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);
 179                                         if (skb->tc_classid > 0) {
 180                                                 res->classid = skb->tc_classid;
 181                                                 skb->tc_classid = 0;
 182                                         }
 183
 184                                         if (pol_res >= 0)
 185                                                 return pol_res;
 186                                 } else
 187 #else
 188 #ifdef CONFIG_NET_CLS_POLICE
 189                                 if (n->police) {
 190                                         int pol_res = tcf_police(skb, n->police);
 191                                         if (pol_res >= 0)
 192                                                 return pol_res;
 193                                 } else
 194 #endif
 195 #endif
 196                                         return 0;
 197                         }
 198                         n = n->next;
 199                         goto next_knode;
 200                 }
 201
 202                 /* PUSH */
 203                 if (sdepth >= TC_U32_MAXDEPTH)
 204                         goto deadloop;
 205                 stack[sdepth].knode = n;
 206                 stack[sdepth].ptr = ptr;
 207                 sdepth++;
 208
 209                 ht = n->ht_down;
 210                 sel = 0;
 211                 if (ht->divisor)
 212                         sel = ht->divisor&u32_hash_fold(*(u32*)(ptr+n->sel.hoff), &n->sel,n->fshift);
 213
 214                 if (!(n->sel.flags&(TC_U32_VAROFFSET|TC_U32_OFFSET|TC_U32_EAT)))
 215                         goto next_ht;
 216
 217                 if (n->sel.flags&(TC_U32_OFFSET|TC_U32_VAROFFSET)) {
 218                         off2 = n->sel.off + 3;
 219                         if (n->sel.flags&TC_U32_VAROFFSET)
 220                                 off2 += ntohs(n->sel.offmask & *(u16*)(ptr+n->sel.offoff)) >>n->sel.offshift;
 221                         off2 &= ~3;
 222                 }
 223                 if (n->sel.flags&TC_U32_EAT) {
 224                         ptr += off2;
 225                         off2 = 0;
 226                 }
 227
 228                 if (ptr < skb->tail)
 229                         goto next_ht;
 230         }
 231
 232         /* POP */
 233         if (sdepth--) {
 234                 n = stack[sdepth].knode;
 235                 ht = n->ht_up;
 236                 ptr = stack[sdepth].ptr;
 237                 goto check_terminal;
 238         }
 239         return -1;
 240
 241 deadloop:
 242         if (net_ratelimit())
 243                 printk("cls_u32: dead loop\n");
 244         return -1;
 245 }
 246
 247 static __inline__ struct tc_u_hnode *
 248 u32_lookup_ht(struct tc_u_common *tp_c, u32 handle)
 249 {
 250         struct tc_u_hnode *ht;
 251
 252         for (ht = tp_c->hlist; ht; ht = ht->next)
 253                 if (ht->handle == handle)
 254                         break;
 255
 256         return ht;
 257 }
 258
 259 static __inline__ struct tc_u_knode *
 260 u32_lookup_key(struct tc_u_hnode *ht, u32 handle)
 261 {
 262         unsigned sel;
 263         struct tc_u_knode *n = NULL;
 264
 265         sel = TC_U32_HASH(handle);
 266         if (sel > ht->divisor)
 267                 goto out;
 268
 269         for (n = ht->ht[sel]; n; n = n->next)
 270                 if (n->handle == handle)
 271                         break;
 272 out:
 273         return n;
 274 }
 275
 276
 277 static unsigned long u32_get(struct tcf_proto *tp, u32 handle)
 278 {
 279         struct tc_u_hnode *ht;
 280         struct tc_u_common *tp_c = tp->data;
 281
 282         if (TC_U32_HTID(handle) == TC_U32_ROOT)
 283                 ht = tp->root;
 284         else
 285                 ht = u32_lookup_ht(tp_c, TC_U32_HTID(handle));
