net/sched/act_bpf.c

   1 /*
   2  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  */
   9
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/kernel.h>
  13 #include <linux/skbuff.h>
  14 #include <linux/rtnetlink.h>
  15 #include <linux/filter.h>
  16 #include <linux/bpf.h>
  17
  18 #include <net/netlink.h>
  19 #include <net/pkt_sched.h>
  20
  21 #include <linux/tc_act/tc_bpf.h>
  22 #include <net/tc_act/tc_bpf.h>
  23
  24 #define BPF_TAB_MASK            15
  25 #define ACT_BPF_NAME_LEN        256
  26
  27 struct tcf_bpf_cfg {
  28         struct bpf_prog *filter;
  29         struct sock_filter *bpf_ops;
  30         const char *bpf_name;
  31         u32 bpf_fd;
  32         u16 bpf_num_ops;
  33         bool is_ebpf;
  34 };
  35
  36 static int bpf_net_id;
  37
  38 static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act,
  39                    struct tcf_result *res)
  40 {
  41         struct tcf_bpf *prog = act->priv;
  42         struct bpf_prog *filter;
  43         int action, filter_res;
  44         bool at_ingress = G_TC_AT(skb->tc_verd) & AT_INGRESS;
  45
  46         if (unlikely(!skb_mac_header_was_set(skb)))
  47                 return TC_ACT_UNSPEC;
  48
  49         tcf_lastuse_update(&prog->tcf_tm);
  50         bstats_cpu_update(this_cpu_ptr(prog->common.cpu_bstats), skb);
  51
  52         rcu_read_lock();
  53         filter = rcu_dereference(prog->filter);
  54         if (at_ingress) {
  55                 __skb_push(skb, skb->mac_len);
  56                 bpf_compute_data_end(skb);
  57                 filter_res = BPF_PROG_RUN(filter, skb);
  58                 __skb_pull(skb, skb->mac_len);
  59         } else {
  60                 bpf_compute_data_end(skb);
  61                 filter_res = BPF_PROG_RUN(filter, skb);
  62         }
  63         rcu_read_unlock();
  64
  65         /* A BPF program may overwrite the default action opcode.
  66          * Similarly as in cls_bpf, if filter_res == -1 we use the
  67          * default action specified from tc.
  68          *
  69          * In case a different well-known TC_ACT opcode has been
  70          * returned, it will overwrite the default one.
  71          *
  72          * For everything else that is unkown, TC_ACT_UNSPEC is
  73          * returned.
  74          */
  75         switch (filter_res) {
  76         case TC_ACT_PIPE:
  77         case TC_ACT_RECLASSIFY:
  78         case TC_ACT_OK:
  79         case TC_ACT_REDIRECT:
  80                 action = filter_res;
  81                 break;
  82         case TC_ACT_SHOT:
  83                 action = filter_res;
  84                 qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats));
  85                 break;
  86         case TC_ACT_UNSPEC:
  87                 action = prog->tcf_action;
  88                 break;
  89         default:
  90                 action = TC_ACT_UNSPEC;
  91                 break;
  92         }
  93
  94         return action;
  95 }
  96
  97 static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog)
  98 {
  99         return !prog->bpf_ops;
 100 }
 101
 102 static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog,
 103                                  struct sk_buff *skb)
 104 {
 105         struct nlattr *nla;
 106
 107         if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops))
 108                 return -EMSGSIZE;
 109
 110         nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops *
 111                           sizeof(struct sock_filter));
 112         if (nla == NULL)
 113                 return -EMSGSIZE;
 114
 115         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 116
 117         return 0;
 118 }
 119
 120 static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog,
 121                                   struct sk_buff *skb)
 122 {
 123         if (nla_put_u32(skb, TCA_ACT_BPF_FD, prog->bpf_fd))
 124                 return -EMSGSIZE;
 125
 126         if (prog->bpf_name &&
 127             nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name))
 128                 return -EMSGSIZE;
 129
 130         return 0;
 131 }
 132
 133 static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act,
 134                         int bind, int ref)
 135 {
 136         unsigned char *tp = skb_tail_pointer(skb);
 137         struct tcf_bpf *prog = act->priv;
 138         struct tc_act_bpf opt = {
 139                 .index   = prog->tcf_index,
 140                 .refcnt  = prog->tcf_refcnt - ref,
 141                 .bindcnt = prog->tcf_bindcnt - bind,
 142                 .action  = prog->tcf_action,
 143         };
 144         struct tcf_t tm;
 145         int ret;
 146
 147         if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt))
 148                 goto nla_put_failure;
