net/sched/cls_u32.c

   1 /*
   2  * net/sched/cls_u32.c  Ugly (or Universal) 32bit key Packet Classifier.
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  * Authors:     Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
  10  *
  11  *      The filters are packed to hash tables of key nodes
  12  *      with a set of 32bit key/mask pairs at every node.
  13  *      Nodes reference next level hash tables etc.
  14  *
  15  *      This scheme is the best universal classifier I managed to
  16  *      invent; it is not super-fast, but it is not slow (provided you
  17  *      program it correctly), and general enough.  And its relative
  18  *      speed grows as the number of rules becomes larger.
  19  *
  20  *      It seems that it represents the best middle point between
  21  *      speed and manageability both by human and by machine.
  22  *
  23  *      It is especially useful for link sharing combined with QoS;
  24  *      pure RSVP doesn't need such a general approach and can use
  25  *      much simpler (and faster) schemes, sort of cls_rsvp.c.
  26  *
  27  *      JHS: We should remove the CONFIG_NET_CLS_IND from here
  28  *      eventually when the meta match extension is made available
  29  *
  30  *      nfmark match added by Catalin(ux aka Dino) BOIE <catab at umbrella.ro>
  31  */
  32
  33 #include <linux/module.h>
  34 #include <linux/slab.h>
  35 #include <linux/types.h>
  36 #include <linux/kernel.h>
  37 #include <linux/string.h>
  38 #include <linux/errno.h>
  39 #include <linux/percpu.h>
  40 #include <linux/rtnetlink.h>
  41 #include <linux/skbuff.h>
  42 #include <linux/bitmap.h>
  43 #include <net/netlink.h>
  44 #include <net/act_api.h>
  45 #include <net/pkt_cls.h>
  46 #include <linux/netdevice.h>
  47
  48 struct tc_u_knode {
  49         struct tc_u_knode __rcu *next;
  50         u32                     handle;
  51         struct tc_u_hnode __rcu *ht_up;
  52         struct tcf_exts         exts;
  53 #ifdef CONFIG_NET_CLS_IND
  54         int                     ifindex;
  55 #endif
  56         u8                      fshift;
  57         struct tcf_result       res;
  58         struct tc_u_hnode __rcu *ht_down;
  59 #ifdef CONFIG_CLS_U32_PERF
  60         struct tc_u32_pcnt __percpu *pf;
  61 #endif
  62         u32                     flags;
  63 #ifdef CONFIG_CLS_U32_MARK
  64         u32                     val;
  65         u32                     mask;
  66         u32 __percpu            *pcpu_success;
  67 #endif
  68         struct tcf_proto        *tp;
  69         struct rcu_head         rcu;
  70         /* The 'sel' field MUST be the last field in structure to allow for
  71          * tc_u32_keys allocated at end of structure.
  72          */
  73         struct tc_u32_sel       sel;
  74 };
  75
  76 struct tc_u_hnode {
  77         struct tc_u_hnode __rcu *next;
  78         u32                     handle;
  79         u32                     prio;
  80         struct tc_u_common      *tp_c;
  81         int                     refcnt;
  82         unsigned int            divisor;
  83         struct rcu_head         rcu;
  84         /* The 'ht' field MUST be the last field in structure to allow for
  85          * more entries allocated at end of structure.
  86          */
  87         struct tc_u_knode __rcu *ht[1];
  88 };
  89
  90 struct tc_u_common {
  91         struct tc_u_hnode __rcu *hlist;
  92         struct Qdisc            *q;
  93         int                     refcnt;
  94         u32                     hgenerator;
  95         struct rcu_head         rcu;
  96 };
  97
  98 static inline unsigned int u32_hash_fold(__be32 key,
  99                                          const struct tc_u32_sel *sel,
 100                                          u8 fshift)
 101 {
 102         unsigned int h = ntohl(key & sel->hmask) >> fshift;
 103
 104         return h;
 105 }
 106
 107 static int u32_classify(struct sk_buff *skb, const struct tcf_proto *tp, struct tcf_result *res)
 108 {
 109         struct {
 110                 struct tc_u_knode *knode;
 111                 unsigned int      off;
 112         } stack[TC_U32_MAXDEPTH];
 113
 114         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 115         unsigned int off = skb_network_offset(skb);
 116         struct tc_u_knode *n;
 117         int sdepth = 0;
 118         int off2 = 0;
 119         int sel = 0;
 120 #ifdef CONFIG_CLS_U32_PERF
 121         int j;
 122 #endif
 123         int i, r;
 124
 125 next_ht:
 126         n = rcu_dereference_bh(ht->ht[sel]);
 127
 128 next_knode:
 129         if (n) {
 130                 struct tc_u32_key *key = n->sel.keys;
 131
 132 #ifdef CONFIG_CLS_U32_PERF
 133                 __this_cpu_inc(n->pf->rcnt);
 134                 j = 0;
 135 #endif
 136
 137                 if (tc_skip_sw(n->flags)) {
 138                         n = rcu_dereference_bh(n->next);
 139                         goto next_knode;
 140                 }
 141
 142 #ifdef CONFIG_CLS_U32_MARK
 143                 if ((skb->mark & n->mask) != n->val) {
 144                         n = rcu_dereference_bh(n->next);
 145                         goto next_knode;
 146                 } else {
 147                         __this_cpu_inc(*n->pcpu_success);
 148                 }
 149 #endif
 150
 151                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 152                         int toff = off + key->off + (off2 & key->offmask);
 153                         __be32 *data, hdata;
 154
 155                         if (skb_headroom(skb) + toff > INT_MAX)
 156                                 goto out;
 157
