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->divisor = 0;
 338         root_ht->refcnt++;
 339         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 340         root_ht->prio = tp->prio;
 341
 342         if (tp_c == NULL) {
 343                 tp_c = kzalloc(sizeof(*tp_c), GFP_KERNEL);
 344                 if (tp_c == NULL) {
 345                         kfree(root_ht);
 346                         return -ENOBUFS;
 347                 }
 348                 tp_c->q = tp->q;
 349                 tp->q->u32_node = tp_c;
 350         }
 351
 352         tp_c->refcnt++;
 353         RCU_INIT_POINTER(root_ht->next, tp_c->hlist);
 354         rcu_assign_pointer(tp_c->hlist, root_ht);
 355         root_ht->tp_c = tp_c;
 356
 357         rcu_assign_pointer(tp->root, root_ht);
 358         tp->data = tp_c;
 359         return 0;
 360 }
 361
 362 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n,
 363                            bool free_pf)
 364 {
 365         tcf_exts_destroy(&n->exts);
 366         if (n->ht_down)
 367                 n->ht_down->refcnt--;
 368 #ifdef CONFIG_CLS_U32_PERF
 369         if (free_pf)
 370                 free_percpu(n->pf);
 371 #endif
 372 #ifdef CONFIG_CLS_U32_MARK
 373         if (free_pf)
 374                 free_percpu(n->pcpu_success);
 375 #endif
 376         kfree(n);
 377         return 0;
 378 }
 379
 380 /* u32_delete_key_rcu should be called when free'ing a copied
 381  * version of a tc_u_knode obtained from u32_init_knode(). When
 382  * copies are obtained from u32_init_knode() the statistics are
 383  * shared between the old and new copies to allow readers to
 384  * continue to update the statistics during the copy. To support
 385  * this the u32_delete_key_rcu variant does not free the percpu
 386  * statistics.
 387  */
 388 static void u32_delete_key_rcu(struct rcu_head *rcu)
 389 {
 390         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 391
 392         u32_destroy_key(key->tp, key, false);
 393 }
 394
 395 /* u32_delete_key_freepf_rcu is the rcu callback variant
 396  * that free's the entire structure including the statistics
 397  * percpu variables. Only use this if the key is not a copy
 398  * returned by u32_init_knode(). See u32_delete_key_rcu()
 399  * for the variant that should be used with keys return from
 400  * u32_init_knode()
 401  */
 402 static void u32_delete_key_freepf_rcu(struct rcu_head *rcu)
 403 {
 404         struct tc_u_knode *key = container_of(rcu, struct tc_u_knode, rcu);
 405
 406         u32_destroy_key(key->tp, key, true);
 407 }
 408
 409 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode *key)
 410 {
 411         struct tc_u_knode __rcu **kp;
 412         struct tc_u_knode *pkp;
 413         struct tc_u_hnode *ht = rtnl_dereference(key->ht_up);
 414
 415         if (ht) {
 416                 kp = &ht->ht[TC_U32_HASH(key->handle)];
 417                 for (pkp = rtnl_dereference(*kp); pkp;
 418                      kp = &pkp->next, pkp = rtnl_dereference(*kp)) {
 419                         if (pkp == key) {
 420                                 RCU_INIT_POINTER(*kp, key->next);
 421
 422                                 tcf_unbind_filter(tp, &key->res);
 423                                 call_rcu(&key->rcu, u32_delete_key_freepf_rcu);
 424                                 return 0;
 425                         }
 426                 }
 427         }
 428         WARN_ON(1);
 429         return 0;
 430 }
 431
 432 static void u32_remove_hw_knode(struct tcf_proto *tp, u32 handle)
 433 {
 434         struct net_device *dev = tp->q->dev_queue->dev;
 435         struct tc_cls_u32_offload u32_offload = {0};
 436         struct tc_to_netdev offload;
 437
 438         offload.type = TC_SETUP_CLSU32;
 439         offload.cls_u32 = &u32_offload;
 440
 441         if (tc_should_offload(dev, tp, 0)) {
 442                 offload.cls_u32->command = TC_CLSU32_DELETE_KNODE;
 443                 offload.cls_u32->knode.handle = handle;
