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 const struct tcf_ext_map tcindex_ext_map = {
  54         .police = TCA_TCINDEX_POLICE,
  55         .action = TCA_TCINDEX_ACT
  56 };
  57
  58 static inline int
  59 tcindex_filter_is_set(struct tcindex_filter_result *r)
  60 {
  61         return tcf_exts_is_predicative(&r->exts) || r->res.classid;
  62 }
  63
  64 static struct tcindex_filter_result *
  65 tcindex_lookup(struct tcindex_data *p, u16 key)
  66 {
  67         struct tcindex_filter *f;
  68
  69         if (p->perfect)
  70                 return tcindex_filter_is_set(p->perfect + key) ?
  71                         p->perfect + key : NULL;
  72         else if (p->h) {
  73                 for (f = p->h[key % p->hash]; f; f = f->next)
  74                         if (f->key == key)
  75                                 return &f->result;
  76         }
  77
  78         return NULL;
  79 }
  80
  81
  82 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  83                             struct tcf_result *res)
  84 {
  85         struct tcindex_data *p = PRIV(tp);
  86         struct tcindex_filter_result *f;
  87         int key = (skb->tc_index & p->mask) >> p->shift;
  88
  89         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
  90                  skb, tp, res, p);
  91
  92         f = tcindex_lookup(p, key);
  93         if (!f) {
  94                 if (!p->fall_through)
  95                         return -1;
  96                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
  97                 res->class = 0;
  98                 pr_debug("alg 0x%x\n", res->classid);
  99                 return 0;
 100         }
 101         *res = f->res;
 102         pr_debug("map 0x%x\n", res->classid);
 103
 104         return tcf_exts_exec(skb, &f->exts, res);
 105 }
 106
 107
 108 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
 109 {
 110         struct tcindex_data *p = PRIV(tp);
 111         struct tcindex_filter_result *r;
 112
 113         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
 114         if (p->perfect && handle >= p->alloc_hash)
 115                 return 0;
 116         r = tcindex_lookup(p, handle);
 117         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 118 }
 119
 120
 121 static void tcindex_put(struct tcf_proto *tp, unsigned long f)
 122 {
 123         pr_debug("tcindex_put(tp %p,f 0x%lx)\n", tp, f);
 124 }
 125
 126
 127 static int tcindex_init(struct tcf_proto *tp)
 128 {
 129         struct tcindex_data *p;
 130
 131         pr_debug("tcindex_init(tp %p)\n", tp);
 132         p = kzalloc(sizeof(struct tcindex_data), GFP_KERNEL);
 133         if (!p)
 134                 return -ENOMEM;
 135
 136         p->mask = 0xffff;
 137         p->hash = DEFAULT_HASH_SIZE;
 138         p->fall_through = 1;
 139
 140         tp->root = p;
 141         return 0;
 142 }
 143
 144
 145 static int
 146 __tcindex_delete(struct tcf_proto *tp, unsigned long arg, int lock)
 147 {
 148         struct tcindex_data *p = PRIV(tp);
 149         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 150         struct tcindex_filter *f = NULL;
 151
 152         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p,f %p\n", tp, arg, p, f);
 153         if (p->perfect) {
 154                 if (!r->res.class)
 155                         return -ENOENT;
 156         } else {
 157                 int i;
 158                 struct tcindex_filter **walk = NULL;
 159
 160                 for (i = 0; i < p->hash; i++)
 161                         for (walk = p->h+i; *walk; walk = &(*walk)->next)
 162                                 if (&(*walk)->result == r)
 163                                         goto found;
 164                 return -ENOENT;
 165
 166 found:
 167                 f = *walk;
 168                 if (lock)
 169                         tcf_tree_lock(tp);
 170                 *walk = f->next;
 171                 if (lock)
 172                         tcf_tree_unlock(tp);
 173         }
 174         tcf_unbind_filter(tp, &r->res);
 175         tcf_exts_destroy(tp, &r->exts);
 176         kfree(f);
 177         return 0;
 178 }
 179
 180 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 181 {
 182         return __tcindex_delete(tp, arg, 1);
 183 }
 184
 185 static inline int
 186 valid_perfect_hash(struct tcindex_data *p)
 187 {
 188         return  p->hash > (p->mask >> p->shift);
 189 }
 190
 191 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 192         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 193         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 194         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 195         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 196         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 197 };
 198
 199 static int
 200 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 201                   u32 handle, struct tcindex_data *p,
 202                   struct tcindex_filter_result *r, struct nlattr **tb,
 203                  struct nlattr *est)
 204 {
 205         int err, balloc = 0;
 206         struct tcindex_filter_result new_filter_result, *old_r = r;
 207         struct tcindex_filter_result cr;
 208         struct tcindex_data cp;
 209         struct tcindex_filter *f = NULL; /* make gcc behave */
 210         struct tcf_exts e;
 211
 212         err = tcf_exts_validate(net, tp, tb, est, &e, &tcindex_ext_map);
 213         if (err < 0)
 214                 return err;
 215
 216         memcpy(&cp, p, sizeof(cp));
 217         memset(&new_filter_result, 0, sizeof(new_filter_result));
 218         tcf_exts_init(&new_filter_result.exts);
 219
 220         if (old_r)
 221                 memcpy(&cr, r, sizeof(cr));
 222         else {
 223                 memset(&cr, 0, sizeof(cr));
 224                 tcf_exts_init(&cr.exts);
 225         }
 226
 227         if (tb[TCA_TCINDEX_HASH])
 228                 cp.hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 229
 230         if (tb[TCA_TCINDEX_MASK])
 231                 cp.mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 232
 233         if (tb[TCA_TCINDEX_SHIFT])
 234                 cp.shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 235
 236         err = -EBUSY;
 237         /* Hash already allocated, make sure that we still meet the
 238          * requirements for the allocated hash.
