net/core/filter.c

   1 /*
   2  * Linux Socket Filter - Kernel level socket filtering
   3  *
   4  * Based on the design of the Berkeley Packet Filter. The new
   5  * internal format has been designed by PLUMgrid:
   6  *
   7  *      Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
   8  *
   9  * Authors:
  10  *
  11  *      Jay Schulist <jschlst@samba.org>
  12  *      Alexei Starovoitov <ast@plumgrid.com>
  13  *      Daniel Borkmann <dborkman@redhat.com>
  14  *
  15  * This program is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU General Public License
  17  * as published by the Free Software Foundation; either version
  18  * 2 of the License, or (at your option) any later version.
  19  *
  20  * Andi Kleen - Fix a few bad bugs and races.
  21  * Kris Katterjohn - Added many additional checks in bpf_check_classic()
  22  */
  23
  24 #include <linux/module.h>
  25 #include <linux/types.h>
  26 #include <linux/mm.h>
  27 #include <linux/fcntl.h>
  28 #include <linux/socket.h>
  29 #include <linux/sock_diag.h>
  30 #include <linux/in.h>
  31 #include <linux/inet.h>
  32 #include <linux/netdevice.h>
  33 #include <linux/if_packet.h>
  34 #include <linux/if_arp.h>
  35 #include <linux/gfp.h>
  36 #include <net/ip.h>
  37 #include <net/protocol.h>
  38 #include <net/netlink.h>
  39 #include <linux/skbuff.h>
  40 #include <net/sock.h>
  41 #include <net/flow_dissector.h>
  42 #include <linux/errno.h>
  43 #include <linux/timer.h>
  44 #include <linux/uaccess.h>
  45 #include <asm/unaligned.h>
  46 #include <asm/cmpxchg.h>
  47 #include <linux/filter.h>
  48 #include <linux/ratelimit.h>
  49 #include <linux/seccomp.h>
  50 #include <linux/if_vlan.h>
  51 #include <linux/bpf.h>
  52 #include <net/sch_generic.h>
  53 #include <net/cls_cgroup.h>
  54 #include <net/dst_metadata.h>
  55 #include <net/dst.h>
  56 #include <net/sock_reuseport.h>
  57 #include <net/busy_poll.h>
  58 #include <net/tcp.h>
  59 #include <linux/bpf_trace.h>
  60
  61 /**
  62  *      sk_filter_trim_cap - run a packet through a socket filter
  63  *      @sk: sock associated with &sk_buff
  64  *      @skb: buffer to filter
  65  *      @cap: limit on how short the eBPF program may trim the packet
  66  *
  67  * Run the eBPF program and then cut skb->data to correct size returned by
  68  * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
  69  * than pkt_len we keep whole skb->data. This is the socket level
  70  * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
  71  * be accepted or -EPERM if the packet should be tossed.
  72  *
  73  */
  74 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
  75 {
  76         int err;
  77         struct sk_filter *filter;
  78
  79         /*
  80          * If the skb was allocated from pfmemalloc reserves, only
  81          * allow SOCK_MEMALLOC sockets to use it as this socket is
  82          * helping free memory
  83          */
  84         if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) {
  85                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PFMEMALLOCDROP);
  86                 return -ENOMEM;
  87         }
  88         err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
  89         if (err)
  90                 return err;
  91
  92         err = security_sock_rcv_skb(sk, skb);
  93         if (err)
  94                 return err;
  95
  96         rcu_read_lock();
  97         filter = rcu_dereference(sk->sk_filter);
  98         if (filter) {
  99                 struct sock *save_sk = skb->sk;
 100                 unsigned int pkt_len;
 101
 102                 skb->sk = sk;
 103                 pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
 104                 skb->sk = save_sk;
 105                 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
 106         }
 107         rcu_read_unlock();
 108
 109         return err;
 110 }
 111 EXPORT_SYMBOL(sk_filter_trim_cap);
 112
 113 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
 114 {
 115         return skb_get_poff(skb);
 116 }
 117
 118 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
 119 {
 120         struct nlattr *nla;
 121
 122         if (skb_is_nonlinear(skb))
 123                 return 0;
 124
 125         if (skb->len < sizeof(struct nlattr))
 126                 return 0;
 127
 128         if (a > skb->len - sizeof(struct nlattr))
 129                 return 0;
 130
 131         nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
 132         if (nla)
 133                 return (void *) nla - (void *) skb->data;
 134
 135         return 0;
 136 }
 137
 138 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
 139 {
 140         struct nlattr *nla;
 141
 142         if (skb_is_nonlinear(skb))
 143                 return 0;
 144
 145         if (skb->len < sizeof(struct nlattr))
 146                 return 0;
 147
 148         if (a > skb->len - sizeof(struct nlattr))
 149                 return 0;
 150
 151         nla = (struct nlattr *) &skb->data[a];
 152         if (nla->nla_len > skb->len - a)
 153                 return 0;
 154
 155         nla = nla_find_nested(nla, x);
 156         if (nla)
 157                 return (void *) nla - (void *) skb->data;
 158
 159         return 0;
 160 }
 161
 162 BPF_CALL_0(__get_raw_cpu_id)
 163 {
 164         return raw_smp_processor_id();
 165 }
 166
 167 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
 168         .func           = __get_raw_cpu_id,
 169         .gpl_only       = false,
 170         .ret_type       = RET_INTEGER,
 171 };
 172
 173 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
 174                               struct bpf_insn *insn_buf)
 175 {
 176         struct bpf_insn *insn = insn_buf;
 177
 178         switch (skb_field) {
 179         case SKF_AD_MARK:
 180                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
 181
 182                 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
 183                                       offsetof(struct sk_buff, mark));
 184                 break;
 185
 186         case SKF_AD_PKTTYPE:
 187                 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
 188                 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
 189 #ifdef __BIG_ENDIAN_BITFIELD
 190                 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
 191 #endif
 192                 break;
 193
 194         case SKF_AD_QUEUE:
 195                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
 196
 197                 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
 198                                       offsetof(struct sk_buff, queue_mapping));
 199                 break;
 200
 201         case SKF_AD_VLAN_TAG:
 202         case SKF_AD_VLAN_TAG_PRESENT:
 203                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
 204                 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
 205
 206                 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
 207                 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
 208                                       offsetof(struct sk_buff, vlan_tci));
 209                 if (skb_field == SKF_AD_VLAN_TAG) {
 210                         *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
 211                                                 ~VLAN_TAG_PRESENT);
 212                 } else {
 213                         /* dst_reg >>= 12 */
 214                         *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
 215                         /* dst_reg &= 1 */
 216                         *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
 217                 }
 218                 break;
 219         }
 220
 221         return insn - insn_buf;
 222 }
 223
 224 static bool convert_bpf_extensions(struct sock_filter *fp,
 225                                    struct bpf_insn **insnp)
 226 {
 227         struct bpf_insn *insn = *insnp;
 228         u32 cnt;
 229
 230         switch (fp->k) {
 231         case SKF_AD_OFF + SKF_AD_PROTOCOL:
 232                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
 233
 234                 /* A = *(u16 *) (CTX + offsetof(protocol)) */
 235                 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
 236                                       offsetof(struct sk_buff, protocol));
 237                 /* A = ntohs(A) [emitting a nop or swap16] */
 238                 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
 239                 break;
 240
 241         case SKF_AD_OFF + SKF_AD_PKTTYPE:
 242                 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
 243                 insn += cnt - 1;
 244                 break;
 245
 246         case SKF_AD_OFF + SKF_AD_IFINDEX:
 247         case SKF_AD_OFF + SKF_AD_HATYPE:
 248                 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
 249                 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
 250
 251                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
 252                                       BPF_REG_TMP, BPF_REG_CTX,
 253                                       offsetof(struct sk_buff, dev));
 254                 /* if (tmp != 0) goto pc + 1 */
 255                 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
 256                 *insn++ = BPF_EXIT_INSN();
 257                 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
 258                         *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
 259                                             offsetof(struct net_device, ifindex));
 260                 else
 261                         *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
 262                                             offsetof(struct net_device, type));
 263                 break;
 264
 265         case SKF_AD_OFF + SKF_AD_MARK:
 266                 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
 267                 insn += cnt - 1;
 268                 break;
 269
 270         case SKF_AD_OFF + SKF_AD_RXHASH:
 271                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
 272
 273                 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
 274                                     offsetof(struct sk_buff, hash));
 275                 break;
 276
 277         case SKF_AD_OFF + SKF_AD_QUEUE:
 278                 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
 279                 insn += cnt - 1;
 280                 break;
 281
 282         case SKF_AD_OFF + SKF_AD_VLAN_TAG:
 283                 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
 284                                          BPF_REG_A, BPF_REG_CTX, insn);
 285                 insn += cnt - 1;
 286                 break;
 287
 288         case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
 289                 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
 290                                          BPF_REG_A, BPF_REG_CTX, insn);
 291                 insn += cnt - 1;
 292                 break;
 293
 294         case SKF_AD_OFF + SKF_AD_VLAN_TPID:
 295                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
 296
 297                 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
 298                 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
 299                                       offsetof(struct sk_buff, vlan_proto));
 300                 /* A = ntohs(A) [emitting a nop or swap16] */
 301                 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
 302                 break;
 303
 304         case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
 305         case SKF_AD_OFF + SKF_AD_NLATTR:
 306         case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
 307         case SKF_AD_OFF + SKF_AD_CPU:
 308         case SKF_AD_OFF + SKF_AD_RANDOM:
 309                 /* arg1 = CTX */
 310                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
 311                 /* arg2 = A */
 312                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
 313                 /* arg3 = X */
 314                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
 315                 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
 316                 switch (fp->k) {
 317                 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
 318                         *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
 319                         break;
 320                 case SKF_AD_OFF + SKF_AD_NLATTR:
 321                         *insn = BPF_EMIT_CALL(__skb_get_nlattr);
 322                         break;
 323                 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
 324                         *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
 325                         break;
 326                 case SKF_AD_OFF + SKF_AD_CPU:
 327                         *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
 328                         break;
 329                 case SKF_AD_OFF + SKF_AD_RANDOM:
 330                         *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
 331                         bpf_user_rnd_init_once();
 332                         break;
 333                 }
 334                 break;
 335
 336         case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
 337                 /* A ^= X */
 338                 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
 339                 break;
 340
 341         default:
 342                 /* This is just a dummy call to avoid letting the compiler
 343                  * evict __bpf_call_base() as an optimization. Placed here
 344                  * where no-one bothers.
 345                  */
 346                 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
 347                 return false;
 348         }
 349
 350         *insnp = insn;
 351         return true;
 352 }
 353
 354 /**
 355  *      bpf_convert_filter - convert filter program
 356  *      @prog: the user passed filter program
 357  *      @len: the length of the user passed filter program
 358  *      @new_prog: allocated 'struct bpf_prog' or NULL
 359  *      @new_len: pointer to store length of converted program
 360  *
 361  * Remap 'sock_filter' style classic BPF (cBPF) instruction set to 'bpf_insn'
 362  * style extended BPF (eBPF).
 363  * Conversion workflow:
 364  *
 365  * 1) First pass for calculating the new program length:
 366  *   bpf_convert_filter(old_prog, old_len, NULL, &new_len)
 367  *
 368  * 2) 2nd pass to remap in two passes: 1st pass finds new
 369  *    jump offsets, 2nd pass remapping:
 370  *   bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
 371  */
 372 static int bpf_convert_filter(struct sock_filter *prog, int len,
 373                               struct bpf_prog *new_prog, int *new_len)
 374 {
 375         int new_flen = 0, pass = 0, target, i, stack_off;
 376         struct bpf_insn *new_insn, *first_insn = NULL;
 377         struct sock_filter *fp;
 378         int *addrs = NULL;
 379         u8 bpf_src;
 380
 381         BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
 382         BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
 383
 384         if (len <= 0 || len > BPF_MAXINSNS)
 385                 return -EINVAL;
 386
 387         if (new_prog) {
 388                 first_insn = new_prog->insnsi;
 389                 addrs = kcalloc(len, sizeof(*addrs),
 390                                 GFP_KERNEL | __GFP_NOWARN);
 391                 if (!addrs)
 392                         return -ENOMEM;
 393         }
 394
 395 do_pass:
 396         new_insn = first_insn;
 397         fp = prog;
 398
 399         /* Classic BPF related prologue emission. */
 400         if (new_prog) {
 401                 /* Classic BPF expects A and X to be reset first. These need
 402                  * to be guaranteed to be the first two instructions.
 403                  */
 404                 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
 405                 *new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
 406
 407                 /* All programs must keep CTX in callee saved BPF_REG_CTX.
 408                  * In eBPF case it's done by the compiler, here we need to
 409                  * do this ourself. Initial CTX is present in BPF_REG_ARG1.
 410                  */
 411                 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
 412         } else {
 413                 new_insn += 3;
 414         }
 415
 416         for (i = 0; i < len; fp++, i++) {
 417                 struct bpf_insn tmp_insns[6] = { };
 418                 struct bpf_insn *insn = tmp_insns;
 419
 420                 if (addrs)
 421                         addrs[i] = new_insn - first_insn;
 422
 423                 switch (fp->code) {
 424                 /* All arithmetic insns and skb loads map as-is. */
 425                 case BPF_ALU | BPF_ADD | BPF_X:
 426                 case BPF_ALU | BPF_ADD | BPF_K:
 427                 case BPF_ALU | BPF_SUB | BPF_X:
 428                 case BPF_ALU | BPF_SUB | BPF_K:
 429                 case BPF_ALU | BPF_AND | BPF_X:
 430                 case BPF_ALU | BPF_AND | BPF_K:
 431                 case BPF_ALU | BPF_OR | BPF_X:
 432                 case BPF_ALU | BPF_OR | BPF_K:
 433                 case BPF_ALU | BPF_LSH | BPF_X:
 434                 case BPF_ALU | BPF_LSH | BPF_K:
 435                 case BPF_ALU | BPF_RSH | BPF_X:
 436                 case BPF_ALU | BPF_RSH | BPF_K:
 437                 case BPF_ALU | BPF_XOR | BPF_X:
 438                 case BPF_ALU | BPF_XOR | BPF_K:
 439                 case BPF_ALU | BPF_MUL | BPF_X:
 440                 case BPF_ALU | BPF_MUL | BPF_K:
 441                 case BPF_ALU | BPF_DIV | BPF_X:
 442                 case BPF_ALU | BPF_DIV | BPF_K:
 443                 case BPF_ALU | BPF_MOD | BPF_X:
 444                 case BPF_ALU | BPF_MOD | BPF_K:
 445                 case BPF_ALU | BPF_NEG:
 446                 case BPF_LD | BPF_ABS | BPF_W:
 447                 case BPF_LD | BPF_ABS | BPF_H:
 448                 case BPF_LD | BPF_ABS | BPF_B:
 449                 case BPF_LD | BPF_IND | BPF_W:
 450                 case BPF_LD | BPF_IND | BPF_H:
 451                 case BPF_LD | BPF_IND | BPF_B:
 452                         /* Check for overloaded BPF extension and
 453                          * directly convert it if found, otherwise
 454                          * just move on with mapping.
 455                          */
 456                         if (BPF_CLASS(fp->code) == BPF_LD &&
 457                             BPF_MODE(fp->code) == BPF_ABS &&
 458                             convert_bpf_extensions(fp, &insn))
 459                                 break;
 460
 461                         *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
 462                         break;
 463
 464                 /* Jump transformation cannot use BPF block macros
 465                  * everywhere as offset calculation and target updates
 466                  * require a bit more work than the rest, i.e. jump
 467                  * opcodes map as-is, but offsets need adjustment.
 468                  */
 469
 470 #define BPF_EMIT_JMP                                                    \
 471         do {                                                            \
 472                 if (target >= len || target < 0)                        \
 473                         goto err;                                       \
 474                 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0;   \
 475                 /* Adjust pc relative offset for 2nd or 3rd insn. */    \
 476                 insn->off -= insn - tmp_insns;                          \
 477         } while (0)
 478
 479                 case BPF_JMP | BPF_JA:
 480                         target = i + fp->k + 1;
 481                         insn->code = fp->code;
 482                         BPF_EMIT_JMP;
 483                         break;
 484
 485                 case BPF_JMP | BPF_JEQ | BPF_K:
 486                 case BPF_JMP | BPF_JEQ | BPF_X:
 487                 case BPF_JMP | BPF_JSET | BPF_K:
 488                 case BPF_JMP | BPF_JSET | BPF_X:
 489                 case BPF_JMP | BPF_JGT | BPF_K:
 490                 case BPF_JMP | BPF_JGT | BPF_X:
 491                 case BPF_JMP | BPF_JGE | BPF_K:
 492                 case BPF_JMP | BPF_JGE | BPF_X:
 493                         if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
 494                                 /* BPF immediates are signed, zero extend
 495                                  * immediate into tmp register and use it
 496                                  * in compare insn.
 497                                  */
 498                                 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
 499
 500                                 insn->dst_reg = BPF_REG_A;
 501                                 insn->src_reg = BPF_REG_TMP;
 502                                 bpf_src = BPF_X;
 503                         } else {
 504                                 insn->dst_reg = BPF_REG_A;
 505                                 insn->imm = fp->k;
 506                                 bpf_src = BPF_SRC(fp->code);
 507                                 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
 508                         }
 509
 510                         /* Common case where 'jump_false' is next insn. */
 511                         if (fp->jf == 0) {
 512                                 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 513                                 target = i + fp->jt + 1;
 514                                 BPF_EMIT_JMP;
 515                                 break;
 516                         }
 517
 518                         /* Convert some jumps when 'jump_true' is next insn. */
 519                         if (fp->jt == 0) {
 520                                 switch (BPF_OP(fp->code)) {
 521                                 case BPF_JEQ:
 522                                         insn->code = BPF_JMP | BPF_JNE | bpf_src;
 523                                         break;
 524                                 case BPF_JGT:
 525                                         insn->code = BPF_JMP | BPF_JLE | bpf_src;
 526                                         break;
 527                                 case BPF_JGE:
 528                                         insn->code = BPF_JMP | BPF_JLT | bpf_src;
 529                                         break;
 530                                 default:
 531                                         goto jmp_rest;
 532                                 }
 533
 534                                 target = i + fp->jf + 1;
 535                                 BPF_EMIT_JMP;
 536                                 break;
 537                         }
 538 jmp_rest:
 539                         /* Other jumps are mapped into two insns: Jxx and JA. */
 540                         target = i + fp->jt + 1;
 541                         insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 542                         BPF_EMIT_JMP;
 543                         insn++;
 544
 545                         insn->code = BPF_JMP | BPF_JA;
 546                         target = i + fp->jf + 1;
 547                         BPF_EMIT_JMP;
 548                         break;
 549
 550                 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
 551                 case BPF_LDX | BPF_MSH | BPF_B:
 552                         /* tmp = A */
 553                         *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
 554                         /* A = BPF_R0 = *(u8 *) (skb->data + K) */
 555                         *insn++ = BPF_LD_ABS(BPF_B, fp->k);
 556                         /* A &= 0xf */
 557                         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
 558                         /* A <<= 2 */
 559                         *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
 560                         /* X = A */
 561                         *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 562                         /* A = tmp */
 563                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
 564                         break;
 565
 566                 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
 567                  * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
 568                  */
 569                 case BPF_RET | BPF_A:
 570                 case BPF_RET | BPF_K:
 571                         if (BPF_RVAL(fp->code) == BPF_K)
 572                                 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
 573                                                         0, fp->k);
 574                         *insn = BPF_EXIT_INSN();
 575                         break;
 576
 577                 /* Store to stack. */
 578                 case BPF_ST:
 579                 case BPF_STX:
 580                         stack_off = fp->k * 4  + 4;
 581                         *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
 582                                             BPF_ST ? BPF_REG_A : BPF_REG_X,
 583                                             -stack_off);
 584                         /* check_load_and_stores() verifies that classic BPF can
 585                          * load from stack only after write, so tracking
 586                          * stack_depth for ST|STX insns is enough
 587                          */
 588                         if (new_prog && new_prog->aux->stack_depth < stack_off)
 589                                 new_prog->aux->stack_depth = stack_off;
 590                         break;
 591
 592                 /* Load from stack. */
 593                 case BPF_LD | BPF_MEM:
 594                 case BPF_LDX | BPF_MEM:
 595                         stack_off = fp->k * 4  + 4;
 596                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD  ?
