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  */
  31
  32 #include <asm/uaccess.h>
  33 #include <asm/system.h>
  34 #include <asm/bitops.h>
  35 #include <linux/config.h>
  36 #include <linux/module.h>
  37 #include <linux/types.h>
  38 #include <linux/kernel.h>
  39 #include <linux/sched.h>
  40 #include <linux/string.h>
  41 #include <linux/mm.h>
  42 #include <linux/socket.h>
  43 #include <linux/sockios.h>
  44 #include <linux/in.h>
  45 #include <linux/errno.h>
  46 #include <linux/interrupt.h>
  47 #include <linux/if_ether.h>
  48 #include <linux/inet.h>
  49 #include <linux/netdevice.h>
  50 #include <linux/etherdevice.h>
  51 #include <linux/notifier.h>
  52 #include <linux/rtnetlink.h>
  53 #include <net/ip.h>
  54 #include <net/route.h>
  55 #include <linux/skbuff.h>
  56 #include <net/sock.h>
  57 #include <net/pkt_sched.h>
  58
  59
  60 struct tc_u_knode
  61 {
  62         struct tc_u_knode       *next;
  63         u32                     handle;
  64         struct tc_u_hnode       *ht_up;
  65 #ifdef CONFIG_NET_CLS_ACT
  66         struct tc_action        *action;
  67 #else
  68 #ifdef CONFIG_NET_CLS_POLICE
  69         struct tcf_police       *police;
  70 #endif
  71 #endif
  72 #ifdef CONFIG_NET_CLS_IND
  73         char                     indev[IFNAMSIZ];
  74 #endif
  75         u8                      fshift;
  76         struct tcf_result       res;
  77         struct tc_u_hnode       *ht_down;
  78 #ifdef CONFIG_CLS_U32_PERF
  79         struct tc_u32_pcnt      *pf;
  80 #endif
  81         struct tc_u32_sel       sel;
  82 };
  83
  84 struct tc_u_hnode
  85 {
  86         struct tc_u_hnode       *next;
  87         u32                     handle;
  88         struct tc_u_common      *tp_c;
  89         int                     refcnt;
  90         unsigned                divisor;
  91         u32                     hgenerator;
  92         struct tc_u_knode       *ht[1];
  93 };
  94
  95 struct tc_u_common
  96 {
  97         struct tc_u_common      *next;
  98         struct tc_u_hnode       *hlist;
  99         struct Qdisc            *q;
 100         int                     refcnt;
 101         u32                     hgenerator;
 102 };
 103
 104 static struct tc_u_common *u32_list;
 105
 106 static __inline__ unsigned u32_hash_fold(u32 key, struct tc_u32_sel *sel, u8 fshift)
 107 {
 108         unsigned h = (key & sel->hmask)>>fshift;
 109
 110         return h;
 111 }
 112
 113 static int u32_classify(struct sk_buff *skb, struct tcf_proto *tp, struct tcf_result *res)
 114 {
 115         struct {
 116                 struct tc_u_knode *knode;
 117                 u8                *ptr;
 118         } stack[TC_U32_MAXDEPTH];
 119
 120         struct tc_u_hnode *ht = (struct tc_u_hnode*)tp->root;
 121         u8 *ptr = skb->nh.raw;
 122         struct tc_u_knode *n;
 123         int sdepth = 0;
 124         int off2 = 0;
 125         int sel = 0;
 126 #ifdef CONFIG_CLS_U32_PERF
 127         int j;
 128 #endif
 129         int i;
 130
 131 next_ht:
 132         n = ht->ht[sel];
 133
 134 next_knode:
 135         if (n) {
 136                 struct tc_u32_key *key = n->sel.keys;
 137
 138 #ifdef CONFIG_CLS_U32_PERF
 139                 n->pf->rcnt +=1;
 140                 j = 0;
 141 #endif
 142                 for (i = n->sel.nkeys; i>0; i--, key++) {
 143
 144                         if ((*(u32*)(ptr+key->off+(off2&key->offmask))^key->val)&key->mask) {
 145                                 n = n->next;
 146                                 goto next_knode;
 147                         }
 148 #ifdef CONFIG_CLS_U32_PERF
 149                         n->pf->kcnts[j] +=1;
 150                         j++;
 151 #endif
 152                 }
 153                 if (n->ht_down == NULL) {
 154 check_terminal:
 155                         if (n->sel.flags&TC_U32_TERMINAL) {
 156
 157                                 *res = n->res;
 158 #ifdef CONFIG_NET_CLS_IND
