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 union { 53 struct work_struct work; 54 struct rcu_head rcu; 55 }; 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_TRAP: 77 case TC_ACT_REDIRECT: 78 case TC_ACT_UNSPEC: 79 return code; 80 default: 81 return TC_ACT_UNSPEC; 82 } 83 } 84 85 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp, 86 struct tcf_result *res) 87 { 88 struct cls_bpf_head *head = rcu_dereference_bh(tp->root); 89 bool at_ingress = skb_at_tc_ingress(skb); 90 struct cls_bpf_prog *prog; 91 int ret = -1; 92 93 /* Needed here for accessing maps. */ 94 rcu_read_lock(); 95 list_for_each_entry_rcu(prog, &head->plist, link) { 96 int filter_res; 97 98 qdisc_skb_cb(skb)->tc_classid = prog->res.classid; 99 100 if (tc_skip_sw(prog->gen_flags)) { 101 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0; 102 } else if (at_ingress) { 103 /* It is safe to push/pull even if skb_shared() */ 104 __skb_push(skb, skb->mac_len); 105 bpf_compute_data_end(skb); 106 filter_res = BPF_PROG_RUN(prog->filter, skb); 107 __skb_pull(skb, skb->mac_len); 108 } else { 109 bpf_compute_data_end(skb); 110 filter_res = BPF_PROG_RUN(prog->filter, skb); 111 } 112 113 if (prog->exts_integrated) { 114 res->class = 0; 115 res->classid = TC_H_MAJ(prog->res.classid) | 116 qdisc_skb_cb(skb)->tc_classid; 117 118 ret = cls_bpf_exec_opcode(filter_res); 119 if (ret == TC_ACT_UNSPEC) 120 continue; 121 break; 122 } 123 124 if (filter_res == 0) 125 continue; 126 if (filter_res != -1) { 127 res->class = 0; 128 res->classid = filter_res; 129 } else { 130 *res = prog->res; 131 } 132 133 ret = tcf_exts_exec(skb, &prog->exts, res); 134 if (ret < 0) 135 continue; 136 137 break; 138 } 139 rcu_read_unlock(); 140 141 return ret; 142 } 143 144 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog) 145 { 146 return !prog->bpf_ops; 147 } 148 149 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog, 150 enum tc_clsbpf_command cmd) 151 { 152 struct net_device *dev = tp->q->dev_queue->dev; 153 struct tc_cls_bpf_offload cls_bpf = {}; 154 int err; 155 156 tc_cls_common_offload_init(&cls_bpf.common, tp); 157 cls_bpf.command = cmd; 158 cls_bpf.exts = &prog->exts; 159 cls_bpf.prog = prog->filter; 160 cls_bpf.name = prog->bpf_name; 161 cls_bpf.exts_integrated = prog->exts_integrated; 162 cls_bpf.gen_flags = prog->gen_flags; 163 164 err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSBPF, &cls_bpf); 165 if (!err && (cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE)) 166 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW; 167 168 return err; 169 } 170 171 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog, 172 struct cls_bpf_prog *oldprog) 173 { 174 struct net_device *dev = tp->q->dev_queue->dev; 175 struct cls_bpf_prog *obj = prog; 176 enum tc_clsbpf_command cmd; 177 bool skip_sw; 178 int ret; 179 180 skip_sw = tc_skip_sw(prog->gen_flags) || 181 (oldprog && tc_skip_sw(oldprog->gen_flags)); 182 183 if (oldprog && oldprog->offloaded) { 184 if (tc_should_offload(dev, prog->gen_flags)) { 185 cmd = TC_CLSBPF_REPLACE; 186 } else if (!tc_skip_sw(prog->gen_flags)) { 187 obj = oldprog; 188 cmd = TC_CLSBPF_DESTROY; 189 } else { 190 return -EINVAL; 191 } 192 } else { 193 if (!tc_should_offload(dev, prog->gen_flags)) 194 return skip_sw ? -EINVAL : 0; 195 cmd = TC_CLSBPF_ADD; 196 } 197 198 ret = cls_bpf_offload_cmd(tp, obj, cmd); 199 if (ret) 200 return skip_sw ? ret : 0; 201 202 obj->offloaded = true; 203 if (oldprog) 204 oldprog->offloaded = false; 205 206 return 0; 207 } 208 209 static void cls_bpf_stop_offload(struct tcf_proto *tp, 210 struct cls_bpf_prog *prog) 211 { 212 int err; 213 214 if (!prog->offloaded) 215 return; 216 217 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY); 218 if (err) { 219 pr_err("Stopping hardware offload failed: %d\n", err); 220 return; 221 } 222 223 