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         struct tcf_exts exts;
  42         u32 handle;
  43         union {
  44                 u32 bpf_fd;
  45                 u16 bpf_num_ops;
  46         };
  47         struct sock_filter *bpf_ops;
  48         const char *bpf_name;
  49         struct tcf_proto *tp;
  50         struct rcu_head rcu;
  51 };
  52
  53 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
  54         [TCA_BPF_CLASSID]       = { .type = NLA_U32 },
  55         [TCA_BPF_FD]            = { .type = NLA_U32 },
  56         [TCA_BPF_NAME]          = { .type = NLA_NUL_STRING, .len = CLS_BPF_NAME_LEN },
  57         [TCA_BPF_OPS_LEN]       = { .type = NLA_U16 },
  58         [TCA_BPF_OPS]           = { .type = NLA_BINARY,
  59                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
  60 };
  61
  62 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
  63                             struct tcf_result *res)
  64 {
  65         struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
  66         struct cls_bpf_prog *prog;
  67 #ifdef CONFIG_NET_CLS_ACT
  68         bool at_ingress = G_TC_AT(skb->tc_verd) & AT_INGRESS;
  69 #else
  70         bool at_ingress = false;
  71 #endif
  72         int ret = -1;
  73
  74         if (unlikely(!skb_mac_header_was_set(skb)))
  75                 return -1;
  76
  77         /* Needed here for accessing maps. */
  78         rcu_read_lock();
  79         list_for_each_entry_rcu(prog, &head->plist, link) {
  80                 int filter_res;
  81
  82                 if (at_ingress) {
  83                         /* It is safe to push/pull even if skb_shared() */
  84                         __skb_push(skb, skb->mac_len);
  85                         filter_res = BPF_PROG_RUN(prog->filter, skb);
  86                         __skb_pull(skb, skb->mac_len);
  87                 } else {
  88                         filter_res = BPF_PROG_RUN(prog->filter, skb);
  89                 }
  90
  91                 if (filter_res == 0)
  92                         continue;
  93
  94                 *res = prog->res;
  95                 if (filter_res != -1)
  96                         res->classid = filter_res;
  97
  98                 ret = tcf_exts_exec(skb, &prog->exts, res);
  99                 if (ret < 0)
 100                         continue;
 101
 102                 break;
 103         }
 104         rcu_read_unlock();
 105
 106         return ret;
 107 }
 108
 109 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
 110 {
 111         return !prog->bpf_ops;
 112 }
 113
 114 static int cls_bpf_init(struct tcf_proto *tp)
 115 {
 116         struct cls_bpf_head *head;
 117
 118         head = kzalloc(sizeof(*head), GFP_KERNEL);
 119         if (head == NULL)
 120                 return -ENOBUFS;
 121
 122         INIT_LIST_HEAD_RCU(&head->plist);
 123         rcu_assign_pointer(tp->root, head);
 124
 125         return 0;
 126 }
 127
 128 static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
 129 {
 130         tcf_exts_destroy(&prog->exts);
 131
 132         if (cls_bpf_is_ebpf(prog))
 133                 bpf_prog_put(prog->filter);
 134         else
 135                 bpf_prog_destroy(prog->filter);
 136
 137         kfree(prog->bpf_name);
 138         kfree(prog->bpf_ops);
 139         kfree(prog);
 140 }
 141
 142 static void __cls_bpf_delete_prog(struct rcu_head *rcu)
 143 {
 144         struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
 145
 146         cls_bpf_delete_prog(prog->tp, prog);
 147 }
 148
 149 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
 150 {
 151         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
 152
 153         list_del_rcu(&prog->link);
 154         tcf_unbind_filter(tp, &prog->res);
 155         call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 156
 157         return 0;
 158 }
 159
 160 static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
 161 {
 162         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 163         struct cls_bpf_prog *prog, *tmp;
 164
 165         if (!force && !list_empty(&head->plist))
 166                 return false;
 167
 168         list_for_each_entry_safe(prog, tmp, &head->plist, link) {
 169                 list_del_rcu(&prog->link);
 170                 tcf_unbind_filter(tp, &prog->res);
 171                 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
 172         }
 173
 174         RCU_INIT_POINTER(tp->root, NULL);
 175         kfree_rcu(head, rcu);
 176         return true;
 177 }
 178
 179 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
 180 {
 181         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 182         struct cls_bpf_prog *prog;
 183         unsigned long ret = 0UL;
 184
 185         if (head == NULL)
 186                 return 0UL;
 187
 188         list_for_each_entry(prog, &head->plist, link) {
 189                 if (prog->handle == handle) {
 190                         ret = (unsigned long) prog;
 191                         break;
 192                 }
 193         }
 194
 195         return ret;
 196 }
 197
 198 static int cls_bpf_prog_from_ops(struct nlattr **tb,