 158                         data = skb_header_pointer(skb, toff, 4, &hdata);
 159                         if (!data)
 160                                 goto out;
 161                         if ((*data ^ key->val) & key->mask) {
 162                                 n = rcu_dereference_bh(n->next);
 163                                 goto next_knode;
 164                         }
 165 #ifdef CONFIG_CLS_U32_PERF
 166                         __this_cpu_inc(n->pf->kcnts[j]);
 167                         j++;
 168 #endif
 169                 }
 170
 171                 ht = rcu_dereference_bh(n->ht_down);
 172                 if (!ht) {
 173 check_terminal:
 174                         if (n->sel.flags & TC_U32_TERMINAL) {
 175
 176                                 *res = n->res;
 177 #ifdef CONFIG_NET_CLS_IND
 178                                 if (!tcf_match_indev(skb, n->ifindex)) {
 179                                         n = rcu_dereference_bh(n->next);
 180                                         goto next_knode;
 181                                 }
 182 #endif
 183 #ifdef CONFIG_CLS_U32_PERF
 184                                 __this_cpu_inc(n->pf->rhit);
 185 #endif
 186                                 r = tcf_exts_exec(skb, &n->exts, res);
 187                                 if (r < 0) {
 188                                         n = rcu_dereference_bh(n->next);
 189                                         goto next_knode;
 190                                 }
 191
 192                                 return r;
 193                         }
 194                         n = rcu_dereference_bh(n->next);
 195                         goto next_knode;
 196                 }
 197
 198                 /* PUSH */
 199                 if (sdepth >= TC_U32_MAXDEPTH)
 200                         goto deadloop;
 201                 stack[sdepth].knode = n;
 202                 stack[sdepth].off = off;
 203                 sdepth++;
 204
 205                 ht = rcu_dereference_bh(n->ht_down);
 206                 sel = 0;
 207                 if (ht->divisor) {
 208                         __be32 *data, hdata;
 209
 210                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 211                                                   &hdata);
 212                         if (!data)
 213                                 goto out;
 214                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 215                                                           n->fshift);
 216                 }
 217                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 218                         goto next_ht;
 219
 220                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 221                         off2 = n->sel.off + 3;
 222                         if (n->sel.flags & TC_U32_VAROFFSET) {
 223                                 __be16 *data, hdata;
 224
 225                                 data = skb_header_pointer(skb,
 226                                                           off + n->sel.offoff,
 227                                                           2, &hdata);
 228                                 if (!data)
 229                                         goto out;
 230                                 off2 += ntohs(n->sel.offmask & *data) >>
 231                                         n->sel.offshift;
 232                         }
 233                         off2 &= ~3;
 234                 }
 235                 if (n->sel.flags & TC_U32_EAT) {
 236                         off += off2;
 237                         off2 = 0;
 238                 }
 239
 240                 if (off < skb->len)
 241                         goto next_ht;
 242         }
 243
 244         /* POP */
 245         if (sdepth--) {
 246                 n = stack[sdepth].knode;
 247                 ht = rcu_dereference_bh(n->ht_up);
 248                 off = stack[sdepth].off;
 249                 goto check_terminal;
 250         }
 251 out:
 252         return -1;
 253
 254 deadloop:
 255         net_warn_ratelimited("cls_u32: dead loop\n");
 256         return -1;
 257 }
 258
 259 static struct tc_u_hnode *
 260 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 261 {
 262         struct tc_u_hnode *ht;
 263
 264         for (ht = rtnl_dereference(tp_c->hlist);
 265              ht;
 266              ht = rtnl_dereference(ht->next))
 267                 if (ht->handle == handle)
 268                         break;
 269
 270         return ht;
 271 }
 272
 273 static struct tc_u_knode *
 274 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 275 {
 276         unsigned int sel;
 277         struct tc_u_knode *n = NULL;
 278
 279         sel = TC_U32_HASH(handle);
 280         if (sel > ht->divisor)
 281                 goto out;
 282
 283         for (n = rtnl_dereference(ht->ht[sel]);
 284              n;
 285              n = rtnl_dereference(n->next))
 286                 if (n->handle == handle)
 287                         break;
 288 out:
 289         return n;
 290 }
 291
 292
 293 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 294 {
 295         struct tc_u_hnode *ht;
 296         struct tc_u_common *tp_c = tp->data;
 297
 298         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 299                 ht = rtnl_dereference(tp->root);
 300         else
 301                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 302
 303         if (!ht)
 304                 return 0;
 305
 306         if (TC_U32_KEY(handle) == 0)
 307                 return (unsigned long)ht;
 308
 309         return (unsigned long)u32_lookup_key(ht, handle);
 310 }
 311
 312 static u32 gen_new_htid(struct tc_u_common *tp_c)