prog->offloaded = false; 224 } 225 226 static void cls_bpf_offload_update_stats(struct tcf_proto *tp, 227 struct cls_bpf_prog *prog) 228 { 229 if (!prog->offloaded) 230 return; 231 232 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS); 233 } 234 235 static int cls_bpf_init(struct tcf_proto *tp) 236 { 237 struct cls_bpf_head *head; 238 239 head = kzalloc(sizeof(*head), GFP_KERNEL); 240 if (head == NULL) 241 return -ENOBUFS; 242 243 INIT_LIST_HEAD_RCU(&head->plist); 244 rcu_assign_pointer(tp->root, head); 245 246 return 0; 247 } 248 249 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog) 250 { 251 tcf_exts_destroy(&prog->exts); 252 tcf_exts_put_net(&prog->exts); 253 254 if (cls_bpf_is_ebpf(prog)) 255 bpf_prog_put(prog->filter); 256 else 257 bpf_prog_destroy(prog->filter); 258 259 kfree(prog->bpf_name); 260 kfree(prog->bpf_ops); 261 kfree(prog); 262 } 263 264 static void cls_bpf_delete_prog_work(struct work_struct *work) 265 { 266 struct cls_bpf_prog *prog = container_of(work, struct cls_bpf_prog, work); 267 268 rtnl_lock(); 269 __cls_bpf_delete_prog(prog); 270 rtnl_unlock(); 271 } 272 273 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu) 274 { 275 struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu); 276 277 INIT_WORK(&prog->work, cls_bpf_delete_prog_work); 278 tcf_queue_work(&prog->work); 279 } 280 281 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog) 282 { 283 cls_bpf_stop_offload(tp, prog); 284 list_del_rcu(&prog->link); 285 tcf_unbind_filter(tp, &prog->res); 286 if (tcf_exts_get_net(&prog->exts)) 287 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu); 288 else 289 __cls_bpf_delete_prog(prog); 290 } 291 292 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last) 293 { 294 struct cls_bpf_head *head = rtnl_dereference(tp->root); 295 296 __cls_bpf_delete(tp, arg); 297 *last = list_empty(&head->plist); 298 return 0; 299 } 300 301 static void cls_bpf_destroy(struct tcf_proto *tp) 302 { 303 struct cls_bpf_head *head = rtnl_dereference(tp->root); 304 struct cls_bpf_prog *prog, *tmp; 305 306 list_for_each_entry_safe(prog, tmp, &head->plist, link) 307 __cls_bpf_delete(tp, prog); 308 309 kfree_rcu(head, rcu); 310 } 311 312 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle) 313 { 314 struct cls_bpf_head *head = rtnl_dereference(tp->root); 315 struct cls_bpf_prog *prog; 316 317 list_for_each_entry(prog, &head->plist, link) { 318 if (prog->handle == handle) 319 return prog; 320 } 321 322 return NULL; 323 } 324 325 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog) 326 { 327 struct sock_filter *bpf_ops; 328 struct sock_fprog_kern fprog_tmp; 329 struct bpf_prog *fp; 330 u16 bpf_size, bpf_num_ops; 331 int ret; 332 333 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]); 334 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0) 335 return -EINVAL; 336 337 bpf_size = bpf_num_ops * sizeof(*bpf_ops); 338 if (bpf_size != nla_len(tb[TCA_BPF_OPS])) 339 return -EINVAL; 340 341 bpf_ops = kzalloc(bpf_size, GFP_KERNEL); 342 if (bpf_ops == NULL) 343 return -ENOMEM; 344 345 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size); 346 347 fprog_tmp.len = bpf_num_ops; 348 fprog_tmp.filter = bpf_ops; 349 350 ret = bpf_prog_create(&fp, &fprog_tmp); 351 if (ret < 0) { 352 kfree(bpf_ops); 353 return ret; 354 } 355 356 prog->bpf_ops = bpf_ops; 357 prog->bpf_num_ops = bpf_num_ops; 358 prog->bpf_name = NULL; 359 prog->filter = fp; 360 361 return 0; 362 } 363 364 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog, 365 const struct tcf_proto *tp) 366 { 367 struct bpf_prog *fp; 368 char *name = NULL; 369 u32 bpf_fd; 370 371 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]); 372 373 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS); 374 if (IS_ERR(fp)) 375 return PTR_ERR(fp); 376 377 if (tb[TCA_BPF_NAME]) { 378 