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 tcindex_filter_is_set(struct tcindex_filter_result *r)
  54 {
  55         return tcf_exts_has_actions(&r->exts) || r->res.classid;
  56 }
  57
  58 static struct tcindex_filter_result *tcindex_lookup(struct tcindex_data *p,
  59                                                     u16 key)
  60 {
  61         if (p->perfect) {
  62                 struct tcindex_filter_result *f = p->perfect + key;
  63
  64                 return tcindex_filter_is_set(f) ? f : NULL;
  65         } else if (p->h) {
  66                 struct tcindex_filter __rcu **fp;
  67                 struct tcindex_filter *f;
  68
  69                 fp = &p->h[key % p->hash];
  70                 for (f = rcu_dereference_bh_rtnl(*fp);
  71                      f;
  72                      fp = &f->next, f = rcu_dereference_bh_rtnl(*fp))
  73                         if (f->key == key)
  74                                 return &f->result;
  75         }
  76
  77         return NULL;
  78 }
  79
  80
  81 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  82                             struct tcf_result *res)
  83 {
  84         struct tcindex_data *p = rcu_dereference_bh(tp->root);
  85         struct tcindex_filter_result *f;
  86         int key = (skb->tc_index & p->mask) >> p->shift;
  87
  88         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
  89                  skb, tp, res, p);
  90
  91         f = tcindex_lookup(p, key);
  92         if (!f) {
  93                 if (!p->fall_through)
  94                         return -1;
  95                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
  96                 res->class = 0;
  97                 pr_debug("alg 0x%x\n", res->classid);
  98                 return 0;
  99         }
 100         *res = f->res;
 101         pr_debug("map 0x%x\n", res->classid);
 102
 103         return tcf_exts_exec(skb, &f->exts, res);
 104 }
 105
 106
 107 static void *tcindex_get(struct tcf_proto *tp, u32 handle)
 108 {
 109         struct tcindex_data *p = rtnl_dereference(tp->root);
 110         struct tcindex_filter_result *r;
 111
 112         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
 113         if (p->perfect && handle >= p->alloc_hash)
 114                 return NULL;
 115         r = tcindex_lookup(p, handle);
 116         return r && tcindex_filter_is_set(r) ? r : NULL;
 117 }
 118
 119 static int tcindex_init(struct tcf_proto *tp)
 120 {
 121         struct tcindex_data *p;
 122
 123         pr_debug("tcindex_init(tp %p)\n", tp);
 124         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
 125         if (!p)
 126                 return -ENOMEM;
 127
 128         p->mask = 0xffff;
 129         p->hash = DEFAULT_HASH_SIZE;
 130         p->fall_through = 1;
 131
 132         rcu_assign_pointer(tp->root, p);
 133         return 0;
 134 }
 135
 136 static void tcindex_destroy_rexts(struct rcu_head *head)
 137 {
 138         struct tcindex_filter_result *r;
 139
 140         r = container_of(head, struct tcindex_filter_result, rcu);
 141         tcf_exts_destroy(&r->exts);
 142 }
 143
 144 static void tcindex_destroy_fexts(struct rcu_head *head)
 145 {
 146         struct tcindex_filter *f = container_of(head, struct tcindex_filter,
 147                                                 rcu);
 148
 149         tcf_exts_destroy(&f->result.exts);
 150         kfree(f);
 151 }
 152
 153 static int tcindex_delete(struct tcf_proto *tp, void *arg, bool *last)
 154 {
 155         struct tcindex_data *p = rtnl_dereference(tp->root);
 156         struct tcindex_filter_result *r = arg;
 157         struct tcindex_filter __rcu **walk;
 158         struct tcindex_filter *f = NULL;
 159
 160         pr_debug("tcindex_delete(tp %p,arg %p),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
 190         *last = false;
 191         return 0;
 192 }
 193
 194 static int tcindex_destroy_element(struct tcf_proto *tp,
 195                                    void *arg, struct tcf_walker *walker)
 196 {
 197         bool last;
 198
 199         return tcindex_delete(tp, arg, &last);
 200 }
 201
 202 static void __tcindex_destroy(struct rcu_head *head)
 203 {
 204         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 205
 206         kfree(p->perfect);
 207         kfree(p->h);
 208         kfree(p);
 209 }
 210
 211 static inline int
 212 valid_perfect_hash(struct tcindex_data *p)
 213 {
 214         return  p->hash > (p->mask >> p->shift);
 215 }
 216
 217 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 218         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 219         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 220         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 221         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 222         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 223 };
 224
 225 static int tcindex_filter_result_init(struct tcindex_filter_result *r)
 226 {
 227         memset(r, 0, sizeof(*r));
 228         return tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 229 }
 230