 313 {
 314         int i = 0x800;
 315
 316         /* hgenerator only used inside rtnl lock it is safe to increment
 317          * without read _copy_ update semantics
 318          */
 319         do {
 320                 if (++tp_c->hgenerator == 0x7FF)
 321                         tp_c->hgenerator = 1;
 322         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 323
 324         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 325 }
 326
 327 static int u32_init(struct tcf_proto *tp)
 328 {
 329         struct tc_u_hnode *root_ht;
 330         struct tc_u_common *tp_c;
 331
 332         tp_c = tp->q->u32_node;
 333
 334         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 335         if (root_ht == NULL)
 336                 return -ENOBUFS;
 337
 338         root_ht->divisor = 0;
 339         root_ht->refcnt++;
 340         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 341         root_ht->prio = tp->prio;
 342
 343         if (tp_c == NULL) {
 344                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 345                 if (tp_c == NULL) {
 346                         kfree(root_ht);
 347                         return -ENOBUFS;
 348                 }
 349                 tp_c->q = tp->q;
 350                 tp->q->u32_node = tp_c;
 351         }
 352
 353         tp_c->refcnt++;
 354         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 355         rcu_assign_pointer(tp_c->hlist, root_ht);
 356         root_ht->tp_c = tp_c;
 357
 358         rcu_assign_pointer(tp->root, root_ht);
 359         tp->data = tp_c;
 360         return 0;
 361 }
 362
 363 static int u32_destroy_key(struct tcf_proto *tp,
 364                            struct tc_u_knode *n,
 365                            bool free_pf)
 366 {
 367         tcf_exts_destroy(&n->exts);
 368         if (n->ht_down)
 369                 n->ht_down->refcnt--;
 370 #ifdef CONFIG_CLS_U32_PERF
 371         if (free_pf)
 372                 free_percpu(n->pf);
 373 #endif
 374 #ifdef CONFIG_CLS_U32_MARK
 375         if (free_pf)
 376                 free_percpu(n->pcpu_success);
 377 #endif
 378         kfree(n);
 379         return 0;
 380 }
 381
 382 /* u32_delete_key_rcu should be called when free'ing a copied
 383  * version of a tc_u_knode obtained from u32_init_knode(). When
 384  * copies are obtained from u32_init_knode() the statistics are
 385  * shared between the old and new copies to allow readers to
 386  * continue to update the statistics during the copy. To support
 387  * this the u32_delete_key_rcu variant does not free the percpu
 388  * statistics.
 389  */
 390 static void u32_delete_key_rcu(struct rcu_head *rcu)
 391 {
 392         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 393
 394         u32_destroy_key(key->tp, key, false);
 395 }
 396
 397 /* u32_delete_key_freepf_rcu is the rcu callback variant
 398  * that free's the entire structure including the statistics
 399  * percpu variables. Only use this if the key is not a copy
 400  * returned by u32_init_knode(). See u32_delete_key_rcu()
 401  * for the variant that should be used with keys return from
 402  * u32_init_knode()
 403  */
 404 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 405 {
 406         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 407
 408         u32_destroy_key(key->tp, key, true);
 409 }
 410
 411 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 412 {
 413         struct tc_u_knode __rcu **kp;
 414         struct tc_u_knode *pkp;
 415         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 416
 417         if (ht) {
 418                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 419                 for (pkp = rtnl_dereference(*kp); pkp;
 420                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 421                         if (pkp == key) {
 422                                 RCU_INIT_POINTER(*kp, key->next);
 423
 424                                 tcf_unbind_filter(tp, &key->res);
 425                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 426                                 return 0;
 427                         }
 428                 }
 429         }
 430         WARN_ON(1);
 431         return 0;
 432 }
 433
 434 static void u32_remove_hw_knode(struct tcf_proto *tp, u32 handle)
 435 {
 436         struct net_device *dev = tp->q->dev_queue->dev;
 437         struct tc_cls_u32_offload u32_offload = {0};
 438         struct tc_to_netdev offload;
 439
 440         offload.type = TC_SETUP_CLSU32;
 441         offload.cls_u32 = &u32_offload;
 442
 443         if (tc_should_offload(dev, 0)) {
 444                 offload.cls_u32->command = TC_CLSU32_DELETE_KNODE;
 445                 offload.cls_u32->knode.handle = handle;
 446                 dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 447                                               tp->protocol, &offload);
 448         }
 449 }
 450
 451 static int u32_replace_hw_hnode(struct tcf_proto *tp,
 452                                  struct tc_u_hnode *h,
 453                                  u32 flags)
 454 {
 455         struct net_device *dev = tp->q->dev_queue->dev;
 456         struct tc_cls_u32_offload u32_offload = {0};
 457         struct tc_to_netdev offload;
 458         int err;
 459
 460         offload.type = TC_SETUP_CLSU32;
 461         offload.cls_u32 = &u32_offload;
 462
 463         if (tc_should_offload(dev, flags)) {
 464                 offload.cls_u32->command = TC_CLSU32_NEW_HNODE;
 465                 offload.cls_u32->hnode.divisor = h->divisor;
 466                 offload.cls_u32->hnode.handle = h->handle;
