specified");
1032                 err = -EINVAL;
1033                 goto erridr;
1034         }
1035
1036         s = nla_data(tb[TCA_U32_SEL]);
1037         sel_size = struct_size(s, keys, s->nkeys);
1038         if (nla_len(tb[TCA_U32_SEL]) < sel_size) {
1039                 err = -EINVAL;
1040                 goto erridr;
1041         }
1042
1043         n = kzalloc(offsetof(typeof(*n), sel) + sel_size, GFP_KERNEL);
1044         if (n == NULL) {
1045                 err = -ENOBUFS;
1046                 goto erridr;
1047         }
1048
1049 #ifdef CONFIG_CLS_U32_PERF
1050         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
1051         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
1052         if (!n->pf) {
1053                 err = -ENOBUFS;
1054                 goto errfree;
1055         }
1056 #endif
1057
1058         memcpy(&n->sel, s, sel_size);
1059         RCU_INIT_POINTER(n->ht_up, ht);
1060         n->handle = handle;
1061         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
1062         n->flags = flags;
1063
1064         err = tcf_exts_init(&n->exts, net, TCA_U32_ACT, TCA_U32_POLICE);
1065         if (err < 0)
1066                 goto errout;
1067
1068 #ifdef CONFIG_CLS_U32_MARK
1069         n->pcpu_success = alloc_percpu(u32);
1070         if (!n->pcpu_success) {
1071                 err = -ENOMEM;
1072                 goto errout;
1073         }
1074
1075         if (tb[TCA_U32_MARK]) {
1076                 struct tc_u32_mark *mark;
1077
1078                 mark = nla_data(tb[TCA_U32_MARK]);
1079                 n->val = mark->val;
1080                 n->mask = mark->mask;
1081         }
1082 #endif
1083
1084         err = u32_set_parms(net, tp, base, n, tb, tca[TCA_RATE], ovr,
1085                             extack);
1086         if (err == 0) {
1087                 struct tc_u_knode __rcu **ins;
1088                 struct tc_u_knode *pins;
1089
1090                 err = u32_replace_hw_knode(tp, n, flags, extack);
1091                 if (err)
1092                         goto errhw;
1093
1094                 if (!tc_in_hw(n->flags))
1095                         n->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1096
1097                 ins = &ht->ht[TC_U32_HASH(handle)];
1098                 for (pins = rtnl_dereference(*ins); pins;
1099                      ins = &pins->next, pins = rtnl_dereference(*ins))
1100                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
1101                                 break;
1102
1103                 RCU_INIT_POINTER(n->next, pins);
1104                 rcu_assign_pointer(*ins, n);
1105                 tp_c->knodes++;
1106                 *arg = n;
1107                 return 0;
1108         }
1109
1110 errhw:
1111 #ifdef CONFIG_CLS_U32_MARK
1112         free_percpu(n->pcpu_success);
1113 #endif
1114
1115 errout:
1116         tcf_exts_destroy(&n->exts);
1117 #ifdef CONFIG_CLS_U32_PERF
1118 errfree:
1119         free_percpu(n->pf);
1120 #endif
1121         kfree(n);
1122 erridr:
1123         idr_remove(&ht->handle_idr, handle);
1124         return err;
1125 }
1126
1127 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1128                      bool rtnl_held)
1129 {
1130         struct tc_u_common *tp_c = tp->data;
1131         struct tc_u_hnode *ht;
1132         struct tc_u_knode *n;
1133         unsigned int h;
1134
1135         if (arg->stop)
1136                 return;
1137
1138         for (ht = rtnl_dereference(tp_c->hlist);
1139              ht;
1140              ht = rtnl_dereference(ht->next)) {
1141                 if (ht->prio != tp->prio)
1142                         continue;
1143                 if (arg->count >= arg->skip) {
1144                         if (arg->fn(tp, ht, arg) < 0) {
1145                                 arg->stop = 1;
1146                                 return;
1147                         }
1148                 }
1149                 arg->count++;
1150                 for (h = 0; h <= ht->divisor; h++) {
1151                         for (n = rtnl_dereference(ht->ht[h]);
1152                              n;
1153                              n = rtnl_dereference(n->next)) {
1154                                 if (arg->count < arg->skip) {
1155                                         arg->count++;
1156                                         continue;
1157                                 }
1158                                 if (arg->fn(tp, n, arg) < 0) {
1159                                         arg->stop = 1;
1160                                         return;
1161                                 }
1162                                 arg->count++;
1163                         }
1164                 }
1165         }
1166 }
1167
1168 static int u32_reoffload_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht,
1169                                bool add, tc_setup_cb_t *cb, void *cb_priv,
1170                                struct netlink_ext_ack *extack)
1171 {
1172         struct tc_cls_u32_offload cls_u32 = {};
1173         int err;
1174
1175         tc_cls_common_offload_init(&cls_u32.common, tp, ht->flags, extack);
1176         cls_u32.command = add ? TC_CLSU32_NEW_HNODE : TC_CLSU32_DELETE_HNODE;
1177         cls_u32.hnode.divisor = ht->divisor;
1178         cls_u32.hnode.handle = ht->handle;
1179         cls_u32.hnode.prio = ht->prio;
1180
1181         err = cb(TC_SETUP_CLSU32, &cls_u32, cb_priv);
1182         if (err && add && tc_skip_sw(ht->flags))
1183                 return err;
1184
1185         return 0;
1186 }
1187
1188 static int u32_reoffload_knode(struct tcf_proto *tp, struct tc_u_knode *n,
1189                                bool add, tc_setup_cb_t *cb, void *cb_priv,
1190                                struct netlink_ext_ack *extack)
1191 {
1192         struct tc_u_hnode *ht = rtnl_dereference(n->ht_down);
1193         struct tcf_block *block = tp->chain->block;
1194         struct tc_cls_u32_offload cls_u32 = {};
1195         int err;
1196
1197         tc_cls_common_offload_init(&cls_u32.common, tp, n->flags, extack);
1198         cls_u32.command = add ?
