xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision ac73d4bf)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * BPF JIT compiler
4  *
5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6  * 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 <linux/init.h>
15 #include <asm/extable.h>
16 #include <asm/set_memory.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/text-patching.h>
19 
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
21 {
22 	if (len == 1)
23 		*ptr = bytes;
24 	else if (len == 2)
25 		*(u16 *)ptr = bytes;
26 	else {
27 		*(u32 *)ptr = bytes;
28 		barrier();
29 	}
30 	return ptr + len;
31 }
32 
33 #define EMIT(bytes, len) \
34 	do { prog = emit_code(prog, bytes, len); } while (0)
35 
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)
40 
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)
49 
50 #ifdef CONFIG_X86_KERNEL_IBT
51 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
52 #else
53 #define EMIT_ENDBR()
54 #endif
55 
56 static bool is_imm8(int value)
57 {
58 	return value <= 127 && value >= -128;
59 }
60 
61 static bool is_simm32(s64 value)
62 {
63 	return value == (s64)(s32)value;
64 }
65 
66 static bool is_uimm32(u64 value)
67 {
68 	return value == (u64)(u32)value;
69 }
70 
71 /* mov dst, src */
72 #define EMIT_mov(DST, SRC)								 \
73 	do {										 \
74 		if (DST != SRC)								 \
75 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
76 	} while (0)
77 
78 static int bpf_size_to_x86_bytes(int bpf_size)
79 {
80 	if (bpf_size == BPF_W)
81 		return 4;
82 	else if (bpf_size == BPF_H)
83 		return 2;
84 	else if (bpf_size == BPF_B)
85 		return 1;
86 	else if (bpf_size == BPF_DW)
87 		return 4; /* imm32 */
88 	else
89 		return 0;
90 }
91 
92 /*
93  * List of x86 cond jumps opcodes (. + s8)
94  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
95  */
96 #define X86_JB  0x72
97 #define X86_JAE 0x73
98 #define X86_JE  0x74
99 #define X86_JNE 0x75
100 #define X86_JBE 0x76
101 #define X86_JA  0x77
102 #define X86_JL  0x7C
103 #define X86_JGE 0x7D
104 #define X86_JLE 0x7E
105 #define X86_JG  0x7F
106 
107 /* Pick a register outside of BPF range for JIT internal work */
108 #define AUX_REG (MAX_BPF_JIT_REG + 1)
109 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
110 
111 /*
112  * The following table maps BPF registers to x86-64 registers.
113  *
114  * x86-64 register R12 is unused, since if used as base address
115  * register in load/store instructions, it always needs an
116  * extra byte of encoding and is callee saved.
117  *
118  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
119  * trampoline. x86-64 register R10 is used for blinding (if enabled).
120  */
121 static const int reg2hex[] = {
122 	[BPF_REG_0] = 0,  /* RAX */
123 	[BPF_REG_1] = 7,  /* RDI */
124 	[BPF_REG_2] = 6,  /* RSI */
125 	[BPF_REG_3] = 2,  /* RDX */
126 	[BPF_REG_4] = 1,  /* RCX */
127 	[BPF_REG_5] = 0,  /* R8  */
128 	[BPF_REG_6] = 3,  /* RBX callee saved */
129 	[BPF_REG_7] = 5,  /* R13 callee saved */
130 	[BPF_REG_8] = 6,  /* R14 callee saved */
131 	[BPF_REG_9] = 7,  /* R15 callee saved */
132 	[BPF_REG_FP] = 5, /* RBP readonly */
133 	[BPF_REG_AX] = 2, /* R10 temp register */
134 	[AUX_REG] = 3,    /* R11 temp register */
135 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
136 };
137 
138 static const int reg2pt_regs[] = {
139 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
140 	[BPF_REG_1] = offsetof(struct pt_regs, di),
141 	[BPF_REG_2] = offsetof(struct pt_regs, si),
142 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
143 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
144 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
145 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
146 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
147 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
148 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
149 };
150 
151 /*
152  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
153  * which need extra byte of encoding.
154  * rax,rcx,...,rbp have simpler encoding
155  */
156 static bool is_ereg(u32 reg)
157 {
158 	return (1 << reg) & (BIT(BPF_REG_5) |
159 			     BIT(AUX_REG) |
160 			     BIT(BPF_REG_7) |
161 			     BIT(BPF_REG_8) |
162 			     BIT(BPF_REG_9) |
163 			     BIT(X86_REG_R9) |
164 			     BIT(BPF_REG_AX));
165 }
166 
167 /*
168  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
169  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
170  * of encoding. al,cl,dl,bl have simpler encoding.
171  */
172 static bool is_ereg_8l(u32 reg)
173 {
174 	return is_ereg(reg) ||
175 	    (1 << reg) & (BIT(BPF_REG_1) |
176 			  BIT(BPF_REG_2) |
177 			  BIT(BPF_REG_FP));
178 }
179 
180 static bool is_axreg(u32 reg)
181 {
182 	return reg == BPF_REG_0;
183 }
184 
185 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
186 static u8 add_1mod(u8 byte, u32 reg)
187 {
188 	if (is_ereg(reg))
189 		byte |= 1;
190 	return byte;
191 }
192 
193 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
194 {
195 	if (is_ereg(r1))
196 		byte |= 1;
197 	if (is_ereg(r2))
198 		byte |= 4;
199 	return byte;
200 }
201 
202 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
203 static u8 add_1reg(u8 byte, u32 dst_reg)
204 {
205 	return byte + reg2hex[dst_reg];
206 }
207 
208 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
209 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
210 {
211 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
212 }
213 
214 /* Some 1-byte opcodes for binary ALU operations */
215 static u8 simple_alu_opcodes[] = {
216 	[BPF_ADD] = 0x01,
217 	[BPF_SUB] = 0x29,
218 	[BPF_AND] = 0x21,
219 	[BPF_OR] = 0x09,
220 	[BPF_XOR] = 0x31,
221 	[BPF_LSH] = 0xE0,
222 	[BPF_RSH] = 0xE8,
223 	[BPF_ARSH] = 0xF8,
224 };
225 
226 static void jit_fill_hole(void *area, unsigned int size)
227 {
228 	/* Fill whole space with INT3 instructions */
229 	memset(area, 0xcc, size);
230 }
231 
232 int bpf_arch_text_invalidate(void *dst, size_t len)
233 {
234 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
235 }
236 
237 struct jit_context {
238 	int cleanup_addr; /* Epilogue code offset */
239 
240 	/*
241 	 * Program specific offsets of labels in the code; these rely on the
242 	 * JIT doing at least 2 passes, recording the position on the first
243 	 * pass, only to generate the correct offset on the second pass.
244 	 */
245 	int tail_call_direct_label;
246 	int tail_call_indirect_label;
247 };
248 
249 /* Maximum number of bytes emitted while JITing one eBPF insn */
250 #define BPF_MAX_INSN_SIZE	128
251 #define BPF_INSN_SAFETY		64
252 
253 /* Number of bytes emit_patch() needs to generate instructions */
254 #define X86_PATCH_SIZE		5
255 /* Number of bytes that will be skipped on tailcall */
256 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
257 
258 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
259 {
260 	u8 *prog = *pprog;
261 
262 	if (callee_regs_used[0])
263 		EMIT1(0x53);         /* push rbx */
264 	if (callee_regs_used[1])
265 		EMIT2(0x41, 0x55);   /* push r13 */
266 	if (callee_regs_used[2])
267 		EMIT2(0x41, 0x56);   /* push r14 */
268 	if (callee_regs_used[3])
269 		EMIT2(0x41, 0x57);   /* push r15 */
270 	*pprog = prog;
271 }
272 
273 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
274 {
275 	u8 *prog = *pprog;
276 
277 	if (callee_regs_used[3])
278 		EMIT2(0x41, 0x5F);   /* pop r15 */
279 	if (callee_regs_used[2])
280 		EMIT2(0x41, 0x5E);   /* pop r14 */
281 	if (callee_regs_used[1])
282 		EMIT2(0x41, 0x5D);   /* pop r13 */
283 	if (callee_regs_used[0])
284 		EMIT1(0x5B);         /* pop rbx */
285 	*pprog = prog;
286 }
287 
288 /*
289  * Emit x86-64 prologue code for BPF program.
290  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
291  * while jumping to another program
292  */
293 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
294 			  bool tail_call_reachable, bool is_subprog)
295 {
296 	u8 *prog = *pprog;
297 
298 	/* BPF trampoline can be made to work without these nops,
299 	 * but let's waste 5 bytes for now and optimize later
300 	 */
301 	EMIT_ENDBR();
302 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
303 	prog += X86_PATCH_SIZE;
304 	if (!ebpf_from_cbpf) {
305 		if (tail_call_reachable && !is_subprog)
306 			EMIT2(0x31, 0xC0); /* xor eax, eax */
307 		else
308 			EMIT2(0x66, 0x90); /* nop2 */
309 	}
310 	EMIT1(0x55);             /* push rbp */
311 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
312 
313 	/* X86_TAIL_CALL_OFFSET is here */
314 	EMIT_ENDBR();
315 
316 	/* sub rsp, rounded_stack_depth */
317 	if (stack_depth)
318 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
319 	if (tail_call_reachable)
320 		EMIT1(0x50);         /* push rax */
321 	*pprog = prog;
322 }
323 
324 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
325 {
326 	u8 *prog = *pprog;
327 	s64 offset;
328 
329 	offset = func - (ip + X86_PATCH_SIZE);
330 	if (!is_simm32(offset)) {
331 		pr_err("Target call %p is out of range\n", func);
332 		return -ERANGE;
333 	}
334 	EMIT1_off32(opcode, offset);
335 	*pprog = prog;
336 	return 0;
337 }
338 
339 static int emit_call(u8 **pprog, void *func, void *ip)
340 {
341 	return emit_patch(pprog, func, ip, 0xE8);
342 }
343 
344 static int emit_jump(u8 **pprog, void *func, void *ip)
345 {
346 	return emit_patch(pprog, func, ip, 0xE9);
347 }
348 
349 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
350 				void *old_addr, void *new_addr)
351 {
352 	const u8 *nop_insn = x86_nops[5];
353 	u8 old_insn[X86_PATCH_SIZE];
354 	u8 new_insn[X86_PATCH_SIZE];
355 	u8 *prog;
356 	int ret;
357 
358 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
359 	if (old_addr) {
360 		prog = old_insn;
361 		ret = t == BPF_MOD_CALL ?
