net/sched/cls_route.c

   1 /*
   2  * net/sched/cls_route.c        ROUTE4 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
  12 #include <linux/module.h>
  13 #include <linux/slab.h>
  14 #include <linux/types.h>
  15 #include <linux/kernel.h>
  16 #include <linux/string.h>
  17 #include <linux/errno.h>
  18 #include <linux/skbuff.h>
  19 #include <net/dst.h>
  20 #include <net/route.h>
  21 #include <net/netlink.h>
  22 #include <net/act_api.h>
  23 #include <net/pkt_cls.h>
  24
  25 /*
  26  * 1. For now we assume that route tags < 256.
  27  *    It allows to use direct table lookups, instead of hash tables.
  28  * 2. For now we assume that "from TAG" and "fromdev DEV" statements
  29  *    are mutually  exclusive.
  30  * 3. "to TAG from ANY" has higher priority, than "to ANY from XXX"
  31  */
  32 struct route4_fastmap {
  33         struct route4_filter            *filter;
  34         u32                             id;
  35         int                             iif;
  36 };
  37
  38 struct route4_head {
  39         struct route4_fastmap           fastmap[16];
  40         struct route4_bucket __rcu      *table[256 + 1];
  41         struct rcu_head                 rcu;
  42 };
  43
  44 struct route4_bucket {
  45         /* 16 FROM buckets + 16 IIF buckets + 1 wildcard bucket */
  46         struct route4_filter __rcu      *ht[16 + 16 + 1];
  47         struct rcu_head                 rcu;
  48 };
  49
  50 struct route4_filter {
  51         struct route4_filter __rcu      *next;
  52         u32                     id;
  53         int                     iif;
  54
  55         struct tcf_result       res;
  56         struct tcf_exts         exts;
  57         u32                     handle;
  58         struct route4_bucket    *bkt;
  59         struct tcf_proto        *tp;
  60         union {
  61                 struct work_struct      work;
  62                 struct rcu_head         rcu;
  63         };
  64 };
  65
  66 #define ROUTE4_FAILURE ((struct route4_filter *)(-1L))
  67
  68 static inline int route4_fastmap_hash(u32 id, int iif)
  69 {
  70         return id & 0xF;
  71 }
  72
  73 static DEFINE_SPINLOCK(fastmap_lock);
  74 static void
  75 route4_reset_fastmap(struct route4_head *head)
  76 {
  77         spin_lock_bh(&fastmap_lock);
  78         memset(head->fastmap, 0, sizeof(head->fastmap));
  79         spin_unlock_bh(&fastmap_lock);
  80 }
  81
  82 static void
  83 route4_set_fastmap(struct route4_head *head, u32 id, int iif,
  84                    struct route4_filter *f)
  85 {
  86         int h = route4_fastmap_hash(id, iif);
  87
  88         /* fastmap updates must look atomic to aling id, iff, filter */
  89         spin_lock_bh(&fastmap_lock);
  90         head->fastmap[h].id = id;
  91         head->fastmap[h].iif = iif;
  92         head->fastmap[h].filter = f;
  93         spin_unlock_bh(&fastmap_lock);
  94 }
  95
  96 static inline int route4_hash_to(u32 id)
  97 {
  98         return id & 0xFF;
  99 }
 100
 101 static inline int route4_hash_from(u32 id)
 102 {
 103         return (id >> 16) & 0xF;
 104 }
 105
 106 static inline int route4_hash_iif(int iif)
 107 {
 108         return 16 + ((iif >> 16) & 0xF);
 109 }
 110
 111 static inline int route4_hash_wild(void)
 112 {
 113         return 32;
 114 }
 115
 116 #define ROUTE4_APPLY_RESULT()                                   \
 117 {                                                               \
 118         *res = f->res;                                          \
 119         if (tcf_exts_has_actions(&f->exts)) {                   \
 120                 int r = tcf_exts_exec(skb, &f->exts, res);      \
 121                 if (r < 0) {                                    \
