xref: /linux/arch/x86/net/bpf_jit_comp.c (revision f879306834818ebd1722a4372079610cdd466fec)
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 <asm/extable.h>
15 #include <asm/ftrace.h>
16 #include <asm/set_memory.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/text-patching.h>
19 #include <asm/unwind.h>
20 #include <asm/cfi.h>
21 
22 static bool all_callee_regs_used[4] = {true, true, true, true};
23 
24 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
25 {
26 	if (len == 1)
27 		*ptr = bytes;
28 	else if (len == 2)
29 		*(u16 *)ptr = bytes;
30 	else {
31 		*(u32 *)ptr = bytes;
32 		barrier();
33 	}
34 	return ptr + len;
35 }
36 
37 #define EMIT(bytes, len) \
38 	do { prog = emit_code(prog, bytes, len); } while (0)
39 
40 #define EMIT1(b1)		EMIT(b1, 1)
41 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
42 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
43 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
44 
45 #define EMIT1_off32(b1, off) \
46 	do { EMIT1(b1); EMIT(off, 4); } while (0)
47 #define EMIT2_off32(b1, b2, off) \
48 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
49 #define EMIT3_off32(b1, b2, b3, off) \
50 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
51 #define EMIT4_off32(b1, b2, b3, b4, off) \
52 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
53 
54 #ifdef CONFIG_X86_KERNEL_IBT
55 #define EMIT_ENDBR()		EMIT(gen_endbr(), 4)
56 #define EMIT_ENDBR_POISON()	EMIT(gen_endbr_poison(), 4)
57 #else
58 #define EMIT_ENDBR()
59 #define EMIT_ENDBR_POISON()
60 #endif
61 
62 static bool is_imm8(int value)
63 {
64 	return value <= 127 && value >= -128;
65 }
66 
67 static bool is_simm32(s64 value)
68 {
69 	return value == (s64)(s32)value;
70 }
71 
72 static bool is_uimm32(u64 value)
73 {
74 	return value == (u64)(u32)value;
75 }
76 
77 /* mov dst, src */
78 #define EMIT_mov(DST, SRC)								 \
79 	do {										 \
80 		if (DST != SRC)								 \
81 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
82 	} while (0)
83 
84 static int bpf_size_to_x86_bytes(int bpf_size)
85 {
86 	if (bpf_size == BPF_W)
87 		return 4;
88 	else if (bpf_size == BPF_H)
89 		return 2;
90 	else if (bpf_size == BPF_B)
91 		return 1;
92 	else if (bpf_size == BPF_DW)
93 		return 4; /* imm32 */
94 	else
95 		return 0;
96 }
97 
98 /*
99  * List of x86 cond jumps opcodes (. + s8)
100  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
101  */
102 #define X86_JB  0x72
103 #define X86_JAE 0x73
104 #define X86_JE  0x74
105 #define X86_JNE 0x75
106 #define X86_JBE 0x76
107 #define X86_JA  0x77
108 #define X86_JL  0x7C
109 #define X86_JGE 0x7D
110 #define X86_JLE 0x7E
111 #define X86_JG  0x7F
112 
113 /* Pick a register outside of BPF range for JIT internal work */
114 #define AUX_REG (MAX_BPF_JIT_REG + 1)
115 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
116 #define X86_REG_R12 (MAX_BPF_JIT_REG + 3)
117 
118 /*
119  * The following table maps BPF registers to x86-64 registers.
120  *
121  * x86-64 register R12 is unused, since if used as base address
122  * register in load/store instructions, it always needs an
123  * extra byte of encoding and is callee saved.
124  *
125  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
126  * trampoline. x86-64 register R10 is used for blinding (if enabled).
127  */
128 static const int reg2hex[] = {
129 	[BPF_REG_0] = 0,  /* RAX */
130 	[BPF_REG_1] = 7,  /* RDI */
131 	[BPF_REG_2] = 6,  /* RSI */
132 	[BPF_REG_3] = 2,  /* RDX */
133 	[BPF_REG_4] = 1,  /* RCX */
134 	[BPF_REG_5] = 0,  /* R8  */
135 	[BPF_REG_6] = 3,  /* RBX callee saved */
136 	[BPF_REG_7] = 5,  /* R13 callee saved */
137 	[BPF_REG_8] = 6,  /* R14 callee saved */
138 	[BPF_REG_9] = 7,  /* R15 callee saved */
139 	[BPF_REG_FP] = 5, /* RBP readonly */
140 	[BPF_REG_AX] = 2, /* R10 temp register */
141 	[AUX_REG] = 3,    /* R11 temp register */
142 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
143 	[X86_REG_R12] = 4, /* R12 callee saved */
144 };
145 
146 static const int reg2pt_regs[] = {
147 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
148 	[BPF_REG_1] = offsetof(struct pt_regs, di),
149 	[BPF_REG_2] = offsetof(struct pt_regs, si),
150 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
151 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
152 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
153 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
154 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
155 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
156 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
157 };
158 
159 /*
160  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
161  * which need extra byte of encoding.
162  * rax,rcx,...,rbp have simpler encoding
163  */
164 static bool is_ereg(u32 reg)
165 {
166 	return (1 << reg) & (BIT(BPF_REG_5) |
167 			     BIT(AUX_REG) |
168 			     BIT(BPF_REG_7) |
169 			     BIT(BPF_REG_8) |
170 			     BIT(BPF_REG_9) |
171 			     BIT(X86_REG_R9) |
172 			     BIT(X86_REG_R12) |
173 			     BIT(BPF_REG_AX));
174 }
175 
176 /*
177  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
178  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
179  * of encoding. al,cl,dl,bl have simpler encoding.