362 		      emit_call(&prog, old_addr, ip) :
363 		      emit_jump(&prog, old_addr, ip);
364 		if (ret)
365 			return ret;
366 	}
367 
368 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
369 	if (new_addr) {
370 		prog = new_insn;
371 		ret = t == BPF_MOD_CALL ?
372 		      emit_call(&prog, new_addr, ip) :
373 		      emit_jump(&prog, new_addr, ip);
374 		if (ret)
375 			return ret;
376 	}
377 
378 	ret = -EBUSY;
379 	mutex_lock(&text_mutex);
380 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
381 		goto out;
382 	ret = 1;
383 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
384 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
385 		ret = 0;
386 	}
387 out:
388 	mutex_unlock(&text_mutex);
389 	return ret;
390 }
391 
392 int __init bpf_arch_init_dispatcher_early(void *ip)
393 {
394 	const u8 *nop_insn = x86_nops[5];
395 
396 	if (is_endbr(*(u32 *)ip))
397 		ip += ENDBR_INSN_SIZE;
398 
399 	if (memcmp(ip, nop_insn, X86_PATCH_SIZE))
400 		text_poke_early(ip, nop_insn, X86_PATCH_SIZE);
401 	return 0;
402 }
403 
404 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
405 		       void *old_addr, void *new_addr)
406 {
407 	if (!is_kernel_text((long)ip) &&
408 	    !is_bpf_text_address((long)ip))
409 		/* BPF poking in modules is not supported */
410 		return -EINVAL;
411 
412 	/*
413 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
414 	 * with an ENDBR instruction.
415 	 */
416 	if (is_endbr(*(u32 *)ip))
417 		ip += ENDBR_INSN_SIZE;
418 
419 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
420 }
421 
422 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
423 
424 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
425 {
426 	u8 *prog = *pprog;
427 
428 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
429 		EMIT_LFENCE();
430 		EMIT2(0xFF, 0xE0 + reg);
431 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
432 		OPTIMIZER_HIDE_VAR(reg);
433 		emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
434 	} else {
435 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
436 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
437 			EMIT1(0xCC);		/* int3 */
438 	}
439 
440 	*pprog = prog;
441 }
442 
443 static void emit_return(u8 **pprog, u8 *ip)
444 {
445 	u8 *prog = *pprog;
446 
447 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
448 		emit_jump(&prog, &__x86_return_thunk, ip);
449 	} else {
450 		EMIT1(0xC3);		/* ret */
451 		if (IS_ENABLED(CONFIG_SLS))
452 			EMIT1(0xCC);	/* int3 */
453 	}
454 
455 	*pprog = prog;
456 }
457 
458 /*
459  * Generate the following code:
460  *
461  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
462  *   if (index >= array->map.max_entries)
463  *     goto out;
464  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
465  *     goto out;
466  *   prog = array->ptrs[index];
467  *   if (prog == NULL)
468  *     goto out;
469  *   goto *(prog->bpf_func + prologue_size);
470  * out:
471  */
472 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
473 					u32 stack_depth, u8 *ip,
474 					struct jit_context *ctx)
475 {
476 	int tcc_off = -4 - round_up(stack_depth, 8);
477 	u8 *prog = *pprog, *start = *pprog;
478 	int offset;
479 
480 	/*
481 	 * rdi - pointer to ctx
482 	 * rsi - pointer to bpf_array
483 	 * rdx - index in bpf_array
484 	 */
485 
486 	/*
487 	 * if (index >= array->map.max_entries)
488 	 *	goto out;
489 	 */
490 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
491 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
492 	      offsetof(struct bpf_array, map.max_entries));
493 
494 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
495 	EMIT2(X86_JBE, offset);                   /* jbe out */
496 
497 	/*
498 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
499 	 *	goto out;
500 	 */
501 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
502 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
503 
504 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
505 	EMIT2(X86_JAE, offset);                   /* jae out */
506 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
507 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
508 
509 	/* prog = array->ptrs[index]; */
510 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
511 		    offsetof(struct bpf_array, ptrs));
512 
513 	/*
514 	 * if (prog == NULL)
515 	 *	goto out;
516 	 */
517 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
518 
519 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
520 	EMIT2(X86_JE, offset);                    /* je out */
521 
522 	pop_callee_regs(&prog, callee_regs_used);
523 
524 	EMIT1(0x58);                              /* pop rax */
525 	if (stack_depth)
526 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
527 			    round_up(stack_depth, 8));
528 
529 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
530 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
531 	      offsetof(struct bpf_prog, bpf_func));
532 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
533 	      X86_TAIL_CALL_OFFSET);
534 	/*
535 	 * Now we're ready to jump into next BPF program
536 	 * rdi == ctx (1st arg)
537 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
538 	 */
539 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
540 
541 	/* out: */
542 	ctx->tail_call_indirect_label = prog - start;
543 	*pprog = prog;
544 }
545 
546 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
547 				      u8 **pprog, u8 *ip,
548 				      bool *callee_regs_used, u32 stack_depth,
549 				      struct jit_context *ctx)
550 {
551 	int tcc_off = -4 - round_up(stack_depth, 8);
552 	u8 *prog = *pprog, *start = *pprog;
553 	int offset;
554 
555 	/*
556 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
557 	 *	goto out;
558 	 */
559 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
560 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
561 
562 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
563 	EMIT2(X86_JAE, offset);                       /* jae out */
564 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
565 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
566 
567 	poke->tailcall_bypass = ip + (prog - start);
568 	poke->adj_off = X86_TAIL_CALL_OFFSET;
569 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
570 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
571 
572 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
573 		  poke->tailcall_bypass);
574 
575 	pop_callee_regs(&prog, callee_regs_used);
576 	EMIT1(0x58);                                  /* pop rax */
577 	if (stack_depth)
578 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
579 
580 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
581 	prog += X86_PATCH_SIZE;
582 
583 	/* out: */
584 	ctx->tail_call_direct_label = prog - start;
585 
586 	*pprog = prog;
587 }
588 
589 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
590 {
591 	struct bpf_jit_poke_descriptor *poke;
592 	struct bpf_array *array;
593 	struct bpf_prog *target;
594 	int i, ret;
595 
596 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
597 		poke = &prog->aux->poke_tab[i];
598 		if (poke->aux && poke->aux != prog->aux)
599 			continue;
600 
601 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
602 
603 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
604 			continue;
605 
606 		array = container_of(poke->tail_call.map, struct bpf_array, map);
607 		mutex_lock(&array->aux->poke_mutex);
608 		target = array->ptrs[poke->tail_call.key];
609 		if (target) {
610 			ret = __bpf_arch_text_poke(poke->tailcall_target,
611 						   BPF_MOD_JUMP, NULL,
612 						   (u8 *)target->bpf_func +
613 						   poke->adj_off);
614 			BUG_ON(ret < 0);
615 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
616 						   BPF_MOD_JUMP,
617 						   (u8 *)poke->tailcall_target +
618 						   X86_PATCH_SIZE, NULL);
619 			BUG_ON(ret < 0);
620 		}
621 		WRITE_ONCE(poke->tailcall_target_stable, true);
622 		mutex_unlock(&array->aux->poke_mutex);
623 	}
624 }
625 
626 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
627 			   u32 dst_reg, const u32 imm32)
628 {
629 	u8 *prog = *pprog;
630 	u8 b1, b2, b3;
631 
632 	/*
633 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
634 	 * (which zero-extends imm32) to save 2 bytes.
635 	 */
636 	if (sign_propagate && (s32)imm32 < 0) {
637 		/* 'mov %rax, imm32' sign extends imm32 */
638 		b1 = add_1mod(0x48, dst_reg);
639 		b2 = 0xC7;
640 		b3 = 0xC0;
641 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
642 		goto done;
643 	}
644 
645 	/*
646 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
647 	 * to save 3 bytes.
648 	 */
649 	if (imm32 == 0) {
650 		if (is_ereg(dst_reg))
651 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
652 		b2 = 0x31; /* xor */
653 		b3 = 0xC0;
654 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
655 		goto done;
656 	}
657 
658 	/* mov %eax, imm32 */
659 	if (is_ereg(dst_reg))
660 		EMIT1(add_1mod(0x40, dst_reg));
661 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
662 done:
663 	*pprog = prog;
664 }
665 
666 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
667 			   const u32 imm32_hi, const u32 imm32_lo)
668 {
669 	u8 *prog = *pprog;
670 
671 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
672 		/*
673 		 * For emitting plain u32, where sign bit must not be
674 		 * propagated LLVM tends to load imm64 over mov32
675 		 * directly, so save couple of bytes by just doing
676 		 * 'mov %eax, imm32' instead.
677 		 */
678 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
679 	} else {
680 		/* movabsq rax, imm64 */
681 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
682 		EMIT(imm32_lo, 4);
683 		EMIT(imm32_hi, 4);
684 	}
685 
686 	*pprog = prog;
687 }
688 
689 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
690 {
691 	u8 *prog = *pprog;
692 
693 	if (is64) {
694 		/* mov dst, src */
695 		EMIT_mov(dst_reg, src_reg);
696 	} else {
697 		/* mov32 dst, src */
698 		if (is_ereg(dst_reg) || is_ereg(src_reg))
699 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
700 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
701 	}
702 
703 	*pprog = prog;
704 }
705 
706 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
707 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
708 {
709 	u8 *prog = *pprog;
710 
711 	if (is_imm8(off)) {
712 		/* 1-byte signed displacement.
