1 // SPDX-License-Identifier: GPL-2.0-only
3 * bpf_jit_comp.c: BPF JIT compiler
5 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6 * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
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 #include <asm/asm-prototypes.h>
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
33 #define EMIT(bytes, len) \
34 do { prog = emit_code(prog, bytes, len); } while (0)
36 #define EMIT1(b1) EMIT(b1, 1)
37 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2)
38 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
39 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
41 #define EMIT1_off32(b1, off) \
42 do { EMIT1(b1); EMIT(off, 4); } while (0)
43 #define EMIT2_off32(b1, b2, off) \
44 do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
45 #define EMIT3_off32(b1, b2, b3, off) \
46 do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
47 #define EMIT4_off32(b1, b2, b3, b4, off) \
48 do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
50 static bool is_imm8(int value)
52 return value <= 127 && value >= -128;
55 static bool is_simm32(s64 value)
57 return value == (s64)(s32)value;
60 static bool is_uimm32(u64 value)
62 return value == (u64)(u32)value;
66 #define EMIT_mov(DST, SRC) \
69 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
72 static int bpf_size_to_x86_bytes(int bpf_size)
74 if (bpf_size == BPF_W)
76 else if (bpf_size == BPF_H)
78 else if (bpf_size == BPF_B)
80 else if (bpf_size == BPF_DW)
87 * List of x86 cond jumps opcodes (. + s8)
88 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
101 /* Pick a register outside of BPF range for JIT internal work */
102 #define AUX_REG (MAX_BPF_JIT_REG + 1)
103 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
106 * The following table maps BPF registers to x86-64 registers.
108 * x86-64 register R12 is unused, since if used as base address
109 * register in load/store instructions, it always needs an
110 * extra byte of encoding and is callee saved.
112 * x86-64 register R9 is not used by BPF programs, but can be used by BPF
113 * trampoline. x86-64 register R10 is used for blinding (if enabled).
115 static const int reg2hex[] = {
116 [BPF_REG_0] = 0, /* RAX */
117 [BPF_REG_1] = 7, /* RDI */
118 [BPF_REG_2] = 6, /* RSI */
119 [BPF_REG_3] = 2, /* RDX */
120 [BPF_REG_4] = 1, /* RCX */
121 [BPF_REG_5] = 0, /* R8 */
122 [BPF_REG_6] = 3, /* RBX callee saved */
123 [BPF_REG_7] = 5, /* R13 callee saved */
124 [BPF_REG_8] = 6, /* R14 callee saved */
125 [BPF_REG_9] = 7, /* R15 callee saved */
126 [BPF_REG_FP] = 5, /* RBP readonly */
127 [BPF_REG_AX] = 2, /* R10 temp register */
128 [AUX_REG] = 3, /* R11 temp register */
129 [X86_REG_R9] = 1, /* R9 register, 6th function argument */
132 static const int reg2pt_regs[] = {
133 [BPF_REG_0] = offsetof(struct pt_regs, ax),
134 [BPF_REG_1] = offsetof(struct pt_regs, di),
135 [BPF_REG_2] = offsetof(struct pt_regs, si),
136 [BPF_REG_3] = offsetof(struct pt_regs, dx),
137 [BPF_REG_4] = offsetof(struct pt_regs, cx),
138 [BPF_REG_5] = offsetof(struct pt_regs, r8),
139 [BPF_REG_6] = offsetof(struct pt_regs, bx),
140 [BPF_REG_7] = offsetof(struct pt_regs, r13),
141 [BPF_REG_8] = offsetof(struct pt_regs, r14),
142 [BPF_REG_9] = offsetof(struct pt_regs, r15),
146 * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
147 * which need extra byte of encoding.
148 * rax,rcx,...,rbp have simpler encoding
150 static bool is_ereg(u32 reg)
152 return (1 << reg) & (BIT(BPF_REG_5) |
162 * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
163 * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
164 * of encoding. al,cl,dl,bl have simpler encoding.
166 static bool is_ereg_8l(u32 reg)
168 return is_ereg(reg) ||
169 (1 << reg) & (BIT(BPF_REG_1) |
174 static bool is_axreg(u32 reg)
176 return reg == BPF_REG_0;
179 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
180 static u8 add_1mod(u8 byte, u32 reg)
187 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
196 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
197 static u8 add_1reg(u8 byte, u32 dst_reg)
199 return byte + reg2hex[dst_reg];
202 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
203 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
205 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
208 /* Some 1-byte opcodes for binary ALU operations */
209 static u8 simple_alu_opcodes[] = {
220 static void jit_fill_hole(void *area, unsigned int size)
222 /* Fill whole space with INT3 instructions */
223 memset(area, 0xcc, size);
227 int cleanup_addr; /* Epilogue code offset */
230 /* Maximum number of bytes emitted while JITing one eBPF insn */
231 #define BPF_MAX_INSN_SIZE 128
232 #define BPF_INSN_SAFETY 64
234 /* Number of bytes emit_patch() needs to generate instructions */
235 #define X86_PATCH_SIZE 5
236 /* Number of bytes that will be skipped on tailcall */
237 #define X86_TAIL_CALL_OFFSET 11
239 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
243 if (callee_regs_used[0])
244 EMIT1(0x53); /* push rbx */
245 if (callee_regs_used[1])
246 EMIT2(0x41, 0x55); /* push r13 */
247 if (callee_regs_used[2])
248 EMIT2(0x41, 0x56); /* push r14 */
249 if (callee_regs_used[3])
250 EMIT2(0x41, 0x57); /* push r15 */
254 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
258 if (callee_regs_used[3])
259 EMIT2(0x41, 0x5F); /* pop r15 */
260 if (callee_regs_used[2])
261 EMIT2(0x41, 0x5E); /* pop r14 */
262 if (callee_regs_used[1])
263 EMIT2(0x41, 0x5D); /* pop r13 */
264 if (callee_regs_used[0])
265 EMIT1(0x5B); /* pop rbx */
270 * Emit x86-64 prologue code for BPF program.
271 * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
272 * while jumping to another program
274 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
275 bool tail_call_reachable, bool is_subprog)
279 /* BPF trampoline can be made to work without these nops,
280 * but let's waste 5 bytes for now and optimize later
282 memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
283 prog += X86_PATCH_SIZE;
284 if (!ebpf_from_cbpf) {
285 if (tail_call_reachable && !is_subprog)
286 EMIT2(0x31, 0xC0); /* xor eax, eax */
288 EMIT2(0x66, 0x90); /* nop2 */
290 EMIT1(0x55); /* push rbp */
291 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
292 /* sub rsp, rounded_stack_depth */
294 EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
295 if (tail_call_reachable)
296 EMIT1(0x50); /* push rax */
300 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
305 offset = func - (ip + X86_PATCH_SIZE);
306 if (!is_simm32(offset)) {
307 pr_err("Target call %p is out of range\n", func);
310 EMIT1_off32(opcode, offset);
315 static int emit_call(u8 **pprog, void *func, void *ip)
317 return emit_patch(pprog, func, ip, 0xE8);
320 static int emit_jump(u8 **pprog, void *func, void *ip)
322 return emit_patch(pprog, func, ip, 0xE9);
325 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
326 void *old_addr, void *new_addr,
327 const bool text_live)
329 const u8 *nop_insn = x86_nops[5];
330 u8 old_insn[X86_PATCH_SIZE];
331 u8 new_insn[X86_PATCH_SIZE];
335 memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
338 ret = t == BPF_MOD_CALL ?
