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,
 108                         struct tcf_result *res)
 109 {
 110         struct {
 111                 struct tc_u_knode *knode;
 112                 unsigned int      off;
 113         } stack[TC_U32_MAXDEPTH];
 114
 115         struct tc_u_hnode *ht = rcu_dereference_bh(tp->root);
 116         unsigned int off = skb_network_offset(skb);
 117         struct tc_u_knode *n;
 118         int sdepth = 0;
 119         int off2 = 0;
 120         int sel = 0;
 121 #ifdef CONFIG_CLS_U32_PERF
 122         int j;
 123 #endif
 124         int i, r;
 125
 126 next_ht:
 127         n = rcu_dereference_bh(ht->ht[sel]);
 128
 129 next_knode:
 130         if (n) {
 131                 struct tc_u32_key *key = n->sel.keys;
 132
 133 #ifdef CONFIG_CLS_U32_PERF
 134                 __this_cpu_inc(n->pf->rcnt);
 135                 j = 0;
 136 #endif
 137
 138                 if (tc_skip_sw(n->flags)) {
 139                         n = rcu_dereference_bh(n->next);
 140                         goto next_knode;
 141                 }
 142
 143 #ifdef CONFIG_CLS_U32_MARK
 144                 if ((skb->mark & n->mask) != n->val) {
 145                         n = rcu_dereference_bh(n->next);
 146                         goto next_knode;
 147                 } else {
 148                         __this_cpu_inc(*n->pcpu_success);
 149                 }
 150 #endif
 151
 152                 for (i = n->sel.nkeys; i > 0; i--, key++) {
 153                         int toff = off + key->off + (off2 & key->offmask);
 154                         __be32 *data, hdata;
 155
 156                         if (skb_headroom(skb) + toff > INT_MAX)
 157                                 goto out;
 158
 159                         data = skb_header_pointer(skb, toff, 4, &hdata);
 160                         if (!data)
 161                                 goto out;
 162                         if ((*data ^ key->val) & key->mask) {
 163                                 n = rcu_dereference_bh(n->next);
 164                                 goto next_knode;
 165                         }
 166 #ifdef CONFIG_CLS_U32_PERF
 167                         __this_cpu_inc(n->pf->kcnts[j]);
 168                         j++;
 169 #endif
 170                 }
 171
 172                 ht = rcu_dereference_bh(n->ht_down);
 173                 if (!ht) {
 174 check_terminal:
 175                         if (n->sel.flags & TC_U32_TERMINAL) {
 176
 177                                 *res = n->res;
 178 #ifdef CONFIG_NET_CLS_IND
 179                                 if (!tcf_match_indev(skb, n->ifindex)) {
 180                                         n = rcu_dereference_bh(n->next);
 181                                         goto next_knode;
 182                                 }
 183 #endif
 184 #ifdef CONFIG_CLS_U32_PERF
 185                                 __this_cpu_inc(n->pf->rhit);
 186 #endif
 187                                 r = tcf_exts_exec(skb, &n->exts, res);
 188                                 if (r < 0) {
 189                                         n = rcu_dereference_bh(n->next);
 190                                         goto next_knode;
 191                                 }
 192
 193                                 return r;
 194                         }
 195                         n = rcu_dereference_bh(n->next);
 196                         goto next_knode;
 197                 }
 198
 199                 /* PUSH */
 200                 if (sdepth >= TC_U32_MAXDEPTH)
 201                         goto deadloop;
 202                 stack[sdepth].knode = n;
 203                 stack[sdepth].off = off;
 204                 sdepth++;
 205
 206                 ht = rcu_dereference_bh(n->ht_down);
 207                 sel = 0;
 208                 if (ht->divisor) {
 209                         __be32 *data, hdata;
 210
 211                         data = skb_header_pointer(skb, off + n->sel.hoff, 4,
 212                                                   &hdata);
 213                         if (!data)
 214                                 goto out;
 215                         sel = ht->divisor & u32_hash_fold(*data, &n->sel,
 216                                                           n->fshift);
 217                 }
 218                 if (!(n->sel.flags & (TC_U32_VAROFFSET | TC_U32_OFFSET | TC_U32_EAT)))
 219                         goto next_ht;
 220
 221                 if (n->sel.flags & (TC_U32_OFFSET | TC_U32_VAROFFSET)) {
 222                         off2 = n->sel.off + 3;
 223                         if (n->sel.flags & TC_U32_VAROFFSET) {
 224                                 __be16 *data, hdata;
 225
 226                                 data = skb_header_pointer(skb,
 227                                                           off + n->sel.offoff,