 467                 offload.cls_u32->hnode.prio = h->prio;
 468
 469                 err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 470                                                     tp->protocol, &offload);
 471                 if (tc_skip_sw(flags))
 472                         return err;
 473         }
 474
 475         return 0;
 476 }
 477
 478 static void u32_clear_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h)
 479 {
 480         struct net_device *dev = tp->q->dev_queue->dev;
 481         struct tc_cls_u32_offload u32_offload = {0};
 482         struct tc_to_netdev offload;
 483
 484         offload.type = TC_SETUP_CLSU32;
 485         offload.cls_u32 = &u32_offload;
 486
 487         if (tc_should_offload(dev, 0)) {
 488                 offload.cls_u32->command = TC_CLSU32_DELETE_HNODE;
 489                 offload.cls_u32->hnode.divisor = h->divisor;
 490                 offload.cls_u32->hnode.handle = h->handle;
 491                 offload.cls_u32->hnode.prio = h->prio;
 492
 493                 dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 494                                               tp->protocol, &offload);
 495         }
 496 }
 497
 498 static int u32_replace_hw_knode(struct tcf_proto *tp,
 499                                  struct tc_u_knode *n,
 500                                  u32 flags)
 501 {
 502         struct net_device *dev = tp->q->dev_queue->dev;
 503         struct tc_cls_u32_offload u32_offload = {0};
 504         struct tc_to_netdev offload;
 505         int err;
 506
 507         offload.type = TC_SETUP_CLSU32;
 508         offload.cls_u32 = &u32_offload;
 509
 510         if (tc_should_offload(dev, flags)) {
 511                 offload.cls_u32->command = TC_CLSU32_REPLACE_KNODE;
 512                 offload.cls_u32->knode.handle = n->handle;
 513                 offload.cls_u32->knode.fshift = n->fshift;
 514 #ifdef CONFIG_CLS_U32_MARK
 515                 offload.cls_u32->knode.val = n->val;
 516                 offload.cls_u32->knode.mask = n->mask;
 517 #else
 518                 offload.cls_u32->knode.val = 0;
 519                 offload.cls_u32->knode.mask = 0;
 520 #endif
 521                 offload.cls_u32->knode.sel = &n->sel;
 522                 offload.cls_u32->knode.exts = &n->exts;
 523                 if (n->ht_down)
 524                         offload.cls_u32->knode.link_handle = n->ht_down->handle;
 525
 526                 err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 527                                                     tp->protocol, &offload);
 528                 if (tc_skip_sw(flags))
 529                         return err;
 530         }
 531
 532         return 0;
 533 }
 534
 535 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 536 {
 537         struct tc_u_knode *n;
 538         unsigned int h;
 539
 540         for (h = 0; h <= ht->divisor; h++) {
 541                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 542                         RCU_INIT_POINTER(ht->ht[h],
 543                                          rtnl_dereference(n->next));
 544                         tcf_unbind_filter(tp, &n->res);
 545                         u32_remove_hw_knode(tp, n->handle);
 546                         call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 547                 }
 548         }
 549 }
 550
 551 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 552 {
 553         struct tc_u_common *tp_c = tp->data;
 554         struct tc_u_hnode __rcu **hn;
 555         struct tc_u_hnode *phn;
 556
 557         WARN_ON(ht->refcnt);
 558
 559         u32_clear_hnode(tp, ht);
 560
 561         hn = &tp_c->hlist;
 562         for (phn = rtnl_dereference(*hn);
 563              phn;
 564              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 565                 if (phn == ht) {
 566                         u32_clear_hw_hnode(tp, ht);
 567                         RCU_INIT_POINTER(*hn, ht->next);
 568                         kfree_rcu(ht, rcu);
 569                         return 0;
 570                 }
 571         }
 572
 573         return -ENOENT;
 574 }
 575
 576 static bool ht_empty(struct tc_u_hnode *ht)
 577 {
 578         unsigned int h;
 579
 580         for (h = 0; h <= ht->divisor; h++)
 581                 if (rcu_access_pointer(ht->ht[h]))
 582                         return false;
 583
 584         return true;
 585 }
 586
 587 static bool u32_destroy(struct tcf_proto *tp, bool force)
 588 {
 589         struct tc_u_common *tp_c = tp->data;
 590         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 591
 592         WARN_ON(root_ht == NULL);
 593
 594         if (!force) {
 595                 if (root_ht) {
 596                         if (root_ht->refcnt > 1)
 597                                 return false;
 598                         if (root_ht->refcnt == 1) {
 599                                 if (!ht_empty(root_ht))
 600                                         return false;
 601                         }
 602                 }
 603
 604                 if (tp_c->refcnt > 1)
 605                         return false;
 606
 607                 if (tp_c->refcnt == 1) {
 608                         struct tc_u_hnode *ht;
 609
 610                         for (ht = rtnl_dereference(tp_c->hlist);
 611                              ht;
 612                              ht = rtnl_dereference(ht->next))
 613                                 if (!ht_empty(ht))
 614                                         return false;
 615                 }
 616         }
 617
 618         if (root_ht && --root_ht->refcnt == 0)
 619                 u32_destroy_hnode(tp, root_ht);