 286
 287         if (!ht)
 288                 return 0;
 289
 290         if (TC_U32_KEY(handle) == 0)
 291                 return (unsigned long)ht;
 292
 293         return (unsigned long)u32_lookup_key(ht, handle);
 294 }
 295
 296 static void u32_put(struct tcf_proto *tp, unsigned long f)
 297 {
 298 }
 299
 300 static u32 gen_new_htid(struct tc_u_common *tp_c)
 301 {
 302         int i = 0x800;
 303
 304         do {
 305                 if (++tp_c->hgenerator == 0x7FF)
 306                         tp_c->hgenerator = 1;
 307         } while (--i>0 && u32_lookup_ht(tp_c, (tp_c->hgenerator|0x800)<<20));
 308
 309         return i > 0 ? (tp_c->hgenerator|0x800)<<20 : 0;
 310 }
 311
 312 static int u32_init(struct tcf_proto *tp)
 313 {
 314         struct tc_u_hnode *root_ht;
 315         struct tc_u_common *tp_c;
 316
 317         for (tp_c = u32_list; tp_c; tp_c = tp_c->next)
 318                 if (tp_c->q == tp->q)
 319                         break;
 320
 321         root_ht = kmalloc(sizeof(*root_ht), GFP_KERNEL);
 322         if (root_ht == NULL)
 323                 return -ENOBUFS;
 324
 325         memset(root_ht, 0, sizeof(*root_ht));
 326         root_ht->divisor = 0;
 327         root_ht->refcnt++;
 328         root_ht->handle = tp_c ? gen_new_htid(tp_c) : 0x80000000;
 329
 330         if (tp_c == NULL) {
 331                 tp_c = kmalloc(sizeof(*tp_c), GFP_KERNEL);
 332                 if (tp_c == NULL) {
 333                         kfree(root_ht);
 334                         return -ENOBUFS;
 335                 }
 336                 memset(tp_c, 0, sizeof(*tp_c));
 337                 tp_c->q = tp->q;
 338                 tp_c->next = u32_list;
 339                 u32_list = tp_c;
 340         }
 341
 342         tp_c->refcnt++;
 343         root_ht->next = tp_c->hlist;
 344         tp_c->hlist = root_ht;
 345         root_ht->tp_c = tp_c;
 346
 347         tp->root = root_ht;
 348         tp->data = tp_c;
 349         return 0;
 350 }
 351
 352 static int u32_destroy_key(struct tcf_proto *tp, struct tc_u_knode *n)
 353 {
 354         unsigned long cl;
 355
 356         if ((cl = __cls_set_class(&n->res.class, 0)) != 0)
 357                 tp->q->ops->cl_ops->unbind_tcf(tp->q, cl);
 358 #ifdef CONFIG_NET_CLS_ACT
 359         if (n->action) {
 360                 tcf_action_destroy(n->action, TCA_ACT_UNBIND);
 361         }
 362 #else
 363 #ifdef CONFIG_NET_CLS_POLICE
 364         tcf_police_release(n->police, TCA_ACT_UNBIND);
 365 #endif
 366 #endif
 367         if (n->ht_down)
 368                 n->ht_down->refcnt--;
 369 #ifdef CONFIG_CLS_U32_PERF
 370         if (n && (NULL != n->pf))
 371                 kfree(n->pf);
 372 #endif
 373         kfree(n);
 374         return 0;
 375 }
 376
 377 static int u32_delete_key(struct tcf_proto *tp, struct tc_u_knode* key)
 378 {
 379         struct tc_u_knode **kp;
 380         struct tc_u_hnode *ht = key->ht_up;
 381
 382         if (ht) {
 383                 for (kp = &ht->ht[TC_U32_HASH(key->handle)]; *kp; kp = &(*kp)->next) {
 384                         if (*kp == key) {
 385                                 tcf_tree_lock(tp);
 386                                 *kp = key->next;
 387                                 tcf_tree_unlock(tp);
 388
 389                                 u32_destroy_key(tp, key);
 390                                 return 0;
 391                         }
 392                 }
 393         }
 394         BUG_TRAP(0);
 395         return 0;
 396 }
 397
 398 static void u32_clear_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 399 {
 400         struct tc_u_knode *n;
 401         unsigned h;
 402
 403         for (h=0; h<=ht->divisor; h++) {
 404                 while ((n = ht->ht[h]) != NULL) {
 405                         ht->ht[h] = n->next;
 406
