net/sched/cls_tcindex.c

   1 /*
   2  * net/sched/cls_tcindex.c      Packet classifier for skb->tc_index
   3  *
   4  * Written 1998,1999 by Werner Almesberger, EPFL ICA
   5  */
   6
   7 #include <linux/module.h>
   8 #include <linux/types.h>
   9 #include <linux/kernel.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/errno.h>
  12 #include <linux/slab.h>
  13 #include <net/act_api.h>
  14 #include <net/netlink.h>
  15 #include <net/pkt_cls.h>
  16
  17 /*
  18  * Passing parameters to the root seems to be done more awkwardly than really
  19  * necessary. At least, u32 doesn't seem to use such dirty hacks. To be
  20  * verified. FIXME.
  21  */
  22
  23 #define PERFECT_HASH_THRESHOLD  64      /* use perfect hash if not bigger */
  24 #define DEFAULT_HASH_SIZE       64      /* optimized for diffserv */
  25
  26
  27 struct tcindex_filter_result {
  28         struct tcf_exts         exts;
  29         struct tcf_result       res;
  30         struct rcu_head         rcu;
  31 };
  32
  33 struct tcindex_filter {
  34         u16 key;
  35         struct tcindex_filter_result result;
  36         struct tcindex_filter __rcu *next;
  37         struct rcu_head rcu;
  38 };
  39
  40
  41 struct tcindex_data {
  42         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  43         struct tcindex_filter __rcu **h; /* imperfect hash; */
  44         struct tcf_proto *tp;
  45         u16 mask;               /* AND key with mask */
  46         u32 shift;              /* shift ANDed key to the right */
  47         u32 hash;               /* hash table size; 0 if undefined */
  48         u32 alloc_hash;         /* allocated size */
  49         u32 fall_through;       /* 0: only classify if explicit match */
  50         struct rcu_head rcu;
  51 };
  52
  53 static inline int
  54 tcindex_filter_is_set(struct tcindex_filter_result *r)
  55 {
  56         return tcf_exts_is_predicative(&r->exts) || r->res.classid;
  57 }
  58
  59 static struct tcindex_filter_result *
  60 tcindex_lookup(struct tcindex_data *p, u16 key)
  61 {
  62         if (p->perfect) {
  63                 struct tcindex_filter_result *f = p->perfect + key;
  64
  65                 return tcindex_filter_is_set(f) ? f : NULL;
  66         } else if (p->h) {
  67                 struct tcindex_filter __rcu **fp;
  68                 struct tcindex_filter *f;
  69
  70                 fp = &p->h[key % p->hash];
  71                 for (f = rcu_dereference_bh_rtnl(*fp);
  72                      f;
  73                      fp = &f->next, f = rcu_dereference_bh_rtnl(*fp))
  74                         if (f->key == key)
  75                                 return &f->result;
  76         }
  77
  78         return NULL;
  79 }
  80
  81
  82 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  83                             struct tcf_result *res)
  84 {
  85         struct tcindex_data *p = rcu_dereference_bh(tp->root);
  86         struct tcindex_filter_result *f;
  87         int key = (skb->tc_index & p->mask) >> p->shift;
  88
  89         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
  90                  skb, tp, res, p);
  91
  92         f = tcindex_lookup(p, key);
  93         if (!f) {
  94                 if (!p->fall_through)
  95                         return -1;
  96                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
  97                 res->class = 0;
  98                 pr_debug("alg 0x%x\n", res->classid);
  99                 return 0;
 100         }
 101         *res = f->res;
 102         pr_debug("map 0x%x\n", res->classid);
 103
 104         return tcf_exts_exec(skb, &f->exts, res);
 105 }
 106
 107
 108 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
 109 {
 110         struct tcindex_data *p = rtnl_dereference(tp->root);
 111         struct tcindex_filter_result *r;
 112
 113         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
 114         if (p->perfect && handle >= p->alloc_hash)
 115                 return 0;
 116         r = tcindex_lookup(p, handle);
 117         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 118 }
 119
 120 static int tcindex_init(struct tcf_proto *tp)
 121 {
 122         struct tcindex_data *p;
 123
 124         pr_debug("tcindex_init(tp %p)\n", tp);
 125         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
 126         if (!p)
 127                 return -ENOMEM;
 128
 129         p->mask = 0xffff;
 130         p->hash = DEFAULT_HASH_SIZE;
 131         p->fall_through = 1;
 132
 133         rcu_assign_pointer(tp->root, p);
