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         struct tcf_exts exts;
  43         u32 handle;
  44         union {
  45                 u32 bpf_fd;
  46                 u16 bpf_num_ops;
  47         };
  48         struct sock_filter *bpf_ops;
  49         const char *bpf_name;
  50         struct tcf_proto *tp;
  51         struct rcu_head rcu;
  52 };
  53
  54 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
  55         [TCA_BPF_CLASSID]       = { .type = NLA_U32 },
  56         [TCA_BPF_FLAGS]         = { .type = NLA_U32 },
  57         [TCA_BPF_FD]            = { .type = NLA_U32 },
  58         [TCA_BPF_NAME]          = { .type = NLA_NUL_STRING, .len = CLS_BPF_NAME_LEN },
  59         [TCA_BPF_OPS_LEN]       = { .type = NLA_U16 },
  60         [TCA_BPF_OPS]           = { .type = NLA_BINARY,
  61                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
  62 };
  63
  64 static int cls_bpf_exec_opcode(int code)
  65 {
  66         switch (code) {
  67         case TC_ACT_OK:
  68         case TC_ACT_SHOT:
  69         case TC_ACT_STOLEN:
  70         case TC_ACT_REDIRECT:
  71         case TC_ACT_UNSPEC:
  72                 return code;
  73         default:
  74                 return TC_ACT_UNSPEC;
  75         }
  76 }
  77
  78 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  79                             struct tcf_result *res)
  80 {
  81         struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
  82         bool at_ingress = skb_at_tc_ingress(skb);
  83         struct cls_bpf_prog *prog;
  84         int ret = -1;
  85
  86         if (unlikely(!skb_mac_header_was_set(skb)))
  87                 return -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 (at_ingress) {
  97                         /* It is safe to push/pull even if skb_shared() */
  98                         __skb_push(skb, skb->mac_len);
  99                         filter_res = BPF_PROG_RUN(prog->filter, skb);
 100                         __skb_pull(skb, skb->mac_len);
 101                 } else {
 102                         filter_res = BPF_PROG_RUN(prog->filter, skb);
 103                 }
 104
 105                 if (prog->exts_integrated) {
 106                         res->class   = 0;
 107                         res->classid = TC_H_MAJ(prog->res.classid) |
 108                                        qdisc_skb_cb(skb)->tc_classid;
 109
 110                         ret = cls_bpf_exec_opcode(filter_res);
 111                         if (ret == TC_ACT_UNSPEC)
 112                                 continue;
 113                         break;
 114                 }
 115
 116                 if (filter_res == 0)
 117                         continue;
 118                 if (filter_res != -1) {
 119                         res->class   = 0;
 120                         res->classid = filter_res;
 121                 } else {
 122                         *res = prog->res;
 123                 }
 124
 125                 ret = tcf_exts_exec(skb, &prog->exts, res);
 126                 if (ret < 0)
 127                         continue;
 128
 129                 break;
 130         }
 131         rcu_read_unlock();
 132
 133         return ret;
 134 }
 135
 136 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
 137 {
 138         return !prog->bpf_ops;
 139 }
 140
 141 static int cls_bpf_init(struct tcf_proto *tp)
 142 {
 143         struct cls_bpf_head *head;
 144
 145         head = kzalloc(sizeof(*head), GFP_KERNEL);
 146         if (head == NULL)
 147                 return -ENOBUFS;
 148
 149         INIT_LIST_HEAD_RCU(&head->plist);
 150         rcu_assign_pointer(tp->root, head);
 151
 152         return 0;
 153 }
 154
 155 static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
 156 {
 157         tcf_exts_destroy(&prog->exts);
 158
 159         if (cls_bpf_is_ebpf(prog))
 160                 bpf_prog_put(prog->filter);
 161         else
 162                 bpf_prog_destroy(prog->filter);
 163
 164         kfree(prog->bpf_name);
 165         kfree(prog->bpf_ops);
 166         kfree(prog);
 167 }
 168
 169 static void __cls_bpf_delete_prog(struct rcu_head *rcu)
 170 {
 171         struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
 172
 173         cls_bpf_delete_prog(prog->tp, prog);
 174 }
 175
 176 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
 177 {
 178         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
 179
 180         list_del_rcu(&prog->link);
 181         tcf_unbind_filter(tp, &prog->res);
 182         call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 183
 184         return 0;
 185 }
 186
 187 static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
 188 {
 189         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 190         struct cls_bpf_prog *prog, *tmp;
 191
 192         if (!force && !list_empty(&head->plist))
 193                 return false;
 194
 195         list_for_each_entry_safe(prog, tmp, &head->plist, link) {
 196                 list_del_rcu(&prog->link);
 197                 tcf_unbind_filter(tp, &prog->res);
 198                 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 199         }
 200
 201         RCU_INIT_POINTER(tp->root, NULL);
 202         kfree_rcu(head, rcu);
 203         return true;
 204 }
 205
 206 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
 207 {
 208         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 209         struct cls_bpf_prog *prog;
 210         unsigned long ret = 0UL;
 211
 212         if (head == NULL)
 213                 return 0UL;
 214
 215         list_for_each_entry(prog, &head->plist, link) {