 231 static void __tcindex_partial_destroy(struct rcu_head *head)
 232 {
 233         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 234
 235         kfree(p->perfect);
 236         kfree(p);
 237 }
 238
 239 static void tcindex_free_perfect_hash(struct tcindex_data *cp)
 240 {
 241         int i;
 242
 243         for (i = 0; i < cp->hash; i++)
 244                 tcf_exts_destroy(&cp->perfect[i].exts);
 245         kfree(cp->perfect);
 246 }
 247
 248 static int tcindex_alloc_perfect_hash(struct tcindex_data *cp)
 249 {
 250         int i, err = 0;
 251
 252         cp->perfect = kcalloc(cp->hash, sizeof(struct tcindex_filter_result),
 253                               GFP_KERNEL);
 254         if (!cp->perfect)
 255                 return -ENOMEM;
 256
 257         for (i = 0; i < cp->hash; i++) {
 258                 err = tcf_exts_init(&cp->perfect[i].exts,
 259                                     TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 260                 if (err < 0)
 261                         goto errout;
 262         }
 263
 264         return 0;
 265
 266 errout:
 267         tcindex_free_perfect_hash(cp);
 268         return err;
 269 }
 270
 271 static int
 272 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 273                   u32 handle, struct tcindex_data *p,
 274                   struct tcindex_filter_result *r, struct nlattr **tb,
 275                   struct nlattr *est, bool ovr)
 276 {
 277         struct tcindex_filter_result new_filter_result, *old_r = r;
 278         struct tcindex_filter_result cr;
 279         struct tcindex_data *cp = NULL, *oldp;
 280         struct tcindex_filter *f = NULL; /* make gcc behave */
 281         int err, balloc = 0;
 282         struct tcf_exts e;
 283
 284         err = tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 285         if (err < 0)
 286                 return err;
 287         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 288         if (err < 0)
 289                 goto errout;
 290
 291         err = -ENOMEM;
 292         /* tcindex_data attributes must look atomic to classifier/lookup so
 293          * allocate new tcindex data and RCU assign it onto root. Keeping
 294          * perfect hash and hash pointers from old data.
 295          */
 296         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
 297         if (!cp)
 298                 goto errout;
 299
 300         cp->mask = p->mask;
 301         cp->shift = p->shift;
 302         cp->hash = p->hash;
 303         cp->alloc_hash = p->alloc_hash;
 304         cp->fall_through = p->fall_through;
 305         cp->tp = tp;
 306
 307         if (p->perfect) {
 308                 int i;
 309
 310                 if (tcindex_alloc_perfect_hash(cp) < 0)
 311                         goto errout;
 312                 for (i = 0; i < cp->hash; i++)
 313                         cp->perfect[i].res = p->perfect[i].res;
 314                 balloc = 1;
 315         }
 316         cp->h = p->h;
 317
 318         err = tcindex_filter_result_init(&new_filter_result);
 319         if (err < 0)
 320                 goto errout1;
 321         err = tcindex_filter_result_init(&cr);
 322         if (err < 0)
 323                 goto errout1;
 324         if (old_r)
 325                 cr.res = r->res;
 326
 327         if (tb[TCA_TCINDEX_HASH])
 328                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 329
 330         if (tb[TCA_TCINDEX_MASK])
 331                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 332
 333         if (tb[TCA_TCINDEX_SHIFT])
 334                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 335
 336         err = -EBUSY;
 337
 338         /* Hash already allocated, make sure that we still meet the
 339          * requirements for the allocated hash.
 340          */
 341         if (cp->perfect) {
 342                 if (!valid_perfect_hash(cp) ||
 343                     cp->hash > cp->alloc_hash)
 344                         goto errout_alloc;
 345         } else if (cp->h && cp->hash != cp->alloc_hash) {
 346                 goto errout_alloc;
 347         }
 348
 349         err = -EINVAL;
 350         if (tb[TCA_TCINDEX_FALL_THROUGH])
 351                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 352
 353         if (!cp->hash) {
 354                 /* Hash not specified, use perfect hash if the upper limit
 355                  * of the hashing index is below the threshold.
 356                  */
 357                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
 358                         cp->hash = (cp->mask >> cp->shift) + 1;
 359                 else
 360                         cp->hash = DEFAULT_HASH_SIZE;
 361         }
 362
 363         if (!cp->perfect && !cp->h)
 364                 cp->alloc_hash = cp->hash;
 365
 366         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 367          * but then, we'd fail handles that may become valid after some future
 368          * mask change. While this is extremely unlikely to ever matter,
 369          * the check below is safer (and also more backwards-compatible).
