kernel/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         char *bpf_name;
  31         u32 bpf_fd;
  32         u16 bpf_num_ops;
  33 };
  34
  35 static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act,
  36                    struct tcf_result *res)
  37 {
  38         struct tcf_bpf *prog = act->priv;
  39         int action, filter_res;
  40
  41         if (unlikely(!skb_mac_header_was_set(skb)))
  42                 return TC_ACT_UNSPEC;
  43
  44         spin_lock(&prog->tcf_lock);
  45
  46         prog->tcf_tm.lastuse = jiffies;
  47         bstats_update(&prog->tcf_bstats, skb);
  48
  49         /* Needed here for accessing maps. */
  50         rcu_read_lock();
  51         filter_res = BPF_PROG_RUN(prog->filter, skb);
  52         rcu_read_unlock();
  53
  54         /* A BPF program may overwrite the default action opcode.
  55          * Similarly as in cls_bpf, if filter_res == -1 we use the
  56          * default action specified from tc.
  57          *
  58          * In case a different well-known TC_ACT opcode has been
  59          * returned, it will overwrite the default one.
  60          *
  61          * For everything else that is unkown, TC_ACT_UNSPEC is
  62          * returned.
  63          */
  64         switch (filter_res) {
  65         case TC_ACT_PIPE:
  66         case TC_ACT_RECLASSIFY:
  67         case TC_ACT_OK:
  68                 action = filter_res;
  69                 break;
  70         case TC_ACT_SHOT:
  71                 action = filter_res;
  72                 prog->tcf_qstats.drops++;
  73                 break;
  74         case TC_ACT_UNSPEC:
  75                 action = prog->tcf_action;
  76                 break;
  77         default:
  78                 action = TC_ACT_UNSPEC;
  79                 break;
  80         }
  81
  82         spin_unlock(&prog->tcf_lock);
  83         return action;
  84 }
  85
  86 static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog)
  87 {
  88         return !prog->bpf_ops;
  89 }
  90
  91 static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog,
  92                                  struct sk_buff *skb)
  93 {
  94         struct nlattr *nla;
  95
  96         if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops))
  97                 return -EMSGSIZE;
  98
  99         nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops *
 100                           sizeof(struct sock_filter));
 101         if (nla == NULL)
 102                 return -EMSGSIZE;
 103
 104         memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
 105
 106         return 0;
 107 }
 108
 109 static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog,
 110                                   struct sk_buff *skb)
 111 {
 112         if (nla_put_u32(skb, TCA_ACT_BPF_FD, prog->bpf_fd))
 113                 return -EMSGSIZE;
 114
 115         if (prog->bpf_name &&
 116             nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name))
 117                 return -EMSGSIZE;
 118
 119         return 0;
 120 }
 121
 122 static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act,
 123                         int bind, int ref)
 124 {
 125         unsigned char *tp = skb_tail_pointer(skb);
 126         struct tcf_bpf *prog = act->priv;
 127         struct tc_act_bpf opt = {
 128                 .index   = prog->tcf_index,
 129                 .refcnt  = prog->tcf_refcnt - ref,
 130                 .bindcnt = prog->tcf_bindcnt - bind,
 131                 .action  = prog->tcf_action,
 132         };
 133         struct tcf_t tm;
 134         int ret;
 135
 136         if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt))
 137                 goto nla_put_failure;
 138
 139         if (tcf_bpf_is_ebpf(prog))
 140                 ret = tcf_bpf_dump_ebpf_info(prog, skb);
 141         else
 142                 ret = tcf_bpf_dump_bpf_info(prog, skb);
 143         if (ret)
 144                 goto nla_put_failure;
 145
 146         tm.install = jiffies_to_clock_t(jiffies - prog->tcf_tm.install);
 147         tm.lastuse = jiffies_to_clock_t(jiffies - prog->tcf_tm.lastuse);
 148         tm.expires = jiffies_to_clock_t(prog->tcf_tm.expires);
 149
 150         if (nla_put(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm))
 151                 goto nla_put_failure;
 152
 153         return skb->len;
 154
 155 nla_put_failure:
 156         nlmsg_trim(skb, tp);
 157         return -1;
 158 }
 159
 160 static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = {
 161         [TCA_ACT_BPF_PARMS]     = { .len = sizeof(struct tc_act_bpf) },
 162         [TCA_ACT_BPF_FD]        = { .type = NLA_U32 },
 163         [TCA_ACT_BPF_NAME]      = { .type = NLA_NUL_STRING, .len = ACT_BPF_NAME_LEN },
 164         [TCA_ACT_BPF_OPS_LEN]   = { .type = NLA_U16 },
 165         [TCA_ACT_BPF_OPS]       = { .type = NLA_BINARY,
 166                                     .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
 167 };
 168
 169 static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 170 {
 171         struct sock_filter *bpf_ops;
 172         struct sock_fprog_kern fprog_tmp;
 173         struct bpf_prog *fp;
 174         u16 bpf_size, bpf_num_ops;
 175         int ret;
 176
 177         bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]);
 178         if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
 179                 return -EINVAL;
 180
 181         bpf_size = bpf_num_ops * sizeof(*bpf_ops);
 182         if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS]))
 183                 return -EINVAL;
 184
