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 <net/act_api.h>
  13 #include <net/netlink.h>
  14 #include <net/pkt_cls.h>
  15
  16 /*
  17  * Passing parameters to the root seems to be done more awkwardly than really
  18  * necessary. At least, u32 doesn't seem to use such dirty hacks. To be
  19  * verified. FIXME.
  20  */
  21
  22 #define PERFECT_HASH_THRESHOLD  64      /* use perfect hash if not bigger */
  23 #define DEFAULT_HASH_SIZE       64      /* optimized for diffserv */
  24
  25
  26 #define PRIV(tp)        ((struct tcindex_data *) (tp)->root)
  27
  28
  29 struct tcindex_filter_result {
  30         struct tcf_exts         exts;
  31         struct tcf_result       res;
  32 };
  33
  34 struct tcindex_filter {
  35         u16 key;
  36         struct tcindex_filter_result result;
  37         struct tcindex_filter *next;
  38 };
  39
  40
  41 struct tcindex_data {
  42         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  43         struct tcindex_filter **h; /* imperfect hash; only used if !perfect;
  44                                       NULL if unused */
  45         u16 mask;               /* AND key with mask */
  46         int shift;              /* shift ANDed key to the right */
  47         int hash;               /* hash table size; 0 if undefined */
  48         int alloc_hash;         /* allocated size */
  49         int fall_through;       /* 0: only classify if explicit match */
  50 };
  51
  52 static const struct tcf_ext_map tcindex_ext_map = {
  53         .police = TCA_TCINDEX_POLICE,
  54         .action = TCA_TCINDEX_ACT
  55 };
  56
  57 static inline int
  58 tcindex_filter_is_set(struct tcindex_filter_result *r)
  59 {
  60         return tcf_exts_is_predicative(&r->exts) || r->res.classid;
  61 }
  62
  63 static struct tcindex_filter_result *
  64 tcindex_lookup(struct tcindex_data *p, u16 key)
  65 {
  66         struct tcindex_filter *f;
  67
  68         if (p->perfect)
  69                 return tcindex_filter_is_set(p->perfect + key) ?
  70                         p->perfect + key : NULL;
  71         else if (p->h) {
  72                 for (f = p->h[key % p->hash]; f; f = f->next)
  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, struct tcf_proto *tp,
  82                             struct tcf_result *res)
  83 {
  84         struct tcindex_data *p = PRIV(tp);
  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 = PRIV(tp);
 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         tp->root = p;
 140         return 0;
 141 }
 142
 143
 144 static int
 145 __tcindex_delete(struct tcf_proto *tp, unsigned long arg, int lock)
 146 {
 147         struct tcindex_data *p = PRIV(tp);
 148         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 149         struct tcindex_filter *f = NULL;
 150
 151         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p,f %p\n", tp, arg, p, f);
 152         if (p->perfect) {
 153                 if (!r->res.class)
 154                         return -ENOENT;
 155         } else {
 156                 int i;
 157                 struct tcindex_filter **walk = NULL;
 158
 159                 for (i = 0; i < p->hash; i++)
 160                         for (walk = p->h+i; *walk; walk = &(*walk)->next)
 161                                 if (&(*walk)->result == r)
 162                                         goto found;
 163                 return -ENOENT;
 164
 165 found:
 166                 f = *walk;
 167                 if (lock)
 168                         tcf_tree_lock(tp);
 169                 *walk = f->next;
 170                 if (lock)
 171                         tcf_tree_unlock(tp);
 172         }
 173         tcf_unbind_filter(tp, &r->res);
 174         tcf_exts_destroy(tp, &r->exts);
 175         kfree(f);
 176         return 0;
 177 }
 178
 179 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 180 {
 181         return __tcindex_delete(tp, arg, 1);
 182 }
 183
 184 static inline int
 185 valid_perfect_hash(struct tcindex_data *p)
 186 {
 187         return  p->hash > (p->mask >> p->shift);
 188 }
 189
 190 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 191         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 192         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 193         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 194         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 195         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 196 };
 197
 198 static int
 199 tcindex_set_parms(struct tcf_proto *tp, unsigned long base, u32 handle,
 200                   struct tcindex_data *p, struct tcindex_filter_result *r,
 201                   struct nlattr **tb, struct nlattr *est)
 202 {
 203         int err, balloc = 0;
 204         struct tcindex_filter_result new_filter_result, *old_r = r;
 205         struct tcindex_filter_result cr;
 206         struct tcindex_data cp;
 207         struct tcindex_filter *f = NULL; /* make gcc behave */
 208         struct tcf_exts e;
 209
 210         err = tcf_exts_validate(tp, tb, est, &e, &tcindex_ext_map);
 211         if (err < 0)
 212                 return err;
 213
 214         memcpy(&cp, p, sizeof(cp));
 215         memset(&new_filter_result, 0, sizeof(new_filter_result));
 216
 217         if (old_r)
 218                 memcpy(&cr, r, sizeof(cr));
 219         else
 220                 memset(&cr, 0, sizeof(cr));
 221
 222         if (tb[TCA_TCINDEX_HASH])
 223                 cp.hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 224
 225         if (tb[TCA_TCINDEX_MASK])
 226                 cp.mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 227
 228         if (tb[TCA_TCINDEX_SHIFT])
 229                 cp.shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 230
 231         err = -EBUSY;
 232         /* Hash already allocated, make sure that we still meet the
 233          * requirements for the allocated hash.