1199                 TC_CLSU32_REPLACE_KNODE : TC_CLSU32_DELETE_KNODE;
1200         cls_u32.knode.handle = n->handle;
1201
1202         if (add) {
1203                 cls_u32.knode.fshift = n->fshift;
1204 #ifdef CONFIG_CLS_U32_MARK
1205                 cls_u32.knode.val = n->val;
1206                 cls_u32.knode.mask = n->mask;
1207 #else
1208                 cls_u32.knode.val = 0;
1209                 cls_u32.knode.mask = 0;
1210 #endif
1211                 cls_u32.knode.sel = &n->sel;
1212                 cls_u32.knode.res = &n->res;
1213                 cls_u32.knode.exts = &n->exts;
1214                 if (n->ht_down)
1215                         cls_u32.knode.link_handle = ht->handle;
1216         }
1217
1218         err = cb(TC_SETUP_CLSU32, &cls_u32, cb_priv);
1219         if (err) {
1220                 if (add && tc_skip_sw(n->flags))
1221                         return err;
1222                 return 0;
1223         }
1224
1225         tc_cls_offload_cnt_update(block, &n->in_hw_count, &n->flags, add);
1226
1227         return 0;
1228 }
1229
1230 static int u32_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
1231                          void *cb_priv, struct netlink_ext_ack *extack)
1232 {
1233         struct tc_u_common *tp_c = tp->data;
1234         struct tc_u_hnode *ht;
1235         struct tc_u_knode *n;
1236         unsigned int h;
1237         int err;
1238
1239         for (ht = rtnl_dereference(tp_c->hlist);
1240              ht;
1241              ht = rtnl_dereference(ht->next)) {
1242                 if (ht->prio != tp->prio)
1243                         continue;
1244
1245                 /* When adding filters to a new dev, try to offload the
1246                  * hashtable first. When removing, do the filters before the
1247                  * hashtable.
1248                  */
1249                 if (add && !tc_skip_hw(ht->flags)) {
1250                         err = u32_reoffload_hnode(tp, ht, add, cb, cb_priv,
1251                                                   extack);
1252                         if (err)
1253                                 return err;
1254                 }
1255
1256                 for (h = 0; h <= ht->divisor; h++) {
1257                         for (n = rtnl_dereference(ht->ht[h]);
1258                              n;
1259                              n = rtnl_dereference(n->next)) {
1260                                 if (tc_skip_hw(n->flags))
1261                                         continue;
1262
1263                                 err = u32_reoffload_knode(tp, n, add, cb,
1264                                                           cb_priv, extack);
1265                                 if (err)
1266                                         return err;
1267                         }
1268                 }
1269
1270                 if (!add && !tc_skip_hw(ht->flags))
1271                         u32_reoffload_hnode(tp, ht, add, cb, cb_priv, extack);
1272         }
1273
1274         return 0;
1275 }
1276
1277 static void u32_bind_class(void *fh, u32 classid, unsigned long cl)
1278 {
1279         struct tc_u_knode *n = fh;
1280
1281         if (n && n->res.classid == classid)
1282                 n->res.class = cl;
1283 }
1284
1285 static int u32_dump(struct net *net, struct tcf_proto *tp, void *fh,
1286                     struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
1287 {
1288         struct tc_u_knode *n = fh;
1289         struct tc_u_hnode *ht_up, *ht_down;
1290         struct nlattr *nest;
1291
1292         if (n == NULL)
1293                 return skb->len;
1294
1295         t->tcm_handle = n->handle;
1296
1297         nest = nla_nest_start(skb, TCA_OPTIONS);
1298         if (nest == NULL)
1299                 goto nla_put_failure;
1300
1301         if (TC_U32_KEY(n->handle) == 0) {
1302                 struct tc_u_hnode *ht = fh;
1303                 u32 divisor = ht->divisor + 1;
1304
1305                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
1306                         goto nla_put_failure;
1307         } else {
1308 #ifdef CONFIG_CLS_U32_PERF
1309                 struct tc_u32_pcnt *gpf;
1310                 int cpu;
1311 #endif
1312
1313                 if (nla_put(skb, TCA_U32_SEL,
1314                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
1315                             &n->sel))
1316                         goto nla_put_failure;
1317
1318                 ht_up = rtnl_dereference(n->ht_up);
1319                 if (ht_up) {
1320                         u32 htid = n->handle & 0xFFFFF000;
1321                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
1322                                 goto nla_put_failure;
1323                 }
1324                 if (n->res.classid &&
1325                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
1326                         goto nla_put_failure;
1327