 407                         u32_destroy_key(tp, n);
 408                 }
 409         }
 410 }
 411
 412 static int u32_destroy_hnode(struct tcf_proto *tp, struct tc_u_hnode *ht)
 413 {
 414         struct tc_u_common *tp_c = tp->data;
 415         struct tc_u_hnode **hn;
 416
 417         BUG_TRAP(!ht->refcnt);
 418
 419         u32_clear_hnode(tp, ht);
 420
 421         for (hn = &tp_c->hlist; *hn; hn = &(*hn)->next) {
 422                 if (*hn == ht) {
 423                         *hn = ht->next;
 424                         kfree(ht);
 425                         return 0;
 426                 }
 427         }
 428
 429         BUG_TRAP(0);
 430         return -ENOENT;
 431 }
 432
 433 static void u32_destroy(struct tcf_proto *tp)
 434 {
 435         struct tc_u_common *tp_c = tp->data;
 436         struct tc_u_hnode *root_ht = xchg(&tp->root, NULL);
 437
 438         BUG_TRAP(root_ht != NULL);
 439
 440         if (root_ht && --root_ht->refcnt == 0)
 441                 u32_destroy_hnode(tp, root_ht);
 442
 443         if (--tp_c->refcnt == 0) {
 444                 struct tc_u_hnode *ht;
 445                 struct tc_u_common **tp_cp;
 446
 447                 for (tp_cp = &u32_list; *tp_cp; tp_cp = &(*tp_cp)->next) {
 448                         if (*tp_cp == tp_c) {
 449                                 *tp_cp = tp_c->next;
 450                                 break;
 451                         }
 452                 }
 453
 454                 for (ht=tp_c->hlist; ht; ht = ht->next)
 455                         u32_clear_hnode(tp, ht);
 456
 457                 while ((ht = tp_c->hlist) != NULL) {
 458                         tp_c->hlist = ht->next;
 459
 460                         BUG_TRAP(ht->refcnt == 0);
 461
 462                         kfree(ht);
 463                 };
 464
 465                 kfree(tp_c);
 466         }
 467
 468         tp->data = NULL;
 469 }
 470
 471 static int u32_delete(struct tcf_proto *tp, unsigned long arg)
 472 {
 473         struct tc_u_hnode *ht = (struct tc_u_hnode*)arg;
 474
 475         if (ht == NULL)
 476                 return 0;
 477
 478         if (TC_U32_KEY(ht->handle))
 479                 return u32_delete_key(tp, (struct tc_u_knode*)ht);
 480
 481         if (tp->root == ht)
 482                 return -EINVAL;
 483
 484         if (--ht->refcnt == 0)
 485                 u32_destroy_hnode(tp, ht);
 486
 487         return 0;
 488 }
 489
 490 static u32 gen_new_kid(struct tc_u_hnode *ht, u32 handle)
 491 {
 492         struct tc_u_knode *n;
 493         unsigned i = 0x7FF;
 494
 495         for (n=ht->ht[TC_U32_HASH(handle)]; n; n = n->next)
 496                 if (i < TC_U32_NODE(n->handle))
 497                         i = TC_U32_NODE(n->handle);
 498         i++;
 499
 500         return handle|(i>0xFFF ? 0xFFF : i);
 501 }
 502
 503 static int u32_set_parms(struct Qdisc *q, unsigned long base,
 504                          struct tc_u_hnode *ht,
 505                          struct tc_u_knode *n, struct rtattr **tb,
 506                          struct rtattr *est)
 507 {
 508 #ifdef CONFIG_NET_CLS_ACT
 509         struct tc_action *act = NULL;
 510         int ret;
 511 #endif
 512         if (tb[TCA_U32_LINK-1]) {
 513                 u32 handle = *(u32*)RTA_DATA(tb[TCA_U32_LINK-1]);
 514                 struct tc_u_hnode *ht_down = NULL;
 515
 516                 if (TC_U32_KEY(handle))
 517                         return -EINVAL;
 518
 519                 if (handle) {
 520                         ht_down = u32_lookup_ht(ht->tp_c, handle);
 521
 522                         if (ht_down == NULL)
 523                                 return -EINVAL;
 524                         ht_down->refcnt++;
 525                 }
 526
 527                 sch_tree_lock(q);
 528                 ht_down = xchg(&n->ht_down, ht_down);