 444                 dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 445                                               tp->protocol, &offload);
 446         }
 447 }
 448
 449 static int u32_replace_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h,
 450                                 u32 flags)
 451 {
 452         struct net_device *dev = tp->q->dev_queue->dev;
 453         struct tc_cls_u32_offload u32_offload = {0};
 454         struct tc_to_netdev offload;
 455         int err;
 456
 457         if (!tc_should_offload(dev, tp, flags))
 458                 return tc_skip_sw(flags) ? -EINVAL : 0;
 459
 460         offload.type = TC_SETUP_CLSU32;
 461         offload.cls_u32 = &u32_offload;
 462
 463         offload.cls_u32->command = TC_CLSU32_NEW_HNODE;
 464         offload.cls_u32->hnode.divisor = h->divisor;
 465         offload.cls_u32->hnode.handle = h->handle;
 466         offload.cls_u32->hnode.prio = h->prio;
 467
 468         err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 469                                             tp->protocol, &offload);
 470         if (tc_skip_sw(flags))
 471                 return err;
 472
 473         return 0;
 474 }
 475
 476 static void u32_clear_hw_hnode(struct tcf_proto *tp, struct tc_u_hnode *h)
 477 {
 478         struct net_device *dev = tp->q->dev_queue->dev;
 479         struct tc_cls_u32_offload u32_offload = {0};
 480         struct tc_to_netdev offload;
 481
 482         offload.type = TC_SETUP_CLSU32;
 483         offload.cls_u32 = &u32_offload;
 484
 485         if (tc_should_offload(dev, tp, 0)) {
 486                 offload.cls_u32->command = TC_CLSU32_DELETE_HNODE;
 487                 offload.cls_u32->hnode.divisor = h->divisor;
 488                 offload.cls_u32->hnode.handle = h->handle;
 489                 offload.cls_u32->hnode.prio = h->prio;
 490
 491                 dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 492                                               tp->protocol, &offload);
 493         }
 494 }
 495
 496 static int u32_replace_hw_knode(struct tcf_proto *tp, struct tc_u_knode *n,
 497                                 u32 flags)
 498 {
 499         struct tc_u_hnode *ht = rtnl_dereference(n->ht_down);
 500         struct net_device *dev = tp->q->dev_queue->dev;
 501         struct tc_cls_u32_offload u32_offload = {0};
 502         struct tc_to_netdev offload;
 503         int err;
 504
 505         offload.type = TC_SETUP_CLSU32;
 506         offload.cls_u32 = &u32_offload;
 507
 508         if (!tc_should_offload(dev, tp, flags))
 509                 return tc_skip_sw(flags) ? -EINVAL : 0;
 510
 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 = ht->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         return 0;
 532 }
 533
 534 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 535 {
 536         struct tc_u_knode *n;
 537         unsigned int h;
 538
 539         for (h = 0; h <= ht->divisor; h++) {
 540                 while ((n = rtnl_dereference(ht->ht[h])) != NULL) {
 541                         RCU_INIT_POINTER(ht->ht[h],
 542                                          rtnl_dereference(n->next));
 543                         tcf_unbind_filter(tp, &n->res);
 544                         u32_remove_hw_knode(tp, n->handle);
 545                         call_rcu(&n->rcu, u32_delete_key_freepf_rcu);
 546                 }
 547         }
 548 }
 549
 550 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 551 {
 552         struct tc_u_common *tp_c = tp->data;
 553         struct tc_u_hnode __rcu **hn;
 554         struct tc_u_hnode *phn;
 555
 556         WARN_ON(ht->refcnt);
 557
 558         u32_clear_hnode(tp, ht);
 559
 560         hn = &tp_c->hlist;
 561         for (phn = rtnl_dereference(*hn);
 562              phn;
 563              hn = &phn->next, phn = rtnl_dereference(*hn)) {
 564                 if (phn == ht) {
 565                         u32_clear_hw_hnode(tp, ht);
 566                         RCU_INIT_POINTER(*hn, ht->next);
