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 };
  31
  32 struct tcindex_filter {
  33         u16 key;
  34         struct tcindex_filter_result result;
  35         struct tcindex_filter __rcu *next;
  36         struct rcu_head rcu;
  37 };
  38
  39
  40 struct tcindex_data {
  41         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  42         struct tcindex_filter __rcu **h; /* imperfect hash; */
  43         struct tcf_proto *tp;
  44         u16 mask;               /* AND key with mask */
  45         u32 shift;              /* shift ANDed key to the right */
  46         u32 hash;               /* hash table size; 0 if undefined */
  47         u32 alloc_hash;         /* allocated size */
  48         u32 fall_through;       /* 0: only classify if explicit match */
  49         struct rcu_head rcu;
  50 };
  51
  52 static inline int
  53 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 *
  59 tcindex_lookup(struct tcindex_data *p, 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 int
 137 tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 138 {
 139         struct tcindex_data *p = rtnl_dereference(tp->root);
 140         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 141         struct tcindex_filter __rcu **walk;
 142         struct tcindex_filter *f = NULL;
 143
 144         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p\n", tp, arg, p);
 145         if (p->perfect) {
 146                 if (!r->res.class)
 147                         return -ENOENT;
 148         } else {
 149                 int i;
 150
 151                 for (i = 0; i < p->hash; i++) {
 152                         walk = p->h + i;
 153                         for (f = rtnl_dereference(*walk); f;
 154                              walk = &f->next, f = rtnl_dereference(*walk)) {
 155                                 if (&f->result == r)
 156                                         goto found;
 157                         }
 158                 }
 159                 return -ENOENT;
 160
 161 found:
 162                 rcu_assign_pointer(*walk, rtnl_dereference(f->next));
 163         }
 164         tcf_unbind_filter(tp, &r->res);
 165         tcf_exts_destroy(&r->exts);
 166         if (f)
 167                 kfree_rcu(f, rcu);
 168         return 0;
 169 }
 170
 171 static int tcindex_destroy_element(struct tcf_proto *tp,
 172                                    unsigned long arg,
 173                                    struct tcf_walker *walker)
 174 {
 175         return tcindex_delete(tp, arg);
 176 }
 177
 178 static void __tcindex_destroy(struct rcu_head *head)
 179 {
 180         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 181
 182         kfree(p->perfect);
 183         kfree(p->h);
 184         kfree(p);
 185 }
 186
 187 static inline int
 188 valid_perfect_hash(struct tcindex_data *p)
 189 {
 190         return  p->hash > (p->mask >> p->shift);
 191 }
 192
 193 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 194         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 195         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 196         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 197         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 198         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 199 };
 200
 201 static void tcindex_filter_result_init(struct tcindex_filter_result *r)
 202 {
 203         memset(r, 0, sizeof(*r));
 204         tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 205 }
 206
 207 static void __tcindex_partial_destroy(struct rcu_head *head)
 208 {
 209         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 210
 211         kfree(p->perfect);
 212         kfree(p);
 213 }
 214
 215 static int
 216 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 217                   u32 handle, struct tcindex_data *p,
 218                   struct tcindex_filter_result *r, struct nlattr **tb,
 219                   struct nlattr *est, bool ovr)
 220 {
 221         int err, balloc = 0;
 222         struct tcindex_filter_result new_filter_result, *old_r = r;
 223         struct tcindex_filter_result cr;
 224         struct tcindex_data *cp, *oldp;
 225         struct tcindex_filter *f = NULL; /* make gcc behave */
 226         struct tcf_exts e;
 227
 228         tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 229         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 230         if (err < 0)
 231                 return err;
 232
 233         err = -ENOMEM;
 234         /* tcindex_data attributes must look atomic to classifier/lookup so
 235          * allocate new tcindex data and RCU assign it onto root. Keeping
 236          * perfect hash and hash pointers from old data.
 237          */
 238         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
 239         if (!cp)
 240                 goto errout;
 241
 242         cp->mask = p->mask;
 243         cp->shift = p->shift;
 244         cp->hash = p->hash;
 245         cp->alloc_hash = p->alloc_hash;
 246         cp->fall_through = p->fall_through;
 247         cp->tp = tp;
 248
 249         if (p->perfect) {
 250                 int i;
 251
 252                 cp->perfect = kmemdup(p->perfect,
 253                                       sizeof(*r) * cp->hash, GFP_KERNEL);
 254                 if (!cp->perfect)
 255                         goto errout;
 256                 for (i = 0; i < cp->hash; i++)
 257                         tcf_exts_init(&cp->perfect[i].exts,
 258                                       TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 259                 balloc = 1;
 260         }
 261         cp->h = p->h;
 262
 263         tcindex_filter_result_init(&new_filter_result);
 264         tcindex_filter_result_init(&cr);
 265         if (old_r)
 266                 cr.res = r->res;
 267
 268         if (tb[TCA_TCINDEX_HASH])
 269                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 270
 271         if (tb[TCA_TCINDEX_MASK])
 272                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 273
 274         if (tb[TCA_TCINDEX_SHIFT])
 275                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 276
 277         err = -EBUSY;
 278
 279         /* Hash already allocated, make sure that we still meet the
 280          * requirements for the allocated hash.