 529                 sch_tree_unlock(q);
 530
 531                 if (ht_down)
 532                         ht_down->refcnt--;
 533         }
 534         if (tb[TCA_U32_CLASSID-1]) {
 535                 unsigned long cl;
 536
 537                 n->res.classid = *(u32*)RTA_DATA(tb[TCA_U32_CLASSID-1]);
 538                 sch_tree_lock(q);
 539                 cl = __cls_set_class(&n->res.class, q->ops->cl_ops->bind_tcf(q, base, n->res.classid));
 540                 sch_tree_unlock(q);
 541                 if (cl)
 542                         q->ops->cl_ops->unbind_tcf(q, cl);
 543         }
 544 #ifdef CONFIG_NET_CLS_ACT
 545         /*backward compatibility */
 546         if (tb[TCA_U32_POLICE-1])
 547         {
 548                 act = kmalloc(sizeof(*act),GFP_KERNEL);
 549                 if (NULL == act)
 550                         return -ENOMEM;
 551
 552                 memset(act,0,sizeof(*act));
 553                 ret = tcf_action_init_1(tb[TCA_U32_POLICE-1], est,act,"police", TCA_ACT_NOREPLACE, TCA_ACT_BIND);
 554                 if (0 > ret){
 555                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 556                         return ret;
 557                 }
 558                 act->type = TCA_OLD_COMPAT;
 559
 560                 sch_tree_lock(q);
 561                 act = xchg(&n->action, act);
 562                 sch_tree_unlock(q);
 563
 564                 tcf_action_destroy(act, TCA_ACT_UNBIND);
 565
 566         }
 567
 568         if(tb[TCA_U32_ACT-1]) {
 569                 act = kmalloc(sizeof(*act),GFP_KERNEL);
 570                 if (NULL == act)
 571                         return -ENOMEM;
 572                 memset(act,0,sizeof(*act));
 573                 ret = tcf_action_init(tb[TCA_U32_ACT-1], est,act,NULL,TCA_ACT_NOREPLACE, TCA_ACT_BIND);
 574                 if (0 > ret) {
 575                         tcf_action_destroy(act, TCA_ACT_UNBIND);
 576                         return ret;
 577                 }
 578
 579                 sch_tree_lock(q);
 580                 act = xchg(&n->action, act);
 581                 sch_tree_unlock(q);
 582
 583                 tcf_action_destroy(act, TCA_ACT_UNBIND);
 584         }
 585
 586
 587 #else
 588 #ifdef CONFIG_NET_CLS_POLICE
 589         if (tb[TCA_U32_POLICE-1]) {
 590                 struct tcf_police *police = tcf_police_locate(tb[TCA_U32_POLICE-1], est);
 591                 sch_tree_lock(q);
 592                 police = xchg(&n->police, police);
 593                 sch_tree_unlock(q);
 594                 tcf_police_release(police, TCA_ACT_UNBIND);
 595         }
 596 #endif
 597 #endif
 598 #ifdef CONFIG_NET_CLS_IND
 599         n->indev[0] = 0;
 600         if(tb[TCA_U32_INDEV-1]) {
 601                 struct rtattr *input_dev = tb[TCA_U32_INDEV-1];
 602                 if (RTA_PAYLOAD(input_dev) >= IFNAMSIZ) {
 603                         printk("cls_u32: bad indev name %s\n",(char*)RTA_DATA(input_dev));
 604                         /* should we clear state first? */
 605                         return  -EINVAL;
 606                 }
 607                 sprintf(n->indev, "%s", (char*)RTA_DATA(input_dev));
 608                 printk("got IND %s\n",n->indev);
 609         }
 610 #endif
 611
 612         return 0;
 613 }
 614
 615 static int u32_change(struct tcf_proto *tp, unsigned long base, u32 handle,
 616                       struct rtattr **tca,
 617                       unsigned long *arg)
 618 {
 619         struct tc_u_common *tp_c = tp->data;
 620         struct tc_u_hnode *ht;
 621         struct tc_u_knode *n;
 622         struct tc_u32_sel *s;
 623         struct rtattr *opt = tca[TCA_OPTIONS-1];
 624         struct rtattr *tb[TCA_U32_MAX];
 625         u32 htid;
 626         int err;
 627
 628         if (opt == NULL)
 629                 return handle ? -EINVAL : 0;
 630