180  */
181 static bool is_ereg_8l(u32 reg)
182 {
183 	return is_ereg(reg) ||
184 	    (1 << reg) & (BIT(BPF_REG_1) |
185 			  BIT(BPF_REG_2) |
186 			  BIT(BPF_REG_FP));
187 }
188 
189 static bool is_axreg(u32 reg)
190 {
191 	return reg == BPF_REG_0;
192 }
193 
194 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
195 static u8 add_1mod(u8 byte, u32 reg)
196 {
197 	if (is_ereg(reg))
198 		byte |= 1;
199 	return byte;
200 }
201 
202 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
203 {
204 	if (is_ereg(r1))
205 		byte |= 1;
206 	if (is_ereg(r2))
207 		byte |= 4;
208 	return byte;
209 }
210 
211 static u8 add_3mod(u8 byte, u32 r1, u32 r2, u32 index)
212 {
213 	if (is_ereg(r1))
214 		byte |= 1;
215 	if (is_ereg(index))
216 		byte |= 2;
217 	if (is_ereg(r2))
218 		byte |= 4;
219 	return byte;
220 }
221 
222 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
223 static u8 add_1reg(u8 byte, u32 dst_reg)
224 {
225 	return byte + reg2hex[dst_reg];
226 }
227 
228 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
229 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
230 {
231 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
232 }
233 
234 /* Some 1-byte opcodes for binary ALU operations */
235 static u8 simple_alu_opcodes[] = {
236 	[BPF_ADD] = 0x01,
237 	[BPF_SUB] = 0x29,
238 	[BPF_AND] = 0x21,
239 	[BPF_OR] = 0x09,
240 	[BPF_XOR] = 0x31,
241 	[BPF_LSH] = 0xE0,
242 	[BPF_RSH] = 0xE8,
243 	[BPF_ARSH] = 0xF8,
244 };
245 
246 static void jit_fill_hole(void *area, unsigned int size)
247 {
248 	/* Fill whole space with INT3 instructions */
249 	memset(area, 0xcc, size);
250 }
251 
252 int bpf_arch_text_invalidate(void *dst, size_t len)
253 {
254 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
255 }
256 
257 struct jit_context {
258 	int cleanup_addr; /* Epilogue code offset */
259 
260 	/*
261 	 * Program specific offsets of labels in the code; these rely on the
262 	 * JIT doing at least 2 passes, recording the position on the first
263 	 * pass, only to generate the correct offset on the second pass.
264 	 */
265 	int tail_call_direct_label;
266 	int tail_call_indirect_label;
267 };
268 
269 /* Maximum number of bytes emitted while JITing one eBPF insn */
270 #define BPF_MAX_INSN_SIZE	128
271 #define BPF_INSN_SAFETY		64
272 
273 /* Number of bytes emit_patch() needs to generate instructions */
274 #define X86_PATCH_SIZE		5
275 /* Number of bytes that will be skipped on tailcall */
276 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
277 
278 static void push_r12(u8 **pprog)
279 {
280 	u8 *prog = *pprog;
281 
282 	EMIT2(0x41, 0x54);   /* push r12 */
283 	*pprog = prog;
284 }
285 
286 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
287 {
288 	u8 *prog = *pprog;
289 
290 	if (callee_regs_used[0])
291 		EMIT1(0x53);         /* push rbx */
292 	if (callee_regs_used[1])
293 		EMIT2(0x41, 0x55);   /* push r13 */
294 	if (callee_regs_used[2])
295 		EMIT2(0x41, 0x56);   /* push r14 */
296 	if (callee_regs_used[3])
297 		EMIT2(0x41, 0x57);   /* push r15 */
298 	*pprog = prog;
299 }
300 
301 static void pop_r12(u8 **pprog)
302 {
303 	u8 *prog = *pprog;
304 
305 	EMIT2(0x41, 0x5C);   /* pop r12 */
306 	*pprog = prog;
307 }
308 
309 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
310 {
311 	u8 *prog = *pprog;
312 
313 	if (callee_regs_used[3])
314 		EMIT2(0x41, 0x5F);   /* pop r15 */
315 	if (callee_regs_used[2])
316 		EMIT2(0x41, 0x5E);   /* pop r14 */
317 	if (callee_regs_used[1])
318 		EMIT2(0x41, 0x5D);   /* pop r13 */
319 	if (callee_regs_used[0])
320 		EMIT1(0x5B);         /* pop rbx */
321 	*pprog = prog;
322 }
323 
324 static void emit_nops(u8 **pprog, int len)
325 {
326 	u8 *prog = *pprog;
327 	int i, noplen;
328 
329 	while (len > 0) {
330 		noplen = len;
331 
332 		if (noplen > ASM_NOP_MAX)
333 			noplen = ASM_NOP_MAX;
334 
335 		for (i = 0; i < noplen; i++)
336 			EMIT1(x86_nops[noplen][i]);
337 		len -= noplen;
338 	}
339 
340 	*pprog = prog;
341 }
342 
343 /*
344  * Emit the various CFI preambles, see asm/cfi.h and the comments about FineIBT
345  * in arch/x86/kernel/alternative.c
346  */
347 
348 static void emit_fineibt(u8 **pprog, u32 hash)
349 {
350 	u8 *prog = *pprog;
351 
352 	EMIT_ENDBR();
353 	EMIT3_off32(0x41, 0x81, 0xea, hash);		/* subl $hash, %r10d	*/
354 	EMIT2(0x74, 0x07);				/* jz.d8 +7		*/
355 	EMIT2(0x0f, 0x0b);				/* ud2			*/
356 	EMIT1(0x90);					/* nop			*/
357 	EMIT_ENDBR_POISON();
358 
359 	*pprog = prog;
360 }
361 
362 static void emit_kcfi(u8 **pprog, u32 hash)
363 {
364 	u8 *prog = *pprog;
365 
366 	EMIT1_off32(0xb8, hash);			/* movl $hash, %eax	*/
367 #ifdef CONFIG_CALL_PADDING
368 	EMIT1(0x90);
369 	EMIT1(0x90);
370 	EMIT1(0x90);
371 	EMIT1(0x90);
372 	EMIT1(0x90);
373 	EMIT1(0x90);
374 	EMIT1(0x90);
375 	EMIT1(0x90);
376 	EMIT1(0x90);
377 	EMIT1(0x90);
378 	EMIT1(0x90);
379 #endif
380 	EMIT_ENDBR();
381 
382 	*pprog = prog;
383 }
384 
385 static void emit_cfi(u8 **pprog, u32 hash)
386 {
387 	u8 *prog = *pprog;
388 
389 	switch (cfi_mode) {
390 	case CFI_FINEIBT:
391 		emit_fineibt(&prog, hash);
392 		break;
393 
394 	case CFI_KCFI:
395 		emit_kcfi(&prog, hash);
396 		break;
397 
398 	default:
399 		EMIT_ENDBR();
400 		break;
401 	}
402 
403 	*pprog = prog;
404 }
405 
406 /*
407  * Emit x86-64 prologue code for BPF program.
408  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
409  * while jumping to another program
410  */
411 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
412 			  bool tail_call_reachable, bool is_subprog,
413 			  bool is_exception_cb)
414 {
415 	u8 *prog = *pprog;
416 
417 	emit_cfi(&prog, is_subprog ? cfi_bpf_subprog_hash : cfi_bpf_hash);
418 	/* BPF trampoline can be made to work without these nops,
419 	 * but let's waste 5 bytes for now and optimize later
420 	 */
421 	emit_nops(&prog, X86_PATCH_SIZE);
422 	if (!ebpf_from_cbpf) {
423 		if (tail_call_reachable && !is_subprog)
424 			/* When it's the entry of the whole tailcall context,
425 			 * zeroing rax means initialising tail_call_cnt.
