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                 int wlen = skb_network_offset(skb);
 203
 204                 switch (tc_skb_protocol(skb)) {
 205                 case htons(ETH_P_IP):
 206                         wlen += sizeof(struct iphdr);
 207                         if (!pskb_may_pull(skb, wlen) ||
 208                             skb_try_make_writable(skb, wlen))
 209                                 goto drop;
 210
 211                         skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
 212                                 & ~INET_ECN_MASK;
 213                         break;
 214
 215                 case htons(ETH_P_IPV6):
 216                         wlen += sizeof(struct ipv6hdr);
 217                         if (!pskb_may_pull(skb, wlen) ||
 218                             skb_try_make_writable(skb, wlen))
 219                                 goto drop;
 220
 221                         skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
 222                                 & ~INET_ECN_MASK;
 223                         break;
 224                 default:
 225                         skb->tc_index = 0;
 226                         break;
 227                 }
 228         }
 229
 230         if (TC_H_MAJ(skb->priority) == sch->handle)
 231                 skb->tc_index = TC_H_MIN(skb->priority);
 232         else {
 233                 struct tcf_result res;
 234                 struct tcf_proto *fl = rcu_dereference_bh(p->filter_list);
 235                 int result = tc_classify(skb, fl, &res, false);
 236
 237                 pr_debug("result %d class 0x%04x\n", result, res.classid);
 238
 239                 switch (result) {
 240 #ifdef CONFIG_NET_CLS_ACT
 241                 case TC_ACT_QUEUED:
 242                 case TC_ACT_STOLEN:
 243                         kfree_skb(skb);
 244                         return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
 245
 246                 case TC_ACT_SHOT:
 247                         goto drop;
 248 #endif
 249                 case TC_ACT_OK:
 250                         skb->tc_index = TC_H_MIN(res.classid);
 251                         break;
 252
 253                 default:
 254                         if (p->default_index != NO_DEFAULT_INDEX)
 255                                 skb->tc_index = p->default_index;
 256                         break;
 257                 }
 258         }
 259
 260         err = qdisc_enqueue(skb, p->q);
 261         if (err != NET_XMIT_SUCCESS) {
 262                 if (net_xmit_drop_count(err))
 263                         qdisc_qstats_drop(sch);
 264                 return err;
 265         }
 266
 267         qdisc_qstats_backlog_inc(sch, skb);
 268         sch->q.qlen++;
 269
 270         return NET_XMIT_SUCCESS;
 271
 272 drop:
 273         qdisc_drop(skb, sch);
 274         return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
 275 }
 276
 277 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 278 {
 279         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 280         struct sk_buff *skb;
 281         u32 index;
 282
 283         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 284
 285         skb = p->q->ops->dequeue(p->q);
 286         if (skb == NULL)
 287                 return NULL;
 288
 289         qdisc_bstats_update(sch, skb);
 290         qdisc_qstats_backlog_dec(sch, skb);
 291         sch->q.qlen--;
 292
 293         index = skb->tc_index & (p->indices - 1);
 294         pr_debug("index %d->%d\n", skb->tc_index, index);
 295
 296         switch (tc_skb_protocol(skb)) {
 297         case htons(ETH_P_IP):
 298                 ipv4_change_dsfield(ip_hdr(skb), p->mv[index].mask,
 299                                     p->mv[index].value);
 300                         break;
 301         case htons(ETH_P_IPV6):
 302                 ipv6_change_dsfield(ipv6_hdr(skb), p->mv[index].mask,
 303                                     p->mv[index].value);
 304                         break;
 305         default:
 306                 /*
 307                  * Only complain if a change was actually attempted.
 308                  * This way, we can send non-IP traffic through dsmark
 309                  * and don't need yet another qdisc as a bypass.
 310                  */
 311                 if (p->mv[index].mask != 0xff || p->mv[index].value)
 312                         pr_warn("%s: unsupported protocol %d\n",
 313                                 __func__, ntohs(tc_skb_protocol(skb)));
 314                 break;
 315         }
 316
 317         return skb;
 318 }
 319
 320 static struct sk_buff *dsmark_peek(struct Qdisc *sch)
 321 {
 322         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 323
 324         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 325
 326         return p->q->ops->peek(p->q);
 327 }
 328
 329 static unsigned int dsmark_drop(struct Qdisc *sch)
 330 {
 331         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 332         unsigned int len;
 333
 334         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 335
 336         if (p->q->ops->drop == NULL)
 337                 return 0;
 338
 339         len = p->q->ops->drop(p->q);
 340         if (len)
 341                 sch->q.qlen--;
 342
 343         return len;
 344 }
 345
 346 static int dsmark_init(struct Qdisc *sch, struct nlattr *opt)
 347 {
 348         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 349         struct nlattr *tb[TCA_DSMARK_MAX + 1];
 350         int err = -EINVAL;
 351         u32 default_index = NO_DEFAULT_INDEX;
 352         u16 indices;
 353         int i;
 354
 355         pr_debug("%s(sch %p,[qdisc %p],opt %p)\n", __func__, sch, p, opt);
 356
 357         if (!opt)