 631         if (rtattr_parse(tb, TCA_U32_MAX, RTA_DATA(opt), RTA_PAYLOAD(opt)) < 0)
 632                 return -EINVAL;
 633
 634         if ((n = (struct tc_u_knode*)*arg) != NULL) {
 635                 if (TC_U32_KEY(n->handle) == 0)
 636                         return -EINVAL;
 637
 638                 return u32_set_parms(tp->q, base, n->ht_up, n, tb, tca[TCA_RATE-1]);
 639         }
 640
 641         if (tb[TCA_U32_DIVISOR-1]) {
 642                 unsigned divisor = *(unsigned*)RTA_DATA(tb[TCA_U32_DIVISOR-1]);
 643
 644                 if (--divisor > 0x100)
 645                         return -EINVAL;
 646                 if (TC_U32_KEY(handle))
 647                         return -EINVAL;
 648                 if (handle == 0) {
 649                         handle = gen_new_htid(tp->data);
 650                         if (handle == 0)
 651                                 return -ENOMEM;
 652                 }
 653                 ht = kmalloc(sizeof(*ht) + divisor*sizeof(void*), GFP_KERNEL);
 654                 if (ht == NULL)
 655                         return -ENOBUFS;
 656                 memset(ht, 0, sizeof(*ht) + divisor*sizeof(void*));
 657                 ht->tp_c = tp_c;
 658                 ht->refcnt = 0;
 659                 ht->divisor = divisor;
 660                 ht->handle = handle;
 661                 ht->next = tp_c->hlist;
 662                 tp_c->hlist = ht;
 663                 *arg = (unsigned long)ht;
 664                 return 0;
 665         }
 666
 667         if (tb[TCA_U32_HASH-1]) {
 668                 htid = *(unsigned*)RTA_DATA(tb[TCA_U32_HASH-1]);
 669                 if (TC_U32_HTID(htid) == TC_U32_ROOT) {
 670                         ht = tp->root;
 671                         htid = ht->handle;
 672                 } else {
 673                         ht = u32_lookup_ht(tp->data, TC_U32_HTID(htid));
 674                         if (ht == NULL)
 675                                 return -EINVAL;
 676                 }
 677         } else {
 678                 ht = tp->root;
 679                 htid = ht->handle;
 680         }
 681
 682         if (ht->divisor < TC_U32_HASH(htid))
 683                 return -EINVAL;
 684
 685         if (handle) {
 686                 if (TC_U32_HTID(handle) && TC_U32_HTID(handle^htid))
 687                         return -EINVAL;
 688                 handle = htid | TC_U32_NODE(handle);
 689         } else
 690                 handle = gen_new_kid(ht, htid);
 691
 692         if (tb[TCA_U32_SEL-1] == 0 ||
 693             RTA_PAYLOAD(tb[TCA_U32_SEL-1]) < sizeof(struct tc_u32_sel))
 694                 return -EINVAL;
 695
 696         s = RTA_DATA(tb[TCA_U32_SEL-1]);
 697
 698         n = kmalloc(sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key), GFP_KERNEL);
 699         if (n == NULL)
 700                 return -ENOBUFS;
 701
 702         memset(n, 0, sizeof(*n) + s->nkeys*sizeof(struct tc_u32_key));
 703 #ifdef CONFIG_CLS_U32_PERF
 704         n->pf = kmalloc(sizeof(struct tc_u32_pcnt) + s->nkeys*sizeof(__u64), GFP_KERNEL);
 705         if (n->pf == NULL) {
 706                 kfree(n);
 707                 return -ENOBUFS;
 708         }
 709         memset(n->pf, 0, sizeof(struct tc_u32_pcnt) + s->nkeys*sizeof(__u64));
 710 #endif
 711
 712         memcpy(&n->sel, s, sizeof(*s) + s->nkeys*sizeof(struct tc_u32_key));
 713         n->ht_up = ht;
 714         n->handle = handle;
 715 {
 716         u8 i = 0;
 717         u32 mask = s->hmask;
 718         if (mask) {
 719                 while (!(mask & 1)) {
 720                         i++;
 721                         mask>>=1;
 722                 }
 723         }
 724         n->fshift = i;
 725 }
 726         err = u32_set_parms(tp->q, base, ht, n, tb, tca[TCA_RATE-1]);
 727         if (err == 0) {
 728                 struct tc_u_knode **ins;