426 			 */
427 			EMIT2(0x31, 0xC0); /* xor eax, eax */
428 		else
429 			/* Keep the same instruction layout. */
430 			EMIT2(0x66, 0x90); /* nop2 */
431 	}
432 	/* Exception callback receives FP as third parameter */
433 	if (is_exception_cb) {
434 		EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */
435 		EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */
436 		/* The main frame must have exception_boundary as true, so we
437 		 * first restore those callee-saved regs from stack, before
438 		 * reusing the stack frame.
439 		 */
440 		pop_callee_regs(&prog, all_callee_regs_used);
441 		pop_r12(&prog);
442 		/* Reset the stack frame. */
443 		EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */
444 	} else {
445 		EMIT1(0x55);             /* push rbp */
446 		EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
447 	}
448 
449 	/* X86_TAIL_CALL_OFFSET is here */
450 	EMIT_ENDBR();
451 
452 	/* sub rsp, rounded_stack_depth */
453 	if (stack_depth)
454 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
455 	if (tail_call_reachable)
456 		EMIT1(0x50);         /* push rax */
457 	*pprog = prog;
458 }
459 
460 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
461 {
462 	u8 *prog = *pprog;
463 	s64 offset;
464 
465 	offset = func - (ip + X86_PATCH_SIZE);
466 	if (!is_simm32(offset)) {
467 		pr_err("Target call %p is out of range\n", func);
468 		return -ERANGE;
469 	}
470 	EMIT1_off32(opcode, offset);
471 	*pprog = prog;
472 	return 0;
473 }
474 
475 static int emit_call(u8 **pprog, void *func, void *ip)
476 {
477 	return emit_patch(pprog, func, ip, 0xE8);
478 }
479 
480 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
481 {
482 	OPTIMIZER_HIDE_VAR(func);
483 	ip += x86_call_depth_emit_accounting(pprog, func, ip);
484 	return emit_patch(pprog, func, ip, 0xE8);
485 }
486 
487 static int emit_jump(u8 **pprog, void *func, void *ip)
488 {
489 	return emit_patch(pprog, func, ip, 0xE9);
490 }
491 
492 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
493 				void *old_addr, void *new_addr)
494 {
495 	const u8 *nop_insn = x86_nops[5];
496 	u8 old_insn[X86_PATCH_SIZE];
497 	u8 new_insn[X86_PATCH_SIZE];
498 	u8 *prog;
499 	int ret;
500 
501 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
502 	if (old_addr) {
503 		prog = old_insn;
504 		ret = t == BPF_MOD_CALL ?
505 		      emit_call(&prog, old_addr, ip) :
506 		      emit_jump(&prog, old_addr, ip);
507 		if (ret)
508 			return ret;
509 	}
510 
511 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
512 	if (new_addr) {
513 		prog = new_insn;
514 		ret = t == BPF_MOD_CALL ?
515 		      emit_call(&prog, new_addr, ip) :
516 		      emit_jump(&prog, new_addr, ip);
517 		if (ret)
518 			return ret;
519 	}
520 
521 	ret = -EBUSY;
522 	mutex_lock(&text_mutex);
523 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
524 		goto out;
525 	ret = 1;
526 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
527 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
528 		ret = 0;
529 	}
530 out:
531 	mutex_unlock(&text_mutex);
532 	return ret;
533 }
534 
535 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
536 		       void *old_addr, void *new_addr)
537 {
538 	if (!is_kernel_text((long)ip) &&
539 	    !is_bpf_text_address((long)ip))
540 		/* BPF poking in modules is not supported */
541 		return -EINVAL;
542 
543 	/*
544 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
545 	 * with an ENDBR instruction.
546 	 */
547 	if (is_endbr(*(u32 *)ip))
548 		ip += ENDBR_INSN_SIZE;
549 
550 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
551 }
552 
553 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
554 
555 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
556 {
557 	u8 *prog = *pprog;
558 
559 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
560 		EMIT_LFENCE();
561 		EMIT2(0xFF, 0xE0 + reg);
562 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
563 		OPTIMIZER_HIDE_VAR(reg);
564 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
565 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
566 		else
567 			emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
568 	} else {
569 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
570 		if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) || IS_ENABLED(CONFIG_MITIGATION_SLS))
571 			EMIT1(0xCC);		/* int3 */
572 	}
573 
574 	*pprog = prog;
575 }
576 
577 static void emit_return(u8 **pprog, u8 *ip)
578 {
579 	u8 *prog = *pprog;
580 
581 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
582 		emit_jump(&prog, x86_return_thunk, ip);
583 	} else {
584 		EMIT1(0xC3);		/* ret */
585 		if (IS_ENABLED(CONFIG_MITIGATION_SLS))
586 			EMIT1(0xCC);	/* int3 */
587 	}
588 
589 	*pprog = prog;
590 }
591 
592 /*
593  * Generate the following code:
594  *
595  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
596  *   if (index >= array->map.max_entries)
597  *     goto out;
598  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
599  *     goto out;
600  *   prog = array->ptrs[index];
601  *   if (prog == NULL)
602  *     goto out;
603  *   goto *(prog->bpf_func + prologue_size);
604  * out:
605  */
606 static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog,
607 					u8 **pprog, bool *callee_regs_used,
608 					u32 stack_depth, u8 *ip,
609 					struct jit_context *ctx)
610 {
611 	int tcc_off = -4 - round_up(stack_depth, 8);
612 	u8 *prog = *pprog, *start = *pprog;
613 	int offset;
614 
615 	/*
616 	 * rdi - pointer to ctx
617 	 * rsi - pointer to bpf_array
618 	 * rdx - index in bpf_array
619 	 */
620 
621 	/*
622 	 * if (index >= array->map.max_entries)
623 	 *	goto out;
624 	 */
625 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
626 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
627 	      offsetof(struct bpf_array, map.max_entries));
628 
629 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
630 	EMIT2(X86_JBE, offset);                   /* jbe out */
631 
632 	/*
633 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
634 	 *	goto out;
635 	 */
636 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
637 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
638 
639 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
640 	EMIT2(X86_JAE, offset);                   /* jae out */
641 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
642 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
643 
644 	/* prog = array->ptrs[index]; */
645 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
646 		    offsetof(struct bpf_array, ptrs));
647 
648 	/*
649 	 * if (prog == NULL)
650 	 *	goto out;
651 	 */
652 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
653 
654 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
655 	EMIT2(X86_JE, offset);                    /* je out */
656 
657 	if (bpf_prog->aux->exception_boundary) {
658 		pop_callee_regs(&prog, all_callee_regs_used);
659 		pop_r12(&prog);
660 	} else {
661 		pop_callee_regs(&prog, callee_regs_used);
662 		if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
663 			pop_r12(&prog);
664 	}
665 
666 	EMIT1(0x58);                              /* pop rax */
