kernel/bpf/core.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Linux Socket Filter - Kernel level socket filtering
   4  *
   5  * Based on the design of the Berkeley Packet Filter. The new
   6  * internal format has been designed by PLUMgrid:
   7  *
   8  *      Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
   9  *
  10  * Authors:
  11  *
  12  *      Jay Schulist <jschlst@samba.org>
  13  *      Alexei Starovoitov <ast@plumgrid.com>
  14  *      Daniel Borkmann <dborkman@redhat.com>
  15  *
  16  * Andi Kleen - Fix a few bad bugs and races.
  17  * Kris Katterjohn - Added many additional checks in bpf_check_classic()
  18  */
  19
  20 #include <uapi/linux/btf.h>
  21 #include <linux/filter.h>
  22 #include <linux/skbuff.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/random.h>
  25 #include <linux/moduleloader.h>
  26 #include <linux/bpf.h>
  27 #include <linux/btf.h>
  28 #include <linux/objtool.h>
  29 #include <linux/rbtree_latch.h>
  30 #include <linux/kallsyms.h>
  31 #include <linux/rcupdate.h>
  32 #include <linux/perf_event.h>
  33 #include <linux/extable.h>
  34 #include <linux/log2.h>
  35 #include <linux/bpf_verifier.h>
  36 #include <linux/nodemask.h>
  37 #include <linux/nospec.h>
  38 #include <linux/bpf_mem_alloc.h>
  39 #include <linux/memcontrol.h>
  40
  41 #include <asm/barrier.h>
  42 #include <asm/unaligned.h>
  43
  44 /* Registers */
  45 #define BPF_R0  regs[BPF_REG_0]
  46 #define BPF_R1  regs[BPF_REG_1]
  47 #define BPF_R2  regs[BPF_REG_2]
  48 #define BPF_R3  regs[BPF_REG_3]
  49 #define BPF_R4  regs[BPF_REG_4]
  50 #define BPF_R5  regs[BPF_REG_5]
  51 #define BPF_R6  regs[BPF_REG_6]
  52 #define BPF_R7  regs[BPF_REG_7]
  53 #define BPF_R8  regs[BPF_REG_8]
  54 #define BPF_R9  regs[BPF_REG_9]
  55 #define BPF_R10 regs[BPF_REG_10]
  56
  57 /* Named registers */
  58 #define DST     regs[insn->dst_reg]
  59 #define SRC     regs[insn->src_reg]
  60 #define FP      regs[BPF_REG_FP]
  61 #define AX      regs[BPF_REG_AX]
  62 #define ARG1    regs[BPF_REG_ARG1]
  63 #define CTX     regs[BPF_REG_CTX]
  64 #define OFF     insn->off
  65 #define IMM     insn->imm
  66
  67 struct bpf_mem_alloc bpf_global_ma;
  68 bool bpf_global_ma_set;
  69
  70 /* No hurry in this branch
  71  *
  72  * Exported for the bpf jit load helper.
  73  */
  74 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
  75 {
  76         u8 *ptr = NULL;
  77
  78         if (k >= SKF_NET_OFF) {
  79                 ptr = skb_network_header(skb) + k - SKF_NET_OFF;
  80         } else if (k >= SKF_LL_OFF) {
  81                 if (unlikely(!skb_mac_header_was_set(skb)))
  82                         return NULL;
  83                 ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
  84         }
  85         if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
  86                 return ptr;
  87
  88         return NULL;
  89 }
  90
  91 struct bpf_prog *bpf_prog_alloc_no_stats(unsigned int size, gfp_t gfp_extra_flags)
  92 {
  93         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
  94         struct bpf_prog_aux *aux;
  95         struct bpf_prog *fp;
  96
  97         size = round_up(size, PAGE_SIZE);
  98         fp = __vmalloc(size, gfp_flags);
  99         if (fp == NULL)
 100                 return NULL;
 101
 102         aux = kzalloc(sizeof(*aux), bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
 103         if (aux == NULL) {
 104                 vfree(fp);
 105                 return NULL;
 106         }
 107         fp->active = alloc_percpu_gfp(int, bpf_memcg_flags(GFP_KERNEL | gfp_extra_flags));
 108         if (!fp->active) {
 109                 vfree(fp);
 110                 kfree(aux);
 111                 return NULL;
 112         }
 113
 114         fp->pages = size / PAGE_SIZE;
 115         fp->aux = aux;
 116         fp->aux->prog = fp;
 117         fp->jit_requested = ebpf_jit_enabled();
 118         fp->blinding_requested = bpf_jit_blinding_enabled(fp);
 119 #ifdef CONFIG_CGROUP_BPF
 120         aux->cgroup_atype = CGROUP_BPF_ATTACH_TYPE_INVALID;
 121 #endif
 122
 123         INIT_LIST_HEAD_RCU(&fp->aux->ksym.lnode);
 124         mutex_init(&fp->aux->used_maps_mutex);
 125         mutex_init(&fp->aux->dst_mutex);
 126
 127         return fp;
 128 }
 129
 130 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
 131 {
 132         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
 133         struct bpf_prog *prog;
 134         int cpu;
 135
 136         prog = bpf_prog_alloc_no_stats(size, gfp_extra_flags);
 137         if (!prog)
 138                 return NULL;
 139
 140         prog->stats = alloc_percpu_gfp(struct bpf_prog_stats, gfp_flags);
 141         if (!prog->stats) {
 142                 free_percpu(prog->active);
 143                 kfree(prog->aux);
 144                 vfree(prog);
 145                 return NULL;
 146         }
 147
 148         for_each_possible_cpu(cpu) {
 149                 struct bpf_prog_stats *pstats;
 150
 151                 pstats = per_cpu_ptr(prog->stats, cpu);
 152                 u64_stats_init(&pstats->syncp);
 153         }
 154         return prog;
 155 }
 156 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
 157
 158 int bpf_prog_alloc_jited_linfo(struct bpf_prog *prog)
 159 {
 160         if (!prog->aux->nr_linfo || !prog->jit_requested)
 161                 return 0;
 162
 163         prog->aux->jited_linfo = kvcalloc(prog->aux->nr_linfo,
 164                                           sizeof(*prog->aux->jited_linfo),
 165                                           bpf_memcg_flags(GFP_KERNEL | __GFP_NOWARN));
 166         if (!prog->aux->jited_linfo)
 167                 return -ENOMEM;
 168
 169         return 0;
 170 }
 171
 172 void bpf_prog_jit_attempt_done(struct bpf_prog *prog)
 173 {
 174         if (prog->aux->jited_linfo &&
 175             (!prog->jited || !prog->aux->jited_linfo[0])) {
 176                 kvfree(prog->aux->jited_linfo);
 177                 prog->aux->jited_linfo = NULL;
 178         }
 179
 180         kfree(prog->aux->kfunc_tab);
 181         prog->aux->kfunc_tab = NULL;
 182 }
 183
 184 /* The jit engine is responsible to provide an array
 185  * for insn_off to the jited_off mapping (insn_to_jit_off).
 186  *
 187  * The idx to this array is the insn_off.  Hence, the insn_off
 188  * here is relative to the prog itself instead of the main prog.
 189  * This array has one entry for each xlated bpf insn.
 190  *
 191  * jited_off is the byte off to the end of the jited insn.
 192  *
 193  * Hence, with
 194  * insn_start:
 195  *      The first bpf insn off of the prog.  The insn off
 196  *      here is relative to the main prog.
 197  *      e.g. if prog is a subprog, insn_start > 0
 198  * linfo_idx:
 199  *      The prog's idx to prog->aux->linfo and jited_linfo
 200  *
 201  * jited_linfo[linfo_idx] = prog->bpf_func
 202  *
 203  * For i > linfo_idx,
 204  *
 205  * jited_linfo[i] = prog->bpf_func +
 206  *      insn_to_jit_off[linfo[i].insn_off - insn_start - 1]
 207  */
 208 void bpf_prog_fill_jited_linfo(struct bpf_prog *prog,
 209                                const u32 *insn_to_jit_off)
 210 {
 211         u32 linfo_idx, insn_start, insn_end, nr_linfo, i;
 212         const struct bpf_line_info *linfo;
 213         void **jited_linfo;
 214
 215         if (!prog->aux->jited_linfo)
 216                 /* Userspace did not provide linfo */
 217                 return;
 218
 219         linfo_idx = prog->aux->linfo_idx;
 220         linfo = &prog->aux->linfo[linfo_idx];
 221         insn_start = linfo[0].insn_off;
 222         insn_end = insn_start + prog->len;
 223
 224         jited_linfo = &prog->aux->jited_linfo[linfo_idx];
 225         jited_linfo[0] = prog->bpf_func;
 226
 227         nr_linfo = prog->aux->nr_linfo - linfo_idx;
 228
 229         for (i = 1; i < nr_linfo && linfo[i].insn_off < insn_end; i++)
 230                 /* The verifier ensures that linfo[i].insn_off is
 231                  * strictly increasing
 232                  */
 233                 jited_linfo[i] = prog->bpf_func +
 234                         insn_to_jit_off[linfo[i].insn_off - insn_start - 1];
 235 }
 236
 237 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
 238                                   gfp_t gfp_extra_flags)
 239 {
 240         gfp_t gfp_flags = bpf_memcg_flags(GFP_KERNEL | __GFP_ZERO | gfp_extra_flags);
 241         struct bpf_prog *fp;
 242         u32 pages;
 243
 244         size = round_up(size, PAGE_SIZE);
 245         pages = size / PAGE_SIZE;
 246         if (pages <= fp_old->pages)
 247                 return fp_old;
 248
 249         fp = __vmalloc(size, gfp_flags);
 250         if (fp) {
 251                 memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
 252                 fp->pages = pages;
 253                 fp->aux->prog = fp;
 254
 255                 /* We keep fp->aux from fp_old around in the new
 256                  * reallocated structure.
 257                  */
 258                 fp_old->aux = NULL;
 259                 fp_old->stats = NULL;
 260                 fp_old->active = NULL;
 261                 __bpf_prog_free(fp_old);
 262         }
 263
 264         return fp;
 265 }
 266
 267 void __bpf_prog_free(struct bpf_prog *fp)
 268 {
 269         if (fp->aux) {
 270                 mutex_destroy(&fp->aux->used_maps_mutex);
 271                 mutex_destroy(&fp->aux->dst_mutex);
 272                 kfree(fp->aux->poke_tab);
 273                 kfree(fp->aux);
 274         }
 275         free_percpu(fp->stats);
 276         free_percpu(fp->active);
 277         vfree(fp);
 278 }
 279
 280 int bpf_prog_calc_tag(struct bpf_prog *fp)
 281 {
 282         const u32 bits_offset = SHA1_BLOCK_SIZE - sizeof(__be64);
 283         u32 raw_size = bpf_prog_tag_scratch_size(fp);
 284         u32 digest[SHA1_DIGEST_WORDS];
 285         u32 ws[SHA1_WORKSPACE_WORDS];
 286         u32 i, bsize, psize, blocks;
 287         struct bpf_insn *dst;
 288         bool was_ld_map;
 289         u8 *raw, *todo;
 290         __be32 *result;
 291         __be64 *bits;
 292
 293         raw = vmalloc(raw_size);
 294         if (!raw)
 295                 return -ENOMEM;
 296
 297         sha1_init(digest);
 298         memset(ws, 0, sizeof(ws));
 299
 300         /* We need to take out the map fd for the digest calculation
 301          * since they are unstable from user space side.
 302          */
 303         dst = (void *)raw;
 304         for (i = 0, was_ld_map = false; i < fp->len; i++) {
 305                 dst[i] = fp->insnsi[i];
 306                 if (!was_ld_map &&
 307                     dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
 308                     (dst[i].src_reg == BPF_PSEUDO_MAP_FD ||
 309                      dst[i].src_reg == BPF_PSEUDO_MAP_VALUE)) {
 310                         was_ld_map = true;
 311                         dst[i].imm = 0;
 312                 } else if (was_ld_map &&
 313                            dst[i].code == 0 &&
 314                            dst[i].dst_reg == 0 &&
 315                            dst[i].src_reg == 0 &&
 316                            dst[i].off == 0) {
 317                         was_ld_map = false;
 318                         dst[i].imm = 0;
 319                 } else {
 320                         was_ld_map = false;
 321                 }
 322         }
 323
 324         psize = bpf_prog_insn_size(fp);
 325         memset(&raw[psize], 0, raw_size - psize);
 326         raw[psize++] = 0x80;
 327
 328         bsize  = round_up(psize, SHA1_BLOCK_SIZE);
 329         blocks = bsize / SHA1_BLOCK_SIZE;
 330         todo   = raw;
 331         if (bsize - psize >= sizeof(__be64)) {
 332                 bits = (__be64 *)(todo + bsize - sizeof(__be64));
 333         } else {
 334                 bits = (__be64 *)(todo + bsize + bits_offset);
 335                 blocks++;
 336         }
 337         *bits = cpu_to_be64((psize - 1) << 3);
 338
 339         while (blocks--) {
 340                 sha1_transform(digest, todo, ws);
 341                 todo += SHA1_BLOCK_SIZE;
 342         }
 343
 344         result = (__force __be32 *)digest;
 345         for (i = 0; i < SHA1_DIGEST_WORDS; i++)
 346                 result[i] = cpu_to_be32(digest[i]);
 347         memcpy(fp->tag, result, sizeof(fp->tag));
 348
 349         vfree(raw);
 350         return 0;
 351 }
 352
 353 static int bpf_adj_delta_to_imm(struct bpf_insn *insn, u32 pos, s32 end_old,
 354                                 s32 end_new, s32 curr, const bool probe_pass)
 355 {
 356         const s64 imm_min = S32_MIN, imm_max = S32_MAX;
 357         s32 delta = end_new - end_old;
 358         s64 imm = insn->imm;
 359
 360         if (curr < pos && curr + imm + 1 >= end_old)
 361                 imm += delta;
 362         else if (curr >= end_new && curr + imm + 1 < end_new)
 363                 imm -= delta;
 364         if (imm < imm_min || imm > imm_max)
 365                 return -ERANGE;
 366         if (!probe_pass)
 367                 insn->imm = imm;
 368         return 0;
 369 }
 370
 371 static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, s32 end_old,
 372                                 s32 end_new, s32 curr, const bool probe_pass)
 373 {
 374         const s32 off_min = S16_MIN, off_max = S16_MAX;
 375         s32 delta = end_new - end_old;
 376         s32 off;
 377
 378         if (insn->code == (BPF_JMP32 | BPF_JA))
 379                 off = insn->imm;
 380         else
 381                 off = insn->off;
 382
 383         if (curr < pos && curr + off + 1 >= end_old)
 384                 off += delta;
 385         else if (curr >= end_new && curr + off + 1 < end_new)
 386                 off -= delta;
 387         if (off < off_min || off > off_max)
 388                 return -ERANGE;
 389         if (!probe_pass) {
 390                 if (insn->code == (BPF_JMP32 | BPF_JA))
 391                         insn->imm = off;
 392                 else
 393                         insn->off = off;
 394         }
 395         return 0;
 396 }
 397
 398 static int bpf_adj_branches(struct bpf_prog *prog, u32 pos, s32 end_old,
 399                             s32 end_new, const bool probe_pass)
 400 {
 401         u32 i, insn_cnt = prog->len + (probe_pass ? end_new - end_old : 0);
 402         struct bpf_insn *insn = prog->insnsi;
 403         int ret = 0;
 404
 405         for (i = 0; i < insn_cnt; i++, insn++) {
 406                 u8 code;
 407
 408                 /* In the probing pass we still operate on the original,
 409                  * unpatched image in order to check overflows before we
 410                  * do any other adjustments. Therefore skip the patchlet.