 620
 621         if (--tp_c->refcnt == 0) {
 622                 struct tc_u_hnode *ht;
 623
 624                 tp->q->u32_node = NULL;
 625
 626                 for (ht = rtnl_dereference(tp_c->hlist);
 627                      ht;
 628                      ht = rtnl_dereference(ht->next)) {
 629                         ht->refcnt--;
 630                         u32_clear_hnode(tp, ht);
 631                 }
 632
 633                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 634                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 635                         kfree_rcu(ht, rcu);
 636                 }
 637
 638                 kfree(tp_c);
 639         }
 640
 641         tp->data = NULL;
 642         return true;
 643 }
 644
 645 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 646 {
 647         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 648         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 649
 650         if (ht == NULL)
 651                 return 0;
 652
 653         if (TC_U32_KEY(ht->handle)) {
 654                 u32_remove_hw_knode(tp, ht->handle);
 655                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 656         }
 657
 658         if (root_ht == ht)
 659                 return -EINVAL;
 660
 661         if (ht->refcnt == 1) {
 662                 ht->refcnt--;
 663                 u32_destroy_hnode(tp, ht);
 664         } else {
 665                 return -EBUSY;
 666         }
 667
 668         return 0;
 669 }
 670
 671 #define NR_U32_NODE (1<<12)
 672 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 673 {
 674         struct tc_u_knode *n;
 675         unsigned long i;
 676         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 677                                         GFP_KERNEL);
 678         if (!bitmap)
 679                 return handle | 0xFFF;
 680
 681         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 682              n;
 683              n = rtnl_dereference(n->next))
 684                 set_bit(TC_U32_NODE(n->handle), bitmap);
 685
 686         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 687         if (i >= NR_U32_NODE)
 688                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 689
 690         kfree(bitmap);
 691         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 692 }
 693
 694 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 695         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 696         [TCA_U32_HASH]          = { .type = NLA_U32 },
 697         [TCA_U32_LINK]          = { .type = NLA_U32 },
 698         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 699         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 700         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 701         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 702         [TCA_U32_FLAGS]         = { .type = NLA_U32 },
 703 };
 704
 705 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 706                          unsigned long base, struct tc_u_hnode *ht,
 707                          struct tc_u_knode *n, struct nlattr **tb,
 708                          struct nlattr *est, bool ovr)
 709 {
 710         int err;
 711         struct tcf_exts e;
 712
 713         tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 714         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 715         if (err < 0)
 716                 return err;
 717
 718         err = -EINVAL;
 719         if (tb[TCA_U32_LINK]) {
 720                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 721                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 722
 723                 if (TC_U32_KEY(handle))
 724                         goto errout;
 725
 726                 if (handle) {
 727                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 728
 729                         if (ht_down == NULL)
 730                                 goto errout;
 731                         ht_down->refcnt++;
 732                 }
 733
 734                 ht_old = rtnl_dereference(n->ht_down);
 735                 rcu_assign_pointer(n->ht_down, ht_down);
 736
 737                 if (ht_old)
 738                         ht_old->refcnt--;
 739         }
 740         if (tb[TCA_U32_CLASSID]) {
 741                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 742                 tcf_bind_filter(tp, &n->res, base);
 743         }
 744
 745 #ifdef CONFIG_NET_CLS_IND
 746         if (tb[TCA_U32_INDEV]) {
 747                 int ret;
 748                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 749                 if (ret < 0)
 750                         goto errout;
 751                 n->ifindex = ret;
 752         }
 753 #endif
 754         tcf_exts_change(tp, &n->exts, &e);
 755
 756         return 0;
 757 errout:
 758         tcf_exts_destroy(&e);
 759         return err;
 760 }
 761
 762 static void u32_replace_knode(struct tcf_proto *tp,
 763                               struct tc_u_common *tp_c,
 764                               struct tc_u_knode *n)
 765 {
 766         struct tc_u_knode __rcu **ins;
 767         struct tc_u_knode *pins;
 768         struct tc_u_hnode *ht;
 769
 770         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 771                 ht = rtnl_dereference(tp->root);
 772         else
 773                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 774
 775         ins = &ht->ht[TC_U32_HASH(n->handle)];
 776
 777         /* The node must always exist for it to be replaced if this is not the
 778          * case then something went very wrong elsewhere.
