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 /*
  18  * Not quite sure if we need all the xchgs Alexey uses when accessing things.
  19  * Can always add them later ... :)
  20  */
  21
  22 /*
  23  * Passing parameters to the root seems to be done more awkwardly than really
  24  * necessary. At least, u32 doesn't seem to use such dirty hacks. To be
  25  * verified. FIXME.
  26  */
  27
  28 #define PERFECT_HASH_THRESHOLD  64      /* use perfect hash if not bigger */
  29 #define DEFAULT_HASH_SIZE       64      /* optimized for diffserv */
  30
  31
  32 #if 1 /* control */
  33 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
  34 #else
  35 #define DPRINTK(format,args...)
  36 #endif
  37
  38 #if 0 /* data */
  39 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
  40 #else
  41 #define D2PRINTK(format,args...)
  42 #endif
  43
  44
  45 #define PRIV(tp)        ((struct tcindex_data *) (tp)->root)
  46
  47
  48 struct tcindex_filter_result {
  49         struct tcf_exts         exts;
  50         struct tcf_result       res;
  51 };
  52
  53 struct tcindex_filter {
  54         u16 key;
  55         struct tcindex_filter_result result;
  56         struct tcindex_filter *next;
  57 };
  58
  59
  60 struct tcindex_data {
  61         struct tcindex_filter_result *perfect; /* perfect hash; NULL if none */
  62         struct tcindex_filter **h; /* imperfect hash; only used if !perfect;
  63                                       NULL if unused */
  64         u16 mask;               /* AND key with mask */
  65         int shift;              /* shift ANDed key to the right */
  66         int hash;               /* hash table size; 0 if undefined */
  67         int alloc_hash;         /* allocated size */
  68         int fall_through;       /* 0: only classify if explicit match */
  69 };
  70
  71 static struct tcf_ext_map tcindex_ext_map = {
  72         .police = TCA_TCINDEX_POLICE,
  73         .action = TCA_TCINDEX_ACT
  74 };
  75
  76 static inline int
  77 tcindex_filter_is_set(struct tcindex_filter_result *r)
  78 {
  79         return tcf_exts_is_predicative(&r->exts) || r->res.classid;
  80 }
  81
  82 static struct tcindex_filter_result *
  83 tcindex_lookup(struct tcindex_data *p, u16 key)
  84 {
  85         struct tcindex_filter *f;
  86
  87         if (p->perfect)
  88                 return tcindex_filter_is_set(p->perfect + key) ?
  89                         p->perfect + key : NULL;
  90         else if (p->h) {
  91                 for (f = p->h[key % p->hash]; f; f = f->next)
  92                         if (f->key == key)
  93                                 return &f->result;
  94         }
  95
  96         return NULL;
  97 }
  98
  99
 100 static int tcindex_classify(struct sk_buff *skb, struct tcf_proto *tp,
 101                             struct tcf_result *res)
 102 {
 103         struct tcindex_data *p = PRIV(tp);
 104         struct tcindex_filter_result *f;
 105         int key = (skb->tc_index & p->mask) >> p->shift;
 106
 107         D2PRINTK("tcindex_classify(skb %p,tp %p,res %p),p %p\n",skb,tp,res,p);
 108
 109         f = tcindex_lookup(p, key);
 110         if (!f) {
 111                 if (!p->fall_through)
 112                         return -1;
 113                 res->classid = TC_H_MAKE(TC_H_MAJ(tp->q->handle), key);
 114                 res->class = 0;
 115                 D2PRINTK("alg 0x%x\n",res->classid);
 116                 return 0;
 117         }
 118         *res = f->res;
 119         D2PRINTK("map 0x%x\n",res->classid);
 120
 121         return tcf_exts_exec(skb, &f->exts, res);
 122 }
 123
 124
 125 static unsigned long tcindex_get(struct tcf_proto *tp, u32 handle)
 126 {
 127         struct tcindex_data *p = PRIV(tp);
 128         struct tcindex_filter_result *r;
 129
 130         DPRINTK("tcindex_get(tp %p,handle 0x%08x)\n",tp,handle);
 131         if (p->perfect && handle >= p->alloc_hash)
 132                 return 0;
 133         r = tcindex_lookup(p, handle);
 134         return r && tcindex_filter_is_set(r) ? (unsigned long) r : 0UL;