 411                  */
 412                 if (probe_pass && i == pos) {
 413                         i = end_new;
 414                         insn = prog->insnsi + end_old;
 415                 }
 416                 if (bpf_pseudo_func(insn)) {
 417                         ret = bpf_adj_delta_to_imm(insn, pos, end_old,
 418                                                    end_new, i, probe_pass);
 419                         if (ret)
 420                                 return ret;
 421                         continue;
 422                 }
 423                 code = insn->code;
 424                 if ((BPF_CLASS(code) != BPF_JMP &&
 425                      BPF_CLASS(code) != BPF_JMP32) ||
 426                     BPF_OP(code) == BPF_EXIT)
 427                         continue;
 428                 /* Adjust offset of jmps if we cross patch boundaries. */
 429                 if (BPF_OP(code) == BPF_CALL) {
 430                         if (insn->src_reg != BPF_PSEUDO_CALL)
 431                                 continue;
 432                         ret = bpf_adj_delta_to_imm(insn, pos, end_old,
 433                                                    end_new, i, probe_pass);
 434                 } else {
 435                         ret = bpf_adj_delta_to_off(insn, pos, end_old,
 436                                                    end_new, i, probe_pass);
 437                 }
 438                 if (ret)
 439                         break;
 440         }
 441
 442         return ret;
 443 }
 444
 445 static void bpf_adj_linfo(struct bpf_prog *prog, u32 off, u32 delta)
 446 {
 447         struct bpf_line_info *linfo;
 448         u32 i, nr_linfo;
 449
 450         nr_linfo = prog->aux->nr_linfo;
 451         if (!nr_linfo || !delta)
 452                 return;
 453
 454         linfo = prog->aux->linfo;
 455
 456         for (i = 0; i < nr_linfo; i++)
 457                 if (off < linfo[i].insn_off)
 458                         break;
 459
 460         /* Push all off < linfo[i].insn_off by delta */
 461         for (; i < nr_linfo; i++)
 462                 linfo[i].insn_off += delta;
 463 }
 464
 465 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
 466                                        const struct bpf_insn *patch, u32 len)
 467 {
 468         u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
 469         const u32 cnt_max = S16_MAX;
 470         struct bpf_prog *prog_adj;
 471         int err;
 472
 473         /* Since our patchlet doesn't expand the image, we're done. */
 474         if (insn_delta == 0) {
 475                 memcpy(prog->insnsi + off, patch, sizeof(*patch));
 476                 return prog;
 477         }
 478
 479         insn_adj_cnt = prog->len + insn_delta;
 480
 481         /* Reject anything that would potentially let the insn->off
 482          * target overflow when we have excessive program expansions.
 483          * We need to probe here before we do any reallocation where
 484          * we afterwards may not fail anymore.
 485          */
 486         if (insn_adj_cnt > cnt_max &&
 487             (err = bpf_adj_branches(prog, off, off + 1, off + len, true)))
 488                 return ERR_PTR(err);
 489
 490         /* Several new instructions need to be inserted. Make room
 491          * for them. Likely, there's no need for a new allocation as
 492          * last page could have large enough tailroom.
 493          */
 494         prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
 495                                     GFP_USER);
 496         if (!prog_adj)
 497                 return ERR_PTR(-ENOMEM);
 498
 499         prog_adj->len = insn_adj_cnt;
 500
 501         /* Patching happens in 3 steps:
 502          *
 503          * 1) Move over tail of insnsi from next instruction onwards,
 504          *    so we can patch the single target insn with one or more
 505          *    new ones (patching is always from 1 to n insns, n > 0).
 506          * 2) Inject new instructions at the target location.
 507          * 3) Adjust branch offsets if necessary.
 508          */
 509         insn_rest = insn_adj_cnt - off - len;
 510
 511         memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
 512                 sizeof(*patch) * insn_rest);
 513         memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
 514
 515         /* We are guaranteed to not fail at this point, otherwise
 516          * the ship has sailed to reverse to the original state. An
 517          * overflow cannot happen at this point.
 518          */
 519         BUG_ON(bpf_adj_branches(prog_adj, off, off + 1, off + len, false));
 520
 521         bpf_adj_linfo(prog_adj, off, insn_delta);
 522
 523         return prog_adj;
 524 }
 525
 526 int bpf_remove_insns(struct bpf_prog *prog, u32 off, u32 cnt)
 527 {
 528         /* Branch offsets can't overflow when program is shrinking, no need
 529          * to call bpf_adj_branches(..., true) here
 530          */
 531         memmove(prog->insnsi + off, prog->insnsi + off + cnt,
 532                 sizeof(struct bpf_insn) * (prog->len - off - cnt));
 533         prog->len -= cnt;
 534
 535         return WARN_ON_ONCE(bpf_adj_branches(prog, off, off + cnt, off, false));
 536 }
 537
 538 static void bpf_prog_kallsyms_del_subprogs(struct bpf_prog *fp)
 539 {
 540         int i;
 541
 542         for (i = 0; i < fp->aux->func_cnt; i++)
 543                 bpf_prog_kallsyms_del(fp->aux->func[i]);
 544 }
 545
 546 void bpf_prog_kallsyms_del_all(struct bpf_prog *fp)
 547 {
 548         bpf_prog_kallsyms_del_subprogs(fp);
 549         bpf_prog_kallsyms_del(fp);
 550 }
 551
 552 #ifdef CONFIG_BPF_JIT
 553 /* All BPF JIT sysctl knobs here. */
 554 int bpf_jit_enable   __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
 555 int bpf_jit_kallsyms __read_mostly = IS_BUILTIN(CONFIG_BPF_JIT_DEFAULT_ON);
 556 int bpf_jit_harden   __read_mostly;
 557 long bpf_jit_limit   __read_mostly;
 558 long bpf_jit_limit_max __read_mostly;
 559
 560 static void
 561 bpf_prog_ksym_set_addr(struct bpf_prog *prog)
 562 {
 563         WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
 564
 565         prog->aux->ksym.start = (unsigned long) prog->bpf_func;
 566         prog->aux->ksym.end   = prog->aux->ksym.start + prog->jited_len;
 567 }
 568
 569 static void
 570 bpf_prog_ksym_set_name(struct bpf_prog *prog)
 571 {
 572         char *sym = prog->aux->ksym.name;
 573         const char *end = sym + KSYM_NAME_LEN;
 574         const struct btf_type *type;
 575         const char *func_name;
 576
 577         BUILD_BUG_ON(sizeof("bpf_prog_") +
 578                      sizeof(prog->tag) * 2 +
 579                      /* name has been null terminated.
 580                       * We should need +1 for the '_' preceding
 581                       * the name.  However, the null character
 582                       * is double counted between the name and the
 583                       * sizeof("bpf_prog_") above, so we omit
 584                       * the +1 here.
 585                       */
 586                      sizeof(prog->aux->name) > KSYM_NAME_LEN);
 587
 588         sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
 589         sym  = bin2hex(sym, prog->tag, sizeof(prog->tag));
 590
 591         /* prog->aux->name will be ignored if full btf name is available */
 592         if (prog->aux->func_info_cnt) {
 593                 type = btf_type_by_id(prog->aux->btf,
 594                                       prog->aux->func_info[prog->aux->func_idx].type_id);
 595                 func_name = btf_name_by_offset(prog->aux->btf, type->name_off);
 596                 snprintf(sym, (size_t)(end - sym), "_%s", func_name);
 597                 return;
 598         }
 599
 600         if (prog->aux->name[0])
 601                 snprintf(sym, (size_t)(end - sym), "_%s", prog->aux->name);
 602         else
 603                 *sym = 0;
 604 }
 605
 606 static unsigned long bpf_get_ksym_start(struct latch_tree_node *n)
 607 {
 608         return container_of(n, struct bpf_ksym, tnode)->start;
 609 }
 610
 611 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
 612                                           struct latch_tree_node *b)
 613 {
 614         return bpf_get_ksym_start(a) < bpf_get_ksym_start(b);
 615 }
 616
 617 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
 618 {
 619         unsigned long val = (unsigned long)key;
 620         const struct bpf_ksym *ksym;
 621
 622         ksym = container_of(n, struct bpf_ksym, tnode);
 623
 624         if (val < ksym->start)
 625                 return -1;
 626         if (val >= ksym->end)
 627                 return  1;
 628
 629         return 0;
 630 }
 631
 632 static const struct latch_tree_ops bpf_tree_ops = {
 633         .less   = bpf_tree_less,
 634         .comp   = bpf_tree_comp,
 635 };
 636
 637 static DEFINE_SPINLOCK(bpf_lock);
 638 static LIST_HEAD(bpf_kallsyms);
 639 static struct latch_tree_root bpf_tree __cacheline_aligned;
 640
 641 void bpf_ksym_add(struct bpf_ksym *ksym)
 642 {
 643         spin_lock_bh(&bpf_lock);
 644         WARN_ON_ONCE(!list_empty(&ksym->lnode));
 645         list_add_tail_rcu(&ksym->lnode, &bpf_kallsyms);
 646         latch_tree_insert(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
 647         spin_unlock_bh(&bpf_lock);
 648 }
 649
 650 static void __bpf_ksym_del(struct bpf_ksym *ksym)
 651 {
 652         if (list_empty(&ksym->lnode))
 653                 return;
 654
 655         latch_tree_erase(&ksym->tnode, &bpf_tree, &bpf_tree_ops);
 656         list_del_rcu(&ksym->lnode);
 657 }
 658
 659 void bpf_ksym_del(struct bpf_ksym *ksym)
 660 {
 661         spin_lock_bh(&bpf_lock);
 662         __bpf_ksym_del(ksym);
 663         spin_unlock_bh(&bpf_lock);
 664 }
 665
 666 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
 667 {
 668         return fp->jited && !bpf_prog_was_classic(fp);
 669 }
 670
 671 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
 672 {
 673         if (!bpf_prog_kallsyms_candidate(fp) ||
 674             !bpf_capable())
 675                 return;
 676
 677         bpf_prog_ksym_set_addr(fp);
 678         bpf_prog_ksym_set_name(fp);
 679         fp->aux->ksym.prog = true;
 680
 681         bpf_ksym_add(&fp->aux->ksym);
 682 }
 683
 684 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
 685 {
 686         if (!bpf_prog_kallsyms_candidate(fp))
 687                 return;
 688
 689         bpf_ksym_del(&fp->aux->ksym);
 690 }
 691
 692 static struct bpf_ksym *bpf_ksym_find(unsigned long addr)
 693 {
 694         struct latch_tree_node *n;
 695
 696         n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
 697         return n ? container_of(n, struct bpf_ksym, tnode) : NULL;
 698 }
 699
 700 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
 701                                  unsigned long *off, char *sym)
 702 {
 703         struct bpf_ksym *ksym;
 704         char *ret = NULL;
 705
 706         rcu_read_lock();
 707         ksym = bpf_ksym_find(addr);
 708         if (ksym) {
 709                 unsigned long symbol_start = ksym->start;
 710                 unsigned long symbol_end = ksym->end;
 711
 712                 strncpy(sym, ksym->name, KSYM_NAME_LEN);
 713
 714                 ret = sym;
 715                 if (size)
 716                         *size = symbol_end - symbol_start;
 717                 if (off)
 718                         *off  = addr - symbol_start;
 719         }
 720         rcu_read_unlock();
 721
 722         return ret;
 723 }
 724
 725 bool is_bpf_text_address(unsigned long addr)
 726 {
 727         bool ret;
 728
 729         rcu_read_lock();
 730         ret = bpf_ksym_find(addr) != NULL;
 731         rcu_read_unlock();
 732
 733         return ret;
 734 }
 735
 736 static struct bpf_prog *bpf_prog_ksym_find(unsigned long addr)
 737 {
 738         struct bpf_ksym *ksym = bpf_ksym_find(addr);
 739
 740         return ksym && ksym->prog ?