 122                         dont_cache = 1;                         \
 123                         continue;                               \
 124                 }                                               \
 125                 return r;                                       \
 126         } else if (!dont_cache)                                 \
 127                 route4_set_fastmap(head, id, iif, f);           \
 128         return 0;                                               \
 129 }
 130
 131 static int route4_classify(struct sk_buff *skb, const struct tcf_proto *tp,
 132                            struct tcf_result *res)
 133 {
 134         struct route4_head *head = rcu_dereference_bh(tp->root);
 135         struct dst_entry *dst;
 136         struct route4_bucket *b;
 137         struct route4_filter *f;
 138         u32 id, h;
 139         int iif, dont_cache = 0;
 140
 141         dst = skb_dst(skb);
 142         if (!dst)
 143                 goto failure;
 144
 145         id = dst->tclassid;
 146
 147         iif = inet_iif(skb);
 148
 149         h = route4_fastmap_hash(id, iif);
 150
 151         spin_lock(&fastmap_lock);
 152         if (id == head->fastmap[h].id &&
 153             iif == head->fastmap[h].iif &&
 154             (f = head->fastmap[h].filter) != NULL) {
 155                 if (f == ROUTE4_FAILURE) {
 156                         spin_unlock(&fastmap_lock);
 157                         goto failure;
 158                 }
 159
 160                 *res = f->res;
 161                 spin_unlock(&fastmap_lock);
 162                 return 0;
 163         }
 164         spin_unlock(&fastmap_lock);
 165
 166         h = route4_hash_to(id);
 167
 168 restart:
 169         b = rcu_dereference_bh(head->table[h]);
 170         if (b) {
 171                 for (f = rcu_dereference_bh(b->ht[route4_hash_from(id)]);
 172                      f;
 173                      f = rcu_dereference_bh(f->next))
 174                         if (f->id == id)
 175                                 ROUTE4_APPLY_RESULT();
 176
 177                 for (f = rcu_dereference_bh(b->ht[route4_hash_iif(iif)]);
 178                      f;
 179                      f = rcu_dereference_bh(f->next))
 180                         if (f->iif == iif)
 181                                 ROUTE4_APPLY_RESULT();
 182
 183                 for (f = rcu_dereference_bh(b->ht[route4_hash_wild()]);
 184                      f;
 185                      f = rcu_dereference_bh(f->next))
 186                         ROUTE4_APPLY_RESULT();
 187         }
 188         if (h < 256) {
 189                 h = 256;
 190                 id &= ~0xFFFF;
 191                 goto restart;
 192         }
 193
 194         if (!dont_cache)
 195                 route4_set_fastmap(head, id, iif, ROUTE4_FAILURE);
 196 failure:
 197         return -1;
 198 }
 199
 200 static inline u32 to_hash(u32 id)
 201 {
 202         u32 h = id & 0xFF;
 203
 204         if (id & 0x8000)
 205                 h += 256;
 206         return h;
 207 }
 208
 209 static inline u32 from_hash(u32 id)
 210 {
 211         id &= 0xFFFF;
 212         if (id == 0xFFFF)
 213                 return 32;
 214         if (!(id & 0x8000)) {
 215                 if (id > 255)
 216                         return 256;
 217                 return id & 0xF;
 218         }
 219         return 16 + (id & 0xF);
 220 }
 221
 222 static void *route4_get(struct tcf_proto *tp, u32 handle)
 223 {
 224         struct route4_head *head = rtnl_dereference(tp->root);
 225         struct route4_bucket *b;
 226         struct route4_filter *f;
 227         unsigned int h1, h2;
 228
 229         h1 = to_hash(handle);
 230         if (h1 > 256)