 597                                             BPF_REG_A : BPF_REG_X, BPF_REG_FP,
 598                                             -stack_off);
 599                         break;
 600
 601                 /* A = K or X = K */
 602                 case BPF_LD | BPF_IMM:
 603                 case BPF_LDX | BPF_IMM:
 604                         *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
 605                                               BPF_REG_A : BPF_REG_X, fp->k);
 606                         break;
 607
 608                 /* X = A */
 609                 case BPF_MISC | BPF_TAX:
 610                         *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 611                         break;
 612
 613                 /* A = X */
 614                 case BPF_MISC | BPF_TXA:
 615                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
 616                         break;
 617
 618                 /* A = skb->len or X = skb->len */
 619                 case BPF_LD | BPF_W | BPF_LEN:
 620                 case BPF_LDX | BPF_W | BPF_LEN:
 621                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
 622                                             BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
 623                                             offsetof(struct sk_buff, len));
 624                         break;
 625
 626                 /* Access seccomp_data fields. */
 627                 case BPF_LDX | BPF_ABS | BPF_W:
 628                         /* A = *(u32 *) (ctx + K) */
 629                         *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
 630                         break;
 631
 632                 /* Unknown instruction. */
 633                 default:
 634                         goto err;
 635                 }
 636
 637                 insn++;
 638                 if (new_prog)
 639                         memcpy(new_insn, tmp_insns,
 640                                sizeof(*insn) * (insn - tmp_insns));
 641                 new_insn += insn - tmp_insns;
 642         }
 643
 644         if (!new_prog) {
 645                 /* Only calculating new length. */
 646                 *new_len = new_insn - first_insn;
 647                 return 0;
 648         }
 649
 650         pass++;
 651         if (new_flen != new_insn - first_insn) {
 652                 new_flen = new_insn - first_insn;
 653                 if (pass > 2)
 654                         goto err;
 655                 goto do_pass;
 656         }
 657
 658         kfree(addrs);
 659         BUG_ON(*new_len != new_flen);
 660         return 0;
 661 err:
 662         kfree(addrs);
 663         return -EINVAL;
 664 }
 665
 666 /* Security:
 667  *
 668  * As we dont want to clear mem[] array for each packet going through
 669  * __bpf_prog_run(), we check that filter loaded by user never try to read
 670  * a cell if not previously written, and we check all branches to be sure
 671  * a malicious user doesn't try to abuse us.
 672  */
 673 static int check_load_and_stores(const struct sock_filter *filter, int flen)
 674 {
 675         u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
 676         int pc, ret = 0;
 677
 678         BUILD_BUG_ON(BPF_MEMWORDS > 16);
 679
 680         masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
 681         if (!masks)
 682                 return -ENOMEM;
 683
 684         memset(masks, 0xff, flen * sizeof(*masks));
 685
 686         for (pc = 0; pc < flen; pc++) {
 687                 memvalid &= masks[pc];
 688
 689                 switch (filter[pc].code) {
 690                 case BPF_ST:
 691                 case BPF_STX:
 692                         memvalid |= (1 << filter[pc].k);
 693                         break;
 694                 case BPF_LD | BPF_MEM:
 695                 case BPF_LDX | BPF_MEM:
 696                         if (!(memvalid & (1 << filter[pc].k))) {
 697                                 ret = -EINVAL;
 698                                 goto error;
 699                         }
 700                         break;
 701                 case BPF_JMP | BPF_JA:
 702                         /* A jump must set masks on target */
 703                         masks[pc + 1 + filter[pc].k] &= memvalid;
 704                         memvalid = ~0;
 705                         break;
 706                 case BPF_JMP | BPF_JEQ | BPF_K:
 707                 case BPF_JMP | BPF_JEQ | BPF_X:
 708                 case BPF_JMP | BPF_JGE | BPF_K:
 709                 case BPF_JMP | BPF_JGE | BPF_X:
 710                 case BPF_JMP | BPF_JGT | BPF_K:
 711                 case BPF_JMP | BPF_JGT | BPF_X:
 712                 case BPF_JMP | BPF_JSET | BPF_K:
 713                 case BPF_JMP | BPF_JSET | BPF_X:
 714                         /* A jump must set masks on targets */
 715                         masks[pc + 1 + filter[pc].jt] &= memvalid;
 716                         masks[pc + 1 + filter[pc].jf] &= memvalid;
 717                         memvalid = ~0;
 718                         break;
 719                 }
 720         }
 721 error:
 722         kfree(masks);
 723         return ret;
 724 }
 725
 726 static bool chk_code_allowed(u16 code_to_probe)
 727 {
 728         static const bool codes[] = {
 729                 /* 32 bit ALU operations */
 730                 [BPF_ALU | BPF_ADD | BPF_K] = true,
 731                 [BPF_ALU | BPF_ADD | BPF_X] = true,
 732                 [BPF_ALU | BPF_SUB | BPF_K] = true,
 733                 [BPF_ALU | BPF_SUB | BPF_X] = true,
 734                 [BPF_ALU | BPF_MUL | BPF_K] = true,
 735                 [BPF_ALU | BPF_MUL | BPF_X] = true,
 736                 [BPF_ALU | BPF_DIV | BPF_K] = true,
 737                 [BPF_ALU | BPF_DIV | BPF_X] = true,
 738                 [BPF_ALU | BPF_MOD | BPF_K] = true,
 739                 [BPF_ALU | BPF_MOD | BPF_X] = true,
 740                 [BPF_ALU | BPF_AND | BPF_K] = true,
 741                 [BPF_ALU | BPF_AND | BPF_X] = true,
 742                 [BPF_ALU | BPF_OR | BPF_K] = true,
 743                 [BPF_ALU | BPF_OR | BPF_X] = true,
 744                 [BPF_ALU | BPF_XOR | BPF_K] = true,
 745                 [BPF_ALU | BPF_XOR | BPF_X] = true,
 746                 [BPF_ALU | BPF_LSH | BPF_K] = true,
 747                 [BPF_ALU | BPF_LSH | BPF_X] = true,
 748                 [BPF_ALU | BPF_RSH | BPF_K] = true,
 749                 [BPF_ALU | BPF_RSH | BPF_X] = true,
 750                 [BPF_ALU | BPF_NEG] = true,
 751                 /* Load instructions */
 752                 [BPF_LD | BPF_W | BPF_ABS] = true,
 753                 [BPF_LD | BPF_H | BPF_ABS] = true,
 754                 [BPF_LD | BPF_B | BPF_ABS] = true,
 755                 [BPF_LD | BPF_W | BPF_LEN] = true,
 756                 [BPF_LD | BPF_W | BPF_IND] = true,
 757                 [BPF_LD | BPF_H | BPF_IND] = true,
 758                 [BPF_LD | BPF_B | BPF_IND] = true,
 759                 [BPF_LD | BPF_IMM] = true,
 760                 [BPF_LD | BPF_MEM] = true,
 761                 [BPF_LDX | BPF_W | BPF_LEN] = true,
 762                 [BPF_LDX | BPF_B | BPF_MSH] = true,
 763                 [BPF_LDX | BPF_IMM] = true,
 764                 [BPF_LDX | BPF_MEM] = true,
 765                 /* Store instructions */
 766                 [BPF_ST] = true,
 767                 [BPF_STX] = true,
 768                 /* Misc instructions */
 769                 [BPF_MISC | BPF_TAX] = true,
 770                 [BPF_MISC | BPF_TXA] = true,
 771                 /* Return instructions */
 772                 [BPF_RET | BPF_K] = true,
 773                 [BPF_RET | BPF_A] = true,
 774                 /* Jump instructions */
 775                 [BPF_JMP | BPF_JA] = true,
 776                 [BPF_JMP | BPF_JEQ | BPF_K] = true,
 777                 [BPF_JMP | BPF_JEQ | BPF_X] = true,
 778                 [BPF_JMP | BPF_JGE | BPF_K] = true,
 779                 [BPF_JMP | BPF_JGE | BPF_X] = true,
 780                 [BPF_JMP | BPF_JGT | BPF_K] = true,
 781                 [BPF_JMP | BPF_JGT | BPF_X] = true,
 782                 [BPF_JMP | BPF_JSET | BPF_K] = true,
 783                 [BPF_JMP | BPF_JSET | BPF_X] = true,
 784         };
 785
 786         if (code_to_probe >= ARRAY_SIZE(codes))
 787                 return false;
 788
 789         return codes[code_to_probe];
 790 }
 791
 792 static bool bpf_check_basics_ok(const struct sock_filter *filter,
 793                                 unsigned int flen)
 794 {
 795         if (filter == NULL)
 796                 return false;
 797         if (flen == 0 || flen > BPF_MAXINSNS)
 798                 return false;
 799
 800         return true;
 801 }
 802
 803 /**
 804  *      bpf_check_classic - verify socket filter code
 805  *      @filter: filter to verify
 806  *      @flen: length of filter
 807  *
 808  * Check the user's filter code. If we let some ugly
 809  * filter code slip through kaboom! The filter must contain
 810  * no references or jumps that are out of range, no illegal
 811  * instructions, and must end with a RET instruction.
 812  *
 813  * All jumps are forward as they are not signed.
 814  *
 815  * Returns 0 if the rule set is legal or -EINVAL if not.
 816  */
 817 static int bpf_check_classic(const struct sock_filter *filter,
 818                              unsigned int flen)
 819 {
 820         bool anc_found;
 821         int pc;
 822
 823         /* Check the filter code now */
 824         for (pc = 0; pc < flen; pc++) {
 825                 const struct sock_filter *ftest = &filter[pc];
 826
 827                 /* May we actually operate on this code? */
 828                 if (!chk_code_allowed(ftest->code))
 829                         return -EINVAL;
 830
 831                 /* Some instructions need special checks */
 832                 switch (ftest->code) {
 833                 case BPF_ALU | BPF_DIV | BPF_K:
 834                 case BPF_ALU | BPF_MOD | BPF_K:
 835                         /* Check for division by zero */
 836                         if (ftest->k == 0)
 837                                 return -EINVAL;
 838                         break;
 839                 case BPF_ALU | BPF_LSH | BPF_K:
 840                 case BPF_ALU | BPF_RSH | BPF_K:
 841                         if (ftest->k >= 32)
 842                                 return -EINVAL;
 843                         break;
 844                 case BPF_LD | BPF_MEM:
 845                 case BPF_LDX | BPF_MEM:
 846                 case BPF_ST:
 847                 case BPF_STX:
 848                         /* Check for invalid memory addresses */
 849                         if (ftest->k >= BPF_MEMWORDS)
 850                                 return -EINVAL;
 851                         break;
 852                 case BPF_JMP | BPF_JA:
 853                         /* Note, the large ftest->k might cause loops.
 854                          * Compare this with conditional jumps below,
 855                          * where offsets are limited. --ANK (981016)
 856                          */
 857                         if (ftest->k >= (unsigned int)(flen - pc - 1))
 858                                 return -EINVAL;
 859                         break;
 860                 case BPF_JMP | BPF_JEQ | BPF_K:
 861                 case BPF_JMP | BPF_JEQ | BPF_X:
 862                 case BPF_JMP | BPF_JGE | BPF_K:
 863                 case BPF_JMP | BPF_JGE | BPF_X:
 864                 case BPF_JMP | BPF_JGT | BPF_K:
 865                 case BPF_JMP | BPF_JGT | BPF_X:
 866                 case BPF_JMP | BPF_JSET | BPF_K:
 867                 case BPF_JMP | BPF_JSET | BPF_X:
 868                         /* Both conditionals must be safe */
 869                         if (pc + ftest->jt + 1 >= flen ||
 870                             pc + ftest->jf + 1 >= flen)
 871                                 return -EINVAL;
 872                         break;
 873                 case BPF_LD | BPF_W | BPF_ABS:
 874                 case BPF_LD | BPF_H | BPF_ABS:
 875                 case BPF_LD | BPF_B | BPF_ABS:
 876                         anc_found = false;
 877                         if (bpf_anc_helper(ftest) & BPF_ANC)
 878                                 anc_found = true;
 879                         /* Ancillary operation unknown or unsupported */
 880                         if (anc_found == false && ftest->k >= SKF_AD_OFF)
 881                                 return -EINVAL;
 882                 }
 883         }
 884
 885         /* Last instruction must be a RET code */
 886         switch (filter[flen - 1].code) {
 887         case BPF_RET | BPF_K:
 888         case BPF_RET | BPF_A:
 889                 return check_load_and_stores(filter, flen);
 890         }
 891
 892         return -EINVAL;
 893 }
 894
 895 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
 896                                       const struct sock_fprog *fprog)
 897 {
 898         unsigned int fsize = bpf_classic_proglen(fprog);
 899         struct sock_fprog_kern *fkprog;
 900
 901         fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
 902         if (!fp->orig_prog)
 903                 return -ENOMEM;
 904
 905         fkprog = fp->orig_prog;
 906         fkprog->len = fprog->len;
 907
 908         fkprog->filter = kmemdup(fp->insns, fsize,
 909                                  GFP_KERNEL | __GFP_NOWARN);
 910         if (!fkprog->filter) {
 911                 kfree(fp->orig_prog);
 912                 return -ENOMEM;
 913         }
 914
 915         return 0;
 916 }
 917
 918 static void bpf_release_orig_filter(struct bpf_prog *fp)
 919 {
 920         struct sock_fprog_kern *fprog = fp->orig_prog;
 921
 922         if (fprog) {
 923                 kfree(fprog->filter);
 924                 kfree(fprog);
 925         }
 926 }
 927
 928 static void __bpf_prog_release(struct bpf_prog *prog)
 929 {
 930         if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
 931                 bpf_prog_put(prog);
 932         } else {
 933                 bpf_release_orig_filter(prog);
 934                 bpf_prog_free(prog);
 935         }
 936 }
 937
 938 static void __sk_filter_release(struct sk_filter *fp)
 939 {
 940         __bpf_prog_release(fp->prog);
 941         kfree(fp);
 942 }
 943
 944 /**
 945  *      sk_filter_release_rcu - Release a socket filter by rcu_head
 946  *      @rcu: rcu_head that contains the sk_filter to free
 947  */
 948 static void sk_filter_release_rcu(struct rcu_head *rcu)
 949 {
 950         struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
 951
 952         __sk_filter_release(fp);
 953 }
 954
 955 /**
 956  *      sk_filter_release - release a socket filter
 957  *      @fp: filter to remove
 958  *
 959  *      Remove a filter from a socket and release its resources.
 960  */
 961 static void sk_filter_release(struct sk_filter *fp)
 962 {
 963         if (refcount_dec_and_test(&fp->refcnt))
 964                 call_rcu(&fp->rcu, sk_filter_release_rcu);
 965 }
 966
 967 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
 968 {
 969         u32 filter_size = bpf_prog_size(fp->prog->len);
 970
 971         atomic_sub(filter_size, &sk->sk_omem_alloc);
 972         sk_filter_release(fp);
 973 }
 974
 975 /* try to charge the socket memory if there is space available
 976  * return true on success
 977  */
 978 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 979 {
 980         u32 filter_size = bpf_prog_size(fp->prog->len);
 981
 982         /* same check as in sock_kmalloc() */
 983         if (filter_size <= sysctl_optmem_max &&
 984             atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
 985                 atomic_add(filter_size, &sk->sk_omem_alloc);
 986                 return true;
 987         }
 988         return false;
 989 }
 990
 991 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 992 {
 993         if (!refcount_inc_not_zero(&fp->refcnt))
 994                 return false;
 995
 996         if (!__sk_filter_charge(sk, fp)) {
 997                 sk_filter_release(fp);
 998                 return false;
 999         }
1000         return true;
1001 }
1002
1003 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
1004 {
1005         struct sock_filter *old_prog;
1006         struct bpf_prog *old_fp;
1007         int err, new_len, old_len = fp->len;
1008
1009         /* We are free to overwrite insns et al right here as it
1010          * won't be used at this point in time anymore internally
1011          * after the migration to the internal BPF instruction
1012          * representation.
1013          */
1014         BUILD_BUG_ON(sizeof(struct sock_filter) !=
1015                      sizeof(struct bpf_insn));
1016
1017         /* Conversion cannot happen on overlapping memory areas,
1018          * so we need to keep the user BPF around until the 2nd
1019          * pass. At this time, the user BPF is stored in fp->insns.
1020          */
1021         old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1022                            GFP_KERNEL | __GFP_NOWARN);
1023         if (!old_prog) {
1024                 err = -ENOMEM;
1025                 goto out_err;
1026         }
1027
1028         /* 1st pass: calculate the new program length. */
1029         err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1030         if (err)
1031                 goto out_err_free;
1032
1033         /* Expand fp for appending the new filter representation. */
1034         old_fp = fp;
1035         fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1036         if (!fp) {
1037                 /* The old_fp is still around in case we couldn't
1038                  * allocate new memory, so uncharge on that one.
1039                  */
1040                 fp = old_fp;
1041                 err = -ENOMEM;
1042                 goto out_err_free;
1043         }
1044
1045         fp->len = new_len;
1046
1047         /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1048         err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1049         if (err)
1050                 /* 2nd bpf_convert_filter() can fail only if it fails
1051                  * to allocate memory, remapping must succeed. Note,
1052                  * that at this time old_fp has already been released
1053                  * by krealloc().
1054                  */
1055                 goto out_err_free;
1056
1057         fp = bpf_prog_select_runtime(fp, &err);
1058         if (err)
1059                 goto out_err_free;
1060
1061         kfree(old_prog);
1062         return fp;
1063
1064 out_err_free:
1065         kfree(old_prog);
1066 out_err:
1067         __bpf_prog_release(fp);
1068         return ERR_PTR(err);
1069 }
1070
1071 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1072                                            bpf_aux_classic_check_t trans)
1073 {
1074         int err;
1075
1076         fp->bpf_func = NULL;
1077         fp->jited = 0;
1078
1079         err = bpf_check_classic(fp->insns, fp->len);
1080         if (err) {
1081                 __bpf_prog_release(fp);
1082                 return ERR_PTR(err);
1083         }
1084
1085         /* There might be additional checks and transformations
1086          * needed on classic filters, f.e. in case of seccomp.
1087          */
1088         if (trans) {
1089                 err = trans(fp->insns, fp->len);
1090                 if (err) {
1091                         __bpf_prog_release(fp);
1092                         return ERR_PTR(err);
1093                 }
1094         }
1095
1096         /* Probe if we can JIT compile the filter and if so, do
1097          * the compilation of the filter.
