net/sched/sch_dsmark.c

   1 /* net/sched/sch_dsmark.c - Differentiated Services field marker */
   2
   3 /* Written 1998-2000 by Werner Almesberger, EPFL ICA */
   4
   5
   6 #include <linux/config.h>
   7 #include <linux/module.h>
   8 #include <linux/types.h>
   9 #include <linux/string.h>
  10 #include <linux/errno.h>
  11 #include <linux/skbuff.h>
  12 #include <linux/netdevice.h> /* for pkt_sched */
  13 #include <linux/rtnetlink.h>
  14 #include <net/pkt_sched.h>
  15 #include <net/dsfield.h>
  16 #include <net/inet_ecn.h>
  17 #include <asm/byteorder.h>
  18
  19
  20 #if 1 /* control */
  21 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
  22 #else
  23 #define DPRINTK(format,args...)
  24 #endif
  25
  26 #if 0 /* data */
  27 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
  28 #else
  29 #define D2PRINTK(format,args...)
  30 #endif
  31
  32
  33 #define PRIV(sch) ((struct dsmark_qdisc_data *) (sch)->data)
  34
  35
  36 /*
  37  * classid      class           marking
  38  * -------      -----           -------
  39  *   n/a          0             n/a
  40  *   x:0          1             use entry [0]
  41  *   ...         ...            ...
  42  *   x:y y>0     y+1            use entry [y]
  43  *   ...         ...            ...
  44  * x:indices-1  indices         use entry [indices-1]
  45  *   ...         ...            ...
  46  *   x:y         y+1            use entry [y & (indices-1)]
  47  *   ...         ...            ...
  48  * 0xffff       0x10000         use entry [indices-1]
  49  */
  50
  51
  52 #define NO_DEFAULT_INDEX        (1 << 16)
  53
  54 struct dsmark_qdisc_data {
  55         struct Qdisc            *q;
  56         struct tcf_proto        *filter_list;
  57         __u8                    *mask;  /* "owns" the array */
  58         __u8                    *value;
  59         __u16                   indices;
  60         __u32                   default_index;  /* index range is 0...0xffff */
  61         int                     set_tc_index;
  62 };
  63
  64
  65 /* ------------------------- Class/flow operations ------------------------- */
  66
  67
  68 static int dsmark_graft(struct Qdisc *sch,unsigned long arg,
  69     struct Qdisc *new,struct Qdisc **old)
  70 {
  71         struct dsmark_qdisc_data *p = PRIV(sch);
  72
  73         DPRINTK("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",sch,p,new,
  74             old);
  75         if (!new)
  76                 new = &noop_qdisc;
  77         sch_tree_lock(sch);
  78         *old = xchg(&p->q,new);
  79         if (*old)
  80                 qdisc_reset(*old);
  81         sch->q.qlen = 0;
  82         sch_tree_unlock(sch); /* @@@ move up ? */
  83         return 0;
  84 }
  85
  86
  87 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
  88 {
  89         struct dsmark_qdisc_data *p = PRIV(sch);
  90
  91         return p->q;
  92 }
  93
  94
  95 static unsigned long dsmark_get(struct Qdisc *sch,u32 classid)
  96 {
  97         struct dsmark_qdisc_data *p __attribute__((unused)) = PRIV(sch);
  98
  99         DPRINTK("dsmark_get(sch %p,[qdisc %p],classid %x)\n",sch,p,classid);
 100         return TC_H_MIN(classid)+1;
 101 }
 102
 103
 104 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
 105     unsigned long parent, u32 classid)
 106 {
 107         return dsmark_get(sch,classid);
 108 }
 109
 110
 111 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
 112 {
 113 }
 114
 115
 116 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
 117     struct rtattr **tca, unsigned long *arg)
 118 {
 119         struct dsmark_qdisc_data *p = PRIV(sch);
 120         struct rtattr *opt = tca[TCA_OPTIONS-1];
 121         struct rtattr *tb[TCA_DSMARK_MAX];
 122
 123         DPRINTK("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
 124             "arg 0x%lx\n",sch,p,classid,parent,*arg);
 125         if (*arg > p->indices)
 126                 return -ENOENT;
 127         if (!opt || rtattr_parse(tb, TCA_DSMARK_MAX, RTA_DATA(opt),
 128                                  RTA_PAYLOAD(opt)))
 129                 return -EINVAL;
 130         if (tb[TCA_DSMARK_MASK-1]) {
 131                 if (!RTA_PAYLOAD(tb[TCA_DSMARK_MASK-1]))
 132                         return -EINVAL;
 133                 p->mask[*arg-1] = *(__u8 *) RTA_DATA(tb[TCA_DSMARK_MASK-1]);
 134         }
 135         if (tb[TCA_DSMARK_VALUE-1]) {
 136                 if (!RTA_PAYLOAD(tb[TCA_DSMARK_VALUE-1]))
 137                         return -EINVAL;
 138                 p->value[*arg-1] = *(__u8 *) RTA_DATA(tb[TCA_DSMARK_VALUE-1]);
 139         }
 140         return 0;
 141 }
 142
 143
 144 static int dsmark_delete(struct Qdisc *sch,unsigned long arg)
 145 {
 146         struct dsmark_qdisc_data *p = PRIV(sch);
 147
 148         if (!arg || arg > p->indices)
 149                 return -EINVAL;
 150         p->mask[arg-1] = 0xff;
 151         p->value[arg-1] = 0;
 152         return 0;
 153 }
 154
 155
