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