net/sched/cls_bpf.c

   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
  31 struct cls_bpf_head {
  32         struct list_head plist;
  33         u32 hgen;
  34         struct rcu_head rcu;
  35 };
  36
  37 struct cls_bpf_prog {
  38         struct bpf_prog *filter;
  39         struct list_head link;
  40         struct tcf_result res;
  41         bool exts_integrated;
  42         bool offloaded;
  43         struct tcf_exts exts;
  44         u32 handle;
  45         union {
  46                 u32 bpf_fd;
  47                 u16 bpf_num_ops;
  48         };
  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_FD]            = { .type = NLA_U32 },
  59         [TCA_BPF_NAME]          = { .type = NLA_NUL_STRING,
  60                                     .len = CLS_BPF_NAME_LEN },
  61         [TCA_BPF_OPS_LEN]       = { .type = NLA_U16 },
  62         [TCA_BPF_OPS]           = { .type = NLA_BINARY,
  63                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
  64 };
  65
  66 static int cls_bpf_exec_opcode(int code)
  67 {
  68         switch (code) {
  69         case TC_ACT_OK:
  70         case TC_ACT_SHOT:
  71         case TC_ACT_STOLEN:
  72         case TC_ACT_REDIRECT:
  73         case TC_ACT_UNSPEC:
  74                 return code;
  75         default:
  76                 return TC_ACT_UNSPEC;
  77         }
  78 }
  79
  80 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  81                             struct tcf_result *res)
  82 {
  83         struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
  84         bool at_ingress = skb_at_tc_ingress(skb);
  85         struct cls_bpf_prog *prog;
  86         int ret = -1;
  87
  88         /* Needed here for accessing maps. */
  89         rcu_read_lock();
  90         list_for_each_entry_rcu(prog, &head->plist, link) {
  91                 int filter_res;
  92
  93                 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
  94
  95                 if (at_ingress) {
  96                         /* It is safe to push/pull even if skb_shared() */
  97                         __skb_push(skb, skb->mac_len);
  98                         bpf_compute_data_end(skb);
  99                         filter_res = BPF_PROG_RUN(prog->filter, skb);
 100                         __skb_pull(skb, skb->mac_len);
 101                 } else {
 102                         bpf_compute_data_end(skb);
 103                         filter_res = BPF_PROG_RUN(prog->filter, skb);
 104                 }
 105
 106                 if (prog->exts_integrated) {
 107                         res->class   = 0;
 108                         res->classid = TC_H_MAJ(prog->res.classid) |
 109                                        qdisc_skb_cb(skb)->tc_classid;
 110
 111                         ret = cls_bpf_exec_opcode(filter_res);
 112                         if (ret == TC_ACT_UNSPEC)
 113                                 continue;
 114                         break;
 115                 }
 116
 117                 if (filter_res == 0)
 118                         continue;
 119                 if (filter_res != -1) {
 120                         res->class   = 0;
 121                         res->classid = filter_res;
 122                 } else {
 123                         *res = prog->res;
 124                 }
 125
 126                 ret = tcf_exts_exec(skb, &prog->exts, res);
 127                 if (ret < 0)
 128                         continue;
 129
 130                 break;
 131         }
 132         rcu_read_unlock();
 133
 134         return ret;
 135 }
 136
 137 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
 138 {
 139         return !prog->bpf_ops;
 140 }
 141
 142 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
 143                                enum tc_clsbpf_command cmd)
 144 {
 145         struct net_device *dev = tp->q->dev_queue->dev;
 146         struct tc_cls_bpf_offload bpf_offload = {};
 147         struct tc_to_netdev offload;
 148
 149         offload.type = TC_SETUP_CLSBPF;
 150         offload.cls_bpf = &bpf_offload;
 151
 152         bpf_offload.command = cmd;
 153         bpf_offload.exts = &prog->exts;
 154         bpf_offload.prog = prog->filter;
 155         bpf_offload.name = prog->bpf_name;
 156         bpf_offload.exts_integrated = prog->exts_integrated;
 157
 158         return dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
 159                                              tp->protocol, &offload);
 160 }
 161
 162 static void cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
 163                             struct cls_bpf_prog *oldprog)
 164 {
 165         struct net_device *dev = tp->q->dev_queue->dev;
 166         struct cls_bpf_prog *obj = prog;
 167         enum tc_clsbpf_command cmd;
 168
 169         if (oldprog && oldprog->offloaded) {
 170                 if (tc_should_offload(dev, tp, 0)) {
 171                         cmd = TC_CLSBPF_REPLACE;
