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