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/inet_common.h>
  37 #include <net/ip.h>
  38 #include <net/protocol.h>
  39 #include <net/netlink.h>
  40 #include <linux/skbuff.h>
  41 #include <net/sock.h>
  42 #include <net/flow_dissector.h>
  43 #include <linux/errno.h>
  44 #include <linux/timer.h>
  45 #include <linux/uaccess.h>
  46 #include <asm/unaligned.h>
  47 #include <asm/cmpxchg.h>
  48 #include <linux/filter.h>
  49 #include <linux/ratelimit.h>
  50 #include <linux/seccomp.h>
  51 #include <linux/if_vlan.h>
  52 #include <linux/bpf.h>
  53 #include <net/sch_generic.h>
  54 #include <net/cls_cgroup.h>
  55 #include <net/dst_metadata.h>
  56 #include <net/dst.h>
  57 #include <net/sock_reuseport.h>
  58 #include <net/busy_poll.h>
  59 #include <net/tcp.h>
  60 #include <linux/bpf_trace.h>
  61
  62 /**
  63  *      sk_filter_trim_cap - run a packet through a socket filter
  64  *      @sk: sock associated with &sk_buff
  65  *      @skb: buffer to filter
  66  *      @cap: limit on how short the eBPF program may trim the packet
  67  *
  68  * Run the eBPF program and then cut skb->data to correct size returned by
  69  * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
  70  * than pkt_len we keep whole skb->data. This is the socket level
  71  * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
  72  * be accepted or -EPERM if the packet should be tossed.
  73  *
  74  */
  75 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
  76 {
  77         int err;
  78         struct sk_filter *filter;
  79
  80         /*
  81          * If the skb was allocated from pfmemalloc reserves, only
  82          * allow SOCK_MEMALLOC sockets to use it as this socket is
  83          * helping free memory
  84          */
  85         if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC)) {
  86                 NET_INC_STATS(sock_net(sk), LINUX_MIB_PFMEMALLOCDROP);
  87                 return -ENOMEM;
  88         }
  89         err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
  90         if (err)
  91                 return err;
  92
  93         err = security_sock_rcv_skb(sk, skb);
  94         if (err)
  95                 return err;
  96
  97         rcu_read_lock();
  98         filter = rcu_dereference(sk->sk_filter);
  99         if (filter) {
 100                 struct sock *save_sk = skb->sk;
 101                 unsigned int pkt_len;
 102
 103                 skb->sk = sk;
 104                 pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
 105                 skb->sk = save_sk;
 106                 err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
 107         }
 108         rcu_read_unlock();
 109
 110         return err;
 111 }
 112 EXPORT_SYMBOL(sk_filter_trim_cap);
 113
 114 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
 115 {
 116         return skb_get_poff(skb);
 117 }
 118
 119 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
 120 {
 121         struct nlattr *nla;
 122
 123         if (skb_is_nonlinear(skb))
 124                 return 0;
 125
 126         if (skb->len < sizeof(struct nlattr))
 127                 return 0;
 128
 129         if (a > skb->len - sizeof(struct nlattr))
 130                 return 0;
 131
 132         nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
 133         if (nla)
 134                 return (void *) nla - (void *) skb->data;
 135
 136         return 0;
 137 }
 138
 139 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
 140 {
 141         struct nlattr *nla;
 142
 143         if (skb_is_nonlinear(skb))
 144                 return 0;
 145
 146         if (skb->len < sizeof(struct nlattr))
 147                 return 0;
 148
 149         if (a > skb->len - sizeof(struct nlattr))
 150                 return 0;
 151
 152         nla = (struct nlattr *) &skb->data[a];
 153         if (nla->nla_len > skb->len - a)
 154                 return 0;
 155
 156         nla = nla_find_nested(nla, x);
 157         if (nla)
 158                 return (void *) nla - (void *) skb->data;
 159
 160         return 0;
 161 }
 162
 163 BPF_CALL_0(__get_raw_cpu_id)
 164 {
 165         return raw_smp_processor_id();
 166 }
 167
 168 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
 169         .func           = __get_raw_cpu_id,
 170         .gpl_only       = false,
 171         .ret_type       = RET_INTEGER,
 172 };
 173
 174 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
 175                               struct bpf_insn *insn_buf)
 176 {
 177         struct bpf_insn *insn = insn_buf;
 178
 179         switch (skb_field) {
 180         case SKF_AD_MARK:
 181                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
 182
 183                 *insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
 184                                       offsetof(struct sk_buff, mark));
 185                 break;
 186
 187         case SKF_AD_PKTTYPE:
 188                 *insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
 189                 *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
 190 #ifdef __BIG_ENDIAN_BITFIELD
 191                 *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
 192 #endif
 193                 break;
 194
 195         case SKF_AD_QUEUE:
 196                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
 197
 198                 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
 199                                       offsetof(struct sk_buff, queue_mapping));
 200                 break;
 201
 202         case SKF_AD_VLAN_TAG:
 203         case SKF_AD_VLAN_TAG_PRESENT:
 204                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
 205                 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
 206
 207                 /* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
 208                 *insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
 209                                       offsetof(struct sk_buff, vlan_tci));
 210                 if (skb_field == SKF_AD_VLAN_TAG) {
 211                         *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
 212                                                 ~VLAN_TAG_PRESENT);
 213                 } else {
 214                         /* dst_reg >>= 12 */
 215                         *insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
 216                         /* dst_reg &= 1 */
 217                         *insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
 218                 }
 219                 break;
 220         }
 221
 222         return insn - insn_buf;
 223 }
 224
 225 static bool convert_bpf_extensions(struct sock_filter *fp,
 226                                    struct bpf_insn **insnp)
 227 {
 228         struct bpf_insn *insn = *insnp;
 229         u32 cnt;
 230
 231         switch (fp->k) {
 232         case SKF_AD_OFF + SKF_AD_PROTOCOL:
 233                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
 234
 235                 /* A = *(u16 *) (CTX + offsetof(protocol)) */
 236                 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
 237                                       offsetof(struct sk_buff, protocol));
 238                 /* A = ntohs(A) [emitting a nop or swap16] */
 239                 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
 240                 break;
 241
 242         case SKF_AD_OFF + SKF_AD_PKTTYPE:
 243                 cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
 244                 insn += cnt - 1;
 245                 break;
 246
 247         case SKF_AD_OFF + SKF_AD_IFINDEX:
 248         case SKF_AD_OFF + SKF_AD_HATYPE:
 249                 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
 250                 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
 251
 252                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
 253                                       BPF_REG_TMP, BPF_REG_CTX,
 254                                       offsetof(struct sk_buff, dev));
 255                 /* if (tmp != 0) goto pc + 1 */
 256                 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
 257                 *insn++ = BPF_EXIT_INSN();
 258                 if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
 259                         *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
 260                                             offsetof(struct net_device, ifindex));
 261                 else
 262                         *insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
 263                                             offsetof(struct net_device, type));
 264                 break;
 265
 266         case SKF_AD_OFF + SKF_AD_MARK:
 267                 cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
 268                 insn += cnt - 1;
 269                 break;
 270
 271         case SKF_AD_OFF + SKF_AD_RXHASH:
 272                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
 273
 274                 *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
 275                                     offsetof(struct sk_buff, hash));
 276                 break;
 277
 278         case SKF_AD_OFF + SKF_AD_QUEUE:
 279                 cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
 280                 insn += cnt - 1;
 281                 break;
 282
 283         case SKF_AD_OFF + SKF_AD_VLAN_TAG:
 284                 cnt = convert_skb_access(SKF_AD_VLAN_TAG,
 285                                          BPF_REG_A, BPF_REG_CTX, insn);
 286                 insn += cnt - 1;
 287                 break;
 288
 289         case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
 290                 cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
 291                                          BPF_REG_A, BPF_REG_CTX, insn);
 292                 insn += cnt - 1;
 293                 break;
 294
 295         case SKF_AD_OFF + SKF_AD_VLAN_TPID:
 296                 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
 297
 298                 /* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
 299                 *insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
 300                                       offsetof(struct sk_buff, vlan_proto));
 301                 /* A = ntohs(A) [emitting a nop or swap16] */
 302                 *insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
 303                 break;
 304
 305         case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
 306         case SKF_AD_OFF + SKF_AD_NLATTR:
 307         case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
 308         case SKF_AD_OFF + SKF_AD_CPU:
 309         case SKF_AD_OFF + SKF_AD_RANDOM:
 310                 /* arg1 = CTX */
 311                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
 312                 /* arg2 = A */
 313                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
 314                 /* arg3 = X */
 315                 *insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
 316                 /* Emit call(arg1=CTX, arg2=A, arg3=X) */
 317                 switch (fp->k) {
 318                 case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
 319                         *insn = BPF_EMIT_CALL(__skb_get_pay_offset);
 320                         break;
 321                 case SKF_AD_OFF + SKF_AD_NLATTR:
 322                         *insn = BPF_EMIT_CALL(__skb_get_nlattr);
 323                         break;
 324                 case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
 325                         *insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
 326                         break;
 327                 case SKF_AD_OFF + SKF_AD_CPU:
 328                         *insn = BPF_EMIT_CALL(__get_raw_cpu_id);
 329                         break;
 330                 case SKF_AD_OFF + SKF_AD_RANDOM:
 331                         *insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
 332                         bpf_user_rnd_init_once();
 333                         break;
 334                 }
 335                 break;
 336
 337         case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
 338                 /* A ^= X */
 339                 *insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
 340                 break;
 341
 342         default:
 343                 /* This is just a dummy call to avoid letting the compiler
 344                  * evict __bpf_call_base() as an optimization. Placed here
 345                  * where no-one bothers.
 346                  */
 347                 BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
 348                 return false;
 349         }
 350
 351         *insnp = insn;
 352         return true;
 353 }
 354
 355 /**
 356  *      bpf_convert_filter - convert filter program
 357  *      @prog: the user passed filter program
 358  *      @len: the length of the user passed filter program
 359  *      @new_prog: allocated 'struct bpf_prog' or NULL
 360  *      @new_len: pointer to store length of converted program
 361  *
 362  * Remap 'sock_filter' style classic BPF (cBPF) instruction set to 'bpf_insn'
 363  * style extended BPF (eBPF).
 364  * Conversion workflow:
 365  *
 366  * 1) First pass for calculating the new program length:
 367  *   bpf_convert_filter(old_prog, old_len, NULL, &new_len)
 368  *
 369  * 2) 2nd pass to remap in two passes: 1st pass finds new
 370  *    jump offsets, 2nd pass remapping:
 371  *   bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
 372  */
 373 static int bpf_convert_filter(struct sock_filter *prog, int len,
 374                               struct bpf_prog *new_prog, int *new_len)
 375 {
 376         int new_flen = 0, pass = 0, target, i, stack_off;
 377         struct bpf_insn *new_insn, *first_insn = NULL;
 378         struct sock_filter *fp;
 379         int *addrs = NULL;
 380         u8 bpf_src;
 381
 382         BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
 383         BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
 384
 385         if (len <= 0 || len > BPF_MAXINSNS)
 386                 return -EINVAL;
 387
 388         if (new_prog) {
 389                 first_insn = new_prog->insnsi;
 390                 addrs = kcalloc(len, sizeof(*addrs),
 391                                 GFP_KERNEL | __GFP_NOWARN);
 392                 if (!addrs)
 393                         return -ENOMEM;
 394         }
 395
 396 do_pass:
 397         new_insn = first_insn;
 398         fp = prog;
 399
 400         /* Classic BPF related prologue emission. */
 401         if (new_prog) {
 402                 /* Classic BPF expects A and X to be reset first. These need
 403                  * to be guaranteed to be the first two instructions.
 404                  */
 405                 *new_insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
 406                 *new_insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
 407
 408                 /* All programs must keep CTX in callee saved BPF_REG_CTX.
 409                  * In eBPF case it's done by the compiler, here we need to
 410                  * do this ourself. Initial CTX is present in BPF_REG_ARG1.
 411                  */
 412                 *new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
 413         } else {
 414                 new_insn += 3;
 415         }
 416
 417         for (i = 0; i < len; fp++, i++) {
 418                 struct bpf_insn tmp_insns[6] = { };
 419                 struct bpf_insn *insn = tmp_insns;
 420
 421                 if (addrs)
 422                         addrs[i] = new_insn - first_insn;
 423
 424                 switch (fp->code) {
 425                 /* All arithmetic insns and skb loads map as-is. */
 426                 case BPF_ALU | BPF_ADD | BPF_X:
 427                 case BPF_ALU | BPF_ADD | BPF_K:
 428                 case BPF_ALU | BPF_SUB | BPF_X:
 429                 case BPF_ALU | BPF_SUB | BPF_K:
 430                 case BPF_ALU | BPF_AND | BPF_X:
 431                 case BPF_ALU | BPF_AND | BPF_K:
 432                 case BPF_ALU | BPF_OR | BPF_X:
 433                 case BPF_ALU | BPF_OR | BPF_K:
 434                 case BPF_ALU | BPF_LSH | BPF_X:
 435                 case BPF_ALU | BPF_LSH | BPF_K:
 436                 case BPF_ALU | BPF_RSH | BPF_X:
 437                 case BPF_ALU | BPF_RSH | BPF_K:
 438                 case BPF_ALU | BPF_XOR | BPF_X:
 439                 case BPF_ALU | BPF_XOR | BPF_K:
 440                 case BPF_ALU | BPF_MUL | BPF_X:
 441                 case BPF_ALU | BPF_MUL | BPF_K:
 442                 case BPF_ALU | BPF_DIV | BPF_X:
 443                 case BPF_ALU | BPF_DIV | BPF_K:
 444                 case BPF_ALU | BPF_MOD | BPF_X:
 445                 case BPF_ALU | BPF_MOD | BPF_K:
 446                 case BPF_ALU | BPF_NEG:
 447                 case BPF_LD | BPF_ABS | BPF_W:
 448                 case BPF_LD | BPF_ABS | BPF_H:
 449                 case BPF_LD | BPF_ABS | BPF_B:
 450                 case BPF_LD | BPF_IND | BPF_W:
 451                 case BPF_LD | BPF_IND | BPF_H:
 452                 case BPF_LD | BPF_IND | BPF_B:
 453                         /* Check for overloaded BPF extension and
 454                          * directly convert it if found, otherwise
 455                          * just move on with mapping.
 456                          */
 457                         if (BPF_CLASS(fp->code) == BPF_LD &&
 458                             BPF_MODE(fp->code) == BPF_ABS &&
 459                             convert_bpf_extensions(fp, &insn))
 460                                 break;
 461
 462                         if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) ||
 463                             fp->code == (BPF_ALU | BPF_MOD | BPF_X)) {
 464                                 *insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X);
 465                                 /* Error with exception code on div/mod by 0.
 466                                  * For cBPF programs, this was always return 0.
 467                                  */
 468                                 *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_X, 0, 2);
 469                                 *insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
 470                                 *insn++ = BPF_EXIT_INSN();
 471                         }
 472
 473                         *insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
 474                         break;
 475
 476                 /* Jump transformation cannot use BPF block macros
 477                  * everywhere as offset calculation and target updates
 478                  * require a bit more work than the rest, i.e. jump
 479                  * opcodes map as-is, but offsets need adjustment.
 480                  */
 481
 482 #define BPF_EMIT_JMP                                                    \
 483         do {                                                            \
 484                 if (target >= len || target < 0)                        \
 485                         goto err;                                       \
 486                 insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0;   \
 487                 /* Adjust pc relative offset for 2nd or 3rd insn. */    \
 488                 insn->off -= insn - tmp_insns;                          \
 489         } while (0)
 490
 491                 case BPF_JMP | BPF_JA:
 492                         target = i + fp->k + 1;
 493                         insn->code = fp->code;
 494                         BPF_EMIT_JMP;
 495                         break;
 496
 497                 case BPF_JMP | BPF_JEQ | BPF_K:
 498                 case BPF_JMP | BPF_JEQ | BPF_X:
 499                 case BPF_JMP | BPF_JSET | BPF_K:
 500                 case BPF_JMP | BPF_JSET | BPF_X:
 501                 case BPF_JMP | BPF_JGT | BPF_K:
 502                 case BPF_JMP | BPF_JGT | BPF_X:
 503                 case BPF_JMP | BPF_JGE | BPF_K:
 504                 case BPF_JMP | BPF_JGE | BPF_X:
 505                         if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
 506                                 /* BPF immediates are signed, zero extend
 507                                  * immediate into tmp register and use it
 508                                  * in compare insn.
 509                                  */
 510                                 *insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
 511
 512                                 insn->dst_reg = BPF_REG_A;
 513                                 insn->src_reg = BPF_REG_TMP;
 514                                 bpf_src = BPF_X;
 515                         } else {
 516                                 insn->dst_reg = BPF_REG_A;
 517                                 insn->imm = fp->k;
 518                                 bpf_src = BPF_SRC(fp->code);
 519                                 insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
 520                         }
 521
 522                         /* Common case where 'jump_false' is next insn. */
 523                         if (fp->jf == 0) {
 524                                 insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 525                                 target = i + fp->jt + 1;
 526                                 BPF_EMIT_JMP;
 527                                 break;
 528                         }
 529
 530                         /* Convert some jumps when 'jump_true' is next insn. */
 531                         if (fp->jt == 0) {
 532                                 switch (BPF_OP(fp->code)) {
 533                                 case BPF_JEQ:
 534                                         insn->code = BPF_JMP | BPF_JNE | bpf_src;
 535                                         break;
 536                                 case BPF_JGT:
 537                                         insn->code = BPF_JMP | BPF_JLE | bpf_src;
 538                                         break;
 539                                 case BPF_JGE:
 540                                         insn->code = BPF_JMP | BPF_JLT | bpf_src;
 541                                         break;
 542                                 default:
 543                                         goto jmp_rest;
 544                                 }
 545
 546                                 target = i + fp->jf + 1;
 547                                 BPF_EMIT_JMP;
 548                                 break;
 549                         }
 550 jmp_rest:
 551                         /* Other jumps are mapped into two insns: Jxx and JA. */
 552                         target = i + fp->jt + 1;
 553                         insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
 554                         BPF_EMIT_JMP;
 555                         insn++;
 556
 557                         insn->code = BPF_JMP | BPF_JA;
 558                         target = i + fp->jf + 1;
 559                         BPF_EMIT_JMP;
 560                         break;
 561
 562                 /* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
 563                 case BPF_LDX | BPF_MSH | BPF_B:
 564                         /* tmp = A */
 565                         *insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
 566                         /* A = BPF_R0 = *(u8 *) (skb->data + K) */
 567                         *insn++ = BPF_LD_ABS(BPF_B, fp->k);
 568                         /* A &= 0xf */
 569                         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
 570                         /* A <<= 2 */
 571                         *insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
 572                         /* X = A */
 573                         *insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 574                         /* A = tmp */
 575                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
 576                         break;
 577
 578                 /* RET_K is remaped into 2 insns. RET_A case doesn't need an
 579                  * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
 580                  */
 581                 case BPF_RET | BPF_A:
 582                 case BPF_RET | BPF_K:
 583                         if (BPF_RVAL(fp->code) == BPF_K)
 584                                 *insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
 585                                                         0, fp->k);
 586                         *insn = BPF_EXIT_INSN();
 587                         break;
 588
 589                 /* Store to stack. */
 590                 case BPF_ST:
 591                 case BPF_STX:
 592                         stack_off = fp->k * 4  + 4;
 593                         *insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
 594                                             BPF_ST ? BPF_REG_A : BPF_REG_X,
 595                                             -stack_off);
 596                         /* check_load_and_stores() verifies that classic BPF can
 597                          * load from stack only after write, so tracking
 598                          * stack_depth for ST|STX insns is enough
 599                          */
 600                         if (new_prog && new_prog->aux->stack_depth < stack_off)
 601                                 new_prog->aux->stack_depth = stack_off;
 602                         break;
 603
 604                 /* Load from stack. */
 605                 case BPF_LD | BPF_MEM:
 606                 case BPF_LDX | BPF_MEM:
 607                         stack_off = fp->k * 4  + 4;
 608                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD  ?
 609                                             BPF_REG_A : BPF_REG_X, BPF_REG_FP,
 610                                             -stack_off);
 611                         break;
 612
 613                 /* A = K or X = K */
 614                 case BPF_LD | BPF_IMM:
 615                 case BPF_LDX | BPF_IMM:
 616                         *insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
 617                                               BPF_REG_A : BPF_REG_X, fp->k);
 618                         break;
 619
 620                 /* X = A */
 621                 case BPF_MISC | BPF_TAX:
 622                         *insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
 623                         break;
 624
 625                 /* A = X */
 626                 case BPF_MISC | BPF_TXA:
 627                         *insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
 628                         break;
 629
 630                 /* A = skb->len or X = skb->len */
 631                 case BPF_LD | BPF_W | BPF_LEN:
 632                 case BPF_LDX | BPF_W | BPF_LEN:
 633                         *insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
 634                                             BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
 635                                             offsetof(struct sk_buff, len));
 636                         break;
 637
 638                 /* Access seccomp_data fields. */
 639                 case BPF_LDX | BPF_ABS | BPF_W:
 640                         /* A = *(u32 *) (ctx + K) */
 641                         *insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
 642                         break;
 643
 644                 /* Unknown instruction. */
 645                 default:
 646                         goto err;
 647                 }
 648
 649                 insn++;
 650                 if (new_prog)
 651                         memcpy(new_insn, tmp_insns,
 652                                sizeof(*insn) * (insn - tmp_insns));
 653                 new_insn += insn - tmp_insns;
 654         }
 655
 656         if (!new_prog) {
 657                 /* Only calculating new length. */
 658                 *new_len = new_insn - first_insn;
 659                 return 0;
 660         }
 661
 662         pass++;
 663         if (new_flen != new_insn - first_insn) {
 664                 new_flen = new_insn - first_insn;
 665                 if (pass > 2)
 666                         goto err;
 667                 goto do_pass;
 668         }
 669
 670         kfree(addrs);
 671         BUG_ON(*new_len != new_flen);
 672         return 0;
 673 err:
 674         kfree(addrs);
 675         return -EINVAL;
 676 }
 677
 678 /* Security:
 679  *
 680  * As we dont want to clear mem[] array for each packet going through
 681  * __bpf_prog_run(), we check that filter loaded by user never try to read
 682  * a cell if not previously written, and we check all branches to be sure
 683  * a malicious user doesn't try to abuse us.
