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