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                 filter_res = BPF_PROG_RUN(filter, skb);
  57                 __skb_pull(skb, skb->mac_len);
  58         } else {
  59                 filter_res = BPF_PROG_RUN(filter, skb);
  60         }
  61         rcu_read_unlock();
  62
  63         /* A BPF program may overwrite the default action opcode.
  64          * Similarly as in cls_bpf, if filter_res == -1 we use the
  65          * default action specified from tc.
  66          *
  67          * In case a different well-known TC_ACT opcode has been
  68          * returned, it will overwrite the default one.
  69          *
  70          * For everything else that is unkown, TC_ACT_UNSPEC is
  71          * returned.
  72          */
  73         switch (filter_res) {
  74         case TC_ACT_PIPE:
  75         case TC_ACT_RECLASSIFY:
  76         case TC_ACT_OK:
  77         case TC_ACT_REDIRECT:
  78                 action = filter_res;
  79                 break;
  80         case TC_ACT_SHOT:
  81                 action = filter_res;
  82                 qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats));
  83                 break;
  84         case TC_ACT_UNSPEC:
  85                 action = prog->tcf_action;
  86                 break;
  87         default:
  88                 action = TC_ACT_UNSPEC;
  89                 break;
  90         }
  91
  92         return action;
  93 }
  94
  95 static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog)
  96 {
  97         return !prog->bpf_ops;
  98 }
  99
 100 static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog,
 101                                  struct sk_buff *skb)
 102 {
 103         struct nlattr *nla;
 104
 105         if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops))
 106                 return -EMSGSIZE;
 107
 108         nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops *
 109                           sizeof(struct sock_filter));
 110         if (nla == NULL)
 111                 return -EMSGSIZE;
 112
 113         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 114
 115         return 0;
 116 }
 117
 118 static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog,
 119                                   struct sk_buff *skb)
 120 {
 121         if (nla_put_u32(skb, TCA_ACT_BPF_FD, prog->bpf_fd))
 122                 return -EMSGSIZE;
 123
 124         if (prog->bpf_name &&
 125             nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name))
 126                 return -EMSGSIZE;
 127
 128         return 0;
 129 }
 130
 131 static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act,
 132                         int bind, int ref)
 133 {
 134         unsigned char *tp = skb_tail_pointer(skb);
 135         struct tcf_bpf *prog = act->priv;
 136         struct tc_act_bpf opt = {
 137                 .index   = prog->tcf_index,
 138                 .refcnt  = prog->tcf_refcnt - ref,
 139                 .bindcnt = prog->tcf_bindcnt - bind,
 140                 .action  = prog->tcf_action,
 141         };
 142         struct tcf_t tm;
 143         int ret;
 144
 145         if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt))
 146                 goto nla_put_failure;
 147
 148         if (tcf_bpf_is_ebpf(prog))
 149                 ret = tcf_bpf_dump_ebpf_info(prog, skb);
 150         else
 151                 ret = tcf_bpf_dump_bpf_info(prog, skb);
 152         if (ret)
 153                 goto nla_put_failure;
 154
 155         tm.install = jiffies_to_clock_t(jiffies - prog->tcf_tm.install);
 156         tm.lastuse = jiffies_to_clock_t(jiffies - prog->tcf_tm.lastuse);
 157         tm.expires = jiffies_to_clock_t(prog->tcf_tm.expires);
 158
 159         if (nla_put(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm))
 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, .len = ACT_BPF_NAME_LEN },
 173         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 174         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 175                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 176 };
 177
 178 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 179 {
 180         struct sock_filter *bpf_ops;