 199                                  struct cls_bpf_prog *prog, u32 classid)
 200 {
 201         struct sock_filter *bpf_ops;
 202         struct sock_fprog_kern fprog_tmp;
 203         struct bpf_prog *fp;
 204         u16 bpf_size, bpf_num_ops;
 205         int ret;
 206
 207         bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
 208         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 209                 return -EINVAL;
 210
 211         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 212         if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
 213                 return -EINVAL;
 214
 215         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 216         if (bpf_ops == NULL)
 217                 return -ENOMEM;
 218
 219         memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
 220
 221         fprog_tmp.len = bpf_num_ops;
 222         fprog_tmp.filter = bpf_ops;
 223
 224         ret = bpf_prog_create(&fp, &fprog_tmp);
 225         if (ret < 0) {
 226                 kfree(bpf_ops);
 227                 return ret;
 228         }
 229
 230         prog->bpf_ops = bpf_ops;
 231         prog->bpf_num_ops = bpf_num_ops;
 232         prog->bpf_name = NULL;
 233
 234         prog->filter = fp;
 235         prog->res.classid = classid;
 236
 237         return 0;
 238 }
 239
 240 static int cls_bpf_prog_from_efd(struct nlattr **tb,
 241                                  struct cls_bpf_prog *prog, u32 classid)
 242 {
 243         struct bpf_prog *fp;
 244         char *name = NULL;
 245         u32 bpf_fd;
 246
 247         bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
 248
 249         fp = bpf_prog_get(bpf_fd);
 250         if (IS_ERR(fp))
 251                 return PTR_ERR(fp);
 252
 253         if (fp->type != BPF_PROG_TYPE_SCHED_CLS) {
 254                 bpf_prog_put(fp);
 255                 return -EINVAL;
 256         }
 257
 258         if (tb[TCA_BPF_NAME]) {
 259                 name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
 260                                nla_len(tb[TCA_BPF_NAME]),
 261                                GFP_KERNEL);
 262                 if (!name) {
 263                         bpf_prog_put(fp);
 264                         return -ENOMEM;
 265                 }
 266         }
 267
 268         prog->bpf_ops = NULL;
 269         prog->bpf_fd = bpf_fd;
 270         prog->bpf_name = name;
 271
 272         prog->filter = fp;
 273         prog->res.classid = classid;
 274
 275         return 0;
 276 }
 277
 278 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
 279                                    struct cls_bpf_prog *prog,
 280                                    unsigned long base, struct nlattr **tb,
 281                                    struct nlattr *est, bool ovr)
 282 {
 283         struct tcf_exts exts;
 284         bool is_bpf, is_ebpf;
 285         u32 classid;
 286         int ret;
 287
 288         is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
 289         is_ebpf = tb[TCA_BPF_FD];
 290
 291         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf) ||
 292             !tb[TCA_BPF_CLASSID])
 293                 return -EINVAL;
 294
 295         tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 296         ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
 297         if (ret < 0)
 298                 return ret;
 299
 300         classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
 301
 302         ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog, classid) :
 303                        cls_bpf_prog_from_efd(tb, prog, classid);
 304         if (ret < 0) {
 305                 tcf_exts_destroy(&exts);
 306                 return ret;
 307         }
 308
 309         tcf_bind_filter(tp, &prog->res, base);
 310         tcf_exts_change(tp, &prog->exts, &exts);
 311
 312         return 0;
 313 }
 314
 315 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
 316                                    struct cls_bpf_head *head)
 317 {
 318         unsigned int i = 0x80000000;
 319         u32 handle;
 320
 321         do {
 322                 if (++head->hgen == 0x7FFFFFFF)
 323                         head->hgen = 1;
 324         } while (--i > 0 && cls_bpf_get(tp, head->hgen));
 325
 326         if (unlikely(i == 0)) {
 327                 pr_err("Insufficient number of handles\n");
 328                 handle = 0;
 329         } else {
 330                 handle = head->hgen;
 331         }
 332
 333         return handle;
 334 }
 335
 336 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
 337                           struct tcf_proto *tp, unsigned long base,
 338                           u32 handle, struct nlattr **tca,
 339                           unsigned long *arg, bool ovr)
 340 {
 341         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 342         struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
 343         struct nlattr *tb[TCA_BPF_MAX + 1];
 344         struct cls_bpf_prog *prog;
 345         int ret;
 346
 347         if (tca[TCA_OPTIONS] == NULL)
 348                 return -EINVAL;
 349