 779          */
 780         for (pins = rtnl_dereference(*ins); ;
 781              ins = &pins->next, pins = rtnl_dereference(*ins))
 782                 if (pins->handle == n->handle)
 783                         break;
 784
 785         RCU_INIT_POINTER(n->next, pins->next);
 786         rcu_assign_pointer(*ins, n);
 787 }
 788
 789 static struct tc_u_knode *u32_init_knode(struct tcf_proto *tp,
 790                                          struct tc_u_knode *n)
 791 {
 792         struct tc_u_knode *new;
 793         struct tc_u32_sel *s = &n->sel;
 794
 795         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 796                       GFP_KERNEL);
 797
 798         if (!new)
 799                 return NULL;
 800
 801         RCU_INIT_POINTER(new->next, n->next);
 802         new->handle = n->handle;
 803         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 804
 805 #ifdef CONFIG_NET_CLS_IND
 806         new->ifindex = n->ifindex;
 807 #endif
 808         new->fshift = n->fshift;
 809         new->res = n->res;
 810         new->flags = n->flags;
 811         RCU_INIT_POINTER(new->ht_down, n->ht_down);
 812
 813         /* bump reference count as long as we hold pointer to structure */
 814         if (new->ht_down)
 815                 new->ht_down->refcnt++;
 816
 817 #ifdef CONFIG_CLS_U32_PERF
 818         /* Statistics may be incremented by readers during update
 819          * so we must keep them in tact. When the node is later destroyed
 820          * a special destroy call must be made to not free the pf memory.
 821          */
 822         new->pf = n->pf;
 823 #endif
 824
 825 #ifdef CONFIG_CLS_U32_MARK
 826         new->val = n->val;
 827         new->mask = n->mask;
 828         /* Similarly success statistics must be moved as pointers */
 829         new->pcpu_success = n->pcpu_success;
 830 #endif
 831         new->tp = tp;
 832         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 833
 834         tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE);
 835
 836         return new;
 837 }
 838
 839 static int u32_change(struct net *net, struct sk_buff *in_skb,
 840                       struct tcf_proto *tp, unsigned long base, u32 handle,
 841                       struct nlattr **tca,
 842                       unsigned long *arg, bool ovr)
 843 {
 844         struct tc_u_common *tp_c = tp->data;
 845         struct tc_u_hnode *ht;
 846         struct tc_u_knode *n;
 847         struct tc_u32_sel *s;
 848         struct nlattr *opt = tca[TCA_OPTIONS];
 849         struct nlattr *tb[TCA_U32_MAX + 1];
 850         u32 htid, flags = 0;
 851         int err;
 852 #ifdef CONFIG_CLS_U32_PERF
 853         size_t size;
 854 #endif
 855
 856         if (opt == NULL)
 857                 return handle ? -EINVAL : 0;
 858
 859         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 860         if (err < 0)
 861                 return err;
 862
 863         if (tb[TCA_U32_FLAGS]) {
 864                 flags = nla_get_u32(tb[TCA_U32_FLAGS]);
 865                 if (!tc_flags_valid(flags))
 866                         return err;
 867         }
 868
 869         n = (struct tc_u_knode *)*arg;
 870         if (n) {
 871                 struct tc_u_knode *new;
 872
 873                 if (TC_U32_KEY(n->handle) == 0)
 874                         return -EINVAL;
 875
 876                 if (n->flags != flags)
 877                         return -EINVAL;
 878
 879                 new = u32_init_knode(tp, n);
 880                 if (!new)
 881                         return -ENOMEM;
 882
 883                 err = u32_set_parms(net, tp, base,
 884                                     rtnl_dereference(n->ht_up), new, tb,
 885                                     tca[TCA_RATE], ovr);
 886
 887                 if (err) {
 888                         u32_destroy_key(tp, new, false);
 889                         return err;
 890                 }
 891
 892                 err = u32_replace_hw_knode(tp, new, flags);
 893                 if (err) {
 894                         u32_destroy_key(tp, new, false);
 895                         return err;
 896                 }
 897
 898                 u32_replace_knode(tp, tp_c, new);
 899                 tcf_unbind_filter(tp, &n->res);
 900                 call_rcu(&n->rcu, u32_delete_key_rcu);
 901                 return 0;
 902         }
 903
 904         if (tb[TCA_U32_DIVISOR]) {
 905                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 906
 907                 if (--divisor > 0x100)
 908                         return -EINVAL;
 909                 if (TC_U32_KEY(handle))
 910                         return -EINVAL;
 911                 if (handle == 0) {
 912                         handle = gen_new_htid(tp->data);
 913                         if (handle == 0)
 914                                 return -ENOMEM;
 915                 }
 916                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 917                 if (ht == NULL)
 918                         return -ENOBUFS;
 919                 ht->tp_c = tp_c;
 920                 ht->refcnt = 1;
 921                 ht->divisor = divisor;
 922                 ht->handle = handle;
 923                 ht->prio = tp->prio;
 924                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 925                 rcu_assign_pointer(tp_c->hlist, ht);
 926                 *arg = (unsigned long)ht;
 927
 928                 u32_replace_hw_hnode(tp, ht, flags);
 929                 return 0;