 741                container_of(ksym, struct bpf_prog_aux, ksym)->prog :
 742                NULL;
 743 }
 744
 745 const struct exception_table_entry *search_bpf_extables(unsigned long addr)
 746 {
 747         const struct exception_table_entry *e = NULL;
 748         struct bpf_prog *prog;
 749
 750         rcu_read_lock();
 751         prog = bpf_prog_ksym_find(addr);
 752         if (!prog)
 753                 goto out;
 754         if (!prog->aux->num_exentries)
 755                 goto out;
 756
 757         e = search_extable(prog->aux->extable, prog->aux->num_exentries, addr);
 758 out:
 759         rcu_read_unlock();
 760         return e;
 761 }
 762
 763 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
 764                     char *sym)
 765 {
 766         struct bpf_ksym *ksym;
 767         unsigned int it = 0;
 768         int ret = -ERANGE;
 769
 770         if (!bpf_jit_kallsyms_enabled())
 771                 return ret;
 772
 773         rcu_read_lock();
 774         list_for_each_entry_rcu(ksym, &bpf_kallsyms, lnode) {
 775                 if (it++ != symnum)
 776                         continue;
 777
 778                 strncpy(sym, ksym->name, KSYM_NAME_LEN);
 779
 780                 *value = ksym->start;
 781                 *type  = BPF_SYM_ELF_TYPE;
 782
 783                 ret = 0;
 784                 break;
 785         }
 786         rcu_read_unlock();
 787
 788         return ret;
 789 }
 790
 791 int bpf_jit_add_poke_descriptor(struct bpf_prog *prog,
 792                                 struct bpf_jit_poke_descriptor *poke)
 793 {
 794         struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
 795         static const u32 poke_tab_max = 1024;
 796         u32 slot = prog->aux->size_poke_tab;
 797         u32 size = slot + 1;
 798
 799         if (size > poke_tab_max)
 800                 return -ENOSPC;
 801         if (poke->tailcall_target || poke->tailcall_target_stable ||
 802             poke->tailcall_bypass || poke->adj_off || poke->bypass_addr)
 803                 return -EINVAL;
 804
 805         switch (poke->reason) {
 806         case BPF_POKE_REASON_TAIL_CALL:
 807                 if (!poke->tail_call.map)
 808                         return -EINVAL;
 809                 break;
 810         default:
 811                 return -EINVAL;
 812         }
 813
 814         tab = krealloc(tab, size * sizeof(*poke), GFP_KERNEL);
 815         if (!tab)
 816                 return -ENOMEM;
 817
 818         memcpy(&tab[slot], poke, sizeof(*poke));
 819         prog->aux->size_poke_tab = size;
 820         prog->aux->poke_tab = tab;
 821
 822         return slot;
 823 }
 824
 825 /*
 826  * BPF program pack allocator.
 827  *
 828  * Most BPF programs are pretty small. Allocating a hole page for each
 829  * program is sometime a waste. Many small bpf program also adds pressure
 830  * to instruction TLB. To solve this issue, we introduce a BPF program pack
 831  * allocator. The prog_pack allocator uses HPAGE_PMD_SIZE page (2MB on x86)
 832  * to host BPF programs.
 833  */
 834 #define BPF_PROG_CHUNK_SHIFT    6
 835 #define BPF_PROG_CHUNK_SIZE     (1 << BPF_PROG_CHUNK_SHIFT)
 836 #define BPF_PROG_CHUNK_MASK     (~(BPF_PROG_CHUNK_SIZE - 1))
 837
 838 struct bpf_prog_pack {
 839         struct list_head list;
 840         void *ptr;
 841         unsigned long bitmap[];
 842 };
 843
 844 void bpf_jit_fill_hole_with_zero(void *area, unsigned int size)
 845 {
 846         memset(area, 0, size);
 847 }
 848
 849 #define BPF_PROG_SIZE_TO_NBITS(size)    (round_up(size, BPF_PROG_CHUNK_SIZE) / BPF_PROG_CHUNK_SIZE)
 850
 851 static DEFINE_MUTEX(pack_mutex);
 852 static LIST_HEAD(pack_list);
 853
 854 /* PMD_SIZE is not available in some special config, e.g. ARCH=arm with
 855  * CONFIG_MMU=n. Use PAGE_SIZE in these cases.
 856  */
 857 #ifdef PMD_SIZE
 858 #define BPF_PROG_PACK_SIZE (PMD_SIZE * num_possible_nodes())
 859 #else
 860 #define BPF_PROG_PACK_SIZE PAGE_SIZE
 861 #endif
 862
 863 #define BPF_PROG_CHUNK_COUNT (BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE)
 864
 865 static struct bpf_prog_pack *alloc_new_pack(bpf_jit_fill_hole_t bpf_fill_ill_insns)
 866 {
 867         struct bpf_prog_pack *pack;
 868
 869         pack = kzalloc(struct_size(pack, bitmap, BITS_TO_LONGS(BPF_PROG_CHUNK_COUNT)),
 870                        GFP_KERNEL);
 871         if (!pack)
 872                 return NULL;
 873         pack->ptr = module_alloc(BPF_PROG_PACK_SIZE);
 874         if (!pack->ptr) {
 875                 kfree(pack);
 876                 return NULL;
 877         }
 878         bpf_fill_ill_insns(pack->ptr, BPF_PROG_PACK_SIZE);
 879         bitmap_zero(pack->bitmap, BPF_PROG_PACK_SIZE / BPF_PROG_CHUNK_SIZE);
 880         list_add_tail(&pack->list, &pack_list);
 881
 882         set_vm_flush_reset_perms(pack->ptr);
 883         set_memory_rox((unsigned long)pack->ptr, BPF_PROG_PACK_SIZE / PAGE_SIZE);
 884         return pack;
 885 }
 886
 887 void *bpf_prog_pack_alloc(u32 size, bpf_jit_fill_hole_t bpf_fill_ill_insns)
 888 {
 889         unsigned int nbits = BPF_PROG_SIZE_TO_NBITS(size);
 890         struct bpf_prog_pack *pack;
 891         unsigned long pos;
 892         void *ptr = NULL;
 893
 894         mutex_lock(&pack_mutex);
 895         if (size > BPF_PROG_PACK_SIZE) {
 896                 size = round_up(size, PAGE_SIZE);
 897                 ptr = module_alloc(size);
 898                 if (ptr) {
 899                         bpf_fill_ill_insns(ptr, size);
 900                         set_vm_flush_reset_perms(ptr);
 901                         set_memory_rox((unsigned long)ptr, size / PAGE_SIZE);
 902                 }
 903                 goto out;
 904         }
 905         list_for_each_entry(pack, &pack_list, list) {
 906                 pos = bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
 907                                                  nbits, 0);
 908                 if (pos < BPF_PROG_CHUNK_COUNT)
 909                         goto found_free_area;
 910         }
 911
 912         pack = alloc_new_pack(bpf_fill_ill_insns);
 913         if (!pack)
 914                 goto out;
 915
 916         pos = 0;
 917
 918 found_free_area:
 919         bitmap_set(pack->bitmap, pos, nbits);
 920         ptr = (void *)(pack->ptr) + (pos << BPF_PROG_CHUNK_SHIFT);
 921
 922 out:
 923         mutex_unlock(&pack_mutex);
 924         return ptr;
 925 }
 926
 927 void bpf_prog_pack_free(struct bpf_binary_header *hdr)
 928 {
 929         struct bpf_prog_pack *pack = NULL, *tmp;
 930         unsigned int nbits;
 931         unsigned long pos;
 932
 933         mutex_lock(&pack_mutex);
 934         if (hdr->size > BPF_PROG_PACK_SIZE) {
 935                 module_memfree(hdr);
 936                 goto out;
 937         }
 938
 939         list_for_each_entry(tmp, &pack_list, list) {
 940                 if ((void *)hdr >= tmp->ptr && (tmp->ptr + BPF_PROG_PACK_SIZE) > (void *)hdr) {
 941                         pack = tmp;
 942                         break;
 943                 }
 944         }
 945
 946         if (WARN_ONCE(!pack, "bpf_prog_pack bug\n"))
 947                 goto out;
 948
 949         nbits = BPF_PROG_SIZE_TO_NBITS(hdr->size);
 950         pos = ((unsigned long)hdr - (unsigned long)pack->ptr) >> BPF_PROG_CHUNK_SHIFT;
 951
 952         WARN_ONCE(bpf_arch_text_invalidate(hdr, hdr->size),
 953                   "bpf_prog_pack bug: missing bpf_arch_text_invalidate?\n");
 954
 955         bitmap_clear(pack->bitmap, pos, nbits);
 956         if (bitmap_find_next_zero_area(pack->bitmap, BPF_PROG_CHUNK_COUNT, 0,
 957                                        BPF_PROG_CHUNK_COUNT, 0) == 0) {
 958                 list_del(&pack->list);
 959                 module_memfree(pack->ptr);
 960                 kfree(pack);
 961         }
 962 out:
 963         mutex_unlock(&pack_mutex);
 964 }
 965
 966 static atomic_long_t bpf_jit_current;
 967
 968 /* Can be overridden by an arch's JIT compiler if it has a custom,
 969  * dedicated BPF backend memory area, or if neither of the two
 970  * below apply.
 971  */
 972 u64 __weak bpf_jit_alloc_exec_limit(void)
 973 {
 974 #if defined(MODULES_VADDR)
 975         return MODULES_END - MODULES_VADDR;
 976 #else
 977         return VMALLOC_END - VMALLOC_START;
 978 #endif
 979 }
 980
 981 static int __init bpf_jit_charge_init(void)
 982 {
 983         /* Only used as heuristic here to derive limit. */
 984         bpf_jit_limit_max = bpf_jit_alloc_exec_limit();
 985         bpf_jit_limit = min_t(u64, round_up(bpf_jit_limit_max >> 1,
 986                                             PAGE_SIZE), LONG_MAX);
 987         return 0;
 988 }
 989 pure_initcall(bpf_jit_charge_init);
 990
 991 int bpf_jit_charge_modmem(u32 size)
 992 {
 993         if (atomic_long_add_return(size, &bpf_jit_current) > READ_ONCE(bpf_jit_limit)) {
 994                 if (!bpf_capable()) {
 995                         atomic_long_sub(size, &bpf_jit_current);
 996                         return -EPERM;
 997                 }
 998         }
 999
1000         return 0;
1001 }
1002
1003 void bpf_jit_uncharge_modmem(u32 size)
1004 {
1005         atomic_long_sub(size, &bpf_jit_current);
1006 }
1007
1008 void *__weak bpf_jit_alloc_exec(unsigned long size)
1009 {
1010         return module_alloc(size);
1011 }
1012
1013 void __weak bpf_jit_free_exec(void *addr)
1014 {
1015         module_memfree(addr);
1016 }
1017
1018 struct bpf_binary_header *
1019 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
1020                      unsigned int alignment,
1021                      bpf_jit_fill_hole_t bpf_fill_ill_insns)
1022 {
1023         struct bpf_binary_header *hdr;
1024         u32 size, hole, start;
1025
1026         WARN_ON_ONCE(!is_power_of_2(alignment) ||
1027                      alignment > BPF_IMAGE_ALIGNMENT);
1028
1029         /* Most of BPF filters are really small, but if some of them
1030          * fill a page, allow at least 128 extra bytes to insert a
1031          * random section of illegal instructions.
1032          */
1033         size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
1034
1035         if (bpf_jit_charge_modmem(size))
1036                 return NULL;
1037         hdr = bpf_jit_alloc_exec(size);
1038         if (!hdr) {
1039                 bpf_jit_uncharge_modmem(size);
1040                 return NULL;
1041         }
1042
1043         /* Fill space with illegal/arch-dep instructions. */
1044         bpf_fill_ill_insns(hdr, size);
1045
1046         hdr->size = size;
1047         hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
1048                      PAGE_SIZE - sizeof(*hdr));
1049         start = get_random_u32_below(hole) & ~(alignment - 1);
1050
1051         /* Leave a random number of instructions before BPF code. */
1052         *image_ptr = &hdr->image[start];
1053
1054         return hdr;
1055 }
1056
1057 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
1058 {
1059         u32 size = hdr->size;
1060
1061         bpf_jit_free_exec(hdr);
1062         bpf_jit_uncharge_modmem(size);
1063 }
1064
1065 /* Allocate jit binary from bpf_prog_pack allocator.
1066  * Since the allocated memory is RO+X, the JIT engine cannot write directly
1067  * to the memory. To solve this problem, a RW buffer is also allocated at
1068  * as the same time. The JIT engine should calculate offsets based on the
1069  * RO memory address, but write JITed program to the RW buffer. Once the
1070  * JIT engine finishes, it calls bpf_jit_binary_pack_finalize, which copies
1071  * the JITed program to the RO memory.
1072  */
1073 struct bpf_binary_header *
1074 bpf_jit_binary_pack_alloc(unsigned int proglen, u8 **image_ptr,
1075                           unsigned int alignment,
1076                           struct bpf_binary_header **rw_header,
1077                           u8 **rw_image,
1078                           bpf_jit_fill_hole_t bpf_fill_ill_insns)
1079 {
1080         struct bpf_binary_header *ro_header;
1081         u32 size, hole, start;
1082
1083         WARN_ON_ONCE(!is_power_of_2(alignment) ||
1084                      alignment > BPF_IMAGE_ALIGNMENT);
1085
1086         /* add 16 bytes for a random section of illegal instructions */
1087         size = round_up(proglen + sizeof(*ro_header) + 16, BPF_PROG_CHUNK_SIZE);
1088
1089         if (bpf_jit_charge_modmem(size))
1090                 return NULL;
1091         ro_header = bpf_prog_pack_alloc(size, bpf_fill_ill_insns);
1092         if (!ro_header) {
1093                 bpf_jit_uncharge_modmem(size);
1094                 return NULL;
1095         }
1096
1097         *rw_header = kvmalloc(size, GFP_KERNEL);
1098         if (!*rw_header) {
1099                 bpf_arch_text_copy(&ro_header->size, &size, sizeof(size));
1100                 bpf_prog_pack_free(ro_header);
1101                 bpf_jit_uncharge_modmem(size);
1102                 return NULL;
1103         }
1104
1105         /* Fill space with illegal/arch-dep instructions. */
1106         bpf_fill_ill_insns(*rw_header, size);
1107         (*rw_header)->size = size;
1108
1109         hole = min_t(unsigned int, size - (proglen + sizeof(*ro_header)),
1110                      BPF_PROG_CHUNK_SIZE - sizeof(*ro_header));
1111         start = get_random_u32_below(hole) & ~(alignment - 1);
1112
1113         *image_ptr = &ro_header->image[start];
1114         *rw_image = &(*rw_header)->image[start];
1115
1116         return ro_header;
1117 }
1118
1119 /* Copy JITed text from rw_header to its final location, the ro_header. */
1120 int bpf_jit_binary_pack_finalize(struct bpf_prog *prog,
1121                                  struct bpf_binary_header *ro_header,
1122                                  struct bpf_binary_header *rw_header)
1123 {
1124         void *ptr;
1125
1126         ptr = bpf_arch_text_copy(ro_header, rw_header, rw_header->size);
1127
1128         kvfree(rw_header);
1129
1130         if (IS_ERR(ptr)) {
1131                 bpf_prog_pack_free(ro_header);
1132                 return PTR_ERR(ptr);
1133         }
1134         return 0;
1135 }
1136
1137 /* bpf_jit_binary_pack_free is called in two different scenarios:
1138  *   1) when the program is freed after;
1139  *   2) when the JIT engine fails (before bpf_jit_binary_pack_finalize).