 567                         kfree_rcu(ht, rcu);
 568                         return 0;
 569                 }
 570         }
 571
 572         return -ENOENT;
 573 }
 574
 575 static bool ht_empty(struct tc_u_hnode *ht)
 576 {
 577         unsigned int h;
 578
 579         for (h = 0; h <= ht->divisor; h++)
 580                 if (rcu_access_pointer(ht->ht[h]))
 581                         return false;
 582
 583         return true;
 584 }
 585
 586 static bool u32_destroy(struct tcf_proto *tp, bool force)
 587 {
 588         struct tc_u_common *tp_c = tp->data;
 589         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 590
 591         WARN_ON(root_ht == NULL);
 592
 593         if (!force) {
 594                 if (root_ht) {
 595                         if (root_ht->refcnt > 1)
 596                                 return false;
 597                         if (root_ht->refcnt == 1) {
 598                                 if (!ht_empty(root_ht))
 599                                         return false;
 600                         }
 601                 }
 602
 603                 if (tp_c->refcnt > 1)
 604                         return false;
 605
 606                 if (tp_c->refcnt == 1) {
 607                         struct tc_u_hnode *ht;
 608
 609                         for (ht = rtnl_dereference(tp_c->hlist);
 610                              ht;
 611                              ht = rtnl_dereference(ht->next))
 612                                 if (!ht_empty(ht))
 613                                         return false;
 614                 }
 615         }
 616
 617         if (root_ht && --root_ht->refcnt == 0)
 618                 u32_destroy_hnode(tp, root_ht);
 619
 620         if (--tp_c->refcnt == 0) {
 621                 struct tc_u_hnode *ht;
 622
 623                 tp->q->u32_node = NULL;
 624
 625                 for (ht = rtnl_dereference(tp_c->hlist);
 626                      ht;
 627                      ht = rtnl_dereference(ht->next)) {
 628                         ht->refcnt--;
 629                         u32_clear_hnode(tp, ht);
 630                 }
 631
 632                 while ((ht = rtnl_dereference(tp_c->hlist)) != NULL) {
 633                         RCU_INIT_POINTER(tp_c->hlist, ht->next);
 634                         kfree_rcu(ht, rcu);
 635                 }
 636
 637                 kfree(tp_c);
 638         }
 639
 640         tp->data = NULL;
 641         return true;
 642 }
 643
 644 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 645 {
 646         struct tc_u_hnode *ht = (struct tc_u_hnode *)arg;
 647         struct tc_u_hnode *root_ht = rtnl_dereference(tp->root);
 648
 649         if (ht == NULL)
 650                 return 0;
 651
 652         if (TC_U32_KEY(ht->handle)) {
 653                 u32_remove_hw_knode(tp, ht->handle);
 654                 return u32_delete_key(tp, (struct tc_u_knode *)ht);
 655         }
 656
 657         if (root_ht == ht)
 658                 return -EINVAL;
 659
 660         if (ht->refcnt == 1) {
 661                 ht->refcnt--;
 662                 u32_destroy_hnode(tp, ht);
 663         } else {
 664                 return -EBUSY;
 665         }
 666
 667         return 0;
 668 }
 669
 670 #define NR_U32_NODE (1<<12)
 671 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 672 {
 673         struct tc_u_knode *n;
 674         unsigned long i;
 675         unsigned long *bitmap = kzalloc(BITS_TO_LONGS(NR_U32_NODE) * sizeof(unsigned long),
 676                                         GFP_KERNEL);
 677         if (!bitmap)
 678                 return handle | 0xFFF;
 679
 680         for (n = rtnl_dereference(ht->ht[TC_U32_HASH(handle)]);
 681              n;
 682              n = rtnl_dereference(n->next))
 683                 set_bit(TC_U32_NODE(n->handle), bitmap);
 684
 685         i = find_next_zero_bit(bitmap, NR_U32_NODE, 0x800);
 686         if (i >= NR_U32_NODE)
 687                 i = find_next_zero_bit(bitmap, NR_U32_NODE, 1);