 159                                 /* yes, i know it sucks but the feature is
 160                                 ** optional dammit! - JHS */
 161                                 if (0 != n->indev[0]) {
 162                                         if  (NULL == skb->input_dev) {
 163                                                 n = n->next;
 164                                                 goto next_knode;
 165                                         } else {
 166                                                 if (0 != strcmp(n->indev, skb->input_dev->name)) {
 167                                                         n = n->next;
 168                                                         goto next_knode;
 169                                                 }
 170                                         }
 171                                 }
 172 #endif
 173 #ifdef CONFIG_CLS_U32_PERF
 174                                 n->pf->rhit +=1;
 175 #endif
 176 #ifdef CONFIG_NET_CLS_ACT
 177                                 if (n->action) {
 178                                         int pol_res = tcf_action_exec(skb, n->action, res);
 179                                         if (pol_res >= 0)
 180                                                 return pol_res;
 181                                 } else
 182 #else
 183 #ifdef CONFIG_NET_CLS_POLICE
 184                                 if (n->police) {
 185                                         int pol_res = tcf_police(skb, n->police);
 186                                         if (pol_res >= 0)
 187                                                 return pol_res;
 188                                 } else
 189 #endif
 190 #endif
 191                                         return 0;
 192                         }
 193                         n = n->next;
 194                         goto next_knode;
 195                 }
 196
 197                 /* PUSH */
 198                 if (sdepth >= TC_U32_MAXDEPTH)
 199                         goto deadloop;
 200                 stack[sdepth].knode = n;
 201                 stack[sdepth].ptr = ptr;
 202                 sdepth++;
 203
 204                 ht = n->ht_down;
 205                 sel = 0;
 206                 if (ht->divisor)
 207                         sel = ht->divisor&u32_hash_fold(*(u32*)(ptr+n->sel.hoff), &n->sel,n->fshift);
 208
 209                 if (!(n->sel.flags&(TC_U32_VAROFFSET|TC_U32_OFFSET|TC_U32_EAT)))
 210                         goto next_ht;
 211
 212                 if (n->sel.flags&(TC_U32_OFFSET|TC_U32_VAROFFSET)) {
 213                         off2 = n->sel.off + 3;
 214                         if (n->sel.flags&TC_U32_VAROFFSET)
 215                                 off2 += ntohs(n->sel.offmask & *(u16*)(ptr+n->sel.offoff)) >>n->sel.offshift;
 216                         off2 &= ~3;
 217                 }
 218                 if (n->sel.flags&TC_U32_EAT) {
 219                         ptr += off2;
 220                         off2 = 0;
 221                 }
 222
 223                 if (ptr < skb->tail)
 224                         goto next_ht;
 225         }
 226
 227         /* POP */
 228         if (sdepth--) {
 229                 n = stack[sdepth].knode;
 230                 ht = n->ht_up;
 231                 ptr = stack[sdepth].ptr;
 232                 goto check_terminal;
 233         }
 234         return -1;
 235
 236 deadloop:
 237         if (net_ratelimit())
 238                 printk("cls_u32: dead loop\n");
 239         return -1;
 240 }
 241
 242 static __inline__ struct tc_u_hnode *
 243 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 244 {
 245         struct tc_u_hnode *ht;
 246
 247         for (ht = tp_c->hlist; ht; ht = ht->next)
 248                 if (ht->handle == handle)
 249                         break;
 250
 251         return ht;
 252 }
 253
 254 static __inline__ struct tc_u_knode *
