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
 120 static void tcindex_put(struct tcf_proto *tp, unsigned long f)
 121 {
 122         pr_debug("tcindex_put(tp %p,f 0x%lx)\n", tp, f);
 123 }
 124
 125
 126 static int tcindex_init(struct tcf_proto *tp)
 127 {
 128         struct tcindex_data *p;
 129
 130         pr_debug("tcindex_init(tp %p)\n", tp);
 131         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
 132         if (!p)
 133                 return -ENOMEM;
 134
 135         p->mask = 0xffff;
 136         p->hash = DEFAULT_HASH_SIZE;
 137         p->fall_through = 1;
 138
 139         rcu_assign_pointer(tp->root, p);
 140         return 0;
 141 }
 142
 143 static int
 144 tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 145 {
 146         struct tcindex_data *p = rtnl_dereference(tp->root);
 147         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 148         struct tcindex_filter __rcu **walk;
 149         struct tcindex_filter *f = NULL;
 150
 151         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p\n", tp, arg, p);
 152         if (p->perfect) {
 153                 if (!r->res.class)
 154                         return -ENOENT;
 155         } else {
 156                 int i;
 157
 158                 for (i = 0; i < p->hash; i++) {
 159                         walk = p->h + i;
 160                         for (f = rtnl_dereference(*walk); f;
 161                              walk = &f->next, f = rtnl_dereference(*walk)) {
 162                                 if (&f->result == r)
 163                                         goto found;
 164                         }
 165                 }
 166                 return -ENOENT;
 167
 168 found:
 169                 rcu_assign_pointer(*walk, rtnl_dereference(f->next));
 170         }
 171         tcf_unbind_filter(tp, &r->res);
 172         tcf_exts_destroy(&r->exts);
 173         if (f)
 174                 kfree_rcu(f, rcu);
 175         return 0;
 176 }
 177
 178 static int tcindex_destroy_element(struct tcf_proto *tp,
 179                                    unsigned long arg,
 180                                    struct tcf_walker *walker)
 181 {
 182         return tcindex_delete(tp, arg);
 183 }
 184
 185 static void __tcindex_destroy(struct rcu_head *head)
 186 {
 187         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 188
 189         kfree(p->perfect);
 190         kfree(p->h);
 191         kfree(p);
 192 }
 193
 194 static inline int
 195 valid_perfect_hash(struct tcindex_data *p)
 196 {
 197         return  p->hash > (p->mask >> p->shift);
 198 }
 199
 200 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 201         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 202         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 203         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 204         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 205         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 206 };
 207
 208 static void tcindex_filter_result_init(struct tcindex_filter_result *r)
 209 {
 210         memset(r, 0, sizeof(*r));
 211         tcf_exts_init(&r->exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 212 }
 213
 214 static void __tcindex_partial_destroy(struct rcu_head *head)
 215 {
 216         struct tcindex_data *p = container_of(head, struct tcindex_data, rcu);
 217
 218         kfree(p->perfect);
 219         kfree(p);
 220 }
 221
 222 static int
 223 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 224                   u32 handle, struct tcindex_data *p,
 225                   struct tcindex_filter_result *r, struct nlattr **tb,
 226                   struct nlattr *est, bool ovr)
 227 {
 228         int err, balloc = 0;
 229         struct tcindex_filter_result new_filter_result, *old_r = r;
 230         struct tcindex_filter_result cr;
 231         struct tcindex_data *cp, *oldp;
 232         struct tcindex_filter *f = NULL; /* make gcc behave */
 233         struct tcf_exts e;
 234
 235         tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 236         err = tcf_exts_validate(net, tp, tb, est, &e, ovr);
 237         if (err < 0)
 238                 return err;
 239
 240         err = -ENOMEM;
 241         /* tcindex_data attributes must look atomic to classifier/lookup so
 242          * allocate new tcindex data and RCU assign it onto root. Keeping
 243          * perfect hash and hash pointers from old data.
 244          */
 245         cp = kzalloc(sizeof(*cp), GFP_KERNEL);
 246         if (!cp)
 247                 goto errout;
 248
 249         cp->mask = p->mask;
 250         cp->shift = p->shift;
 251         cp->hash = p->hash;
 252         cp->alloc_hash = p->alloc_hash;
 253         cp->fall_through = p->fall_through;
 254         cp->tp = tp;
 255
 256         if (p->perfect) {
 257                 int i;
 258
 259                 cp->perfect = kmemdup(p->perfect,
 260                                       sizeof(*r) * cp->hash, GFP_KERNEL);
 261                 if (!cp->perfect)
 262                         goto errout;
 263                 for (i = 0; i < cp->hash; i++)
 264                         tcf_exts_init(&cp->perfect[i].exts,
 265                                       TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 266                 balloc = 1;
 267         }
 268         cp->h = p->h;
 269
 270         tcindex_filter_result_init(&new_filter_result);
 271         tcindex_filter_result_init(&cr);
 272         if (old_r)
 273                 cr.res = r->res;
 274
 275         if (tb[TCA_TCINDEX_HASH])
 276                 cp->hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 277
 278         if (tb[TCA_TCINDEX_MASK])
 279                 cp->mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 280
 281         if (tb[TCA_TCINDEX_SHIFT])
 282                 cp->shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 283
 284         err = -EBUSY;
 285
 286         /* Hash already allocated, make sure that we still meet the
 287          * requirements for the allocated hash.