 729                 for (ins = &ht->ht[TC_U32_HASH(handle)]; *ins; ins = &(*ins)->next)
 730                         if (TC_U32_NODE(handle) < TC_U32_NODE((*ins)->handle))
 731                                 break;
 732
 733                 n->next = *ins;
 734                 wmb();
 735                 *ins = n;
 736
 737                 *arg = (unsigned long)n;
 738                 return 0;
 739         }
 740 #ifdef CONFIG_CLS_U32_PERF
 741         if (n && (NULL != n->pf))
 742                 kfree(n->pf);
 743 #endif
 744         kfree(n);
 745         return err;
 746 }
 747
 748 static void u32_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 749 {
 750         struct tc_u_common *tp_c = tp->data;
 751         struct tc_u_hnode *ht;
 752         struct tc_u_knode *n;
 753         unsigned h;
 754
 755         if (arg->stop)
 756                 return;
 757
 758         for (ht = tp_c->hlist; ht; ht = ht->next) {
 759                 if (arg->count >= arg->skip) {
 760                         if (arg->fn(tp, (unsigned long)ht, arg) < 0) {
 761                                 arg->stop = 1;
 762                                 return;
 763                         }
 764                 }
 765                 arg->count++;
 766                 for (h = 0; h <= ht->divisor; h++) {
 767                         for (n = ht->ht[h]; n; n = n->next) {
 768                                 if (arg->count < arg->skip) {
 769                                         arg->count++;
 770                                         continue;
 771                                 }
 772                                 if (arg->fn(tp, (unsigned long)n, arg) < 0) {
 773                                         arg->stop = 1;
 774                                         return;
 775                                 }
 776                                 arg->count++;
 777                         }
 778                 }
 779         }
 780 }
 781
 782 static int u32_dump(struct tcf_proto *tp, unsigned long fh,
 783                      struct sk_buff *skb, struct tcmsg *t)
 784 {
 785         struct tc_u_knode *n = (struct tc_u_knode*)fh;
 786         unsigned char    *b = skb->tail;
 787         struct rtattr *rta;
 788
 789         if (n == NULL)
 790                 return skb->len;
 791
 792         t->tcm_handle = n->handle;
 793
 794         rta = (struct rtattr*)b;
 795         RTA_PUT(skb, TCA_OPTIONS, 0, NULL);
 796
 797         if (TC_U32_KEY(n->handle) == 0) {
 798                 struct tc_u_hnode *ht = (struct tc_u_hnode*)fh;
 799                 u32 divisor = ht->divisor+1;
 800                 RTA_PUT(skb, TCA_U32_DIVISOR, 4, &divisor);
 801         } else {
 802                 RTA_PUT(skb, TCA_U32_SEL,
 803                         sizeof(n->sel) + n->sel.nkeys*sizeof(struct tc_u32_key),
 804                         &n->sel);
 805                 if (n->ht_up) {
 806                         u32 htid = n->handle & 0xFFFFF000;
 807                         RTA_PUT(skb, TCA_U32_HASH, 4, &htid);
 808                 }
 809                 if (n->res.classid)
 810                         RTA_PUT(skb, TCA_U32_CLASSID, 4, &n->res.classid);
 811                 if (n->ht_down)
 812                         RTA_PUT(skb, TCA_U32_LINK, 4, &n->ht_down->handle);
 813 #ifdef CONFIG_NET_CLS_ACT
 814                 /* again for backward compatible mode - we want
 815                 *  to work with both old and new modes of entering
 816                 *  tc data even if iproute2  was newer - jhs
 817                 */
 818                 if (n->action) {
 819                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 820
 821                         if (n->action->type != TCA_OLD_COMPAT) {
 822                                 RTA_PUT(skb, TCA_U32_ACT, 0, NULL);