 255 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 256 {
 257         unsigned sel;
 258         struct tc_u_knode *n = NULL;
 259
 260         sel = TC_U32_HASH(handle);
 261         if (sel > ht->divisor)
 262                 goto out;
 263
 264         for (n = ht->ht[sel]; n; n = n->next)
 265                 if (n->handle == handle)
 266                         break;
 267 out:
 268         return n;
 269 }
 270
 271
 272 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 273 {
 274         struct tc_u_hnode *ht;
 275         struct tc_u_common *tp_c = tp->data;
 276
 277         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 278                 ht = tp->root;
 279         else
 280                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 281
 282         if (!ht)
 283                 return 0;
 284
 285         if (TC_U32_KEY(handle) == 0)
 286                 return (unsigned long)ht;
 287
 288         return (unsigned long)u32_lookup_key(ht, handle);
 289 }
 290
 291 static void u32_put(struct tcf_proto *tp, unsigned long f)
 292 {
 293 }
 294
 295 static u32 gen_new_htid(struct tc_u_common *tp_c)
 296 {
 297         int i = 0x800;
 298
 299         do {
 300                 if (++tp_c->hgenerator == 0x7FF)
 301                         tp_c->hgenerator = 1;
 302         } while (--i>0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 303
 304         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 305 }
 306
 307 static int u32_init(struct tcf_proto *tp)
 308 {
 309         struct tc_u_hnode *root_ht;
 310         struct tc_u_common *tp_c;
 311
 312         for (tp_c = u32_list; tp_c; tp_c = tp_c->next)
 313                 if (tp_c->q == tp->q)
 314                         break;
 315
 316         root_ht = kmalloc(sizeof(*root_ht), GFP_KERNEL);
 317         if (root_ht == NULL)
 318                 return -ENOBUFS;
 319
 320         memset(root_ht, 0, sizeof(*root_ht));
 321         root_ht->divisor = 0;
 322         root_ht->refcnt++;
 323         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 324
 325         if (tp_c == NULL) {
 326                 tp_c = kmalloc(sizeof(*tp_c), GFP_KERNEL);
 327                 if (tp_c == NULL) {
 328                         kfree(root_ht);
 329                         return -ENOBUFS;
 330                 }
 331                 memset(tp_c, 0, sizeof(*tp_c));
 332                 tp_c->q = tp->q;
 333                 tp_c->next = u32_list;
 334                 u32_list = tp_c;
 335         }
 336
 337         tp_c->refcnt++;
 338         root_ht->next = tp_c->hlist;
 339         tp_c->hlist = root_ht;
 340         root_ht->tp_c = tp_c;
 341
 342         tp->root = root_ht;
 343         tp->data = tp_c;
 344         return 0;
 345 }
 346
 347 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
 348 {
 349         unsigned long cl;
 350
 351         if ((cl = __cls_set_class(&n->res.class, 0)) != 0)
 352                 tp->q->ops->cl_ops->unbind_tcf(tp->q, cl);
 353 #ifdef CONFIG_NET_CLS_ACT
 354         if (n->action) {
 355                 tcf_action_destroy(n->action, TCA_ACT_UNBIND);
 356         }
 357 #else
 358 #ifdef CONFIG_NET_CLS_POLICE
 359         tcf_police_release(n->police, TCA_ACT_UNBIND);
 360 #endif
 361 #endif
 362         if (n->ht_down)
 363                 n->ht_down->refcnt--;
 364 #ifdef CONFIG_CLS_U32_PERF
 365         if (n && (NULL != n->pf))
 366                 kfree(n->pf);
 367 #endif
 368         kfree(n);
 369         return 0;
 370 }
 371
 372 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode* key)
 373 {
 374         struct tc_u_knode **kp;
 375         struct tc_u_hnode *ht = key->ht_up;
 376
 377         if (ht) {
 378                 for (kp = &ht->ht[TC_U32_HASH(key->handle)]; *kp; kp = &(*kp)->next) {
 379                         if (*kp == key) {
 380                                 tcf_tree_lock(tp);
 381                                 *kp = key->next;