1098          */
1099         bpf_jit_compile(fp);
1100
1101         /* JIT compiler couldn't process this filter, so do the
1102          * internal BPF translation for the optimized interpreter.
1103          */
1104         if (!fp->jited)
1105                 fp = bpf_migrate_filter(fp);
1106
1107         return fp;
1108 }
1109
1110 /**
1111  *      bpf_prog_create - create an unattached filter
1112  *      @pfp: the unattached filter that is created
1113  *      @fprog: the filter program
1114  *
1115  * Create a filter independent of any socket. We first run some
1116  * sanity checks on it to make sure it does not explode on us later.
1117  * If an error occurs or there is insufficient memory for the filter
1118  * a negative errno code is returned. On success the return is zero.
1119  */
1120 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1121 {
1122         unsigned int fsize = bpf_classic_proglen(fprog);
1123         struct bpf_prog *fp;
1124
1125         /* Make sure new filter is there and in the right amounts. */
1126         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1127                 return -EINVAL;
1128
1129         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1130         if (!fp)
1131                 return -ENOMEM;
1132
1133         memcpy(fp->insns, fprog->filter, fsize);
1134
1135         fp->len = fprog->len;
1136         /* Since unattached filters are not copied back to user
1137          * space through sk_get_filter(), we do not need to hold
1138          * a copy here, and can spare us the work.
1139          */
1140         fp->orig_prog = NULL;
1141
1142         /* bpf_prepare_filter() already takes care of freeing
1143          * memory in case something goes wrong.
1144          */
1145         fp = bpf_prepare_filter(fp, NULL);
1146         if (IS_ERR(fp))
1147                 return PTR_ERR(fp);
1148
1149         *pfp = fp;
1150         return 0;
1151 }
1152 EXPORT_SYMBOL_GPL(bpf_prog_create);
1153
1154 /**
1155  *      bpf_prog_create_from_user - create an unattached filter from user buffer
1156  *      @pfp: the unattached filter that is created
1157  *      @fprog: the filter program
1158  *      @trans: post-classic verifier transformation handler
1159  *      @save_orig: save classic BPF program
1160  *
1161  * This function effectively does the same as bpf_prog_create(), only
1162  * that it builds up its insns buffer from user space provided buffer.
1163  * It also allows for passing a bpf_aux_classic_check_t handler.
1164  */
1165 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1166                               bpf_aux_classic_check_t trans, bool save_orig)
1167 {
1168         unsigned int fsize = bpf_classic_proglen(fprog);
1169         struct bpf_prog *fp;
1170         int err;
1171
1172         /* Make sure new filter is there and in the right amounts. */
1173         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1174                 return -EINVAL;
1175
1176         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1177         if (!fp)
1178                 return -ENOMEM;
1179
1180         if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1181                 __bpf_prog_free(fp);
1182                 return -EFAULT;
1183         }
1184
1185         fp->len = fprog->len;
1186         fp->orig_prog = NULL;
1187
1188         if (save_orig) {
1189                 err = bpf_prog_store_orig_filter(fp, fprog);
1190                 if (err) {
1191                         __bpf_prog_free(fp);
1192                         return -ENOMEM;
1193                 }
1194         }
1195
1196         /* bpf_prepare_filter() already takes care of freeing
1197          * memory in case something goes wrong.
1198          */
1199         fp = bpf_prepare_filter(fp, trans);
1200         if (IS_ERR(fp))
1201                 return PTR_ERR(fp);
1202
1203         *pfp = fp;
1204         return 0;
1205 }
1206 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1207
1208 void bpf_prog_destroy(struct bpf_prog *fp)
1209 {
1210         __bpf_prog_release(fp);
1211 }
1212 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1213
1214 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1215 {
1216         struct sk_filter *fp, *old_fp;
1217
1218         fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1219         if (!fp)
1220                 return -ENOMEM;
1221
1222         fp->prog = prog;
1223
1224         if (!__sk_filter_charge(sk, fp)) {
1225                 kfree(fp);
1226                 return -ENOMEM;
1227         }
1228         refcount_set(&fp->refcnt, 1);
1229
1230         old_fp = rcu_dereference_protected(sk->sk_filter,
1231                                            lockdep_sock_is_held(sk));
1232         rcu_assign_pointer(sk->sk_filter, fp);
1233
1234         if (old_fp)
1235                 sk_filter_uncharge(sk, old_fp);
1236
1237         return 0;
1238 }
1239
1240 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1241 {
1242         struct bpf_prog *old_prog;
1243         int err;
1244
1245         if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1246                 return -ENOMEM;
1247
1248         if (sk_unhashed(sk) && sk->sk_reuseport) {
1249                 err = reuseport_alloc(sk);
1250                 if (err)
1251                         return err;
1252         } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1253                 /* The socket wasn't bound with SO_REUSEPORT */
1254                 return -EINVAL;
1255         }
1256
1257         old_prog = reuseport_attach_prog(sk, prog);
1258         if (old_prog)
1259                 bpf_prog_destroy(old_prog);
1260
1261         return 0;
1262 }
1263
1264 static
1265 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1266 {
1267         unsigned int fsize = bpf_classic_proglen(fprog);
1268         struct bpf_prog *prog;
1269         int err;
1270
1271         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1272                 return ERR_PTR(-EPERM);
1273
1274         /* Make sure new filter is there and in the right amounts. */
1275         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1276                 return ERR_PTR(-EINVAL);
1277
1278         prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1279         if (!prog)
1280                 return ERR_PTR(-ENOMEM);
1281
1282         if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1283                 __bpf_prog_free(prog);
1284                 return ERR_PTR(-EFAULT);
1285         }
1286
1287         prog->len = fprog->len;
1288
1289         err = bpf_prog_store_orig_filter(prog, fprog);
1290         if (err) {
1291                 __bpf_prog_free(prog);
1292                 return ERR_PTR(-ENOMEM);
1293         }
1294
1295         /* bpf_prepare_filter() already takes care of freeing
1296          * memory in case something goes wrong.
1297          */
1298         return bpf_prepare_filter(prog, NULL);
1299 }
1300
1301 /**
1302  *      sk_attach_filter - attach a socket filter
1303  *      @fprog: the filter program
1304  *      @sk: the socket to use
1305  *
1306  * Attach the user's filter code. We first run some sanity checks on
1307  * it to make sure it does not explode on us later. If an error
1308  * occurs or there is insufficient memory for the filter a negative
1309  * errno code is returned. On success the return is zero.
1310  */
1311 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1312 {
1313         struct bpf_prog *prog = __get_filter(fprog, sk);
1314         int err;
1315
1316         if (IS_ERR(prog))
1317                 return PTR_ERR(prog);
1318
1319         err = __sk_attach_prog(prog, sk);
1320         if (err < 0) {
1321                 __bpf_prog_release(prog);
1322                 return err;
1323         }
1324
1325         return 0;
1326 }
1327 EXPORT_SYMBOL_GPL(sk_attach_filter);
1328
1329 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1330 {
1331         struct bpf_prog *prog = __get_filter(fprog, sk);
1332         int err;
1333
1334         if (IS_ERR(prog))
1335                 return PTR_ERR(prog);
1336
1337         err = __reuseport_attach_prog(prog, sk);
1338         if (err < 0) {
1339                 __bpf_prog_release(prog);
1340                 return err;
1341         }
1342
1343         return 0;
1344 }
1345
1346 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1347 {
1348         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1349                 return ERR_PTR(-EPERM);
1350
1351         return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1352 }
1353
1354 int sk_attach_bpf(u32 ufd, struct sock *sk)
1355 {
1356         struct bpf_prog *prog = __get_bpf(ufd, sk);
1357         int err;
1358
1359         if (IS_ERR(prog))
1360                 return PTR_ERR(prog);
1361
1362         err = __sk_attach_prog(prog, sk);
1363         if (err < 0) {
1364                 bpf_prog_put(prog);
1365                 return err;
1366         }
1367
1368         return 0;
1369 }
1370
1371 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1372 {
1373         struct bpf_prog *prog = __get_bpf(ufd, sk);
1374         int err;
1375
1376         if (IS_ERR(prog))
1377                 return PTR_ERR(prog);
1378
1379         err = __reuseport_attach_prog(prog, sk);
1380         if (err < 0) {
1381                 bpf_prog_put(prog);
1382                 return err;
1383         }
1384
1385         return 0;
1386 }
1387
1388 struct bpf_scratchpad {
1389         union {
1390                 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1391                 u8     buff[MAX_BPF_STACK];
1392         };
1393 };
1394
1395 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1396
1397 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1398                                           unsigned int write_len)
1399 {
1400         return skb_ensure_writable(skb, write_len);
1401 }
1402
1403 static inline int bpf_try_make_writable(struct sk_buff *skb,
1404                                         unsigned int write_len)
1405 {
1406         int err = __bpf_try_make_writable(skb, write_len);
1407
1408         bpf_compute_data_pointers(skb);
1409         return err;
1410 }
1411
1412 static int bpf_try_make_head_writable(struct sk_buff *skb)
1413 {
1414         return bpf_try_make_writable(skb, skb_headlen(skb));
1415 }
1416
1417 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1418 {
1419         if (skb_at_tc_ingress(skb))
1420                 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1421 }
1422
1423 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1424 {
1425         if (skb_at_tc_ingress(skb))
1426                 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1427 }
1428
1429 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1430            const void *, from, u32, len, u64, flags)
1431 {
1432         void *ptr;
1433
1434         if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1435                 return -EINVAL;
1436         if (unlikely(offset > 0xffff))
1437                 return -EFAULT;
1438         if (unlikely(bpf_try_make_writable(skb, offset + len)))
1439                 return -EFAULT;
1440
1441         ptr = skb->data + offset;
1442         if (flags & BPF_F_RECOMPUTE_CSUM)
1443                 __skb_postpull_rcsum(skb, ptr, len, offset);
1444
1445         memcpy(ptr, from, len);
1446
1447         if (flags & BPF_F_RECOMPUTE_CSUM)
1448                 __skb_postpush_rcsum(skb, ptr, len, offset);
1449         if (flags & BPF_F_INVALIDATE_HASH)
1450                 skb_clear_hash(skb);
1451
1452         return 0;
1453 }
1454
1455 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1456         .func           = bpf_skb_store_bytes,
1457         .gpl_only       = false,
1458         .ret_type       = RET_INTEGER,
1459         .arg1_type      = ARG_PTR_TO_CTX,
1460         .arg2_type      = ARG_ANYTHING,
1461         .arg3_type      = ARG_PTR_TO_MEM,
1462         .arg4_type      = ARG_CONST_SIZE,
1463         .arg5_type      = ARG_ANYTHING,
1464 };
1465
1466 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1467            void *, to, u32, len)
1468 {
1469         void *ptr;
1470
1471         if (unlikely(offset > 0xffff))
1472                 goto err_clear;
1473
1474         ptr = skb_header_pointer(skb, offset, len, to);
1475         if (unlikely(!ptr))
1476                 goto err_clear;
1477         if (ptr != to)
1478                 memcpy(to, ptr, len);
1479
1480         return 0;
1481 err_clear:
1482         memset(to, 0, len);
1483         return -EFAULT;
1484 }
1485
1486 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1487         .func           = bpf_skb_load_bytes,
1488         .gpl_only       = false,
1489         .ret_type       = RET_INTEGER,
1490         .arg1_type      = ARG_PTR_TO_CTX,
1491         .arg2_type      = ARG_ANYTHING,
1492         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
1493         .arg4_type      = ARG_CONST_SIZE,
1494 };
1495
1496 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1497 {
1498         /* Idea is the following: should the needed direct read/write
1499          * test fail during runtime, we can pull in more data and redo
1500          * again, since implicitly, we invalidate previous checks here.
1501          *
1502          * Or, since we know how much we need to make read/writeable,
1503          * this can be done once at the program beginning for direct
1504          * access case. By this we overcome limitations of only current
1505          * headroom being accessible.
1506          */
1507         return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1508 }
1509
1510 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1511         .func           = bpf_skb_pull_data,
1512         .gpl_only       = false,
1513         .ret_type       = RET_INTEGER,
1514         .arg1_type      = ARG_PTR_TO_CTX,
1515         .arg2_type      = ARG_ANYTHING,
1516 };
1517
1518 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1519            u64, from, u64, to, u64, flags)
1520 {
1521         __sum16 *ptr;
1522
1523         if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1524                 return -EINVAL;
1525         if (unlikely(offset > 0xffff || offset & 1))
1526                 return -EFAULT;
1527         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1528                 return -EFAULT;
1529
1530         ptr = (__sum16 *)(skb->data + offset);
1531         switch (flags & BPF_F_HDR_FIELD_MASK) {
1532         case 0:
1533                 if (unlikely(from != 0))
1534                         return -EINVAL;
1535
1536                 csum_replace_by_diff(ptr, to);
1537                 break;
1538         case 2:
1539                 csum_replace2(ptr, from, to);
1540                 break;
1541         case 4:
1542                 csum_replace4(ptr, from, to);
1543                 break;
1544         default:
1545                 return -EINVAL;
1546         }
1547
1548         return 0;
1549 }
1550
1551 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1552         .func           = bpf_l3_csum_replace,
1553         .gpl_only       = false,
1554         .ret_type       = RET_INTEGER,
1555         .arg1_type      = ARG_PTR_TO_CTX,
1556         .arg2_type      = ARG_ANYTHING,
1557         .arg3_type      = ARG_ANYTHING,
1558         .arg4_type      = ARG_ANYTHING,
1559         .arg5_type      = ARG_ANYTHING,
1560 };
1561
1562 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1563            u64, from, u64, to, u64, flags)
1564 {
1565         bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1566         bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1567         bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1568         __sum16 *ptr;
1569
1570         if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1571                                BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1572                 return -EINVAL;
1573         if (unlikely(offset > 0xffff || offset & 1))
1574                 return -EFAULT;
1575         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1576                 return -EFAULT;
1577
1578         ptr = (__sum16 *)(skb->data + offset);
1579         if (is_mmzero && !do_mforce && !*ptr)
1580                 return 0;
1581
1582         switch (flags & BPF_F_HDR_FIELD_MASK) {
1583         case 0:
1584                 if (unlikely(from != 0))
1585                         return -EINVAL;
1586
1587                 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1588                 break;
1589         case 2:
1590                 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1591                 break;
1592         case 4:
1593                 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1594                 break;
1595         default:
1596                 return -EINVAL;
1597         }
1598
1599         if (is_mmzero && !*ptr)
1600                 *ptr = CSUM_MANGLED_0;
1601         return 0;
1602 }
1603
1604 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1605         .func           = bpf_l4_csum_replace,
1606         .gpl_only       = false,
1607         .ret_type       = RET_INTEGER,
1608         .arg1_type      = ARG_PTR_TO_CTX,
1609         .arg2_type      = ARG_ANYTHING,
1610         .arg3_type      = ARG_ANYTHING,
1611         .arg4_type      = ARG_ANYTHING,
1612         .arg5_type      = ARG_ANYTHING,
1613 };
1614
1615 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1616            __be32 *, to, u32, to_size, __wsum, seed)
1617 {
1618         struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1619         u32 diff_size = from_size + to_size;
1620         int i, j = 0;
1621
1622         /* This is quite flexible, some examples:
1623          *
1624          * from_size == 0, to_size > 0,  seed := csum --> pushing data
1625          * from_size > 0,  to_size == 0, seed := csum --> pulling data
1626          * from_size > 0,  to_size > 0,  seed := 0    --> diffing data
1627          *
1628          * Even for diffing, from_size and to_size don't need to be equal.
1629          */
1630         if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1631                      diff_size > sizeof(sp->diff)))
1632                 return -EINVAL;
1633
1634         for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1635                 sp->diff[j] = ~from[i];
1636         for (i = 0; i <   to_size / sizeof(__be32); i++, j++)
1637                 sp->diff[j] = to[i];
1638
1639         return csum_partial(sp->diff, diff_size, seed);
1640 }
1641
1642 static const struct bpf_func_proto bpf_csum_diff_proto = {
1643         .func           = bpf_csum_diff,
1644         .gpl_only       = false,
1645         .pkt_access     = true,
1646         .ret_type       = RET_INTEGER,
1647         .arg1_type      = ARG_PTR_TO_MEM_OR_NULL,
1648         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
1649         .arg3_type      = ARG_PTR_TO_MEM_OR_NULL,
1650         .arg4_type      = ARG_CONST_SIZE_OR_ZERO,
1651         .arg5_type      = ARG_ANYTHING,
1652 };
1653
1654 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1655 {
1656         /* The interface is to be used in combination with bpf_csum_diff()
1657          * for direct packet writes. csum rotation for alignment as well
1658          * as emulating csum_sub() can be done from the eBPF program.
1659          */
1660         if (skb->ip_summed == CHECKSUM_COMPLETE)
1661                 return (skb->csum = csum_add(skb->csum, csum));
1662
1663         return -ENOTSUPP;
1664 }
1665
1666 static const struct bpf_func_proto bpf_csum_update_proto = {
1667         .func           = bpf_csum_update,
1668         .gpl_only       = false,
1669         .ret_type       = RET_INTEGER,
1670         .arg1_type      = ARG_PTR_TO_CTX,
1671         .arg2_type      = ARG_ANYTHING,
1672 };
1673
1674 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1675 {
1676         return dev_forward_skb(dev, skb);
1677 }
1678
1679 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1680                                       struct sk_buff *skb)
1681 {
1682         int ret = ____dev_forward_skb(dev, skb);
1683
1684         if (likely(!ret)) {
1685                 skb->dev = dev;
1686                 ret = netif_rx(skb);
1687         }
1688
1689         return ret;
1690 }
1691
1692 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1693 {
1694         int ret;
1695
1696         if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1697                 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1698                 kfree_skb(skb);
1699                 return -ENETDOWN;
1700         }
1701
1702         skb->dev = dev;
1703
1704         __this_cpu_inc(xmit_recursion);
1705         ret = dev_queue_xmit(skb);
1706         __this_cpu_dec(xmit_recursion);
1707
1708         return ret;
1709 }
1710
1711 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1712                                  u32 flags)
1713 {
1714         /* skb->mac_len is not set on normal egress */
1715         unsigned int mlen = skb->network_header - skb->mac_header;
1716
1717         __skb_pull(skb, mlen);
1718
1719         /* At ingress, the mac header has already been pulled once.
1720          * At egress, skb_pospull_rcsum has to be done in case that
1721          * the skb is originated from ingress (i.e. a forwarded skb)
1722          * to ensure that rcsum starts at net header.
1723          */
1724         if (!skb_at_tc_ingress(skb))
1725                 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1726         skb_pop_mac_header(skb);
1727         skb_reset_mac_len(skb);
1728         return flags & BPF_F_INGRESS ?
1729                __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1730 }
1731
1732 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1733                                  u32 flags)
1734 {
1735         /* Verify that a link layer header is carried */
1736         if (unlikely(skb->mac_header >= skb->network_header)) {
1737                 kfree_skb(skb);
1738                 return -ERANGE;
1739         }
1740
1741         bpf_push_mac_rcsum(skb);
1742         return flags & BPF_F_INGRESS ?
