arch/x86/net/bpf_jit_comp.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * bpf_jit_comp.c: BPF JIT compiler
   4  *
   5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
   6  * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   7  */
   8 #include <linux/netdevice.h>
   9 #include <linux/filter.h>
  10 #include <linux/if_vlan.h>
  11 #include <linux/bpf.h>
  12 #include <linux/memory.h>
  13 #include <linux/sort.h>
  14 #include <asm/extable.h>
  15 #include <asm/set_memory.h>
  16 #include <asm/nospec-branch.h>
  17 #include <asm/text-patching.h>
  18
  19 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
  20 {
  21         if (len == 1)
  22                 *ptr = bytes;
  23         else if (len == 2)
  24                 *(u16 *)ptr = bytes;
  25         else {
  26                 *(u32 *)ptr = bytes;
  27                 barrier();
  28         }
  29         return ptr + len;
  30 }
  31
  32 #define EMIT(bytes, len) \
  33         do { prog = emit_code(prog, bytes, len); } while (0)
  34
  35 #define EMIT1(b1)               EMIT(b1, 1)
  36 #define EMIT2(b1, b2)           EMIT((b1) + ((b2) << 8), 2)
  37 #define EMIT3(b1, b2, b3)       EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
  38 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
  39
  40 #define EMIT1_off32(b1, off) \
  41         do { EMIT1(b1); EMIT(off, 4); } while (0)
  42 #define EMIT2_off32(b1, b2, off) \
  43         do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
  44 #define EMIT3_off32(b1, b2, b3, off) \
  45         do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
  46 #define EMIT4_off32(b1, b2, b3, b4, off) \
  47         do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
  48
  49 static bool is_imm8(int value)
  50 {
  51         return value <= 127 && value >= -128;
  52 }
  53
  54 static bool is_simm32(s64 value)
  55 {
  56         return value == (s64)(s32)value;
  57 }
  58
  59 static bool is_uimm32(u64 value)
  60 {
  61         return value == (u64)(u32)value;
  62 }
  63
  64 /* mov dst, src */
  65 #define EMIT_mov(DST, SRC)                                                               \
  66         do {                                                                             \
  67                 if (DST != SRC)                                                          \
  68                         EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
  69         } while (0)
  70
  71 static int bpf_size_to_x86_bytes(int bpf_size)
  72 {
  73         if (bpf_size == BPF_W)
  74                 return 4;
  75         else if (bpf_size == BPF_H)
  76                 return 2;
  77         else if (bpf_size == BPF_B)
  78                 return 1;
  79         else if (bpf_size == BPF_DW)
  80                 return 4; /* imm32 */
  81         else
  82                 return 0;
  83 }
  84
  85 /*
  86  * List of x86 cond jumps opcodes (. + s8)
  87  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
  88  */
  89 #define X86_JB  0x72
  90 #define X86_JAE 0x73
  91 #define X86_JE  0x74
  92 #define X86_JNE 0x75
  93 #define X86_JBE 0x76
  94 #define X86_JA  0x77
  95 #define X86_JL  0x7C
  96 #define X86_JGE 0x7D
  97 #define X86_JLE 0x7E
  98 #define X86_JG  0x7F
  99
 100 /* Pick a register outside of BPF range for JIT internal work */
 101 #define AUX_REG (MAX_BPF_JIT_REG + 1)
 102 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
 103
 104 /*
 105  * The following table maps BPF registers to x86-64 registers.
 106  *
 107  * x86-64 register R12 is unused, since if used as base address
 108  * register in load/store instructions, it always needs an
 109  * extra byte of encoding and is callee saved.
 110  *
 111  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
 112  * trampoline. x86-64 register R10 is used for blinding (if enabled).
 113  */
 114 static const int reg2hex[] = {
 115         [BPF_REG_0] = 0,  /* RAX */
 116         [BPF_REG_1] = 7,  /* RDI */
 117         [BPF_REG_2] = 6,  /* RSI */
 118         [BPF_REG_3] = 2,  /* RDX */
 119         [BPF_REG_4] = 1,  /* RCX */
 120         [BPF_REG_5] = 0,  /* R8  */
 121         [BPF_REG_6] = 3,  /* RBX callee saved */
 122         [BPF_REG_7] = 5,  /* R13 callee saved */
 123         [BPF_REG_8] = 6,  /* R14 callee saved */
 124         [BPF_REG_9] = 7,  /* R15 callee saved */
 125         [BPF_REG_FP] = 5, /* RBP readonly */
 126         [BPF_REG_AX] = 2, /* R10 temp register */
 127         [AUX_REG] = 3,    /* R11 temp register */
 128         [X86_REG_R9] = 1, /* R9 register, 6th function argument */
 129 };
 130
 131 static const int reg2pt_regs[] = {
 132         [BPF_REG_0] = offsetof(struct pt_regs, ax),
 133         [BPF_REG_1] = offsetof(struct pt_regs, di),
 134         [BPF_REG_2] = offsetof(struct pt_regs, si),
 135         [BPF_REG_3] = offsetof(struct pt_regs, dx),
 136         [BPF_REG_4] = offsetof(struct pt_regs, cx),
 137         [BPF_REG_5] = offsetof(struct pt_regs, r8),
 138         [BPF_REG_6] = offsetof(struct pt_regs, bx),
 139         [BPF_REG_7] = offsetof(struct pt_regs, r13),
 140         [BPF_REG_8] = offsetof(struct pt_regs, r14),
 141         [BPF_REG_9] = offsetof(struct pt_regs, r15),
 142 };
 143
 144 /*
 145  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
 146  * which need extra byte of encoding.
 147  * rax,rcx,...,rbp have simpler encoding
 148  */
 149 static bool is_ereg(u32 reg)
 150 {
 151         return (1 << reg) & (BIT(BPF_REG_5) |
 152                              BIT(AUX_REG) |
 153                              BIT(BPF_REG_7) |
 154                              BIT(BPF_REG_8) |
 155                              BIT(BPF_REG_9) |
 156                              BIT(X86_REG_R9) |
 157                              BIT(BPF_REG_AX));
 158 }
 159
 160 /*
 161  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
 162  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
 163  * of encoding. al,cl,dl,bl have simpler encoding.
 164  */
 165 static bool is_ereg_8l(u32 reg)
 166 {
 167         return is_ereg(reg) ||
 168             (1 << reg) & (BIT(BPF_REG_1) |
 169                           BIT(BPF_REG_2) |
 170                           BIT(BPF_REG_FP));
 171 }
 172
 173 static bool is_axreg(u32 reg)
 174 {
 175         return reg == BPF_REG_0;
 176 }
 177
 178 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
 179 static u8 add_1mod(u8 byte, u32 reg)
 180 {
 181         if (is_ereg(reg))
 182                 byte |= 1;
 183         return byte;
 184 }
 185
 186 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
 187 {
 188         if (is_ereg(r1))
 189                 byte |= 1;
 190         if (is_ereg(r2))
 191                 byte |= 4;
 192         return byte;
 193 }
 194
 195 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
 196 static u8 add_1reg(u8 byte, u32 dst_reg)
 197 {
 198         return byte + reg2hex[dst_reg];
 199 }
 200
 201 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
 202 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
 203 {
 204         return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
 205 }
 206
 207 /* Some 1-byte opcodes for binary ALU operations */
 208 static u8 simple_alu_opcodes[] = {
 209         [BPF_ADD] = 0x01,
 210         [BPF_SUB] = 0x29,
 211         [BPF_AND] = 0x21,
 212         [BPF_OR] = 0x09,
 213         [BPF_XOR] = 0x31,
 214         [BPF_LSH] = 0xE0,
 215         [BPF_RSH] = 0xE8,
 216         [BPF_ARSH] = 0xF8,
 217 };
 218
 219 static void jit_fill_hole(void *area, unsigned int size)
 220 {
 221         /* Fill whole space with INT3 instructions */
 222         memset(area, 0xcc, size);
 223 }
 224
 225 struct jit_context {
 226         int cleanup_addr; /* Epilogue code offset */
 227
 228         /*
 229          * Program specific offsets of labels in the code; these rely on the
 230          * JIT doing at least 2 passes, recording the position on the first
 231          * pass, only to generate the correct offset on the second pass.
 232          */
 233         int tail_call_direct_label;
 234         int tail_call_indirect_label;
 235 };
 236
 237 /* Maximum number of bytes emitted while JITing one eBPF insn */
 238 #define BPF_MAX_INSN_SIZE       128
 239 #define BPF_INSN_SAFETY         64
 240
 241 /* Number of bytes emit_patch() needs to generate instructions */
 242 #define X86_PATCH_SIZE          5
 243 /* Number of bytes that will be skipped on tailcall */
 244 #define X86_TAIL_CALL_OFFSET    11
 245
 246 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
 247 {
 248         u8 *prog = *pprog;
 249
 250         if (callee_regs_used[0])
 251                 EMIT1(0x53);         /* push rbx */
 252         if (callee_regs_used[1])
 253                 EMIT2(0x41, 0x55);   /* push r13 */
 254         if (callee_regs_used[2])
 255                 EMIT2(0x41, 0x56);   /* push r14 */
 256         if (callee_regs_used[3])
 257                 EMIT2(0x41, 0x57);   /* push r15 */
 258         *pprog = prog;
 259 }
 260
 261 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
 262 {
 263         u8 *prog = *pprog;
 264
 265         if (callee_regs_used[3])
 266                 EMIT2(0x41, 0x5F);   /* pop r15 */
 267         if (callee_regs_used[2])
 268                 EMIT2(0x41, 0x5E);   /* pop r14 */
 269         if (callee_regs_used[1])
 270                 EMIT2(0x41, 0x5D);   /* pop r13 */
 271         if (callee_regs_used[0])
 272                 EMIT1(0x5B);         /* pop rbx */
 273         *pprog = prog;
 274 }
 275
 276 /*
 277  * Emit x86-64 prologue code for BPF program.