 382                                 tcf_tree_unlock(tp);
 383
 384                                 u32_destroy_key(tp, key);
 385                                 return 0;
 386                         }
 387                 }
 388         }
 389         BUG_TRAP(0);
 390         return 0;
 391 }
 392
 393 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 394 {
 395         struct tc_u_knode *n;
 396         unsigned h;
 397
 398         for (h=0; h<=ht->divisor; h++) {
 399                 while ((n = ht->ht[h]) != NULL) {
 400                         ht->ht[h] = n->next;
 401
 402                         u32_destroy_key(tp, n);
 403                 }
 404         }
 405 }
 406
 407 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 408 {
 409         struct tc_u_common *tp_c = tp->data;
 410         struct tc_u_hnode **hn;
 411
 412         BUG_TRAP(!ht->refcnt);
 413
 414         u32_clear_hnode(tp, ht);
 415
 416         for (hn = &tp_c->hlist; *hn; hn = &(*hn)->next) {
 417                 if (*hn == ht) {
 418                         *hn = ht->next;
 419                         kfree(ht);
 420                         return 0;
 421                 }
 422         }
 423
 424         BUG_TRAP(0);
 425         return -ENOENT;
 426 }
 427
 428 static void u32_destroy(struct tcf_proto *tp)
 429 {
 430         struct tc_u_common *tp_c = tp->data;
 431         struct tc_u_hnode *root_ht = xchg(&tp->root, NULL);
 432
 433         BUG_TRAP(root_ht != NULL);
 434
 435         if (root_ht && --root_ht->refcnt == 0)
 436                 u32_destroy_hnode(tp, root_ht);
 437
 438         if (--tp_c->refcnt == 0) {
 439                 struct tc_u_hnode *ht;
 440                 struct tc_u_common **tp_cp;
 441
 442                 for (tp_cp = &u32_list; *tp_cp; tp_cp = &(*tp_cp)->next) {
 443                         if (*tp_cp == tp_c) {
 444                                 *tp_cp = tp_c->next;
 445                                 break;
 446                         }
 447                 }
 448
 449                 for (ht=tp_c->hlist; ht; ht = ht->next)
 450                         u32_clear_hnode(tp, ht);
 451
 452                 while ((ht = tp_c->hlist) != NULL) {
 453                         tp_c->hlist = ht->next;
 454
 455                         BUG_TRAP(ht->refcnt == 0);
 456
 457                         kfree(ht);
 458                 };
 459
 460                 kfree(tp_c);
 461         }
 462
 463         tp->data = NULL;
 464 }
 465
 466 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 467 {
 468         struct tc_u_hnode *ht = (struct tc_u_hnode*)arg;
 469
 470         if (ht == NULL)
 471                 return 0;
 472
 473         if (TC_U32_KEY(ht->handle))
 474                 return u32_delete_key(tp, (struct tc_u_knode*)ht);
 475
 476         if (tp->root == ht)
 477                 return -EINVAL;
 478
 479         if (--ht->refcnt == 0)
 480                 u32_destroy_hnode(tp, ht);
 481
 482         return 0;
 483 }
 484
 485 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 486 {
 487         struct tc_u_knode *n;
 488         unsigned i = 0x7FF;
 489
 490         for (n=ht->ht[TC_U32_HASH(handle)]; n; n = n->next)
 491                 if (i < TC_U32_NODE(n->handle))
 492                         i = TC_U32_NODE(n->handle);
 493         i++;
 494
 495         return handle|(i>0xFFF ? 0xFFF : i);
 496 }
 497
 498 static int u32_set_parms(struct Qdisc *q, unsigned long base,
 499                          struct tc_u_hnode *ht,
 500                          struct tc_u_knode *n, struct rtattr **tb,
 501                          struct rtattr *est)
 502 {
 503 #ifdef CONFIG_NET_CLS_ACT
 504         struct tc_action *act = NULL;
 505         int ret;
 506 #endif
 507         if (tb[TCA_U32_LINK-1]) {
 508                 u32 handle = *(u32*)RTA_DATA(tb[TCA_U32_LINK-1]);
 509                 struct tc_u_hnode *ht_down = NULL;
 510
 511                 if (TC_U32_KEY(handle))