1140  * For case 2), we need to free both the RO memory and the RW buffer.
1141  *
1142  * bpf_jit_binary_pack_free requires proper ro_header->size. However,
1143  * bpf_jit_binary_pack_alloc does not set it. Therefore, ro_header->size
1144  * must be set with either bpf_jit_binary_pack_finalize (normal path) or
1145  * bpf_arch_text_copy (when jit fails).
1146  */
1147 void bpf_jit_binary_pack_free(struct bpf_binary_header *ro_header,
1148                               struct bpf_binary_header *rw_header)
1149 {
1150         u32 size = ro_header->size;
1151
1152         bpf_prog_pack_free(ro_header);
1153         kvfree(rw_header);
1154         bpf_jit_uncharge_modmem(size);
1155 }
1156
1157 struct bpf_binary_header *
1158 bpf_jit_binary_pack_hdr(const struct bpf_prog *fp)
1159 {
1160         unsigned long real_start = (unsigned long)fp->bpf_func;
1161         unsigned long addr;
1162
1163         addr = real_start & BPF_PROG_CHUNK_MASK;
1164         return (void *)addr;
1165 }
1166
1167 static inline struct bpf_binary_header *
1168 bpf_jit_binary_hdr(const struct bpf_prog *fp)
1169 {
1170         unsigned long real_start = (unsigned long)fp->bpf_func;
1171         unsigned long addr;
1172
1173         addr = real_start & PAGE_MASK;
1174         return (void *)addr;
1175 }
1176
1177 /* This symbol is only overridden by archs that have different
1178  * requirements than the usual eBPF JITs, f.e. when they only
1179  * implement cBPF JIT, do not set images read-only, etc.
1180  */
1181 void __weak bpf_jit_free(struct bpf_prog *fp)
1182 {
1183         if (fp->jited) {
1184                 struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
1185
1186                 bpf_jit_binary_free(hdr);
1187                 WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
1188         }
1189
1190         bpf_prog_unlock_free(fp);
1191 }
1192
1193 int bpf_jit_get_func_addr(const struct bpf_prog *prog,
1194                           const struct bpf_insn *insn, bool extra_pass,
1195                           u64 *func_addr, bool *func_addr_fixed)
1196 {
1197         s16 off = insn->off;
1198         s32 imm = insn->imm;
1199         u8 *addr;
1200         int err;
1201
1202         *func_addr_fixed = insn->src_reg != BPF_PSEUDO_CALL;
1203         if (!*func_addr_fixed) {
1204                 /* Place-holder address till the last pass has collected
1205                  * all addresses for JITed subprograms in which case we
1206                  * can pick them up from prog->aux.
1207                  */
1208                 if (!extra_pass)
1209                         addr = NULL;
1210                 else if (prog->aux->func &&
1211                          off >= 0 && off < prog->aux->func_cnt)
1212                         addr = (u8 *)prog->aux->func[off]->bpf_func;
1213                 else
1214                         return -EINVAL;
1215         } else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
1216                    bpf_jit_supports_far_kfunc_call()) {
1217                 err = bpf_get_kfunc_addr(prog, insn->imm, insn->off, &addr);
1218                 if (err)
1219                         return err;
1220         } else {
1221                 /* Address of a BPF helper call. Since part of the core
1222                  * kernel, it's always at a fixed location. __bpf_call_base
1223                  * and the helper with imm relative to it are both in core
1224                  * kernel.
1225                  */
1226                 addr = (u8 *)__bpf_call_base + imm;
1227         }
1228
1229         *func_addr = (unsigned long)addr;
1230         return 0;
1231 }
1232
1233 static int bpf_jit_blind_insn(const struct bpf_insn *from,
1234                               const struct bpf_insn *aux,
1235                               struct bpf_insn *to_buff,
1236                               bool emit_zext)
1237 {
1238         struct bpf_insn *to = to_buff;
1239         u32 imm_rnd = get_random_u32();
1240         s16 off;
1241
1242         BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
1243         BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
1244
1245         /* Constraints on AX register:
1246          *
1247          * AX register is inaccessible from user space. It is mapped in
1248          * all JITs, and used here for constant blinding rewrites. It is
1249          * typically "stateless" meaning its contents are only valid within
1250          * the executed instruction, but not across several instructions.
1251          * There are a few exceptions however which are further detailed
1252          * below.
1253          *
1254          * Constant blinding is only used by JITs, not in the interpreter.
1255          * The interpreter uses AX in some occasions as a local temporary
1256          * register e.g. in DIV or MOD instructions.
1257          *
1258          * In restricted circumstances, the verifier can also use the AX
1259          * register for rewrites as long as they do not interfere with
1260          * the above cases!
1261          */
1262         if (from->dst_reg == BPF_REG_AX || from->src_reg == BPF_REG_AX)
1263                 goto out;
1264
1265         if (from->imm == 0 &&
1266             (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
1267              from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
1268                 *to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
1269                 goto out;
1270         }
1271
1272         switch (from->code) {
1273         case BPF_ALU | BPF_ADD | BPF_K:
1274         case BPF_ALU | BPF_SUB | BPF_K:
1275         case BPF_ALU | BPF_AND | BPF_K:
1276         case BPF_ALU | BPF_OR  | BPF_K:
1277         case BPF_ALU | BPF_XOR | BPF_K:
1278         case BPF_ALU | BPF_MUL | BPF_K:
1279         case BPF_ALU | BPF_MOV | BPF_K:
1280         case BPF_ALU | BPF_DIV | BPF_K:
1281         case BPF_ALU | BPF_MOD | BPF_K:
1282                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1283                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1284                 *to++ = BPF_ALU32_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1285                 break;
1286
1287         case BPF_ALU64 | BPF_ADD | BPF_K:
1288         case BPF_ALU64 | BPF_SUB | BPF_K:
1289         case BPF_ALU64 | BPF_AND | BPF_K:
1290         case BPF_ALU64 | BPF_OR  | BPF_K:
1291         case BPF_ALU64 | BPF_XOR | BPF_K:
1292         case BPF_ALU64 | BPF_MUL | BPF_K:
1293         case BPF_ALU64 | BPF_MOV | BPF_K:
1294         case BPF_ALU64 | BPF_DIV | BPF_K:
1295         case BPF_ALU64 | BPF_MOD | BPF_K:
1296                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1297                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1298                 *to++ = BPF_ALU64_REG_OFF(from->code, from->dst_reg, BPF_REG_AX, from->off);
1299                 break;
1300
1301         case BPF_JMP | BPF_JEQ  | BPF_K:
1302         case BPF_JMP | BPF_JNE  | BPF_K:
1303         case BPF_JMP | BPF_JGT  | BPF_K:
1304         case BPF_JMP | BPF_JLT  | BPF_K:
1305         case BPF_JMP | BPF_JGE  | BPF_K:
1306         case BPF_JMP | BPF_JLE  | BPF_K:
1307         case BPF_JMP | BPF_JSGT | BPF_K:
1308         case BPF_JMP | BPF_JSLT | BPF_K:
1309         case BPF_JMP | BPF_JSGE | BPF_K:
1310         case BPF_JMP | BPF_JSLE | BPF_K:
1311         case BPF_JMP | BPF_JSET | BPF_K:
1312                 /* Accommodate for extra offset in case of a backjump. */
1313                 off = from->off;
1314                 if (off < 0)
1315                         off -= 2;
1316                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1317                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1318                 *to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
1319                 break;
1320
1321         case BPF_JMP32 | BPF_JEQ  | BPF_K:
1322         case BPF_JMP32 | BPF_JNE  | BPF_K:
1323         case BPF_JMP32 | BPF_JGT  | BPF_K:
1324         case BPF_JMP32 | BPF_JLT  | BPF_K:
1325         case BPF_JMP32 | BPF_JGE  | BPF_K:
1326         case BPF_JMP32 | BPF_JLE  | BPF_K:
1327         case BPF_JMP32 | BPF_JSGT | BPF_K:
1328         case BPF_JMP32 | BPF_JSLT | BPF_K:
1329         case BPF_JMP32 | BPF_JSGE | BPF_K:
1330         case BPF_JMP32 | BPF_JSLE | BPF_K:
1331         case BPF_JMP32 | BPF_JSET | BPF_K:
1332                 /* Accommodate for extra offset in case of a backjump. */
1333                 off = from->off;
1334                 if (off < 0)
1335                         off -= 2;
1336                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1337                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1338                 *to++ = BPF_JMP32_REG(from->code, from->dst_reg, BPF_REG_AX,
1339                                       off);
1340                 break;
1341
1342         case BPF_LD | BPF_IMM | BPF_DW:
1343                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
1344                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1345                 *to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
1346                 *to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
1347                 break;
1348         case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
1349                 *to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
1350                 *to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1351                 if (emit_zext)
1352                         *to++ = BPF_ZEXT_REG(BPF_REG_AX);
1353                 *to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
1354                 break;
1355
1356         case BPF_ST | BPF_MEM | BPF_DW:
1357         case BPF_ST | BPF_MEM | BPF_W:
1358         case BPF_ST | BPF_MEM | BPF_H:
1359         case BPF_ST | BPF_MEM | BPF_B:
1360                 *to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
1361                 *to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
1362                 *to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
1363                 break;
1364         }
1365 out:
1366         return to - to_buff;
1367 }
1368
1369 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
1370                                               gfp_t gfp_extra_flags)
1371 {
1372         gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
1373         struct bpf_prog *fp;
1374
1375         fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags);
1376         if (fp != NULL) {
1377                 /* aux->prog still points to the fp_other one, so
1378                  * when promoting the clone to the real program,
1379                  * this still needs to be adapted.
1380                  */
1381                 memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
1382         }
1383
1384         return fp;
1385 }
1386
1387 static void bpf_prog_clone_free(struct bpf_prog *fp)
1388 {
1389         /* aux was stolen by the other clone, so we cannot free
1390          * it from this path! It will be freed eventually by the
1391          * other program on release.
1392          *
1393          * At this point, we don't need a deferred release since
1394          * clone is guaranteed to not be locked.
1395          */
1396         fp->aux = NULL;
1397         fp->stats = NULL;
1398         fp->active = NULL;
1399         __bpf_prog_free(fp);
1400 }
1401
1402 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
1403 {
1404         /* We have to repoint aux->prog to self, as we don't
1405          * know whether fp here is the clone or the original.
1406          */
1407         fp->aux->prog = fp;
1408         bpf_prog_clone_free(fp_other);
1409 }
1410
1411 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
1412 {
1413         struct bpf_insn insn_buff[16], aux[2];
1414         struct bpf_prog *clone, *tmp;
1415         int insn_delta, insn_cnt;
1416         struct bpf_insn *insn;
1417         int i, rewritten;
1418
1419         if (!prog->blinding_requested || prog->blinded)
1420                 return prog;
1421
1422         clone = bpf_prog_clone_create(prog, GFP_USER);
1423         if (!clone)
1424                 return ERR_PTR(-ENOMEM);
1425
1426         insn_cnt = clone->len;
1427         insn = clone->insnsi;
1428
1429         for (i = 0; i < insn_cnt; i++, insn++) {
1430                 if (bpf_pseudo_func(insn)) {
1431                         /* ld_imm64 with an address of bpf subprog is not
1432                          * a user controlled constant. Don't randomize it,
1433                          * since it will conflict with jit_subprogs() logic.
1434                          */
1435                         insn++;
1436                         i++;
1437                         continue;
1438                 }
1439
1440                 /* We temporarily need to hold the original ld64 insn
1441                  * so that we can still access the first part in the
1442                  * second blinding run.
1443                  */
1444                 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
1445                     insn[1].code == 0)
1446                         memcpy(aux, insn, sizeof(aux));
1447
1448                 rewritten = bpf_jit_blind_insn(insn, aux, insn_buff,
1449                                                 clone->aux->verifier_zext);
1450                 if (!rewritten)
1451                         continue;
1452
1453                 tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
1454                 if (IS_ERR(tmp)) {
1455                         /* Patching may have repointed aux->prog during
1456                          * realloc from the original one, so we need to
1457                          * fix it up here on error.
1458                          */
1459                         bpf_jit_prog_release_other(prog, clone);
1460                         return tmp;
1461                 }
1462
1463                 clone = tmp;
1464                 insn_delta = rewritten - 1;
1465
1466                 /* Walk new program and skip insns we just inserted. */
1467                 insn = clone->insnsi + i + insn_delta;
1468                 insn_cnt += insn_delta;
1469                 i        += insn_delta;
1470         }
1471
1472         clone->blinded = 1;
1473         return clone;
1474 }
1475 #endif /* CONFIG_BPF_JIT */
1476
1477 /* Base function for offset calculation. Needs to go into .text section,
1478  * therefore keeping it non-static as well; will also be used by JITs
1479  * anyway later on, so do not let the compiler omit it. This also needs
1480  * to go into kallsyms for correlation from e.g. bpftool, so naming
1481  * must not change.
