net/core/lwt_bpf.c

   1 /* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of version 2 of the GNU General Public
   5  * License as published by the Free Software Foundation.
   6  *
   7  * This program is distributed in the hope that it will be useful, but
   8  * WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  10  * General Public License for more details.
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/skbuff.h>
  16 #include <linux/types.h>
  17 #include <linux/bpf.h>
  18 #include <net/lwtunnel.h>
  19
  20 struct bpf_lwt_prog {
  21         struct bpf_prog *prog;
  22         char *name;
  23 };
  24
  25 struct bpf_lwt {
  26         struct bpf_lwt_prog in;
  27         struct bpf_lwt_prog out;
  28         struct bpf_lwt_prog xmit;
  29         int family;
  30 };
  31
  32 #define MAX_PROG_NAME 256
  33
  34 static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
  35 {
  36         return (struct bpf_lwt *)lwt->data;
  37 }
  38
  39 #define NO_REDIRECT false
  40 #define CAN_REDIRECT true
  41
  42 static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
  43                        struct dst_entry *dst, bool can_redirect)
  44 {
  45         int ret;
  46
  47         /* Preempt disable is needed to protect per-cpu redirect_info between
  48          * BPF prog and skb_do_redirect(). The call_rcu in bpf_prog_put() and
  49          * access to maps strictly require a rcu_read_lock() for protection,
  50          * mixing with BH RCU lock doesn't work.
  51          */
  52         preempt_disable();
  53         bpf_compute_data_pointers(skb);
  54         ret = bpf_prog_run_save_cb(lwt->prog, skb);
  55
  56         switch (ret) {
  57         case BPF_OK:
  58                 break;
  59
  60         case BPF_REDIRECT:
  61                 if (unlikely(!can_redirect)) {
  62                         pr_warn_once("Illegal redirect return code in prog %s\n",
  63                                      lwt->name ? : "<unknown>");
  64                         ret = BPF_OK;
  65                 } else {
  66                         ret = skb_do_redirect(skb);
  67                         if (ret == 0)
  68                                 ret = BPF_REDIRECT;
  69                 }
  70                 break;
  71
  72         case BPF_DROP:
  73                 kfree_skb(skb);
  74                 ret = -EPERM;
  75                 break;
  76
  77         default:
  78                 pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret);
  79                 kfree_skb(skb);
  80                 ret = -EINVAL;
  81                 break;
  82         }
  83
  84         preempt_enable();
  85
  86         return ret;
  87 }
  88
  89 static int bpf_input(struct sk_buff *skb)
  90 {
  91         struct dst_entry *dst = skb_dst(skb);
  92         struct bpf_lwt *bpf;
  93         int ret;
  94
  95         bpf = bpf_lwt_lwtunnel(dst->lwtstate);
  96         if (bpf->in.prog) {
  97                 ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
  98                 if (ret < 0)
  99                         return ret;
 100         }
 101
 102         if (unlikely(!dst->lwtstate->orig_input)) {
 103                 pr_warn_once("orig_input not set on dst for prog %s\n",
 104                              bpf->out.name);
 105                 kfree_skb(skb);
 106                 return -EINVAL;
 107         }
 108
 109         return dst->lwtstate->orig_input(skb);
 110 }
 111
 112 static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
 113 {
 114         struct dst_entry *dst = skb_dst(skb);
 115         struct bpf_lwt *bpf;
 116         int ret;
 117
 118         bpf = bpf_lwt_lwtunnel(dst->lwtstate);
 119         if (bpf->out.prog) {
 120                 ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
 121                 if (ret < 0)
 122                         return ret;
 123         }
 124
 125         if (unlikely(!dst->lwtstate->orig_output)) {
 126                 pr_warn_once("orig_output not set on dst for prog %s\n",
 127                              bpf->out.name);
 128                 kfree_skb(skb);
 129                 return -EINVAL;
 130         }
 131
 132         return dst->lwtstate->orig_output(net, sk, skb);
 133 }
 134
 135 static int xmit_check_hhlen(struct sk_buff *skb)
 136 {
 137         int hh_len = skb_dst(skb)->dev->hard_header_len;
 138
 139         if (skb_headroom(skb) < hh_len) {
 140                 int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
 141
 142                 if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
 143                         return -ENOMEM;
 144         }
 145
 146         return 0;
 147 }
 148
 149 static int bpf_xmit(struct sk_buff *skb)
 150 {
 151         struct dst_entry *dst = skb_dst(skb);
 152         struct bpf_lwt *bpf;
 153
 154         bpf = bpf_lwt_lwtunnel(dst->lwtstate);
 155         if (bpf->xmit.prog) {
 156                 int ret;
 157
 158                 ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT);
 159                 switch (ret) {
 160                 case BPF_OK:
 161                         /* If the header was expanded, headroom might be too
 162                          * small for L2 header to come, expand as needed.
