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