 228                                                           2, &hdata);
 229                                 if (!data)
 230                                         goto out;
 231                                 off2 += ntohs(n->sel.offmask & *data) >>
 232                                         n->sel.offshift;
 233                         }
 234                         off2 &= ~3;
 235                 }
 236                 if (n->sel.flags & TC_U32_EAT) {
 237                         off += off2;
 238                         off2 = 0;
 239                 }
 240
 241                 if (off < skb->len)
 242                         goto next_ht;
 243         }
 244
 245         /* POP */
 246         if (sdepth--) {
 247                 n = stack[sdepth].knode;
 248                 ht = rcu_dereference_bh(n->ht_up);
 249                 off = stack[sdepth].off;
 250                 goto check_terminal;
 251         }
 252 out:
 253         return -1;
 254
 255 deadloop:
 256         net_warn_ratelimited("cls_u32: dead loop\n");
 257         return -1;
 258 }
 259
 260 static struct tc_u_hnode *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 *u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 274 {
 275         unsigned int sel;
 276         struct tc_u_knode *n = NULL;
 277
 278         sel = TC_U32_HASH(handle);
 279         if (sel > ht->divisor)
 280                 goto out;
 281
 282         for (n = rtnl_dereference(ht->ht[sel]);
 283              n;
 284              n = rtnl_dereference(n->next))
 285                 if (n->handle == handle)
 286                         break;
 287 out:
 288         return n;
 289 }
 290
 291
 292 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 293 {
 294         struct tc_u_hnode *ht;
 295         struct tc_u_common *tp_c = tp->data;
 296
 297         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 298                 ht = rtnl_dereference(tp->root);
 299         else
 300                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 301
 302         if (!ht)
 303                 return 0;
 304
 305         if (TC_U32_KEY(handle) == 0)
 306                 return (unsigned long)ht;
 307
 308         return (unsigned long)u32_lookup_key(ht, handle);
 309 }
 310
 311 static u32 gen_new_htid(struct tc_u_common *tp_c)
 312 {
 313         int i = 0x800;
 314
 315         /* hgenerator only used inside rtnl lock it is safe to increment
 316          * without read _copy_ update semantics
 317          */
 318         do {
 319                 if (++tp_c->hgenerator == 0x7FF)
 320                         tp_c->hgenerator = 1;
 321         } while (--i > 0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 322
 323         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 324 }
 325
 326 static int u32_init(struct tcf_proto *tp)
 327 {
 328         struct tc_u_hnode *root_ht;
 329         struct tc_u_common *tp_c;
 330
 331         tp_c = tp->q->u32_node;
 332
 333         root_ht = kzalloc(sizeof(*root_ht), GFP_KERNEL);
 334         if (root_ht == NULL)
 335                 return -ENOBUFS;
 336
 337         root_ht->refcnt++;
 338         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 339         root_ht->prio = tp->prio;
 340
 341         if (tp_c == NULL) {
 342                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 343                 if (tp_c == NULL) {
 344                         kfree(root_ht);
 345                         return -ENOBUFS;
 346                 }
 347                 tp_c->q = tp->q;
 348                 tp->q->u32_node = tp_c;
 349         }
 350
 351         tp_c->refcnt++;
 352         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 353         rcu_assign_pointer(tp_c->hlist, root_ht);
 354         root_ht->tp_c = tp_c;
 355
 356         rcu_assign_pointer(tp->root, root_ht);
 357         tp->data = tp_c;
 358         return 0;
 359 }
 360
 361 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n,
 362                            bool free_pf)
 363 {
 364         tcf_exts_destroy(&n->exts);
 365         if (n->ht_down)
 366                 n->ht_down->refcnt--;
 367 #ifdef CONFIG_CLS_U32_PERF
 368         if (free_pf)
 369                 free_percpu(n->pf);
 370 #endif
 371 #ifdef CONFIG_CLS_U32_MARK
 372         if (free_pf)
 373                 free_percpu(n->pcpu_success);
 374 #endif
 375         kfree(n);
 376         return 0;
 377 }
 378
 379 /* u32_delete_key_rcu should be called when free'ing a copied
 380  * version of a tc_u_knode obtained from u32_init_knode(). When
 381  * copies are obtained from u32_init_knode() the statistics are
 382  * shared between the old and new copies to allow readers to
 383  * continue to update the statistics during the copy. To support
 384  * this the u32_delete_key_rcu variant does not free the percpu
 385  * statistics.