 163                          */
 164                         ret = xmit_check_hhlen(skb);
 165                         if (unlikely(ret))
 166                                 return ret;
 167
 168                         return LWTUNNEL_XMIT_CONTINUE;
 169                 case BPF_REDIRECT:
 170                         return LWTUNNEL_XMIT_DONE;
 171                 default:
 172                         return ret;
 173                 }
 174         }
 175
 176         return LWTUNNEL_XMIT_CONTINUE;
 177 }
 178
 179 static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog)
 180 {
 181         if (prog->prog)
 182                 bpf_prog_put(prog->prog);
 183
 184         kfree(prog->name);
 185 }
 186
 187 static void bpf_destroy_state(struct lwtunnel_state *lwt)
 188 {
 189         struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
 190
 191         bpf_lwt_prog_destroy(&bpf->in);
 192         bpf_lwt_prog_destroy(&bpf->out);
 193         bpf_lwt_prog_destroy(&bpf->xmit);
 194 }
 195
 196 static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = {
 197         [LWT_BPF_PROG_FD]   = { .type = NLA_U32, },
 198         [LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING,
 199                                 .len = MAX_PROG_NAME },
 200 };
 201
 202 static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog,
 203                           enum bpf_prog_type type)
 204 {
 205         struct nlattr *tb[LWT_BPF_PROG_MAX + 1];
 206         struct bpf_prog *p;
 207         int ret;
 208         u32 fd;
 209
 210         ret = nla_parse_nested(tb, LWT_BPF_PROG_MAX, attr, bpf_prog_policy,
 211                                NULL);
 212         if (ret < 0)
 213                 return ret;
 214
 215         if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME])
 216                 return -EINVAL;
 217
 218         prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC);
 219         if (!prog->name)
 220                 return -ENOMEM;
 221
 222         fd = nla_get_u32(tb[LWT_BPF_PROG_FD]);
 223         p = bpf_prog_get_type(fd, type);
 224         if (IS_ERR(p))
 225                 return PTR_ERR(p);
 226
 227         prog->prog = p;
 228
 229         return 0;
 230 }
 231
 232 static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = {
 233         [LWT_BPF_IN]            = { .type = NLA_NESTED, },
 234         [LWT_BPF_OUT]           = { .type = NLA_NESTED, },
 235         [LWT_BPF_XMIT]          = { .type = NLA_NESTED, },
 236         [LWT_BPF_XMIT_HEADROOM] = { .type = NLA_U32 },
 237 };
 238
 239 static int bpf_build_state(struct nlattr *nla,
 240                            unsigned int family, const void *cfg,
 241                            struct lwtunnel_state **ts,
 242                            struct netlink_ext_ack *extack)
 243 {
 244         struct nlattr *tb[LWT_BPF_MAX + 1];
 245         struct lwtunnel_state *newts;
 246         struct bpf_lwt *bpf;
 247         int ret;
 248
 249         if (family != AF_INET && family != AF_INET6)
 250                 return -EAFNOSUPPORT;
 251
 252         ret = nla_parse_nested(tb, LWT_BPF_MAX, nla, bpf_nl_policy, extack);
 253         if (ret < 0)
 254                 return ret;
 255
 256         if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT])
 257                 return -EINVAL;
 258
 259         newts = lwtunnel_state_alloc(sizeof(*bpf));
 260         if (!newts)
 261                 return -ENOMEM;
 262
 263         newts->type = LWTUNNEL_ENCAP_BPF;
 264         bpf = bpf_lwt_lwtunnel(newts);
 265
 266         if (tb[LWT_BPF_IN]) {
 267                 newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT;
 268                 ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in,
 269                                      BPF_PROG_TYPE_LWT_IN);
 270                 if (ret  < 0)
 271                         goto errout;
 272         }
 273
 274         if (tb[LWT_BPF_OUT]) {
 275                 newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT;
 276                 ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out,