 172                 } else {
 173                         obj = oldprog;
 174                         cmd = TC_CLSBPF_DESTROY;
 175                 }
 176         } else {
 177                 if (!tc_should_offload(dev, tp, 0))
 178                         return;
 179                 cmd = TC_CLSBPF_ADD;
 180         }
 181
 182         if (cls_bpf_offload_cmd(tp, obj, cmd))
 183                 return;
 184
 185         obj->offloaded = true;
 186         if (oldprog)
 187                 oldprog->offloaded = false;
 188 }
 189
 190 static void cls_bpf_stop_offload(struct tcf_proto *tp,
 191                                  struct cls_bpf_prog *prog)
 192 {
 193         int err;
 194
 195         if (!prog->offloaded)
 196                 return;
 197
 198         err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
 199         if (err) {
 200                 pr_err("Stopping hardware offload failed: %d\n", err);
 201                 return;
 202         }
 203
 204         prog->offloaded = false;
 205 }
 206
 207 static int cls_bpf_init(struct tcf_proto *tp)
 208 {
 209         struct cls_bpf_head *head;
 210
 211         head = kzalloc(sizeof(*head), GFP_KERNEL);
 212         if (head == NULL)
 213                 return -ENOBUFS;
 214
 215         INIT_LIST_HEAD_RCU(&head->plist);
 216         rcu_assign_pointer(tp->root, head);
 217
 218         return 0;
 219 }
 220
 221 static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
 222 {
 223         tcf_exts_destroy(&prog->exts);
 224
 225         if (cls_bpf_is_ebpf(prog))
 226                 bpf_prog_put(prog->filter);
 227         else
 228                 bpf_prog_destroy(prog->filter);
 229
 230         kfree(prog->bpf_name);
 231         kfree(prog->bpf_ops);
 232         kfree(prog);
 233 }
 234
 235 static void __cls_bpf_delete_prog(struct rcu_head *rcu)
 236 {
 237         struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
 238
 239         cls_bpf_delete_prog(prog->tp, prog);
 240 }
 241
 242 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
 243 {
 244         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
 245
 246         cls_bpf_stop_offload(tp, prog);
 247         list_del_rcu(&prog->link);
 248         tcf_unbind_filter(tp, &prog->res);
 249         call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 250
 251         return 0;
 252 }
 253
 254 static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
 255 {
 256         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 257         struct cls_bpf_prog *prog, *tmp;
 258
 259         if (!force && !list_empty(&head->plist))
 260                 return false;
 261
 262         list_for_each_entry_safe(prog, tmp, &head->plist, link) {
 263                 cls_bpf_stop_offload(tp, prog);
 264                 list_del_rcu(&prog->link);
 265                 tcf_unbind_filter(tp, &prog->res);
 266                 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 267         }
 268
 269         RCU_INIT_POINTER(tp->root, NULL);
 270         kfree_rcu(head, rcu);
 271         return true;
 272 }
 273
 274 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
 275 {
 276         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 277         struct cls_bpf_prog *prog;
 278         unsigned long ret = 0UL;
 279
 280         if (head == NULL)
 281                 return 0UL;
 282
 283         list_for_each_entry(prog, &head->plist, link) {
 284                 if (prog->handle == handle) {
 285                         ret = (unsigned long) prog;
 286                         break;
 287                 }
 288         }
 289
 290         return ret;
 291 }
 292
 293 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
 294 {
 295         struct sock_filter *bpf_ops;
 296         struct sock_fprog_kern fprog_tmp;
 297         struct bpf_prog *fp;
 298         u16 bpf_size, bpf_num_ops;
 299         int ret;
 300
 301         bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
 302         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 303                 return -EINVAL;
 304
 305         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 306         if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
 307                 return -EINVAL;
 308
 309         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 310         if (bpf_ops == NULL)
 311                 return -ENOMEM;
 312
 313         memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
 314
 315         fprog_tmp.len = bpf_num_ops;
 316         fprog_tmp.filter = bpf_ops;
 317
 318         ret = bpf_prog_create(&fp, &fprog_tmp);
 319         if (ret < 0) {
 320                 kfree(bpf_ops);
 321                 return ret;
 322         }
 323
 324         prog->bpf_ops = bpf_ops;
 325         prog->bpf_num_ops = bpf_num_ops;
 326         prog->bpf_name = NULL;
 327         prog->filter = fp;
 328
 329         return 0;
 330 }
 331
 332 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
