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