 277                                      BPF_PROG_TYPE_LWT_OUT);
 278                 if (ret < 0)
 279                         goto errout;
 280         }
 281
 282         if (tb[LWT_BPF_XMIT]) {
 283                 newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT;
 284                 ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit,
 285                                      BPF_PROG_TYPE_LWT_XMIT);
 286                 if (ret < 0)
 287                         goto errout;
 288         }
 289
 290         if (tb[LWT_BPF_XMIT_HEADROOM]) {
 291                 u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]);
 292
 293                 if (headroom > LWT_BPF_MAX_HEADROOM) {
 294                         ret = -ERANGE;
 295                         goto errout;
 296                 }
 297
 298                 newts->headroom = headroom;
 299         }
 300
 301         bpf->family = family;
 302         *ts = newts;
 303
 304         return 0;
 305
 306 errout:
 307         bpf_destroy_state(newts);
 308         kfree(newts);
 309         return ret;
 310 }
 311
 312 static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr,
 313                              struct bpf_lwt_prog *prog)
 314 {
 315         struct nlattr *nest;
 316
 317         if (!prog->prog)
 318                 return 0;
 319
 320         nest = nla_nest_start(skb, attr);
 321         if (!nest)
 322                 return -EMSGSIZE;
 323
 324         if (prog->name &&
 325             nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name))
 326                 return -EMSGSIZE;
 327
 328         return nla_nest_end(skb, nest);
 329 }
 330
 331 static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt)
 332 {
 333         struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
 334
 335         if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 ||
 336             bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 ||
 337             bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0)
 338                 return -EMSGSIZE;
 339
 340         return 0;
 341 }
 342
 343 static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate)
 344 {
 345         int nest_len = nla_total_size(sizeof(struct nlattr)) +
 346                        nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */
 347                        0;
 348
 349         return nest_len + /* LWT_BPF_IN */
 350                nest_len + /* LWT_BPF_OUT */
 351                nest_len + /* LWT_BPF_XMIT */
 352                0;
 353 }
 354
 355 static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b)
 356 {
 357         /* FIXME:
 358          * The LWT state is currently rebuilt for delete requests which
 359          * results in a new bpf_prog instance. Comparing names for now.
 360          */
 361         if (!a->name && !b->name)
 362                 return 0;
 363
 364         if (!a->name || !b->name)
 365                 return 1;
 366
 367         return strcmp(a->name, b->name);
 368 }
 369
 370 static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b)
 371 {
 372         struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a);
 373         struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b);
 374
 375         return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) ||
 376                bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) ||
 377                bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit);
 378 }
 379
 380 static const struct lwtunnel_encap_ops bpf_encap_ops = {
 381         .build_state    = bpf_build_state,
 382         .destroy_state  = bpf_destroy_state,
 383         .input          = bpf_input,
 384         .output         = bpf_output,
 385         .xmit           = bpf_xmit,
 386         .fill_encap     = bpf_fill_encap_info,
 387         .get_encap_size = bpf_encap_nlsize,
 388         .cmp_encap      = bpf_encap_cmp,
 389         .owner          = THIS_MODULE,
 390 };
 391
 392 static int __init bpf_lwt_init(void)
 393 {
 394         return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
 395 }
 396
 397 subsys_initcall(bpf_lwt_init)