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                         if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) ||
 462                             fp->code == (BPF_ALU | BPF_MOD | BPF_X))
 463                                 *insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X);
 464
 465                         *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
 466                         break;
 467
 468                 /* Jump transformation cannot use BPF block macros
 469                  * everywhere as offset calculation and target updates
 470                  * require a bit more work than the rest, i.e. jump
 471                  * opcodes map as-is, but offsets need adjustment.
 472                  */
 473
 474 #define BPF_EMIT_JMP                                                    \
 475         do {                                                            \
 476                 if (target >= len || target < 0)                        \
 477                         goto err;                                       \
 478                 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0;   \
 479                 /* Adjust pc relative offset for 2nd or 3rd insn. */    \
 480                 insn->off -= insn - tmp_insns;                          \
 481         } while (0)
 482
 483                 case BPF_JMP | BPF_JA:
 484                         target = i + fp->k + 1;
 485                         insn->code = fp->code;
 486                         BPF_EMIT_JMP;
 487                         break;
 488
 489                 case BPF_JMP | BPF_JEQ | BPF_K:
 490                 case BPF_JMP | BPF_JEQ | BPF_X:
 491                 case BPF_JMP | BPF_JSET | BPF_K:
 492                 case BPF_JMP | BPF_JSET | BPF_X:
 493                 case BPF_JMP | BPF_JGT | BPF_K:
 494                 case BPF_JMP | BPF_JGT | BPF_X:
 495                 case BPF_JMP | BPF_JGE | BPF_K:
 496                 case BPF_JMP | BPF_JGE | BPF_X:
 497                         if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
 498                                 /* BPF immediates are signed, zero extend
 499                                  * immediate into tmp register and use it
 500                                  * in compare insn.
 501                                  */
 502                                 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
 503
 504                                 insn->dst_reg = BPF_REG_A;
 505                                 insn->src_reg = BPF_REG_TMP;
 506                                 bpf_src = BPF_X;
 507                         } else {
 508                                 insn->dst_reg = BPF_REG_A;
 509                                 insn->imm = fp->k;
 510                                 bpf_src = BPF_SRC(fp->code);
 511                                 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
 512                         }
 513
 514                         /* Common case where 'jump_false' is next insn. */
 515                         if (fp->jf == 0) {
 516                                 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 517                                 target = i + fp->jt + 1;
 518                                 BPF_EMIT_JMP;
 519                                 break;
 520                         }
 521
 522                         /* Convert some jumps when 'jump_true' is next insn. */
 523                         if (fp->jt == 0) {
 524                                 switch (BPF_OP(fp->code)) {
 525                                 case BPF_JEQ:
 526                                         insn->code = BPF_JMP | BPF_JNE | bpf_src;
 527                                         break;
 528                                 case BPF_JGT:
 529                                         insn->code = BPF_JMP | BPF_JLE | bpf_src;
 530                                         break;
 531                                 case BPF_JGE:
 532                                         insn->code = BPF_JMP | BPF_JLT | bpf_src;
 533                                         break;
 534                                 default:
 535                                         goto jmp_rest;
 536                                 }
 537
 538                                 target = i + fp->jf + 1;
 539                                 BPF_EMIT_JMP;
 540                                 break;
 541                         }
 542 jmp_rest:
 543                         /* Other jumps are mapped into two insns: Jxx and JA. */
 544                         target = i + fp->jt + 1;
 545                         insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 546                         BPF_EMIT_JMP;
 547                         insn++;
 548
 549                         insn->code = BPF_JMP | BPF_JA;
 550                         target = i + fp->jf + 1;
 551                         BPF_EMIT_JMP;
 552                         break;
 553
 554                 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
 555                 case BPF_LDX | BPF_MSH | BPF_B:
 556                         /* tmp = A */
 557                         *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
 558                         /* A = BPF_R0 = *(u8 *) (skb->data + K) */
 559                         *insn++ = BPF_LD_ABS(BPF_B, fp->k);
 560                         /* A &= 0xf */
 561                         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
 562                         /* A <<= 2 */
 563                         *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
 564                         /* X = A */
 565                         *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 566                         /* A = tmp */
 567                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
 568                         break;
 569
 570                 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
 571                  * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
 572                  */
 573                 case BPF_RET | BPF_A:
 574                 case BPF_RET | BPF_K:
 575                         if (BPF_RVAL(fp->code) == BPF_K)
 576                                 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
 577                                                         0, fp->k);
 578                         *insn = BPF_EXIT_INSN();
 579                         break;
 580
 581                 /* Store to stack. */
 582                 case BPF_ST:
 583                 case BPF_STX:
 584                         stack_off = fp->k * 4  + 4;
 585                         *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
 586                                             BPF_ST ? BPF_REG_A : BPF_REG_X,
 587                                             -stack_off);
 588                         /* check_load_and_stores() verifies that classic BPF can
 589                          * load from stack only after write, so tracking
 590                          * stack_depth for ST|STX insns is enough
 591                          */
 592                         if (new_prog && new_prog->aux->stack_depth < stack_off)
 593                                 new_prog->aux->stack_depth = stack_off;
 594                         break;
 595
 596                 /* Load from stack. */
 597                 case BPF_LD | BPF_MEM:
 598                 case BPF_LDX | BPF_MEM:
 599                         stack_off = fp->k * 4  + 4;
 600                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD  ?
 601                                             BPF_REG_A : BPF_REG_X, BPF_REG_FP,
 602                                             -stack_off);
 603                         break;
 604
 605                 /* A = K or X = K */
 606                 case BPF_LD | BPF_IMM:
 607                 case BPF_LDX | BPF_IMM:
 608                         *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
 609                                               BPF_REG_A : BPF_REG_X, fp->k);
 610                         break;
 611
 612                 /* X = A */
 613                 case BPF_MISC | BPF_TAX:
 614                         *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 615                         break;
 616
 617                 /* A = X */
 618                 case BPF_MISC | BPF_TXA:
 619                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
 620                         break;
 621
 622                 /* A = skb->len or X = skb->len */
 623                 case BPF_LD | BPF_W | BPF_LEN:
 624                 case BPF_LDX | BPF_W | BPF_LEN:
 625                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
 626                                             BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
 627                                             offsetof(struct sk_buff, len));
 628                         break;
 629
 630                 /* Access seccomp_data fields. */
 631                 case BPF_LDX | BPF_ABS | BPF_W:
 632                         /* A = *(u32 *) (ctx + K) */
 633                         *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
 634                         break;
 635
 636                 /* Unknown instruction. */
 637                 default:
 638                         goto err;
 639                 }
 640
 641                 insn++;
 642                 if (new_prog)
 643                         memcpy(new_insn, tmp_insns,
 644                                sizeof(*insn) * (insn - tmp_insns));
 645                 new_insn += insn - tmp_insns;
 646         }
 647
 648         if (!new_prog) {
 649                 /* Only calculating new length. */
 650                 *new_len = new_insn - first_insn;
 651                 return 0;
 652         }
 653
 654         pass++;
 655         if (new_flen != new_insn - first_insn) {
 656                 new_flen = new_insn - first_insn;
 657                 if (pass > 2)
 658                         goto err;
 659                 goto do_pass;
 660         }
 661
 662         kfree(addrs);
 663         BUG_ON(*new_len != new_flen);
 664         return 0;
 665 err:
 666         kfree(addrs);
 667         return -EINVAL;
 668 }
 669
 670 /* Security:
 671  *
 672  * As we dont want to clear mem[] array for each packet going through
 673  * __bpf_prog_run(), we check that filter loaded by user never try to read
 674  * a cell if not previously written, and we check all branches to be sure
 675  * a malicious user doesn't try to abuse us.
 676  */
 677 static int check_load_and_stores(const struct sock_filter *filter, int flen)
 678 {
 679         u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
 680         int pc, ret = 0;
 681
 682         BUILD_BUG_ON(BPF_MEMWORDS > 16);
 683
 684         masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
 685         if (!masks)
 686                 return -ENOMEM;
 687
 688         memset(masks, 0xff, flen * sizeof(*masks));
 689
 690         for (pc = 0; pc < flen; pc++) {
 691                 memvalid &= masks[pc];
 692
 693                 switch (filter[pc].code) {
 694                 case BPF_ST:
 695                 case BPF_STX:
 696                         memvalid |= (1 << filter[pc].k);
 697                         break;
 698                 case BPF_LD | BPF_MEM:
 699                 case BPF_LDX | BPF_MEM:
 700                         if (!(memvalid & (1 << filter[pc].k))) {
 701                                 ret = -EINVAL;
 702                                 goto error;
 703                         }
 704                         break;
 705                 case BPF_JMP | BPF_JA:
 706                         /* A jump must set masks on target */
 707                         masks[pc + 1 + filter[pc].k] &= memvalid;
 708                         memvalid = ~0;
 709                         break;
 710                 case BPF_JMP | BPF_JEQ | BPF_K:
 711                 case BPF_JMP | BPF_JEQ | BPF_X:
 712                 case BPF_JMP | BPF_JGE | BPF_K:
 713                 case BPF_JMP | BPF_JGE | BPF_X:
 714                 case BPF_JMP | BPF_JGT | BPF_K:
 715                 case BPF_JMP | BPF_JGT | BPF_X:
 716                 case BPF_JMP | BPF_JSET | BPF_K:
 717                 case BPF_JMP | BPF_JSET | BPF_X:
 718                         /* A jump must set masks on targets */
 719                         masks[pc + 1 + filter[pc].jt] &= memvalid;
 720                         masks[pc + 1 + filter[pc].jf] &= memvalid;
 721                         memvalid = ~0;
 722                         break;
 723                 }
 724         }
 725 error:
 726         kfree(masks);
 727         return ret;
 728 }
 729
 730 static bool chk_code_allowed(u16 code_to_probe)
 731 {
 732         static const bool codes[] = {
 733                 /* 32 bit ALU operations */
 734                 [BPF_ALU | BPF_ADD | BPF_K] = true,
 735                 [BPF_ALU | BPF_ADD | BPF_X] = true,
 736                 [BPF_ALU | BPF_SUB | BPF_K] = true,
 737                 [BPF_ALU | BPF_SUB | BPF_X] = true,
 738                 [BPF_ALU | BPF_MUL | BPF_K] = true,
 739                 [BPF_ALU | BPF_MUL | BPF_X] = true,
 740                 [BPF_ALU | BPF_DIV | BPF_K] = true,
 741                 [BPF_ALU | BPF_DIV | BPF_X] = true,
 742                 [BPF_ALU | BPF_MOD | BPF_K] = true,
 743                 [BPF_ALU | BPF_MOD | BPF_X] = true,
 744                 [BPF_ALU | BPF_AND | BPF_K] = true,
 745                 [BPF_ALU | BPF_AND | BPF_X] = true,
 746                 [BPF_ALU | BPF_OR | BPF_K] = true,
 747                 [BPF_ALU | BPF_OR | BPF_X] = true,
 748                 [BPF_ALU | BPF_XOR | BPF_K] = true,
 749                 [BPF_ALU | BPF_XOR | BPF_X] = true,
 750                 [BPF_ALU | BPF_LSH | BPF_K] = true,
 751                 [BPF_ALU | BPF_LSH | BPF_X] = true,
 752                 [BPF_ALU | BPF_RSH | BPF_K] = true,
 753                 [BPF_ALU | BPF_RSH | BPF_X] = true,
 754                 [BPF_ALU | BPF_NEG] = true,
 755                 /* Load instructions */
 756                 [BPF_LD | BPF_W | BPF_ABS] = true,
 757                 [BPF_LD | BPF_H | BPF_ABS] = true,
 758                 [BPF_LD | BPF_B | BPF_ABS] = true,
 759                 [BPF_LD | BPF_W | BPF_LEN] = true,
 760                 [BPF_LD | BPF_W | BPF_IND] = true,
 761                 [BPF_LD | BPF_H | BPF_IND] = true,
 762                 [BPF_LD | BPF_B | BPF_IND] = true,
 763                 [BPF_LD | BPF_IMM] = true,
 764                 [BPF_LD | BPF_MEM] = true,
 765                 [BPF_LDX | BPF_W | BPF_LEN] = true,
 766                 [BPF_LDX | BPF_B | BPF_MSH] = true,
 767                 [BPF_LDX | BPF_IMM] = true,
 768                 [BPF_LDX | BPF_MEM] = true,
 769                 /* Store instructions */
 770                 [BPF_ST] = true,
 771                 [BPF_STX] = true,
 772                 /* Misc instructions */
 773                 [BPF_MISC | BPF_TAX] = true,
 774                 [BPF_MISC | BPF_TXA] = true,
 775                 /* Return instructions */
 776                 [BPF_RET | BPF_K] = true,
 777                 [BPF_RET | BPF_A] = true,
 778                 /* Jump instructions */
 779                 [BPF_JMP | BPF_JA] = true,
 780                 [BPF_JMP | BPF_JEQ | BPF_K] = true,
 781                 [BPF_JMP | BPF_JEQ | BPF_X] = true,
 782                 [BPF_JMP | BPF_JGE | BPF_K] = true,
 783                 [BPF_JMP | BPF_JGE | BPF_X] = true,
 784                 [BPF_JMP | BPF_JGT | BPF_K] = true,
 785                 [BPF_JMP | BPF_JGT | BPF_X] = true,
 786                 [BPF_JMP | BPF_JSET | BPF_K] = true,
 787                 [BPF_JMP | BPF_JSET | BPF_X] = true,
 788         };
 789
 790         if (code_to_probe >= ARRAY_SIZE(codes))
 791                 return false;
 792
 793         return codes[code_to_probe];
 794 }
 795
 796 static bool bpf_check_basics_ok(const struct sock_filter *filter,
 797                                 unsigned int flen)
 798 {
 799         if (filter == NULL)
 800                 return false;
 801         if (flen == 0 || flen > BPF_MAXINSNS)
 802                 return false;
 803
 804         return true;
 805 }
 806
 807 /**
 808  *      bpf_check_classic - verify socket filter code
 809  *      @filter: filter to verify
 810  *      @flen: length of filter
 811  *
 812  * Check the user's filter code. If we let some ugly
 813  * filter code slip through kaboom! The filter must contain
 814  * no references or jumps that are out of range, no illegal
 815  * instructions, and must end with a RET instruction.
 816  *
 817  * All jumps are forward as they are not signed.
 818  *
 819  * Returns 0 if the rule set is legal or -EINVAL if not.
 820  */
 821 static int bpf_check_classic(const struct sock_filter *filter,
 822                              unsigned int flen)
 823 {
 824         bool anc_found;
 825         int pc;
 826
 827         /* Check the filter code now */
 828         for (pc = 0; pc < flen; pc++) {
 829                 const struct sock_filter *ftest = &filter[pc];
 830
 831                 /* May we actually operate on this code? */
 832                 if (!chk_code_allowed(ftest->code))
 833                         return -EINVAL;
 834
 835                 /* Some instructions need special checks */
 836                 switch (ftest->code) {
 837                 case BPF_ALU | BPF_DIV | BPF_K:
 838                 case BPF_ALU | BPF_MOD | BPF_K:
 839                         /* Check for division by zero */
 840                         if (ftest->k == 0)
 841                                 return -EINVAL;
 842                         break;
 843                 case BPF_ALU | BPF_LSH | BPF_K:
 844                 case BPF_ALU | BPF_RSH | BPF_K:
 845                         if (ftest->k >= 32)
 846                                 return -EINVAL;
 847                         break;
 848                 case BPF_LD | BPF_MEM:
 849                 case BPF_LDX | BPF_MEM:
 850                 case BPF_ST:
 851                 case BPF_STX:
 852                         /* Check for invalid memory addresses */
 853                         if (ftest->k >= BPF_MEMWORDS)
 854                                 return -EINVAL;
 855                         break;
 856                 case BPF_JMP | BPF_JA:
 857                         /* Note, the large ftest->k might cause loops.
