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