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_is_predicative(&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 unsigned long 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 0;
 115         r = tcindex_lookup(p, handle);
 116         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 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, 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 int tcindex_filter_result_init(struct tcindex_filter_result *r)
 223 {
 224         memset(r, 0, sizeof(*r));
 225         return 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 void tcindex_free_perfect_hash(struct tcindex_data *cp)
 237 {
 238         int i;
 239
 240         for (i = 0; i < cp->hash; i++)
 241                 tcf_exts_destroy(&cp->perfect[i].exts);
 242         kfree(cp->perfect);
 243 }
 244
 245 static int tcindex_alloc_perfect_hash(struct tcindex_data *cp)
 246 {
 247         int i, err = 0;
 248
 249         cp->perfect = kcalloc(cp->hash, sizeof(struct tcindex_filter_result),
 250                               GFP_KERNEL | __GFP_NOWARN);
 251         if (!cp->perfect)
 252                 return -ENOMEM;
 253
 254         for (i = 0; i < cp->hash; i++) {
 255                 err = tcf_exts_init(&cp->perfect[i].exts,
 256                                     TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 257                 if (err < 0)
 258                         goto errout;
 259         }
 260
 261         return 0;
 262
 263 errout:
 264         tcindex_free_perfect_hash(cp);
 265         return err;
 266 }
 267
 268 static int
 269 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 270                   u32 handle, struct tcindex_data *p,
 271                   struct tcindex_filter_result *r, struct nlattr **tb,
 272                   struct nlattr *est, bool ovr)
 273 {
 274         struct tcindex_filter_result new_filter_result, *old_r = r;
 275         struct tcindex_filter_result cr;
 276         struct tcindex_data *cp = NULL, *oldp;
 277         struct tcindex_filter *f = NULL; /* make gcc behave */
 278         int err, balloc = 0;
 279         struct tcf_exts e;
 280
 281         err = tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 282         if (err < 0)
 283                 return err;
 284         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 285         if (err < 0)
 286                 goto errout;
 287
 288         err = -ENOMEM;
 289         /* tcindex_data attributes must look atomic to classifier/lookup so
 290          * allocate new tcindex data and RCU assign it onto root. Keeping
 291          * perfect hash and hash pointers from old data.
 292          */
 293         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
 294         if (!cp)
 295                 goto errout;
 296
 297         cp->mask = p->mask;
 298         cp->shift = p->shift;
 299         cp->hash = p->hash;
 300         cp->alloc_hash = p->alloc_hash;
 301         cp->fall_through = p->fall_through;
 302         cp->tp = tp;
 303
 304         if (tb[TCA_TCINDEX_HASH])
 305                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 306
 307         if (tb[TCA_TCINDEX_MASK])
 308                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 309
 310         if (tb[TCA_TCINDEX_SHIFT]) {
 311                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 312                 if (cp->shift > 16) {
 313                         err = -EINVAL;
 314                         goto errout;
 315                 }
 316         }
 317         if (!cp->hash) {
 318                 /* Hash not specified, use perfect hash if the upper limit
 319                  * of the hashing index is below the threshold.
 320                  */
 321                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
 322                         cp->hash = (cp->mask >> cp->shift) + 1;
 323                 else
 324                         cp->hash = DEFAULT_HASH_SIZE;
 325         }
 326
 327         if (p->perfect) {
 328                 int i;
 329
 330                 if (tcindex_alloc_perfect_hash(cp) < 0)
 331                         goto errout;
 332                 cp->alloc_hash = cp->hash;
 333                 for (i = 0; i < min(cp->hash, p->hash); i++)
 334                         cp->perfect[i].res = p->perfect[i].res;
 335                 balloc = 1;
 336         }
 337         cp->h = p->h;
 338
 339         err = tcindex_filter_result_init(&new_filter_result);
 340         if (err < 0)
 341                 goto errout1;
 342         err = tcindex_filter_result_init(&cr);
 343         if (err < 0)
 344                 goto errout1;
 345         if (old_r)
 346                 cr.res = r->res;
 347
 348         err = -EBUSY;
 349
 350         /* Hash already allocated, make sure that we still meet the
 351          * requirements for the allocated hash.
 352          */
 353         if (cp->perfect) {
 354                 if (!valid_perfect_hash(cp) ||
 355                     cp->hash > cp->alloc_hash)
 356                         goto errout_alloc;
 357         } else if (cp->h && cp->hash != cp->alloc_hash) {
 358                 goto errout_alloc;
 359         }
 360
 361         err = -EINVAL;
 362         if (tb[TCA_TCINDEX_FALL_THROUGH])
 363                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 364
 365         if (!cp->perfect && !cp->h)
 366                 cp->alloc_hash = cp->hash;
 367
 368         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 369          * but then, we'd fail handles that may become valid after some future
 370          * mask change. While this is extremely unlikely to ever matter,
 371          * the check below is safer (and also more backwards-compatible).