 858                          * Compare this with conditional jumps below,
 859                          * where offsets are limited. --ANK (981016)
 860                          */
 861                         if (ftest->k >= (unsigned int)(flen - pc - 1))
 862                                 return -EINVAL;
 863                         break;
 864                 case BPF_JMP | BPF_JEQ | BPF_K:
 865                 case BPF_JMP | BPF_JEQ | BPF_X:
 866                 case BPF_JMP | BPF_JGE | BPF_K:
 867                 case BPF_JMP | BPF_JGE | BPF_X:
 868                 case BPF_JMP | BPF_JGT | BPF_K:
 869                 case BPF_JMP | BPF_JGT | BPF_X:
 870                 case BPF_JMP | BPF_JSET | BPF_K:
 871                 case BPF_JMP | BPF_JSET | BPF_X:
 872                         /* Both conditionals must be safe */
 873                         if (pc + ftest->jt + 1 >= flen ||
 874                             pc + ftest->jf + 1 >= flen)
 875                                 return -EINVAL;
 876                         break;
 877                 case BPF_LD | BPF_W | BPF_ABS:
 878                 case BPF_LD | BPF_H | BPF_ABS:
 879                 case BPF_LD | BPF_B | BPF_ABS:
 880                         anc_found = false;
 881                         if (bpf_anc_helper(ftest) & BPF_ANC)
 882                                 anc_found = true;
 883                         /* Ancillary operation unknown or unsupported */
 884                         if (anc_found == false && ftest->k >= SKF_AD_OFF)
 885                                 return -EINVAL;
 886                 }
 887         }
 888
 889         /* Last instruction must be a RET code */
 890         switch (filter[flen - 1].code) {
 891         case BPF_RET | BPF_K:
 892         case BPF_RET | BPF_A:
 893                 return check_load_and_stores(filter, flen);
 894         }
 895
 896         return -EINVAL;
 897 }
 898
 899 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
 900                                       const struct sock_fprog *fprog)
 901 {
 902         unsigned int fsize = bpf_classic_proglen(fprog);
 903         struct sock_fprog_kern *fkprog;
 904
 905         fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
 906         if (!fp->orig_prog)
 907                 return -ENOMEM;
 908
 909         fkprog = fp->orig_prog;
 910         fkprog->len = fprog->len;
 911
 912         fkprog->filter = kmemdup(fp->insns, fsize,
 913                                  GFP_KERNEL | __GFP_NOWARN);
 914         if (!fkprog->filter) {
 915                 kfree(fp->orig_prog);
 916                 return -ENOMEM;
 917         }
 918
 919         return 0;
 920 }
 921
 922 static void bpf_release_orig_filter(struct bpf_prog *fp)
 923 {
 924         struct sock_fprog_kern *fprog = fp->orig_prog;
 925
 926         if (fprog) {
 927                 kfree(fprog->filter);
 928                 kfree(fprog);
 929         }
 930 }
 931
 932 static void __bpf_prog_release(struct bpf_prog *prog)
 933 {
 934         if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
 935                 bpf_prog_put(prog);
 936         } else {
 937                 bpf_release_orig_filter(prog);
 938                 bpf_prog_free(prog);
 939         }
 940 }
 941
 942 static void __sk_filter_release(struct sk_filter *fp)
 943 {
 944         __bpf_prog_release(fp->prog);
 945         kfree(fp);
 946 }
 947
 948 /**
 949  *      sk_filter_release_rcu - Release a socket filter by rcu_head
 950  *      @rcu: rcu_head that contains the sk_filter to free
 951  */
 952 static void sk_filter_release_rcu(struct rcu_head *rcu)
 953 {
 954         struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
 955
 956         __sk_filter_release(fp);
 957 }
 958
 959 /**
 960  *      sk_filter_release - release a socket filter
 961  *      @fp: filter to remove
 962  *
 963  *      Remove a filter from a socket and release its resources.
 964  */
 965 static void sk_filter_release(struct sk_filter *fp)
 966 {
 967         if (refcount_dec_and_test(&fp->refcnt))
 968                 call_rcu(&fp->rcu, sk_filter_release_rcu);
 969 }
 970
 971 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
 972 {
 973         u32 filter_size = bpf_prog_size(fp->prog->len);
 974
 975         atomic_sub(filter_size, &sk->sk_omem_alloc);
 976         sk_filter_release(fp);
 977 }
 978
 979 /* try to charge the socket memory if there is space available
 980  * return true on success
 981  */
 982 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 983 {
 984         u32 filter_size = bpf_prog_size(fp->prog->len);
 985
 986         /* same check as in sock_kmalloc() */
 987         if (filter_size <= sysctl_optmem_max &&
 988             atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
 989                 atomic_add(filter_size, &sk->sk_omem_alloc);
 990                 return true;
 991         }
 992         return false;
 993 }
 994
 995 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 996 {
 997         if (!refcount_inc_not_zero(&fp->refcnt))
 998                 return false;
 999
1000         if (!__sk_filter_charge(sk, fp)) {
1001                 sk_filter_release(fp);
1002                 return false;
1003         }
1004         return true;
1005 }
1006
1007 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
1008 {
1009         struct sock_filter *old_prog;
1010         struct bpf_prog *old_fp;
1011         int err, new_len, old_len = fp->len;
1012
1013         /* We are free to overwrite insns et al right here as it
1014          * won't be used at this point in time anymore internally
1015          * after the migration to the internal BPF instruction
1016          * representation.
1017          */
1018         BUILD_BUG_ON(sizeof(struct sock_filter) !=
1019                      sizeof(struct bpf_insn));
1020
1021         /* Conversion cannot happen on overlapping memory areas,
1022          * so we need to keep the user BPF around until the 2nd
1023          * pass. At this time, the user BPF is stored in fp->insns.
1024          */
1025         old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1026                            GFP_KERNEL | __GFP_NOWARN);
1027         if (!old_prog) {
1028                 err = -ENOMEM;
1029                 goto out_err;
1030         }
1031
1032         /* 1st pass: calculate the new program length. */
1033         err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1034         if (err)
1035                 goto out_err_free;
1036
1037         /* Expand fp for appending the new filter representation. */
1038         old_fp = fp;
1039         fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1040         if (!fp) {
1041                 /* The old_fp is still around in case we couldn't
1042                  * allocate new memory, so uncharge on that one.
1043                  */
1044                 fp = old_fp;
1045                 err = -ENOMEM;
1046                 goto out_err_free;
1047         }
1048
1049         fp->len = new_len;
1050
1051         /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1052         err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1053         if (err)
1054                 /* 2nd bpf_convert_filter() can fail only if it fails
1055                  * to allocate memory, remapping must succeed. Note,
1056                  * that at this time old_fp has already been released
1057                  * by krealloc().
1058                  */
1059                 goto out_err_free;
1060
1061         fp = bpf_prog_select_runtime(fp, &err);
1062         if (err)
1063                 goto out_err_free;
1064
1065         kfree(old_prog);
1066         return fp;
1067
1068 out_err_free:
1069         kfree(old_prog);
1070 out_err:
1071         __bpf_prog_release(fp);
1072         return ERR_PTR(err);
1073 }
1074
1075 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1076                                            bpf_aux_classic_check_t trans)
1077 {
1078         int err;
1079
1080         fp->bpf_func = NULL;
1081         fp->jited = 0;
1082
1083         err = bpf_check_classic(fp->insns, fp->len);
1084         if (err) {
1085                 __bpf_prog_release(fp);
1086                 return ERR_PTR(err);
1087         }
1088
1089         /* There might be additional checks and transformations
1090          * needed on classic filters, f.e. in case of seccomp.
1091          */
1092         if (trans) {
1093                 err = trans(fp->insns, fp->len);
1094                 if (err) {
1095                         __bpf_prog_release(fp);
1096                         return ERR_PTR(err);
1097                 }
1098         }
1099
1100         /* Probe if we can JIT compile the filter and if so, do
1101          * the compilation of the filter.
1102          */
1103         bpf_jit_compile(fp);
1104
1105         /* JIT compiler couldn't process this filter, so do the
1106          * internal BPF translation for the optimized interpreter.
1107          */
1108         if (!fp->jited)
1109                 fp = bpf_migrate_filter(fp);
1110
1111         return fp;
1112 }
1113
1114 /**
1115  *      bpf_prog_create - create an unattached filter
1116  *      @pfp: the unattached filter that is created
1117  *      @fprog: the filter program
1118  *
1119  * Create a filter independent of any socket. We first run some
1120  * sanity checks on it to make sure it does not explode on us later.
1121  * If an error occurs or there is insufficient memory for the filter
1122  * a negative errno code is returned. On success the return is zero.
1123  */
1124 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1125 {
1126         unsigned int fsize = bpf_classic_proglen(fprog);
1127         struct bpf_prog *fp;
1128
1129         /* Make sure new filter is there and in the right amounts. */
1130         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1131                 return -EINVAL;
1132
1133         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1134         if (!fp)
1135                 return -ENOMEM;
1136
1137         memcpy(fp->insns, fprog->filter, fsize);
1138
1139         fp->len = fprog->len;
1140         /* Since unattached filters are not copied back to user
1141          * space through sk_get_filter(), we do not need to hold
1142          * a copy here, and can spare us the work.
1143          */
1144         fp->orig_prog = NULL;
1145
1146         /* bpf_prepare_filter() already takes care of freeing
1147          * memory in case something goes wrong.
1148          */
1149         fp = bpf_prepare_filter(fp, NULL);
1150         if (IS_ERR(fp))
1151                 return PTR_ERR(fp);
1152
1153         *pfp = fp;
1154         return 0;
1155 }
1156 EXPORT_SYMBOL_GPL(bpf_prog_create);
1157
1158 /**
1159  *      bpf_prog_create_from_user - create an unattached filter from user buffer
1160  *      @pfp: the unattached filter that is created
1161  *      @fprog: the filter program
1162  *      @trans: post-classic verifier transformation handler
1163  *      @save_orig: save classic BPF program
1164  *
1165  * This function effectively does the same as bpf_prog_create(), only
1166  * that it builds up its insns buffer from user space provided buffer.
1167  * It also allows for passing a bpf_aux_classic_check_t handler.
1168  */
1169 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1170                               bpf_aux_classic_check_t trans, bool save_orig)
1171 {
1172         unsigned int fsize = bpf_classic_proglen(fprog);
1173         struct bpf_prog *fp;
1174         int err;
1175
1176         /* Make sure new filter is there and in the right amounts. */
1177         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1178                 return -EINVAL;
1179
1180         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1181         if (!fp)
1182                 return -ENOMEM;
1183
1184         if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1185                 __bpf_prog_free(fp);
1186                 return -EFAULT;
1187         }
1188
1189         fp->len = fprog->len;
1190         fp->orig_prog = NULL;
1191
1192         if (save_orig) {
1193                 err = bpf_prog_store_orig_filter(fp, fprog);
1194                 if (err) {
1195                         __bpf_prog_free(fp);
1196                         return -ENOMEM;
1197                 }
1198         }
1199
1200         /* bpf_prepare_filter() already takes care of freeing
1201          * memory in case something goes wrong.
1202          */
1203         fp = bpf_prepare_filter(fp, trans);
1204         if (IS_ERR(fp))
1205                 return PTR_ERR(fp);
1206
1207         *pfp = fp;
1208         return 0;
1209 }
1210 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1211
1212 void bpf_prog_destroy(struct bpf_prog *fp)
1213 {
1214         __bpf_prog_release(fp);
1215 }
1216 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1217
1218 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1219 {
1220         struct sk_filter *fp, *old_fp;
1221
1222         fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1223         if (!fp)
1224                 return -ENOMEM;
1225
1226         fp->prog = prog;
1227
1228         if (!__sk_filter_charge(sk, fp)) {
1229                 kfree(fp);
1230                 return -ENOMEM;
1231         }
1232         refcount_set(&fp->refcnt, 1);
1233
1234         old_fp = rcu_dereference_protected(sk->sk_filter,
1235                                            lockdep_sock_is_held(sk));
1236         rcu_assign_pointer(sk->sk_filter, fp);
1237
1238         if (old_fp)
1239                 sk_filter_uncharge(sk, old_fp);
1240
1241         return 0;
1242 }
1243
1244 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1245 {
1246         struct bpf_prog *old_prog;
1247         int err;
1248
1249         if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1250                 return -ENOMEM;
1251
1252         if (sk_unhashed(sk) && sk->sk_reuseport) {
1253                 err = reuseport_alloc(sk);
1254                 if (err)
1255                         return err;
1256         } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1257                 /* The socket wasn't bound with SO_REUSEPORT */
1258                 return -EINVAL;
1259         }
1260
1261         old_prog = reuseport_attach_prog(sk, prog);
1262         if (old_prog)
1263                 bpf_prog_destroy(old_prog);
1264
1265         return 0;
1266 }
1267
1268 static
1269 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1270 {
1271         unsigned int fsize = bpf_classic_proglen(fprog);
1272         struct bpf_prog *prog;
1273         int err;
1274
1275         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1276                 return ERR_PTR(-EPERM);
1277
1278         /* Make sure new filter is there and in the right amounts. */
1279         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1280                 return ERR_PTR(-EINVAL);
1281
1282         prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1283         if (!prog)
1284                 return ERR_PTR(-ENOMEM);
1285
1286         if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1287                 __bpf_prog_free(prog);
1288                 return ERR_PTR(-EFAULT);
1289         }
1290
1291         prog->len = fprog->len;
1292
1293         err = bpf_prog_store_orig_filter(prog, fprog);
1294         if (err) {
1295                 __bpf_prog_free(prog);
1296                 return ERR_PTR(-ENOMEM);
1297         }
1298
1299         /* bpf_prepare_filter() already takes care of freeing
1300          * memory in case something goes wrong.
1301          */
1302         return bpf_prepare_filter(prog, NULL);
1303 }
1304
1305 /**
1306  *      sk_attach_filter - attach a socket filter
1307  *      @fprog: the filter program
1308  *      @sk: the socket to use
1309  *
1310  * Attach the user's filter code. We first run some sanity checks on
1311  * it to make sure it does not explode on us later. If an error
1312  * occurs or there is insufficient memory for the filter a negative
1313  * errno code is returned. On success the return is zero.
1314  */
1315 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1316 {
1317         struct bpf_prog *prog = __get_filter(fprog, sk);
1318         int err;
1319
1320         if (IS_ERR(prog))
1321                 return PTR_ERR(prog);
1322
1323         err = __sk_attach_prog(prog, sk);
1324         if (err < 0) {
1325                 __bpf_prog_release(prog);
1326                 return err;
1327         }
1328
1329         return 0;
1330 }
1331 EXPORT_SYMBOL_GPL(sk_attach_filter);
1332
1333 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1334 {
1335         struct bpf_prog *prog = __get_filter(fprog, sk);
1336         int err;
1337
1338         if (IS_ERR(prog))
1339                 return PTR_ERR(prog);
1340
1341         err = __reuseport_attach_prog(prog, sk);
1342         if (err < 0) {
1343                 __bpf_prog_release(prog);
1344                 return err;
1345         }
1346
1347         return 0;
1348 }
1349
1350 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1351 {
1352         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1353                 return ERR_PTR(-EPERM);
1354
1355         return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1356 }
1357
1358 int sk_attach_bpf(u32 ufd, struct sock *sk)
1359 {
1360         struct bpf_prog *prog = __get_bpf(ufd, sk);
1361         int err;
1362
1363         if (IS_ERR(prog))
1364                 return PTR_ERR(prog);
1365
1366         err = __sk_attach_prog(prog, sk);
1367         if (err < 0) {
1368                 bpf_prog_put(prog);
1369                 return err;
1370         }
1371
1372         return 0;
1373 }
1374
1375 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1376 {
1377         struct bpf_prog *prog = __get_bpf(ufd, sk);
1378         int err;
1379
1380         if (IS_ERR(prog))
1381                 return PTR_ERR(prog);
1382
1383         err = __reuseport_attach_prog(prog, sk);
1384         if (err < 0) {
1385                 bpf_prog_put(prog);
1386                 return err;
1387         }
1388
1389         return 0;
1390 }
1391
1392 struct bpf_scratchpad {
1393         union {
1394                 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1395                 u8     buff[MAX_BPF_STACK];
1396         };
1397 };
1398
1399 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1400
1401 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1402                                           unsigned int write_len)
1403 {
1404         return skb_ensure_writable(skb, write_len);
1405 }
1406
1407 static inline int bpf_try_make_writable(struct sk_buff *skb,
1408                                         unsigned int write_len)
1409 {
1410         int err = __bpf_try_make_writable(skb, write_len);
1411
1412         bpf_compute_data_pointers(skb);
1413         return err;
1414 }
1415
1416 static int bpf_try_make_head_writable(struct sk_buff *skb)
1417 {
1418         return bpf_try_make_writable(skb, skb_headlen(skb));
1419 }
1420
1421 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1422 {
1423         if (skb_at_tc_ingress(skb))
1424                 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1425 }
1426
1427 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1428 {
1429         if (skb_at_tc_ingress(skb))
1430                 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1431 }
1432
1433 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1434            const void *, from, u32, len, u64, flags)
1435 {
1436         void *ptr;
1437
1438         if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1439                 return -EINVAL;
1440         if (unlikely(offset > 0xffff))
1441                 return -EFAULT;
1442         if (unlikely(bpf_try_make_writable(skb, offset + len)))
1443                 return -EFAULT;
1444
1445         ptr = skb->data + offset;
1446         if (flags & BPF_F_RECOMPUTE_CSUM)
1447                 __skb_postpull_rcsum(skb, ptr, len, offset);
1448
1449         memcpy(ptr, from, len);
1450
1451         if (flags & BPF_F_RECOMPUTE_CSUM)
1452                 __skb_postpush_rcsum(skb, ptr, len, offset);
1453         if (flags & BPF_F_INVALIDATE_HASH)
1454                 skb_clear_hash(skb);
1455
1456         return 0;
1457 }
1458
1459 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1460         .func           = bpf_skb_store_bytes,
1461         .gpl_only       = false,
1462         .ret_type       = RET_INTEGER,
1463         .arg1_type      = ARG_PTR_TO_CTX,
1464         .arg2_type      = ARG_ANYTHING,
1465         .arg3_type      = ARG_PTR_TO_MEM,
1466         .arg4_type      = ARG_CONST_SIZE,
1467         .arg5_type      = ARG_ANYTHING,
1468 };
1469
1470 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1471            void *, to, u32, len)
1472 {
1473         void *ptr;
1474
1475         if (unlikely(offset > 0xffff))
1476                 goto err_clear;
1477
1478         ptr = skb_header_pointer(skb, offset, len, to);
1479         if (unlikely(!ptr))
1480                 goto err_clear;
1481         if (ptr != to)
1482                 memcpy(to, ptr, len);
1483
1484         return 0;
1485 err_clear:
1486         memset(to, 0, len);
1487         return -EFAULT;
1488 }
1489
1490 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1491         .func           = bpf_skb_load_bytes,
1492         .gpl_only       = false,
1493         .ret_type       = RET_INTEGER,
1494         .arg1_type      = ARG_PTR_TO_CTX,
1495         .arg2_type      = ARG_ANYTHING,
1496         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
1497         .arg4_type      = ARG_CONST_SIZE,
1498 };
1499
1500 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1501 {
1502         /* Idea is the following: should the needed direct read/write
1503          * test fail during runtime, we can pull in more data and redo
1504          * again, since implicitly, we invalidate previous checks here.