 688
 689         kfree(bitmap);
 690         return handle | (i >= NR_U32_NODE ? 0xFFF : i);
 691 }
 692
 693 static const struct nla_policy u32_policy[TCA_U32_MAX + 1] = {
 694         [TCA_U32_CLASSID]       = { .type = NLA_U32 },
 695         [TCA_U32_HASH]          = { .type = NLA_U32 },
 696         [TCA_U32_LINK]          = { .type = NLA_U32 },
 697         [TCA_U32_DIVISOR]       = { .type = NLA_U32 },
 698         [TCA_U32_SEL]           = { .len = sizeof(struct tc_u32_sel) },
 699         [TCA_U32_INDEV]         = { .type = NLA_STRING, .len = IFNAMSIZ },
 700         [TCA_U32_MARK]          = { .len = sizeof(struct tc_u32_mark) },
 701         [TCA_U32_FLAGS]         = { .type = NLA_U32 },
 702 };
 703
 704 static int u32_set_parms(struct net *net, struct tcf_proto *tp,
 705                          unsigned long base, struct tc_u_hnode *ht,
 706                          struct tc_u_knode *n, struct nlattr **tb,
 707                          struct nlattr *est, bool ovr)
 708 {
 709         struct tcf_exts e;
 710         int err;
 711
 712         err = tcf_exts_init(&e, TCA_U32_ACT, TCA_U32_POLICE);
 713         if (err < 0)
 714                 return err;
 715         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 716         if (err < 0)
 717                 goto errout;
 718
 719         err = -EINVAL;
 720         if (tb[TCA_U32_LINK]) {
 721                 u32 handle = nla_get_u32(tb[TCA_U32_LINK]);
 722                 struct tc_u_hnode *ht_down = NULL, *ht_old;
 723
 724                 if (TC_U32_KEY(handle))
 725                         goto errout;
 726
 727                 if (handle) {
 728                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 729
 730                         if (ht_down == NULL)
 731                                 goto errout;
 732                         ht_down->refcnt++;
 733                 }
 734
 735                 ht_old = rtnl_dereference(n->ht_down);
 736                 rcu_assign_pointer(n->ht_down, ht_down);
 737
 738                 if (ht_old)
 739                         ht_old->refcnt--;
 740         }
 741         if (tb[TCA_U32_CLASSID]) {
 742                 n->res.classid = nla_get_u32(tb[TCA_U32_CLASSID]);
 743                 tcf_bind_filter(tp, &n->res, base);
 744         }
 745
 746 #ifdef CONFIG_NET_CLS_IND
 747         if (tb[TCA_U32_INDEV]) {
 748                 int ret;
 749                 ret = tcf_change_indev(net, tb[TCA_U32_INDEV]);
 750                 if (ret < 0)
 751                         goto errout;
 752                 n->ifindex = ret;
 753         }
 754 #endif
 755         tcf_exts_change(tp, &n->exts, &e);
 756
 757         return 0;
 758 errout:
 759         tcf_exts_destroy(&e);
 760         return err;
 761 }
 762
 763 static void u32_replace_knode(struct tcf_proto *tp, 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_hnode *ht = rtnl_dereference(n->ht_down);
 793         struct tc_u32_sel *s = &n->sel;
 794         struct tc_u_knode *new;
 795
 796         new = kzalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key),
 797                       GFP_KERNEL);
 798
 799         if (!new)
 800                 return NULL;
 801
 802         RCU_INIT_POINTER(new->next, n->next);
 803         new->handle = n->handle;
 804         RCU_INIT_POINTER(new->ht_up, n->ht_up);
 805
 806 #ifdef CONFIG_NET_CLS_IND
 807         new->ifindex = n->ifindex;
 808 #endif
 809         new->fshift = n->fshift;
 810         new->res = n->res;
 811         new->flags = n->flags;
 812         RCU_INIT_POINTER(new->ht_down, ht);
 813
 814         /* bump reference count as long as we hold pointer to structure */
 815         if (ht)
 816                 ht->refcnt++;
 817
 818 #ifdef CONFIG_CLS_U32_PERF
 819         /* Statistics may be incremented by readers during update
 820          * so we must keep them in tact. When the node is later destroyed
 821          * a special destroy call must be made to not free the pf memory.