 370          */
 371         if (cp->perfect || valid_perfect_hash(cp))
 372                 if (handle >= cp->alloc_hash)
 373                         goto errout_alloc;
 374
 375
 376         err = -ENOMEM;
 377         if (!cp->perfect && !cp->h) {
 378                 if (valid_perfect_hash(cp)) {
 379                         if (tcindex_alloc_perfect_hash(cp) < 0)
 380                                 goto errout_alloc;
 381                         balloc = 1;
 382                 } else {
 383                         struct tcindex_filter __rcu **hash;
 384
 385                         hash = kcalloc(cp->hash,
 386                                        sizeof(struct tcindex_filter *),
 387                                        GFP_KERNEL);
 388
 389                         if (!hash)
 390                                 goto errout_alloc;
 391
 392                         cp->h = hash;
 393                         balloc = 2;
 394                 }
 395         }
 396
 397         if (cp->perfect)
 398                 r = cp->perfect + handle;
 399         else
 400                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
 401
 402         if (r == &new_filter_result) {
 403                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 404                 if (!f)
 405                         goto errout_alloc;
 406                 f->key = handle;
 407                 f->next = NULL;
 408                 err = tcindex_filter_result_init(&f->result);
 409                 if (err < 0) {
 410                         kfree(f);
 411                         goto errout_alloc;
 412                 }
 413         }
 414
 415         if (tb[TCA_TCINDEX_CLASSID]) {
 416                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 417                 tcf_bind_filter(tp, &cr.res, base);
 418         }
 419
 420         if (old_r)
 421                 tcf_exts_change(&r->exts, &e);
 422         else
 423                 tcf_exts_change(&cr.exts, &e);
 424
 425         if (old_r && old_r != r) {
 426                 err = tcindex_filter_result_init(old_r);
 427                 if (err < 0) {
 428                         kfree(f);
 429                         goto errout_alloc;
 430                 }
 431         }
 432
 433         oldp = p;
 434         r->res = cr.res;
 435         rcu_assign_pointer(tp->root, cp);
 436
 437         if (r == &new_filter_result) {
 438                 struct tcindex_filter *nfp;
 439                 struct tcindex_filter __rcu **fp;
 440
 441                 tcf_exts_change(&f->result.exts, &r->exts);
 442
 443                 fp = cp->h + (handle % cp->hash);
 444                 for (nfp = rtnl_dereference(*fp);
 445                      nfp;
 446                      fp = &nfp->next, nfp = rtnl_dereference(*fp))
 447                                 ; /* nothing */
 448
 449                 rcu_assign_pointer(*fp, f);
 450         }
 451
 452         if (oldp)
 453                 call_rcu(&oldp->rcu, __tcindex_partial_destroy);
 454         return 0;
 455
 456 errout_alloc:
 457         if (balloc == 1)
 458                 tcindex_free_perfect_hash(cp);
 459         else if (balloc == 2)
 460                 kfree(cp->h);
 461 errout1:
 462         tcf_exts_destroy(&cr.exts);
 463         tcf_exts_destroy(&new_filter_result.exts);
 464 errout:
 465         kfree(cp);
 466         tcf_exts_destroy(&e);
 467         return err;
 468 }
 469
 470 static int
 471 tcindex_change(struct net *net, struct sk_buff *in_skb,
 472                struct tcf_proto *tp, unsigned long base, u32 handle,
 473                struct nlattr **tca, void **arg, bool ovr)
 474 {
 475         struct nlattr *opt = tca[TCA_OPTIONS];
 476         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 477         struct tcindex_data *p = rtnl_dereference(tp->root);
 478         struct tcindex_filter_result *r = *arg;
 479         int err;
 480
 481         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 482             "p %p,r %p,*arg %p\n",
 483             tp, handle, tca, arg, opt, p, r, arg ? *arg : NULL);
 484
 485         if (!opt)
 486                 return 0;
 487
 488         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy, NULL);
 489         if (err < 0)
 490                 return err;
 491
 492         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 493                                  tca[TCA_RATE], ovr);
 494 }
 495
 496 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 497 {
 498         struct tcindex_data *p = rtnl_dereference(tp->root);
 499         struct tcindex_filter *f, *next;
 500         int i;
 501
 502         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 503         if (p->perfect) {
 504                 for (i = 0; i < p->hash; i++) {
 505                         if (!p->perfect[i].res.class)
 506                                 continue;
 507                         if (walker->count >= walker->skip) {
 508                                 if (walker->fn(tp, p->perfect + i, walker) < 0) {
 509                                         walker->stop = 1;
 510                                         return;