 278  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
 279  * while jumping to another program
 280  */
 281 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
 282                           bool tail_call_reachable, bool is_subprog)
 283 {
 284         u8 *prog = *pprog;
 285
 286         /* BPF trampoline can be made to work without these nops,
 287          * but let's waste 5 bytes for now and optimize later
 288          */
 289         memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
 290         prog += X86_PATCH_SIZE;
 291         if (!ebpf_from_cbpf) {
 292                 if (tail_call_reachable && !is_subprog)
 293                         EMIT2(0x31, 0xC0); /* xor eax, eax */
 294                 else
 295                         EMIT2(0x66, 0x90); /* nop2 */
 296         }
 297         EMIT1(0x55);             /* push rbp */
 298         EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
 299         /* sub rsp, rounded_stack_depth */
 300         if (stack_depth)
 301                 EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
 302         if (tail_call_reachable)
 303                 EMIT1(0x50);         /* push rax */
 304         *pprog = prog;
 305 }
 306
 307 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
 308 {
 309         u8 *prog = *pprog;
 310         s64 offset;
 311
 312         offset = func - (ip + X86_PATCH_SIZE);
 313         if (!is_simm32(offset)) {
 314                 pr_err("Target call %p is out of range\n", func);
 315                 return -ERANGE;
 316         }
 317         EMIT1_off32(opcode, offset);
 318         *pprog = prog;
 319         return 0;
 320 }
 321
 322 static int emit_call(u8 **pprog, void *func, void *ip)
 323 {
 324         return emit_patch(pprog, func, ip, 0xE8);
 325 }
 326
 327 static int emit_jump(u8 **pprog, void *func, void *ip)
 328 {
 329         return emit_patch(pprog, func, ip, 0xE9);
 330 }
 331
 332 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
 333                                 void *old_addr, void *new_addr,
 334                                 const bool text_live)
 335 {
 336         const u8 *nop_insn = x86_nops[5];
 337         u8 old_insn[X86_PATCH_SIZE];
 338         u8 new_insn[X86_PATCH_SIZE];
 339         u8 *prog;
 340         int ret;
 341
 342         memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
 343         if (old_addr) {
 344                 prog = old_insn;
 345                 ret = t == BPF_MOD_CALL ?
 346                       emit_call(&prog, old_addr, ip) :
 347                       emit_jump(&prog, old_addr, ip);
 348                 if (ret)
 349                         return ret;
 350         }
 351
 352         memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
 353         if (new_addr) {
 354                 prog = new_insn;
 355                 ret = t == BPF_MOD_CALL ?
 356                       emit_call(&prog, new_addr, ip) :
 357                       emit_jump(&prog, new_addr, ip);
 358                 if (ret)
 359                         return ret;
 360         }
 361
 362         ret = -EBUSY;
 363         mutex_lock(&text_mutex);
 364         if (memcmp(ip, old_insn, X86_PATCH_SIZE))
 365                 goto out;
 366         ret = 1;
 367         if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
 368                 if (text_live)
 369                         text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
 370                 else
 371                         memcpy(ip, new_insn, X86_PATCH_SIZE);
 372                 ret = 0;
 373         }
 374 out:
 375         mutex_unlock(&text_mutex);
 376         return ret;
 377 }
 378
 379 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
 380                        void *old_addr, void *new_addr)
 381 {
 382         if (!is_kernel_text((long)ip) &&
 383             !is_bpf_text_address((long)ip))
 384                 /* BPF poking in modules is not supported */
 385                 return -EINVAL;
 386
 387         return __bpf_arch_text_poke(ip, t, old_addr, new_addr, true);
 388 }
 389
 390 #define EMIT_LFENCE()   EMIT3(0x0F, 0xAE, 0xE8)
 391
 392 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
 393 {
 394         u8 *prog = *pprog;
 395
 396 #ifdef CONFIG_RETPOLINE
 397         if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
 398                 EMIT_LFENCE();
 399                 EMIT2(0xFF, 0xE0 + reg);
 400         } else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
 401                 emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
 402         } else
 403 #endif
 404         EMIT2(0xFF, 0xE0 + reg);
 405
 406         *pprog = prog;
 407 }
 408
 409 static void emit_return(u8 **pprog, u8 *ip)
 410 {
 411         u8 *prog = *pprog;
 412
 413         if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
 414                 emit_jump(&prog, &__x86_return_thunk, ip);
 415         } else {
 416                 EMIT1(0xC3);            /* ret */
 417                 if (IS_ENABLED(CONFIG_SLS))
 418                         EMIT1(0xCC);    /* int3 */
 419         }
 420
 421         *pprog = prog;
 422 }
 423
 424 /*
 425  * Generate the following code:
 426  *
 427  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
 428  *   if (index >= array->map.max_entries)
 429  *     goto out;
 430  *   if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
 431  *     goto out;
 432  *   prog = array->ptrs[index];
 433  *   if (prog == NULL)
 434  *     goto out;
 435  *   goto *(prog->bpf_func + prologue_size);
 436  * out:
 437  */
 438 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
 439                                         u32 stack_depth, u8 *ip,
 440                                         struct jit_context *ctx)
 441 {
 442         int tcc_off = -4 - round_up(stack_depth, 8);
 443         u8 *prog = *pprog, *start = *pprog;
 444         int offset;
 445
 446         /*
 447          * rdi - pointer to ctx
 448          * rsi - pointer to bpf_array
 449          * rdx - index in bpf_array
 450          */
 451
 452         /*
 453          * if (index >= array->map.max_entries)
 454          *      goto out;
 455          */
 456         EMIT2(0x89, 0xD2);                        /* mov edx, edx */
 457         EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
 458               offsetof(struct bpf_array, map.max_entries));
 459
 460         offset = ctx->tail_call_indirect_label - (prog + 2 - start);
 461         EMIT2(X86_JBE, offset);                   /* jbe out */
 462
 463         /*
 464          * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
 465          *      goto out;
 466          */
 467         EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
 468         EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
 469
 470         offset = ctx->tail_call_indirect_label - (prog + 2 - start);
 471         EMIT2(X86_JA, offset);                    /* ja out */
 472         EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
 473         EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
 474
 475         /* prog = array->ptrs[index]; */
 476         EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
 477                     offsetof(struct bpf_array, ptrs));
 478
 479         /*
 480          * if (prog == NULL)
 481          *      goto out;
 482          */
 483         EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
 484
 485         offset = ctx->tail_call_indirect_label - (prog + 2 - start);
 486         EMIT2(X86_JE, offset);                    /* je out */
 487
 488         pop_callee_regs(&prog, callee_regs_used);
 489
 490         EMIT1(0x58);                              /* pop rax */
 491         if (stack_depth)
 492                 EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
 493                             round_up(stack_depth, 8));
 494
 495         /* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
 496         EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
 497               offsetof(struct bpf_prog, bpf_func));
 498         EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
 499               X86_TAIL_CALL_OFFSET);
 500         /*
 501          * Now we're ready to jump into next BPF program
 502          * rdi == ctx (1st arg)
 503          * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
 504          */
 505         emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
 506
 507         /* out: */
 508         ctx->tail_call_indirect_label = prog - start;
 509         *pprog = prog;
 510 }
 511
 512 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
 513                                       u8 **pprog, u8 *ip,
 514                                       bool *callee_regs_used, u32 stack_depth,
 515                                       struct jit_context *ctx)
 516 {
 517         int tcc_off = -4 - round_up(stack_depth, 8);
 518         u8 *prog = *pprog, *start = *pprog;
 519         int offset;
 520
 521         /*
 522          * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
 523          *      goto out;
 524          */
 525         EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
 526         EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
 527
 528         offset = ctx->tail_call_direct_label - (prog + 2 - start);
 529         EMIT2(X86_JA, offset);                        /* ja out */
 530         EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
 531         EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
 532
 533         poke->tailcall_bypass = ip + (prog - start);
 534         poke->adj_off = X86_TAIL_CALL_OFFSET;
 535         poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
 536         poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
 537
 538         emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
 539                   poke->tailcall_bypass);
 540
 541         pop_callee_regs(&prog, callee_regs_used);
 542         EMIT1(0x58);                                  /* pop rax */
 543         if (stack_depth)
 544                 EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
 545
 546         memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
 547         prog += X86_PATCH_SIZE;
 548
 549         /* out: */
 550         ctx->tail_call_direct_label = prog - start;
 551
 552         *pprog = prog;
 553 }
 554
 555 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
 556 {
 557         struct bpf_jit_poke_descriptor *poke;
 558         struct bpf_array *array;
 559         struct bpf_prog *target;
 560         int i, ret;
 561
 562         for (i = 0; i < prog->aux->size_poke_tab; i++) {
 563                 poke = &prog->aux->poke_tab[i];
 564                 if (poke->aux && poke->aux != prog->aux)
 565                         continue;
 566
 567                 WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
 568
 569                 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
 570                         continue;
 571
 572                 array = container_of(poke->tail_call.map, struct bpf_array, map);
 573                 mutex_lock(&array->aux->poke_mutex);
 574                 target = array->ptrs[poke->tail_call.key];
 575                 if (target) {
 576                         /* Plain memcpy is used when image is not live yet
 577                          * and still not locked as read-only. Once poke
 578                          * location is active (poke->tailcall_target_stable),
 579                          * any parallel bpf_arch_text_poke() might occur
 580                          * still on the read-write image until we finally
 581                          * locked it as read-only. Both modifications on
 582                          * the given image are under text_mutex to avoid
 583                          * interference.
 584                          */
 585                         ret = __bpf_arch_text_poke(poke->tailcall_target,
 586                                                    BPF_MOD_JUMP, NULL,
 587                                                    (u8 *)target->bpf_func +
 588                                                    poke->adj_off, false);
 589                         BUG_ON(ret < 0);
 590                         ret = __bpf_arch_text_poke(poke->tailcall_bypass,
 591                                                    BPF_MOD_JUMP,
 592                                                    (u8 *)poke->tailcall_target +
 593                                                    X86_PATCH_SIZE, NULL, false);
 594                         BUG_ON(ret < 0);
 595                 }
 596                 WRITE_ONCE(poke->tailcall_target_stable, true);
 597                 mutex_unlock(&array->aux->poke_mutex);
 598         }
 599 }
 600
 601 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
 602                            u32 dst_reg, const u32 imm32)
 603 {
 604         u8 *prog = *pprog;
 605         u8 b1, b2, b3;
 606
 607         /*
 608          * Optimization: if imm32 is positive, use 'mov %eax, imm32'
 609          * (which zero-extends imm32) to save 2 bytes.
 610          */
 611         if (sign_propagate && (s32)imm32 < 0) {
 612                 /* 'mov %rax, imm32' sign extends imm32 */
 613                 b1 = add_1mod(0x48, dst_reg);
 614                 b2 = 0xC7;
 615                 b3 = 0xC0;
 616                 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
 617                 goto done;
 618         }
 619
 620         /*
 621          * Optimization: if imm32 is zero, use 'xor %eax, %eax'
 622          * to save 3 bytes.
 623          */
 624         if (imm32 == 0) {
 625                 if (is_ereg(dst_reg))
 626                         EMIT1(add_2mod(0x40, dst_reg, dst_reg));
 627                 b2 = 0x31; /* xor */
 628                 b3 = 0xC0;
 629                 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
 630                 goto done;
 631         }
 632
 633         /* mov %eax, imm32 */
 634         if (is_ereg(dst_reg))
 635                 EMIT1(add_1mod(0x40, dst_reg));
 636         EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
 637 done:
 638         *pprog = prog;
 639 }
 640
 641 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
 642                            const u32 imm32_hi, const u32 imm32_lo)
 643 {
 644         u8 *prog = *pprog;
 645
 646         if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
 647                 /*
 648                  * For emitting plain u32, where sign bit must not be
 649                  * propagated LLVM tends to load imm64 over mov32
 650                  * directly, so save couple of bytes by just doing
 651                  * 'mov %eax, imm32' instead.