1328                 ht_down = rtnl_dereference(n->ht_down);
1329                 if (ht_down &&
1330                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
1331                         goto nla_put_failure;
1332
1333                 if (n->flags && nla_put_u32(skb, TCA_U32_FLAGS, n->flags))
1334                         goto nla_put_failure;
1335
1336 #ifdef CONFIG_CLS_U32_MARK
1337                 if ((n->val || n->mask)) {
1338                         struct tc_u32_mark mark = {.val = n->val,
1339                                                    .mask = n->mask,
1340                                                    .success = 0};
1341                         int cpum;
1342
1343                         for_each_possible_cpu(cpum) {
1344                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
1345
1346                                 mark.success += cnt;
1347                         }
1348
1349                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
1350                                 goto nla_put_failure;
1351                 }
1352 #endif
1353
1354                 if (tcf_exts_dump(skb, &n->exts) < 0)
1355                         goto nla_put_failure;
1356
1357 #ifdef CONFIG_NET_CLS_IND
1358                 if (n->ifindex) {
1359                         struct net_device *dev;
1360                         dev = __dev_get_by_index(net, n->ifindex);
1361                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
1362                                 goto nla_put_failure;
1363                 }
1364 #endif
1365 #ifdef CONFIG_CLS_U32_PERF
1366                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1367                               n->sel.nkeys * sizeof(u64),
1368                               GFP_KERNEL);
1369                 if (!gpf)
1370                         goto nla_put_failure;
1371
1372                 for_each_possible_cpu(cpu) {
1373                         int i;
1374                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1375
1376                         gpf->rcnt += pf->rcnt;
1377                         gpf->rhit += pf->rhit;
1378                         for (i = 0; i < n->sel.nkeys; i++)
1379                                 gpf->kcnts[i] += pf->kcnts[i];
1380                 }
1381
1382                 if (nla_put_64bit(skb, TCA_U32_PCNT,
1383                                   sizeof(struct tc_u32_pcnt) +
1384                                   n->sel.nkeys * sizeof(u64),
1385                                   gpf, TCA_U32_PAD)) {
1386                         kfree(gpf);
1387                         goto nla_put_failure;
1388                 }
1389                 kfree(gpf);
1390 #endif
1391         }
1392
1393         nla_nest_end(skb, nest);
1394
1395         if (TC_U32_KEY(n->handle))
1396                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1397                         goto nla_put_failure;
1398         return skb->len;
1399
1400 nla_put_failure:
1401         nla_nest_cancel(skb, nest);
1402         return -1;
1403 }
1404
1405 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1406         .kind           =       "u32",
1407         .classify       =       u32_classify,
1408         .init           =       u32_init,
1409         .destroy        =       u32_destroy,
1410         .get            =       u32_get,
1411         .change         =       u32_change,
1412         .delete         =       u32_delete,
1413         .walk           =       u32_walk,
1414         .reoffload      =       u32_reoffload,
1415         .dump           =       u32_dump,
1416         .bind_class     =       u32_bind_class,
1417         .owner          =       THIS_MODULE,
1418 };
1419
1420 static int __init init_u32(void)
1421 {
1422         int i, ret;
1423
1424         pr_info("u32 classifier\n");
1425 #ifdef CONFIG_CLS_U32_PERF
1426         pr_info("    Performance counters on\n");
1427 #endif
1428 #ifdef CONFIG_NET_CLS_IND
1429         pr_info("    input device check on\n");
1430 #endif
1431 #ifdef CONFIG_NET_CLS_ACT
1432         pr_info("    Actions configured\n");
1433 #endif
1434         tc_u_common_hash = kvmalloc_array(U32_HASH_SIZE,
1435                                           sizeof(struct hlist_head),
1436                                           GFP_KERNEL);
1437         if (!tc_u_common_hash)
1438                 return -ENOMEM;
1439
1440         for (i = 0; i < U32_HASH_SIZE; i++)
1441                 INIT_HLIST_HEAD(&tc_u_common_hash[i]);
1442
1443         ret = register_tcf_proto_ops(&cls_u32_ops);
1444         if (ret)
1445                 kvfree(tc_u_common_hash);
1446         return ret;
1447 }
1448
1449 static void __exit exit_u32(void)
1450 {
1451         unregister_tcf_proto_ops(&cls_u32_ops);
1452         kvfree(tc_u_common_hash);
1453 }
1454
1455 module_init(init_u32)
1456 module_exit(exit_u32)
1457 MODULE_LICENSE("GPL");