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         tm.install = jiffies_to_clock_t(jiffies - prog->tcf_tm.install);
 158         tm.lastuse = jiffies_to_clock_t(jiffies - prog->tcf_tm.lastuse);
 159         tm.expires = jiffies_to_clock_t(prog->tcf_tm.expires);
 160
 161         if (nla_put_64bit(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm,
 162                           TCA_ACT_BPF_PAD))
 163                 goto nla_put_failure;
 164
 165         return skb->len;
 166
 167 nla_put_failure:
 168         nlmsg_trim(skb, tp);
 169         return -1;
 170 }
 171
 172 static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = {
 173         [TCA_ACT_BPF_PARMS]     = { .len = sizeof(struct tc_act_bpf) },
 174         [TCA_ACT_BPF_FD]        = { .type = NLA_U32 },
 175         [TCA_ACT_BPF_NAME]      = { .type = NLA_NUL_STRING, .len = ACT_BPF_NAME_LEN },
 176         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 177         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 178                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 179 };
 180
 181 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 182 {
 183         struct sock_filter *bpf_ops;
 184         struct sock_fprog_kern fprog_tmp;
 185         struct bpf_prog *fp;
 186         u16 bpf_size, bpf_num_ops;
 187         int ret;
 188
 189         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 190         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 191                 return -EINVAL;
 192
 193         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 194         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 195                 return -EINVAL;
 196
 197         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 198         if (bpf_ops == NULL)
 199                 return -ENOMEM;
 200
 201         memcpy(bpf_ops, nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size);
 202
 203         fprog_tmp.len = bpf_num_ops;
 204         fprog_tmp.filter = bpf_ops;
 205
 206         ret = bpf_prog_create(&fp, &fprog_tmp);
 207         if (ret < 0) {
 208                 kfree(bpf_ops);
 209                 return ret;
 210         }
 211
 212         cfg->bpf_ops = bpf_ops;
 213         cfg->bpf_num_ops = bpf_num_ops;
 214         cfg->filter = fp;
 215         cfg->is_ebpf = false;
 216
 217         return 0;
 218 }
 219
 220 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 221 {
 222         struct bpf_prog *fp;
 223         char *name = NULL;
 224         u32 bpf_fd;
 225
 226         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 227
 228         fp = bpf_prog_get(bpf_fd);
 229         if (IS_ERR(fp))
 230                 return PTR_ERR(fp);
 231
 232         if (fp->type != BPF_PROG_TYPE_SCHED_ACT) {
 233                 bpf_prog_put(fp);
 234                 return -EINVAL;
 235         }
 236
 237         if (tb[TCA_ACT_BPF_NAME]) {
 238                 name = kmemdup(nla_data(tb[TCA_ACT_BPF_NAME]),
 239                                nla_len(tb[TCA_ACT_BPF_NAME]),
 240                                GFP_KERNEL);
 241                 if (!name) {
 242                         bpf_prog_put(fp);
 243                         return -ENOMEM;
 244                 }
 245         }
 246
 247         cfg->bpf_fd = bpf_fd;
 248         cfg->bpf_name = name;
 249         cfg->filter = fp;
 250         cfg->is_ebpf = true;
 251
 252         return 0;
 253 }
 254
 255 static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg)
 256 {
 257         if (cfg->is_ebpf)
 258                 bpf_prog_put(cfg->filter);
 259         else
 260                 bpf_prog_destroy(cfg->filter);
 261
 262         kfree(cfg->bpf_ops);
 263         kfree(cfg->bpf_name);
 264 }
 265
 266 static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
 267                                   struct tcf_bpf_cfg *cfg)
 268 {
 269         cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
 270         /* updates to prog->filter are prevented, since it's called either
 271          * with rtnl lock or during final cleanup in rcu callback
 272          */
 273         cfg->filter = rcu_dereference_protected(prog->filter, 1);
 274
 275         cfg->bpf_ops = prog->bpf_ops;
 276         cfg->bpf_name = prog->bpf_name;
 277 }
 278
 279 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 280                         struct nlattr *est, struct tc_action *act,
 281                         int replace, int bind)
 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
 291         if (!nla)
 292                 return -EINVAL;
 293
 294         ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy);
 295         if (ret < 0)
 296                 return ret;