 930         }
 931
 932         if (tb[TCA_U32_HASH]) {
 933                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 934                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 935                         ht = rtnl_dereference(tp->root);
 936                         htid = ht->handle;
 937                 } else {
 938                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 939                         if (ht == NULL)
 940                                 return -EINVAL;
 941                 }
 942         } else {
 943                 ht = rtnl_dereference(tp->root);
 944                 htid = ht->handle;
 945         }
 946
 947         if (ht->divisor < TC_U32_HASH(htid))
 948                 return -EINVAL;
 949
 950         if (handle) {
 951                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 952                         return -EINVAL;
 953                 handle = htid | TC_U32_NODE(handle);
 954         } else
 955                 handle = gen_new_kid(ht, htid);
 956
 957         if (tb[TCA_U32_SEL] == NULL)
 958                 return -EINVAL;
 959
 960         s = nla_data(tb[TCA_U32_SEL]);
 961
 962         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 963         if (n == NULL)
 964                 return -ENOBUFS;
 965
 966 #ifdef CONFIG_CLS_U32_PERF
 967         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 968         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 969         if (!n->pf) {
 970                 kfree(n);
 971                 return -ENOBUFS;
 972         }
 973 #endif
 974
 975         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 976         RCU_INIT_POINTER(n->ht_up, ht);
 977         n->handle = handle;
 978         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 979         n->flags = flags;
 980         tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 981         n->tp = tp;
 982
 983 #ifdef CONFIG_CLS_U32_MARK
 984         n->pcpu_success = alloc_percpu(u32);
 985         if (!n->pcpu_success) {
 986                 err = -ENOMEM;
 987                 goto errout;
 988         }
 989
 990         if (tb[TCA_U32_MARK]) {
 991                 struct tc_u32_mark *mark;
 992
 993                 mark = nla_data(tb[TCA_U32_MARK]);
 994                 n->val = mark->val;
 995                 n->mask = mark->mask;
 996         }
 997 #endif
 998
 999         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
1000         if (err == 0) {
1001                 struct tc_u_knode __rcu **ins;
1002                 struct tc_u_knode *pins;
1003
1004                 err = u32_replace_hw_knode(tp, n, flags);
1005                 if (err)
1006                         goto errhw;
1007
1008                 ins = &ht->ht[TC_U32_HASH(handle)];
1009                 for (pins = rtnl_dereference(*ins); pins;
1010                      ins = &pins->next, pins = rtnl_dereference(*ins))
1011                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
1012                                 break;
1013
1014                 RCU_INIT_POINTER(n->next, pins);
1015                 rcu_assign_pointer(*ins, n);
1016                 *arg = (unsigned long)n;
1017                 return 0;
1018         }
1019
1020 errhw:
1021 #ifdef CONFIG_CLS_U32_MARK
1022         free_percpu(n->pcpu_success);
1023 errout:
1024 #endif
1025
1026 #ifdef CONFIG_CLS_U32_PERF
1027         free_percpu(n->pf);
1028 #endif
1029         kfree(n);
1030         return err;
1031 }
1032
1033 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1034 {
1035         struct tc_u_common *tp_c = tp->data;
1036         struct tc_u_hnode *ht;
1037         struct tc_u_knode *n;
1038         unsigned int h;
1039
1040         if (arg->stop)
1041                 return;
1042
1043         for (ht = rtnl_dereference(tp_c->hlist);
1044              ht;
1045              ht = rtnl_dereference(ht->next)) {
1046                 if (ht->prio != tp->prio)
1047                         continue;
1048                 if (arg->count >= arg->skip) {
1049                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
1050                                 arg->stop = 1;
1051                                 return;
1052                         }
1053                 }
1054                 arg->count++;
1055                 for (h = 0; h <= ht->divisor; h++) {
1056                         for (n = rtnl_dereference(ht->ht[h]);
1057                              n;
1058                              n = rtnl_dereference(n->next)) {
1059                                 if (arg->count < arg->skip) {
1060                                         arg->count++;
1061                                         continue;
1062                                 }
1063                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
1064                                         arg->stop = 1;
1065                                         return;
1066                                 }
1067                                 arg->count++;
1068                         }
1069                 }
1070         }
1071 }
1072
1073 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
1074                      struct sk_buff *skb, struct tcmsg *t)
1075 {
1076         struct tc_u_knode *n = (struct tc_u_knode *)fh;
1077         struct tc_u_hnode *ht_up, *ht_down;
1078         struct nlattr *nest;
1079
1080         if (n == NULL)
1081                 return skb->len;
1082
1083         t->tcm_handle = n->handle;
1084