 386  */
 387 static void u32_delete_key_rcu(struct rcu_head *rcu)
 388 {
 389         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 390
 391         u32_destroy_key(key->tp, key, false);
 392 }
 393
 394 /* u32_delete_key_freepf_rcu is the rcu callback variant
 395  * that free's the entire structure including the statistics
 396  * percpu variables. Only use this if the key is not a copy
 397  * returned by u32_init_knode(). See u32_delete_key_rcu()
 398  * for the variant that should be used with keys return from
 399  * u32_init_knode()
 400  */
 401 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 402 {
 403         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 404
 405         u32_destroy_key(key->tp, key, true);
 406 }
 407
 408 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 409 {
 410         struct tc_u_knode __rcu **kp;
 411         struct tc_u_knode *pkp;
 412         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 413
 414         if (ht) {
 415                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 416                 for (pkp = rtnl_dereference(*kp); pkp;
 417                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 418                         if (pkp == key) {
 419                                 RCU_INIT_POINTER(*kp, key->next);
 420
 421                                 tcf_unbind_filter(tp, &key->res);
 422                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 423                                 return 0;
 424                         }
 425                 }
 426         }
 427         WARN_ON(1);
 428         return 0;
 429 }
 430
 431 static void u32_remove_hw_knode(struct tcf_proto *tp, u32 handle)
 432 {
 433         struct net_device *dev = tp->q->dev_queue->dev;
 434         struct tc_cls_u32_offload u32_offload = {0};
 435         struct tc_to_netdev offload;
 436
 437         offload.cls_u32 = &u32_offload;
 438
 439         if (tc_should_offload(dev, tp, 0)) {
 440                 offload.cls_u32->command = TC_CLSU32_DELETE_KNODE;
 441                 offload.cls_u32->knode.handle = handle;
 442                 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32,
 443                                               tp->q->handle, tp->chain->index,
 444                                               tp->protocol, &offload);
 445         }
 446 }
 447
 448 static int u32_replace_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h,
 449                                 u32 flags)
 450 {
 451         struct net_device *dev = tp->q->dev_queue->dev;
 452         struct tc_cls_u32_offload u32_offload = {0};
 453         struct tc_to_netdev offload;
 454         int err;
 455
 456         if (!tc_should_offload(dev, tp, flags))
 457                 return tc_skip_sw(flags) ? -EINVAL : 0;
 458
 459         offload.cls_u32 = &u32_offload;
 460
 461         offload.cls_u32->command = TC_CLSU32_NEW_HNODE;
 462         offload.cls_u32->hnode.divisor = h->divisor;
 463         offload.cls_u32->hnode.handle = h->handle;
 464         offload.cls_u32->hnode.prio = h->prio;
 465
 466         err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, tp->q->handle,
 467                                             tp->chain->index, tp->protocol,
 468                                             &offload);
 469         if (tc_skip_sw(flags))
 470                 return err;
 471
 472         return 0;
 473 }
 474
 475 static void u32_clear_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h)
 476 {
 477         struct net_device *dev = tp->q->dev_queue->dev;
 478         struct tc_cls_u32_offload u32_offload = {0};
 479         struct tc_to_netdev offload;
 480
 481         offload.cls_u32 = &u32_offload;
 482
 483         if (tc_should_offload(dev, tp, 0)) {
 484                 offload.cls_u32->command = TC_CLSU32_DELETE_HNODE;
 485                 offload.cls_u32->hnode.divisor = h->divisor;
 486                 offload.cls_u32->hnode.handle = h->handle;
 487                 offload.cls_u32->hnode.prio = h->prio;
 488
 489                 dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32,
 490                                               tp->q->handle, tp->chain->index,
 491                                               tp->protocol, &offload);
 492         }
 493 }
 494
 495 static int u32_replace_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 496                                 u32 flags)
 497 {
 498         struct net_device *dev = tp->q->dev_queue->dev;
 499         struct tc_cls_u32_offload u32_offload = {0};
 500         struct tc_to_netdev offload;
 501         int err;
 502
 503         offload.cls_u32 = &u32_offload;
 504
 505         if (!tc_should_offload(dev, tp, flags))
 506                 return tc_skip_sw(flags) ? -EINVAL : 0;
 507
 508         offload.cls_u32->command = TC_CLSU32_REPLACE_KNODE;
 509         offload.cls_u32->knode.handle = n->handle;
 510         offload.cls_u32->knode.fshift = n->fshift;
 511 #ifdef CONFIG_CLS_U32_MARK
 512         offload.cls_u32->knode.val = n->val;
 513         offload.cls_u32->knode.mask = n->mask;
 514 #else
 515         offload.cls_u32->knode.val = 0;
 516         offload.cls_u32->knode.mask = 0;
 517 #endif
 518         offload.cls_u32->knode.sel = &n->sel;
 519         offload.cls_u32->knode.exts = &n->exts;