 239          */
 240         if (cp.perfect) {
 241                 if (!valid_perfect_hash(&cp) ||
 242                     cp.hash > cp.alloc_hash)
 243                         goto errout;
 244         } else if (cp.h && cp.hash != cp.alloc_hash)
 245                 goto errout;
 246
 247         err = -EINVAL;
 248         if (tb[TCA_TCINDEX_FALL_THROUGH])
 249                 cp.fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 250
 251         if (!cp.hash) {
 252                 /* Hash not specified, use perfect hash if the upper limit
 253                  * of the hashing index is below the threshold.
 254                  */
 255                 if ((cp.mask >> cp.shift) < PERFECT_HASH_THRESHOLD)
 256                         cp.hash = (cp.mask >> cp.shift) + 1;
 257                 else
 258                         cp.hash = DEFAULT_HASH_SIZE;
 259         }
 260
 261         if (!cp.perfect && !cp.h)
 262                 cp.alloc_hash = cp.hash;
 263
 264         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 265          * but then, we'd fail handles that may become valid after some future
 266          * mask change. While this is extremely unlikely to ever matter,
 267          * the check below is safer (and also more backwards-compatible).
 268          */
 269         if (cp.perfect || valid_perfect_hash(&cp))
 270                 if (handle >= cp.alloc_hash)
 271                         goto errout;
 272
 273
 274         err = -ENOMEM;
 275         if (!cp.perfect && !cp.h) {
 276                 if (valid_perfect_hash(&cp)) {
 277                         cp.perfect = kcalloc(cp.hash, sizeof(*r), GFP_KERNEL);
 278                         if (!cp.perfect)
 279                                 goto errout;
 280                         balloc = 1;
 281                 } else {
 282                         cp.h = kcalloc(cp.hash, sizeof(f), GFP_KERNEL);
 283                         if (!cp.h)
 284                                 goto errout;
 285                         balloc = 2;
 286                 }
 287         }
 288
 289         if (cp.perfect)
 290                 r = cp.perfect + handle;
 291         else
 292                 r = tcindex_lookup(&cp, handle) ? : &new_filter_result;
 293
 294         if (r == &new_filter_result) {
 295                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 296                 if (!f)
 297                         goto errout_alloc;
 298         }
 299
 300         if (tb[TCA_TCINDEX_CLASSID]) {
 301                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 302                 tcf_bind_filter(tp, &cr.res, base);
 303         }
 304
 305         tcf_exts_change(tp, &cr.exts, &e);
 306
 307         tcf_tree_lock(tp);
 308         if (old_r && old_r != r)
 309                 memset(old_r, 0, sizeof(*old_r));
 310
 311         memcpy(p, &cp, sizeof(cp));
 312         memcpy(r, &cr, sizeof(cr));
 313
 314         if (r == &new_filter_result) {
 315                 struct tcindex_filter **fp;
 316
 317                 f->key = handle;
 318                 f->result = new_filter_result;
 319                 f->next = NULL;
 320                 for (fp = p->h+(handle % p->hash); *fp; fp = &(*fp)->next)
 321                         /* nothing */;
 322                 *fp = f;
 323         }
 324         tcf_tree_unlock(tp);
 325
 326         return 0;
 327
 328 errout_alloc:
 329         if (balloc == 1)
 330                 kfree(cp.perfect);
 331         else if (balloc == 2)
 332                 kfree(cp.h);
 333 errout:
 334         tcf_exts_destroy(tp, &e);
 335         return err;
 336 }
 337
 338 static int
 339 tcindex_change(struct net *net, struct sk_buff *in_skb,
 340                struct tcf_proto *tp, unsigned long base, u32 handle,
 341                struct nlattr **tca, unsigned long *arg)
 342 {
 343         struct nlattr *opt = tca[TCA_OPTIONS];
 344         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 345         struct tcindex_data *p = PRIV(tp);
 346         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 347         int err;
 348
 349         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 350             "p %p,r %p,*arg 0x%lx\n",
 351             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 352
 353         if (!opt)
 354                 return 0;
 355
 356         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 357         if (err < 0)
 358                 return err;
 359
 360         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 361                                  tca[TCA_RATE]);
 362 }
 363
 364
 365 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 366 {
 367         struct tcindex_data *p = PRIV(tp);
 368         struct tcindex_filter *f, *next;
 369         int i;
 370
 371         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 372         if (p->perfect) {
 373                 for (i = 0; i < p->hash; i++) {
 374                         if (!p->perfect[i].res.class)
 375                                 continue;
 376                         if (walker->count >= walker->skip) {
 377                                 if (walker->fn(tp,
 378                                     (unsigned long) (p->perfect+i), walker)
