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 ACT_BPF_NAME_LEN        256
  25
  26 struct tcf_bpf_cfg {
  27         struct bpf_prog *filter;
  28         struct sock_filter *bpf_ops;
  29         const char *bpf_name;
  30         u16 bpf_num_ops;
  31         bool is_ebpf;
  32 };
  33
  34 static unsigned int bpf_net_id;
  35 static struct tc_action_ops act_bpf_ops;
  36
  37 static int tcf_bpf_act(struct sk_buff *skb, const struct tc_action *act,
  38                        struct tcf_result *res)
  39 {
  40         bool at_ingress = skb_at_tc_ingress(skb);
  41         struct tcf_bpf *prog = to_bpf(act);
  42         struct bpf_prog *filter;
  43         int action, filter_res;
  44
  45         tcf_lastuse_update(&prog->tcf_tm);
  46         bstats_cpu_update(this_cpu_ptr(prog->common.cpu_bstats), skb);
  47
  48         rcu_read_lock();
  49         filter = rcu_dereference(prog->filter);
  50         if (at_ingress) {
  51                 __skb_push(skb, skb->mac_len);
  52                 bpf_compute_data_pointers(skb);
  53                 filter_res = BPF_PROG_RUN(filter, skb);
  54                 __skb_pull(skb, skb->mac_len);
  55         } else {
  56                 bpf_compute_data_pointers(skb);
  57                 filter_res = BPF_PROG_RUN(filter, skb);
  58         }
  59         rcu_read_unlock();
  60
  61         /* A BPF program may overwrite the default action opcode.
  62          * Similarly as in cls_bpf, if filter_res == -1 we use the
  63          * default action specified from tc.
  64          *
  65          * In case a different well-known TC_ACT opcode has been
  66          * returned, it will overwrite the default one.
  67          *
  68          * For everything else that is unkown, TC_ACT_UNSPEC is
  69          * returned.
  70          */
  71         switch (filter_res) {
  72         case TC_ACT_PIPE:
  73         case TC_ACT_RECLASSIFY:
  74         case TC_ACT_OK:
  75         case TC_ACT_REDIRECT:
  76                 action = filter_res;
  77                 break;
  78         case TC_ACT_SHOT:
  79                 action = filter_res;
  80                 qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats));
  81                 break;
  82         case TC_ACT_UNSPEC:
  83                 action = prog->tcf_action;
  84                 break;
  85         default:
  86                 action = TC_ACT_UNSPEC;
  87                 break;
  88         }
  89
  90         return action;
  91 }
  92
  93 static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog)
  94 {
  95         return !prog->bpf_ops;
  96 }
  97
  98 static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog,
  99                                  struct sk_buff *skb)
 100 {
 101         struct nlattr *nla;
 102
 103         if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops))
 104                 return -EMSGSIZE;
 105
 106         nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops *
 107                           sizeof(struct sock_filter));
 108         if (nla == NULL)
 109                 return -EMSGSIZE;
 110
 111         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 112
 113         return 0;
 114 }
 115
 116 static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog,
 117                                   struct sk_buff *skb)
 118 {
 119         struct nlattr *nla;
 120
 121         if (prog->bpf_name &&
 122             nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name))
 123                 return -EMSGSIZE;
 124
 125         if (nla_put_u32(skb, TCA_ACT_BPF_ID, prog->filter->aux->id))
 126                 return -EMSGSIZE;
 127
 128         nla = nla_reserve(skb, TCA_ACT_BPF_TAG, sizeof(prog->filter->tag));
 129         if (nla == NULL)
 130                 return -EMSGSIZE;
 131
 132         memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
 133
 134         return 0;
 135 }
 136
 137 static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act,
 138                         int bind, int ref)
 139 {
 140         unsigned char *tp = skb_tail_pointer(skb);
 141         struct tcf_bpf *prog = to_bpf(act);
 142         struct tc_act_bpf opt = {
 143                 .index   = prog->tcf_index,
 144                 .refcnt  = refcount_read(&prog->tcf_refcnt) - ref,
 145                 .bindcnt = atomic_read(&prog->tcf_bindcnt) - bind,
 146         };
 147         struct tcf_t tm;
 148         int ret;
 149
 150         spin_lock_bh(&prog->tcf_lock);
 151         opt.action = prog->tcf_action;
 152         if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt))
 153                 goto nla_put_failure;
 154
 155         if (tcf_bpf_is_ebpf(prog))
 156                 ret = tcf_bpf_dump_ebpf_info(prog, skb);
 157         else
 158                 ret = tcf_bpf_dump_bpf_info(prog, skb);
 159         if (ret)
 160                 goto nla_put_failure;