 684  */
 685 static int check_load_and_stores(const struct sock_filter *filter, int flen)
 686 {
 687         u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
 688         int pc, ret = 0;
 689
 690         BUILD_BUG_ON(BPF_MEMWORDS > 16);
 691
 692         masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
 693         if (!masks)
 694                 return -ENOMEM;
 695
 696         memset(masks, 0xff, flen * sizeof(*masks));
 697
 698         for (pc = 0; pc < flen; pc++) {
 699                 memvalid &= masks[pc];
 700
 701                 switch (filter[pc].code) {
 702                 case BPF_ST:
 703                 case BPF_STX:
 704                         memvalid |= (1 << filter[pc].k);
 705                         break;
 706                 case BPF_LD | BPF_MEM:
 707                 case BPF_LDX | BPF_MEM:
 708                         if (!(memvalid & (1 << filter[pc].k))) {
 709                                 ret = -EINVAL;
 710                                 goto error;
 711                         }
 712                         break;
 713                 case BPF_JMP | BPF_JA:
 714                         /* A jump must set masks on target */
 715                         masks[pc + 1 + filter[pc].k] &= memvalid;
 716                         memvalid = ~0;
 717                         break;
 718                 case BPF_JMP | BPF_JEQ | BPF_K:
 719                 case BPF_JMP | BPF_JEQ | BPF_X:
 720                 case BPF_JMP | BPF_JGE | BPF_K:
 721                 case BPF_JMP | BPF_JGE | BPF_X:
 722                 case BPF_JMP | BPF_JGT | BPF_K:
 723                 case BPF_JMP | BPF_JGT | BPF_X:
 724                 case BPF_JMP | BPF_JSET | BPF_K:
 725                 case BPF_JMP | BPF_JSET | BPF_X:
 726                         /* A jump must set masks on targets */
 727                         masks[pc + 1 + filter[pc].jt] &= memvalid;
 728                         masks[pc + 1 + filter[pc].jf] &= memvalid;
 729                         memvalid = ~0;
 730                         break;
 731                 }
 732         }
 733 error:
 734         kfree(masks);
 735         return ret;
 736 }
 737
 738 static bool chk_code_allowed(u16 code_to_probe)
 739 {
 740         static const bool codes[] = {
 741                 /* 32 bit ALU operations */
 742                 [BPF_ALU | BPF_ADD | BPF_K] = true,
 743                 [BPF_ALU | BPF_ADD | BPF_X] = true,
 744                 [BPF_ALU | BPF_SUB | BPF_K] = true,
 745                 [BPF_ALU | BPF_SUB | BPF_X] = true,
 746                 [BPF_ALU | BPF_MUL | BPF_K] = true,
 747                 [BPF_ALU | BPF_MUL | BPF_X] = true,
 748                 [BPF_ALU | BPF_DIV | BPF_K] = true,
 749                 [BPF_ALU | BPF_DIV | BPF_X] = true,
 750                 [BPF_ALU | BPF_MOD | BPF_K] = true,
 751                 [BPF_ALU | BPF_MOD | BPF_X] = true,
 752                 [BPF_ALU | BPF_AND | BPF_K] = true,
 753                 [BPF_ALU | BPF_AND | BPF_X] = true,
 754                 [BPF_ALU | BPF_OR | BPF_K] = true,
 755                 [BPF_ALU | BPF_OR | BPF_X] = true,
 756                 [BPF_ALU | BPF_XOR | BPF_K] = true,
 757                 [BPF_ALU | BPF_XOR | BPF_X] = true,
 758                 [BPF_ALU | BPF_LSH | BPF_K] = true,
 759                 [BPF_ALU | BPF_LSH | BPF_X] = true,
 760                 [BPF_ALU | BPF_RSH | BPF_K] = true,
 761                 [BPF_ALU | BPF_RSH | BPF_X] = true,
 762                 [BPF_ALU | BPF_NEG] = true,
 763                 /* Load instructions */
 764                 [BPF_LD | BPF_W | BPF_ABS] = true,
 765                 [BPF_LD | BPF_H | BPF_ABS] = true,
 766                 [BPF_LD | BPF_B | BPF_ABS] = true,
 767                 [BPF_LD | BPF_W | BPF_LEN] = true,
 768                 [BPF_LD | BPF_W | BPF_IND] = true,
 769                 [BPF_LD | BPF_H | BPF_IND] = true,
 770                 [BPF_LD | BPF_B | BPF_IND] = true,
 771                 [BPF_LD | BPF_IMM] = true,
 772                 [BPF_LD | BPF_MEM] = true,
 773                 [BPF_LDX | BPF_W | BPF_LEN] = true,
 774                 [BPF_LDX | BPF_B | BPF_MSH] = true,
 775                 [BPF_LDX | BPF_IMM] = true,
 776                 [BPF_LDX | BPF_MEM] = true,
 777                 /* Store instructions */
 778                 [BPF_ST] = true,
 779                 [BPF_STX] = true,
 780                 /* Misc instructions */
 781                 [BPF_MISC | BPF_TAX] = true,
 782                 [BPF_MISC | BPF_TXA] = true,
 783                 /* Return instructions */
 784                 [BPF_RET | BPF_K] = true,
 785                 [BPF_RET | BPF_A] = true,
 786                 /* Jump instructions */
 787                 [BPF_JMP | BPF_JA] = true,
 788                 [BPF_JMP | BPF_JEQ | BPF_K] = true,
 789                 [BPF_JMP | BPF_JEQ | BPF_X] = true,
 790                 [BPF_JMP | BPF_JGE | BPF_K] = true,
 791                 [BPF_JMP | BPF_JGE | BPF_X] = true,
 792                 [BPF_JMP | BPF_JGT | BPF_K] = true,
 793                 [BPF_JMP | BPF_JGT | BPF_X] = true,
 794                 [BPF_JMP | BPF_JSET | BPF_K] = true,
 795                 [BPF_JMP | BPF_JSET | BPF_X] = true,
 796         };
 797
 798         if (code_to_probe >= ARRAY_SIZE(codes))
 799                 return false;
 800
 801         return codes[code_to_probe];
 802 }
 803
 804 static bool bpf_check_basics_ok(const struct sock_filter *filter,
 805                                 unsigned int flen)
 806 {
 807         if (filter == NULL)
 808                 return false;
 809         if (flen == 0 || flen > BPF_MAXINSNS)
 810                 return false;
 811
 812         return true;
 813 }
 814
 815 /**
 816  *      bpf_check_classic - verify socket filter code
 817  *      @filter: filter to verify
 818  *      @flen: length of filter
 819  *
 820  * Check the user's filter code. If we let some ugly
 821  * filter code slip through kaboom! The filter must contain
 822  * no references or jumps that are out of range, no illegal
 823  * instructions, and must end with a RET instruction.
 824  *
 825  * All jumps are forward as they are not signed.
 826  *
 827  * Returns 0 if the rule set is legal or -EINVAL if not.
 828  */
 829 static int bpf_check_classic(const struct sock_filter *filter,
 830                              unsigned int flen)
 831 {
 832         bool anc_found;
 833         int pc;
 834
 835         /* Check the filter code now */
 836         for (pc = 0; pc < flen; pc++) {
 837                 const struct sock_filter *ftest = &filter[pc];
 838
 839                 /* May we actually operate on this code? */
 840                 if (!chk_code_allowed(ftest->code))
 841                         return -EINVAL;
 842
 843                 /* Some instructions need special checks */
 844                 switch (ftest->code) {
 845                 case BPF_ALU | BPF_DIV | BPF_K:
 846                 case BPF_ALU | BPF_MOD | BPF_K:
 847                         /* Check for division by zero */
 848                         if (ftest->k == 0)
 849                                 return -EINVAL;
 850                         break;
 851                 case BPF_ALU | BPF_LSH | BPF_K:
 852                 case BPF_ALU | BPF_RSH | BPF_K:
 853                         if (ftest->k >= 32)
 854                                 return -EINVAL;
 855                         break;
 856                 case BPF_LD | BPF_MEM:
 857                 case BPF_LDX | BPF_MEM:
 858                 case BPF_ST:
 859                 case BPF_STX:
 860                         /* Check for invalid memory addresses */
 861                         if (ftest->k >= BPF_MEMWORDS)
 862                                 return -EINVAL;
 863                         break;
 864                 case BPF_JMP | BPF_JA:
 865                         /* Note, the large ftest->k might cause loops.
 866                          * Compare this with conditional jumps below,
 867                          * where offsets are limited. --ANK (981016)
 868                          */
 869                         if (ftest->k >= (unsigned int)(flen - pc - 1))
 870                                 return -EINVAL;
 871                         break;
 872                 case BPF_JMP | BPF_JEQ | BPF_K:
 873                 case BPF_JMP | BPF_JEQ | BPF_X:
 874                 case BPF_JMP | BPF_JGE | BPF_K:
 875                 case BPF_JMP | BPF_JGE | BPF_X:
 876                 case BPF_JMP | BPF_JGT | BPF_K:
 877                 case BPF_JMP | BPF_JGT | BPF_X:
 878                 case BPF_JMP | BPF_JSET | BPF_K:
 879                 case BPF_JMP | BPF_JSET | BPF_X:
 880                         /* Both conditionals must be safe */
 881                         if (pc + ftest->jt + 1 >= flen ||
 882                             pc + ftest->jf + 1 >= flen)
 883                                 return -EINVAL;
 884                         break;
 885                 case BPF_LD | BPF_W | BPF_ABS:
 886                 case BPF_LD | BPF_H | BPF_ABS:
 887                 case BPF_LD | BPF_B | BPF_ABS:
 888                         anc_found = false;
 889                         if (bpf_anc_helper(ftest) & BPF_ANC)
 890                                 anc_found = true;
 891                         /* Ancillary operation unknown or unsupported */
 892                         if (anc_found == false && ftest->k >= SKF_AD_OFF)
 893                                 return -EINVAL;
 894                 }
 895         }
 896
 897         /* Last instruction must be a RET code */
 898         switch (filter[flen - 1].code) {
 899         case BPF_RET | BPF_K:
 900         case BPF_RET | BPF_A:
 901                 return check_load_and_stores(filter, flen);
 902         }
 903
 904         return -EINVAL;
 905 }
 906
 907 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
 908                                       const struct sock_fprog *fprog)
 909 {
 910         unsigned int fsize = bpf_classic_proglen(fprog);
 911         struct sock_fprog_kern *fkprog;
 912
 913         fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
 914         if (!fp->orig_prog)
 915                 return -ENOMEM;
 916
 917         fkprog = fp->orig_prog;
 918         fkprog->len = fprog->len;
 919
 920         fkprog->filter = kmemdup(fp->insns, fsize,
 921                                  GFP_KERNEL | __GFP_NOWARN);
 922         if (!fkprog->filter) {
 923                 kfree(fp->orig_prog);
 924                 return -ENOMEM;
 925         }
 926
 927         return 0;
 928 }
 929
 930 static void bpf_release_orig_filter(struct bpf_prog *fp)
 931 {
 932         struct sock_fprog_kern *fprog = fp->orig_prog;
 933
 934         if (fprog) {
 935                 kfree(fprog->filter);
 936                 kfree(fprog);
 937         }
 938 }
 939
 940 static void __bpf_prog_release(struct bpf_prog *prog)
 941 {
 942         if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
 943                 bpf_prog_put(prog);
 944         } else {
 945                 bpf_release_orig_filter(prog);
 946                 bpf_prog_free(prog);
 947         }
 948 }
 949
 950 static void __sk_filter_release(struct sk_filter *fp)
 951 {
 952         __bpf_prog_release(fp->prog);
 953         kfree(fp);
 954 }
 955
 956 /**
 957  *      sk_filter_release_rcu - Release a socket filter by rcu_head
 958  *      @rcu: rcu_head that contains the sk_filter to free
 959  */
 960 static void sk_filter_release_rcu(struct rcu_head *rcu)
 961 {
 962         struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
 963
 964         __sk_filter_release(fp);
 965 }
 966
 967 /**
 968  *      sk_filter_release - release a socket filter
 969  *      @fp: filter to remove
 970  *
 971  *      Remove a filter from a socket and release its resources.
 972  */
 973 static void sk_filter_release(struct sk_filter *fp)
 974 {
 975         if (refcount_dec_and_test(&fp->refcnt))
 976                 call_rcu(&fp->rcu, sk_filter_release_rcu);
 977 }
 978
 979 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
 980 {
 981         u32 filter_size = bpf_prog_size(fp->prog->len);
 982
 983         atomic_sub(filter_size, &sk->sk_omem_alloc);
 984         sk_filter_release(fp);
 985 }
 986
 987 /* try to charge the socket memory if there is space available
 988  * return true on success
 989  */
 990 static bool __sk_filter_charge(struct sock *sk, struct sk_filter *fp)
 991 {
 992         u32 filter_size = bpf_prog_size(fp->prog->len);
 993
 994         /* same check as in sock_kmalloc() */
 995         if (filter_size <= sysctl_optmem_max &&
 996             atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
 997                 atomic_add(filter_size, &sk->sk_omem_alloc);
 998                 return true;
 999         }
1000         return false;
1001 }
1002
1003 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
1004 {
1005         if (!refcount_inc_not_zero(&fp->refcnt))
1006                 return false;
1007
1008         if (!__sk_filter_charge(sk, fp)) {
1009                 sk_filter_release(fp);
1010                 return false;
1011         }
1012         return true;
1013 }
1014
1015 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
1016 {
1017         struct sock_filter *old_prog;
1018         struct bpf_prog *old_fp;
1019         int err, new_len, old_len = fp->len;
1020
1021         /* We are free to overwrite insns et al right here as it
1022          * won't be used at this point in time anymore internally
1023          * after the migration to the internal BPF instruction
1024          * representation.
1025          */
1026         BUILD_BUG_ON(sizeof(struct sock_filter) !=
1027                      sizeof(struct bpf_insn));
1028
1029         /* Conversion cannot happen on overlapping memory areas,
1030          * so we need to keep the user BPF around until the 2nd
1031          * pass. At this time, the user BPF is stored in fp->insns.
1032          */
1033         old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
1034                            GFP_KERNEL | __GFP_NOWARN);
1035         if (!old_prog) {
1036                 err = -ENOMEM;
1037                 goto out_err;
1038         }
1039
1040         /* 1st pass: calculate the new program length. */
1041         err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
1042         if (err)
1043                 goto out_err_free;
1044
1045         /* Expand fp for appending the new filter representation. */
1046         old_fp = fp;
1047         fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1048         if (!fp) {
1049                 /* The old_fp is still around in case we couldn't
1050                  * allocate new memory, so uncharge on that one.
1051                  */
1052                 fp = old_fp;
1053                 err = -ENOMEM;
1054                 goto out_err_free;
1055         }
1056
1057         fp->len = new_len;
1058
1059         /* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1060         err = bpf_convert_filter(old_prog, old_len, fp, &new_len);
1061         if (err)
1062                 /* 2nd bpf_convert_filter() can fail only if it fails
1063                  * to allocate memory, remapping must succeed. Note,
1064                  * that at this time old_fp has already been released
1065                  * by krealloc().
1066                  */
1067                 goto out_err_free;
1068
1069         fp = bpf_prog_select_runtime(fp, &err);
1070         if (err)
1071                 goto out_err_free;
1072
1073         kfree(old_prog);
1074         return fp;
1075
1076 out_err_free:
1077         kfree(old_prog);
1078 out_err:
1079         __bpf_prog_release(fp);
1080         return ERR_PTR(err);
1081 }
1082
1083 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1084                                            bpf_aux_classic_check_t trans)
1085 {
1086         int err;
1087
1088         fp->bpf_func = NULL;
1089         fp->jited = 0;
1090
1091         err = bpf_check_classic(fp->insns, fp->len);
1092         if (err) {
1093                 __bpf_prog_release(fp);
1094                 return ERR_PTR(err);
1095         }
1096
1097         /* There might be additional checks and transformations
1098          * needed on classic filters, f.e. in case of seccomp.
1099          */
1100         if (trans) {
1101                 err = trans(fp->insns, fp->len);
1102                 if (err) {
1103                         __bpf_prog_release(fp);
1104                         return ERR_PTR(err);
1105                 }
1106         }
1107
1108         /* Probe if we can JIT compile the filter and if so, do
1109          * the compilation of the filter.
1110          */
1111         bpf_jit_compile(fp);
1112
1113         /* JIT compiler couldn't process this filter, so do the
1114          * internal BPF translation for the optimized interpreter.
1115          */
1116         if (!fp->jited)
1117                 fp = bpf_migrate_filter(fp);
1118
1119         return fp;
1120 }
1121
1122 /**
1123  *      bpf_prog_create - create an unattached filter
1124  *      @pfp: the unattached filter that is created
1125  *      @fprog: the filter program
1126  *
1127  * Create a filter independent of any socket. We first run some
1128  * sanity checks on it to make sure it does not explode on us later.
1129  * If an error occurs or there is insufficient memory for the filter
1130  * a negative errno code is returned. On success the return is zero.
1131  */
1132 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1133 {
1134         unsigned int fsize = bpf_classic_proglen(fprog);
1135         struct bpf_prog *fp;
1136
1137         /* Make sure new filter is there and in the right amounts. */
1138         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1139                 return -EINVAL;
1140
1141         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1142         if (!fp)
1143                 return -ENOMEM;
1144
1145         memcpy(fp->insns, fprog->filter, fsize);
1146
1147         fp->len = fprog->len;
1148         /* Since unattached filters are not copied back to user
1149          * space through sk_get_filter(), we do not need to hold
1150          * a copy here, and can spare us the work.
1151          */
1152         fp->orig_prog = NULL;
1153
1154         /* bpf_prepare_filter() already takes care of freeing
1155          * memory in case something goes wrong.
1156          */
1157         fp = bpf_prepare_filter(fp, NULL);
1158         if (IS_ERR(fp))
1159                 return PTR_ERR(fp);
1160
1161         *pfp = fp;
1162         return 0;
1163 }
1164 EXPORT_SYMBOL_GPL(bpf_prog_create);
1165
1166 /**
1167  *      bpf_prog_create_from_user - create an unattached filter from user buffer
1168  *      @pfp: the unattached filter that is created
1169  *      @fprog: the filter program
1170  *      @trans: post-classic verifier transformation handler
1171  *      @save_orig: save classic BPF program
1172  *
1173  * This function effectively does the same as bpf_prog_create(), only
1174  * that it builds up its insns buffer from user space provided buffer.
1175  * It also allows for passing a bpf_aux_classic_check_t handler.
1176  */
1177 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1178                               bpf_aux_classic_check_t trans, bool save_orig)
1179 {
1180         unsigned int fsize = bpf_classic_proglen(fprog);
1181         struct bpf_prog *fp;
1182         int err;
1183
1184         /* Make sure new filter is there and in the right amounts. */
1185         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1186                 return -EINVAL;
1187
1188         fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1189         if (!fp)
1190                 return -ENOMEM;
1191
1192         if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1193                 __bpf_prog_free(fp);
1194                 return -EFAULT;
1195         }
1196
1197         fp->len = fprog->len;
1198         fp->orig_prog = NULL;
1199
1200         if (save_orig) {
1201                 err = bpf_prog_store_orig_filter(fp, fprog);
1202                 if (err) {
1203                         __bpf_prog_free(fp);
1204                         return -ENOMEM;
1205                 }
1206         }
1207
1208         /* bpf_prepare_filter() already takes care of freeing
1209          * memory in case something goes wrong.
1210          */
1211         fp = bpf_prepare_filter(fp, trans);
1212         if (IS_ERR(fp))
1213                 return PTR_ERR(fp);
1214
1215         *pfp = fp;
1216         return 0;
1217 }
1218 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1219
1220 void bpf_prog_destroy(struct bpf_prog *fp)
1221 {
1222         __bpf_prog_release(fp);
1223 }
1224 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1225
1226 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1227 {
1228         struct sk_filter *fp, *old_fp;
1229
1230         fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1231         if (!fp)
1232                 return -ENOMEM;
1233
1234         fp->prog = prog;
1235
1236         if (!__sk_filter_charge(sk, fp)) {
1237                 kfree(fp);
1238                 return -ENOMEM;
1239         }
1240         refcount_set(&fp->refcnt, 1);
1241
1242         old_fp = rcu_dereference_protected(sk->sk_filter,
1243                                            lockdep_sock_is_held(sk));
1244         rcu_assign_pointer(sk->sk_filter, fp);
1245
1246         if (old_fp)
1247                 sk_filter_uncharge(sk, old_fp);
1248
1249         return 0;
1250 }
1251
1252 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1253 {
1254         struct bpf_prog *old_prog;
1255         int err;
1256
1257         if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1258                 return -ENOMEM;
1259
1260         if (sk_unhashed(sk) && sk->sk_reuseport) {
1261                 err = reuseport_alloc(sk);
1262                 if (err)
1263                         return err;
1264         } else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1265                 /* The socket wasn't bound with SO_REUSEPORT */
1266                 return -EINVAL;
1267         }
1268
1269         old_prog = reuseport_attach_prog(sk, prog);
1270         if (old_prog)
1271                 bpf_prog_destroy(old_prog);
1272
1273         return 0;
1274 }
1275
1276 static
1277 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1278 {
1279         unsigned int fsize = bpf_classic_proglen(fprog);
1280         struct bpf_prog *prog;
1281         int err;
1282
1283         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1284                 return ERR_PTR(-EPERM);
1285
1286         /* Make sure new filter is there and in the right amounts. */
1287         if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1288                 return ERR_PTR(-EINVAL);
1289
1290         prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1291         if (!prog)
1292                 return ERR_PTR(-ENOMEM);
1293
1294         if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1295                 __bpf_prog_free(prog);
1296                 return ERR_PTR(-EFAULT);
1297         }
1298
1299         prog->len = fprog->len;
1300
1301         err = bpf_prog_store_orig_filter(prog, fprog);
1302         if (err) {
1303                 __bpf_prog_free(prog);
1304                 return ERR_PTR(-ENOMEM);
1305         }
1306
1307         /* bpf_prepare_filter() already takes care of freeing
1308          * memory in case something goes wrong.
1309          */
1310         return bpf_prepare_filter(prog, NULL);
1311 }
1312
1313 /**
1314  *      sk_attach_filter - attach a socket filter
1315  *      @fprog: the filter program
1316  *      @sk: the socket to use
1317  *
1318  * Attach the user's filter code. We first run some sanity checks on
1319  * it to make sure it does not explode on us later. If an error
1320  * occurs or there is insufficient memory for the filter a negative
1321  * errno code is returned. On success the return is zero.
1322  */
1323 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1324 {
1325         struct bpf_prog *prog = __get_filter(fprog, sk);
1326         int err;
1327
1328         if (IS_ERR(prog))
1329                 return PTR_ERR(prog);
1330
1331         err = __sk_attach_prog(prog, sk);
1332         if (err < 0) {
1333                 __bpf_prog_release(prog);
1334                 return err;
1335         }
1336
1337         return 0;
1338 }
1339 EXPORT_SYMBOL_GPL(sk_attach_filter);
1340
1341 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1342 {
1343         struct bpf_prog *prog = __get_filter(fprog, sk);
1344         int err;
1345
1346         if (IS_ERR(prog))
1347                 return PTR_ERR(prog);
1348
1349         err = __reuseport_attach_prog(prog, sk);
1350         if (err < 0) {
1351                 __bpf_prog_release(prog);
1352                 return err;
1353         }
1354
1355         return 0;
1356 }
1357
1358 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1359 {
1360         if (sock_flag(sk, SOCK_FILTER_LOCKED))
1361                 return ERR_PTR(-EPERM);
1362
1363         return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1364 }
1365
1366 int sk_attach_bpf(u32 ufd, struct sock *sk)
1367 {
1368         struct bpf_prog *prog = __get_bpf(ufd, sk);
1369         int err;
1370
1371         if (IS_ERR(prog))
1372                 return PTR_ERR(prog);
1373
1374         err = __sk_attach_prog(prog, sk);
1375         if (err < 0) {
1376                 bpf_prog_put(prog);
1377                 return err;
1378         }
1379
1380         return 0;
1381 }
1382
1383 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1384 {
1385         struct bpf_prog *prog = __get_bpf(ufd, sk);
1386         int err;
1387
1388         if (IS_ERR(prog))
1389                 return PTR_ERR(prog);
1390
1391         err = __reuseport_attach_prog(prog, sk);
1392         if (err < 0) {
1393                 bpf_prog_put(prog);
1394                 return err;
1395         }
1396
1397         return 0;
1398 }
1399
1400 struct bpf_scratchpad {
1401         union {
1402                 __be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1403                 u8     buff[MAX_BPF_STACK];
1404         };
1405 };
1406
1407 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1408
1409 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1410                                           unsigned int write_len)
1411 {
1412         return skb_ensure_writable(skb, write_len);
1413 }
1414
1415 static inline int bpf_try_make_writable(struct sk_buff *skb,
1416                                         unsigned int write_len)
1417 {
1418         int err = __bpf_try_make_writable(skb, write_len);
1419
1420         bpf_compute_data_pointers(skb);
1421         return err;
1422 }
1423
1424 static int bpf_try_make_head_writable(struct sk_buff *skb)
1425 {
1426         return bpf_try_make_writable(skb, skb_headlen(skb));
1427 }
1428
1429 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1430 {
1431         if (skb_at_tc_ingress(skb))
1432                 skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1433 }
1434
1435 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1436 {
1437         if (skb_at_tc_ingress(skb))
1438                 skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1439 }
1440
1441 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1442            const void *, from, u32, len, u64, flags)
1443 {
1444         void *ptr;
1445
1446         if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1447                 return -EINVAL;
1448         if (unlikely(offset > 0xffff))
1449                 return -EFAULT;
1450         if (unlikely(bpf_try_make_writable(skb, offset + len)))
1451                 return -EFAULT;
1452
1453         ptr = skb->data + offset;
1454         if (flags & BPF_F_RECOMPUTE_CSUM)
1455                 __skb_postpull_rcsum(skb, ptr, len, offset);
1456
1457         memcpy(ptr, from, len);
1458
1459         if (flags & BPF_F_RECOMPUTE_CSUM)
1460                 __skb_postpush_rcsum(skb, ptr, len, offset);
1461         if (flags & BPF_F_INVALIDATE_HASH)
1462                 skb_clear_hash(skb);
1463
1464         return 0;
1465 }
1466
1467 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1468         .func           = bpf_skb_store_bytes,
1469         .gpl_only       = false,
1470         .ret_type       = RET_INTEGER,
1471         .arg1_type      = ARG_PTR_TO_CTX,
1472         .arg2_type      = ARG_ANYTHING,
1473         .arg3_type      = ARG_PTR_TO_MEM,
1474         .arg4_type      = ARG_CONST_SIZE,
1475         .arg5_type      = ARG_ANYTHING,
1476 };
1477
1478 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1479            void *, to, u32, len)
1480 {
1481         void *ptr;
1482
1483         if (unlikely(offset > 0xffff))
1484                 goto err_clear;
1485
1486         ptr = skb_header_pointer(skb, offset, len, to);
1487         if (unlikely(!ptr))
1488                 goto err_clear;
1489         if (ptr != to)
1490                 memcpy(to, ptr, len);
1491
1492         return 0;
1493 err_clear:
1494         memset(to, 0, len);
1495         return -EFAULT;
1496 }
1497
1498 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1499         .func           = bpf_skb_load_bytes,
1500         .gpl_only       = false,
1501         .ret_type       = RET_INTEGER,
1502         .arg1_type      = ARG_PTR_TO_CTX,
1503         .arg2_type      = ARG_ANYTHING,
1504         .arg3_type      = ARG_PTR_TO_UNINIT_MEM,
1505         .arg4_type      = ARG_CONST_SIZE,
1506 };
1507
1508 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1509 {
1510         /* Idea is the following: should the needed direct read/write
1511          * test fail during runtime, we can pull in more data and redo
1512          * again, since implicitly, we invalidate previous checks here.
1513          *
1514          * Or, since we know how much we need to make read/writeable,
1515          * this can be done once at the program beginning for direct
1516          * access case. By this we overcome limitations of only current
1517          * headroom being accessible.