667 	if (stack_depth)
668 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
669 			    round_up(stack_depth, 8));
670 
671 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
672 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
673 	      offsetof(struct bpf_prog, bpf_func));
674 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
675 	      X86_TAIL_CALL_OFFSET);
676 	/*
677 	 * Now we're ready to jump into next BPF program
678 	 * rdi == ctx (1st arg)
679 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
680 	 */
681 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
682 
683 	/* out: */
684 	ctx->tail_call_indirect_label = prog - start;
685 	*pprog = prog;
686 }
687 
688 static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog,
689 				      struct bpf_jit_poke_descriptor *poke,
690 				      u8 **pprog, u8 *ip,
691 				      bool *callee_regs_used, u32 stack_depth,
692 				      struct jit_context *ctx)
693 {
694 	int tcc_off = -4 - round_up(stack_depth, 8);
695 	u8 *prog = *pprog, *start = *pprog;
696 	int offset;
697 
698 	/*
699 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
700 	 *	goto out;
701 	 */
702 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
703 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
704 
705 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
706 	EMIT2(X86_JAE, offset);                       /* jae out */
707 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
708 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
709 
710 	poke->tailcall_bypass = ip + (prog - start);
711 	poke->adj_off = X86_TAIL_CALL_OFFSET;
712 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
713 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
714 
715 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
716 		  poke->tailcall_bypass);
717 
718 	if (bpf_prog->aux->exception_boundary) {
719 		pop_callee_regs(&prog, all_callee_regs_used);
720 		pop_r12(&prog);
721 	} else {
722 		pop_callee_regs(&prog, callee_regs_used);
723 		if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
724 			pop_r12(&prog);
725 	}
726 
727 	EMIT1(0x58);                                  /* pop rax */
728 	if (stack_depth)
729 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
730 
731 	emit_nops(&prog, X86_PATCH_SIZE);
732 
733 	/* out: */
734 	ctx->tail_call_direct_label = prog - start;
735 
736 	*pprog = prog;
737 }
738 
739 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
740 {
741 	struct bpf_jit_poke_descriptor *poke;
742 	struct bpf_array *array;
743 	struct bpf_prog *target;
744 	int i, ret;
745 
746 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
747 		poke = &prog->aux->poke_tab[i];
748 		if (poke->aux && poke->aux != prog->aux)
749 			continue;
750 
751 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
752 
753 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
754 			continue;
755 
756 		array = container_of(poke->tail_call.map, struct bpf_array, map);
757 		mutex_lock(&array->aux->poke_mutex);
758 		target = array->ptrs[poke->tail_call.key];
759 		if (target) {
760 			ret = __bpf_arch_text_poke(poke->tailcall_target,
761 						   BPF_MOD_JUMP, NULL,
762 						   (u8 *)target->bpf_func +
763 						   poke->adj_off);
764 			BUG_ON(ret < 0);
765 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
766 						   BPF_MOD_JUMP,
767 						   (u8 *)poke->tailcall_target +
768 						   X86_PATCH_SIZE, NULL);
769 			BUG_ON(ret < 0);
770 		}
771 		WRITE_ONCE(poke->tailcall_target_stable, true);
772 		mutex_unlock(&array->aux->poke_mutex);
773 	}
774 }
775 
776 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
777 			   u32 dst_reg, const u32 imm32)
778 {
779 	u8 *prog = *pprog;
780 	u8 b1, b2, b3;
781 
782 	/*
783 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
784 	 * (which zero-extends imm32) to save 2 bytes.
785 	 */
786 	if (sign_propagate && (s32)imm32 < 0) {
787 		/* 'mov %rax, imm32' sign extends imm32 */
788 		b1 = add_1mod(0x48, dst_reg);
789 		b2 = 0xC7;
790 		b3 = 0xC0;
791 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
792 		goto done;
793 	}
794 
795 	/*
796 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
797 	 * to save 3 bytes.
798 	 */
799 	if (imm32 == 0) {
800 		if (is_ereg(dst_reg))
801 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
802 		b2 = 0x31; /* xor */
803 		b3 = 0xC0;
804 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
805 		goto done;
806 	}
807 
808 	/* mov %eax, imm32 */
809 	if (is_ereg(dst_reg))
810 		EMIT1(add_1mod(0x40, dst_reg));
811 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
812 done:
813 	*pprog = prog;
814 }
815 
816 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
817 			   const u32 imm32_hi, const u32 imm32_lo)
818 {
819 	u64 imm64 = ((u64)imm32_hi << 32) | (u32)imm32_lo;
820 	u8 *prog = *pprog;
821 
822 	if (is_uimm32(imm64)) {
823 		/*
824 		 * For emitting plain u32, where sign bit must not be
825 		 * propagated LLVM tends to load imm64 over mov32
826 		 * directly, so save couple of bytes by just doing
827 		 * 'mov %eax, imm32' instead.
828 		 */
829 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
830 	} else if (is_simm32(imm64)) {
831 		emit_mov_imm32(&prog, true, dst_reg, imm32_lo);
832 	} else {
833 		/* movabsq rax, imm64 */
834 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
835 		EMIT(imm32_lo, 4);
836 		EMIT(imm32_hi, 4);
837 	}
838 
839 	*pprog = prog;
840 }
841 
842 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
843 {
844 	u8 *prog = *pprog;
845 
846 	if (is64) {
847 		/* mov dst, src */
848 		EMIT_mov(dst_reg, src_reg);
849 	} else {
850 		/* mov32 dst, src */
851 		if (is_ereg(dst_reg) || is_ereg(src_reg))
852 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
853 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
854 	}
855 
856 	*pprog = prog;
857 }
858 
859 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
860 			   u32 src_reg)
861 {
862 	u8 *prog = *pprog;
863 
864 	if (is64) {
865 		/* movs[b,w,l]q dst, src */
866 		if (num_bits == 8)
867 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
868 			      add_2reg(0xC0, src_reg, dst_reg));
869 		else if (num_bits == 16)
870 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
871 			      add_2reg(0xC0, src_reg, dst_reg));
872 		else if (num_bits == 32)
873 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
874 			      add_2reg(0xC0, src_reg, dst_reg));
875 	} else {
876 		/* movs[b,w]l dst, src */
877 		if (num_bits == 8) {
878 			EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
879 			      add_2reg(0xC0, src_reg, dst_reg));
880 		} else if (num_bits == 16) {
881 			if (is_ereg(dst_reg) || is_ereg(src_reg))
882 				EMIT1(add_2mod(0x40, src_reg, dst_reg));
883 			EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
884 			      add_2reg(0xC0, src_reg, dst_reg));
885 		}
886 	}
887 
888 	*pprog = prog;
889 }
890 
891 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
892 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
893 {
894 	u8 *prog = *pprog;
895 
896 	if (is_imm8(off)) {
897 		/* 1-byte signed displacement.