 161
 162         tcf_tm_dump(&tm, &prog->tcf_tm);
 163         if (nla_put_64bit(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm,
 164                           TCA_ACT_BPF_PAD))
 165                 goto nla_put_failure;
 166
 167         spin_unlock_bh(&prog->tcf_lock);
 168         return skb->len;
 169
 170 nla_put_failure:
 171         spin_unlock_bh(&prog->tcf_lock);
 172         nlmsg_trim(skb, tp);
 173         return -1;
 174 }
 175
 176 static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = {
 177         [TCA_ACT_BPF_PARMS]     = { .len = sizeof(struct tc_act_bpf) },
 178         [TCA_ACT_BPF_FD]        = { .type = NLA_U32 },
 179         [TCA_ACT_BPF_NAME]      = { .type = NLA_NUL_STRING,
 180                                     .len = ACT_BPF_NAME_LEN },
 181         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 182         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 183                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 184 };
 185
 186 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 187 {
 188         struct sock_filter *bpf_ops;
 189         struct sock_fprog_kern fprog_tmp;
 190         struct bpf_prog *fp;
 191         u16 bpf_size, bpf_num_ops;
 192         int ret;
 193
 194         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 195         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 196                 return -EINVAL;
 197
 198         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 199         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 200                 return -EINVAL;
 201
 202         bpf_ops = kmemdup(nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size, GFP_KERNEL);
 203         if (bpf_ops == NULL)
 204                 return -ENOMEM;
 205
 206         fprog_tmp.len = bpf_num_ops;
 207         fprog_tmp.filter = bpf_ops;
 208
 209         ret = bpf_prog_create(&fp, &fprog_tmp);
 210         if (ret < 0) {
 211                 kfree(bpf_ops);
 212                 return ret;
 213         }
 214
 215         cfg->bpf_ops = bpf_ops;
 216         cfg->bpf_num_ops = bpf_num_ops;
 217         cfg->filter = fp;
 218         cfg->is_ebpf = false;
 219
 220         return 0;
 221 }
 222
 223 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 224 {
 225         struct bpf_prog *fp;
 226         char *name = NULL;
 227         u32 bpf_fd;
 228
 229         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 230
 231         fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_ACT);
 232         if (IS_ERR(fp))
 233                 return PTR_ERR(fp);
 234
 235         if (tb[TCA_ACT_BPF_NAME]) {
 236                 name = nla_memdup(tb[TCA_ACT_BPF_NAME], GFP_KERNEL);
 237                 if (!name) {
 238                         bpf_prog_put(fp);
 239                         return -ENOMEM;
 240                 }
 241         }
 242
 243         cfg->bpf_name = name;
 244         cfg->filter = fp;
 245         cfg->is_ebpf = true;
 246
 247         return 0;
 248 }
 249
 250 static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg)
 251 {
 252         struct bpf_prog *filter = cfg->filter;
 253
 254         if (filter) {
 255                 if (cfg->is_ebpf)
 256                         bpf_prog_put(filter);
 257                 else
 258                         bpf_prog_destroy(filter);
 259         }
 260
 261         kfree(cfg->bpf_ops);
 262         kfree(cfg->bpf_name);
 263 }
 264
 265 static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
 266                                   struct tcf_bpf_cfg *cfg)
 267 {
 268         cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
 269         /* updates to prog->filter are prevented, since it's called either
 270          * with tcf lock or during final cleanup in rcu callback
 271          */
 272         cfg->filter = rcu_dereference_protected(prog->filter, 1);
 273
 274         cfg->bpf_ops = prog->bpf_ops;
 275         cfg->bpf_name = prog->bpf_name;
 276 }
 277
 278 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 279                         struct nlattr *est, struct tc_action **act,
 280                         int replace, int bind, bool rtnl_held,
 281                         struct netlink_ext_ack *extack)
 282 {
 283         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 284         struct nlattr *tb[TCA_ACT_BPF_MAX + 1];
 285         struct tcf_bpf_cfg cfg, old;
 286         struct tc_act_bpf *parm;
 287         struct tcf_bpf *prog;
 288         bool is_bpf, is_ebpf;
 289         int ret, res = 0;
 290         u32 index;
 291
 292         if (!nla)
 293                 return -EINVAL;
 294
 295         ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy, NULL);
 296         if (ret < 0)
 297                 return ret;
 298
 299         if (!tb[TCA_ACT_BPF_PARMS])
 300                 return -EINVAL;
 301
