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 #include <linux/idr.h> 21 22 #include <net/rtnetlink.h> 23 #include <net/pkt_cls.h> 24 #include <net/sock.h> 25 26 MODULE_LICENSE("GPL"); 27 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>"); 28 MODULE_DESCRIPTION("TC BPF based classifier"); 29 30 #define CLS_BPF_NAME_LEN 256 31 #define CLS_BPF_SUPPORTED_GEN_FLAGS \ 32 (TCA_CLS_FLAGS_SKIP_HW | TCA_CLS_FLAGS_SKIP_SW) 33 34 struct cls_bpf_head { 35 struct list_head plist; 36 struct idr handle_idr; 37 struct rcu_head rcu; 38 }; 39 40 struct cls_bpf_prog { 41 struct bpf_prog *filter; 42 struct list_head link; 43 struct tcf_result res; 44 bool exts_integrated; 45 bool offloaded; 46 u32 gen_flags; 47 struct tcf_exts exts; 48 u32 handle; 49 u16 bpf_num_ops; 50 struct sock_filter *bpf_ops; 51 const char *bpf_name; 52 struct tcf_proto *tp; 53 union { 54 struct work_struct work; 55 struct rcu_head rcu; 56 }; 57 }; 58 59 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = { 60 [TCA_BPF_CLASSID] = { .type = NLA_U32 }, 61 [TCA_BPF_FLAGS] = { .type = NLA_U32 }, 62 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 }, 63 [TCA_BPF_FD] = { .type = NLA_U32 }, 64 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING, 65 .len = CLS_BPF_NAME_LEN }, 66 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 }, 67 [TCA_BPF_OPS] = { .type = NLA_BINARY, 68 .len = sizeof(struct sock_filter) * BPF_MAXINSNS }, 69 }; 70 71 static int cls_bpf_exec_opcode(int code) 72 { 73 switch (code) { 74 case TC_ACT_OK: 75 case TC_ACT_SHOT: 76 case TC_ACT_STOLEN: 77 case TC_ACT_TRAP: 78 case TC_ACT_REDIRECT: 79 case TC_ACT_UNSPEC: 80 return code; 81 default: 82 return TC_ACT_UNSPEC; 83 } 84 } 85 86 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp, 87 struct tcf_result *res) 88 { 89 struct cls_bpf_head *head = rcu_dereference_bh(tp->root); 90 bool at_ingress = skb_at_tc_ingress(skb); 91 struct cls_bpf_prog *prog; 92 int ret = -1; 93 94 /* Needed here for accessing maps. */ 95 rcu_read_lock(); 96 list_for_each_entry_rcu(prog, &head->plist, link) { 97 int filter_res; 98 99 qdisc_skb_cb(skb)->tc_classid = prog->res.classid; 100 101 if (tc_skip_sw(prog->gen_flags)) { 102 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0; 103 } else if (at_ingress) { 104 /* It is safe to push/pull even if skb_shared() */ 105 __skb_push(skb, skb->mac_len); 106 bpf_compute_data_pointers(skb); 107 filter_res = BPF_PROG_RUN(prog->filter, skb); 108 __skb_pull(skb, skb->mac_len); 109 } else { 110 bpf_compute_data_pointers(skb); 111 filter_res = BPF_PROG_RUN(prog->filter, skb); 112 } 113 114 if (prog->exts_integrated) { 115 res->class = 0; 116 res->classid = TC_H_MAJ(prog->res.classid) | 117 qdisc_skb_cb(skb)->tc_classid; 118 119 ret = cls_bpf_exec_opcode(filter_res); 120 if (ret == TC_ACT_UNSPEC) 121 continue; 122 break; 123 } 124 125 if (filter_res == 0) 126 continue; 127 if (filter_res != -1) { 128 res->class = 0; 129 res->classid = filter_res; 130 } else { 131 *res = prog->res; 132 } 133 134 ret = tcf_exts_exec(skb, &prog->exts, res); 135 if (ret < 0) 136 continue; 137 138 break; 139 } 140 rcu_read_unlock(); 141 142 return ret; 143 } 144 145 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog) 146 { 147 return !prog->bpf_ops; 148 } 149 150 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog, 151 enum tc_clsbpf_command cmd) 152 { 153 bool addorrep = cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE; 154 struct tcf_block *block = tp->chain->block; 155 bool skip_sw = tc_skip_sw(prog->gen_flags); 156 struct tc_cls_bpf_offload cls_bpf = {}; 157 int err; 158 159 tc_cls_common_offload_init(&cls_bpf.common, tp); 160 cls_bpf.command = cmd; 161 cls_bpf.exts = &prog->exts; 162 cls_bpf.prog = prog->filter; 163 cls_bpf.name = prog->bpf_name; 164 cls_bpf.exts_integrated = prog->exts_integrated; 165 cls_bpf.gen_flags = prog->gen_flags; 166 167 err = tc_setup_cb_call(block, NULL, TC_SETUP_CLSBPF, &cls_bpf, skip_sw); 168 if (addorrep) { 169 if (err < 0) { 170 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY); 171 return err; 172 } else if (err > 0) { 173 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW; 174 } 175 } 176 177 if (addorrep && skip_sw && !(prog->gen_flags & TCA_CLS_FLAGS_IN_HW)) 178 return -EINVAL; 179 180 return 0; 181 } 182 183 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog, 184 struct cls_bpf_prog *oldprog) 185 { 186 struct cls_bpf_prog *obj = prog; 187 enum tc_clsbpf_command cmd; 188 bool skip_sw; 189 int ret; 190 191 skip_sw = tc_skip_sw(prog->gen_flags) || 192 (oldprog && tc_skip_sw(oldprog->gen_flags)); 193 194 if (oldprog && oldprog->offloaded) { 195 if (!tc_skip_hw(prog->gen_flags)) { 196 cmd = TC_CLSBPF_REPLACE; 197 } else if (!tc_skip_sw(prog->gen_flags)) { 198 obj = oldprog; 199 cmd = TC_CLSBPF_DESTROY; 200 } else { 201 return -EINVAL; 202 } 203 } else { 204 if (tc_skip_hw(prog->gen_flags)) 205 return skip_sw ? -EINVAL : 0; 206 cmd = TC_CLSBPF_ADD; 207 } 208 209 ret = cls_bpf_offload_cmd(tp, obj, cmd); 210 if (ret) 211 return ret; 212 213 obj->offloaded = true; 214 if (oldprog) 215 oldprog->offloaded = false; 216 217 return 0; 218 } 219 220 static void cls_bpf_stop_offload(struct tcf_proto *tp, 221 struct cls_bpf_prog *prog) 222 { 223 int err; 224 225 if (!prog->offloaded) 226 return; 227 228 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY); 229 if (err) { 230 pr_err("Stopping hardware offload failed: %d\n", err); 231 return; 232 } 233 234 prog->offloaded = false; 235 } 236 237 static void cls_bpf_offload_update_stats(struct tcf_proto *tp, 238 struct cls_bpf_prog *prog) 239 { 240 if (!prog->offloaded) 241 return; 242 243 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS); 244 } 245 246 static int cls_bpf_init(struct tcf_proto *tp) 247 { 248 struct cls_bpf_head *head; 249 250 head = kzalloc(sizeof(*head), GFP_KERNEL); 251 if (head == NULL) 252 return -ENOBUFS; 253 254 INIT_LIST_HEAD_RCU(&head->plist); 255 idr_init(&head->handle_idr); 256 rcu_assign_pointer(tp->root, head); 257 258 return 0; 259 } 260 261 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog) 262 { 263 tcf_exts_destroy(&prog->exts); 264 tcf_exts_put_net(&prog->exts); 265 266 if (cls_bpf_is_ebpf(prog)) 267 bpf_prog_put(prog->filter); 268 else 269 bpf_prog_destroy(prog->filter); 270 271 kfree(prog->bpf_name); 272 kfree(prog->bpf_ops); 273 kfree(prog); 274 } 275 276 static void cls_bpf_delete_prog_work(struct work_struct *work) 277 { 278 struct cls_bpf_prog *prog = container_of(work, struct cls_bpf_prog, work); 279 280 rtnl_lock(); 281 __cls_bpf_delete_prog(prog); 282 rtnl_unlock(); 283 } 284 285 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu) 286 { 287 struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu); 288 289 INIT_WORK(&prog->work, cls_bpf_delete_prog_work); 290 tcf_queue_work(&prog->work); 