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 *next;
  36 };
  37
  38
  39 struct tcindex_data {
  40         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  41         struct tcindex_filter **h; /* imperfect hash; only used if !perfect;
  42                                       NULL if unused */
  43         u16 mask;               /* AND key with mask */
  44         int shift;              /* shift ANDed key to the right */
  45         int hash;               /* hash table size; 0 if undefined */
  46         int alloc_hash;         /* allocated size */
  47         int fall_through;       /* 0: only classify if explicit match */
  48 };
  49
  50 static inline int
  51 tcindex_filter_is_set(struct tcindex_filter_result *r)
  52 {
  53         return tcf_exts_is_predicative(&r->exts) || r->res.classid;
  54 }
  55
  56 static struct tcindex_filter_result *
  57 tcindex_lookup(struct tcindex_data *p, u16 key)
  58 {
  59         struct tcindex_filter *f;
  60
  61         if (p->perfect)
  62                 return tcindex_filter_is_set(p->perfect + key) ?
  63                         p->perfect + key : NULL;
  64         else if (p->h) {
  65                 for (f = p->h[key % p->hash]; f; f = f->next)
  66                         if (f->key == key)
  67                                 return &f->result;
  68         }
  69
  70         return NULL;
  71 }
  72
  73
  74 static int tcindex_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  75                             struct tcf_result *res)
  76 {
  77         struct tcindex_data *p = tp->root;
  78         struct tcindex_filter_result *f;
  79         int key = (skb->tc_index & p->mask) >> p->shift;
  80
  81         pr_debug("tcindex_classify(skb %p,tp %p,res %p),p %p\n",
  82                  skb, tp, res, p);
  83
  84         f = tcindex_lookup(p, key);
  85         if (!f) {
  86                 if (!p->fall_through)
  87                         return -1;
  88                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
  89                 res->class = 0;
  90                 pr_debug("alg 0x%x\n", res->classid);
  91                 return 0;
  92         }
  93         *res = f->res;
  94         pr_debug("map 0x%x\n", res->classid);
  95
  96         return tcf_exts_exec(skb, &f->exts, res);
  97 }
  98
  99
 100 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
 101 {
 102         struct tcindex_data *p = tp->root;
 103         struct tcindex_filter_result *r;
 104
 105         pr_debug("tcindex_get(tp %p,handle 0x%08x)\n", tp, handle);
 106         if (p->perfect && handle >= p->alloc_hash)
 107                 return 0;
 108         r = tcindex_lookup(p, handle);
 109         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 110 }
 111
 112
 113 static void tcindex_put(struct tcf_proto *tp, unsigned long f)
 114 {
 115         pr_debug("tcindex_put(tp %p,f 0x%lx)\n", tp, f);
 116 }
 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         tp->root = p;
 133         return 0;
 134 }
 135
 136
 137 static int
 138 __tcindex_delete(struct tcf_proto *tp, unsigned long arg, int lock)
 139 {
 140         struct tcindex_data *p = tp->root;
 141         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 142         struct tcindex_filter *f = NULL;
 143
 144         pr_debug("tcindex_delete(tp %p,arg 0x%lx),p %p,f %p\n", tp, arg, p, f);
 145         if (p->perfect) {
 146                 if (!r->res.class)
 147                         return -ENOENT;
 148         } else {
 149                 int i;
 150                 struct tcindex_filter **walk = NULL;
 151
 152                 for (i = 0; i < p->hash; i++)
 153                         for (walk = p->h+i; *walk; walk = &(*walk)->next)
 154                                 if (&(*walk)->result == r)
 155                                         goto found;
 156                 return -ENOENT;
 157
 158 found:
 159                 f = *walk;
 160                 if (lock)
