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 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_work rwork; 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_pointers(skb); 103 filter_res = BPF_PROG_RUN(prog->filter, skb); 104 __skb_pull(skb, skb->mac_len); 105 } else { 106 bpf_compute_data_pointers(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 struct cls_bpf_prog *oldprog, 148 struct netlink_ext_ack *extack) 149 { 150 struct tcf_block *block = tp->chain->block; 151 struct tc_cls_bpf_offload cls_bpf = {}; 152 struct cls_bpf_prog *obj; 153 bool skip_sw; 154 int err; 155 156 skip_sw = prog && tc_skip_sw(prog->gen_flags); 157 obj = prog ?: oldprog; 158 159 tc_cls_common_offload_init(&cls_bpf.common, tp, obj->gen_flags, 160 extack); 161 cls_bpf.command = TC_CLSBPF_OFFLOAD; 162 cls_bpf.exts = &obj->exts; 163 cls_bpf.prog = prog ? prog->filter : NULL; 164 cls_bpf.oldprog = oldprog ? oldprog->filter : NULL; 165 cls_bpf.name = obj->bpf_name; 166 cls_bpf.exts_integrated = obj->exts_integrated; 167 168 if (oldprog) 169 tcf_block_offload_dec(block, &oldprog->gen_flags); 170 171 err = tc_setup_cb_call(block, NULL, TC_SETUP_CLSBPF, &cls_bpf, skip_sw); 172 if (prog) { 173 if (err < 0) { 174 cls_bpf_offload_cmd(tp, oldprog, prog, extack); 175 return err; 176 } else if (err > 0) { 177 tcf_block_offload_inc(block, &prog->gen_flags); 178 } 179 } 180 181 if (prog && skip_sw && !(prog->gen_flags & TCA_CLS_FLAGS_IN_HW)) 182 return -EINVAL; 183 184 return 0; 185 } 186 187 static u32 cls_bpf_flags(u32 flags) 188 { 189 return flags & CLS_BPF_SUPPORTED_GEN_FLAGS; 190 } 191 192 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog, 193 struct cls_bpf_prog *oldprog, 194 struct netlink_ext_ack *extack) 195 { 196 if (prog && oldprog && 197 cls_bpf_flags(prog->gen_flags) != 198 cls_bpf_flags(oldprog->gen_flags)) 199 return -EINVAL; 200 201 if (prog && tc_skip_hw(prog->gen_flags)) 202 prog = NULL; 203 if (oldprog && tc_skip_hw(oldprog->gen_flags)) 204 oldprog = NULL; 205 if (!prog && !oldprog) 206 return 0; 207 208 return cls_bpf_offload_cmd(tp, prog, oldprog, extack); 209 } 210 211 static void cls_bpf_stop_offload(struct tcf_proto *tp, 212 struct cls_bpf_prog *prog, 213 struct netlink_ext_ack *extack) 214 { 215 int err; 216 217 err = cls_bpf_offload_cmd(tp, NULL, prog, extack); 218 if (err) 219 pr_err("Stopping hardware offload failed: %d\n", err); 220 } 221 222 static void cls_bpf_offload_update_stats(struct tcf_proto *tp, 223 struct cls_bpf_prog *prog) 224 { 225 struct tcf_block *block = tp->chain->block; 226 struct tc_cls_bpf_offload cls_bpf = {}; 227 228 tc_cls_common_offload_init(&cls_bpf.common, tp, prog->gen_flags, NULL); 229 cls_bpf.command = TC_CLSBPF_STATS; 230 cls_bpf.exts = &prog->exts; 231 cls_bpf.prog = prog->filter; 232 cls_bpf.name = prog->bpf_name; 233 cls_bpf.exts_integrated = prog->exts_integrated; 234 235 tc_setup_cb_call(block, NULL, TC_SETUP_CLSBPF, &cls_bpf, false); 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 idr_init(&head->handle_idr); 248 rcu_assign_pointer(tp->root, head); 249 250 return 0; 251 } 252 253 static void cls_bpf_free_parms(struct cls_bpf_prog *prog) 254 { 255 if (cls_bpf_is_ebpf(prog)) 256 bpf_prog_put(prog->filter); 257 else 258 bpf_prog_destroy(prog->filter); 259 260 kfree(prog->bpf_name); 261 kfree(prog->bpf_ops); 262 } 263 264 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog) 265 { 266 tcf_exts_destroy(&prog->exts); 267 tcf_exts_put_net(&prog->exts); 268 269 cls_bpf_free_parms(prog); 