 512                         return -EINVAL;
 513
 514                 if (handle) {
 515                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 516
 517                         if (ht_down == NULL)
 518                                 return -EINVAL;
 519                         ht_down->refcnt++;
 520                 }
 521
 522                 sch_tree_lock(q);
 523                 ht_down = xchg(&n->ht_down, ht_down);
 524                 sch_tree_unlock(q);
 525
 526                 if (ht_down)
 527                         ht_down->refcnt--;
 528         }
 529         if (tb[TCA_U32_CLASSID-1]) {
 530                 unsigned long cl;
 531
 532                 n->res.classid = *(u32*)RTA_DATA(tb[TCA_U32_CLASSID-1]);
 533                 sch_tree_lock(q);
 534                 cl = __cls_set_class(&n->res.class, q->ops->cl_ops->bind_tcf(q, base, n->res.classid));
 535                 sch_tree_unlock(q);
 536                 if (cl)
 537                         q->ops->cl_ops->unbind_tcf(q, cl);
 538         }
 539 #ifdef CONFIG_NET_CLS_ACT
 540         /*backward compatibility */
 541         if (tb[TCA_U32_POLICE-1])
 542         {
 543                 act = kmalloc(sizeof(*act),GFP_KERNEL);
 544                 if (NULL == act)
 545                         return -ENOMEM;
 546
 547                 memset(act,0,sizeof(*act));
 548                 ret = tcf_action_init_1(tb[TCA_U32_POLICE-1], est,act,"police", TCA_ACT_NOREPLACE, TCA_ACT_BIND);
 549                 if (0 > ret){
 550                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 551                         return ret;
 552                 }
 553                 act->type = TCA_OLD_COMPAT;
 554
 555                 sch_tree_lock(q);
 556                 act = xchg(&n->action, act);
 557                 sch_tree_unlock(q);
 558
 559                 tcf_action_destroy(act, TCA_ACT_UNBIND);
 560
 561         }
 562
 563         if(tb[TCA_U32_ACT-1]) {
 564                 act = kmalloc(sizeof(*act),GFP_KERNEL);
 565                 if (NULL == act)
 566                         return -ENOMEM;
 567                 memset(act,0,sizeof(*act));
 568                 ret = tcf_action_init(tb[TCA_U32_ACT-1], est,act,NULL,TCA_ACT_NOREPLACE, TCA_ACT_BIND);
 569                 if (0 > ret) {
 570                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 571                         return ret;
 572                 }
 573
 574                 sch_tree_lock(q);
 575                 act = xchg(&n->action, act);
 576                 sch_tree_unlock(q);
 577
 578                 tcf_action_destroy(act, TCA_ACT_UNBIND);
 579         }
 580
 581
 582 #else
 583 #ifdef CONFIG_NET_CLS_POLICE
 584         if (tb[TCA_U32_POLICE-1]) {
 585                 struct tcf_police *police = tcf_police_locate(tb[TCA_U32_POLICE-1], est);
 586                 sch_tree_lock(q);
 587                 police = xchg(&n->police, police);
 588                 sch_tree_unlock(q);
 589                 tcf_police_release(police, TCA_ACT_UNBIND);
 590         }
 591 #endif
 592 #endif
 593 #ifdef CONFIG_NET_CLS_IND
 594         n->indev[0] = 0;
 595         if(tb[TCA_U32_INDEV-1]) {
 596                 struct rtattr *input_dev = tb[TCA_U32_INDEV-1];
 597                 if (RTA_PAYLOAD(input_dev) >= IFNAMSIZ) {
 598                         printk("cls_u32: bad indev name %s\n",(char*)RTA_DATA(input_dev));
 599                         /* should we clear state first? */
 600                         return  -EINVAL;
 601                 }
 602                 sprintf(n->indev, "%s", (char*)RTA_DATA(input_dev));
 603                 printk("got IND %s\n",n->indev);
 604         }
 605 #endif
 606
 607         return 0;
 608 }
 609
 610 static int u32_change(struct tcf_proto *tp, unsigned long base, u32 handle,
 611                       struct rtattr **tca,
 612                       unsigned long *arg)
 613 {
 614         struct tc_u_common *tp_c = tp->data;