339 emit_call(&prog, old_addr, ip) :
340 emit_jump(&prog, old_addr, ip);
345 memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
348 ret = t == BPF_MOD_CALL ?
349 emit_call(&prog, new_addr, ip) :
350 emit_jump(&prog, new_addr, ip);
356 mutex_lock(&text_mutex);
357 if (memcmp(ip, old_insn, X86_PATCH_SIZE))
360 if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
362 text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
364 memcpy(ip, new_insn, X86_PATCH_SIZE);
368 mutex_unlock(&text_mutex);
372 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
373 void *old_addr, void *new_addr)
375 if (!is_kernel_text((long)ip) &&
376 !is_bpf_text_address((long)ip))
377 /* BPF poking in modules is not supported */
380 return __bpf_arch_text_poke(ip, t, old_addr, new_addr, true);
383 static int get_pop_bytes(bool *callee_regs_used)
387 if (callee_regs_used[3])
389 if (callee_regs_used[2])
391 if (callee_regs_used[1])
393 if (callee_regs_used[0])
400 * Generate the following code:
402 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
403 * if (index >= array->map.max_entries)
405 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
407 * prog = array->ptrs[index];
410 * goto *(prog->bpf_func + prologue_size);
413 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
416 int tcc_off = -4 - round_up(stack_depth, 8);
423 /* count the additional bytes used for popping callee regs from stack
424 * that need to be taken into account for each of the offsets that
425 * are used for bailing out of the tail call
427 pop_bytes = get_pop_bytes(callee_regs_used);
439 * rdi - pointer to ctx
440 * rsi - pointer to bpf_array
441 * rdx - index in bpf_array
445 * if (index >= array->map.max_entries)
448 EMIT2(0x89, 0xD2); /* mov edx, edx */
449 EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */
450 offsetof(struct bpf_array, map.max_entries));
451 #define OFFSET1 (off1 + RETPOLINE_RCX_BPF_JIT_SIZE) /* Number of bytes to jump */
452 EMIT2(X86_JBE, OFFSET1); /* jbe out */
455 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
458 EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */
459 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
460 #define OFFSET2 (off2 + RETPOLINE_RCX_BPF_JIT_SIZE)
461 EMIT2(X86_JA, OFFSET2); /* ja out */
462 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
463 EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */
465 /* prog = array->ptrs[index]; */
466 EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6, /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
467 offsetof(struct bpf_array, ptrs));
473 EMIT3(0x48, 0x85, 0xC9); /* test rcx,rcx */
474 #define OFFSET3 (off3 + RETPOLINE_RCX_BPF_JIT_SIZE)
475 EMIT2(X86_JE, OFFSET3); /* je out */
478 pop_callee_regs(pprog, callee_regs_used);
481 EMIT1(0x58); /* pop rax */
483 EMIT3_off32(0x48, 0x81, 0xC4, /* add rsp, sd */
484 round_up(stack_depth, 8));
486 /* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
487 EMIT4(0x48, 0x8B, 0x49, /* mov rcx, qword ptr [rcx + 32] */
488 offsetof(struct bpf_prog, bpf_func));
489 EMIT4(0x48, 0x83, 0xC1, /* add rcx, X86_TAIL_CALL_OFFSET */
490 X86_TAIL_CALL_OFFSET);
492 * Now we're ready to jump into next BPF program
493 * rdi == ctx (1st arg)
494 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
496 RETPOLINE_RCX_BPF_JIT();
502 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
503 u8 **pprog, int addr, u8 *image,
504 bool *callee_regs_used, u32 stack_depth)
506 int tcc_off = -4 - round_up(stack_depth, 8);
512 /* count the additional bytes used for popping callee regs to stack
513 * that need to be taken into account for jump offset that is used for
514 * bailing out from of the tail call when limit is reached
516 pop_bytes = get_pop_bytes(callee_regs_used);
523 * - sub rsp, $val if depth > 0
526 poke_off = X86_PATCH_SIZE + pop_bytes + 1;
533 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
536 EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */
537 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
538 EMIT2(X86_JA, off1); /* ja out */
539 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
540 EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */
542 poke->tailcall_bypass = image + (addr - poke_off - X86_PATCH_SIZE);
543 poke->adj_off = X86_TAIL_CALL_OFFSET;
544 poke->tailcall_target = image + (addr - X86_PATCH_SIZE);
545 poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
547 emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
548 poke->tailcall_bypass);
551 pop_callee_regs(pprog, callee_regs_used);
553 EMIT1(0x58); /* pop rax */
555 EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
557 memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
558 prog += X86_PATCH_SIZE;
564 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
566 struct bpf_jit_poke_descriptor *poke;
567 struct bpf_array *array;
568 struct bpf_prog *target;
571 for (i = 0; i < prog->aux->size_poke_tab; i++) {
572 poke = &prog->aux->poke_tab[i];
573 if (poke->aux && poke->aux != prog->aux)
576 WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
578 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
581 array = container_of(poke->tail_call.map, struct bpf_array, map);
582 mutex_lock(&array->aux->poke_mutex);
583 target = array->ptrs[poke->tail_call.key];
585 /* Plain memcpy is used when image is not live yet
586 * and still not locked as read-only. Once poke
587 * location is active (poke->tailcall_target_stable),
588 * any parallel bpf_arch_text_poke() might occur
589 * still on the read-write image until we finally
590 * locked it as read-only. Both modifications on
591 * the given image are under text_mutex to avoid
594 ret = __bpf_arch_text_poke(poke->tailcall_target,
596 (u8 *)target->bpf_func +
597 poke->adj_off, false);
599 ret = __bpf_arch_text_poke(poke->tailcall_bypass,
601 (u8 *)poke->tailcall_target +
602 X86_PATCH_SIZE, NULL, false);
605 WRITE_ONCE(poke->tailcall_target_stable, true);
606 mutex_unlock(&array->aux->poke_mutex);
610 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
611 u32 dst_reg, const u32 imm32)
617 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
618 * (which zero-extends imm32) to save 2 bytes.
620 if (sign_propagate && (s32)imm32 < 0) {
621 /* 'mov %rax, imm32' sign extends imm32 */
622 b1 = add_1mod(0x48, dst_reg);
625 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
630 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
634 if (is_ereg(dst_reg))
635 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
638 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
642 /* mov %eax, imm32 */
643 if (is_ereg(dst_reg))
644 EMIT1(add_1mod(0x40, dst_reg));
645 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
650 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
651 const u32 imm32_hi, const u32 imm32_lo)
655 if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
657 * For emitting plain u32, where sign bit must not be
658 * propagated LLVM tends to load imm64 over mov32
659 * directly, so save couple of bytes by just doing
660 * 'mov %eax, imm32' instead.
662 emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
664 /* movabsq %rax, imm64 */
665 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
673 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
679 EMIT_mov(dst_reg, src_reg);
682 if (is_ereg(dst_reg) || is_ereg(src_reg))
683 EMIT1(add_2mod(0x40, dst_reg, src_reg));
684 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
690 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
691 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
696 /* 1-byte signed displacement.