 520         if (n->ht_down)
 521                 offload.cls_u32->knode.link_handle = n->ht_down->handle;
 522
 523         err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSU32, tp->q->handle,
 524                                             tp->chain->index, tp->protocol,
 525                                             &offload);
 526
 527         if (!err)
 528                 n->flags |= TCA_CLS_FLAGS_IN_HW;
 529
 530         if (tc_skip_sw(flags))
 531                 return err;
 532
 533         return 0;
 534 }
 535
 536 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 537 {
 538         struct tc_u_knode *n;
 539         unsigned int h;
 540
 541         for (h = 0; h <= ht->divisor; h++) {
 542                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 543                         RCU_INIT_POINTER(ht->ht[h],
 544                                          rtnl_dereference(n->next));
 545                         tcf_unbind_filter(tp, &n->res);
 546                         u32_remove_hw_knode(tp, n->handle);
 547                         call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 548                 }
 549         }
 550 }
 551
 552 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 553 {
 554         struct tc_u_common *tp_c = tp->data;
 555         struct tc_u_hnode __rcu **hn;
 556         struct tc_u_hnode *phn;
 557
 558         WARN_ON(ht->refcnt);
 559
 560         u32_clear_hnode(tp, ht);
 561
 562         hn = &tp_c->hlist;
 563         for (phn = rtnl_dereference(*hn);
 564              phn;
 565              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 566                 if (phn == ht) {
 567                         u32_clear_hw_hnode(tp, ht);
 568                         RCU_INIT_POINTER(*hn, ht->next);
 569                         kfree_rcu(ht, rcu);
 570                         return 0;
 571                 }
 572         }
 573
 574         return -ENOENT;
 575 }
 576
 577 static bool ht_empty(struct tc_u_hnode *ht)
 578 {
 579         unsigned int h;
 580
 581         for (h = 0; h <= ht->divisor; h++)
 582                 if (rcu_access_pointer(ht->ht[h]))
 583                         return false;
 584
 585         return true;
 586 }
 587
 588 static void u32_destroy(struct tcf_proto *tp)
 589 {
 590         struct tc_u_common *tp_c = tp->data;
 591         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 592
 593         WARN_ON(root_ht == NULL);
 594
 595         if (root_ht && --root_ht->refcnt == 0)
 596                 u32_destroy_hnode(tp, root_ht);
 597
 598         if (--tp_c->refcnt == 0) {
 599                 struct tc_u_hnode *ht;
 600
 601                 tp->q->u32_node = NULL;
 602
 603                 for (ht = rtnl_dereference(tp_c->hlist);
 604                      ht;
 605                      ht = rtnl_dereference(ht->next)) {
 606                         ht->refcnt--;
 607                         u32_clear_hnode(tp, ht);
 608                 }
 609
 610                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 611                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 612                         kfree_rcu(ht, rcu);
 613                 }
 614
 615                 kfree(tp_c);
 616         }
 617
 618         tp->data = NULL;
 619 }
 620
 621 static int u32_delete(struct tcf_proto *tp, unsigned long arg, bool *last)
 622 {
 623         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 624         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 625         struct tc_u_common *tp_c = tp->data;
 626         int ret = 0;
 627
 628         if (ht == NULL)
 629                 goto out;
 630
 631         if (TC_U32_KEY(ht->handle)) {
 632                 u32_remove_hw_knode(tp, ht->handle);
 633                 ret = u32_delete_key(tp, (struct tc_u_knode *)ht);
 634                 goto out;
 635         }
 636
 637         if (root_ht == ht)
 638                 return -EINVAL;
 639
 640         if (ht->refcnt == 1) {
 641                 ht->refcnt--;
 642                 u32_destroy_hnode(tp, ht);
 643         } else {
 644                 return -EBUSY;
 645         }
 646
 647 out:
 648         *last = true;
 649         if (root_ht) {
 650                 if (root_ht->refcnt > 1) {
 651                         *last = false;
 652                         goto ret;
 653                 }
 654                 if (root_ht->refcnt == 1) {
 655                         if (!ht_empty(root_ht)) {
 656                                 *last = false;
 657                                 goto ret;
 658                         }
 659                 }
 660         }
 661
 662         if (tp_c->refcnt > 1) {
 663                 *last = false;
 664                 goto ret;
 665         }
 666
 667         if (tp_c->refcnt == 1) {
 668                 struct tc_u_hnode *ht;
 669
 670                 for (ht = rtnl_dereference(tp_c->hlist);
 671                      ht;
 672                      ht = rtnl_dereference(ht->next))
 673                         if (!ht_empty(ht)) {
 674                                 *last = false;
 675                                 break;
 676                         }
 677         }
 678
 679 ret:
 680         return ret;
 681 }
 682
 683 #define NR_U32_NODE (1<<12)
 684 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 685 {
