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 #define PRIV(tp)        ((struct tcindex_data *) (tp)->root)
  28
  29
  30 struct tcindex_filter_result {
  31         struct tcf_exts         exts;
  32         struct tcf_result       res;
  33 };
  34
  35 struct tcindex_filter {
  36         u16 key;
  37         struct tcindex_filter_result result;
  38         struct tcindex_filter *next;
  39 };
  40
  41
  42 struct tcindex_data {
  43         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  44         struct tcindex_filter **h; /* imperfect hash; only used if !perfect;
  45                                       NULL if unused */
  46         u16 mask;               /* AND key with mask */
  47         int shift;              /* shift ANDed key to the right */
  48         int hash;               /* hash table size; 0 if undefined */
  49         int alloc_hash;         /* allocated size */
  50         int fall_through;       /* 0: only classify if explicit match */
  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         struct tcindex_filter *f;
  63
  64         if (p->perfect)
  65                 return tcindex_filter_is_set(p->perfect + key) ?
  66                         p->perfect + key : NULL;
  67         else if (p->h) {
  68                 for (f = p->h[key % p->hash]; f; f = f->next)
  69                         if (f->key == key)
  70                                 return &f->result;
  71         }
  72
  73         return NULL;
  74 }
  75
  76
  77 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  78                             struct tcf_result *res)
  79 {
  80         struct tcindex_data *p = PRIV(tp);
  81         struct tcindex_filter_result *f;
  82         int key = (skb->tc_index & p->mask) >> p->shift;
  83
  84         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
  85                  skb, tp, res, p);
  86
  87         f = tcindex_lookup(p, key);
  88         if (!f) {
  89                 if (!p->fall_through)
  90                         return -1;
  91                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
  92                 res->class = 0;
  93                 pr_debug("alg 0x%x\n", res->classid);
  94                 return 0;
  95         }
  96         *res = f->res;
  97         pr_debug("map 0x%x\n", res->classid);
  98
  99         return tcf_exts_exec(skb, &f->exts, res);
 100 }
 101
 102
 103 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
 104 {
 105         struct tcindex_data *p = PRIV(tp);
 106         struct tcindex_filter_result *r;
 107
 108         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
 109         if (p->perfect && handle >= p->alloc_hash)
 110                 return 0;
 111         r = tcindex_lookup(p, handle);
 112         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 113 }
 114
 115
 116 static void tcindex_put(struct tcf_proto *tp, unsigned long f)
 117 {
 118         pr_debug("tcindex_put(tp %p,f 0x%lx)\n", tp, f);
 119 }
 120
 121
 122 static int tcindex_init(struct tcf_proto *tp)
 123 {
 124         struct tcindex_data *p;
 125
 126         pr_debug("tcindex_init(tp %p)\n", tp);
 127         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
 128         if (!p)
 129                 return -ENOMEM;
 130
 131         p->mask = 0xffff;
 132         p->hash = DEFAULT_HASH_SIZE;
 133         p->fall_through = 1;
 134
 135         tp->root = p;
 136         return 0;
 137 }
 138
 139
 140 static int
 141 __tcindex_delete(struct tcf_proto *tp, unsigned long arg, int lock)
 142 {
 143         struct tcindex_data *p = PRIV(tp);
 144         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 145         struct tcindex_filter *f = NULL;
 146
 147         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p,f %p\n", tp, arg, p, f);
 148         if (p->perfect) {
 149                 if (!r->res.class)
 150                         return -ENOENT;
 151         } else {
 152                 int i;
 153                 struct tcindex_filter **walk = NULL;
 154
 155                 for (i = 0; i < p->hash; i++)
 156                         for (walk = p->h+i; *walk; walk = &(*walk)->next)
 157                                 if (&(*walk)->result == r)
 158                                         goto found;
 159                 return -ENOENT;
 160
 161 found:
 162                 f = *walk;
 163                 if (lock)
 164                         tcf_tree_lock(tp);
 165                 *walk = f->next;
 166                 if (lock)
 167                         tcf_tree_unlock(tp);
 168         }
 169         tcf_unbind_filter(tp, &r->res);
 170         tcf_exts_destroy(tp, &r->exts);
 171         kfree(f);
 172         return 0;
 173 }
 174
 175 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 176 {
 177         return __tcindex_delete(tp, arg, 1);
 178 }
 179
 180 static inline int
 181 valid_perfect_hash(struct tcindex_data *p)
 182 {
 183         return  p->hash > (p->mask >> p->shift);
 184 }
 185
 186 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 187         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 188         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 189         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 190         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 191         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 192 };
 193
 194 static int
 195 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 196                   u32 handle, struct tcindex_data *p,
 197                   struct tcindex_filter_result *r, struct nlattr **tb,
 198                  struct nlattr *est)
 199 {
 200         int err, balloc = 0;
 201         struct tcindex_filter_result new_filter_result, *old_r = r;
 202         struct tcindex_filter_result cr;
 203         struct tcindex_data cp;
 204         struct tcindex_filter *f = NULL; /* make gcc behave */
 205         struct tcf_exts e;
 206
 207         tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 208         err = tcf_exts_validate(net, tp, tb, est, &e);
 209         if (err < 0)
 210                 return err;
 211
 212         memcpy(&cp, p, sizeof(cp));
 213         memset(&new_filter_result, 0, sizeof(new_filter_result));
 214         tcf_exts_init(&new_filter_result.exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 215
 216         if (old_r)
 217                 memcpy(&cr, r, sizeof(cr));
 218         else {
 219                 memset(&cr, 0, sizeof(cr));
 220                 tcf_exts_init(&cr.exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 221         }
 222
 223         if (tb[TCA_TCINDEX_HASH])
 224                 cp.hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 225
 226         if (tb[TCA_TCINDEX_MASK])
 227                 cp.mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 228
 229         if (tb[TCA_TCINDEX_SHIFT])
 230                 cp.shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 231
 232         err = -EBUSY;
 233         /* Hash already allocated, make sure that we still meet the
 234          * requirements for the allocated hash.