 822          */
 823         new->pf = n->pf;
 824 #endif
 825
 826 #ifdef CONFIG_CLS_U32_MARK
 827         new->val = n->val;
 828         new->mask = n->mask;
 829         /* Similarly success statistics must be moved as pointers */
 830         new->pcpu_success = n->pcpu_success;
 831 #endif
 832         new->tp = tp;
 833         memcpy(&new->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 834
 835         if (tcf_exts_init(&new->exts, TCA_U32_ACT, TCA_U32_POLICE)) {
 836                 kfree(new);
 837                 return NULL;
 838         }
 839
 840         return new;
 841 }
 842
 843 static int u32_change(struct net *net, struct sk_buff *in_skb,
 844                       struct tcf_proto *tp, unsigned long base, u32 handle,
 845                       struct nlattr **tca, unsigned long *arg, bool ovr)
 846 {
 847         struct tc_u_common *tp_c = tp->data;
 848         struct tc_u_hnode *ht;
 849         struct tc_u_knode *n;
 850         struct tc_u32_sel *s;
 851         struct nlattr *opt = tca[TCA_OPTIONS];
 852         struct nlattr *tb[TCA_U32_MAX + 1];
 853         u32 htid, flags = 0;
 854         size_t sel_size;
 855         int err;
 856 #ifdef CONFIG_CLS_U32_PERF
 857         size_t size;
 858 #endif
 859
 860         if (opt == NULL)
 861                 return handle ? -EINVAL : 0;
 862
 863         err = nla_parse_nested(tb, TCA_U32_MAX, opt, u32_policy);
 864         if (err < 0)
 865                 return err;
 866
 867         if (tb[TCA_U32_FLAGS]) {
 868                 flags = nla_get_u32(tb[TCA_U32_FLAGS]);
 869                 if (!tc_flags_valid(flags))
 870                         return -EINVAL;
 871         }
 872
 873         n = (struct tc_u_knode *)*arg;
 874         if (n) {
 875                 struct tc_u_knode *new;
 876
 877                 if (TC_U32_KEY(n->handle) == 0)
 878                         return -EINVAL;
 879
 880                 if (n->flags != flags)
 881                         return -EINVAL;
 882
 883                 new = u32_init_knode(tp, n);
 884                 if (!new)
 885                         return -ENOMEM;
 886
 887                 err = u32_set_parms(net, tp, base,
 888                                     rtnl_dereference(n->ht_up), new, tb,
 889                                     tca[TCA_RATE], ovr);
 890
 891                 if (err) {
 892                         u32_destroy_key(tp, new, false);
 893                         return err;
 894                 }
 895
 896                 err = u32_replace_hw_knode(tp, new, flags);
 897                 if (err) {
 898                         u32_destroy_key(tp, new, false);
 899                         return err;
 900                 }
 901
 902                 u32_replace_knode(tp, tp_c, new);
 903                 tcf_unbind_filter(tp, &n->res);
 904                 call_rcu(&n->rcu, u32_delete_key_rcu);
 905                 return 0;
 906         }
 907
 908         if (tb[TCA_U32_DIVISOR]) {
 909                 unsigned int divisor = nla_get_u32(tb[TCA_U32_DIVISOR]);
 910
 911                 if (--divisor > 0x100)
 912                         return -EINVAL;
 913                 if (TC_U32_KEY(handle))
 914                         return -EINVAL;
 915                 if (handle == 0) {
 916                         handle = gen_new_htid(tp->data);
 917                         if (handle == 0)
 918                                 return -ENOMEM;
 919                 }
 920                 ht = kzalloc(sizeof(*ht) + divisor*sizeof(void *), GFP_KERNEL);
 921                 if (ht == NULL)
 922                         return -ENOBUFS;
 923                 ht->tp_c = tp_c;
 924                 ht->refcnt = 1;
 925                 ht->divisor = divisor;
 926                 ht->handle = handle;
 927                 ht->prio = tp->prio;
 928
 929                 err = u32_replace_hw_hnode(tp, ht, flags);