 686         struct tc_u_knode *n;
 687         unsigned long i;
 688         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 689                                         GFP_KERNEL);
 690         if (!bitmap)
 691                 return handle | 0xFFF;
 692
 693         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 694              n;
 695              n = rtnl_dereference(n->next))
 696                 set_bit(TC_U32_NODE(n->handle), bitmap);
 697
 698         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 699         if (i >= NR_U32_NODE)
 700                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 701
 702         kfree(bitmap);
 703         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 704 }
 705
 706 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 707         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 708         [TCA_U32_HASH]          = { .type = NLA_U32 },
 709         [TCA_U32_LINK]          = { .type = NLA_U32 },
 710         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 711         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 712         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 713         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 714         [TCA_U32_FLAGS]         = { .type = NLA_U32 },
 715 };
 716
 717 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 718                          unsigned long base, struct tc_u_hnode *ht,
 719                          struct tc_u_knode *n, struct nlattr **tb,
 720                          struct nlattr *est, bool ovr)
 721 {
 722         int err;
 723
 724         err = tcf_exts_validate(net, tp, tb, est, &n->exts, ovr);
 725         if (err < 0)
 726                 return err;
 727
 728         if (tb[TCA_U32_LINK]) {
 729                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 730                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 731
 732                 if (TC_U32_KEY(handle))
 733                         return -EINVAL;
 734
 735                 if (handle) {
 736                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 737
 738                         if (ht_down == NULL)
 739                                 return -EINVAL;
 740                         ht_down->refcnt++;
 741                 }
 742
 743                 ht_old = rtnl_dereference(n->ht_down);
 744                 rcu_assign_pointer(n->ht_down, ht_down);
 745
 746                 if (ht_old)
 747                         ht_old->refcnt--;
 748         }
 749         if (tb[TCA_U32_CLASSID]) {
 750                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 751                 tcf_bind_filter(tp, &n->res, base);
 752         }
 753
 754 #ifdef CONFIG_NET_CLS_IND
 755         if (tb[TCA_U32_INDEV]) {
 756                 int ret;
 757                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 758                 if (ret < 0)
 759                         return -EINVAL;
 760                 n->ifindex = ret;
 761         }
 762 #endif
 763         return 0;
 764 }
 765
 766 static void u32_replace_knode(struct tcf_proto *tp, struct tc_u_common *tp_c,
 767                               struct tc_u_knode *n)
 768 {
 769         struct tc_u_knode __rcu **ins;
 770         struct tc_u_knode *pins;
 771         struct tc_u_hnode *ht;
 772
 773         if (TC_U32_HTID(n->handle) == TC_U32_ROOT)
 774                 ht = rtnl_dereference(tp->root);
 775         else
 776                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(n->handle));
 777
 778         ins = &ht->ht[TC_U32_HASH(n->handle)];
 779
 780         /* The node must always exist for it to be replaced if this is not the
 781          * case then something went very wrong elsewhere.
 782          */
 783         for (pins = rtnl_dereference(*ins); ;
 784              ins = &pins->next, pins = rtnl_dereference(*ins))
 785                 if (pins->handle == n->handle)
 786                         break;
 787
 788         RCU_INIT_POINTER(n->next, pins->next);
 789         rcu_assign_pointer(*ins, n);
 790 }
 791
 792 static struct tc_u_knode *u32_init_knode(struct tcf_proto *tp,
 793                                          struct tc_u_knode *n)
 794 {
 795         struct tc_u_knode *new;
 796         struct tc_u32_sel *s = &n->sel;
 797
 798         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 799                       GFP_KERNEL);
 800
 801         if (!new)
 802                 return NULL;
 803
 804         RCU_INIT_POINTER(new->next, n->next);
 805         new->handle = n->handle;
 806         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 807
 808 #ifdef CONFIG_NET_CLS_IND
 809         new->ifindex = n->ifindex;
 810 #endif
 811         new->fshift = n->fshift;
 812         new->res = n->res;
 813         new->flags = n->flags;
 814         RCU_INIT_POINTER(new->ht_down, n->ht_down);
 815
 816         /* bump reference count as long as we hold pointer to structure */
 817         if (new->ht_down)
 818                 new->ht_down->refcnt++;
 819
 820 #ifdef CONFIG_CLS_U32_PERF
 821         /* Statistics may be incremented by readers during update
 822          * so we must keep them in tact. When the node is later destroyed
 823          * a special destroy call must be made to not free the pf memory.