1505          *
1506          * Or, since we know how much we need to make read/writeable,
1507          * this can be done once at the program beginning for direct
1508          * access case. By this we overcome limitations of only current
1509          * headroom being accessible.
1510          */
1511         return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1512 }
1513
1514 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1515         .func           = bpf_skb_pull_data,
1516         .gpl_only       = false,
1517         .ret_type       = RET_INTEGER,
1518         .arg1_type      = ARG_PTR_TO_CTX,
1519         .arg2_type      = ARG_ANYTHING,
1520 };
1521
1522 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1523            u64, from, u64, to, u64, flags)
1524 {
1525         __sum16 *ptr;
1526
1527         if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1528                 return -EINVAL;
1529         if (unlikely(offset > 0xffff || offset & 1))
1530                 return -EFAULT;
1531         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1532                 return -EFAULT;
1533
1534         ptr = (__sum16 *)(skb->data + offset);
1535         switch (flags & BPF_F_HDR_FIELD_MASK) {
1536         case 0:
1537                 if (unlikely(from != 0))
1538                         return -EINVAL;
1539
1540                 csum_replace_by_diff(ptr, to);
1541                 break;
1542         case 2:
1543                 csum_replace2(ptr, from, to);
1544                 break;
1545         case 4:
1546                 csum_replace4(ptr, from, to);
1547                 break;
1548         default:
1549                 return -EINVAL;
1550         }
1551
1552         return 0;
1553 }
1554
1555 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1556         .func           = bpf_l3_csum_replace,
1557         .gpl_only       = false,
1558         .ret_type       = RET_INTEGER,
1559         .arg1_type      = ARG_PTR_TO_CTX,
1560         .arg2_type      = ARG_ANYTHING,
1561         .arg3_type      = ARG_ANYTHING,
1562         .arg4_type      = ARG_ANYTHING,
1563         .arg5_type      = ARG_ANYTHING,
1564 };
1565
1566 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1567            u64, from, u64, to, u64, flags)
1568 {
1569         bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1570         bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1571         bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1572         __sum16 *ptr;
1573
1574         if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1575                                BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1576                 return -EINVAL;
1577         if (unlikely(offset > 0xffff || offset & 1))
1578                 return -EFAULT;
1579         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1580                 return -EFAULT;
1581
1582         ptr = (__sum16 *)(skb->data + offset);
1583         if (is_mmzero && !do_mforce && !*ptr)
1584                 return 0;
1585
1586         switch (flags & BPF_F_HDR_FIELD_MASK) {
1587         case 0:
1588                 if (unlikely(from != 0))
1589                         return -EINVAL;
1590
1591                 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1592                 break;
1593         case 2:
1594                 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1595                 break;
1596         case 4:
1597                 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1598                 break;
1599         default:
1600                 return -EINVAL;
1601         }
1602
1603         if (is_mmzero && !*ptr)
1604                 *ptr = CSUM_MANGLED_0;
1605         return 0;
1606 }
1607
1608 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1609         .func           = bpf_l4_csum_replace,
1610         .gpl_only       = false,
1611         .ret_type       = RET_INTEGER,
1612         .arg1_type      = ARG_PTR_TO_CTX,
1613         .arg2_type      = ARG_ANYTHING,
1614         .arg3_type      = ARG_ANYTHING,
1615         .arg4_type      = ARG_ANYTHING,
1616         .arg5_type      = ARG_ANYTHING,
1617 };
1618
1619 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1620            __be32 *, to, u32, to_size, __wsum, seed)
1621 {
1622         struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1623         u32 diff_size = from_size + to_size;
1624         int i, j = 0;
1625
1626         /* This is quite flexible, some examples:
1627          *
1628          * from_size == 0, to_size > 0,  seed := csum --> pushing data
1629          * from_size > 0,  to_size == 0, seed := csum --> pulling data
1630          * from_size > 0,  to_size > 0,  seed := 0    --> diffing data
1631          *
1632          * Even for diffing, from_size and to_size don't need to be equal.
1633          */
1634         if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1635                      diff_size > sizeof(sp->diff)))
1636                 return -EINVAL;
1637
1638         for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1639                 sp->diff[j] = ~from[i];
1640         for (i = 0; i <   to_size / sizeof(__be32); i++, j++)
1641                 sp->diff[j] = to[i];
1642
1643         return csum_partial(sp->diff, diff_size, seed);
1644 }
1645
1646 static const struct bpf_func_proto bpf_csum_diff_proto = {
1647         .func           = bpf_csum_diff,
1648         .gpl_only       = false,
1649         .pkt_access     = true,
1650         .ret_type       = RET_INTEGER,
1651         .arg1_type      = ARG_PTR_TO_MEM_OR_NULL,
1652         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
1653         .arg3_type      = ARG_PTR_TO_MEM_OR_NULL,
1654         .arg4_type      = ARG_CONST_SIZE_OR_ZERO,
1655         .arg5_type      = ARG_ANYTHING,
1656 };
1657
1658 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1659 {
1660         /* The interface is to be used in combination with bpf_csum_diff()
1661          * for direct packet writes. csum rotation for alignment as well
1662          * as emulating csum_sub() can be done from the eBPF program.
1663          */
1664         if (skb->ip_summed == CHECKSUM_COMPLETE)
1665                 return (skb->csum = csum_add(skb->csum, csum));
1666
1667         return -ENOTSUPP;
1668 }
1669
1670 static const struct bpf_func_proto bpf_csum_update_proto = {
1671         .func           = bpf_csum_update,
1672         .gpl_only       = false,
1673         .ret_type       = RET_INTEGER,
1674         .arg1_type      = ARG_PTR_TO_CTX,
1675         .arg2_type      = ARG_ANYTHING,
1676 };
1677
1678 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1679 {
1680         return dev_forward_skb(dev, skb);
1681 }
1682
1683 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1684                                       struct sk_buff *skb)
1685 {
1686         int ret = ____dev_forward_skb(dev, skb);
1687
1688         if (likely(!ret)) {
1689                 skb->dev = dev;
1690                 ret = netif_rx(skb);
1691         }
1692
1693         return ret;
1694 }
1695
1696 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1697 {
1698         int ret;
1699
1700         if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1701                 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1702                 kfree_skb(skb);
1703                 return -ENETDOWN;
1704         }
1705
1706         skb->dev = dev;
1707
1708         __this_cpu_inc(xmit_recursion);
1709         ret = dev_queue_xmit(skb);
1710         __this_cpu_dec(xmit_recursion);
1711
1712         return ret;
1713 }
1714
1715 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1716                                  u32 flags)
1717 {
1718         /* skb->mac_len is not set on normal egress */
1719         unsigned int mlen = skb->network_header - skb->mac_header;
1720
1721         __skb_pull(skb, mlen);
1722
1723         /* At ingress, the mac header has already been pulled once.
1724          * At egress, skb_pospull_rcsum has to be done in case that
1725          * the skb is originated from ingress (i.e. a forwarded skb)
1726          * to ensure that rcsum starts at net header.
1727          */
1728         if (!skb_at_tc_ingress(skb))
1729                 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1730         skb_pop_mac_header(skb);
1731         skb_reset_mac_len(skb);
1732         return flags & BPF_F_INGRESS ?
1733                __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1734 }
1735
1736 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1737                                  u32 flags)
1738 {
1739         /* Verify that a link layer header is carried */
1740         if (unlikely(skb->mac_header >= skb->network_header)) {
1741                 kfree_skb(skb);
1742                 return -ERANGE;
1743         }
1744
1745         bpf_push_mac_rcsum(skb);
1746         return flags & BPF_F_INGRESS ?
1747                __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1748 }
1749
1750 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1751                           u32 flags)
1752 {
1753         if (dev_is_mac_header_xmit(dev))
1754                 return __bpf_redirect_common(skb, dev, flags);
1755         else
1756                 return __bpf_redirect_no_mac(skb, dev, flags);
1757 }
1758
1759 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1760 {
1761         struct net_device *dev;
1762         struct sk_buff *clone;
1763         int ret;
1764
1765         if (unlikely(flags & ~(BPF_F_INGRESS)))
1766                 return -EINVAL;
1767
1768         dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1769         if (unlikely(!dev))
1770                 return -EINVAL;
1771
1772         clone = skb_clone(skb, GFP_ATOMIC);
1773         if (unlikely(!clone))
1774                 return -ENOMEM;
1775
1776         /* For direct write, we need to keep the invariant that the skbs
1777          * we're dealing with need to be uncloned. Should uncloning fail
1778          * here, we need to free the just generated clone to unclone once
1779          * again.
1780          */
1781         ret = bpf_try_make_head_writable(skb);
1782         if (unlikely(ret)) {
1783                 kfree_skb(clone);
1784                 return -ENOMEM;
1785         }
1786
1787         return __bpf_redirect(clone, dev, flags);
1788 }
1789
1790 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1791         .func           = bpf_clone_redirect,
1792         .gpl_only       = false,
1793         .ret_type       = RET_INTEGER,
1794         .arg1_type      = ARG_PTR_TO_CTX,
1795         .arg2_type      = ARG_ANYTHING,
1796         .arg3_type      = ARG_ANYTHING,
1797 };
1798
1799 struct redirect_info {
1800         u32 ifindex;
1801         u32 flags;
1802         struct bpf_map *map;
1803         struct bpf_map *map_to_flush;
1804         unsigned long   map_owner;
1805 };
1806
1807 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1808
1809 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1810 {
1811         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1812
1813         if (unlikely(flags & ~(BPF_F_INGRESS)))
1814                 return TC_ACT_SHOT;
1815
1816         ri->ifindex = ifindex;
1817         ri->flags = flags;
1818
1819         return TC_ACT_REDIRECT;
1820 }
1821
1822 int skb_do_redirect(struct sk_buff *skb)
1823 {
1824         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1825         struct net_device *dev;
1826
1827         dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1828         ri->ifindex = 0;
1829         if (unlikely(!dev)) {
1830                 kfree_skb(skb);
1831                 return -EINVAL;
1832         }
1833
1834         return __bpf_redirect(skb, dev, ri->flags);
1835 }
1836
1837 static const struct bpf_func_proto bpf_redirect_proto = {
1838         .func           = bpf_redirect,
1839         .gpl_only       = false,
1840         .ret_type       = RET_INTEGER,
1841         .arg1_type      = ARG_ANYTHING,
1842         .arg2_type      = ARG_ANYTHING,
1843 };
1844
1845 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
1846            struct bpf_map *, map, u32, key, u64, flags)
1847 {
1848         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1849
1850         /* If user passes invalid input drop the packet. */
1851         if (unlikely(flags))
1852                 return SK_DROP;
1853
1854         tcb->bpf.key = key;
1855         tcb->bpf.flags = flags;
1856         tcb->bpf.map = map;
1857
1858         return SK_PASS;
1859 }
1860
1861 struct sock *do_sk_redirect_map(struct sk_buff *skb)
1862 {
1863         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1864         struct sock *sk = NULL;
1865
1866         if (tcb->bpf.map) {
1867                 sk = __sock_map_lookup_elem(tcb->bpf.map, tcb->bpf.key);
1868
1869                 tcb->bpf.key = 0;
1870                 tcb->bpf.map = NULL;
1871         }
1872
1873         return sk;
1874 }
1875
1876 static const struct bpf_func_proto bpf_sk_redirect_map_proto = {
1877         .func           = bpf_sk_redirect_map,
1878         .gpl_only       = false,
1879         .ret_type       = RET_INTEGER,
1880         .arg1_type      = ARG_PTR_TO_CTX,
1881         .arg2_type      = ARG_CONST_MAP_PTR,
1882         .arg3_type      = ARG_ANYTHING,
1883         .arg4_type      = ARG_ANYTHING,
1884 };
1885
1886 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1887 {
1888         return task_get_classid(skb);
1889 }
1890
1891 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1892         .func           = bpf_get_cgroup_classid,
1893         .gpl_only       = false,
1894         .ret_type       = RET_INTEGER,
1895         .arg1_type      = ARG_PTR_TO_CTX,
1896 };
1897
1898 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1899 {
1900         return dst_tclassid(skb);
1901 }
1902
1903 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1904         .func           = bpf_get_route_realm,
1905         .gpl_only       = false,
1906         .ret_type       = RET_INTEGER,
1907         .arg1_type      = ARG_PTR_TO_CTX,
1908 };
1909
1910 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1911 {
1912         /* If skb_clear_hash() was called due to mangling, we can
1913          * trigger SW recalculation here. Later access to hash
1914          * can then use the inline skb->hash via context directly
1915          * instead of calling this helper again.
1916          */
1917         return skb_get_hash(skb);
1918 }
1919
1920 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1921         .func           = bpf_get_hash_recalc,
1922         .gpl_only       = false,
1923         .ret_type       = RET_INTEGER,
1924         .arg1_type      = ARG_PTR_TO_CTX,
1925 };
1926
1927 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1928 {
1929         /* After all direct packet write, this can be used once for
1930          * triggering a lazy recalc on next skb_get_hash() invocation.
1931          */
1932         skb_clear_hash(skb);
1933         return 0;
1934 }
1935
1936 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1937         .func           = bpf_set_hash_invalid,
1938         .gpl_only       = false,
1939         .ret_type       = RET_INTEGER,
1940         .arg1_type      = ARG_PTR_TO_CTX,
1941 };
1942
1943 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
1944 {
1945         /* Set user specified hash as L4(+), so that it gets returned
1946          * on skb_get_hash() call unless BPF prog later on triggers a
1947          * skb_clear_hash().
1948          */
1949         __skb_set_sw_hash(skb, hash, true);
1950         return 0;
1951 }
1952
1953 static const struct bpf_func_proto bpf_set_hash_proto = {
1954         .func           = bpf_set_hash,
1955         .gpl_only       = false,
1956         .ret_type       = RET_INTEGER,
1957         .arg1_type      = ARG_PTR_TO_CTX,
1958         .arg2_type      = ARG_ANYTHING,
1959 };
1960
1961 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1962            u16, vlan_tci)
1963 {
1964         int ret;
1965
1966         if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1967                      vlan_proto != htons(ETH_P_8021AD)))
1968                 vlan_proto = htons(ETH_P_8021Q);
1969
1970         bpf_push_mac_rcsum(skb);
1971         ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1972         bpf_pull_mac_rcsum(skb);
1973
1974         bpf_compute_data_pointers(skb);
1975         return ret;
1976 }
1977
1978 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1979         .func           = bpf_skb_vlan_push,
1980         .gpl_only       = false,
1981         .ret_type       = RET_INTEGER,
1982         .arg1_type      = ARG_PTR_TO_CTX,
1983         .arg2_type      = ARG_ANYTHING,
1984         .arg3_type      = ARG_ANYTHING,
1985 };
1986 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1987
1988 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
1989 {
1990         int ret;
1991
1992         bpf_push_mac_rcsum(skb);
1993         ret = skb_vlan_pop(skb);
1994         bpf_pull_mac_rcsum(skb);
1995
1996         bpf_compute_data_pointers(skb);
1997         return ret;
1998 }
1999
2000 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
2001         .func           = bpf_skb_vlan_pop,
2002         .gpl_only       = false,
2003         .ret_type       = RET_INTEGER,
2004         .arg1_type      = ARG_PTR_TO_CTX,
2005 };
2006 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
2007
2008 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
2009 {
2010         /* Caller already did skb_cow() with len as headroom,
2011          * so no need to do it here.
2012          */
2013         skb_push(skb, len);
2014         memmove(skb->data, skb->data + len, off);
2015         memset(skb->data + off, 0, len);
2016
2017         /* No skb_postpush_rcsum(skb, skb->data + off, len)
2018          * needed here as it does not change the skb->csum
2019          * result for checksum complete when summing over
2020          * zeroed blocks.
2021          */
2022         return 0;
2023 }
2024
2025 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
2026 {
2027         /* skb_ensure_writable() is not needed here, as we're
2028          * already working on an uncloned skb.
2029          */
2030         if (unlikely(!pskb_may_pull(skb, off + len)))
2031                 return -ENOMEM;
2032
2033         skb_postpull_rcsum(skb, skb->data + off, len);
2034         memmove(skb->data + len, skb->data, off);
2035         __skb_pull(skb, len);
2036
2037         return 0;
2038 }
2039
2040 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
2041 {
2042         bool trans_same = skb->transport_header == skb->network_header;
2043         int ret;
2044
2045         /* There's no need for __skb_push()/__skb_pull() pair to
2046          * get to the start of the mac header as we're guaranteed
2047          * to always start from here under eBPF.
2048          */
2049         ret = bpf_skb_generic_push(skb, off, len);
2050         if (likely(!ret)) {
2051                 skb->mac_header -= len;
2052                 skb->network_header -= len;
2053                 if (trans_same)
2054                         skb->transport_header = skb->network_header;
2055         }
2056
2057         return ret;
2058 }
2059
2060 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
2061 {
2062         bool trans_same = skb->transport_header == skb->network_header;
2063         int ret;
2064
2065         /* Same here, __skb_push()/__skb_pull() pair not needed. */
2066         ret = bpf_skb_generic_pop(skb, off, len);
2067         if (likely(!ret)) {
2068                 skb->mac_header += len;
2069                 skb->network_header += len;
2070                 if (trans_same)
2071                         skb->transport_header = skb->network_header;
2072         }
2073
2074         return ret;
2075 }
2076
2077 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2078 {
2079         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2080         u32 off = skb_mac_header_len(skb);
2081         int ret;
2082
2083         ret = skb_cow(skb, len_diff);
2084         if (unlikely(ret < 0))
2085                 return ret;
2086
2087         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2088         if (unlikely(ret < 0))
2089                 return ret;
2090
2091         if (skb_is_gso(skb)) {
2092                 /* SKB_GSO_TCPV4 needs to be changed into
2093                  * SKB_GSO_TCPV6.