 930                 if (err) {
 931                         kfree(ht);
 932                         return err;
 933                 }
 934
 935                 RCU_INIT_POINTER(ht->next, tp_c->hlist);
 936                 rcu_assign_pointer(tp_c->hlist, ht);
 937                 *arg = (unsigned long)ht;
 938
 939                 return 0;
 940         }
 941
 942         if (tb[TCA_U32_HASH]) {
 943                 htid = nla_get_u32(tb[TCA_U32_HASH]);
 944                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 945                         ht = rtnl_dereference(tp->root);
 946                         htid = ht->handle;
 947                 } else {
 948                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 949                         if (ht == NULL)
 950                                 return -EINVAL;
 951                 }
 952         } else {
 953                 ht = rtnl_dereference(tp->root);
 954                 htid = ht->handle;
 955         }
 956
 957         if (ht->divisor < TC_U32_HASH(htid))
 958                 return -EINVAL;
 959
 960         if (handle) {
 961                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 962                         return -EINVAL;
 963                 handle = htid | TC_U32_NODE(handle);
 964         } else
 965                 handle = gen_new_kid(ht, htid);
 966
 967         if (tb[TCA_U32_SEL] == NULL)
 968                 return -EINVAL;
 969
 970         s = nla_data(tb[TCA_U32_SEL]);
 971         sel_size = sizeof(*s) + sizeof(*s->keys) * s->nkeys;
 972         if (nla_len(tb[TCA_U32_SEL]) < sel_size)
 973                 return -EINVAL;
 974
 975         n = kzalloc(offsetof(typeof(*n), sel) + sel_size, GFP_KERNEL);
 976         if (n == NULL)
 977                 return -ENOBUFS;
 978
 979 #ifdef CONFIG_CLS_U32_PERF
 980         size = sizeof(struct tc_u32_pcnt) + s->nkeys * sizeof(u64);
 981         n->pf = __alloc_percpu(size, __alignof__(struct tc_u32_pcnt));
 982         if (!n->pf) {
 983                 kfree(n);
 984                 return -ENOBUFS;
 985         }
 986 #endif
 987
 988         memcpy(&n->sel, s, sel_size);
 989         RCU_INIT_POINTER(n->ht_up, ht);
 990         n->handle = handle;
 991         n->fshift = s->hmask ? ffs(ntohl(s->hmask)) - 1 : 0;
 992         n->flags = flags;
 993         n->tp = tp;
 994
 995         err = tcf_exts_init(&n->exts, TCA_U32_ACT, TCA_U32_POLICE);
 996         if (err < 0)
 997                 goto errout;
 998
 999 #ifdef CONFIG_CLS_U32_MARK
1000         n->pcpu_success = alloc_percpu(u32);
1001         if (!n->pcpu_success) {
1002                 err = -ENOMEM;
1003                 goto errout;
1004         }
1005
1006         if (tb[TCA_U32_MARK]) {
1007                 struct tc_u32_mark *mark;
1008
1009                 mark = nla_data(tb[TCA_U32_MARK]);
1010                 n->val = mark->val;
1011                 n->mask = mark->mask;
1012         }
1013 #endif
1014
1015         err = u32_set_parms(net, tp, base, ht, n, tb, tca[TCA_RATE], ovr);
1016         if (err == 0) {
1017                 struct tc_u_knode __rcu **ins;
1018                 struct tc_u_knode *pins;
1019
1020                 err = u32_replace_hw_knode(tp, n, flags);
1021                 if (err)
1022                         goto errhw;
1023
1024                 ins = &ht->ht[TC_U32_HASH(handle)];
1025                 for (pins = rtnl_dereference(*ins); pins;
1026                      ins = &pins->next, pins = rtnl_dereference(*ins))
1027                         if (TC_U32_NODE(handle) < TC_U32_NODE(pins->handle))
1028                                 break;
1029
1030                 RCU_INIT_POINTER(n->next, pins);
1031                 rcu_assign_pointer(*ins, n);
1032                 *arg = (unsigned long)n;
1033                 return 0;
1034         }
1035
1036 errhw:
1037 #ifdef CONFIG_CLS_U32_MARK
1038         free_percpu(n->pcpu_success);