 379                                      < 0) {
 380                                         walker->stop = 1;
 381                                         return;
 382                                 }
 383                         }
 384                         walker->count++;
 385                 }
 386         }
 387         if (!p->h)
 388                 return;
 389         for (i = 0; i < p->hash; i++) {
 390                 for (f = p->h[i]; f; f = next) {
 391                         next = f->next;
 392                         if (walker->count >= walker->skip) {
 393                                 if (walker->fn(tp, (unsigned long) &f->result,
 394                                     walker) < 0) {
 395                                         walker->stop = 1;
 396                                         return;
 397                                 }
 398                         }
 399                         walker->count++;
 400                 }
 401         }
 402 }
 403
 404
 405 static int tcindex_destroy_element(struct tcf_proto *tp,
 406     unsigned long arg, struct tcf_walker *walker)
 407 {
 408         return __tcindex_delete(tp, arg, 0);
 409 }
 410
 411
 412 static void tcindex_destroy(struct tcf_proto *tp)
 413 {
 414         struct tcindex_data *p = PRIV(tp);
 415         struct tcf_walker walker;
 416
 417         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 418         walker.count = 0;
 419         walker.skip = 0;
 420         walker.fn = &tcindex_destroy_element;
 421         tcindex_walk(tp, &walker);
 422         kfree(p->perfect);
 423         kfree(p->h);
 424         kfree(p);
 425         tp->root = NULL;
 426 }
 427
 428
 429 static int tcindex_dump(struct tcf_proto *tp, unsigned long fh,
 430     struct sk_buff *skb, struct tcmsg *t)
 431 {
 432         struct tcindex_data *p = PRIV(tp);
 433         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 434         unsigned char *b = skb_tail_pointer(skb);
 435         struct nlattr *nest;
 436
 437         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p,b %p\n",
 438                  tp, fh, skb, t, p, r, b);
 439         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 440
 441         nest = nla_nest_start(skb, TCA_OPTIONS);
 442         if (nest == NULL)
 443                 goto nla_put_failure;
 444
 445         if (!fh) {
 446                 t->tcm_handle = ~0; /* whatever ... */
 447                 if (nla_put_u32(skb, TCA_TCINDEX_HASH, p->hash) ||
 448                     nla_put_u16(skb, TCA_TCINDEX_MASK, p->mask) ||
 449                     nla_put_u32(skb, TCA_TCINDEX_SHIFT, p->shift) ||
 450                     nla_put_u32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through))
 451                         goto nla_put_failure;
 452                 nla_nest_end(skb, nest);
 453         } else {
 454                 if (p->perfect) {
 455                         t->tcm_handle = r-p->perfect;
 456                 } else {
 457                         struct tcindex_filter *f;
 458                         int i;
 459
 460                         t->tcm_handle = 0;
 461                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 462                                 for (f = p->h[i]; !t->tcm_handle && f;
 463                                      f = f->next) {
 464                                         if (&f->result == r)
 465                                                 t->tcm_handle = f->key;
 466                                 }
 467                         }
 468                 }
 469                 pr_debug("handle = %d\n", t->tcm_handle);
 470                 if (r->res.class &&
 471                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 472                         goto nla_put_failure;
 473
 474                 if (tcf_exts_dump(skb, &r->exts, &tcindex_ext_map) < 0)
 475                         goto nla_put_failure;
 476                 nla_nest_end(skb, nest);
 477
 478                 if (tcf_exts_dump_stats(skb, &r->exts, &tcindex_ext_map) < 0)
 479                         goto nla_put_failure;
 480         }
 481
 482         return skb->len;
 483
 484 nla_put_failure:
 485         nlmsg_trim(skb, b);
 486         return -1;
 487 }
 488
 489 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 490         .kind           =       "tcindex",
 491         .classify       =       tcindex_classify,
 492         .init           =       tcindex_init,
 493         .destroy        =       tcindex_destroy,
 494         .get            =       tcindex_get,
 495         .put            =       tcindex_put,
 496         .change         =       tcindex_change,
 497         .delete         =       tcindex_delete,
 498         .walk           =       tcindex_walk,
 499         .dump           =       tcindex_dump,
 500         .owner          =       THIS_MODULE,
 501 };
 502
 503 static int __init init_tcindex(void)
 504 {
 505         return register_tcf_proto_ops(&cls_tcindex_ops);
 506 }
 507
 508 static void __exit exit_tcindex(void)
 509 {
 510         unregister_tcf_proto_ops(&cls_tcindex_ops);
 511 }
 512
 513 module_init(init_tcindex)
 514 module_exit(exit_tcindex)
 515 MODULE_LICENSE("GPL");