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 #define CLS_BPF_SUPPORTED_GEN_FLAGS \ 31 (TCA_CLS_FLAGS_SKIP_HW | TCA_CLS_FLAGS_SKIP_SW) 32 33 struct cls_bpf_head { 34 struct list_head plist; 35 u32 hgen; 36 struct rcu_head rcu; 37 }; 38 39 struct cls_bpf_prog { 40 struct bpf_prog *filter; 41 struct list_head link; 42 struct tcf_result res; 43 bool exts_integrated; 44 bool offloaded; 45 u32 gen_flags; 46 struct tcf_exts exts; 47 u32 handle; 48 u16 bpf_num_ops; 49 struct sock_filter *bpf_ops; 50 const char *bpf_name; 51 struct tcf_proto *tp; 52 struct rcu_head rcu; 53 }; 54 55 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = { 56 [TCA_BPF_CLASSID] = { .type = NLA_U32 }, 57 [TCA_BPF_FLAGS] = { .type = NLA_U32 }, 58 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 }, 59 [TCA_BPF_FD] = { .type = NLA_U32 }, 60 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING, 61 .len = CLS_BPF_NAME_LEN }, 62 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 }, 63 [TCA_BPF_OPS] = { .type = NLA_BINARY, 64 .len = sizeof(struct sock_filter) * BPF_MAXINSNS }, 65 }; 66 67 static int cls_bpf_exec_opcode(int code) 68 { 69 switch (code) { 70 case TC_ACT_OK: 71 case TC_ACT_SHOT: 72 case TC_ACT_STOLEN: 73 case TC_ACT_TRAP: 74 case TC_ACT_REDIRECT: 75 case TC_ACT_UNSPEC: 76 return code; 77 default: 78 return TC_ACT_UNSPEC; 79 } 80 } 81 82 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp, 83 struct tcf_result *res) 84 { 85 struct cls_bpf_head *head = rcu_dereference_bh(tp->root); 86 bool at_ingress = skb_at_tc_ingress(skb); 87 struct cls_bpf_prog *prog; 88 int ret = -1; 89 90 /* Needed here for accessing maps. */ 91 rcu_read_lock(); 92 list_for_each_entry_rcu(prog, &head->plist, link) { 93 int filter_res; 94 95 qdisc_skb_cb(skb)->tc_classid = prog->res.classid; 96 97 if (tc_skip_sw(prog->gen_flags)) { 98 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0; 99 } else if (at_ingress) { 100 /* It is safe to push/pull even if skb_shared() */ 101 __skb_push(skb, skb->mac_len); 102 bpf_compute_data_end(skb); 103 filter_res = BPF_PROG_RUN(prog->filter, skb); 104 __skb_pull(skb, skb->mac_len); 105 } else { 106 bpf_compute_data_end(skb); 107 filter_res = BPF_PROG_RUN(prog->filter, skb); 108 } 109 110 if (prog->exts_integrated) { 111 res->class = 0; 112 res->classid = TC_H_MAJ(prog->res.classid) | 113 qdisc_skb_cb(skb)->tc_classid; 114 115 ret = cls_bpf_exec_opcode(filter_res); 116 if (ret == TC_ACT_UNSPEC) 117 continue; 118 break; 119 } 120 121 if (filter_res == 0) 122 continue; 123 if (filter_res != -1) { 124 res->class = 0; 125 res->classid = filter_res; 126 } else { 127 *res = prog->res; 128 } 129 130 ret = tcf_exts_exec(skb, &prog->exts, res); 131 if (ret < 0) 132 continue; 133 134 break; 135 } 136 rcu_read_unlock(); 137 138 return ret; 139 } 140 141 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog) 142 { 143 return !prog->bpf_ops; 144 } 145 146 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog, 147 enum tc_clsbpf_command cmd) 148 { 149 struct net_device *dev = tp->q->dev_queue->dev; 150 struct tc_cls_bpf_offload bpf_offload = {}; 151 struct tc_to_netdev offload; 152 int err; 153 154 offload.type = TC_SETUP_CLSBPF; 155 offload.cls_bpf = &bpf_offload; 156 157 bpf_offload.command = cmd; 158 bpf_offload.exts = &prog->exts; 159 bpf_offload.prog = prog->filter; 160 bpf_offload.name = prog->bpf_name; 161 bpf_offload.exts_integrated = prog->exts_integrated; 162 bpf_offload.gen_flags = prog->gen_flags; 163 164 err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle, 165 tp->chain->index, 166 tp->protocol, &offload); 167 168 if (!err && (cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE)) 169 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW; 170 171 return err; 172 } 173 174 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog, 175 struct cls_bpf_prog *oldprog) 176 { 177 struct net_device *dev = tp->q->dev_queue->dev; 178 struct cls_bpf_prog *obj = prog; 179 enum tc_clsbpf_command cmd; 180 bool skip_sw; 181 int ret; 182 