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