270 kfree(prog); 271 } 272 273 static void cls_bpf_delete_prog_work(struct work_struct *work) 274 { 275 struct cls_bpf_prog *prog = container_of(to_rcu_work(work), 276 struct cls_bpf_prog, 277 rwork); 278 rtnl_lock(); 279 __cls_bpf_delete_prog(prog); 280 rtnl_unlock(); 281 } 282 283 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog, 284 struct netlink_ext_ack *extack) 285 { 286 struct cls_bpf_head *head = rtnl_dereference(tp->root); 287 288 idr_remove(&head->handle_idr, prog->handle); 289 cls_bpf_stop_offload(tp, prog, extack); 290 list_del_rcu(&prog->link); 291 tcf_unbind_filter(tp, &prog->res); 292 if (tcf_exts_get_net(&prog->exts)) 293 tcf_queue_work(&prog->rwork, cls_bpf_delete_prog_work); 294 else 295 __cls_bpf_delete_prog(prog); 296 } 297 298 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last, 299 struct netlink_ext_ack *extack) 300 { 301 struct cls_bpf_head *head = rtnl_dereference(tp->root); 302 303 __cls_bpf_delete(tp, arg, extack); 304 *last = list_empty(&head->plist); 305 return 0; 306 } 307 308 static void cls_bpf_destroy(struct tcf_proto *tp, 309 struct netlink_ext_ack *extack) 310 { 311 struct cls_bpf_head *head = rtnl_dereference(tp->root); 312 struct cls_bpf_prog *prog, *tmp; 313 314 list_for_each_entry_safe(prog, tmp, &head->plist, link) 315 __cls_bpf_delete(tp, prog, extack); 316 317 idr_destroy(&head->handle_idr); 318 kfree_rcu(head, rcu); 319 } 320 321 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle) 322 { 323 struct cls_bpf_head *head = rtnl_dereference(tp->root); 324 struct cls_bpf_prog *prog; 325 326 list_for_each_entry(prog, &head->plist, link) { 327 if (prog->handle == handle) 328 return prog; 329 } 330 331 return NULL; 332 } 333 334 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog) 335 { 336 struct sock_filter *bpf_ops; 337 struct sock_fprog_kern fprog_tmp; 338 struct bpf_prog *fp; 339 u16 bpf_size, bpf_num_ops; 340 int ret; 341 342 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]); 343 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0) 344 return -EINVAL; 345 346 bpf_size = bpf_num_ops * sizeof(*bpf_ops); 347 if (bpf_size != nla_len(tb[TCA_BPF_OPS])) 348 return -EINVAL; 349 350 bpf_ops = kzalloc(bpf_size, GFP_KERNEL); 351 if (bpf_ops == NULL) 352 return -ENOMEM; 353 354 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size); 355 356 fprog_tmp.len = bpf_num_ops; 357 fprog_tmp.filter = bpf_ops; 358 359 ret = bpf_prog_create(&fp, &fprog_tmp); 360 if (ret < 0) { 361 kfree(bpf_ops); 362 return ret; 363 } 364 365 prog->bpf_ops = bpf_ops; 366 prog->bpf_num_ops = bpf_num_ops; 367 prog->bpf_name = NULL; 368 prog->filter = fp; 369 370 return 0; 371 } 372 373 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog, 374 u32 gen_flags, const struct tcf_proto *tp) 375 { 376 struct bpf_prog *fp; 377 char *name = NULL; 378 bool skip_sw; 379 u32 bpf_fd; 380 381 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]); 382 skip_sw = gen_flags & TCA_CLS_FLAGS_SKIP_SW; 383 384 fp = bpf_prog_get_type_dev(bpf_fd, BPF_PROG_TYPE_SCHED_CLS, skip_sw); 385 if (IS_ERR(fp)) 386 return PTR_ERR(fp); 387 388 if (tb[TCA_BPF_NAME]) { 389 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL); 390 if (!name) { 391 bpf_prog_put(fp); 392 return -ENOMEM; 393 } 394 } 395 396 prog->bpf_ops = NULL; 397 prog->bpf_name = name; 398 prog->filter = fp; 399 400 if (fp->dst_needed) 401 tcf_block_netif_keep_dst(tp->chain->block); 402 403 return 0; 404 } 405 406 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp, 