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