 281          */
 282         if (cp->perfect) {
 283                 if (!valid_perfect_hash(cp) ||
 284                     cp->hash > cp->alloc_hash)
 285                         goto errout_alloc;
 286         } else if (cp->h && cp->hash != cp->alloc_hash) {
 287                 goto errout_alloc;
 288         }
 289
 290         err = -EINVAL;
 291         if (tb[TCA_TCINDEX_FALL_THROUGH])
 292                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 293
 294         if (!cp->hash) {
 295                 /* Hash not specified, use perfect hash if the upper limit
 296                  * of the hashing index is below the threshold.
 297                  */
 298                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
 299                         cp->hash = (cp->mask >> cp->shift) + 1;
 300                 else
 301                         cp->hash = DEFAULT_HASH_SIZE;
 302         }
 303
 304         if (!cp->perfect && !cp->h)
 305                 cp->alloc_hash = cp->hash;
 306
 307         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 308          * but then, we'd fail handles that may become valid after some future
 309          * mask change. While this is extremely unlikely to ever matter,
 310          * the check below is safer (and also more backwards-compatible).
 311          */
 312         if (cp->perfect || valid_perfect_hash(cp))
 313                 if (handle >= cp->alloc_hash)
 314                         goto errout_alloc;
 315
 316
 317         err = -ENOMEM;
 318         if (!cp->perfect && !cp->h) {
 319                 if (valid_perfect_hash(cp)) {
 320                         int i;
 321
 322                         cp->perfect = kcalloc(cp->hash, sizeof(*r), GFP_KERNEL);
 323                         if (!cp->perfect)
 324                                 goto errout_alloc;
 325                         for (i = 0; i < cp->hash; i++)
 326                                 tcf_exts_init(&cp->perfect[i].exts,
 327                                               TCA_TCINDEX_ACT,
 328                                               TCA_TCINDEX_POLICE);
 329                         balloc = 1;
 330                 } else {
 331                         struct tcindex_filter __rcu **hash;
 332
 333                         hash = kcalloc(cp->hash,
 334                                        sizeof(struct tcindex_filter *),
 335                                        GFP_KERNEL);
 336
 337                         if (!hash)
 338                                 goto errout_alloc;
 339
 340                         cp->h = hash;
 341                         balloc = 2;
 342                 }
 343         }
 344
 345         if (cp->perfect)
 346                 r = cp->perfect + handle;
 347         else
 348                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
 349
 350         if (r == &new_filter_result) {
 351                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 352                 if (!f)
 353                         goto errout_alloc;
 354                 f->key = handle;
 355                 tcindex_filter_result_init(&f->result);
 356                 f->next = NULL;
 357         }
 358
 359         if (tb[TCA_TCINDEX_CLASSID]) {
 360                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 361                 tcf_bind_filter(tp, &cr.res, base);
 362         }
 363
 364         if (old_r)
 365                 tcf_exts_change(tp, &r->exts, &e);
 366         else
 367                 tcf_exts_change(tp, &cr.exts, &e);
 368
 369         if (old_r && old_r != r)
 370                 tcindex_filter_result_init(old_r);
 371
 372         oldp = p;
 373         r->res = cr.res;
 374         rcu_assign_pointer(tp->root, cp);
 375
 376         if (r == &new_filter_result) {
 377                 struct tcindex_filter *nfp;
 378                 struct tcindex_filter __rcu **fp;
 379
 380                 tcf_exts_change(tp, &f->result.exts, &r->exts);
 381
 382                 fp = cp->h + (handle % cp->hash);
 383                 for (nfp = rtnl_dereference(*fp);
 384                      nfp;
 385                      fp = &nfp->next, nfp = rtnl_dereference(*fp))
 386                                 ; /* nothing */
 387
 388                 rcu_assign_pointer(*fp, f);
 389         }
 390
 391         if (oldp)
 392                 call_rcu(&oldp->rcu, __tcindex_partial_destroy);
 393         return 0;
 394
 395 errout_alloc:
 396         if (balloc == 1)
 397                 kfree(cp->perfect);
 398         else if (balloc == 2)
 399                 kfree(cp->h);
 400 errout:
 401         kfree(cp);
 402         tcf_exts_destroy(&e);
 403         return err;
 404 }
 405
 406 static int
 407 tcindex_change(struct net *net, struct sk_buff *in_skb,
 408                struct tcf_proto *tp, unsigned long base, u32 handle,
 409                struct nlattr **tca, unsigned long *arg, bool ovr)
 410 {
 411         struct nlattr *opt = tca[TCA_OPTIONS];
 412         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 413         struct tcindex_data *p = rtnl_dereference(tp->root);
 414         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 415         int err;
 416
 417         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 418             "p %p,r %p,*arg 0x%lx\n",
 419             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 420
 421         if (!opt)
 422                 return 0;
 423
 424         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 425         if (err < 0)
 426                 return err;
 427
 428         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 429                                  tca[TCA_RATE], ovr);
 430 }
 431