1085         nest = nla_nest_start(skb, TCA_OPTIONS);
1086         if (nest == NULL)
1087                 goto nla_put_failure;
1088
1089         if (TC_U32_KEY(n->handle) == 0) {
1090                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
1091                 u32 divisor = ht->divisor + 1;
1092
1093                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
1094                         goto nla_put_failure;
1095         } else {
1096 #ifdef CONFIG_CLS_U32_PERF
1097                 struct tc_u32_pcnt *gpf;
1098                 int cpu;
1099 #endif
1100
1101                 if (nla_put(skb, TCA_U32_SEL,
1102                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
1103                             &n->sel))
1104                         goto nla_put_failure;
1105
1106                 ht_up = rtnl_dereference(n->ht_up);
1107                 if (ht_up) {
1108                         u32 htid = n->handle & 0xFFFFF000;
1109                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
1110                                 goto nla_put_failure;
1111                 }
1112                 if (n->res.classid &&
1113                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
1114                         goto nla_put_failure;
1115
1116                 ht_down = rtnl_dereference(n->ht_down);
1117                 if (ht_down &&
1118                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
1119                         goto nla_put_failure;
1120
1121                 if (n->flags && nla_put_u32(skb, TCA_U32_FLAGS, n->flags))
1122                         goto nla_put_failure;
1123
1124 #ifdef CONFIG_CLS_U32_MARK
1125                 if ((n->val || n->mask)) {
1126                         struct tc_u32_mark mark = {.val = n->val,
1127                                                    .mask = n->mask,
1128                                                    .success = 0};
1129                         int cpum;
1130
1131                         for_each_possible_cpu(cpum) {
1132                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
1133
1134                                 mark.success += cnt;
1135                         }
1136
1137                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
1138                                 goto nla_put_failure;
1139                 }
1140 #endif
1141
1142                 if (tcf_exts_dump(skb, &n->exts) < 0)
1143                         goto nla_put_failure;
1144
1145 #ifdef CONFIG_NET_CLS_IND
1146                 if (n->ifindex) {
1147                         struct net_device *dev;
1148                         dev = __dev_get_by_index(net, n->ifindex);
1149                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
1150                                 goto nla_put_failure;
1151                 }
1152 #endif
1153 #ifdef CONFIG_CLS_U32_PERF
1154                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1155                               n->sel.nkeys * sizeof(u64),
1156                               GFP_KERNEL);
1157                 if (!gpf)
1158                         goto nla_put_failure;
1159
1160                 for_each_possible_cpu(cpu) {
1161                         int i;
1162                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1163
1164                         gpf->rcnt += pf->rcnt;
1165                         gpf->rhit += pf->rhit;
1166                         for (i = 0; i < n->sel.nkeys; i++)
1167                                 gpf->kcnts[i] += pf->kcnts[i];
1168                 }
1169
1170                 if (nla_put_64bit(skb, TCA_U32_PCNT,
1171                                   sizeof(struct tc_u32_pcnt) +
1172                                   n->sel.nkeys * sizeof(u64),
1173                                   gpf, TCA_U32_PAD)) {
1174                         kfree(gpf);
1175                         goto nla_put_failure;
1176                 }
1177                 kfree(gpf);
1178 #endif
1179         }
1180
1181         nla_nest_end(skb, nest);
1182
1183         if (TC_U32_KEY(n->handle))
1184                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1185                         goto nla_put_failure;
1186         return skb->len;
1187
1188 nla_put_failure:
1189         nla_nest_cancel(skb, nest);
1190         return -1;
1191 }
1192
1193 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1194         .kind           =       "u32",
1195         .classify       =       u32_classify,
1196         .init           =       u32_init,
1197         .destroy        =       u32_destroy,
1198         .get            =       u32_get,
1199         .change         =       u32_change,
1200         .delete         =       u32_delete,
1201         .walk           =       u32_walk,
1202         .dump           =       u32_dump,
1203         .owner          =       THIS_MODULE,
1204 };
1205
1206 static int __init init_u32(void)
1207 {
1208         pr_info("u32 classifier\n");
1209 #ifdef CONFIG_CLS_U32_PERF
1210         pr_info("    Performance counters on\n");
1211 #endif
1212 #ifdef CONFIG_NET_CLS_IND
1213         pr_info("    input device check on\n");
1214 #endif
1215 #ifdef CONFIG_NET_CLS_ACT
1216         pr_info("    Actions configured\n");
1217 #endif
1218         return register_tcf_proto_ops(&cls_u32_ops);
1219 }
1220
1221 static void __exit exit_u32(void)
1222 {
1223         unregister_tcf_proto_ops(&cls_u32_ops);
1224 }
1225
1226 module_init(init_u32)
1227 module_exit(exit_u32)
1228 MODULE_LICENSE("GPL");