698 * If off == 0 we could skip this and save one extra byte, but
699 * special case of x86 R13 which always needs an offset is not
702 EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
704 /* 4-byte signed displacement */
705 EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
711 * Emit a REX byte if it will be necessary to address these registers
713 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
718 EMIT1(add_2mod(0x48, dst_reg, src_reg));
719 else if (is_ereg(dst_reg) || is_ereg(src_reg))
720 EMIT1(add_2mod(0x40, dst_reg, src_reg));
725 * Similar version of maybe_emit_mod() for a single register
727 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
732 EMIT1(add_1mod(0x48, reg));
733 else if (is_ereg(reg))
734 EMIT1(add_1mod(0x40, reg));
738 /* LDX: dst_reg = *(u8*)(src_reg + off) */
739 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
745 /* Emit 'movzx rax, byte ptr [rax + off]' */
746 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
749 /* Emit 'movzx rax, word ptr [rax + off]' */
750 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
753 /* Emit 'mov eax, dword ptr [rax+0x14]' */
754 if (is_ereg(dst_reg) || is_ereg(src_reg))
755 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
760 /* Emit 'mov rax, qword ptr [rax+0x14]' */
761 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
764 emit_insn_suffix(&prog, src_reg, dst_reg, off);
768 /* STX: *(u8*)(dst_reg + off) = src_reg */
769 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
775 /* Emit 'mov byte ptr [rax + off], al' */
776 if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
777 /* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
778 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
783 if (is_ereg(dst_reg) || is_ereg(src_reg))
784 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
789 if (is_ereg(dst_reg) || is_ereg(src_reg))
790 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
795 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
798 emit_insn_suffix(&prog, dst_reg, src_reg, off);
802 static int emit_atomic(u8 **pprog, u8 atomic_op,
803 u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
807 EMIT1(0xF0); /* lock prefix */
809 maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
818 /* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
819 EMIT1(simple_alu_opcodes[atomic_op]);
821 case BPF_ADD | BPF_FETCH:
822 /* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
826 /* src_reg = atomic_xchg(dst_reg + off, src_reg); */
830 /* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
834 pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
838 emit_insn_suffix(&prog, dst_reg, src_reg, off);
844 static bool ex_handler_bpf(const struct exception_table_entry *x,
845 struct pt_regs *regs, int trapnr,
846 unsigned long error_code, unsigned long fault_addr)
848 u32 reg = x->fixup >> 8;
850 /* jump over faulting load and clear dest register */
851 *(unsigned long *)((void *)regs + reg) = 0;
852 regs->ip += x->fixup & 0xff;
856 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
857 bool *regs_used, bool *tail_call_seen)
861 for (i = 1; i <= insn_cnt; i++, insn++) {
862 if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
863 *tail_call_seen = true;
864 if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
866 if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
868 if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
870 if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
875 static void emit_nops(u8 **pprog, int len)
883 if (noplen > ASM_NOP_MAX)
884 noplen = ASM_NOP_MAX;
886 for (i = 0; i < noplen; i++)
887 EMIT1(x86_nops[noplen][i]);
894 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
896 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
897 int oldproglen, struct jit_context *ctx, bool jmp_padding)
899 bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
900 struct bpf_insn *insn = bpf_prog->insnsi;
901 bool callee_regs_used[4] = {};
902 int insn_cnt = bpf_prog->len;
903 bool tail_call_seen = false;
904 bool seen_exit = false;
905 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
907 int ilen, proglen = 0;
911 detect_reg_usage(insn, insn_cnt, callee_regs_used,
914 /* tail call's presence in current prog implies it is reachable */
915 tail_call_reachable |= tail_call_seen;
917 emit_prologue(&prog, bpf_prog->aux->stack_depth,
918 bpf_prog_was_classic(bpf_prog), tail_call_reachable,
919 bpf_prog->aux->func_idx != 0);
920 push_callee_regs(&prog, callee_regs_used);
924 memcpy(image + proglen, temp, ilen);
929 for (i = 1; i <= insn_cnt; i++, insn++) {
930 const s32 imm32 = insn->imm;
931 u32 dst_reg = insn->dst_reg;
932 u32 src_reg = insn->src_reg;
940 switch (insn->code) {
942 case BPF_ALU | BPF_ADD | BPF_X:
943 case BPF_ALU | BPF_SUB | BPF_X:
944 case BPF_ALU | BPF_AND | BPF_X:
945 case BPF_ALU | BPF_OR | BPF_X:
946 case BPF_ALU | BPF_XOR | BPF_X:
947 case BPF_ALU64 | BPF_ADD | BPF_X:
948 case BPF_ALU64 | BPF_SUB | BPF_X:
949 case BPF_ALU64 | BPF_AND | BPF_X:
950 case BPF_ALU64 | BPF_OR | BPF_X:
951 case BPF_ALU64 | BPF_XOR | BPF_X:
952 maybe_emit_mod(&prog, dst_reg, src_reg,
953 BPF_CLASS(insn->code) == BPF_ALU64);
954 b2 = simple_alu_opcodes[BPF_OP(insn->code)];
955 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
958 case BPF_ALU64 | BPF_MOV | BPF_X:
959 case BPF_ALU | BPF_MOV | BPF_X:
961 BPF_CLASS(insn->code) == BPF_ALU64,
966 case BPF_ALU | BPF_NEG:
967 case BPF_ALU64 | BPF_NEG:
968 maybe_emit_1mod(&prog, dst_reg,
969 BPF_CLASS(insn->code) == BPF_ALU64);
970 EMIT2(0xF7, add_1reg(0xD8, dst_reg));
973 case BPF_ALU | BPF_ADD | BPF_K:
974 case BPF_ALU | BPF_SUB | BPF_K:
975 case BPF_ALU | BPF_AND | BPF_K:
976 case BPF_ALU | BPF_OR | BPF_K:
977 case BPF_ALU | BPF_XOR | BPF_K:
978 case BPF_ALU64 | BPF_ADD | BPF_K:
979 case BPF_ALU64 | BPF_SUB | BPF_K:
980 case BPF_ALU64 | BPF_AND | BPF_K:
981 case BPF_ALU64 | BPF_OR | BPF_K:
982 case BPF_ALU64 | BPF_XOR | BPF_K:
983 maybe_emit_1mod(&prog, dst_reg,
984 BPF_CLASS(insn->code) == BPF_ALU64);
987 * b3 holds 'normal' opcode, b2 short form only valid
988 * in case dst is eax/rax.