1482  */
1483 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1484 {
1485         return 0;
1486 }
1487 EXPORT_SYMBOL_GPL(__bpf_call_base);
1488
1489 /* All UAPI available opcodes. */
1490 #define BPF_INSN_MAP(INSN_2, INSN_3)            \
1491         /* 32 bit ALU operations. */            \
1492         /*   Register based. */                 \
1493         INSN_3(ALU, ADD,  X),                   \
1494         INSN_3(ALU, SUB,  X),                   \
1495         INSN_3(ALU, AND,  X),                   \
1496         INSN_3(ALU, OR,   X),                   \
1497         INSN_3(ALU, LSH,  X),                   \
1498         INSN_3(ALU, RSH,  X),                   \
1499         INSN_3(ALU, XOR,  X),                   \
1500         INSN_3(ALU, MUL,  X),                   \
1501         INSN_3(ALU, MOV,  X),                   \
1502         INSN_3(ALU, ARSH, X),                   \
1503         INSN_3(ALU, DIV,  X),                   \
1504         INSN_3(ALU, MOD,  X),                   \
1505         INSN_2(ALU, NEG),                       \
1506         INSN_3(ALU, END, TO_BE),                \
1507         INSN_3(ALU, END, TO_LE),                \
1508         /*   Immediate based. */                \
1509         INSN_3(ALU, ADD,  K),                   \
1510         INSN_3(ALU, SUB,  K),                   \
1511         INSN_3(ALU, AND,  K),                   \
1512         INSN_3(ALU, OR,   K),                   \
1513         INSN_3(ALU, LSH,  K),                   \
1514         INSN_3(ALU, RSH,  K),                   \
1515         INSN_3(ALU, XOR,  K),                   \
1516         INSN_3(ALU, MUL,  K),                   \
1517         INSN_3(ALU, MOV,  K),                   \
1518         INSN_3(ALU, ARSH, K),                   \
1519         INSN_3(ALU, DIV,  K),                   \
1520         INSN_3(ALU, MOD,  K),                   \
1521         /* 64 bit ALU operations. */            \
1522         /*   Register based. */                 \
1523         INSN_3(ALU64, ADD,  X),                 \
1524         INSN_3(ALU64, SUB,  X),                 \
1525         INSN_3(ALU64, AND,  X),                 \
1526         INSN_3(ALU64, OR,   X),                 \
1527         INSN_3(ALU64, LSH,  X),                 \
1528         INSN_3(ALU64, RSH,  X),                 \
1529         INSN_3(ALU64, XOR,  X),                 \
1530         INSN_3(ALU64, MUL,  X),                 \
1531         INSN_3(ALU64, MOV,  X),                 \
1532         INSN_3(ALU64, ARSH, X),                 \
1533         INSN_3(ALU64, DIV,  X),                 \
1534         INSN_3(ALU64, MOD,  X),                 \
1535         INSN_2(ALU64, NEG),                     \
1536         INSN_3(ALU64, END, TO_LE),              \
1537         /*   Immediate based. */                \
1538         INSN_3(ALU64, ADD,  K),                 \
1539         INSN_3(ALU64, SUB,  K),                 \
1540         INSN_3(ALU64, AND,  K),                 \
1541         INSN_3(ALU64, OR,   K),                 \
1542         INSN_3(ALU64, LSH,  K),                 \
1543         INSN_3(ALU64, RSH,  K),                 \
1544         INSN_3(ALU64, XOR,  K),                 \
1545         INSN_3(ALU64, MUL,  K),                 \
1546         INSN_3(ALU64, MOV,  K),                 \
1547         INSN_3(ALU64, ARSH, K),                 \
1548         INSN_3(ALU64, DIV,  K),                 \
1549         INSN_3(ALU64, MOD,  K),                 \
1550         /* Call instruction. */                 \
1551         INSN_2(JMP, CALL),                      \
1552         /* Exit instruction. */                 \
1553         INSN_2(JMP, EXIT),                      \
1554         /* 32-bit Jump instructions. */         \
1555         /*   Register based. */                 \
1556         INSN_3(JMP32, JEQ,  X),                 \
1557         INSN_3(JMP32, JNE,  X),                 \
1558         INSN_3(JMP32, JGT,  X),                 \
1559         INSN_3(JMP32, JLT,  X),                 \
1560         INSN_3(JMP32, JGE,  X),                 \
1561         INSN_3(JMP32, JLE,  X),                 \
1562         INSN_3(JMP32, JSGT, X),                 \
1563         INSN_3(JMP32, JSLT, X),                 \
1564         INSN_3(JMP32, JSGE, X),                 \
1565         INSN_3(JMP32, JSLE, X),                 \
1566         INSN_3(JMP32, JSET, X),                 \
1567         /*   Immediate based. */                \
1568         INSN_3(JMP32, JEQ,  K),                 \
1569         INSN_3(JMP32, JNE,  K),                 \
1570         INSN_3(JMP32, JGT,  K),                 \
1571         INSN_3(JMP32, JLT,  K),                 \
1572         INSN_3(JMP32, JGE,  K),                 \
1573         INSN_3(JMP32, JLE,  K),                 \
1574         INSN_3(JMP32, JSGT, K),                 \
1575         INSN_3(JMP32, JSLT, K),                 \
1576         INSN_3(JMP32, JSGE, K),                 \
1577         INSN_3(JMP32, JSLE, K),                 \
1578         INSN_3(JMP32, JSET, K),                 \
1579         /* Jump instructions. */                \
1580         /*   Register based. */                 \
1581         INSN_3(JMP, JEQ,  X),                   \
1582         INSN_3(JMP, JNE,  X),                   \
1583         INSN_3(JMP, JGT,  X),                   \
1584         INSN_3(JMP, JLT,  X),                   \
1585         INSN_3(JMP, JGE,  X),                   \
1586         INSN_3(JMP, JLE,  X),                   \
1587         INSN_3(JMP, JSGT, X),                   \
1588         INSN_3(JMP, JSLT, X),                   \
1589         INSN_3(JMP, JSGE, X),                   \
1590         INSN_3(JMP, JSLE, X),                   \
1591         INSN_3(JMP, JSET, X),                   \
1592         /*   Immediate based. */                \
1593         INSN_3(JMP, JEQ,  K),                   \
1594         INSN_3(JMP, JNE,  K),                   \
1595         INSN_3(JMP, JGT,  K),                   \
1596         INSN_3(JMP, JLT,  K),                   \
1597         INSN_3(JMP, JGE,  K),                   \
1598         INSN_3(JMP, JLE,  K),                   \
1599         INSN_3(JMP, JSGT, K),                   \
1600         INSN_3(JMP, JSLT, K),                   \
1601         INSN_3(JMP, JSGE, K),                   \
1602         INSN_3(JMP, JSLE, K),                   \
1603         INSN_3(JMP, JSET, K),                   \
1604         INSN_2(JMP, JA),                        \
1605         INSN_2(JMP32, JA),                      \
1606         /* Store instructions. */               \
1607         /*   Register based. */                 \
1608         INSN_3(STX, MEM,  B),                   \
1609         INSN_3(STX, MEM,  H),                   \
1610         INSN_3(STX, MEM,  W),                   \
1611         INSN_3(STX, MEM,  DW),                  \
1612         INSN_3(STX, ATOMIC, W),                 \
1613         INSN_3(STX, ATOMIC, DW),                \
1614         /*   Immediate based. */                \
1615         INSN_3(ST, MEM, B),                     \
1616         INSN_3(ST, MEM, H),                     \
1617         INSN_3(ST, MEM, W),                     \
1618         INSN_3(ST, MEM, DW),                    \
1619         /* Load instructions. */                \
1620         /*   Register based. */                 \
1621         INSN_3(LDX, MEM, B),                    \
1622         INSN_3(LDX, MEM, H),                    \
1623         INSN_3(LDX, MEM, W),                    \
1624         INSN_3(LDX, MEM, DW),                   \
1625         INSN_3(LDX, MEMSX, B),                  \
1626         INSN_3(LDX, MEMSX, H),                  \
1627         INSN_3(LDX, MEMSX, W),                  \
1628         /*   Immediate based. */                \
1629         INSN_3(LD, IMM, DW)
1630
1631 bool bpf_opcode_in_insntable(u8 code)
1632 {
1633 #define BPF_INSN_2_TBL(x, y)    [BPF_##x | BPF_##y] = true
1634 #define BPF_INSN_3_TBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = true
1635         static const bool public_insntable[256] = {
1636                 [0 ... 255] = false,
1637                 /* Now overwrite non-defaults ... */
1638                 BPF_INSN_MAP(BPF_INSN_2_TBL, BPF_INSN_3_TBL),
1639                 /* UAPI exposed, but rewritten opcodes. cBPF carry-over. */
1640                 [BPF_LD | BPF_ABS | BPF_B] = true,
1641                 [BPF_LD | BPF_ABS | BPF_H] = true,
1642                 [BPF_LD | BPF_ABS | BPF_W] = true,
1643                 [BPF_LD | BPF_IND | BPF_B] = true,
1644                 [BPF_LD | BPF_IND | BPF_H] = true,
1645                 [BPF_LD | BPF_IND | BPF_W] = true,
1646         };
1647 #undef BPF_INSN_3_TBL
1648 #undef BPF_INSN_2_TBL
1649         return public_insntable[code];
1650 }
1651
1652 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1653 /**
1654  *      ___bpf_prog_run - run eBPF program on a given context
1655  *      @regs: is the array of MAX_BPF_EXT_REG eBPF pseudo-registers
1656  *      @insn: is the array of eBPF instructions
1657  *
1658  * Decode and execute eBPF instructions.
1659  *
1660  * Return: whatever value is in %BPF_R0 at program exit
1661  */
1662 static u64 ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn)
1663 {
1664 #define BPF_INSN_2_LBL(x, y)    [BPF_##x | BPF_##y] = &&x##_##y
1665 #define BPF_INSN_3_LBL(x, y, z) [BPF_##x | BPF_##y | BPF_##z] = &&x##_##y##_##z
1666         static const void * const jumptable[256] __annotate_jump_table = {
1667                 [0 ... 255] = &&default_label,
1668                 /* Now overwrite non-defaults ... */
1669                 BPF_INSN_MAP(BPF_INSN_2_LBL, BPF_INSN_3_LBL),
1670                 /* Non-UAPI available opcodes. */
1671                 [BPF_JMP | BPF_CALL_ARGS] = &&JMP_CALL_ARGS,
1672                 [BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
1673                 [BPF_ST  | BPF_NOSPEC] = &&ST_NOSPEC,
1674                 [BPF_LDX | BPF_PROBE_MEM | BPF_B] = &&LDX_PROBE_MEM_B,
1675                 [BPF_LDX | BPF_PROBE_MEM | BPF_H] = &&LDX_PROBE_MEM_H,
1676                 [BPF_LDX | BPF_PROBE_MEM | BPF_W] = &&LDX_PROBE_MEM_W,
1677                 [BPF_LDX | BPF_PROBE_MEM | BPF_DW] = &&LDX_PROBE_MEM_DW,
1678                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_B] = &&LDX_PROBE_MEMSX_B,
1679                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_H] = &&LDX_PROBE_MEMSX_H,
1680                 [BPF_LDX | BPF_PROBE_MEMSX | BPF_W] = &&LDX_PROBE_MEMSX_W,
1681         };
1682 #undef BPF_INSN_3_LBL
1683 #undef BPF_INSN_2_LBL
1684         u32 tail_call_cnt = 0;
1685
1686 #define CONT     ({ insn++; goto select_insn; })
1687 #define CONT_JMP ({ insn++; goto select_insn; })
1688
1689 select_insn:
1690         goto *jumptable[insn->code];
1691
1692         /* Explicitly mask the register-based shift amounts with 63 or 31
1693          * to avoid undefined behavior. Normally this won't affect the
1694          * generated code, for example, in case of native 64 bit archs such
1695          * as x86-64 or arm64, the compiler is optimizing the AND away for
1696          * the interpreter. In case of JITs, each of the JIT backends compiles
1697          * the BPF shift operations to machine instructions which produce
1698          * implementation-defined results in such a case; the resulting
1699          * contents of the register may be arbitrary, but program behaviour
1700          * as a whole remains defined. In other words, in case of JIT backends,
1701          * the AND must /not/ be added to the emitted LSH/RSH/ARSH translation.