713 		 *
714 		 * If off == 0 we could skip this and save one extra byte, but
715 		 * special case of x86 R13 which always needs an offset is not
716 		 * worth the hassle
717 		 */
718 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
719 	} else {
720 		/* 4-byte signed displacement */
721 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
722 	}
723 	*pprog = prog;
724 }
725 
726 /*
727  * Emit a REX byte if it will be necessary to address these registers
728  */
729 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
730 {
731 	u8 *prog = *pprog;
732 
733 	if (is64)
734 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
735 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
736 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
737 	*pprog = prog;
738 }
739 
740 /*
741  * Similar version of maybe_emit_mod() for a single register
742  */
743 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
744 {
745 	u8 *prog = *pprog;
746 
747 	if (is64)
748 		EMIT1(add_1mod(0x48, reg));
749 	else if (is_ereg(reg))
750 		EMIT1(add_1mod(0x40, reg));
751 	*pprog = prog;
752 }
753 
754 /* LDX: dst_reg = *(u8*)(src_reg + off) */
755 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
756 {
757 	u8 *prog = *pprog;
758 
759 	switch (size) {
760 	case BPF_B:
761 		/* Emit 'movzx rax, byte ptr [rax + off]' */
762 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
763 		break;
764 	case BPF_H:
765 		/* Emit 'movzx rax, word ptr [rax + off]' */
766 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
767 		break;
768 	case BPF_W:
769 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
770 		if (is_ereg(dst_reg) || is_ereg(src_reg))
771 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
772 		else
773 			EMIT1(0x8B);
774 		break;
775 	case BPF_DW:
776 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
777 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
778 		break;
779 	}
780 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
781 	*pprog = prog;
782 }
783 
784 /* STX: *(u8*)(dst_reg + off) = src_reg */
785 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
786 {
787 	u8 *prog = *pprog;
788 
789 	switch (size) {
790 	case BPF_B:
791 		/* Emit 'mov byte ptr [rax + off], al' */
792 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
793 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
794 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
795 		else
796 			EMIT1(0x88);
797 		break;
798 	case BPF_H:
799 		if (is_ereg(dst_reg) || is_ereg(src_reg))
800 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
801 		else
802 			EMIT2(0x66, 0x89);
803 		break;
804 	case BPF_W:
805 		if (is_ereg(dst_reg) || is_ereg(src_reg))
806 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
807 		else
808 			EMIT1(0x89);
809 		break;
810 	case BPF_DW:
811 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
812 		break;
813 	}
814 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
815 	*pprog = prog;
816 }
817 
818 static int emit_atomic(u8 **pprog, u8 atomic_op,
819 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
820 {
821 	u8 *prog = *pprog;
822 
823 	EMIT1(0xF0); /* lock prefix */
824 
825 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
826 
827 	/* emit opcode */
828 	switch (atomic_op) {
829 	case BPF_ADD:
830 	case BPF_AND:
831 	case BPF_OR:
832 	case BPF_XOR:
833 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
834 		EMIT1(simple_alu_opcodes[atomic_op]);
835 		break;
836 	case BPF_ADD | BPF_FETCH:
837 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
838 		EMIT2(0x0F, 0xC1);
839 		break;
840 	case BPF_XCHG:
841 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
842 		EMIT1(0x87);
843 		break;
844 	case BPF_CMPXCHG:
845 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
846 		EMIT2(0x0F, 0xB1);
847 		break;
848 	default:
849 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
850 		return -EFAULT;
851 	}
852 
853 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
854 
855 	*pprog = prog;
856 	return 0;
857 }
858 
859 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
860 {
861 	u32 reg = x->fixup >> 8;
862 
863 	/* jump over faulting load and clear dest register */
864 	*(unsigned long *)((void *)regs + reg) = 0;
865 	regs->ip += x->fixup & 0xff;
866 	return true;
867 }
868 
869 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
870 			     bool *regs_used, bool *tail_call_seen)
871 {
872 	int i;
873 
874 	for (i = 1; i <= insn_cnt; i++, insn++) {
875 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
876 			*tail_call_seen = true;
877 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
878 			regs_used[0] = true;
879 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
880 			regs_used[1] = true;
881 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
882 			regs_used[2] = true;
883 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
884 			regs_used[3] = true;
885 	}
886 }
887 
888 static void emit_nops(u8 **pprog, int len)
889 {
890 	u8 *prog = *pprog;
891 	int i, noplen;
892 
893 	while (len > 0) {
894 		noplen = len;
895 
896 		if (noplen > ASM_NOP_MAX)
897 			noplen = ASM_NOP_MAX;
898 
899 		for (i = 0; i < noplen; i++)
900 			EMIT1(x86_nops[noplen][i]);
901 		len -= noplen;
902 	}
903 
904 	*pprog = prog;
905 }
906 
907 /* emit the 3-byte VEX prefix
908  *
909  * r: same as rex.r, extra bit for ModRM reg field
910  * x: same as rex.x, extra bit for SIB index field
911  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
912  * m: opcode map select, encoding escape bytes e.g. 0x0f38
913  * w: same as rex.w (32 bit or 64 bit) or opcode specific
914  * src_reg2: additional source reg (encoded as BPF reg)
915  * l: vector length (128 bit or 256 bit) or reserved
916  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
917  */
918 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
919 		      bool w, u8 src_reg2, bool l, u8 pp)
920 {
921 	u8 *prog = *pprog;
922 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
923 	u8 b1, b2;
924 	u8 vvvv = reg2hex[src_reg2];
925 
926 	/* reg2hex gives only the lower 3 bit of vvvv */
927 	if (is_ereg(src_reg2))
928 		vvvv |= 1 << 3;
929 
930 	/*
931 	 * 2nd byte of 3-byte VEX prefix
932 	 * ~ means bit inverted encoding
933 	 *
934 	 *    7                           0
935 	 *  +---+---+---+---+---+---+---+---+
936 	 *  |~R |~X |~B |         m         |
937 	 *  +---+---+---+---+---+---+---+---+
938 	 */
939 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
940 	/*
941 	 * 3rd byte of 3-byte VEX prefix
942 	 *
943 	 *    7                           0
944 	 *  +---+---+---+---+---+---+---+---+
945 	 *  | W |     ~vvvv     | L |   pp  |
946 	 *  +---+---+---+---+---+---+---+---+
947 	 */
948 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
949 
950 	EMIT3(b0, b1, b2);
951 	*pprog = prog;
952 }
953 
954 /* emit BMI2 shift instruction */
955 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
956 {
957 	u8 *prog = *pprog;
958 	bool r = is_ereg(dst_reg);
959 	u8 m = 2; /* escape code 0f38 */
960 
961 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
962 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
963 	*pprog = prog;
964 }
965 
966 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
967 
968 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
969 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
970 {
971 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
972 	struct bpf_insn *insn = bpf_prog->insnsi;
973 	bool callee_regs_used[4] = {};
974 	int insn_cnt = bpf_prog->len;
975 	bool tail_call_seen = false;
976 	bool seen_exit = false;
977 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
978 	int i, excnt = 0;
979 	int ilen, proglen = 0;
980 	u8 *prog = temp;
981 	int err;
982 
983 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
984 			 &tail_call_seen);
985 
986 	/* tail call's presence in current prog implies it is reachable */
987 	tail_call_reachable |= tail_call_seen;
988 
989 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
990 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
991 		      bpf_prog->aux->func_idx != 0);
992 	push_callee_regs(&prog, callee_regs_used);
993 
994 	ilen = prog - temp;
995 	if (rw_image)
996 		memcpy(rw_image + proglen, temp, ilen);
997 	proglen += ilen;
998 	addrs[0] = proglen;
999 	prog = temp;
1000 
1001 	for (i = 1; i <= insn_cnt; i++, insn++) {
1002 		const s32 imm32 = insn->imm;
1003 		u32 dst_reg = insn->dst_reg;
1004 		u32 src_reg = insn->src_reg;
1005 		u8 b2 = 0, b3 = 0;
1006 		u8 *start_of_ldx;
1007 		s64 jmp_offset;
1008 		u8 jmp_cond;
1009 		u8 *func;
1010 		int nops;
1011 
1012 		switch (insn->code) {
1013 			/* ALU */
1014 		case BPF_ALU | BPF_ADD | BPF_X:
1015 		case BPF_ALU | BPF_SUB | BPF_X:
1016 		case BPF_ALU | BPF_AND | BPF_X:
1017 		case BPF_ALU | BPF_OR | BPF_X:
1018 		case BPF_ALU | BPF_XOR | BPF_X:
1019 		case BPF_ALU64 | BPF_ADD | BPF_X:
1020 		case BPF_ALU64 | BPF_SUB | BPF_X:
1021 		case BPF_ALU64 | BPF_AND | BPF_X:
1022 		case BPF_ALU64 | BPF_OR | BPF_X:
1023 		case BPF_ALU64 | BPF_XOR | BPF_X:
1024 			maybe_emit_mod(&prog, dst_reg, src_reg,
1025 				       BPF_CLASS(insn->code) == BPF_ALU64);
1026 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1027 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1028 			break;
1029 
1030 		case BPF_ALU64 | BPF_MOV | BPF_X:
1031 		case BPF_ALU | BPF_MOV | BPF_X:
1032 			emit_mov_reg(&prog,
1033 				     BPF_CLASS(insn->code) == BPF_ALU64,
1034 				     dst_reg, src_reg);
1035 			break;
1036 
1037 			/* neg dst */
1038 		case BPF_ALU | BPF_NEG:
1039 		case BPF_ALU64 | BPF_NEG:
1040 			maybe_emit_1mod(&prog, dst_reg,
1041 					BPF_CLASS(insn->code) == BPF_ALU64);
1042 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1043 			break;
1044 
1045 		case BPF_ALU | BPF_ADD | BPF_K:
1046 		case BPF_ALU | BPF_SUB | BPF_K:
1047 		case BPF_ALU | BPF_AND | BPF_K:
1048 		case BPF_ALU | BPF_OR | BPF_K:
1049 		case BPF_ALU | BPF_XOR | BPF_K:
1050 		case BPF_ALU64 | BPF_ADD | BPF_K:
1051 		case BPF_ALU64 | BPF_SUB | BPF_K:
1052 		case BPF_ALU64 | BPF_AND | BPF_K:
1053 		case BPF_ALU64 | BPF_OR | BPF_K:
1054 		case BPF_ALU64 | BPF_XOR | BPF_K:
1055 			maybe_emit_1mod(&prog, dst_reg,
1056 					BPF_CLASS(insn->code) == BPF_ALU64);
1057 
1058 			/*
1059 			 * b3 holds 'normal' opcode, b2 short form only valid
1060 			 * in case dst is eax/rax.
