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/module.h>
   7 #include <linux/init.h>
   8 #include <linux/slab.h>
   9 #include <linux/types.h>
  10 #include <linux/string.h>
  11 #include <linux/errno.h>
  12 #include <linux/skbuff.h>
  13 #include <linux/rtnetlink.h>
  14 #include <linux/bitops.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  * classid      class           marking
  22  * -------      -----           -------
  23  *   n/a          0             n/a
  24  *   x:0          1             use entry [0]
  25  *   ...         ...            ...
  26  *   x:y y>0     y+1            use entry [y]
  27  *   ...         ...            ...
  28  * x:indices-1  indices         use entry [indices-1]
  29  *   ...         ...            ...
  30  *   x:y         y+1            use entry [y & (indices-1)]
  31  *   ...         ...            ...
  32  * 0xffff       0x10000         use entry [indices-1]
  33  */
  34
  35
  36 #define NO_DEFAULT_INDEX        (1 << 16)
  37
  38 struct mask_value {
  39         u8                      mask;
  40         u8                      value;
  41 };
  42
  43 struct dsmark_qdisc_data {
  44         struct Qdisc            *q;
  45         struct tcf_proto __rcu  *filter_list;
  46         struct mask_value       *mv;
  47         u16                     indices;
  48         u8                      set_tc_index;
  49         u32                     default_index;  /* index range is 0...0xffff */
  50 #define DSMARK_EMBEDDED_SZ      16
  51         struct mask_value       embedded[DSMARK_EMBEDDED_SZ];
  52 };
  53
  54 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
  55 {
  56         return index <= p->indices && index > 0;
  57 }
  58
  59 /* ------------------------- Class/flow operations ------------------------- */
  60
  61 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
  62                         struct Qdisc *new, struct Qdisc **old)
  63 {
  64         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  65
  66         pr_debug("%s(sch %p,[qdisc %p],new %p,old %p)\n",
  67                  __func__, sch, p, new, old);
  68
  69         if (new == NULL) {
  70                 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
  71                                         sch->handle);
  72                 if (new == NULL)
  73                         new = &noop_qdisc;
  74         }
  75
  76         *old = qdisc_replace(sch, new, &p->q);
  77         return 0;
  78 }
  79
  80 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
  81 {
  82         struct dsmark_qdisc_data *p = qdisc_priv(sch);
  83         return p->q;
  84 }
  85
  86 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
  87 {
  88         pr_debug("%s(sch %p,[qdisc %p],classid %x)\n",
  89                  __func__, sch, qdisc_priv(sch), classid);
  90
  91         return TC_H_MIN(classid) + 1;
  92 }
  93
  94 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
  95                                         unsigned long parent, u32 classid)
  96 {
  97         return dsmark_get(sch, classid);
  98 }
  99
 100 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
 101 {
 102 }
 103
 104 static const struct nla_policy dsmark_policy[TCA_DSMARK_MAX + 1] = {
 105         [TCA_DSMARK_INDICES]            = { .type = NLA_U16 },
 106         [TCA_DSMARK_DEFAULT_INDEX]      = { .type = NLA_U16 },
 107         [TCA_DSMARK_SET_TC_INDEX]       = { .type = NLA_FLAG },
 108         [TCA_DSMARK_MASK]               = { .type = NLA_U8 },
 109         [TCA_DSMARK_VALUE]              = { .type = NLA_U8 },
 110 };
 111
 112 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
 113                          struct nlattr **tca, unsigned long *arg)
 114 {
 115         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 116         struct nlattr *opt = tca[TCA_OPTIONS];
 117         struct nlattr *tb[TCA_DSMARK_MAX + 1];
 118         int err = -EINVAL;
 119
 120         pr_debug("%s(sch %p,[qdisc %p],classid %x,parent %x), arg 0x%lx\n",
 121                  __func__, sch, p, classid, parent, *arg);
 122
 123         if (!dsmark_valid_index(p, *arg)) {
 124                 err = -ENOENT;
 125                 goto errout;
 126         }
 127
 128         if (!opt)
 129                 goto errout;
 130
 131         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
 132         if (err < 0)
 133                 goto errout;
 134
 135         if (tb[TCA_DSMARK_VALUE])
 136                 p->mv[*arg - 1].value = nla_get_u8(tb[TCA_DSMARK_VALUE]);
 137
 138         if (tb[TCA_DSMARK_MASK])
 139                 p->mv[*arg - 1].mask = nla_get_u8(tb[TCA_DSMARK_MASK]);
 140
 141         err = 0;
 142
 143 errout:
 144         return err;
 145 }
 146
 147 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
 148 {
 149         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 150
 151         if (!dsmark_valid_index(p, arg))
 152                 return -EINVAL;
 153
 154         p->mv[arg - 1].mask = 0xff;
 155         p->mv[arg - 1].value = 0;
 156
 157         return 0;
 158 }
 159
 160 static void dsmark_walk(struct Qdisc *sch, struct qdisc_walker *walker)
 161 {