1702          */
1703         /* ALU (shifts) */
1704 #define SHT(OPCODE, OP)                                 \
1705         ALU64_##OPCODE##_X:                             \
1706                 DST = DST OP (SRC & 63);                \
1707                 CONT;                                   \
1708         ALU_##OPCODE##_X:                               \
1709                 DST = (u32) DST OP ((u32) SRC & 31);    \
1710                 CONT;                                   \
1711         ALU64_##OPCODE##_K:                             \
1712                 DST = DST OP IMM;                       \
1713                 CONT;                                   \
1714         ALU_##OPCODE##_K:                               \
1715                 DST = (u32) DST OP (u32) IMM;           \
1716                 CONT;
1717         /* ALU (rest) */
1718 #define ALU(OPCODE, OP)                                 \
1719         ALU64_##OPCODE##_X:                             \
1720                 DST = DST OP SRC;                       \
1721                 CONT;                                   \
1722         ALU_##OPCODE##_X:                               \
1723                 DST = (u32) DST OP (u32) SRC;           \
1724                 CONT;                                   \
1725         ALU64_##OPCODE##_K:                             \
1726                 DST = DST OP IMM;                       \
1727                 CONT;                                   \
1728         ALU_##OPCODE##_K:                               \
1729                 DST = (u32) DST OP (u32) IMM;           \
1730                 CONT;
1731         ALU(ADD,  +)
1732         ALU(SUB,  -)
1733         ALU(AND,  &)
1734         ALU(OR,   |)
1735         ALU(XOR,  ^)
1736         ALU(MUL,  *)
1737         SHT(LSH, <<)
1738         SHT(RSH, >>)
1739 #undef SHT
1740 #undef ALU
1741         ALU_NEG:
1742                 DST = (u32) -DST;
1743                 CONT;
1744         ALU64_NEG:
1745                 DST = -DST;
1746                 CONT;
1747         ALU_MOV_X:
1748                 switch (OFF) {
1749                 case 0:
1750                         DST = (u32) SRC;
1751                         break;
1752                 case 8:
1753                         DST = (u32)(s8) SRC;
1754                         break;
1755                 case 16:
1756                         DST = (u32)(s16) SRC;
1757                         break;
1758                 }
1759                 CONT;
1760         ALU_MOV_K:
1761                 DST = (u32) IMM;
1762                 CONT;
1763         ALU64_MOV_X:
1764                 switch (OFF) {
1765                 case 0:
1766                         DST = SRC;
1767                         break;
1768                 case 8:
1769                         DST = (s8) SRC;
1770                         break;
1771                 case 16:
1772                         DST = (s16) SRC;
1773                         break;
1774                 case 32:
1775                         DST = (s32) SRC;
1776                         break;
1777                 }
1778                 CONT;
1779         ALU64_MOV_K:
1780                 DST = IMM;
1781                 CONT;
1782         LD_IMM_DW:
1783                 DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
1784                 insn++;
1785                 CONT;
1786         ALU_ARSH_X:
1787                 DST = (u64) (u32) (((s32) DST) >> (SRC & 31));
1788                 CONT;
1789         ALU_ARSH_K:
1790                 DST = (u64) (u32) (((s32) DST) >> IMM);
1791                 CONT;
1792         ALU64_ARSH_X:
1793                 (*(s64 *) &DST) >>= (SRC & 63);
1794                 CONT;
1795         ALU64_ARSH_K:
1796                 (*(s64 *) &DST) >>= IMM;
1797                 CONT;
1798         ALU64_MOD_X:
1799                 switch (OFF) {
1800                 case 0:
1801                         div64_u64_rem(DST, SRC, &AX);
1802                         DST = AX;
1803                         break;
1804                 case 1:
1805                         AX = div64_s64(DST, SRC);
1806                         DST = DST - AX * SRC;
1807                         break;
1808                 }
1809                 CONT;
1810         ALU_MOD_X:
1811                 switch (OFF) {
1812                 case 0:
1813                         AX = (u32) DST;
1814                         DST = do_div(AX, (u32) SRC);
1815                         break;
1816                 case 1:
1817                         AX = abs((s32)DST);
1818                         AX = do_div(AX, abs((s32)SRC));
1819                         if ((s32)DST < 0)
1820                                 DST = (u32)-AX;
1821                         else
1822                                 DST = (u32)AX;
1823                         break;
1824                 }
1825                 CONT;
1826         ALU64_MOD_K:
1827                 switch (OFF) {
1828                 case 0:
1829                         div64_u64_rem(DST, IMM, &AX);
1830                         DST = AX;
1831                         break;
1832                 case 1:
1833                         AX = div64_s64(DST, IMM);
1834                         DST = DST - AX * IMM;
1835                         break;
1836                 }
1837                 CONT;
1838         ALU_MOD_K:
1839                 switch (OFF) {
1840                 case 0:
1841                         AX = (u32) DST;
1842                         DST = do_div(AX, (u32) IMM);
1843                         break;
1844                 case 1:
1845                         AX = abs((s32)DST);
1846                         AX = do_div(AX, abs((s32)IMM));
1847                         if ((s32)DST < 0)
1848                                 DST = (u32)-AX;
1849                         else
1850                                 DST = (u32)AX;
1851                         break;
1852                 }
1853                 CONT;
1854         ALU64_DIV_X:
1855                 switch (OFF) {
1856                 case 0:
1857                         DST = div64_u64(DST, SRC);
1858                         break;
1859                 case 1:
1860                         DST = div64_s64(DST, SRC);
1861                         break;
1862                 }
1863                 CONT;
1864         ALU_DIV_X:
1865                 switch (OFF) {
1866                 case 0:
1867                         AX = (u32) DST;
1868                         do_div(AX, (u32) SRC);
1869                         DST = (u32) AX;
1870                         break;
1871                 case 1:
1872                         AX = abs((s32)DST);
1873                         do_div(AX, abs((s32)SRC));
1874                         if (((s32)DST < 0) == ((s32)SRC < 0))
1875                                 DST = (u32)AX;
1876                         else
1877                                 DST = (u32)-AX;
1878                         break;
1879                 }
1880                 CONT;
1881         ALU64_DIV_K:
1882                 switch (OFF) {
1883                 case 0:
1884                         DST = div64_u64(DST, IMM);
1885                         break;
1886                 case 1:
1887                         DST = div64_s64(DST, IMM);
1888                         break;
1889                 }
1890                 CONT;
1891         ALU_DIV_K:
1892                 switch (OFF) {
1893                 case 0:
1894                         AX = (u32) DST;
1895                         do_div(AX, (u32) IMM);
1896                         DST = (u32) AX;
1897                         break;
1898                 case 1:
1899                         AX = abs((s32)DST);
1900                         do_div(AX, abs((s32)IMM));
1901                         if (((s32)DST < 0) == ((s32)IMM < 0))
1902                                 DST = (u32)AX;
1903                         else
1904                                 DST = (u32)-AX;
1905                         break;
1906                 }
1907                 CONT;
1908         ALU_END_TO_BE:
1909                 switch (IMM) {
1910                 case 16:
1911                         DST = (__force u16) cpu_to_be16(DST);
1912                         break;
1913                 case 32:
1914                         DST = (__force u32) cpu_to_be32(DST);
1915                         break;
1916                 case 64:
1917                         DST = (__force u64) cpu_to_be64(DST);
1918                         break;
1919                 }
1920                 CONT;
1921         ALU_END_TO_LE:
1922                 switch (IMM) {
1923                 case 16:
1924                         DST = (__force u16) cpu_to_le16(DST);
1925                         break;
1926                 case 32:
1927                         DST = (__force u32) cpu_to_le32(DST);
1928                         break;
1929                 case 64:
1930                         DST = (__force u64) cpu_to_le64(DST);
1931                         break;
1932                 }
1933                 CONT;
1934         ALU64_END_TO_LE:
1935                 switch (IMM) {
1936                 case 16:
1937                         DST = (__force u16) __swab16(DST);
1938                         break;
1939                 case 32:
1940                         DST = (__force u32) __swab32(DST);
1941                         break;
1942                 case 64:
1943                         DST = (__force u64) __swab64(DST);
1944                         break;
1945                 }
1946                 CONT;
1947
1948         /* CALL */
1949         JMP_CALL:
1950                 /* Function call scratches BPF_R1-BPF_R5 registers,
1951                  * preserves BPF_R6-BPF_R9, and stores return value
1952                  * into BPF_R0.
1953                  */
1954                 BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1955                                                        BPF_R4, BPF_R5);
1956                 CONT;
1957
1958         JMP_CALL_ARGS:
1959                 BPF_R0 = (__bpf_call_base_args + insn->imm)(BPF_R1, BPF_R2,
1960                                                             BPF_R3, BPF_R4,
1961                                                             BPF_R5,
1962                                                             insn + insn->off + 1);
1963                 CONT;
1964
1965         JMP_TAIL_CALL: {
1966                 struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1967                 struct bpf_array *array = container_of(map, struct bpf_array, map);
1968                 struct bpf_prog *prog;
1969                 u32 index = BPF_R3;
1970
1971                 if (unlikely(index >= array->map.max_entries))
1972                         goto out;
1973
1974                 if (unlikely(tail_call_cnt >= MAX_TAIL_CALL_CNT))
1975                         goto out;
1976
1977                 tail_call_cnt++;
1978
1979                 prog = READ_ONCE(array->ptrs[index]);
1980                 if (!prog)
1981                         goto out;
1982
1983                 /* ARG1 at this point is guaranteed to point to CTX from
1984                  * the verifier side due to the fact that the tail call is
1985                  * handled like a helper, that is, bpf_tail_call_proto,
1986                  * where arg1_type is ARG_PTR_TO_CTX.
1987                  */
1988                 insn = prog->insnsi;
1989                 goto select_insn;
1990 out:
1991                 CONT;
1992         }
1993         JMP_JA:
1994                 insn += insn->off;
1995                 CONT;
1996         JMP32_JA:
1997                 insn += insn->imm;
1998                 CONT;
1999         JMP_EXIT:
2000                 return BPF_R0;
2001         /* JMP */
2002 #define COND_JMP(SIGN, OPCODE, CMP_OP)                          \
2003         JMP_##OPCODE##_X:                                       \
2004                 if ((SIGN##64) DST CMP_OP (SIGN##64) SRC) {     \
2005                         insn += insn->off;                      \
2006                         CONT_JMP;                               \
2007                 }                                               \
2008                 CONT;                                           \
2009         JMP32_##OPCODE##_X:                                     \
2010                 if ((SIGN##32) DST CMP_OP (SIGN##32) SRC) {     \
2011                         insn += insn->off;                      \
2012                         CONT_JMP;                               \
2013                 }                                               \
2014                 CONT;                                           \
2015         JMP_##OPCODE##_K:                                       \
2016                 if ((SIGN##64) DST CMP_OP (SIGN##64) IMM) {     \
2017                         insn += insn->off;                      \
2018                         CONT_JMP;                               \
2019                 }                                               \
2020                 CONT;                                           \
2021         JMP32_##OPCODE##_K:                                     \
2022                 if ((SIGN##32) DST CMP_OP (SIGN##32) IMM) {     \
2023                         insn += insn->off;                      \
2024                         CONT_JMP;                               \
2025                 }                                               \
2026                 CONT;
2027         COND_JMP(u, JEQ, ==)
2028         COND_JMP(u, JNE, !=)
2029         COND_JMP(u, JGT, >)
2030         COND_JMP(u, JLT, <)
2031         COND_JMP(u, JGE, >=)
2032         COND_JMP(u, JLE, <=)
2033         COND_JMP(u, JSET, &)
2034         COND_JMP(s, JSGT, >)
2035         COND_JMP(s, JSLT, <)
2036         COND_JMP(s, JSGE, >=)
2037         COND_JMP(s, JSLE, <=)
2038 #undef COND_JMP
2039         /* ST, STX and LDX*/
2040         ST_NOSPEC:
2041                 /* Speculation barrier for mitigating Speculative Store Bypass.
2042                  * In case of arm64, we rely on the firmware mitigation as
2043                  * controlled via the ssbd kernel parameter. Whenever the
2044                  * mitigation is enabled, it works for all of the kernel code
2045                  * with no need to provide any additional instructions here.
2046                  * In case of x86, we use 'lfence' insn for mitigation. We
2047                  * reuse preexisting logic from Spectre v1 mitigation that
2048                  * happens to produce the required code on x86 for v4 as well.
2049                  */
2050                 barrier_nospec();
2051                 CONT;
2052 #define LDST(SIZEOP, SIZE)                                              \
2053         STX_MEM_##SIZEOP:                                               \
2054                 *(SIZE *)(unsigned long) (DST + insn->off) = SRC;       \
2055                 CONT;                                                   \
2056         ST_MEM_##SIZEOP:                                                \
2057                 *(SIZE *)(unsigned long) (DST + insn->off) = IMM;       \
2058                 CONT;                                                   \
2059         LDX_MEM_##SIZEOP:                                               \
2060                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
2061                 CONT;                                                   \
2062         LDX_PROBE_MEM_##SIZEOP:                                         \
2063                 bpf_probe_read_kernel_common(&DST, sizeof(SIZE),        \
2064                               (const void *)(long) (SRC + insn->off));  \
2065                 DST = *((SIZE *)&DST);                                  \
2066                 CONT;
2067
2068         LDST(B,   u8)
2069         LDST(H,  u16)
2070         LDST(W,  u32)
2071         LDST(DW, u64)
2072 #undef LDST
2073
2074 #define LDSX(SIZEOP, SIZE)                                              \
2075         LDX_MEMSX_##SIZEOP:                                             \
2076                 DST = *(SIZE *)(unsigned long) (SRC + insn->off);       \
2077                 CONT;                                                   \
2078         LDX_PROBE_MEMSX_##SIZEOP:                                       \
2079                 bpf_probe_read_kernel_common(&DST, sizeof(SIZE),                \
2080                                       (const void *)(long) (SRC + insn->off));  \
2081                 DST = *((SIZE *)&DST);                                  \
2082                 CONT;
2083
2084         LDSX(B,   s8)
2085         LDSX(H,  s16)
2086         LDSX(W,  s32)
2087 #undef LDSX
2088
2089 #define ATOMIC_ALU_OP(BOP, KOP)                                         \
2090                 case BOP:                                               \
2091                         if (BPF_SIZE(insn->code) == BPF_W)              \
2092                                 atomic_##KOP((u32) SRC, (atomic_t *)(unsigned long) \
2093                                              (DST + insn->off));        \
2094                         else                                            \
2095                                 atomic64_##KOP((u64) SRC, (atomic64_t *)(unsigned long) \
2096                                                (DST + insn->off));      \
2097                         break;                                          \
2098                 case BOP | BPF_FETCH:                                   \
2099                         if (BPF_SIZE(insn->code) == BPF_W)              \
2100                                 SRC = (u32) atomic_fetch_##KOP(         \
2101                                         (u32) SRC,                      \
2102                                         (atomic_t *)(unsigned long) (DST + insn->off)); \
2103                         else                                            \
2104                                 SRC = (u64) atomic64_fetch_##KOP(       \
2105                                         (u64) SRC,                      \
2106                                         (atomic64_t *)(unsigned long) (DST + insn->off)); \
2107                         break;
2108
2109         STX_ATOMIC_DW:
2110         STX_ATOMIC_W:
2111                 switch (IMM) {
2112                 ATOMIC_ALU_OP(BPF_ADD, add)
2113                 ATOMIC_ALU_OP(BPF_AND, and)
2114                 ATOMIC_ALU_OP(BPF_OR, or)
2115                 ATOMIC_ALU_OP(BPF_XOR, xor)
2116 #undef ATOMIC_ALU_OP
2117
2118                 case BPF_XCHG:
2119                         if (BPF_SIZE(insn->code) == BPF_W)
2120                                 SRC = (u32) atomic_xchg(
2121                                         (atomic_t *)(unsigned long) (DST + insn->off),
2122                                         (u32) SRC);
2123                         else
2124                                 SRC = (u64) atomic64_xchg(
2125                                         (atomic64_t *)(unsigned long) (DST + insn->off),
2126                                         (u64) SRC);
2127                         break;
2128                 case BPF_CMPXCHG:
2129                         if (BPF_SIZE(insn->code) == BPF_W)
2130                                 BPF_R0 = (u32) atomic_cmpxchg(
2131                                         (atomic_t *)(unsigned long) (DST + insn->off),
2132                                         (u32) BPF_R0, (u32) SRC);
2133                         else
2134                                 BPF_R0 = (u64) atomic64_cmpxchg(
2135                                         (atomic64_t *)(unsigned long) (DST + insn->off),
2136                                         (u64) BPF_R0, (u64) SRC);
2137                         break;
2138
2139                 default:
2140                         goto default_label;
2141                 }
2142                 CONT;
2143
2144         default_label:
2145                 /* If we ever reach this, we have a bug somewhere. Die hard here
2146                  * instead of just returning 0; we could be somewhere in a subprog,
2147                  * so execution could continue otherwise which we do /not/ want.