 333                                  const struct tcf_proto *tp)
 334 {
 335         struct bpf_prog *fp;
 336         char *name = NULL;
 337         u32 bpf_fd;
 338
 339         bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
 340
 341         fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
 342         if (IS_ERR(fp))
 343                 return PTR_ERR(fp);
 344
 345         if (tb[TCA_BPF_NAME]) {
 346                 name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
 347                                nla_len(tb[TCA_BPF_NAME]),
 348                                GFP_KERNEL);
 349                 if (!name) {
 350                         bpf_prog_put(fp);
 351                         return -ENOMEM;
 352                 }
 353         }
 354
 355         prog->bpf_ops = NULL;
 356         prog->bpf_fd = bpf_fd;
 357         prog->bpf_name = name;
 358         prog->filter = fp;
 359
 360         if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
 361                 netif_keep_dst(qdisc_dev(tp->q));
 362
 363         return 0;
 364 }
 365
 366 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
 367                                    struct cls_bpf_prog *prog,
 368                                    unsigned long base, struct nlattr **tb,
 369                                    struct nlattr *est, bool ovr)
 370 {
 371         bool is_bpf, is_ebpf, have_exts = false;
 372         struct tcf_exts exts;
 373         int ret;
 374
 375         is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
 376         is_ebpf = tb[TCA_BPF_FD];
 377         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
 378                 return -EINVAL;
 379
 380         ret = tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 381         if (ret < 0)
 382                 return ret;
 383         ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
 384         if (ret < 0)
 385                 goto errout;
 386
 387         if (tb[TCA_BPF_FLAGS]) {
 388                 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
 389
 390                 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
 391                         ret = -EINVAL;
 392                         goto errout;
 393                 }
 394
 395                 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
 396         }
 397
 398         prog->exts_integrated = have_exts;
 399
 400         ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
 401                        cls_bpf_prog_from_efd(tb, prog, tp);
 402         if (ret < 0)
 403                 goto errout;
 404
 405         if (tb[TCA_BPF_CLASSID]) {
 406                 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
 407                 tcf_bind_filter(tp, &prog->res, base);
 408         }
 409
 410         tcf_exts_change(tp, &prog->exts, &exts);
 411         return 0;
 412
 413 errout:
 414         tcf_exts_destroy(&exts);
 415         return ret;
 416 }
 417
 418 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
 419                                    struct cls_bpf_head *head)
 420 {
 421         unsigned int i = 0x80000000;
 422         u32 handle;
 423
 424         do {
 425                 if (++head->hgen == 0x7FFFFFFF)
 426                         head->hgen = 1;
 427         } while (--i > 0 && cls_bpf_get(tp, head->hgen));
 428
 429         if (unlikely(i == 0)) {
 430                 pr_err("Insufficient number of handles\n");
 431                 handle = 0;
 432         } else {
 433                 handle = head->hgen;
 434         }
 435
 436         return handle;
 437 }
 438
 439 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
 440                           struct tcf_proto *tp, unsigned long base,
 441                           u32 handle, struct nlattr **tca,
 442                           unsigned long *arg, bool ovr)
 443 {
 444         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 445         struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
 446         struct nlattr *tb[TCA_BPF_MAX + 1];
 447         struct cls_bpf_prog *prog;
 448         int ret;
 449
 450         if (tca[TCA_OPTIONS] == NULL)
 451                 return -EINVAL;
 452
 453         ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
 454         if (ret < 0)
 455                 return ret;
 456
 457         prog = kzalloc(sizeof(*prog), GFP_KERNEL);
 458         if (!prog)
 459                 return -ENOBUFS;
 460
 461         ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 462         if (ret < 0)
 463                 goto errout;
 464
 465         if (oldprog) {
 466                 if (handle && oldprog->handle != handle) {
 467                         ret = -EINVAL;
 468                         goto errout;
 469                 }
 470         }
 471
 472         if (handle == 0)
 473                 prog->handle = cls_bpf_grab_new_handle(tp, head);
 474         else