183 skip_sw = tc_skip_sw(prog->gen_flags) || 184 (oldprog && tc_skip_sw(oldprog->gen_flags)); 185 186 if (oldprog && oldprog->offloaded) { 187 if (tc_should_offload(dev, tp, prog->gen_flags)) { 188 cmd = TC_CLSBPF_REPLACE; 189 } else if (!tc_skip_sw(prog->gen_flags)) { 190 obj = oldprog; 191 cmd = TC_CLSBPF_DESTROY; 192 } else { 193 return -EINVAL; 194 } 195 } else { 196 if (!tc_should_offload(dev, tp, prog->gen_flags)) 197 return skip_sw ? -EINVAL : 0; 198 cmd = TC_CLSBPF_ADD; 199 } 200 201 ret = cls_bpf_offload_cmd(tp, obj, cmd); 202 if (ret) 203 return skip_sw ? ret : 0; 204 205 obj->offloaded = true; 206 if (oldprog) 207 oldprog->offloaded = false; 208 209 return 0; 210 } 211 212 static void cls_bpf_stop_offload(struct tcf_proto *tp, 213 struct cls_bpf_prog *prog) 214 { 215 int err; 216 217 if (!prog->offloaded) 218 return; 219 220 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY); 221 if (err) { 222 pr_err("Stopping hardware offload failed: %d\n", err); 223 return; 224 } 225 226 prog->offloaded = false; 227 } 228 229 static void cls_bpf_offload_update_stats(struct tcf_proto *tp, 230 struct cls_bpf_prog *prog) 231 { 232 if (!prog->offloaded) 233 return; 234 235 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS); 236 } 237 238 static int cls_bpf_init(struct tcf_proto *tp) 239 { 240 struct cls_bpf_head *head; 241 242 head = kzalloc(sizeof(*head), GFP_KERNEL); 243 if (head == NULL) 244 return -ENOBUFS; 245 246 INIT_LIST_HEAD_RCU(&head->plist); 247 rcu_assign_pointer(tp->root, head); 248 249 return 0; 250 } 251 252 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog) 253 { 254 tcf_exts_destroy(&prog->exts); 255 256 if (cls_bpf_is_ebpf(prog)) 257 bpf_prog_put(prog->filter); 258 else 259 bpf_prog_destroy(prog->filter); 260 261 kfree(prog->bpf_name); 262 kfree(prog->bpf_ops); 263 kfree(prog); 264 } 265 266 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu) 267 { 268 __cls_bpf_delete_prog(container_of(rcu, struct cls_bpf_prog, rcu)); 269 } 270 271 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog) 272 { 273 cls_bpf_stop_offload(tp, prog); 274 list_del_rcu(&prog->link); 275 tcf_unbind_filter(tp, &prog->res); 276 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu); 277 } 278 279 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg, bool *last) 280 { 281 struct cls_bpf_head *head = rtnl_dereference(tp->root); 282 283 __cls_bpf_delete(tp, (struct cls_bpf_prog *) arg); 284 *last = list_empty(&head->plist); 285 return 0; 286 } 287 288 static void cls_bpf_destroy(struct tcf_proto *tp) 289 { 290 struct cls_bpf_head *head = rtnl_dereference(tp->root); 291 struct cls_bpf_prog *prog, *tmp; 292 293 list_for_each_entry_safe(prog, tmp, &head->plist, link) 294 __cls_bpf_delete(tp, prog); 295 296 kfree_rcu(head, rcu); 297 } 298 299 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle) 300 { 301 struct cls_bpf_head *head = rtnl_dereference(tp->root); 302 struct cls_bpf_prog *prog; 303 unsigned long ret = 0UL; 304 305 list_for_each_entry(prog, &head->plist, link) { 306 if (prog->handle == handle) { 307 ret = (unsigned long) prog; 308 break; 309 } 310 } 311 312 return ret; 313 } 314 315 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog) 316 { 317 struct sock_filter *bpf_ops; 318 struct sock_fprog_kern fprog_tmp; 319 struct bpf_prog *fp; 320 u16 bpf_size, bpf_num_ops; 321 int ret; 322 323 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]); 324 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0) 325 return -EINVAL; 326 327 bpf_size = bpf_num_ops * sizeof(*bpf_ops); 328 if (bpf_size != nla_len(tb[TCA_BPF_OPS])) 329 return -EINVAL; 330 331 bpf_ops = kzalloc(bpf_size, GFP_KERNEL); 332 if (bpf_ops == NULL) 333 return -ENOMEM; 334 335 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size); 336 337 fprog_tmp.len = bpf_num_ops; 338 fprog_tmp.filter = bpf_ops; 339 340 ret = bpf_prog_create(&fp, &fprog_tmp); 341 if (ret < 0) { 342 kfree(bpf_ops); 343 return ret; 344 } 345 346 prog->bpf_ops = bpf_ops; 347 prog->bpf_num_ops = bpf_num_ops; 348 prog->bpf_name = NULL; 349 prog->filter = fp; 350 351 return 0; 352 } 353 354 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog, 355 const struct tcf_proto *tp) 356 { 357 struct bpf_prog *fp; 358 char *name = NULL; 359 u32 bpf_fd; 360 361 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]); 362 363 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS); 364 if (IS_ERR(fp)) 365 return PTR_ERR(fp); 366 367 if (tb[TCA_BPF_NAME]) { 368 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL); 369 if (!name) { 370 bpf_prog_put(fp); 371 return -ENOMEM; 372 } 373 } 374 375 prog->bpf_ops = NULL; 376 prog->bpf_name = name; 377 prog->filter = fp; 378 379 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS)) 380 netif_keep_dst(qdisc_dev(tp->q)); 381 382 return 0; 383 } 384 385 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp, 386 struct cls_bpf_prog *prog, 387 unsigned long base, struct nlattr **tb, 388 struct nlattr *est, bool ovr) 389 { 390 bool is_bpf, is_ebpf, have_exts = false; 391 struct tcf_exts exts; 392 u32 gen_flags = 0; 393 int ret; 394 395 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS]; 396 is_ebpf = tb[TCA_BPF_FD]; 397 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) 398 return -EINVAL; 399 400 ret = tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE); 401 if (ret < 0) 402 return ret; 403 ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr); 404 if (ret < 0) 405 goto errout; 406 407 if (tb[TCA_BPF_FLAGS]) { 408 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]); 409 410 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) { 411 ret = -EINVAL; 412 goto errout; 413 } 414 415 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT; 416 } 417 if (tb[TCA_BPF_FLAGS_GEN]) { 418 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]); 419 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS || 420 !tc_flags_valid(gen_flags)) { 421 ret = -EINVAL; 422 goto errout; 423 } 424 } 425 426 prog->exts_integrated = have_exts; 427 prog->gen_flags = gen_flags; 428 429 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) : 430 cls_bpf_prog_from_efd(tb, prog, tp); 431 if (ret < 0) 432 goto errout; 433 434 if (tb[TCA_BPF_CLASSID]) { 435 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]); 436 tcf_bind_filter(tp, &prog->res, base); 437 } 438 439 tcf_exts_change(tp, &prog->exts, &exts); 440 return 0; 441 442 errout: 443 tcf_exts_destroy(&exts); 444 return ret; 445 } 446 447 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp, 448 struct cls_bpf_head *head) 449 { 450 unsigned int i = 0x80000000; 451 u32 handle; 452 453 do { 454 if (++head->hgen == 0x7FFFFFFF) 455 head->hgen = 1; 456 } while (--i > 0 && cls_bpf_get(tp, head->hgen)); 457 458 if (unlikely(i == 0)) { 459 pr_err("Insufficient number of handles\n"); 460 handle = 0; 461 } else { 462 handle = head->hgen; 463 } 464 465 return handle; 466 } 467 468 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb, 469 struct tcf_proto *tp, unsigned long base, 470 u32 handle, struct nlattr **tca, 471 unsigned long *arg, bool ovr) 472 { 473 struct cls_bpf_head *head = rtnl_dereference(tp->root); 474 struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg; 475 struct nlattr *tb[TCA_BPF_MAX + 1]; 476 struct cls_bpf_prog *prog; 477 int ret; 478 479 if (tca[TCA_OPTIONS] == NULL) 480 return -EINVAL; 481 482 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy, 483 NULL); 484 if (ret < 0) 485 return ret; 486 487 prog = kzalloc(sizeof(*prog), GFP_KERNEL); 488 if (!prog) 489 return -ENOBUFS; 490 491 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE); 492 if (ret < 0) 493 goto errout; 494 495 if (oldprog) { 496 if (handle && oldprog->handle != handle) { 497 ret = -EINVAL; 498 goto errout; 499 } 500 } 501 502 if (handle == 0) 503 prog->handle = cls_bpf_grab_new_handle(tp, head); 504 else 505 prog->handle = handle; 506 if (prog->handle == 0) { 507 ret = -EINVAL; 508 goto errout; 509 } 510 511 ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE], 