2094                  */
2095                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2096                         skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
2097                         skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV6;
2098                 }
2099
2100                 /* Due to IPv6 header, MSS needs to be downgraded. */
2101                 skb_shinfo(skb)->gso_size -= len_diff;
2102                 /* Header must be checked, and gso_segs recomputed. */
2103                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2104                 skb_shinfo(skb)->gso_segs = 0;
2105         }
2106
2107         skb->protocol = htons(ETH_P_IPV6);
2108         skb_clear_hash(skb);
2109
2110         return 0;
2111 }
2112
2113 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2114 {
2115         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2116         u32 off = skb_mac_header_len(skb);
2117         int ret;
2118
2119         ret = skb_unclone(skb, GFP_ATOMIC);
2120         if (unlikely(ret < 0))
2121                 return ret;
2122
2123         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2124         if (unlikely(ret < 0))
2125                 return ret;
2126
2127         if (skb_is_gso(skb)) {
2128                 /* SKB_GSO_TCPV6 needs to be changed into
2129                  * SKB_GSO_TCPV4.
2130                  */
2131                 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2132                         skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2133                         skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV4;
2134                 }
2135
2136                 /* Due to IPv4 header, MSS can be upgraded. */
2137                 skb_shinfo(skb)->gso_size += len_diff;
2138                 /* Header must be checked, and gso_segs recomputed. */
2139                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2140                 skb_shinfo(skb)->gso_segs = 0;
2141         }
2142
2143         skb->protocol = htons(ETH_P_IP);
2144         skb_clear_hash(skb);
2145
2146         return 0;
2147 }
2148
2149 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2150 {
2151         __be16 from_proto = skb->protocol;
2152
2153         if (from_proto == htons(ETH_P_IP) &&
2154               to_proto == htons(ETH_P_IPV6))
2155                 return bpf_skb_proto_4_to_6(skb);
2156
2157         if (from_proto == htons(ETH_P_IPV6) &&
2158               to_proto == htons(ETH_P_IP))
2159                 return bpf_skb_proto_6_to_4(skb);
2160
2161         return -ENOTSUPP;
2162 }
2163
2164 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2165            u64, flags)
2166 {
2167         int ret;
2168
2169         if (unlikely(flags))
2170                 return -EINVAL;
2171
2172         /* General idea is that this helper does the basic groundwork
2173          * needed for changing the protocol, and eBPF program fills the
2174          * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2175          * and other helpers, rather than passing a raw buffer here.
2176          *
2177          * The rationale is to keep this minimal and without a need to
2178          * deal with raw packet data. F.e. even if we would pass buffers
2179          * here, the program still needs to call the bpf_lX_csum_replace()
2180          * helpers anyway. Plus, this way we keep also separation of
2181          * concerns, since f.e. bpf_skb_store_bytes() should only take
2182          * care of stores.
2183          *
2184          * Currently, additional options and extension header space are
2185          * not supported, but flags register is reserved so we can adapt
2186          * that. For offloads, we mark packet as dodgy, so that headers
2187          * need to be verified first.
2188          */
2189         ret = bpf_skb_proto_xlat(skb, proto);
2190         bpf_compute_data_pointers(skb);
2191         return ret;
2192 }
2193
2194 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2195         .func           = bpf_skb_change_proto,
2196         .gpl_only       = false,
2197         .ret_type       = RET_INTEGER,
2198         .arg1_type      = ARG_PTR_TO_CTX,
2199         .arg2_type      = ARG_ANYTHING,
2200         .arg3_type      = ARG_ANYTHING,
2201 };
2202
2203 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2204 {
2205         /* We only allow a restricted subset to be changed for now. */
2206         if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2207                      !skb_pkt_type_ok(pkt_type)))
2208                 return -EINVAL;
2209
2210         skb->pkt_type = pkt_type;
2211         return 0;
2212 }
2213
2214 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2215         .func           = bpf_skb_change_type,
2216         .gpl_only       = false,
2217         .ret_type       = RET_INTEGER,
2218         .arg1_type      = ARG_PTR_TO_CTX,
2219         .arg2_type      = ARG_ANYTHING,
2220 };
2221
2222 static u32 bpf_skb_net_base_len(const struct sk_buff *skb)
2223 {
2224         switch (skb->protocol) {
2225         case htons(ETH_P_IP):
2226                 return sizeof(struct iphdr);
2227         case htons(ETH_P_IPV6):
2228                 return sizeof(struct ipv6hdr);
2229         default:
2230                 return ~0U;
2231         }
2232 }
2233
2234 static int bpf_skb_net_grow(struct sk_buff *skb, u32 len_diff)
2235 {
2236         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2237         int ret;
2238
2239         ret = skb_cow(skb, len_diff);
2240         if (unlikely(ret < 0))
2241                 return ret;
2242
2243         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2244         if (unlikely(ret < 0))
2245                 return ret;
2246
2247         if (skb_is_gso(skb)) {
2248                 /* Due to header grow, MSS needs to be downgraded. */
2249                 skb_shinfo(skb)->gso_size -= len_diff;
2250                 /* Header must be checked, and gso_segs recomputed. */
2251                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2252                 skb_shinfo(skb)->gso_segs = 0;
2253         }
2254
2255         return 0;
2256 }
2257
2258 static int bpf_skb_net_shrink(struct sk_buff *skb, u32 len_diff)
2259 {
2260         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2261         int ret;
2262
2263         ret = skb_unclone(skb, GFP_ATOMIC);
2264         if (unlikely(ret < 0))
2265                 return ret;
2266
2267         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2268         if (unlikely(ret < 0))
2269                 return ret;
2270
2271         if (skb_is_gso(skb)) {
2272                 /* Due to header shrink, MSS can be upgraded. */
2273                 skb_shinfo(skb)->gso_size += len_diff;
2274                 /* Header must be checked, and gso_segs recomputed. */
2275                 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2276                 skb_shinfo(skb)->gso_segs = 0;
2277         }
2278
2279         return 0;
2280 }
2281
2282 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2283 {
2284         return skb->dev->mtu + skb->dev->hard_header_len;
2285 }
2286
2287 static int bpf_skb_adjust_net(struct sk_buff *skb, s32 len_diff)
2288 {
2289         bool trans_same = skb->transport_header == skb->network_header;
2290         u32 len_cur, len_diff_abs = abs(len_diff);
2291         u32 len_min = bpf_skb_net_base_len(skb);
2292         u32 len_max = __bpf_skb_max_len(skb);
2293         __be16 proto = skb->protocol;
2294         bool shrink = len_diff < 0;
2295         int ret;
2296
2297         if (unlikely(len_diff_abs > 0xfffU))
2298                 return -EFAULT;
2299         if (unlikely(proto != htons(ETH_P_IP) &&
2300                      proto != htons(ETH_P_IPV6)))
2301                 return -ENOTSUPP;
2302
2303         len_cur = skb->len - skb_network_offset(skb);
2304         if (skb_transport_header_was_set(skb) && !trans_same)
2305                 len_cur = skb_network_header_len(skb);
2306         if ((shrink && (len_diff_abs >= len_cur ||
2307                         len_cur - len_diff_abs < len_min)) ||
2308             (!shrink && (skb->len + len_diff_abs > len_max &&
2309                          !skb_is_gso(skb))))
2310                 return -ENOTSUPP;
2311
2312         ret = shrink ? bpf_skb_net_shrink(skb, len_diff_abs) :
2313                        bpf_skb_net_grow(skb, len_diff_abs);
2314
2315         bpf_compute_data_pointers(skb);
2316         return ret;
2317 }
2318
2319 BPF_CALL_4(bpf_skb_adjust_room, struct sk_buff *, skb, s32, len_diff,
2320            u32, mode, u64, flags)
2321 {
2322         if (unlikely(flags))
2323                 return -EINVAL;
2324         if (likely(mode == BPF_ADJ_ROOM_NET))
2325                 return bpf_skb_adjust_net(skb, len_diff);
2326
2327         return -ENOTSUPP;
2328 }
2329
2330 static const struct bpf_func_proto bpf_skb_adjust_room_proto = {
2331         .func           = bpf_skb_adjust_room,
2332         .gpl_only       = false,
2333         .ret_type       = RET_INTEGER,
2334         .arg1_type      = ARG_PTR_TO_CTX,
2335         .arg2_type      = ARG_ANYTHING,
2336         .arg3_type      = ARG_ANYTHING,
2337         .arg4_type      = ARG_ANYTHING,
2338 };
2339
2340 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2341 {
2342         u32 min_len = skb_network_offset(skb);
2343
2344         if (skb_transport_header_was_set(skb))
2345                 min_len = skb_transport_offset(skb);
2346         if (skb->ip_summed == CHECKSUM_PARTIAL)
2347                 min_len = skb_checksum_start_offset(skb) +
2348                           skb->csum_offset + sizeof(__sum16);
2349         return min_len;
2350 }
2351
2352 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2353 {
2354         unsigned int old_len = skb->len;
2355         int ret;
2356
2357         ret = __skb_grow_rcsum(skb, new_len);
2358         if (!ret)
2359                 memset(skb->data + old_len, 0, new_len - old_len);
2360         return ret;
2361 }
2362
2363 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2364 {
2365         return __skb_trim_rcsum(skb, new_len);
2366 }
2367
2368 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2369            u64, flags)
2370 {
2371         u32 max_len = __bpf_skb_max_len(skb);
2372         u32 min_len = __bpf_skb_min_len(skb);
2373         int ret;
2374
2375         if (unlikely(flags || new_len > max_len || new_len < min_len))
2376                 return -EINVAL;
2377         if (skb->encapsulation)
2378                 return -ENOTSUPP;
2379
2380         /* The basic idea of this helper is that it's performing the
2381          * needed work to either grow or trim an skb, and eBPF program
2382          * rewrites the rest via helpers like bpf_skb_store_bytes(),
2383          * bpf_lX_csum_replace() and others rather than passing a raw
2384          * buffer here. This one is a slow path helper and intended
2385          * for replies with control messages.
2386          *
2387          * Like in bpf_skb_change_proto(), we want to keep this rather
2388          * minimal and without protocol specifics so that we are able
2389          * to separate concerns as in bpf_skb_store_bytes() should only
2390          * be the one responsible for writing buffers.
2391          *
2392          * It's really expected to be a slow path operation here for
2393          * control message replies, so we're implicitly linearizing,
2394          * uncloning and drop offloads from the skb by this.
2395          */
2396         ret = __bpf_try_make_writable(skb, skb->len);
2397         if (!ret) {
2398                 if (new_len > skb->len)
2399                         ret = bpf_skb_grow_rcsum(skb, new_len);
2400                 else if (new_len < skb->len)
2401                         ret = bpf_skb_trim_rcsum(skb, new_len);
2402                 if (!ret && skb_is_gso(skb))
2403                         skb_gso_reset(skb);
2404         }
2405
2406         bpf_compute_data_pointers(skb);
2407         return ret;
2408 }
2409
2410 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2411         .func           = bpf_skb_change_tail,
2412         .gpl_only       = false,
2413         .ret_type       = RET_INTEGER,
2414         .arg1_type      = ARG_PTR_TO_CTX,
2415         .arg2_type      = ARG_ANYTHING,
2416         .arg3_type      = ARG_ANYTHING,
2417 };
2418
2419 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2420            u64, flags)
2421 {
2422         u32 max_len = __bpf_skb_max_len(skb);
2423         u32 new_len = skb->len + head_room;
2424         int ret;
2425
2426         if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2427                      new_len < skb->len))
2428                 return -EINVAL;
2429
2430         ret = skb_cow(skb, head_room);
2431         if (likely(!ret)) {
2432                 /* Idea for this helper is that we currently only
2433                  * allow to expand on mac header. This means that
2434                  * skb->protocol network header, etc, stay as is.
2435                  * Compared to bpf_skb_change_tail(), we're more
2436                  * flexible due to not needing to linearize or
2437                  * reset GSO. Intention for this helper is to be
2438                  * used by an L3 skb that needs to push mac header
2439                  * for redirection into L2 device.