291 } 292 293 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog) 294 { 295 struct cls_bpf_head *head = rtnl_dereference(tp->root); 296 297 idr_remove_ext(&head->handle_idr, prog->handle); 298 cls_bpf_stop_offload(tp, prog); 299 list_del_rcu(&prog->link); 300 tcf_unbind_filter(tp, &prog->res); 301 if (tcf_exts_get_net(&prog->exts)) 302 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu); 303 else 304 __cls_bpf_delete_prog(prog); 305 } 306 307 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last) 308 { 309 struct cls_bpf_head *head = rtnl_dereference(tp->root); 310 311 __cls_bpf_delete(tp, arg); 312 *last = list_empty(&head->plist); 313 return 0; 314 } 315 316 static void cls_bpf_destroy(struct tcf_proto *tp) 317 { 318 struct cls_bpf_head *head = rtnl_dereference(tp->root); 319 struct cls_bpf_prog *prog, *tmp; 320 321 list_for_each_entry_safe(prog, tmp, &head->plist, link) 322 __cls_bpf_delete(tp, prog); 323 324 idr_destroy(&head->handle_idr); 325 kfree_rcu(head, rcu); 326 } 327 328 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle) 329 { 330 struct cls_bpf_head *head = rtnl_dereference(tp->root); 331 struct cls_bpf_prog *prog; 332 333 list_for_each_entry(prog, &head->plist, link) { 334 if (prog->handle == handle) 335 return prog; 336 } 337 338 return NULL; 339 } 340 341 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog) 342 { 343 struct sock_filter *bpf_ops; 344 struct sock_fprog_kern fprog_tmp; 345 struct bpf_prog *fp; 346 u16 bpf_size, bpf_num_ops; 347 int ret; 348 349 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]); 350 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0) 351 return -EINVAL; 352 353 bpf_size = bpf_num_ops * sizeof(*bpf_ops); 354 if (bpf_size != nla_len(tb[TCA_BPF_OPS])) 355 return -EINVAL; 356 357 bpf_ops = kzalloc(bpf_size, GFP_KERNEL); 358 if (bpf_ops == NULL) 359 return -ENOMEM; 360 361 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size); 362 363 fprog_tmp.len = bpf_num_ops; 364 fprog_tmp.filter = bpf_ops; 365 366 ret = bpf_prog_create(&fp, &fprog_tmp); 367 if (ret < 0) { 368 kfree(bpf_ops); 369 return ret; 370 } 371 372 prog->bpf_ops = bpf_ops; 373 prog->bpf_num_ops = bpf_num_ops; 374 prog->bpf_name = NULL; 375 prog->filter = fp; 376 377 return 0; 378 } 379 380 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog, 381 u32 gen_flags, const struct tcf_proto *tp) 382 { 383 struct bpf_prog *fp; 384 char *name = NULL; 385 bool skip_sw; 386 u32 bpf_fd; 387 388 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]); 389 skip_sw = gen_flags & TCA_CLS_FLAGS_SKIP_SW; 390 391 fp = bpf_prog_get_type_dev(bpf_fd, BPF_PROG_TYPE_SCHED_CLS, skip_sw); 392 if (IS_ERR(fp)) 393 return PTR_ERR(fp); 394 395 if (tb[TCA_BPF_NAME]) { 396 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL); 397 if (!name) { 398 bpf_prog_put(fp); 399 return -ENOMEM; 400 } 401 } 402 403 prog->bpf_ops = NULL; 404 prog->bpf_name = name; 405 prog->filter = fp; 406 407 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS)) 408 netif_keep_dst(qdisc_dev(tp->q)); 409 410 return 0; 411 } 412 413 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp, 414 struct cls_bpf_prog *prog, unsigned long base, 415 struct nlattr **tb, struct nlattr *est, bool ovr) 416 { 417 bool is_bpf, is_ebpf, have_exts = false; 418 u32 gen_flags = 0; 419 int ret; 420 421 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS]; 422 