 161                         tcf_tree_lock(tp);
 162                 *walk = f->next;
 163                 if (lock)
 164                         tcf_tree_unlock(tp);
 165         }
 166         tcf_unbind_filter(tp, &r->res);
 167         tcf_exts_destroy(tp, &r->exts);
 168         kfree(f);
 169         return 0;
 170 }
 171
 172 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 173 {
 174         return __tcindex_delete(tp, arg, 1);
 175 }
 176
 177 static inline int
 178 valid_perfect_hash(struct tcindex_data *p)
 179 {
 180         return  p->hash > (p->mask >> p->shift);
 181 }
 182
 183 static const struct nla_policy tcindex_policy[TCA_TCINDEX_MAX + 1] = {
 184         [TCA_TCINDEX_HASH]              = { .type = NLA_U32 },
 185         [TCA_TCINDEX_MASK]              = { .type = NLA_U16 },
 186         [TCA_TCINDEX_SHIFT]             = { .type = NLA_U32 },
 187         [TCA_TCINDEX_FALL_THROUGH]      = { .type = NLA_U32 },
 188         [TCA_TCINDEX_CLASSID]           = { .type = NLA_U32 },
 189 };
 190
 191 static int
 192 tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
 193                   u32 handle, struct tcindex_data *p,
 194                   struct tcindex_filter_result *r, struct nlattr **tb,
 195                  struct nlattr *est)
 196 {
 197         int err, balloc = 0;
 198         struct tcindex_filter_result new_filter_result, *old_r = r;
 199         struct tcindex_filter_result cr;
 200         struct tcindex_data cp;
 201         struct tcindex_filter *f = NULL; /* make gcc behave */
 202         struct tcf_exts e;
 203
 204         tcf_exts_init(&e, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 205         err = tcf_exts_validate(net, tp, tb, est, &e);
 206         if (err < 0)
 207                 return err;
 208
 209         memcpy(&cp, p, sizeof(cp));
 210         memset(&new_filter_result, 0, sizeof(new_filter_result));
 211         tcf_exts_init(&new_filter_result.exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 212
 213         if (old_r)
 214                 memcpy(&cr, r, sizeof(cr));
 215         else {
 216                 memset(&cr, 0, sizeof(cr));
 217                 tcf_exts_init(&cr.exts, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
 218         }
 219
 220         if (tb[TCA_TCINDEX_HASH])
 221                 cp.hash = nla_get_u32(tb[TCA_TCINDEX_HASH]);
 222
 223         if (tb[TCA_TCINDEX_MASK])
 224                 cp.mask = nla_get_u16(tb[TCA_TCINDEX_MASK]);
 225
 226         if (tb[TCA_TCINDEX_SHIFT])
 227                 cp.shift = nla_get_u32(tb[TCA_TCINDEX_SHIFT]);
 228
 229         err = -EBUSY;
 230         /* Hash already allocated, make sure that we still meet the
 231          * requirements for the allocated hash.
 232          */
 233         if (cp.perfect) {
 234                 if (!valid_perfect_hash(&cp) ||
 235                     cp.hash > cp.alloc_hash)
 236                         goto errout;
 237         } else if (cp.h && cp.hash != cp.alloc_hash)
 238                 goto errout;
 239
 240         err = -EINVAL;
 241         if (tb[TCA_TCINDEX_FALL_THROUGH])
 242                 cp.fall_through = nla_get_u32(tb[TCA_TCINDEX_FALL_THROUGH]);
 243
 244         if (!cp.hash) {
 245                 /* Hash not specified, use perfect hash if the upper limit
 246                  * of the hashing index is below the threshold.
 247                  */
 248                 if ((cp.mask >> cp.shift) < PERFECT_HASH_THRESHOLD)
 249                         cp.hash = (cp.mask >> cp.shift) + 1;
 250                 else
 251                         cp.hash = DEFAULT_HASH_SIZE;
 252         }
 253
 254         if (!cp.perfect && !cp.h)
 255                 cp.alloc_hash = cp.hash;
 256
 257         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 258          * but then, we'd fail handles that may become valid after some future
 259          * mask change. While this is extremely unlikely to ever matter,
 260          * the check below is safer (and also more backwards-compatible).