 149
 150         if (tcf_bpf_is_ebpf(prog))
 151                 ret = tcf_bpf_dump_ebpf_info(prog, skb);
 152         else
 153                 ret = tcf_bpf_dump_bpf_info(prog, skb);
 154         if (ret)
 155                 goto nla_put_failure;
 156
 157         tcf_tm_dump(&tm, &prog->tcf_tm);
 158         if (nla_put_64bit(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm,
 159                           TCA_ACT_BPF_PAD))
 160                 goto nla_put_failure;
 161
 162         return skb->len;
 163
 164 nla_put_failure:
 165         nlmsg_trim(skb, tp);
 166         return -1;
 167 }
 168
 169 static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = {
 170         [TCA_ACT_BPF_PARMS]     = { .len = sizeof(struct tc_act_bpf) },
 171         [TCA_ACT_BPF_FD]        = { .type = NLA_U32 },
 172         [TCA_ACT_BPF_NAME]      = { .type = NLA_NUL_STRING,
 173                                     .len = ACT_BPF_NAME_LEN },
 174         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 175         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 176                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 177 };
 178
 179 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 180 {
 181         struct sock_filter *bpf_ops;
 182         struct sock_fprog_kern fprog_tmp;
 183         struct bpf_prog *fp;
 184         u16 bpf_size, bpf_num_ops;
 185         int ret;
 186
 187         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 188         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 189                 return -EINVAL;
 190
 191         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 192         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 193                 return -EINVAL;
 194
 195         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 196         if (bpf_ops == NULL)
 197                 return -ENOMEM;
 198
 199         memcpy(bpf_ops, nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size);
 200
 201         fprog_tmp.len = bpf_num_ops;
 202         fprog_tmp.filter = bpf_ops;
 203
 204         ret = bpf_prog_create(&fp, &fprog_tmp);
 205         if (ret < 0) {
 206                 kfree(bpf_ops);
 207                 return ret;
 208         }
 209
 210         cfg->bpf_ops = bpf_ops;
 211         cfg->bpf_num_ops = bpf_num_ops;
 212         cfg->filter = fp;
 213         cfg->is_ebpf = false;
 214
 215         return 0;
 216 }
 217
 218 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 219 {
 220         struct bpf_prog *fp;
 221         char *name = NULL;
 222         u32 bpf_fd;
 223
 224         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 225
 226         fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_ACT);
 227         if (IS_ERR(fp))
 228                 return PTR_ERR(fp);
 229
 230         if (tb[TCA_ACT_BPF_NAME]) {
 231                 name = kmemdup(nla_data(tb[TCA_ACT_BPF_NAME]),
 232                                nla_len(tb[TCA_ACT_BPF_NAME]),
 233                                GFP_KERNEL);
 234                 if (!name) {
 235                         bpf_prog_put(fp);
 236                         return -ENOMEM;
 237                 }
 238         }
 239
 240         cfg->bpf_fd = bpf_fd;
 241         cfg->bpf_name = name;
 242         cfg->filter = fp;
 243         cfg->is_ebpf = true;
 244
 245         return 0;
 246 }
 247
 248 static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg)
 249 {
 250         if (cfg->is_ebpf)
 251                 bpf_prog_put(cfg->filter);
 252         else
 253                 bpf_prog_destroy(cfg->filter);
 254
 255         kfree(cfg->bpf_ops);
 256         kfree(cfg->bpf_name);
 257 }
 258
 259 static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
 260                                   struct tcf_bpf_cfg *cfg)
 261 {
 262         cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
 263         /* updates to prog->filter are prevented, since it's called either
 264          * with rtnl lock or during final cleanup in rcu callback
 265          */
 266         cfg->filter = rcu_dereference_protected(prog->filter, 1);
 267
 268         cfg->bpf_ops = prog->bpf_ops;
 269         cfg->bpf_name = prog->bpf_name;
 270 }
 271
 272 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 273                         struct nlattr *est, struct tc_action *act,
 274                         int replace, int bind)
 275 {
 276         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 277         struct nlattr *tb[TCA_ACT_BPF_MAX + 1];
 278         struct tcf_bpf_cfg cfg, old;
 279         struct tc_act_bpf *parm;
 280         struct tcf_bpf *prog;
 281         bool is_bpf, is_ebpf;
 282         int ret, res = 0;
 283
 284         if (!nla)
 285                 return -EINVAL;
 286