 185         bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
 186         if (bpf_ops == NULL)
 187                 return -ENOMEM;
 188
 189         memcpy(bpf_ops, nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size);
 190
 191         fprog_tmp.len = bpf_num_ops;
 192         fprog_tmp.filter = bpf_ops;
 193
 194         ret = bpf_prog_create(&fp, &fprog_tmp);
 195         if (ret < 0) {
 196                 kfree(bpf_ops);
 197                 return ret;
 198         }
 199
 200         cfg->bpf_ops = bpf_ops;
 201         cfg->bpf_num_ops = bpf_num_ops;
 202         cfg->filter = fp;
 203
 204         return 0;
 205 }
 206
 207 static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg)
 208 {
 209         struct bpf_prog *fp;
 210         char *name = NULL;
 211         u32 bpf_fd;
 212
 213         bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]);
 214
 215         fp = bpf_prog_get(bpf_fd);
 216         if (IS_ERR(fp))
 217                 return PTR_ERR(fp);
 218
 219         if (fp->type != BPF_PROG_TYPE_SCHED_ACT) {
 220                 bpf_prog_put(fp);
 221                 return -EINVAL;
 222         }
 223
 224         if (tb[TCA_ACT_BPF_NAME]) {
 225                 name = kmemdup(nla_data(tb[TCA_ACT_BPF_NAME]),
 226                                nla_len(tb[TCA_ACT_BPF_NAME]),
 227                                GFP_KERNEL);
 228                 if (!name) {
 229                         bpf_prog_put(fp);
 230                         return -ENOMEM;
 231                 }
 232         }
 233
 234         cfg->bpf_fd = bpf_fd;
 235         cfg->bpf_name = name;
 236         cfg->filter = fp;
 237
 238         return 0;
 239 }
 240
 241 static int tcf_bpf_init(struct net *net, struct nlattr *nla,
 242                         struct nlattr *est, struct tc_action *act,
 243                         int replace, int bind)
 244 {
 245         struct nlattr *tb[TCA_ACT_BPF_MAX + 1];
 246         struct tc_act_bpf *parm;
 247         struct tcf_bpf *prog;
 248         struct tcf_bpf_cfg cfg;
 249         bool is_bpf, is_ebpf;
 250         int ret;
 251
 252         if (!nla)
 253                 return -EINVAL;
 254
 255         ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy);
 256         if (ret < 0)
 257                 return ret;
 258
 259         is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS];
 260         is_ebpf = tb[TCA_ACT_BPF_FD];
 261
 262         if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf) ||
 263             !tb[TCA_ACT_BPF_PARMS])
 264                 return -EINVAL;
 265
 266         parm = nla_data(tb[TCA_ACT_BPF_PARMS]);
 267
 268         memset(&cfg, 0, sizeof(cfg));
 269
 270         ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) :
 271                        tcf_bpf_init_from_efd(tb, &cfg);
 272         if (ret < 0)
 273                 return ret;
 274
 275         if (!tcf_hash_check(parm->index, act, bind)) {
 276                 ret = tcf_hash_create(parm->index, est, act,
 277                                       sizeof(*prog), bind);
 278                 if (ret < 0)
 279                         goto destroy_fp;
 280
 281                 ret = ACT_P_CREATED;
 282         } else {
 283                 /* Don't override defaults. */
 284                 if (bind)
 285                         goto destroy_fp;
 286
 287                 tcf_hash_release(act, bind);
 288                 if (!replace) {
 289                         ret = -EEXIST;
 290                         goto destroy_fp;
 291                 }
 292         }
 293
 294         prog = to_bpf(act);
 295         spin_lock_bh(&prog->tcf_lock);
 296
 297         prog->bpf_ops = cfg.bpf_ops;
 298         prog->bpf_name = cfg.bpf_name;
 299
 300         if (cfg.bpf_num_ops)
 301                 prog->bpf_num_ops = cfg.bpf_num_ops;
 302         if (cfg.bpf_fd)
 303                 prog->bpf_fd = cfg.bpf_fd;
 304
 305         prog->tcf_action = parm->action;
 306         prog->filter = cfg.filter;
 307
 308         spin_unlock_bh(&prog->tcf_lock);
 309
 310         if (ret == ACT_P_CREATED)
 311                 tcf_hash_insert(act);
 312
 313         return ret;
 314
 315 destroy_fp:
 316         if (is_ebpf)
 317                 bpf_prog_put(cfg.filter);
 318         else
 319                 bpf_prog_destroy(cfg.filter);
 320
 321         kfree(cfg.bpf_ops);
 322         kfree(cfg.bpf_name);
 323
 324         return ret;
 325 }
 326
 327 static void tcf_bpf_cleanup(struct tc_action *act, int bind)
 328 {
 329         const struct tcf_bpf *prog = act->priv;
 330
 331         if (tcf_bpf_is_ebpf(prog))
 332                 bpf_prog_put(prog->filter);
 333         else
 334                 bpf_prog_destroy(prog->filter);
 335 }
 336
 337 static struct tc_action_ops act_bpf_ops __read_mostly = {
 338         .kind           =       "bpf",
 339         .type           =       TCA_ACT_BPF,
 340         .owner          =       THIS_MODULE,
 341         .act            =       tcf_bpf,
 342         .dump           =       tcf_bpf_dump,
 343         .cleanup        =       tcf_bpf_cleanup,
 344         .init           =       tcf_bpf_init,
 345 };
 346
 347 static int __init bpf_init_module(void)
 348 {
 349         return tcf_register_action(&act_bpf_ops, BPF_TAB_MASK);
 350 }
 351
 352 static void __exit bpf_cleanup_module(void)
 353 {
 354         tcf_unregister_action(&act_bpf_ops);
 355 }
 356
 357 module_init(bpf_init_module);
 358 module_exit(bpf_cleanup_module);
 359
 360 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
 361 MODULE_DESCRIPTION("TC BPF based action");
 362 MODULE_LICENSE("GPL v2");