 134         return 0;
 135 }
 136
 137 static void tcindex_destroy_rexts(struct rcu_head *head)
 138 {
 139         struct tcindex_filter_result *r;
 140
 141         r = container_of(head, struct tcindex_filter_result, rcu);
 142         tcf_exts_destroy(&r->exts);
 143 }
 144
 145 static void tcindex_destroy_fexts(struct rcu_head *head)
 146 {
 147         struct tcindex_filter *f = container_of(head, struct tcindex_filter, rcu);
 148
 149         tcf_exts_destroy(&f->result.exts);
 150         kfree(f);
 151 }
 152
 153 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 154 {
 155         struct tcindex_data *p = rtnl_dereference(tp->root);
 156         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 157         struct tcindex_filter __rcu **walk;
 158         struct tcindex_filter *f = NULL;
 159
 160         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p\n", tp, arg, p);
 161         if (p->perfect) {
 162                 if (!r->res.class)
 163                         return -ENOENT;
 164         } else {
 165                 int i;
 166
 167                 for (i = 0; i < p->hash; i++) {
 168                         walk = p->h + i;
 169                         for (f = rtnl_dereference(*walk); f;
 170                              walk = &f->next, f = rtnl_dereference(*walk)) {
 171                                 if (&f->result == r)
 172                                         goto found;
 173                         }
 174                 }
 175                 return -ENOENT;
 176
 177 found:
 178                 rcu_assign_pointer(*walk, rtnl_dereference(f->next));
 179         }
 180         tcf_unbind_filter(tp, &r->res);
 181         /* all classifiers are required to call tcf_exts_destroy() after rcu
 182          * grace period, since converted-to-rcu actions are relying on that
 183          * in cleanup() callback
 184          */
 185         if (f)
 186                 call_rcu(&f->rcu, tcindex_destroy_fexts);
 187         else
 188                 call_rcu(&r->rcu, tcindex_destroy_rexts);
 189         return 0;
 190 }
 191
 192 static int tcindex_destroy_element(struct tcf_proto *tp,
 193                                    unsigned long arg,
 194                                    struct tcf_walker *walker)
 195 {
 196         return tcindex_delete(tp, arg);
 197 }
 198
 199 static void __tcindex_destroy(struct rcu_head *head)
 200 {
 201         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 202
 203         kfree(p->perfect);
 204         kfree(p->h);
 205         kfree(p);
 206 }
 207
 208 static inline int
 209 valid_perfect_hash(struct tcindex_data *p)
 210 {
 211         return  p->hash > (p->mask >> p->shift);
 212 }
 213
 214 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 215         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 216         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 217         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 218         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 219         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 220 };
 221
 222 static void tcindex_filter_result_init(struct tcindex_filter_result *r)
 223 {
 224         memset(r, 0, sizeof(*r));
 225         tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 226 }
 227
 228 static void __tcindex_partial_destroy(struct rcu_head *head)
 229 {
 230         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 231
 232         kfree(p->perfect);
 233         kfree(p);
 234 }
 235
 236 static int
 237 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 238                   u32 handle, struct tcindex_data *p,
 239                   struct tcindex_filter_result *r, struct nlattr **tb,
 240                   struct nlattr *est, bool ovr)
 241 {
 242         int err, balloc = 0;
 243         struct tcindex_filter_result new_filter_result, *old_r = r;
 244         struct tcindex_filter_result cr;
 245         struct tcindex_data *cp, *oldp;
 246         struct tcindex_filter *f = NULL; /* make gcc behave */
 247         struct tcf_exts e;
 248
 249         tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 250         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 251         if (err < 0)
 252                 return err;
 253
 254         err = -ENOMEM;
 255         /* tcindex_data attributes must look atomic to classifier/lookup so
 256          * allocate new tcindex data and RCU assign it onto root. Keeping
 257          * perfect hash and hash pointers from old data.