1061 			 */
1062 			switch (BPF_OP(insn->code)) {
1063 			case BPF_ADD:
1064 				b3 = 0xC0;
1065 				b2 = 0x05;
1066 				break;
1067 			case BPF_SUB:
1068 				b3 = 0xE8;
1069 				b2 = 0x2D;
1070 				break;
1071 			case BPF_AND:
1072 				b3 = 0xE0;
1073 				b2 = 0x25;
1074 				break;
1075 			case BPF_OR:
1076 				b3 = 0xC8;
1077 				b2 = 0x0D;
1078 				break;
1079 			case BPF_XOR:
1080 				b3 = 0xF0;
1081 				b2 = 0x35;
1082 				break;
1083 			}
1084 
1085 			if (is_imm8(imm32))
1086 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1087 			else if (is_axreg(dst_reg))
1088 				EMIT1_off32(b2, imm32);
1089 			else
1090 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1091 			break;
1092 
1093 		case BPF_ALU64 | BPF_MOV | BPF_K:
1094 		case BPF_ALU | BPF_MOV | BPF_K:
1095 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1096 				       dst_reg, imm32);
1097 			break;
1098 
1099 		case BPF_LD | BPF_IMM | BPF_DW:
1100 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1101 			insn++;
1102 			i++;
1103 			break;
1104 
1105 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1106 		case BPF_ALU | BPF_MOD | BPF_X:
1107 		case BPF_ALU | BPF_DIV | BPF_X:
1108 		case BPF_ALU | BPF_MOD | BPF_K:
1109 		case BPF_ALU | BPF_DIV | BPF_K:
1110 		case BPF_ALU64 | BPF_MOD | BPF_X:
1111 		case BPF_ALU64 | BPF_DIV | BPF_X:
1112 		case BPF_ALU64 | BPF_MOD | BPF_K:
1113 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1114 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1115 
1116 			if (dst_reg != BPF_REG_0)
1117 				EMIT1(0x50); /* push rax */
1118 			if (dst_reg != BPF_REG_3)
1119 				EMIT1(0x52); /* push rdx */
1120 
1121 			if (BPF_SRC(insn->code) == BPF_X) {
1122 				if (src_reg == BPF_REG_0 ||
1123 				    src_reg == BPF_REG_3) {
1124 					/* mov r11, src_reg */
1125 					EMIT_mov(AUX_REG, src_reg);
1126 					src_reg = AUX_REG;
1127 				}
1128 			} else {
1129 				/* mov r11, imm32 */
1130 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1131 				src_reg = AUX_REG;
1132 			}
1133 
1134 			if (dst_reg != BPF_REG_0)
1135 				/* mov rax, dst_reg */
1136 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1137 
1138 			/*
1139 			 * xor edx, edx
1140 			 * equivalent to 'xor rdx, rdx', but one byte less
1141 			 */
1142 			EMIT2(0x31, 0xd2);
1143 
1144 			/* div src_reg */
1145 			maybe_emit_1mod(&prog, src_reg, is64);
1146 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1147 
1148 			if (BPF_OP(insn->code) == BPF_MOD &&
1149 			    dst_reg != BPF_REG_3)
1150 				/* mov dst_reg, rdx */
1151 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1152 			else if (BPF_OP(insn->code) == BPF_DIV &&
1153 				 dst_reg != BPF_REG_0)
1154 				/* mov dst_reg, rax */
1155 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1156 
1157 			if (dst_reg != BPF_REG_3)
1158 				EMIT1(0x5A); /* pop rdx */
1159 			if (dst_reg != BPF_REG_0)
1160 				EMIT1(0x58); /* pop rax */
1161 			break;
1162 		}
1163 
1164 		case BPF_ALU | BPF_MUL | BPF_K:
1165 		case BPF_ALU64 | BPF_MUL | BPF_K:
1166 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1167 				       BPF_CLASS(insn->code) == BPF_ALU64);
1168 
1169 			if (is_imm8(imm32))
1170 				/* imul dst_reg, dst_reg, imm8 */
1171 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1172 				      imm32);
1173 			else
1174 				/* imul dst_reg, dst_reg, imm32 */
1175 				EMIT2_off32(0x69,
1176 					    add_2reg(0xC0, dst_reg, dst_reg),
1177 					    imm32);
1178 			break;
1179 
1180 		case BPF_ALU | BPF_MUL | BPF_X:
1181 		case BPF_ALU64 | BPF_MUL | BPF_X:
1182 			maybe_emit_mod(&prog, src_reg, dst_reg,
1183 				       BPF_CLASS(insn->code) == BPF_ALU64);
1184 
1185 			/* imul dst_reg, src_reg */
1186 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1187 			break;
1188 
1189 			/* Shifts */
1190 		case BPF_ALU | BPF_LSH | BPF_K:
1191 		case BPF_ALU | BPF_RSH | BPF_K:
1192 		case BPF_ALU | BPF_ARSH | BPF_K:
1193 		case BPF_ALU64 | BPF_LSH | BPF_K:
1194 		case BPF_ALU64 | BPF_RSH | BPF_K:
1195 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1196 			maybe_emit_1mod(&prog, dst_reg,
1197 					BPF_CLASS(insn->code) == BPF_ALU64);
1198 
1199 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1200 			if (imm32 == 1)
1201 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1202 			else
1203 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1204 			break;
1205 
1206 		case BPF_ALU | BPF_LSH | BPF_X:
1207 		case BPF_ALU | BPF_RSH | BPF_X:
1208 		case BPF_ALU | BPF_ARSH | BPF_X:
1209 		case BPF_ALU64 | BPF_LSH | BPF_X:
1210 		case BPF_ALU64 | BPF_RSH | BPF_X:
1211 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1212 			/* BMI2 shifts aren't better when shift count is already in rcx */
1213 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1214 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1215 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1216 				u8 op;
1217 
1218 				switch (BPF_OP(insn->code)) {
1219 				case BPF_LSH:
1220 					op = 1; /* prefix 0x66 */
1221 					break;
1222 				case BPF_RSH:
1223 					op = 3; /* prefix 0xf2 */
1224 					break;
1225 				case BPF_ARSH:
1226 					op = 2; /* prefix 0xf3 */
1227 					break;
1228 				}
1229 
1230 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1231 
1232 				break;
1233 			}
1234 
1235 			if (src_reg != BPF_REG_4) { /* common case */
1236 				/* Check for bad case when dst_reg == rcx */
1237 				if (dst_reg == BPF_REG_4) {
1238 					/* mov r11, dst_reg */
1239 					EMIT_mov(AUX_REG, dst_reg);
1240 					dst_reg = AUX_REG;
1241 				} else {
1242 					EMIT1(0x51); /* push rcx */
1243 				}
1244 				/* mov rcx, src_reg */
1245 				EMIT_mov(BPF_REG_4, src_reg);
1246 			}
1247 
1248 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1249 			maybe_emit_1mod(&prog, dst_reg,
1250 					BPF_CLASS(insn->code) == BPF_ALU64);
1251 
1252 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1253 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1254 
1255 			if (src_reg != BPF_REG_4) {
1256 				if (insn->dst_reg == BPF_REG_4)
1257 					/* mov dst_reg, r11 */
1258 					EMIT_mov(insn->dst_reg, AUX_REG);
1259 				else
1260 					EMIT1(0x59); /* pop rcx */
1261 			}
1262 
1263 			break;
1264 
1265 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1266 			switch (imm32) {
1267 			case 16:
1268 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1269 				EMIT1(0x66);
1270 				if (is_ereg(dst_reg))
1271 					EMIT1(0x41);
1272 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1273 
1274 				/* Emit 'movzwl eax, ax' */
1275 				if (is_ereg(dst_reg))
1276 					EMIT3(0x45, 0x0F, 0xB7);
1277 				else
1278 					EMIT2(0x0F, 0xB7);
1279 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1280 				break;
1281 			case 32:
1282 				/* Emit 'bswap eax' to swap lower 4 bytes */
1283 				if (is_ereg(dst_reg))
1284 					EMIT2(0x41, 0x0F);
1285 				else
1286 					EMIT1(0x0F);
1287 				EMIT1(add_1reg(0xC8, dst_reg));
1288 				break;
1289 			case 64:
1290 				/* Emit 'bswap rax' to swap 8 bytes */
1291 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1292 				      add_1reg(0xC8, dst_reg));
1293 				break;
1294 			}
1295 			break;
1296 
1297 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1298 			switch (imm32) {
1299 			case 16:
1300 				/*
1301 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1302 				 * into 64 bit
1303 				 */
1304 				if (is_ereg(dst_reg))
1305 					EMIT3(0x45, 0x0F, 0xB7);
1306 				else
1307 					EMIT2(0x0F, 0xB7);
1308 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1309 				break;
1310 			case 32:
1311 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1312 				if (is_ereg(dst_reg))
1313 					EMIT1(0x45);
1314 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1315 				break;
1316 			case 64:
1317 				/* nop */
1318 				break;
1319 			}
1320 			break;
1321 
1322 			/* speculation barrier */
1323 		case BPF_ST | BPF_NOSPEC:
1324 			EMIT_LFENCE();
1325 			break;
1326 
1327 			/* ST: *(u8*)(dst_reg + off) = imm */
1328 		case BPF_ST | BPF_MEM | BPF_B:
1329 			if (is_ereg(dst_reg))
1330 				EMIT2(0x41, 0xC6);
1331 			else
1332 				EMIT1(0xC6);
1333 			goto st;
1334 		case BPF_ST | BPF_MEM | BPF_H:
1335 			if (is_ereg(dst_reg))
1336 				EMIT3(0x66, 0x41, 0xC7);
1337 			else
1338 				EMIT2(0x66, 0xC7);
1339 			goto st;
1340 		case BPF_ST | BPF_MEM | BPF_W:
1341 			if (is_ereg(dst_reg))
1342 				EMIT2(0x41, 0xC7);
1343 			else
1344 				EMIT1(0xC7);
1345 			goto st;
1346 		case BPF_ST | BPF_MEM | BPF_DW:
1347 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1348 
1349 st:			if (is_imm8(insn->off))
1350 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1351 			else
1352 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1353 
1354 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1355 			break;
1356 
1357 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1358 		case BPF_STX | BPF_MEM | BPF_B:
1359 		case BPF_STX | BPF_MEM | BPF_H:
1360 		case BPF_STX | BPF_MEM | BPF_W:
1361 		case BPF_STX | BPF_MEM | BPF_DW:
1362 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1363 			break;
1364 
1365 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1366 		case BPF_LDX | BPF_MEM | BPF_B:
1367 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1368 		case BPF_LDX | BPF_MEM | BPF_H:
1369 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1370 		case BPF_LDX | BPF_MEM | BPF_W:
1371 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1372 		case BPF_LDX | BPF_MEM | BPF_DW:
1373 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1374 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1375 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1376 				 * add abs(insn->off) to the limit to make sure that negative
1377 				 * offset won't be an issue.
1378 				 * insn->off is s16, so it won't affect valid pointers.
1379 				 */
1380 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1381 				u8 *end_of_jmp1, *end_of_jmp2;
1382 
1383 				/* Conservatively check that src_reg + insn->off is a kernel address:
1384 				 * 1. src_reg + insn->off >= limit
1385 				 * 2. src_reg + insn->off doesn't become small positive.
