sched/act_bpf.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/kernel.h>
   9 #include <linux/skbuff.h>
  10 #include <linux/rtnetlink.h>
  11 #include <linux/filter.h>
  12 #include <linux/bpf.h>
  13
  14 #include <net/netlink.h>
  15 #include <net/sock.h>
  16 #include <net/pkt_sched.h>
  17 #include <net/pkt_cls.h>
  18
  19 #include <linux/tc_act/tc_bpf.h>
  20 #include <net/tc_act/tc_bpf.h>
  21
  22 #define ACT_BPF_NAME_LEN        256
  23
  24 struct tcf_bpf_cfg {
  25         struct bpf_prog *filter;
  26         struct sock_filter *bpf_ops;
  27         const char *bpf_name;
  28         u16 bpf_num_ops;
  29         bool is_ebpf;
  30 };
  31
  32 static struct tc_action_ops act_bpf_ops;
  33
  34 static int tcf_bpf_act(struct sk_buff *skb, const struct tc_action *act,
  35                        struct tcf_result *res)
  36 {
  37         bool at_ingress = skb_at_tc_ingress(skb);
  38         struct tcf_bpf *prog = to_bpf(act);
  39         struct bpf_prog *filter;
  40         int action, filter_res;
  41
  42         tcf_lastuse_update(&prog->tcf_tm);
  43         bstats_update(this_cpu_ptr(prog->common.cpu_bstats), skb);
  44
  45         filter = rcu_dereference(prog->filter);
  46         if (at_ingress) {
  47                 __skb_push(skb, skb->mac_len);
  48                 bpf_compute_data_pointers(skb);
  49                 filter_res = bpf_prog_run(filter, skb);
  50                 __skb_pull(skb, skb->mac_len);
  51         } else {
  52                 bpf_compute_data_pointers(skb);
  53                 filter_res = bpf_prog_run(filter, skb);
  54         }
  55         if (unlikely(!skb->tstamp && skb->mono_delivery_time))
  56                 skb->mono_delivery_time = 0;
  57         if (skb_sk_is_prefetched(skb) && filter_res != TC_ACT_OK)
  58                 skb_orphan(skb);
  59
  60         /* A BPF program may overwrite the default action opcode.
  61          * Similarly as in cls_bpf, if filter_res == -1 we use the
  62          * default action specified from tc.
  63          *
  64          * In case a different well-known TC_ACT opcode has been
  65          * returned, it will overwrite the default one.
  66          *
  67          * For everything else that is unknown, TC_ACT_UNSPEC is
  68          * returned.
  69          */
  70         switch (filter_res) {
  71         case TC_ACT_PIPE:
  72         case TC_ACT_RECLASSIFY:
  73         case TC_ACT_OK:
  74         case TC_ACT_REDIRECT:
  75                 action = filter_res;
  76                 break;
  77         case TC_ACT_SHOT:
  78                 action = filter_res;
  79                 qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats));
  80                 break;
  81         case TC_ACT_UNSPEC:
  82                 action = prog->tcf_action;
  83                 break;
  84         default:
  85                 action = TC_ACT_UNSPEC;
  86                 break;
  87         }
  88
  89         return action;
  90 }
  91
  92 static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog)
  93 {
  94         return !prog->bpf_ops;
  95 }
  96
  97 static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog,
  98                                  struct sk_buff *skb)
  99 {
 100         struct nlattr *nla;
 101
 102         if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops))
 103                 return -EMSGSIZE;
 104
 105         nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops *
 106                           sizeof(struct sock_filter));
 107         if (nla == NULL)
 108                 return -EMSGSIZE;
 109
 110         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 111
 112         return 0;
 113 }
 114
 115 static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog,
 116                                   struct sk_buff *skb)
 117 {
 118         struct nlattr *nla;
 119
 120         if (prog->bpf_name &&
 121             nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name))
 122                 return -EMSGSIZE;
 123
 124         if (nla_put_u32(skb, TCA_ACT_BPF_ID, prog->filter->aux->id))
 125                 return -EMSGSIZE;
 126
 127         nla = nla_reserve(skb, TCA_ACT_BPF_TAG, sizeof(prog->filter->tag));
 128         if (nla == NULL)
 129                 return -EMSGSIZE;
 130
 131         memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
 132
 133         return 0;
 134 }
 135
 136 static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act,
 137                         int bind, int ref)
 138 {
 139         unsigned char *tp = skb_tail_pointer(skb);
 140         struct tcf_bpf *prog = to_bpf(act);
 141         struct tc_act_bpf opt = {
 142                 .index   = prog->tcf_index,
 143                 .refcnt  = refcount_read(&prog->tcf_refcnt) - ref,