 156 static void dsmark_walk(struct Qdisc *sch,struct qdisc_walker *walker)
 157 {
 158         struct dsmark_qdisc_data *p = PRIV(sch);
 159         int i;
 160
 161         DPRINTK("dsmark_walk(sch %p,[qdisc %p],walker %p)\n",sch,p,walker);
 162         if (walker->stop)
 163                 return;
 164         for (i = 0; i < p->indices; i++) {
 165                 if (p->mask[i] == 0xff && !p->value[i])
 166                         continue;
 167                 if (walker->count >= walker->skip) {
 168                         if (walker->fn(sch, i+1, walker) < 0) {
 169                                 walker->stop = 1;
 170                                 break;
 171                         }
 172                 }
 173                 walker->count++;
 174         }
 175 }
 176
 177
 178 static struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,unsigned long cl)
 179 {
 180         struct dsmark_qdisc_data *p = PRIV(sch);
 181
 182         return &p->filter_list;
 183 }
 184
 185
 186 /* --------------------------- Qdisc operations ---------------------------- */
 187
 188
 189 static int dsmark_enqueue(struct sk_buff *skb,struct Qdisc *sch)
 190 {
 191         struct dsmark_qdisc_data *p = PRIV(sch);
 192         struct tcf_result res;
 193         int result;
 194         int ret = NET_XMIT_POLICED;
 195
 196         D2PRINTK("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n",skb,sch,p);
 197         if (p->set_tc_index) {
 198                 switch (skb->protocol) {
 199                         case __constant_htons(ETH_P_IP):
 200                                 skb->tc_index = ipv4_get_dsfield(skb->nh.iph)
 201                                         & ~INET_ECN_MASK;
 202                                 break;
 203                         case __constant_htons(ETH_P_IPV6):
 204                                 skb->tc_index = ipv6_get_dsfield(skb->nh.ipv6h)
 205                                         & ~INET_ECN_MASK;
 206                                 break;
 207                         default:
 208                                 skb->tc_index = 0;
 209                                 break;
 210                 };
 211         }
 212         result = TC_POLICE_OK; /* be nice to gcc */
 213         if (TC_H_MAJ(skb->priority) == sch->handle) {
 214                 skb->tc_index = TC_H_MIN(skb->priority);
 215         } else {
 216                 result = tc_classify(skb,p->filter_list,&res);
 217                 D2PRINTK("result %d class 0x%04x\n",result,res.classid);
 218                 switch (result) {
 219 #ifdef CONFIG_NET_CLS_POLICE
 220                         case TC_POLICE_SHOT:
 221                                 kfree_skb(skb);
 222                                 break;
 223 #if 0
 224                         case TC_POLICE_RECLASSIFY:
 225                                 /* FIXME: what to do here ??? */
 226 #endif
 227 #endif
 228                         case TC_POLICE_OK:
 229                                 skb->tc_index = TC_H_MIN(res.classid);
 230                                 break;
 231                         case TC_POLICE_UNSPEC:
 232                                 /* fall through */
 233                         default:
 234                                 if (p->default_index != NO_DEFAULT_INDEX)
 235                                         skb->tc_index = p->default_index;
 236                                 break;
 237                 };
 238         }
 239         if (
 240 #ifdef CONFIG_NET_CLS_POLICE
 241             result == TC_POLICE_SHOT ||
 242 #endif
 243
 244             ((ret = p->q->enqueue(skb,p->q)) != 0)) {
 245                 sch->stats.drops++;
 246                 return ret;
 247         }
 248         sch->stats.bytes += skb->len;
 249         sch->stats.packets++;
 250         sch->q.qlen++;
 251         return ret;
 252 }
 253
 254
 255 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 256 {
 257         struct dsmark_qdisc_data *p = PRIV(sch);
 258         struct sk_buff *skb;
 259         int index;
 260
 261         D2PRINTK("dsmark_dequeue(sch %p,[qdisc %p])\n",sch,p);
 262         skb = p->q->ops->dequeue(p->q);
 263         if (!skb)
 264                 return NULL;
 265         sch->q.qlen--;
 266         index = skb->tc_index & (p->indices-1);
 267         D2PRINTK("index %d->%d\n",skb->tc_index,index);
 268         switch (skb->protocol) {
 269                 case __constant_htons(ETH_P_IP):
 270                         ipv4_change_dsfield(skb->nh.iph,
 271                             p->mask[index],p->value[index]);
 272                         break;
 273                 case __constant_htons(ETH_P_IPV6):
 274                         ipv6_change_dsfield(skb->nh.ipv6h,
 275                             p->mask[index],p->value[index]);
 276                         break;
 277                 default:
 278                         /*
 279                          * Only complain if a change was actually attempted.