 358                 goto errout;
 359
 360         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
 361         if (err < 0)
 362                 goto errout;
 363
 364         err = -EINVAL;
 365         indices = nla_get_u16(tb[TCA_DSMARK_INDICES]);
 366
 367         if (hweight32(indices) != 1)
 368                 goto errout;
 369
 370         if (tb[TCA_DSMARK_DEFAULT_INDEX])
 371                 default_index = nla_get_u16(tb[TCA_DSMARK_DEFAULT_INDEX]);
 372
 373         if (indices <= DSMARK_EMBEDDED_SZ)
 374                 p->mv = p->embedded;
 375         else
 376                 p->mv = kmalloc_array(indices, sizeof(*p->mv), GFP_KERNEL);
 377         if (!p->mv) {
 378                 err = -ENOMEM;
 379                 goto errout;
 380         }
 381         for (i = 0; i < indices; i++) {
 382                 p->mv[i].mask = 0xff;
 383                 p->mv[i].value = 0;
 384         }
 385         p->indices = indices;
 386         p->default_index = default_index;
 387         p->set_tc_index = nla_get_flag(tb[TCA_DSMARK_SET_TC_INDEX]);
 388
 389         p->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, sch->handle);
 390         if (p->q == NULL)
 391                 p->q = &noop_qdisc;
 392
 393         pr_debug("%s: qdisc %p\n", __func__, p->q);
 394
 395         err = 0;
 396 errout:
 397         return err;
 398 }
 399
 400 static void dsmark_reset(struct Qdisc *sch)
 401 {
 402         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 403
 404         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 405         qdisc_reset(p->q);
 406         sch->qstats.backlog = 0;
 407         sch->q.qlen = 0;
 408 }
 409
 410 static void dsmark_destroy(struct Qdisc *sch)
 411 {
 412         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 413
 414         pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
 415
 416         tcf_destroy_chain(&p->filter_list);
 417         qdisc_destroy(p->q);
 418         if (p->mv != p->embedded)
 419                 kfree(p->mv);
 420 }
 421
 422 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
 423                              struct sk_buff *skb, struct tcmsg *tcm)
 424 {
 425         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 426         struct nlattr *opts = NULL;
 427
 428         pr_debug("%s(sch %p,[qdisc %p],class %ld\n", __func__, sch, p, cl);
 429
 430         if (!dsmark_valid_index(p, cl))
 431                 return -EINVAL;
 432
 433         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl - 1);
 434         tcm->tcm_info = p->q->handle;
 435
 436         opts = nla_nest_start(skb, TCA_OPTIONS);
 437         if (opts == NULL)
 438                 goto nla_put_failure;
 439         if (nla_put_u8(skb, TCA_DSMARK_MASK, p->mv[cl - 1].mask) ||
 440             nla_put_u8(skb, TCA_DSMARK_VALUE, p->mv[cl - 1].value))
 441                 goto nla_put_failure;
 442
 443         return nla_nest_end(skb, opts);
 444
 445 nla_put_failure:
 446         nla_nest_cancel(skb, opts);
 447         return -EMSGSIZE;
 448 }
 449
 450 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
 451 {
 452         struct dsmark_qdisc_data *p = qdisc_priv(sch);
 453         struct nlattr *opts = NULL;
 454
 455         opts = nla_nest_start(skb, TCA_OPTIONS);
 456         if (opts == NULL)
 457                 goto nla_put_failure;
 458         if (nla_put_u16(skb, TCA_DSMARK_INDICES, p->indices))
 459                 goto nla_put_failure;
 460
 461         if (p->default_index != NO_DEFAULT_INDEX &&
 462             nla_put_u16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index))
 463                 goto nla_put_failure;
 464
 465         if (p->set_tc_index &&
 466             nla_put_flag(skb, TCA_DSMARK_SET_TC_INDEX))
 467                 goto nla_put_failure;
 468
 469         return nla_nest_end(skb, opts);
 470
 471 nla_put_failure:
 472         nla_nest_cancel(skb, opts);
 473         return -EMSGSIZE;
 474 }
 475
 476 static const struct Qdisc_class_ops dsmark_class_ops = {
 477         .graft          =       dsmark_graft,
 478         .leaf           =       dsmark_leaf,
 479         .get            =       dsmark_get,
 480         .put            =       dsmark_put,
 481         .change         =       dsmark_change,
 482         .delete         =       dsmark_delete,
 483         .walk           =       dsmark_walk,
 484         .tcf_chain      =       dsmark_find_tcf,
 485         .bind_tcf       =       dsmark_bind_filter,
 486         .unbind_tcf     =       dsmark_put,
 487         .dump           =       dsmark_dump_class,
 488 };
 489
 490 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
 491         .next           =       NULL,
 492         .cl_ops         =       &dsmark_class_ops,
 493         .id             =       "dsmark",
 494         .priv_size      =       sizeof(struct dsmark_qdisc_data),
 495         .enqueue        =       dsmark_enqueue,
 496         .dequeue        =       dsmark_dequeue,
 497         .peek           =       dsmark_peek,
 498         .drop           =       dsmark_drop,
 499         .init           =       dsmark_init,
 500         .reset          =       dsmark_reset,
 501         .destroy        =       dsmark_destroy,
 502         .change         =       NULL,
 503         .dump           =       dsmark_dump,
 504         .owner          =       THIS_MODULE,
 505 };
 506
 507 static int __init dsmark_module_init(void)
 508 {
 509         return register_qdisc(&dsmark_qdisc_ops);
 510 }
 511
 512 static void __exit dsmark_module_exit(void)
 513 {
 514         unregister_qdisc(&dsmark_qdisc_ops);
 515 }
 516
 517 module_init(dsmark_module_init)
 518 module_exit(dsmark_module_exit)
 519
 520 MODULE_LICENSE("GPL");