 475                 prog->handle = handle;
 476         if (prog->handle == 0) {
 477                 ret = -EINVAL;
 478                 goto errout;
 479         }
 480
 481         ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE],
 482                                       ovr);
 483         if (ret < 0)
 484                 goto errout;
 485
 486         cls_bpf_offload(tp, prog, oldprog);
 487
 488         if (oldprog) {
 489                 list_replace_rcu(&oldprog->link, &prog->link);
 490                 tcf_unbind_filter(tp, &oldprog->res);
 491                 call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
 492         } else {
 493                 list_add_rcu(&prog->link, &head->plist);
 494         }
 495
 496         *arg = (unsigned long) prog;
 497         return 0;
 498
 499 errout:
 500         tcf_exts_destroy(&prog->exts);
 501         kfree(prog);
 502         return ret;
 503 }
 504
 505 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
 506                                  struct sk_buff *skb)
 507 {
 508         struct nlattr *nla;
 509
 510         if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
 511                 return -EMSGSIZE;
 512
 513         nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
 514                           sizeof(struct sock_filter));
 515         if (nla == NULL)
 516                 return -EMSGSIZE;
 517
 518         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 519
 520         return 0;
 521 }
 522
 523 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
 524                                   struct sk_buff *skb)
 525 {
 526         if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
 527                 return -EMSGSIZE;
 528
 529         if (prog->bpf_name &&
 530             nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
 531                 return -EMSGSIZE;
 532
 533         return 0;
 534 }
 535
 536 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 537                         struct sk_buff *skb, struct tcmsg *tm)
 538 {
 539         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
 540         struct nlattr *nest;
 541         u32 bpf_flags = 0;
 542         int ret;
 543
 544         if (prog == NULL)
 545                 return skb->len;
 546
 547         tm->tcm_handle = prog->handle;
 548
 549         nest = nla_nest_start(skb, TCA_OPTIONS);
 550         if (nest == NULL)
 551                 goto nla_put_failure;
 552
 553         if (prog->res.classid &&
 554             nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
 555                 goto nla_put_failure;
 556
 557         if (cls_bpf_is_ebpf(prog))
 558                 ret = cls_bpf_dump_ebpf_info(prog, skb);
 559         else
 560                 ret = cls_bpf_dump_bpf_info(prog, skb);
 561         if (ret)
 562                 goto nla_put_failure;
 563
 564         if (tcf_exts_dump(skb, &prog->exts) < 0)
 565                 goto nla_put_failure;
 566
 567         if (prog->exts_integrated)
 568                 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
 569         if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
 570                 goto nla_put_failure;
 571
 572         nla_nest_end(skb, nest);
 573
 574         if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
 575                 goto nla_put_failure;
 576
 577         return skb->len;
 578
 579 nla_put_failure:
 580         nla_nest_cancel(skb, nest);
 581         return -1;
 582 }
 583
 584 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 585 {
 586         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 587         struct cls_bpf_prog *prog;
 588
 589         list_for_each_entry(prog, &head->plist, link) {
 590                 if (arg->count < arg->skip)
 591                         goto skip;
 592                 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
 593                         arg->stop = 1;
 594                         break;
 595                 }
 596 skip:
 597                 arg->count++;
 598         }
 599 }
 600
 601 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
 602         .kind           =       "bpf",
 603         .owner          =       THIS_MODULE,
 604         .classify       =       cls_bpf_classify,
 605         .init           =       cls_bpf_init,
 606         .destroy        =       cls_bpf_destroy,
 607         .get            =       cls_bpf_get,
 608         .change         =       cls_bpf_change,
 609         .delete         =       cls_bpf_delete,
 610         .walk           =       cls_bpf_walk,
 611         .dump           =       cls_bpf_dump,
 612 };
 613
 614 static int __init cls_bpf_init_mod(void)
 615 {
 616         return register_tcf_proto_ops(&cls_bpf_ops);
 617 }
 618
 619 static void __exit cls_bpf_exit_mod(void)
 620 {
 621         unregister_tcf_proto_ops(&cls_bpf_ops);
 622 }
 623
 624 module_init(cls_bpf_init_mod);
 625 module_exit(cls_bpf_exit_mod);