net/sched/cls_bpf.c

   1 /*
   2  * Berkeley Packet Filter based traffic classifier
   3  *
   4  * Might be used to classify traffic through flexible, user-defined and
   5  * possibly JIT-ed BPF filters for traffic control as an alternative to
   6  * ematches.
   7  *
   8  * (C) 2013 Daniel Borkmann <dborkman@redhat.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2 as
  12  * published by the Free Software Foundation.
  13  */
  14
  15 #include <linux/module.h>
  16 #include <linux/types.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/filter.h>
  19 #include <linux/bpf.h>
  20
  21 #include <net/rtnetlink.h>
  22 #include <net/pkt_cls.h>
  23 #include <net/sock.h>
  24
  25 MODULE_LICENSE("GPL");
  26 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
  27 MODULE_DESCRIPTION("TC BPF based classifier");
  28
  29 #define CLS_BPF_NAME_LEN        256
  30
  31 struct cls_bpf_head {
  32         struct list_head plist;
  33         u32 hgen;
  34         struct rcu_head rcu;
  35 };
  36
  37 struct cls_bpf_prog {
  38         struct bpf_prog *filter;
  39         struct list_head link;
  40         struct tcf_result res;
  41         bool exts_integrated;
  42         struct tcf_exts exts;
  43         u32 handle;
  44         union {
  45                 u32 bpf_fd;
  46                 u16 bpf_num_ops;
  47         };
  48         struct sock_filter *bpf_ops;
  49         const char *bpf_name;
  50         struct tcf_proto *tp;
  51         struct rcu_head rcu;
  52 };
  53
  54 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
  55         [TCA_BPF_CLASSID]       = { .type = NLA_U32 },
  56         [TCA_BPF_FLAGS]         = { .type = NLA_U32 },
  57         [TCA_BPF_FD]            = { .type = NLA_U32 },
  58         [TCA_BPF_NAME]          = { .type = NLA_NUL_STRING,
  59                                     .len = CLS_BPF_NAME_LEN },
  60         [TCA_BPF_OPS_LEN]       = { .type = NLA_U16 },
  61         [TCA_BPF_OPS]           = { .type = NLA_BINARY,
  62                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
  63 };
  64
  65 static int cls_bpf_exec_opcode(int code)
  66 {
  67         switch (code) {
  68         case TC_ACT_OK:
  69         case TC_ACT_SHOT:
  70         case TC_ACT_STOLEN:
  71         case TC_ACT_REDIRECT:
  72         case TC_ACT_UNSPEC:
  73                 return code;
  74         default:
  75                 return TC_ACT_UNSPEC;
  76         }
  77 }
  78
  79 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  80                             struct tcf_result *res)
  81 {
  82         struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
  83         bool at_ingress = skb_at_tc_ingress(skb);
  84         struct cls_bpf_prog *prog;
  85         int ret = -1;
  86
  87         /* Needed here for accessing maps. */
  88         rcu_read_lock();
  89         list_for_each_entry_rcu(prog, &head->plist, link) {
  90                 int filter_res;
  91
  92                 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
  93
  94                 if (at_ingress) {
  95                         /* It is safe to push/pull even if skb_shared() */
  96                         __skb_push(skb, skb->mac_len);
  97                         bpf_compute_data_end(skb);
  98                         filter_res = BPF_PROG_RUN(prog->filter, skb);
  99                         __skb_pull(skb, skb->mac_len);
 100                 } else {
 101                         bpf_compute_data_end(skb);
 102                         filter_res = BPF_PROG_RUN(prog->filter, skb);
 103                 }
 104
 105                 if (prog->exts_integrated) {
 106                         res->class   = 0;
 107                         res->classid = TC_H_MAJ(prog->res.classid) |
 108                                        qdisc_skb_cb(skb)->tc_classid;
 109
 110                         ret = cls_bpf_exec_opcode(filter_res);
 111                         if (ret == TC_ACT_UNSPEC)
 112                                 continue;
 113                         break;
 114                 }
 115
 116                 if (filter_res == 0)
 117                         continue;
 118                 if (filter_res != -1) {
 119                         res->class   = 0;
 120                         res->classid = filter_res;
 121                 } else {
 122                         *res = prog->res;
 123                 }
 124
 125                 ret = tcf_exts_exec(skb, &prog->exts, res);
 126                 if (ret < 0)
 127                         continue;
 128
 129                 break;
 130         }
 131         rcu_read_unlock();