 231                 return NULL;
 232
 233         h2 = from_hash(handle >> 16);
 234         if (h2 > 32)
 235                 return NULL;
 236
 237         b = rtnl_dereference(head->table[h1]);
 238         if (b) {
 239                 for (f = rtnl_dereference(b->ht[h2]);
 240                      f;
 241                      f = rtnl_dereference(f->next))
 242                         if (f->handle == handle)
 243                                 return f;
 244         }
 245         return NULL;
 246 }
 247
 248 static int route4_init(struct tcf_proto *tp)
 249 {
 250         struct route4_head *head;
 251
 252         head = kzalloc(sizeof(struct route4_head), GFP_KERNEL);
 253         if (head == NULL)
 254                 return -ENOBUFS;
 255
 256         rcu_assign_pointer(tp->root, head);
 257         return 0;
 258 }
 259
 260 static void route4_delete_filter_work(struct work_struct *work)
 261 {
 262         struct route4_filter *f = container_of(work, struct route4_filter, work);
 263
 264         rtnl_lock();
 265         tcf_exts_destroy(&f->exts);
 266         kfree(f);
 267         rtnl_unlock();
 268 }
 269
 270 static void route4_delete_filter(struct rcu_head *head)
 271 {
 272         struct route4_filter *f = container_of(head, struct route4_filter, rcu);
 273
 274         INIT_WORK(&f->work, route4_delete_filter_work);
 275         tcf_queue_work(&f->work);
 276 }
 277
 278 static void route4_destroy(struct tcf_proto *tp)
 279 {
 280         struct route4_head *head = rtnl_dereference(tp->root);
 281         int h1, h2;
 282
 283         if (head == NULL)
 284                 return;
 285
 286         for (h1 = 0; h1 <= 256; h1++) {
 287                 struct route4_bucket *b;
 288
 289                 b = rtnl_dereference(head->table[h1]);
 290                 if (b) {
 291                         for (h2 = 0; h2 <= 32; h2++) {
 292                                 struct route4_filter *f;
 293
 294                                 while ((f = rtnl_dereference(b->ht[h2])) != NULL) {
 295                                         struct route4_filter *next;
 296
 297                                         next = rtnl_dereference(f->next);
 298                                         RCU_INIT_POINTER(b->ht[h2], next);
 299                                         tcf_unbind_filter(tp, &f->res);
 300                                         call_rcu(&f->rcu, route4_delete_filter);
 301                                 }
 302                         }
 303                         RCU_INIT_POINTER(head->table[h1], NULL);
 304                         kfree_rcu(b, rcu);
 305                 }
 306         }
 307         kfree_rcu(head, rcu);
 308 }
 309
 310 static int route4_delete(struct tcf_proto *tp, void *arg, bool *last)
 311 {
 312         struct route4_head *head = rtnl_dereference(tp->root);
 313         struct route4_filter *f = arg;
 314         struct route4_filter __rcu **fp;
 315         struct route4_filter *nf;
 316         struct route4_bucket *b;
 317         unsigned int h = 0;
 318         int i, h1;
 319
 320         if (!head || !f)
 321                 return -EINVAL;
 322
 323         h = f->handle;
 324         b = f->bkt;
 325
 326         fp = &b->ht[from_hash(h >> 16)];
 327         for (nf = rtnl_dereference(*fp); nf;
 328              fp = &nf->next, nf = rtnl_dereference(*fp)) {
 329                 if (nf == f) {
 330                         /* unlink it */
 331                         RCU_INIT_POINTER(*fp, rtnl_dereference(f->next));
 332
 333                         /* Remove any fastmap lookups that might ref filter
 334                          * notice we unlink'd the filter so we can't get it
 335                          * back in the fastmap.
