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