 615         struct tc_u_hnode *ht;
 616         struct tc_u_knode *n;
 617         struct tc_u32_sel *s;
 618         struct rtattr *opt = tca[TCA_OPTIONS-1];
 619         struct rtattr *tb[TCA_U32_MAX];
 620         u32 htid;
 621         int err;
 622
 623         if (opt == NULL)
 624                 return handle ? -EINVAL : 0;
 625
 626         if (rtattr_parse(tb, TCA_U32_MAX, RTA_DATA(opt), RTA_PAYLOAD(opt)) < 0)
 627                 return -EINVAL;
 628
 629         if ((n = (struct tc_u_knode*)*arg) != NULL) {
 630                 if (TC_U32_KEY(n->handle) == 0)
 631                         return -EINVAL;
 632
 633                 return u32_set_parms(tp->q, base, n->ht_up, n, tb, tca[TCA_RATE-1]);
 634         }
 635
 636         if (tb[TCA_U32_DIVISOR-1]) {
 637                 unsigned divisor = *(unsigned*)RTA_DATA(tb[TCA_U32_DIVISOR-1]);
 638
 639                 if (--divisor > 0x100)
 640                         return -EINVAL;
 641                 if (TC_U32_KEY(handle))
 642                         return -EINVAL;
 643                 if (handle == 0) {
 644                         handle = gen_new_htid(tp->data);
 645                         if (handle == 0)
 646                                 return -ENOMEM;
 647                 }
 648                 ht = kmalloc(sizeof(*ht) + divisor*sizeof(void*), GFP_KERNEL);
 649                 if (ht == NULL)
 650                         return -ENOBUFS;
 651                 memset(ht, 0, sizeof(*ht) + divisor*sizeof(void*));
 652                 ht->tp_c = tp_c;
 653                 ht->refcnt = 0;
 654                 ht->divisor = divisor;
 655                 ht->handle = handle;
 656                 ht->next = tp_c->hlist;
 657                 tp_c->hlist = ht;
 658                 *arg = (unsigned long)ht;
 659                 return 0;
 660         }
 661
 662         if (tb[TCA_U32_HASH-1]) {
 663                 htid = *(unsigned*)RTA_DATA(tb[TCA_U32_HASH-1]);
 664                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 665                         ht = tp->root;
 666                         htid = ht->handle;
 667                 } else {
 668                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 669                         if (ht == NULL)
 670                                 return -EINVAL;
 671                 }
 672         } else {
 673                 ht = tp->root;
 674                 htid = ht->handle;
 675         }
 676
 677         if (ht->divisor < TC_U32_HASH(htid))
 678                 return -EINVAL;
 679
 680         if (handle) {
 681                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 682                         return -EINVAL;
 683                 handle = htid | TC_U32_NODE(handle);
 684         } else
 685                 handle = gen_new_kid(ht, htid);
 686
 687         if (tb[TCA_U32_SEL-1] == 0 ||
 688             RTA_PAYLOAD(tb[TCA_U32_SEL-1]) < sizeof(struct tc_u32_sel))
 689                 return -EINVAL;
 690
 691         s = RTA_DATA(tb[TCA_U32_SEL-1]);
 692
 693         n = kmalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 694         if (n == NULL)
 695                 return -ENOBUFS;
 696
 697         memset(n, 0, sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key));
 698 #ifdef CONFIG_CLS_U32_PERF
 699         n->pf = kmalloc(sizeof(struct tc_u32_pcnt) + s->nkeys*sizeof(__u64), GFP_KERNEL);
 700         if (n->pf == NULL) {
 701                 kfree(n);
 702                 return -ENOBUFS;
 703         }
 704         memset(n->pf, 0, sizeof(struct tc_u32_pcnt) + s->nkeys*sizeof(__u64));
 705 #endif
 706
 707         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 708         n->ht_up = ht;
 709         n->handle = handle;
 710 {
 711         u8 i = 0;
 712         u32 mask = s->hmask;
 713         if (mask) {
 714                 while (!(mask & 1)) {
 715                         i++;
 716                         mask>>=1;