 511                                 }
 512                         }
 513                         walker->count++;
 514                 }
 515         }
 516         if (!p->h)
 517                 return;
 518         for (i = 0; i < p->hash; i++) {
 519                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
 520                         next = rtnl_dereference(f->next);
 521                         if (walker->count >= walker->skip) {
 522                                 if (walker->fn(tp, &f->result, walker) < 0) {
 523                                         walker->stop = 1;
 524                                         return;
 525                                 }
 526                         }
 527                         walker->count++;
 528                 }
 529         }
 530 }
 531
 532 static void tcindex_destroy(struct tcf_proto *tp)
 533 {
 534         struct tcindex_data *p = rtnl_dereference(tp->root);
 535         struct tcf_walker walker;
 536
 537         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 538         walker.count = 0;
 539         walker.skip = 0;
 540         walker.fn = tcindex_destroy_element;
 541         tcindex_walk(tp, &walker);
 542
 543         call_rcu(&p->rcu, __tcindex_destroy);
 544 }
 545
 546
 547 static int tcindex_dump(struct net *net, struct tcf_proto *tp, void *fh,
 548                         struct sk_buff *skb, struct tcmsg *t)
 549 {
 550         struct tcindex_data *p = rtnl_dereference(tp->root);
 551         struct tcindex_filter_result *r = fh;
 552         struct nlattr *nest;
 553
 554         pr_debug("tcindex_dump(tp %p,fh %p,skb %p,t %p),p %p,r %p\n",
 555                  tp, fh, skb, t, p, r);
 556         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 557
 558         nest = nla_nest_start(skb, TCA_OPTIONS);
 559         if (nest == NULL)
 560                 goto nla_put_failure;
 561
 562         if (!fh) {
 563                 t->tcm_handle = ~0; /* whatever ... */
 564                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 565                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 566                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 567                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 568                         goto nla_put_failure;
 569                 nla_nest_end(skb, nest);
 570         } else {
 571                 if (p->perfect) {
 572                         t->tcm_handle = r - p->perfect;
 573                 } else {
 574                         struct tcindex_filter *f;
 575                         struct tcindex_filter __rcu **fp;
 576                         int i;
 577
 578                         t->tcm_handle = 0;
 579                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 580                                 fp = &p->h[i];
 581                                 for (f = rtnl_dereference(*fp);
 582                                      !t->tcm_handle && f;
 583                                      fp = &f->next, f = rtnl_dereference(*fp)) {
 584                                         if (&f->result == r)
 585                                                 t->tcm_handle = f->key;
 586                                 }
 587                         }
 588                 }
 589                 pr_debug("handle = %d\n", t->tcm_handle);
 590                 if (r->res.class &&
 591                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 592                         goto nla_put_failure;
 593
 594                 if (tcf_exts_dump(skb, &r->exts) < 0)
 595                         goto nla_put_failure;
 596                 nla_nest_end(skb, nest);
 597
 598                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 599                         goto nla_put_failure;
 600         }
 601
 602         return skb->len;
 603
 604 nla_put_failure:
 605         nla_nest_cancel(skb, nest);
 606         return -1;
 607 }
 608
 609 static void tcindex_bind_class(void *fh, u32 classid, unsigned long cl)
 610 {
 611         struct tcindex_filter_result *r = fh;
 612
 613         if (r && r->res.classid == classid)
 614                 r->res.class = cl;
 615 }
 616
 617 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 618         .kind           =       "tcindex",
 619         .classify       =       tcindex_classify,
 620         .init           =       tcindex_init,
 621         .destroy        =       tcindex_destroy,
 622         .get            =       tcindex_get,
 623         .change         =       tcindex_change,
 624         .delete         =       tcindex_delete,
 625         .walk           =       tcindex_walk,
 626         .dump           =       tcindex_dump,
 627         .bind_class     =       tcindex_bind_class,
 628         .owner          =       THIS_MODULE,
 629 };
 630
 631 static int __init init_tcindex(void)
 632 {
 633         return register_tcf_proto_ops(&cls_tcindex_ops);
 634 }
 635
 636 static void __exit exit_tcindex(void)
 637 {
 638         unregister_tcf_proto_ops(&cls_tcindex_ops);
 639 }
 640
 641 module_init(init_tcindex)
 642 module_exit(exit_tcindex)
 643 MODULE_LICENSE("GPL");