407 struct cls_bpf_prog *prog, unsigned long base, 408 struct nlattr **tb, struct nlattr *est, bool ovr, 409 struct netlink_ext_ack *extack) 410 { 411 bool is_bpf, is_ebpf, have_exts = false; 412 u32 gen_flags = 0; 413 int ret; 414 415 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS]; 416 is_ebpf = tb[TCA_BPF_FD]; 417 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) 418 return -EINVAL; 419 420 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr, extack); 421 if (ret < 0) 422 return ret; 423 424 if (tb[TCA_BPF_FLAGS]) { 425 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]); 426 427 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) 428 return -EINVAL; 429 430 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT; 431 } 432 if (tb[TCA_BPF_FLAGS_GEN]) { 433 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]); 434 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS || 435 !tc_flags_valid(gen_flags)) 436 return -EINVAL; 437 } 438 439 prog->exts_integrated = have_exts; 440 prog->gen_flags = gen_flags; 441 442 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) : 443 cls_bpf_prog_from_efd(tb, prog, gen_flags, tp); 444 if (ret < 0) 445 return ret; 446 447 if (tb[TCA_BPF_CLASSID]) { 448 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]); 449 tcf_bind_filter(tp, &prog->res, base); 450 } 451 452 return 0; 453 } 454 455 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb, 456 struct tcf_proto *tp, unsigned long base, 457 u32 handle, struct nlattr **tca, 458 void **arg, bool ovr, struct netlink_ext_ack *extack) 459 { 460 struct cls_bpf_head *head = rtnl_dereference(tp->root); 461 struct cls_bpf_prog *oldprog = *arg; 462 struct nlattr *tb[TCA_BPF_MAX + 1]; 463 struct cls_bpf_prog *prog; 464 int ret; 465 466 if (tca[TCA_OPTIONS] == NULL) 467 return -EINVAL; 468 469 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy, 470 NULL); 471 if (ret < 0) 472 return ret; 473 474 prog = kzalloc(sizeof(*prog), GFP_KERNEL); 475 if (!prog) 476 return -ENOBUFS; 477 478 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE); 479 if (ret < 0) 480 goto errout; 481 482 if (oldprog) { 483 if (handle && oldprog->handle != handle) { 484 ret = -EINVAL; 485 goto errout; 486 } 487 } 488 489 if (handle == 0) { 490 handle = 1; 491 ret = idr_alloc_u32(&head->handle_idr, prog, &handle, 492 INT_MAX, GFP_KERNEL); 493 } else if (!oldprog) { 494 ret = idr_alloc_u32(&head->handle_idr, prog, &handle, 495 handle, GFP_KERNEL); 496 } 497 498 if (ret) 499 goto errout; 500 prog->handle = handle; 501 502 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr, 503 extack); 504 if (ret < 0) 505 goto errout_idr; 506 507 ret = cls_bpf_offload(tp, prog, oldprog, extack); 508 if (ret) 509 goto errout_parms; 510 511 if (!tc_in_hw(prog->gen_flags)) 512 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW; 513 514 if (oldprog) { 515 idr_replace(&head->handle_idr, prog, handle); 516 list_replace_rcu(&oldprog->link, &prog->link); 517 tcf_unbind_filter(tp, &oldprog->res); 518 tcf_exts_get_net(&oldprog->exts); 519 tcf_queue_work(&oldprog->rwork, cls_bpf_delete_prog_work); 520 } else { 521 list_add_rcu(&prog->link, &head->plist); 522 } 523 524 *arg = prog; 525 return 0; 526 527 errout_parms: 528 cls_bpf_free_parms(prog); 529 errout_idr: 530 if (!oldprog) 531 idr_remove(&head->handle_idr, prog->handle); 532 errout: 533 tcf_exts_destroy(&prog->exts); 534 kfree(prog); 535 return ret; 536 } 537 538 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog, 