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL); 379 if (!name) { 380 bpf_prog_put(fp); 381 return -ENOMEM; 382 } 383 } 384 385 prog->bpf_ops = NULL; 386 prog->bpf_name = name; 387 prog->filter = fp; 388 389 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS)) 390 netif_keep_dst(qdisc_dev(tp->q)); 391 392 return 0; 393 } 394 395 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp, 396 struct cls_bpf_prog *prog, unsigned long base, 397 struct nlattr **tb, struct nlattr *est, bool ovr) 398 { 399 bool is_bpf, is_ebpf, have_exts = false; 400 u32 gen_flags = 0; 401 int ret; 402 403 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS]; 404 is_ebpf = tb[TCA_BPF_FD]; 405 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) 406 return -EINVAL; 407 408 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr); 409 if (ret < 0) 410 return ret; 411 412 if (tb[TCA_BPF_FLAGS]) { 413 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]); 414 415 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) 416 return -EINVAL; 417 418 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT; 419 } 420 if (tb[TCA_BPF_FLAGS_GEN]) { 421 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]); 422 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS || 423 !tc_flags_valid(gen_flags)) 424 return -EINVAL; 425 } 426 427 prog->exts_integrated = have_exts; 428 prog->gen_flags = gen_flags; 429 430 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) : 431 cls_bpf_prog_from_efd(tb, prog, tp); 432 if (ret < 0) 433 return ret; 434 435 if (tb[TCA_BPF_CLASSID]) { 436 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]); 437 tcf_bind_filter(tp, &prog->res, base); 438 } 439 440 return 0; 441 } 442 443 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp, 444 struct cls_bpf_head *head) 445 { 446 unsigned int i = 0x80000000; 447 u32 handle; 448 449 do { 450 if (++head->hgen == 0x7FFFFFFF) 451 head->hgen = 1; 452 } while (--i > 0 && cls_bpf_get(tp, head->hgen)); 453 454 if (unlikely(i == 0)) { 455 pr_err("Insufficient number of handles\n"); 456 handle = 0; 457 } else { 458 handle = head->hgen; 459 } 460 461 return handle; 462 } 463 464 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb, 465 struct tcf_proto *tp, unsigned long base, 466 u32 handle, struct nlattr **tca, 467 void **arg, bool ovr) 468 { 469 struct cls_bpf_head *head = rtnl_dereference(tp->root); 470 struct cls_bpf_prog *oldprog = *arg; 471 struct nlattr *tb[TCA_BPF_MAX + 1]; 472 struct cls_bpf_prog *prog; 473 int ret; 474 475 if (tca[TCA_OPTIONS] == NULL) 476 return -EINVAL; 477 478 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy, 479 NULL); 480 if (ret < 0) 481 return ret; 482 483 prog = kzalloc(sizeof(*prog), GFP_KERNEL); 484 if (!prog) 485 return -ENOBUFS; 486 487 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE); 488 if (ret < 0) 489 goto errout; 490 491 if (oldprog) { 492 if (handle && oldprog->handle != handle) { 493 ret = -EINVAL; 494 goto errout; 495 } 496 } 497 498 if (handle == 0) 499 prog->handle = cls_bpf_grab_new_handle(tp, head); 500 else 501 prog->handle = handle; 502 if (prog->handle == 0) { 503 ret = -EINVAL; 504 goto errout; 505 } 506 507 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr); 508 if (ret < 0) 509 goto errout; 510 511 ret = cls_bpf_offload(tp, prog, oldprog); 512 if (ret) { 513 __cls_bpf_delete_prog(prog); 514 return ret; 515 } 516 517 if (!tc_in_hw(prog->gen_flags)) 518 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW; 519 520 if (oldprog) { 521 list_replace_rcu(&oldprog->link, &prog->link); 522 tcf_unbind_filter(tp, &oldprog->res); 523 tcf_exts_get_net(&oldprog->exts); 524 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu); 525 } else { 526 list_add_rcu(&prog->link, &head->plist); 527 } 528 529 *arg = prog; 530 return 0; 531 532 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