1039 #endif
1040
1041 errout:
1042         tcf_exts_destroy(&n->exts);
1043 #ifdef CONFIG_CLS_U32_PERF
1044         free_percpu(n->pf);
1045 #endif
1046         kfree(n);
1047         return err;
1048 }
1049
1050 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
1051 {
1052         struct tc_u_common *tp_c = tp->data;
1053         struct tc_u_hnode *ht;
1054         struct tc_u_knode *n;
1055         unsigned int h;
1056
1057         if (arg->stop)
1058                 return;
1059
1060         for (ht = rtnl_dereference(tp_c->hlist);
1061              ht;
1062              ht = rtnl_dereference(ht->next)) {
1063                 if (ht->prio != tp->prio)
1064                         continue;
1065                 if (arg->count >= arg->skip) {
1066                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
1067                                 arg->stop = 1;
1068                                 return;
1069                         }
1070                 }
1071                 arg->count++;
1072                 for (h = 0; h <= ht->divisor; h++) {
1073                         for (n = rtnl_dereference(ht->ht[h]);
1074                              n;
1075                              n = rtnl_dereference(n->next)) {
1076                                 if (arg->count < arg->skip) {
1077                                         arg->count++;
1078                                         continue;
1079                                 }
1080                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
1081                                         arg->stop = 1;
1082                                         return;
1083                                 }
1084                                 arg->count++;
1085                         }
1086                 }
1087         }
1088 }
1089
1090 static int u32_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
1091                     struct sk_buff *skb, struct tcmsg *t)
1092 {
1093         struct tc_u_knode *n = (struct tc_u_knode *)fh;
1094         struct tc_u_hnode *ht_up, *ht_down;
1095         struct nlattr *nest;
1096
1097         if (n == NULL)
1098                 return skb->len;
1099
1100         t->tcm_handle = n->handle;
1101
1102         nest = nla_nest_start(skb, TCA_OPTIONS);
1103         if (nest == NULL)
1104                 goto nla_put_failure;
1105
1106         if (TC_U32_KEY(n->handle) == 0) {
1107                 struct tc_u_hnode *ht = (struct tc_u_hnode *)fh;
1108                 u32 divisor = ht->divisor + 1;
1109
1110                 if (nla_put_u32(skb, TCA_U32_DIVISOR, divisor))
1111                         goto nla_put_failure;
1112         } else {
1113 #ifdef CONFIG_CLS_U32_PERF
1114                 struct tc_u32_pcnt *gpf;
1115                 int cpu;
1116 #endif
1117
1118                 if (nla_put(skb, TCA_U32_SEL,
1119                             sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
1120                             &n->sel))
1121                         goto nla_put_failure;
1122
1123                 ht_up = rtnl_dereference(n->ht_up);
1124                 if (ht_up) {
1125                         u32 htid = n->handle & 0xFFFFF000;
1126                         if (nla_put_u32(skb, TCA_U32_HASH, htid))
1127                                 goto nla_put_failure;
1128                 }
1129                 if (n->res.classid &&
1130                     nla_put_u32(skb, TCA_U32_CLASSID, n->res.classid))
1131                         goto nla_put_failure;
1132
1133                 ht_down = rtnl_dereference(n->ht_down);
1134                 if (ht_down &&
1135                     nla_put_u32(skb, TCA_U32_LINK, ht_down->handle))
1136                         goto nla_put_failure;
1137
1138                 if (n->flags && nla_put_u32(skb, TCA_U32_FLAGS, n->flags))
1139                         goto nla_put_failure;
1140
1141 #ifdef CONFIG_CLS_U32_MARK