is_ebpf = tb[TCA_BPF_FD]; 423 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) 424 return -EINVAL; 425 426 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr); 427 if (ret < 0) 428 return ret; 429 430 if (tb[TCA_BPF_FLAGS]) { 431 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]); 432 433 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) 434 return -EINVAL; 435 436 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT; 437 } 438 if (tb[TCA_BPF_FLAGS_GEN]) { 439 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]); 440 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS || 441 !tc_flags_valid(gen_flags)) 442 return -EINVAL; 443 } 444 445 prog->exts_integrated = have_exts; 446 prog->gen_flags = gen_flags; 447 448 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) : 449 cls_bpf_prog_from_efd(tb, prog, gen_flags, tp); 450 if (ret < 0) 451 return ret; 452 453 if (tb[TCA_BPF_CLASSID]) { 454 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]); 455 tcf_bind_filter(tp, &prog->res, base); 456 } 457 458 return 0; 459 } 460 461 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb, 462 struct tcf_proto *tp, unsigned long base, 463 u32 handle, struct nlattr **tca, 464 void **arg, bool ovr) 465 { 466 struct cls_bpf_head *head = rtnl_dereference(tp->root); 467 struct cls_bpf_prog *oldprog = *arg; 468 struct nlattr *tb[TCA_BPF_MAX + 1]; 469 struct cls_bpf_prog *prog; 470 unsigned long idr_index; 471 int ret; 472 473 if (tca[TCA_OPTIONS] == NULL) 474 return -EINVAL; 475 476 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy, 477 NULL); 478 if (ret < 0) 479 return ret; 480 481 prog = kzalloc(sizeof(*prog), GFP_KERNEL); 482 if (!prog) 483 return -ENOBUFS; 484 485 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE); 486 if (ret < 0) 487 goto errout; 488 489 if (oldprog) { 490 if (handle && oldprog->handle != handle) { 491 ret = -EINVAL; 492 goto errout; 493 } 494 } 495 496 if (handle == 0) { 497 ret = idr_alloc_ext(&head->handle_idr, prog, &idr_index, 498 1, 0x7FFFFFFF, GFP_KERNEL); 499 if (ret) 500 goto errout; 501 prog->handle = idr_index; 502 } else { 503 if (!oldprog) { 504 ret = idr_alloc_ext(&head->handle_idr, prog, &idr_index, 505 handle, handle + 1, GFP_KERNEL); 506 if (ret) 507 goto errout; 508 } 509 prog->handle = handle; 510 } 511 512 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr); 513 if (ret < 0) 514 goto errout_idr; 515 516 ret = cls_bpf_offload(tp, prog, oldprog); 517 if (ret) { 518 if (!oldprog) 519 idr_remove_ext(&head->handle_idr, prog->handle); 520 __cls_bpf_delete_prog(prog); 521 return ret; 522 } 523 524 if (!tc_in_hw(prog->gen_flags)) 525 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW; 526 527 if (oldprog) { 528 idr_replace_ext(&head->handle_idr, prog, handle); 529 list_replace_rcu(&oldprog->link, &prog->link); 530 tcf_unbind_filter(tp, &oldprog->res); 531 tcf_exts_get_net(&oldprog->exts); 532 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu); 533 } else { 534 list_add_rcu(&prog->link, &head->plist); 535 } 536 537 *arg = prog; 538 return 0; 539 540 errout_idr: 541 if (!oldprog) 542 idr_remove_ext(&head->handle_idr, prog->handle); 543 errout: 544 tcf_exts_destroy(&prog->exts); 545 kfree(prog); 546 return ret; 547 } 548 549 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog, 550 struct sk_buff *skb) 551 { 552 struct nlattr *nla; 553 554 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops)) 