 235          */
 236         if (cp.perfect) {
 237                 if (!valid_perfect_hash(&cp) ||
 238                     cp.hash > cp.alloc_hash)
 239                         goto errout;
 240         } else if (cp.h && cp.hash != cp.alloc_hash)
 241                 goto errout;
 242
 243         err = -EINVAL;
 244         if (tb[TCA_TCINDEX_FALL_THROUGH])
 245                 cp.fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 246
 247         if (!cp.hash) {
 248                 /* Hash not specified, use perfect hash if the upper limit
 249                  * of the hashing index is below the threshold.
 250                  */
 251                 if ((cp.mask >> cp.shift) < PERFECT_HASH_THRESHOLD)
 252                         cp.hash = (cp.mask >> cp.shift) + 1;
 253                 else
 254                         cp.hash = DEFAULT_HASH_SIZE;
 255         }
 256
 257         if (!cp.perfect && !cp.h)
 258                 cp.alloc_hash = cp.hash;
 259
 260         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 261          * but then, we'd fail handles that may become valid after some future
 262          * mask change. While this is extremely unlikely to ever matter,
 263          * the check below is safer (and also more backwards-compatible).
 264          */
 265         if (cp.perfect || valid_perfect_hash(&cp))
 266                 if (handle >= cp.alloc_hash)
 267                         goto errout;
 268
 269
 270         err = -ENOMEM;
 271         if (!cp.perfect && !cp.h) {
 272                 if (valid_perfect_hash(&cp)) {
 273                         cp.perfect = kcalloc(cp.hash, sizeof(*r), GFP_KERNEL);
 274                         if (!cp.perfect)
 275                                 goto errout;
 276                         balloc = 1;
 277                 } else {
 278                         cp.h = kcalloc(cp.hash, sizeof(f), GFP_KERNEL);
 279                         if (!cp.h)
 280                                 goto errout;
 281                         balloc = 2;
 282                 }
 283         }
 284
 285         if (cp.perfect)
 286                 r = cp.perfect + handle;
 287         else
 288                 r = tcindex_lookup(&cp, handle) ? : &new_filter_result;
 289
 290         if (r == &new_filter_result) {
 291                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 292                 if (!f)
 293                         goto errout_alloc;
 294         }
 295
 296         if (tb[TCA_TCINDEX_CLASSID]) {
 297                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 298                 tcf_bind_filter(tp, &cr.res, base);
 299         }
 300
 301         tcf_exts_change(tp, &cr.exts, &e);
 302
 303         tcf_tree_lock(tp);
 304         if (old_r && old_r != r)
 305                 memset(old_r, 0, sizeof(*old_r));
 306
 307         memcpy(p, &cp, sizeof(cp));
 308         memcpy(r, &cr, sizeof(cr));
 309
 310         if (r == &new_filter_result) {
 311                 struct tcindex_filter **fp;
 312
 313                 f->key = handle;
 314                 f->result = new_filter_result;
 315                 f->next = NULL;
 316                 for (fp = p->h+(handle % p->hash); *fp; fp = &(*fp)->next)
 317                         /* nothing */;
 318                 *fp = f;
 319         }
 320         tcf_tree_unlock(tp);
 321
 322         return 0;
 323
 324 errout_alloc:
 325         if (balloc == 1)
 326                 kfree(cp.perfect);
 327         else if (balloc == 2)
 328                 kfree(cp.h);
 329 errout:
 330         tcf_exts_destroy(tp, &e);
 331         return err;
 332 }
 333
 334 static int
 335 tcindex_change(struct net *net, struct sk_buff *in_skb,
 336                struct tcf_proto *tp, unsigned long base, u32 handle,
 337                struct nlattr **tca, unsigned long *arg)
 338 {
 339         struct nlattr *opt = tca[TCA_OPTIONS];
 340         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 341         struct tcindex_data *p = PRIV(tp);
 342         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 343         int err;
 344
 345         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 346             "p %p,r %p,*arg 0x%lx\n",
 347             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 348
 349         if (!opt)
 350                 return 0;
 351
 352         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 353         if (err < 0)
 354                 return err;
 355
 356         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 357                                  tca[TCA_RATE]);
 358 }
 359
 360
 361 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 362 {
 363         struct tcindex_data *p = PRIV(tp);
 364         struct tcindex_filter *f, *next;
 365         int i;
 366
 367         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 368         if (p->perfect) {
 369                 for (i = 0; i < p->hash; i++) {
 370                         if (!p->perfect[i].res.class)
 371                                 continue;
 372                         if (walker->count >= walker->skip) {
 373                                 if (walker->fn(tp,
 374                                     (unsigned long) (p->perfect+i), walker)