1386 				 * Cannot do src_reg + insn->off >= limit in one branch,
1387 				 * since it needs two spare registers, but JIT has only one.
1388 				 */
1389 
1390 				/* movabsq r11, limit */
1391 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1392 				EMIT((u32)limit, 4);
1393 				EMIT(limit >> 32, 4);
1394 				/* cmp src_reg, r11 */
1395 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1396 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1397 				/* if unsigned '<' goto end_of_jmp2 */
1398 				EMIT2(X86_JB, 0);
1399 				end_of_jmp1 = prog;
1400 
1401 				/* mov r11, src_reg */
1402 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1403 				/* add r11, insn->off */
1404 				maybe_emit_1mod(&prog, AUX_REG, true);
1405 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1406 				/* jmp if not carry to start_of_ldx
1407 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1408 				 * that has to be rejected.
1409 				 */
1410 				EMIT2(0x73 /* JNC */, 0);
1411 				end_of_jmp2 = prog;
1412 
1413 				/* xor dst_reg, dst_reg */
1414 				emit_mov_imm32(&prog, false, dst_reg, 0);
1415 				/* jmp byte_after_ldx */
1416 				EMIT2(0xEB, 0);
1417 
1418 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1419 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1420 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1421 				start_of_ldx = prog;
1422 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1423 			}
1424 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1425 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1426 				struct exception_table_entry *ex;
1427 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1428 				s64 delta;
1429 
1430 				/* populate jmp_offset for JMP above */
1431 				start_of_ldx[-1] = prog - start_of_ldx;
1432 
1433 				if (!bpf_prog->aux->extable)
1434 					break;
1435 
1436 				if (excnt >= bpf_prog->aux->num_exentries) {
1437 					pr_err("ex gen bug\n");
1438 					return -EFAULT;
1439 				}
1440 				ex = &bpf_prog->aux->extable[excnt++];
1441 
1442 				delta = _insn - (u8 *)&ex->insn;
1443 				if (!is_simm32(delta)) {
1444 					pr_err("extable->insn doesn't fit into 32-bit\n");
1445 					return -EFAULT;
1446 				}
1447 				/* switch ex to rw buffer for writes */
1448 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1449 
1450 				ex->insn = delta;
1451 
1452 				ex->data = EX_TYPE_BPF;
1453 
1454 				if (dst_reg > BPF_REG_9) {
1455 					pr_err("verifier error\n");
1456 					return -EFAULT;
1457 				}
1458 				/*
1459 				 * Compute size of x86 insn and its target dest x86 register.
1460 				 * ex_handler_bpf() will use lower 8 bits to adjust
1461 				 * pt_regs->ip to jump over this x86 instruction
1462 				 * and upper bits to figure out which pt_regs to zero out.
1463 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1464 				 * of 4 bytes will be ignored and rbx will be zero inited.
1465 				 */
1466 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1467 			}
1468 			break;
1469 
1470 		case BPF_STX | BPF_ATOMIC | BPF_W:
1471 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1472 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1473 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1474 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1475 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1476 				u32 real_src_reg = src_reg;
1477 				u32 real_dst_reg = dst_reg;
1478 				u8 *branch_target;
1479 
1480 				/*
1481 				 * Can't be implemented with a single x86 insn.
1482 				 * Need to do a CMPXCHG loop.
1483 				 */
1484 
1485 				/* Will need RAX as a CMPXCHG operand so save R0 */
1486 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1487 				if (src_reg == BPF_REG_0)
1488 					real_src_reg = BPF_REG_AX;
1489 				if (dst_reg == BPF_REG_0)
1490 					real_dst_reg = BPF_REG_AX;
1491 
1492 				branch_target = prog;
1493 				/* Load old value */
1494 				emit_ldx(&prog, BPF_SIZE(insn->code),
1495 					 BPF_REG_0, real_dst_reg, insn->off);
1496 				/*
1497 				 * Perform the (commutative) operation locally,
1498 				 * put the result in the AUX_REG.
1499 				 */
1500 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1501 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1502 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1503 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1504 				/* Attempt to swap in new value */
1505 				err = emit_atomic(&prog, BPF_CMPXCHG,
1506 						  real_dst_reg, AUX_REG,
1507 						  insn->off,
1508 						  BPF_SIZE(insn->code));
1509 				if (WARN_ON(err))
1510 					return err;
1511 				/*
1512 				 * ZF tells us whether we won the race. If it's
1513 				 * cleared we need to try again.
1514 				 */
1515 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1516 				/* Return the pre-modification value */
1517 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1518 				/* Restore R0 after clobbering RAX */
1519 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1520 				break;
1521 			}
1522 
1523 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1524 					  insn->off, BPF_SIZE(insn->code));
1525 			if (err)
1526 				return err;
1527 			break;
1528 
1529 			/* call */
1530 		case BPF_JMP | BPF_CALL:
1531 			func = (u8 *) __bpf_call_base + imm32;
1532 			if (tail_call_reachable) {
1533 				/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1534 				EMIT3_off32(0x48, 0x8B, 0x85,
1535 					    -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1536 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1537 					return -EINVAL;
1538 			} else {
1539 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1540 					return -EINVAL;
1541 			}
1542 			break;
1543 
1544 		case BPF_JMP | BPF_TAIL_CALL:
1545 			if (imm32)
1546 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1547 							  &prog, image + addrs[i - 1],
1548 							  callee_regs_used,
1549 							  bpf_prog->aux->stack_depth,
1550 							  ctx);
1551 			else
1552 				emit_bpf_tail_call_indirect(&prog,
1553 							    callee_regs_used,
1554 							    bpf_prog->aux->stack_depth,
1555 							    image + addrs[i - 1],
1556 							    ctx);
1557 			break;
1558 
1559 			/* cond jump */
1560 		case BPF_JMP | BPF_JEQ | BPF_X:
1561 		case BPF_JMP | BPF_JNE | BPF_X:
1562 		case BPF_JMP | BPF_JGT | BPF_X:
1563 		case BPF_JMP | BPF_JLT | BPF_X:
1564 		case BPF_JMP | BPF_JGE | BPF_X:
1565 		case BPF_JMP | BPF_JLE | BPF_X:
1566 		case BPF_JMP | BPF_JSGT | BPF_X:
1567 		case BPF_JMP | BPF_JSLT | BPF_X:
1568 		case BPF_JMP | BPF_JSGE | BPF_X:
1569 		case BPF_JMP | BPF_JSLE | BPF_X:
1570 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1571 		case BPF_JMP32 | BPF_JNE | BPF_X:
1572 		case BPF_JMP32 | BPF_JGT | BPF_X:
1573 		case BPF_JMP32 | BPF_JLT | BPF_X:
1574 		case BPF_JMP32 | BPF_JGE | BPF_X:
1575 		case BPF_JMP32 | BPF_JLE | BPF_X:
1576 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1577 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1578 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1579 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1580 			/* cmp dst_reg, src_reg */
1581 			maybe_emit_mod(&prog, dst_reg, src_reg,
1582 				       BPF_CLASS(insn->code) == BPF_JMP);
1583 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1584 			goto emit_cond_jmp;
1585 
1586 		case BPF_JMP | BPF_JSET | BPF_X:
1587 		case BPF_JMP32 | BPF_JSET | BPF_X:
1588 			/* test dst_reg, src_reg */
1589 			maybe_emit_mod(&prog, dst_reg, src_reg,
1590 				       BPF_CLASS(insn->code) == BPF_JMP);
1591 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1592 			goto emit_cond_jmp;
1593 
1594 		case BPF_JMP | BPF_JSET | BPF_K:
1595 		case BPF_JMP32 | BPF_JSET | BPF_K:
1596 			/* test dst_reg, imm32 */
1597 			maybe_emit_1mod(&prog, dst_reg,
1598 					BPF_CLASS(insn->code) == BPF_JMP);
1599 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1600 			goto emit_cond_jmp;
1601 
1602 		case BPF_JMP | BPF_JEQ | BPF_K:
1603 		case BPF_JMP | BPF_JNE | BPF_K:
1604 		case BPF_JMP | BPF_JGT | BPF_K:
1605 		case BPF_JMP | BPF_JLT | BPF_K:
1606 		case BPF_JMP | BPF_JGE | BPF_K:
1607 		case BPF_JMP | BPF_JLE | BPF_K:
1608 		case BPF_JMP | BPF_JSGT | BPF_K:
1609 		case BPF_JMP | BPF_JSLT | BPF_K:
1610 		case BPF_JMP | BPF_JSGE | BPF_K:
1611 		case BPF_JMP | BPF_JSLE | BPF_K:
1612 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1613 		case BPF_JMP32 | BPF_JNE | BPF_K:
1614 		case BPF_JMP32 | BPF_JGT | BPF_K:
1615 		case BPF_JMP32 | BPF_JLT | BPF_K:
1616 		case BPF_JMP32 | BPF_JGE | BPF_K:
1617 		case BPF_JMP32 | BPF_JLE | BPF_K:
1618 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1619 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1620 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1621 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1622 			/* test dst_reg, dst_reg to save one extra byte */
1623 			if (imm32 == 0) {
1624 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1625 					       BPF_CLASS(insn->code) == BPF_JMP);
1626 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1627 				goto emit_cond_jmp;
1628 			}
1629 
1630 			/* cmp dst_reg, imm8/32 */
1631 			maybe_emit_1mod(&prog, dst_reg,
1632 					BPF_CLASS(insn->code) == BPF_JMP);
1633 
1634 			if (is_imm8(imm32))
1635 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1636 			else
1637 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1638 
1639 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1640 			switch (BPF_OP(insn->code)) {
1641 			case BPF_JEQ:
1642 				jmp_cond = X86_JE;
1643 				break;
1644 			case BPF_JSET:
1645 			case BPF_JNE:
1646 				jmp_cond = X86_JNE;
1647 				break;
1648 			case BPF_JGT:
1649 				/* GT is unsigned '>', JA in x86 */
1650 				jmp_cond = X86_JA;
1651 				break;
1652 			case BPF_JLT:
1653 				/* LT is unsigned '<', JB in x86 */
1654 				jmp_cond = X86_JB;
1655 				break;
1656 			case BPF_JGE:
1657 				/* GE is unsigned '>=', JAE in x86 */
1658 				jmp_cond = X86_JAE;
1659 				break;
1660 			case BPF_JLE:
1661 				/* LE is unsigned '<=', JBE in x86 */
1662 				jmp_cond = X86_JBE;
1663 				break;
1664 			case BPF_JSGT:
1665 				/* Signed '>', GT in x86 */
1666 				jmp_cond = X86_JG;
1667 				break;
1668 			case BPF_JSLT:
1669 				/* Signed '<', LT in x86 */
1670 				jmp_cond = X86_JL;
1671 				break;
1672 			case BPF_JSGE:
1673 				/* Signed '>=', GE in x86 */
1674 				jmp_cond = X86_JGE;
1675 				break;
1676 			case BPF_JSLE:
1677 				/* Signed '<=', LE in x86 */
1678 				jmp_cond = X86_JLE;
1679 				break;
1680 			default: /* to silence GCC warning */
1681 				return -EFAULT;
1682 			}
1683 			jmp_offset = addrs[i + insn->off] - addrs[i];
1684 			if (is_imm8(jmp_offset)) {
1685 				if (jmp_padding) {
1686 					/* To keep the jmp_offset valid, the extra bytes are
1687 					 * padded before the jump insn, so we subtract the
1688 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1689 					 *
1690 					 * If the previous pass already emits an imm8
1691 					 * jmp_cond, then this BPF insn won't shrink, so
1692 					 * "nops" is 0.