 135 }
 136
 137
 138 static void tcindex_put(struct tcf_proto *tp, unsigned long f)
 139 {
 140         DPRINTK("tcindex_put(tp %p,f 0x%lx)\n",tp,f);
 141 }
 142
 143
 144 static int tcindex_init(struct tcf_proto *tp)
 145 {
 146         struct tcindex_data *p;
 147
 148         DPRINTK("tcindex_init(tp %p)\n",tp);
 149         p = kzalloc(sizeof(struct tcindex_data),GFP_KERNEL);
 150         if (!p)
 151                 return -ENOMEM;
 152
 153         p->mask = 0xffff;
 154         p->hash = DEFAULT_HASH_SIZE;
 155         p->fall_through = 1;
 156
 157         tp->root = p;
 158         return 0;
 159 }
 160
 161
 162 static int
 163 __tcindex_delete(struct tcf_proto *tp, unsigned long arg, int lock)
 164 {
 165         struct tcindex_data *p = PRIV(tp);
 166         struct tcindex_filter_result *r = (struct tcindex_filter_result *) arg;
 167         struct tcindex_filter *f = NULL;
 168
 169         DPRINTK("tcindex_delete(tp %p,arg 0x%lx),p %p,f %p\n",tp,arg,p,f);
 170         if (p->perfect) {
 171                 if (!r->res.class)
 172                         return -ENOENT;
 173         } else {
 174                 int i;
 175                 struct tcindex_filter **walk = NULL;
 176
 177                 for (i = 0; i < p->hash; i++)
 178                         for (walk = p->h+i; *walk; walk = &(*walk)->next)
 179                                 if (&(*walk)->result == r)
 180                                         goto found;
 181                 return -ENOENT;
 182
 183 found:
 184                 f = *walk;
 185                 if (lock)
 186                         tcf_tree_lock(tp);
 187                 *walk = f->next;
 188                 if (lock)
 189                         tcf_tree_unlock(tp);
 190         }
 191         tcf_unbind_filter(tp, &r->res);
 192         tcf_exts_destroy(tp, &r->exts);
 193         kfree(f);
 194         return 0;
 195 }
 196
 197 static int tcindex_delete(struct tcf_proto *tp, unsigned long arg)
 198 {
 199         return __tcindex_delete(tp, arg, 1);
 200 }
 201
 202 static inline int
 203 valid_perfect_hash(struct tcindex_data *p)
 204 {
 205         return  p->hash > (p->mask >> p->shift);
 206 }
 207
 208 static int
 209 tcindex_set_parms(struct tcf_proto *tp, unsigned long base, u32 handle,
 210                   struct tcindex_data *p, struct tcindex_filter_result *r,
 211                   struct rtattr **tb, struct rtattr *est)
 212 {
 213         int err, balloc = 0;
 214         struct tcindex_filter_result new_filter_result, *old_r = r;
 215         struct tcindex_filter_result cr;
 216         struct tcindex_data cp;
 217         struct tcindex_filter *f = NULL; /* make gcc behave */
 218         struct tcf_exts e;
 219
 220         err = tcf_exts_validate(tp, tb, est, &e, &tcindex_ext_map);
 221         if (err < 0)
 222                 return err;
 223
 224         memcpy(&cp, p, sizeof(cp));
 225         memset(&new_filter_result, 0, sizeof(new_filter_result));
 226
 227         if (old_r)
 228                 memcpy(&cr, r, sizeof(cr));
 229         else
 230                 memset(&cr, 0, sizeof(cr));
 231
 232         err = -EINVAL;
 233         if (tb[TCA_TCINDEX_HASH-1]) {
 234                 if (RTA_PAYLOAD(tb[TCA_TCINDEX_HASH-1]) < sizeof(u32))
 235                         goto errout;
 236                 cp.hash = *(u32 *) RTA_DATA(tb[TCA_TCINDEX_HASH-1]);
 237         }
 238
 239         if (tb[TCA_TCINDEX_MASK-1]) {
 240                 if (RTA_PAYLOAD(tb[TCA_TCINDEX_MASK-1]) < sizeof(u16))
 241                         goto errout;
 242                 cp.mask = *(u16 *) RTA_DATA(tb[TCA_TCINDEX_MASK-1]);
 243         }
 244
 245         if (tb[TCA_TCINDEX_SHIFT-1]) {
 246                 if (RTA_PAYLOAD(tb[TCA_TCINDEX_SHIFT-1]) < sizeof(int))
 247                         goto errout;
 248                 cp.shift = *(int *) RTA_DATA(tb[TCA_TCINDEX_SHIFT-1]);
 249         }
 250
 251         err = -EBUSY;
 252         /* Hash already allocated, make sure that we still meet the
 253          * requirements for the allocated hash.