1743                __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1744 }
1745
1746 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1747                           u32 flags)
1748 {
1749         if (dev_is_mac_header_xmit(dev))
1750                 return __bpf_redirect_common(skb, dev, flags);
1751         else
1752                 return __bpf_redirect_no_mac(skb, dev, flags);
1753 }
1754
1755 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1756 {
1757         struct net_device *dev;
1758         struct sk_buff *clone;
1759         int ret;
1760
1761         if (unlikely(flags & ~(BPF_F_INGRESS)))
1762                 return -EINVAL;
1763
1764         dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1765         if (unlikely(!dev))
1766                 return -EINVAL;
1767
1768         clone = skb_clone(skb, GFP_ATOMIC);
1769         if (unlikely(!clone))
1770                 return -ENOMEM;
1771
1772         /* For direct write, we need to keep the invariant that the skbs
1773          * we're dealing with need to be uncloned. Should uncloning fail
1774          * here, we need to free the just generated clone to unclone once
1775          * again.
1776          */
1777         ret = bpf_try_make_head_writable(skb);
1778         if (unlikely(ret)) {
1779                 kfree_skb(clone);
1780                 return -ENOMEM;
1781         }
1782
1783         return __bpf_redirect(clone, dev, flags);
1784 }
1785
1786 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1787         .func           = bpf_clone_redirect,
1788         .gpl_only       = false,
1789         .ret_type       = RET_INTEGER,
1790         .arg1_type      = ARG_PTR_TO_CTX,
1791         .arg2_type      = ARG_ANYTHING,
1792         .arg3_type      = ARG_ANYTHING,
1793 };
1794
1795 struct redirect_info {
1796         u32 ifindex;
1797         u32 flags;
1798         struct bpf_map *map;
1799         struct bpf_map *map_to_flush;
1800         unsigned long   map_owner;
1801 };
1802
1803 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1804
1805 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1806 {
1807         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1808
1809         if (unlikely(flags & ~(BPF_F_INGRESS)))
1810                 return TC_ACT_SHOT;
1811
1812         ri->ifindex = ifindex;
1813         ri->flags = flags;
1814
1815         return TC_ACT_REDIRECT;
1816 }
1817
1818 int skb_do_redirect(struct sk_buff *skb)
1819 {
1820         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1821         struct net_device *dev;
1822
1823         dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1824         ri->ifindex = 0;
1825         if (unlikely(!dev)) {
1826                 kfree_skb(skb);
1827                 return -EINVAL;
1828         }
1829
1830         return __bpf_redirect(skb, dev, ri->flags);
1831 }
1832
1833 static const struct bpf_func_proto bpf_redirect_proto = {
1834         .func           = bpf_redirect,
1835         .gpl_only       = false,
1836         .ret_type       = RET_INTEGER,
1837         .arg1_type      = ARG_ANYTHING,
1838         .arg2_type      = ARG_ANYTHING,
1839 };
1840
1841 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
1842            struct bpf_map *, map, u32, key, u64, flags)
1843 {
1844         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1845
1846         /* If user passes invalid input drop the packet. */
1847         if (unlikely(flags))
1848                 return SK_DROP;
1849
1850         tcb->bpf.key = key;
1851         tcb->bpf.flags = flags;
1852         tcb->bpf.map = map;
1853
1854         return SK_PASS;
1855 }
1856
1857 struct sock *do_sk_redirect_map(struct sk_buff *skb)
1858 {
1859         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1860         struct sock *sk = NULL;
1861
1862         if (tcb->bpf.map) {
1863                 sk = __sock_map_lookup_elem(tcb->bpf.map, tcb->bpf.key);
1864
1865                 tcb->bpf.key = 0;
1866                 tcb->bpf.map = NULL;
1867         }
1868
1869         return sk;
1870 }
1871
1872 static const struct bpf_func_proto bpf_sk_redirect_map_proto = {
1873         .func           = bpf_sk_redirect_map,
1874         .gpl_only       = false,
1875         .ret_type       = RET_INTEGER,
1876         .arg1_type      = ARG_PTR_TO_CTX,
1877         .arg2_type      = ARG_CONST_MAP_PTR,
1878         .arg3_type      = ARG_ANYTHING,
1879         .arg4_type      = ARG_ANYTHING,
1880 };
1881
1882 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1883 {
1884         return task_get_classid(skb);
1885 }
1886
1887 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1888         .func           = bpf_get_cgroup_classid,
1889         .gpl_only       = false,
1890         .ret_type       = RET_INTEGER,
1891         .arg1_type      = ARG_PTR_TO_CTX,
1892 };
1893
1894 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1895 {
1896         return dst_tclassid(skb);
1897 }
1898
1899 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1900         .func           = bpf_get_route_realm,
1901         .gpl_only       = false,
1902         .ret_type       = RET_INTEGER,
1903         .arg1_type      = ARG_PTR_TO_CTX,
1904 };
1905
1906 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1907 {
1908         /* If skb_clear_hash() was called due to mangling, we can
1909          * trigger SW recalculation here. Later access to hash
1910          * can then use the inline skb->hash via context directly
1911          * instead of calling this helper again.
1912          */
1913         return skb_get_hash(skb);
1914 }
1915
1916 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1917         .func           = bpf_get_hash_recalc,
1918         .gpl_only       = false,
1919         .ret_type       = RET_INTEGER,
1920         .arg1_type      = ARG_PTR_TO_CTX,
1921 };
1922
1923 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1924 {
1925         /* After all direct packet write, this can be used once for
1926          * triggering a lazy recalc on next skb_get_hash() invocation.
1927          */
1928         skb_clear_hash(skb);
1929         return 0;
1930 }
1931
1932 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1933         .func           = bpf_set_hash_invalid,
1934         .gpl_only       = false,
1935         .ret_type       = RET_INTEGER,
1936         .arg1_type      = ARG_PTR_TO_CTX,
1937 };
1938
1939 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
1940 {
1941         /* Set user specified hash as L4(+), so that it gets returned
1942          * on skb_get_hash() call unless BPF prog later on triggers a
1943          * skb_clear_hash().
1944          */
1945         __skb_set_sw_hash(skb, hash, true);
1946         return 0;
1947 }
1948
1949 static const struct bpf_func_proto bpf_set_hash_proto = {
1950         .func           = bpf_set_hash,
1951         .gpl_only       = false,
1952         .ret_type       = RET_INTEGER,
1953         .arg1_type      = ARG_PTR_TO_CTX,
1954         .arg2_type      = ARG_ANYTHING,
1955 };
1956
1957 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1958            u16, vlan_tci)
1959 {
1960         int ret;
1961
1962         if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1963                      vlan_proto != htons(ETH_P_8021AD)))
1964                 vlan_proto = htons(ETH_P_8021Q);
1965
1966         bpf_push_mac_rcsum(skb);
1967         ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1968         bpf_pull_mac_rcsum(skb);
1969
1970         bpf_compute_data_pointers(skb);
1971         return ret;
1972 }
1973
1974 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1975         .func           = bpf_skb_vlan_push,
1976         .gpl_only       = false,
1977         .ret_type       = RET_INTEGER,
1978         .arg1_type      = ARG_PTR_TO_CTX,
1979         .arg2_type      = ARG_ANYTHING,
1980         .arg3_type      = ARG_ANYTHING,
1981 };
1982 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1983
1984 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
1985 {
1986         int ret;
1987
1988         bpf_push_mac_rcsum(skb);
1989         ret = skb_vlan_pop(skb);
1990         bpf_pull_mac_rcsum(skb);
1991
1992         bpf_compute_data_pointers(skb);
1993         return ret;
1994 }
1995
1996 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1997         .func           = bpf_skb_vlan_pop,
1998         .gpl_only       = false,
1999         .ret_type       = RET_INTEGER,
2000         .arg1_type      = ARG_PTR_TO_CTX,
2001 };
2002 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
2003
2004 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
2005 {
2006         /* Caller already did skb_cow() with len as headroom,
2007          * so no need to do it here.
2008          */
2009         skb_push(skb, len);
2010         memmove(skb->data, skb->data + len, off);
2011         memset(skb->data + off, 0, len);
2012
2013         /* No skb_postpush_rcsum(skb, skb->data + off, len)
2014          * needed here as it does not change the skb->csum
2015          * result for checksum complete when summing over
2016          * zeroed blocks.
2017          */
2018         return 0;
2019 }
2020
2021 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
2022 {
2023         /* skb_ensure_writable() is not needed here, as we're
2024          * already working on an uncloned skb.
2025          */
2026         if (unlikely(!pskb_may_pull(skb, off + len)))
2027                 return -ENOMEM;
2028
2029         skb_postpull_rcsum(skb, skb->data + off, len);
2030         memmove(skb->data + len, skb->data, off);
2031         __skb_pull(skb, len);
2032
2033         return 0;
2034 }
2035
2036 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
2037 {
2038         bool trans_same = skb->transport_header == skb->network_header;
2039         int ret;
2040
2041         /* There's no need for __skb_push()/__skb_pull() pair to
2042          * get to the start of the mac header as we're guaranteed
2043          * to always start from here under eBPF.
2044          */
2045         ret = bpf_skb_generic_push(skb, off, len);
2046         if (likely(!ret)) {
2047                 skb->mac_header -= len;
2048                 skb->network_header -= len;
2049                 if (trans_same)
2050                         skb->transport_header = skb->network_header;
2051         }
2052
2053         return ret;
2054 }
2055
2056 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
2057 {
2058         bool trans_same = skb->transport_header == skb->network_header;
2059         int ret;
2060
2061         /* Same here, __skb_push()/__skb_pull() pair not needed. */
2062         ret = bpf_skb_generic_pop(skb, off, len);
2063         if (likely(!ret)) {
2064                 skb->mac_header += len;
2065                 skb->network_header += len;
2066                 if (trans_same)
2067                         skb->transport_header = skb->network_header;
2068         }
2069
2070         return ret;
2071 }
2072
2073 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2074 {
2075         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2076         u32 off = skb_mac_header_len(skb);
2077         int ret;
2078
2079         ret = skb_cow(skb, len_diff);
2080         if (unlikely(ret < 0))
2081                 return ret;
2082
2083         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2084         if (unlikely(ret < 0))
2085                 return ret;
2086
2087         if (skb_is_gso(skb)) {
2088                 /* SKB_GSO_TCPV4 needs to be changed into
2089                  * SKB_GSO_TCPV6.
2090                  */
2091                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2092                         skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
2093                         skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV6;
2094                 }
2095
2096                 /* Due to IPv6 header, MSS needs to be downgraded. */
2097                 skb_shinfo(skb)->gso_size -= len_diff;
2098                 /* Header must be checked, and gso_segs recomputed. */
2099                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2100                 skb_shinfo(skb)->gso_segs = 0;
2101         }
2102
2103         skb->protocol = htons(ETH_P_IPV6);
2104         skb_clear_hash(skb);
2105
2106         return 0;
2107 }
2108
2109 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2110 {
2111         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2112         u32 off = skb_mac_header_len(skb);
2113         int ret;
2114
2115         ret = skb_unclone(skb, GFP_ATOMIC);
2116         if (unlikely(ret < 0))
2117                 return ret;
2118
2119         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2120         if (unlikely(ret < 0))
2121                 return ret;
2122
2123         if (skb_is_gso(skb)) {
2124                 /* SKB_GSO_TCPV6 needs to be changed into
2125                  * SKB_GSO_TCPV4.
2126                  */
2127                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2128                         skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2129                         skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV4;
2130                 }
2131
2132                 /* Due to IPv4 header, MSS can be upgraded. */
2133                 skb_shinfo(skb)->gso_size += len_diff;
2134                 /* Header must be checked, and gso_segs recomputed. */
2135                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2136                 skb_shinfo(skb)->gso_segs = 0;
2137         }
2138
2139         skb->protocol = htons(ETH_P_IP);
2140         skb_clear_hash(skb);
2141
2142         return 0;
2143 }
2144
2145 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2146 {
2147         __be16 from_proto = skb->protocol;
2148
2149         if (from_proto == htons(ETH_P_IP) &&
2150               to_proto == htons(ETH_P_IPV6))
2151                 return bpf_skb_proto_4_to_6(skb);
2152
2153         if (from_proto == htons(ETH_P_IPV6) &&
2154               to_proto == htons(ETH_P_IP))
2155                 return bpf_skb_proto_6_to_4(skb);
2156
2157         return -ENOTSUPP;
2158 }
2159
2160 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2161            u64, flags)
2162 {
2163         int ret;
2164
2165         if (unlikely(flags))
2166                 return -EINVAL;
2167
2168         /* General idea is that this helper does the basic groundwork
2169          * needed for changing the protocol, and eBPF program fills the
2170          * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2171          * and other helpers, rather than passing a raw buffer here.
2172          *
2173          * The rationale is to keep this minimal and without a need to
2174          * deal with raw packet data. F.e. even if we would pass buffers
2175          * here, the program still needs to call the bpf_lX_csum_replace()
2176          * helpers anyway. Plus, this way we keep also separation of
2177          * concerns, since f.e. bpf_skb_store_bytes() should only take
2178          * care of stores.
2179          *
2180          * Currently, additional options and extension header space are
2181          * not supported, but flags register is reserved so we can adapt
2182          * that. For offloads, we mark packet as dodgy, so that headers
2183          * need to be verified first.
2184          */
2185         ret = bpf_skb_proto_xlat(skb, proto);
2186         bpf_compute_data_pointers(skb);
2187         return ret;
2188 }
2189
2190 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2191         .func           = bpf_skb_change_proto,
2192         .gpl_only       = false,
2193         .ret_type       = RET_INTEGER,
2194         .arg1_type      = ARG_PTR_TO_CTX,
2195         .arg2_type      = ARG_ANYTHING,
2196         .arg3_type      = ARG_ANYTHING,
2197 };
2198
2199 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2200 {
2201         /* We only allow a restricted subset to be changed for now. */
2202         if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2203                      !skb_pkt_type_ok(pkt_type)))
2204                 return -EINVAL;
2205
2206         skb->pkt_type = pkt_type;
2207         return 0;
2208 }
2209
2210 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2211         .func           = bpf_skb_change_type,
2212         .gpl_only       = false,
2213         .ret_type       = RET_INTEGER,
2214         .arg1_type      = ARG_PTR_TO_CTX,
2215         .arg2_type      = ARG_ANYTHING,
2216 };
2217
2218 static u32 bpf_skb_net_base_len(const struct sk_buff *skb)
2219 {
2220         switch (skb->protocol) {
2221         case htons(ETH_P_IP):
2222                 return sizeof(struct iphdr);
2223         case htons(ETH_P_IPV6):
2224                 return sizeof(struct ipv6hdr);
2225         default:
2226                 return ~0U;
2227         }
2228 }
2229
2230 static int bpf_skb_net_grow(struct sk_buff *skb, u32 len_diff)
2231 {
2232         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2233         int ret;
2234
2235         ret = skb_cow(skb, len_diff);
2236         if (unlikely(ret < 0))
2237                 return ret;
2238
2239         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2240         if (unlikely(ret < 0))
2241                 return ret;
2242
2243         if (skb_is_gso(skb)) {
2244                 /* Due to header grow, MSS needs to be downgraded. */
2245                 skb_shinfo(skb)->gso_size -= len_diff;
2246                 /* Header must be checked, and gso_segs recomputed. */
2247                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2248                 skb_shinfo(skb)->gso_segs = 0;
2249         }
2250
2251         return 0;
2252 }
2253
2254 static int bpf_skb_net_shrink(struct sk_buff *skb, u32 len_diff)
2255 {
2256         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2257         int ret;
2258
2259         ret = skb_unclone(skb, GFP_ATOMIC);
2260         if (unlikely(ret < 0))
2261                 return ret;
2262
2263         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2264         if (unlikely(ret < 0))
2265                 return ret;
2266
2267         if (skb_is_gso(skb)) {
2268                 /* Due to header shrink, MSS can be upgraded. */
2269                 skb_shinfo(skb)->gso_size += len_diff;
2270                 /* Header must be checked, and gso_segs recomputed. */
2271                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2272                 skb_shinfo(skb)->gso_segs = 0;
2273         }
2274
2275         return 0;
2276 }
2277
2278 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2279 {
2280         return skb->dev->mtu + skb->dev->hard_header_len;
2281 }
2282
2283 static int bpf_skb_adjust_net(struct sk_buff *skb, s32 len_diff)
2284 {
2285         bool trans_same = skb->transport_header == skb->network_header;
2286         u32 len_cur, len_diff_abs = abs(len_diff);
2287         u32 len_min = bpf_skb_net_base_len(skb);
2288         u32 len_max = __bpf_skb_max_len(skb);
2289         __be16 proto = skb->protocol;
2290         bool shrink = len_diff < 0;
2291         int ret;
2292
2293         if (unlikely(len_diff_abs > 0xfffU))
2294                 return -EFAULT;
2295         if (unlikely(proto != htons(ETH_P_IP) &&
2296                      proto != htons(ETH_P_IPV6)))
2297                 return -ENOTSUPP;
2298
2299         len_cur = skb->len - skb_network_offset(skb);
2300         if (skb_transport_header_was_set(skb) && !trans_same)
2301                 len_cur = skb_network_header_len(skb);
2302         if ((shrink && (len_diff_abs >= len_cur ||
2303                         len_cur - len_diff_abs < len_min)) ||
2304             (!shrink && (skb->len + len_diff_abs > len_max &&
2305                          !skb_is_gso(skb))))
2306                 return -ENOTSUPP;
2307
2308         ret = shrink ? bpf_skb_net_shrink(skb, len_diff_abs) :
2309                        bpf_skb_net_grow(skb, len_diff_abs);
2310
2311         bpf_compute_data_pointers(skb);
2312         return ret;
2313 }
2314
2315 BPF_CALL_4(bpf_skb_adjust_room, struct sk_buff *, skb, s32, len_diff,
2316            u32, mode, u64, flags)
2317 {
2318         if (unlikely(flags))
2319                 return -EINVAL;
2320         if (likely(mode == BPF_ADJ_ROOM_NET))
2321                 return bpf_skb_adjust_net(skb, len_diff);
2322
2323         return -ENOTSUPP;
2324 }
2325
2326 static const struct bpf_func_proto bpf_skb_adjust_room_proto = {
2327         .func           = bpf_skb_adjust_room,
2328         .gpl_only       = false,
2329         .ret_type       = RET_INTEGER,
2330         .arg1_type      = ARG_PTR_TO_CTX,
2331         .arg2_type      = ARG_ANYTHING,
2332         .arg3_type      = ARG_ANYTHING,
2333         .arg4_type      = ARG_ANYTHING,
2334 };
2335
2336 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2337 {
2338         u32 min_len = skb_network_offset(skb);
2339
2340         if (skb_transport_header_was_set(skb))
2341                 min_len = skb_transport_offset(skb);
2342         if (skb->ip_summed == CHECKSUM_PARTIAL)
2343                 min_len = skb_checksum_start_offset(skb) +
2344                           skb->csum_offset + sizeof(__sum16);
2345         return min_len;
2346 }
2347
2348 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2349 {
2350         unsigned int old_len = skb->len;
2351         int ret;
2352
2353         ret = __skb_grow_rcsum(skb, new_len);
2354         if (!ret)
2355                 memset(skb->data + old_len, 0, new_len - old_len);
2356         return ret;
2357 }
2358
2359 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2360 {
2361         return __skb_trim_rcsum(skb, new_len);
2362 }
2363
2364 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2365            u64, flags)
2366 {
2367         u32 max_len = __bpf_skb_max_len(skb);
2368         u32 min_len = __bpf_skb_min_len(skb);
2369         int ret;
2370
2371         if (unlikely(flags || new_len > max_len || new_len < min_len))
2372                 return -EINVAL;
2373         if (skb->encapsulation)
2374                 return -ENOTSUPP;
2375
2376         /* The basic idea of this helper is that it's performing the
2377          * needed work to either grow or trim an skb, and eBPF program
2378          * rewrites the rest via helpers like bpf_skb_store_bytes(),
2379          * bpf_lX_csum_replace() and others rather than passing a raw
2380          * buffer here. This one is a slow path helper and intended
2381          * for replies with control messages.