 717                 }
 718         }
 719         n->fshift = i;
 720 }
 721         err = u32_set_parms(tp->q, base, ht, n, tb, tca[TCA_RATE-1]);
 722         if (err == 0) {
 723                 struct tc_u_knode **ins;
 724                 for (ins = &ht->ht[TC_U32_HASH(handle)]; *ins; ins = &(*ins)->next)
 725                         if (TC_U32_NODE(handle) < TC_U32_NODE((*ins)->handle))
 726                                 break;
 727
 728                 n->next = *ins;
 729                 wmb();
 730                 *ins = n;
 731
 732                 *arg = (unsigned long)n;
 733                 return 0;
 734         }
 735 #ifdef CONFIG_CLS_U32_PERF
 736         if (n && (NULL != n->pf))
 737                 kfree(n->pf);
 738 #endif
 739         kfree(n);
 740         return err;
 741 }
 742
 743 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 744 {
 745         struct tc_u_common *tp_c = tp->data;
 746         struct tc_u_hnode *ht;
 747         struct tc_u_knode *n;
 748         unsigned h;
 749
 750         if (arg->stop)
 751                 return;
 752
 753         for (ht = tp_c->hlist; ht; ht = ht->next) {
 754                 if (arg->count >= arg->skip) {
 755                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 756                                 arg->stop = 1;
 757                                 return;
 758                         }
 759                 }
 760                 arg->count++;
 761                 for (h = 0; h <= ht->divisor; h++) {
 762                         for (n = ht->ht[h]; n; n = n->next) {
 763                                 if (arg->count < arg->skip) {
 764                                         arg->count++;
 765                                         continue;
 766                                 }
 767                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 768                                         arg->stop = 1;
 769                                         return;
 770                                 }
 771                                 arg->count++;
 772                         }
 773                 }
 774         }
 775 }
 776
 777 static int u32_dump(struct tcf_proto *tp, unsigned long fh,
 778                      struct sk_buff *skb, struct tcmsg *t)
 779 {
 780         struct tc_u_knode *n = (struct tc_u_knode*)fh;
 781         unsigned char    *b = skb->tail;
 782         struct rtattr *rta;
 783
 784         if (n == NULL)
 785                 return skb->len;
 786
 787         t->tcm_handle = n->handle;
 788
 789         rta = (struct rtattr*)b;
 790         RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
 791
 792         if (TC_U32_KEY(n->handle) == 0) {
 793                 struct tc_u_hnode *ht = (struct tc_u_hnode*)fh;
 794                 u32 divisor = ht->divisor+1;
 795                 RTA_PUT(skb, TCA_U32_DIVISOR, 4, &divisor);
 796         } else {
 797                 RTA_PUT(skb, TCA_U32_SEL,
 798                         sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 799                         &n->sel);
 800                 if (n->ht_up) {
 801                         u32 htid = n->handle & 0xFFFFF000;
 802                         RTA_PUT(skb, TCA_U32_HASH, 4, &htid);
 803                 }
 804                 if (n->res.classid)
 805                         RTA_PUT(skb, TCA_U32_CLASSID, 4, &n->res.classid);
 806                 if (n->ht_down)
 807                         RTA_PUT(skb, TCA_U32_LINK, 4, &n->ht_down->handle);
 808 #ifdef CONFIG_NET_CLS_ACT
 809                 /* again for backward compatible mode - we want
 810                 *  to work with both old and new modes of entering
 811                 *  tc data even if iproute2  was newer - jhs
 812                 */
 813                 if (n->action) {
 814                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 815
 816                         if (n->action->type != TCA_OLD_COMPAT) {