898 		 *
899 		 * If off == 0 we could skip this and save one extra byte, but
900 		 * special case of x86 R13 which always needs an offset is not
901 		 * worth the hassle
902 		 */
903 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
904 	} else {
905 		/* 4-byte signed displacement */
906 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
907 	}
908 	*pprog = prog;
909 }
910 
911 static void emit_insn_suffix_SIB(u8 **pprog, u32 ptr_reg, u32 val_reg, u32 index_reg, int off)
912 {
913 	u8 *prog = *pprog;
914 
915 	if (is_imm8(off)) {
916 		EMIT3(add_2reg(0x44, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
917 	} else {
918 		EMIT2_off32(add_2reg(0x84, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
919 	}
920 	*pprog = prog;
921 }
922 
923 /*
924  * Emit a REX byte if it will be necessary to address these registers
925  */
926 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
927 {
928 	u8 *prog = *pprog;
929 
930 	if (is64)
931 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
932 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
933 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
934 	*pprog = prog;
935 }
936 
937 /*
938  * Similar version of maybe_emit_mod() for a single register
939  */
940 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
941 {
942 	u8 *prog = *pprog;
943 
944 	if (is64)
945 		EMIT1(add_1mod(0x48, reg));
946 	else if (is_ereg(reg))
947 		EMIT1(add_1mod(0x40, reg));
948 	*pprog = prog;
949 }
950 
951 /* LDX: dst_reg = *(u8*)(src_reg + off) */
952 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
953 {
954 	u8 *prog = *pprog;
955 
956 	switch (size) {
957 	case BPF_B:
958 		/* Emit 'movzx rax, byte ptr [rax + off]' */
959 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
960 		break;
961 	case BPF_H:
962 		/* Emit 'movzx rax, word ptr [rax + off]' */
963 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
964 		break;
965 	case BPF_W:
966 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
967 		if (is_ereg(dst_reg) || is_ereg(src_reg))
968 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
969 		else
970 			EMIT1(0x8B);
971 		break;
972 	case BPF_DW:
973 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
974 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
975 		break;
976 	}
977 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
978 	*pprog = prog;
979 }
980 
981 /* LDSX: dst_reg = *(s8*)(src_reg + off) */
982 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
983 {
984 	u8 *prog = *pprog;
985 
986 	switch (size) {
987 	case BPF_B:
988 		/* Emit 'movsx rax, byte ptr [rax + off]' */
989 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
990 		break;
991 	case BPF_H:
992 		/* Emit 'movsx rax, word ptr [rax + off]' */
993 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
994 		break;
995 	case BPF_W:
996 		/* Emit 'movsx rax, dword ptr [rax+0x14]' */
997 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
998 		break;
999 	}
1000 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
1001 	*pprog = prog;
1002 }
1003 
1004 static void emit_ldx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1005 {
1006 	u8 *prog = *pprog;
1007 
1008 	switch (size) {
1009 	case BPF_B:
1010 		/* movzx rax, byte ptr [rax + r12 + off] */
1011 		EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB6);
1012 		break;
1013 	case BPF_H:
1014 		/* movzx rax, word ptr [rax + r12 + off] */
1015 		EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB7);
1016 		break;
1017 	case BPF_W:
1018 		/* mov eax, dword ptr [rax + r12 + off] */
1019 		EMIT2(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x8B);
1020 		break;
1021 	case BPF_DW:
1022 		/* mov rax, qword ptr [rax + r12 + off] */
1023 		EMIT2(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x8B);
1024 		break;
1025 	}
1026 	emit_insn_suffix_SIB(&prog, src_reg, dst_reg, index_reg, off);
1027 	*pprog = prog;
1028 }
1029 
1030 static void emit_ldx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1031 {
1032 	emit_ldx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
1033 }
1034 
1035 /* STX: *(u8*)(dst_reg + off) = src_reg */
1036 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1037 {
1038 	u8 *prog = *pprog;
1039 
1040 	switch (size) {
1041 	case BPF_B:
1042 		/* Emit 'mov byte ptr [rax + off], al' */
1043 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
1044 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
1045 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
1046 		else
1047 			EMIT1(0x88);
1048 		break;
1049 	case BPF_H:
1050 		if (is_ereg(dst_reg) || is_ereg(src_reg))
1051 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
1052 		else
1053 			EMIT2(0x66, 0x89);
1054 		break;
1055 	case BPF_W:
1056 		if (is_ereg(dst_reg) || is_ereg(src_reg))
1057 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
1058 		else
1059 			EMIT1(0x89);
1060 		break;
1061 	case BPF_DW:
1062 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
1063 		break;
1064 	}
1065 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1066 	*pprog = prog;
1067 }
1068 
1069 /* STX: *(u8*)(dst_reg + index_reg + off) = src_reg */
1070 static void emit_stx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1071 {
1072 	u8 *prog = *pprog;
1073 
1074 	switch (size) {
1075 	case BPF_B:
1076 		/* mov byte ptr [rax + r12 + off], al */
1077 		EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x88);
1078 		break;
1079 	case BPF_H:
1080 		/* mov word ptr [rax + r12 + off], ax */
1081 		EMIT3(0x66, add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
1082 		break;
1083 	case BPF_W:
1084 		/* mov dword ptr [rax + r12 + 1], eax */
1085 		EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
1086 		break;
1087 	case BPF_DW:
1088 		/* mov qword ptr [rax + r12 + 1], rax */
1089 		EMIT2(add_3mod(0x48, dst_reg, src_reg, index_reg), 0x89);
1090 		break;
1091 	}
1092 	emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off);
1093 	*pprog = prog;
1094 }
1095 
1096 static void emit_stx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1097 {
1098 	emit_stx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
1099 }
1100 
1101 /* ST: *(u8*)(dst_reg + index_reg + off) = imm32 */
1102 static void emit_st_index(u8 **pprog, u32 size, u32 dst_reg, u32 index_reg, int off, int imm)
1103 {
1104 	u8 *prog = *pprog;
1105 
1106 	switch (size) {
1107 	case BPF_B:
1108 		/* mov byte ptr [rax + r12 + off], imm8 */
1109 		EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC6);
1110 		break;
1111 	case BPF_H:
1112 		/* mov word ptr [rax + r12 + off], imm16 */
1113 		EMIT3(0x66, add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
1114 		break;
1115 	case BPF_W:
1116 		/* mov dword ptr [rax + r12 + 1], imm32 */
1117 		EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
1118 		break;
1119 	case BPF_DW:
1120 		/* mov qword ptr [rax + r12 + 1], imm32 */
1121 		EMIT2(add_3mod(0x48, dst_reg, 0, index_reg), 0xC7);
1122 		break;
1123 	}