 181         struct sock_fprog_kern fprog_tmp;
 182         struct bpf_prog *fp;
 183         u16 bpf_size, bpf_num_ops;
 184         int ret;
 185
 186         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 187         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 188                 return -EINVAL;
 189
 190         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 191         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 192                 return -EINVAL;
 193
 194         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 195         if (bpf_ops == NULL)
 196                 return -ENOMEM;
 197
 198         memcpy(bpf_ops, nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size);
 199
 200         fprog_tmp.len = bpf_num_ops;
 201         fprog_tmp.filter = bpf_ops;
 202
 203         ret = bpf_prog_create(&fp, &fprog_tmp);
 204         if (ret < 0) {
 205                 kfree(bpf_ops);
 206                 return ret;
 207         }
 208
 209         cfg->bpf_ops = bpf_ops;
 210         cfg->bpf_num_ops = bpf_num_ops;
 211         cfg->filter = fp;
 212         cfg->is_ebpf = false;
 213
 214         return 0;
 215 }
 216
 217 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 218 {
 219         struct bpf_prog *fp;
 220         char *name = NULL;
 221         u32 bpf_fd;
 222
 223         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 224
 225         fp = bpf_prog_get(bpf_fd);
 226         if (IS_ERR(fp))
 227                 return PTR_ERR(fp);
 228
 229         if (fp->type != BPF_PROG_TYPE_SCHED_ACT) {
 230                 bpf_prog_put(fp);
 231                 return -EINVAL;
 232         }
 233
 234         if (tb[TCA_ACT_BPF_NAME]) {
 235                 name = kmemdup(nla_data(tb[TCA_ACT_BPF_NAME]),
 236                                nla_len(tb[TCA_ACT_BPF_NAME]),
 237                                GFP_KERNEL);
 238                 if (!name) {
 239                         bpf_prog_put(fp);
 240                         return -ENOMEM;
 241                 }
 242         }
 243
 244         cfg->bpf_fd = bpf_fd;
 245         cfg->bpf_name = name;
 246         cfg->filter = fp;
 247         cfg->is_ebpf = true;
 248
 249         return 0;
 250 }
 251
 252 static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg)
 253 {
 254         if (cfg->is_ebpf)
 255                 bpf_prog_put(cfg->filter);
 256         else
 257                 bpf_prog_destroy(cfg->filter);
 258
 259         kfree(cfg->bpf_ops);
 260         kfree(cfg->bpf_name);
 261 }
 262
 263 static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
 264                                   struct tcf_bpf_cfg *cfg)
 265 {
 266         cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
 267         /* updates to prog->filter are prevented, since it's called either
 268          * with rtnl lock or during final cleanup in rcu callback
 269          */
 270         cfg->filter = rcu_dereference_protected(prog->filter, 1);
 271
 272         cfg->bpf_ops = prog->bpf_ops;
 273         cfg->bpf_name = prog->bpf_name;
 274 }
 275
 276 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 277                         struct nlattr *est, struct tc_action *act,
 278                         int replace, int bind)
 279 {
 280         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 281         struct nlattr *tb[TCA_ACT_BPF_MAX + 1];
 282         struct tcf_bpf_cfg cfg, old;
 283         struct tc_act_bpf *parm;
 284         struct tcf_bpf *prog;
 285         bool is_bpf, is_ebpf;
 286         int ret, res = 0;
 287
 288         if (!nla)
 289                 return -EINVAL;
 290
 291         ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy);
 292         if (ret < 0)
 293                 return ret;
 294
 295         if (!tb[TCA_ACT_BPF_PARMS])
 296                 return -EINVAL;
 297
 298         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 299
 300         if (!tcf_hash_check(tn, parm->index, act, bind)) {
 301                 ret = tcf_hash_create(tn, parm->index, est, act,
 302                                       sizeof(*prog), bind, true);
 303                 if (ret < 0)
 304                         return ret;
 305