1693 					 *
1694 					 * On the other hand, if the previous pass emits an
1695 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1696 					 * keep the image from shrinking further.
1697 					 *
1698 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1699 					 *     is 2 bytes, so the size difference is 4 bytes.
1700 					 */
1701 					nops = INSN_SZ_DIFF - 2;
1702 					if (nops != 0 && nops != 4) {
1703 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1704 						       nops);
1705 						return -EFAULT;
1706 					}
1707 					emit_nops(&prog, nops);
1708 				}
1709 				EMIT2(jmp_cond, jmp_offset);
1710 			} else if (is_simm32(jmp_offset)) {
1711 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1712 			} else {
1713 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1714 				return -EFAULT;
1715 			}
1716 
1717 			break;
1718 
1719 		case BPF_JMP | BPF_JA:
1720 			if (insn->off == -1)
1721 				/* -1 jmp instructions will always jump
1722 				 * backwards two bytes. Explicitly handling
1723 				 * this case avoids wasting too many passes
1724 				 * when there are long sequences of replaced
1725 				 * dead code.
1726 				 */
1727 				jmp_offset = -2;
1728 			else
1729 				jmp_offset = addrs[i + insn->off] - addrs[i];
1730 
1731 			if (!jmp_offset) {
1732 				/*
1733 				 * If jmp_padding is enabled, the extra nops will
1734 				 * be inserted. Otherwise, optimize out nop jumps.
1735 				 */
1736 				if (jmp_padding) {
1737 					/* There are 3 possible conditions.
1738 					 * (1) This BPF_JA is already optimized out in
1739 					 *     the previous run, so there is no need
1740 					 *     to pad any extra byte (0 byte).
1741 					 * (2) The previous pass emits an imm8 jmp,
1742 					 *     so we pad 2 bytes to match the previous
1743 					 *     insn size.
1744 					 * (3) Similarly, the previous pass emits an
1745 					 *     imm32 jmp, and 5 bytes is padded.
1746 					 */
1747 					nops = INSN_SZ_DIFF;
1748 					if (nops != 0 && nops != 2 && nops != 5) {
1749 						pr_err("unexpected nop jump padding: %d bytes\n",
1750 						       nops);
1751 						return -EFAULT;
1752 					}
1753 					emit_nops(&prog, nops);
1754 				}
1755 				break;
1756 			}
1757 emit_jmp:
1758 			if (is_imm8(jmp_offset)) {
1759 				if (jmp_padding) {
1760 					/* To avoid breaking jmp_offset, the extra bytes
1761 					 * are padded before the actual jmp insn, so
1762 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1763 					 *
1764 					 * If the previous pass already emits an imm8
1765 					 * jmp, there is nothing to pad (0 byte).
1766 					 *
1767 					 * If it emits an imm32 jmp (5 bytes) previously
1768 					 * and now an imm8 jmp (2 bytes), then we pad
1769 					 * (5 - 2 = 3) bytes to stop the image from
1770 					 * shrinking further.
1771 					 */
1772 					nops = INSN_SZ_DIFF - 2;
1773 					if (nops != 0 && nops != 3) {
1774 						pr_err("unexpected jump padding: %d bytes\n",
1775 						       nops);
1776 						return -EFAULT;
1777 					}
1778 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1779 				}
1780 				EMIT2(0xEB, jmp_offset);
1781 			} else if (is_simm32(jmp_offset)) {
1782 				EMIT1_off32(0xE9, jmp_offset);
1783 			} else {
1784 				pr_err("jmp gen bug %llx\n", jmp_offset);
1785 				return -EFAULT;
1786 			}
1787 			break;
1788 
1789 		case BPF_JMP | BPF_EXIT:
1790 			if (seen_exit) {
1791 				jmp_offset = ctx->cleanup_addr - addrs[i];
1792 				goto emit_jmp;
1793 			}
1794 			seen_exit = true;
1795 			/* Update cleanup_addr */
1796 			ctx->cleanup_addr = proglen;
1797 			pop_callee_regs(&prog, callee_regs_used);
1798 			EMIT1(0xC9);         /* leave */
1799 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1800 			break;
1801 
1802 		default:
1803 			/*
1804 			 * By design x86-64 JIT should support all BPF instructions.
1805 			 * This error will be seen if new instruction was added
1806 			 * to the interpreter, but not to the JIT, or if there is
1807 			 * junk in bpf_prog.
1808 			 */
1809 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1810 			return -EINVAL;
1811 		}
1812 
1813 		ilen = prog - temp;
1814 		if (ilen > BPF_MAX_INSN_SIZE) {
1815 			pr_err("bpf_jit: fatal insn size error\n");
1816 			return -EFAULT;
1817 		}
1818 
1819 		if (image) {
1820 			/*
1821 			 * When populating the image, assert that:
1822 			 *
1823 			 *  i) We do not write beyond the allocated space, and
1824 			 * ii) addrs[i] did not change from the prior run, in order
1825 			 *     to validate assumptions made for computing branch
1826 			 *     displacements.
1827 			 */
1828 			if (unlikely(proglen + ilen > oldproglen ||
1829 				     proglen + ilen != addrs[i])) {
1830 				pr_err("bpf_jit: fatal error\n");
1831 				return -EFAULT;
1832 			}
1833 			memcpy(rw_image + proglen, temp, ilen);
1834 		}
1835 		proglen += ilen;
1836 		addrs[i] = proglen;
1837 		prog = temp;
1838 	}
1839 
1840 	if (image && excnt != bpf_prog->aux->num_exentries) {
1841 		pr_err("extable is not populated\n");
1842 		return -EFAULT;
1843 	}
1844 	return proglen;
1845 }
1846 
1847 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1848 		      int stack_size)
1849 {
1850 	int i, j, arg_size, nr_regs;
1851 	/* Store function arguments to stack.
1852 	 * For a function that accepts two pointers the sequence will be:
1853 	 * mov QWORD PTR [rbp-0x10],rdi
1854 	 * mov QWORD PTR [rbp-0x8],rsi
1855 	 */
1856 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1857 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1858 			nr_regs = (m->arg_size[i] + 7) / 8;
1859 			arg_size = 8;
1860 		} else {
1861 			nr_regs = 1;
1862 			arg_size = m->arg_size[i];
1863 		}
1864 
1865 		while (nr_regs) {
1866 			emit_stx(prog, bytes_to_bpf_size(arg_size),
1867 				 BPF_REG_FP,
1868 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1869 				 -(stack_size - j * 8));
1870 			nr_regs--;
1871 			j++;
1872 		}
1873 	}
1874 }
1875 
1876 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1877 			 int stack_size)
1878 {
1879 	int i, j, arg_size, nr_regs;
1880 
1881 	/* Restore function arguments from stack.
1882 	 * For a function that accepts two pointers the sequence will be:
1883 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1884 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1885 	 */
1886 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1887 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1888 			nr_regs = (m->arg_size[i] + 7) / 8;
1889 			arg_size = 8;
1890 		} else {
1891 			nr_regs = 1;
1892 			arg_size = m->arg_size[i];
1893 		}
1894 
1895 		while (nr_regs) {
1896 			emit_ldx(prog, bytes_to_bpf_size(arg_size),
1897 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1898 				 BPF_REG_FP,
1899 				 -(stack_size - j * 8));
1900 			nr_regs--;
1901 			j++;
1902 		}
1903 	}
1904 }
1905 
1906 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1907 			   struct bpf_tramp_link *l, int stack_size,
1908 			   int run_ctx_off, bool save_ret)
1909 {
1910 	u8 *prog = *pprog;
1911 	u8 *jmp_insn;
1912 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1913 	struct bpf_prog *p = l->link.prog;
1914 	u64 cookie = l->cookie;
1915 
1916 	/* mov rdi, cookie */
1917 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
1918 
1919 	/* Prepare struct bpf_tramp_run_ctx.
1920 	 *
1921 	 * bpf_tramp_run_ctx is already preserved by
1922 	 * arch_prepare_bpf_trampoline().
1923 	 *
1924 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
1925 	 */
1926 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
1927 
1928 	/* arg1: mov rdi, progs[i] */
1929 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1930 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
1931 	EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
1932 
1933 	if (emit_call(&prog, bpf_trampoline_enter(p), prog))
1934 		return -EINVAL;
1935 	/* remember prog start time returned by __bpf_prog_enter */
1936 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1937 
1938 	/* if (__bpf_prog_enter*(prog) == 0)
1939 	 *	goto skip_exec_of_prog;
1940 	 */
1941 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1942 	/* emit 2 nops that will be replaced with JE insn */
1943 	jmp_insn = prog;
1944 	emit_nops(&prog, 2);
1945 
1946 	/* arg1: lea rdi, [rbp - stack_size] */
1947 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1948 	/* arg2: progs[i]->insnsi for interpreter */
1949 	if (!p->jited)
1950 		emit_mov_imm64(&prog, BPF_REG_2,
1951 			       (long) p->insnsi >> 32,
1952 			       (u32) (long) p->insnsi);
1953 	/* call JITed bpf program or interpreter */
1954 	if (emit_call(&prog, p->bpf_func, prog))
1955 		return -EINVAL;
1956 
1957 	/*
1958 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1959 	 * of the previous call which is then passed on the stack to
1960 	 * the next BPF program.