 336                          */
 337                         route4_reset_fastmap(head);
 338
 339                         /* Delete it */
 340                         tcf_unbind_filter(tp, &f->res);
 341                         call_rcu(&f->rcu, route4_delete_filter);
 342
 343                         /* Strip RTNL protected tree */
 344                         for (i = 0; i <= 32; i++) {
 345                                 struct route4_filter *rt;
 346
 347                                 rt = rtnl_dereference(b->ht[i]);
 348                                 if (rt)
 349                                         goto out;
 350                         }
 351
 352                         /* OK, session has no flows */
 353                         RCU_INIT_POINTER(head->table[to_hash(h)], NULL);
 354                         kfree_rcu(b, rcu);
 355                         break;
 356                 }
 357         }
 358
 359 out:
 360         *last = true;
 361         for (h1 = 0; h1 <= 256; h1++) {
 362                 if (rcu_access_pointer(head->table[h1])) {
 363                         *last = false;
 364                         break;
 365                 }
 366         }
 367
 368         return 0;
 369 }
 370
 371 static const struct nla_policy route4_policy[TCA_ROUTE4_MAX + 1] = {
 372         [TCA_ROUTE4_CLASSID]    = { .type = NLA_U32 },
 373         [TCA_ROUTE4_TO]         = { .type = NLA_U32 },
 374         [TCA_ROUTE4_FROM]       = { .type = NLA_U32 },
 375         [TCA_ROUTE4_IIF]        = { .type = NLA_U32 },
 376 };
 377
 378 static int route4_set_parms(struct net *net, struct tcf_proto *tp,
 379                             unsigned long base, struct route4_filter *f,
 380                             u32 handle, struct route4_head *head,
 381                             struct nlattr **tb, struct nlattr *est, int new,
 382                             bool ovr)
 383 {
 384         u32 id = 0, to = 0, nhandle = 0x8000;
 385         struct route4_filter *fp;
 386         unsigned int h1;
 387         struct route4_bucket *b;
 388         int err;
 389
 390         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr);
 391         if (err < 0)
 392                 return err;
 393
 394         if (tb[TCA_ROUTE4_TO]) {
 395                 if (new && handle & 0x8000)
 396                         return -EINVAL;
 397                 to = nla_get_u32(tb[TCA_ROUTE4_TO]);
 398                 if (to > 0xFF)
 399                         return -EINVAL;
 400                 nhandle = to;
 401         }
 402
 403         if (tb[TCA_ROUTE4_FROM]) {
 404                 if (tb[TCA_ROUTE4_IIF])
 405                         return -EINVAL;
 406                 id = nla_get_u32(tb[TCA_ROUTE4_FROM]);
 407                 if (id > 0xFF)
 408                         return -EINVAL;
 409                 nhandle |= id << 16;
 410         } else if (tb[TCA_ROUTE4_IIF]) {
 411                 id = nla_get_u32(tb[TCA_ROUTE4_IIF]);
 412                 if (id > 0x7FFF)
 413                         return -EINVAL;
 414                 nhandle |= (id | 0x8000) << 16;
 415         } else
 416                 nhandle |= 0xFFFF << 16;
 417
 418         if (handle && new) {
 419                 nhandle |= handle & 0x7F00;
 420                 if (nhandle != handle)
 421                         return -EINVAL;
 422         }
 423
 424         h1 = to_hash(nhandle);
 425         b = rtnl_dereference(head->table[h1]);
 426         if (!b) {
 427                 b = kzalloc(sizeof(struct route4_bucket), GFP_KERNEL);
 428                 if (b == NULL)
 429                         return -ENOBUFS;
 430
 431                 rcu_assign_pointer(head->table[h1], b);
 432         } else {
 433                 unsigned int h2 = from_hash(nhandle >> 16);
 434
 435                 for (fp = rtnl_dereference(b->ht[h2]);
 436                      fp;
 437                      fp = rtnl_dereference(fp->next))
 438                         if (fp->handle == f->handle)
 439                                 return -EEXIST;
 440         }
 441
 442         if (tb[TCA_ROUTE4_TO])