 280                          * This way, we can send non-IP traffic through dsmark
 281                          * and don't need yet another qdisc as a bypass.
 282                          */
 283                         if (p->mask[index] != 0xff || p->value[index])
 284                                 printk(KERN_WARNING "dsmark_dequeue: "
 285                                        "unsupported protocol %d\n",
 286                                        htons(skb->protocol));
 287                         break;
 288         };
 289         return skb;
 290 }
 291
 292
 293 static int dsmark_requeue(struct sk_buff *skb,struct Qdisc *sch)
 294 {
 295         int ret;
 296         struct dsmark_qdisc_data *p = PRIV(sch);
 297
 298         D2PRINTK("dsmark_requeue(skb %p,sch %p,[qdisc %p])\n",skb,sch,p);
 299         if ((ret = p->q->ops->requeue(skb, p->q)) == 0) {
 300                 sch->q.qlen++;
 301                 return 0;
 302         }
 303         sch->stats.drops++;
 304         return ret;
 305 }
 306
 307
 308 static unsigned int dsmark_drop(struct Qdisc *sch)
 309 {
 310         struct dsmark_qdisc_data *p = PRIV(sch);
 311         unsigned int len;
 312
 313         DPRINTK("dsmark_reset(sch %p,[qdisc %p])\n",sch,p);
 314         if (!p->q->ops->drop)
 315                 return 0;
 316         if (!(len = p->q->ops->drop(p->q)))
 317                 return 0;
 318         sch->q.qlen--;
 319         return len;
 320 }
 321
 322
 323 int dsmark_init(struct Qdisc *sch,struct rtattr *opt)
 324 {
 325         struct dsmark_qdisc_data *p = PRIV(sch);
 326         struct rtattr *tb[TCA_DSMARK_MAX];
 327         __u16 tmp;
 328
 329         DPRINTK("dsmark_init(sch %p,[qdisc %p],opt %p)\n",sch,p,opt);
 330         if (!opt ||
 331             rtattr_parse(tb,TCA_DSMARK_MAX,RTA_DATA(opt),RTA_PAYLOAD(opt)) < 0 ||
 332             !tb[TCA_DSMARK_INDICES-1] ||
 333             RTA_PAYLOAD(tb[TCA_DSMARK_INDICES-1]) < sizeof(__u16))
 334                 return -EINVAL;
 335         p->indices = *(__u16 *) RTA_DATA(tb[TCA_DSMARK_INDICES-1]);
 336         if (!p->indices)
 337                 return -EINVAL;
 338         for (tmp = p->indices; tmp != 1; tmp >>= 1) {
 339                 if (tmp & 1)
 340                         return -EINVAL;
 341         }
 342         p->default_index = NO_DEFAULT_INDEX;
 343         if (tb[TCA_DSMARK_DEFAULT_INDEX-1]) {
 344                 if (RTA_PAYLOAD(tb[TCA_DSMARK_DEFAULT_INDEX-1]) < sizeof(__u16))
 345                         return -EINVAL;
 346                 p->default_index =
 347                     *(__u16 *) RTA_DATA(tb[TCA_DSMARK_DEFAULT_INDEX-1]);
 348         }
 349         p->set_tc_index = !!tb[TCA_DSMARK_SET_TC_INDEX-1];
 350         p->mask = kmalloc(p->indices*2,GFP_KERNEL);
 351         if (!p->mask)
 352                 return -ENOMEM;
 353         p->value = p->mask+p->indices;
 354         memset(p->mask,0xff,p->indices);
 355         memset(p->value,0,p->indices);
 356         if (!(p->q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops)))
 357                 p->q = &noop_qdisc;
 358         DPRINTK("dsmark_init: qdisc %p\n",&p->q);
 359         MOD_INC_USE_COUNT;
 360         return 0;
 361 }
 362
 363
 364 static void dsmark_reset(struct Qdisc *sch)
 365 {
 366         struct dsmark_qdisc_data *p = PRIV(sch);
 367
 368         DPRINTK("dsmark_reset(sch %p,[qdisc %p])\n",sch,p);
 369         qdisc_reset(p->q);
 370         sch->q.qlen = 0;
 371 }
 372
 373
 374 static void dsmark_destroy(struct Qdisc *sch)
 375 {
 376         struct dsmark_qdisc_data *p = PRIV(sch);
 377         struct tcf_proto *tp;
 378
 379         DPRINTK("dsmark_destroy(sch %p,[qdisc %p])\n",sch,p);
 380         while (p->filter_list) {