1961 	 *
1962 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1963 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1964 	 */
1965 	if (save_ret)
1966 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1967 
1968 	/* replace 2 nops with JE insn, since jmp target is known */
1969 	jmp_insn[0] = X86_JE;
1970 	jmp_insn[1] = prog - jmp_insn - 2;
1971 
1972 	/* arg1: mov rdi, progs[i] */
1973 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1974 	/* arg2: mov rsi, rbx <- start time in nsec */
1975 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1976 	/* arg3: lea rdx, [rbp - run_ctx_off] */
1977 	EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
1978 	if (emit_call(&prog, bpf_trampoline_exit(p), prog))
1979 		return -EINVAL;
1980 
1981 	*pprog = prog;
1982 	return 0;
1983 }
1984 
1985 static void emit_align(u8 **pprog, u32 align)
1986 {
1987 	u8 *target, *prog = *pprog;
1988 
1989 	target = PTR_ALIGN(prog, align);
1990 	if (target != prog)
1991 		emit_nops(&prog, target - prog);
1992 
1993 	*pprog = prog;
1994 }
1995 
1996 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1997 {
1998 	u8 *prog = *pprog;
1999 	s64 offset;
2000 
2001 	offset = func - (ip + 2 + 4);
2002 	if (!is_simm32(offset)) {
2003 		pr_err("Target %p is out of range\n", func);
2004 		return -EINVAL;
2005 	}
2006 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2007 	*pprog = prog;
2008 	return 0;
2009 }
2010 
2011 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2012 		      struct bpf_tramp_links *tl, int stack_size,
2013 		      int run_ctx_off, bool save_ret)
2014 {
2015 	int i;
2016 	u8 *prog = *pprog;
2017 
2018 	for (i = 0; i < tl->nr_links; i++) {
2019 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2020 				    run_ctx_off, save_ret))
2021 			return -EINVAL;
2022 	}
2023 	*pprog = prog;
2024 	return 0;
2025 }
2026 
2027 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2028 			      struct bpf_tramp_links *tl, int stack_size,
2029 			      int run_ctx_off, u8 **branches)
2030 {
2031 	u8 *prog = *pprog;
2032 	int i;
2033 
2034 	/* The first fmod_ret program will receive a garbage return value.
2035 	 * Set this to 0 to avoid confusing the program.
2036 	 */
2037 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2038 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2039 	for (i = 0; i < tl->nr_links; i++) {
2040 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
2041 			return -EINVAL;
2042 
2043 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2044 		 * if (*(u64 *)(rbp - 8) !=  0)
2045 		 *	goto do_fexit;
2046 		 */
2047 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2048 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2049 
2050 		/* Save the location of the branch and Generate 6 nops
2051 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2052 		 * are replaced with a conditional jump once do_fexit (i.e. the
2053 		 * start of the fexit invocation) is finalized.
2054 		 */
2055 		branches[i] = prog;
2056 		emit_nops(&prog, 4 + 2);
2057 	}
2058 
2059 	*pprog = prog;
2060 	return 0;
2061 }
2062 
2063 /* Example:
2064  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2065  * its 'struct btf_func_model' will be nr_args=2
2066  * The assembly code when eth_type_trans is executing after trampoline:
2067  *
2068  * push rbp
2069  * mov rbp, rsp
2070  * sub rsp, 16                     // space for skb and dev
2071  * push rbx                        // temp regs to pass start time
2072  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2073  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2074  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2075  * mov rbx, rax                    // remember start time in bpf stats are enabled
2076  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2077  * call addr_of_jited_FENTRY_prog
2078  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2079  * mov rsi, rbx                    // prog start time
2080  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2081  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2082  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2083  * pop rbx
2084  * leave
2085  * ret
2086  *
2087  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2088  * replaced with 'call generated_bpf_trampoline'. When it returns
2089  * eth_type_trans will continue executing with original skb and dev pointers.
2090  *
2091  * The assembly code when eth_type_trans is called from trampoline:
2092  *
2093  * push rbp
2094  * mov rbp, rsp
2095  * sub rsp, 24                     // space for skb, dev, return value
2096  * push rbx                        // temp regs to pass start time
2097  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2098  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2099  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2100  * mov rbx, rax                    // remember start time if bpf stats are enabled
2101  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2102  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2103  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2104  * mov rsi, rbx                    // prog start time
2105  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2106  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2107  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2108  * call eth_type_trans+5           // execute body of eth_type_trans
2109  * mov qword ptr [rbp - 8], rax    // save return value
2110  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2111  * mov rbx, rax                    // remember start time in bpf stats are enabled
2112  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2113  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2114  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2115  * mov rsi, rbx                    // prog start time
2116  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2117  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2118  * pop rbx
2119  * leave
2120  * add rsp, 8                      // skip eth_type_trans's frame
2121  * ret                             // return to its caller
2122  */
2123 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2124 				const struct btf_func_model *m, u32 flags,
2125 				struct bpf_tramp_links *tlinks,
2126 				void *func_addr)
2127 {
2128 	int ret, i, nr_args = m->nr_args, extra_nregs = 0;
2129 	int regs_off, ip_off, args_off, stack_size = nr_args * 8, run_ctx_off;
2130 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2131 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2132 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2133 	void *orig_call = func_addr;
2134 	u8 **branches = NULL;
2135 	u8 *prog;
2136 	bool save_ret;
2137 
2138 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2139 	if (nr_args > 6)
2140 		return -ENOTSUPP;
2141 
2142 	for (i = 0; i < MAX_BPF_FUNC_ARGS; i++) {
2143 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2144 			extra_nregs += (m->arg_size[i] + 7) / 8 - 1;
2145 	}
2146 	if (nr_args + extra_nregs > 6)
2147 		return -ENOTSUPP;
2148 	stack_size += extra_nregs * 8;
2149 
2150 	/* Generated trampoline stack layout:
2151 	 *
2152 	 * RBP + 8         [ return address  ]
2153 	 * RBP + 0         [ RBP             ]
2154 	 *
2155 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2156 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2157 	 *
2158 	 *                 [ reg_argN        ]  always
2159 	 *                 [ ...             ]
2160 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2161 	 *
2162 	 * RBP - args_off  [ arg regs count  ]  always
2163 	 *
2164 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2165 	 *
2166 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2167 	 */
2168 
2169 	/* room for return value of orig_call or fentry prog */
2170 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2171 	if (save_ret)
2172 		stack_size += 8;
2173 
2174 	regs_off = stack_size;
2175 
2176 	/* args count  */
2177 	stack_size += 8;
2178 	args_off = stack_size;
2179 
2180 	if (flags & BPF_TRAMP_F_IP_ARG)
2181 		stack_size += 8; /* room for IP address argument */
2182 
2183 	ip_off = stack_size;
2184 
2185 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2186 	run_ctx_off = stack_size;
2187 
2188 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2189 		/* skip patched call instruction and point orig_call to actual
2190 		 * body of the kernel function.
2191 		 */
2192 		if (is_endbr(*(u32 *)orig_call))
2193 			orig_call += ENDBR_INSN_SIZE;
2194 		orig_call += X86_PATCH_SIZE;
2195 	}
2196 
2197 	prog = image;
2198 
2199 	EMIT_ENDBR();
2200 	EMIT1(0x55);		 /* push rbp */
2201 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2202 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2203 	EMIT1(0x53);		 /* push rbx */
2204 
2205 	/* Store number of argument registers of the traced function:
2206 	 *   mov rax, nr_args + extra_nregs
2207 	 *   mov QWORD PTR [rbp - args_off], rax
2208 	 */
2209 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args + extra_nregs);
2210 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2211 
2212 	if (flags & BPF_TRAMP_F_IP_ARG) {
2213 		/* Store IP address of the traced function:
2214 		 * movabsq rax, func_addr
2215 		 * mov QWORD PTR [rbp - ip_off], rax
2216 		 */
2217 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2218 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2219 	}
2220 
2221 	save_regs(m, &prog, nr_args, regs_off);
2222 
2223 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2224 		/* arg1: mov rdi, im */
2225 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2226 		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2227 			ret = -EINVAL;
2228 			goto cleanup;
2229 		}
2230 	}
2231 
2232 	if (fentry->nr_links)
2233 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2234 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2235 			return -EINVAL;
2236 
2237 	if (fmod_ret->nr_links) {
2238 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2239 				   GFP_KERNEL);
2240 		if (!branches)
2241 			return -ENOMEM;
2242 
2243 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2244 				       run_ctx_off, branches)) {
2245 			ret = -EINVAL;
2246 			goto cleanup;
2247 		}
2248 	}
2249 
2250 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2251 		restore_regs(m, &prog, nr_args, regs_off);
2252 
2253 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2254 			emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2255 			EMIT2(0xff, 0xd0); /* call *rax */
2256 		} else {
2257 			/* call original function */
2258 			if (emit_call(&prog, orig_call, prog)) {
2259 				ret = -EINVAL;
2260 				goto cleanup;
2261 			}
2262 		}
2263 		/* remember return value in a stack for bpf prog to access */
2264 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2265 		im->ip_after_call = prog;
2266 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2267 		prog += X86_PATCH_SIZE;
2268 	}
2269 
2270 	if (fmod_ret->nr_links) {
2271 		/* From Intel 64 and IA-32 Architectures Optimization
2272 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2273 		 * Coding Rule 11: All branch targets should be 16-byte
2274 		 * aligned.
2275 		 */
2276 		emit_align(&prog, 16);
2277 		/* Update the branches saved in invoke_bpf_mod_ret with the
2278 		 * aligned address of do_fexit.
2279 		 */
2280 		for (i = 0; i < fmod_ret->nr_links; i++)
2281 			emit_cond_near_jump(&branches[i], prog, branches[i],
2282 					    X86_JNE);
2283 	}
2284 
2285 	if (fexit->nr_links)
2286 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2287 			ret = -EINVAL;
2288 			goto cleanup;
2289 		}
2290 
2291 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2292 		restore_regs(m, &prog, nr_args, regs_off);
2293 
2294 	/* This needs to be done regardless. If there were fmod_ret programs,
2295 	 * the return value is only updated on the stack and still needs to be
2296 	 * restored to R0.