 261          */
 262         if (cp.perfect || valid_perfect_hash(&cp))
 263                 if (handle >= cp.alloc_hash)
 264                         goto errout;
 265
 266
 267         err = -ENOMEM;
 268         if (!cp.perfect && !cp.h) {
 269                 if (valid_perfect_hash(&cp)) {
 270                         cp.perfect = kcalloc(cp.hash, sizeof(*r), GFP_KERNEL);
 271                         if (!cp.perfect)
 272                                 goto errout;
 273                         balloc = 1;
 274                 } else {
 275                         cp.h = kcalloc(cp.hash, sizeof(f), GFP_KERNEL);
 276                         if (!cp.h)
 277                                 goto errout;
 278                         balloc = 2;
 279                 }
 280         }
 281
 282         if (cp.perfect)
 283                 r = cp.perfect + handle;
 284         else
 285                 r = tcindex_lookup(&cp, handle) ? : &new_filter_result;
 286
 287         if (r == &new_filter_result) {
 288                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 289                 if (!f)
 290                         goto errout_alloc;
 291         }
 292
 293         if (tb[TCA_TCINDEX_CLASSID]) {
 294                 cr.res.classid = nla_get_u32(tb[TCA_TCINDEX_CLASSID]);
 295                 tcf_bind_filter(tp, &cr.res, base);
 296         }
 297
 298         tcf_exts_change(tp, &cr.exts, &e);
 299
 300         tcf_tree_lock(tp);
 301         if (old_r && old_r != r)
 302                 memset(old_r, 0, sizeof(*old_r));
 303
 304         memcpy(p, &cp, sizeof(cp));
 305         memcpy(r, &cr, sizeof(cr));
 306
 307         if (r == &new_filter_result) {
 308                 struct tcindex_filter **fp;
 309
 310                 f->key = handle;
 311                 f->result = new_filter_result;
 312                 f->next = NULL;
 313                 for (fp = p->h+(handle % p->hash); *fp; fp = &(*fp)->next)
 314                         /* nothing */;
 315                 *fp = f;
 316         }
 317         tcf_tree_unlock(tp);
 318
 319         return 0;
 320
 321 errout_alloc:
 322         if (balloc == 1)
 323                 kfree(cp.perfect);
 324         else if (balloc == 2)
 325                 kfree(cp.h);
 326 errout:
 327         tcf_exts_destroy(tp, &e);
 328         return err;
 329 }
 330
 331 static int
 332 tcindex_change(struct net *net, struct sk_buff *in_skb,
 333                struct tcf_proto *tp, unsigned long base, u32 handle,
 334                struct nlattr **tca, unsigned long *arg)
 335 {
 336         struct nlattr *opt = tca[TCA_OPTIONS];
 337         struct nlattr *tb[TCA_TCINDEX_MAX + 1];
 338         struct tcindex_data *p = tp->root;
 339         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 340         int err;
 341
 342         pr_debug("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 343             "p %p,r %p,*arg 0x%lx\n",
 344             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 345
 346         if (!opt)
 347                 return 0;
 348
 349         err = nla_parse_nested(tb, TCA_TCINDEX_MAX, opt, tcindex_policy);
 350         if (err < 0)
 351                 return err;
 352
 353         return tcindex_set_parms(net, tp, base, handle, p, r, tb,
 354                                  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 = tp->root;
 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 = tp->root;
 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 net *net, struct tcf_proto *tp, unsigned long fh,
 423     struct sk_buff *skb, struct tcmsg *t)
 424 {
 425         struct tcindex_data *p = tp->root;
 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                 if (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                         goto nla_put_failure;
 445                 nla_nest_end(skb, nest);
 446         } else {
 447                 if (p->perfect) {
 448                         t->tcm_handle = r-p->perfect;
 449                 } else {
 450                         struct tcindex_filter *f;
 451                         int i;
 452
 453                         t->tcm_handle = 0;
 454                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 455                                 for (f = p->h[i]; !t->tcm_handle && f;
 456                                      f = f->next) {
 457                                         if (&f->result == r)
 458                                                 t->tcm_handle = f->key;
 459                                 }
 460                         }
 461                 }
 462                 pr_debug("handle = %d\n", t->tcm_handle);
 463                 if (r->res.class &&
 464                     nla_put_u32(skb, TCA_TCINDEX_CLASSID, r->res.classid))
 465                         goto nla_put_failure;
 466
 467                 if (tcf_exts_dump(skb, &r->exts) < 0)
 468                         goto nla_put_failure;
 469                 nla_nest_end(skb, nest);
 470
 471                 if (tcf_exts_dump_stats(skb, &r->exts) < 0)
 472                         goto nla_put_failure;
 473         }
 474
 475         return skb->len;
 476
 477 nla_put_failure:
 478         nlmsg_trim(skb, b);
 479         return -1;
 480 }
 481
 482 static struct tcf_proto_ops cls_tcindex_ops __read_mostly = {
 483         .kind           =       "tcindex",
 484         .classify       =       tcindex_classify,
 485         .init           =       tcindex_init,
 486         .destroy        =       tcindex_destroy,
 487         .get            =       tcindex_get,
 488         .put            =       tcindex_put,
 489         .change         =       tcindex_change,
 490         .delete         =       tcindex_delete,
 491         .walk           =       tcindex_walk,
 492         .dump           =       tcindex_dump,
 493         .owner          =       THIS_MODULE,
 494 };
 495
 496 static int __init init_tcindex(void)
 497 {
 498         return register_tcf_proto_ops(&cls_tcindex_ops);
 499 }
 500
 501 static void __exit exit_tcindex(void)
 502 {
 503         unregister_tcf_proto_ops(&cls_tcindex_ops);
 504 }
 505
 506 module_init(init_tcindex)
 507 module_exit(exit_tcindex)
 508 MODULE_LICENSE("GPL");