2382          *
2383          * Like in bpf_skb_change_proto(), we want to keep this rather
2384          * minimal and without protocol specifics so that we are able
2385          * to separate concerns as in bpf_skb_store_bytes() should only
2386          * be the one responsible for writing buffers.
2387          *
2388          * It's really expected to be a slow path operation here for
2389          * control message replies, so we're implicitly linearizing,
2390          * uncloning and drop offloads from the skb by this.
2391          */
2392         ret = __bpf_try_make_writable(skb, skb->len);
2393         if (!ret) {
2394                 if (new_len > skb->len)
2395                         ret = bpf_skb_grow_rcsum(skb, new_len);
2396                 else if (new_len < skb->len)
2397                         ret = bpf_skb_trim_rcsum(skb, new_len);
2398                 if (!ret && skb_is_gso(skb))
2399                         skb_gso_reset(skb);
2400         }
2401
2402         bpf_compute_data_pointers(skb);
2403         return ret;
2404 }
2405
2406 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2407         .func           = bpf_skb_change_tail,
2408         .gpl_only       = false,
2409         .ret_type       = RET_INTEGER,
2410         .arg1_type      = ARG_PTR_TO_CTX,
2411         .arg2_type      = ARG_ANYTHING,
2412         .arg3_type      = ARG_ANYTHING,
2413 };
2414
2415 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2416            u64, flags)
2417 {
2418         u32 max_len = __bpf_skb_max_len(skb);
2419         u32 new_len = skb->len + head_room;
2420         int ret;
2421
2422         if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2423                      new_len < skb->len))
2424                 return -EINVAL;
2425
2426         ret = skb_cow(skb, head_room);
2427         if (likely(!ret)) {
2428                 /* Idea for this helper is that we currently only
2429                  * allow to expand on mac header. This means that
2430                  * skb->protocol network header, etc, stay as is.
2431                  * Compared to bpf_skb_change_tail(), we're more
2432                  * flexible due to not needing to linearize or
2433                  * reset GSO. Intention for this helper is to be
2434                  * used by an L3 skb that needs to push mac header
2435                  * for redirection into L2 device.
2436                  */
2437                 __skb_push(skb, head_room);
2438                 memset(skb->data, 0, head_room);
2439                 skb_reset_mac_header(skb);
2440         }
2441
2442         bpf_compute_data_pointers(skb);
2443         return 0;
2444 }
2445
2446 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2447         .func           = bpf_skb_change_head,
2448         .gpl_only       = false,
2449         .ret_type       = RET_INTEGER,
2450         .arg1_type      = ARG_PTR_TO_CTX,
2451         .arg2_type      = ARG_ANYTHING,
2452         .arg3_type      = ARG_ANYTHING,
2453 };
2454
2455 static unsigned long xdp_get_metalen(const struct xdp_buff *xdp)
2456 {
2457         return xdp_data_meta_unsupported(xdp) ? 0 :
2458                xdp->data - xdp->data_meta;
2459 }
2460
2461 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2462 {
2463         unsigned long metalen = xdp_get_metalen(xdp);
2464         void *data_start = xdp->data_hard_start + metalen;
2465         void *data = xdp->data + offset;
2466
2467         if (unlikely(data < data_start ||
2468                      data > xdp->data_end - ETH_HLEN))
2469                 return -EINVAL;
2470
2471         if (metalen)
2472                 memmove(xdp->data_meta + offset,
2473                         xdp->data_meta, metalen);
2474         xdp->data_meta += offset;
2475         xdp->data = data;
2476
2477         return 0;
2478 }
2479
2480 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2481         .func           = bpf_xdp_adjust_head,
2482         .gpl_only       = false,
2483         .ret_type       = RET_INTEGER,
2484         .arg1_type      = ARG_PTR_TO_CTX,
2485         .arg2_type      = ARG_ANYTHING,
2486 };
2487
2488 BPF_CALL_2(bpf_xdp_adjust_meta, struct xdp_buff *, xdp, int, offset)
2489 {
2490         void *meta = xdp->data_meta + offset;
2491         unsigned long metalen = xdp->data - meta;
2492
2493         if (xdp_data_meta_unsupported(xdp))
2494                 return -ENOTSUPP;
2495         if (unlikely(meta < xdp->data_hard_start ||
2496                      meta > xdp->data))
2497                 return -EINVAL;
2498         if (unlikely((metalen & (sizeof(__u32) - 1)) ||
2499                      (metalen > 32)))
2500                 return -EACCES;
2501
2502         xdp->data_meta = meta;
2503
2504         return 0;
2505 }
2506
2507 static const struct bpf_func_proto bpf_xdp_adjust_meta_proto = {
2508         .func           = bpf_xdp_adjust_meta,
2509         .gpl_only       = false,
2510         .ret_type       = RET_INTEGER,
2511         .arg1_type      = ARG_PTR_TO_CTX,
2512         .arg2_type      = ARG_ANYTHING,
2513 };
2514
2515 static int __bpf_tx_xdp(struct net_device *dev,
2516                         struct bpf_map *map,
2517                         struct xdp_buff *xdp,
2518                         u32 index)
2519 {
2520         int err;
2521
2522         if (!dev->netdev_ops->ndo_xdp_xmit) {
2523                 return -EOPNOTSUPP;
2524         }
2525
2526         err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
2527         if (err)
2528                 return err;
2529         dev->netdev_ops->ndo_xdp_flush(dev);
2530         return 0;
2531 }
2532
2533 static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
2534                             struct bpf_map *map,
2535                             struct xdp_buff *xdp,
2536                             u32 index)
2537 {
2538         int err;
2539
2540         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2541                 struct net_device *dev = fwd;
2542
2543                 if (!dev->netdev_ops->ndo_xdp_xmit)
2544                         return -EOPNOTSUPP;
2545
2546                 err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
2547                 if (err)
2548                         return err;
2549                 __dev_map_insert_ctx(map, index);
2550
2551         } else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
2552                 struct bpf_cpu_map_entry *rcpu = fwd;
2553
2554                 err = cpu_map_enqueue(rcpu, xdp, dev_rx);
2555                 if (err)
2556                         return err;
2557                 __cpu_map_insert_ctx(map, index);
2558         }
2559         return 0;
2560 }
2561
2562 void xdp_do_flush_map(void)
2563 {
2564         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2565         struct bpf_map *map = ri->map_to_flush;
2566
2567         ri->map_to_flush = NULL;
2568         if (map) {
2569                 switch (map->map_type) {
2570                 case BPF_MAP_TYPE_DEVMAP:
2571                         __dev_map_flush(map);
2572                         break;
2573                 case BPF_MAP_TYPE_CPUMAP:
2574                         __cpu_map_flush(map);
2575                         break;
2576                 default:
2577                         break;
2578                 }
2579         }
2580 }
2581 EXPORT_SYMBOL_GPL(xdp_do_flush_map);
2582
2583 static void *__xdp_map_lookup_elem(struct bpf_map *map, u32 index)
2584 {
2585         switch (map->map_type) {
2586         case BPF_MAP_TYPE_DEVMAP:
2587                 return __dev_map_lookup_elem(map, index);
2588         case BPF_MAP_TYPE_CPUMAP:
2589                 return __cpu_map_lookup_elem(map, index);
2590         default:
2591                 return NULL;
2592         }
2593 }
2594
2595 static inline bool xdp_map_invalid(const struct bpf_prog *xdp_prog,
2596                                    unsigned long aux)
2597 {
2598         return (unsigned long)xdp_prog->aux != aux;
2599 }
2600
2601 static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
2602                                struct bpf_prog *xdp_prog)
2603 {
2604         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2605         unsigned long map_owner = ri->map_owner;
2606         struct bpf_map *map = ri->map;
2607         u32 index = ri->ifindex;
2608         void *fwd = NULL;
2609         int err;
2610
2611         ri->ifindex = 0;
2612         ri->map = NULL;
2613         ri->map_owner = 0;
2614
2615         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2616                 err = -EFAULT;
2617                 map = NULL;
2618                 goto err;
2619         }
2620
2621         fwd = __xdp_map_lookup_elem(map, index);
2622         if (!fwd) {
2623                 err = -EINVAL;
2624                 goto err;
2625         }
2626         if (ri->map_to_flush && ri->map_to_flush != map)
2627                 xdp_do_flush_map();
2628
2629         err = __bpf_tx_xdp_map(dev, fwd, map, xdp, index);
2630         if (unlikely(err))
2631                 goto err;
2632
2633         ri->map_to_flush = map;
2634         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2635         return 0;
2636 err:
2637         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2638         return err;
2639 }
2640
2641 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
2642                     struct bpf_prog *xdp_prog)
2643 {
2644         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2645         struct net_device *fwd;
2646         u32 index = ri->ifindex;
2647         int err;
2648
2649         if (ri->map)
2650                 return xdp_do_redirect_map(dev, xdp, xdp_prog);
2651
2652         fwd = dev_get_by_index_rcu(dev_net(dev), index);
2653         ri->ifindex = 0;
2654         if (unlikely(!fwd)) {
2655                 err = -EINVAL;
2656                 goto err;
2657         }
2658
2659         err = __bpf_tx_xdp(fwd, NULL, xdp, 0);
2660         if (unlikely(err))
2661                 goto err;
2662
2663         _trace_xdp_redirect(dev, xdp_prog, index);
2664         return 0;
2665 err:
2666         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2667         return err;
2668 }
2669 EXPORT_SYMBOL_GPL(xdp_do_redirect);
2670
2671 static int __xdp_generic_ok_fwd_dev(struct sk_buff *skb, struct net_device *fwd)
2672 {
2673         unsigned int len;
2674
2675         if (unlikely(!(fwd->flags & IFF_UP)))
2676                 return -ENETDOWN;
2677
2678         len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
2679         if (skb->len > len)
2680                 return -EMSGSIZE;
2681
2682         return 0;
2683 }
2684
2685 int xdp_do_generic_redirect_map(struct net_device *dev, struct sk_buff *skb,
2686                                 struct bpf_prog *xdp_prog)
2687 {
2688         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2689         unsigned long map_owner = ri->map_owner;
2690         struct bpf_map *map = ri->map;
2691         struct net_device *fwd = NULL;
2692         u32 index = ri->ifindex;
2693         int err = 0;
2694
2695         ri->ifindex = 0;
2696         ri->map = NULL;
2697         ri->map_owner = 0;
2698
2699         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2700                 err = -EFAULT;
2701                 map = NULL;
2702                 goto err;
2703         }
2704         fwd = __xdp_map_lookup_elem(map, index);
2705         if (unlikely(!fwd)) {
2706                 err = -EINVAL;
2707                 goto err;
2708         }
2709
2710         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2711                 if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
2712                         goto err;
2713                 skb->dev = fwd;
2714         } else {
2715                 /* TODO: Handle BPF_MAP_TYPE_CPUMAP */
2716                 err = -EBADRQC;
2717                 goto err;
2718         }
2719
2720         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2721         return 0;
2722 err:
2723         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2724         return err;
2725 }
2726
2727 int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
2728                             struct bpf_prog *xdp_prog)
2729 {
2730         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2731         u32 index = ri->ifindex;
2732         struct net_device *fwd;
2733         int err = 0;
2734
2735         if (ri->map)
2736                 return xdp_do_generic_redirect_map(dev, skb, xdp_prog);
2737
2738         ri->ifindex = 0;
2739         fwd = dev_get_by_index_rcu(dev_net(dev), index);
2740         if (unlikely(!fwd)) {
2741                 err = -EINVAL;
2742                 goto err;
2743         }
2744
2745         if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
2746                 goto err;
2747
2748         skb->dev = fwd;
2749         _trace_xdp_redirect(dev, xdp_prog, index);
2750         return 0;
2751 err:
2752         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2753         return err;
2754 }
2755 EXPORT_SYMBOL_GPL(xdp_do_generic_redirect);
2756
2757 BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
2758 {
2759         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2760
2761         if (unlikely(flags))
2762                 return XDP_ABORTED;
2763
2764         ri->ifindex = ifindex;
2765         ri->flags = flags;
2766         ri->map = NULL;
2767         ri->map_owner = 0;
2768
2769         return XDP_REDIRECT;
2770 }
2771
2772 static const struct bpf_func_proto bpf_xdp_redirect_proto = {
2773         .func           = bpf_xdp_redirect,
2774         .gpl_only       = false,
2775         .ret_type       = RET_INTEGER,
2776         .arg1_type      = ARG_ANYTHING,
2777         .arg2_type      = ARG_ANYTHING,
2778 };
2779
2780 BPF_CALL_4(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex, u64, flags,
2781            unsigned long, map_owner)
2782 {
2783         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2784
2785         if (unlikely(flags))
2786                 return XDP_ABORTED;
2787
2788         ri->ifindex = ifindex;
2789         ri->flags = flags;
2790         ri->map = map;
2791         ri->map_owner = map_owner;
2792
2793         return XDP_REDIRECT;
2794 }
2795
2796 /* Note, arg4 is hidden from users and populated by the verifier
2797  * with the right pointer.
2798  */
2799 static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
2800         .func           = bpf_xdp_redirect_map,
2801         .gpl_only       = false,
2802         .ret_type       = RET_INTEGER,
2803         .arg1_type      = ARG_CONST_MAP_PTR,
2804         .arg2_type      = ARG_ANYTHING,
2805         .arg3_type      = ARG_ANYTHING,
2806 };
2807
2808 bool bpf_helper_changes_pkt_data(void *func)
2809 {
2810         if (func == bpf_skb_vlan_push ||
2811             func == bpf_skb_vlan_pop ||
2812             func == bpf_skb_store_bytes ||
2813             func == bpf_skb_change_proto ||
2814             func == bpf_skb_change_head ||
2815             func == bpf_skb_change_tail ||
2816             func == bpf_skb_adjust_room ||
2817             func == bpf_skb_pull_data ||
2818             func == bpf_clone_redirect ||
2819             func == bpf_l3_csum_replace ||
2820             func == bpf_l4_csum_replace ||
2821             func == bpf_xdp_adjust_head ||
2822             func == bpf_xdp_adjust_meta)
2823                 return true;
2824
2825         return false;
2826 }
2827
2828 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2829                                   unsigned long off, unsigned long len)
2830 {
2831         void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2832
2833         if (unlikely(!ptr))
2834                 return len;
2835         if (ptr != dst_buff)
2836                 memcpy(dst_buff, ptr, len);
2837
2838         return 0;
2839 }
2840
2841 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2842            u64, flags, void *, meta, u64, meta_size)
2843 {
2844         u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2845
2846         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2847                 return -EINVAL;
2848         if (unlikely(skb_size > skb->len))
2849                 return -EFAULT;
2850
2851         return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2852                                 bpf_skb_copy);
2853 }
2854
2855 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2856         .func           = bpf_skb_event_output,
2857         .gpl_only       = true,
2858         .ret_type       = RET_INTEGER,
2859         .arg1_type      = ARG_PTR_TO_CTX,
2860         .arg2_type      = ARG_CONST_MAP_PTR,
2861         .arg3_type      = ARG_ANYTHING,
2862         .arg4_type      = ARG_PTR_TO_MEM,
2863         .arg5_type      = ARG_CONST_SIZE,
2864 };
2865
2866 static unsigned short bpf_tunnel_key_af(u64 flags)
2867 {
2868         return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2869 }
2870
2871 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2872            u32, size, u64, flags)
2873 {
2874         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2875         u8 compat[sizeof(struct bpf_tunnel_key)];
2876         void *to_orig = to;
2877         int err;
2878
2879         if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2880                 err = -EINVAL;
2881                 goto err_clear;
2882         }
2883         if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2884                 err = -EPROTO;
2885                 goto err_clear;
2886         }
2887         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2888                 err = -EINVAL;
2889                 switch (size) {
2890                 case offsetof(struct bpf_tunnel_key, tunnel_label):
2891                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2892                         goto set_compat;
2893                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2894                         /* Fixup deprecated structure layouts here, so we have
2895                          * a common path later on.
2896                          */
2897                         if (ip_tunnel_info_af(info) != AF_INET)
2898                                 goto err_clear;
2899 set_compat:
2900                         to = (struct bpf_tunnel_key *)compat;
2901                         break;
2902                 default:
2903                         goto err_clear;
2904                 }
2905         }
2906
2907         to->tunnel_id = be64_to_cpu(info->key.tun_id);
2908         to->tunnel_tos = info->key.tos;
2909         to->tunnel_ttl = info->key.ttl;
2910
2911         if (flags & BPF_F_TUNINFO_IPV6) {
2912                 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2913                        sizeof(to->remote_ipv6));
2914                 to->tunnel_label = be32_to_cpu(info->key.label);
2915         } else {
2916                 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2917         }
2918
2919         if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2920                 memcpy(to_orig, to, size);
2921
2922         return 0;
2923 err_clear:
2924         memset(to_orig, 0, size);
2925         return err;
2926 }
2927
2928 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2929         .func           = bpf_skb_get_tunnel_key,
2930         .gpl_only       = false,
2931         .ret_type       = RET_INTEGER,
2932         .arg1_type      = ARG_PTR_TO_CTX,
2933         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
2934         .arg3_type      = ARG_CONST_SIZE,
2935         .arg4_type      = ARG_ANYTHING,
2936 };
2937
2938 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2939 {
2940         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2941         int err;
2942
2943         if (unlikely(!info ||
2944                      !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2945                 err = -ENOENT;
2946                 goto err_clear;
2947         }
2948         if (unlikely(size < info->options_len)) {
2949                 err = -ENOMEM;
2950                 goto err_clear;
2951         }
2952
2953         ip_tunnel_info_opts_get(to, info);
2954         if (size > info->options_len)
2955                 memset(to + info->options_len, 0, size - info->options_len);
2956
2957         return info->options_len;
2958 err_clear:
2959         memset(to, 0, size);
2960         return err;
2961 }
2962
2963 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2964         .func           = bpf_skb_get_tunnel_opt,
2965         .gpl_only       = false,
2966         .ret_type       = RET_INTEGER,
2967         .arg1_type      = ARG_PTR_TO_CTX,
2968         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
2969         .arg3_type      = ARG_CONST_SIZE,
2970 };
2971
2972 static struct metadata_dst __percpu *md_dst;
2973
2974 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2975            const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2976 {
2977         struct metadata_dst *md = this_cpu_ptr(md_dst);
2978         u8 compat[sizeof(struct bpf_tunnel_key)];
2979         struct ip_tunnel_info *info;
2980
2981         if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2982                                BPF_F_DONT_FRAGMENT)))
2983                 return -EINVAL;
2984         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2985                 switch (size) {
2986                 case offsetof(struct bpf_tunnel_key, tunnel_label):
2987                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2988                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2989                         /* Fixup deprecated structure layouts here, so we have
2990                          * a common path later on.