 162         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 163         int i;
 164
 165         pr_debug("%s(sch %p,[qdisc %p],walker %p)\n",
 166                  __func__, sch, p, walker);
 167
 168         if (walker->stop)
 169                 return;
 170
 171         for (i = 0; i < p->indices; i++) {
 172                 if (p->mv[i].mask == 0xff && !p->mv[i].value)
 173                         goto ignore;
 174                 if (walker->count >= walker->skip) {
 175                         if (walker->fn(sch, i + 1, walker) < 0) {
 176                                 walker->stop = 1;
 177                                 break;
 178                         }
 179                 }
 180 ignore:
 181                 walker->count++;
 182         }
 183 }
 184
 185 static inline struct tcf_proto __rcu **dsmark_find_tcf(struct Qdisc *sch,
 186                                                        unsigned long cl)
 187 {
 188         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 189         return &p->filter_list;
 190 }
 191
 192 /* --------------------------- Qdisc operations ---------------------------- */
 193
 194 static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch)
 195 {
 196         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 197         int err;
 198
 199         pr_debug("%s(skb %p,sch %p,[qdisc %p])\n", __func__, skb, sch, p);
 200
 201         if (p->set_tc_index) {
 202                 switch (tc_skb_protocol(skb)) {
 203                 case htons(ETH_P_IP):
 204                         if (skb_cow_head(skb, sizeof(struct iphdr)))
 205                                 goto drop;
 206
 207                         skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
 208                                 & ~INET_ECN_MASK;
 209                         break;
 210
 211                 case htons(ETH_P_IPV6):
 212                         if (skb_cow_head(skb, sizeof(struct ipv6hdr)))
 213                                 goto drop;
 214
 215                         skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
 216                                 & ~INET_ECN_MASK;
 217                         break;
 218                 default:
 219                         skb->tc_index = 0;
 220                         break;
 221                 }
 222         }
 223
 224         if (TC_H_MAJ(skb->priority) == sch->handle)
 225                 skb->tc_index = TC_H_MIN(skb->priority);
 226         else {
 227                 struct tcf_result res;
 228                 struct tcf_proto *fl = rcu_dereference_bh(p->filter_list);
 229                 int result = tc_classify(skb, fl, &res, false);
 230
 231                 pr_debug("result %d class 0x%04x\n", result, res.classid);
 232
 233                 switch (result) {
 234 #ifdef CONFIG_NET_CLS_ACT
 235                 case TC_ACT_QUEUED:
 236                 case TC_ACT_STOLEN:
 237                         kfree_skb(skb);
 238                         return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 239
 240                 case TC_ACT_SHOT:
 241                         goto drop;
 242 #endif
 243                 case TC_ACT_OK:
 244                         skb->tc_index = TC_H_MIN(res.classid);
 245                         break;
 246
 247                 default:
 248                         if (p->default_index != NO_DEFAULT_INDEX)
 249                                 skb->tc_index = p->default_index;
 250                         break;
 251                 }
 252         }
 253
 254         err = qdisc_enqueue(skb, p->q);
 255         if (err != NET_XMIT_SUCCESS) {
 256                 if (net_xmit_drop_count(err))
 257                         qdisc_qstats_drop(sch);
 258                 return err;
 259         }
 260
 261         qdisc_qstats_backlog_inc(sch, skb);
 262         sch->q.qlen++;
 263
 264         return NET_XMIT_SUCCESS;
 265
 266 drop:
 267         qdisc_drop(skb, sch);
 268         return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 269 }
 270
 271 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 272 {
 273         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 274         struct sk_buff *skb;
 275         u32 index;
 276
 277         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 278
 279         skb = qdisc_dequeue_peeked(p->q);
 280         if (skb == NULL)
 281                 return NULL;
 282
 283         qdisc_bstats_update(sch, skb);
 284         qdisc_qstats_backlog_dec(sch, skb);
 285         sch->q.qlen--;
 286
 287         index = skb->tc_index & (p->indices - 1);
 288         pr_debug("index %d->%d\n", skb->tc_index, index);
 289
 290         switch (tc_skb_protocol(skb)) {
 291         case htons(ETH_P_IP):
 292                 ipv4_change_dsfield(ip_hdr(skb), p->mv[index].mask,
 293                                     p->mv[index].value);
 294                         break;
 295         case htons(ETH_P_IPV6):
 296                 ipv6_change_dsfield(ipv6_hdr(skb), p->mv[index].mask,
 297                                     p->mv[index].value);
 298                         break;
 299         default:
 300                 /*
 301                  * Only complain if a change was actually attempted.