 372          */
 373         if (cp->perfect || valid_perfect_hash(cp))
 374                 if (handle >= cp->alloc_hash)
 375                         goto errout_alloc;
 376
 377
 378         err = -ENOMEM;
 379         if (!cp->perfect && !cp->h) {
 380                 if (valid_perfect_hash(cp)) {
 381                         if (tcindex_alloc_perfect_hash(cp) < 0)
 382                                 goto errout_alloc;
 383                         balloc = 1;
 384                 } else {
 385                         struct tcindex_filter __rcu **hash;
 386
 387                         hash = kcalloc(cp->hash,
 388                                        sizeof(struct tcindex_filter *),
 389                                        GFP_KERNEL);
 390
 391                         if (!hash)
 392                                 goto errout_alloc;
 393
 394                         cp->h = hash;
 395                         balloc = 2;
 396                 }
 397         }
 398
 399         if (cp->perfect)
 400                 r = cp->perfect + handle;
 401         else
 402                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
 403
 404         if (r == &new_filter_result) {
 405                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 406                 if (!f)
 407                         goto errout_alloc;
 408                 f->key = handle;
 409                 f->next = NULL;
 410                 err = tcindex_filter_result_init(&f->result);
 411                 if (err < 0) {
 412                         kfree(f);
 413                         goto errout_alloc;
 414                 }
 415         }
 416
 417         if (tb[TCA_TCINDEX_CLASSID]) {
 418                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 419                 tcf_bind_filter(tp, &cr.res, base);
 420         }
 421
 422         if (old_r && old_r != r) {
 423                 err = tcindex_filter_result_init(old_r);
 424                 if (err < 0) {
 425                         kfree(f);
 426                         goto errout_alloc;
 427                 }
 428         }
 429
 430         oldp = p;
 431         r->res = cr.res;
 432         tcf_exts_change(tp, &r->exts, &e);
 433
 434         rcu_assign_pointer(tp->root, cp);
 435
 436         if (r == &new_filter_result) {
 437                 struct tcindex_filter *nfp;
 438                 struct tcindex_filter __rcu **fp;
 439
 440                 f->result.res = r->res;
 441                 tcf_exts_change(tp, &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, unsigned long *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 = (struct tcindex_filter_result *) *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 0x%lx\n",
 483             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 484
 485         if (!opt)
 486                 return 0;
 487
 488         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 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,
 509                                     (unsigned long) (p->perfect+i), walker)
 510                                      < 0) {
 511                                         walker->stop = 1;
 512                                         return;
 513                                 }
 514                         }
 515                         walker->count++;
 516                 }
 517         }
 518         if (!p->h)
 519                 return;
 520         for (i = 0; i < p->hash; i++) {
 521                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
 522                         next = rtnl_dereference(f->next);
 523                         if (walker->count >= walker->skip) {
 524                                 if (walker->fn(tp, (unsigned long) &f->result,
 525                                     walker) < 0) {
 526                                         walker->stop = 1;
 527                                         return;
 528                                 }
 529                         }
 530                         walker->count++;
 531                 }
 532         }
 533 }
 534
 535 static bool tcindex_destroy(struct tcf_proto *tp, bool force)
 536 {
 537         struct tcindex_data *p = rtnl_dereference(tp->root);
 538         struct tcf_walker walker;
 539
 540         if (!force)
 541                 return false;
 542
 543         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 544         walker.count = 0;
 545         walker.skip = 0;
 546         walker.fn = tcindex_destroy_element;
 547         tcindex_walk(tp, &walker);
 548
 549         call_rcu(&p->rcu, __tcindex_destroy);
 550         return true;
 551 }
 552
 553
 554 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 555                         struct sk_buff *skb, struct tcmsg *t)
 556 {
 557         struct tcindex_data *p = rtnl_dereference(tp->root);
 558         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 559         struct nlattr *nest;
 560
 561         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p\n",
 562                  tp, fh, skb, t, p, r);
 563         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 564
 565         nest = nla_nest_start(skb, TCA_OPTIONS);
 566         if (nest == NULL)
 567                 goto nla_put_failure;
 568
 569         if (!fh) {
 570                 t->tcm_handle = ~0; /* whatever ... */
 571                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 572                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 573                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 574                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 575                         goto nla_put_failure;
 576                 nla_nest_end(skb, nest);
 577         } else {
 578                 if (p->perfect) {
 579                         t->tcm_handle = r - p->perfect;
 580                 } else {
 581                         struct tcindex_filter *f;
 582                         struct tcindex_filter __rcu **fp;
 583                         int i;
 584
 585                         t->tcm_handle = 0;
 586                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 587                                 fp = &p->h[i];
 588                                 for (f = rtnl_dereference(*fp);
 589                                      !t->tcm_handle && f;
 590                                      fp = &f->next, f = rtnl_dereference(*fp)) {
 591                                         if (&f->result == r)
 592                                                 t->tcm_handle = f->key;
 593                                 }
 594                         }
 595                 }
 596                 pr_debug("handle = %d\n", t->tcm_handle);
 597                 if (r->res.class &&
 598                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 599                         goto nla_put_failure;
 600
 601                 if (tcf_exts_dump(skb, &r->exts) < 0)
 602                         goto nla_put_failure;
 603                 nla_nest_end(skb, nest);
 604
 605                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 606                         goto nla_put_failure;
 607         }
 608
 609         return skb->len;
 610
 611 nla_put_failure:
 612         nla_nest_cancel(skb, nest);
 613         return -1;
 614 }
 615
 616 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 617         .kind           =       "tcindex",
 618         .classify       =       tcindex_classify,
 619         .init           =       tcindex_init,
 620         .destroy        =       tcindex_destroy,
 621         .get            =       tcindex_get,
 622         .change         =       tcindex_change,
 623         .delete         =       tcindex_delete,
 624         .walk           =       tcindex_walk,
 625         .dump           =       tcindex_dump,
 626         .owner          =       THIS_MODULE,
 627 };
 628
 629 static int __init init_tcindex(void)
 630 {
 631         return register_tcf_proto_ops(&cls_tcindex_ops);
 632 }
 633
 634 static void __exit exit_tcindex(void)
 635 {
 636         unregister_tcf_proto_ops(&cls_tcindex_ops);
 637 }
 638
 639 module_init(init_tcindex)
 640 module_exit(exit_tcindex)
 641 MODULE_LICENSE("GPL");