 306                 res = ACT_P_CREATED;
 307         } else {
 308                 /* Don't override defaults. */
 309                 if (bind)
 310                         return 0;
 311
 312                 tcf_hash_release(act, bind);
 313                 if (!replace)
 314                         return -EEXIST;
 315         }
 316
 317         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 318         is_ebpf = tb[TCA_ACT_BPF_FD];
 319
 320         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) {
 321                 ret = -EINVAL;
 322                 goto out;
 323         }
 324
 325         memset(&cfg, 0, sizeof(cfg));
 326
 327         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 328                        tcf_bpf_init_from_efd(tb, &cfg);
 329         if (ret < 0)
 330                 goto out;
 331
 332         prog = to_bpf(act);
 333         ASSERT_RTNL();
 334
 335         if (res != ACT_P_CREATED)
 336                 tcf_bpf_prog_fill_cfg(prog, &old);
 337
 338         prog->bpf_ops = cfg.bpf_ops;
 339         prog->bpf_name = cfg.bpf_name;
 340
 341         if (cfg.bpf_num_ops)
 342                 prog->bpf_num_ops = cfg.bpf_num_ops;
 343         if (cfg.bpf_fd)
 344                 prog->bpf_fd = cfg.bpf_fd;
 345
 346         prog->tcf_action = parm->action;
 347         rcu_assign_pointer(prog->filter, cfg.filter);
 348
 349         if (res == ACT_P_CREATED) {
 350                 tcf_hash_insert(tn, act);
 351         } else {
 352                 /* make sure the program being replaced is no longer executing */
 353                 synchronize_rcu();
 354                 tcf_bpf_cfg_cleanup(&old);
 355         }
 356
 357         return res;
 358 out:
 359         if (res == ACT_P_CREATED)
 360                 tcf_hash_cleanup(act, est);
 361
 362         return ret;
 363 }
 364
 365 static void tcf_bpf_cleanup(struct tc_action *act, int bind)
 366 {
 367         struct tcf_bpf_cfg tmp;
 368
 369         tcf_bpf_prog_fill_cfg(act->priv, &tmp);
 370         tcf_bpf_cfg_cleanup(&tmp);
 371 }
 372
 373 static int tcf_bpf_walker(struct net *net, struct sk_buff *skb,
 374                           struct netlink_callback *cb, int type,
 375                           struct tc_action *a)
 376 {
 377         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 378
 379         return tcf_generic_walker(tn, skb, cb, type, a);
 380 }
 381
 382 static int tcf_bpf_search(struct net *net, struct tc_action *a, u32 index)
 383 {
 384         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 385
 386         return tcf_hash_search(tn, a, index);
 387 }
 388
 389 static struct tc_action_ops act_bpf_ops __read_mostly = {
 390         .kind           =       "bpf",
 391         .type           =       TCA_ACT_BPF,
 392         .owner          =       THIS_MODULE,
 393         .act            =       tcf_bpf,
 394         .dump           =       tcf_bpf_dump,
 395         .cleanup        =       tcf_bpf_cleanup,
 396         .init           =       tcf_bpf_init,
 397         .walk           =       tcf_bpf_walker,
 398         .lookup         =       tcf_bpf_search,
 399 };
 400
 401 static __net_init int bpf_init_net(struct net *net)
 402 {
 403         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 404
 405         return tc_action_net_init(tn, &act_bpf_ops, BPF_TAB_MASK);
 406 }
 407
 408 static void __net_exit bpf_exit_net(struct net *net)
 409 {
 410         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 411
 412         tc_action_net_exit(tn);
 413 }
 414
 415 static struct pernet_operations bpf_net_ops = {
 416         .init = bpf_init_net,
 417         .exit = bpf_exit_net,
 418         .id   = &bpf_net_id,
 419         .size = sizeof(struct tc_action_net),
 420 };
 421
 422 static int __init bpf_init_module(void)
 423 {
 424         return tcf_register_action(&act_bpf_ops, &bpf_net_ops);
 425 }
 426
 427 static void __exit bpf_cleanup_module(void)
 428 {
 429         tcf_unregister_action(&act_bpf_ops, &bpf_net_ops);
 430 }
 431
 432 module_init(bpf_init_module);
 433 module_exit(bpf_cleanup_module);
 434
 435 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 436 MODULE_DESCRIPTION("TC BPF based action");
 437 MODULE_LICENSE("GPL v2");