 302         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 303         index = parm->index;
 304         ret = tcf_idr_check_alloc(tn, &index, act, bind);
 305         if (!ret) {
 306                 ret = tcf_idr_create(tn, index, est, act,
 307                                      &act_bpf_ops, bind, true);
 308                 if (ret < 0) {
 309                         tcf_idr_cleanup(tn, index);
 310                         return ret;
 311                 }
 312
 313                 res = ACT_P_CREATED;
 314         } else if (ret > 0) {
 315                 /* Don't override defaults. */
 316                 if (bind)
 317                         return 0;
 318
 319                 if (!replace) {
 320                         tcf_idr_release(*act, bind);
 321                         return -EEXIST;
 322                 }
 323         } else {
 324                 return ret;
 325         }
 326
 327         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 328         is_ebpf = tb[TCA_ACT_BPF_FD];
 329
 330         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) {
 331                 ret = -EINVAL;
 332                 goto out;
 333         }
 334
 335         memset(&cfg, 0, sizeof(cfg));
 336
 337         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 338                        tcf_bpf_init_from_efd(tb, &cfg);
 339         if (ret < 0)
 340                 goto out;
 341
 342         prog = to_bpf(*act);
 343
 344         spin_lock_bh(&prog->tcf_lock);
 345         if (res != ACT_P_CREATED)
 346                 tcf_bpf_prog_fill_cfg(prog, &old);
 347
 348         prog->bpf_ops = cfg.bpf_ops;
 349         prog->bpf_name = cfg.bpf_name;
 350
 351         if (cfg.bpf_num_ops)
 352                 prog->bpf_num_ops = cfg.bpf_num_ops;
 353
 354         prog->tcf_action = parm->action;
 355         rcu_assign_pointer(prog->filter, cfg.filter);
 356         spin_unlock_bh(&prog->tcf_lock);
 357
 358         if (res == ACT_P_CREATED) {
 359                 tcf_idr_insert(tn, *act);
 360         } else {
 361                 /* make sure the program being replaced is no longer executing */
 362                 synchronize_rcu();
 363                 tcf_bpf_cfg_cleanup(&old);
 364         }
 365
 366         return res;
 367 out:
 368         tcf_idr_release(*act, bind);
 369
 370         return ret;
 371 }
 372
 373 static void tcf_bpf_cleanup(struct tc_action *act)
 374 {
 375         struct tcf_bpf_cfg tmp;
 376
 377         tcf_bpf_prog_fill_cfg(to_bpf(act), &tmp);
 378         tcf_bpf_cfg_cleanup(&tmp);
 379 }
 380
 381 static int tcf_bpf_walker(struct net *net, struct sk_buff *skb,
 382                           struct netlink_callback *cb, int type,
 383                           const struct tc_action_ops *ops,
 384                           struct netlink_ext_ack *extack)
 385 {
 386         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 387
 388         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 389 }
 390
 391 static int tcf_bpf_search(struct net *net, struct tc_action **a, u32 index,
 392                           struct netlink_ext_ack *extack)
 393 {
 394         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 395
 396         return tcf_idr_search(tn, a, index);
 397 }
 398
 399 static struct tc_action_ops act_bpf_ops __read_mostly = {
 400         .kind           =       "bpf",
 401         .type           =       TCA_ACT_BPF,
 402         .owner          =       THIS_MODULE,
 403         .act            =       tcf_bpf_act,
 404         .dump           =       tcf_bpf_dump,
 405         .cleanup        =       tcf_bpf_cleanup,
 406         .init           =       tcf_bpf_init,
 407         .walk           =       tcf_bpf_walker,
 408         .lookup         =       tcf_bpf_search,
 409         .size           =       sizeof(struct tcf_bpf),
 410 };
 411
 412 static __net_init int bpf_init_net(struct net *net)
 413 {
 414         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 415
 416         return tc_action_net_init(net, tn, &act_bpf_ops);
 417 }
 418
 419 static void __net_exit bpf_exit_net(struct list_head *net_list)
 420 {
 421         tc_action_net_exit(net_list, bpf_net_id);
 422 }
 423
 424 static struct pernet_operations bpf_net_ops = {
 425         .init = bpf_init_net,
 426         .exit_batch = bpf_exit_net,
 427         .id   = &bpf_net_id,
 428         .size = sizeof(struct tc_action_net),
 429 };
 430
 431 static int __init bpf_init_module(void)
 432 {
 433         return tcf_register_action(&act_bpf_ops, &bpf_net_ops);
 434 }
 435
 436 static void __exit bpf_cleanup_module(void)
 437 {
 438         tcf_unregister_action(&act_bpf_ops, &bpf_net_ops);
 439 }
 440
 441 module_init(bpf_init_module);
 442 module_exit(bpf_cleanup_module);
 443
 444 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 445 MODULE_DESCRIPTION("TC BPF based action");
 446 MODULE_LICENSE("GPL v2");