1124 	emit_insn_suffix_SIB(&prog, dst_reg, 0, index_reg, off);
1125 	EMIT(imm, bpf_size_to_x86_bytes(size));
1126 	*pprog = prog;
1127 }
1128 
1129 static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
1130 {
1131 	emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
1132 }
1133 
1134 static int emit_atomic(u8 **pprog, u8 atomic_op,
1135 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
1136 {
1137 	u8 *prog = *pprog;
1138 
1139 	EMIT1(0xF0); /* lock prefix */
1140 
1141 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
1142 
1143 	/* emit opcode */
1144 	switch (atomic_op) {
1145 	case BPF_ADD:
1146 	case BPF_AND:
1147 	case BPF_OR:
1148 	case BPF_XOR:
1149 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
1150 		EMIT1(simple_alu_opcodes[atomic_op]);
1151 		break;
1152 	case BPF_ADD | BPF_FETCH:
1153 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
1154 		EMIT2(0x0F, 0xC1);
1155 		break;
1156 	case BPF_XCHG:
1157 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
1158 		EMIT1(0x87);
1159 		break;
1160 	case BPF_CMPXCHG:
1161 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
1162 		EMIT2(0x0F, 0xB1);
1163 		break;
1164 	default:
1165 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
1166 		return -EFAULT;
1167 	}
1168 
1169 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1170 
1171 	*pprog = prog;
1172 	return 0;
1173 }
1174 
1175 static int emit_atomic_index(u8 **pprog, u8 atomic_op, u32 size,
1176 			     u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1177 {
1178 	u8 *prog = *pprog;
1179 
1180 	EMIT1(0xF0); /* lock prefix */
1181 	switch (size) {
1182 	case BPF_W:
1183 		EMIT1(add_3mod(0x40, dst_reg, src_reg, index_reg));
1184 		break;
1185 	case BPF_DW:
1186 		EMIT1(add_3mod(0x48, dst_reg, src_reg, index_reg));
1187 		break;
1188 	default:
1189 		pr_err("bpf_jit: 1 and 2 byte atomics are not supported\n");
1190 		return -EFAULT;
1191 	}
1192 
1193 	/* emit opcode */
1194 	switch (atomic_op) {
1195 	case BPF_ADD:
1196 	case BPF_AND:
1197 	case BPF_OR:
1198 	case BPF_XOR:
1199 		/* lock *(u32/u64*)(dst_reg + idx_reg + off) <op>= src_reg */
1200 		EMIT1(simple_alu_opcodes[atomic_op]);
1201 		break;
1202 	case BPF_ADD | BPF_FETCH:
1203 		/* src_reg = atomic_fetch_add(dst_reg + idx_reg + off, src_reg); */
1204 		EMIT2(0x0F, 0xC1);
1205 		break;
1206 	case BPF_XCHG:
1207 		/* src_reg = atomic_xchg(dst_reg + idx_reg + off, src_reg); */
1208 		EMIT1(0x87);
1209 		break;
1210 	case BPF_CMPXCHG:
1211 		/* r0 = atomic_cmpxchg(dst_reg + idx_reg + off, r0, src_reg); */
1212 		EMIT2(0x0F, 0xB1);
1213 		break;
1214 	default:
1215 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
1216 		return -EFAULT;
1217 	}
1218 	emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off);
1219 	*pprog = prog;
1220 	return 0;
1221 }
1222 
1223 #define DONT_CLEAR 1
1224 
1225 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
1226 {
1227 	u32 reg = x->fixup >> 8;
1228 
1229 	/* jump over faulting load and clear dest register */
1230 	if (reg != DONT_CLEAR)
1231 		*(unsigned long *)((void *)regs + reg) = 0;
1232 	regs->ip += x->fixup & 0xff;
1233 	return true;
1234 }
1235 
1236 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
1237 			     bool *regs_used)
1238 {
1239 	int i;
1240 
1241 	for (i = 1; i <= insn_cnt; i++, insn++) {
1242 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
1243 			regs_used[0] = true;
1244 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
1245 			regs_used[1] = true;
1246 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
1247 			regs_used[2] = true;
1248 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
1249 			regs_used[3] = true;
1250 	}
1251 }
1252 
1253 /* emit the 3-byte VEX prefix
1254  *
1255  * r: same as rex.r, extra bit for ModRM reg field
1256  * x: same as rex.x, extra bit for SIB index field
1257  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
1258  * m: opcode map select, encoding escape bytes e.g. 0x0f38
1259  * w: same as rex.w (32 bit or 64 bit) or opcode specific
1260  * src_reg2: additional source reg (encoded as BPF reg)
1261  * l: vector length (128 bit or 256 bit) or reserved
1262  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
1263  */
1264 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
1265 		      bool w, u8 src_reg2, bool l, u8 pp)
1266 {
1267 	u8 *prog = *pprog;
1268 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
1269 	u8 b1, b2;
1270 	u8 vvvv = reg2hex[src_reg2];
1271 
1272 	/* reg2hex gives only the lower 3 bit of vvvv */
1273 	if (is_ereg(src_reg2))
1274 		vvvv |= 1 << 3;
1275 
1276 	/*
1277 	 * 2nd byte of 3-byte VEX prefix
1278 	 * ~ means bit inverted encoding
1279 	 *
1280 	 *    7                           0
1281 	 *  +---+---+---+---+---+---+---+---+
1282 	 *  |~R |~X |~B |         m         |
1283 	 *  +---+---+---+---+---+---+---+---+
1284 	 */
1285 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
1286 	/*
1287 	 * 3rd byte of 3-byte VEX prefix
1288 	 *
1289 	 *    7                           0
1290 	 *  +---+---+---+---+---+---+---+---+
1291 	 *  | W |     ~vvvv     | L |   pp  |
1292 	 *  +---+---+---+---+---+---+---+---+
1293 	 */
1294 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1295 
1296 	EMIT3(b0, b1, b2);
1297 	*pprog = prog;
1298 }
1299 
1300 /* emit BMI2 shift instruction */
1301 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1302 {
1303 	u8 *prog = *pprog;
1304 	bool r = is_ereg(dst_reg);
1305 	u8 m = 2; /* escape code 0f38 */
1306 
1307 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1308 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1309 	*pprog = prog;
1310 }
1311 
1312 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1313 
1314 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1315 #define RESTORE_TAIL_CALL_CNT(stack)				\
1316 	EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
1317 
1318 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
1319 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
1320 {
1321 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1322 	struct bpf_insn *insn = bpf_prog->insnsi;
1323 	bool callee_regs_used[4] = {};
1324 	int insn_cnt = bpf_prog->len;
1325 	bool seen_exit = false;
1326 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1327 	u64 arena_vm_start, user_vm_start;
1328 	int i, excnt = 0;
1329 	int ilen, proglen = 0;
1330 	u8 *prog = temp;
1331 	int err;
1332 
1333 	arena_vm_start = bpf_arena_get_kern_vm_start(bpf_prog->aux->arena);
1334 	user_vm_start = bpf_arena_get_user_vm_start(bpf_prog->aux->arena);
1335 
1336 	detect_reg_usage(insn, insn_cnt, callee_regs_used);
1337 
1338 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
1339 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1340 		      bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
1341 	/* Exception callback will clobber callee regs for its own use, and
1342 	 * restore the original callee regs from main prog's stack frame.