2991                          */
2992                         memcpy(compat, from, size);
2993                         memset(compat + size, 0, sizeof(compat) - size);
2994                         from = (const struct bpf_tunnel_key *) compat;
2995                         break;
2996                 default:
2997                         return -EINVAL;
2998                 }
2999         }
3000         if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
3001                      from->tunnel_ext))
3002                 return -EINVAL;
3003
3004         skb_dst_drop(skb);
3005         dst_hold((struct dst_entry *) md);
3006         skb_dst_set(skb, (struct dst_entry *) md);
3007
3008         info = &md->u.tun_info;
3009         info->mode = IP_TUNNEL_INFO_TX;
3010
3011         info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
3012         if (flags & BPF_F_DONT_FRAGMENT)
3013                 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
3014
3015         info->key.tun_id = cpu_to_be64(from->tunnel_id);
3016         info->key.tos = from->tunnel_tos;
3017         info->key.ttl = from->tunnel_ttl;
3018
3019         if (flags & BPF_F_TUNINFO_IPV6) {
3020                 info->mode |= IP_TUNNEL_INFO_IPV6;
3021                 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
3022                        sizeof(from->remote_ipv6));
3023                 info->key.label = cpu_to_be32(from->tunnel_label) &
3024                                   IPV6_FLOWLABEL_MASK;
3025         } else {
3026                 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
3027                 if (flags & BPF_F_ZERO_CSUM_TX)
3028                         info->key.tun_flags &= ~TUNNEL_CSUM;
3029         }
3030
3031         return 0;
3032 }
3033
3034 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
3035         .func           = bpf_skb_set_tunnel_key,
3036         .gpl_only       = false,
3037         .ret_type       = RET_INTEGER,
3038         .arg1_type      = ARG_PTR_TO_CTX,
3039         .arg2_type      = ARG_PTR_TO_MEM,
3040         .arg3_type      = ARG_CONST_SIZE,
3041         .arg4_type      = ARG_ANYTHING,
3042 };
3043
3044 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
3045            const u8 *, from, u32, size)
3046 {
3047         struct ip_tunnel_info *info = skb_tunnel_info(skb);
3048         const struct metadata_dst *md = this_cpu_ptr(md_dst);
3049
3050         if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
3051                 return -EINVAL;
3052         if (unlikely(size > IP_TUNNEL_OPTS_MAX))
3053                 return -ENOMEM;
3054
3055         ip_tunnel_info_opts_set(info, from, size);
3056
3057         return 0;
3058 }
3059
3060 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
3061         .func           = bpf_skb_set_tunnel_opt,
3062         .gpl_only       = false,
3063         .ret_type       = RET_INTEGER,
3064         .arg1_type      = ARG_PTR_TO_CTX,
3065         .arg2_type      = ARG_PTR_TO_MEM,
3066         .arg3_type      = ARG_CONST_SIZE,
3067 };
3068
3069 static const struct bpf_func_proto *
3070 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
3071 {
3072         if (!md_dst) {
3073                 struct metadata_dst __percpu *tmp;
3074
3075                 tmp = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
3076                                                 METADATA_IP_TUNNEL,
3077                                                 GFP_KERNEL);
3078                 if (!tmp)
3079                         return NULL;
3080                 if (cmpxchg(&md_dst, NULL, tmp))
3081                         metadata_dst_free_percpu(tmp);
3082         }
3083
3084         switch (which) {
3085         case BPF_FUNC_skb_set_tunnel_key:
3086                 return &bpf_skb_set_tunnel_key_proto;
3087         case BPF_FUNC_skb_set_tunnel_opt:
3088                 return &bpf_skb_set_tunnel_opt_proto;
3089         default:
3090                 return NULL;
3091         }
3092 }
3093
3094 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
3095            u32, idx)
3096 {
3097         struct bpf_array *array = container_of(map, struct bpf_array, map);
3098         struct cgroup *cgrp;
3099         struct sock *sk;
3100
3101         sk = skb_to_full_sk(skb);
3102         if (!sk || !sk_fullsock(sk))
3103                 return -ENOENT;
3104         if (unlikely(idx >= array->map.max_entries))
3105                 return -E2BIG;
3106
3107         cgrp = READ_ONCE(array->ptrs[idx]);
3108         if (unlikely(!cgrp))
3109                 return -EAGAIN;
3110
3111         return sk_under_cgroup_hierarchy(sk, cgrp);
3112 }
3113
3114 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
3115         .func           = bpf_skb_under_cgroup,
3116         .gpl_only       = false,
3117         .ret_type       = RET_INTEGER,
3118         .arg1_type      = ARG_PTR_TO_CTX,
3119         .arg2_type      = ARG_CONST_MAP_PTR,
3120         .arg3_type      = ARG_ANYTHING,
3121 };
3122
3123 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
3124                                   unsigned long off, unsigned long len)
3125 {
3126         memcpy(dst_buff, src_buff + off, len);
3127         return 0;
3128 }
3129
3130 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
3131            u64, flags, void *, meta, u64, meta_size)
3132 {
3133         u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
3134
3135         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
3136                 return -EINVAL;
3137         if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
3138                 return -EFAULT;
3139
3140         return bpf_event_output(map, flags, meta, meta_size, xdp->data,
3141                                 xdp_size, bpf_xdp_copy);
3142 }
3143
3144 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
3145         .func           = bpf_xdp_event_output,
3146         .gpl_only       = true,
3147         .ret_type       = RET_INTEGER,
3148         .arg1_type      = ARG_PTR_TO_CTX,
3149         .arg2_type      = ARG_CONST_MAP_PTR,
3150         .arg3_type      = ARG_ANYTHING,
3151         .arg4_type      = ARG_PTR_TO_MEM,
3152         .arg5_type      = ARG_CONST_SIZE,
3153 };
3154
3155 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
3156 {
3157         return skb->sk ? sock_gen_cookie(skb->sk) : 0;
3158 }
3159
3160 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
3161         .func           = bpf_get_socket_cookie,
3162         .gpl_only       = false,
3163         .ret_type       = RET_INTEGER,
3164         .arg1_type      = ARG_PTR_TO_CTX,
3165 };
3166
3167 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
3168 {
3169         struct sock *sk = sk_to_full_sk(skb->sk);
3170         kuid_t kuid;
3171
3172         if (!sk || !sk_fullsock(sk))
3173                 return overflowuid;
3174         kuid = sock_net_uid(sock_net(sk), sk);
3175         return from_kuid_munged(sock_net(sk)->user_ns, kuid);
3176 }
3177
3178 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
3179         .func           = bpf_get_socket_uid,
3180         .gpl_only       = false,
3181         .ret_type       = RET_INTEGER,
3182         .arg1_type      = ARG_PTR_TO_CTX,
3183 };
3184
3185 BPF_CALL_5(bpf_setsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3186            int, level, int, optname, char *, optval, int, optlen)
3187 {
3188         struct sock *sk = bpf_sock->sk;
3189         int ret = 0;
3190         int val;
3191
3192         if (!sk_fullsock(sk))
3193                 return -EINVAL;
3194
3195         if (level == SOL_SOCKET) {
3196                 if (optlen != sizeof(int))
3197                         return -EINVAL;
3198                 val = *((int *)optval);
3199
3200                 /* Only some socketops are supported */
3201                 switch (optname) {
3202                 case SO_RCVBUF:
3203                         sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
3204                         sk->sk_rcvbuf = max_t(int, val * 2, SOCK_MIN_RCVBUF);
3205                         break;
3206                 case SO_SNDBUF:
3207                         sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
3208                         sk->sk_sndbuf = max_t(int, val * 2, SOCK_MIN_SNDBUF);
3209                         break;
3210                 case SO_MAX_PACING_RATE:
3211                         sk->sk_max_pacing_rate = val;
3212                         sk->sk_pacing_rate = min(sk->sk_pacing_rate,
3213                                                  sk->sk_max_pacing_rate);
3214                         break;
3215                 case SO_PRIORITY:
3216                         sk->sk_priority = val;
3217                         break;
3218                 case SO_RCVLOWAT:
3219                         if (val < 0)
3220                                 val = INT_MAX;
3221                         sk->sk_rcvlowat = val ? : 1;
3222                         break;
3223                 case SO_MARK:
3224                         sk->sk_mark = val;
3225                         break;
3226                 default:
3227                         ret = -EINVAL;
3228                 }
3229 #ifdef CONFIG_INET
3230         } else if (level == SOL_TCP &&
3231                    sk->sk_prot->setsockopt == tcp_setsockopt) {
3232                 if (optname == TCP_CONGESTION) {
3233                         char name[TCP_CA_NAME_MAX];
3234                         bool reinit = bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN;
3235
3236                         strncpy(name, optval, min_t(long, optlen,
3237                                                     TCP_CA_NAME_MAX-1));
3238                         name[TCP_CA_NAME_MAX-1] = 0;
3239                         ret = tcp_set_congestion_control(sk, name, false, reinit);
3240                 } else {
3241                         struct tcp_sock *tp = tcp_sk(sk);
3242
3243                         if (optlen != sizeof(int))
3244                                 return -EINVAL;
3245
3246                         val = *((int *)optval);
3247                         /* Only some options are supported */
3248                         switch (optname) {
3249                         case TCP_BPF_IW:
3250                                 if (val <= 0 || tp->data_segs_out > 0)
3251                                         ret = -EINVAL;
3252                                 else
3253                                         tp->snd_cwnd = val;
3254                                 break;
3255                         case TCP_BPF_SNDCWND_CLAMP:
3256                                 if (val <= 0) {
3257                                         ret = -EINVAL;
3258                                 } else {
3259                                         tp->snd_cwnd_clamp = val;
3260                                         tp->snd_ssthresh = val;
3261                                 }
3262                                 break;
3263                         default:
3264                                 ret = -EINVAL;
3265                         }
3266                 }
3267 #endif
3268         } else {
3269                 ret = -EINVAL;
3270         }
3271         return ret;
3272 }
3273
3274 static const struct bpf_func_proto bpf_setsockopt_proto = {
3275         .func           = bpf_setsockopt,
3276         .gpl_only       = false,
3277         .ret_type       = RET_INTEGER,
3278         .arg1_type      = ARG_PTR_TO_CTX,
3279         .arg2_type      = ARG_ANYTHING,
3280         .arg3_type      = ARG_ANYTHING,
3281         .arg4_type      = ARG_PTR_TO_MEM,
3282         .arg5_type      = ARG_CONST_SIZE,
3283 };
3284
3285 BPF_CALL_5(bpf_getsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3286            int, level, int, optname, char *, optval, int, optlen)
3287 {
3288         struct sock *sk = bpf_sock->sk;
3289
3290         if (!sk_fullsock(sk))
3291                 goto err_clear;
3292
3293 #ifdef CONFIG_INET
3294         if (level == SOL_TCP && sk->sk_prot->getsockopt == tcp_getsockopt) {
3295                 if (optname == TCP_CONGESTION) {
3296                         struct inet_connection_sock *icsk = inet_csk(sk);
3297
3298                         if (!icsk->icsk_ca_ops || optlen <= 1)
3299                                 goto err_clear;
3300                         strncpy(optval, icsk->icsk_ca_ops->name, optlen);
3301                         optval[optlen - 1] = 0;
3302                 } else {
3303                         goto err_clear;
3304                 }
3305         } else {
3306                 goto err_clear;
3307         }
3308         return 0;
3309 #endif
3310 err_clear:
3311         memset(optval, 0, optlen);
3312         return -EINVAL;
3313 }
3314
3315 static const struct bpf_func_proto bpf_getsockopt_proto = {
3316         .func           = bpf_getsockopt,
3317         .gpl_only       = false,
3318         .ret_type       = RET_INTEGER,
3319         .arg1_type      = ARG_PTR_TO_CTX,
3320         .arg2_type      = ARG_ANYTHING,
3321         .arg3_type      = ARG_ANYTHING,
3322         .arg4_type      = ARG_PTR_TO_UNINIT_MEM,
3323         .arg5_type      = ARG_CONST_SIZE,
3324 };
3325
3326 static const struct bpf_func_proto *
3327 bpf_base_func_proto(enum bpf_func_id func_id)
3328 {
3329         switch (func_id) {
3330         case BPF_FUNC_map_lookup_elem:
3331                 return &bpf_map_lookup_elem_proto;
3332         case BPF_FUNC_map_update_elem:
3333                 return &bpf_map_update_elem_proto;
3334         case BPF_FUNC_map_delete_elem:
3335                 return &bpf_map_delete_elem_proto;
3336         case BPF_FUNC_get_prandom_u32:
3337                 return &bpf_get_prandom_u32_proto;
3338         case BPF_FUNC_get_smp_processor_id:
3339                 return &bpf_get_raw_smp_processor_id_proto;
3340         case BPF_FUNC_get_numa_node_id:
3341                 return &bpf_get_numa_node_id_proto;
3342         case BPF_FUNC_tail_call:
3343                 return &bpf_tail_call_proto;
3344         case BPF_FUNC_ktime_get_ns:
3345                 return &bpf_ktime_get_ns_proto;
3346         case BPF_FUNC_trace_printk:
3347                 if (capable(CAP_SYS_ADMIN))
3348                         return bpf_get_trace_printk_proto();
3349         default:
3350                 return NULL;
3351         }
3352 }
3353
3354 static const struct bpf_func_proto *
3355 sock_filter_func_proto(enum bpf_func_id func_id)
3356 {
3357         switch (func_id) {
3358         /* inet and inet6 sockets are created in a process
3359          * context so there is always a valid uid/gid
3360          */
3361         case BPF_FUNC_get_current_uid_gid:
3362                 return &bpf_get_current_uid_gid_proto;
3363         default:
3364                 return bpf_base_func_proto(func_id);
3365         }
3366 }
3367
3368 static const struct bpf_func_proto *
3369 sk_filter_func_proto(enum bpf_func_id func_id)
3370 {
3371         switch (func_id) {
3372         case BPF_FUNC_skb_load_bytes:
3373                 return &bpf_skb_load_bytes_proto;
3374         case BPF_FUNC_get_socket_cookie:
3375                 return &bpf_get_socket_cookie_proto;
3376         case BPF_FUNC_get_socket_uid:
3377                 return &bpf_get_socket_uid_proto;
3378         default:
3379                 return bpf_base_func_proto(func_id);
3380         }
3381 }
3382
3383 static const struct bpf_func_proto *
3384 tc_cls_act_func_proto(enum bpf_func_id func_id)
3385 {
3386         switch (func_id) {
3387         case BPF_FUNC_skb_store_bytes:
3388                 return &bpf_skb_store_bytes_proto;
3389         case BPF_FUNC_skb_load_bytes:
3390                 return &bpf_skb_load_bytes_proto;
3391         case BPF_FUNC_skb_pull_data:
3392                 return &bpf_skb_pull_data_proto;
3393         case BPF_FUNC_csum_diff:
3394                 return &bpf_csum_diff_proto;
3395         case BPF_FUNC_csum_update:
3396                 return &bpf_csum_update_proto;
3397         case BPF_FUNC_l3_csum_replace:
3398                 return &bpf_l3_csum_replace_proto;
3399         case BPF_FUNC_l4_csum_replace:
3400                 return &bpf_l4_csum_replace_proto;
3401         case BPF_FUNC_clone_redirect:
3402                 return &bpf_clone_redirect_proto;
3403         case BPF_FUNC_get_cgroup_classid:
3404                 return &bpf_get_cgroup_classid_proto;
3405         case BPF_FUNC_skb_vlan_push:
3406                 return &bpf_skb_vlan_push_proto;
3407         case BPF_FUNC_skb_vlan_pop:
3408                 return &bpf_skb_vlan_pop_proto;
3409         case BPF_FUNC_skb_change_proto:
3410                 return &bpf_skb_change_proto_proto;
3411         case BPF_FUNC_skb_change_type:
3412                 return &bpf_skb_change_type_proto;
3413         case BPF_FUNC_skb_adjust_room:
3414                 return &bpf_skb_adjust_room_proto;
3415         case BPF_FUNC_skb_change_tail:
3416                 return &bpf_skb_change_tail_proto;
3417         case BPF_FUNC_skb_get_tunnel_key:
3418                 return &bpf_skb_get_tunnel_key_proto;
3419         case BPF_FUNC_skb_set_tunnel_key:
3420                 return bpf_get_skb_set_tunnel_proto(func_id);
3421         case BPF_FUNC_skb_get_tunnel_opt:
3422                 return &bpf_skb_get_tunnel_opt_proto;
3423         case BPF_FUNC_skb_set_tunnel_opt:
3424                 return bpf_get_skb_set_tunnel_proto(func_id);
3425         case BPF_FUNC_redirect:
3426                 return &bpf_redirect_proto;
3427         case BPF_FUNC_get_route_realm:
3428                 return &bpf_get_route_realm_proto;
3429         case BPF_FUNC_get_hash_recalc:
3430                 return &bpf_get_hash_recalc_proto;
3431         case BPF_FUNC_set_hash_invalid:
3432                 return &bpf_set_hash_invalid_proto;
3433         case BPF_FUNC_set_hash:
3434                 return &bpf_set_hash_proto;
3435         case BPF_FUNC_perf_event_output:
3436                 return &bpf_skb_event_output_proto;
3437         case BPF_FUNC_get_smp_processor_id:
3438                 return &bpf_get_smp_processor_id_proto;
3439         case BPF_FUNC_skb_under_cgroup:
3440                 return &bpf_skb_under_cgroup_proto;
3441         case BPF_FUNC_get_socket_cookie:
3442                 return &bpf_get_socket_cookie_proto;
3443         case BPF_FUNC_get_socket_uid:
3444                 return &bpf_get_socket_uid_proto;
3445         default:
3446                 return bpf_base_func_proto(func_id);
3447         }
3448 }
3449
3450 static const struct bpf_func_proto *
3451 xdp_func_proto(enum bpf_func_id func_id)
3452 {
3453         switch (func_id) {
3454         case BPF_FUNC_perf_event_output:
3455                 return &bpf_xdp_event_output_proto;
3456         case BPF_FUNC_get_smp_processor_id:
3457                 return &bpf_get_smp_processor_id_proto;
3458         case BPF_FUNC_xdp_adjust_head:
3459                 return &bpf_xdp_adjust_head_proto;
3460         case BPF_FUNC_xdp_adjust_meta:
3461                 return &bpf_xdp_adjust_meta_proto;
3462         case BPF_FUNC_redirect:
3463                 return &bpf_xdp_redirect_proto;
3464         case BPF_FUNC_redirect_map:
3465                 return &bpf_xdp_redirect_map_proto;
3466         default:
3467                 return bpf_base_func_proto(func_id);
3468         }
3469 }
3470
3471 static const struct bpf_func_proto *
3472 lwt_inout_func_proto(enum bpf_func_id func_id)
3473 {
3474         switch (func_id) {
3475         case BPF_FUNC_skb_load_bytes:
3476                 return &bpf_skb_load_bytes_proto;
3477         case BPF_FUNC_skb_pull_data:
3478                 return &bpf_skb_pull_data_proto;
3479         case BPF_FUNC_csum_diff:
3480                 return &bpf_csum_diff_proto;
3481         case BPF_FUNC_get_cgroup_classid:
3482                 return &bpf_get_cgroup_classid_proto;
3483         case BPF_FUNC_get_route_realm:
3484                 return &bpf_get_route_realm_proto;
3485         case BPF_FUNC_get_hash_recalc:
3486                 return &bpf_get_hash_recalc_proto;
3487         case BPF_FUNC_perf_event_output:
3488                 return &bpf_skb_event_output_proto;
3489         case BPF_FUNC_get_smp_processor_id:
3490                 return &bpf_get_smp_processor_id_proto;
3491         case BPF_FUNC_skb_under_cgroup:
3492                 return &bpf_skb_under_cgroup_proto;
3493         default:
3494                 return bpf_base_func_proto(func_id);
3495         }
3496 }
3497
3498 static const struct bpf_func_proto *
3499         sock_ops_func_proto(enum bpf_func_id func_id)
3500 {
3501         switch (func_id) {
3502         case BPF_FUNC_setsockopt:
3503                 return &bpf_setsockopt_proto;
3504         case BPF_FUNC_getsockopt:
3505                 return &bpf_getsockopt_proto;
3506         case BPF_FUNC_sock_map_update:
3507                 return &bpf_sock_map_update_proto;
3508         default:
3509                 return bpf_base_func_proto(func_id);
3510         }
3511 }
3512
3513 static const struct bpf_func_proto *sk_skb_func_proto(enum bpf_func_id func_id)
3514 {
3515         switch (func_id) {
3516         case BPF_FUNC_skb_store_bytes:
3517                 return &bpf_skb_store_bytes_proto;
3518         case BPF_FUNC_skb_load_bytes:
3519                 return &bpf_skb_load_bytes_proto;
3520         case BPF_FUNC_skb_pull_data:
3521                 return &bpf_skb_pull_data_proto;
3522         case BPF_FUNC_skb_change_tail:
3523                 return &bpf_skb_change_tail_proto;
3524         case BPF_FUNC_skb_change_head:
3525                 return &bpf_skb_change_head_proto;
3526         case BPF_FUNC_get_socket_cookie:
3527                 return &bpf_get_socket_cookie_proto;
3528         case BPF_FUNC_get_socket_uid:
3529                 return &bpf_get_socket_uid_proto;
3530         case BPF_FUNC_sk_redirect_map:
3531                 return &bpf_sk_redirect_map_proto;
3532         default:
3533                 return bpf_base_func_proto(func_id);
3534         }
3535 }
3536
3537 static const struct bpf_func_proto *
3538 lwt_xmit_func_proto(enum bpf_func_id func_id)
3539 {
3540         switch (func_id) {
3541         case BPF_FUNC_skb_get_tunnel_key:
3542                 return &bpf_skb_get_tunnel_key_proto;
3543         case BPF_FUNC_skb_set_tunnel_key:
3544                 return bpf_get_skb_set_tunnel_proto(func_id);
3545         case BPF_FUNC_skb_get_tunnel_opt:
3546                 return &bpf_skb_get_tunnel_opt_proto;
3547         case BPF_FUNC_skb_set_tunnel_opt:
3548                 return bpf_get_skb_set_tunnel_proto(func_id);
3549         case BPF_FUNC_redirect:
3550                 return &bpf_redirect_proto;
3551         case BPF_FUNC_clone_redirect:
3552                 return &bpf_clone_redirect_proto;
3553         case BPF_FUNC_skb_change_tail:
3554                 return &bpf_skb_change_tail_proto;
3555         case BPF_FUNC_skb_change_head:
3556                 return &bpf_skb_change_head_proto;
3557         case BPF_FUNC_skb_store_bytes:
3558                 return &bpf_skb_store_bytes_proto;
3559         case BPF_FUNC_csum_update:
3560                 return &bpf_csum_update_proto;
3561         case BPF_FUNC_l3_csum_replace:
3562                 return &bpf_l3_csum_replace_proto;
3563         case BPF_FUNC_l4_csum_replace:
3564                 return &bpf_l4_csum_replace_proto;
3565         case BPF_FUNC_set_hash_invalid:
3566                 return &bpf_set_hash_invalid_proto;
3567         default:
3568                 return lwt_inout_func_proto(func_id);
3569         }
3570 }
3571
3572 static bool bpf_skb_is_valid_access(int off, int size, enum bpf_access_type type,
3573                                     struct bpf_insn_access_aux *info)
3574 {
3575         const int size_default = sizeof(__u32);
3576
3577         if (off < 0 || off >= sizeof(struct __sk_buff))
3578                 return false;
3579
3580         /* The verifier guarantees that size > 0. */
3581         if (off % size != 0)
3582                 return false;
3583
3584         switch (off) {
3585         case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3586                 if (off + size > offsetofend(struct __sk_buff, cb[4]))
3587                         return false;
3588                 break;
3589         case bpf_ctx_range_till(struct __sk_buff, remote_ip6[0], remote_ip6[3]):
3590         case bpf_ctx_range_till(struct __sk_buff, local_ip6[0], local_ip6[3]):
3591         case bpf_ctx_range_till(struct __sk_buff, remote_ip4, remote_ip4):
3592         case bpf_ctx_range_till(struct __sk_buff, local_ip4, local_ip4):
3593         case bpf_ctx_range(struct __sk_buff, data):
3594         case bpf_ctx_range(struct __sk_buff, data_meta):
3595         case bpf_ctx_range(struct __sk_buff, data_end):
3596                 if (size != size_default)
3597                         return false;
3598                 break;
3599         default:
3600                 /* Only narrow read access allowed for now. */
3601                 if (type == BPF_WRITE) {
3602                         if (size != size_default)
3603                                 return false;
3604                 } else {
3605                         bpf_ctx_record_field_size(info, size_default);
3606                         if (!bpf_ctx_narrow_access_ok(off, size, size_default))
3607                                 return false;
3608                 }
3609         }
3610
3611         return true;
3612 }
3613
3614 static bool sk_filter_is_valid_access(int off, int size,
3615                                       enum bpf_access_type type,
3616                                       struct bpf_insn_access_aux *info)
3617 {
3618         switch (off) {
3619         case bpf_ctx_range(struct __sk_buff, tc_classid):
3620         case bpf_ctx_range(struct __sk_buff, data):
3621         case bpf_ctx_range(struct __sk_buff, data_meta):
3622         case bpf_ctx_range(struct __sk_buff, data_end):
3623         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3624                 return false;
3625         }
3626
3627         if (type == BPF_WRITE) {
3628                 switch (off) {
3629                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3630                         break;
3631                 default:
3632                         return false;
3633                 }
3634         }
3635
3636         return bpf_skb_is_valid_access(off, size, type, info);
3637 }
3638
3639 static bool lwt_is_valid_access(int off, int size,
3640                                 enum bpf_access_type type,
3641                                 struct bpf_insn_access_aux *info)
3642 {
3643         switch (off) {
3644         case bpf_ctx_range(struct __sk_buff, tc_classid):
3645         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3646         case bpf_ctx_range(struct __sk_buff, data_meta):
3647                 return false;
3648         }
3649
3650         if (type == BPF_WRITE) {
3651                 switch (off) {
3652                 case bpf_ctx_range(struct __sk_buff, mark):
3653                 case bpf_ctx_range(struct __sk_buff, priority):
3654                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3655                         break;
3656                 default:
3657                         return false;
3658                 }
3659         }
3660
3661         switch (off) {
3662         case bpf_ctx_range(struct __sk_buff, data):
3663                 info->reg_type = PTR_TO_PACKET;
3664                 break;
3665         case bpf_ctx_range(struct __sk_buff, data_end):
3666                 info->reg_type = PTR_TO_PACKET_END;
3667                 break;
3668         }
3669
3670         return bpf_skb_is_valid_access(off, size, type, info);
3671 }
3672
3673 static bool sock_filter_is_valid_access(int off, int size,
3674                                         enum bpf_access_type type,
3675                                         struct bpf_insn_access_aux *info)
3676 {
3677         if (type == BPF_WRITE) {
3678                 switch (off) {
3679                 case offsetof(struct bpf_sock, bound_dev_if):
3680                 case offsetof(struct bpf_sock, mark):
3681                 case offsetof(struct bpf_sock, priority):
3682                         break;
3683                 default:
3684                         return false;
3685                 }
3686         }
3687
3688         if (off < 0 || off + size > sizeof(struct bpf_sock))
3689                 return false;
3690         /* The verifier guarantees that size > 0. */
3691         if (off % size != 0)
3692                 return false;
3693         if (size != sizeof(__u32))
3694                 return false;
3695
3696         return true;
3697 }
3698
3699 static int bpf_unclone_prologue(struct bpf_insn *insn_buf, bool direct_write,
3700                                 const struct bpf_prog *prog, int drop_verdict)
3701 {
3702         struct bpf_insn *insn = insn_buf;
3703
3704         if (!direct_write)
3705                 return 0;
3706
3707         /* if (!skb->cloned)
3708          *       goto start;
3709          *
3710          * (Fast-path, otherwise approximation that we might be
3711          *  a clone, do the rest in helper.)
3712          */
3713         *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
3714         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
3715         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
3716
3717         /* ret = bpf_skb_pull_data(skb, 0); */
3718         *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
3719         *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
3720         *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
3721                                BPF_FUNC_skb_pull_data);
3722         /* if (!ret)
3723          *      goto restore;
3724          * return TC_ACT_SHOT;
3725          */
3726         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
3727         *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, drop_verdict);
3728         *insn++ = BPF_EXIT_INSN();
3729
3730         /* restore: */
3731         *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
3732         /* start: */
3733         *insn++ = prog->insnsi[0];
3734
3735         return insn - insn_buf;
3736 }
3737
3738 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
3739                                const struct bpf_prog *prog)
3740 {
3741         return bpf_unclone_prologue(insn_buf, direct_write, prog, TC_ACT_SHOT);
3742 }
3743
3744 static bool tc_cls_act_is_valid_access(int off, int size,
3745                                        enum bpf_access_type type,
3746                                        struct bpf_insn_access_aux *info)
3747 {
3748         if (type == BPF_WRITE) {
3749                 switch (off) {
3750                 case bpf_ctx_range(struct __sk_buff, mark):
3751                 case bpf_ctx_range(struct __sk_buff, tc_index):
3752                 case bpf_ctx_range(struct __sk_buff, priority):
3753                 case bpf_ctx_range(struct __sk_buff, tc_classid):
3754                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3755                         break;
3756                 default:
3757                         return false;
3758                 }
3759         }
3760
3761         switch (off) {
3762         case bpf_ctx_range(struct __sk_buff, data):
3763                 info->reg_type = PTR_TO_PACKET;
3764                 break;
3765         case bpf_ctx_range(struct __sk_buff, data_meta):
3766                 info->reg_type = PTR_TO_PACKET_META;
3767                 break;
3768         case bpf_ctx_range(struct __sk_buff, data_end):
3769                 info->reg_type = PTR_TO_PACKET_END;
3770                 break;
3771         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3772                 return false;
3773         }
3774
3775         return bpf_skb_is_valid_access(off, size, type, info);
3776 }
3777
3778 static bool __is_valid_xdp_access(int off, int size)
3779 {
3780         if (off < 0 || off >= sizeof(struct xdp_md))
3781                 return false;
3782         if (off % size != 0)
3783                 return false;
3784         if (size != sizeof(__u32))
3785                 return false;
3786
3787         return true;
3788 }
3789
3790 static bool xdp_is_valid_access(int off, int size,
3791                                 enum bpf_access_type type,
3792                                 struct bpf_insn_access_aux *info)
3793 {
3794         if (type == BPF_WRITE)
3795                 return false;
3796
3797         switch (off) {
3798         case offsetof(struct xdp_md, data):
3799                 info->reg_type = PTR_TO_PACKET;
3800                 break;
3801         case offsetof(struct xdp_md, data_meta):
3802                 info->reg_type = PTR_TO_PACKET_META;
3803                 break;
3804         case offsetof(struct xdp_md, data_end):
3805                 info->reg_type = PTR_TO_PACKET_END;
3806                 break;
3807         }
3808
3809         return __is_valid_xdp_access(off, size);
3810 }
3811
3812 void bpf_warn_invalid_xdp_action(u32 act)
3813 {
3814         const u32 act_max = XDP_REDIRECT;
3815
3816         WARN_ONCE(1, "%s XDP return value %u, expect packet loss!\n",
3817                   act > act_max ? "Illegal" : "Driver unsupported",
3818                   act);
3819 }
3820 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
3821
3822 static bool __is_valid_sock_ops_access(int off, int size)
3823 {
3824         if (off < 0 || off >= sizeof(struct bpf_sock_ops))
3825                 return false;
3826         /* The verifier guarantees that size > 0. */
3827         if (off % size != 0)
3828                 return false;
3829         if (size != sizeof(__u32))
3830                 return false;
3831
3832         return true;
3833 }
3834
3835 static bool sock_ops_is_valid_access(int off, int size,
3836                                      enum bpf_access_type type,
3837                                      struct bpf_insn_access_aux *info)
3838 {
3839         if (type == BPF_WRITE) {
3840                 switch (off) {
3841                 case offsetof(struct bpf_sock_ops, op) ...
3842                      offsetof(struct bpf_sock_ops, replylong[3]):
3843                         break;
3844                 default:
3845                         return false;
3846                 }
3847         }
3848
3849         return __is_valid_sock_ops_access(off, size);
3850 }
3851
3852 static int sk_skb_prologue(struct bpf_insn *insn_buf, bool direct_write,
3853                            const struct bpf_prog *prog)
3854 {
3855         return bpf_unclone_prologue(insn_buf, direct_write, prog, SK_DROP);
3856 }
3857
3858 static bool sk_skb_is_valid_access(int off, int size,
3859                                    enum bpf_access_type type,
3860                                    struct bpf_insn_access_aux *info)
3861 {
3862         switch (off) {
3863         case bpf_ctx_range(struct __sk_buff, tc_classid):
3864         case bpf_ctx_range(struct __sk_buff, data_meta):
3865                 return false;
3866         }
3867
3868         if (type == BPF_WRITE) {
3869                 switch (off) {
3870                 case bpf_ctx_range(struct __sk_buff, tc_index):
3871                 case bpf_ctx_range(struct __sk_buff, priority):
3872                         break;
3873                 default:
3874                         return false;
3875                 }
3876         }
3877
3878         switch (off) {
3879         case bpf_ctx_range(struct __sk_buff, mark):
3880                 return false;
3881         case bpf_ctx_range(struct __sk_buff, data):
3882                 info->reg_type = PTR_TO_PACKET;
3883                 break;
3884         case bpf_ctx_range(struct __sk_buff, data_end):
3885                 info->reg_type = PTR_TO_PACKET_END;
3886                 break;
3887         }
3888
3889         return bpf_skb_is_valid_access(off, size, type, info);
3890 }
3891
3892 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
3893                                   const struct bpf_insn *si,
3894                                   struct bpf_insn *insn_buf,
3895                                   struct bpf_prog *prog, u32 *target_size)
3896 {
3897         struct bpf_insn *insn = insn_buf;
3898         int off;
3899
3900         switch (si->off) {
3901         case offsetof(struct __sk_buff, len):
3902                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3903                                       bpf_target_off(struct sk_buff, len, 4,
3904                                                      target_size));
3905                 break;
3906
3907         case offsetof(struct __sk_buff, protocol):
3908                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3909                                       bpf_target_off(struct sk_buff, protocol, 2,
3910                                                      target_size));
3911                 break;
3912
3913         case offsetof(struct __sk_buff, vlan_proto):
3914                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3915                                       bpf_target_off(struct sk_buff, vlan_proto, 2,
3916                                                      target_size));
3917                 break;
3918
3919         case offsetof(struct __sk_buff, priority):
3920                 if (type == BPF_WRITE)
3921                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3922                                               bpf_target_off(struct sk_buff, priority, 4,
3923                                                              target_size));
3924                 else
3925                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3926                                               bpf_target_off(struct sk_buff, priority, 4,
3927                                                              target_size));
3928                 break;
3929
3930         case offsetof(struct __sk_buff, ingress_ifindex):
3931                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3932                                       bpf_target_off(struct sk_buff, skb_iif, 4,
3933                                                      target_size));
3934                 break;
3935
3936         case offsetof(struct __sk_buff, ifindex):
3937                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3938                                       si->dst_reg, si->src_reg,
3939                                       offsetof(struct sk_buff, dev));
3940                 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
3941                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3942                                       bpf_target_off(struct net_device, ifindex, 4,
3943                                                      target_size));
3944                 break;
3945
3946         case offsetof(struct __sk_buff, hash):
3947                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3948                                       bpf_target_off(struct sk_buff, hash, 4,
3949                                                      target_size));
3950                 break;
3951
3952         case offsetof(struct __sk_buff, mark):
3953                 if (type == BPF_WRITE)
3954                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3955                                               bpf_target_off(struct sk_buff, mark, 4,
3956                                                              target_size));
3957                 else
3958                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3959                                               bpf_target_off(struct sk_buff, mark, 4,
3960                                                              target_size));
3961                 break;
3962
3963         case offsetof(struct __sk_buff, pkt_type):
3964                 *target_size = 1;
3965                 *insn++ = BPF_LDX_MEM(BPF_B, si->dst_reg, si->src_reg,
3966                                       PKT_TYPE_OFFSET());
3967                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, PKT_TYPE_MAX);
3968 #ifdef __BIG_ENDIAN_BITFIELD
3969                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 5);
3970 #endif
3971                 break;
3972
3973         case offsetof(struct __sk_buff, queue_mapping):
3974                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3975                                       bpf_target_off(struct sk_buff, queue_mapping, 2,
3976                                                      target_size));
3977                 break;
3978
3979         case offsetof(struct __sk_buff, vlan_present):
3980         case offsetof(struct __sk_buff, vlan_tci):
3981                 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
3982
3983                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3984                                       bpf_target_off(struct sk_buff, vlan_tci, 2,
3985                                                      target_size));
3986                 if (si->off == offsetof(struct __sk_buff, vlan_tci)) {
3987                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg,
3988                                                 ~VLAN_TAG_PRESENT);
3989                 } else {
3990                         *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 12);
3991                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, 1);
3992                 }
3993                 break;
3994
3995         case offsetof(struct __sk_buff, cb[0]) ...