2440                  */
2441                 __skb_push(skb, head_room);
2442                 memset(skb->data, 0, head_room);
2443                 skb_reset_mac_header(skb);
2444         }
2445
2446         bpf_compute_data_pointers(skb);
2447         return 0;
2448 }
2449
2450 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2451         .func           = bpf_skb_change_head,
2452         .gpl_only       = false,
2453         .ret_type       = RET_INTEGER,
2454         .arg1_type      = ARG_PTR_TO_CTX,
2455         .arg2_type      = ARG_ANYTHING,
2456         .arg3_type      = ARG_ANYTHING,
2457 };
2458
2459 static unsigned long xdp_get_metalen(const struct xdp_buff *xdp)
2460 {
2461         return xdp_data_meta_unsupported(xdp) ? 0 :
2462                xdp->data - xdp->data_meta;
2463 }
2464
2465 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2466 {
2467         unsigned long metalen = xdp_get_metalen(xdp);
2468         void *data_start = xdp->data_hard_start + metalen;
2469         void *data = xdp->data + offset;
2470
2471         if (unlikely(data < data_start ||
2472                      data > xdp->data_end - ETH_HLEN))
2473                 return -EINVAL;
2474
2475         if (metalen)
2476                 memmove(xdp->data_meta + offset,
2477                         xdp->data_meta, metalen);
2478         xdp->data_meta += offset;
2479         xdp->data = data;
2480
2481         return 0;
2482 }
2483
2484 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2485         .func           = bpf_xdp_adjust_head,
2486         .gpl_only       = false,
2487         .ret_type       = RET_INTEGER,
2488         .arg1_type      = ARG_PTR_TO_CTX,
2489         .arg2_type      = ARG_ANYTHING,
2490 };
2491
2492 BPF_CALL_2(bpf_xdp_adjust_meta, struct xdp_buff *, xdp, int, offset)
2493 {
2494         void *meta = xdp->data_meta + offset;
2495         unsigned long metalen = xdp->data - meta;
2496
2497         if (xdp_data_meta_unsupported(xdp))
2498                 return -ENOTSUPP;
2499         if (unlikely(meta < xdp->data_hard_start ||
2500                      meta > xdp->data))
2501                 return -EINVAL;
2502         if (unlikely((metalen & (sizeof(__u32) - 1)) ||
2503                      (metalen > 32)))
2504                 return -EACCES;
2505
2506         xdp->data_meta = meta;
2507
2508         return 0;
2509 }
2510
2511 static const struct bpf_func_proto bpf_xdp_adjust_meta_proto = {
2512         .func           = bpf_xdp_adjust_meta,
2513         .gpl_only       = false,
2514         .ret_type       = RET_INTEGER,
2515         .arg1_type      = ARG_PTR_TO_CTX,
2516         .arg2_type      = ARG_ANYTHING,
2517 };
2518
2519 static int __bpf_tx_xdp(struct net_device *dev,
2520                         struct bpf_map *map,
2521                         struct xdp_buff *xdp,
2522                         u32 index)
2523 {
2524         int err;
2525
2526         if (!dev->netdev_ops->ndo_xdp_xmit) {
2527                 return -EOPNOTSUPP;
2528         }
2529
2530         err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
2531         if (err)
2532                 return err;
2533         dev->netdev_ops->ndo_xdp_flush(dev);
2534         return 0;
2535 }
2536
2537 static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
2538                             struct bpf_map *map,
2539                             struct xdp_buff *xdp,
2540                             u32 index)
2541 {
2542         int err;
2543
2544         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2545                 struct net_device *dev = fwd;
2546
2547                 if (!dev->netdev_ops->ndo_xdp_xmit)
2548                         return -EOPNOTSUPP;
2549
2550                 err = dev->netdev_ops->ndo_xdp_xmit(dev, xdp);
2551                 if (err)
2552                         return err;
2553                 __dev_map_insert_ctx(map, index);
2554
2555         } else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
2556                 struct bpf_cpu_map_entry *rcpu = fwd;
2557
2558                 err = cpu_map_enqueue(rcpu, xdp, dev_rx);
2559                 if (err)
2560                         return err;
2561                 __cpu_map_insert_ctx(map, index);
2562         }
2563         return 0;
2564 }
2565
2566 void xdp_do_flush_map(void)
2567 {
2568         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2569         struct bpf_map *map = ri->map_to_flush;
2570
2571         ri->map_to_flush = NULL;
2572         if (map) {
2573                 switch (map->map_type) {
2574                 case BPF_MAP_TYPE_DEVMAP:
2575                         __dev_map_flush(map);
2576                         break;
2577                 case BPF_MAP_TYPE_CPUMAP:
2578                         __cpu_map_flush(map);
2579                         break;
2580                 default:
2581                         break;
2582                 }
2583         }
2584 }
2585 EXPORT_SYMBOL_GPL(xdp_do_flush_map);
2586
2587 static void *__xdp_map_lookup_elem(struct bpf_map *map, u32 index)
2588 {
2589         switch (map->map_type) {
2590         case BPF_MAP_TYPE_DEVMAP:
2591                 return __dev_map_lookup_elem(map, index);
2592         case BPF_MAP_TYPE_CPUMAP:
2593                 return __cpu_map_lookup_elem(map, index);
2594         default:
2595                 return NULL;
2596         }
2597 }
2598
2599 static inline bool xdp_map_invalid(const struct bpf_prog *xdp_prog,
2600                                    unsigned long aux)
2601 {
2602         return (unsigned long)xdp_prog->aux != aux;
2603 }
2604
2605 static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
2606                                struct bpf_prog *xdp_prog)
2607 {
2608         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2609         unsigned long map_owner = ri->map_owner;
2610         struct bpf_map *map = ri->map;
2611         u32 index = ri->ifindex;
2612         void *fwd = NULL;
2613         int err;
2614
2615         ri->ifindex = 0;
2616         ri->map = NULL;
2617         ri->map_owner = 0;
2618
2619         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2620                 err = -EFAULT;
2621                 map = NULL;
2622                 goto err;
2623         }
2624
2625         fwd = __xdp_map_lookup_elem(map, index);
2626         if (!fwd) {
2627                 err = -EINVAL;
2628                 goto err;
2629         }
2630         if (ri->map_to_flush && ri->map_to_flush != map)
2631                 xdp_do_flush_map();
2632
2633         err = __bpf_tx_xdp_map(dev, fwd, map, xdp, index);
2634         if (unlikely(err))
2635                 goto err;
2636
2637         ri->map_to_flush = map;
2638         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2639         return 0;
2640 err:
2641         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2642         return err;
2643 }
2644
2645 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
2646                     struct bpf_prog *xdp_prog)
2647 {
2648         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2649         struct net_device *fwd;
2650         u32 index = ri->ifindex;
2651         int err;
2652
2653         if (ri->map)
2654                 return xdp_do_redirect_map(dev, xdp, xdp_prog);
2655
2656         fwd = dev_get_by_index_rcu(dev_net(dev), index);
2657         ri->ifindex = 0;
2658         if (unlikely(!fwd)) {
2659                 err = -EINVAL;
2660                 goto err;
2661         }
2662
2663         err = __bpf_tx_xdp(fwd, NULL, xdp, 0);
2664         if (unlikely(err))
2665                 goto err;
2666
2667         _trace_xdp_redirect(dev, xdp_prog, index);
2668         return 0;
2669 err:
2670         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2671         return err;
2672 }
2673 EXPORT_SYMBOL_GPL(xdp_do_redirect);
2674
2675 static int __xdp_generic_ok_fwd_dev(struct sk_buff *skb, struct net_device *fwd)
2676 {
2677         unsigned int len;
2678
2679         if (unlikely(!(fwd->flags & IFF_UP)))
2680                 return -ENETDOWN;
2681
2682         len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
2683         if (skb->len > len)
2684                 return -EMSGSIZE;
2685
2686         return 0;
2687 }
2688
2689 int xdp_do_generic_redirect_map(struct net_device *dev, struct sk_buff *skb,
2690                                 struct bpf_prog *xdp_prog)
2691 {
2692         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2693         unsigned long map_owner = ri->map_owner;
2694         struct bpf_map *map = ri->map;
2695         struct net_device *fwd = NULL;
2696         u32 index = ri->ifindex;
2697         int err = 0;
2698
2699         ri->ifindex = 0;
2700         ri->map = NULL;
2701         ri->map_owner = 0;
2702
2703         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2704                 err = -EFAULT;
2705                 map = NULL;
2706                 goto err;
2707         }
2708         fwd = __xdp_map_lookup_elem(map, index);
2709         if (unlikely(!fwd)) {
2710                 err = -EINVAL;
2711                 goto err;
2712         }
2713
2714         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2715                 if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
2716                         goto err;
2717                 skb->dev = fwd;
2718         } else {
2719                 /* TODO: Handle BPF_MAP_TYPE_CPUMAP */
2720                 err = -EBADRQC;
2721                 goto err;
2722         }
2723
2724         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2725         return 0;
2726 err:
2727         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2728         return err;
2729 }
2730
2731 int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
2732                             struct bpf_prog *xdp_prog)
2733 {
2734         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2735         u32 index = ri->ifindex;
2736         struct net_device *fwd;
2737         int err = 0;
2738
2739         if (ri->map)
2740                 return xdp_do_generic_redirect_map(dev, skb, xdp_prog);
2741
2742         ri->ifindex = 0;
2743         fwd = dev_get_by_index_rcu(dev_net(dev), index);
2744         if (unlikely(!fwd)) {
2745                 err = -EINVAL;
2746                 goto err;
2747         }
2748
2749         if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
2750                 goto err;
2751
2752         skb->dev = fwd;
2753         _trace_xdp_redirect(dev, xdp_prog, index);
2754         return 0;
2755 err:
2756         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2757         return err;
2758 }
2759 EXPORT_SYMBOL_GPL(xdp_do_generic_redirect);
2760
2761 BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
2762 {
2763         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2764
2765         if (unlikely(flags))
2766                 return XDP_ABORTED;
2767
2768         ri->ifindex = ifindex;
2769         ri->flags = flags;
2770         ri->map = NULL;
2771         ri->map_owner = 0;
2772
2773         return XDP_REDIRECT;
2774 }
2775
2776 static const struct bpf_func_proto bpf_xdp_redirect_proto = {
2777         .func           = bpf_xdp_redirect,
2778         .gpl_only       = false,
2779         .ret_type       = RET_INTEGER,
2780         .arg1_type      = ARG_ANYTHING,
2781         .arg2_type      = ARG_ANYTHING,
2782 };
2783
2784 BPF_CALL_4(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex, u64, flags,
2785            unsigned long, map_owner)
2786 {
2787         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2788
2789         if (unlikely(flags))
2790                 return XDP_ABORTED;
2791
2792         ri->ifindex = ifindex;
2793         ri->flags = flags;
2794         ri->map = map;
2795         ri->map_owner = map_owner;
2796
2797         return XDP_REDIRECT;
2798 }
2799
2800 /* Note, arg4 is hidden from users and populated by the verifier
2801  * with the right pointer.
2802  */
2803 static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
2804         .func           = bpf_xdp_redirect_map,
2805         .gpl_only       = false,
2806         .ret_type       = RET_INTEGER,
2807         .arg1_type      = ARG_CONST_MAP_PTR,
2808         .arg2_type      = ARG_ANYTHING,
2809         .arg3_type      = ARG_ANYTHING,
2810 };
2811
2812 bool bpf_helper_changes_pkt_data(void *func)
2813 {
2814         if (func == bpf_skb_vlan_push ||
2815             func == bpf_skb_vlan_pop ||
2816             func == bpf_skb_store_bytes ||
2817             func == bpf_skb_change_proto ||
2818             func == bpf_skb_change_head ||
2819             func == bpf_skb_change_tail ||
2820             func == bpf_skb_adjust_room ||
2821             func == bpf_skb_pull_data ||
2822             func == bpf_clone_redirect ||
2823             func == bpf_l3_csum_replace ||
2824             func == bpf_l4_csum_replace ||
2825             func == bpf_xdp_adjust_head ||
2826             func == bpf_xdp_adjust_meta)
2827                 return true;
2828
2829         return false;
2830 }
2831
2832 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2833                                   unsigned long off, unsigned long len)
2834 {
2835         void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2836
2837         if (unlikely(!ptr))
2838                 return len;
2839         if (ptr != dst_buff)
2840                 memcpy(dst_buff, ptr, len);
2841
2842         return 0;
2843 }
2844
2845 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2846            u64, flags, void *, meta, u64, meta_size)
2847 {
2848         u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2849
2850         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2851                 return -EINVAL;
2852         if (unlikely(skb_size > skb->len))
2853                 return -EFAULT;
2854
2855         return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2856                                 bpf_skb_copy);
2857 }
2858
2859 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2860         .func           = bpf_skb_event_output,
2861         .gpl_only       = true,
2862         .ret_type       = RET_INTEGER,
2863         .arg1_type      = ARG_PTR_TO_CTX,
2864         .arg2_type      = ARG_CONST_MAP_PTR,
2865         .arg3_type      = ARG_ANYTHING,
2866         .arg4_type      = ARG_PTR_TO_MEM,
2867         .arg5_type      = ARG_CONST_SIZE,
2868 };
2869
2870 static unsigned short bpf_tunnel_key_af(u64 flags)
2871 {
2872         return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2873 }
2874
2875 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2876            u32, size, u64, flags)
2877 {
2878         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2879         u8 compat[sizeof(struct bpf_tunnel_key)];
2880         void *to_orig = to;
2881         int err;
2882
2883         if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2884                 err = -EINVAL;
2885                 goto err_clear;
2886         }
2887         if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2888                 err = -EPROTO;
2889                 goto err_clear;
2890         }
2891         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2892                 err = -EINVAL;
2893                 switch (size) {
2894                 case offsetof(struct bpf_tunnel_key, tunnel_label):
2895                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2896                         goto set_compat;
2897                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2898                         /* Fixup deprecated structure layouts here, so we have
2899                          * a common path later on.
2900                          */
2901                         if (ip_tunnel_info_af(info) != AF_INET)
2902                                 goto err_clear;
2903 set_compat:
2904                         to = (struct bpf_tunnel_key *)compat;
2905                         break;
2906                 default:
2907                         goto err_clear;
2908                 }
2909         }
2910
2911         to->tunnel_id = be64_to_cpu(info->key.tun_id);
2912         to->tunnel_tos = info->key.tos;
2913         to->tunnel_ttl = info->key.ttl;
2914
2915         if (flags & BPF_F_TUNINFO_IPV6) {
2916                 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2917                        sizeof(to->remote_ipv6));
2918                 to->tunnel_label = be32_to_cpu(info->key.label);
2919         } else {
2920                 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2921         }
2922
2923         if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2924                 memcpy(to_orig, to, size);
2925
2926         return 0;
2927 err_clear:
2928         memset(to_orig, 0, size);
2929         return err;
2930 }
2931
2932 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2933         .func           = bpf_skb_get_tunnel_key,
2934         .gpl_only       = false,
2935         .ret_type       = RET_INTEGER,
2936         .arg1_type      = ARG_PTR_TO_CTX,
2937         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
2938         .arg3_type      = ARG_CONST_SIZE,
2939         .arg4_type      = ARG_ANYTHING,
2940 };
2941
2942 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2943 {
2944         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2945         int err;
2946
2947         if (unlikely(!info ||
2948                      !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2949                 err = -ENOENT;
2950                 goto err_clear;
2951         }
2952         if (unlikely(size < info->options_len)) {
2953                 err = -ENOMEM;
2954                 goto err_clear;
2955         }
2956
2957         ip_tunnel_info_opts_get(to, info);
2958         if (size > info->options_len)
2959                 memset(to + info->options_len, 0, size - info->options_len);
2960
2961         return info->options_len;
2962 err_clear:
2963         memset(to, 0, size);
2964         return err;
2965 }
2966
2967 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2968         .func           = bpf_skb_get_tunnel_opt,
2969         .gpl_only       = false,
2970         .ret_type       = RET_INTEGER,
2971         .arg1_type      = ARG_PTR_TO_CTX,
2972         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
2973         .arg3_type      = ARG_CONST_SIZE,
2974 };
2975
2976 static struct metadata_dst __percpu *md_dst;
2977
2978 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2979            const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2980 {
2981         struct metadata_dst *md = this_cpu_ptr(md_dst);
2982         u8 compat[sizeof(struct bpf_tunnel_key)];
2983         struct ip_tunnel_info *info;
2984
2985         if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2986                                BPF_F_DONT_FRAGMENT)))
2987                 return -EINVAL;
2988         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2989                 switch (size) {
2990                 case offsetof(struct bpf_tunnel_key, tunnel_label):
2991                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
2992                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2993                         /* Fixup deprecated structure layouts here, so we have
2994                          * a common path later on.
2995                          */
2996                         memcpy(compat, from, size);
2997                         memset(compat + size, 0, sizeof(compat) - size);
2998                         from = (const struct bpf_tunnel_key *) compat;
2999                         break;
3000                 default:
3001                         return -EINVAL;
3002                 }
3003         }
3004         if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
3005                      from->tunnel_ext))
3006                 return -EINVAL;
3007
3008         skb_dst_drop(skb);
3009         dst_hold((struct dst_entry *) md);
3010         skb_dst_set(skb, (struct dst_entry *) md);
3011
3012         info = &md->u.tun_info;
3013         info->mode = IP_TUNNEL_INFO_TX;
3014
3015         info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
3016         if (flags & BPF_F_DONT_FRAGMENT)
3017                 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
3018
3019         info->key.tun_id = cpu_to_be64(from->tunnel_id);
3020         info->key.tos = from->tunnel_tos;
3021         info->key.ttl = from->tunnel_ttl;
3022
3023         if (flags & BPF_F_TUNINFO_IPV6) {
3024                 info->mode |= IP_TUNNEL_INFO_IPV6;
3025                 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
3026                        sizeof(from->remote_ipv6));
3027                 info->key.label = cpu_to_be32(from->tunnel_label) &
3028                                   IPV6_FLOWLABEL_MASK;
3029         } else {
3030                 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
3031                 if (flags & BPF_F_ZERO_CSUM_TX)
3032                         info->key.tun_flags &= ~TUNNEL_CSUM;
3033         }
3034
3035         return 0;
3036 }
3037
3038 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
3039         .func           = bpf_skb_set_tunnel_key,
3040         .gpl_only       = false,
3041         .ret_type       = RET_INTEGER,
3042         .arg1_type      = ARG_PTR_TO_CTX,
3043         .arg2_type      = ARG_PTR_TO_MEM,
3044         .arg3_type      = ARG_CONST_SIZE,
3045         .arg4_type      = ARG_ANYTHING,
3046 };
3047
3048 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
3049            const u8 *, from, u32, size)
3050 {
3051         struct ip_tunnel_info *info = skb_tunnel_info(skb);
3052         const struct metadata_dst *md = this_cpu_ptr(md_dst);
3053
3054         if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
3055                 return -EINVAL;
3056         if (unlikely(size > IP_TUNNEL_OPTS_MAX))
3057                 return -ENOMEM;
3058
3059         ip_tunnel_info_opts_set(info, from, size);
3060
3061         return 0;
3062 }
3063
3064 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
3065         .func           = bpf_skb_set_tunnel_opt,
3066         .gpl_only       = false,
3067         .ret_type       = RET_INTEGER,
3068         .arg1_type      = ARG_PTR_TO_CTX,
3069         .arg2_type      = ARG_PTR_TO_MEM,
3070         .arg3_type      = ARG_CONST_SIZE,
3071 };
3072
3073 static const struct bpf_func_proto *
3074 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
3075 {
3076         if (!md_dst) {
3077                 struct metadata_dst __percpu *tmp;
3078
3079                 tmp = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
3080                                                 METADATA_IP_TUNNEL,
3081                                                 GFP_KERNEL);
3082                 if (!tmp)
3083                         return NULL;
3084                 if (cmpxchg(&md_dst, NULL, tmp))
3085                         metadata_dst_free_percpu(tmp);
3086         }
3087
3088         switch (which) {
3089         case BPF_FUNC_skb_set_tunnel_key:
3090                 return &bpf_skb_set_tunnel_key_proto;
3091         case BPF_FUNC_skb_set_tunnel_opt:
3092                 return &bpf_skb_set_tunnel_opt_proto;
3093         default:
3094                 return NULL;
3095         }
3096 }
3097
3098 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
3099            u32, idx)
3100 {
3101         struct bpf_array *array = container_of(map, struct bpf_array, map);
3102         struct cgroup *cgrp;
3103         struct sock *sk;
3104
3105         sk = skb_to_full_sk(skb);
3106         if (!sk || !sk_fullsock(sk))
3107                 return -ENOENT;
3108         if (unlikely(idx >= array->map.max_entries))
3109                 return -E2BIG;
3110
3111         cgrp = READ_ONCE(array->ptrs[idx]);
3112         if (unlikely(!cgrp))
3113                 return -EAGAIN;
3114
3115         return sk_under_cgroup_hierarchy(sk, cgrp);
3116 }
3117
3118 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
3119         .func           = bpf_skb_under_cgroup,
3120         .gpl_only       = false,
3121         .ret_type       = RET_INTEGER,
3122         .arg1_type      = ARG_PTR_TO_CTX,