 652                  */
 653                 emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
 654         } else {
 655                 /* movabsq %rax, imm64 */
 656                 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
 657                 EMIT(imm32_lo, 4);
 658                 EMIT(imm32_hi, 4);
 659         }
 660
 661         *pprog = prog;
 662 }
 663
 664 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
 665 {
 666         u8 *prog = *pprog;
 667
 668         if (is64) {
 669                 /* mov dst, src */
 670                 EMIT_mov(dst_reg, src_reg);
 671         } else {
 672                 /* mov32 dst, src */
 673                 if (is_ereg(dst_reg) || is_ereg(src_reg))
 674                         EMIT1(add_2mod(0x40, dst_reg, src_reg));
 675                 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
 676         }
 677
 678         *pprog = prog;
 679 }
 680
 681 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
 682 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
 683 {
 684         u8 *prog = *pprog;
 685
 686         if (is_imm8(off)) {
 687                 /* 1-byte signed displacement.
 688                  *
 689                  * If off == 0 we could skip this and save one extra byte, but
 690                  * special case of x86 R13 which always needs an offset is not
 691                  * worth the hassle
 692                  */
 693                 EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
 694         } else {
 695                 /* 4-byte signed displacement */
 696                 EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
 697         }
 698         *pprog = prog;
 699 }
 700
 701 /*
 702  * Emit a REX byte if it will be necessary to address these registers
 703  */
 704 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
 705 {
 706         u8 *prog = *pprog;
 707
 708         if (is64)
 709                 EMIT1(add_2mod(0x48, dst_reg, src_reg));
 710         else if (is_ereg(dst_reg) || is_ereg(src_reg))
 711                 EMIT1(add_2mod(0x40, dst_reg, src_reg));
 712         *pprog = prog;
 713 }
 714
 715 /*
 716  * Similar version of maybe_emit_mod() for a single register
 717  */
 718 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
 719 {
 720         u8 *prog = *pprog;
 721
 722         if (is64)
 723                 EMIT1(add_1mod(0x48, reg));
 724         else if (is_ereg(reg))
 725                 EMIT1(add_1mod(0x40, reg));
 726         *pprog = prog;
 727 }
 728
 729 /* LDX: dst_reg = *(u8*)(src_reg + off) */
 730 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
 731 {
 732         u8 *prog = *pprog;
 733
 734         switch (size) {
 735         case BPF_B:
 736                 /* Emit 'movzx rax, byte ptr [rax + off]' */
 737                 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
 738                 break;
 739         case BPF_H:
 740                 /* Emit 'movzx rax, word ptr [rax + off]' */
 741                 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
 742                 break;
 743         case BPF_W:
 744                 /* Emit 'mov eax, dword ptr [rax+0x14]' */
 745                 if (is_ereg(dst_reg) || is_ereg(src_reg))
 746                         EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
 747                 else
 748                         EMIT1(0x8B);
 749                 break;
 750         case BPF_DW:
 751                 /* Emit 'mov rax, qword ptr [rax+0x14]' */
 752                 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
 753                 break;
 754         }
 755         emit_insn_suffix(&prog, src_reg, dst_reg, off);
 756         *pprog = prog;
 757 }
 758
 759 /* STX: *(u8*)(dst_reg + off) = src_reg */
 760 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
 761 {
 762         u8 *prog = *pprog;
 763
 764         switch (size) {
 765         case BPF_B:
 766                 /* Emit 'mov byte ptr [rax + off], al' */
 767                 if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
 768                         /* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
 769                         EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
 770                 else
 771                         EMIT1(0x88);
 772                 break;
 773         case BPF_H:
 774                 if (is_ereg(dst_reg) || is_ereg(src_reg))
 775                         EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
 776                 else
 777                         EMIT2(0x66, 0x89);
 778                 break;
 779         case BPF_W:
 780                 if (is_ereg(dst_reg) || is_ereg(src_reg))
 781                         EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
 782                 else
 783                         EMIT1(0x89);
 784                 break;
 785         case BPF_DW:
 786                 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
 787                 break;
 788         }
 789         emit_insn_suffix(&prog, dst_reg, src_reg, off);
 790         *pprog = prog;
 791 }
 792
 793 static int emit_atomic(u8 **pprog, u8 atomic_op,
 794                        u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
 795 {
 796         u8 *prog = *pprog;
 797
 798         EMIT1(0xF0); /* lock prefix */
 799
 800         maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
 801
 802         /* emit opcode */
 803         switch (atomic_op) {
 804         case BPF_ADD:
 805         case BPF_SUB:
 806         case BPF_AND:
 807         case BPF_OR:
 808         case BPF_XOR:
 809                 /* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
 810                 EMIT1(simple_alu_opcodes[atomic_op]);
 811                 break;
 812         case BPF_ADD | BPF_FETCH:
 813                 /* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
 814                 EMIT2(0x0F, 0xC1);
 815                 break;
 816         case BPF_XCHG:
 817                 /* src_reg = atomic_xchg(dst_reg + off, src_reg); */
 818                 EMIT1(0x87);
 819                 break;
 820         case BPF_CMPXCHG:
 821                 /* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
 822                 EMIT2(0x0F, 0xB1);
 823                 break;
 824         default:
 825                 pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
 826                 return -EFAULT;
 827         }
 828
 829         emit_insn_suffix(&prog, dst_reg, src_reg, off);
 830
 831         *pprog = prog;
 832         return 0;
 833 }
 834
 835 static bool ex_handler_bpf(const struct exception_table_entry *x,
 836                            struct pt_regs *regs, int trapnr,
 837                            unsigned long error_code, unsigned long fault_addr)
 838 {
 839         u32 reg = x->fixup >> 8;
 840
 841         /* jump over faulting load and clear dest register */
 842         *(unsigned long *)((void *)regs + reg) = 0;
 843         regs->ip += x->fixup & 0xff;
 844         return true;
 845 }
 846
 847 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
 848                              bool *regs_used, bool *tail_call_seen)
 849 {
 850         int i;
 851
 852         for (i = 1; i <= insn_cnt; i++, insn++) {
 853                 if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
 854                         *tail_call_seen = true;
 855                 if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
 856                         regs_used[0] = true;
 857                 if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
 858                         regs_used[1] = true;
 859                 if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
 860                         regs_used[2] = true;
 861                 if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
 862                         regs_used[3] = true;
 863         }
 864 }
 865
 866 static void emit_nops(u8 **pprog, int len)
 867 {
 868         u8 *prog = *pprog;
 869         int i, noplen;
 870
 871         while (len > 0) {
 872                 noplen = len;
 873
 874                 if (noplen > ASM_NOP_MAX)
 875                         noplen = ASM_NOP_MAX;
 876
 877                 for (i = 0; i < noplen; i++)
 878                         EMIT1(x86_nops[noplen][i]);
 879                 len -= noplen;
 880         }
 881
 882         *pprog = prog;
 883 }
 884
 885 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
 886
 887 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
 888                   int oldproglen, struct jit_context *ctx, bool jmp_padding)
 889 {
 890         bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
 891         struct bpf_insn *insn = bpf_prog->insnsi;
 892         bool callee_regs_used[4] = {};
 893         int insn_cnt = bpf_prog->len;
 894         bool tail_call_seen = false;
 895         bool seen_exit = false;
 896         u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
 897         int i, excnt = 0;
 898         int ilen, proglen = 0;
 899         u8 *prog = temp;
 900         int err;
 901
 902         detect_reg_usage(insn, insn_cnt, callee_regs_used,
 903                          &tail_call_seen);
 904
 905         /* tail call's presence in current prog implies it is reachable */
 906         tail_call_reachable |= tail_call_seen;
 907
 908         emit_prologue(&prog, bpf_prog->aux->stack_depth,
 909                       bpf_prog_was_classic(bpf_prog), tail_call_reachable,
 910                       bpf_prog->aux->func_idx != 0);
 911         push_callee_regs(&prog, callee_regs_used);
 912
 913         ilen = prog - temp;
 914         if (image)
 915                 memcpy(image + proglen, temp, ilen);
 916         proglen += ilen;
 917         addrs[0] = proglen;
 918         prog = temp;
 919
 920         for (i = 1; i <= insn_cnt; i++, insn++) {
 921                 const s32 imm32 = insn->imm;
 922                 u32 dst_reg = insn->dst_reg;
 923                 u32 src_reg = insn->src_reg;
 924                 u8 b2 = 0, b3 = 0;
 925                 u8 *start_of_ldx;
 926                 s64 jmp_offset;
 927                 u8 jmp_cond;
 928                 u8 *func;
 929                 int nops;
 930
 931                 switch (insn->code) {
 932                         /* ALU */
 933                 case BPF_ALU | BPF_ADD | BPF_X:
 934                 case BPF_ALU | BPF_SUB | BPF_X:
 935                 case BPF_ALU | BPF_AND | BPF_X:
 936                 case BPF_ALU | BPF_OR | BPF_X:
 937                 case BPF_ALU | BPF_XOR | BPF_X:
 938                 case BPF_ALU64 | BPF_ADD | BPF_X:
 939                 case BPF_ALU64 | BPF_SUB | BPF_X:
 940                 case BPF_ALU64 | BPF_AND | BPF_X:
 941                 case BPF_ALU64 | BPF_OR | BPF_X:
 942                 case BPF_ALU64 | BPF_XOR | BPF_X:
 943                         maybe_emit_mod(&prog, dst_reg, src_reg,
 944                                        BPF_CLASS(insn->code) == BPF_ALU64);
 945                         b2 = simple_alu_opcodes[BPF_OP(insn->code)];
 946                         EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
 947                         break;
 948
 949                 case BPF_ALU64 | BPF_MOV | BPF_X:
 950                 case BPF_ALU | BPF_MOV | BPF_X:
 951                         emit_mov_reg(&prog,
 952                                      BPF_CLASS(insn->code) == BPF_ALU64,
 953                                      dst_reg, src_reg);
 954                         break;
 955
 956                         /* neg dst */
 957                 case BPF_ALU | BPF_NEG:
 958                 case BPF_ALU64 | BPF_NEG:
 959                         maybe_emit_1mod(&prog, dst_reg,
 960                                         BPF_CLASS(insn->code) == BPF_ALU64);
 961                         EMIT2(0xF7, add_1reg(0xD8, dst_reg));
 962                         break;
 963
 964                 case BPF_ALU | BPF_ADD | BPF_K:
 965                 case BPF_ALU | BPF_SUB | BPF_K:
 966                 case BPF_ALU | BPF_AND | BPF_K:
 967                 case BPF_ALU | BPF_OR | BPF_K:
 968                 case BPF_ALU | BPF_XOR | BPF_K:
 969                 case BPF_ALU64 | BPF_ADD | BPF_K:
 970                 case BPF_ALU64 | BPF_SUB | BPF_K:
 971                 case BPF_ALU64 | BPF_AND | BPF_K:
 972                 case BPF_ALU64 | BPF_OR | BPF_K:
 973                 case BPF_ALU64 | BPF_XOR | BPF_K:
 974                         maybe_emit_1mod(&prog, dst_reg,
 975                                         BPF_CLASS(insn->code) == BPF_ALU64);
 976
 977                         /*
 978                          * b3 holds 'normal' opcode, b2 short form only valid
 979                          * in case dst is eax/rax.