1518          */
1519         return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1520 }
1521
1522 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1523         .func           = bpf_skb_pull_data,
1524         .gpl_only       = false,
1525         .ret_type       = RET_INTEGER,
1526         .arg1_type      = ARG_PTR_TO_CTX,
1527         .arg2_type      = ARG_ANYTHING,
1528 };
1529
1530 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1531            u64, from, u64, to, u64, flags)
1532 {
1533         __sum16 *ptr;
1534
1535         if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1536                 return -EINVAL;
1537         if (unlikely(offset > 0xffff || offset & 1))
1538                 return -EFAULT;
1539         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1540                 return -EFAULT;
1541
1542         ptr = (__sum16 *)(skb->data + offset);
1543         switch (flags & BPF_F_HDR_FIELD_MASK) {
1544         case 0:
1545                 if (unlikely(from != 0))
1546                         return -EINVAL;
1547
1548                 csum_replace_by_diff(ptr, to);
1549                 break;
1550         case 2:
1551                 csum_replace2(ptr, from, to);
1552                 break;
1553         case 4:
1554                 csum_replace4(ptr, from, to);
1555                 break;
1556         default:
1557                 return -EINVAL;
1558         }
1559
1560         return 0;
1561 }
1562
1563 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1564         .func           = bpf_l3_csum_replace,
1565         .gpl_only       = false,
1566         .ret_type       = RET_INTEGER,
1567         .arg1_type      = ARG_PTR_TO_CTX,
1568         .arg2_type      = ARG_ANYTHING,
1569         .arg3_type      = ARG_ANYTHING,
1570         .arg4_type      = ARG_ANYTHING,
1571         .arg5_type      = ARG_ANYTHING,
1572 };
1573
1574 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1575            u64, from, u64, to, u64, flags)
1576 {
1577         bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1578         bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1579         bool do_mforce = flags & BPF_F_MARK_ENFORCE;
1580         __sum16 *ptr;
1581
1582         if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_MARK_ENFORCE |
1583                                BPF_F_PSEUDO_HDR | BPF_F_HDR_FIELD_MASK)))
1584                 return -EINVAL;
1585         if (unlikely(offset > 0xffff || offset & 1))
1586                 return -EFAULT;
1587         if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1588                 return -EFAULT;
1589
1590         ptr = (__sum16 *)(skb->data + offset);
1591         if (is_mmzero && !do_mforce && !*ptr)
1592                 return 0;
1593
1594         switch (flags & BPF_F_HDR_FIELD_MASK) {
1595         case 0:
1596                 if (unlikely(from != 0))
1597                         return -EINVAL;
1598
1599                 inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1600                 break;
1601         case 2:
1602                 inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1603                 break;
1604         case 4:
1605                 inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1606                 break;
1607         default:
1608                 return -EINVAL;
1609         }
1610
1611         if (is_mmzero && !*ptr)
1612                 *ptr = CSUM_MANGLED_0;
1613         return 0;
1614 }
1615
1616 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1617         .func           = bpf_l4_csum_replace,
1618         .gpl_only       = false,
1619         .ret_type       = RET_INTEGER,
1620         .arg1_type      = ARG_PTR_TO_CTX,
1621         .arg2_type      = ARG_ANYTHING,
1622         .arg3_type      = ARG_ANYTHING,
1623         .arg4_type      = ARG_ANYTHING,
1624         .arg5_type      = ARG_ANYTHING,
1625 };
1626
1627 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1628            __be32 *, to, u32, to_size, __wsum, seed)
1629 {
1630         struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1631         u32 diff_size = from_size + to_size;
1632         int i, j = 0;
1633
1634         /* This is quite flexible, some examples:
1635          *
1636          * from_size == 0, to_size > 0,  seed := csum --> pushing data
1637          * from_size > 0,  to_size == 0, seed := csum --> pulling data
1638          * from_size > 0,  to_size > 0,  seed := 0    --> diffing data
1639          *
1640          * Even for diffing, from_size and to_size don't need to be equal.
1641          */
1642         if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1643                      diff_size > sizeof(sp->diff)))
1644                 return -EINVAL;
1645
1646         for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1647                 sp->diff[j] = ~from[i];
1648         for (i = 0; i <   to_size / sizeof(__be32); i++, j++)
1649                 sp->diff[j] = to[i];
1650
1651         return csum_partial(sp->diff, diff_size, seed);
1652 }
1653
1654 static const struct bpf_func_proto bpf_csum_diff_proto = {
1655         .func           = bpf_csum_diff,
1656         .gpl_only       = false,
1657         .pkt_access     = true,
1658         .ret_type       = RET_INTEGER,
1659         .arg1_type      = ARG_PTR_TO_MEM_OR_NULL,
1660         .arg2_type      = ARG_CONST_SIZE_OR_ZERO,
1661         .arg3_type      = ARG_PTR_TO_MEM_OR_NULL,
1662         .arg4_type      = ARG_CONST_SIZE_OR_ZERO,
1663         .arg5_type      = ARG_ANYTHING,
1664 };
1665
1666 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1667 {
1668         /* The interface is to be used in combination with bpf_csum_diff()
1669          * for direct packet writes. csum rotation for alignment as well
1670          * as emulating csum_sub() can be done from the eBPF program.
1671          */
1672         if (skb->ip_summed == CHECKSUM_COMPLETE)
1673                 return (skb->csum = csum_add(skb->csum, csum));
1674
1675         return -ENOTSUPP;
1676 }
1677
1678 static const struct bpf_func_proto bpf_csum_update_proto = {
1679         .func           = bpf_csum_update,
1680         .gpl_only       = false,
1681         .ret_type       = RET_INTEGER,
1682         .arg1_type      = ARG_PTR_TO_CTX,
1683         .arg2_type      = ARG_ANYTHING,
1684 };
1685
1686 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1687 {
1688         return dev_forward_skb(dev, skb);
1689 }
1690
1691 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1692                                       struct sk_buff *skb)
1693 {
1694         int ret = ____dev_forward_skb(dev, skb);
1695
1696         if (likely(!ret)) {
1697                 skb->dev = dev;
1698                 ret = netif_rx(skb);
1699         }
1700
1701         return ret;
1702 }
1703
1704 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1705 {
1706         int ret;
1707
1708         if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1709                 net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1710                 kfree_skb(skb);
1711                 return -ENETDOWN;
1712         }
1713
1714         skb->dev = dev;
1715
1716         __this_cpu_inc(xmit_recursion);
1717         ret = dev_queue_xmit(skb);
1718         __this_cpu_dec(xmit_recursion);
1719
1720         return ret;
1721 }
1722
1723 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1724                                  u32 flags)
1725 {
1726         /* skb->mac_len is not set on normal egress */
1727         unsigned int mlen = skb->network_header - skb->mac_header;
1728
1729         __skb_pull(skb, mlen);
1730
1731         /* At ingress, the mac header has already been pulled once.
1732          * At egress, skb_pospull_rcsum has to be done in case that
1733          * the skb is originated from ingress (i.e. a forwarded skb)
1734          * to ensure that rcsum starts at net header.
1735          */
1736         if (!skb_at_tc_ingress(skb))
1737                 skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1738         skb_pop_mac_header(skb);
1739         skb_reset_mac_len(skb);
1740         return flags & BPF_F_INGRESS ?
1741                __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1742 }
1743
1744 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1745                                  u32 flags)
1746 {
1747         /* Verify that a link layer header is carried */
1748         if (unlikely(skb->mac_header >= skb->network_header)) {
1749                 kfree_skb(skb);
1750                 return -ERANGE;
1751         }
1752
1753         bpf_push_mac_rcsum(skb);
1754         return flags & BPF_F_INGRESS ?
1755                __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1756 }
1757
1758 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1759                           u32 flags)
1760 {
1761         if (dev_is_mac_header_xmit(dev))
1762                 return __bpf_redirect_common(skb, dev, flags);
1763         else
1764                 return __bpf_redirect_no_mac(skb, dev, flags);
1765 }
1766
1767 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1768 {
1769         struct net_device *dev;
1770         struct sk_buff *clone;
1771         int ret;
1772
1773         if (unlikely(flags & ~(BPF_F_INGRESS)))
1774                 return -EINVAL;
1775
1776         dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1777         if (unlikely(!dev))
1778                 return -EINVAL;
1779
1780         clone = skb_clone(skb, GFP_ATOMIC);
1781         if (unlikely(!clone))
1782                 return -ENOMEM;
1783
1784         /* For direct write, we need to keep the invariant that the skbs
1785          * we're dealing with need to be uncloned. Should uncloning fail
1786          * here, we need to free the just generated clone to unclone once
1787          * again.
1788          */
1789         ret = bpf_try_make_head_writable(skb);
1790         if (unlikely(ret)) {
1791                 kfree_skb(clone);
1792                 return -ENOMEM;
1793         }
1794
1795         return __bpf_redirect(clone, dev, flags);
1796 }
1797
1798 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1799         .func           = bpf_clone_redirect,
1800         .gpl_only       = false,
1801         .ret_type       = RET_INTEGER,
1802         .arg1_type      = ARG_PTR_TO_CTX,
1803         .arg2_type      = ARG_ANYTHING,
1804         .arg3_type      = ARG_ANYTHING,
1805 };
1806
1807 struct redirect_info {
1808         u32 ifindex;
1809         u32 flags;
1810         struct bpf_map *map;
1811         struct bpf_map *map_to_flush;
1812         unsigned long   map_owner;
1813 };
1814
1815 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1816
1817 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1818 {
1819         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1820
1821         if (unlikely(flags & ~(BPF_F_INGRESS)))
1822                 return TC_ACT_SHOT;
1823
1824         ri->ifindex = ifindex;
1825         ri->flags = flags;
1826
1827         return TC_ACT_REDIRECT;
1828 }
1829
1830 int skb_do_redirect(struct sk_buff *skb)
1831 {
1832         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1833         struct net_device *dev;
1834
1835         dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1836         ri->ifindex = 0;
1837         if (unlikely(!dev)) {
1838                 kfree_skb(skb);
1839                 return -EINVAL;
1840         }
1841
1842         return __bpf_redirect(skb, dev, ri->flags);
1843 }
1844
1845 static const struct bpf_func_proto bpf_redirect_proto = {
1846         .func           = bpf_redirect,
1847         .gpl_only       = false,
1848         .ret_type       = RET_INTEGER,
1849         .arg1_type      = ARG_ANYTHING,
1850         .arg2_type      = ARG_ANYTHING,
1851 };
1852
1853 BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
1854            struct bpf_map *, map, u32, key, u64, flags)
1855 {
1856         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1857
1858         /* If user passes invalid input drop the packet. */
1859         if (unlikely(flags & ~(BPF_F_INGRESS)))
1860                 return SK_DROP;
1861
1862         tcb->bpf.key = key;
1863         tcb->bpf.flags = flags;
1864         tcb->bpf.map = map;
1865
1866         return SK_PASS;
1867 }
1868
1869 struct sock *do_sk_redirect_map(struct sk_buff *skb)
1870 {
1871         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1872         struct sock *sk = NULL;
1873
1874         if (tcb->bpf.map) {
1875                 sk = __sock_map_lookup_elem(tcb->bpf.map, tcb->bpf.key);
1876
1877                 tcb->bpf.key = 0;
1878                 tcb->bpf.map = NULL;
1879         }
1880
1881         return sk;
1882 }
1883
1884 static const struct bpf_func_proto bpf_sk_redirect_map_proto = {
1885         .func           = bpf_sk_redirect_map,
1886         .gpl_only       = false,
1887         .ret_type       = RET_INTEGER,
1888         .arg1_type      = ARG_PTR_TO_CTX,
1889         .arg2_type      = ARG_CONST_MAP_PTR,
1890         .arg3_type      = ARG_ANYTHING,
1891         .arg4_type      = ARG_ANYTHING,
1892 };
1893
1894 BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg_buff *, msg,
1895            struct bpf_map *, map, u32, key, u64, flags)
1896 {
1897         /* If user passes invalid input drop the packet. */
1898         if (unlikely(flags & ~(BPF_F_INGRESS)))
1899                 return SK_DROP;
1900
1901         msg->key = key;
1902         msg->flags = flags;
1903         msg->map = map;
1904
1905         return SK_PASS;
1906 }
1907
1908 struct sock *do_msg_redirect_map(struct sk_msg_buff *msg)
1909 {
1910         struct sock *sk = NULL;
1911
1912         if (msg->map) {
1913                 sk = __sock_map_lookup_elem(msg->map, msg->key);
1914
1915                 msg->key = 0;
1916                 msg->map = NULL;
1917         }
1918
1919         return sk;
1920 }
1921
1922 static const struct bpf_func_proto bpf_msg_redirect_map_proto = {
1923         .func           = bpf_msg_redirect_map,
1924         .gpl_only       = false,
1925         .ret_type       = RET_INTEGER,
1926         .arg1_type      = ARG_PTR_TO_CTX,
1927         .arg2_type      = ARG_CONST_MAP_PTR,
1928         .arg3_type      = ARG_ANYTHING,
1929         .arg4_type      = ARG_ANYTHING,
1930 };
1931
1932 BPF_CALL_2(bpf_msg_apply_bytes, struct sk_msg_buff *, msg, u32, bytes)
1933 {
1934         msg->apply_bytes = bytes;
1935         return 0;
1936 }
1937
1938 static const struct bpf_func_proto bpf_msg_apply_bytes_proto = {
1939         .func           = bpf_msg_apply_bytes,
1940         .gpl_only       = false,
1941         .ret_type       = RET_INTEGER,
1942         .arg1_type      = ARG_PTR_TO_CTX,
1943         .arg2_type      = ARG_ANYTHING,
1944 };
1945
1946 BPF_CALL_2(bpf_msg_cork_bytes, struct sk_msg_buff *, msg, u32, bytes)
1947 {
1948         msg->cork_bytes = bytes;
1949         return 0;
1950 }
1951
1952 static const struct bpf_func_proto bpf_msg_cork_bytes_proto = {
1953         .func           = bpf_msg_cork_bytes,
1954         .gpl_only       = false,
1955         .ret_type       = RET_INTEGER,
1956         .arg1_type      = ARG_PTR_TO_CTX,
1957         .arg2_type      = ARG_ANYTHING,
1958 };
1959
1960 BPF_CALL_4(bpf_msg_pull_data,
1961            struct sk_msg_buff *, msg, u32, start, u32, end, u64, flags)
1962 {
1963         unsigned int len = 0, offset = 0, copy = 0;
1964         struct scatterlist *sg = msg->sg_data;
1965         int first_sg, last_sg, i, shift;
1966         unsigned char *p, *to, *from;
1967         int bytes = end - start;
1968         struct page *page;
1969
1970         if (unlikely(flags || end <= start))
1971                 return -EINVAL;
1972
1973         /* First find the starting scatterlist element */
1974         i = msg->sg_start;
1975         do {
1976                 len = sg[i].length;
1977                 offset += len;
1978                 if (start < offset + len)
1979                         break;
1980                 i++;
1981                 if (i == MAX_SKB_FRAGS)
1982                         i = 0;
1983         } while (i != msg->sg_end);
1984
1985         if (unlikely(start >= offset + len))
1986                 return -EINVAL;
1987
1988         if (!msg->sg_copy[i] && bytes <= len)
1989                 goto out;
1990
1991         first_sg = i;
1992
1993         /* At this point we need to linearize multiple scatterlist
1994          * elements or a single shared page. Either way we need to
1995          * copy into a linear buffer exclusively owned by BPF. Then
1996          * place the buffer in the scatterlist and fixup the original
1997          * entries by removing the entries now in the linear buffer
1998          * and shifting the remaining entries. For now we do not try
1999          * to copy partial entries to avoid complexity of running out
2000          * of sg_entry slots. The downside is reading a single byte
2001          * will copy the entire sg entry.
2002          */
2003         do {
2004                 copy += sg[i].length;
2005                 i++;
2006                 if (i == MAX_SKB_FRAGS)
2007                         i = 0;
2008                 if (bytes < copy)
2009                         break;
2010         } while (i != msg->sg_end);
2011         last_sg = i;
2012
2013         if (unlikely(copy < end - start))
2014                 return -EINVAL;
2015
2016         page = alloc_pages(__GFP_NOWARN | GFP_ATOMIC, get_order(copy));
2017         if (unlikely(!page))
2018                 return -ENOMEM;
2019         p = page_address(page);
2020         offset = 0;
2021
2022         i = first_sg;
2023         do {
2024                 from = sg_virt(&sg[i]);
2025                 len = sg[i].length;
2026                 to = p + offset;
2027
2028                 memcpy(to, from, len);
2029                 offset += len;
2030                 sg[i].length = 0;
2031                 put_page(sg_page(&sg[i]));
2032
2033                 i++;
2034                 if (i == MAX_SKB_FRAGS)
2035                         i = 0;
2036         } while (i != last_sg);
2037
2038         sg[first_sg].length = copy;
2039         sg_set_page(&sg[first_sg], page, copy, 0);
2040
2041         /* To repair sg ring we need to shift entries. If we only
2042          * had a single entry though we can just replace it and
2043          * be done. Otherwise walk the ring and shift the entries.
2044          */
2045         shift = last_sg - first_sg - 1;
2046         if (!shift)
2047                 goto out;
2048
2049         i = first_sg + 1;
2050         do {
2051                 int move_from;
2052
2053                 if (i + shift >= MAX_SKB_FRAGS)
2054                         move_from = i + shift - MAX_SKB_FRAGS;
2055                 else
2056                         move_from = i + shift;
2057
2058                 if (move_from == msg->sg_end)
2059                         break;
2060
2061                 sg[i] = sg[move_from];
2062                 sg[move_from].length = 0;
2063                 sg[move_from].page_link = 0;
2064                 sg[move_from].offset = 0;
2065
2066                 i++;
2067                 if (i == MAX_SKB_FRAGS)
2068                         i = 0;
2069         } while (1);
2070         msg->sg_end -= shift;
2071         if (msg->sg_end < 0)
2072                 msg->sg_end += MAX_SKB_FRAGS;
2073 out:
2074         msg->data = sg_virt(&sg[i]) + start - offset;
2075         msg->data_end = msg->data + bytes;
2076
2077         return 0;
2078 }
2079
2080 static const struct bpf_func_proto bpf_msg_pull_data_proto = {
2081         .func           = bpf_msg_pull_data,
2082         .gpl_only       = false,
2083         .ret_type       = RET_INTEGER,
2084         .arg1_type      = ARG_PTR_TO_CTX,
2085         .arg2_type      = ARG_ANYTHING,
2086         .arg3_type      = ARG_ANYTHING,
2087         .arg4_type      = ARG_ANYTHING,
2088 };
2089
2090 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
2091 {
2092         return task_get_classid(skb);
2093 }
2094
2095 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
2096         .func           = bpf_get_cgroup_classid,
2097         .gpl_only       = false,
2098         .ret_type       = RET_INTEGER,
2099         .arg1_type      = ARG_PTR_TO_CTX,
2100 };
2101
2102 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
2103 {
2104         return dst_tclassid(skb);
2105 }
2106
2107 static const struct bpf_func_proto bpf_get_route_realm_proto = {
2108         .func           = bpf_get_route_realm,
2109         .gpl_only       = false,
2110         .ret_type       = RET_INTEGER,
2111         .arg1_type      = ARG_PTR_TO_CTX,
2112 };
2113
2114 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
2115 {
2116         /* If skb_clear_hash() was called due to mangling, we can
2117          * trigger SW recalculation here. Later access to hash
2118          * can then use the inline skb->hash via context directly
2119          * instead of calling this helper again.
2120          */
2121         return skb_get_hash(skb);
2122 }
2123
2124 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
2125         .func           = bpf_get_hash_recalc,
2126         .gpl_only       = false,
2127         .ret_type       = RET_INTEGER,
2128         .arg1_type      = ARG_PTR_TO_CTX,
2129 };
2130
2131 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
2132 {
2133         /* After all direct packet write, this can be used once for
2134          * triggering a lazy recalc on next skb_get_hash() invocation.
2135          */
2136         skb_clear_hash(skb);
2137         return 0;
2138 }
2139
2140 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
2141         .func           = bpf_set_hash_invalid,
2142         .gpl_only       = false,
2143         .ret_type       = RET_INTEGER,
2144         .arg1_type      = ARG_PTR_TO_CTX,
2145 };
2146
2147 BPF_CALL_2(bpf_set_hash, struct sk_buff *, skb, u32, hash)
2148 {
2149         /* Set user specified hash as L4(+), so that it gets returned
2150          * on skb_get_hash() call unless BPF prog later on triggers a
2151          * skb_clear_hash().
2152          */
2153         __skb_set_sw_hash(skb, hash, true);
2154         return 0;
2155 }
2156
2157 static const struct bpf_func_proto bpf_set_hash_proto = {
2158         .func           = bpf_set_hash,
2159         .gpl_only       = false,
2160         .ret_type       = RET_INTEGER,
2161         .arg1_type      = ARG_PTR_TO_CTX,
2162         .arg2_type      = ARG_ANYTHING,
2163 };
2164
2165 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
2166            u16, vlan_tci)
2167 {
2168         int ret;
2169
2170         if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
2171                      vlan_proto != htons(ETH_P_8021AD)))
2172                 vlan_proto = htons(ETH_P_8021Q);
2173
2174         bpf_push_mac_rcsum(skb);
2175         ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
2176         bpf_pull_mac_rcsum(skb);
2177
2178         bpf_compute_data_pointers(skb);
2179         return ret;
2180 }
2181
2182 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
2183         .func           = bpf_skb_vlan_push,
2184         .gpl_only       = false,
2185         .ret_type       = RET_INTEGER,
2186         .arg1_type      = ARG_PTR_TO_CTX,
2187         .arg2_type      = ARG_ANYTHING,
2188         .arg3_type      = ARG_ANYTHING,
2189 };
2190 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
2191
2192 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
2193 {
2194         int ret;
2195
2196         bpf_push_mac_rcsum(skb);
2197         ret = skb_vlan_pop(skb);
2198         bpf_pull_mac_rcsum(skb);
2199
2200         bpf_compute_data_pointers(skb);
2201         return ret;
2202 }
2203
2204 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
2205         .func           = bpf_skb_vlan_pop,
2206         .gpl_only       = false,
2207         .ret_type       = RET_INTEGER,
2208         .arg1_type      = ARG_PTR_TO_CTX,
2209 };
2210 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
2211
2212 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
2213 {
2214         /* Caller already did skb_cow() with len as headroom,
2215          * so no need to do it here.
2216          */
2217         skb_push(skb, len);
2218         memmove(skb->data, skb->data + len, off);
2219         memset(skb->data + off, 0, len);
2220
2221         /* No skb_postpush_rcsum(skb, skb->data + off, len)
2222          * needed here as it does not change the skb->csum
2223          * result for checksum complete when summing over
2224          * zeroed blocks.
2225          */
2226         return 0;
2227 }
2228
2229 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
2230 {
2231         /* skb_ensure_writable() is not needed here, as we're
2232          * already working on an uncloned skb.
2233          */
2234         if (unlikely(!pskb_may_pull(skb, off + len)))
2235                 return -ENOMEM;
2236
2237         skb_postpull_rcsum(skb, skb->data + off, len);
2238         memmove(skb->data + len, skb->data, off);
2239         __skb_pull(skb, len);
2240
2241         return 0;
2242 }
2243
2244 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
2245 {
2246         bool trans_same = skb->transport_header == skb->network_header;
2247         int ret;
2248
2249         /* There's no need for __skb_push()/__skb_pull() pair to
2250          * get to the start of the mac header as we're guaranteed
2251          * to always start from here under eBPF.
2252          */
2253         ret = bpf_skb_generic_push(skb, off, len);
2254         if (likely(!ret)) {
2255                 skb->mac_header -= len;
2256                 skb->network_header -= len;
2257                 if (trans_same)
2258                         skb->transport_header = skb->network_header;
2259         }
2260
2261         return ret;
2262 }
2263
2264 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
2265 {
2266         bool trans_same = skb->transport_header == skb->network_header;
2267         int ret;
2268
2269         /* Same here, __skb_push()/__skb_pull() pair not needed. */
2270         ret = bpf_skb_generic_pop(skb, off, len);
2271         if (likely(!ret)) {
2272                 skb->mac_header += len;
2273                 skb->network_header += len;
2274                 if (trans_same)
2275                         skb->transport_header = skb->network_header;
2276         }
2277
2278         return ret;
2279 }
2280
2281 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
2282 {
2283         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2284         u32 off = skb_mac_header_len(skb);
2285         int ret;
2286
2287         /* SCTP uses GSO_BY_FRAGS, thus cannot adjust it. */
2288         if (skb_is_gso(skb) && unlikely(skb_is_gso_sctp(skb)))
2289                 return -ENOTSUPP;
2290
2291         ret = skb_cow(skb, len_diff);
2292         if (unlikely(ret < 0))
2293                 return ret;
2294
2295         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2296         if (unlikely(ret < 0))
2297                 return ret;
2298
2299         if (skb_is_gso(skb)) {
2300                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2301
2302                 /* SKB_GSO_TCPV4 needs to be changed into
2303                  * SKB_GSO_TCPV6.
2304                  */
2305                 if (shinfo->gso_type & SKB_GSO_TCPV4) {
2306                         shinfo->gso_type &= ~SKB_GSO_TCPV4;
2307                         shinfo->gso_type |=  SKB_GSO_TCPV6;
2308                 }
2309
2310                 /* Due to IPv6 header, MSS needs to be downgraded. */
2311                 skb_decrease_gso_size(shinfo, len_diff);
2312                 /* Header must be checked, and gso_segs recomputed. */
2313                 shinfo->gso_type |= SKB_GSO_DODGY;
2314                 shinfo->gso_segs = 0;
2315         }
2316
2317         skb->protocol = htons(ETH_P_IPV6);
2318         skb_clear_hash(skb);
2319
2320         return 0;
2321 }
2322
2323 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2324 {
2325         const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2326         u32 off = skb_mac_header_len(skb);
2327         int ret;
2328
2329         /* SCTP uses GSO_BY_FRAGS, thus cannot adjust it. */
2330         if (skb_is_gso(skb) && unlikely(skb_is_gso_sctp(skb)))
2331                 return -ENOTSUPP;
2332
2333         ret = skb_unclone(skb, GFP_ATOMIC);
2334         if (unlikely(ret < 0))
2335                 return ret;
2336
2337         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2338         if (unlikely(ret < 0))
2339                 return ret;
2340
2341         if (skb_is_gso(skb)) {
2342                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2343
2344                 /* SKB_GSO_TCPV6 needs to be changed into
2345                  * SKB_GSO_TCPV4.