 258          */
 259         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
 260         if (!cp)
 261                 goto errout;
 262
 263         cp->mask = p->mask;
 264         cp->shift = p->shift;
 265         cp->hash = p->hash;
 266         cp->alloc_hash = p->alloc_hash;
 267         cp->fall_through = p->fall_through;
 268         cp->tp = tp;
 269
 270         if (tb[TCA_TCINDEX_HASH])
 271                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 272
 273         if (tb[TCA_TCINDEX_MASK])
 274                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 275
 276         if (tb[TCA_TCINDEX_SHIFT]) {
 277                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 278                 if (cp->shift > 16) {
 279                         err = -EINVAL;
 280                         goto errout;
 281                 }
 282         }
 283         if (!cp->hash) {
 284                 /* Hash not specified, use perfect hash if the upper limit
 285                  * of the hashing index is below the threshold.
 286                  */
 287                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
 288                         cp->hash = (cp->mask >> cp->shift) + 1;
 289                 else
 290                         cp->hash = DEFAULT_HASH_SIZE;
 291         }
 292
 293         if (p->perfect) {
 294                 int i;
 295
 296                 cp->perfect = kmemdup(p->perfect,
 297                                       sizeof(*r) * cp->hash, GFP_KERNEL);
 298                 if (!cp->perfect)
 299                         goto errout;
 300                 cp->alloc_hash = cp->hash;
 301                 for (i = 0; i < min(cp->hash, p->hash); i++)
 302                         tcf_exts_init(&cp->perfect[i].exts,
 303                                       TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 304                 balloc = 1;
 305         }
 306         cp->h = p->h;
 307
 308         tcindex_filter_result_init(&new_filter_result);
 309         tcindex_filter_result_init(&cr);
 310         if (old_r)
 311                 cr.res = r->res;
 312
 313         err = -EBUSY;
 314
 315         /* Hash already allocated, make sure that we still meet the
 316          * requirements for the allocated hash.
 317          */
 318         if (cp->perfect) {
 319                 if (!valid_perfect_hash(cp) ||
 320                     cp->hash > cp->alloc_hash)
 321                         goto errout_alloc;
 322         } else if (cp->h && cp->hash != cp->alloc_hash) {
 323                 goto errout_alloc;
 324         }
 325
 326         err = -EINVAL;
 327         if (tb[TCA_TCINDEX_FALL_THROUGH])
 328                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 329
 330         if (!cp->perfect && !cp->h)
 331                 cp->alloc_hash = cp->hash;
 332
 333         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 334          * but then, we'd fail handles that may become valid after some future
 335          * mask change. While this is extremely unlikely to ever matter,
 336          * the check below is safer (and also more backwards-compatible).