2297 	 */
2298 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2299 		im->ip_epilogue = prog;
2300 		/* arg1: mov rdi, im */
2301 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2302 		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2303 			ret = -EINVAL;
2304 			goto cleanup;
2305 		}
2306 	}
2307 	/* restore return value of orig_call or fentry prog back into RAX */
2308 	if (save_ret)
2309 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2310 
2311 	EMIT1(0x5B); /* pop rbx */
2312 	EMIT1(0xC9); /* leave */
2313 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2314 		/* skip our return address and return to parent */
2315 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2316 	emit_return(&prog, prog);
2317 	/* Make sure the trampoline generation logic doesn't overflow */
2318 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2319 		ret = -EFAULT;
2320 		goto cleanup;
2321 	}
2322 	ret = prog - (u8 *)image;
2323 
2324 cleanup:
2325 	kfree(branches);
2326 	return ret;
2327 }
2328 
2329 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2330 {
2331 	u8 *jg_reloc, *prog = *pprog;
2332 	int pivot, err, jg_bytes = 1;
2333 	s64 jg_offset;
2334 
2335 	if (a == b) {
2336 		/* Leaf node of recursion, i.e. not a range of indices
2337 		 * anymore.
2338 		 */
2339 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2340 		if (!is_simm32(progs[a]))
2341 			return -1;
2342 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2343 			    progs[a]);
2344 		err = emit_cond_near_jump(&prog,	/* je func */
2345 					  (void *)progs[a], image + (prog - buf),
2346 					  X86_JE);
2347 		if (err)
2348 			return err;
2349 
2350 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2351 
2352 		*pprog = prog;
2353 		return 0;
2354 	}
2355 
2356 	/* Not a leaf node, so we pivot, and recursively descend into
2357 	 * the lower and upper ranges.
2358 	 */
2359 	pivot = (b - a) / 2;
2360 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2361 	if (!is_simm32(progs[a + pivot]))
2362 		return -1;
2363 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2364 
2365 	if (pivot > 2) {				/* jg upper_part */
2366 		/* Require near jump. */
2367 		jg_bytes = 4;
2368 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2369 	} else {
2370 		EMIT2(X86_JG, 0);
2371 	}
2372 	jg_reloc = prog;
2373 
2374 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2375 				  progs, image, buf);
2376 	if (err)
2377 		return err;
2378 
2379 	/* From Intel 64 and IA-32 Architectures Optimization
2380 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2381 	 * Coding Rule 11: All branch targets should be 16-byte
2382 	 * aligned.
2383 	 */
2384 	emit_align(&prog, 16);
2385 	jg_offset = prog - jg_reloc;
2386 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2387 
2388 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2389 				  b, progs, image, buf);
2390 	if (err)
2391 		return err;
2392 
2393 	*pprog = prog;
2394 	return 0;
2395 }
2396 
2397 static int cmp_ips(const void *a, const void *b)
2398 {
2399 	const s64 *ipa = a;
2400 	const s64 *ipb = b;
2401 
2402 	if (*ipa > *ipb)
2403 		return 1;
2404 	if (*ipa < *ipb)
2405 		return -1;
2406 	return 0;
2407 }
2408 
2409 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2410 {
2411 	u8 *prog = buf;
2412 
2413 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2414 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2415 }
2416 
2417 struct x64_jit_data {
2418 	struct bpf_binary_header *rw_header;
2419 	struct bpf_binary_header *header;
2420 	int *addrs;
2421 	u8 *image;
2422 	int proglen;
2423 	struct jit_context ctx;
2424 };
2425 
2426 #define MAX_PASSES 20
2427 #define PADDING_PASSES (MAX_PASSES - 5)
2428 
2429 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2430 {
2431 	struct bpf_binary_header *rw_header = NULL;
2432 	struct bpf_binary_header *header = NULL;
2433 	struct bpf_prog *tmp, *orig_prog = prog;
2434 	struct x64_jit_data *jit_data;
2435 	int proglen, oldproglen = 0;
2436 	struct jit_context ctx = {};
2437 	bool tmp_blinded = false;
2438 	bool extra_pass = false;
2439 	bool padding = false;
2440 	u8 *rw_image = NULL;
2441 	u8 *image = NULL;
2442 	int *addrs;
2443 	int pass;
2444 	int i;
2445 
2446 	if (!prog->jit_requested)
2447 		return orig_prog;
2448 
2449 	tmp = bpf_jit_blind_constants(prog);
2450 	/*
2451 	 * If blinding was requested and we failed during blinding,
2452 	 * we must fall back to the interpreter.
2453 	 */
2454 	if (IS_ERR(tmp))
2455 		return orig_prog;
2456 	if (tmp != prog) {
2457 		tmp_blinded = true;
2458 		prog = tmp;
2459 	}
2460 
2461 	jit_data = prog->aux->jit_data;
2462 	if (!jit_data) {
2463 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2464 		if (!jit_data) {
2465 			prog = orig_prog;
2466 			goto out;
2467 		}
2468 		prog->aux->jit_data = jit_data;
2469 	}
2470 	addrs = jit_data->addrs;
2471 	if (addrs) {
2472 		ctx = jit_data->ctx;
2473 		oldproglen = jit_data->proglen;
2474 		image = jit_data->image;
2475 		header = jit_data->header;
2476 		rw_header = jit_data->rw_header;
2477 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2478 		extra_pass = true;
2479 		padding = true;
2480 		goto skip_init_addrs;
2481 	}
2482 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2483 	if (!addrs) {
2484 		prog = orig_prog;
2485 		goto out_addrs;
2486 	}
2487 
2488 	/*
2489 	 * Before first pass, make a rough estimation of addrs[]
2490 	 * each BPF instruction is translated to less than 64 bytes
2491 	 */
2492 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2493 		proglen += 64;
2494 		addrs[i] = proglen;
2495 	}
2496 	ctx.cleanup_addr = proglen;
2497 skip_init_addrs:
2498 
2499 	/*
2500 	 * JITed image shrinks with every pass and the loop iterates
2501 	 * until the image stops shrinking. Very large BPF programs
2502 	 * may converge on the last pass. In such case do one more
2503 	 * pass to emit the final image.
2504 	 */
2505 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2506 		if (!padding && pass >= PADDING_PASSES)
2507 			padding = true;
2508 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2509 		if (proglen <= 0) {
2510 out_image:
2511 			image = NULL;
2512 			if (header) {
2513 				bpf_arch_text_copy(&header->size, &rw_header->size,
2514 						   sizeof(rw_header->size));
2515 				bpf_jit_binary_pack_free(header, rw_header);
2516 			}
2517 			/* Fall back to interpreter mode */
2518 			prog = orig_prog;
2519 			if (extra_pass) {
2520 				prog->bpf_func = NULL;
2521 				prog->jited = 0;
2522 				prog->jited_len = 0;
2523 			}
2524 			goto out_addrs;
2525 		}
2526 		if (image) {
2527 			if (proglen != oldproglen) {
2528 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2529 				       proglen, oldproglen);
2530 				goto out_image;
2531 			}
2532 			break;
2533 		}
2534 		if (proglen == oldproglen) {
2535 			/*
2536 			 * The number of entries in extable is the number of BPF_LDX
2537 			 * insns that access kernel memory via "pointer to BTF type".
2538 			 * The verifier changed their opcode from LDX|MEM|size
2539 			 * to LDX|PROBE_MEM|size to make JITing easier.
2540 			 */
2541 			u32 align = __alignof__(struct exception_table_entry);
2542 			u32 extable_size = prog->aux->num_exentries *
2543 				sizeof(struct exception_table_entry);
2544 
2545 			/* allocate module memory for x86 insns and extable */
2546 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2547 							   &image, align, &rw_header, &rw_image,
2548 							   jit_fill_hole);
2549 			if (!header) {
2550 				prog = orig_prog;
2551 				goto out_addrs;
2552 			}
2553 			prog->aux->extable = (void *) image + roundup(proglen, align);
2554 		}
2555 		oldproglen = proglen;
2556 		cond_resched();
2557 	}
2558 
2559 	if (bpf_jit_enable > 1)
2560 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2561 
2562 	if (image) {
2563 		if (!prog->is_func || extra_pass) {
2564 			/*
2565 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2566 			 *   1) header is not pointing to proper module memory;
2567 			 *   2) the arch doesn't support bpf_arch_text_copy().
2568 			 *
2569 			 * Both cases are serious bugs and justify WARN_ON.
2570 			 */
2571 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2572 				/* header has been freed */
2573 				header = NULL;
2574 				goto out_image;
2575 			}
2576 
2577 			bpf_tail_call_direct_fixup(prog);
2578 		} else {
2579 			jit_data->addrs = addrs;
2580 			jit_data->ctx = ctx;
2581 			jit_data->proglen = proglen;
2582 			jit_data->image = image;
2583 			jit_data->header = header;
2584 			jit_data->rw_header = rw_header;
2585 		}
2586 		prog->bpf_func = (void *)image;
2587 		prog->jited = 1;
2588 		prog->jited_len = proglen;
2589 	} else {
2590 		prog = orig_prog;
2591 	}
2592 
2593 	if (!image || !prog->is_func || extra_pass) {
2594 		if (image)
2595 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2596 out_addrs:
2597 		kvfree(addrs);
2598 		kfree(jit_data);
2599 		prog->aux->jit_data = NULL;
2600 	}
2601 out:
2602 	if (tmp_blinded)
2603 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2604 					   tmp : orig_prog);
2605 	return prog;
2606 }
2607 
2608 bool bpf_jit_supports_kfunc_call(void)
2609 {
2610 	return true;
2611 }
2612 
2613 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2614 {
2615 	if (text_poke_copy(dst, src, len) == NULL)
2616 		return ERR_PTR(-EINVAL);
2617 	return dst;
2618 }
2619 
2620 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2621 bool bpf_jit_supports_subprog_tailcalls(void)
2622 {
2623 	return true;
2624 }
2625 
2626 void bpf_jit_free(struct bpf_prog *prog)
2627 {
2628 	if (prog->jited) {
2629 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2630 		struct bpf_binary_header *hdr;
2631 
2632 		/*
2633 		 * If we fail the final pass of JIT (from jit_subprogs),
2634 		 * the program may not be finalized yet. Call finalize here
2635 		 * before freeing it.
2636 		 */
2637 		if (jit_data) {
2638 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2639 						     jit_data->rw_header);
2640 			kvfree(jit_data->addrs);
2641 			kfree(jit_data);
2642 		}
2643 		hdr = bpf_jit_binary_pack_hdr(prog);
2644 		bpf_jit_binary_pack_free(hdr, NULL);
2645 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2646 	}
2647 
2648 	bpf_prog_unlock_free(prog);
2649 }
2650