 132
 133         return ret;
 134 }
 135
 136 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
 137 {
 138         return !prog->bpf_ops;
 139 }
 140
 141 static int cls_bpf_init(struct tcf_proto *tp)
 142 {
 143         struct cls_bpf_head *head;
 144
 145         head = kzalloc(sizeof(*head), GFP_KERNEL);
 146         if (head == NULL)
 147                 return -ENOBUFS;
 148
 149         INIT_LIST_HEAD_RCU(&head->plist);
 150         rcu_assign_pointer(tp->root, head);
 151
 152         return 0;
 153 }
 154
 155 static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
 156 {
 157         tcf_exts_destroy(&prog->exts);
 158
 159         if (cls_bpf_is_ebpf(prog))
 160                 bpf_prog_put(prog->filter);
 161         else
 162                 bpf_prog_destroy(prog->filter);
 163
 164         kfree(prog->bpf_name);
 165         kfree(prog->bpf_ops);
 166         kfree(prog);
 167 }
 168
 169 static void __cls_bpf_delete_prog(struct rcu_head *rcu)
 170 {
 171         struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
 172
 173         cls_bpf_delete_prog(prog->tp, prog);
 174 }
 175
 176 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
 177 {
 178         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
 179
 180         list_del_rcu(&prog->link);
 181         tcf_unbind_filter(tp, &prog->res);
 182         call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 183
 184         return 0;
 185 }
 186
 187 static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
 188 {
 189         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 190         struct cls_bpf_prog *prog, *tmp;
 191
 192         if (!force && !list_empty(&head->plist))
 193                 return false;
 194
 195         list_for_each_entry_safe(prog, tmp, &head->plist, link) {
 196                 list_del_rcu(&prog->link);
 197                 tcf_unbind_filter(tp, &prog->res);
 198                 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 199         }
 200
 201         RCU_INIT_POINTER(tp->root, NULL);
 202         kfree_rcu(head, rcu);
 203         return true;
 204 }
 205
 206 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
 207 {
 208         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 209         struct cls_bpf_prog *prog;
 210         unsigned long ret = 0UL;
 211
 212         if (head == NULL)
 213                 return 0UL;
 214
 215         list_for_each_entry(prog, &head->plist, link) {
 216                 if (prog->handle == handle) {
 217                         ret = (unsigned long) prog;
 218                         break;
 219                 }
 220         }
 221
 222         return ret;
 223 }
 224
 225 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
 226 {
 227         struct sock_filter *bpf_ops;
 228         struct sock_fprog_kern fprog_tmp;
 229         struct bpf_prog *fp;
 230         u16 bpf_size, bpf_num_ops;
 231         int ret;
 232
 233         bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
 234         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 235                 return -EINVAL;
 236
 237         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 238         if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
 239                 return -EINVAL;
 240
 241         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 242         if (bpf_ops == NULL)
 243                 return -ENOMEM;
 244
 245         memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
 246
 247         fprog_tmp.len = bpf_num_ops;
 248         fprog_tmp.filter = bpf_ops;
 249
 250         ret = bpf_prog_create(&fp, &fprog_tmp);
 251         if (ret < 0) {
 252                 kfree(bpf_ops);
 253                 return ret;
 254         }
 255
 256         prog->bpf_ops = bpf_ops;
 257         prog->bpf_num_ops = bpf_num_ops;
 258         prog->bpf_name = NULL;
 259         prog->filter = fp;
 260
 261         return 0;
 262 }
 263
 264 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
 265                                  const struct tcf_proto *tp)
 266 {
 267         struct bpf_prog *fp;
 268         char *name = NULL;
 269         u32 bpf_fd;
 270
 271         bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
 272
 273         fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
 274         if (IS_ERR(fp))
 275                 return PTR_ERR(fp);
 276
 277         if (tb[TCA_BPF_NAME]) {
 278                 name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
 279                                nla_len(tb[TCA_BPF_NAME]),
 280                                GFP_KERNEL);