990 switch (BPF_OP(insn->code)) {
1014 EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1015 else if (is_axreg(dst_reg))
1016 EMIT1_off32(b2, imm32);
1018 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1021 case BPF_ALU64 | BPF_MOV | BPF_K:
1022 case BPF_ALU | BPF_MOV | BPF_K:
1023 emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1027 case BPF_LD | BPF_IMM | BPF_DW:
1028 emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1033 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1034 case BPF_ALU | BPF_MOD | BPF_X:
1035 case BPF_ALU | BPF_DIV | BPF_X:
1036 case BPF_ALU | BPF_MOD | BPF_K:
1037 case BPF_ALU | BPF_DIV | BPF_K:
1038 case BPF_ALU64 | BPF_MOD | BPF_X:
1039 case BPF_ALU64 | BPF_DIV | BPF_X:
1040 case BPF_ALU64 | BPF_MOD | BPF_K:
1041 case BPF_ALU64 | BPF_DIV | BPF_K:
1042 EMIT1(0x50); /* push rax */
1043 EMIT1(0x52); /* push rdx */
1045 if (BPF_SRC(insn->code) == BPF_X)
1046 /* mov r11, src_reg */
1047 EMIT_mov(AUX_REG, src_reg);
1049 /* mov r11, imm32 */
1050 EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1052 /* mov rax, dst_reg */
1053 EMIT_mov(BPF_REG_0, dst_reg);
1057 * equivalent to 'xor rdx, rdx', but one byte less
1061 if (BPF_CLASS(insn->code) == BPF_ALU64)
1063 EMIT3(0x49, 0xF7, 0xF3);
1066 EMIT3(0x41, 0xF7, 0xF3);
1068 if (BPF_OP(insn->code) == BPF_MOD)
1070 EMIT3(0x49, 0x89, 0xD3);
1073 EMIT3(0x49, 0x89, 0xC3);
1075 EMIT1(0x5A); /* pop rdx */
1076 EMIT1(0x58); /* pop rax */
1078 /* mov dst_reg, r11 */
1079 EMIT_mov(dst_reg, AUX_REG);
1082 case BPF_ALU | BPF_MUL | BPF_K:
1083 case BPF_ALU | BPF_MUL | BPF_X:
1084 case BPF_ALU64 | BPF_MUL | BPF_K:
1085 case BPF_ALU64 | BPF_MUL | BPF_X:
1087 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1089 if (dst_reg != BPF_REG_0)
1090 EMIT1(0x50); /* push rax */
1091 if (dst_reg != BPF_REG_3)
1092 EMIT1(0x52); /* push rdx */
1094 /* mov r11, dst_reg */
1095 EMIT_mov(AUX_REG, dst_reg);
1097 if (BPF_SRC(insn->code) == BPF_X)
1098 emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
1100 emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
1103 EMIT1(add_1mod(0x48, AUX_REG));
1104 else if (is_ereg(AUX_REG))
1105 EMIT1(add_1mod(0x40, AUX_REG));
1107 EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
1109 if (dst_reg != BPF_REG_3)
1110 EMIT1(0x5A); /* pop rdx */
1111 if (dst_reg != BPF_REG_0) {
1112 /* mov dst_reg, rax */
1113 EMIT_mov(dst_reg, BPF_REG_0);
1114 EMIT1(0x58); /* pop rax */
1119 case BPF_ALU | BPF_LSH | BPF_K:
1120 case BPF_ALU | BPF_RSH | BPF_K:
1121 case BPF_ALU | BPF_ARSH | BPF_K:
1122 case BPF_ALU64 | BPF_LSH | BPF_K:
1123 case BPF_ALU64 | BPF_RSH | BPF_K:
1124 case BPF_ALU64 | BPF_ARSH | BPF_K:
1125 maybe_emit_1mod(&prog, dst_reg,
1126 BPF_CLASS(insn->code) == BPF_ALU64);
1128 b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1130 EMIT2(0xD1, add_1reg(b3, dst_reg));
1132 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1135 case BPF_ALU | BPF_LSH | BPF_X:
1136 case BPF_ALU | BPF_RSH | BPF_X:
1137 case BPF_ALU | BPF_ARSH | BPF_X:
1138 case BPF_ALU64 | BPF_LSH | BPF_X:
1139 case BPF_ALU64 | BPF_RSH | BPF_X:
1140 case BPF_ALU64 | BPF_ARSH | BPF_X:
1142 /* Check for bad case when dst_reg == rcx */
1143 if (dst_reg == BPF_REG_4) {
1144 /* mov r11, dst_reg */
1145 EMIT_mov(AUX_REG, dst_reg);
1149 if (src_reg != BPF_REG_4) { /* common case */
1150 EMIT1(0x51); /* push rcx */
1152 /* mov rcx, src_reg */
1153 EMIT_mov(BPF_REG_4, src_reg);
1156 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1157 maybe_emit_1mod(&prog, dst_reg,
1158 BPF_CLASS(insn->code) == BPF_ALU64);
1160 b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1161 EMIT2(0xD3, add_1reg(b3, dst_reg));
1163 if (src_reg != BPF_REG_4)
1164 EMIT1(0x59); /* pop rcx */
1166 if (insn->dst_reg == BPF_REG_4)
1167 /* mov dst_reg, r11 */
1168 EMIT_mov(insn->dst_reg, AUX_REG);
1171 case BPF_ALU | BPF_END | BPF_FROM_BE:
1174 /* Emit 'ror %ax, 8' to swap lower 2 bytes */
1176 if (is_ereg(dst_reg))
1178 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1180 /* Emit 'movzwl eax, ax' */
1181 if (is_ereg(dst_reg))
1182 EMIT3(0x45, 0x0F, 0xB7);
1185 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1188 /* Emit 'bswap eax' to swap lower 4 bytes */
1189 if (is_ereg(dst_reg))
1193 EMIT1(add_1reg(0xC8, dst_reg));
1196 /* Emit 'bswap rax' to swap 8 bytes */
1197 EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1198 add_1reg(0xC8, dst_reg));
1203 case BPF_ALU | BPF_END | BPF_FROM_LE:
1207 * Emit 'movzwl eax, ax' to zero extend 16-bit
1210 if (is_ereg(dst_reg))
1211 EMIT3(0x45, 0x0F, 0xB7);
1214 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1217 /* Emit 'mov eax, eax' to clear upper 32-bits */
1218 if (is_ereg(dst_reg))
1220 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1228 /* speculation barrier */
1229 case BPF_ST | BPF_NOSPEC:
1230 if (boot_cpu_has(X86_FEATURE_XMM2))
1232 EMIT3(0x0F, 0xAE, 0xE8);
1235 /* ST: *(u8*)(dst_reg + off) = imm */
1236 case BPF_ST | BPF_MEM | BPF_B:
1237 if (is_ereg(dst_reg))
1242 case BPF_ST | BPF_MEM | BPF_H:
1243 if (is_ereg(dst_reg))
1244 EMIT3(0x66, 0x41, 0xC7);
1248 case BPF_ST | BPF_MEM | BPF_W:
1249 if (is_ereg(dst_reg))
1254 case BPF_ST | BPF_MEM | BPF_DW:
1255 EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1257 st: if (is_imm8(insn->off))
1258 EMIT2(add_1reg(0x40, dst_reg), insn->off);
1260 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1262 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1265 /* STX: *(u8*)(dst_reg + off) = src_reg */
1266 case BPF_STX | BPF_MEM | BPF_B:
1267 case BPF_STX | BPF_MEM | BPF_H:
1268 case BPF_STX | BPF_MEM | BPF_W:
1269 case BPF_STX | BPF_MEM | BPF_DW:
1270 emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1273 /* LDX: dst_reg = *(u8*)(src_reg + off) */
1274 case BPF_LDX | BPF_MEM | BPF_B:
1275 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1276 case BPF_LDX | BPF_MEM | BPF_H:
1277 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1278 case BPF_LDX | BPF_MEM | BPF_W:
1279 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1280 case BPF_LDX | BPF_MEM | BPF_DW:
1281 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1282 if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1283 /* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1284 * add abs(insn->off) to the limit to make sure that negative
1285 * offset won't be an issue.
1286 * insn->off is s16, so it won't affect valid pointers.
1288 u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1289 u8 *end_of_jmp1, *end_of_jmp2;
1291 /* Conservatively check that src_reg + insn->off is a kernel address:
1292 * 1. src_reg + insn->off >= limit
1293 * 2. src_reg + insn->off doesn't become small positive.