3123         .arg2_type      = ARG_CONST_MAP_PTR,
3124         .arg3_type      = ARG_ANYTHING,
3125 };
3126
3127 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
3128                                   unsigned long off, unsigned long len)
3129 {
3130         memcpy(dst_buff, src_buff + off, len);
3131         return 0;
3132 }
3133
3134 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
3135            u64, flags, void *, meta, u64, meta_size)
3136 {
3137         u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
3138
3139         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
3140                 return -EINVAL;
3141         if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
3142                 return -EFAULT;
3143
3144         return bpf_event_output(map, flags, meta, meta_size, xdp->data,
3145                                 xdp_size, bpf_xdp_copy);
3146 }
3147
3148 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
3149         .func           = bpf_xdp_event_output,
3150         .gpl_only       = true,
3151         .ret_type       = RET_INTEGER,
3152         .arg1_type      = ARG_PTR_TO_CTX,
3153         .arg2_type      = ARG_CONST_MAP_PTR,
3154         .arg3_type      = ARG_ANYTHING,
3155         .arg4_type      = ARG_PTR_TO_MEM,
3156         .arg5_type      = ARG_CONST_SIZE,
3157 };
3158
3159 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
3160 {
3161         return skb->sk ? sock_gen_cookie(skb->sk) : 0;
3162 }
3163
3164 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
3165         .func           = bpf_get_socket_cookie,
3166         .gpl_only       = false,
3167         .ret_type       = RET_INTEGER,
3168         .arg1_type      = ARG_PTR_TO_CTX,
3169 };
3170
3171 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
3172 {
3173         struct sock *sk = sk_to_full_sk(skb->sk);
3174         kuid_t kuid;
3175
3176         if (!sk || !sk_fullsock(sk))
3177                 return overflowuid;
3178         kuid = sock_net_uid(sock_net(sk), sk);
3179         return from_kuid_munged(sock_net(sk)->user_ns, kuid);
3180 }
3181
3182 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
3183         .func           = bpf_get_socket_uid,
3184         .gpl_only       = false,
3185         .ret_type       = RET_INTEGER,
3186         .arg1_type      = ARG_PTR_TO_CTX,
3187 };
3188
3189 BPF_CALL_5(bpf_setsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3190            int, level, int, optname, char *, optval, int, optlen)
3191 {
3192         struct sock *sk = bpf_sock->sk;
3193         int ret = 0;
3194         int val;
3195
3196         if (!sk_fullsock(sk))
3197                 return -EINVAL;
3198
3199         if (level == SOL_SOCKET) {
3200                 if (optlen != sizeof(int))
3201                         return -EINVAL;
3202                 val = *((int *)optval);
3203
3204                 /* Only some socketops are supported */
3205                 switch (optname) {
3206                 case SO_RCVBUF:
3207                         sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
3208                         sk->sk_rcvbuf = max_t(int, val * 2, SOCK_MIN_RCVBUF);
3209                         break;
3210                 case SO_SNDBUF:
3211                         sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
3212                         sk->sk_sndbuf = max_t(int, val * 2, SOCK_MIN_SNDBUF);
3213                         break;
3214                 case SO_MAX_PACING_RATE:
3215                         sk->sk_max_pacing_rate = val;
3216                         sk->sk_pacing_rate = min(sk->sk_pacing_rate,
3217                                                  sk->sk_max_pacing_rate);
3218                         break;
3219                 case SO_PRIORITY:
3220                         sk->sk_priority = val;
3221                         break;
3222                 case SO_RCVLOWAT:
3223                         if (val < 0)
3224                                 val = INT_MAX;
3225                         sk->sk_rcvlowat = val ? : 1;
3226                         break;
3227                 case SO_MARK:
3228                         sk->sk_mark = val;
3229                         break;
3230                 default:
3231                         ret = -EINVAL;
3232                 }
3233 #ifdef CONFIG_INET
3234         } else if (level == SOL_TCP &&
3235                    sk->sk_prot->setsockopt == tcp_setsockopt) {
3236                 if (optname == TCP_CONGESTION) {
3237                         char name[TCP_CA_NAME_MAX];
3238                         bool reinit = bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN;
3239
3240                         strncpy(name, optval, min_t(long, optlen,
3241                                                     TCP_CA_NAME_MAX-1));
3242                         name[TCP_CA_NAME_MAX-1] = 0;
3243                         ret = tcp_set_congestion_control(sk, name, false, reinit);
3244                 } else {
3245                         struct tcp_sock *tp = tcp_sk(sk);
3246
3247                         if (optlen != sizeof(int))
3248                                 return -EINVAL;
3249
3250                         val = *((int *)optval);
3251                         /* Only some options are supported */
3252                         switch (optname) {
3253                         case TCP_BPF_IW:
3254                                 if (val <= 0 || tp->data_segs_out > 0)
3255                                         ret = -EINVAL;
3256                                 else
3257                                         tp->snd_cwnd = val;
3258                                 break;
3259                         case TCP_BPF_SNDCWND_CLAMP:
3260                                 if (val <= 0) {
3261                                         ret = -EINVAL;
3262                                 } else {
3263                                         tp->snd_cwnd_clamp = val;
3264                                         tp->snd_ssthresh = val;
3265                                 }
3266                                 break;
3267                         default:
3268                                 ret = -EINVAL;
3269                         }
3270                 }
3271 #endif
3272         } else {
3273                 ret = -EINVAL;
3274         }
3275         return ret;
3276 }
3277
3278 static const struct bpf_func_proto bpf_setsockopt_proto = {
3279         .func           = bpf_setsockopt,
3280         .gpl_only       = false,
3281         .ret_type       = RET_INTEGER,
3282         .arg1_type      = ARG_PTR_TO_CTX,
3283         .arg2_type      = ARG_ANYTHING,
3284         .arg3_type      = ARG_ANYTHING,
3285         .arg4_type      = ARG_PTR_TO_MEM,
3286         .arg5_type      = ARG_CONST_SIZE,
3287 };
3288
3289 BPF_CALL_5(bpf_getsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3290            int, level, int, optname, char *, optval, int, optlen)
3291 {
3292         struct sock *sk = bpf_sock->sk;
3293
3294         if (!sk_fullsock(sk))
3295                 goto err_clear;
3296
3297 #ifdef CONFIG_INET
3298         if (level == SOL_TCP && sk->sk_prot->getsockopt == tcp_getsockopt) {
3299                 if (optname == TCP_CONGESTION) {
3300                         struct inet_connection_sock *icsk = inet_csk(sk);
3301
3302                         if (!icsk->icsk_ca_ops || optlen <= 1)
3303                                 goto err_clear;
3304                         strncpy(optval, icsk->icsk_ca_ops->name, optlen);
3305                         optval[optlen - 1] = 0;
3306                 } else {
3307                         goto err_clear;
3308                 }
3309         } else {
3310                 goto err_clear;
3311         }
3312         return 0;
3313 #endif
3314 err_clear:
3315         memset(optval, 0, optlen);
3316         return -EINVAL;
3317 }
3318
3319 static const struct bpf_func_proto bpf_getsockopt_proto = {
3320         .func           = bpf_getsockopt,
3321         .gpl_only       = false,
3322         .ret_type       = RET_INTEGER,
3323         .arg1_type      = ARG_PTR_TO_CTX,
3324         .arg2_type      = ARG_ANYTHING,
3325         .arg3_type      = ARG_ANYTHING,
3326         .arg4_type      = ARG_PTR_TO_UNINIT_MEM,
3327         .arg5_type      = ARG_CONST_SIZE,
3328 };
3329
3330 static const struct bpf_func_proto *
3331 bpf_base_func_proto(enum bpf_func_id func_id)
3332 {
3333         switch (func_id) {
3334         case BPF_FUNC_map_lookup_elem:
3335                 return &bpf_map_lookup_elem_proto;
3336         case BPF_FUNC_map_update_elem:
3337                 return &bpf_map_update_elem_proto;
3338         case BPF_FUNC_map_delete_elem:
3339                 return &bpf_map_delete_elem_proto;
3340         case BPF_FUNC_get_prandom_u32:
3341                 return &bpf_get_prandom_u32_proto;
3342         case BPF_FUNC_get_smp_processor_id:
3343                 return &bpf_get_raw_smp_processor_id_proto;
3344         case BPF_FUNC_get_numa_node_id:
3345                 return &bpf_get_numa_node_id_proto;
3346         case BPF_FUNC_tail_call:
3347                 return &bpf_tail_call_proto;
3348         case BPF_FUNC_ktime_get_ns:
3349                 return &bpf_ktime_get_ns_proto;
3350         case BPF_FUNC_trace_printk:
3351                 if (capable(CAP_SYS_ADMIN))
3352                         return bpf_get_trace_printk_proto();
3353         default:
3354                 return NULL;
3355         }
3356 }
3357
3358 static const struct bpf_func_proto *
3359 sock_filter_func_proto(enum bpf_func_id func_id)
3360 {
3361         switch (func_id) {
3362         /* inet and inet6 sockets are created in a process
3363          * context so there is always a valid uid/gid
3364          */
3365         case BPF_FUNC_get_current_uid_gid:
3366                 return &bpf_get_current_uid_gid_proto;
3367         default:
3368                 return bpf_base_func_proto(func_id);
3369         }
3370 }
3371
3372 static const struct bpf_func_proto *
3373 sk_filter_func_proto(enum bpf_func_id func_id)
3374 {
3375         switch (func_id) {
3376         case BPF_FUNC_skb_load_bytes:
3377                 return &bpf_skb_load_bytes_proto;
3378         case BPF_FUNC_get_socket_cookie:
3379                 return &bpf_get_socket_cookie_proto;
3380         case BPF_FUNC_get_socket_uid:
3381                 return &bpf_get_socket_uid_proto;
3382         default:
3383                 return bpf_base_func_proto(func_id);
3384         }
3385 }
3386
3387 static const struct bpf_func_proto *
3388 tc_cls_act_func_proto(enum bpf_func_id func_id)
3389 {
3390         switch (func_id) {
3391         case BPF_FUNC_skb_store_bytes:
3392                 return &bpf_skb_store_bytes_proto;
3393         case BPF_FUNC_skb_load_bytes:
3394                 return &bpf_skb_load_bytes_proto;
3395         case BPF_FUNC_skb_pull_data:
3396                 return &bpf_skb_pull_data_proto;
3397         case BPF_FUNC_csum_diff:
3398                 return &bpf_csum_diff_proto;
3399         case BPF_FUNC_csum_update:
3400                 return &bpf_csum_update_proto;
3401         case BPF_FUNC_l3_csum_replace:
3402                 return &bpf_l3_csum_replace_proto;
3403         case BPF_FUNC_l4_csum_replace:
3404                 return &bpf_l4_csum_replace_proto;
3405         case BPF_FUNC_clone_redirect:
3406                 return &bpf_clone_redirect_proto;
3407         case BPF_FUNC_get_cgroup_classid:
3408                 return &bpf_get_cgroup_classid_proto;
3409         case BPF_FUNC_skb_vlan_push:
3410                 return &bpf_skb_vlan_push_proto;
3411         case BPF_FUNC_skb_vlan_pop:
3412                 return &bpf_skb_vlan_pop_proto;
3413         case BPF_FUNC_skb_change_proto:
3414                 return &bpf_skb_change_proto_proto;
3415         case BPF_FUNC_skb_change_type:
3416                 return &bpf_skb_change_type_proto;
3417         case BPF_FUNC_skb_adjust_room:
3418                 return &bpf_skb_adjust_room_proto;
3419         case BPF_FUNC_skb_change_tail:
3420                 return &bpf_skb_change_tail_proto;
3421         case BPF_FUNC_skb_get_tunnel_key:
3422                 return &bpf_skb_get_tunnel_key_proto;
3423         case BPF_FUNC_skb_set_tunnel_key:
3424                 return bpf_get_skb_set_tunnel_proto(func_id);
3425         case BPF_FUNC_skb_get_tunnel_opt:
3426                 return &bpf_skb_get_tunnel_opt_proto;
3427         case BPF_FUNC_skb_set_tunnel_opt:
3428                 return bpf_get_skb_set_tunnel_proto(func_id);
3429         case BPF_FUNC_redirect:
3430                 return &bpf_redirect_proto;
3431         case BPF_FUNC_get_route_realm:
3432                 return &bpf_get_route_realm_proto;
3433         case BPF_FUNC_get_hash_recalc:
3434                 return &bpf_get_hash_recalc_proto;
3435         case BPF_FUNC_set_hash_invalid:
3436                 return &bpf_set_hash_invalid_proto;
3437         case BPF_FUNC_set_hash:
3438                 return &bpf_set_hash_proto;
3439         case BPF_FUNC_perf_event_output:
3440                 return &bpf_skb_event_output_proto;
3441         case BPF_FUNC_get_smp_processor_id:
3442                 return &bpf_get_smp_processor_id_proto;
3443         case BPF_FUNC_skb_under_cgroup:
3444                 return &bpf_skb_under_cgroup_proto;
3445         case BPF_FUNC_get_socket_cookie:
3446                 return &bpf_get_socket_cookie_proto;
3447         case BPF_FUNC_get_socket_uid:
3448                 return &bpf_get_socket_uid_proto;
3449         default:
3450                 return bpf_base_func_proto(func_id);
3451         }
3452 }
3453
3454 static const struct bpf_func_proto *
3455 xdp_func_proto(enum bpf_func_id func_id)
3456 {
3457         switch (func_id) {
3458         case BPF_FUNC_perf_event_output:
3459                 return &bpf_xdp_event_output_proto;
3460         case BPF_FUNC_get_smp_processor_id:
3461                 return &bpf_get_smp_processor_id_proto;
3462         case BPF_FUNC_xdp_adjust_head:
3463                 return &bpf_xdp_adjust_head_proto;
3464         case BPF_FUNC_xdp_adjust_meta:
3465                 return &bpf_xdp_adjust_meta_proto;
3466         case BPF_FUNC_redirect:
3467                 return &bpf_xdp_redirect_proto;
3468         case BPF_FUNC_redirect_map:
3469                 return &bpf_xdp_redirect_map_proto;
3470         default:
3471                 return bpf_base_func_proto(func_id);
3472         }
3473 }
3474
3475 static const struct bpf_func_proto *
3476 lwt_inout_func_proto(enum bpf_func_id func_id)
3477 {
3478         switch (func_id) {
3479         case BPF_FUNC_skb_load_bytes:
3480                 return &bpf_skb_load_bytes_proto;
3481         case BPF_FUNC_skb_pull_data:
3482                 return &bpf_skb_pull_data_proto;
3483         case BPF_FUNC_csum_diff:
3484                 return &bpf_csum_diff_proto;
3485         case BPF_FUNC_get_cgroup_classid:
3486                 return &bpf_get_cgroup_classid_proto;
3487         case BPF_FUNC_get_route_realm:
3488                 return &bpf_get_route_realm_proto;
3489         case BPF_FUNC_get_hash_recalc:
3490                 return &bpf_get_hash_recalc_proto;
3491         case BPF_FUNC_perf_event_output:
3492                 return &bpf_skb_event_output_proto;
3493         case BPF_FUNC_get_smp_processor_id:
3494                 return &bpf_get_smp_processor_id_proto;
3495         case BPF_FUNC_skb_under_cgroup:
3496                 return &bpf_skb_under_cgroup_proto;
3497         default:
3498                 return bpf_base_func_proto(func_id);
3499         }
3500 }
3501
3502 static const struct bpf_func_proto *
3503         sock_ops_func_proto(enum bpf_func_id func_id)
3504 {
3505         switch (func_id) {
3506         case BPF_FUNC_setsockopt:
3507                 return &bpf_setsockopt_proto;
3508         case BPF_FUNC_getsockopt:
3509                 return &bpf_getsockopt_proto;
3510         case BPF_FUNC_sock_map_update:
3511                 return &bpf_sock_map_update_proto;
3512         default:
3513                 return bpf_base_func_proto(func_id);
3514         }
3515 }
3516
3517 static const struct bpf_func_proto *sk_skb_func_proto(enum bpf_func_id func_id)
3518 {
3519         switch (func_id) {
3520         case BPF_FUNC_skb_store_bytes:
3521                 return &bpf_skb_store_bytes_proto;
3522         case BPF_FUNC_skb_load_bytes:
3523                 return &bpf_skb_load_bytes_proto;
3524         case BPF_FUNC_skb_pull_data:
3525                 return &bpf_skb_pull_data_proto;
3526         case BPF_FUNC_skb_change_tail:
3527                 return &bpf_skb_change_tail_proto;
3528         case BPF_FUNC_skb_change_head:
3529                 return &bpf_skb_change_head_proto;
3530         case BPF_FUNC_get_socket_cookie:
3531                 return &bpf_get_socket_cookie_proto;
3532         case BPF_FUNC_get_socket_uid:
3533                 return &bpf_get_socket_uid_proto;
3534         case BPF_FUNC_sk_redirect_map:
3535                 return &bpf_sk_redirect_map_proto;
3536         default:
3537                 return bpf_base_func_proto(func_id);
3538         }
3539 }
3540
3541 static const struct bpf_func_proto *
3542 lwt_xmit_func_proto(enum bpf_func_id func_id)
3543 {
3544         switch (func_id) {
3545         case BPF_FUNC_skb_get_tunnel_key:
3546                 return &bpf_skb_get_tunnel_key_proto;
3547         case BPF_FUNC_skb_set_tunnel_key:
3548                 return bpf_get_skb_set_tunnel_proto(func_id);
3549         case BPF_FUNC_skb_get_tunnel_opt:
3550                 return &bpf_skb_get_tunnel_opt_proto;
3551         case BPF_FUNC_skb_set_tunnel_opt:
3552                 return bpf_get_skb_set_tunnel_proto(func_id);
3553         case BPF_FUNC_redirect:
3554                 return &bpf_redirect_proto;
3555         case BPF_FUNC_clone_redirect:
3556                 return &bpf_clone_redirect_proto;
3557         case BPF_FUNC_skb_change_tail:
3558                 return &bpf_skb_change_tail_proto;
3559         case BPF_FUNC_skb_change_head:
3560                 return &bpf_skb_change_head_proto;
3561         case BPF_FUNC_skb_store_bytes:
3562                 return &bpf_skb_store_bytes_proto;
3563         case BPF_FUNC_csum_update:
3564                 return &bpf_csum_update_proto;
3565         case BPF_FUNC_l3_csum_replace:
3566                 return &bpf_l3_csum_replace_proto;
3567         case BPF_FUNC_l4_csum_replace:
3568                 return &bpf_l4_csum_replace_proto;
3569         case BPF_FUNC_set_hash_invalid:
3570                 return &bpf_set_hash_invalid_proto;
3571         default:
3572                 return lwt_inout_func_proto(func_id);
3573         }
3574 }
3575
3576 static bool bpf_skb_is_valid_access(int off, int size, enum bpf_access_type type,
3577                                     struct bpf_insn_access_aux *info)
3578 {
3579         const int size_default = sizeof(__u32);
3580
3581         if (off < 0 || off >= sizeof(struct __sk_buff))
3582                 return false;
3583
3584         /* The verifier guarantees that size > 0. */
3585         if (off % size != 0)
3586                 return false;
3587
3588         switch (off) {
3589         case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3590                 if (off + size > offsetofend(struct __sk_buff, cb[4]))
3591                         return false;
3592                 break;
3593         case bpf_ctx_range_till(struct __sk_buff, remote_ip6[0], remote_ip6[3]):
3594         case bpf_ctx_range_till(struct __sk_buff, local_ip6[0], local_ip6[3]):
3595         case bpf_ctx_range_till(struct __sk_buff, remote_ip4, remote_ip4):
3596         case bpf_ctx_range_till(struct __sk_buff, local_ip4, local_ip4):
3597         case bpf_ctx_range(struct __sk_buff, data):
3598         case bpf_ctx_range(struct __sk_buff, data_meta):
3599         case bpf_ctx_range(struct __sk_buff, data_end):
3600                 if (size != size_default)
3601                         return false;
3602                 break;
3603         default:
3604                 /* Only narrow read access allowed for now. */
3605                 if (type == BPF_WRITE) {
3606                         if (size != size_default)
3607                                 return false;
3608                 } else {
3609                         bpf_ctx_record_field_size(info, size_default);
3610                         if (!bpf_ctx_narrow_access_ok(off, size, size_default))
3611                                 return false;
3612                 }
3613         }
3614
3615         return true;
3616 }
3617
3618 static bool sk_filter_is_valid_access(int off, int size,
3619                                       enum bpf_access_type type,
3620                                       struct bpf_insn_access_aux *info)
3621 {
3622         switch (off) {
3623         case bpf_ctx_range(struct __sk_buff, tc_classid):
3624         case bpf_ctx_range(struct __sk_buff, data):
3625         case bpf_ctx_range(struct __sk_buff, data_meta):
3626         case bpf_ctx_range(struct __sk_buff, data_end):
3627         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3628                 return false;
3629         }
3630
3631         if (type == BPF_WRITE) {
3632                 switch (off) {
3633                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3634                         break;
3635                 default:
3636                         return false;
3637                 }
3638         }
3639
3640         return bpf_skb_is_valid_access(off, size, type, info);
3641 }
3642
3643 static bool lwt_is_valid_access(int off, int size,
3644                                 enum bpf_access_type type,
3645                                 struct bpf_insn_access_aux *info)
3646 {
3647         switch (off) {
3648         case bpf_ctx_range(struct __sk_buff, tc_classid):
3649         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3650         case bpf_ctx_range(struct __sk_buff, data_meta):
3651                 return false;
3652         }
3653
3654         if (type == BPF_WRITE) {
3655                 switch (off) {
3656                 case bpf_ctx_range(struct __sk_buff, mark):
3657                 case bpf_ctx_range(struct __sk_buff, priority):
3658                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3659                         break;
3660                 default:
3661                         return false;
3662                 }
3663         }
3664
3665         switch (off) {
3666         case bpf_ctx_range(struct __sk_buff, data):
3667                 info->reg_type = PTR_TO_PACKET;
3668                 break;
3669         case bpf_ctx_range(struct __sk_buff, data_end):
3670                 info->reg_type = PTR_TO_PACKET_END;
3671                 break;
3672         }
3673
3674         return bpf_skb_is_valid_access(off, size, type, info);
3675 }
3676
3677 static bool sock_filter_is_valid_access(int off, int size,
3678                                         enum bpf_access_type type,
3679                                         struct bpf_insn_access_aux *info)
3680 {
3681         if (type == BPF_WRITE) {
3682                 switch (off) {
3683                 case offsetof(struct bpf_sock, bound_dev_if):
3684                 case offsetof(struct bpf_sock, mark):
3685                 case offsetof(struct bpf_sock, priority):
3686                         break;
3687                 default:
3688                         return false;
3689                 }
3690         }
3691
3692         if (off < 0 || off + size > sizeof(struct bpf_sock))
3693                 return false;
3694         /* The verifier guarantees that size > 0. */
3695         if (off % size != 0)
3696                 return false;
3697         if (size != sizeof(__u32))
3698                 return false;
3699
3700         return true;
3701 }
3702
3703 static int bpf_unclone_prologue(struct bpf_insn *insn_buf, bool direct_write,
3704                                 const struct bpf_prog *prog, int drop_verdict)
3705 {
3706         struct bpf_insn *insn = insn_buf;
3707
3708         if (!direct_write)
3709                 return 0;
3710
3711         /* if (!skb->cloned)
3712          *       goto start;
3713          *
3714          * (Fast-path, otherwise approximation that we might be
3715          *  a clone, do the rest in helper.)