 337          */
 338         if (cp->perfect || valid_perfect_hash(cp))
 339                 if (handle >= cp->alloc_hash)
 340                         goto errout_alloc;
 341
 342
 343         err = -ENOMEM;
 344         if (!cp->perfect && !cp->h) {
 345                 if (valid_perfect_hash(cp)) {
 346                         int i;
 347
 348                         cp->perfect = kcalloc(cp->hash, sizeof(*r), GFP_KERNEL);
 349                         if (!cp->perfect)
 350                                 goto errout_alloc;
 351                         for (i = 0; i < cp->hash; i++)
 352                                 tcf_exts_init(&cp->perfect[i].exts,
 353                                               TCA_TCINDEX_ACT,
 354                                               TCA_TCINDEX_POLICE);
 355                         balloc = 1;
 356                 } else {
 357                         struct tcindex_filter __rcu **hash;
 358
 359                         hash = kcalloc(cp->hash,
 360                                        sizeof(struct tcindex_filter *),
 361                                        GFP_KERNEL);
 362
 363                         if (!hash)
 364                                 goto errout_alloc;
 365
 366                         cp->h = hash;
 367                         balloc = 2;
 368                 }
 369         }
 370
 371         if (cp->perfect)
 372                 r = cp->perfect + handle;
 373         else
 374                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
 375
 376         if (r == &new_filter_result) {
 377                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 378                 if (!f)
 379                         goto errout_alloc;
 380                 f->key = handle;
 381                 tcindex_filter_result_init(&f->result);
 382                 f->next = NULL;
 383         }
 384
 385         if (tb[TCA_TCINDEX_CLASSID]) {
 386                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 387                 tcf_bind_filter(tp, &cr.res, base);
 388         }
 389
 390         if (old_r && old_r != r)
 391                 tcindex_filter_result_init(old_r);
 392
 393         oldp = p;
 394         r->res = cr.res;
 395         tcf_exts_change(tp, &r->exts, &e);
 396
 397         rcu_assign_pointer(tp->root, cp);
 398
 399         if (r == &new_filter_result) {
 400                 struct tcindex_filter *nfp;
 401                 struct tcindex_filter __rcu **fp;
 402
 403                 f->result.res = r->res;
 404                 tcf_exts_change(tp, &f->result.exts, &r->exts);
 405
 406                 fp = cp->h + (handle % cp->hash);
 407                 for (nfp = rtnl_dereference(*fp);
 408                      nfp;
 409                      fp = &nfp->next, nfp = rtnl_dereference(*fp))
 410                                 ; /* nothing */
 411
 412                 rcu_assign_pointer(*fp, f);
 413         }
 414
 415         if (oldp)
 416                 call_rcu(&oldp->rcu, __tcindex_partial_destroy);
 417         return 0;
 418
 419 errout_alloc:
 420         if (balloc == 1)
 421                 kfree(cp->perfect);
 422         else if (balloc == 2)
 423                 kfree(cp->h);
 424 errout:
 425         kfree(cp);
 426         tcf_exts_destroy(&e);
 427         return err;
 428 }
 429
 430 static int
 431 tcindex_change(struct net *net, struct sk_buff *in_skb,
 432                struct tcf_proto *tp, unsigned long base, u32 handle,
 433                struct nlattr **tca, unsigned long *arg, bool ovr)
 434 {
 435         struct nlattr *opt = tca[TCA_OPTIONS];
 436         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 437         struct tcindex_data *p = rtnl_dereference(tp->root);
 438         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 439         int err;
 440
 441         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 442             "p %p,r %p,*arg 0x%lx\n",
 443             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 444
 445         if (!opt)
 446                 return 0;
 447
 448         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 449         if (err < 0)
 450                 return err;
 451
 452         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 453                                  tca[TCA_RATE], ovr);
 454 }
 455
 456 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 457 {
 458         struct tcindex_data *p = rtnl_dereference(tp->root);
 459         struct tcindex_filter *f, *next;
 460         int i;
 461
 462         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 463         if (p->perfect) {
 464                 for (i = 0; i < p->hash; i++) {
 465                         if (!p->perfect[i].res.class)
 466                                 continue;
 467                         if (walker->count >= walker->skip) {
 468                                 if (walker->fn(tp,
 469                                     (unsigned long) (p->perfect+i), walker)
 470                                      < 0) {
 471                                         walker->stop = 1;