 302                  * This way, we can send non-IP traffic through dsmark
 303                  * and don't need yet another qdisc as a bypass.
 304                  */
 305                 if (p->mv[index].mask != 0xff || p->mv[index].value)
 306                         pr_warn("%s: unsupported protocol %d\n",
 307                                 __func__, ntohs(tc_skb_protocol(skb)));
 308                 break;
 309         }
 310
 311         return skb;
 312 }
 313
 314 static struct sk_buff *dsmark_peek(struct Qdisc *sch)
 315 {
 316         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 317
 318         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 319
 320         return p->q->ops->peek(p->q);
 321 }
 322
 323 static unsigned int dsmark_drop(struct Qdisc *sch)
 324 {
 325         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 326         unsigned int len;
 327
 328         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 329
 330         if (p->q->ops->drop == NULL)
 331                 return 0;
 332
 333         len = p->q->ops->drop(p->q);
 334         if (len)
 335                 sch->q.qlen--;
 336
 337         return len;
 338 }
 339
 340 static int dsmark_init(struct Qdisc *sch, struct nlattr *opt)
 341 {
 342         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 343         struct nlattr *tb[TCA_DSMARK_MAX + 1];
 344         int err = -EINVAL;
 345         u32 default_index = NO_DEFAULT_INDEX;
 346         u16 indices;
 347         int i;
 348
 349         pr_debug("%s(sch %p,[qdisc %p],opt %p)\n", __func__, sch, p, opt);
 350
 351         if (!opt)
 352                 goto errout;
 353
 354         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
 355         if (err < 0)
 356                 goto errout;
 357
 358         err = -EINVAL;
 359         indices = nla_get_u16(tb[TCA_DSMARK_INDICES]);
 360
 361         if (hweight32(indices) != 1)
 362                 goto errout;
 363
 364         if (tb[TCA_DSMARK_DEFAULT_INDEX])
 365                 default_index = nla_get_u16(tb[TCA_DSMARK_DEFAULT_INDEX]);
 366
 367         if (indices <= DSMARK_EMBEDDED_SZ)
 368                 p->mv = p->embedded;
 369         else
 370                 p->mv = kmalloc_array(indices, sizeof(*p->mv), GFP_KERNEL);
 371         if (!p->mv) {
 372                 err = -ENOMEM;
 373                 goto errout;
 374         }
 375         for (i = 0; i < indices; i++) {
 376                 p->mv[i].mask = 0xff;
 377                 p->mv[i].value = 0;
 378         }
 379         p->indices = indices;
 380         p->default_index = default_index;
 381         p->set_tc_index = nla_get_flag(tb[TCA_DSMARK_SET_TC_INDEX]);
 382
 383         p->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, sch->handle);
 384         if (p->q == NULL)
 385                 p->q = &noop_qdisc;
 386
 387         pr_debug("%s: qdisc %p\n", __func__, p->q);
 388
 389         err = 0;
 390 errout:
 391         return err;
 392 }
 393
 394 static void dsmark_reset(struct Qdisc *sch)
 395 {
 396         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 397
 398         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 399         qdisc_reset(p->q);
 400         sch->qstats.backlog = 0;
 401         sch->q.qlen = 0;
 402 }
 403
 404 static void dsmark_destroy(struct Qdisc *sch)
 405 {
 406         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 407
 408         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 409
 410         tcf_destroy_chain(&p->filter_list);
 411         qdisc_destroy(p->q);
 412         if (p->mv != p->embedded)
 413                 kfree(p->mv);
 414 }
 415