1343 	 */
1344 	if (bpf_prog->aux->exception_boundary) {
1345 		/* We also need to save r12, which is not mapped to any BPF
1346 		 * register, as we throw after entry into the kernel, which may
1347 		 * overwrite r12.
1348 		 */
1349 		push_r12(&prog);
1350 		push_callee_regs(&prog, all_callee_regs_used);
1351 	} else {
1352 		if (arena_vm_start)
1353 			push_r12(&prog);
1354 		push_callee_regs(&prog, callee_regs_used);
1355 	}
1356 	if (arena_vm_start)
1357 		emit_mov_imm64(&prog, X86_REG_R12,
1358 			       arena_vm_start >> 32, (u32) arena_vm_start);
1359 
1360 	ilen = prog - temp;
1361 	if (rw_image)
1362 		memcpy(rw_image + proglen, temp, ilen);
1363 	proglen += ilen;
1364 	addrs[0] = proglen;
1365 	prog = temp;
1366 
1367 	for (i = 1; i <= insn_cnt; i++, insn++) {
1368 		const s32 imm32 = insn->imm;
1369 		u32 dst_reg = insn->dst_reg;
1370 		u32 src_reg = insn->src_reg;
1371 		u8 b2 = 0, b3 = 0;
1372 		u8 *start_of_ldx;
1373 		s64 jmp_offset;
1374 		s16 insn_off;
1375 		u8 jmp_cond;
1376 		u8 *func;
1377 		int nops;
1378 
1379 		switch (insn->code) {
1380 			/* ALU */
1381 		case BPF_ALU | BPF_ADD | BPF_X:
1382 		case BPF_ALU | BPF_SUB | BPF_X:
1383 		case BPF_ALU | BPF_AND | BPF_X:
1384 		case BPF_ALU | BPF_OR | BPF_X:
1385 		case BPF_ALU | BPF_XOR | BPF_X:
1386 		case BPF_ALU64 | BPF_ADD | BPF_X:
1387 		case BPF_ALU64 | BPF_SUB | BPF_X:
1388 		case BPF_ALU64 | BPF_AND | BPF_X:
1389 		case BPF_ALU64 | BPF_OR | BPF_X:
1390 		case BPF_ALU64 | BPF_XOR | BPF_X:
1391 			maybe_emit_mod(&prog, dst_reg, src_reg,
1392 				       BPF_CLASS(insn->code) == BPF_ALU64);
1393 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1394 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1395 			break;
1396 
1397 		case BPF_ALU64 | BPF_MOV | BPF_X:
1398 			if (insn_is_cast_user(insn)) {
1399 				if (dst_reg != src_reg)
1400 					/* 32-bit mov */
1401 					emit_mov_reg(&prog, false, dst_reg, src_reg);
1402 				/* shl dst_reg, 32 */
1403 				maybe_emit_1mod(&prog, dst_reg, true);
1404 				EMIT3(0xC1, add_1reg(0xE0, dst_reg), 32);
1405 
1406 				/* or dst_reg, user_vm_start */
1407 				maybe_emit_1mod(&prog, dst_reg, true);
1408 				if (is_axreg(dst_reg))
1409 					EMIT1_off32(0x0D,  user_vm_start >> 32);
1410 				else
1411 					EMIT2_off32(0x81, add_1reg(0xC8, dst_reg),  user_vm_start >> 32);
1412 
1413 				/* rol dst_reg, 32 */
1414 				maybe_emit_1mod(&prog, dst_reg, true);
1415 				EMIT3(0xC1, add_1reg(0xC0, dst_reg), 32);
1416 
1417 				/* xor r11, r11 */
1418 				EMIT3(0x4D, 0x31, 0xDB);
1419 
1420 				/* test dst_reg32, dst_reg32; check if lower 32-bit are zero */
1421 				maybe_emit_mod(&prog, dst_reg, dst_reg, false);
1422 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1423 
1424 				/* cmove r11, dst_reg; if so, set dst_reg to zero */
1425 				/* WARNING: Intel swapped src/dst register encoding in CMOVcc !!! */
1426 				maybe_emit_mod(&prog, AUX_REG, dst_reg, true);
1427 				EMIT3(0x0F, 0x44, add_2reg(0xC0, AUX_REG, dst_reg));
1428 				break;
1429 			} else if (insn_is_mov_percpu_addr(insn)) {
1430 				/* mov <dst>, <src> (if necessary) */
1431 				EMIT_mov(dst_reg, src_reg);
1432 #ifdef CONFIG_SMP
1433 				/* add <dst>, gs:[<off>] */
1434 				EMIT2(0x65, add_1mod(0x48, dst_reg));
1435 				EMIT3(0x03, add_2reg(0x04, 0, dst_reg), 0x25);
1436 				EMIT((u32)(unsigned long)&this_cpu_off, 4);
1437 #endif
1438 				break;
1439 			}
1440 			fallthrough;
1441 		case BPF_ALU | BPF_MOV | BPF_X:
1442 			if (insn->off == 0)
1443 				emit_mov_reg(&prog,
1444 					     BPF_CLASS(insn->code) == BPF_ALU64,
1445 					     dst_reg, src_reg);
1446 			else
1447 				emit_movsx_reg(&prog, insn->off,
1448 					       BPF_CLASS(insn->code) == BPF_ALU64,
1449 					       dst_reg, src_reg);
1450 			break;
1451 
1452 			/* neg dst */
1453 		case BPF_ALU | BPF_NEG:
1454 		case BPF_ALU64 | BPF_NEG:
1455 			maybe_emit_1mod(&prog, dst_reg,
1456 					BPF_CLASS(insn->code) == BPF_ALU64);
1457 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1458 			break;
1459 
1460 		case BPF_ALU | BPF_ADD | BPF_K:
1461 		case BPF_ALU | BPF_SUB | BPF_K:
1462 		case BPF_ALU | BPF_AND | BPF_K:
1463 		case BPF_ALU | BPF_OR | BPF_K:
1464 		case BPF_ALU | BPF_XOR | BPF_K:
1465 		case BPF_ALU64 | BPF_ADD | BPF_K:
1466 		case BPF_ALU64 | BPF_SUB | BPF_K:
1467 		case BPF_ALU64 | BPF_AND | BPF_K:
1468 		case BPF_ALU64 | BPF_OR | BPF_K:
1469 		case BPF_ALU64 | BPF_XOR | BPF_K:
1470 			maybe_emit_1mod(&prog, dst_reg,
1471 					BPF_CLASS(insn->code) == BPF_ALU64);
1472 
1473 			/*
1474 			 * b3 holds 'normal' opcode, b2 short form only valid
1475 			 * in case dst is eax/rax.