539 struct sk_buff *skb) 540 { 541 struct nlattr *nla; 542 543 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops)) 544 return -EMSGSIZE; 545 546 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops * 547 sizeof(struct sock_filter)); 548 if (nla == NULL) 549 return -EMSGSIZE; 550 551 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla)); 552 553 return 0; 554 } 555 556 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog, 557 struct sk_buff *skb) 558 { 559 struct nlattr *nla; 560 561 if (prog->bpf_name && 562 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name)) 563 return -EMSGSIZE; 564 565 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id)) 566 return -EMSGSIZE; 567 568 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag)); 569 if (nla == NULL) 570 return -EMSGSIZE; 571 572 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla)); 573 574 return 0; 575 } 576 577 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, void *fh, 578 struct sk_buff *skb, struct tcmsg *tm) 579 { 580 struct cls_bpf_prog *prog = fh; 581 struct nlattr *nest; 582 u32 bpf_flags = 0; 583 int ret; 584 585 if (prog == NULL) 586 return skb->len; 587 588 tm->tcm_handle = prog->handle; 589 590 cls_bpf_offload_update_stats(tp, prog); 591 592 nest = nla_nest_start(skb, TCA_OPTIONS); 593 if (nest == NULL) 594 goto nla_put_failure; 595 596 if (prog->res.classid && 597 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid)) 598 goto nla_put_failure; 599 600 if (cls_bpf_is_ebpf(prog)) 601 ret = cls_bpf_dump_ebpf_info(prog, skb); 602 else 603 ret = cls_bpf_dump_bpf_info(prog, skb); 604 if (ret) 605 goto nla_put_failure; 606 607 if (tcf_exts_dump(skb, &prog->exts) < 0) 608 goto nla_put_failure; 609 610 if (prog->exts_integrated) 611 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT; 612 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags)) 613 goto nla_put_failure; 614 if (prog->gen_flags && 615 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags)) 616 goto nla_put_failure; 617 618 nla_nest_end(skb, nest); 619 620 if (tcf_exts_dump_stats(skb, &prog->exts) < 0) 621 goto nla_put_failure; 622 623 return skb->len; 624 625 nla_put_failure: 626 nla_nest_cancel(skb, nest); 627 return -1; 628 } 629 630 static void cls_bpf_bind_class(void *fh, u32 classid, unsigned long cl) 631 { 632 struct cls_bpf_prog *prog = fh; 633 634 if (prog && prog->res.classid == classid) 635 prog->res.class = cl; 636 } 637 638 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg) 639 { 640 struct cls_bpf_head *head = rtnl_dereference(tp->root); 641 struct cls_bpf_prog *prog; 642 643 list_for_each_entry(prog, &head->plist, link) { 644 if (arg->count < arg->skip) 645 goto skip; 646 if (arg->fn(tp, prog, arg) < 0) { 647 arg->stop = 1; 648 break; 649 } 650 skip: 651 arg->count++; 652 } 653 } 654 655 static struct tcf_proto_ops cls_bpf_ops __read_mostly = { 656 .kind = "bpf", 657 .owner = THIS_MODULE, 658 .classify = cls_bpf_classify, 659 .init = cls_bpf_init, 660 .destroy = cls_bpf_destroy, 661 .get = cls_bpf_get, 662 .change = cls_bpf_change, 663 .delete = cls_bpf_delete, 664 .walk = cls_bpf_walk, 665 .dump = cls_bpf_dump, 666 .bind_class = cls_bpf_bind_class, 667 }; 668 669 static int __init cls_bpf_init_mod(void) 670 { 671 return register_tcf_proto_ops(&cls_bpf_ops); 672 } 673 674 static void __exit cls_bpf_exit_mod(void) 675 { 676 unregister_tcf_proto_ops(&cls_bpf_ops); 677 } 678 679 module_init(cls_bpf_init_mod); 680 module_exit(cls_bpf_exit_mod); 681