2148                  *
2149                  * Note, verifier whitelists all opcodes in bpf_opcode_in_insntable().
2150                  */
2151                 pr_warn("BPF interpreter: unknown opcode %02x (imm: 0x%x)\n",
2152                         insn->code, insn->imm);
2153                 BUG_ON(1);
2154                 return 0;
2155 }
2156
2157 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
2158 #define DEFINE_BPF_PROG_RUN(stack_size) \
2159 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
2160 { \
2161         u64 stack[stack_size / sizeof(u64)]; \
2162         u64 regs[MAX_BPF_EXT_REG] = {}; \
2163 \
2164         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2165         ARG1 = (u64) (unsigned long) ctx; \
2166         return ___bpf_prog_run(regs, insn); \
2167 }
2168
2169 #define PROG_NAME_ARGS(stack_size) __bpf_prog_run_args##stack_size
2170 #define DEFINE_BPF_PROG_RUN_ARGS(stack_size) \
2171 static u64 PROG_NAME_ARGS(stack_size)(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5, \
2172                                       const struct bpf_insn *insn) \
2173 { \
2174         u64 stack[stack_size / sizeof(u64)]; \
2175         u64 regs[MAX_BPF_EXT_REG]; \
2176 \
2177         FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
2178         BPF_R1 = r1; \
2179         BPF_R2 = r2; \
2180         BPF_R3 = r3; \
2181         BPF_R4 = r4; \
2182         BPF_R5 = r5; \
2183         return ___bpf_prog_run(regs, insn); \
2184 }
2185
2186 #define EVAL1(FN, X) FN(X)
2187 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
2188 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
2189 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
2190 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
2191 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
2192
2193 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
2194 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
2195 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
2196
2197 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 32, 64, 96, 128, 160, 192);
2198 EVAL6(DEFINE_BPF_PROG_RUN_ARGS, 224, 256, 288, 320, 352, 384);
2199 EVAL4(DEFINE_BPF_PROG_RUN_ARGS, 416, 448, 480, 512);
2200
2201 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
2202
2203 static unsigned int (*interpreters[])(const void *ctx,
2204                                       const struct bpf_insn *insn) = {
2205 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2206 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2207 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2208 };
2209 #undef PROG_NAME_LIST
2210 #define PROG_NAME_LIST(stack_size) PROG_NAME_ARGS(stack_size),
2211 static __maybe_unused
2212 u64 (*interpreters_args[])(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5,
2213                            const struct bpf_insn *insn) = {
2214 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
2215 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
2216 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
2217 };
2218 #undef PROG_NAME_LIST
2219
2220 #ifdef CONFIG_BPF_SYSCALL
2221 void bpf_patch_call_args(struct bpf_insn *insn, u32 stack_depth)
2222 {
2223         stack_depth = max_t(u32, stack_depth, 1);
2224         insn->off = (s16) insn->imm;
2225         insn->imm = interpreters_args[(round_up(stack_depth, 32) / 32) - 1] -
2226                 __bpf_call_base_args;
2227         insn->code = BPF_JMP | BPF_CALL_ARGS;
2228 }
2229 #endif
2230 #else
2231 static unsigned int __bpf_prog_ret0_warn(const void *ctx,
2232                                          const struct bpf_insn *insn)
2233 {
2234         /* If this handler ever gets executed, then BPF_JIT_ALWAYS_ON
2235          * is not working properly, so warn about it!
2236          */
2237         WARN_ON_ONCE(1);
2238         return 0;
2239 }
2240 #endif
2241
2242 bool bpf_prog_map_compatible(struct bpf_map *map,
2243                              const struct bpf_prog *fp)
2244 {
2245         enum bpf_prog_type prog_type = resolve_prog_type(fp);
2246         bool ret;
2247
2248         if (fp->kprobe_override)
2249                 return false;
2250
2251         /* XDP programs inserted into maps are not guaranteed to run on
2252          * a particular netdev (and can run outside driver context entirely
2253          * in the case of devmap and cpumap). Until device checks
2254          * are implemented, prohibit adding dev-bound programs to program maps.
2255          */
2256         if (bpf_prog_is_dev_bound(fp->aux))
2257                 return false;
2258
2259         spin_lock(&map->owner.lock);
2260         if (!map->owner.type) {
2261                 /* There's no owner yet where we could check for
2262                  * compatibility.
2263                  */
2264                 map->owner.type  = prog_type;
2265                 map->owner.jited = fp->jited;
2266                 map->owner.xdp_has_frags = fp->aux->xdp_has_frags;
2267                 ret = true;
2268         } else {
2269                 ret = map->owner.type  == prog_type &&
2270                       map->owner.jited == fp->jited &&
2271                       map->owner.xdp_has_frags == fp->aux->xdp_has_frags;
2272         }
2273         spin_unlock(&map->owner.lock);
2274
2275         return ret;
2276 }
2277
2278 static int bpf_check_tail_call(const struct bpf_prog *fp)
2279 {
2280         struct bpf_prog_aux *aux = fp->aux;
2281         int i, ret = 0;
2282
2283         mutex_lock(&aux->used_maps_mutex);
2284         for (i = 0; i < aux->used_map_cnt; i++) {
2285                 struct bpf_map *map = aux->used_maps[i];
2286
2287                 if (!map_type_contains_progs(map))
2288                         continue;
2289
2290                 if (!bpf_prog_map_compatible(map, fp)) {
2291                         ret = -EINVAL;
2292                         goto out;
2293                 }
2294         }
2295
2296 out:
2297         mutex_unlock(&aux->used_maps_mutex);
2298         return ret;
2299 }
2300
2301 static void bpf_prog_select_func(struct bpf_prog *fp)
2302 {
2303 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2304         u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
2305
2306         fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
2307 #else
2308         fp->bpf_func = __bpf_prog_ret0_warn;
2309 #endif
2310 }
2311
2312 /**
2313  *      bpf_prog_select_runtime - select exec runtime for BPF program
2314  *      @fp: bpf_prog populated with BPF program
2315  *      @err: pointer to error variable
2316  *
2317  * Try to JIT eBPF program, if JIT is not available, use interpreter.
2318  * The BPF program will be executed via bpf_prog_run() function.
2319  *
2320  * Return: the &fp argument along with &err set to 0 for success or
2321  * a negative errno code on failure
2322  */
2323 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
2324 {
2325         /* In case of BPF to BPF calls, verifier did all the prep
2326          * work with regards to JITing, etc.
2327          */
2328         bool jit_needed = false;
2329
2330         if (fp->bpf_func)
2331                 goto finalize;
2332
2333         if (IS_ENABLED(CONFIG_BPF_JIT_ALWAYS_ON) ||
2334             bpf_prog_has_kfunc_call(fp))
2335                 jit_needed = true;
2336
2337         bpf_prog_select_func(fp);
2338
2339         /* eBPF JITs can rewrite the program in case constant
2340          * blinding is active. However, in case of error during
2341          * blinding, bpf_int_jit_compile() must always return a
2342          * valid program, which in this case would simply not
2343          * be JITed, but falls back to the interpreter.
2344          */
2345         if (!bpf_prog_is_offloaded(fp->aux)) {
2346                 *err = bpf_prog_alloc_jited_linfo(fp);
2347                 if (*err)
2348                         return fp;
2349
2350                 fp = bpf_int_jit_compile(fp);
2351                 bpf_prog_jit_attempt_done(fp);
2352                 if (!fp->jited && jit_needed) {
2353                         *err = -ENOTSUPP;
2354                         return fp;
2355                 }
2356         } else {
2357                 *err = bpf_prog_offload_compile(fp);
2358                 if (*err)
2359                         return fp;
2360         }
2361
2362 finalize:
2363         bpf_prog_lock_ro(fp);
2364
2365         /* The tail call compatibility check can only be done at
2366          * this late stage as we need to determine, if we deal
2367          * with JITed or non JITed program concatenations and not
2368          * all eBPF JITs might immediately support all features.
2369          */
2370         *err = bpf_check_tail_call(fp);
2371
2372         return fp;
2373 }
2374 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
2375
2376 static unsigned int __bpf_prog_ret1(const void *ctx,
2377                                     const struct bpf_insn *insn)
2378 {
2379         return 1;
2380 }
2381
2382 static struct bpf_prog_dummy {
2383         struct bpf_prog prog;
2384 } dummy_bpf_prog = {
2385         .prog = {
2386                 .bpf_func = __bpf_prog_ret1,
2387         },
2388 };
2389
2390 struct bpf_empty_prog_array bpf_empty_prog_array = {
2391         .null_prog = NULL,
2392 };
2393 EXPORT_SYMBOL(bpf_empty_prog_array);
2394
2395 struct bpf_prog_array *bpf_prog_array_alloc(u32 prog_cnt, gfp_t flags)
2396 {
2397         if (prog_cnt)
2398                 return kzalloc(sizeof(struct bpf_prog_array) +
2399                                sizeof(struct bpf_prog_array_item) *
2400                                (prog_cnt + 1),
2401                                flags);
2402
2403         return &bpf_empty_prog_array.hdr;
2404 }
2405
2406 void bpf_prog_array_free(struct bpf_prog_array *progs)
2407 {
2408         if (!progs || progs == &bpf_empty_prog_array.hdr)
2409                 return;
2410         kfree_rcu(progs, rcu);
2411 }
2412
2413 static void __bpf_prog_array_free_sleepable_cb(struct rcu_head *rcu)
2414 {
2415         struct bpf_prog_array *progs;
2416
2417         /* If RCU Tasks Trace grace period implies RCU grace period, there is
2418          * no need to call kfree_rcu(), just call kfree() directly.
2419          */
2420         progs = container_of(rcu, struct bpf_prog_array, rcu);
2421         if (rcu_trace_implies_rcu_gp())
2422                 kfree(progs);
2423         else
2424                 kfree_rcu(progs, rcu);
2425 }
2426
2427 void bpf_prog_array_free_sleepable(struct bpf_prog_array *progs)
2428 {
2429         if (!progs || progs == &bpf_empty_prog_array.hdr)
2430                 return;
2431         call_rcu_tasks_trace(&progs->rcu, __bpf_prog_array_free_sleepable_cb);
2432 }
2433
2434 int bpf_prog_array_length(struct bpf_prog_array *array)
2435 {
2436         struct bpf_prog_array_item *item;
2437         u32 cnt = 0;
2438
2439         for (item = array->items; item->prog; item++)
2440                 if (item->prog != &dummy_bpf_prog.prog)
2441                         cnt++;
2442         return cnt;
2443 }
2444
2445 bool bpf_prog_array_is_empty(struct bpf_prog_array *array)
2446 {
2447         struct bpf_prog_array_item *item;
2448
2449         for (item = array->items; item->prog; item++)
2450                 if (item->prog != &dummy_bpf_prog.prog)
2451                         return false;
2452         return true;
2453 }
2454
2455 static bool bpf_prog_array_copy_core(struct bpf_prog_array *array,
2456                                      u32 *prog_ids,
2457                                      u32 request_cnt)
2458 {
2459         struct bpf_prog_array_item *item;
2460         int i = 0;
2461
2462         for (item = array->items; item->prog; item++) {
2463                 if (item->prog == &dummy_bpf_prog.prog)
2464                         continue;
2465                 prog_ids[i] = item->prog->aux->id;
2466                 if (++i == request_cnt) {
2467                         item++;
2468                         break;
2469                 }
2470         }
2471
2472         return !!(item->prog);
2473 }
2474
2475 int bpf_prog_array_copy_to_user(struct bpf_prog_array *array,
2476                                 __u32 __user *prog_ids, u32 cnt)
2477 {
2478         unsigned long err = 0;
2479         bool nospc;
2480         u32 *ids;
2481
2482         /* users of this function are doing:
2483          * cnt = bpf_prog_array_length();
2484          * if (cnt > 0)
2485          *     bpf_prog_array_copy_to_user(..., cnt);
2486          * so below kcalloc doesn't need extra cnt > 0 check.
2487          */
2488         ids = kcalloc(cnt, sizeof(u32), GFP_USER | __GFP_NOWARN);
2489         if (!ids)
2490                 return -ENOMEM;
2491         nospc = bpf_prog_array_copy_core(array, ids, cnt);
2492         err = copy_to_user(prog_ids, ids, cnt * sizeof(u32));
2493         kfree(ids);
2494         if (err)
2495                 return -EFAULT;
2496         if (nospc)
2497                 return -ENOSPC;
2498         return 0;
2499 }
2500
2501 void bpf_prog_array_delete_safe(struct bpf_prog_array *array,
2502                                 struct bpf_prog *old_prog)
2503 {
2504         struct bpf_prog_array_item *item;
2505
2506         for (item = array->items; item->prog; item++)
2507                 if (item->prog == old_prog) {
2508                         WRITE_ONCE(item->prog, &dummy_bpf_prog.prog);
2509                         break;
2510                 }
2511 }
2512
2513 /**
2514  * bpf_prog_array_delete_safe_at() - Replaces the program at the given
2515  *                                   index into the program array with
2516  *                                   a dummy no-op program.
2517  * @array: a bpf_prog_array
2518  * @index: the index of the program to replace
2519  *
2520  * Skips over dummy programs, by not counting them, when calculating
2521  * the position of the program to replace.
2522  *
2523  * Return:
2524  * * 0          - Success
2525  * * -EINVAL    - Invalid index value. Must be a non-negative integer.
2526  * * -ENOENT    - Index out of range
2527  */
2528 int bpf_prog_array_delete_safe_at(struct bpf_prog_array *array, int index)
2529 {
2530         return bpf_prog_array_update_at(array, index, &dummy_bpf_prog.prog);
2531 }
2532
2533 /**
2534  * bpf_prog_array_update_at() - Updates the program at the given index
2535  *                              into the program array.