 216                 if (prog->handle == handle) {
 217                         ret = (unsigned long) prog;
 218                         break;
 219                 }
 220         }
 221
 222         return ret;
 223 }
 224
 225 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
 226 {
 227         struct sock_filter *bpf_ops;
 228         struct sock_fprog_kern fprog_tmp;
 229         struct bpf_prog *fp;
 230         u16 bpf_size, bpf_num_ops;
 231         int ret;
 232
 233         bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
 234         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 235                 return -EINVAL;
 236
 237         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 238         if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
 239                 return -EINVAL;
 240
 241         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 242         if (bpf_ops == NULL)
 243                 return -ENOMEM;
 244
 245         memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
 246
 247         fprog_tmp.len = bpf_num_ops;
 248         fprog_tmp.filter = bpf_ops;
 249
 250         ret = bpf_prog_create(&fp, &fprog_tmp);
 251         if (ret < 0) {
 252                 kfree(bpf_ops);
 253                 return ret;
 254         }
 255
 256         prog->bpf_ops = bpf_ops;
 257         prog->bpf_num_ops = bpf_num_ops;
 258         prog->bpf_name = NULL;
 259         prog->filter = fp;
 260
 261         return 0;
 262 }
 263
 264 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
 265                                  const struct tcf_proto *tp)
 266 {
 267         struct bpf_prog *fp;
 268         char *name = NULL;
 269         u32 bpf_fd;
 270
 271         bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
 272
 273         fp = bpf_prog_get(bpf_fd);
 274         if (IS_ERR(fp))
 275                 return PTR_ERR(fp);
 276
 277         if (fp->type != BPF_PROG_TYPE_SCHED_CLS) {
 278                 bpf_prog_put(fp);
 279                 return -EINVAL;
 280         }
 281
 282         if (tb[TCA_BPF_NAME]) {
 283                 name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
 284                                nla_len(tb[TCA_BPF_NAME]),
 285                                GFP_KERNEL);
 286                 if (!name) {
 287                         bpf_prog_put(fp);
 288                         return -ENOMEM;
 289                 }
 290         }
 291
 292         prog->bpf_ops = NULL;
 293         prog->bpf_fd = bpf_fd;
 294         prog->bpf_name = name;
 295         prog->filter = fp;
 296
 297         if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
 298                 netif_keep_dst(qdisc_dev(tp->q));
 299
 300         return 0;
 301 }
 302
 303 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
 304                                    struct cls_bpf_prog *prog,
 305                                    unsigned long base, struct nlattr **tb,
 306                                    struct nlattr *est, bool ovr)
 307 {
 308         bool is_bpf, is_ebpf, have_exts = false;
 309         struct tcf_exts exts;
 310         int ret;
 311
 312         is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
 313         is_ebpf = tb[TCA_BPF_FD];
 314         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
 315                 return -EINVAL;
 316
 317         tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 318         ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
 319         if (ret < 0)
 320                 return ret;
 321
 322         if (tb[TCA_BPF_FLAGS]) {
 323                 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
 324
 325                 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
 326                         tcf_exts_destroy(&exts);
 327                         return -EINVAL;
 328                 }
 329
 330                 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
 331         }
 332
 333         prog->exts_integrated = have_exts;
 334
 335         ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
 336                        cls_bpf_prog_from_efd(tb, prog, tp);
 337         if (ret < 0) {
 338                 tcf_exts_destroy(&exts);
 339                 return ret;
 340         }
 341
 342         if (tb[TCA_BPF_CLASSID]) {
 343                 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
 344                 tcf_bind_filter(tp, &prog->res, base);
 345         }
 346
 347         tcf_exts_change(tp, &prog->exts, &exts);
 348         return 0;
 349 }
 350
 351 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
 352                                    struct cls_bpf_head *head)
 353 {
 354         unsigned int i = 0x80000000;
 355         u32 handle;
 356
 357         do {
 358                 if (++head->hgen == 0x7FFFFFFF)
 359                         head->hgen = 1;
 360         } while (--i > 0 && cls_bpf_get(tp, head->hgen));
 361
 362         if (unlikely(i == 0)) {
 363                 pr_err("Insufficient number of handles\n");
 364                 handle = 0;
 365         } else {
 366                 handle = head->hgen;
 367         }
 368
 369         return handle;
 370 }
 371
 372 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
 373                           struct tcf_proto *tp, unsigned long base,
 374                           u32 handle, struct nlattr **tca,
 375                           unsigned long *arg, bool ovr)
 376 {
 377         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 378         struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
 379         struct nlattr *tb[TCA_BPF_MAX + 1];