 472                                         return;
 473                                 }
 474                         }
 475                         walker->count++;
 476                 }
 477         }
 478         if (!p->h)
 479                 return;
 480         for (i = 0; i < p->hash; i++) {
 481                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
 482                         next = rtnl_dereference(f->next);
 483                         if (walker->count >= walker->skip) {
 484                                 if (walker->fn(tp, (unsigned long) &f->result,
 485                                     walker) < 0) {
 486                                         walker->stop = 1;
 487                                         return;
 488                                 }
 489                         }
 490                         walker->count++;
 491                 }
 492         }
 493 }
 494
 495 static bool tcindex_destroy(struct tcf_proto *tp, bool force)
 496 {
 497         struct tcindex_data *p = rtnl_dereference(tp->root);
 498         struct tcf_walker walker;
 499
 500         if (!force)
 501                 return false;
 502
 503         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 504         walker.count = 0;
 505         walker.skip = 0;
 506         walker.fn = tcindex_destroy_element;
 507         tcindex_walk(tp, &walker);
 508
 509         call_rcu(&p->rcu, __tcindex_destroy);
 510         return true;
 511 }
 512
 513
 514 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 515     struct sk_buff *skb, struct tcmsg *t)
 516 {
 517         struct tcindex_data *p = rtnl_dereference(tp->root);
 518         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 519         struct nlattr *nest;
 520
 521         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p\n",
 522                  tp, fh, skb, t, p, r);
 523         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 524
 525         nest = nla_nest_start(skb, TCA_OPTIONS);
 526         if (nest == NULL)
 527                 goto nla_put_failure;
 528
 529         if (!fh) {
 530                 t->tcm_handle = ~0; /* whatever ... */
 531                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 532                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 533                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 534                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 535                         goto nla_put_failure;
 536                 nla_nest_end(skb, nest);
 537         } else {
 538                 if (p->perfect) {
 539                         t->tcm_handle = r - p->perfect;
 540                 } else {
 541                         struct tcindex_filter *f;
 542                         struct tcindex_filter __rcu **fp;
 543                         int i;
 544
 545                         t->tcm_handle = 0;
 546                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 547                                 fp = &p->h[i];
 548                                 for (f = rtnl_dereference(*fp);
 549                                      !t->tcm_handle && f;
 550                                      fp = &f->next, f = rtnl_dereference(*fp)) {
 551                                         if (&f->result == r)
 552                                                 t->tcm_handle = f->key;
 553                                 }
 554                         }
 555                 }
 556                 pr_debug("handle = %d\n", t->tcm_handle);
 557                 if (r->res.class &&
 558                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 559                         goto nla_put_failure;
 560
 561                 if (tcf_exts_dump(skb, &r->exts) < 0)
 562                         goto nla_put_failure;
 563                 nla_nest_end(skb, nest);
 564
 565                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 566                         goto nla_put_failure;
 567         }
 568
 569         return skb->len;
 570
 571 nla_put_failure:
 572         nla_nest_cancel(skb, nest);
 573         return -1;
 574 }
 575
 576 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 577         .kind           =       "tcindex",
 578         .classify       =       tcindex_classify,
 579         .init           =       tcindex_init,
 580         .destroy        =       tcindex_destroy,
 581         .get            =       tcindex_get,
 582         .change         =       tcindex_change,
 583         .delete         =       tcindex_delete,
 584         .walk           =       tcindex_walk,
 585         .dump           =       tcindex_dump,
 586         .owner          =       THIS_MODULE,
 587 };
 588
 589 static int __init init_tcindex(void)
 590 {
 591         return register_tcf_proto_ops(&cls_tcindex_ops);
 592 }
 593
 594 static void __exit exit_tcindex(void)
 595 {
 596         unregister_tcf_proto_ops(&cls_tcindex_ops);
 597 }
 598
 599 module_init(init_tcindex)
 600 module_exit(exit_tcindex)
 601 MODULE_LICENSE("GPL");