 297
 298         if (!tb[TCA_ACT_BPF_PARMS])
 299                 return -EINVAL;
 300
 301         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 302
 303         if (!tcf_hash_check(tn, parm->index, act, bind)) {
 304                 ret = tcf_hash_create(tn, parm->index, est, act,
 305                                       sizeof(*prog), bind, true);
 306                 if (ret < 0)
 307                         return ret;
 308
 309                 res = ACT_P_CREATED;
 310         } else {
 311                 /* Don't override defaults. */
 312                 if (bind)
 313                         return 0;
 314
 315                 tcf_hash_release(act, bind);
 316                 if (!replace)
 317                         return -EEXIST;
 318         }
 319
 320         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 321         is_ebpf = tb[TCA_ACT_BPF_FD];
 322
 323         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) {
 324                 ret = -EINVAL;
 325                 goto out;
 326         }
 327
 328         memset(&cfg, 0, sizeof(cfg));
 329
 330         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 331                        tcf_bpf_init_from_efd(tb, &cfg);
 332         if (ret < 0)
 333                 goto out;
 334
 335         prog = to_bpf(act);
 336         ASSERT_RTNL();
 337
 338         if (res != ACT_P_CREATED)
 339                 tcf_bpf_prog_fill_cfg(prog, &old);
 340
 341         prog->bpf_ops = cfg.bpf_ops;
 342         prog->bpf_name = cfg.bpf_name;
 343
 344         if (cfg.bpf_num_ops)
 345                 prog->bpf_num_ops = cfg.bpf_num_ops;
 346         if (cfg.bpf_fd)
 347                 prog->bpf_fd = cfg.bpf_fd;
 348
 349         prog->tcf_action = parm->action;
 350         rcu_assign_pointer(prog->filter, cfg.filter);
 351
 352         if (res == ACT_P_CREATED) {
 353                 tcf_hash_insert(tn, act);
 354         } else {
 355                 /* make sure the program being replaced is no longer executing */
 356                 synchronize_rcu();
 357                 tcf_bpf_cfg_cleanup(&old);
 358         }
 359
 360         return res;
 361 out:
 362         if (res == ACT_P_CREATED)
 363                 tcf_hash_cleanup(act, est);
 364
 365         return ret;
 366 }
 367
 368 static void tcf_bpf_cleanup(struct tc_action *act, int bind)
 369 {
 370         struct tcf_bpf_cfg tmp;
 371
 372         tcf_bpf_prog_fill_cfg(act->priv, &tmp);
 373         tcf_bpf_cfg_cleanup(&tmp);
 374 }
 375
 376 static int tcf_bpf_walker(struct net *net, struct sk_buff *skb,
 377                           struct netlink_callback *cb, int type,
 378                           struct tc_action *a)
 379 {
 380         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 381
 382         return tcf_generic_walker(tn, skb, cb, type, a);
 383 }
 384
 385 static int tcf_bpf_search(struct net *net, struct tc_action *a, u32 index)
 386 {
 387         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 388
 389         return tcf_hash_search(tn, a, index);
 390 }
 391
 392 static struct tc_action_ops act_bpf_ops __read_mostly = {
 393         .kind           =       "bpf",
 394         .type           =       TCA_ACT_BPF,
 395         .owner          =       THIS_MODULE,
 396         .act            =       tcf_bpf,
 397         .dump           =       tcf_bpf_dump,
 398         .cleanup        =       tcf_bpf_cleanup,
 399         .init           =       tcf_bpf_init,
 400         .walk           =       tcf_bpf_walker,
 401         .lookup         =       tcf_bpf_search,
 402 };
 403
 404 static __net_init int bpf_init_net(struct net *net)
 405 {
 406         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 407
 408         return tc_action_net_init(tn, &act_bpf_ops, BPF_TAB_MASK);
 409 }
 410
 411 static void __net_exit bpf_exit_net(struct net *net)
 412 {
 413         struct tc_action_net *tn = net_generic(net, bpf_net_id);
 414
 415         tc_action_net_exit(tn);
 416 }
 417
 418 static struct pernet_operations bpf_net_ops = {
 419         .init = bpf_init_net,
 420         .exit = bpf_exit_net,
 421         .id   = &bpf_net_id,
 422         .size = sizeof(struct tc_action_net),
 423 };
 424
 425 static int __init bpf_init_module(void)
 426 {
 427         return tcf_register_action(&act_bpf_ops, &bpf_net_ops);
 428 }
 429
 430 static void __exit bpf_cleanup_module(void)
 431 {
 432         tcf_unregister_action(&act_bpf_ops, &bpf_net_ops);
 433 }
 434
 435 module_init(bpf_init_module);
 436 module_exit(bpf_cleanup_module);
 437
 438 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 439 MODULE_DESCRIPTION("TC BPF based action");
 440 MODULE_LICENSE("GPL v2");