 443                 f->id = to;
 444
 445         if (tb[TCA_ROUTE4_FROM])
 446                 f->id = to | id<<16;
 447         else if (tb[TCA_ROUTE4_IIF])
 448                 f->iif = id;
 449
 450         f->handle = nhandle;
 451         f->bkt = b;
 452         f->tp = tp;
 453
 454         if (tb[TCA_ROUTE4_CLASSID]) {
 455                 f->res.classid = nla_get_u32(tb[TCA_ROUTE4_CLASSID]);
 456                 tcf_bind_filter(tp, &f->res, base);
 457         }
 458
 459         return 0;
 460 }
 461
 462 static int route4_change(struct net *net, struct sk_buff *in_skb,
 463                          struct tcf_proto *tp, unsigned long base, u32 handle,
 464                          struct nlattr **tca, void **arg, bool ovr)
 465 {
 466         struct route4_head *head = rtnl_dereference(tp->root);
 467         struct route4_filter __rcu **fp;
 468         struct route4_filter *fold, *f1, *pfp, *f = NULL;
 469         struct route4_bucket *b;
 470         struct nlattr *opt = tca[TCA_OPTIONS];
 471         struct nlattr *tb[TCA_ROUTE4_MAX + 1];
 472         unsigned int h, th;
 473         int err;
 474         bool new = true;
 475
 476         if (opt == NULL)
 477                 return handle ? -EINVAL : 0;
 478
 479         err = nla_parse_nested(tb, TCA_ROUTE4_MAX, opt, route4_policy, NULL);
 480         if (err < 0)
 481                 return err;
 482
 483         fold = *arg;
 484         if (fold && handle && fold->handle != handle)
 485                         return -EINVAL;
 486
 487         err = -ENOBUFS;
 488         f = kzalloc(sizeof(struct route4_filter), GFP_KERNEL);
 489         if (!f)
 490                 goto errout;
 491
 492         err = tcf_exts_init(&f->exts, TCA_ROUTE4_ACT, TCA_ROUTE4_POLICE);
 493         if (err < 0)
 494                 goto errout;
 495
 496         if (fold) {
 497                 f->id = fold->id;
 498                 f->iif = fold->iif;
 499                 f->res = fold->res;
 500                 f->handle = fold->handle;
 501
 502                 f->tp = fold->tp;
 503                 f->bkt = fold->bkt;
 504                 new = false;
 505         }
 506
 507         err = route4_set_parms(net, tp, base, f, handle, head, tb,
 508                                tca[TCA_RATE], new, ovr);
 509         if (err < 0)
 510                 goto errout;
 511
 512         h = from_hash(f->handle >> 16);
 513         fp = &f->bkt->ht[h];
 514         for (pfp = rtnl_dereference(*fp);
 515              (f1 = rtnl_dereference(*fp)) != NULL;
 516              fp = &f1->next)
 517                 if (f->handle < f1->handle)
 518                         break;
 519
 520         netif_keep_dst(qdisc_dev(tp->q));
 521         rcu_assign_pointer(f->next, f1);
 522         rcu_assign_pointer(*fp, f);
 523
 524         if (fold && fold->handle && f->handle != fold->handle) {
 525                 th = to_hash(fold->handle);
 526                 h = from_hash(fold->handle >> 16);
 527                 b = rtnl_dereference(head->table[th]);
 528                 if (b) {
 529                         fp = &b->ht[h];
 530                         for (pfp = rtnl_dereference(*fp); pfp;
 531                              fp = &pfp->next, pfp = rtnl_dereference(*fp)) {
 532                                 if (pfp == f) {
 533                                         *fp = f->next;
 534                                         break;
 535                                 }
 536                         }
 537                 }
 538         }
 539
 540         route4_reset_fastmap(head);
 541         *arg = f;
 542         if (fold) {
 543                 tcf_unbind_filter(tp, &fold->res);
 544                 call_rcu(&fold->rcu, route4_delete_filter);
 545         }
 546         return 0;
 547
 548 errout:
 549         if (f)
 550                 tcf_exts_destroy(&f->exts);
 551         kfree(f);
 552         return err;
 553 }
 554
 555 static void route4_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 556 {
 557         struct route4_head *head = rtnl_dereference(tp->root);
 558         unsigned int h, h1;
 559
 560         if (head == NULL)