 234          */
 235         if (cp.perfect) {
 236                 if (!valid_perfect_hash(&cp) ||
 237                     cp.hash > cp.alloc_hash)
 238                         goto errout;
 239         } else if (cp.h && cp.hash != cp.alloc_hash)
 240                 goto errout;
 241
 242         err = -EINVAL;
 243         if (tb[TCA_TCINDEX_FALL_THROUGH])
 244                 cp.fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 245
 246         if (!cp.hash) {
 247                 /* Hash not specified, use perfect hash if the upper limit
 248                  * of the hashing index is below the threshold.
 249                  */
 250                 if ((cp.mask >> cp.shift) < PERFECT_HASH_THRESHOLD)
 251                         cp.hash = (cp.mask >> cp.shift)+1;
 252                 else
 253                         cp.hash = DEFAULT_HASH_SIZE;
 254         }
 255
 256         if (!cp.perfect && !cp.h)
 257                 cp.alloc_hash = cp.hash;
 258
 259         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 260          * but then, we'd fail handles that may become valid after some future
 261          * mask change. While this is extremely unlikely to ever matter,
 262          * the check below is safer (and also more backwards-compatible).
 263          */
 264         if (cp.perfect || valid_perfect_hash(&cp))
 265                 if (handle >= cp.alloc_hash)
 266                         goto errout;
 267
 268
 269         err = -ENOMEM;
 270         if (!cp.perfect && !cp.h) {
 271                 if (valid_perfect_hash(&cp)) {
 272                         cp.perfect = kcalloc(cp.hash, sizeof(*r), GFP_KERNEL);
 273                         if (!cp.perfect)
 274                                 goto errout;
 275                         balloc = 1;
 276                 } else {
 277                         cp.h = kcalloc(cp.hash, sizeof(f), GFP_KERNEL);
 278                         if (!cp.h)
 279                                 goto errout;
 280                         balloc = 2;
 281                 }
 282         }
 283
 284         if (cp.perfect)
 285                 r = cp.perfect + handle;
 286         else
 287                 r = tcindex_lookup(&cp, handle) ? : &new_filter_result;
 288
 289         if (r == &new_filter_result) {
 290                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 291                 if (!f)
 292                         goto errout_alloc;
 293         }
 294
 295         if (tb[TCA_TCINDEX_CLASSID]) {
 296                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 297                 tcf_bind_filter(tp, &cr.res, base);
 298         }
 299
 300         tcf_exts_change(tp, &cr.exts, &e);
 301
 302         tcf_tree_lock(tp);
 303         if (old_r && old_r != r)
 304                 memset(old_r, 0, sizeof(*old_r));
 305
 306         memcpy(p, &cp, sizeof(cp));
 307         memcpy(r, &cr, sizeof(cr));
 308
 309         if (r == &new_filter_result) {
 310                 struct tcindex_filter **fp;
 311
 312                 f->key = handle;
 313                 f->result = new_filter_result;
 314                 f->next = NULL;
 315                 for (fp = p->h+(handle % p->hash); *fp; fp = &(*fp)->next)
 316                         /* nothing */;
 317                 *fp = f;
 318         }
 319         tcf_tree_unlock(tp);
 320
 321         return 0;
 322
 323 errout_alloc:
 324         if (balloc == 1)
 325                 kfree(cp.perfect);
 326         else if (balloc == 2)
 327                 kfree(cp.h);
 328 errout:
 329         tcf_exts_destroy(tp, &e);
 330         return err;
 331 }
 332
 333 static int
 334 tcindex_change(struct tcf_proto *tp, unsigned long base, u32 handle,
 335                struct nlattr **tca, unsigned long *arg)
 336 {
 337         struct nlattr *opt = tca[TCA_OPTIONS];
 338         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 339         struct tcindex_data *p = PRIV(tp);
 340         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 341         int err;
 342
 343         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 344             "p %p,r %p,*arg 0x%lx\n",
 345             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 346
 347         if (!opt)
 348                 return 0;
 349
 350         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 351         if (err < 0)
 352                 return err;
 353
 354         return tcindex_set_parms(tp, base, handle, p, r, tb, tca[TCA_RATE]);
 355 }
 356
 357
 358 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 359 {
 360         struct tcindex_data *p = PRIV(tp);
 361         struct tcindex_filter *f, *next;
 362         int i;
 363
 364         pr_debug("tcindex_walk(tp %p,walker %p),p %p\n", tp, walker, p);
 365         if (p->perfect) {
 366                 for (i = 0; i < p->hash; i++) {
 367                         if (!p->perfect[i].res.class)
 368                                 continue;
 369                         if (walker->count >= walker->skip) {
 370                                 if (walker->fn(tp,
 371                                     (unsigned long) (p->perfect+i), walker)