 254          */
 255         if (cp.perfect) {
 256                 if (!valid_perfect_hash(&cp) ||
 257                     cp.hash > cp.alloc_hash)
 258                         goto errout;
 259         } else if (cp.h && cp.hash != cp.alloc_hash)
 260                 goto errout;
 261
 262         err = -EINVAL;
 263         if (tb[TCA_TCINDEX_FALL_THROUGH-1]) {
 264                 if (RTA_PAYLOAD(tb[TCA_TCINDEX_FALL_THROUGH-1]) < sizeof(u32))
 265                         goto errout;
 266                 cp.fall_through =
 267                         *(u32 *) RTA_DATA(tb[TCA_TCINDEX_FALL_THROUGH-1]);
 268         }
 269
 270         if (!cp.hash) {
 271                 /* Hash not specified, use perfect hash if the upper limit
 272                  * of the hashing index is below the threshold.
 273                  */
 274                 if ((cp.mask >> cp.shift) < PERFECT_HASH_THRESHOLD)
 275                         cp.hash = (cp.mask >> cp.shift)+1;
 276                 else
 277                         cp.hash = DEFAULT_HASH_SIZE;
 278         }
 279
 280         if (!cp.perfect && !cp.h)
 281                 cp.alloc_hash = cp.hash;
 282
 283         /* Note: this could be as restrictive as if (handle & ~(mask >> shift))
 284          * but then, we'd fail handles that may become valid after some future
 285          * mask change. While this is extremely unlikely to ever matter,
 286          * the check below is safer (and also more backwards-compatible).
 287          */
 288         if (cp.perfect || valid_perfect_hash(&cp))
 289                 if (handle >= cp.alloc_hash)
 290                         goto errout;
 291
 292
 293         err = -ENOMEM;
 294         if (!cp.perfect && !cp.h) {
 295                 if (valid_perfect_hash(&cp)) {
 296                         cp.perfect = kcalloc(cp.hash, sizeof(*r), GFP_KERNEL);
 297                         if (!cp.perfect)
 298                                 goto errout;
 299                         balloc = 1;
 300                 } else {
 301                         cp.h = kcalloc(cp.hash, sizeof(f), GFP_KERNEL);
 302                         if (!cp.h)
 303                                 goto errout;
 304                         balloc = 2;
 305                 }
 306         }
 307
 308         if (cp.perfect)
 309                 r = cp.perfect + handle;
 310         else
 311                 r = tcindex_lookup(&cp, handle) ? : &new_filter_result;
 312
 313         if (r == &new_filter_result) {
 314                 f = kzalloc(sizeof(*f), GFP_KERNEL);
 315                 if (!f)
 316                         goto errout_alloc;
 317         }
 318
 319         if (tb[TCA_TCINDEX_CLASSID-1]) {
 320                 cr.res.classid = *(u32 *) RTA_DATA(tb[TCA_TCINDEX_CLASSID-1]);
 321                 tcf_bind_filter(tp, &cr.res, base);
 322         }
 323
 324         tcf_exts_change(tp, &cr.exts, &e);
 325
 326         tcf_tree_lock(tp);
 327         if (old_r && old_r != r)
 328                 memset(old_r, 0, sizeof(*old_r));
 329
 330         memcpy(p, &cp, sizeof(cp));
 331         memcpy(r, &cr, sizeof(cr));
 332
 333         if (r == &new_filter_result) {
 334                 struct tcindex_filter **fp;
 335
 336                 f->key = handle;
 337                 f->result = new_filter_result;
 338                 f->next = NULL;
 339                 for (fp = p->h+(handle % p->hash); *fp; fp = &(*fp)->next)
 340                         /* nothing */;
 341                 *fp = f;
 342         }