3716          */
3717         *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
3718         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
3719         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
3720
3721         /* ret = bpf_skb_pull_data(skb, 0); */
3722         *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
3723         *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
3724         *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
3725                                BPF_FUNC_skb_pull_data);
3726         /* if (!ret)
3727          *      goto restore;
3728          * return TC_ACT_SHOT;
3729          */
3730         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
3731         *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, drop_verdict);
3732         *insn++ = BPF_EXIT_INSN();
3733
3734         /* restore: */
3735         *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
3736         /* start: */
3737         *insn++ = prog->insnsi[0];
3738
3739         return insn - insn_buf;
3740 }
3741
3742 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
3743                                const struct bpf_prog *prog)
3744 {
3745         return bpf_unclone_prologue(insn_buf, direct_write, prog, TC_ACT_SHOT);
3746 }
3747
3748 static bool tc_cls_act_is_valid_access(int off, int size,
3749                                        enum bpf_access_type type,
3750                                        struct bpf_insn_access_aux *info)
3751 {
3752         if (type == BPF_WRITE) {
3753                 switch (off) {
3754                 case bpf_ctx_range(struct __sk_buff, mark):
3755                 case bpf_ctx_range(struct __sk_buff, tc_index):
3756                 case bpf_ctx_range(struct __sk_buff, priority):
3757                 case bpf_ctx_range(struct __sk_buff, tc_classid):
3758                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
3759                         break;
3760                 default:
3761                         return false;
3762                 }
3763         }
3764
3765         switch (off) {
3766         case bpf_ctx_range(struct __sk_buff, data):
3767                 info->reg_type = PTR_TO_PACKET;
3768                 break;
3769         case bpf_ctx_range(struct __sk_buff, data_meta):
3770                 info->reg_type = PTR_TO_PACKET_META;
3771                 break;
3772         case bpf_ctx_range(struct __sk_buff, data_end):
3773                 info->reg_type = PTR_TO_PACKET_END;
3774                 break;
3775         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
3776                 return false;
3777         }
3778
3779         return bpf_skb_is_valid_access(off, size, type, info);
3780 }
3781
3782 static bool __is_valid_xdp_access(int off, int size)
3783 {
3784         if (off < 0 || off >= sizeof(struct xdp_md))
3785                 return false;
3786         if (off % size != 0)
3787                 return false;
3788         if (size != sizeof(__u32))
3789                 return false;
3790
3791         return true;
3792 }
3793
3794 static bool xdp_is_valid_access(int off, int size,
3795                                 enum bpf_access_type type,
3796                                 struct bpf_insn_access_aux *info)
3797 {
3798         if (type == BPF_WRITE)
3799                 return false;
3800
3801         switch (off) {
3802         case offsetof(struct xdp_md, data):
3803                 info->reg_type = PTR_TO_PACKET;
3804                 break;
3805         case offsetof(struct xdp_md, data_meta):
3806                 info->reg_type = PTR_TO_PACKET_META;
3807                 break;
3808         case offsetof(struct xdp_md, data_end):
3809                 info->reg_type = PTR_TO_PACKET_END;
3810                 break;
3811         }
3812
3813         return __is_valid_xdp_access(off, size);
3814 }
3815
3816 void bpf_warn_invalid_xdp_action(u32 act)
3817 {
3818         const u32 act_max = XDP_REDIRECT;
3819
3820         WARN_ONCE(1, "%s XDP return value %u, expect packet loss!\n",
3821                   act > act_max ? "Illegal" : "Driver unsupported",
3822                   act);
3823 }
3824 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
3825
3826 static bool __is_valid_sock_ops_access(int off, int size)
3827 {
3828         if (off < 0 || off >= sizeof(struct bpf_sock_ops))
3829                 return false;
3830         /* The verifier guarantees that size > 0. */
3831         if (off % size != 0)
3832                 return false;
3833         if (size != sizeof(__u32))
3834                 return false;
3835
3836         return true;
3837 }
3838
3839 static bool sock_ops_is_valid_access(int off, int size,
3840                                      enum bpf_access_type type,
3841                                      struct bpf_insn_access_aux *info)
3842 {
3843         if (type == BPF_WRITE) {
3844                 switch (off) {
3845                 case offsetof(struct bpf_sock_ops, op) ...
3846                      offsetof(struct bpf_sock_ops, replylong[3]):
3847                         break;
3848                 default:
3849                         return false;
3850                 }
3851         }
3852
3853         return __is_valid_sock_ops_access(off, size);
3854 }
3855
3856 static int sk_skb_prologue(struct bpf_insn *insn_buf, bool direct_write,
3857                            const struct bpf_prog *prog)
3858 {
3859         return bpf_unclone_prologue(insn_buf, direct_write, prog, SK_DROP);
3860 }
3861
3862 static bool sk_skb_is_valid_access(int off, int size,
3863                                    enum bpf_access_type type,
3864                                    struct bpf_insn_access_aux *info)
3865 {
3866         switch (off) {
3867         case bpf_ctx_range(struct __sk_buff, tc_classid):
3868         case bpf_ctx_range(struct __sk_buff, data_meta):
3869                 return false;
3870         }
3871
3872         if (type == BPF_WRITE) {
3873                 switch (off) {
3874                 case bpf_ctx_range(struct __sk_buff, tc_index):
3875                 case bpf_ctx_range(struct __sk_buff, priority):
3876                         break;
3877                 default:
3878                         return false;
3879                 }
3880         }
3881
3882         switch (off) {
3883         case bpf_ctx_range(struct __sk_buff, mark):
3884                 return false;
3885         case bpf_ctx_range(struct __sk_buff, data):
3886                 info->reg_type = PTR_TO_PACKET;
3887                 break;
3888         case bpf_ctx_range(struct __sk_buff, data_end):
3889                 info->reg_type = PTR_TO_PACKET_END;
3890                 break;
3891         }
3892
3893         return bpf_skb_is_valid_access(off, size, type, info);
3894 }
3895
3896 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
3897                                   const struct bpf_insn *si,
3898                                   struct bpf_insn *insn_buf,
3899                                   struct bpf_prog *prog, u32 *target_size)
3900 {
3901         struct bpf_insn *insn = insn_buf;
3902         int off;
3903
3904         switch (si->off) {
3905         case offsetof(struct __sk_buff, len):
3906                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3907                                       bpf_target_off(struct sk_buff, len, 4,
3908                                                      target_size));
3909                 break;
3910
3911         case offsetof(struct __sk_buff, protocol):
3912                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3913                                       bpf_target_off(struct sk_buff, protocol, 2,
3914                                                      target_size));
3915                 break;
3916
3917         case offsetof(struct __sk_buff, vlan_proto):
3918                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3919                                       bpf_target_off(struct sk_buff, vlan_proto, 2,
3920                                                      target_size));
3921                 break;
3922
3923         case offsetof(struct __sk_buff, priority):
3924                 if (type == BPF_WRITE)
3925                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3926                                               bpf_target_off(struct sk_buff, priority, 4,
3927                                                              target_size));
3928                 else
3929                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3930                                               bpf_target_off(struct sk_buff, priority, 4,
3931                                                              target_size));
3932                 break;
3933
3934         case offsetof(struct __sk_buff, ingress_ifindex):
3935                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3936                                       bpf_target_off(struct sk_buff, skb_iif, 4,
3937                                                      target_size));
3938                 break;
3939
3940         case offsetof(struct __sk_buff, ifindex):
3941                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
3942                                       si->dst_reg, si->src_reg,
3943                                       offsetof(struct sk_buff, dev));
3944                 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
3945                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
3946                                       bpf_target_off(struct net_device, ifindex, 4,
3947                                                      target_size));
3948                 break;
3949
3950         case offsetof(struct __sk_buff, hash):
3951                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3952                                       bpf_target_off(struct sk_buff, hash, 4,
3953                                                      target_size));
3954                 break;
3955
3956         case offsetof(struct __sk_buff, mark):
3957                 if (type == BPF_WRITE)
3958                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
3959                                               bpf_target_off(struct sk_buff, mark, 4,
3960                                                              target_size));
3961                 else
3962                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
3963                                               bpf_target_off(struct sk_buff, mark, 4,
3964                                                              target_size));
3965                 break;
3966
3967         case offsetof(struct __sk_buff, pkt_type):
3968                 *target_size = 1;
3969                 *insn++ = BPF_LDX_MEM(BPF_B, si->dst_reg, si->src_reg,
3970                                       PKT_TYPE_OFFSET());
3971                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, PKT_TYPE_MAX);
3972 #ifdef __BIG_ENDIAN_BITFIELD
3973                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 5);
3974 #endif
3975                 break;
3976
3977         case offsetof(struct __sk_buff, queue_mapping):
3978                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3979                                       bpf_target_off(struct sk_buff, queue_mapping, 2,
3980                                                      target_size));
3981                 break;
3982
3983         case offsetof(struct __sk_buff, vlan_present):
3984         case offsetof(struct __sk_buff, vlan_tci):
3985                 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
3986
3987                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
3988                                       bpf_target_off(struct sk_buff, vlan_tci, 2,
3989                                                      target_size));
3990                 if (si->off == offsetof(struct __sk_buff, vlan_tci)) {
3991                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg,
3992                                                 ~VLAN_TAG_PRESENT);
3993                 } else {
3994                         *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 12);
3995                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, 1);
3996                 }
3997                 break;
3998
3999         case offsetof(struct __sk_buff, cb[0]) ...
4000              offsetofend(struct __sk_buff, cb[4]) - 1:
4001                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
4002                 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
4003                               offsetof(struct qdisc_skb_cb, data)) %
4004                              sizeof(__u64));
4005
4006                 prog->cb_access = 1;
4007                 off  = si->off;
4008                 off -= offsetof(struct __sk_buff, cb[0]);
4009                 off += offsetof(struct sk_buff, cb);
4010                 off += offsetof(struct qdisc_skb_cb, data);
4011                 if (type == BPF_WRITE)
4012                         *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
4013                                               si->src_reg, off);
4014                 else
4015                         *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
4016                                               si->src_reg, off);
4017                 break;
4018
4019         case offsetof(struct __sk_buff, tc_classid):
4020                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
4021
4022                 off  = si->off;
4023                 off -= offsetof(struct __sk_buff, tc_classid);
4024                 off += offsetof(struct sk_buff, cb);
4025                 off += offsetof(struct qdisc_skb_cb, tc_classid);
4026                 *target_size = 2;
4027                 if (type == BPF_WRITE)
4028                         *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
4029                                               si->src_reg, off);
4030                 else
4031                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
4032                                               si->src_reg, off);
4033                 break;
4034
4035         case offsetof(struct __sk_buff, data):
4036                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
4037                                       si->dst_reg, si->src_reg,
4038                                       offsetof(struct sk_buff, data));
4039                 break;
4040
4041         case offsetof(struct __sk_buff, data_meta):
4042                 off  = si->off;
4043                 off -= offsetof(struct __sk_buff, data_meta);
4044                 off += offsetof(struct sk_buff, cb);
4045                 off += offsetof(struct bpf_skb_data_end, data_meta);
4046                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4047                                       si->src_reg, off);
4048                 break;
4049
4050         case offsetof(struct __sk_buff, data_end):
4051                 off  = si->off;
4052                 off -= offsetof(struct __sk_buff, data_end);
4053                 off += offsetof(struct sk_buff, cb);
4054                 off += offsetof(struct bpf_skb_data_end, data_end);
4055                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4056                                       si->src_reg, off);
4057                 break;
4058
4059         case offsetof(struct __sk_buff, tc_index):
4060 #ifdef CONFIG_NET_SCHED
4061                 if (type == BPF_WRITE)
4062                         *insn++ = BPF_STX_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                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4067                                               bpf_target_off(struct sk_buff, tc_index, 2,
4068                                                              target_size));
4069 #else
4070                 *target_size = 2;
4071                 if (type == BPF_WRITE)
4072                         *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
4073                 else
4074                         *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4075 #endif
4076                 break;
4077
4078         case offsetof(struct __sk_buff, napi_id):
4079 #if defined(CONFIG_NET_RX_BUSY_POLL)
4080                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4081                                       bpf_target_off(struct sk_buff, napi_id, 4,
4082                                                      target_size));
4083                 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
4084                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4085 #else
4086                 *target_size = 4;
4087                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4088 #endif
4089                 break;
4090         case offsetof(struct __sk_buff, family):
4091                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4092
4093                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4094                                       si->dst_reg, si->src_reg,
4095                                       offsetof(struct sk_buff, sk));
4096                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4097                                       bpf_target_off(struct sock_common,
4098                                                      skc_family,
4099                                                      2, target_size));
4100                 break;
4101         case offsetof(struct __sk_buff, remote_ip4):
4102                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4103
4104                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4105                                       si->dst_reg, si->src_reg,
4106                                       offsetof(struct sk_buff, sk));
4107                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4108                                       bpf_target_off(struct sock_common,
4109                                                      skc_daddr,
4110                                                      4, target_size));
4111                 break;
4112         case offsetof(struct __sk_buff, local_ip4):
4113                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4114                                           skc_rcv_saddr) != 4);
4115
4116                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4117                                       si->dst_reg, si->src_reg,
4118                                       offsetof(struct sk_buff, sk));
4119                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4120                                       bpf_target_off(struct sock_common,
4121                                                      skc_rcv_saddr,
4122                                                      4, target_size));
4123                 break;
4124         case offsetof(struct __sk_buff, remote_ip6[0]) ...