 144                 .bindcnt = atomic_read(&prog->tcf_bindcnt) - bind,
 145         };
 146         struct tcf_t tm;
 147         int ret;
 148
 149         spin_lock_bh(&prog->tcf_lock);
 150         opt.action = prog->tcf_action;
 151         if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt))
 152                 goto nla_put_failure;
 153
 154         if (tcf_bpf_is_ebpf(prog))
 155                 ret = tcf_bpf_dump_ebpf_info(prog, skb);
 156         else
 157                 ret = tcf_bpf_dump_bpf_info(prog, skb);
 158         if (ret)
 159                 goto nla_put_failure;
 160
 161         tcf_tm_dump(&tm, &prog->tcf_tm);
 162         if (nla_put_64bit(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm,
 163                           TCA_ACT_BPF_PAD))
 164                 goto nla_put_failure;
 165
 166         spin_unlock_bh(&prog->tcf_lock);
 167         return skb->len;
 168
 169 nla_put_failure:
 170         spin_unlock_bh(&prog->tcf_lock);
 171         nlmsg_trim(skb, tp);
 172         return -1;
 173 }
 174
 175 static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = {
 176         [TCA_ACT_BPF_PARMS]     = { .len = sizeof(struct tc_act_bpf) },
 177         [TCA_ACT_BPF_FD]        = { .type = NLA_U32 },
 178         [TCA_ACT_BPF_NAME]      = { .type = NLA_NUL_STRING,
 179                                     .len = ACT_BPF_NAME_LEN },
 180         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 181         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 182                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 183 };
 184
 185 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 186 {
 187         struct sock_filter *bpf_ops;
 188         struct sock_fprog_kern fprog_tmp;
 189         struct bpf_prog *fp;
 190         u16 bpf_size, bpf_num_ops;
 191         int ret;
 192
 193         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 194         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 195                 return -EINVAL;
 196
 197         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 198         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 199                 return -EINVAL;
 200
 201         bpf_ops = kmemdup(nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size, GFP_KERNEL);
 202         if (bpf_ops == NULL)
 203                 return -ENOMEM;
 204
 205         fprog_tmp.len = bpf_num_ops;
 206         fprog_tmp.filter = bpf_ops;
 207
 208         ret = bpf_prog_create(&fp, &fprog_tmp);
 209         if (ret < 0) {
 210                 kfree(bpf_ops);
 211                 return ret;
 212         }
 213
 214         cfg->bpf_ops = bpf_ops;
 215         cfg->bpf_num_ops = bpf_num_ops;
 216         cfg->filter = fp;
 217         cfg->is_ebpf = false;
 218
 219         return 0;
 220 }
 221
 222 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 223 {
 224         struct bpf_prog *fp;
 225         char *name = NULL;
 226         u32 bpf_fd;
 227
 228         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 229
 230         fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_ACT);
 231         if (IS_ERR(fp))
 232                 return PTR_ERR(fp);
 233
 234         if (tb[TCA_ACT_BPF_NAME]) {
 235                 name = nla_memdup(tb[TCA_ACT_BPF_NAME], GFP_KERNEL);
 236                 if (!name) {
 237                         bpf_prog_put(fp);
 238                         return -ENOMEM;
 239                 }
 240         }
 241
 242         cfg->bpf_name = name;
 243         cfg->filter = fp;
 244         cfg->is_ebpf = true;
 245
 246         return 0;
 247 }
 248
 249 static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg)
 250 {
 251         struct bpf_prog *filter = cfg->filter;
 252
 253         if (filter) {
 254                 if (cfg->is_ebpf)
 255                         bpf_prog_put(filter);
 256                 else
 257                         bpf_prog_destroy(filter);
 258         }
 259
 260         kfree(cfg->bpf_ops);
 261         kfree(cfg->bpf_name);
 262 }
 263
 264 static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog,
 265                                   struct tcf_bpf_cfg *cfg)
 266 {
 267         cfg->is_ebpf = tcf_bpf_is_ebpf(prog);
 268         /* updates to prog->filter are prevented, since it's called either
 269          * with tcf lock or during final cleanup in rcu callback
 270          */
 271         cfg->filter = rcu_dereference_protected(prog->filter, 1);
 272
 273         cfg->bpf_ops = prog->bpf_ops;
 274         cfg->bpf_name = prog->bpf_name;
 275 }
 276
 277 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 278                         struct nlattr *est, struct tc_action **act,
 279                         struct tcf_proto *tp, u32 flags,