2346                  */
2347                 if (shinfo->gso_type & SKB_GSO_TCPV6) {
2348                         shinfo->gso_type &= ~SKB_GSO_TCPV6;
2349                         shinfo->gso_type |=  SKB_GSO_TCPV4;
2350                 }
2351
2352                 /* Due to IPv4 header, MSS can be upgraded. */
2353                 skb_increase_gso_size(shinfo, len_diff);
2354                 /* Header must be checked, and gso_segs recomputed. */
2355                 shinfo->gso_type |= SKB_GSO_DODGY;
2356                 shinfo->gso_segs = 0;
2357         }
2358
2359         skb->protocol = htons(ETH_P_IP);
2360         skb_clear_hash(skb);
2361
2362         return 0;
2363 }
2364
2365 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2366 {
2367         __be16 from_proto = skb->protocol;
2368
2369         if (from_proto == htons(ETH_P_IP) &&
2370               to_proto == htons(ETH_P_IPV6))
2371                 return bpf_skb_proto_4_to_6(skb);
2372
2373         if (from_proto == htons(ETH_P_IPV6) &&
2374               to_proto == htons(ETH_P_IP))
2375                 return bpf_skb_proto_6_to_4(skb);
2376
2377         return -ENOTSUPP;
2378 }
2379
2380 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2381            u64, flags)
2382 {
2383         int ret;
2384
2385         if (unlikely(flags))
2386                 return -EINVAL;
2387
2388         /* General idea is that this helper does the basic groundwork
2389          * needed for changing the protocol, and eBPF program fills the
2390          * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2391          * and other helpers, rather than passing a raw buffer here.
2392          *
2393          * The rationale is to keep this minimal and without a need to
2394          * deal with raw packet data. F.e. even if we would pass buffers
2395          * here, the program still needs to call the bpf_lX_csum_replace()
2396          * helpers anyway. Plus, this way we keep also separation of
2397          * concerns, since f.e. bpf_skb_store_bytes() should only take
2398          * care of stores.
2399          *
2400          * Currently, additional options and extension header space are
2401          * not supported, but flags register is reserved so we can adapt
2402          * that. For offloads, we mark packet as dodgy, so that headers
2403          * need to be verified first.
2404          */
2405         ret = bpf_skb_proto_xlat(skb, proto);
2406         bpf_compute_data_pointers(skb);
2407         return ret;
2408 }
2409
2410 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2411         .func           = bpf_skb_change_proto,
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_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2420 {
2421         /* We only allow a restricted subset to be changed for now. */
2422         if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2423                      !skb_pkt_type_ok(pkt_type)))
2424                 return -EINVAL;
2425
2426         skb->pkt_type = pkt_type;
2427         return 0;
2428 }
2429
2430 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2431         .func           = bpf_skb_change_type,
2432         .gpl_only       = false,
2433         .ret_type       = RET_INTEGER,
2434         .arg1_type      = ARG_PTR_TO_CTX,
2435         .arg2_type      = ARG_ANYTHING,
2436 };
2437
2438 static u32 bpf_skb_net_base_len(const struct sk_buff *skb)
2439 {
2440         switch (skb->protocol) {
2441         case htons(ETH_P_IP):
2442                 return sizeof(struct iphdr);
2443         case htons(ETH_P_IPV6):
2444                 return sizeof(struct ipv6hdr);
2445         default:
2446                 return ~0U;
2447         }
2448 }
2449
2450 static int bpf_skb_net_grow(struct sk_buff *skb, u32 len_diff)
2451 {
2452         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2453         int ret;
2454
2455         /* SCTP uses GSO_BY_FRAGS, thus cannot adjust it. */
2456         if (skb_is_gso(skb) && unlikely(skb_is_gso_sctp(skb)))
2457                 return -ENOTSUPP;
2458
2459         ret = skb_cow(skb, len_diff);
2460         if (unlikely(ret < 0))
2461                 return ret;
2462
2463         ret = bpf_skb_net_hdr_push(skb, off, len_diff);
2464         if (unlikely(ret < 0))
2465                 return ret;
2466
2467         if (skb_is_gso(skb)) {
2468                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2469
2470                 /* Due to header grow, MSS needs to be downgraded. */
2471                 skb_decrease_gso_size(shinfo, len_diff);
2472                 /* Header must be checked, and gso_segs recomputed. */
2473                 shinfo->gso_type |= SKB_GSO_DODGY;
2474                 shinfo->gso_segs = 0;
2475         }
2476
2477         return 0;
2478 }
2479
2480 static int bpf_skb_net_shrink(struct sk_buff *skb, u32 len_diff)
2481 {
2482         u32 off = skb_mac_header_len(skb) + bpf_skb_net_base_len(skb);
2483         int ret;
2484
2485         /* SCTP uses GSO_BY_FRAGS, thus cannot adjust it. */
2486         if (skb_is_gso(skb) && unlikely(skb_is_gso_sctp(skb)))
2487                 return -ENOTSUPP;
2488
2489         ret = skb_unclone(skb, GFP_ATOMIC);
2490         if (unlikely(ret < 0))
2491                 return ret;
2492
2493         ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2494         if (unlikely(ret < 0))
2495                 return ret;
2496
2497         if (skb_is_gso(skb)) {
2498                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2499
2500                 /* Due to header shrink, MSS can be upgraded. */
2501                 skb_increase_gso_size(shinfo, len_diff);
2502                 /* Header must be checked, and gso_segs recomputed. */
2503                 shinfo->gso_type |= SKB_GSO_DODGY;
2504                 shinfo->gso_segs = 0;
2505         }
2506
2507         return 0;
2508 }
2509
2510 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2511 {
2512         return skb->dev->mtu + skb->dev->hard_header_len;
2513 }
2514
2515 static int bpf_skb_adjust_net(struct sk_buff *skb, s32 len_diff)
2516 {
2517         bool trans_same = skb->transport_header == skb->network_header;
2518         u32 len_cur, len_diff_abs = abs(len_diff);
2519         u32 len_min = bpf_skb_net_base_len(skb);
2520         u32 len_max = __bpf_skb_max_len(skb);
2521         __be16 proto = skb->protocol;
2522         bool shrink = len_diff < 0;
2523         int ret;
2524
2525         if (unlikely(len_diff_abs > 0xfffU))
2526                 return -EFAULT;
2527         if (unlikely(proto != htons(ETH_P_IP) &&
2528                      proto != htons(ETH_P_IPV6)))
2529                 return -ENOTSUPP;
2530
2531         len_cur = skb->len - skb_network_offset(skb);
2532         if (skb_transport_header_was_set(skb) && !trans_same)
2533                 len_cur = skb_network_header_len(skb);
2534         if ((shrink && (len_diff_abs >= len_cur ||
2535                         len_cur - len_diff_abs < len_min)) ||
2536             (!shrink && (skb->len + len_diff_abs > len_max &&
2537                          !skb_is_gso(skb))))
2538                 return -ENOTSUPP;
2539
2540         ret = shrink ? bpf_skb_net_shrink(skb, len_diff_abs) :
2541                        bpf_skb_net_grow(skb, len_diff_abs);
2542
2543         bpf_compute_data_pointers(skb);
2544         return ret;
2545 }
2546
2547 BPF_CALL_4(bpf_skb_adjust_room, struct sk_buff *, skb, s32, len_diff,
2548            u32, mode, u64, flags)
2549 {
2550         if (unlikely(flags))
2551                 return -EINVAL;
2552         if (likely(mode == BPF_ADJ_ROOM_NET))
2553                 return bpf_skb_adjust_net(skb, len_diff);
2554
2555         return -ENOTSUPP;
2556 }
2557
2558 static const struct bpf_func_proto bpf_skb_adjust_room_proto = {
2559         .func           = bpf_skb_adjust_room,
2560         .gpl_only       = false,
2561         .ret_type       = RET_INTEGER,
2562         .arg1_type      = ARG_PTR_TO_CTX,
2563         .arg2_type      = ARG_ANYTHING,
2564         .arg3_type      = ARG_ANYTHING,
2565         .arg4_type      = ARG_ANYTHING,
2566 };
2567
2568 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2569 {
2570         u32 min_len = skb_network_offset(skb);
2571
2572         if (skb_transport_header_was_set(skb))
2573                 min_len = skb_transport_offset(skb);
2574         if (skb->ip_summed == CHECKSUM_PARTIAL)
2575                 min_len = skb_checksum_start_offset(skb) +
2576                           skb->csum_offset + sizeof(__sum16);
2577         return min_len;
2578 }
2579
2580 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2581 {
2582         unsigned int old_len = skb->len;
2583         int ret;
2584
2585         ret = __skb_grow_rcsum(skb, new_len);
2586         if (!ret)
2587                 memset(skb->data + old_len, 0, new_len - old_len);
2588         return ret;
2589 }
2590
2591 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2592 {
2593         return __skb_trim_rcsum(skb, new_len);
2594 }
2595
2596 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2597            u64, flags)
2598 {
2599         u32 max_len = __bpf_skb_max_len(skb);
2600         u32 min_len = __bpf_skb_min_len(skb);
2601         int ret;
2602
2603         if (unlikely(flags || new_len > max_len || new_len < min_len))
2604                 return -EINVAL;
2605         if (skb->encapsulation)
2606                 return -ENOTSUPP;
2607
2608         /* The basic idea of this helper is that it's performing the
2609          * needed work to either grow or trim an skb, and eBPF program
2610          * rewrites the rest via helpers like bpf_skb_store_bytes(),
2611          * bpf_lX_csum_replace() and others rather than passing a raw
2612          * buffer here. This one is a slow path helper and intended
2613          * for replies with control messages.
2614          *
2615          * Like in bpf_skb_change_proto(), we want to keep this rather
2616          * minimal and without protocol specifics so that we are able
2617          * to separate concerns as in bpf_skb_store_bytes() should only
2618          * be the one responsible for writing buffers.
2619          *
2620          * It's really expected to be a slow path operation here for
2621          * control message replies, so we're implicitly linearizing,
2622          * uncloning and drop offloads from the skb by this.
2623          */
2624         ret = __bpf_try_make_writable(skb, skb->len);
2625         if (!ret) {
2626                 if (new_len > skb->len)
2627                         ret = bpf_skb_grow_rcsum(skb, new_len);
2628                 else if (new_len < skb->len)
2629                         ret = bpf_skb_trim_rcsum(skb, new_len);
2630                 if (!ret && skb_is_gso(skb))
2631                         skb_gso_reset(skb);
2632         }
2633
2634         bpf_compute_data_pointers(skb);
2635         return ret;
2636 }
2637
2638 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2639         .func           = bpf_skb_change_tail,
2640         .gpl_only       = false,
2641         .ret_type       = RET_INTEGER,
2642         .arg1_type      = ARG_PTR_TO_CTX,
2643         .arg2_type      = ARG_ANYTHING,
2644         .arg3_type      = ARG_ANYTHING,
2645 };
2646
2647 BPF_CALL_3(bpf_skb_change_head, struct sk_buff *, skb, u32, head_room,
2648            u64, flags)
2649 {
2650         u32 max_len = __bpf_skb_max_len(skb);
2651         u32 new_len = skb->len + head_room;
2652         int ret;
2653
2654         if (unlikely(flags || (!skb_is_gso(skb) && new_len > max_len) ||
2655                      new_len < skb->len))
2656                 return -EINVAL;
2657
2658         ret = skb_cow(skb, head_room);
2659         if (likely(!ret)) {
2660                 /* Idea for this helper is that we currently only
2661                  * allow to expand on mac header. This means that
2662                  * skb->protocol network header, etc, stay as is.
2663                  * Compared to bpf_skb_change_tail(), we're more
2664                  * flexible due to not needing to linearize or
2665                  * reset GSO. Intention for this helper is to be
2666                  * used by an L3 skb that needs to push mac header
2667                  * for redirection into L2 device.
2668                  */
2669                 __skb_push(skb, head_room);
2670                 memset(skb->data, 0, head_room);
2671                 skb_reset_mac_header(skb);
2672         }
2673
2674         bpf_compute_data_pointers(skb);
2675         return 0;
2676 }
2677
2678 static const struct bpf_func_proto bpf_skb_change_head_proto = {
2679         .func           = bpf_skb_change_head,
2680         .gpl_only       = false,
2681         .ret_type       = RET_INTEGER,
2682         .arg1_type      = ARG_PTR_TO_CTX,
2683         .arg2_type      = ARG_ANYTHING,
2684         .arg3_type      = ARG_ANYTHING,
2685 };
2686
2687 static unsigned long xdp_get_metalen(const struct xdp_buff *xdp)
2688 {
2689         return xdp_data_meta_unsupported(xdp) ? 0 :
2690                xdp->data - xdp->data_meta;
2691 }
2692
2693 BPF_CALL_2(bpf_xdp_adjust_head, struct xdp_buff *, xdp, int, offset)
2694 {
2695         void *xdp_frame_end = xdp->data_hard_start + sizeof(struct xdp_frame);
2696         unsigned long metalen = xdp_get_metalen(xdp);
2697         void *data_start = xdp_frame_end + metalen;
2698         void *data = xdp->data + offset;
2699
2700         if (unlikely(data < data_start ||
2701                      data > xdp->data_end - ETH_HLEN))
2702                 return -EINVAL;
2703
2704         if (metalen)
2705                 memmove(xdp->data_meta + offset,
2706                         xdp->data_meta, metalen);
2707         xdp->data_meta += offset;
2708         xdp->data = data;
2709
2710         return 0;
2711 }
2712
2713 static const struct bpf_func_proto bpf_xdp_adjust_head_proto = {
2714         .func           = bpf_xdp_adjust_head,
2715         .gpl_only       = false,
2716         .ret_type       = RET_INTEGER,
2717         .arg1_type      = ARG_PTR_TO_CTX,
2718         .arg2_type      = ARG_ANYTHING,
2719 };
2720
2721 BPF_CALL_2(bpf_xdp_adjust_tail, struct xdp_buff *, xdp, int, offset)
2722 {
2723         void *data_end = xdp->data_end + offset;
2724
2725         /* only shrinking is allowed for now. */
2726         if (unlikely(offset >= 0))
2727                 return -EINVAL;
2728
2729         if (unlikely(data_end < xdp->data + ETH_HLEN))
2730                 return -EINVAL;
2731
2732         xdp->data_end = data_end;
2733
2734         return 0;
2735 }
2736
2737 static const struct bpf_func_proto bpf_xdp_adjust_tail_proto = {
2738         .func           = bpf_xdp_adjust_tail,
2739         .gpl_only       = false,
2740         .ret_type       = RET_INTEGER,
2741         .arg1_type      = ARG_PTR_TO_CTX,
2742         .arg2_type      = ARG_ANYTHING,
2743 };
2744
2745 BPF_CALL_2(bpf_xdp_adjust_meta, struct xdp_buff *, xdp, int, offset)
2746 {
2747         void *xdp_frame_end = xdp->data_hard_start + sizeof(struct xdp_frame);
2748         void *meta = xdp->data_meta + offset;
2749         unsigned long metalen = xdp->data - meta;
2750
2751         if (xdp_data_meta_unsupported(xdp))
2752                 return -ENOTSUPP;
2753         if (unlikely(meta < xdp_frame_end ||
2754                      meta > xdp->data))
2755                 return -EINVAL;
2756         if (unlikely((metalen & (sizeof(__u32) - 1)) ||
2757                      (metalen > 32)))
2758                 return -EACCES;
2759
2760         xdp->data_meta = meta;
2761
2762         return 0;
2763 }
2764
2765 static const struct bpf_func_proto bpf_xdp_adjust_meta_proto = {
2766         .func           = bpf_xdp_adjust_meta,
2767         .gpl_only       = false,
2768         .ret_type       = RET_INTEGER,
2769         .arg1_type      = ARG_PTR_TO_CTX,
2770         .arg2_type      = ARG_ANYTHING,
2771 };
2772
2773 static int __bpf_tx_xdp(struct net_device *dev,
2774                         struct bpf_map *map,
2775                         struct xdp_buff *xdp,
2776                         u32 index)
2777 {
2778         struct xdp_frame *xdpf;
2779         int err;
2780
2781         if (!dev->netdev_ops->ndo_xdp_xmit) {
2782                 return -EOPNOTSUPP;
2783         }
2784
2785         xdpf = convert_to_xdp_frame(xdp);
2786         if (unlikely(!xdpf))
2787                 return -EOVERFLOW;
2788
2789         err = dev->netdev_ops->ndo_xdp_xmit(dev, xdpf);
2790         if (err)
2791                 return err;
2792         dev->netdev_ops->ndo_xdp_flush(dev);
2793         return 0;
2794 }
2795
2796 static int __bpf_tx_xdp_map(struct net_device *dev_rx, void *fwd,
2797                             struct bpf_map *map,
2798                             struct xdp_buff *xdp,
2799                             u32 index)
2800 {
2801         int err;
2802
2803         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2804                 struct net_device *dev = fwd;
2805                 struct xdp_frame *xdpf;
2806
2807                 if (!dev->netdev_ops->ndo_xdp_xmit)
2808                         return -EOPNOTSUPP;
2809
2810                 xdpf = convert_to_xdp_frame(xdp);
2811                 if (unlikely(!xdpf))
2812                         return -EOVERFLOW;
2813
2814                 /* TODO: move to inside map code instead, for bulk support
2815                  * err = dev_map_enqueue(dev, xdp);
2816                  */
2817                 err = dev->netdev_ops->ndo_xdp_xmit(dev, xdpf);
2818                 if (err)
2819                         return err;
2820                 __dev_map_insert_ctx(map, index);
2821
2822         } else if (map->map_type == BPF_MAP_TYPE_CPUMAP) {
2823                 struct bpf_cpu_map_entry *rcpu = fwd;
2824
2825                 err = cpu_map_enqueue(rcpu, xdp, dev_rx);
2826                 if (err)
2827                         return err;
2828                 __cpu_map_insert_ctx(map, index);
2829         }
2830         return 0;
2831 }
2832
2833 void xdp_do_flush_map(void)
2834 {
2835         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2836         struct bpf_map *map = ri->map_to_flush;
2837
2838         ri->map_to_flush = NULL;
2839         if (map) {
2840                 switch (map->map_type) {
2841                 case BPF_MAP_TYPE_DEVMAP:
2842                         __dev_map_flush(map);
2843                         break;
2844                 case BPF_MAP_TYPE_CPUMAP:
2845                         __cpu_map_flush(map);
2846                         break;
2847                 default:
2848                         break;
2849                 }
2850         }
2851 }
2852 EXPORT_SYMBOL_GPL(xdp_do_flush_map);
2853
2854 static void *__xdp_map_lookup_elem(struct bpf_map *map, u32 index)
2855 {
2856         switch (map->map_type) {
2857         case BPF_MAP_TYPE_DEVMAP:
2858                 return __dev_map_lookup_elem(map, index);
2859         case BPF_MAP_TYPE_CPUMAP:
2860                 return __cpu_map_lookup_elem(map, index);
2861         default:
2862                 return NULL;
2863         }
2864 }
2865
2866 static inline bool xdp_map_invalid(const struct bpf_prog *xdp_prog,
2867                                    unsigned long aux)
2868 {
2869         return (unsigned long)xdp_prog->aux != aux;
2870 }
2871
2872 static int xdp_do_redirect_map(struct net_device *dev, struct xdp_buff *xdp,
2873                                struct bpf_prog *xdp_prog)
2874 {
2875         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2876         unsigned long map_owner = ri->map_owner;
2877         struct bpf_map *map = ri->map;
2878         u32 index = ri->ifindex;
2879         void *fwd = NULL;
2880         int err;
2881
2882         ri->ifindex = 0;
2883         ri->map = NULL;
2884         ri->map_owner = 0;
2885
2886         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2887                 err = -EFAULT;
2888                 map = NULL;
2889                 goto err;
2890         }
2891
2892         fwd = __xdp_map_lookup_elem(map, index);
2893         if (!fwd) {
2894                 err = -EINVAL;
2895                 goto err;
2896         }
2897         if (ri->map_to_flush && ri->map_to_flush != map)
2898                 xdp_do_flush_map();
2899
2900         err = __bpf_tx_xdp_map(dev, fwd, map, xdp, index);
2901         if (unlikely(err))
2902                 goto err;
2903
2904         ri->map_to_flush = map;
2905         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2906         return 0;
2907 err:
2908         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2909         return err;
2910 }
2911
2912 int xdp_do_redirect(struct net_device *dev, struct xdp_buff *xdp,
2913                     struct bpf_prog *xdp_prog)
2914 {
2915         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2916         struct net_device *fwd;
2917         u32 index = ri->ifindex;
2918         int err;
2919
2920         if (ri->map)
2921                 return xdp_do_redirect_map(dev, xdp, xdp_prog);
2922
2923         fwd = dev_get_by_index_rcu(dev_net(dev), index);
2924         ri->ifindex = 0;
2925         if (unlikely(!fwd)) {
2926                 err = -EINVAL;
2927                 goto err;
2928         }
2929
2930         err = __bpf_tx_xdp(fwd, NULL, xdp, 0);
2931         if (unlikely(err))
2932                 goto err;
2933
2934         _trace_xdp_redirect(dev, xdp_prog, index);
2935         return 0;
2936 err:
2937         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
2938         return err;
2939 }
2940 EXPORT_SYMBOL_GPL(xdp_do_redirect);
2941
2942 static int __xdp_generic_ok_fwd_dev(struct sk_buff *skb, struct net_device *fwd)
2943 {
2944         unsigned int len;
2945
2946         if (unlikely(!(fwd->flags & IFF_UP)))
2947                 return -ENETDOWN;
2948
2949         len = fwd->mtu + fwd->hard_header_len + VLAN_HLEN;
2950         if (skb->len > len)
2951                 return -EMSGSIZE;
2952
2953         return 0;
2954 }
2955
2956 static int xdp_do_generic_redirect_map(struct net_device *dev,
2957                                        struct sk_buff *skb,
2958                                        struct bpf_prog *xdp_prog)
2959 {
2960         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
2961         unsigned long map_owner = ri->map_owner;
2962         struct bpf_map *map = ri->map;
2963         struct net_device *fwd = NULL;
2964         u32 index = ri->ifindex;
2965         int err = 0;
2966
2967         ri->ifindex = 0;
2968         ri->map = NULL;
2969         ri->map_owner = 0;
2970
2971         if (unlikely(xdp_map_invalid(xdp_prog, map_owner))) {
2972                 err = -EFAULT;
2973                 map = NULL;
2974                 goto err;
2975         }
2976         fwd = __xdp_map_lookup_elem(map, index);
2977         if (unlikely(!fwd)) {
2978                 err = -EINVAL;
2979                 goto err;
2980         }
2981
2982         if (map->map_type == BPF_MAP_TYPE_DEVMAP) {
2983                 if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
2984                         goto err;
2985                 skb->dev = fwd;
2986         } else {
2987                 /* TODO: Handle BPF_MAP_TYPE_CPUMAP */
2988                 err = -EBADRQC;
2989                 goto err;
2990         }
2991
2992         _trace_xdp_redirect_map(dev, xdp_prog, fwd, map, index);
2993         return 0;
2994 err:
2995         _trace_xdp_redirect_map_err(dev, xdp_prog, fwd, map, index, err);
2996         return err;
2997 }
2998
2999 int xdp_do_generic_redirect(struct net_device *dev, struct sk_buff *skb,
3000                             struct bpf_prog *xdp_prog)
3001 {
3002         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
3003         u32 index = ri->ifindex;
3004         struct net_device *fwd;
3005         int err = 0;
3006
3007         if (ri->map)
3008                 return xdp_do_generic_redirect_map(dev, skb, xdp_prog);
3009
3010         ri->ifindex = 0;
3011         fwd = dev_get_by_index_rcu(dev_net(dev), index);
3012         if (unlikely(!fwd)) {
3013                 err = -EINVAL;
3014                 goto err;
3015         }
3016
3017         if (unlikely((err = __xdp_generic_ok_fwd_dev(skb, fwd))))
3018                 goto err;
3019
3020         skb->dev = fwd;
3021         _trace_xdp_redirect(dev, xdp_prog, index);
3022         return 0;
3023 err:
3024         _trace_xdp_redirect_err(dev, xdp_prog, index, err);
3025         return err;
3026 }
3027 EXPORT_SYMBOL_GPL(xdp_do_generic_redirect);
3028
3029 BPF_CALL_2(bpf_xdp_redirect, u32, ifindex, u64, flags)
3030 {
3031         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
3032
3033         if (unlikely(flags))
3034                 return XDP_ABORTED;
3035
3036         ri->ifindex = ifindex;
3037         ri->flags = flags;
3038         ri->map = NULL;
3039         ri->map_owner = 0;
3040
3041         return XDP_REDIRECT;
3042 }
3043
3044 static const struct bpf_func_proto bpf_xdp_redirect_proto = {
3045         .func           = bpf_xdp_redirect,
3046         .gpl_only       = false,
3047         .ret_type       = RET_INTEGER,
3048         .arg1_type      = ARG_ANYTHING,
3049         .arg2_type      = ARG_ANYTHING,
3050 };
3051
3052 BPF_CALL_4(bpf_xdp_redirect_map, struct bpf_map *, map, u32, ifindex, u64, flags,
3053            unsigned long, map_owner)
3054 {
3055         struct redirect_info *ri = this_cpu_ptr(&redirect_info);
3056
3057         if (unlikely(flags))
3058                 return XDP_ABORTED;
3059
3060         ri->ifindex = ifindex;
3061         ri->flags = flags;
3062         ri->map = map;
3063         ri->map_owner = map_owner;
3064
3065         return XDP_REDIRECT;
3066 }
3067
3068 /* Note, arg4 is hidden from users and populated by the verifier
3069  * with the right pointer.