1294 * Cannot do src_reg + insn->off >= limit in one branch,
1295 * since it needs two spare registers, but JIT has only one.
1298 /* movabsq r11, limit */
1299 EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1300 EMIT((u32)limit, 4);
1301 EMIT(limit >> 32, 4);
1302 /* cmp src_reg, r11 */
1303 maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1304 EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1305 /* if unsigned '<' goto end_of_jmp2 */
1309 /* mov r11, src_reg */
1310 emit_mov_reg(&prog, true, AUX_REG, src_reg);
1311 /* add r11, insn->off */
1312 maybe_emit_1mod(&prog, AUX_REG, true);
1313 EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1314 /* jmp if not carry to start_of_ldx
1315 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1316 * that has to be rejected.
1318 EMIT2(0x73 /* JNC */, 0);
1321 /* xor dst_reg, dst_reg */
1322 emit_mov_imm32(&prog, false, dst_reg, 0);
1323 /* jmp byte_after_ldx */
1326 /* populate jmp_offset for JB above to jump to xor dst_reg */
1327 end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1328 /* populate jmp_offset for JNC above to jump to start_of_ldx */
1329 start_of_ldx = prog;
1330 end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1332 emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1333 if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1334 struct exception_table_entry *ex;
1335 u8 *_insn = image + proglen + (start_of_ldx - temp);
1338 /* populate jmp_offset for JMP above */
1339 start_of_ldx[-1] = prog - start_of_ldx;
1341 if (!bpf_prog->aux->extable)
1344 if (excnt >= bpf_prog->aux->num_exentries) {
1345 pr_err("ex gen bug\n");
1348 ex = &bpf_prog->aux->extable[excnt++];
1350 delta = _insn - (u8 *)&ex->insn;
1351 if (!is_simm32(delta)) {
1352 pr_err("extable->insn doesn't fit into 32-bit\n");
1357 delta = (u8 *)ex_handler_bpf - (u8 *)&ex->handler;
1358 if (!is_simm32(delta)) {
1359 pr_err("extable->handler doesn't fit into 32-bit\n");
1362 ex->handler = delta;
1364 if (dst_reg > BPF_REG_9) {
1365 pr_err("verifier error\n");
1369 * Compute size of x86 insn and its target dest x86 register.
1370 * ex_handler_bpf() will use lower 8 bits to adjust
1371 * pt_regs->ip to jump over this x86 instruction
1372 * and upper bits to figure out which pt_regs to zero out.
1373 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1374 * of 4 bytes will be ignored and rbx will be zero inited.
1376 ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1380 case BPF_STX | BPF_ATOMIC | BPF_W:
1381 case BPF_STX | BPF_ATOMIC | BPF_DW:
1382 if (insn->imm == (BPF_AND | BPF_FETCH) ||
1383 insn->imm == (BPF_OR | BPF_FETCH) ||
1384 insn->imm == (BPF_XOR | BPF_FETCH)) {
1385 bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1386 u32 real_src_reg = src_reg;
1387 u32 real_dst_reg = dst_reg;
1391 * Can't be implemented with a single x86 insn.
1392 * Need to do a CMPXCHG loop.
1395 /* Will need RAX as a CMPXCHG operand so save R0 */
1396 emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1397 if (src_reg == BPF_REG_0)
1398 real_src_reg = BPF_REG_AX;
1399 if (dst_reg == BPF_REG_0)
1400 real_dst_reg = BPF_REG_AX;
1402 branch_target = prog;
1403 /* Load old value */
1404 emit_ldx(&prog, BPF_SIZE(insn->code),
1405 BPF_REG_0, real_dst_reg, insn->off);
1407 * Perform the (commutative) operation locally,
1408 * put the result in the AUX_REG.
1410 emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1411 maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1412 EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1413 add_2reg(0xC0, AUX_REG, real_src_reg));
1414 /* Attempt to swap in new value */
1415 err = emit_atomic(&prog, BPF_CMPXCHG,
1416 real_dst_reg, AUX_REG,
1418 BPF_SIZE(insn->code));
1422 * ZF tells us whether we won the race. If it's
1423 * cleared we need to try again.
1425 EMIT2(X86_JNE, -(prog - branch_target) - 2);
1426 /* Return the pre-modification value */
1427 emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1428 /* Restore R0 after clobbering RAX */
1429 emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1433 err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1434 insn->off, BPF_SIZE(insn->code));
1440 case BPF_JMP | BPF_CALL:
1441 func = (u8 *) __bpf_call_base + imm32;
1442 if (tail_call_reachable) {
1443 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1444 EMIT3_off32(0x48, 0x8B, 0x85,
1445 -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1446 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1449 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1454 case BPF_JMP | BPF_TAIL_CALL:
1456 emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1457 &prog, addrs[i], image,
1459 bpf_prog->aux->stack_depth);
1461 emit_bpf_tail_call_indirect(&prog,
1463 bpf_prog->aux->stack_depth);
1467 case BPF_JMP | BPF_JEQ | BPF_X:
1468 case BPF_JMP | BPF_JNE | BPF_X:
1469 case BPF_JMP | BPF_JGT | BPF_X:
1470 case BPF_JMP | BPF_JLT | BPF_X:
1471 case BPF_JMP | BPF_JGE | BPF_X:
1472 case BPF_JMP | BPF_JLE | BPF_X:
1473 case BPF_JMP | BPF_JSGT | BPF_X:
1474 case BPF_JMP | BPF_JSLT | BPF_X:
1475 case BPF_JMP | BPF_JSGE | BPF_X:
1476 case BPF_JMP | BPF_JSLE | BPF_X:
1477 case BPF_JMP32 | BPF_JEQ | BPF_X:
1478 case BPF_JMP32 | BPF_JNE | BPF_X:
1479 case BPF_JMP32 | BPF_JGT | BPF_X:
1480 case BPF_JMP32 | BPF_JLT | BPF_X:
1481 case BPF_JMP32 | BPF_JGE | BPF_X:
1482 case BPF_JMP32 | BPF_JLE | BPF_X:
1483 case BPF_JMP32 | BPF_JSGT | BPF_X:
1484 case BPF_JMP32 | BPF_JSLT | BPF_X:
1485 case BPF_JMP32 | BPF_JSGE | BPF_X:
1486 case BPF_JMP32 | BPF_JSLE | BPF_X:
1487 /* cmp dst_reg, src_reg */
1488 maybe_emit_mod(&prog, dst_reg, src_reg,
1489 BPF_CLASS(insn->code) == BPF_JMP);
1490 EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1493 case BPF_JMP | BPF_JSET | BPF_X:
1494 case BPF_JMP32 | BPF_JSET | BPF_X:
1495 /* test dst_reg, src_reg */
1496 maybe_emit_mod(&prog, dst_reg, src_reg,
1497 BPF_CLASS(insn->code) == BPF_JMP);
1498 EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1501 case BPF_JMP | BPF_JSET | BPF_K:
1502 case BPF_JMP32 | BPF_JSET | BPF_K:
1503 /* test dst_reg, imm32 */
1504 maybe_emit_1mod(&prog, dst_reg,
1505 BPF_CLASS(insn->code) == BPF_JMP);
1506 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1509 case BPF_JMP | BPF_JEQ | BPF_K:
1510 case BPF_JMP | BPF_JNE | BPF_K:
1511 case BPF_JMP | BPF_JGT | BPF_K:
1512 case BPF_JMP | BPF_JLT | BPF_K:
1513 case BPF_JMP | BPF_JGE | BPF_K:
1514 case BPF_JMP | BPF_JLE | BPF_K:
1515 case BPF_JMP | BPF_JSGT | BPF_K:
1516 case BPF_JMP | BPF_JSLT | BPF_K:
1517 case BPF_JMP | BPF_JSGE | BPF_K:
1518 case BPF_JMP | BPF_JSLE | BPF_K:
1519 case BPF_JMP32 | BPF_JEQ | BPF_K:
1520 case BPF_JMP32 | BPF_JNE | BPF_K:
1521 case BPF_JMP32 | BPF_JGT | BPF_K:
1522 case BPF_JMP32 | BPF_JLT | BPF_K:
1523 case BPF_JMP32 | BPF_JGE | BPF_K:
1524 case BPF_JMP32 | BPF_JLE | BPF_K:
1525 case BPF_JMP32 | BPF_JSGT | BPF_K:
1526 case BPF_JMP32 | BPF_JSLT | BPF_K:
1527 case BPF_JMP32 | BPF_JSGE | BPF_K:
1528 case BPF_JMP32 | BPF_JSLE | BPF_K:
1529 /* test dst_reg, dst_reg to save one extra byte */
1531 maybe_emit_mod(&prog, dst_reg, dst_reg,
1532 BPF_CLASS(insn->code) == BPF_JMP);
1533 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1537 /* cmp dst_reg, imm8/32 */
1538 maybe_emit_1mod(&prog, dst_reg,
1539 BPF_CLASS(insn->code) == BPF_JMP);
1542 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1544 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1546 emit_cond_jmp: /* Convert BPF opcode to x86 */
1547 switch (BPF_OP(insn->code)) {
1556 /* GT is unsigned '>', JA in x86 */
1560 /* LT is unsigned '<', JB in x86 */
1564 /* GE is unsigned '>=', JAE in x86 */
1568 /* LE is unsigned '<=', JBE in x86 */
1572 /* Signed '>', GT in x86 */
1576 /* Signed '<', LT in x86 */
1580 /* Signed '>=', GE in x86 */
1584 /* Signed '<=', LE in x86 */
1587 default: /* to silence GCC warning */
1590 jmp_offset = addrs[i + insn->off] - addrs[i];
1591 if (is_imm8(jmp_offset)) {
1593 /* To keep the jmp_offset valid, the extra bytes are
1594 * padded before the jump insn, so we subtract the
1595 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1597 * If the previous pass already emits an imm8
1598 * jmp_cond, then this BPF insn won't shrink, so
1601 * On the other hand, if the previous pass emits an
1602 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1603 * keep the image from shrinking further.
1605 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1606 * is 2 bytes, so the size difference is 4 bytes.
1608 nops = INSN_SZ_DIFF - 2;
1609 if (nops != 0 && nops != 4) {
1610 pr_err("unexpected jmp_cond padding: %d bytes\n",
1614 emit_nops(&prog, nops);
1616 EMIT2(jmp_cond, jmp_offset);
1617 } else if (is_simm32(jmp_offset)) {
1618 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1620 pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1626 case BPF_JMP | BPF_JA:
1627 if (insn->off == -1)
1628 /* -1 jmp instructions will always jump
1629 * backwards two bytes. Explicitly handling
1630 * this case avoids wasting too many passes
1631 * when there are long sequences of replaced
1636 jmp_offset = addrs[i + insn->off] - addrs[i];
1640 * If jmp_padding is enabled, the extra nops will
1641 * be inserted. Otherwise, optimize out nop jumps.
1644 /* There are 3 possible conditions.
1645 * (1) This BPF_JA is already optimized out in
1646 * the previous run, so there is no need
1647 * to pad any extra byte (0 byte).
1648 * (2) The previous pass emits an imm8 jmp,
1649 * so we pad 2 bytes to match the previous
1651 * (3) Similarly, the previous pass emits an
1652 * imm32 jmp, and 5 bytes is padded.
1654 nops = INSN_SZ_DIFF;
1655 if (nops != 0 && nops != 2 && nops != 5) {
1656 pr_err("unexpected nop jump padding: %d bytes\n",
1660 emit_nops(&prog, nops);
1665 if (is_imm8(jmp_offset)) {
1667 /* To avoid breaking jmp_offset, the extra bytes
1668 * are padded before the actual jmp insn, so
1669 * 2 bytes is subtracted from INSN_SZ_DIFF.
1671 * If the previous pass already emits an imm8
1672 * jmp, there is nothing to pad (0 byte).
1674 * If it emits an imm32 jmp (5 bytes) previously
1675 * and now an imm8 jmp (2 bytes), then we pad
1676 * (5 - 2 = 3) bytes to stop the image from
1677 * shrinking further.
1679 nops = INSN_SZ_DIFF - 2;
1680 if (nops != 0 && nops != 3) {
1681 pr_err("unexpected jump padding: %d bytes\n",
1685 emit_nops(&prog, INSN_SZ_DIFF - 2);
1687 EMIT2(0xEB, jmp_offset);
1688 } else if (is_simm32(jmp_offset)) {
1689 EMIT1_off32(0xE9, jmp_offset);
1691 pr_err("jmp gen bug %llx\n", jmp_offset);
1696 case BPF_JMP | BPF_EXIT:
1698 jmp_offset = ctx->cleanup_addr - addrs[i];
1702 /* Update cleanup_addr */
1703 ctx->cleanup_addr = proglen;
1704 pop_callee_regs(&prog, callee_regs_used);
1705 EMIT1(0xC9); /* leave */
1706 EMIT1(0xC3); /* ret */
1711 * By design x86-64 JIT should support all BPF instructions.
1712 * This error will be seen if new instruction was added
1713 * to the interpreter, but not to the JIT, or if there is
1716 pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1721 if (ilen > BPF_MAX_INSN_SIZE) {
1722 pr_err("bpf_jit: fatal insn size error\n");
1728 * When populating the image, assert that:
1730 * i) We do not write beyond the allocated space, and
1731 * ii) addrs[i] did not change from the prior run, in order
1732 * to validate assumptions made for computing branch
1735 if (unlikely(proglen + ilen > oldproglen ||
1736 proglen + ilen != addrs[i])) {
1737 pr_err("bpf_jit: fatal error\n");
1740 memcpy(image + proglen, temp, ilen);
1747 if (image && excnt != bpf_prog->aux->num_exentries) {
1748 pr_err("extable is not populated\n");
1754 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1758 /* Store function arguments to stack.
1759 * For a function that accepts two pointers the sequence will be:
1760 * mov QWORD PTR [rbp-0x10],rdi
1761 * mov QWORD PTR [rbp-0x8],rsi
1763 for (i = 0; i < min(nr_args, 6); i++)
1764 emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1766 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1767 -(stack_size - i * 8));
1770 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1775 /* Restore function arguments from stack.