 980                          */
 981                         switch (BPF_OP(insn->code)) {
 982                         case BPF_ADD:
 983                                 b3 = 0xC0;
 984                                 b2 = 0x05;
 985                                 break;
 986                         case BPF_SUB:
 987                                 b3 = 0xE8;
 988                                 b2 = 0x2D;
 989                                 break;
 990                         case BPF_AND:
 991                                 b3 = 0xE0;
 992                                 b2 = 0x25;
 993                                 break;
 994                         case BPF_OR:
 995                                 b3 = 0xC8;
 996                                 b2 = 0x0D;
 997                                 break;
 998                         case BPF_XOR:
 999                                 b3 = 0xF0;
1000                                 b2 = 0x35;
1001                                 break;
1002                         }
1003
1004                         if (is_imm8(imm32))
1005                                 EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1006                         else if (is_axreg(dst_reg))
1007                                 EMIT1_off32(b2, imm32);
1008                         else
1009                                 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1010                         break;
1011
1012                 case BPF_ALU64 | BPF_MOV | BPF_K:
1013                 case BPF_ALU | BPF_MOV | BPF_K:
1014                         emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1015                                        dst_reg, imm32);
1016                         break;
1017
1018                 case BPF_LD | BPF_IMM | BPF_DW:
1019                         emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1020                         insn++;
1021                         i++;
1022                         break;
1023
1024                         /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1025                 case BPF_ALU | BPF_MOD | BPF_X:
1026                 case BPF_ALU | BPF_DIV | BPF_X:
1027                 case BPF_ALU | BPF_MOD | BPF_K:
1028                 case BPF_ALU | BPF_DIV | BPF_K:
1029                 case BPF_ALU64 | BPF_MOD | BPF_X:
1030                 case BPF_ALU64 | BPF_DIV | BPF_X:
1031                 case BPF_ALU64 | BPF_MOD | BPF_K:
1032                 case BPF_ALU64 | BPF_DIV | BPF_K:
1033                         EMIT1(0x50); /* push rax */
1034                         EMIT1(0x52); /* push rdx */
1035
1036                         if (BPF_SRC(insn->code) == BPF_X)
1037                                 /* mov r11, src_reg */
1038                                 EMIT_mov(AUX_REG, src_reg);
1039                         else
1040                                 /* mov r11, imm32 */
1041                                 EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1042
1043                         /* mov rax, dst_reg */
1044                         EMIT_mov(BPF_REG_0, dst_reg);
1045
1046                         /*
1047                          * xor edx, edx
1048                          * equivalent to 'xor rdx, rdx', but one byte less
1049                          */
1050                         EMIT2(0x31, 0xd2);
1051
1052                         if (BPF_CLASS(insn->code) == BPF_ALU64)
1053                                 /* div r11 */
1054                                 EMIT3(0x49, 0xF7, 0xF3);
1055                         else
1056                                 /* div r11d */
1057                                 EMIT3(0x41, 0xF7, 0xF3);
1058
1059                         if (BPF_OP(insn->code) == BPF_MOD)
1060                                 /* mov r11, rdx */
1061                                 EMIT3(0x49, 0x89, 0xD3);
1062                         else
1063                                 /* mov r11, rax */
1064                                 EMIT3(0x49, 0x89, 0xC3);
1065
1066                         EMIT1(0x5A); /* pop rdx */
1067                         EMIT1(0x58); /* pop rax */
1068
1069                         /* mov dst_reg, r11 */
1070                         EMIT_mov(dst_reg, AUX_REG);
1071                         break;
1072
1073                 case BPF_ALU | BPF_MUL | BPF_K:
1074                 case BPF_ALU | BPF_MUL | BPF_X:
1075                 case BPF_ALU64 | BPF_MUL | BPF_K:
1076                 case BPF_ALU64 | BPF_MUL | BPF_X:
1077                 {
1078                         bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1079
1080                         if (dst_reg != BPF_REG_0)
1081                                 EMIT1(0x50); /* push rax */
1082                         if (dst_reg != BPF_REG_3)
1083                                 EMIT1(0x52); /* push rdx */
1084
1085                         /* mov r11, dst_reg */
1086                         EMIT_mov(AUX_REG, dst_reg);
1087
1088                         if (BPF_SRC(insn->code) == BPF_X)
1089                                 emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
1090                         else
1091                                 emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
1092
1093                         if (is64)
1094                                 EMIT1(add_1mod(0x48, AUX_REG));
1095                         else if (is_ereg(AUX_REG))
1096                                 EMIT1(add_1mod(0x40, AUX_REG));
1097                         /* mul(q) r11 */
1098                         EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
1099
1100                         if (dst_reg != BPF_REG_3)
1101                                 EMIT1(0x5A); /* pop rdx */
1102                         if (dst_reg != BPF_REG_0) {
1103                                 /* mov dst_reg, rax */
1104                                 EMIT_mov(dst_reg, BPF_REG_0);
1105                                 EMIT1(0x58); /* pop rax */
1106                         }
1107                         break;
1108                 }
1109                         /* Shifts */
1110                 case BPF_ALU | BPF_LSH | BPF_K:
1111                 case BPF_ALU | BPF_RSH | BPF_K:
1112                 case BPF_ALU | BPF_ARSH | BPF_K:
1113                 case BPF_ALU64 | BPF_LSH | BPF_K:
1114                 case BPF_ALU64 | BPF_RSH | BPF_K:
1115                 case BPF_ALU64 | BPF_ARSH | BPF_K:
1116                         maybe_emit_1mod(&prog, dst_reg,
1117                                         BPF_CLASS(insn->code) == BPF_ALU64);
1118
1119                         b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1120                         if (imm32 == 1)
1121                                 EMIT2(0xD1, add_1reg(b3, dst_reg));
1122                         else
1123                                 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1124                         break;
1125
1126                 case BPF_ALU | BPF_LSH | BPF_X:
1127                 case BPF_ALU | BPF_RSH | BPF_X:
1128                 case BPF_ALU | BPF_ARSH | BPF_X:
1129                 case BPF_ALU64 | BPF_LSH | BPF_X:
1130                 case BPF_ALU64 | BPF_RSH | BPF_X:
1131                 case BPF_ALU64 | BPF_ARSH | BPF_X:
1132
1133                         /* Check for bad case when dst_reg == rcx */
1134                         if (dst_reg == BPF_REG_4) {
1135                                 /* mov r11, dst_reg */
1136                                 EMIT_mov(AUX_REG, dst_reg);
1137                                 dst_reg = AUX_REG;
1138                         }
1139
1140                         if (src_reg != BPF_REG_4) { /* common case */
1141                                 EMIT1(0x51); /* push rcx */
1142
1143                                 /* mov rcx, src_reg */
1144                                 EMIT_mov(BPF_REG_4, src_reg);
1145                         }
1146
1147                         /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1148                         maybe_emit_1mod(&prog, dst_reg,
1149                                         BPF_CLASS(insn->code) == BPF_ALU64);
1150
1151                         b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1152                         EMIT2(0xD3, add_1reg(b3, dst_reg));
1153
1154                         if (src_reg != BPF_REG_4)
1155                                 EMIT1(0x59); /* pop rcx */
1156
1157                         if (insn->dst_reg == BPF_REG_4)
1158                                 /* mov dst_reg, r11 */
1159                                 EMIT_mov(insn->dst_reg, AUX_REG);
1160                         break;
1161
1162                 case BPF_ALU | BPF_END | BPF_FROM_BE:
1163                         switch (imm32) {
1164                         case 16:
1165                                 /* Emit 'ror %ax, 8' to swap lower 2 bytes */
1166                                 EMIT1(0x66);
1167                                 if (is_ereg(dst_reg))
1168                                         EMIT1(0x41);
1169                                 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1170
1171                                 /* Emit 'movzwl eax, ax' */
1172                                 if (is_ereg(dst_reg))
1173                                         EMIT3(0x45, 0x0F, 0xB7);
1174                                 else
1175                                         EMIT2(0x0F, 0xB7);
1176                                 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1177                                 break;
1178                         case 32:
1179                                 /* Emit 'bswap eax' to swap lower 4 bytes */
1180                                 if (is_ereg(dst_reg))
1181                                         EMIT2(0x41, 0x0F);
1182                                 else
1183                                         EMIT1(0x0F);
1184                                 EMIT1(add_1reg(0xC8, dst_reg));
1185                                 break;
1186                         case 64:
1187                                 /* Emit 'bswap rax' to swap 8 bytes */
1188                                 EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1189                                       add_1reg(0xC8, dst_reg));
1190                                 break;
1191                         }
1192                         break;
1193
1194                 case BPF_ALU | BPF_END | BPF_FROM_LE:
1195                         switch (imm32) {
1196                         case 16:
1197                                 /*
1198                                  * Emit 'movzwl eax, ax' to zero extend 16-bit
1199                                  * into 64 bit
1200                                  */
1201                                 if (is_ereg(dst_reg))
1202                                         EMIT3(0x45, 0x0F, 0xB7);
1203                                 else
1204                                         EMIT2(0x0F, 0xB7);
1205                                 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1206                                 break;
1207                         case 32:
1208                                 /* Emit 'mov eax, eax' to clear upper 32-bits */
1209                                 if (is_ereg(dst_reg))
1210                                         EMIT1(0x45);
1211                                 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1212                                 break;
1213                         case 64:
1214                                 /* nop */
1215                                 break;
1216                         }
1217                         break;
1218
1219                         /* speculation barrier */
1220                 case BPF_ST | BPF_NOSPEC:
1221                         if (boot_cpu_has(X86_FEATURE_XMM2))
1222                                 EMIT_LFENCE();
1223                         break;
1224
1225                         /* ST: *(u8*)(dst_reg + off) = imm */
1226                 case BPF_ST | BPF_MEM | BPF_B:
1227                         if (is_ereg(dst_reg))
1228                                 EMIT2(0x41, 0xC6);
1229                         else
1230                                 EMIT1(0xC6);
1231                         goto st;
1232                 case BPF_ST | BPF_MEM | BPF_H:
1233                         if (is_ereg(dst_reg))
1234                                 EMIT3(0x66, 0x41, 0xC7);
1235                         else
1236                                 EMIT2(0x66, 0xC7);
1237                         goto st;
1238                 case BPF_ST | BPF_MEM | BPF_W:
1239                         if (is_ereg(dst_reg))
1240                                 EMIT2(0x41, 0xC7);
1241                         else
1242                                 EMIT1(0xC7);
1243                         goto st;
1244                 case BPF_ST | BPF_MEM | BPF_DW:
1245                         EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1246
1247 st:                     if (is_imm8(insn->off))
1248                                 EMIT2(add_1reg(0x40, dst_reg), insn->off);
1249                         else
1250                                 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1251
1252                         EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1253                         break;
1254
1255                         /* STX: *(u8*)(dst_reg + off) = src_reg */
1256                 case BPF_STX | BPF_MEM | BPF_B:
1257                 case BPF_STX | BPF_MEM | BPF_H:
1258                 case BPF_STX | BPF_MEM | BPF_W:
1259                 case BPF_STX | BPF_MEM | BPF_DW:
1260                         emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1261                         break;
1262
1263                         /* LDX: dst_reg = *(u8*)(src_reg + off) */
1264                 case BPF_LDX | BPF_MEM | BPF_B:
1265                 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1266                 case BPF_LDX | BPF_MEM | BPF_H:
1267                 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1268                 case BPF_LDX | BPF_MEM | BPF_W:
1269                 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1270                 case BPF_LDX | BPF_MEM | BPF_DW:
1271                 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1272                         if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1273                                 /* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1274                                  * add abs(insn->off) to the limit to make sure that negative
1275                                  * offset won't be an issue.