 823                                 if (tcf_action_dump(skb,n->action, 0, 0) < 0) {
 824                                         goto rtattr_failure;
 825                                 }
 826                         } else {
 827                                 RTA_PUT(skb, TCA_U32_POLICE, 0, NULL);
 828                                 if (tcf_action_dump_old(skb,n->action,0,0) < 0) {
 829                                         goto rtattr_failure;
 830                                 }
 831                         }
 832
 833                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 834                 }
 835
 836 #else
 837 #ifdef CONFIG_NET_CLS_POLICE
 838                 if (n->police) {
 839                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 840                         RTA_PUT(skb, TCA_U32_POLICE, 0, NULL);
 841
 842                         if (tcf_police_dump(skb, n->police) < 0)
 843                                 goto rtattr_failure;
 844
 845                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 846
 847                 }
 848 #endif
 849 #endif
 850
 851 #ifdef CONFIG_NET_CLS_IND
 852                 if(strlen(n->indev)) {
 853                         struct rtattr * p_rta = (struct rtattr*)skb->tail;
 854                         RTA_PUT(skb, TCA_U32_INDEV, IFNAMSIZ, n->indev);
 855                         p_rta->rta_len = skb->tail - (u8*)p_rta;
 856                 }
 857 #endif
 858 #ifdef CONFIG_CLS_U32_PERF
 859                 RTA_PUT(skb, TCA_U32_PCNT,
 860                 sizeof(struct tc_u32_pcnt) + n->sel.nkeys*sizeof(__u64),
 861                         n->pf);
 862 #endif
 863         }
 864
 865         rta->rta_len = skb->tail - b;
 866 #ifdef CONFIG_NET_CLS_ACT
 867         if (TC_U32_KEY(n->handle) != 0) {
 868                 if (TC_U32_KEY(n->handle) && n->action && n->action->type == TCA_OLD_COMPAT) {
 869                         if (tcf_action_copy_stats(skb,n->action))
 870                                 goto rtattr_failure;
 871                 }
 872         }
 873 #else
 874 #ifdef CONFIG_NET_CLS_POLICE
 875         if (TC_U32_KEY(n->handle) && n->police) {
 876                 if (qdisc_copy_stats(skb, &n->police->stats,
 877                                      n->police->stats_lock))
 878                         goto rtattr_failure;
 879         }
 880 #endif
 881 #endif
 882         return skb->len;
 883
 884 rtattr_failure:
 885         skb_trim(skb, b - skb->data);
 886         return -1;
 887 }
 888
 889 static struct tcf_proto_ops cls_u32_ops = {
 890         .next           =       NULL,
 891         .kind           =       "u32",
 892         .classify       =       u32_classify,
 893         .init           =       u32_init,
 894         .destroy        =       u32_destroy,
 895         .get            =       u32_get,
 896         .put            =       u32_put,
 897         .change         =       u32_change,
 898         .delete         =       u32_delete,
 899         .walk           =       u32_walk,
 900         .dump           =       u32_dump,
 901         .owner          =       THIS_MODULE,
 902 };
 903
 904 static int __init init_u32(void)
 905 {
 906         printk("u32 classifier\n");
 907 #ifdef CONFIG_CLS_U32_PERF
 908         printk("    Perfomance counters on\n");
 909 #endif
 910 #ifdef CONFIG_NET_CLS_POLICE
 911         printk("    OLD policer on \n");
 912 #endif
 913 #ifdef CONFIG_NET_CLS_IND
 914         printk("    input device check on \n");
 915 #endif
 916 #ifdef CONFIG_NET_CLS_ACT
 917         printk("    Actions configured \n");
 918 #endif
 919         return register_tcf_proto_ops(&cls_u32_ops);
 920 }
 921
 922 static void __exit exit_u32(void)
 923 {
 924         unregister_tcf_proto_ops(&cls_u32_ops);
 925 }
 926
 927 module_init(init_u32)
 928 module_exit(exit_u32)
 929 MODULE_LICENSE("GPL");