512 ovr); 513 if (ret < 0) 514 goto errout; 515 516 ret = cls_bpf_offload(tp, prog, oldprog); 517 if (ret) { 518 __cls_bpf_delete_prog(prog); 519 return ret; 520 } 521 522 if (!tc_in_hw(prog->gen_flags)) 523 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW; 524 525 if (oldprog) { 526 list_replace_rcu(&oldprog->link, &prog->link); 527 tcf_unbind_filter(tp, &oldprog->res); 528 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu); 529 } else { 530 list_add_rcu(&prog->link, &head->plist); 531 } 532 533 *arg = (unsigned long) prog; 534 return 0; 535 536 errout: 537 tcf_exts_destroy(&prog->exts); 538 kfree(prog); 539 return ret; 540 } 541 542 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog, 543 struct sk_buff *skb) 544 { 545 struct nlattr *nla; 546 547 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops)) 548 return -EMSGSIZE; 549 550 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops * 551 sizeof(struct sock_filter)); 552 if (nla == NULL) 553 return -EMSGSIZE; 554 555 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla)); 556 557 return 0; 558 } 559 560 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog, 561 struct sk_buff *skb) 562 { 563 struct nlattr *nla; 564 565 if (prog->bpf_name && 566 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name)) 567 return -EMSGSIZE; 568 569 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id)) 570 return -EMSGSIZE; 571 572 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag)); 573 if (nla == NULL) 574 return -EMSGSIZE; 575 576 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla)); 577 578 return 0; 579 } 580 581 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh, 582 struct sk_buff *skb, struct tcmsg *tm) 583 { 584 struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh; 585 struct nlattr *nest; 586 u32 bpf_flags = 0; 587 int ret; 588 589 if (prog == NULL) 590 return skb->len; 591 592 tm->tcm_handle = prog->handle; 593 594 cls_bpf_offload_update_stats(tp, prog); 595 596 nest = nla_nest_start(skb, TCA_OPTIONS); 597 if (nest == NULL) 598 goto nla_put_failure; 599 600 if (prog->res.classid && 601 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid)) 602 goto nla_put_failure; 603 604 if (cls_bpf_is_ebpf(prog)) 605 ret = cls_bpf_dump_ebpf_info(prog, skb); 606 else 607 ret = cls_bpf_dump_bpf_info(prog, skb); 608 if (ret) 609 goto nla_put_failure; 610 611 if (tcf_exts_dump(skb, &prog->exts) < 0) 612 goto nla_put_failure; 613 614 if (prog->exts_integrated) 615 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT; 616 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags)) 617 goto nla_put_failure; 618 if (prog->gen_flags && 619 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags)) 620 goto nla_put_failure; 621 622 nla_nest_end(skb, nest); 623 624 if (tcf_exts_dump_stats(skb, &prog->exts) < 0) 625 goto nla_put_failure; 626 627 return skb->len; 628 629 nla_put_failure: 630 nla_nest_cancel(skb, nest); 631 return -1; 632 } 633 634 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg) 635 { 636 struct cls_bpf_head *head = rtnl_dereference(tp->root); 637 struct cls_bpf_prog *prog; 638 639 list_for_each_entry(prog, &head->plist, link) { 640 if (arg->count < arg->skip) 641 goto skip; 642 if (arg->fn(tp, (unsigned long) prog, arg) < 0) { 643 arg->stop = 1; 644 break; 645 } 646 skip: 647 arg->count++; 648 } 649 } 650 651 static struct tcf_proto_ops cls_bpf_ops __read_mostly = { 652 .kind = "bpf", 653 .owner = THIS_MODULE, 654 .classify = cls_bpf_classify, 655 .init = cls_bpf_init, 656 .destroy = cls_bpf_destroy, 657 .get = cls_bpf_get, 658 .change = cls_bpf_change, 659 .delete = cls_bpf_delete, 660 .walk = cls_bpf_walk, 661 .dump = cls_bpf_dump, 662 }; 663 664 static int __init cls_bpf_init_mod(void) 665 { 666 return register_tcf_proto_ops(&cls_bpf_ops); 667 } 668 669 static void __exit cls_bpf_exit_mod(void) 670 { 671 unregister_tcf_proto_ops(&cls_bpf_ops); 672 } 673 674 module_init(cls_bpf_init_mod); 675 module_exit(cls_bpf_exit_mod); 676