3070  */
3071 static const struct bpf_func_proto bpf_xdp_redirect_map_proto = {
3072         .func           = bpf_xdp_redirect_map,
3073         .gpl_only       = false,
3074         .ret_type       = RET_INTEGER,
3075         .arg1_type      = ARG_CONST_MAP_PTR,
3076         .arg2_type      = ARG_ANYTHING,
3077         .arg3_type      = ARG_ANYTHING,
3078 };
3079
3080 bool bpf_helper_changes_pkt_data(void *func)
3081 {
3082         if (func == bpf_skb_vlan_push ||
3083             func == bpf_skb_vlan_pop ||
3084             func == bpf_skb_store_bytes ||
3085             func == bpf_skb_change_proto ||
3086             func == bpf_skb_change_head ||
3087             func == bpf_skb_change_tail ||
3088             func == bpf_skb_adjust_room ||
3089             func == bpf_skb_pull_data ||
3090             func == bpf_clone_redirect ||
3091             func == bpf_l3_csum_replace ||
3092             func == bpf_l4_csum_replace ||
3093             func == bpf_xdp_adjust_head ||
3094             func == bpf_xdp_adjust_meta ||
3095             func == bpf_msg_pull_data ||
3096             func == bpf_xdp_adjust_tail)
3097                 return true;
3098
3099         return false;
3100 }
3101
3102 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
3103                                   unsigned long off, unsigned long len)
3104 {
3105         void *ptr = skb_header_pointer(skb, off, len, dst_buff);
3106
3107         if (unlikely(!ptr))
3108                 return len;
3109         if (ptr != dst_buff)
3110                 memcpy(dst_buff, ptr, len);
3111
3112         return 0;
3113 }
3114
3115 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
3116            u64, flags, void *, meta, u64, meta_size)
3117 {
3118         u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
3119
3120         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
3121                 return -EINVAL;
3122         if (unlikely(skb_size > skb->len))
3123                 return -EFAULT;
3124
3125         return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
3126                                 bpf_skb_copy);
3127 }
3128
3129 static const struct bpf_func_proto bpf_skb_event_output_proto = {
3130         .func           = bpf_skb_event_output,
3131         .gpl_only       = true,
3132         .ret_type       = RET_INTEGER,
3133         .arg1_type      = ARG_PTR_TO_CTX,
3134         .arg2_type      = ARG_CONST_MAP_PTR,
3135         .arg3_type      = ARG_ANYTHING,
3136         .arg4_type      = ARG_PTR_TO_MEM,
3137         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
3138 };
3139
3140 static unsigned short bpf_tunnel_key_af(u64 flags)
3141 {
3142         return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
3143 }
3144
3145 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
3146            u32, size, u64, flags)
3147 {
3148         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
3149         u8 compat[sizeof(struct bpf_tunnel_key)];
3150         void *to_orig = to;
3151         int err;
3152
3153         if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
3154                 err = -EINVAL;
3155                 goto err_clear;
3156         }
3157         if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
3158                 err = -EPROTO;
3159                 goto err_clear;
3160         }
3161         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
3162                 err = -EINVAL;
3163                 switch (size) {
3164                 case offsetof(struct bpf_tunnel_key, tunnel_label):
3165                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
3166                         goto set_compat;
3167                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
3168                         /* Fixup deprecated structure layouts here, so we have
3169                          * a common path later on.
3170                          */
3171                         if (ip_tunnel_info_af(info) != AF_INET)
3172                                 goto err_clear;
3173 set_compat:
3174                         to = (struct bpf_tunnel_key *)compat;
3175                         break;
3176                 default:
3177                         goto err_clear;
3178                 }
3179         }
3180
3181         to->tunnel_id = be64_to_cpu(info->key.tun_id);
3182         to->tunnel_tos = info->key.tos;
3183         to->tunnel_ttl = info->key.ttl;
3184
3185         if (flags & BPF_F_TUNINFO_IPV6) {
3186                 memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
3187                        sizeof(to->remote_ipv6));
3188                 to->tunnel_label = be32_to_cpu(info->key.label);
3189         } else {
3190                 to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
3191         }
3192
3193         if (unlikely(size != sizeof(struct bpf_tunnel_key)))
3194                 memcpy(to_orig, to, size);
3195
3196         return 0;
3197 err_clear:
3198         memset(to_orig, 0, size);
3199         return err;
3200 }
3201
3202 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
3203         .func           = bpf_skb_get_tunnel_key,
3204         .gpl_only       = false,
3205         .ret_type       = RET_INTEGER,
3206         .arg1_type      = ARG_PTR_TO_CTX,
3207         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
3208         .arg3_type      = ARG_CONST_SIZE,
3209         .arg4_type      = ARG_ANYTHING,
3210 };
3211
3212 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
3213 {
3214         const struct ip_tunnel_info *info = skb_tunnel_info(skb);
3215         int err;
3216
3217         if (unlikely(!info ||
3218                      !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
3219                 err = -ENOENT;
3220                 goto err_clear;
3221         }
3222         if (unlikely(size < info->options_len)) {
3223                 err = -ENOMEM;
3224                 goto err_clear;
3225         }
3226
3227         ip_tunnel_info_opts_get(to, info);
3228         if (size > info->options_len)
3229                 memset(to + info->options_len, 0, size - info->options_len);
3230
3231         return info->options_len;
3232 err_clear:
3233         memset(to, 0, size);
3234         return err;
3235 }
3236
3237 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
3238         .func           = bpf_skb_get_tunnel_opt,
3239         .gpl_only       = false,
3240         .ret_type       = RET_INTEGER,
3241         .arg1_type      = ARG_PTR_TO_CTX,
3242         .arg2_type      = ARG_PTR_TO_UNINIT_MEM,
3243         .arg3_type      = ARG_CONST_SIZE,
3244 };
3245
3246 static struct metadata_dst __percpu *md_dst;
3247
3248 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
3249            const struct bpf_tunnel_key *, from, u32, size, u64, flags)
3250 {
3251         struct metadata_dst *md = this_cpu_ptr(md_dst);
3252         u8 compat[sizeof(struct bpf_tunnel_key)];
3253         struct ip_tunnel_info *info;
3254
3255         if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
3256                                BPF_F_DONT_FRAGMENT | BPF_F_SEQ_NUMBER)))
3257                 return -EINVAL;
3258         if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
3259                 switch (size) {
3260                 case offsetof(struct bpf_tunnel_key, tunnel_label):
3261                 case offsetof(struct bpf_tunnel_key, tunnel_ext):
3262                 case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
3263                         /* Fixup deprecated structure layouts here, so we have
3264                          * a common path later on.
3265                          */
3266                         memcpy(compat, from, size);
3267                         memset(compat + size, 0, sizeof(compat) - size);
3268                         from = (const struct bpf_tunnel_key *) compat;
3269                         break;
3270                 default:
3271                         return -EINVAL;
3272                 }
3273         }
3274         if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
3275                      from->tunnel_ext))
3276                 return -EINVAL;
3277
3278         skb_dst_drop(skb);
3279         dst_hold((struct dst_entry *) md);
3280         skb_dst_set(skb, (struct dst_entry *) md);
3281
3282         info = &md->u.tun_info;
3283         info->mode = IP_TUNNEL_INFO_TX;
3284
3285         info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
3286         if (flags & BPF_F_DONT_FRAGMENT)
3287                 info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
3288         if (flags & BPF_F_ZERO_CSUM_TX)
3289                 info->key.tun_flags &= ~TUNNEL_CSUM;
3290         if (flags & BPF_F_SEQ_NUMBER)
3291                 info->key.tun_flags |= TUNNEL_SEQ;
3292
3293         info->key.tun_id = cpu_to_be64(from->tunnel_id);
3294         info->key.tos = from->tunnel_tos;
3295         info->key.ttl = from->tunnel_ttl;
3296
3297         if (flags & BPF_F_TUNINFO_IPV6) {
3298                 info->mode |= IP_TUNNEL_INFO_IPV6;
3299                 memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
3300                        sizeof(from->remote_ipv6));
3301                 info->key.label = cpu_to_be32(from->tunnel_label) &
3302                                   IPV6_FLOWLABEL_MASK;
3303         } else {
3304                 info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
3305         }
3306
3307         return 0;
3308 }
3309
3310 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
3311         .func           = bpf_skb_set_tunnel_key,
3312         .gpl_only       = false,
3313         .ret_type       = RET_INTEGER,
3314         .arg1_type      = ARG_PTR_TO_CTX,
3315         .arg2_type      = ARG_PTR_TO_MEM,
3316         .arg3_type      = ARG_CONST_SIZE,
3317         .arg4_type      = ARG_ANYTHING,
3318 };
3319
3320 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
3321            const u8 *, from, u32, size)
3322 {
3323         struct ip_tunnel_info *info = skb_tunnel_info(skb);
3324         const struct metadata_dst *md = this_cpu_ptr(md_dst);
3325
3326         if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
3327                 return -EINVAL;
3328         if (unlikely(size > IP_TUNNEL_OPTS_MAX))
3329                 return -ENOMEM;
3330
3331         ip_tunnel_info_opts_set(info, from, size);
3332
3333         return 0;
3334 }
3335
3336 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
3337         .func           = bpf_skb_set_tunnel_opt,
3338         .gpl_only       = false,
3339         .ret_type       = RET_INTEGER,
3340         .arg1_type      = ARG_PTR_TO_CTX,
3341         .arg2_type      = ARG_PTR_TO_MEM,
3342         .arg3_type      = ARG_CONST_SIZE,
3343 };
3344
3345 static const struct bpf_func_proto *
3346 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
3347 {
3348         if (!md_dst) {
3349                 struct metadata_dst __percpu *tmp;
3350
3351                 tmp = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
3352                                                 METADATA_IP_TUNNEL,
3353                                                 GFP_KERNEL);
3354                 if (!tmp)
3355                         return NULL;
3356                 if (cmpxchg(&md_dst, NULL, tmp))
3357                         metadata_dst_free_percpu(tmp);
3358         }
3359
3360         switch (which) {
3361         case BPF_FUNC_skb_set_tunnel_key:
3362                 return &bpf_skb_set_tunnel_key_proto;
3363         case BPF_FUNC_skb_set_tunnel_opt:
3364                 return &bpf_skb_set_tunnel_opt_proto;
3365         default:
3366                 return NULL;
3367         }
3368 }
3369
3370 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
3371            u32, idx)
3372 {
3373         struct bpf_array *array = container_of(map, struct bpf_array, map);
3374         struct cgroup *cgrp;
3375         struct sock *sk;
3376
3377         sk = skb_to_full_sk(skb);
3378         if (!sk || !sk_fullsock(sk))
3379                 return -ENOENT;
3380         if (unlikely(idx >= array->map.max_entries))
3381                 return -E2BIG;
3382
3383         cgrp = READ_ONCE(array->ptrs[idx]);
3384         if (unlikely(!cgrp))
3385                 return -EAGAIN;
3386
3387         return sk_under_cgroup_hierarchy(sk, cgrp);
3388 }
3389
3390 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
3391         .func           = bpf_skb_under_cgroup,
3392         .gpl_only       = false,
3393         .ret_type       = RET_INTEGER,
3394         .arg1_type      = ARG_PTR_TO_CTX,
3395         .arg2_type      = ARG_CONST_MAP_PTR,
3396         .arg3_type      = ARG_ANYTHING,
3397 };
3398
3399 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
3400                                   unsigned long off, unsigned long len)
3401 {
3402         memcpy(dst_buff, src_buff + off, len);
3403         return 0;
3404 }
3405
3406 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
3407            u64, flags, void *, meta, u64, meta_size)
3408 {
3409         u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
3410
3411         if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
3412                 return -EINVAL;
3413         if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
3414                 return -EFAULT;
3415
3416         return bpf_event_output(map, flags, meta, meta_size, xdp->data,
3417                                 xdp_size, bpf_xdp_copy);
3418 }
3419
3420 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
3421         .func           = bpf_xdp_event_output,
3422         .gpl_only       = true,
3423         .ret_type       = RET_INTEGER,
3424         .arg1_type      = ARG_PTR_TO_CTX,
3425         .arg2_type      = ARG_CONST_MAP_PTR,
3426         .arg3_type      = ARG_ANYTHING,
3427         .arg4_type      = ARG_PTR_TO_MEM,
3428         .arg5_type      = ARG_CONST_SIZE_OR_ZERO,
3429 };
3430
3431 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
3432 {
3433         return skb->sk ? sock_gen_cookie(skb->sk) : 0;
3434 }
3435
3436 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
3437         .func           = bpf_get_socket_cookie,
3438         .gpl_only       = false,
3439         .ret_type       = RET_INTEGER,
3440         .arg1_type      = ARG_PTR_TO_CTX,
3441 };
3442
3443 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
3444 {
3445         struct sock *sk = sk_to_full_sk(skb->sk);
3446         kuid_t kuid;
3447
3448         if (!sk || !sk_fullsock(sk))
3449                 return overflowuid;
3450         kuid = sock_net_uid(sock_net(sk), sk);
3451         return from_kuid_munged(sock_net(sk)->user_ns, kuid);
3452 }
3453
3454 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
3455         .func           = bpf_get_socket_uid,
3456         .gpl_only       = false,
3457         .ret_type       = RET_INTEGER,
3458         .arg1_type      = ARG_PTR_TO_CTX,
3459 };
3460
3461 BPF_CALL_5(bpf_setsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3462            int, level, int, optname, char *, optval, int, optlen)
3463 {
3464         struct sock *sk = bpf_sock->sk;
3465         int ret = 0;
3466         int val;
3467
3468         if (!sk_fullsock(sk))
3469                 return -EINVAL;
3470
3471         if (level == SOL_SOCKET) {
3472                 if (optlen != sizeof(int))
3473                         return -EINVAL;
3474                 val = *((int *)optval);
3475
3476                 /* Only some socketops are supported */
3477                 switch (optname) {
3478                 case SO_RCVBUF:
3479                         sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
3480                         sk->sk_rcvbuf = max_t(int, val * 2, SOCK_MIN_RCVBUF);
3481                         break;
3482                 case SO_SNDBUF:
3483                         sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
3484                         sk->sk_sndbuf = max_t(int, val * 2, SOCK_MIN_SNDBUF);
3485                         break;
3486                 case SO_MAX_PACING_RATE:
3487                         sk->sk_max_pacing_rate = val;
3488                         sk->sk_pacing_rate = min(sk->sk_pacing_rate,
3489                                                  sk->sk_max_pacing_rate);
3490                         break;
3491                 case SO_PRIORITY:
3492                         sk->sk_priority = val;
3493                         break;
3494                 case SO_RCVLOWAT:
3495                         if (val < 0)
3496                                 val = INT_MAX;
3497                         sk->sk_rcvlowat = val ? : 1;
3498                         break;
3499                 case SO_MARK:
3500                         sk->sk_mark = val;
3501                         break;
3502                 default:
3503                         ret = -EINVAL;
3504                 }
3505 #ifdef CONFIG_INET
3506         } else if (level == SOL_IP) {
3507                 if (optlen != sizeof(int) || sk->sk_family != AF_INET)
3508                         return -EINVAL;
3509
3510                 val = *((int *)optval);
3511                 /* Only some options are supported */
3512                 switch (optname) {
3513                 case IP_TOS:
3514                         if (val < -1 || val > 0xff) {
3515                                 ret = -EINVAL;
3516                         } else {
3517                                 struct inet_sock *inet = inet_sk(sk);
3518
3519                                 if (val == -1)
3520                                         val = 0;
3521                                 inet->tos = val;
3522                         }
3523                         break;
3524                 default:
3525                         ret = -EINVAL;
3526                 }
3527 #if IS_ENABLED(CONFIG_IPV6)
3528         } else if (level == SOL_IPV6) {
3529                 if (optlen != sizeof(int) || sk->sk_family != AF_INET6)
3530                         return -EINVAL;
3531
3532                 val = *((int *)optval);
3533                 /* Only some options are supported */
3534                 switch (optname) {
3535                 case IPV6_TCLASS:
3536                         if (val < -1 || val > 0xff) {
3537                                 ret = -EINVAL;
3538                         } else {
3539                                 struct ipv6_pinfo *np = inet6_sk(sk);
3540
3541                                 if (val == -1)
3542                                         val = 0;
3543                                 np->tclass = val;
3544                         }
3545                         break;
3546                 default:
3547                         ret = -EINVAL;
3548                 }
3549 #endif
3550         } else if (level == SOL_TCP &&
3551                    sk->sk_prot->setsockopt == tcp_setsockopt) {
3552                 if (optname == TCP_CONGESTION) {
3553                         char name[TCP_CA_NAME_MAX];
3554                         bool reinit = bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN;
3555
3556                         strncpy(name, optval, min_t(long, optlen,
3557                                                     TCP_CA_NAME_MAX-1));
3558                         name[TCP_CA_NAME_MAX-1] = 0;
3559                         ret = tcp_set_congestion_control(sk, name, false,
3560                                                          reinit);
3561                 } else {
3562                         struct tcp_sock *tp = tcp_sk(sk);
3563
3564                         if (optlen != sizeof(int))
3565                                 return -EINVAL;
3566
3567                         val = *((int *)optval);
3568                         /* Only some options are supported */
3569                         switch (optname) {
3570                         case TCP_BPF_IW:
3571                                 if (val <= 0 || tp->data_segs_out > 0)
3572                                         ret = -EINVAL;
3573                                 else
3574                                         tp->snd_cwnd = val;
3575                                 break;
3576                         case TCP_BPF_SNDCWND_CLAMP:
3577                                 if (val <= 0) {
3578                                         ret = -EINVAL;
3579                                 } else {
3580                                         tp->snd_cwnd_clamp = val;
3581                                         tp->snd_ssthresh = val;
3582                                 }
3583                                 break;
3584                         default:
3585                                 ret = -EINVAL;
3586                         }
3587                 }
3588 #endif
3589         } else {
3590                 ret = -EINVAL;
3591         }
3592         return ret;
3593 }
3594
3595 static const struct bpf_func_proto bpf_setsockopt_proto = {
3596         .func           = bpf_setsockopt,
3597         .gpl_only       = false,
3598         .ret_type       = RET_INTEGER,
3599         .arg1_type      = ARG_PTR_TO_CTX,
3600         .arg2_type      = ARG_ANYTHING,
3601         .arg3_type      = ARG_ANYTHING,
3602         .arg4_type      = ARG_PTR_TO_MEM,
3603         .arg5_type      = ARG_CONST_SIZE,
3604 };
3605
3606 BPF_CALL_5(bpf_getsockopt, struct bpf_sock_ops_kern *, bpf_sock,
3607            int, level, int, optname, char *, optval, int, optlen)
3608 {
3609         struct sock *sk = bpf_sock->sk;
3610
3611         if (!sk_fullsock(sk))
3612                 goto err_clear;
3613
3614 #ifdef CONFIG_INET
3615         if (level == SOL_TCP && sk->sk_prot->getsockopt == tcp_getsockopt) {
3616                 if (optname == TCP_CONGESTION) {
3617                         struct inet_connection_sock *icsk = inet_csk(sk);
3618
3619                         if (!icsk->icsk_ca_ops || optlen <= 1)
3620                                 goto err_clear;
3621                         strncpy(optval, icsk->icsk_ca_ops->name, optlen);
3622                         optval[optlen - 1] = 0;
3623                 } else {
3624                         goto err_clear;
3625                 }
3626         } else if (level == SOL_IP) {
3627                 struct inet_sock *inet = inet_sk(sk);
3628
3629                 if (optlen != sizeof(int) || sk->sk_family != AF_INET)
3630                         goto err_clear;
3631
3632                 /* Only some options are supported */
3633                 switch (optname) {
3634                 case IP_TOS:
3635                         *((int *)optval) = (int)inet->tos;
3636                         break;
3637                 default:
3638                         goto err_clear;
3639                 }
3640 #if IS_ENABLED(CONFIG_IPV6)
3641         } else if (level == SOL_IPV6) {
3642                 struct ipv6_pinfo *np = inet6_sk(sk);
3643
3644                 if (optlen != sizeof(int) || sk->sk_family != AF_INET6)
3645                         goto err_clear;
3646
3647                 /* Only some options are supported */
3648                 switch (optname) {
3649                 case IPV6_TCLASS:
3650                         *((int *)optval) = (int)np->tclass;
3651                         break;
3652                 default:
3653                         goto err_clear;
3654                 }
3655 #endif
3656         } else {
3657                 goto err_clear;
3658         }
3659         return 0;
3660 #endif
3661 err_clear:
3662         memset(optval, 0, optlen);
3663         return -EINVAL;
3664 }
3665
3666 static const struct bpf_func_proto bpf_getsockopt_proto = {
3667         .func           = bpf_getsockopt,
3668         .gpl_only       = false,
3669         .ret_type       = RET_INTEGER,
3670         .arg1_type      = ARG_PTR_TO_CTX,
3671         .arg2_type      = ARG_ANYTHING,
3672         .arg3_type      = ARG_ANYTHING,
3673         .arg4_type      = ARG_PTR_TO_UNINIT_MEM,
3674         .arg5_type      = ARG_CONST_SIZE,
3675 };
3676
3677 BPF_CALL_2(bpf_sock_ops_cb_flags_set, struct bpf_sock_ops_kern *, bpf_sock,
3678            int, argval)
3679 {
3680         struct sock *sk = bpf_sock->sk;
3681         int val = argval & BPF_SOCK_OPS_ALL_CB_FLAGS;
3682
3683         if (!IS_ENABLED(CONFIG_INET) || !sk_fullsock(sk))
3684                 return -EINVAL;
3685
3686         if (val)
3687                 tcp_sk(sk)->bpf_sock_ops_cb_flags = val;
3688
3689         return argval & (~BPF_SOCK_OPS_ALL_CB_FLAGS);
3690 }
3691
3692 static const struct bpf_func_proto bpf_sock_ops_cb_flags_set_proto = {
3693         .func           = bpf_sock_ops_cb_flags_set,
3694         .gpl_only       = false,
3695         .ret_type       = RET_INTEGER,
3696         .arg1_type      = ARG_PTR_TO_CTX,
3697         .arg2_type      = ARG_ANYTHING,
3698 };
3699
3700 const struct ipv6_bpf_stub *ipv6_bpf_stub __read_mostly;
3701 EXPORT_SYMBOL_GPL(ipv6_bpf_stub);
3702
3703 BPF_CALL_3(bpf_bind, struct bpf_sock_addr_kern *, ctx, struct sockaddr *, addr,
3704            int, addr_len)
3705 {
3706 #ifdef CONFIG_INET
3707         struct sock *sk = ctx->sk;
3708         int err;
3709
3710         /* Binding to port can be expensive so it's prohibited in the helper.
3711          * Only binding to IP is supported.