 381                 tp = p->filter_list;
 382                 p->filter_list = tp->next;
 383                 tcf_destroy(tp);
 384         }
 385         qdisc_destroy(p->q);
 386         p->q = &noop_qdisc;
 387         kfree(p->mask);
 388         MOD_DEC_USE_COUNT;
 389 }
 390
 391
 392 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
 393     struct sk_buff *skb, struct tcmsg *tcm)
 394 {
 395         struct dsmark_qdisc_data *p = PRIV(sch);
 396         unsigned char *b = skb->tail;
 397         struct rtattr *rta;
 398
 399         DPRINTK("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n",sch,p,cl);
 400         if (!cl || cl > p->indices)
 401                 return -EINVAL;
 402         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle),cl-1);
 403         rta = (struct rtattr *) b;
 404         RTA_PUT(skb,TCA_OPTIONS,0,NULL);
 405         RTA_PUT(skb,TCA_DSMARK_MASK,1,&p->mask[cl-1]);
 406         RTA_PUT(skb,TCA_DSMARK_VALUE,1,&p->value[cl-1]);
 407         rta->rta_len = skb->tail-b;
 408         return skb->len;
 409
 410 rtattr_failure:
 411         skb_trim(skb,b-skb->data);
 412         return -1;
 413 }
 414
 415 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
 416 {
 417         struct dsmark_qdisc_data *p = PRIV(sch);
 418         unsigned char *b = skb->tail;
 419         struct rtattr *rta;
 420
 421         rta = (struct rtattr *) b;
 422         RTA_PUT(skb,TCA_OPTIONS,0,NULL);
 423         RTA_PUT(skb,TCA_DSMARK_INDICES,sizeof(__u16),&p->indices);
 424         if (p->default_index != NO_DEFAULT_INDEX) {
 425                 __u16 tmp = p->default_index;
 426
 427                 RTA_PUT(skb,TCA_DSMARK_DEFAULT_INDEX, sizeof(__u16), &tmp);
 428         }
 429         if (p->set_tc_index)
 430                 RTA_PUT(skb, TCA_DSMARK_SET_TC_INDEX, 0, NULL);
 431         rta->rta_len = skb->tail-b;
 432         return skb->len;
 433
 434 rtattr_failure:
 435         skb_trim(skb,b-skb->data);
 436         return -1;
 437 }
 438
 439 static struct Qdisc_class_ops dsmark_class_ops =
 440 {
 441         dsmark_graft,                   /* graft */
 442         dsmark_leaf,                    /* leaf */
 443         dsmark_get,                     /* get */
 444         dsmark_put,                     /* put */
 445         dsmark_change,                  /* change */
 446         dsmark_delete,                  /* delete */
 447         dsmark_walk,                    /* walk */
 448
 449         dsmark_find_tcf,                /* tcf_chain */
 450         dsmark_bind_filter,             /* bind_tcf */
 451         dsmark_put,                     /* unbind_tcf */
 452
 453         dsmark_dump_class,              /* dump */
 454 };
 455
 456 struct Qdisc_ops dsmark_qdisc_ops =
 457 {
 458         NULL,                           /* next */
 459         &dsmark_class_ops,              /* cl_ops */
 460         "dsmark",
 461         sizeof(struct dsmark_qdisc_data),
 462
 463         dsmark_enqueue,                 /* enqueue */
 464         dsmark_dequeue,                 /* dequeue */
 465         dsmark_requeue,                 /* requeue */
 466         dsmark_drop,                    /* drop */
 467
 468         dsmark_init,                    /* init */
 469         dsmark_reset,                   /* reset */
 470         dsmark_destroy,                 /* destroy */
 471         NULL,                           /* change */
 472
 473         dsmark_dump                     /* dump */
 474 };
 475
 476 #ifdef MODULE
 477 int init_module(void)
 478 {
 479         return register_qdisc(&dsmark_qdisc_ops);
 480 }
 481
 482
 483 void cleanup_module(void)
 484 {
 485         unregister_qdisc(&dsmark_qdisc_ops);
 486 }
 487 #endif
 488 MODULE_LICENSE("GPL");