 343         tcf_tree_unlock(tp);
 344
 345         return 0;
 346
 347 errout_alloc:
 348         if (balloc == 1)
 349                 kfree(cp.perfect);
 350         else if (balloc == 2)
 351                 kfree(cp.h);
 352 errout:
 353         tcf_exts_destroy(tp, &e);
 354         return err;
 355 }
 356
 357 static int
 358 tcindex_change(struct tcf_proto *tp, unsigned long base, u32 handle,
 359                struct rtattr **tca, unsigned long *arg)
 360 {
 361         struct rtattr *opt = tca[TCA_OPTIONS-1];
 362         struct rtattr *tb[TCA_TCINDEX_MAX];
 363         struct tcindex_data *p = PRIV(tp);
 364         struct tcindex_filter_result *r = (struct tcindex_filter_result *) *arg;
 365
 366         DPRINTK("tcindex_change(tp %p,handle 0x%08x,tca %p,arg %p),opt %p,"
 367             "p %p,r %p,*arg 0x%lx\n",
 368             tp, handle, tca, arg, opt, p, r, arg ? *arg : 0L);
 369
 370         if (!opt)
 371                 return 0;
 372
 373         if (rtattr_parse_nested(tb, TCA_TCINDEX_MAX, opt) < 0)
 374                 return -EINVAL;
 375
 376         return tcindex_set_parms(tp, base, handle, p, r, tb, tca[TCA_RATE-1]);
 377 }
 378
 379
 380 static void tcindex_walk(struct tcf_proto *tp, struct tcf_walker *walker)
 381 {
 382         struct tcindex_data *p = PRIV(tp);
 383         struct tcindex_filter *f,*next;
 384         int i;
 385
 386         DPRINTK("tcindex_walk(tp %p,walker %p),p %p\n",tp,walker,p);
 387         if (p->perfect) {
 388                 for (i = 0; i < p->hash; i++) {
 389                         if (!p->perfect[i].res.class)
 390                                 continue;
 391                         if (walker->count >= walker->skip) {
 392                                 if (walker->fn(tp,
 393                                     (unsigned long) (p->perfect+i), walker)
 394                                      < 0) {
 395                                         walker->stop = 1;
 396                                         return;
 397                                 }
 398                         }
 399                         walker->count++;
 400                 }
 401         }
 402         if (!p->h)
 403                 return;
 404         for (i = 0; i < p->hash; i++) {
 405                 for (f = p->h[i]; f; f = next) {
 406                         next = f->next;
 407                         if (walker->count >= walker->skip) {
 408                                 if (walker->fn(tp,(unsigned long) &f->result,
 409                                     walker) < 0) {
 410                                         walker->stop = 1;
 411                                         return;
 412                                 }
 413                         }
 414                         walker->count++;
 415                 }
 416         }
 417 }
 418
 419
 420 static int tcindex_destroy_element(struct tcf_proto *tp,
 421     unsigned long arg, struct tcf_walker *walker)
 422 {
 423         return __tcindex_delete(tp, arg, 0);
 424 }
 425
 426
 427 static void tcindex_destroy(struct tcf_proto *tp)
 428 {
 429         struct tcindex_data *p = PRIV(tp);
 430         struct tcf_walker walker;
 431
 432         DPRINTK("tcindex_destroy(tp %p),p %p\n",tp,p);
 433         walker.count = 0;
 434         walker.skip = 0;
 435         walker.fn = &tcindex_destroy_element;
 436         tcindex_walk(tp,&walker);
 437         kfree(p->perfect);
 438         kfree(p->h);
 439         kfree(p);
 440         tp->root = NULL;
 441 }
 442