 380         struct cls_bpf_prog *prog;
 381         int ret;
 382
 383         if (tca[TCA_OPTIONS] == NULL)
 384                 return -EINVAL;
 385
 386         ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
 387         if (ret < 0)
 388                 return ret;
 389
 390         prog = kzalloc(sizeof(*prog), GFP_KERNEL);
 391         if (!prog)
 392                 return -ENOBUFS;
 393
 394         tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 395
 396         if (oldprog) {
 397                 if (handle && oldprog->handle != handle) {
 398                         ret = -EINVAL;
 399                         goto errout;
 400                 }
 401         }
 402
 403         if (handle == 0)
 404                 prog->handle = cls_bpf_grab_new_handle(tp, head);
 405         else
 406                 prog->handle = handle;
 407         if (prog->handle == 0) {
 408                 ret = -EINVAL;
 409                 goto errout;
 410         }
 411
 412         ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
 413         if (ret < 0)
 414                 goto errout;
 415
 416         if (oldprog) {
 417                 list_replace_rcu(&oldprog->link, &prog->link);
 418                 tcf_unbind_filter(tp, &oldprog->res);
 419                 call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
 420         } else {
 421                 list_add_rcu(&prog->link, &head->plist);
 422         }
 423
 424         *arg = (unsigned long) prog;
 425         return 0;
 426 errout:
 427         kfree(prog);
 428
 429         return ret;
 430 }
 431
 432 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
 433                                  struct sk_buff *skb)
 434 {
 435         struct nlattr *nla;
 436
 437         if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
 438                 return -EMSGSIZE;
 439
 440         nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
 441                           sizeof(struct sock_filter));
 442         if (nla == NULL)
 443                 return -EMSGSIZE;
 444
 445         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 446
 447         return 0;
 448 }
 449
 450 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
 451                                   struct sk_buff *skb)
 452 {
 453         if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
 454                 return -EMSGSIZE;
 455
 456         if (prog->bpf_name &&
 457             nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
 458                 return -EMSGSIZE;
 459
 460         return 0;
 461 }
 462
 463 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 464                         struct sk_buff *skb, struct tcmsg *tm)
 465 {
 466         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
 467         struct nlattr *nest;
 468         u32 bpf_flags = 0;
 469         int ret;
 470
 471         if (prog == NULL)
 472                 return skb->len;
 473
 474         tm->tcm_handle = prog->handle;
 475
 476         nest = nla_nest_start(skb, TCA_OPTIONS);
 477         if (nest == NULL)
 478                 goto nla_put_failure;
 479
 480         if (prog->res.classid &&
 481             nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
 482                 goto nla_put_failure;
 483
 484         if (cls_bpf_is_ebpf(prog))
 485                 ret = cls_bpf_dump_ebpf_info(prog, skb);
 486         else
 487                 ret = cls_bpf_dump_bpf_info(prog, skb);
 488         if (ret)
 489                 goto nla_put_failure;
 490
 491         if (tcf_exts_dump(skb, &prog->exts) < 0)
 492                 goto nla_put_failure;
 493
 494         if (prog->exts_integrated)
 495                 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
 496         if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
 497                 goto nla_put_failure;
 498
 499         nla_nest_end(skb, nest);
 500
 501         if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
 502                 goto nla_put_failure;
 503
 504         return skb->len;
 505
 506 nla_put_failure:
 507         nla_nest_cancel(skb, nest);
 508         return -1;
 509 }
 510
 511 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 512 {
 513         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 514         struct cls_bpf_prog *prog;
 515
 516         list_for_each_entry(prog, &head->plist, link) {
 517                 if (arg->count < arg->skip)
 518                         goto skip;
 519                 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
 520                         arg->stop = 1;
 521                         break;
 522                 }
 523 skip:
 524                 arg->count++;
 525         }
 526 }
 527
 528 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
 529         .kind           =       "bpf",
 530         .owner          =       THIS_MODULE,
 531         .classify       =       cls_bpf_classify,
 532         .init           =       cls_bpf_init,
 533         .destroy        =       cls_bpf_destroy,
 534         .get            =       cls_bpf_get,
 535         .change         =       cls_bpf_change,
 536         .delete         =       cls_bpf_delete,
 537         .walk           =       cls_bpf_walk,
 538         .dump           =       cls_bpf_dump,
 539 };
 540
 541 static int __init cls_bpf_init_mod(void)
 542 {
 543         return register_tcf_proto_ops(&cls_bpf_ops);
 544 }
 545
 546 static void __exit cls_bpf_exit_mod(void)
 547 {
 548         unregister_tcf_proto_ops(&cls_bpf_ops);
 549 }
 550
 551 module_init(cls_bpf_init_mod);
 552 module_exit(cls_bpf_exit_mod);