1142                 if ((n->val || n->mask)) {
1143                         struct tc_u32_mark mark = {.val = n->val,
1144                                                    .mask = n->mask,
1145                                                    .success = 0};
1146                         int cpum;
1147
1148                         for_each_possible_cpu(cpum) {
1149                                 __u32 cnt = *per_cpu_ptr(n->pcpu_success, cpum);
1150
1151                                 mark.success += cnt;
1152                         }
1153
1154                         if (nla_put(skb, TCA_U32_MARK, sizeof(mark), &mark))
1155                                 goto nla_put_failure;
1156                 }
1157 #endif
1158
1159                 if (tcf_exts_dump(skb, &n->exts) < 0)
1160                         goto nla_put_failure;
1161
1162 #ifdef CONFIG_NET_CLS_IND
1163                 if (n->ifindex) {
1164                         struct net_device *dev;
1165                         dev = __dev_get_by_index(net, n->ifindex);
1166                         if (dev && nla_put_string(skb, TCA_U32_INDEV, dev->name))
1167                                 goto nla_put_failure;
1168                 }
1169 #endif
1170 #ifdef CONFIG_CLS_U32_PERF
1171                 gpf = kzalloc(sizeof(struct tc_u32_pcnt) +
1172                               n->sel.nkeys * sizeof(u64),
1173                               GFP_KERNEL);
1174                 if (!gpf)
1175                         goto nla_put_failure;
1176
1177                 for_each_possible_cpu(cpu) {
1178                         int i;
1179                         struct tc_u32_pcnt *pf = per_cpu_ptr(n->pf, cpu);
1180
1181                         gpf->rcnt += pf->rcnt;
1182                         gpf->rhit += pf->rhit;
1183                         for (i = 0; i < n->sel.nkeys; i++)
1184                                 gpf->kcnts[i] += pf->kcnts[i];
1185                 }
1186
1187                 if (nla_put_64bit(skb, TCA_U32_PCNT,
1188                                   sizeof(struct tc_u32_pcnt) +
1189                                   n->sel.nkeys * sizeof(u64),
1190                                   gpf, TCA_U32_PAD)) {
1191                         kfree(gpf);
1192                         goto nla_put_failure;
1193                 }
1194                 kfree(gpf);
1195 #endif
1196         }
1197
1198         nla_nest_end(skb, nest);
1199
1200         if (TC_U32_KEY(n->handle))
1201                 if (tcf_exts_dump_stats(skb, &n->exts) < 0)
1202                         goto nla_put_failure;
1203         return skb->len;
1204
1205 nla_put_failure:
1206         nla_nest_cancel(skb, nest);
1207         return -1;
1208 }
1209
1210 static struct tcf_proto_ops cls_u32_ops __read_mostly = {
1211         .kind           =       "u32",
1212         .classify       =       u32_classify,
1213         .init           =       u32_init,
1214         .destroy        =       u32_destroy,
1215         .get            =       u32_get,
1216         .change         =       u32_change,
1217         .delete         =       u32_delete,
1218         .walk           =       u32_walk,
1219         .dump           =       u32_dump,
1220         .owner          =       THIS_MODULE,
1221 };
1222
1223 static int __init init_u32(void)
1224 {
1225         pr_info("u32 classifier\n");
1226 #ifdef CONFIG_CLS_U32_PERF
1227         pr_info("    Performance counters on\n");
1228 #endif
1229 #ifdef CONFIG_NET_CLS_IND
1230         pr_info("    input device check on\n");
1231 #endif
1232 #ifdef CONFIG_NET_CLS_ACT
1233         pr_info("    Actions configured\n");
1234 #endif
1235         return register_tcf_proto_ops(&cls_u32_ops);
1236 }
1237
1238 static void __exit exit_u32(void)
1239 {
1240         unregister_tcf_proto_ops(&cls_u32_ops);
1241 }
1242
1243 module_init(init_u32)
1244 module_exit(exit_u32)
1245 MODULE_LICENSE("GPL");