 443
 444 static int tcindex_dump(struct tcf_proto *tp, unsigned long fh,
 445     struct sk_buff *skb, struct tcmsg *t)
 446 {
 447         struct tcindex_data *p = PRIV(tp);
 448         struct tcindex_filter_result *r = (struct tcindex_filter_result *) fh;
 449         unsigned char *b = skb_tail_pointer(skb);
 450         struct rtattr *rta;
 451
 452         DPRINTK("tcindex_dump(tp %p,fh 0x%lx,skb %p,t %p),p %p,r %p,b %p\n",
 453             tp,fh,skb,t,p,r,b);
 454         DPRINTK("p->perfect %p p->h %p\n",p->perfect,p->h);
 455         rta = (struct rtattr *) b;
 456         RTA_PUT(skb,TCA_OPTIONS,0,NULL);
 457         if (!fh) {
 458                 t->tcm_handle = ~0; /* whatever ... */
 459                 RTA_PUT(skb,TCA_TCINDEX_HASH,sizeof(p->hash),&p->hash);
 460                 RTA_PUT(skb,TCA_TCINDEX_MASK,sizeof(p->mask),&p->mask);
 461                 RTA_PUT(skb,TCA_TCINDEX_SHIFT,sizeof(p->shift),&p->shift);
 462                 RTA_PUT(skb,TCA_TCINDEX_FALL_THROUGH,sizeof(p->fall_through),
 463                     &p->fall_through);
 464                 rta->rta_len = skb_tail_pointer(skb) - b;
 465         } else {
 466                 if (p->perfect) {
 467                         t->tcm_handle = r-p->perfect;
 468                 } else {
 469                         struct tcindex_filter *f;
 470                         int i;
 471
 472                         t->tcm_handle = 0;
 473                         for (i = 0; !t->tcm_handle && i < p->hash; i++) {
 474                                 for (f = p->h[i]; !t->tcm_handle && f;
 475                                      f = f->next) {
 476                                         if (&f->result == r)
 477                                                 t->tcm_handle = f->key;
 478                                 }
 479                         }
 480                 }
 481                 DPRINTK("handle = %d\n",t->tcm_handle);
 482                 if (r->res.class)
 483                         RTA_PUT(skb, TCA_TCINDEX_CLASSID, 4, &r->res.classid);
 484
 485                 if (tcf_exts_dump(skb, &r->exts, &tcindex_ext_map) < 0)
 486                         goto rtattr_failure;
 487                 rta->rta_len = skb_tail_pointer(skb) - b;
 488
 489                 if (tcf_exts_dump_stats(skb, &r->exts, &tcindex_ext_map) < 0)
 490                         goto rtattr_failure;
 491         }
 492
 493         return skb->len;
 494
 495 rtattr_failure:
 496         nlmsg_trim(skb, b);
 497         return -1;
 498 }
 499
 500 static struct tcf_proto_ops cls_tcindex_ops = {
 501         .next           =       NULL,
 502         .kind           =       "tcindex",
 503         .classify       =       tcindex_classify,
 504         .init           =       tcindex_init,
 505         .destroy        =       tcindex_destroy,
 506         .get            =       tcindex_get,
 507         .put            =       tcindex_put,
 508         .change         =       tcindex_change,
 509         .delete         =       tcindex_delete,
 510         .walk           =       tcindex_walk,
 511         .dump           =       tcindex_dump,
 512         .owner          =       THIS_MODULE,
 513 };
 514
 515 static int __init init_tcindex(void)
 516 {
 517         return register_tcf_proto_ops(&cls_tcindex_ops);
 518 }
 519
 520 static void __exit exit_tcindex(void)
 521 {
 522         unregister_tcf_proto_ops(&cls_tcindex_ops);
 523 }
 524
 525 module_init(init_tcindex)
 526 module_exit(exit_tcindex)
 527 MODULE_LICENSE("GPL");