 288          */
 289         if (cp->perfect) {
 290                 if (!valid_perfect_hash(cp) ||
 291                     cp->hash > cp->alloc_hash)
 292                         goto errout_alloc;
 293         } else if (cp->h && cp->hash != cp->alloc_hash) {
 294                 goto errout_alloc;
 295         }
 296
 297         err = -EINVAL;
 298         if (tb[TCA_TCINDEX_FALL_THROUGH])
 299                 cp->fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 300
 301         if (!cp->hash) {
 302                 /* Hash not specified, use perfect hash if the upper limit
 303                  * of the hashing index is below the threshold.
 304                  */
 305                 if ((cp->mask >> cp->shift) < PERFECT_HASH_THRESHOLD)
 306                         cp->hash = (cp->mask >> cp->shift) + 1;
 307                 else
 308                         cp->hash = DEFAULT_HASH_SIZE;
 309         }
 310
 311         if (!cp->perfect && !cp->h)
 312                 cp->alloc_hash = cp->hash;
 313
 314         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 315          * but then, we'd fail handles that may become valid after some future
 316          * mask change. While this is extremely unlikely to ever matter,
 317          * the check below is safer (and also more backwards-compatible).
 318          */
 319         if (cp->perfect || valid_perfect_hash(cp))
 320                 if (handle >= cp->alloc_hash)
 321                         goto errout_alloc;
 322
 323
 324         err = -ENOMEM;
 325         if (!cp->perfect && !cp->h) {
 326                 if (valid_perfect_hash(cp)) {
 327                         int i;
 328
 329                         cp->perfect = kcalloc(cp->hash, sizeof(*r), GFP_KERNEL);
 330                         if (!cp->perfect)
 331                                 goto errout_alloc;
 332                         for (i = 0; i < cp->hash; i++)
 333                                 tcf_exts_init(&cp->perfect[i].exts,
 334                                               TCA_TCINDEX_ACT,
 335                                               TCA_TCINDEX_POLICE);
 336                         balloc = 1;
 337                 } else {
 338                         struct tcindex_filter __rcu **hash;
 339
 340                         hash = kcalloc(cp->hash,
 341                                        sizeof(struct tcindex_filter *),
 342                                        GFP_KERNEL);
 343
 344                         if (!hash)
 345                                 goto errout_alloc;
 346
 347                         cp->h = hash;
 348                         balloc = 2;
 349                 }
 350         }
 351
 352         if (cp->perfect)
 353                 r = cp->perfect + handle;
 354         else
 355                 r = tcindex_lookup(cp, handle) ? : &new_filter_result;
 356
 357         if (r == &new_filter_result) {
 358                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 359                 if (!f)
 360                         goto errout_alloc;
 361                 f->key = handle;
 362                 tcindex_filter_result_init(&f->result);
 363                 f->next = NULL;
 364         }
 365
 366         if (tb[TCA_TCINDEX_CLASSID]) {
 367                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 368                 tcf_bind_filter(tp, &cr.res, base);
 369         }
 370
 371         if (old_r)
 372                 tcf_exts_change(tp, &r->exts, &e);
 373         else
 374                 tcf_exts_change(tp, &cr.exts, &e);
 375
 376         if (old_r && old_r != r)
 377                 tcindex_filter_result_init(old_r);
 378
 379         oldp = p;
 380         r->res = cr.res;
 381         rcu_assign_pointer(tp->root, cp);
 382
 383         if (r == &new_filter_result) {
 384                 struct tcindex_filter *nfp;
 385                 struct tcindex_filter __rcu **fp;
 386
 387                 tcf_exts_change(tp, &f->result.exts, &r->exts);
 388
 389                 fp = cp->h + (handle % cp->hash);
 390                 for (nfp = rtnl_dereference(*fp);
 391                      nfp;
 392                      fp = &nfp->next, nfp = rtnl_dereference(*fp))
 393                                 ; /* nothing */
 394
 395                 rcu_assign_pointer(*fp, f);
 396         }
 397
 398         if (oldp)
 399                 call_rcu(&oldp->rcu, __tcindex_partial_destroy);
 400         return 0;
 401
 402 errout_alloc:
 403         if (balloc == 1)
 404                 kfree(cp->perfect);
 405         else if (balloc == 2)
 406                 kfree(cp->h);
 407 errout:
 408         kfree(cp);
 409         tcf_exts_destroy(&e);
 410         return err;
 411 }
 412
 413 static int
 414 tcindex_change(struct net *net, struct sk_buff *in_skb,
 415                struct tcf_proto *tp, unsigned long base, u32 handle,
 416                struct nlattr **tca, unsigned long *arg, bool ovr)
 417 {
 418         struct nlattr *opt = tca[TCA_OPTIONS];
 419         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 420         struct tcindex_data *p = rtnl_dereference(tp->root);
 421         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 422         int err;
 423
 424         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 425             "p %p,r %p,*arg 0x%lx\n",
 426             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 427
 428         if (!opt)
 429                 return 0;
 430
 431         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 432         if (err < 0)
 433                 return err;
 434
 435         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 436                                  tca[TCA_RATE], ovr);
 437 }
 438
 439 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 440 {