1776 * For a function that accepts two pointers the sequence will be:
1777 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1778 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1780 for (i = 0; i < min(nr_args, 6); i++)
1781 emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1782 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1784 -(stack_size - i * 8));
1787 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1788 struct bpf_prog *p, int stack_size, bool save_ret)
1793 /* arg1: mov rdi, progs[i] */
1794 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1795 if (emit_call(&prog,
1796 p->aux->sleepable ? __bpf_prog_enter_sleepable :
1797 __bpf_prog_enter, prog))
1799 /* remember prog start time returned by __bpf_prog_enter */
1800 emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1802 /* if (__bpf_prog_enter*(prog) == 0)
1803 * goto skip_exec_of_prog;
1805 EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */
1806 /* emit 2 nops that will be replaced with JE insn */
1808 emit_nops(&prog, 2);
1810 /* arg1: lea rdi, [rbp - stack_size] */
1811 EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1812 /* arg2: progs[i]->insnsi for interpreter */
1814 emit_mov_imm64(&prog, BPF_REG_2,
1815 (long) p->insnsi >> 32,
1816 (u32) (long) p->insnsi);
1817 /* call JITed bpf program or interpreter */
1818 if (emit_call(&prog, p->bpf_func, prog))
1822 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1823 * of the previous call which is then passed on the stack to
1824 * the next BPF program.
1826 * BPF_TRAMP_FENTRY trampoline may need to return the return
1827 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1830 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1832 /* replace 2 nops with JE insn, since jmp target is known */
1833 jmp_insn[0] = X86_JE;
1834 jmp_insn[1] = prog - jmp_insn - 2;
1836 /* arg1: mov rdi, progs[i] */
1837 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1838 /* arg2: mov rsi, rbx <- start time in nsec */
1839 emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1840 if (emit_call(&prog,
1841 p->aux->sleepable ? __bpf_prog_exit_sleepable :
1842 __bpf_prog_exit, prog))
1849 static void emit_align(u8 **pprog, u32 align)
1851 u8 *target, *prog = *pprog;
1853 target = PTR_ALIGN(prog, align);
1855 emit_nops(&prog, target - prog);
1860 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1865 offset = func - (ip + 2 + 4);
1866 if (!is_simm32(offset)) {
1867 pr_err("Target %p is out of range\n", func);
1870 EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1875 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1876 struct bpf_tramp_progs *tp, int stack_size,
1882 for (i = 0; i < tp->nr_progs; i++) {
1883 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size,
1891 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1892 struct bpf_tramp_progs *tp, int stack_size,
1898 /* The first fmod_ret program will receive a garbage return value.
1899 * Set this to 0 to avoid confusing the program.
1901 emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1902 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1903 for (i = 0; i < tp->nr_progs; i++) {
1904 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1907 /* mod_ret prog stored return value into [rbp - 8]. Emit:
1908 * if (*(u64 *)(rbp - 8) != 0)
1911 /* cmp QWORD PTR [rbp - 0x8], 0x0 */
1912 EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1914 /* Save the location of the branch and Generate 6 nops
1915 * (4 bytes for an offset and 2 bytes for the jump) These nops
1916 * are replaced with a conditional jump once do_fexit (i.e. the
1917 * start of the fexit invocation) is finalized.
1920 emit_nops(&prog, 4 + 2);
1927 static bool is_valid_bpf_tramp_flags(unsigned int flags)
1929 if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1930 (flags & BPF_TRAMP_F_SKIP_FRAME))
1934 * BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
1935 * and it must be used alone.
1937 if ((flags & BPF_TRAMP_F_RET_FENTRY_RET) &&
1938 (flags & ~BPF_TRAMP_F_RET_FENTRY_RET))
1945 * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1946 * its 'struct btf_func_model' will be nr_args=2
1947 * The assembly code when eth_type_trans is executing after trampoline:
1951 * sub rsp, 16 // space for skb and dev
1952 * push rbx // temp regs to pass start time
1953 * mov qword ptr [rbp - 16], rdi // save skb pointer to stack
1954 * mov qword ptr [rbp - 8], rsi // save dev pointer to stack
1955 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1956 * mov rbx, rax // remember start time in bpf stats are enabled
1957 * lea rdi, [rbp - 16] // R1==ctx of bpf prog
1958 * call addr_of_jited_FENTRY_prog
1959 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1960 * mov rsi, rbx // prog start time
1961 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1962 * mov rdi, qword ptr [rbp - 16] // restore skb pointer from stack
1963 * mov rsi, qword ptr [rbp - 8] // restore dev pointer from stack
1968 * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1969 * replaced with 'call generated_bpf_trampoline'. When it returns
1970 * eth_type_trans will continue executing with original skb and dev pointers.
1972 * The assembly code when eth_type_trans is called from trampoline:
1976 * sub rsp, 24 // space for skb, dev, return value
1977 * push rbx // temp regs to pass start time
1978 * mov qword ptr [rbp - 24], rdi // save skb pointer to stack
1979 * mov qword ptr [rbp - 16], rsi // save dev pointer to stack
1980 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1981 * mov rbx, rax // remember start time if bpf stats are enabled
1982 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1983 * call addr_of_jited_FENTRY_prog // bpf prog can access skb and dev
1984 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1985 * mov rsi, rbx // prog start time
1986 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1987 * mov rdi, qword ptr [rbp - 24] // restore skb pointer from stack
1988 * mov rsi, qword ptr [rbp - 16] // restore dev pointer from stack
1989 * call eth_type_trans+5 // execute body of eth_type_trans
1990 * mov qword ptr [rbp - 8], rax // save return value
1991 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1992 * mov rbx, rax // remember start time in bpf stats are enabled
1993 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1994 * call addr_of_jited_FEXIT_prog // bpf prog can access skb, dev, return value
1995 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1996 * mov rsi, rbx // prog start time
1997 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1998 * mov rax, qword ptr [rbp - 8] // restore eth_type_trans's return value
2001 * add rsp, 8 // skip eth_type_trans's frame
2002 * ret // return to its caller
2004 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2005 const struct btf_func_model *m, u32 flags,
2006 struct bpf_tramp_progs *tprogs,
2009 int ret, i, nr_args = m->nr_args;
2010 int stack_size = nr_args * 8;
2011 struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
2012 struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
2013 struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
2014 u8 **branches = NULL;
2018 /* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2022 if (!is_valid_bpf_tramp_flags(flags))
2025 /* room for return value of orig_call or fentry prog */
2026 save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2030 if (flags & BPF_TRAMP_F_IP_ARG)
2031 stack_size += 8; /* room for IP address argument */
2033 if (flags & BPF_TRAMP_F_SKIP_FRAME)
2034 /* skip patched call instruction and point orig_call to actual
2035 * body of the kernel function.
2037 orig_call += X86_PATCH_SIZE;
2041 EMIT1(0x55); /* push rbp */
2042 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2043 EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2044 EMIT1(0x53); /* push rbx */
2046 if (flags & BPF_TRAMP_F_IP_ARG) {
2047 /* Store IP address of the traced function:
2048 * mov rax, QWORD PTR [rbp + 8]
2049 * sub rax, X86_PATCH_SIZE
2050 * mov QWORD PTR [rbp - stack_size], rax
2052 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2053 EMIT4(0x48, 0x83, 0xe8, X86_PATCH_SIZE);
2054 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -stack_size);
2056 /* Continue with stack_size for regs storage, stack will
2057 * be correctly restored with 'leave' instruction.