 281                 if (!name) {
 282                         bpf_prog_put(fp);
 283                         return -ENOMEM;
 284                 }
 285         }
 286
 287         prog->bpf_ops = NULL;
 288         prog->bpf_fd = bpf_fd;
 289         prog->bpf_name = name;
 290         prog->filter = fp;
 291
 292         if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
 293                 netif_keep_dst(qdisc_dev(tp->q));
 294
 295         return 0;
 296 }
 297
 298 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
 299                                    struct cls_bpf_prog *prog,
 300                                    unsigned long base, struct nlattr **tb,
 301                                    struct nlattr *est, bool ovr)
 302 {
 303         bool is_bpf, is_ebpf, have_exts = false;
 304         struct tcf_exts exts;
 305         int ret;
 306
 307         is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
 308         is_ebpf = tb[TCA_BPF_FD];
 309         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
 310                 return -EINVAL;
 311
 312         ret = tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 313         if (ret < 0)
 314                 return ret;
 315         ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
 316         if (ret < 0)
 317                 goto errout;
 318
 319         if (tb[TCA_BPF_FLAGS]) {
 320                 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
 321
 322                 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
 323                         ret = -EINVAL;
 324                         goto errout;
 325                 }
 326
 327                 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
 328         }
 329
 330         prog->exts_integrated = have_exts;
 331
 332         ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
 333                        cls_bpf_prog_from_efd(tb, prog, tp);
 334         if (ret < 0)
 335                 goto errout;
 336
 337         if (tb[TCA_BPF_CLASSID]) {
 338                 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
 339                 tcf_bind_filter(tp, &prog->res, base);
 340         }
 341
 342         tcf_exts_change(tp, &prog->exts, &exts);
 343         return 0;
 344
 345 errout:
 346         tcf_exts_destroy(&exts);
 347         return ret;
 348 }
 349
 350 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
 351                                    struct cls_bpf_head *head)
 352 {
 353         unsigned int i = 0x80000000;
 354         u32 handle;
 355
 356         do {
 357                 if (++head->hgen == 0x7FFFFFFF)
 358                         head->hgen = 1;
 359         } while (--i > 0 && cls_bpf_get(tp, head->hgen));
 360
 361         if (unlikely(i == 0)) {
 362                 pr_err("Insufficient number of handles\n");
 363                 handle = 0;
 364         } else {
 365                 handle = head->hgen;
 366         }
 367
 368         return handle;
 369 }
 370
 371 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
 372                           struct tcf_proto *tp, unsigned long base,
 373                           u32 handle, struct nlattr **tca,
 374                           unsigned long *arg, bool ovr)
 375 {
 376         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 377         struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
 378         struct nlattr *tb[TCA_BPF_MAX + 1];
 379         struct cls_bpf_prog *prog;
 380         int ret;
 381
 382         if (tca[TCA_OPTIONS] == NULL)
 383                 return -EINVAL;
 384
 385         ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
 386         if (ret < 0)
 387                 return ret;
 388
 389         prog = kzalloc(sizeof(*prog), GFP_KERNEL);
 390         if (!prog)
 391                 return -ENOBUFS;
 392
 393         ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 394         if (ret < 0)
 395                 goto errout;
 396
 397         if (oldprog) {
 398                 if (handle && oldprog->handle != handle) {
 399                         ret = -EINVAL;
 400                         goto errout;
 401                 }
 402         }
 403
 404         if (handle == 0)
 405                 prog->handle = cls_bpf_grab_new_handle(tp, head);
 406         else
 407                 prog->handle = handle;
 408         if (prog->handle == 0) {
 409                 ret = -EINVAL;
 410                 goto errout;
 411         }
 412
 413         ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE],
 414                                       ovr);