 280                         struct netlink_ext_ack *extack)
 281 {
 282         struct tc_action_net *tn = net_generic(net, act_bpf_ops.net_id);
 283         bool bind = flags & TCA_ACT_FLAGS_BIND;
 284         struct nlattr *tb[TCA_ACT_BPF_MAX + 1];
 285         struct tcf_chain *goto_ch = NULL;
 286         struct tcf_bpf_cfg cfg, old;
 287         struct tc_act_bpf *parm;
 288         struct tcf_bpf *prog;
 289         bool is_bpf, is_ebpf;
 290         int ret, res = 0;
 291         u32 index;
 292
 293         if (!nla)
 294                 return -EINVAL;
 295
 296         ret = nla_parse_nested_deprecated(tb, TCA_ACT_BPF_MAX, nla,
 297                                           act_bpf_policy, NULL);
 298         if (ret < 0)
 299                 return ret;
 300
 301         if (!tb[TCA_ACT_BPF_PARMS])
 302                 return -EINVAL;
 303
 304         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 305         index = parm->index;
 306         ret = tcf_idr_check_alloc(tn, &index, act, bind);
 307         if (!ret) {
 308                 ret = tcf_idr_create(tn, index, est, act,
 309                                      &act_bpf_ops, bind, true, flags);
 310                 if (ret < 0) {
 311                         tcf_idr_cleanup(tn, index);
 312                         return ret;
 313                 }
 314
 315                 res = ACT_P_CREATED;
 316         } else if (ret > 0) {
 317                 /* Don't override defaults. */
 318                 if (bind)
 319                         return 0;
 320
 321                 if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
 322                         tcf_idr_release(*act, bind);
 323                         return -EEXIST;
 324                 }
 325         } else {
 326                 return ret;
 327         }
 328
 329         ret = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
 330         if (ret < 0)
 331                 goto release_idr;
 332
 333         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 334         is_ebpf = tb[TCA_ACT_BPF_FD];
 335
 336         if (is_bpf == is_ebpf) {
 337                 ret = -EINVAL;
 338                 goto put_chain;
 339         }
 340
 341         memset(&cfg, 0, sizeof(cfg));
 342
 343         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 344                        tcf_bpf_init_from_efd(tb, &cfg);
 345         if (ret < 0)
 346                 goto put_chain;
 347
 348         prog = to_bpf(*act);
 349
 350         spin_lock_bh(&prog->tcf_lock);
 351         if (res != ACT_P_CREATED)
 352                 tcf_bpf_prog_fill_cfg(prog, &old);
 353
 354         prog->bpf_ops = cfg.bpf_ops;
 355         prog->bpf_name = cfg.bpf_name;
 356
 357         if (cfg.bpf_num_ops)
 358                 prog->bpf_num_ops = cfg.bpf_num_ops;
 359
 360         goto_ch = tcf_action_set_ctrlact(*act, parm->action, goto_ch);
 361         rcu_assign_pointer(prog->filter, cfg.filter);
 362         spin_unlock_bh(&prog->tcf_lock);
 363
 364         if (goto_ch)
 365                 tcf_chain_put_by_act(goto_ch);
 366
 367         if (res != ACT_P_CREATED) {
 368                 /* make sure the program being replaced is no longer executing */
 369                 synchronize_rcu();
 370                 tcf_bpf_cfg_cleanup(&old);
 371         }
 372
 373         return res;
 374
 375 put_chain:
 376         if (goto_ch)
 377                 tcf_chain_put_by_act(goto_ch);
 378
 379 release_idr:
 380         tcf_idr_release(*act, bind);
 381         return ret;
 382 }
 383
 384 static void tcf_bpf_cleanup(struct tc_action *act)
 385 {
 386         struct tcf_bpf_cfg tmp;
 387
 388         tcf_bpf_prog_fill_cfg(to_bpf(act), &tmp);
 389         tcf_bpf_cfg_cleanup(&tmp);
 390 }
 391
 392 static struct tc_action_ops act_bpf_ops __read_mostly = {
 393         .kind           =       "bpf",
 394         .id             =       TCA_ID_BPF,
 395         .owner          =       THIS_MODULE,
 396         .act            =       tcf_bpf_act,
 397         .dump           =       tcf_bpf_dump,
 398         .cleanup        =       tcf_bpf_cleanup,
 399         .init           =       tcf_bpf_init,
 400         .size           =       sizeof(struct tcf_bpf),
 401 };
 402
 403 static __net_init int bpf_init_net(struct net *net)
 404 {
 405         struct tc_action_net *tn = net_generic(net, act_bpf_ops.net_id);
 406
 407         return tc_action_net_init(net, tn, &act_bpf_ops);
 408 }
 409
 410 static void __net_exit bpf_exit_net(struct list_head *net_list)
 411 {
 412         tc_action_net_exit(net_list, act_bpf_ops.net_id);
 413 }
 414
 415 static struct pernet_operations bpf_net_ops = {
 416         .init = bpf_init_net,
 417         .exit_batch = bpf_exit_net,
 418         .id   = &act_bpf_ops.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");