1276                                  * insn->off is s16, so it won't affect valid pointers.
1277                                  */
1278                                 u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1279                                 u8 *end_of_jmp1, *end_of_jmp2;
1280
1281                                 /* Conservatively check that src_reg + insn->off is a kernel address:
1282                                  * 1. src_reg + insn->off >= limit
1283                                  * 2. src_reg + insn->off doesn't become small positive.
1284                                  * Cannot do src_reg + insn->off >= limit in one branch,
1285                                  * since it needs two spare registers, but JIT has only one.
1286                                  */
1287
1288                                 /* movabsq r11, limit */
1289                                 EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1290                                 EMIT((u32)limit, 4);
1291                                 EMIT(limit >> 32, 4);
1292                                 /* cmp src_reg, r11 */
1293                                 maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1294                                 EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1295                                 /* if unsigned '<' goto end_of_jmp2 */
1296                                 EMIT2(X86_JB, 0);
1297                                 end_of_jmp1 = prog;
1298
1299                                 /* mov r11, src_reg */
1300                                 emit_mov_reg(&prog, true, AUX_REG, src_reg);
1301                                 /* add r11, insn->off */
1302                                 maybe_emit_1mod(&prog, AUX_REG, true);
1303                                 EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1304                                 /* jmp if not carry to start_of_ldx
1305                                  * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1306                                  * that has to be rejected.
1307                                  */
1308                                 EMIT2(0x73 /* JNC */, 0);
1309                                 end_of_jmp2 = prog;
1310
1311                                 /* xor dst_reg, dst_reg */
1312                                 emit_mov_imm32(&prog, false, dst_reg, 0);
1313                                 /* jmp byte_after_ldx */
1314                                 EMIT2(0xEB, 0);
1315
1316                                 /* populate jmp_offset for JB above to jump to xor dst_reg */
1317                                 end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1318                                 /* populate jmp_offset for JNC above to jump to start_of_ldx */
1319                                 start_of_ldx = prog;
1320                                 end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1321                         }
1322                         emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1323                         if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1324                                 struct exception_table_entry *ex;
1325                                 u8 *_insn = image + proglen + (start_of_ldx - temp);
1326                                 s64 delta;
1327
1328                                 /* populate jmp_offset for JMP above */
1329                                 start_of_ldx[-1] = prog - start_of_ldx;
1330
1331                                 if (!bpf_prog->aux->extable)
1332                                         break;
1333
1334                                 if (excnt >= bpf_prog->aux->num_exentries) {
1335                                         pr_err("ex gen bug\n");
1336                                         return -EFAULT;
1337                                 }
1338                                 ex = &bpf_prog->aux->extable[excnt++];
1339
1340                                 delta = _insn - (u8 *)&ex->insn;
1341                                 if (!is_simm32(delta)) {
1342                                         pr_err("extable->insn doesn't fit into 32-bit\n");
1343                                         return -EFAULT;
1344                                 }
1345                                 ex->insn = delta;
1346
1347                                 delta = (u8 *)ex_handler_bpf - (u8 *)&ex->handler;
1348                                 if (!is_simm32(delta)) {
1349                                         pr_err("extable->handler doesn't fit into 32-bit\n");
1350                                         return -EFAULT;
1351                                 }
1352                                 ex->handler = delta;
1353
1354                                 if (dst_reg > BPF_REG_9) {
1355                                         pr_err("verifier error\n");
1356                                         return -EFAULT;
1357                                 }
1358                                 /*
1359                                  * Compute size of x86 insn and its target dest x86 register.
1360                                  * ex_handler_bpf() will use lower 8 bits to adjust
1361                                  * pt_regs->ip to jump over this x86 instruction
1362                                  * and upper bits to figure out which pt_regs to zero out.
1363                                  * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1364                                  * of 4 bytes will be ignored and rbx will be zero inited.
1365                                  */
1366                                 ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1367                         }
1368                         break;
1369
1370                 case BPF_STX | BPF_ATOMIC | BPF_W:
1371                 case BPF_STX | BPF_ATOMIC | BPF_DW:
1372                         if (insn->imm == (BPF_AND | BPF_FETCH) ||
1373                             insn->imm == (BPF_OR | BPF_FETCH) ||
1374                             insn->imm == (BPF_XOR | BPF_FETCH)) {
1375                                 bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1376                                 u32 real_src_reg = src_reg;
1377                                 u32 real_dst_reg = dst_reg;
1378                                 u8 *branch_target;
1379
1380                                 /*
1381                                  * Can't be implemented with a single x86 insn.
1382                                  * Need to do a CMPXCHG loop.
1383                                  */
1384
1385                                 /* Will need RAX as a CMPXCHG operand so save R0 */
1386                                 emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1387                                 if (src_reg == BPF_REG_0)
1388                                         real_src_reg = BPF_REG_AX;
1389                                 if (dst_reg == BPF_REG_0)
1390                                         real_dst_reg = BPF_REG_AX;
1391
1392                                 branch_target = prog;
1393                                 /* Load old value */
1394                                 emit_ldx(&prog, BPF_SIZE(insn->code),
1395                                          BPF_REG_0, real_dst_reg, insn->off);
1396                                 /*
1397                                  * Perform the (commutative) operation locally,
1398                                  * put the result in the AUX_REG.
1399                                  */
1400                                 emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1401                                 maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1402                                 EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1403                                       add_2reg(0xC0, AUX_REG, real_src_reg));
1404                                 /* Attempt to swap in new value */
1405                                 err = emit_atomic(&prog, BPF_CMPXCHG,
1406                                                   real_dst_reg, AUX_REG,
1407                                                   insn->off,
1408                                                   BPF_SIZE(insn->code));
1409                                 if (WARN_ON(err))
1410                                         return err;
1411                                 /*
1412                                  * ZF tells us whether we won the race. If it's
1413                                  * cleared we need to try again.
1414                                  */
1415                                 EMIT2(X86_JNE, -(prog - branch_target) - 2);
1416                                 /* Return the pre-modification value */
1417                                 emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1418                                 /* Restore R0 after clobbering RAX */
1419                                 emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1420                                 break;
1421                         }
1422
1423                         err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1424                                           insn->off, BPF_SIZE(insn->code));
1425                         if (err)
1426                                 return err;
1427                         break;
1428
1429                         /* call */
1430                 case BPF_JMP | BPF_CALL:
1431                         func = (u8 *) __bpf_call_base + imm32;
1432                         if (tail_call_reachable) {
1433                                 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1434                                 EMIT3_off32(0x48, 0x8B, 0x85,
1435                                             -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1436                                 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1437                                         return -EINVAL;
1438                         } else {
1439                                 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1440                                         return -EINVAL;
1441                         }
1442                         break;
1443
1444                 case BPF_JMP | BPF_TAIL_CALL:
1445                         if (imm32)
1446                                 emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1447                                                           &prog, image + addrs[i - 1],
1448                                                           callee_regs_used,
1449                                                           bpf_prog->aux->stack_depth,
1450                                                           ctx);
1451                         else
1452                                 emit_bpf_tail_call_indirect(&prog,
1453                                                             callee_regs_used,
1454                                                             bpf_prog->aux->stack_depth,
1455                                                             image + addrs[i - 1],
1456                                                             ctx);
1457                         break;
1458
1459                         /* cond jump */
1460                 case BPF_JMP | BPF_JEQ | BPF_X:
1461                 case BPF_JMP | BPF_JNE | BPF_X:
1462                 case BPF_JMP | BPF_JGT | BPF_X:
1463                 case BPF_JMP | BPF_JLT | BPF_X:
1464                 case BPF_JMP | BPF_JGE | BPF_X:
1465                 case BPF_JMP | BPF_JLE | BPF_X:
1466                 case BPF_JMP | BPF_JSGT | BPF_X:
1467                 case BPF_JMP | BPF_JSLT | BPF_X:
1468                 case BPF_JMP | BPF_JSGE | BPF_X:
1469                 case BPF_JMP | BPF_JSLE | BPF_X:
1470                 case BPF_JMP32 | BPF_JEQ | BPF_X:
1471                 case BPF_JMP32 | BPF_JNE | BPF_X:
1472                 case BPF_JMP32 | BPF_JGT | BPF_X:
1473                 case BPF_JMP32 | BPF_JLT | BPF_X:
1474                 case BPF_JMP32 | BPF_JGE | BPF_X:
1475                 case BPF_JMP32 | BPF_JLE | BPF_X:
1476                 case BPF_JMP32 | BPF_JSGT | BPF_X:
1477                 case BPF_JMP32 | BPF_JSLT | BPF_X:
1478                 case BPF_JMP32 | BPF_JSGE | BPF_X:
1479                 case BPF_JMP32 | BPF_JSLE | BPF_X:
1480                         /* cmp dst_reg, src_reg */
1481                         maybe_emit_mod(&prog, dst_reg, src_reg,
1482                                        BPF_CLASS(insn->code) == BPF_JMP);
1483                         EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1484                         goto emit_cond_jmp;
1485
1486                 case BPF_JMP | BPF_JSET | BPF_X:
1487                 case BPF_JMP32 | BPF_JSET | BPF_X:
1488                         /* test dst_reg, src_reg */
1489                         maybe_emit_mod(&prog, dst_reg, src_reg,
1490                                        BPF_CLASS(insn->code) == BPF_JMP);
1491                         EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1492                         goto emit_cond_jmp;
1493
1494                 case BPF_JMP | BPF_JSET | BPF_K:
1495                 case BPF_JMP32 | BPF_JSET | BPF_K:
1496                         /* test dst_reg, imm32 */
1497                         maybe_emit_1mod(&prog, dst_reg,
1498                                         BPF_CLASS(insn->code) == BPF_JMP);
1499                         EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1500                         goto emit_cond_jmp;
1501
1502                 case BPF_JMP | BPF_JEQ | BPF_K:
1503                 case BPF_JMP | BPF_JNE | BPF_K:
1504                 case BPF_JMP | BPF_JGT | BPF_K:
1505                 case BPF_JMP | BPF_JLT | BPF_K:
1506                 case BPF_JMP | BPF_JGE | BPF_K:
1507                 case BPF_JMP | BPF_JLE | BPF_K:
1508                 case BPF_JMP | BPF_JSGT | BPF_K:
1509                 case BPF_JMP | BPF_JSLT | BPF_K:
1510                 case BPF_JMP | BPF_JSGE | BPF_K:
1511                 case BPF_JMP | BPF_JSLE | BPF_K:
1512                 case BPF_JMP32 | BPF_JEQ | BPF_K:
1513                 case BPF_JMP32 | BPF_JNE | BPF_K:
1514                 case BPF_JMP32 | BPF_JGT | BPF_K:
1515                 case BPF_JMP32 | BPF_JLT | BPF_K:
1516                 case BPF_JMP32 | BPF_JGE | BPF_K:
1517                 case BPF_JMP32 | BPF_JLE | BPF_K:
1518                 case BPF_JMP32 | BPF_JSGT | BPF_K:
1519                 case BPF_JMP32 | BPF_JSLT | BPF_K:
1520                 case BPF_JMP32 | BPF_JSGE | BPF_K:
1521                 case BPF_JMP32 | BPF_JSLE | BPF_K:
1522                         /* test dst_reg, dst_reg to save one extra byte */
1523                         if (imm32 == 0) {
1524                                 maybe_emit_mod(&prog, dst_reg, dst_reg,
1525                                                BPF_CLASS(insn->code) == BPF_JMP);
1526                                 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1527                                 goto emit_cond_jmp;
1528                         }
1529
1530                         /* cmp dst_reg, imm8/32 */
1531                         maybe_emit_1mod(&prog, dst_reg,
1532                                         BPF_CLASS(insn->code) == BPF_JMP);
1533
1534                         if (is_imm8(imm32))
1535                                 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1536                         else
1537                                 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1538
1539 emit_cond_jmp:          /* Convert BPF opcode to x86 */
1540                         switch (BPF_OP(insn->code)) {
1541                         case BPF_JEQ:
1542                                 jmp_cond = X86_JE;
1543                                 break;
1544                         case BPF_JSET:
1545                         case BPF_JNE:
1546                                 jmp_cond = X86_JNE;
1547                                 break;
1548                         case BPF_JGT:
1549                                 /* GT is unsigned '>', JA in x86 */
1550                                 jmp_cond = X86_JA;
1551                                 break;
1552                         case BPF_JLT:
1553                                 /* LT is unsigned '<', JB in x86 */
1554                                 jmp_cond = X86_JB;
1555                                 break;
1556                         case BPF_JGE:
1557                                 /* GE is unsigned '>=', JAE in x86 */
1558                                 jmp_cond = X86_JAE;
1559                                 break;
1560                         case BPF_JLE:
1561                                 /* LE is unsigned '<=', JBE in x86 */
1562                                 jmp_cond = X86_JBE;
1563                                 break;
1564                         case BPF_JSGT:
1565                                 /* Signed '>', GT in x86 */
1566                                 jmp_cond = X86_JG;
1567                                 break;
1568                         case BPF_JSLT:
1569                                 /* Signed '<', LT in x86 */
1570                                 jmp_cond = X86_JL;
1571                                 break;
1572                         case BPF_JSGE:
1573                                 /* Signed '>=', GE in x86 */
1574                                 jmp_cond = X86_JGE;
1575                                 break;
1576                         case BPF_JSLE:
1577                                 /* Signed '<=', LE in x86 */
1578                                 jmp_cond = X86_JLE;
1579                                 break;
1580                         default: /* to silence GCC warning */
1581                                 return -EFAULT;
1582                         }
1583                         jmp_offset = addrs[i + insn->off] - addrs[i];
1584                         if (is_imm8(jmp_offset)) {
1585                                 if (jmp_padding) {
1586                                         /* To keep the jmp_offset valid, the extra bytes are
1587                                          * padded before the jump insn, so we subtract the
1588                                          * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1589                                          *
1590                                          * If the previous pass already emits an imm8
1591                                          * jmp_cond, then this BPF insn won't shrink, so
1592                                          * "nops" is 0.
1593                                          *
1594                                          * On the other hand, if the previous pass emits an
1595                                          * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1596                                          * keep the image from shrinking further.
1597                                          *
1598                                          * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1599                                          *     is 2 bytes, so the size difference is 4 bytes.
1600                                          */
1601                                         nops = INSN_SZ_DIFF - 2;
1602                                         if (nops != 0 && nops != 4) {
1603                                                 pr_err("unexpected jmp_cond padding: %d bytes\n",
1604                                                        nops);
1605                                                 return -EFAULT;
1606                                         }
1607                                         emit_nops(&prog, nops);
1608                                 }
1609                                 EMIT2(jmp_cond, jmp_offset);
1610                         } else if (is_simm32(jmp_offset)) {
1611                                 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1612                         } else {
1613                                 pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1614                                 return -EFAULT;
1615                         }
1616
1617                         break;
1618
1619                 case BPF_JMP | BPF_JA:
1620                         if (insn->off == -1)
1621                                 /* -1 jmp instructions will always jump
1622                                  * backwards two bytes. Explicitly handling
1623                                  * this case avoids wasting too many passes
1624                                  * when there are long sequences of replaced
1625                                  * dead code.
1626                                  */
1627                                 jmp_offset = -2;
1628                         else
1629                                 jmp_offset = addrs[i + insn->off] - addrs[i];
1630
1631                         if (!jmp_offset) {
1632                                 /*
1633                                  * If jmp_padding is enabled, the extra nops will
1634                                  * be inserted. Otherwise, optimize out nop jumps.
1635                                  */
1636                                 if (jmp_padding) {
1637                                         /* There are 3 possible conditions.
1638                                          * (1) This BPF_JA is already optimized out in
1639                                          *     the previous run, so there is no need
1640                                          *     to pad any extra byte (0 byte).
1641                                          * (2) The previous pass emits an imm8 jmp,
1642                                          *     so we pad 2 bytes to match the previous
1643                                          *     insn size.
1644                                          * (3) Similarly, the previous pass emits an
1645                                          *     imm32 jmp, and 5 bytes is padded.
1646                                          */
1647                                         nops = INSN_SZ_DIFF;
1648                                         if (nops != 0 && nops != 2 && nops != 5) {
1649                                                 pr_err("unexpected nop jump padding: %d bytes\n",
1650                                                        nops);
1651                                                 return -EFAULT;
1652                                         }
1653                                         emit_nops(&prog, nops);
1654                                 }
1655                                 break;
1656                         }
1657 emit_jmp:
1658                         if (is_imm8(jmp_offset)) {
1659                                 if (jmp_padding) {
1660                                         /* To avoid breaking jmp_offset, the extra bytes
1661                                          * are padded before the actual jmp insn, so
1662                                          * 2 bytes is subtracted from INSN_SZ_DIFF.
1663                                          *
1664                                          * If the previous pass already emits an imm8
1665                                          * jmp, there is nothing to pad (0 byte).
1666                                          *
1667                                          * If it emits an imm32 jmp (5 bytes) previously
1668                                          * and now an imm8 jmp (2 bytes), then we pad
1669                                          * (5 - 2 = 3) bytes to stop the image from
1670                                          * shrinking further.
1671                                          */
1672                                         nops = INSN_SZ_DIFF - 2;
1673                                         if (nops != 0 && nops != 3) {
1674                                                 pr_err("unexpected jump padding: %d bytes\n",
1675                                                        nops);
1676                                                 return -EFAULT;
1677                                         }
1678                                         emit_nops(&prog, INSN_SZ_DIFF - 2);
1679                                 }
1680                                 EMIT2(0xEB, jmp_offset);
1681                         } else if (is_simm32(jmp_offset)) {
1682                                 EMIT1_off32(0xE9, jmp_offset);
1683                         } else {
1684                                 pr_err("jmp gen bug %llx\n", jmp_offset);
1685                                 return -EFAULT;
1686                         }
1687                         break;
1688
1689                 case BPF_JMP | BPF_EXIT:
1690                         if (seen_exit) {
1691                                 jmp_offset = ctx->cleanup_addr - addrs[i];
1692                                 goto emit_jmp;
1693                         }
1694                         seen_exit = true;
1695                         /* Update cleanup_addr */
1696                         ctx->cleanup_addr = proglen;
1697                         pop_callee_regs(&prog, callee_regs_used);
1698                         EMIT1(0xC9);         /* leave */
1699                         emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1700                         break;
1701
1702                 default:
1703                         /*
1704                          * By design x86-64 JIT should support all BPF instructions.
1705                          * This error will be seen if new instruction was added
1706                          * to the interpreter, but not to the JIT, or if there is
1707                          * junk in bpf_prog.
1708                          */
1709                         pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1710                         return -EINVAL;
1711                 }
1712
1713                 ilen = prog - temp;
1714                 if (ilen > BPF_MAX_INSN_SIZE) {
1715                         pr_err("bpf_jit: fatal insn size error\n");
1716                         return -EFAULT;
1717                 }
1718
1719                 if (image) {
1720                         /*
1721                          * When populating the image, assert that:
1722                          *
1723                          *  i) We do not write beyond the allocated space, and
1724                          * ii) addrs[i] did not change from the prior run, in order
1725                          *     to validate assumptions made for computing branch
1726                          *     displacements.
1727                          */
1728                         if (unlikely(proglen + ilen > oldproglen ||
1729                                      proglen + ilen != addrs[i])) {
1730                                 pr_err("bpf_jit: fatal error\n");
1731                                 return -EFAULT;
1732                         }
1733                         memcpy(image + proglen, temp, ilen);
1734                 }
1735                 proglen += ilen;
1736                 addrs[i] = proglen;
1737                 prog = temp;
1738         }
1739
1740         if (image && excnt != bpf_prog->aux->num_exentries) {
1741                 pr_err("extable is not populated\n");
1742                 return -EFAULT;
1743         }
1744         return proglen;
1745 }
1746
1747 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1748                       int stack_size)
1749 {
1750         int i;
1751         /* Store function arguments to stack.