2536  * @array: a bpf_prog_array
2537  * @index: the index of the program to update
2538  * @prog: the program to insert into the array
2539  *
2540  * Skips over dummy programs, by not counting them, when calculating
2541  * the position of the program to update.
2542  *
2543  * Return:
2544  * * 0          - Success
2545  * * -EINVAL    - Invalid index value. Must be a non-negative integer.
2546  * * -ENOENT    - Index out of range
2547  */
2548 int bpf_prog_array_update_at(struct bpf_prog_array *array, int index,
2549                              struct bpf_prog *prog)
2550 {
2551         struct bpf_prog_array_item *item;
2552
2553         if (unlikely(index < 0))
2554                 return -EINVAL;
2555
2556         for (item = array->items; item->prog; item++) {
2557                 if (item->prog == &dummy_bpf_prog.prog)
2558                         continue;
2559                 if (!index) {
2560                         WRITE_ONCE(item->prog, prog);
2561                         return 0;
2562                 }
2563                 index--;
2564         }
2565         return -ENOENT;
2566 }
2567
2568 int bpf_prog_array_copy(struct bpf_prog_array *old_array,
2569                         struct bpf_prog *exclude_prog,
2570                         struct bpf_prog *include_prog,
2571                         u64 bpf_cookie,
2572                         struct bpf_prog_array **new_array)
2573 {
2574         int new_prog_cnt, carry_prog_cnt = 0;
2575         struct bpf_prog_array_item *existing, *new;
2576         struct bpf_prog_array *array;
2577         bool found_exclude = false;
2578
2579         /* Figure out how many existing progs we need to carry over to
2580          * the new array.
2581          */
2582         if (old_array) {
2583                 existing = old_array->items;
2584                 for (; existing->prog; existing++) {
2585                         if (existing->prog == exclude_prog) {
2586                                 found_exclude = true;
2587                                 continue;
2588                         }
2589                         if (existing->prog != &dummy_bpf_prog.prog)
2590                                 carry_prog_cnt++;
2591                         if (existing->prog == include_prog)
2592                                 return -EEXIST;
2593                 }
2594         }
2595
2596         if (exclude_prog && !found_exclude)
2597                 return -ENOENT;
2598
2599         /* How many progs (not NULL) will be in the new array? */
2600         new_prog_cnt = carry_prog_cnt;
2601         if (include_prog)
2602                 new_prog_cnt += 1;
2603
2604         /* Do we have any prog (not NULL) in the new array? */
2605         if (!new_prog_cnt) {
2606                 *new_array = NULL;
2607                 return 0;
2608         }
2609
2610         /* +1 as the end of prog_array is marked with NULL */
2611         array = bpf_prog_array_alloc(new_prog_cnt + 1, GFP_KERNEL);
2612         if (!array)
2613                 return -ENOMEM;
2614         new = array->items;
2615
2616         /* Fill in the new prog array */
2617         if (carry_prog_cnt) {
2618                 existing = old_array->items;
2619                 for (; existing->prog; existing++) {
2620                         if (existing->prog == exclude_prog ||
2621                             existing->prog == &dummy_bpf_prog.prog)
2622                                 continue;
2623
2624                         new->prog = existing->prog;
2625                         new->bpf_cookie = existing->bpf_cookie;
2626                         new++;
2627                 }
2628         }
2629         if (include_prog) {
2630                 new->prog = include_prog;
2631                 new->bpf_cookie = bpf_cookie;
2632                 new++;
2633         }
2634         new->prog = NULL;
2635         *new_array = array;
2636         return 0;
2637 }
2638
2639 int bpf_prog_array_copy_info(struct bpf_prog_array *array,
2640                              u32 *prog_ids, u32 request_cnt,
2641                              u32 *prog_cnt)
2642 {
2643         u32 cnt = 0;
2644
2645         if (array)
2646                 cnt = bpf_prog_array_length(array);
2647
2648         *prog_cnt = cnt;
2649
2650         /* return early if user requested only program count or nothing to copy */
2651         if (!request_cnt || !cnt)
2652                 return 0;
2653
2654         /* this function is called under trace/bpf_trace.c: bpf_event_mutex */
2655         return bpf_prog_array_copy_core(array, prog_ids, request_cnt) ? -ENOSPC
2656                                                                      : 0;
2657 }
2658
2659 void __bpf_free_used_maps(struct bpf_prog_aux *aux,
2660                           struct bpf_map **used_maps, u32 len)
2661 {
2662         struct bpf_map *map;
2663         u32 i;
2664
2665         for (i = 0; i < len; i++) {
2666                 map = used_maps[i];
2667                 if (map->ops->map_poke_untrack)
2668                         map->ops->map_poke_untrack(map, aux);
2669                 bpf_map_put(map);
2670         }
2671 }
2672
2673 static void bpf_free_used_maps(struct bpf_prog_aux *aux)
2674 {
2675         __bpf_free_used_maps(aux, aux->used_maps, aux->used_map_cnt);
2676         kfree(aux->used_maps);
2677 }
2678
2679 void __bpf_free_used_btfs(struct bpf_prog_aux *aux,
2680                           struct btf_mod_pair *used_btfs, u32 len)
2681 {
2682 #ifdef CONFIG_BPF_SYSCALL
2683         struct btf_mod_pair *btf_mod;
2684         u32 i;
2685
2686         for (i = 0; i < len; i++) {
2687                 btf_mod = &used_btfs[i];
2688                 if (btf_mod->module)
2689                         module_put(btf_mod->module);
2690                 btf_put(btf_mod->btf);
2691         }
2692 #endif
2693 }
2694
2695 static void bpf_free_used_btfs(struct bpf_prog_aux *aux)
2696 {
2697         __bpf_free_used_btfs(aux, aux->used_btfs, aux->used_btf_cnt);
2698         kfree(aux->used_btfs);
2699 }
2700
2701 static void bpf_prog_free_deferred(struct work_struct *work)
2702 {
2703         struct bpf_prog_aux *aux;
2704         int i;
2705
2706         aux = container_of(work, struct bpf_prog_aux, work);
2707 #ifdef CONFIG_BPF_SYSCALL
2708         bpf_free_kfunc_btf_tab(aux->kfunc_btf_tab);
2709 #endif
2710 #ifdef CONFIG_CGROUP_BPF
2711         if (aux->cgroup_atype != CGROUP_BPF_ATTACH_TYPE_INVALID)
2712                 bpf_cgroup_atype_put(aux->cgroup_atype);
2713 #endif
2714         bpf_free_used_maps(aux);
2715         bpf_free_used_btfs(aux);
2716         if (bpf_prog_is_dev_bound(aux))
2717                 bpf_prog_dev_bound_destroy(aux->prog);
2718 #ifdef CONFIG_PERF_EVENTS
2719         if (aux->prog->has_callchain_buf)
2720                 put_callchain_buffers();
2721 #endif
2722         if (aux->dst_trampoline)
2723                 bpf_trampoline_put(aux->dst_trampoline);
2724         for (i = 0; i < aux->func_cnt; i++) {
2725                 /* We can just unlink the subprog poke descriptor table as
2726                  * it was originally linked to the main program and is also
2727                  * released along with it.
2728                  */
2729                 aux->func[i]->aux->poke_tab = NULL;
2730                 bpf_jit_free(aux->func[i]);
2731         }
2732         if (aux->func_cnt) {
2733                 kfree(aux->func);
2734                 bpf_prog_unlock_free(aux->prog);
2735         } else {
2736                 bpf_jit_free(aux->prog);
2737         }
2738 }
2739
2740 void bpf_prog_free(struct bpf_prog *fp)
2741 {
2742         struct bpf_prog_aux *aux = fp->aux;
2743
2744         if (aux->dst_prog)
2745                 bpf_prog_put(aux->dst_prog);
2746         INIT_WORK(&aux->work, bpf_prog_free_deferred);
2747         schedule_work(&aux->work);
2748 }
2749 EXPORT_SYMBOL_GPL(bpf_prog_free);
2750
2751 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
2752 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
2753
2754 void bpf_user_rnd_init_once(void)
2755 {
2756         prandom_init_once(&bpf_user_rnd_state);
2757 }
2758
2759 BPF_CALL_0(bpf_user_rnd_u32)
2760 {
2761         /* Should someone ever have the rather unwise idea to use some
2762          * of the registers passed into this function, then note that
2763          * this function is called from native eBPF and classic-to-eBPF
2764          * transformations. Register assignments from both sides are
2765          * different, f.e. classic always sets fn(ctx, A, X) here.
2766          */
2767         struct rnd_state *state;
2768         u32 res;
2769
2770         state = &get_cpu_var(bpf_user_rnd_state);
2771         res = prandom_u32_state(state);
2772         put_cpu_var(bpf_user_rnd_state);
2773
2774         return res;
2775 }
2776
2777 BPF_CALL_0(bpf_get_raw_cpu_id)
2778 {
2779         return raw_smp_processor_id();
2780 }
2781
2782 /* Weak definitions of helper functions in case we don't have bpf syscall. */
2783 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
2784 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
2785 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
2786 const struct bpf_func_proto bpf_map_push_elem_proto __weak;
2787 const struct bpf_func_proto bpf_map_pop_elem_proto __weak;
2788 const struct bpf_func_proto bpf_map_peek_elem_proto __weak;
2789 const struct bpf_func_proto bpf_map_lookup_percpu_elem_proto __weak;
2790 const struct bpf_func_proto bpf_spin_lock_proto __weak;
2791 const struct bpf_func_proto bpf_spin_unlock_proto __weak;
2792 const struct bpf_func_proto bpf_jiffies64_proto __weak;
2793
2794 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
2795 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
2796 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
2797 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
2798 const struct bpf_func_proto bpf_ktime_get_boot_ns_proto __weak;
2799 const struct bpf_func_proto bpf_ktime_get_coarse_ns_proto __weak;
2800 const struct bpf_func_proto bpf_ktime_get_tai_ns_proto __weak;
2801
2802 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
2803 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
2804 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
2805 const struct bpf_func_proto bpf_get_current_cgroup_id_proto __weak;
2806 const struct bpf_func_proto bpf_get_current_ancestor_cgroup_id_proto __weak;
2807 const struct bpf_func_proto bpf_get_local_storage_proto __weak;
2808 const struct bpf_func_proto bpf_get_ns_current_pid_tgid_proto __weak;
2809 const struct bpf_func_proto bpf_snprintf_btf_proto __weak;
2810 const struct bpf_func_proto bpf_seq_printf_btf_proto __weak;
2811 const struct bpf_func_proto bpf_set_retval_proto __weak;
2812 const struct bpf_func_proto bpf_get_retval_proto __weak;
2813
2814 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
2815 {
2816         return NULL;
2817 }
2818
2819 const struct bpf_func_proto * __weak bpf_get_trace_vprintk_proto(void)
2820 {
2821         return NULL;
2822 }
2823
2824 u64 __weak
2825 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
2826                  void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
2827 {
2828         return -ENOTSUPP;
2829 }
2830 EXPORT_SYMBOL_GPL(bpf_event_output);
2831
2832 /* Always built-in helper functions. */
2833 const struct bpf_func_proto bpf_tail_call_proto = {
2834         .func           = NULL,
2835         .gpl_only       = false,
2836         .ret_type       = RET_VOID,
2837         .arg1_type      = ARG_PTR_TO_CTX,
2838         .arg2_type      = ARG_CONST_MAP_PTR,
2839         .arg3_type      = ARG_ANYTHING,
2840 };
2841
2842 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
2843  * It is encouraged to implement bpf_int_jit_compile() instead, so that
2844  * eBPF and implicitly also cBPF can get JITed!
2845  */
2846 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
2847 {
2848         return prog;
2849 }
2850
2851 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
2852  * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
2853  */
2854 void __weak bpf_jit_compile(struct bpf_prog *prog)
2855 {
2856 }
2857
2858 bool __weak bpf_helper_changes_pkt_data(void *func)
2859 {
2860         return false;
2861 }
2862
2863 /* Return TRUE if the JIT backend wants verifier to enable sub-register usage
2864  * analysis code and wants explicit zero extension inserted by verifier.
2865  * Otherwise, return FALSE.
2866  *
2867  * The verifier inserts an explicit zero extension after BPF_CMPXCHGs even if
2868  * you don't override this. JITs that don't want these extra insns can detect
2869  * them using insn_is_zext.
2870  */
2871 bool __weak bpf_jit_needs_zext(void)
2872 {
2873         return false;
2874 }
2875
2876 /* Return TRUE if the JIT backend supports mixing bpf2bpf and tailcalls. */
2877 bool __weak bpf_jit_supports_subprog_tailcalls(void)
2878 {
2879         return false;
2880 }
2881
2882 bool __weak bpf_jit_supports_kfunc_call(void)
2883 {
2884         return false;
2885 }
2886
2887 bool __weak bpf_jit_supports_far_kfunc_call(void)
2888 {
2889         return false;
2890 }
2891
2892 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
2893  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
2894  */
2895 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
2896                          int len)
2897 {
2898         return -EFAULT;
2899 }
2900
2901 int __weak bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
2902                               void *addr1, void *addr2)
2903 {
2904         return -ENOTSUPP;
2905 }
2906
2907 void * __weak bpf_arch_text_copy(void *dst, void *src, size_t len)
2908 {
2909         return ERR_PTR(-ENOTSUPP);
2910 }
2911
2912 int __weak bpf_arch_text_invalidate(void *dst, size_t len)
2913 {
2914         return -ENOTSUPP;
2915 }
2916
2917 #ifdef CONFIG_BPF_SYSCALL
2918 static int __init bpf_global_ma_init(void)
2919 {
2920         int ret;
2921
2922         ret = bpf_mem_alloc_init(&bpf_global_ma, 0, false);
2923         bpf_global_ma_set = !ret;
2924         return ret;
2925 }
2926 late_initcall(bpf_global_ma_init);
2927 #endif
2928
2929 DEFINE_STATIC_KEY_FALSE(bpf_stats_enabled_key);
2930 EXPORT_SYMBOL(bpf_stats_enabled_key);
2931
2932 /* All definitions of tracepoints related to BPF. */
2933 #define CREATE_TRACE_POINTS
2934 #include <linux/bpf_trace.h>
2935
2936 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
2937 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_bulk_tx);