 817                                 RTA_PUT(skb, TCA_U32_ACT, 0, NULL);
 818                                 if (tcf_action_dump(skb,n->action, 0, 0) < 0) {
 819                                         goto rtattr_failure;
 820                                 }
 821                         } else {
 822                                 RTA_PUT(skb, TCA_U32_POLICE, 0, NULL);
 823                                 if (tcf_action_dump_old(skb,n->action,0,0) < 0) {
 824                                         goto rtattr_failure;
 825                                 }
 826                         }
 827
 828                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 829                 }
 830
 831 #else
 832 #ifdef CONFIG_NET_CLS_POLICE
 833                 if (n->police) {
 834                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 835                         RTA_PUT(skb, TCA_U32_POLICE, 0, NULL);
 836
 837                         if (tcf_police_dump(skb, n->police) < 0)
 838                                 goto rtattr_failure;
 839
 840                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 841
 842                 }
 843 #endif
 844 #endif
 845
 846 #ifdef CONFIG_NET_CLS_IND
 847                 if(strlen(n->indev)) {
 848                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 849                         RTA_PUT(skb, TCA_U32_INDEV, IFNAMSIZ, n->indev);
 850                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 851                 }
 852 #endif
 853 #ifdef CONFIG_CLS_U32_PERF
 854                 RTA_PUT(skb, TCA_U32_PCNT,
 855                 sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(__u64),
 856                         n->pf);
 857 #endif
 858         }
 859
 860         rta->rta_len = skb->tail - b;
 861 #ifdef CONFIG_NET_CLS_ACT
 862         if (TC_U32_KEY(n->handle) != 0) {
 863                 if (TC_U32_KEY(n->handle) && n->action && n->action->type == TCA_OLD_COMPAT) {
 864                         if (tcf_action_copy_stats(skb,n->action))
 865                                 goto rtattr_failure;
 866                 }
 867         }
 868 #else
 869 #ifdef CONFIG_NET_CLS_POLICE
 870         if (TC_U32_KEY(n->handle) && n->police) {
 871                 if (qdisc_copy_stats(skb, &n->police->stats,
 872                                      n->police->stats_lock))
 873                         goto rtattr_failure;
 874         }
 875 #endif
 876 #endif
 877         return skb->len;
 878
 879 rtattr_failure:
 880         skb_trim(skb, b - skb->data);
 881         return -1;
 882 }
 883
 884 static struct tcf_proto_ops cls_u32_ops = {
 885         .next           =       NULL,
 886         .kind           =       "u32",
 887         .classify       =       u32_classify,
 888         .init           =       u32_init,
 889         .destroy        =       u32_destroy,
 890         .get            =       u32_get,
 891         .put            =       u32_put,
 892         .change         =       u32_change,
 893         .delete         =       u32_delete,
 894         .walk           =       u32_walk,
 895         .dump           =       u32_dump,
 896         .owner          =       THIS_MODULE,
 897 };
 898
 899 static int __init init_u32(void)
 900 {
 901         printk("u32 classifier\n");
 902 #ifdef CONFIG_CLS_U32_PERF
 903         printk("    Perfomance counters on\n");
 904 #endif
 905 #ifdef CONFIG_NET_CLS_POLICE
 906         printk("    OLD policer on \n");
 907 #endif
 908 #ifdef CONFIG_NET_CLS_IND
 909         printk("    input device check on \n");
 910 #endif
 911 #ifdef CONFIG_NET_CLS_ACT
 912         printk("    Actions configured \n");
 913 #endif
 914         return register_tcf_proto_ops(&cls_u32_ops);
 915 }
 916
 917 static void __exit exit_u32(void)
 918 {
 919         unregister_tcf_proto_ops(&cls_u32_ops);
 920 }
 921
 922 module_init(init_u32)
 923 module_exit(exit_u32)
 924 MODULE_LICENSE("GPL");