 441         struct tcindex_data *p = rtnl_dereference(tp->root);
 442         struct tcindex_filter *f, *next;
 443         int i;
 444
 445         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 446         if (p->perfect) {
 447                 for (i = 0; i < p->hash; i++) {
 448                         if (!p->perfect[i].res.class)
 449                                 continue;
 450                         if (walker->count >= walker->skip) {
 451                                 if (walker->fn(tp,
 452                                     (unsigned long) (p->perfect+i), walker)
 453                                      < 0) {
 454                                         walker->stop = 1;
 455                                         return;
 456                                 }
 457                         }
 458                         walker->count++;
 459                 }
 460         }
 461         if (!p->h)
 462                 return;
 463         for (i = 0; i < p->hash; i++) {
 464                 for (f = rtnl_dereference(p->h[i]); f; f = next) {
 465                         next = rtnl_dereference(f->next);
 466                         if (walker->count >= walker->skip) {
 467                                 if (walker->fn(tp, (unsigned long) &f->result,
 468                                     walker) < 0) {
 469                                         walker->stop = 1;
 470                                         return;
 471                                 }
 472                         }
 473                         walker->count++;
 474                 }
 475         }
 476 }
 477
 478 static void tcindex_destroy(struct tcf_proto *tp)
 479 {
 480         struct tcindex_data *p = rtnl_dereference(tp->root);
 481         struct tcf_walker walker;
 482
 483         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 484         walker.count = 0;
 485         walker.skip = 0;
 486         walker.fn = tcindex_destroy_element;
 487         tcindex_walk(tp, &walker);
 488
 489         RCU_INIT_POINTER(tp->root, NULL);
 490         call_rcu(&p->rcu, __tcindex_destroy);
 491 }
 492
 493
 494 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 495     struct sk_buff *skb, struct tcmsg *t)
 496 {
 497         struct tcindex_data *p = rtnl_dereference(tp->root);
 498         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 499         unsigned char *b = skb_tail_pointer(skb);
 500         struct nlattr *nest;
 501
 502         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p,b %p\n",
 503                  tp, fh, skb, t, p, r, b);
 504         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 505
 506         nest = nla_nest_start(skb, TCA_OPTIONS);
 507         if (nest == NULL)
 508                 goto nla_put_failure;
 509
 510         if (!fh) {
 511                 t->tcm_handle = ~0; /* whatever ... */
 512                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 513                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 514                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 515                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 516                         goto nla_put_failure;
 517                 nla_nest_end(skb, nest);
 518         } else {
 519                 if (p->perfect) {
 520                         t->tcm_handle = r - p->perfect;
 521                 } else {
 522                         struct tcindex_filter *f;
 523                         struct tcindex_filter __rcu **fp;
 524                         int i;
 525
 526                         t->tcm_handle = 0;
 527                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 528                                 fp = &p->h[i];
 529                                 for (f = rtnl_dereference(*fp);
 530                                      !t->tcm_handle && f;
 531                                      fp = &f->next, f = rtnl_dereference(*fp)) {
 532                                         if (&f->result == r)
 533                                                 t->tcm_handle = f->key;
 534                                 }
 535                         }
 536                 }
 537                 pr_debug("handle = %d\n", t->tcm_handle);
 538                 if (r->res.class &&
 539                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 540                         goto nla_put_failure;
 541
 542                 if (tcf_exts_dump(skb, &r->exts) < 0)
 543                         goto nla_put_failure;
 544                 nla_nest_end(skb, nest);
 545
 546                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 547                         goto nla_put_failure;
 548         }
 549
 550         return skb->len;
 551
 552 nla_put_failure:
 553         nlmsg_trim(skb, b);
 554         return -1;
 555 }
 556
 557 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 558         .kind           =       "tcindex",
 559         .classify       =       tcindex_classify,
 560         .init           =       tcindex_init,
 561         .destroy        =       tcindex_destroy,
 562         .get            =       tcindex_get,
 563         .put            =       tcindex_put,
 564         .change         =       tcindex_change,
 565         .delete         =       tcindex_delete,
 566         .walk           =       tcindex_walk,
 567         .dump           =       tcindex_dump,
 568         .owner          =       THIS_MODULE,
 569 };
 570
 571 static int __init init_tcindex(void)
 572 {
 573         return register_tcf_proto_ops(&cls_tcindex_ops);
 574 }
 575
 576 static void __exit exit_tcindex(void)
 577 {
 578         unregister_tcf_proto_ops(&cls_tcindex_ops);
 579 }
 580
 581 module_init(init_tcindex)
 582 module_exit(exit_tcindex)
 583 MODULE_LICENSE("GPL");