 824          */
 825         new->pf = n->pf;
 826 #endif
 827
 828 #ifdef CONFIG_CLS_U32_MARK
 829         new->val = n->val;
 830         new->mask = n->mask;
 831         /* Similarly success statistics must be moved as pointers */
 832         new->pcpu_success = n->pcpu_success;
 833 #endif
 834         new->tp = tp;
 835         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 836
 837         if (tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE)) {
 838                 kfree(new);
 839                 return NULL;
 840         }
 841
 842         return new;
 843 }
 844
 845 static int u32_change(struct net *net, struct sk_buff *in_skb,
 846                       struct tcf_proto *tp, unsigned long base, u32 handle,
 847                       struct nlattr **tca, unsigned long *arg, bool ovr)
 848 {
 849         struct tc_u_common *tp_c = tp->data;
 850         struct tc_u_hnode *ht;
 851         struct tc_u_knode *n;
 852         struct tc_u32_sel *s;
 853         struct nlattr *opt = tca[TCA_OPTIONS];
 854         struct nlattr *tb[TCA_U32_MAX + 1];
 855         u32 htid, flags = 0;
 856         int err;
 857 #ifdef CONFIG_CLS_U32_PERF
 858         size_t size;
 859 #endif
 860
 861         if (opt == NULL)
 862                 return handle ? -EINVAL : 0;
 863
 864         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy, NULL);
 865         if (err < 0)
 866                 return err;
 867
 868         if (tb[TCA_U32_FLAGS]) {
 869                 flags = nla_get_u32(tb[TCA_U32_FLAGS]);
 870                 if (!tc_flags_valid(flags))
 871                         return -EINVAL;
 872         }
 873
 874         n = (struct tc_u_knode *)*arg;
 875         if (n) {
 876                 struct tc_u_knode *new;
 877
 878                 if (TC_U32_KEY(n->handle) == 0)
 879                         return -EINVAL;
 880
 881                 if (n->flags != flags)
 882                         return -EINVAL;
 883
 884                 new = u32_init_knode(tp, n);
 885                 if (!new)
 886                         return -ENOMEM;
 887
 888                 err = u32_set_parms(net, tp, base,
 889                                     rtnl_dereference(n->ht_up), new, tb,
 890                                     tca[TCA_RATE], ovr);
 891
 892                 if (err) {
 893                         u32_destroy_key(tp, new, false);
 894                         return err;
 895                 }
 896
 897                 err = u32_replace_hw_knode(tp, new, flags);
 898                 if (err) {
 899                         u32_destroy_key(tp, new, false);
 900                         return err;
 901                 }
 902
 903                 if (!tc_in_hw(new->flags))
 904                         new->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
 905
 906                 u32_replace_knode(tp, tp_c, new);
 907                 tcf_unbind_filter(tp, &n->res);
 908                 call_rcu(&n->rcu, u32_delete_key_rcu);
 909                 return 0;
 910         }
 911
 912         if (tb[TCA_U32_DIVISOR]) {
 913                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 914
 915                 if (--divisor > 0x100)
 916                         return -EINVAL;
 917                 if (TC_U32_KEY(handle))
 918                         return -EINVAL;
 919                 if (handle == 0) {
 920                         handle = gen_new_htid(tp->data);
 921                         if (handle == 0)
 922                                 return -ENOMEM;
 923                 }
 924                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 925                 if (ht == NULL)
 926                         return -ENOBUFS;
 927                 ht->tp_c = tp_c;
 928                 ht->refcnt = 1;
 929                 ht->divisor = divisor;
 930                 ht->handle = handle;
 931                 ht->prio = tp->prio;
 932
 933                 err = u32_replace_hw_hnode(tp, ht, flags);
 934                 if (err) {
 935                         kfree(ht);
 936                         return err;
 937                 }
 938
 939                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 940                 rcu_assign_pointer(tp_c->hlist, ht);
 941                 *arg = (unsigned long)ht;
 942
 943                 return 0;
 944         }
 945
 946         if (tb[TCA_U32_HASH]) {
 947                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 948                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 949                         ht = rtnl_dereference(tp->root);
 950                         htid = ht->handle;
 951                 } else {
 952                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 953                         if (ht == NULL)
 954                                 return -EINVAL;
 955                 }
 956         } else {
 957                 ht = rtnl_dereference(tp->root);
 958                 htid = ht->handle;
 959         }
 960
 961         if (ht->divisor < TC_U32_HASH(htid))
 962                 return -EINVAL;
 963
 964         if (handle) {
 965                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 966                         return -EINVAL;