3712          */
3713         err = -EINVAL;
3714         if (addr->sa_family == AF_INET) {
3715                 if (addr_len < sizeof(struct sockaddr_in))
3716                         return err;
3717                 if (((struct sockaddr_in *)addr)->sin_port != htons(0))
3718                         return err;
3719                 return __inet_bind(sk, addr, addr_len, true, false);
3720 #if IS_ENABLED(CONFIG_IPV6)
3721         } else if (addr->sa_family == AF_INET6) {
3722                 if (addr_len < SIN6_LEN_RFC2133)
3723                         return err;
3724                 if (((struct sockaddr_in6 *)addr)->sin6_port != htons(0))
3725                         return err;
3726                 /* ipv6_bpf_stub cannot be NULL, since it's called from
3727                  * bpf_cgroup_inet6_connect hook and ipv6 is already loaded
3728                  */
3729                 return ipv6_bpf_stub->inet6_bind(sk, addr, addr_len, true, false);
3730 #endif /* CONFIG_IPV6 */
3731         }
3732 #endif /* CONFIG_INET */
3733
3734         return -EAFNOSUPPORT;
3735 }
3736
3737 static const struct bpf_func_proto bpf_bind_proto = {
3738         .func           = bpf_bind,
3739         .gpl_only       = false,
3740         .ret_type       = RET_INTEGER,
3741         .arg1_type      = ARG_PTR_TO_CTX,
3742         .arg2_type      = ARG_PTR_TO_MEM,
3743         .arg3_type      = ARG_CONST_SIZE,
3744 };
3745
3746 static const struct bpf_func_proto *
3747 bpf_base_func_proto(enum bpf_func_id func_id)
3748 {
3749         switch (func_id) {
3750         case BPF_FUNC_map_lookup_elem:
3751                 return &bpf_map_lookup_elem_proto;
3752         case BPF_FUNC_map_update_elem:
3753                 return &bpf_map_update_elem_proto;
3754         case BPF_FUNC_map_delete_elem:
3755                 return &bpf_map_delete_elem_proto;
3756         case BPF_FUNC_get_prandom_u32:
3757                 return &bpf_get_prandom_u32_proto;
3758         case BPF_FUNC_get_smp_processor_id:
3759                 return &bpf_get_raw_smp_processor_id_proto;
3760         case BPF_FUNC_get_numa_node_id:
3761                 return &bpf_get_numa_node_id_proto;
3762         case BPF_FUNC_tail_call:
3763                 return &bpf_tail_call_proto;
3764         case BPF_FUNC_ktime_get_ns:
3765                 return &bpf_ktime_get_ns_proto;
3766         case BPF_FUNC_trace_printk:
3767                 if (capable(CAP_SYS_ADMIN))
3768                         return bpf_get_trace_printk_proto();
3769         default:
3770                 return NULL;
3771         }
3772 }
3773
3774 static const struct bpf_func_proto *
3775 sock_filter_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3776 {
3777         switch (func_id) {
3778         /* inet and inet6 sockets are created in a process
3779          * context so there is always a valid uid/gid
3780          */
3781         case BPF_FUNC_get_current_uid_gid:
3782                 return &bpf_get_current_uid_gid_proto;
3783         default:
3784                 return bpf_base_func_proto(func_id);
3785         }
3786 }
3787
3788 static const struct bpf_func_proto *
3789 sock_addr_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3790 {
3791         switch (func_id) {
3792         /* inet and inet6 sockets are created in a process
3793          * context so there is always a valid uid/gid
3794          */
3795         case BPF_FUNC_get_current_uid_gid:
3796                 return &bpf_get_current_uid_gid_proto;
3797         case BPF_FUNC_bind:
3798                 switch (prog->expected_attach_type) {
3799                 case BPF_CGROUP_INET4_CONNECT:
3800                 case BPF_CGROUP_INET6_CONNECT:
3801                         return &bpf_bind_proto;
3802                 default:
3803                         return NULL;
3804                 }
3805         default:
3806                 return bpf_base_func_proto(func_id);
3807         }
3808 }
3809
3810 static const struct bpf_func_proto *
3811 sk_filter_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3812 {
3813         switch (func_id) {
3814         case BPF_FUNC_skb_load_bytes:
3815                 return &bpf_skb_load_bytes_proto;
3816         case BPF_FUNC_get_socket_cookie:
3817                 return &bpf_get_socket_cookie_proto;
3818         case BPF_FUNC_get_socket_uid:
3819                 return &bpf_get_socket_uid_proto;
3820         default:
3821                 return bpf_base_func_proto(func_id);
3822         }
3823 }
3824
3825 static const struct bpf_func_proto *
3826 tc_cls_act_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3827 {
3828         switch (func_id) {
3829         case BPF_FUNC_skb_store_bytes:
3830                 return &bpf_skb_store_bytes_proto;
3831         case BPF_FUNC_skb_load_bytes:
3832                 return &bpf_skb_load_bytes_proto;
3833         case BPF_FUNC_skb_pull_data:
3834                 return &bpf_skb_pull_data_proto;
3835         case BPF_FUNC_csum_diff:
3836                 return &bpf_csum_diff_proto;
3837         case BPF_FUNC_csum_update:
3838                 return &bpf_csum_update_proto;
3839         case BPF_FUNC_l3_csum_replace:
3840                 return &bpf_l3_csum_replace_proto;
3841         case BPF_FUNC_l4_csum_replace:
3842                 return &bpf_l4_csum_replace_proto;
3843         case BPF_FUNC_clone_redirect:
3844                 return &bpf_clone_redirect_proto;
3845         case BPF_FUNC_get_cgroup_classid:
3846                 return &bpf_get_cgroup_classid_proto;
3847         case BPF_FUNC_skb_vlan_push:
3848                 return &bpf_skb_vlan_push_proto;
3849         case BPF_FUNC_skb_vlan_pop:
3850                 return &bpf_skb_vlan_pop_proto;
3851         case BPF_FUNC_skb_change_proto:
3852                 return &bpf_skb_change_proto_proto;
3853         case BPF_FUNC_skb_change_type:
3854                 return &bpf_skb_change_type_proto;
3855         case BPF_FUNC_skb_adjust_room:
3856                 return &bpf_skb_adjust_room_proto;
3857         case BPF_FUNC_skb_change_tail:
3858                 return &bpf_skb_change_tail_proto;
3859         case BPF_FUNC_skb_get_tunnel_key:
3860                 return &bpf_skb_get_tunnel_key_proto;
3861         case BPF_FUNC_skb_set_tunnel_key:
3862                 return bpf_get_skb_set_tunnel_proto(func_id);
3863         case BPF_FUNC_skb_get_tunnel_opt:
3864                 return &bpf_skb_get_tunnel_opt_proto;
3865         case BPF_FUNC_skb_set_tunnel_opt:
3866                 return bpf_get_skb_set_tunnel_proto(func_id);
3867         case BPF_FUNC_redirect:
3868                 return &bpf_redirect_proto;
3869         case BPF_FUNC_get_route_realm:
3870                 return &bpf_get_route_realm_proto;
3871         case BPF_FUNC_get_hash_recalc:
3872                 return &bpf_get_hash_recalc_proto;
3873         case BPF_FUNC_set_hash_invalid:
3874                 return &bpf_set_hash_invalid_proto;
3875         case BPF_FUNC_set_hash:
3876                 return &bpf_set_hash_proto;
3877         case BPF_FUNC_perf_event_output:
3878                 return &bpf_skb_event_output_proto;
3879         case BPF_FUNC_get_smp_processor_id:
3880                 return &bpf_get_smp_processor_id_proto;
3881         case BPF_FUNC_skb_under_cgroup:
3882                 return &bpf_skb_under_cgroup_proto;
3883         case BPF_FUNC_get_socket_cookie:
3884                 return &bpf_get_socket_cookie_proto;
3885         case BPF_FUNC_get_socket_uid:
3886                 return &bpf_get_socket_uid_proto;
3887         default:
3888                 return bpf_base_func_proto(func_id);
3889         }
3890 }
3891
3892 static const struct bpf_func_proto *
3893 xdp_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3894 {
3895         switch (func_id) {
3896         case BPF_FUNC_perf_event_output:
3897                 return &bpf_xdp_event_output_proto;
3898         case BPF_FUNC_get_smp_processor_id:
3899                 return &bpf_get_smp_processor_id_proto;
3900         case BPF_FUNC_csum_diff:
3901                 return &bpf_csum_diff_proto;
3902         case BPF_FUNC_xdp_adjust_head:
3903                 return &bpf_xdp_adjust_head_proto;
3904         case BPF_FUNC_xdp_adjust_meta:
3905                 return &bpf_xdp_adjust_meta_proto;
3906         case BPF_FUNC_redirect:
3907                 return &bpf_xdp_redirect_proto;
3908         case BPF_FUNC_redirect_map:
3909                 return &bpf_xdp_redirect_map_proto;
3910         case BPF_FUNC_xdp_adjust_tail:
3911                 return &bpf_xdp_adjust_tail_proto;
3912         default:
3913                 return bpf_base_func_proto(func_id);
3914         }
3915 }
3916
3917 static const struct bpf_func_proto *
3918 lwt_inout_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3919 {
3920         switch (func_id) {
3921         case BPF_FUNC_skb_load_bytes:
3922                 return &bpf_skb_load_bytes_proto;
3923         case BPF_FUNC_skb_pull_data:
3924                 return &bpf_skb_pull_data_proto;
3925         case BPF_FUNC_csum_diff:
3926                 return &bpf_csum_diff_proto;
3927         case BPF_FUNC_get_cgroup_classid:
3928                 return &bpf_get_cgroup_classid_proto;
3929         case BPF_FUNC_get_route_realm:
3930                 return &bpf_get_route_realm_proto;
3931         case BPF_FUNC_get_hash_recalc:
3932                 return &bpf_get_hash_recalc_proto;
3933         case BPF_FUNC_perf_event_output:
3934                 return &bpf_skb_event_output_proto;
3935         case BPF_FUNC_get_smp_processor_id:
3936                 return &bpf_get_smp_processor_id_proto;
3937         case BPF_FUNC_skb_under_cgroup:
3938                 return &bpf_skb_under_cgroup_proto;
3939         default:
3940                 return bpf_base_func_proto(func_id);
3941         }
3942 }
3943
3944 static const struct bpf_func_proto *
3945 sock_ops_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3946 {
3947         switch (func_id) {
3948         case BPF_FUNC_setsockopt:
3949                 return &bpf_setsockopt_proto;
3950         case BPF_FUNC_getsockopt:
3951                 return &bpf_getsockopt_proto;
3952         case BPF_FUNC_sock_ops_cb_flags_set:
3953                 return &bpf_sock_ops_cb_flags_set_proto;
3954         case BPF_FUNC_sock_map_update:
3955                 return &bpf_sock_map_update_proto;
3956         default:
3957                 return bpf_base_func_proto(func_id);
3958         }
3959 }
3960
3961 static const struct bpf_func_proto *
3962 sk_msg_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3963 {
3964         switch (func_id) {
3965         case BPF_FUNC_msg_redirect_map:
3966                 return &bpf_msg_redirect_map_proto;
3967         case BPF_FUNC_msg_apply_bytes:
3968                 return &bpf_msg_apply_bytes_proto;
3969         case BPF_FUNC_msg_cork_bytes:
3970                 return &bpf_msg_cork_bytes_proto;
3971         case BPF_FUNC_msg_pull_data:
3972                 return &bpf_msg_pull_data_proto;
3973         default:
3974                 return bpf_base_func_proto(func_id);
3975         }
3976 }
3977
3978 static const struct bpf_func_proto *
3979 sk_skb_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
3980 {
3981         switch (func_id) {
3982         case BPF_FUNC_skb_store_bytes:
3983                 return &bpf_skb_store_bytes_proto;
3984         case BPF_FUNC_skb_load_bytes:
3985                 return &bpf_skb_load_bytes_proto;
3986         case BPF_FUNC_skb_pull_data:
3987                 return &bpf_skb_pull_data_proto;
3988         case BPF_FUNC_skb_change_tail:
3989                 return &bpf_skb_change_tail_proto;
3990         case BPF_FUNC_skb_change_head:
3991                 return &bpf_skb_change_head_proto;
3992         case BPF_FUNC_get_socket_cookie:
3993                 return &bpf_get_socket_cookie_proto;
3994         case BPF_FUNC_get_socket_uid:
3995                 return &bpf_get_socket_uid_proto;
3996         case BPF_FUNC_sk_redirect_map:
3997                 return &bpf_sk_redirect_map_proto;
3998         default:
3999                 return bpf_base_func_proto(func_id);
4000         }
4001 }
4002
4003 static const struct bpf_func_proto *
4004 lwt_xmit_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
4005 {
4006         switch (func_id) {
4007         case BPF_FUNC_skb_get_tunnel_key:
4008                 return &bpf_skb_get_tunnel_key_proto;
4009         case BPF_FUNC_skb_set_tunnel_key:
4010                 return bpf_get_skb_set_tunnel_proto(func_id);
4011         case BPF_FUNC_skb_get_tunnel_opt:
4012                 return &bpf_skb_get_tunnel_opt_proto;
4013         case BPF_FUNC_skb_set_tunnel_opt:
4014                 return bpf_get_skb_set_tunnel_proto(func_id);
4015         case BPF_FUNC_redirect:
4016                 return &bpf_redirect_proto;
4017         case BPF_FUNC_clone_redirect:
4018                 return &bpf_clone_redirect_proto;
4019         case BPF_FUNC_skb_change_tail:
4020                 return &bpf_skb_change_tail_proto;
4021         case BPF_FUNC_skb_change_head:
4022                 return &bpf_skb_change_head_proto;
4023         case BPF_FUNC_skb_store_bytes:
4024                 return &bpf_skb_store_bytes_proto;
4025         case BPF_FUNC_csum_update:
4026                 return &bpf_csum_update_proto;
4027         case BPF_FUNC_l3_csum_replace:
4028                 return &bpf_l3_csum_replace_proto;
4029         case BPF_FUNC_l4_csum_replace:
4030                 return &bpf_l4_csum_replace_proto;
4031         case BPF_FUNC_set_hash_invalid:
4032                 return &bpf_set_hash_invalid_proto;
4033         default:
4034                 return lwt_inout_func_proto(func_id, prog);
4035         }
4036 }
4037
4038 static bool bpf_skb_is_valid_access(int off, int size, enum bpf_access_type type,
4039                                     const struct bpf_prog *prog,
4040                                     struct bpf_insn_access_aux *info)
4041 {
4042         const int size_default = sizeof(__u32);
4043
4044         if (off < 0 || off >= sizeof(struct __sk_buff))
4045                 return false;
4046
4047         /* The verifier guarantees that size > 0. */
4048         if (off % size != 0)
4049                 return false;
4050
4051         switch (off) {
4052         case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
4053                 if (off + size > offsetofend(struct __sk_buff, cb[4]))
4054                         return false;
4055                 break;
4056         case bpf_ctx_range_till(struct __sk_buff, remote_ip6[0], remote_ip6[3]):
4057         case bpf_ctx_range_till(struct __sk_buff, local_ip6[0], local_ip6[3]):
4058         case bpf_ctx_range_till(struct __sk_buff, remote_ip4, remote_ip4):
4059         case bpf_ctx_range_till(struct __sk_buff, local_ip4, local_ip4):
4060         case bpf_ctx_range(struct __sk_buff, data):
4061         case bpf_ctx_range(struct __sk_buff, data_meta):
4062         case bpf_ctx_range(struct __sk_buff, data_end):
4063                 if (size != size_default)
4064                         return false;
4065                 break;
4066         default:
4067                 /* Only narrow read access allowed for now. */
4068                 if (type == BPF_WRITE) {
4069                         if (size != size_default)
4070                                 return false;
4071                 } else {
4072                         bpf_ctx_record_field_size(info, size_default);
4073                         if (!bpf_ctx_narrow_access_ok(off, size, size_default))
4074                                 return false;
4075                 }
4076         }
4077
4078         return true;
4079 }
4080
4081 static bool sk_filter_is_valid_access(int off, int size,
4082                                       enum bpf_access_type type,
4083                                       const struct bpf_prog *prog,
4084                                       struct bpf_insn_access_aux *info)
4085 {
4086         switch (off) {
4087         case bpf_ctx_range(struct __sk_buff, tc_classid):
4088         case bpf_ctx_range(struct __sk_buff, data):
4089         case bpf_ctx_range(struct __sk_buff, data_meta):
4090         case bpf_ctx_range(struct __sk_buff, data_end):
4091         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
4092                 return false;
4093         }
4094
4095         if (type == BPF_WRITE) {
4096                 switch (off) {
4097                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
4098                         break;
4099                 default:
4100                         return false;
4101                 }
4102         }
4103
4104         return bpf_skb_is_valid_access(off, size, type, prog, info);
4105 }
4106
4107 static bool lwt_is_valid_access(int off, int size,
4108                                 enum bpf_access_type type,
4109                                 const struct bpf_prog *prog,
4110                                 struct bpf_insn_access_aux *info)
4111 {
4112         switch (off) {
4113         case bpf_ctx_range(struct __sk_buff, tc_classid):
4114         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
4115         case bpf_ctx_range(struct __sk_buff, data_meta):
4116                 return false;
4117         }
4118
4119         if (type == BPF_WRITE) {
4120                 switch (off) {
4121                 case bpf_ctx_range(struct __sk_buff, mark):
4122                 case bpf_ctx_range(struct __sk_buff, priority):
4123                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
4124                         break;
4125                 default:
4126                         return false;
4127                 }
4128         }
4129
4130         switch (off) {
4131         case bpf_ctx_range(struct __sk_buff, data):
4132                 info->reg_type = PTR_TO_PACKET;
4133                 break;
4134         case bpf_ctx_range(struct __sk_buff, data_end):
4135                 info->reg_type = PTR_TO_PACKET_END;
4136                 break;
4137         }
4138
4139         return bpf_skb_is_valid_access(off, size, type, prog, info);
4140 }
4141
4142
4143 /* Attach type specific accesses */
4144 static bool __sock_filter_check_attach_type(int off,
4145                                             enum bpf_access_type access_type,
4146                                             enum bpf_attach_type attach_type)
4147 {
4148         switch (off) {
4149         case offsetof(struct bpf_sock, bound_dev_if):
4150         case offsetof(struct bpf_sock, mark):
4151         case offsetof(struct bpf_sock, priority):
4152                 switch (attach_type) {
4153                 case BPF_CGROUP_INET_SOCK_CREATE:
4154                         goto full_access;
4155                 default:
4156                         return false;
4157                 }
4158         case bpf_ctx_range(struct bpf_sock, src_ip4):
4159                 switch (attach_type) {
4160                 case BPF_CGROUP_INET4_POST_BIND:
4161                         goto read_only;
4162                 default:
4163                         return false;
4164                 }
4165         case bpf_ctx_range_till(struct bpf_sock, src_ip6[0], src_ip6[3]):
4166                 switch (attach_type) {
4167                 case BPF_CGROUP_INET6_POST_BIND:
4168                         goto read_only;
4169                 default:
4170                         return false;
4171                 }
4172         case bpf_ctx_range(struct bpf_sock, src_port):
4173                 switch (attach_type) {
4174                 case BPF_CGROUP_INET4_POST_BIND:
4175                 case BPF_CGROUP_INET6_POST_BIND:
4176                         goto read_only;
4177                 default:
4178                         return false;
4179                 }
4180         }
4181 read_only:
4182         return access_type == BPF_READ;
4183 full_access:
4184         return true;
4185 }
4186
4187 static bool __sock_filter_check_size(int off, int size,
4188                                      struct bpf_insn_access_aux *info)
4189 {
4190         const int size_default = sizeof(__u32);
4191
4192         switch (off) {
4193         case bpf_ctx_range(struct bpf_sock, src_ip4):
4194         case bpf_ctx_range_till(struct bpf_sock, src_ip6[0], src_ip6[3]):
4195                 bpf_ctx_record_field_size(info, size_default);
4196                 return bpf_ctx_narrow_access_ok(off, size, size_default);
4197         }
4198
4199         return size == size_default;
4200 }
4201
4202 static bool sock_filter_is_valid_access(int off, int size,
4203                                         enum bpf_access_type type,
4204                                         const struct bpf_prog *prog,
4205                                         struct bpf_insn_access_aux *info)
4206 {
4207         if (off < 0 || off >= sizeof(struct bpf_sock))
4208                 return false;
4209         if (off % size != 0)
4210                 return false;
4211         if (!__sock_filter_check_attach_type(off, type,
4212                                              prog->expected_attach_type))
4213                 return false;
4214         if (!__sock_filter_check_size(off, size, info))
4215                 return false;
4216         return true;
4217 }
4218
4219 static int bpf_unclone_prologue(struct bpf_insn *insn_buf, bool direct_write,
4220                                 const struct bpf_prog *prog, int drop_verdict)
4221 {
4222         struct bpf_insn *insn = insn_buf;
4223
4224         if (!direct_write)
4225                 return 0;
4226
4227         /* if (!skb->cloned)
4228          *       goto start;
4229          *
4230          * (Fast-path, otherwise approximation that we might be
4231          *  a clone, do the rest in helper.)
4232          */
4233         *insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
4234         *insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
4235         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
4236
4237         /* ret = bpf_skb_pull_data(skb, 0); */
4238         *insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
4239         *insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
4240         *insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
4241                                BPF_FUNC_skb_pull_data);
4242         /* if (!ret)
4243          *      goto restore;
4244          * return TC_ACT_SHOT;
4245          */
4246         *insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
4247         *insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, drop_verdict);
4248         *insn++ = BPF_EXIT_INSN();
4249
4250         /* restore: */
4251         *insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
4252         /* start: */
4253         *insn++ = prog->insnsi[0];
4254
4255         return insn - insn_buf;
4256 }
4257
4258 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
4259                                const struct bpf_prog *prog)
4260 {
4261         return bpf_unclone_prologue(insn_buf, direct_write, prog, TC_ACT_SHOT);
4262 }
4263
4264 static bool tc_cls_act_is_valid_access(int off, int size,
4265                                        enum bpf_access_type type,
4266                                        const struct bpf_prog *prog,
4267                                        struct bpf_insn_access_aux *info)
4268 {
4269         if (type == BPF_WRITE) {
4270                 switch (off) {
4271                 case bpf_ctx_range(struct __sk_buff, mark):
4272                 case bpf_ctx_range(struct __sk_buff, tc_index):
4273                 case bpf_ctx_range(struct __sk_buff, priority):
4274                 case bpf_ctx_range(struct __sk_buff, tc_classid):
4275                 case bpf_ctx_range_till(struct __sk_buff, cb[0], cb[4]):
4276                         break;
4277                 default:
4278                         return false;
4279                 }
4280         }
4281
4282         switch (off) {
4283         case bpf_ctx_range(struct __sk_buff, data):
4284                 info->reg_type = PTR_TO_PACKET;
4285                 break;
4286         case bpf_ctx_range(struct __sk_buff, data_meta):
4287                 info->reg_type = PTR_TO_PACKET_META;
4288                 break;
4289         case bpf_ctx_range(struct __sk_buff, data_end):
4290                 info->reg_type = PTR_TO_PACKET_END;
4291                 break;
4292         case bpf_ctx_range_till(struct __sk_buff, family, local_port):
4293                 return false;
4294         }
4295
4296         return bpf_skb_is_valid_access(off, size, type, prog, info);
4297 }
4298
4299 static bool __is_valid_xdp_access(int off, int size)
4300 {
4301         if (off < 0 || off >= sizeof(struct xdp_md))
4302                 return false;
4303         if (off % size != 0)
4304                 return false;
4305         if (size != sizeof(__u32))
4306                 return false;
4307
4308         return true;
4309 }
4310
4311 static bool xdp_is_valid_access(int off, int size,
4312                                 enum bpf_access_type type,
4313                                 const struct bpf_prog *prog,
4314                                 struct bpf_insn_access_aux *info)
4315 {
4316         if (type == BPF_WRITE)
4317                 return false;
4318
4319         switch (off) {
4320         case offsetof(struct xdp_md, data):
4321                 info->reg_type = PTR_TO_PACKET;
4322                 break;
4323         case offsetof(struct xdp_md, data_meta):
4324                 info->reg_type = PTR_TO_PACKET_META;
4325                 break;
4326         case offsetof(struct xdp_md, data_end):
4327                 info->reg_type = PTR_TO_PACKET_END;
4328                 break;
4329         }
4330
4331         return __is_valid_xdp_access(off, size);
4332 }
4333
4334 void bpf_warn_invalid_xdp_action(u32 act)
4335 {
4336         const u32 act_max = XDP_REDIRECT;
4337
4338         WARN_ONCE(1, "%s XDP return value %u, expect packet loss!\n",
4339                   act > act_max ? "Illegal" : "Driver unsupported",
4340                   act);
4341 }
4342 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
4343
4344 static bool sock_addr_is_valid_access(int off, int size,
4345                                       enum bpf_access_type type,
4346                                       const struct bpf_prog *prog,
4347                                       struct bpf_insn_access_aux *info)
4348 {
4349         const int size_default = sizeof(__u32);
4350
4351         if (off < 0 || off >= sizeof(struct bpf_sock_addr))
4352                 return false;
4353         if (off % size != 0)
4354                 return false;
4355
4356         /* Disallow access to IPv6 fields from IPv4 contex and vise
4357          * versa.