3996              offsetofend(struct __sk_buff, cb[4]) - 1:
3997                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
3998                 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
3999                               offsetof(struct qdisc_skb_cb, data)) %
4000                              sizeof(__u64));
4001
4002                 prog->cb_access = 1;
4003                 off  = si->off;
4004                 off -= offsetof(struct __sk_buff, cb[0]);
4005                 off += offsetof(struct sk_buff, cb);
4006                 off += offsetof(struct qdisc_skb_cb, data);
4007                 if (type == BPF_WRITE)
4008                         *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
4009                                               si->src_reg, off);
4010                 else
4011                         *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
4012                                               si->src_reg, off);
4013                 break;
4014
4015         case offsetof(struct __sk_buff, tc_classid):
4016                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
4017
4018                 off  = si->off;
4019                 off -= offsetof(struct __sk_buff, tc_classid);
4020                 off += offsetof(struct sk_buff, cb);
4021                 off += offsetof(struct qdisc_skb_cb, tc_classid);
4022                 *target_size = 2;
4023                 if (type == BPF_WRITE)
4024                         *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
4025                                               si->src_reg, off);
4026                 else
4027                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
4028                                               si->src_reg, off);
4029                 break;
4030
4031         case offsetof(struct __sk_buff, data):
4032                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
4033                                       si->dst_reg, si->src_reg,
4034                                       offsetof(struct sk_buff, data));
4035                 break;
4036
4037         case offsetof(struct __sk_buff, data_meta):
4038                 off  = si->off;
4039                 off -= offsetof(struct __sk_buff, data_meta);
4040                 off += offsetof(struct sk_buff, cb);
4041                 off += offsetof(struct bpf_skb_data_end, data_meta);
4042                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4043                                       si->src_reg, off);
4044                 break;
4045
4046         case offsetof(struct __sk_buff, data_end):
4047                 off  = si->off;
4048                 off -= offsetof(struct __sk_buff, data_end);
4049                 off += offsetof(struct sk_buff, cb);
4050                 off += offsetof(struct bpf_skb_data_end, data_end);
4051                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4052                                       si->src_reg, off);
4053                 break;
4054
4055         case offsetof(struct __sk_buff, tc_index):
4056 #ifdef CONFIG_NET_SCHED
4057                 if (type == BPF_WRITE)
4058                         *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg, si->src_reg,
4059                                               bpf_target_off(struct sk_buff, tc_index, 2,
4060                                                              target_size));
4061                 else
4062                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4063                                               bpf_target_off(struct sk_buff, tc_index, 2,
4064                                                              target_size));
4065 #else
4066                 *target_size = 2;
4067                 if (type == BPF_WRITE)
4068                         *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
4069                 else
4070                         *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4071 #endif
4072                 break;
4073
4074         case offsetof(struct __sk_buff, napi_id):
4075 #if defined(CONFIG_NET_RX_BUSY_POLL)
4076                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4077                                       bpf_target_off(struct sk_buff, napi_id, 4,
4078                                                      target_size));
4079                 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
4080                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4081 #else
4082                 *target_size = 4;
4083                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4084 #endif
4085                 break;
4086         case offsetof(struct __sk_buff, family):
4087                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4088
4089                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4090                                       si->dst_reg, si->src_reg,
4091                                       offsetof(struct sk_buff, sk));
4092                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4093                                       bpf_target_off(struct sock_common,
4094                                                      skc_family,
4095                                                      2, target_size));
4096                 break;
4097         case offsetof(struct __sk_buff, remote_ip4):
4098                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4099
4100                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4101                                       si->dst_reg, si->src_reg,
4102                                       offsetof(struct sk_buff, sk));
4103                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4104                                       bpf_target_off(struct sock_common,
4105                                                      skc_daddr,
4106                                                      4, target_size));
4107                 break;
4108         case offsetof(struct __sk_buff, local_ip4):
4109                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4110                                           skc_rcv_saddr) != 4);
4111
4112                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4113                                       si->dst_reg, si->src_reg,
4114                                       offsetof(struct sk_buff, sk));
4115                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4116                                       bpf_target_off(struct sock_common,
4117                                                      skc_rcv_saddr,
4118                                                      4, target_size));
4119                 break;
4120         case offsetof(struct __sk_buff, remote_ip6[0]) ...
4121              offsetof(struct __sk_buff, remote_ip6[3]):
4122 #if IS_ENABLED(CONFIG_IPV6)
4123                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4124                                           skc_v6_daddr.s6_addr32[0]) != 4);
4125
4126                 off = si->off;
4127                 off -= offsetof(struct __sk_buff, remote_ip6[0]);
4128
4129                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4130                                       si->dst_reg, si->src_reg,
4131                                       offsetof(struct sk_buff, sk));
4132                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4133                                       offsetof(struct sock_common,
4134                                                skc_v6_daddr.s6_addr32[0]) +
4135                                       off);
4136 #else
4137                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4138 #endif
4139                 break;
4140         case offsetof(struct __sk_buff, local_ip6[0]) ...
4141              offsetof(struct __sk_buff, local_ip6[3]):
4142 #if IS_ENABLED(CONFIG_IPV6)
4143                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4144                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4145
4146                 off = si->off;
4147                 off -= offsetof(struct __sk_buff, local_ip6[0]);
4148
4149                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4150                                       si->dst_reg, si->src_reg,
4151                                       offsetof(struct sk_buff, sk));
4152                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4153                                       offsetof(struct sock_common,
4154                                                skc_v6_rcv_saddr.s6_addr32[0]) +
4155                                       off);
4156 #else
4157                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4158 #endif
4159                 break;
4160
4161         case offsetof(struct __sk_buff, remote_port):
4162                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4163
4164                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4165                                       si->dst_reg, si->src_reg,
4166                                       offsetof(struct sk_buff, sk));
4167                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4168                                       bpf_target_off(struct sock_common,
4169                                                      skc_dport,
4170                                                      2, target_size));
4171 #ifndef __BIG_ENDIAN_BITFIELD
4172                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4173 #endif
4174                 break;
4175
4176         case offsetof(struct __sk_buff, local_port):
4177                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4178
4179                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4180                                       si->dst_reg, si->src_reg,
4181                                       offsetof(struct sk_buff, sk));
4182                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4183                                       bpf_target_off(struct sock_common,
4184                                                      skc_num, 2, target_size));
4185                 break;
4186         }
4187
4188         return insn - insn_buf;
4189 }
4190
4191 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
4192                                           const struct bpf_insn *si,
4193                                           struct bpf_insn *insn_buf,
4194                                           struct bpf_prog *prog, u32 *target_size)
4195 {
4196         struct bpf_insn *insn = insn_buf;
4197
4198         switch (si->off) {
4199         case offsetof(struct bpf_sock, bound_dev_if):
4200                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
4201
4202                 if (type == BPF_WRITE)
4203                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4204                                         offsetof(struct sock, sk_bound_dev_if));
4205                 else
4206                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4207                                       offsetof(struct sock, sk_bound_dev_if));
4208                 break;
4209
4210         case offsetof(struct bpf_sock, mark):
4211                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_mark) != 4);
4212
4213                 if (type == BPF_WRITE)
4214                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4215                                         offsetof(struct sock, sk_mark));
4216                 else
4217                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4218                                       offsetof(struct sock, sk_mark));
4219                 break;
4220
4221         case offsetof(struct bpf_sock, priority):
4222                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_priority) != 4);
4223
4224                 if (type == BPF_WRITE)
4225                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4226                                         offsetof(struct sock, sk_priority));
4227                 else
4228                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4229                                       offsetof(struct sock, sk_priority));
4230                 break;
4231
4232         case offsetof(struct bpf_sock, family):
4233                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
4234
4235                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4236                                       offsetof(struct sock, sk_family));
4237                 break;
4238
4239         case offsetof(struct bpf_sock, type):
4240                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4241                                       offsetof(struct sock, __sk_flags_offset));
4242                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
4243                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
4244                 break;
4245
4246         case offsetof(struct bpf_sock, protocol):
4247                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4248                                       offsetof(struct sock, __sk_flags_offset));
4249                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
4250                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
4251                 break;
4252         }
4253
4254         return insn - insn_buf;
4255 }
4256
4257 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
4258                                          const struct bpf_insn *si,
4259                                          struct bpf_insn *insn_buf,
4260                                          struct bpf_prog *prog, u32 *target_size)
4261 {
4262         struct bpf_insn *insn = insn_buf;
4263
4264         switch (si->off) {
4265         case offsetof(struct __sk_buff, ifindex):
4266                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
4267                                       si->dst_reg, si->src_reg,
4268                                       offsetof(struct sk_buff, dev));
4269                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4270                                       bpf_target_off(struct net_device, ifindex, 4,
4271                                                      target_size));
4272                 break;
4273         default:
4274                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4275                                               target_size);
4276         }
4277
4278         return insn - insn_buf;
4279 }
4280
4281 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
4282                                   const struct bpf_insn *si,
4283                                   struct bpf_insn *insn_buf,
4284                                   struct bpf_prog *prog, u32 *target_size)
4285 {
4286         struct bpf_insn *insn = insn_buf;
4287
4288         switch (si->off) {
4289         case offsetof(struct xdp_md, data):
4290                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
4291                                       si->dst_reg, si->src_reg,
4292                                       offsetof(struct xdp_buff, data));
4293                 break;
4294         case offsetof(struct xdp_md, data_meta):
4295                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_meta),
4296                                       si->dst_reg, si->src_reg,
4297                                       offsetof(struct xdp_buff, data_meta));
4298                 break;
4299         case offsetof(struct xdp_md, data_end):
4300                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
4301                                       si->dst_reg, si->src_reg,
4302                                       offsetof(struct xdp_buff, data_end));
4303                 break;
4304         }
4305
4306         return insn - insn_buf;
4307 }
4308
4309 static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
4310                                        const struct bpf_insn *si,
4311                                        struct bpf_insn *insn_buf,
4312                                        struct bpf_prog *prog,
4313                                        u32 *target_size)
4314 {
4315         struct bpf_insn *insn = insn_buf;
4316         int off;
4317
4318         switch (si->off) {
4319         case offsetof(struct bpf_sock_ops, op) ...
4320              offsetof(struct bpf_sock_ops, replylong[3]):
4321                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, op) !=
4322                              FIELD_SIZEOF(struct bpf_sock_ops_kern, op));
4323                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, reply) !=
4324                              FIELD_SIZEOF(struct bpf_sock_ops_kern, reply));
4325                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, replylong) !=
4326                              FIELD_SIZEOF(struct bpf_sock_ops_kern, replylong));
4327                 off = si->off;
4328                 off -= offsetof(struct bpf_sock_ops, op);
4329                 off += offsetof(struct bpf_sock_ops_kern, op);
4330                 if (type == BPF_WRITE)
4331                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4332                                               off);
4333                 else
4334                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4335                                               off);
4336                 break;
4337
4338         case offsetof(struct bpf_sock_ops, family):
4339                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4340
4341                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4342                                               struct bpf_sock_ops_kern, sk),
4343                                       si->dst_reg, si->src_reg,
4344                                       offsetof(struct bpf_sock_ops_kern, sk));
4345                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4346                                       offsetof(struct sock_common, skc_family));
4347                 break;
4348
4349         case offsetof(struct bpf_sock_ops, remote_ip4):
4350                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4351
4352                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4353                                                 struct bpf_sock_ops_kern, sk),
4354                                       si->dst_reg, si->src_reg,
4355                                       offsetof(struct bpf_sock_ops_kern, sk));
4356                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4357                                       offsetof(struct sock_common, skc_daddr));
4358                 break;
4359
4360         case offsetof(struct bpf_sock_ops, local_ip4):
4361                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_rcv_saddr) != 4);
4362
4363                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4364                                               struct bpf_sock_ops_kern, sk),
4365                                       si->dst_reg, si->src_reg,
4366                                       offsetof(struct bpf_sock_ops_kern, sk));
4367                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4368                                       offsetof(struct sock_common,
4369                                                skc_rcv_saddr));
4370                 break;
4371
4372         case offsetof(struct bpf_sock_ops, remote_ip6[0]) ...
4373              offsetof(struct bpf_sock_ops, remote_ip6[3]):
4374 #if IS_ENABLED(CONFIG_IPV6)
4375                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4376                                           skc_v6_daddr.s6_addr32[0]) != 4);
4377
4378                 off = si->off;
4379                 off -= offsetof(struct bpf_sock_ops, remote_ip6[0]);
4380                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4381                                                 struct bpf_sock_ops_kern, sk),
4382                                       si->dst_reg, si->src_reg,
4383                                       offsetof(struct bpf_sock_ops_kern, sk));
4384                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4385                                       offsetof(struct sock_common,
4386                                                skc_v6_daddr.s6_addr32[0]) +
4387                                       off);
4388 #else
4389                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4390 #endif
4391                 break;
4392
4393         case offsetof(struct bpf_sock_ops, local_ip6[0]) ...
4394              offsetof(struct bpf_sock_ops, local_ip6[3]):
4395 #if IS_ENABLED(CONFIG_IPV6)
4396                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4397                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4398
4399                 off = si->off;
4400                 off -= offsetof(struct bpf_sock_ops, local_ip6[0]);
4401                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4402                                                 struct bpf_sock_ops_kern, sk),
4403                                       si->dst_reg, si->src_reg,
4404                                       offsetof(struct bpf_sock_ops_kern, sk));
4405                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4406                                       offsetof(struct sock_common,
4407                                                skc_v6_rcv_saddr.s6_addr32[0]) +
4408                                       off);
4409 #else
4410                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4411 #endif
4412                 break;
4413
4414         case offsetof(struct bpf_sock_ops, remote_port):
4415                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4416
4417                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4418                                                 struct bpf_sock_ops_kern, sk),
4419                                       si->dst_reg, si->src_reg,
4420                                       offsetof(struct bpf_sock_ops_kern, sk));
4421                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4422                                       offsetof(struct sock_common, skc_dport));
4423 #ifndef __BIG_ENDIAN_BITFIELD
4424                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4425 #endif
4426                 break;
4427
4428         case offsetof(struct bpf_sock_ops, local_port):
4429                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4430
4431                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4432                                                 struct bpf_sock_ops_kern, sk),
4433                                       si->dst_reg, si->src_reg,
4434                                       offsetof(struct bpf_sock_ops_kern, sk));
4435                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4436                                       offsetof(struct sock_common, skc_num));
4437                 break;
4438         }
4439         return insn - insn_buf;
4440 }
4441
4442 static u32 sk_skb_convert_ctx_access(enum bpf_access_type type,
4443                                      const struct bpf_insn *si,
4444                                      struct bpf_insn *insn_buf,
4445                                      struct bpf_prog *prog, u32 *target_size)
4446 {
4447         struct bpf_insn *insn = insn_buf;
4448         int off;
4449
4450         switch (si->off) {
4451         case offsetof(struct __sk_buff, data_end):
4452                 off  = si->off;
4453                 off -= offsetof(struct __sk_buff, data_end);
4454                 off += offsetof(struct sk_buff, cb);
4455                 off += offsetof(struct tcp_skb_cb, bpf.data_end);
4456                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4457                                       si->src_reg, off);
4458                 break;
4459         default:
4460                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4461                                               target_size);
4462         }
4463
4464         return insn - insn_buf;
4465 }
4466
4467 const struct bpf_verifier_ops sk_filter_verifier_ops = {
4468         .get_func_proto         = sk_filter_func_proto,
4469         .is_valid_access        = sk_filter_is_valid_access,
4470         .convert_ctx_access     = bpf_convert_ctx_access,
4471 };
4472
4473 const struct bpf_prog_ops sk_filter_prog_ops = {
4474 };
4475
4476 const struct bpf_verifier_ops tc_cls_act_verifier_ops = {
4477         .get_func_proto         = tc_cls_act_func_proto,
4478         .is_valid_access        = tc_cls_act_is_valid_access,
4479         .convert_ctx_access     = tc_cls_act_convert_ctx_access,
4480         .gen_prologue           = tc_cls_act_prologue,
4481 };
4482
4483 const struct bpf_prog_ops tc_cls_act_prog_ops = {
4484         .test_run               = bpf_prog_test_run_skb,
4485 };
4486
4487 const struct bpf_verifier_ops xdp_verifier_ops = {
4488         .get_func_proto         = xdp_func_proto,
4489         .is_valid_access        = xdp_is_valid_access,
4490         .convert_ctx_access     = xdp_convert_ctx_access,
4491 };
4492
4493 const struct bpf_prog_ops xdp_prog_ops = {
4494         .test_run               = bpf_prog_test_run_xdp,
4495 };
4496
4497 const struct bpf_verifier_ops cg_skb_verifier_ops = {
4498         .get_func_proto         = sk_filter_func_proto,
4499         .is_valid_access        = sk_filter_is_valid_access,
4500         .convert_ctx_access     = bpf_convert_ctx_access,
4501 };
4502
4503 const struct bpf_prog_ops cg_skb_prog_ops = {
4504         .test_run               = bpf_prog_test_run_skb,
4505 };
4506
4507 const struct bpf_verifier_ops lwt_inout_verifier_ops = {
4508         .get_func_proto         = lwt_inout_func_proto,
4509         .is_valid_access        = lwt_is_valid_access,
4510         .convert_ctx_access     = bpf_convert_ctx_access,
4511 };
4512
4513 const struct bpf_prog_ops lwt_inout_prog_ops = {
4514         .test_run               = bpf_prog_test_run_skb,
4515 };
4516
4517 const struct bpf_verifier_ops lwt_xmit_verifier_ops = {
4518         .get_func_proto         = lwt_xmit_func_proto,
4519         .is_valid_access        = lwt_is_valid_access,
4520         .convert_ctx_access     = bpf_convert_ctx_access,
4521         .gen_prologue           = tc_cls_act_prologue,
4522 };
4523
4524 const struct bpf_prog_ops lwt_xmit_prog_ops = {
4525         .test_run               = bpf_prog_test_run_skb,
4526 };
4527
4528 const struct bpf_verifier_ops cg_sock_verifier_ops = {
4529         .get_func_proto         = sock_filter_func_proto,
4530         .is_valid_access        = sock_filter_is_valid_access,
4531         .convert_ctx_access     = sock_filter_convert_ctx_access,
4532 };
4533
4534 const struct bpf_prog_ops cg_sock_prog_ops = {
4535 };
4536
4537 const struct bpf_verifier_ops sock_ops_verifier_ops = {
4538         .get_func_proto         = sock_ops_func_proto,
4539         .is_valid_access        = sock_ops_is_valid_access,
4540         .convert_ctx_access     = sock_ops_convert_ctx_access,
4541 };
4542
4543 const struct bpf_prog_ops sock_ops_prog_ops = {
4544 };
4545
4546 const struct bpf_verifier_ops sk_skb_verifier_ops = {
4547         .get_func_proto         = sk_skb_func_proto,
4548         .is_valid_access        = sk_skb_is_valid_access,
4549         .convert_ctx_access     = sk_skb_convert_ctx_access,
4550         .gen_prologue           = sk_skb_prologue,
4551 };
4552
4553 const struct bpf_prog_ops sk_skb_prog_ops = {
4554 };
4555
4556 int sk_detach_filter(struct sock *sk)
4557 {
4558         int ret = -ENOENT;
4559         struct sk_filter *filter;
4560
4561         if (sock_flag(sk, SOCK_FILTER_LOCKED))
4562                 return -EPERM;
4563
4564         filter = rcu_dereference_protected(sk->sk_filter,
4565                                            lockdep_sock_is_held(sk));
4566         if (filter) {
4567                 RCU_INIT_POINTER(sk->sk_filter, NULL);
4568                 sk_filter_uncharge(sk, filter);
4569                 ret = 0;
4570         }
4571
4572         return ret;
4573 }
4574 EXPORT_SYMBOL_GPL(sk_detach_filter);
4575
4576 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
4577                   unsigned int len)
4578 {
4579         struct sock_fprog_kern *fprog;
4580         struct sk_filter *filter;
4581         int ret = 0;
4582
4583         lock_sock(sk);
4584         filter = rcu_dereference_protected(sk->sk_filter,
4585                                            lockdep_sock_is_held(sk));
4586         if (!filter)
4587                 goto out;
4588
4589         /* We're copying the filter that has been originally attached,
4590          * so no conversion/decode needed anymore. eBPF programs that
4591          * have no original program cannot be dumped through this.
4592          */
4593         ret = -EACCES;
4594         fprog = filter->prog->orig_prog;
4595         if (!fprog)
4596                 goto out;
4597
4598         ret = fprog->len;
4599         if (!len)
4600                 /* User space only enquires number of filter blocks. */
4601                 goto out;
4602
4603         ret = -EINVAL;
4604         if (len < fprog->len)
4605                 goto out;
4606
4607         ret = -EFAULT;
4608         if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
4609                 goto out;
4610
4611         /* Instead of bytes, the API requests to return the number
4612          * of filter blocks.
4613          */
4614         ret = fprog->len;
4615 out:
4616         release_sock(sk);
4617         return ret;
4618 }