 415         if (ret < 0)
 416                 goto errout;
 417
 418         if (oldprog) {
 419                 list_replace_rcu(&oldprog->link, &prog->link);
 420                 tcf_unbind_filter(tp, &oldprog->res);
 421                 call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
 422         } else {
 423                 list_add_rcu(&prog->link, &head->plist);
 424         }
 425
 426         *arg = (unsigned long) prog;
 427         return 0;
 428
 429 errout:
 430         tcf_exts_destroy(&prog->exts);
 431         kfree(prog);
 432         return ret;
 433 }
 434
 435 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
 436                                  struct sk_buff *skb)
 437 {
 438         struct nlattr *nla;
 439
 440         if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
 441                 return -EMSGSIZE;
 442
 443         nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
 444                           sizeof(struct sock_filter));
 445         if (nla == NULL)
 446                 return -EMSGSIZE;
 447
 448         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 449
 450         return 0;
 451 }
 452
 453 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
 454                                   struct sk_buff *skb)
 455 {
 456         if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
 457                 return -EMSGSIZE;
 458
 459         if (prog->bpf_name &&
 460             nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
 461                 return -EMSGSIZE;
 462
 463         return 0;
 464 }
 465
 466 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 467                         struct sk_buff *skb, struct tcmsg *tm)
 468 {
 469         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
 470         struct nlattr *nest;
 471         u32 bpf_flags = 0;
 472         int ret;
 473
 474         if (prog == NULL)
 475                 return skb->len;
 476
 477         tm->tcm_handle = prog->handle;
 478
 479         nest = nla_nest_start(skb, TCA_OPTIONS);
 480         if (nest == NULL)
 481                 goto nla_put_failure;
 482
 483         if (prog->res.classid &&
 484             nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
 485                 goto nla_put_failure;
 486
 487         if (cls_bpf_is_ebpf(prog))
 488                 ret = cls_bpf_dump_ebpf_info(prog, skb);
 489         else
 490                 ret = cls_bpf_dump_bpf_info(prog, skb);
 491         if (ret)
 492                 goto nla_put_failure;
 493
 494         if (tcf_exts_dump(skb, &prog->exts) < 0)
 495                 goto nla_put_failure;
 496
 497         if (prog->exts_integrated)
 498                 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
 499         if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
 500                 goto nla_put_failure;
 501
 502         nla_nest_end(skb, nest);
 503
 504         if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
 505                 goto nla_put_failure;
 506
 507         return skb->len;
 508
 509 nla_put_failure:
 510         nla_nest_cancel(skb, nest);
 511         return -1;
 512 }
 513
 514 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 515 {
 516         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 517         struct cls_bpf_prog *prog;
 518
 519         list_for_each_entry(prog, &head->plist, link) {
 520                 if (arg->count < arg->skip)
 521                         goto skip;
 522                 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
 523                         arg->stop = 1;
 524                         break;
 525                 }
 526 skip:
 527                 arg->count++;
 528         }
 529 }
 530
 531 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
 532         .kind           =       "bpf",
 533         .owner          =       THIS_MODULE,
 534         .classify       =       cls_bpf_classify,
 535         .init           =       cls_bpf_init,
 536         .destroy        =       cls_bpf_destroy,
 537         .get            =       cls_bpf_get,
 538         .change         =       cls_bpf_change,
 539         .delete         =       cls_bpf_delete,
 540         .walk           =       cls_bpf_walk,
 541         .dump           =       cls_bpf_dump,
 542 };
 543
 544 static int __init cls_bpf_init_mod(void)
 545 {
 546         return register_tcf_proto_ops(&cls_bpf_ops);
 547 }
 548
 549 static void __exit cls_bpf_exit_mod(void)
 550 {
 551         unregister_tcf_proto_ops(&cls_bpf_ops);
 552 }
 553
 554 module_init(cls_bpf_init_mod);
 555 module_exit(cls_bpf_exit_mod);