 372                                      < 0) {
 373                                         walker->stop = 1;
 374                                         return;
 375                                 }
 376                         }
 377                         walker->count++;
 378                 }
 379         }
 380         if (!p->h)
 381                 return;
 382         for (i = 0; i < p->hash; i++) {
 383                 for (f = p->h[i]; f; f = next) {
 384                         next = f->next;
 385                         if (walker->count >= walker->skip) {
 386                                 if (walker->fn(tp, (unsigned long) &f->result,
 387                                     walker) < 0) {
 388                                         walker->stop = 1;
 389                                         return;
 390                                 }
 391                         }
 392                         walker->count++;
 393                 }
 394         }
 395 }
 396
 397
 398 static int tcindex_destroy_element(struct tcf_proto *tp,
 399     unsigned long arg, struct tcf_walker *walker)
 400 {
 401         return __tcindex_delete(tp, arg, 0);
 402 }
 403
 404
 405 static void tcindex_destroy(struct tcf_proto *tp)
 406 {
 407         struct tcindex_data *p = PRIV(tp);
 408         struct tcf_walker walker;
 409
 410         pr_debug("tcindex_destroy(tp %p),p %p\n", tp, p);
 411         walker.count = 0;
 412         walker.skip = 0;
 413         walker.fn = &tcindex_destroy_element;
 414         tcindex_walk(tp, &walker);
 415         kfree(p->perfect);
 416         kfree(p->h);
 417         kfree(p);
 418         tp->root = NULL;
 419 }
 420
 421
 422 static int tcindex_dump(struct tcf_proto *tp, unsigned long fh,
 423     struct sk_buff *skb, struct tcmsg *t)
 424 {
 425         struct tcindex_data *p = PRIV(tp);
 426         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 427         unsigned char *b = skb_tail_pointer(skb);
 428         struct nlattr *nest;
 429
 430         pr_debug("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p,b %p\n",
 431                  tp, fh, skb, t, p, r, b);
 432         pr_debug("p->perfect %p p->h %p\n", p->perfect, p->h);
 433
 434         nest = nla_nest_start(skb, TCA_OPTIONS);
 435         if (nest == NULL)
 436                 goto nla_put_failure;
 437
 438         if (!fh) {
 439                 t->tcm_handle = ~0; /* whatever ... */
 440                 NLA_PUT_U32(skb, TCA_TCINDEX_HASH, p->hash);
 441                 NLA_PUT_U16(skb, TCA_TCINDEX_MASK, p->mask);
 442                 NLA_PUT_U32(skb, TCA_TCINDEX_SHIFT, p->shift);
 443                 NLA_PUT_U32(skb, TCA_TCINDEX_FALL_THROUGH, p->fall_through);
 444                 nla_nest_end(skb, nest);
 445         } else {
 446                 if (p->perfect) {
 447                         t->tcm_handle = r-p->perfect;
 448                 } else {
 449                         struct tcindex_filter *f;
 450                         int i;
 451
 452                         t->tcm_handle = 0;
 453                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 454                                 for (f = p->h[i]; !t->tcm_handle && f;
 455                                      f = f->next) {
 456                                         if (&f->result == r)
 457                                                 t->tcm_handle = f->key;
 458                                 }
 459                         }
 460                 }
 461                 pr_debug("handle = %d\n", t->tcm_handle);
 462                 if (r->res.class)
 463                         NLA_PUT_U32(skb, TCA_TCINDEX_CLASSID, r->res.classid);
 464
 465                 if (tcf_exts_dump(skb, &r->exts, &tcindex_ext_map) < 0)
 466                         goto nla_put_failure;
 467                 nla_nest_end(skb, nest);
 468
 469                 if (tcf_exts_dump_stats(skb, &r->exts, &tcindex_ext_map) < 0)
 470                         goto nla_put_failure;
 471         }
 472
 473         return skb->len;
 474
 475 nla_put_failure:
 476         nlmsg_trim(skb, b);
 477         return -1;
 478 }
 479
 480 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 481         .kind           =       "tcindex",
 482         .classify       =       tcindex_classify,
 483         .init           =       tcindex_init,
 484         .destroy        =       tcindex_destroy,
 485         .get            =       tcindex_get,
 486         .put            =       tcindex_put,
 487         .change         =       tcindex_change,
 488         .delete         =       tcindex_delete,
 489         .walk           =       tcindex_walk,
 490         .dump           =       tcindex_dump,
 491         .owner          =       THIS_MODULE,
 492 };
 493
 494 static int __init init_tcindex(void)
 495 {
 496         return register_tcf_proto_ops(&cls_tcindex_ops);
 497 }
 498
 499 static void __exit exit_tcindex(void)
 500 {
 501         unregister_tcf_proto_ops(&cls_tcindex_ops);
 502 }
 503
 504 module_init(init_tcindex)
 505 module_exit(exit_tcindex)
 506 MODULE_LICENSE("GPL");