4125              offsetof(struct __sk_buff, remote_ip6[3]):
4126 #if IS_ENABLED(CONFIG_IPV6)
4127                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4128                                           skc_v6_daddr.s6_addr32[0]) != 4);
4129
4130                 off = si->off;
4131                 off -= offsetof(struct __sk_buff, remote_ip6[0]);
4132
4133                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4134                                       si->dst_reg, si->src_reg,
4135                                       offsetof(struct sk_buff, sk));
4136                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4137                                       offsetof(struct sock_common,
4138                                                skc_v6_daddr.s6_addr32[0]) +
4139                                       off);
4140 #else
4141                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4142 #endif
4143                 break;
4144         case offsetof(struct __sk_buff, local_ip6[0]) ...
4145              offsetof(struct __sk_buff, local_ip6[3]):
4146 #if IS_ENABLED(CONFIG_IPV6)
4147                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4148                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4149
4150                 off = si->off;
4151                 off -= offsetof(struct __sk_buff, local_ip6[0]);
4152
4153                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4154                                       si->dst_reg, si->src_reg,
4155                                       offsetof(struct sk_buff, sk));
4156                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4157                                       offsetof(struct sock_common,
4158                                                skc_v6_rcv_saddr.s6_addr32[0]) +
4159                                       off);
4160 #else
4161                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4162 #endif
4163                 break;
4164
4165         case offsetof(struct __sk_buff, remote_port):
4166                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4167
4168                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4169                                       si->dst_reg, si->src_reg,
4170                                       offsetof(struct sk_buff, sk));
4171                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4172                                       bpf_target_off(struct sock_common,
4173                                                      skc_dport,
4174                                                      2, target_size));
4175 #ifndef __BIG_ENDIAN_BITFIELD
4176                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4177 #endif
4178                 break;
4179
4180         case offsetof(struct __sk_buff, local_port):
4181                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4182
4183                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4184                                       si->dst_reg, si->src_reg,
4185                                       offsetof(struct sk_buff, sk));
4186                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4187                                       bpf_target_off(struct sock_common,
4188                                                      skc_num, 2, target_size));
4189                 break;
4190         }
4191
4192         return insn - insn_buf;
4193 }
4194
4195 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
4196                                           const struct bpf_insn *si,
4197                                           struct bpf_insn *insn_buf,
4198                                           struct bpf_prog *prog, u32 *target_size)
4199 {
4200         struct bpf_insn *insn = insn_buf;
4201
4202         switch (si->off) {
4203         case offsetof(struct bpf_sock, bound_dev_if):
4204                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
4205
4206                 if (type == BPF_WRITE)
4207                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4208                                         offsetof(struct sock, sk_bound_dev_if));
4209                 else
4210                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4211                                       offsetof(struct sock, sk_bound_dev_if));
4212                 break;
4213
4214         case offsetof(struct bpf_sock, mark):
4215                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_mark) != 4);
4216
4217                 if (type == BPF_WRITE)
4218                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4219                                         offsetof(struct sock, sk_mark));
4220                 else
4221                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4222                                       offsetof(struct sock, sk_mark));
4223                 break;
4224
4225         case offsetof(struct bpf_sock, priority):
4226                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_priority) != 4);
4227
4228                 if (type == BPF_WRITE)
4229                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4230                                         offsetof(struct sock, sk_priority));
4231                 else
4232                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4233                                       offsetof(struct sock, sk_priority));
4234                 break;
4235
4236         case offsetof(struct bpf_sock, family):
4237                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
4238
4239                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4240                                       offsetof(struct sock, sk_family));
4241                 break;
4242
4243         case offsetof(struct bpf_sock, type):
4244                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4245                                       offsetof(struct sock, __sk_flags_offset));
4246                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
4247                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
4248                 break;
4249
4250         case offsetof(struct bpf_sock, protocol):
4251                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4252                                       offsetof(struct sock, __sk_flags_offset));
4253                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
4254                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
4255                 break;
4256         }
4257
4258         return insn - insn_buf;
4259 }
4260
4261 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
4262                                          const struct bpf_insn *si,
4263                                          struct bpf_insn *insn_buf,
4264                                          struct bpf_prog *prog, u32 *target_size)
4265 {
4266         struct bpf_insn *insn = insn_buf;
4267
4268         switch (si->off) {
4269         case offsetof(struct __sk_buff, ifindex):
4270                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
4271                                       si->dst_reg, si->src_reg,
4272                                       offsetof(struct sk_buff, dev));
4273                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4274                                       bpf_target_off(struct net_device, ifindex, 4,
4275                                                      target_size));
4276                 break;
4277         default:
4278                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4279                                               target_size);
4280         }
4281
4282         return insn - insn_buf;
4283 }
4284
4285 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
4286                                   const struct bpf_insn *si,
4287                                   struct bpf_insn *insn_buf,
4288                                   struct bpf_prog *prog, u32 *target_size)
4289 {
4290         struct bpf_insn *insn = insn_buf;
4291
4292         switch (si->off) {
4293         case offsetof(struct xdp_md, data):
4294                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
4295                                       si->dst_reg, si->src_reg,
4296                                       offsetof(struct xdp_buff, data));
4297                 break;
4298         case offsetof(struct xdp_md, data_meta):
4299                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_meta),
4300                                       si->dst_reg, si->src_reg,
4301                                       offsetof(struct xdp_buff, data_meta));
4302                 break;
4303         case offsetof(struct xdp_md, data_end):
4304                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
4305                                       si->dst_reg, si->src_reg,
4306                                       offsetof(struct xdp_buff, data_end));
4307                 break;
4308         }
4309
4310         return insn - insn_buf;
4311 }
4312
4313 static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
4314                                        const struct bpf_insn *si,
4315                                        struct bpf_insn *insn_buf,
4316                                        struct bpf_prog *prog,
4317                                        u32 *target_size)
4318 {
4319         struct bpf_insn *insn = insn_buf;
4320         int off;
4321
4322         switch (si->off) {
4323         case offsetof(struct bpf_sock_ops, op) ...
4324              offsetof(struct bpf_sock_ops, replylong[3]):
4325                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, op) !=
4326                              FIELD_SIZEOF(struct bpf_sock_ops_kern, op));
4327                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, reply) !=
4328                              FIELD_SIZEOF(struct bpf_sock_ops_kern, reply));
4329                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, replylong) !=
4330                              FIELD_SIZEOF(struct bpf_sock_ops_kern, replylong));
4331                 off = si->off;
4332                 off -= offsetof(struct bpf_sock_ops, op);
4333                 off += offsetof(struct bpf_sock_ops_kern, op);
4334                 if (type == BPF_WRITE)
4335                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4336                                               off);
4337                 else
4338                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4339                                               off);
4340                 break;
4341
4342         case offsetof(struct bpf_sock_ops, family):
4343                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4344
4345                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4346                                               struct bpf_sock_ops_kern, sk),
4347                                       si->dst_reg, si->src_reg,
4348                                       offsetof(struct bpf_sock_ops_kern, sk));
4349                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4350                                       offsetof(struct sock_common, skc_family));
4351                 break;
4352
4353         case offsetof(struct bpf_sock_ops, remote_ip4):
4354                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4355
4356                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4357                                                 struct bpf_sock_ops_kern, sk),
4358                                       si->dst_reg, si->src_reg,
4359                                       offsetof(struct bpf_sock_ops_kern, sk));
4360                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4361                                       offsetof(struct sock_common, skc_daddr));
4362                 break;
4363
4364         case offsetof(struct bpf_sock_ops, local_ip4):
4365                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_rcv_saddr) != 4);
4366
4367                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4368                                               struct bpf_sock_ops_kern, sk),
4369                                       si->dst_reg, si->src_reg,
4370                                       offsetof(struct bpf_sock_ops_kern, sk));
4371                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4372                                       offsetof(struct sock_common,
4373                                                skc_rcv_saddr));
4374                 break;
4375
4376         case offsetof(struct bpf_sock_ops, remote_ip6[0]) ...
4377              offsetof(struct bpf_sock_ops, remote_ip6[3]):
4378 #if IS_ENABLED(CONFIG_IPV6)
4379                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4380                                           skc_v6_daddr.s6_addr32[0]) != 4);
4381
4382                 off = si->off;
4383                 off -= offsetof(struct bpf_sock_ops, remote_ip6[0]);
4384                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4385                                                 struct bpf_sock_ops_kern, sk),
4386                                       si->dst_reg, si->src_reg,
4387                                       offsetof(struct bpf_sock_ops_kern, sk));
4388                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4389                                       offsetof(struct sock_common,
4390                                                skc_v6_daddr.s6_addr32[0]) +
4391                                       off);
4392 #else
4393                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4394 #endif
4395                 break;
4396
4397         case offsetof(struct bpf_sock_ops, local_ip6[0]) ...
4398              offsetof(struct bpf_sock_ops, local_ip6[3]):
4399 #if IS_ENABLED(CONFIG_IPV6)
4400                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4401                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4402
4403                 off = si->off;
4404                 off -= offsetof(struct bpf_sock_ops, local_ip6[0]);
4405                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4406                                                 struct bpf_sock_ops_kern, sk),
4407                                       si->dst_reg, si->src_reg,
4408                                       offsetof(struct bpf_sock_ops_kern, sk));
4409                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4410                                       offsetof(struct sock_common,
4411                                                skc_v6_rcv_saddr.s6_addr32[0]) +
4412                                       off);
4413 #else
4414                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4415 #endif
4416                 break;
4417
4418         case offsetof(struct bpf_sock_ops, remote_port):
4419                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4420
4421                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4422                                                 struct bpf_sock_ops_kern, sk),
4423                                       si->dst_reg, si->src_reg,
4424                                       offsetof(struct bpf_sock_ops_kern, sk));
4425                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4426                                       offsetof(struct sock_common, skc_dport));
4427 #ifndef __BIG_ENDIAN_BITFIELD
4428                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4429 #endif
4430                 break;
4431
4432         case offsetof(struct bpf_sock_ops, local_port):
4433                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4434
4435                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
4436                                                 struct bpf_sock_ops_kern, sk),
4437                                       si->dst_reg, si->src_reg,
4438                                       offsetof(struct bpf_sock_ops_kern, sk));
4439                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4440                                       offsetof(struct sock_common, skc_num));
4441                 break;
4442         }
4443         return insn - insn_buf;
4444 }
4445
4446 static u32 sk_skb_convert_ctx_access(enum bpf_access_type type,
4447                                      const struct bpf_insn *si,
4448                                      struct bpf_insn *insn_buf,
4449                                      struct bpf_prog *prog, u32 *target_size)
4450 {
4451         struct bpf_insn *insn = insn_buf;
4452         int off;
4453
4454         switch (si->off) {
4455         case offsetof(struct __sk_buff, data_end):
4456                 off  = si->off;
4457                 off -= offsetof(struct __sk_buff, data_end);
4458                 off += offsetof(struct sk_buff, cb);
4459                 off += offsetof(struct tcp_skb_cb, bpf.data_end);
4460                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4461                                       si->src_reg, off);
4462                 break;
4463         default:
4464                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4465                                               target_size);
4466         }
4467
4468         return insn - insn_buf;
4469 }
4470
4471 const struct bpf_verifier_ops sk_filter_verifier_ops = {
4472         .get_func_proto         = sk_filter_func_proto,
4473         .is_valid_access        = sk_filter_is_valid_access,
4474         .convert_ctx_access     = bpf_convert_ctx_access,
4475 };
4476
4477 const struct bpf_prog_ops sk_filter_prog_ops = {
4478 };
4479
4480 const struct bpf_verifier_ops tc_cls_act_verifier_ops = {
4481         .get_func_proto         = tc_cls_act_func_proto,
4482         .is_valid_access        = tc_cls_act_is_valid_access,
4483         .convert_ctx_access     = tc_cls_act_convert_ctx_access,
4484         .gen_prologue           = tc_cls_act_prologue,
4485 };
4486
4487 const struct bpf_prog_ops tc_cls_act_prog_ops = {
4488         .test_run               = bpf_prog_test_run_skb,
4489 };
4490
4491 const struct bpf_verifier_ops xdp_verifier_ops = {
4492         .get_func_proto         = xdp_func_proto,
4493         .is_valid_access        = xdp_is_valid_access,
4494         .convert_ctx_access     = xdp_convert_ctx_access,
4495 };
4496
4497 const struct bpf_prog_ops xdp_prog_ops = {
4498         .test_run               = bpf_prog_test_run_xdp,
4499 };
4500
4501 const struct bpf_verifier_ops cg_skb_verifier_ops = {
4502         .get_func_proto         = sk_filter_func_proto,
4503         .is_valid_access        = sk_filter_is_valid_access,
4504         .convert_ctx_access     = bpf_convert_ctx_access,
4505 };
4506
4507 const struct bpf_prog_ops cg_skb_prog_ops = {
4508         .test_run               = bpf_prog_test_run_skb,
4509 };
4510
4511 const struct bpf_verifier_ops lwt_inout_verifier_ops = {
4512         .get_func_proto         = lwt_inout_func_proto,
4513         .is_valid_access        = lwt_is_valid_access,
4514         .convert_ctx_access     = bpf_convert_ctx_access,
4515 };
4516
4517 const struct bpf_prog_ops lwt_inout_prog_ops = {
4518         .test_run               = bpf_prog_test_run_skb,
4519 };
4520
4521 const struct bpf_verifier_ops lwt_xmit_verifier_ops = {
4522         .get_func_proto         = lwt_xmit_func_proto,
4523         .is_valid_access        = lwt_is_valid_access,
4524         .convert_ctx_access     = bpf_convert_ctx_access,
4525         .gen_prologue           = tc_cls_act_prologue,
4526 };
4527
4528 const struct bpf_prog_ops lwt_xmit_prog_ops = {
4529         .test_run               = bpf_prog_test_run_skb,
4530 };
4531
4532 const struct bpf_verifier_ops cg_sock_verifier_ops = {
4533         .get_func_proto         = sock_filter_func_proto,
4534         .is_valid_access        = sock_filter_is_valid_access,
4535         .convert_ctx_access     = sock_filter_convert_ctx_access,
4536 };
4537
4538 const struct bpf_prog_ops cg_sock_prog_ops = {
4539 };
4540
4541 const struct bpf_verifier_ops sock_ops_verifier_ops = {
4542         .get_func_proto         = sock_ops_func_proto,
4543         .is_valid_access        = sock_ops_is_valid_access,
4544         .convert_ctx_access     = sock_ops_convert_ctx_access,
4545 };
4546
4547 const struct bpf_prog_ops sock_ops_prog_ops = {
4548 };
4549
4550 const struct bpf_verifier_ops sk_skb_verifier_ops = {
4551         .get_func_proto         = sk_skb_func_proto,
4552         .is_valid_access        = sk_skb_is_valid_access,
4553         .convert_ctx_access     = sk_skb_convert_ctx_access,
4554         .gen_prologue           = sk_skb_prologue,
4555 };
4556
4557 const struct bpf_prog_ops sk_skb_prog_ops = {
4558 };
4559
4560 int sk_detach_filter(struct sock *sk)
4561 {
4562         int ret = -ENOENT;
4563         struct sk_filter *filter;
4564
4565         if (sock_flag(sk, SOCK_FILTER_LOCKED))
4566                 return -EPERM;
4567
4568         filter = rcu_dereference_protected(sk->sk_filter,
4569                                            lockdep_sock_is_held(sk));
4570         if (filter) {
4571                 RCU_INIT_POINTER(sk->sk_filter, NULL);
4572                 sk_filter_uncharge(sk, filter);
4573                 ret = 0;
4574         }
4575
4576         return ret;
4577 }
4578 EXPORT_SYMBOL_GPL(sk_detach_filter);
4579
4580 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
4581                   unsigned int len)
4582 {
4583         struct sock_fprog_kern *fprog;
4584         struct sk_filter *filter;
4585         int ret = 0;
4586
4587         lock_sock(sk);
4588         filter = rcu_dereference_protected(sk->sk_filter,
4589                                            lockdep_sock_is_held(sk));
4590         if (!filter)
4591                 goto out;
4592
4593         /* We're copying the filter that has been originally attached,
4594          * so no conversion/decode needed anymore. eBPF programs that
4595          * have no original program cannot be dumped through this.
4596          */
4597         ret = -EACCES;
4598         fprog = filter->prog->orig_prog;
4599         if (!fprog)
4600                 goto out;
4601
4602         ret = fprog->len;
4603         if (!len)
4604                 /* User space only enquires number of filter blocks. */
4605                 goto out;
4606
4607         ret = -EINVAL;
4608         if (len < fprog->len)
4609                 goto out;
4610
4611         ret = -EFAULT;
4612         if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
4613                 goto out;
4614
4615         /* Instead of bytes, the API requests to return the number
4616          * of filter blocks.
4617          */
4618         ret = fprog->len;
4619 out:
4620         release_sock(sk);
4621         return ret;
4622 }