 287         ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy);
 288         if (ret < 0)
 289                 return ret;
 290
 291         if (!tb[TCA_ACT_BPF_PARMS])
 292                 return -EINVAL;
 293
 294         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 295
 296         if (!tcf_hash_check(tn, parm->index, act, bind)) {
 297                 ret = tcf_hash_create(tn, parm->index, est, act,
 298                                       sizeof(*prog), bind, true);
 299                 if (ret < 0)
 300                         return ret;
 301
 302                 res = ACT_P_CREATED;
 303         } else {
 304                 /* Don't override defaults. */
 305                 if (bind)
 306                         return 0;
 307
 308                 tcf_hash_release(act, bind);
 309                 if (!replace)
 310                         return -EEXIST;
 311         }
 312
 313         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 314         is_ebpf = tb[TCA_ACT_BPF_FD];
 315
 316         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) {
 317                 ret = -EINVAL;
 318                 goto out;
 319         }
 320
 321         memset(&cfg, 0, sizeof(cfg));
 322
 323         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 324                        tcf_bpf_init_from_efd(tb, &cfg);
 325         if (ret < 0)
 326                 goto out;
 327
 328         prog = to_bpf(act);
 329         ASSERT_RTNL();
 330
 331         if (res != ACT_P_CREATED)
 332                 tcf_bpf_prog_fill_cfg(prog, &old);
 333
 334         prog->bpf_ops = cfg.bpf_ops;
 335         prog->bpf_name = cfg.bpf_name;
 336
 337         if (cfg.bpf_num_ops)
 338                 prog->bpf_num_ops = cfg.bpf_num_ops;
 339         if (cfg.bpf_fd)
 340                 prog->bpf_fd = cfg.bpf_fd;
 341
 342         prog->tcf_action = parm->action;
 343         rcu_assign_pointer(prog->filter, cfg.filter);
 344
 345         if (res == ACT_P_CREATED) {
 346                 tcf_hash_insert(tn, act);
 347         } else {
 348                 /* make sure the program being replaced is no longer executing */
 349                 synchronize_rcu();
 350                 tcf_bpf_cfg_cleanup(&old);
 351         }
 352
 353         return res;
 354 out:
 355         if (res == ACT_P_CREATED)
 356                 tcf_hash_cleanup(act, est);
 357
 358         return ret;
 359 }
 360
 361 static void tcf_bpf_cleanup(struct tc_action *act, int bind)
 362 {
 363         struct tcf_bpf_cfg tmp;
 364
 365         tcf_bpf_prog_fill_cfg(act->priv, &tmp);
 366         tcf_bpf_cfg_cleanup(&tmp);
 367 }
 368
 369 static int tcf_bpf_walker(struct net *net, struct sk_buff *skb,
 370                           struct netlink_callback *cb, int type,
 371                           struct tc_action *a)
 372 {
 373         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 374
 375         return tcf_generic_walker(tn, skb, cb, type, a);
 376 }
 377
 378 static int tcf_bpf_search(struct net *net, struct tc_action *a, u32 index)
 379 {
 380         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 381
 382         return tcf_hash_search(tn, a, index);
 383 }
 384
 385 static struct tc_action_ops act_bpf_ops __read_mostly = {
 386         .kind           =       "bpf",
 387         .type           =       TCA_ACT_BPF,
 388         .owner          =       THIS_MODULE,
 389         .act            =       tcf_bpf,
 390         .dump           =       tcf_bpf_dump,
 391         .cleanup        =       tcf_bpf_cleanup,
 392         .init           =       tcf_bpf_init,
 393         .walk           =       tcf_bpf_walker,
 394         .lookup         =       tcf_bpf_search,
 395 };
 396
 397 static __net_init int bpf_init_net(struct net *net)
 398 {
 399         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 400
 401         return tc_action_net_init(tn, &act_bpf_ops, BPF_TAB_MASK);
 402 }
 403
 404 static void __net_exit bpf_exit_net(struct net *net)
 405 {
 406         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 407
 408         tc_action_net_exit(tn);
 409 }
 410
 411 static struct pernet_operations bpf_net_ops = {
 412         .init = bpf_init_net,
 413         .exit = bpf_exit_net,
 414         .id   = &bpf_net_id,
 415         .size = sizeof(struct tc_action_net),
 416 };
 417
 418 static int __init bpf_init_module(void)
 419 {
 420         return tcf_register_action(&act_bpf_ops, &bpf_net_ops);
 421 }
 422
 423 static void __exit bpf_cleanup_module(void)
 424 {
 425         tcf_unregister_action(&act_bpf_ops, &bpf_net_ops);
 426 }
 427
 428 module_init(bpf_init_module);
 429 module_exit(bpf_cleanup_module);
 430
 431 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 432 MODULE_DESCRIPTION("TC BPF based action");
 433 MODULE_LICENSE("GPL v2");