555 return -EMSGSIZE; 556 557 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops * 558 sizeof(struct sock_filter)); 559 if (nla == NULL) 560 return -EMSGSIZE; 561 562 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla)); 563 564 return 0; 565 } 566 567 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog, 568 struct sk_buff *skb) 569 { 570 struct nlattr *nla; 571 572 if (prog->bpf_name && 573 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name)) 574 return -EMSGSIZE; 575 576 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id)) 577 return -EMSGSIZE; 578 579 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag)); 580 if (nla == NULL) 581 return -EMSGSIZE; 582 583 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla)); 584 585 return 0; 586 } 587 588 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, void *fh, 589 struct sk_buff *skb, struct tcmsg *tm) 590 { 591 struct cls_bpf_prog *prog = fh; 592 struct nlattr *nest; 593 u32 bpf_flags = 0; 594 int ret; 595 596 if (prog == NULL) 597 return skb->len; 598 599 tm->tcm_handle = prog->handle; 600 601 cls_bpf_offload_update_stats(tp, prog); 602 603 nest = nla_nest_start(skb, TCA_OPTIONS); 604 if (nest == NULL) 605 goto nla_put_failure; 606 607 if (prog->res.classid && 608 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid)) 609 goto nla_put_failure; 610 611 if (cls_bpf_is_ebpf(prog)) 612 ret = cls_bpf_dump_ebpf_info(prog, skb); 613 else 614 ret = cls_bpf_dump_bpf_info(prog, skb); 615 if (ret) 616 goto nla_put_failure; 617 618 if (tcf_exts_dump(skb, &prog->exts) < 0) 619 goto nla_put_failure; 620 621 if (prog->exts_integrated) 622 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT; 623 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags)) 624 goto nla_put_failure; 625 if (prog->gen_flags && 626 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags)) 627 goto nla_put_failure; 628 629 nla_nest_end(skb, nest); 630 631 if (tcf_exts_dump_stats(skb, &prog->exts) < 0) 632 goto nla_put_failure; 633 634 return skb->len; 635 636 nla_put_failure: 637 nla_nest_cancel(skb, nest); 638 return -1; 639 } 640 641 static void cls_bpf_bind_class(void *fh, u32 classid, unsigned long cl) 642 { 643 struct cls_bpf_prog *prog = fh; 644 645 if (prog && prog->res.classid == classid) 646 prog->res.class = cl; 647 } 648 649 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg) 650 { 651 struct cls_bpf_head *head = rtnl_dereference(tp->root); 652 struct cls_bpf_prog *prog; 653 654 list_for_each_entry(prog, &head->plist, link) { 655 if (arg->count < arg->skip) 656 goto skip; 657 if (arg->fn(tp, prog, arg) < 0) { 658 arg->stop = 1; 659 break; 660 } 661 skip: 662 arg->count++; 663 } 664 } 665 666 static struct tcf_proto_ops cls_bpf_ops __read_mostly = { 667 .kind = "bpf", 668 .owner = THIS_MODULE, 669 .classify = cls_bpf_classify, 670 .init = cls_bpf_init, 671 .destroy = cls_bpf_destroy, 672 .get = cls_bpf_get, 673 .change = cls_bpf_change, 674 .delete = cls_bpf_delete, 675 .walk = cls_bpf_walk, 676 .dump = cls_bpf_dump, 677 .bind_class = cls_bpf_bind_class, 678 }; 679 680 static int __init cls_bpf_init_mod(void) 681 { 682 return register_tcf_proto_ops(&cls_bpf_ops); 683 } 684 685 static void __exit cls_bpf_exit_mod(void) 686 { 687 unregister_tcf_proto_ops(&cls_bpf_ops); 688 } 689 690 module_init(cls_bpf_init_mod); 691 module_exit(cls_bpf_exit_mod); 692