4358          */
4359         switch (off) {
4360         case bpf_ctx_range(struct bpf_sock_addr, user_ip4):
4361                 switch (prog->expected_attach_type) {
4362                 case BPF_CGROUP_INET4_BIND:
4363                 case BPF_CGROUP_INET4_CONNECT:
4364                         break;
4365                 default:
4366                         return false;
4367                 }
4368                 break;
4369         case bpf_ctx_range_till(struct bpf_sock_addr, user_ip6[0], user_ip6[3]):
4370                 switch (prog->expected_attach_type) {
4371                 case BPF_CGROUP_INET6_BIND:
4372                 case BPF_CGROUP_INET6_CONNECT:
4373                         break;
4374                 default:
4375                         return false;
4376                 }
4377                 break;
4378         }
4379
4380         switch (off) {
4381         case bpf_ctx_range(struct bpf_sock_addr, user_ip4):
4382         case bpf_ctx_range_till(struct bpf_sock_addr, user_ip6[0], user_ip6[3]):
4383                 /* Only narrow read access allowed for now. */
4384                 if (type == BPF_READ) {
4385                         bpf_ctx_record_field_size(info, size_default);
4386                         if (!bpf_ctx_narrow_access_ok(off, size, size_default))
4387                                 return false;
4388                 } else {
4389                         if (size != size_default)
4390                                 return false;
4391                 }
4392                 break;
4393         case bpf_ctx_range(struct bpf_sock_addr, user_port):
4394                 if (size != size_default)
4395                         return false;
4396                 break;
4397         default:
4398                 if (type == BPF_READ) {
4399                         if (size != size_default)
4400                                 return false;
4401                 } else {
4402                         return false;
4403                 }
4404         }
4405
4406         return true;
4407 }
4408
4409 static bool sock_ops_is_valid_access(int off, int size,
4410                                      enum bpf_access_type type,
4411                                      const struct bpf_prog *prog,
4412                                      struct bpf_insn_access_aux *info)
4413 {
4414         const int size_default = sizeof(__u32);
4415
4416         if (off < 0 || off >= sizeof(struct bpf_sock_ops))
4417                 return false;
4418
4419         /* The verifier guarantees that size > 0. */
4420         if (off % size != 0)
4421                 return false;
4422
4423         if (type == BPF_WRITE) {
4424                 switch (off) {
4425                 case offsetof(struct bpf_sock_ops, reply):
4426                 case offsetof(struct bpf_sock_ops, sk_txhash):
4427                         if (size != size_default)
4428                                 return false;
4429                         break;
4430                 default:
4431                         return false;
4432                 }
4433         } else {
4434                 switch (off) {
4435                 case bpf_ctx_range_till(struct bpf_sock_ops, bytes_received,
4436                                         bytes_acked):
4437                         if (size != sizeof(__u64))
4438                                 return false;
4439                         break;
4440                 default:
4441                         if (size != size_default)
4442                                 return false;
4443                         break;
4444                 }
4445         }
4446
4447         return true;
4448 }
4449
4450 static int sk_skb_prologue(struct bpf_insn *insn_buf, bool direct_write,
4451                            const struct bpf_prog *prog)
4452 {
4453         return bpf_unclone_prologue(insn_buf, direct_write, prog, SK_DROP);
4454 }
4455
4456 static bool sk_skb_is_valid_access(int off, int size,
4457                                    enum bpf_access_type type,
4458                                    const struct bpf_prog *prog,
4459                                    struct bpf_insn_access_aux *info)
4460 {
4461         switch (off) {
4462         case bpf_ctx_range(struct __sk_buff, tc_classid):
4463         case bpf_ctx_range(struct __sk_buff, data_meta):
4464                 return false;
4465         }
4466
4467         if (type == BPF_WRITE) {
4468                 switch (off) {
4469                 case bpf_ctx_range(struct __sk_buff, tc_index):
4470                 case bpf_ctx_range(struct __sk_buff, priority):
4471                         break;
4472                 default:
4473                         return false;
4474                 }
4475         }
4476
4477         switch (off) {
4478         case bpf_ctx_range(struct __sk_buff, mark):
4479                 return false;
4480         case bpf_ctx_range(struct __sk_buff, data):
4481                 info->reg_type = PTR_TO_PACKET;
4482                 break;
4483         case bpf_ctx_range(struct __sk_buff, data_end):
4484                 info->reg_type = PTR_TO_PACKET_END;
4485                 break;
4486         }
4487
4488         return bpf_skb_is_valid_access(off, size, type, prog, info);
4489 }
4490
4491 static bool sk_msg_is_valid_access(int off, int size,
4492                                    enum bpf_access_type type,
4493                                    const struct bpf_prog *prog,
4494                                    struct bpf_insn_access_aux *info)
4495 {
4496         if (type == BPF_WRITE)
4497                 return false;
4498
4499         switch (off) {
4500         case offsetof(struct sk_msg_md, data):
4501                 info->reg_type = PTR_TO_PACKET;
4502                 break;
4503         case offsetof(struct sk_msg_md, data_end):
4504                 info->reg_type = PTR_TO_PACKET_END;
4505                 break;
4506         }
4507
4508         if (off < 0 || off >= sizeof(struct sk_msg_md))
4509                 return false;
4510         if (off % size != 0)
4511                 return false;
4512         if (size != sizeof(__u64))
4513                 return false;
4514
4515         return true;
4516 }
4517
4518 static u32 bpf_convert_ctx_access(enum bpf_access_type type,
4519                                   const struct bpf_insn *si,
4520                                   struct bpf_insn *insn_buf,
4521                                   struct bpf_prog *prog, u32 *target_size)
4522 {
4523         struct bpf_insn *insn = insn_buf;
4524         int off;
4525
4526         switch (si->off) {
4527         case offsetof(struct __sk_buff, len):
4528                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4529                                       bpf_target_off(struct sk_buff, len, 4,
4530                                                      target_size));
4531                 break;
4532
4533         case offsetof(struct __sk_buff, protocol):
4534                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4535                                       bpf_target_off(struct sk_buff, protocol, 2,
4536                                                      target_size));
4537                 break;
4538
4539         case offsetof(struct __sk_buff, vlan_proto):
4540                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4541                                       bpf_target_off(struct sk_buff, vlan_proto, 2,
4542                                                      target_size));
4543                 break;
4544
4545         case offsetof(struct __sk_buff, priority):
4546                 if (type == BPF_WRITE)
4547                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4548                                               bpf_target_off(struct sk_buff, priority, 4,
4549                                                              target_size));
4550                 else
4551                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4552                                               bpf_target_off(struct sk_buff, priority, 4,
4553                                                              target_size));
4554                 break;
4555
4556         case offsetof(struct __sk_buff, ingress_ifindex):
4557                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4558                                       bpf_target_off(struct sk_buff, skb_iif, 4,
4559                                                      target_size));
4560                 break;
4561
4562         case offsetof(struct __sk_buff, ifindex):
4563                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
4564                                       si->dst_reg, si->src_reg,
4565                                       offsetof(struct sk_buff, dev));
4566                 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 1);
4567                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4568                                       bpf_target_off(struct net_device, ifindex, 4,
4569                                                      target_size));
4570                 break;
4571
4572         case offsetof(struct __sk_buff, hash):
4573                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4574                                       bpf_target_off(struct sk_buff, hash, 4,
4575                                                      target_size));
4576                 break;
4577
4578         case offsetof(struct __sk_buff, mark):
4579                 if (type == BPF_WRITE)
4580                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4581                                               bpf_target_off(struct sk_buff, mark, 4,
4582                                                              target_size));
4583                 else
4584                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4585                                               bpf_target_off(struct sk_buff, mark, 4,
4586                                                              target_size));
4587                 break;
4588
4589         case offsetof(struct __sk_buff, pkt_type):
4590                 *target_size = 1;
4591                 *insn++ = BPF_LDX_MEM(BPF_B, si->dst_reg, si->src_reg,
4592                                       PKT_TYPE_OFFSET());
4593                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, PKT_TYPE_MAX);
4594 #ifdef __BIG_ENDIAN_BITFIELD
4595                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 5);
4596 #endif
4597                 break;
4598
4599         case offsetof(struct __sk_buff, queue_mapping):
4600                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4601                                       bpf_target_off(struct sk_buff, queue_mapping, 2,
4602                                                      target_size));
4603                 break;
4604
4605         case offsetof(struct __sk_buff, vlan_present):
4606         case offsetof(struct __sk_buff, vlan_tci):
4607                 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
4608
4609                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4610                                       bpf_target_off(struct sk_buff, vlan_tci, 2,
4611                                                      target_size));
4612                 if (si->off == offsetof(struct __sk_buff, vlan_tci)) {
4613                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg,
4614                                                 ~VLAN_TAG_PRESENT);
4615                 } else {
4616                         *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, 12);
4617                         *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, 1);
4618                 }
4619                 break;
4620
4621         case offsetof(struct __sk_buff, cb[0]) ...
4622              offsetofend(struct __sk_buff, cb[4]) - 1:
4623                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
4624                 BUILD_BUG_ON((offsetof(struct sk_buff, cb) +
4625                               offsetof(struct qdisc_skb_cb, data)) %
4626                              sizeof(__u64));
4627
4628                 prog->cb_access = 1;
4629                 off  = si->off;
4630                 off -= offsetof(struct __sk_buff, cb[0]);
4631                 off += offsetof(struct sk_buff, cb);
4632                 off += offsetof(struct qdisc_skb_cb, data);
4633                 if (type == BPF_WRITE)
4634                         *insn++ = BPF_STX_MEM(BPF_SIZE(si->code), si->dst_reg,
4635                                               si->src_reg, off);
4636                 else
4637                         *insn++ = BPF_LDX_MEM(BPF_SIZE(si->code), si->dst_reg,
4638                                               si->src_reg, off);
4639                 break;
4640
4641         case offsetof(struct __sk_buff, tc_classid):
4642                 BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, tc_classid) != 2);
4643
4644                 off  = si->off;
4645                 off -= offsetof(struct __sk_buff, tc_classid);
4646                 off += offsetof(struct sk_buff, cb);
4647                 off += offsetof(struct qdisc_skb_cb, tc_classid);
4648                 *target_size = 2;
4649                 if (type == BPF_WRITE)
4650                         *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg,
4651                                               si->src_reg, off);
4652                 else
4653                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg,
4654                                               si->src_reg, off);
4655                 break;
4656
4657         case offsetof(struct __sk_buff, data):
4658                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
4659                                       si->dst_reg, si->src_reg,
4660                                       offsetof(struct sk_buff, data));
4661                 break;
4662
4663         case offsetof(struct __sk_buff, data_meta):
4664                 off  = si->off;
4665                 off -= offsetof(struct __sk_buff, data_meta);
4666                 off += offsetof(struct sk_buff, cb);
4667                 off += offsetof(struct bpf_skb_data_end, data_meta);
4668                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4669                                       si->src_reg, off);
4670                 break;
4671
4672         case offsetof(struct __sk_buff, data_end):
4673                 off  = si->off;
4674                 off -= offsetof(struct __sk_buff, data_end);
4675                 off += offsetof(struct sk_buff, cb);
4676                 off += offsetof(struct bpf_skb_data_end, data_end);
4677                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
4678                                       si->src_reg, off);
4679                 break;
4680
4681         case offsetof(struct __sk_buff, tc_index):
4682 #ifdef CONFIG_NET_SCHED
4683                 if (type == BPF_WRITE)
4684                         *insn++ = BPF_STX_MEM(BPF_H, si->dst_reg, si->src_reg,
4685                                               bpf_target_off(struct sk_buff, tc_index, 2,
4686                                                              target_size));
4687                 else
4688                         *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4689                                               bpf_target_off(struct sk_buff, tc_index, 2,
4690                                                              target_size));
4691 #else
4692                 *target_size = 2;
4693                 if (type == BPF_WRITE)
4694                         *insn++ = BPF_MOV64_REG(si->dst_reg, si->dst_reg);
4695                 else
4696                         *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4697 #endif
4698                 break;
4699
4700         case offsetof(struct __sk_buff, napi_id):
4701 #if defined(CONFIG_NET_RX_BUSY_POLL)
4702                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4703                                       bpf_target_off(struct sk_buff, napi_id, 4,
4704                                                      target_size));
4705                 *insn++ = BPF_JMP_IMM(BPF_JGE, si->dst_reg, MIN_NAPI_ID, 1);
4706                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4707 #else
4708                 *target_size = 4;
4709                 *insn++ = BPF_MOV64_IMM(si->dst_reg, 0);
4710 #endif
4711                 break;
4712         case offsetof(struct __sk_buff, family):
4713                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
4714
4715                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4716                                       si->dst_reg, si->src_reg,
4717                                       offsetof(struct sk_buff, sk));
4718                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4719                                       bpf_target_off(struct sock_common,
4720                                                      skc_family,
4721                                                      2, target_size));
4722                 break;
4723         case offsetof(struct __sk_buff, remote_ip4):
4724                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
4725
4726                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4727                                       si->dst_reg, si->src_reg,
4728                                       offsetof(struct sk_buff, sk));
4729                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4730                                       bpf_target_off(struct sock_common,
4731                                                      skc_daddr,
4732                                                      4, target_size));
4733                 break;
4734         case offsetof(struct __sk_buff, local_ip4):
4735                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4736                                           skc_rcv_saddr) != 4);
4737
4738                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4739                                       si->dst_reg, si->src_reg,
4740                                       offsetof(struct sk_buff, sk));
4741                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4742                                       bpf_target_off(struct sock_common,
4743                                                      skc_rcv_saddr,
4744                                                      4, target_size));
4745                 break;
4746         case offsetof(struct __sk_buff, remote_ip6[0]) ...
4747              offsetof(struct __sk_buff, remote_ip6[3]):
4748 #if IS_ENABLED(CONFIG_IPV6)
4749                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4750                                           skc_v6_daddr.s6_addr32[0]) != 4);
4751
4752                 off = si->off;
4753                 off -= offsetof(struct __sk_buff, remote_ip6[0]);
4754
4755                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4756                                       si->dst_reg, si->src_reg,
4757                                       offsetof(struct sk_buff, sk));
4758                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4759                                       offsetof(struct sock_common,
4760                                                skc_v6_daddr.s6_addr32[0]) +
4761                                       off);
4762 #else
4763                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4764 #endif
4765                 break;
4766         case offsetof(struct __sk_buff, local_ip6[0]) ...
4767              offsetof(struct __sk_buff, local_ip6[3]):
4768 #if IS_ENABLED(CONFIG_IPV6)
4769                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
4770                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
4771
4772                 off = si->off;
4773                 off -= offsetof(struct __sk_buff, local_ip6[0]);
4774
4775                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4776                                       si->dst_reg, si->src_reg,
4777                                       offsetof(struct sk_buff, sk));
4778                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4779                                       offsetof(struct sock_common,
4780                                                skc_v6_rcv_saddr.s6_addr32[0]) +
4781                                       off);
4782 #else
4783                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4784 #endif
4785                 break;
4786
4787         case offsetof(struct __sk_buff, remote_port):
4788                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
4789
4790                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4791                                       si->dst_reg, si->src_reg,
4792                                       offsetof(struct sk_buff, sk));
4793                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4794                                       bpf_target_off(struct sock_common,
4795                                                      skc_dport,
4796                                                      2, target_size));
4797 #ifndef __BIG_ENDIAN_BITFIELD
4798                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
4799 #endif
4800                 break;
4801
4802         case offsetof(struct __sk_buff, local_port):
4803                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
4804
4805                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, sk),
4806                                       si->dst_reg, si->src_reg,
4807                                       offsetof(struct sk_buff, sk));
4808                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
4809                                       bpf_target_off(struct sock_common,
4810                                                      skc_num, 2, target_size));
4811                 break;
4812         }
4813
4814         return insn - insn_buf;
4815 }
4816
4817 static u32 sock_filter_convert_ctx_access(enum bpf_access_type type,
4818                                           const struct bpf_insn *si,
4819                                           struct bpf_insn *insn_buf,
4820                                           struct bpf_prog *prog, u32 *target_size)
4821 {
4822         struct bpf_insn *insn = insn_buf;
4823         int off;
4824
4825         switch (si->off) {
4826         case offsetof(struct bpf_sock, bound_dev_if):
4827                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_bound_dev_if) != 4);
4828
4829                 if (type == BPF_WRITE)
4830                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4831                                         offsetof(struct sock, sk_bound_dev_if));
4832                 else
4833                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4834                                       offsetof(struct sock, sk_bound_dev_if));
4835                 break;
4836
4837         case offsetof(struct bpf_sock, mark):
4838                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_mark) != 4);
4839
4840                 if (type == BPF_WRITE)
4841                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4842                                         offsetof(struct sock, sk_mark));
4843                 else
4844                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4845                                       offsetof(struct sock, sk_mark));
4846                 break;
4847
4848         case offsetof(struct bpf_sock, priority):
4849                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_priority) != 4);
4850
4851                 if (type == BPF_WRITE)
4852                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
4853                                         offsetof(struct sock, sk_priority));
4854                 else
4855                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4856                                       offsetof(struct sock, sk_priority));
4857                 break;
4858
4859         case offsetof(struct bpf_sock, family):
4860                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock, sk_family) != 2);
4861
4862                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->src_reg,
4863                                       offsetof(struct sock, sk_family));
4864                 break;
4865
4866         case offsetof(struct bpf_sock, type):
4867                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4868                                       offsetof(struct sock, __sk_flags_offset));
4869                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
4870                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
4871                 break;
4872
4873         case offsetof(struct bpf_sock, protocol):
4874                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
4875                                       offsetof(struct sock, __sk_flags_offset));
4876                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
4877                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_PROTO_SHIFT);
4878                 break;
4879
4880         case offsetof(struct bpf_sock, src_ip4):
4881                 *insn++ = BPF_LDX_MEM(
4882                         BPF_SIZE(si->code), si->dst_reg, si->src_reg,
4883                         bpf_target_off(struct sock_common, skc_rcv_saddr,
4884                                        FIELD_SIZEOF(struct sock_common,
4885                                                     skc_rcv_saddr),
4886                                        target_size));
4887                 break;
4888
4889         case bpf_ctx_range_till(struct bpf_sock, src_ip6[0], src_ip6[3]):
4890 #if IS_ENABLED(CONFIG_IPV6)
4891                 off = si->off;
4892                 off -= offsetof(struct bpf_sock, src_ip6[0]);
4893                 *insn++ = BPF_LDX_MEM(
4894                         BPF_SIZE(si->code), si->dst_reg, si->src_reg,
4895                         bpf_target_off(
4896                                 struct sock_common,
4897                                 skc_v6_rcv_saddr.s6_addr32[0],
4898                                 FIELD_SIZEOF(struct sock_common,
4899                                              skc_v6_rcv_saddr.s6_addr32[0]),
4900                                 target_size) + off);
4901 #else
4902                 (void)off;
4903                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
4904 #endif
4905                 break;
4906
4907         case offsetof(struct bpf_sock, src_port):
4908                 *insn++ = BPF_LDX_MEM(
4909                         BPF_FIELD_SIZEOF(struct sock_common, skc_num),
4910                         si->dst_reg, si->src_reg,
4911                         bpf_target_off(struct sock_common, skc_num,
4912                                        FIELD_SIZEOF(struct sock_common,
4913                                                     skc_num),
4914                                        target_size));
4915                 break;
4916         }
4917
4918         return insn - insn_buf;
4919 }
4920
4921 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type,
4922                                          const struct bpf_insn *si,
4923                                          struct bpf_insn *insn_buf,
4924                                          struct bpf_prog *prog, u32 *target_size)
4925 {
4926         struct bpf_insn *insn = insn_buf;
4927
4928         switch (si->off) {
4929         case offsetof(struct __sk_buff, ifindex):
4930                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
4931                                       si->dst_reg, si->src_reg,
4932                                       offsetof(struct sk_buff, dev));
4933                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4934                                       bpf_target_off(struct net_device, ifindex, 4,
4935                                                      target_size));
4936                 break;
4937         default:
4938                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
4939                                               target_size);
4940         }
4941
4942         return insn - insn_buf;
4943 }
4944
4945 static u32 xdp_convert_ctx_access(enum bpf_access_type type,
4946                                   const struct bpf_insn *si,
4947                                   struct bpf_insn *insn_buf,
4948                                   struct bpf_prog *prog, u32 *target_size)
4949 {
4950         struct bpf_insn *insn = insn_buf;
4951
4952         switch (si->off) {
4953         case offsetof(struct xdp_md, data):
4954                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
4955                                       si->dst_reg, si->src_reg,
4956                                       offsetof(struct xdp_buff, data));
4957                 break;
4958         case offsetof(struct xdp_md, data_meta):
4959                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_meta),
4960                                       si->dst_reg, si->src_reg,
4961                                       offsetof(struct xdp_buff, data_meta));
4962                 break;
4963         case offsetof(struct xdp_md, data_end):
4964                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
4965                                       si->dst_reg, si->src_reg,
4966                                       offsetof(struct xdp_buff, data_end));
4967                 break;
4968         case offsetof(struct xdp_md, ingress_ifindex):
4969                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, rxq),
4970                                       si->dst_reg, si->src_reg,
4971                                       offsetof(struct xdp_buff, rxq));
4972                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_rxq_info, dev),
4973                                       si->dst_reg, si->dst_reg,
4974                                       offsetof(struct xdp_rxq_info, dev));
4975                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4976                                       offsetof(struct net_device, ifindex));
4977                 break;
4978         case offsetof(struct xdp_md, rx_queue_index):
4979                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, rxq),
4980                                       si->dst_reg, si->src_reg,
4981                                       offsetof(struct xdp_buff, rxq));
4982                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
4983                                       offsetof(struct xdp_rxq_info,
4984                                                queue_index));
4985                 break;
4986         }
4987
4988         return insn - insn_buf;
4989 }
4990
4991 /* SOCK_ADDR_LOAD_NESTED_FIELD() loads Nested Field S.F.NF where S is type of
4992  * context Structure, F is Field in context structure that contains a pointer
4993  * to Nested Structure of type NS that has the field NF.
4994  *
4995  * SIZE encodes the load size (BPF_B, BPF_H, etc). It's up to caller to make
4996  * sure that SIZE is not greater than actual size of S.F.NF.
4997  *
4998  * If offset OFF is provided, the load happens from that offset relative to
4999  * offset of NF.
5000  */
5001 #define SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF(S, NS, F, NF, SIZE, OFF)          \
5002         do {                                                                   \
5003                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), si->dst_reg,     \
5004                                       si->src_reg, offsetof(S, F));            \
5005                 *insn++ = BPF_LDX_MEM(                                         \
5006                         SIZE, si->dst_reg, si->dst_reg,                        \
5007                         bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF),           \
5008                                        target_size)                            \
5009                                 + OFF);                                        \
5010         } while (0)
5011
5012 #define SOCK_ADDR_LOAD_NESTED_FIELD(S, NS, F, NF)                              \
5013         SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF(S, NS, F, NF,                     \
5014                                              BPF_FIELD_SIZEOF(NS, NF), 0)
5015
5016 /* SOCK_ADDR_STORE_NESTED_FIELD_OFF() has semantic similar to
5017  * SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF() but for store operation.
5018  *
5019  * It doesn't support SIZE argument though since narrow stores are not
5020  * supported for now.
5021  *
5022  * In addition it uses Temporary Field TF (member of struct S) as the 3rd
5023  * "register" since two registers available in convert_ctx_access are not
5024  * enough: we can't override neither SRC, since it contains value to store, nor
5025  * DST since it contains pointer to context that may be used by later
5026  * instructions. But we need a temporary place to save pointer to nested
5027  * structure whose field we want to store to.
5028  */
5029 #define SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF, TF)                \
5030         do {                                                                   \
5031                 int tmp_reg = BPF_REG_9;                                       \
5032                 if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg)          \
5033                         --tmp_reg;                                             \
5034                 if (si->src_reg == tmp_reg || si->dst_reg == tmp_reg)          \
5035                         --tmp_reg;                                             \
5036                 *insn++ = BPF_STX_MEM(BPF_DW, si->dst_reg, tmp_reg,            \
5037                                       offsetof(S, TF));                        \
5038                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(S, F), tmp_reg,         \
5039                                       si->dst_reg, offsetof(S, F));            \
5040                 *insn++ = BPF_STX_MEM(                                         \
5041                         BPF_FIELD_SIZEOF(NS, NF), tmp_reg, si->src_reg,        \
5042                         bpf_target_off(NS, NF, FIELD_SIZEOF(NS, NF),           \
5043                                        target_size)                            \
5044                                 + OFF);                                        \
5045                 *insn++ = BPF_LDX_MEM(BPF_DW, tmp_reg, si->dst_reg,            \
5046                                       offsetof(S, TF));                        \
5047         } while (0)
5048
5049 #define SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD_SIZE_OFF(S, NS, F, NF, SIZE, OFF, \
5050                                                       TF)                      \
5051         do {                                                                   \
5052                 if (type == BPF_WRITE) {                                       \
5053                         SOCK_ADDR_STORE_NESTED_FIELD_OFF(S, NS, F, NF, OFF,    \
5054                                                          TF);                  \
5055                 } else {                                                       \
5056                         SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF(                  \
5057                                 S, NS, F, NF, SIZE, OFF);  \
5058                 }                                                              \
5059         } while (0)
5060
5061 #define SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD(S, NS, F, NF, TF)                 \
5062         SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD_SIZE_OFF(                         \
5063                 S, NS, F, NF, BPF_FIELD_SIZEOF(NS, NF), 0, TF)
5064
5065 static u32 sock_addr_convert_ctx_access(enum bpf_access_type type,
5066                                         const struct bpf_insn *si,
5067                                         struct bpf_insn *insn_buf,
5068                                         struct bpf_prog *prog, u32 *target_size)
5069 {
5070         struct bpf_insn *insn = insn_buf;
5071         int off;
5072
5073         switch (si->off) {
5074         case offsetof(struct bpf_sock_addr, user_family):
5075                 SOCK_ADDR_LOAD_NESTED_FIELD(struct bpf_sock_addr_kern,
5076                                             struct sockaddr, uaddr, sa_family);
5077                 break;
5078
5079         case offsetof(struct bpf_sock_addr, user_ip4):
5080                 SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD_SIZE_OFF(
5081                         struct bpf_sock_addr_kern, struct sockaddr_in, uaddr,
5082                         sin_addr, BPF_SIZE(si->code), 0, tmp_reg);
5083                 break;
5084
5085         case bpf_ctx_range_till(struct bpf_sock_addr, user_ip6[0], user_ip6[3]):
5086                 off = si->off;
5087                 off -= offsetof(struct bpf_sock_addr, user_ip6[0]);
5088                 SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD_SIZE_OFF(
5089                         struct bpf_sock_addr_kern, struct sockaddr_in6, uaddr,
5090                         sin6_addr.s6_addr32[0], BPF_SIZE(si->code), off,
5091                         tmp_reg);
5092                 break;
5093
5094         case offsetof(struct bpf_sock_addr, user_port):
5095                 /* To get port we need to know sa_family first and then treat
5096                  * sockaddr as either sockaddr_in or sockaddr_in6.
5097                  * Though we can simplify since port field has same offset and
5098                  * size in both structures.
5099                  * Here we check this invariant and use just one of the
5100                  * structures if it's true.