 561                 arg->stop = 1;
 562
 563         if (arg->stop)
 564                 return;
 565
 566         for (h = 0; h <= 256; h++) {
 567                 struct route4_bucket *b = rtnl_dereference(head->table[h]);
 568
 569                 if (b) {
 570                         for (h1 = 0; h1 <= 32; h1++) {
 571                                 struct route4_filter *f;
 572
 573                                 for (f = rtnl_dereference(b->ht[h1]);
 574                                      f;
 575                                      f = rtnl_dereference(f->next)) {
 576                                         if (arg->count < arg->skip) {
 577                                                 arg->count++;
 578                                                 continue;
 579                                         }
 580                                         if (arg->fn(tp, f, arg) < 0) {
 581                                                 arg->stop = 1;
 582                                                 return;
 583                                         }
 584                                         arg->count++;
 585                                 }
 586                         }
 587                 }
 588         }
 589 }
 590
 591 static int route4_dump(struct net *net, struct tcf_proto *tp, void *fh,
 592                        struct sk_buff *skb, struct tcmsg *t)
 593 {
 594         struct route4_filter *f = fh;
 595         struct nlattr *nest;
 596         u32 id;
 597
 598         if (f == NULL)
 599                 return skb->len;
 600
 601         t->tcm_handle = f->handle;
 602
 603         nest = nla_nest_start(skb, TCA_OPTIONS);
 604         if (nest == NULL)
 605                 goto nla_put_failure;
 606
 607         if (!(f->handle & 0x8000)) {
 608                 id = f->id & 0xFF;
 609                 if (nla_put_u32(skb, TCA_ROUTE4_TO, id))
 610                         goto nla_put_failure;
 611         }
 612         if (f->handle & 0x80000000) {
 613                 if ((f->handle >> 16) != 0xFFFF &&
 614                     nla_put_u32(skb, TCA_ROUTE4_IIF, f->iif))
 615                         goto nla_put_failure;
 616         } else {
 617                 id = f->id >> 16;
 618                 if (nla_put_u32(skb, TCA_ROUTE4_FROM, id))
 619                         goto nla_put_failure;
 620         }
 621         if (f->res.classid &&
 622             nla_put_u32(skb, TCA_ROUTE4_CLASSID, f->res.classid))
 623                 goto nla_put_failure;
 624
 625         if (tcf_exts_dump(skb, &f->exts) < 0)
 626                 goto nla_put_failure;
 627
 628         nla_nest_end(skb, nest);
 629
 630         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
 631                 goto nla_put_failure;
 632
 633         return skb->len;
 634
 635 nla_put_failure:
 636         nla_nest_cancel(skb, nest);
 637         return -1;
 638 }
 639
 640 static void route4_bind_class(void *fh, u32 classid, unsigned long cl)
 641 {
 642         struct route4_filter *f = fh;
 643
 644         if (f && f->res.classid == classid)
 645                 f->res.class = cl;
 646 }
 647
 648 static struct tcf_proto_ops cls_route4_ops __read_mostly = {
 649         .kind           =       "route",
 650         .classify       =       route4_classify,
 651         .init           =       route4_init,
 652         .destroy        =       route4_destroy,
 653         .get            =       route4_get,
 654         .change         =       route4_change,
 655         .delete         =       route4_delete,
 656         .walk           =       route4_walk,
 657         .dump           =       route4_dump,
 658         .bind_class     =       route4_bind_class,
 659         .owner          =       THIS_MODULE,
 660 };
 661
 662 static int __init init_route4(void)
 663 {
 664         return register_tcf_proto_ops(&cls_route4_ops);
 665 }
 666
 667 static void __exit exit_route4(void)
 668 {
 669         unregister_tcf_proto_ops(&cls_route4_ops);
 670 }
 671
 672 module_init(init_route4)
 673 module_exit(exit_route4)
 674 MODULE_LICENSE("GPL");