 350         ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
 351         if (ret < 0)
 352                 return ret;
 353
 354         prog = kzalloc(sizeof(*prog), GFP_KERNEL);
 355         if (!prog)
 356                 return -ENOBUFS;
 357
 358         tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
 359
 360         if (oldprog) {
 361                 if (handle && oldprog->handle != handle) {
 362                         ret = -EINVAL;
 363                         goto errout;
 364                 }
 365         }
 366
 367         if (handle == 0)
 368                 prog->handle = cls_bpf_grab_new_handle(tp, head);
 369         else
 370                 prog->handle = handle;
 371         if (prog->handle == 0) {
 372                 ret = -EINVAL;
 373                 goto errout;
 374         }
 375
 376         ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
 377         if (ret < 0)
 378                 goto errout;
 379
 380         if (oldprog) {
 381                 list_replace_rcu(&oldprog->link, &prog->link);
 382                 tcf_unbind_filter(tp, &oldprog->res);
 383                 call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
 384         } else {
 385                 list_add_rcu(&prog->link, &head->plist);
 386         }
 387
 388         *arg = (unsigned long) prog;
 389         return 0;
 390 errout:
 391         kfree(prog);
 392
 393         return ret;
 394 }
 395
 396 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
 397                                  struct sk_buff *skb)
 398 {
 399         struct nlattr *nla;
 400
 401         if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
 402                 return -EMSGSIZE;
 403
 404         nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
 405                           sizeof(struct sock_filter));
 406         if (nla == NULL)
 407                 return -EMSGSIZE;
 408
 409         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 410
 411         return 0;
 412 }
 413
 414 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
 415                                   struct sk_buff *skb)
 416 {
 417         if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
 418                 return -EMSGSIZE;
 419
 420         if (prog->bpf_name &&
 421             nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
 422                 return -EMSGSIZE;
 423
 424         return 0;
 425 }
 426
 427 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
 428                         struct sk_buff *skb, struct tcmsg *tm)
 429 {
 430         struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
 431         struct nlattr *nest;
 432         int ret;
 433
 434         if (prog == NULL)
 435                 return skb->len;
 436
 437         tm->tcm_handle = prog->handle;
 438
 439         nest = nla_nest_start(skb, TCA_OPTIONS);
 440         if (nest == NULL)
 441                 goto nla_put_failure;
 442
 443         if (nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
 444                 goto nla_put_failure;
 445
 446         if (cls_bpf_is_ebpf(prog))
 447                 ret = cls_bpf_dump_ebpf_info(prog, skb);
 448         else
 449                 ret = cls_bpf_dump_bpf_info(prog, skb);
 450         if (ret)
 451                 goto nla_put_failure;
 452
 453         if (tcf_exts_dump(skb, &prog->exts) < 0)
 454                 goto nla_put_failure;
 455
 456         nla_nest_end(skb, nest);
 457
 458         if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
 459                 goto nla_put_failure;
 460
 461         return skb->len;
 462
 463 nla_put_failure:
 464         nla_nest_cancel(skb, nest);
 465         return -1;
 466 }
 467
 468 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
 469 {
 470         struct cls_bpf_head *head = rtnl_dereference(tp->root);
 471         struct cls_bpf_prog *prog;
 472
 473         list_for_each_entry(prog, &head->plist, link) {
 474                 if (arg->count < arg->skip)
 475                         goto skip;
 476                 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
 477                         arg->stop = 1;
 478                         break;
 479                 }
 480 skip:
 481                 arg->count++;
 482         }
 483 }
 484
 485 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
 486         .kind           =       "bpf",
 487         .owner          =       THIS_MODULE,
 488         .classify       =       cls_bpf_classify,
 489         .init           =       cls_bpf_init,
 490         .destroy        =       cls_bpf_destroy,
 491         .get            =       cls_bpf_get,
 492         .change         =       cls_bpf_change,
 493         .delete         =       cls_bpf_delete,
 494         .walk           =       cls_bpf_walk,
 495         .dump           =       cls_bpf_dump,
 496 };
 497
 498 static int __init cls_bpf_init_mod(void)
 499 {
 500         return register_tcf_proto_ops(&cls_bpf_ops);
 501 }
 502
 503 static void __exit cls_bpf_exit_mod(void)
 504 {
 505         unregister_tcf_proto_ops(&cls_bpf_ops);
 506 }
 507
 508 module_init(cls_bpf_init_mod);
 509 module_exit(cls_bpf_exit_mod);