5101                  */
5102                 BUILD_BUG_ON(offsetof(struct sockaddr_in, sin_port) !=
5103                              offsetof(struct sockaddr_in6, sin6_port));
5104                 BUILD_BUG_ON(FIELD_SIZEOF(struct sockaddr_in, sin_port) !=
5105                              FIELD_SIZEOF(struct sockaddr_in6, sin6_port));
5106                 SOCK_ADDR_LOAD_OR_STORE_NESTED_FIELD(struct bpf_sock_addr_kern,
5107                                                      struct sockaddr_in6, uaddr,
5108                                                      sin6_port, tmp_reg);
5109                 break;
5110
5111         case offsetof(struct bpf_sock_addr, family):
5112                 SOCK_ADDR_LOAD_NESTED_FIELD(struct bpf_sock_addr_kern,
5113                                             struct sock, sk, sk_family);
5114                 break;
5115
5116         case offsetof(struct bpf_sock_addr, type):
5117                 SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF(
5118                         struct bpf_sock_addr_kern, struct sock, sk,
5119                         __sk_flags_offset, BPF_W, 0);
5120                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_TYPE_MASK);
5121                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg, SK_FL_TYPE_SHIFT);
5122                 break;
5123
5124         case offsetof(struct bpf_sock_addr, protocol):
5125                 SOCK_ADDR_LOAD_NESTED_FIELD_SIZE_OFF(
5126                         struct bpf_sock_addr_kern, struct sock, sk,
5127                         __sk_flags_offset, BPF_W, 0);
5128                 *insn++ = BPF_ALU32_IMM(BPF_AND, si->dst_reg, SK_FL_PROTO_MASK);
5129                 *insn++ = BPF_ALU32_IMM(BPF_RSH, si->dst_reg,
5130                                         SK_FL_PROTO_SHIFT);
5131                 break;
5132         }
5133
5134         return insn - insn_buf;
5135 }
5136
5137 static u32 sock_ops_convert_ctx_access(enum bpf_access_type type,
5138                                        const struct bpf_insn *si,
5139                                        struct bpf_insn *insn_buf,
5140                                        struct bpf_prog *prog,
5141                                        u32 *target_size)
5142 {
5143         struct bpf_insn *insn = insn_buf;
5144         int off;
5145
5146         switch (si->off) {
5147         case offsetof(struct bpf_sock_ops, op) ...
5148              offsetof(struct bpf_sock_ops, replylong[3]):
5149                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, op) !=
5150                              FIELD_SIZEOF(struct bpf_sock_ops_kern, op));
5151                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, reply) !=
5152                              FIELD_SIZEOF(struct bpf_sock_ops_kern, reply));
5153                 BUILD_BUG_ON(FIELD_SIZEOF(struct bpf_sock_ops, replylong) !=
5154                              FIELD_SIZEOF(struct bpf_sock_ops_kern, replylong));
5155                 off = si->off;
5156                 off -= offsetof(struct bpf_sock_ops, op);
5157                 off += offsetof(struct bpf_sock_ops_kern, op);
5158                 if (type == BPF_WRITE)
5159                         *insn++ = BPF_STX_MEM(BPF_W, si->dst_reg, si->src_reg,
5160                                               off);
5161                 else
5162                         *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->src_reg,
5163                                               off);
5164                 break;
5165
5166         case offsetof(struct bpf_sock_ops, family):
5167                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_family) != 2);
5168
5169                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5170                                               struct bpf_sock_ops_kern, sk),
5171                                       si->dst_reg, si->src_reg,
5172                                       offsetof(struct bpf_sock_ops_kern, sk));
5173                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
5174                                       offsetof(struct sock_common, skc_family));
5175                 break;
5176
5177         case offsetof(struct bpf_sock_ops, remote_ip4):
5178                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_daddr) != 4);
5179
5180                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5181                                                 struct bpf_sock_ops_kern, sk),
5182                                       si->dst_reg, si->src_reg,
5183                                       offsetof(struct bpf_sock_ops_kern, sk));
5184                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
5185                                       offsetof(struct sock_common, skc_daddr));
5186                 break;
5187
5188         case offsetof(struct bpf_sock_ops, local_ip4):
5189                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_rcv_saddr) != 4);
5190
5191                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5192                                               struct bpf_sock_ops_kern, sk),
5193                                       si->dst_reg, si->src_reg,
5194                                       offsetof(struct bpf_sock_ops_kern, sk));
5195                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
5196                                       offsetof(struct sock_common,
5197                                                skc_rcv_saddr));
5198                 break;
5199
5200         case offsetof(struct bpf_sock_ops, remote_ip6[0]) ...
5201              offsetof(struct bpf_sock_ops, remote_ip6[3]):
5202 #if IS_ENABLED(CONFIG_IPV6)
5203                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
5204                                           skc_v6_daddr.s6_addr32[0]) != 4);
5205
5206                 off = si->off;
5207                 off -= offsetof(struct bpf_sock_ops, remote_ip6[0]);
5208                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5209                                                 struct bpf_sock_ops_kern, sk),
5210                                       si->dst_reg, si->src_reg,
5211                                       offsetof(struct bpf_sock_ops_kern, sk));
5212                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
5213                                       offsetof(struct sock_common,
5214                                                skc_v6_daddr.s6_addr32[0]) +
5215                                       off);
5216 #else
5217                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
5218 #endif
5219                 break;
5220
5221         case offsetof(struct bpf_sock_ops, local_ip6[0]) ...
5222              offsetof(struct bpf_sock_ops, local_ip6[3]):
5223 #if IS_ENABLED(CONFIG_IPV6)
5224                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common,
5225                                           skc_v6_rcv_saddr.s6_addr32[0]) != 4);
5226
5227                 off = si->off;
5228                 off -= offsetof(struct bpf_sock_ops, local_ip6[0]);
5229                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5230                                                 struct bpf_sock_ops_kern, sk),
5231                                       si->dst_reg, si->src_reg,
5232                                       offsetof(struct bpf_sock_ops_kern, sk));
5233                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
5234                                       offsetof(struct sock_common,
5235                                                skc_v6_rcv_saddr.s6_addr32[0]) +
5236                                       off);
5237 #else
5238                 *insn++ = BPF_MOV32_IMM(si->dst_reg, 0);
5239 #endif
5240                 break;
5241
5242         case offsetof(struct bpf_sock_ops, remote_port):
5243                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_dport) != 2);
5244
5245                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5246                                                 struct bpf_sock_ops_kern, sk),
5247                                       si->dst_reg, si->src_reg,
5248                                       offsetof(struct bpf_sock_ops_kern, sk));
5249                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
5250                                       offsetof(struct sock_common, skc_dport));
5251 #ifndef __BIG_ENDIAN_BITFIELD
5252                 *insn++ = BPF_ALU32_IMM(BPF_LSH, si->dst_reg, 16);
5253 #endif
5254                 break;
5255
5256         case offsetof(struct bpf_sock_ops, local_port):
5257                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_num) != 2);
5258
5259                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5260                                                 struct bpf_sock_ops_kern, sk),
5261                                       si->dst_reg, si->src_reg,
5262                                       offsetof(struct bpf_sock_ops_kern, sk));
5263                 *insn++ = BPF_LDX_MEM(BPF_H, si->dst_reg, si->dst_reg,
5264                                       offsetof(struct sock_common, skc_num));
5265                 break;
5266
5267         case offsetof(struct bpf_sock_ops, is_fullsock):
5268                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5269                                                 struct bpf_sock_ops_kern,
5270                                                 is_fullsock),
5271                                       si->dst_reg, si->src_reg,
5272                                       offsetof(struct bpf_sock_ops_kern,
5273                                                is_fullsock));
5274                 break;
5275
5276         case offsetof(struct bpf_sock_ops, state):
5277                 BUILD_BUG_ON(FIELD_SIZEOF(struct sock_common, skc_state) != 1);
5278
5279                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5280                                                 struct bpf_sock_ops_kern, sk),
5281                                       si->dst_reg, si->src_reg,
5282                                       offsetof(struct bpf_sock_ops_kern, sk));
5283                 *insn++ = BPF_LDX_MEM(BPF_B, si->dst_reg, si->dst_reg,
5284                                       offsetof(struct sock_common, skc_state));
5285                 break;
5286
5287         case offsetof(struct bpf_sock_ops, rtt_min):
5288                 BUILD_BUG_ON(FIELD_SIZEOF(struct tcp_sock, rtt_min) !=
5289                              sizeof(struct minmax));
5290                 BUILD_BUG_ON(sizeof(struct minmax) <
5291                              sizeof(struct minmax_sample));
5292
5293                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(
5294                                                 struct bpf_sock_ops_kern, sk),
5295                                       si->dst_reg, si->src_reg,
5296                                       offsetof(struct bpf_sock_ops_kern, sk));
5297                 *insn++ = BPF_LDX_MEM(BPF_W, si->dst_reg, si->dst_reg,
5298                                       offsetof(struct tcp_sock, rtt_min) +
5299                                       FIELD_SIZEOF(struct minmax_sample, t));
5300                 break;
5301
5302 /* Helper macro for adding read access to tcp_sock or sock fields. */
5303 #define SOCK_OPS_GET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ)                         \
5304         do {                                                                  \
5305                 BUILD_BUG_ON(FIELD_SIZEOF(OBJ, OBJ_FIELD) >                   \
5306                              FIELD_SIZEOF(struct bpf_sock_ops, BPF_FIELD));   \
5307                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(                       \
5308                                                 struct bpf_sock_ops_kern,     \
5309                                                 is_fullsock),                 \
5310                                       si->dst_reg, si->src_reg,               \
5311                                       offsetof(struct bpf_sock_ops_kern,      \
5312                                                is_fullsock));                 \
5313                 *insn++ = BPF_JMP_IMM(BPF_JEQ, si->dst_reg, 0, 2);            \
5314                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(                       \
5315                                                 struct bpf_sock_ops_kern, sk),\
5316                                       si->dst_reg, si->src_reg,               \
5317                                       offsetof(struct bpf_sock_ops_kern, sk));\
5318                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(OBJ,                   \
5319                                                        OBJ_FIELD),            \
5320                                       si->dst_reg, si->dst_reg,               \
5321                                       offsetof(OBJ, OBJ_FIELD));              \
5322         } while (0)
5323
5324 /* Helper macro for adding write access to tcp_sock or sock fields.
5325  * The macro is called with two registers, dst_reg which contains a pointer
5326  * to ctx (context) and src_reg which contains the value that should be
5327  * stored. However, we need an additional register since we cannot overwrite
5328  * dst_reg because it may be used later in the program.
5329  * Instead we "borrow" one of the other register. We first save its value
5330  * into a new (temp) field in bpf_sock_ops_kern, use it, and then restore
5331  * it at the end of the macro.
5332  */
5333 #define SOCK_OPS_SET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ)                         \
5334         do {                                                                  \
5335                 int reg = BPF_REG_9;                                          \
5336                 BUILD_BUG_ON(FIELD_SIZEOF(OBJ, OBJ_FIELD) >                   \
5337                              FIELD_SIZEOF(struct bpf_sock_ops, BPF_FIELD));   \
5338                 if (si->dst_reg == reg || si->src_reg == reg)                 \
5339                         reg--;                                                \
5340                 if (si->dst_reg == reg || si->src_reg == reg)                 \
5341                         reg--;                                                \
5342                 *insn++ = BPF_STX_MEM(BPF_DW, si->dst_reg, reg,               \
5343                                       offsetof(struct bpf_sock_ops_kern,      \
5344                                                temp));                        \
5345                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(                       \
5346                                                 struct bpf_sock_ops_kern,     \
5347                                                 is_fullsock),                 \
5348                                       reg, si->dst_reg,                       \
5349                                       offsetof(struct bpf_sock_ops_kern,      \
5350                                                is_fullsock));                 \
5351                 *insn++ = BPF_JMP_IMM(BPF_JEQ, reg, 0, 2);                    \
5352                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(                       \
5353                                                 struct bpf_sock_ops_kern, sk),\
5354                                       reg, si->dst_reg,                       \
5355                                       offsetof(struct bpf_sock_ops_kern, sk));\
5356                 *insn++ = BPF_STX_MEM(BPF_FIELD_SIZEOF(OBJ, OBJ_FIELD),       \
5357                                       reg, si->src_reg,                       \
5358                                       offsetof(OBJ, OBJ_FIELD));              \
5359                 *insn++ = BPF_LDX_MEM(BPF_DW, reg, si->dst_reg,               \
5360                                       offsetof(struct bpf_sock_ops_kern,      \
5361                                                temp));                        \
5362         } while (0)
5363
5364 #define SOCK_OPS_GET_OR_SET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ, TYPE)            \
5365         do {                                                                  \
5366                 if (TYPE == BPF_WRITE)                                        \
5367                         SOCK_OPS_SET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ);        \
5368                 else                                                          \
5369                         SOCK_OPS_GET_FIELD(BPF_FIELD, OBJ_FIELD, OBJ);        \
5370         } while (0)
5371
5372         case offsetof(struct bpf_sock_ops, snd_cwnd):
5373                 SOCK_OPS_GET_FIELD(snd_cwnd, snd_cwnd, struct tcp_sock);
5374                 break;
5375
5376         case offsetof(struct bpf_sock_ops, srtt_us):
5377                 SOCK_OPS_GET_FIELD(srtt_us, srtt_us, struct tcp_sock);
5378                 break;
5379
5380         case offsetof(struct bpf_sock_ops, bpf_sock_ops_cb_flags):
5381                 SOCK_OPS_GET_FIELD(bpf_sock_ops_cb_flags, bpf_sock_ops_cb_flags,
5382                                    struct tcp_sock);
5383                 break;
5384
5385         case offsetof(struct bpf_sock_ops, snd_ssthresh):
5386                 SOCK_OPS_GET_FIELD(snd_ssthresh, snd_ssthresh, struct tcp_sock);
5387                 break;
5388
5389         case offsetof(struct bpf_sock_ops, rcv_nxt):
5390                 SOCK_OPS_GET_FIELD(rcv_nxt, rcv_nxt, struct tcp_sock);
5391                 break;
5392
5393         case offsetof(struct bpf_sock_ops, snd_nxt):
5394                 SOCK_OPS_GET_FIELD(snd_nxt, snd_nxt, struct tcp_sock);
5395                 break;
5396
5397         case offsetof(struct bpf_sock_ops, snd_una):
5398                 SOCK_OPS_GET_FIELD(snd_una, snd_una, struct tcp_sock);
5399                 break;
5400
5401         case offsetof(struct bpf_sock_ops, mss_cache):
5402                 SOCK_OPS_GET_FIELD(mss_cache, mss_cache, struct tcp_sock);
5403                 break;
5404
5405         case offsetof(struct bpf_sock_ops, ecn_flags):
5406                 SOCK_OPS_GET_FIELD(ecn_flags, ecn_flags, struct tcp_sock);
5407                 break;
5408
5409         case offsetof(struct bpf_sock_ops, rate_delivered):
5410                 SOCK_OPS_GET_FIELD(rate_delivered, rate_delivered,
5411                                    struct tcp_sock);
5412                 break;
5413
5414         case offsetof(struct bpf_sock_ops, rate_interval_us):
5415                 SOCK_OPS_GET_FIELD(rate_interval_us, rate_interval_us,
5416                                    struct tcp_sock);
5417                 break;
5418
5419         case offsetof(struct bpf_sock_ops, packets_out):
5420                 SOCK_OPS_GET_FIELD(packets_out, packets_out, struct tcp_sock);
5421                 break;
5422
5423         case offsetof(struct bpf_sock_ops, retrans_out):
5424                 SOCK_OPS_GET_FIELD(retrans_out, retrans_out, struct tcp_sock);
5425                 break;
5426
5427         case offsetof(struct bpf_sock_ops, total_retrans):
5428                 SOCK_OPS_GET_FIELD(total_retrans, total_retrans,
5429                                    struct tcp_sock);
5430                 break;
5431
5432         case offsetof(struct bpf_sock_ops, segs_in):
5433                 SOCK_OPS_GET_FIELD(segs_in, segs_in, struct tcp_sock);
5434                 break;
5435
5436         case offsetof(struct bpf_sock_ops, data_segs_in):
5437                 SOCK_OPS_GET_FIELD(data_segs_in, data_segs_in, struct tcp_sock);
5438                 break;
5439
5440         case offsetof(struct bpf_sock_ops, segs_out):
5441                 SOCK_OPS_GET_FIELD(segs_out, segs_out, struct tcp_sock);
5442                 break;
5443
5444         case offsetof(struct bpf_sock_ops, data_segs_out):
5445                 SOCK_OPS_GET_FIELD(data_segs_out, data_segs_out,
5446                                    struct tcp_sock);
5447                 break;
5448
5449         case offsetof(struct bpf_sock_ops, lost_out):
5450                 SOCK_OPS_GET_FIELD(lost_out, lost_out, struct tcp_sock);
5451                 break;
5452
5453         case offsetof(struct bpf_sock_ops, sacked_out):
5454                 SOCK_OPS_GET_FIELD(sacked_out, sacked_out, struct tcp_sock);
5455                 break;
5456
5457         case offsetof(struct bpf_sock_ops, sk_txhash):
5458                 SOCK_OPS_GET_OR_SET_FIELD(sk_txhash, sk_txhash,
5459                                           struct sock, type);
5460                 break;
5461
5462         case offsetof(struct bpf_sock_ops, bytes_received):
5463                 SOCK_OPS_GET_FIELD(bytes_received, bytes_received,
5464                                    struct tcp_sock);
5465                 break;
5466
5467         case offsetof(struct bpf_sock_ops, bytes_acked):
5468                 SOCK_OPS_GET_FIELD(bytes_acked, bytes_acked, struct tcp_sock);
5469                 break;
5470
5471         }
5472         return insn - insn_buf;
5473 }
5474
5475 static u32 sk_skb_convert_ctx_access(enum bpf_access_type type,
5476                                      const struct bpf_insn *si,
5477                                      struct bpf_insn *insn_buf,
5478                                      struct bpf_prog *prog, u32 *target_size)
5479 {
5480         struct bpf_insn *insn = insn_buf;
5481         int off;
5482
5483         switch (si->off) {
5484         case offsetof(struct __sk_buff, data_end):
5485                 off  = si->off;
5486                 off -= offsetof(struct __sk_buff, data_end);
5487                 off += offsetof(struct sk_buff, cb);
5488                 off += offsetof(struct tcp_skb_cb, bpf.data_end);
5489                 *insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), si->dst_reg,
5490                                       si->src_reg, off);
5491                 break;
5492         default:
5493                 return bpf_convert_ctx_access(type, si, insn_buf, prog,
5494                                               target_size);
5495         }
5496
5497         return insn - insn_buf;
5498 }
5499
5500 static u32 sk_msg_convert_ctx_access(enum bpf_access_type type,
5501                                      const struct bpf_insn *si,
5502                                      struct bpf_insn *insn_buf,
5503                                      struct bpf_prog *prog, u32 *target_size)
5504 {
5505         struct bpf_insn *insn = insn_buf;
5506
5507         switch (si->off) {
5508         case offsetof(struct sk_msg_md, data):
5509                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_msg_buff, data),
5510                                       si->dst_reg, si->src_reg,
5511                                       offsetof(struct sk_msg_buff, data));
5512                 break;
5513         case offsetof(struct sk_msg_md, data_end):
5514                 *insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_msg_buff, data_end),
5515                                       si->dst_reg, si->src_reg,
5516                                       offsetof(struct sk_msg_buff, data_end));
5517                 break;
5518         }
5519
5520         return insn - insn_buf;
5521 }
5522
5523 const struct bpf_verifier_ops sk_filter_verifier_ops = {
5524         .get_func_proto         = sk_filter_func_proto,
5525         .is_valid_access        = sk_filter_is_valid_access,
5526         .convert_ctx_access     = bpf_convert_ctx_access,
5527 };
5528
5529 const struct bpf_prog_ops sk_filter_prog_ops = {
5530         .test_run               = bpf_prog_test_run_skb,
5531 };
5532
5533 const struct bpf_verifier_ops tc_cls_act_verifier_ops = {
5534         .get_func_proto         = tc_cls_act_func_proto,
5535         .is_valid_access        = tc_cls_act_is_valid_access,
5536         .convert_ctx_access     = tc_cls_act_convert_ctx_access,
5537         .gen_prologue           = tc_cls_act_prologue,
5538 };
5539
5540 const struct bpf_prog_ops tc_cls_act_prog_ops = {
5541         .test_run               = bpf_prog_test_run_skb,
5542 };
5543
5544 const struct bpf_verifier_ops xdp_verifier_ops = {
5545         .get_func_proto         = xdp_func_proto,
5546         .is_valid_access        = xdp_is_valid_access,
5547         .convert_ctx_access     = xdp_convert_ctx_access,
5548 };
5549
5550 const struct bpf_prog_ops xdp_prog_ops = {
5551         .test_run               = bpf_prog_test_run_xdp,
5552 };
5553
5554 const struct bpf_verifier_ops cg_skb_verifier_ops = {
5555         .get_func_proto         = sk_filter_func_proto,
5556         .is_valid_access        = sk_filter_is_valid_access,
5557         .convert_ctx_access     = bpf_convert_ctx_access,
5558 };
5559
5560 const struct bpf_prog_ops cg_skb_prog_ops = {
5561         .test_run               = bpf_prog_test_run_skb,
5562 };
5563
5564 const struct bpf_verifier_ops lwt_inout_verifier_ops = {
5565         .get_func_proto         = lwt_inout_func_proto,
5566         .is_valid_access        = lwt_is_valid_access,
5567         .convert_ctx_access     = bpf_convert_ctx_access,
5568 };
5569
5570 const struct bpf_prog_ops lwt_inout_prog_ops = {
5571         .test_run               = bpf_prog_test_run_skb,
5572 };
5573
5574 const struct bpf_verifier_ops lwt_xmit_verifier_ops = {
5575         .get_func_proto         = lwt_xmit_func_proto,
5576         .is_valid_access        = lwt_is_valid_access,
5577         .convert_ctx_access     = bpf_convert_ctx_access,
5578         .gen_prologue           = tc_cls_act_prologue,
5579 };
5580
5581 const struct bpf_prog_ops lwt_xmit_prog_ops = {
5582         .test_run               = bpf_prog_test_run_skb,
5583 };
5584
5585 const struct bpf_verifier_ops cg_sock_verifier_ops = {
5586         .get_func_proto         = sock_filter_func_proto,
5587         .is_valid_access        = sock_filter_is_valid_access,
5588         .convert_ctx_access     = sock_filter_convert_ctx_access,
5589 };
5590
5591 const struct bpf_prog_ops cg_sock_prog_ops = {
5592 };
5593
5594 const struct bpf_verifier_ops cg_sock_addr_verifier_ops = {
5595         .get_func_proto         = sock_addr_func_proto,
5596         .is_valid_access        = sock_addr_is_valid_access,
5597         .convert_ctx_access     = sock_addr_convert_ctx_access,
5598 };
5599
5600 const struct bpf_prog_ops cg_sock_addr_prog_ops = {
5601 };
5602
5603 const struct bpf_verifier_ops sock_ops_verifier_ops = {
5604         .get_func_proto         = sock_ops_func_proto,
5605         .is_valid_access        = sock_ops_is_valid_access,
5606         .convert_ctx_access     = sock_ops_convert_ctx_access,
5607 };
5608
5609 const struct bpf_prog_ops sock_ops_prog_ops = {
5610 };
5611
5612 const struct bpf_verifier_ops sk_skb_verifier_ops = {
5613         .get_func_proto         = sk_skb_func_proto,
5614         .is_valid_access        = sk_skb_is_valid_access,
5615         .convert_ctx_access     = sk_skb_convert_ctx_access,
5616         .gen_prologue           = sk_skb_prologue,
5617 };
5618
5619 const struct bpf_prog_ops sk_skb_prog_ops = {
5620 };
5621
5622 const struct bpf_verifier_ops sk_msg_verifier_ops = {
5623         .get_func_proto         = sk_msg_func_proto,
5624         .is_valid_access        = sk_msg_is_valid_access,
5625         .convert_ctx_access     = sk_msg_convert_ctx_access,
5626 };
5627
5628 const struct bpf_prog_ops sk_msg_prog_ops = {
5629 };
5630
5631 int sk_detach_filter(struct sock *sk)
5632 {
5633         int ret = -ENOENT;
5634         struct sk_filter *filter;
5635
5636         if (sock_flag(sk, SOCK_FILTER_LOCKED))
5637                 return -EPERM;
5638
5639         filter = rcu_dereference_protected(sk->sk_filter,
5640                                            lockdep_sock_is_held(sk));
5641         if (filter) {
5642                 RCU_INIT_POINTER(sk->sk_filter, NULL);
5643                 sk_filter_uncharge(sk, filter);
5644                 ret = 0;
5645         }
5646
5647         return ret;
5648 }
5649 EXPORT_SYMBOL_GPL(sk_detach_filter);
5650
5651 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
5652                   unsigned int len)
5653 {
5654         struct sock_fprog_kern *fprog;
5655         struct sk_filter *filter;
5656         int ret = 0;
5657
5658         lock_sock(sk);
5659         filter = rcu_dereference_protected(sk->sk_filter,
5660                                            lockdep_sock_is_held(sk));
5661         if (!filter)
5662                 goto out;
5663
5664         /* We're copying the filter that has been originally attached,
5665          * so no conversion/decode needed anymore. eBPF programs that
5666          * have no original program cannot be dumped through this.
5667          */
5668         ret = -EACCES;
5669         fprog = filter->prog->orig_prog;
5670         if (!fprog)
5671                 goto out;
5672
5673         ret = fprog->len;
5674         if (!len)
5675                 /* User space only enquires number of filter blocks. */
5676                 goto out;
5677
5678         ret = -EINVAL;
5679         if (len < fprog->len)
5680                 goto out;
5681
5682         ret = -EFAULT;
5683         if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
5684                 goto out;
5685
5686         /* Instead of bytes, the API requests to return the number
5687          * of filter blocks.
5688          */
5689         ret = fprog->len;
5690 out:
5691         release_sock(sk);
5692         return ret;
5693 }