 416 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
 417                              struct sk_buff *skb, struct tcmsg *tcm)
 418 {
 419         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 420         struct nlattr *opts = NULL;
 421
 422         pr_debug("%s(sch %p,[qdisc %p],class %ld\n", __func__, sch, p, cl);
 423
 424         if (!dsmark_valid_index(p, cl))
 425                 return -EINVAL;
 426
 427         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl - 1);
 428         tcm->tcm_info = p->q->handle;
 429
 430         opts = nla_nest_start(skb, TCA_OPTIONS);
 431         if (opts == NULL)
 432                 goto nla_put_failure;
 433         if (nla_put_u8(skb, TCA_DSMARK_MASK, p->mv[cl - 1].mask) ||
 434             nla_put_u8(skb, TCA_DSMARK_VALUE, p->mv[cl - 1].value))
 435                 goto nla_put_failure;
 436
 437         return nla_nest_end(skb, opts);
 438
 439 nla_put_failure:
 440         nla_nest_cancel(skb, opts);
 441         return -EMSGSIZE;
 442 }
 443
 444 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
 445 {
 446         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 447         struct nlattr *opts = NULL;
 448
 449         opts = nla_nest_start(skb, TCA_OPTIONS);
 450         if (opts == NULL)
 451                 goto nla_put_failure;
 452         if (nla_put_u16(skb, TCA_DSMARK_INDICES, p->indices))
 453                 goto nla_put_failure;
 454
 455         if (p->default_index != NO_DEFAULT_INDEX &&
 456             nla_put_u16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index))
 457                 goto nla_put_failure;
 458
 459         if (p->set_tc_index &&
 460             nla_put_flag(skb, TCA_DSMARK_SET_TC_INDEX))
 461                 goto nla_put_failure;
 462
 463         return nla_nest_end(skb, opts);
 464
 465 nla_put_failure:
 466         nla_nest_cancel(skb, opts);
 467         return -EMSGSIZE;
 468 }
 469
 470 static const struct Qdisc_class_ops dsmark_class_ops = {
 471         .graft          =       dsmark_graft,
 472         .leaf           =       dsmark_leaf,
 473         .get            =       dsmark_get,
 474         .put            =       dsmark_put,
 475         .change         =       dsmark_change,
 476         .delete         =       dsmark_delete,
 477         .walk           =       dsmark_walk,
 478         .tcf_chain      =       dsmark_find_tcf,
 479         .bind_tcf       =       dsmark_bind_filter,
 480         .unbind_tcf     =       dsmark_put,
 481         .dump           =       dsmark_dump_class,
 482 };
 483
 484 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
 485         .next           =       NULL,
 486         .cl_ops         =       &dsmark_class_ops,
 487         .id             =       "dsmark",
 488         .priv_size      =       sizeof(struct dsmark_qdisc_data),
 489         .enqueue        =       dsmark_enqueue,
 490         .dequeue        =       dsmark_dequeue,
 491         .peek           =       dsmark_peek,
 492         .drop           =       dsmark_drop,
 493         .init           =       dsmark_init,
 494         .reset          =       dsmark_reset,
 495         .destroy        =       dsmark_destroy,
 496         .change         =       NULL,
 497         .dump           =       dsmark_dump,
 498         .owner          =       THIS_MODULE,
 499 };
 500
 501 static int __init dsmark_module_init(void)
 502 {
 503         return register_qdisc(&dsmark_qdisc_ops);
 504 }
 505
 506 static void __exit dsmark_module_exit(void)
 507 {
 508         unregister_qdisc(&dsmark_qdisc_ops);
 509 }
 510
 511 module_init(dsmark_module_init)
 512 module_exit(dsmark_module_exit)
 513
 514 MODULE_LICENSE("GPL");