 967                 handle = htid | TC_U32_NODE(handle);
 968         } else
 969                 handle = gen_new_kid(ht, htid);
 970
 971         if (tb[TCA_U32_SEL] == NULL)
 972                 return -EINVAL;
 973
 974         s = nla_data(tb[TCA_U32_SEL]);
 975
 976         n = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 977         if (n == NULL)
 978                 return -ENOBUFS;
 979
 980 #ifdef CONFIG_CLS_U32_PERF
 981         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 982         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 983         if (!n->pf) {
 984                 kfree(n);
 985                 return -ENOBUFS;
 986         }
 987 #endif
 988
 989         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 990         RCU_INIT_POINTER(n->ht_up, ht);
 991         n->handle = handle;
 992         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 993         n->flags = flags;
 994         n->tp = tp;
 995
 996         err = tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 997         if (err < 0)
 998                 goto errout;
 999
1000 #ifdef CONFIG_CLS_U32_MARK
1001         n->pcpu_success = alloc_percpu(u32);
1002         if (!n->pcpu_success) {
1003                 err = -ENOMEM;
1004                 goto errout;
1005         }
1006
1007         if (tb[TCA_U32_MARK]) {
1008                 struct tc_u32_mark *mark;
1009
1010                 mark = nla_data(tb[TCA_U32_MARK]);
1011                 n->val = mark->val;
1012                 n->mask = mark->mask;
1013         }
1014 #endif
1015
1016         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
1017         if (err == 0) {
1018                 struct tc_u_knode __rcu **ins;
1019                 struct tc_u_knode *pins;
1020
1021                 err = u32_replace_hw_knode(tp, n, flags);
1022                 if (err)
1023                         goto errhw;
1024
1025                 if (!tc_in_hw(n->flags))
1026                         n->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1027
1028                 ins = &ht->ht[TC_U32_HASH(handle)];
1029                 for (pins = rtnl_dereference(*ins); pins;
1030                      ins = &pins->next, pins = rtnl_dereference(*ins))
1031                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
1032                                 break;
1033
1034                 RCU_INIT_POINTER(n->next, pins);
1035                 rcu_assign_pointer(*ins, n);
1036                 *arg = (unsigned long)n;
1037                 return 0;
1038         }
1039
1040 errhw:
1041 #ifdef CONFIG_CLS_U32_MARK
1042         free_percpu(n->pcpu_success);
1043 #endif
1044
1045 errout:
1046         tcf_exts_destroy(&n->exts);
1047 #ifdef CONFIG_CLS_U32_PERF
1048         free_percpu(n->pf);
1049 #endif
1050         kfree(n);
1051         return err;
1052 }
1053
1054 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1055 {
1056         struct tc_u_common *tp_c = tp->data;
1057         struct tc_u_hnode *ht;
1058         struct tc_u_knode *n;
1059         unsigned int h;
1060
1061         if (arg->stop)
1062                 return;
1063
1064         for (ht = rtnl_dereference(tp_c->hlist);
1065              ht;
1066              ht = rtnl_dereference(ht->next)) {
1067                 if (ht->prio != tp->prio)
1068                         continue;
1069                 if (arg->count >= arg->skip) {
1070                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
1071                                 arg->stop = 1;
1072                                 return;
1073                         }
1074                 }
1075                 arg->count++;
1076                 for (h = 0; h <= ht->divisor; h++) {
1077                         for (n = rtnl_dereference(ht->ht[h]);
1078                              n;
1079                              n = rtnl_dereference(n->next)) {
1080                                 if (arg->count < arg->skip) {
1081                                         arg->count++;
1082                                         continue;
1083                                 }
1084                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
1085                                         arg->stop = 1;
1086                                         return;
1087                                 }
1088                                 arg->count++;
1089                         }
1090                 }
1091         }
1092 }
1093
1094 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
1095                     struct sk_buff *skb, struct tcmsg *t)
1096 {
1097         struct tc_u_knode *n = (struct tc_u_knode *)fh;
1098         struct tc_u_hnode *ht_up, *ht_down;
1099         struct nlattr *nest;
1100
1101         if (n == NULL)
1102                 return skb->len;
1103
1104         t->tcm_handle = n->handle;
1105
1106         nest = nla_nest_start(skb, TCA_OPTIONS);
1107         if (nest == NULL)
1108                 goto nla_put_failure;
1109
1110         if (TC_U32_KEY(n->handle) == 0) {
1111                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
1112                 u32 divisor = ht->divisor + 1;
1113