1752          * For a function that accepts two pointers the sequence will be:
1753          * mov QWORD PTR [rbp-0x10],rdi
1754          * mov QWORD PTR [rbp-0x8],rsi
1755          */
1756         for (i = 0; i < min(nr_args, 6); i++)
1757                 emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1758                          BPF_REG_FP,
1759                          i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1760                          -(stack_size - i * 8));
1761 }
1762
1763 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1764                          int stack_size)
1765 {
1766         int i;
1767
1768         /* Restore function arguments from stack.
1769          * For a function that accepts two pointers the sequence will be:
1770          * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1771          * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1772          */
1773         for (i = 0; i < min(nr_args, 6); i++)
1774                 emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1775                          i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1776                          BPF_REG_FP,
1777                          -(stack_size - i * 8));
1778 }
1779
1780 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1781                            struct bpf_prog *p, int stack_size, bool save_ret)
1782 {
1783         u8 *prog = *pprog;
1784         u8 *jmp_insn;
1785
1786         /* arg1: mov rdi, progs[i] */
1787         emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1788         if (emit_call(&prog,
1789                       p->aux->sleepable ? __bpf_prog_enter_sleepable :
1790                       __bpf_prog_enter, prog))
1791                         return -EINVAL;
1792         /* remember prog start time returned by __bpf_prog_enter */
1793         emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1794
1795         /* if (__bpf_prog_enter*(prog) == 0)
1796          *      goto skip_exec_of_prog;
1797          */
1798         EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1799         /* emit 2 nops that will be replaced with JE insn */
1800         jmp_insn = prog;
1801         emit_nops(&prog, 2);
1802
1803         /* arg1: lea rdi, [rbp - stack_size] */
1804         EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1805         /* arg2: progs[i]->insnsi for interpreter */
1806         if (!p->jited)
1807                 emit_mov_imm64(&prog, BPF_REG_2,
1808                                (long) p->insnsi >> 32,
1809                                (u32) (long) p->insnsi);
1810         /* call JITed bpf program or interpreter */
1811         if (emit_call(&prog, p->bpf_func, prog))
1812                 return -EINVAL;
1813
1814         /*
1815          * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1816          * of the previous call which is then passed on the stack to
1817          * the next BPF program.
1818          *
1819          * BPF_TRAMP_FENTRY trampoline may need to return the return
1820          * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1821          */
1822         if (save_ret)
1823                 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1824
1825         /* replace 2 nops with JE insn, since jmp target is known */
1826         jmp_insn[0] = X86_JE;
1827         jmp_insn[1] = prog - jmp_insn - 2;
1828
1829         /* arg1: mov rdi, progs[i] */
1830         emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1831         /* arg2: mov rsi, rbx <- start time in nsec */
1832         emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1833         if (emit_call(&prog,
1834                       p->aux->sleepable ? __bpf_prog_exit_sleepable :
1835                       __bpf_prog_exit, prog))
1836                         return -EINVAL;
1837
1838         *pprog = prog;
1839         return 0;
1840 }
1841
1842 static void emit_align(u8 **pprog, u32 align)
1843 {
1844         u8 *target, *prog = *pprog;
1845
1846         target = PTR_ALIGN(prog, align);
1847         if (target != prog)
1848                 emit_nops(&prog, target - prog);
1849
1850         *pprog = prog;
1851 }
1852
1853 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1854 {
1855         u8 *prog = *pprog;
1856         s64 offset;
1857
1858         offset = func - (ip + 2 + 4);
1859         if (!is_simm32(offset)) {
1860                 pr_err("Target %p is out of range\n", func);
1861                 return -EINVAL;
1862         }
1863         EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1864         *pprog = prog;
1865         return 0;
1866 }
1867
1868 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1869                       struct bpf_tramp_progs *tp, int stack_size,
1870                       bool save_ret)
1871 {
1872         int i;
1873         u8 *prog = *pprog;
1874
1875         for (i = 0; i < tp->nr_progs; i++) {
1876                 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size,
1877                                     save_ret))
1878                         return -EINVAL;
1879         }
1880         *pprog = prog;
1881         return 0;
1882 }
1883
1884 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1885                               struct bpf_tramp_progs *tp, int stack_size,
1886                               u8 **branches)
1887 {
1888         u8 *prog = *pprog;
1889         int i;
1890
1891         /* The first fmod_ret program will receive a garbage return value.
1892          * Set this to 0 to avoid confusing the program.
1893          */
1894         emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1895         emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1896         for (i = 0; i < tp->nr_progs; i++) {
1897                 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1898                         return -EINVAL;
1899
1900                 /* mod_ret prog stored return value into [rbp - 8]. Emit:
1901                  * if (*(u64 *)(rbp - 8) !=  0)
1902                  *      goto do_fexit;
1903                  */
1904                 /* cmp QWORD PTR [rbp - 0x8], 0x0 */
1905                 EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1906
1907                 /* Save the location of the branch and Generate 6 nops
1908                  * (4 bytes for an offset and 2 bytes for the jump) These nops
1909                  * are replaced with a conditional jump once do_fexit (i.e. the
1910                  * start of the fexit invocation) is finalized.
1911                  */
1912                 branches[i] = prog;
1913                 emit_nops(&prog, 4 + 2);
1914         }
1915
1916         *pprog = prog;
1917         return 0;
1918 }
1919
1920 static bool is_valid_bpf_tramp_flags(unsigned int flags)
1921 {
1922         if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1923             (flags & BPF_TRAMP_F_SKIP_FRAME))
1924                 return false;
1925
1926         /*
1927          * BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
1928          * and it must be used alone.
1929          */
1930         if ((flags & BPF_TRAMP_F_RET_FENTRY_RET) &&
1931             (flags & ~BPF_TRAMP_F_RET_FENTRY_RET))
1932                 return false;
1933
1934         return true;
1935 }
1936
1937 /* Example:
1938  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1939  * its 'struct btf_func_model' will be nr_args=2
1940  * The assembly code when eth_type_trans is executing after trampoline:
1941  *
1942  * push rbp
1943  * mov rbp, rsp
1944  * sub rsp, 16                     // space for skb and dev
1945  * push rbx                        // temp regs to pass start time
1946  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1947  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1948  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1949  * mov rbx, rax                    // remember start time in bpf stats are enabled
1950  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1951  * call addr_of_jited_FENTRY_prog
1952  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1953  * mov rsi, rbx                    // prog start time
1954  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1955  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
1956  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
1957  * pop rbx
1958  * leave
1959  * ret
1960  *
1961  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1962  * replaced with 'call generated_bpf_trampoline'. When it returns
1963  * eth_type_trans will continue executing with original skb and dev pointers.
1964  *
1965  * The assembly code when eth_type_trans is called from trampoline:
1966  *
1967  * push rbp
1968  * mov rbp, rsp
1969  * sub rsp, 24                     // space for skb, dev, return value
1970  * push rbx                        // temp regs to pass start time
1971  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
1972  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
1973  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1974  * mov rbx, rax                    // remember start time if bpf stats are enabled
1975  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1976  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
1977  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1978  * mov rsi, rbx                    // prog start time
1979  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1980  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
1981  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
1982  * call eth_type_trans+5           // execute body of eth_type_trans
1983  * mov qword ptr [rbp - 8], rax    // save return value
1984  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1985  * mov rbx, rax                    // remember start time in bpf stats are enabled
1986  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1987  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
1988  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1989  * mov rsi, rbx                    // prog start time
1990  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1991  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
1992  * pop rbx
1993  * leave
1994  * add rsp, 8                      // skip eth_type_trans's frame
1995  * ret                             // return to its caller
1996  */
1997 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1998                                 const struct btf_func_model *m, u32 flags,
1999                                 struct bpf_tramp_progs *tprogs,
2000                                 void *orig_call)
2001 {
2002         int ret, i, nr_args = m->nr_args;
2003         int stack_size = nr_args * 8;
2004         struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
2005         struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
2006         struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
2007         u8 **branches = NULL;
2008         u8 *prog;
2009         bool save_ret;
2010
2011         /* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2012         if (nr_args > 6)
2013                 return -ENOTSUPP;
2014
2015         if (!is_valid_bpf_tramp_flags(flags))
2016                 return -EINVAL;
2017
2018         /* room for return value of orig_call or fentry prog */
2019         save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2020         if (save_ret)
2021                 stack_size += 8;
2022
2023         if (flags & BPF_TRAMP_F_IP_ARG)
2024                 stack_size += 8; /* room for IP address argument */
2025
2026         if (flags & BPF_TRAMP_F_SKIP_FRAME)
2027                 /* skip patched call instruction and point orig_call to actual
2028                  * body of the kernel function.
2029                  */
2030                 orig_call += X86_PATCH_SIZE;
2031
2032         prog = image;
2033
2034         EMIT1(0x55);             /* push rbp */
2035         EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2036         EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2037         EMIT1(0x53);             /* push rbx */
2038
2039         if (flags & BPF_TRAMP_F_IP_ARG) {
2040                 /* Store IP address of the traced function:
2041                  * mov rax, QWORD PTR [rbp + 8]
2042                  * sub rax, X86_PATCH_SIZE
2043                  * mov QWORD PTR [rbp - stack_size], rax
2044                  */
2045                 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2046                 EMIT4(0x48, 0x83, 0xe8, X86_PATCH_SIZE);
2047                 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -stack_size);
2048
2049                 /* Continue with stack_size for regs storage, stack will
2050                  * be correctly restored with 'leave' instruction.
2051                  */
2052                 stack_size -= 8;
2053         }
2054
2055         save_regs(m, &prog, nr_args, stack_size);
2056
2057         if (flags & BPF_TRAMP_F_CALL_ORIG) {
2058                 /* arg1: mov rdi, im */
2059                 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2060                 if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2061                         ret = -EINVAL;
2062                         goto cleanup;
2063                 }
2064         }
2065
2066         if (fentry->nr_progs)
2067                 if (invoke_bpf(m, &prog, fentry, stack_size,
2068                                flags & BPF_TRAMP_F_RET_FENTRY_RET))
2069                         return -EINVAL;
2070
2071         if (fmod_ret->nr_progs) {
2072                 branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
2073                                    GFP_KERNEL);
2074                 if (!branches)
2075                         return -ENOMEM;
2076
2077                 if (invoke_bpf_mod_ret(m, &prog, fmod_ret, stack_size,
2078                                        branches)) {
2079                         ret = -EINVAL;
2080                         goto cleanup;
2081                 }
2082         }
2083
2084         if (flags & BPF_TRAMP_F_CALL_ORIG) {
2085                 restore_regs(m, &prog, nr_args, stack_size);
2086
2087                 /* call original function */
2088                 if (emit_call(&prog, orig_call, prog)) {
2089                         ret = -EINVAL;
2090                         goto cleanup;
2091                 }
2092                 /* remember return value in a stack for bpf prog to access */
2093                 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2094                 im->ip_after_call = prog;
2095                 memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2096                 prog += X86_PATCH_SIZE;
2097         }
2098
2099         if (fmod_ret->nr_progs) {
2100                 /* From Intel 64 and IA-32 Architectures Optimization
2101                  * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2102                  * Coding Rule 11: All branch targets should be 16-byte
2103                  * aligned.