 432 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 433 {
 434         struct tcindex_data *p = rtnl_dereference(tp->root);
 435         struct tcindex_filter *f, *next;
 436         int i;
 437
 438         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 439         if (p->perfect) {
 440                 for (i = 0; i < p->hash; i++) {
 441                         if (!p->perfect[i].res.class)
 442                                 continue;
 443                         if (walker->count >= walker->skip) {
 444                                 if (walker->fn(tp,
 445                                     (unsigned long) (p->perfect+i), walker)
 446                                      < 0) {
 447                                         walker->stop = 1;
 448                                         return;
 449                                 }
 450                         }
 451                         walker->count++;
 452                 }
 453         }
 454         if (!p->h)
 455                 return;
 456         for (i = 0; i < p->hash; i++) {
 457                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
 458                         next = rtnl_dereference(f->next);
 459                         if (walker->count >= walker->skip) {
 460                                 if (walker->fn(tp, (unsigned long) &f->result,
 461                                     walker) < 0) {
 462                                         walker->stop = 1;
 463                                         return;
 464                                 }
 465                         }
 466                         walker->count++;
 467                 }
 468         }
 469 }
 470
 471 static void tcindex_destroy(struct tcf_proto *tp)
 472 {
 473         struct tcindex_data *p = rtnl_dereference(tp->root);
 474         struct tcf_walker walker;
 475
 476         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 477         walker.count = 0;
 478         walker.skip = 0;
 479         walker.fn = tcindex_destroy_element;
 480         tcindex_walk(tp, &walker);
 481
 482         RCU_INIT_POINTER(tp->root, NULL);
 483         call_rcu(&p->rcu, __tcindex_destroy);
 484 }
 485
 486
 487 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 488     struct sk_buff *skb, struct tcmsg *t)
 489 {
 490         struct tcindex_data *p = rtnl_dereference(tp->root);
 491         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 492         struct nlattr *nest;
 493
 494         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p\n",
 495                  tp, fh, skb, t, p, r);
 496         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 497
 498         nest = nla_nest_start(skb, TCA_OPTIONS);
 499         if (nest == NULL)
 500                 goto nla_put_failure;
 501
 502         if (!fh) {
 503                 t->tcm_handle = ~0; /* whatever ... */
 504                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 505                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 506                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 507                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 508                         goto nla_put_failure;
 509                 nla_nest_end(skb, nest);
 510         } else {
 511                 if (p->perfect) {
 512                         t->tcm_handle = r - p->perfect;
 513                 } else {
 514                         struct tcindex_filter *f;
 515                         struct tcindex_filter __rcu **fp;
 516                         int i;
 517
 518                         t->tcm_handle = 0;
 519                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 520                                 fp = &p->h[i];
 521                                 for (f = rtnl_dereference(*fp);
 522                                      !t->tcm_handle && f;
 523                                      fp = &f->next, f = rtnl_dereference(*fp)) {
 524                                         if (&f->result == r)
 525                                                 t->tcm_handle = f->key;
 526                                 }
 527                         }
 528                 }
 529                 pr_debug("handle = %d\n", t->tcm_handle);
 530                 if (r->res.class &&
 531                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 532                         goto nla_put_failure;
 533
 534                 if (tcf_exts_dump(skb, &r->exts) < 0)
 535                         goto nla_put_failure;
 536                 nla_nest_end(skb, nest);
 537
 538                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 539                         goto nla_put_failure;
 540         }
 541
 542         return skb->len;
 543
 544 nla_put_failure:
 545         nla_nest_cancel(skb, nest);
 546         return -1;
 547 }
 548
 549 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 550         .kind           =       "tcindex",
 551         .classify       =       tcindex_classify,
 552         .init           =       tcindex_init,
 553         .destroy        =       tcindex_destroy,
 554         .get            =       tcindex_get,
 555         .change         =       tcindex_change,
 556         .delete         =       tcindex_delete,
 557         .walk           =       tcindex_walk,
 558         .dump           =       tcindex_dump,
 559         .owner          =       THIS_MODULE,
 560 };
 561
 562 static int __init init_tcindex(void)
 563 {
 564         return register_tcf_proto_ops(&cls_tcindex_ops);
 565 }
 566
 567 static void __exit exit_tcindex(void)
 568 {
 569         unregister_tcf_proto_ops(&cls_tcindex_ops);
 570 }
 571
 572 module_init(init_tcindex)
 573 module_exit(exit_tcindex)
 574 MODULE_LICENSE("GPL");