 375                                      < 0) {
 376                                         walker->stop = 1;
 377                                         return;
 378                                 }
 379                         }
 380                         walker->count++;
 381                 }
 382         }
 383         if (!p->h)
 384                 return;
 385         for (i = 0; i < p->hash; i++) {
 386                 for (f = p->h[i]; f; f = next) {
 387                         next = f->next;
 388                         if (walker->count >= walker->skip) {
 389                                 if (walker->fn(tp, (unsigned long) &f->result,
 390                                     walker) < 0) {
 391                                         walker->stop = 1;
 392                                         return;
 393                                 }
 394                         }
 395                         walker->count++;
 396                 }
 397         }
 398 }
 399
 400
 401 static int tcindex_destroy_element(struct tcf_proto *tp,
 402     unsigned long arg, struct tcf_walker *walker)
 403 {
 404         return __tcindex_delete(tp, arg, 0);
 405 }
 406
 407
 408 static void tcindex_destroy(struct tcf_proto *tp)
 409 {
 410         struct tcindex_data *p = PRIV(tp);
 411         struct tcf_walker walker;
 412
 413         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 414         walker.count = 0;
 415         walker.skip = 0;
 416         walker.fn = &tcindex_destroy_element;
 417         tcindex_walk(tp, &walker);
 418         kfree(p->perfect);
 419         kfree(p->h);
 420         kfree(p);
 421         tp->root = NULL;
 422 }
 423
 424
 425 static int tcindex_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 426     struct sk_buff *skb, struct tcmsg *t)
 427 {
 428         struct tcindex_data *p = PRIV(tp);
 429         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 430         unsigned char *b = skb_tail_pointer(skb);
 431         struct nlattr *nest;
 432
 433         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p,b %p\n",
 434                  tp, fh, skb, t, p, r, b);
 435         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 436
 437         nest = nla_nest_start(skb, TCA_OPTIONS);
 438         if (nest == NULL)
 439                 goto nla_put_failure;
 440
 441         if (!fh) {
 442                 t->tcm_handle = ~0; /* whatever ... */
 443                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 444                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 445                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 446                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 447                         goto nla_put_failure;
 448                 nla_nest_end(skb, nest);
 449         } else {
 450                 if (p->perfect) {
 451                         t->tcm_handle = r-p->perfect;
 452                 } else {
 453                         struct tcindex_filter *f;
 454                         int i;
 455
 456                         t->tcm_handle = 0;
 457                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 458                                 for (f = p->h[i]; !t->tcm_handle && f;
 459                                      f = f->next) {
 460                                         if (&f->result == r)
 461                                                 t->tcm_handle = f->key;
 462                                 }
 463                         }
 464                 }
 465                 pr_debug("handle = %d\n", t->tcm_handle);
 466                 if (r->res.class &&
 467                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 468                         goto nla_put_failure;
 469
 470                 if (tcf_exts_dump(skb, &r->exts) < 0)
 471                         goto nla_put_failure;
 472                 nla_nest_end(skb, nest);
 473
 474                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 475                         goto nla_put_failure;
 476         }
 477
 478         return skb->len;
 479
 480 nla_put_failure:
 481         nlmsg_trim(skb, b);
 482         return -1;
 483 }
 484
 485 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 486         .kind           =       "tcindex",
 487         .classify       =       tcindex_classify,
 488         .init           =       tcindex_init,
 489         .destroy        =       tcindex_destroy,
 490         .get            =       tcindex_get,
 491         .put            =       tcindex_put,
 492         .change         =       tcindex_change,
 493         .delete         =       tcindex_delete,
 494         .walk           =       tcindex_walk,
 495         .dump           =       tcindex_dump,
 496         .owner          =       THIS_MODULE,
 497 };
 498
 499 static int __init init_tcindex(void)
 500 {
 501         return register_tcf_proto_ops(&cls_tcindex_ops);
 502 }
 503
 504 static void __exit exit_tcindex(void)
 505 {
 506         unregister_tcf_proto_ops(&cls_tcindex_ops);
 507 }
 508
 509 module_init(init_tcindex)
 510 module_exit(exit_tcindex)
 511 MODULE_LICENSE("GPL");