1114                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
1115                         goto nla_put_failure;
1116         } else {
1117 #ifdef CONFIG_CLS_U32_PERF
1118                 struct tc_u32_pcnt *gpf;
1119                 int cpu;
1120 #endif
1121
1122                 if (nla_put(skb, TCA_U32_SEL,
1123                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
1124                             &n->sel))
1125                         goto nla_put_failure;
1126
1127                 ht_up = rtnl_dereference(n->ht_up);
1128                 if (ht_up) {
1129                         u32 htid = n->handle & 0xFFFFF000;
1130                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
1131                                 goto nla_put_failure;
1132                 }
1133                 if (n->res.classid &&
1134                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
1135                         goto nla_put_failure;
1136
1137                 ht_down = rtnl_dereference(n->ht_down);
1138                 if (ht_down &&
1139                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
1140                         goto nla_put_failure;
1141
1142                 if (n->flags && nla_put_u32(skb, TCA_U32_FLAGS, n->flags))
1143                         goto nla_put_failure;
1144
1145 #ifdef CONFIG_CLS_U32_MARK
1146                 if ((n->val || n->mask)) {
1147                         struct tc_u32_mark mark = {.val = n->val,
1148                                                    .mask = n->mask,
1149                                                    .success = 0};
1150                         int cpum;
1151
1152                         for_each_possible_cpu(cpum) {
1153                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
1154
1155                                 mark.success += cnt;
1156                         }
1157
1158                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
1159                                 goto nla_put_failure;
1160                 }
1161 #endif
1162
1163                 if (tcf_exts_dump(skb, &n->exts) < 0)
1164                         goto nla_put_failure;
1165
1166 #ifdef CONFIG_NET_CLS_IND
1167                 if (n->ifindex) {
1168                         struct net_device *dev;
1169                         dev = __dev_get_by_index(net, n->ifindex);
1170                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
1171                                 goto nla_put_failure;
1172                 }
1173 #endif
1174 #ifdef CONFIG_CLS_U32_PERF
1175                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1176                               n->sel.nkeys * sizeof(u64),
1177                               GFP_KERNEL);
1178                 if (!gpf)
1179                         goto nla_put_failure;
1180
1181                 for_each_possible_cpu(cpu) {
1182                         int i;
1183                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1184
1185                         gpf->rcnt += pf->rcnt;
1186                         gpf->rhit += pf->rhit;
1187                         for (i = 0; i < n->sel.nkeys; i++)
1188                                 gpf->kcnts[i] += pf->kcnts[i];
1189                 }
1190
1191                 if (nla_put_64bit(skb, TCA_U32_PCNT,
1192                                   sizeof(struct tc_u32_pcnt) +
1193                                   n->sel.nkeys * sizeof(u64),
1194                                   gpf, TCA_U32_PAD)) {
1195                         kfree(gpf);
1196                         goto nla_put_failure;
1197                 }
1198                 kfree(gpf);
1199 #endif
1200         }
1201
1202         nla_nest_end(skb, nest);
1203
1204         if (TC_U32_KEY(n->handle))
1205                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1206                         goto nla_put_failure;
1207         return skb->len;
1208
1209 nla_put_failure:
1210         nla_nest_cancel(skb, nest);
1211         return -1;
1212 }
1213
1214 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1215         .kind           =       "u32",
1216         .classify       =       u32_classify,
1217         .init           =       u32_init,
1218         .destroy        =       u32_destroy,
1219         .get            =       u32_get,
1220         .change         =       u32_change,
1221         .delete         =       u32_delete,
1222         .walk           =       u32_walk,
1223         .dump           =       u32_dump,
1224         .owner          =       THIS_MODULE,
1225 };
1226
1227 static int __init init_u32(void)
1228 {
1229         pr_info("u32 classifier\n");
1230 #ifdef CONFIG_CLS_U32_PERF
1231         pr_info("    Performance counters on\n");
1232 #endif
1233 #ifdef CONFIG_NET_CLS_IND
1234         pr_info("    input device check on\n");
1235 #endif
1236 #ifdef CONFIG_NET_CLS_ACT
1237         pr_info("    Actions configured\n");
1238 #endif
1239         return register_tcf_proto_ops(&cls_u32_ops);
1240 }
1241
1242 static void __exit exit_u32(void)
1243 {
1244         unregister_tcf_proto_ops(&cls_u32_ops);
1245 }
1246
1247 module_init(init_u32)
1248 module_exit(exit_u32)
1249 MODULE_LICENSE("GPL");