2062 save_regs(m, &prog, nr_args, stack_size);
2064 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2065 /* arg1: mov rdi, im */
2066 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2067 if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2073 if (fentry->nr_progs)
2074 if (invoke_bpf(m, &prog, fentry, stack_size,
2075 flags & BPF_TRAMP_F_RET_FENTRY_RET))
2078 if (fmod_ret->nr_progs) {
2079 branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
2084 if (invoke_bpf_mod_ret(m, &prog, fmod_ret, stack_size,
2091 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2092 restore_regs(m, &prog, nr_args, stack_size);
2094 /* call original function */
2095 if (emit_call(&prog, orig_call, prog)) {
2099 /* remember return value in a stack for bpf prog to access */
2100 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2101 im->ip_after_call = prog;
2102 memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2103 prog += X86_PATCH_SIZE;
2106 if (fmod_ret->nr_progs) {
2107 /* From Intel 64 and IA-32 Architectures Optimization
2108 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2109 * Coding Rule 11: All branch targets should be 16-byte
2112 emit_align(&prog, 16);
2113 /* Update the branches saved in invoke_bpf_mod_ret with the
2114 * aligned address of do_fexit.
2116 for (i = 0; i < fmod_ret->nr_progs; i++)
2117 emit_cond_near_jump(&branches[i], prog, branches[i],
2121 if (fexit->nr_progs)
2122 if (invoke_bpf(m, &prog, fexit, stack_size, false)) {
2127 if (flags & BPF_TRAMP_F_RESTORE_REGS)
2128 restore_regs(m, &prog, nr_args, stack_size);
2130 /* This needs to be done regardless. If there were fmod_ret programs,
2131 * the return value is only updated on the stack and still needs to be
2134 if (flags & BPF_TRAMP_F_CALL_ORIG) {
2135 im->ip_epilogue = prog;
2136 /* arg1: mov rdi, im */
2137 emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2138 if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2143 /* restore return value of orig_call or fentry prog back into RAX */
2145 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2147 EMIT1(0x5B); /* pop rbx */
2148 EMIT1(0xC9); /* leave */
2149 if (flags & BPF_TRAMP_F_SKIP_FRAME)
2150 /* skip our return address and return to parent */
2151 EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2152 EMIT1(0xC3); /* ret */
2153 /* Make sure the trampoline generation logic doesn't overflow */
2154 if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2158 ret = prog - (u8 *)image;
2165 static int emit_fallback_jump(u8 **pprog)
2170 #ifdef CONFIG_RETPOLINE
2171 /* Note that this assumes the the compiler uses external
2172 * thunks for indirect calls. Both clang and GCC use the same
2173 * naming convention for external thunks.
2175 err = emit_jump(&prog, __x86_indirect_thunk_rdx, prog);
2177 EMIT2(0xFF, 0xE2); /* jmp rdx */
2183 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2185 u8 *jg_reloc, *prog = *pprog;
2186 int pivot, err, jg_bytes = 1;
2190 /* Leaf node of recursion, i.e. not a range of indices
2193 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
2194 if (!is_simm32(progs[a]))
2196 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2198 err = emit_cond_near_jump(&prog, /* je func */
2199 (void *)progs[a], prog,
2204 err = emit_fallback_jump(&prog); /* jmp thunk/indirect */
2212 /* Not a leaf node, so we pivot, and recursively descend into
2213 * the lower and upper ranges.
2215 pivot = (b - a) / 2;
2216 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
2217 if (!is_simm32(progs[a + pivot]))
2219 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2221 if (pivot > 2) { /* jg upper_part */
2222 /* Require near jump. */
2224 EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2230 err = emit_bpf_dispatcher(&prog, a, a + pivot, /* emit lower_part */
2235 /* From Intel 64 and IA-32 Architectures Optimization
2236 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2237 * Coding Rule 11: All branch targets should be 16-byte
2240 emit_align(&prog, 16);
2241 jg_offset = prog - jg_reloc;
2242 emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2244 err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */
2253 static int cmp_ips(const void *a, const void *b)
2265 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2269 sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2270 return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2273 struct x64_jit_data {
2274 struct bpf_binary_header *header;
2278 struct jit_context ctx;
2281 #define MAX_PASSES 20
2282 #define PADDING_PASSES (MAX_PASSES - 5)
2284 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2286 struct bpf_binary_header *header = NULL;
2287 struct bpf_prog *tmp, *orig_prog = prog;
2288 struct x64_jit_data *jit_data;
2289 int proglen, oldproglen = 0;
2290 struct jit_context ctx = {};
2291 bool tmp_blinded = false;
2292 bool extra_pass = false;
2293 bool padding = false;
2299 if (!prog->jit_requested)
2302 tmp = bpf_jit_blind_constants(prog);
2304 * If blinding was requested and we failed during blinding,
2305 * we must fall back to the interpreter.
2314 jit_data = prog->aux->jit_data;
2316 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2321 prog->aux->jit_data = jit_data;
2323 addrs = jit_data->addrs;
2325 ctx = jit_data->ctx;
2326 oldproglen = jit_data->proglen;
2327 image = jit_data->image;
2328 header = jit_data->header;
2331 goto skip_init_addrs;
2333 addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2340 * Before first pass, make a rough estimation of addrs[]
2341 * each BPF instruction is translated to less than 64 bytes
2343 for (proglen = 0, i = 0; i <= prog->len; i++) {
2347 ctx.cleanup_addr = proglen;
2351 * JITed image shrinks with every pass and the loop iterates
2352 * until the image stops shrinking. Very large BPF programs
2353 * may converge on the last pass. In such case do one more
2354 * pass to emit the final image.
2356 for (pass = 0; pass < MAX_PASSES || image; pass++) {
2357 if (!padding && pass >= PADDING_PASSES)
2359 proglen = do_jit(prog, addrs, image, oldproglen, &ctx, padding);
2364 bpf_jit_binary_free(header);
2369 if (proglen != oldproglen) {
2370 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2371 proglen, oldproglen);
2376 if (proglen == oldproglen) {
2378 * The number of entries in extable is the number of BPF_LDX
2379 * insns that access kernel memory via "pointer to BTF type".
2380 * The verifier changed their opcode from LDX|MEM|size
2381 * to LDX|PROBE_MEM|size to make JITing easier.
2383 u32 align = __alignof__(struct exception_table_entry);
2384 u32 extable_size = prog->aux->num_exentries *
2385 sizeof(struct exception_table_entry);
2387 /* allocate module memory for x86 insns and extable */
2388 header = bpf_jit_binary_alloc(roundup(proglen, align) + extable_size,
2389 &image, align, jit_fill_hole);
2394 prog->aux->extable = (void *) image + roundup(proglen, align);
2396 oldproglen = proglen;
2400 if (bpf_jit_enable > 1)
2401 bpf_jit_dump(prog->len, proglen, pass + 1, image);
2404 if (!prog->is_func || extra_pass) {
2405 bpf_tail_call_direct_fixup(prog);
2406 bpf_jit_binary_lock_ro(header);
2408 jit_data->addrs = addrs;
2409 jit_data->ctx = ctx;
2410 jit_data->proglen = proglen;
2411 jit_data->image = image;
2412 jit_data->header = header;
2414 prog->bpf_func = (void *)image;
2416 prog->jited_len = proglen;
2421 if (!image || !prog->is_func || extra_pass) {
2423 bpf_prog_fill_jited_linfo(prog, addrs + 1);
2427 prog->aux->jit_data = NULL;
2431 bpf_jit_prog_release_other(prog, prog == orig_prog ?
2436 bool bpf_jit_supports_kfunc_call(void)