2104                  */
2105                 emit_align(&prog, 16);
2106                 /* Update the branches saved in invoke_bpf_mod_ret with the
2107                  * aligned address of do_fexit.
2108                  */
2109                 for (i = 0; i < fmod_ret->nr_progs; i++)
2110                         emit_cond_near_jump(&branches[i], prog, branches[i],
2111                                             X86_JNE);
2112         }
2113
2114         if (fexit->nr_progs)
2115                 if (invoke_bpf(m, &prog, fexit, stack_size, false)) {
2116                         ret = -EINVAL;
2117                         goto cleanup;
2118                 }
2119
2120         if (flags & BPF_TRAMP_F_RESTORE_REGS)
2121                 restore_regs(m, &prog, nr_args, stack_size);
2122
2123         /* This needs to be done regardless. If there were fmod_ret programs,
2124          * the return value is only updated on the stack and still needs to be
2125          * restored to R0.
2126          */
2127         if (flags & BPF_TRAMP_F_CALL_ORIG) {
2128                 im->ip_epilogue = prog;
2129                 /* arg1: mov rdi, im */
2130                 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2131                 if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2132                         ret = -EINVAL;
2133                         goto cleanup;
2134                 }
2135         }
2136         /* restore return value of orig_call or fentry prog back into RAX */
2137         if (save_ret)
2138                 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2139
2140         EMIT1(0x5B); /* pop rbx */
2141         EMIT1(0xC9); /* leave */
2142         if (flags & BPF_TRAMP_F_SKIP_FRAME)
2143                 /* skip our return address and return to parent */
2144                 EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2145         emit_return(&prog, prog);
2146         /* Make sure the trampoline generation logic doesn't overflow */
2147         if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2148                 ret = -EFAULT;
2149                 goto cleanup;
2150         }
2151         ret = prog - (u8 *)image;
2152
2153 cleanup:
2154         kfree(branches);
2155         return ret;
2156 }
2157
2158 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2159 {
2160         u8 *jg_reloc, *prog = *pprog;
2161         int pivot, err, jg_bytes = 1;
2162         s64 jg_offset;
2163
2164         if (a == b) {
2165                 /* Leaf node of recursion, i.e. not a range of indices
2166                  * anymore.
2167                  */
2168                 EMIT1(add_1mod(0x48, BPF_REG_3));       /* cmp rdx,func */
2169                 if (!is_simm32(progs[a]))
2170                         return -1;
2171                 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2172                             progs[a]);
2173                 err = emit_cond_near_jump(&prog,        /* je func */
2174                                           (void *)progs[a], prog,
2175                                           X86_JE);
2176                 if (err)
2177                         return err;
2178
2179                 emit_indirect_jump(&prog, 2 /* rdx */, prog);
2180
2181                 *pprog = prog;
2182                 return 0;
2183         }
2184
2185         /* Not a leaf node, so we pivot, and recursively descend into
2186          * the lower and upper ranges.
2187          */
2188         pivot = (b - a) / 2;
2189         EMIT1(add_1mod(0x48, BPF_REG_3));               /* cmp rdx,func */
2190         if (!is_simm32(progs[a + pivot]))
2191                 return -1;
2192         EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2193
2194         if (pivot > 2) {                                /* jg upper_part */
2195                 /* Require near jump. */
2196                 jg_bytes = 4;
2197                 EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2198         } else {
2199                 EMIT2(X86_JG, 0);
2200         }
2201         jg_reloc = prog;
2202
2203         err = emit_bpf_dispatcher(&prog, a, a + pivot,  /* emit lower_part */
2204                                   progs);
2205         if (err)
2206                 return err;
2207
2208         /* From Intel 64 and IA-32 Architectures Optimization
2209          * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2210          * Coding Rule 11: All branch targets should be 16-byte
2211          * aligned.
2212          */
2213         emit_align(&prog, 16);
2214         jg_offset = prog - jg_reloc;
2215         emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2216
2217         err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */
2218                                   b, progs);
2219         if (err)
2220                 return err;
2221
2222         *pprog = prog;
2223         return 0;
2224 }
2225
2226 static int cmp_ips(const void *a, const void *b)
2227 {
2228         const s64 *ipa = a;
2229         const s64 *ipb = b;
2230
2231         if (*ipa > *ipb)
2232                 return 1;
2233         if (*ipa < *ipb)
2234                 return -1;
2235         return 0;
2236 }
2237
2238 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2239 {
2240         u8 *prog = image;
2241
2242         sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2243         return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2244 }
2245
2246 struct x64_jit_data {
2247         struct bpf_binary_header *header;
2248         int *addrs;
2249         u8 *image;
2250         int proglen;
2251         struct jit_context ctx;
2252 };
2253
2254 #define MAX_PASSES 20
2255 #define PADDING_PASSES (MAX_PASSES - 5)
2256
2257 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2258 {
2259         struct bpf_binary_header *header = NULL;
2260         struct bpf_prog *tmp, *orig_prog = prog;
2261         struct x64_jit_data *jit_data;
2262         int proglen, oldproglen = 0;
2263         struct jit_context ctx = {};
2264         bool tmp_blinded = false;
2265         bool extra_pass = false;
2266         bool padding = false;
2267         u8 *image = NULL;
2268         int *addrs;
2269         int pass;
2270         int i;
2271
2272         if (!prog->jit_requested)
2273                 return orig_prog;
2274
2275         tmp = bpf_jit_blind_constants(prog);
2276         /*
2277          * If blinding was requested and we failed during blinding,
2278          * we must fall back to the interpreter.
2279          */
2280         if (IS_ERR(tmp))
2281                 return orig_prog;
2282         if (tmp != prog) {
2283                 tmp_blinded = true;
2284                 prog = tmp;
2285         }
2286
2287         jit_data = prog->aux->jit_data;
2288         if (!jit_data) {
2289                 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2290                 if (!jit_data) {
2291                         prog = orig_prog;
2292                         goto out;
2293                 }
2294                 prog->aux->jit_data = jit_data;
2295         }
2296         addrs = jit_data->addrs;
2297         if (addrs) {
2298                 ctx = jit_data->ctx;
2299                 oldproglen = jit_data->proglen;
2300                 image = jit_data->image;
2301                 header = jit_data->header;
2302                 extra_pass = true;
2303                 padding = true;
2304                 goto skip_init_addrs;
2305         }
2306         addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2307         if (!addrs) {
2308                 prog = orig_prog;
2309                 goto out_addrs;
2310         }
2311
2312         /*
2313          * Before first pass, make a rough estimation of addrs[]
2314          * each BPF instruction is translated to less than 64 bytes
2315          */
2316         for (proglen = 0, i = 0; i <= prog->len; i++) {
2317                 proglen += 64;
2318                 addrs[i] = proglen;
2319         }
2320         ctx.cleanup_addr = proglen;
2321 skip_init_addrs:
2322
2323         /*
2324          * JITed image shrinks with every pass and the loop iterates
2325          * until the image stops shrinking. Very large BPF programs
2326          * may converge on the last pass. In such case do one more
2327          * pass to emit the final image.
2328          */
2329         for (pass = 0; pass < MAX_PASSES || image; pass++) {
2330                 if (!padding && pass >= PADDING_PASSES)
2331                         padding = true;
2332                 proglen = do_jit(prog, addrs, image, oldproglen, &ctx, padding);
2333                 if (proglen <= 0) {
2334 out_image:
2335                         image = NULL;
2336                         if (header)
2337                                 bpf_jit_binary_free(header);
2338                         prog = orig_prog;
2339                         goto out_addrs;
2340                 }
2341                 if (image) {
2342                         if (proglen != oldproglen) {
2343                                 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2344                                        proglen, oldproglen);
2345                                 goto out_image;
2346                         }
2347                         break;
2348                 }
2349                 if (proglen == oldproglen) {
2350                         /*
2351                          * The number of entries in extable is the number of BPF_LDX
2352                          * insns that access kernel memory via "pointer to BTF type".
2353                          * The verifier changed their opcode from LDX|MEM|size
2354                          * to LDX|PROBE_MEM|size to make JITing easier.
2355                          */
2356                         u32 align = __alignof__(struct exception_table_entry);
2357                         u32 extable_size = prog->aux->num_exentries *
2358                                 sizeof(struct exception_table_entry);
2359
2360                         /* allocate module memory for x86 insns and extable */
2361                         header = bpf_jit_binary_alloc(roundup(proglen, align) + extable_size,
2362                                                       &image, align, jit_fill_hole);
2363                         if (!header) {
2364                                 prog = orig_prog;
2365                                 goto out_addrs;
2366                         }
2367                         prog->aux->extable = (void *) image + roundup(proglen, align);
2368                 }
2369                 oldproglen = proglen;
2370                 cond_resched();
2371         }
2372
2373         if (bpf_jit_enable > 1)
2374                 bpf_jit_dump(prog->len, proglen, pass + 1, image);
2375
2376         if (image) {
2377                 if (!prog->is_func || extra_pass) {
2378                         bpf_tail_call_direct_fixup(prog);
2379                         bpf_jit_binary_lock_ro(header);
2380                 } else {
2381                         jit_data->addrs = addrs;
2382                         jit_data->ctx = ctx;
2383                         jit_data->proglen = proglen;
2384                         jit_data->image = image;
2385                         jit_data->header = header;
2386                 }
2387                 prog->bpf_func = (void *)image;
2388                 prog->jited = 1;
2389                 prog->jited_len = proglen;
2390         } else {
2391                 prog = orig_prog;
2392         }
2393
2394         if (!image || !prog->is_func || extra_pass) {
2395                 if (image)
2396                         bpf_prog_fill_jited_linfo(prog, addrs + 1);
2397 out_addrs:
2398                 kvfree(addrs);
2399                 kfree(jit_data);
2400                 prog->aux->jit_data = NULL;
2401         }
2402 out:
2403         if (tmp_blinded)
2404                 bpf_jit_prog_release_other(prog, prog == orig_prog ?
2405                                            tmp : orig_prog);
2406         return prog;
2407 }
2408
2409 bool bpf_jit_supports_kfunc_call(void)
2410 {
2411         return true;
2412 }