1476 			 */
1477 			switch (BPF_OP(insn->code)) {
1478 			case BPF_ADD:
1479 				b3 = 0xC0;
1480 				b2 = 0x05;
1481 				break;
1482 			case BPF_SUB:
1483 				b3 = 0xE8;
1484 				b2 = 0x2D;
1485 				break;
1486 			case BPF_AND:
1487 				b3 = 0xE0;
1488 				b2 = 0x25;
1489 				break;
1490 			case BPF_OR:
1491 				b3 = 0xC8;
1492 				b2 = 0x0D;
1493 				break;
1494 			case BPF_XOR:
1495 				b3 = 0xF0;
1496 				b2 = 0x35;
1497 				break;
1498 			}
1499 
1500 			if (is_imm8(imm32))
1501 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1502 			else if (is_axreg(dst_reg))
1503 				EMIT1_off32(b2, imm32);
1504 			else
1505 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1506 			break;
1507 
1508 		case BPF_ALU64 | BPF_MOV | BPF_K:
1509 		case BPF_ALU | BPF_MOV | BPF_K:
1510 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1511 				       dst_reg, imm32);
1512 			break;
1513 
1514 		case BPF_LD | BPF_IMM | BPF_DW:
1515 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1516 			insn++;
1517 			i++;
1518 			break;
1519 
1520 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1521 		case BPF_ALU | BPF_MOD | BPF_X:
1522 		case BPF_ALU | BPF_DIV | BPF_X:
1523 		case BPF_ALU | BPF_MOD | BPF_K:
1524 		case BPF_ALU | BPF_DIV | BPF_K:
1525 		case BPF_ALU64 | BPF_MOD | BPF_X:
1526 		case BPF_ALU64 | BPF_DIV | BPF_X:
1527 		case BPF_ALU64 | BPF_MOD | BPF_K:
1528 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1529 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1530 
1531 			if (dst_reg != BPF_REG_0)
1532 				EMIT1(0x50); /* push rax */
1533 			if (dst_reg != BPF_REG_3)
1534 				EMIT1(0x52); /* push rdx */
1535 
1536 			if (BPF_SRC(insn->code) == BPF_X) {
1537 				if (src_reg == BPF_REG_0 ||
1538 				    src_reg == BPF_REG_3) {
1539 					/* mov r11, src_reg */
1540 					EMIT_mov(AUX_REG, src_reg);
1541 					src_reg = AUX_REG;
1542 				}
1543 			} else {
1544 				/* mov r11, imm32 */
1545 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1546 				src_reg = AUX_REG;
1547 			}
1548 
1549 			if (dst_reg != BPF_REG_0)
1550 				/* mov rax, dst_reg */
1551 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1552 
1553 			if (insn->off == 0) {
1554 				/*
1555 				 * xor edx, edx
1556 				 * equivalent to 'xor rdx, rdx', but one byte less
1557 				 */
1558 				EMIT2(0x31, 0xd2);
1559 
1560 				/* div src_reg */
1561 				maybe_emit_1mod(&prog, src_reg, is64);
1562 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1563 			} else {
1564 				if (BPF_CLASS(insn->code) == BPF_ALU)
1565 					EMIT1(0x99); /* cdq */
1566 				else
1567 					EMIT2(0x48, 0x99); /* cqo */
1568 
1569 				/* idiv src_reg */
1570 				maybe_emit_1mod(&prog, src_reg, is64);
1571 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1572 			}
1573 
1574 			if (BPF_OP(insn->code) == BPF_MOD &&
1575 			    dst_reg != BPF_REG_3)
1576 				/* mov dst_reg, rdx */
1577 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1578 			else if (BPF_OP(insn->code) == BPF_DIV &&
1579 				 dst_reg != BPF_REG_0)
1580 				/* mov dst_reg, rax */
1581 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1582 
1583 			if (dst_reg != BPF_REG_3)
1584 				EMIT1(0x5A); /* pop rdx */
1585 			if (dst_reg != BPF_REG_0)
1586 				EMIT1(0x58); /* pop rax */
1587 			break;
1588 		}
1589 
1590 		case BPF_ALU | BPF_MUL | BPF_K:
1591 		case BPF_ALU64 | BPF_MUL | BPF_K:
1592 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1593 				       BPF_CLASS(insn->code) == BPF_ALU64);
1594 
1595 			if (is_imm8(imm32))
1596 				/* imul dst_reg, dst_reg, imm8 */
1597 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1598 				      imm32);
1599 			else
1600 				/* imul dst_reg, dst_reg, imm32 */
1601 				EMIT2_off32(0x69,
1602 					    add_2reg(0xC0, dst_reg, dst_reg),
1603 					    imm32);
1604 			break;
1605 
1606 		case BPF_ALU | BPF_MUL | BPF_X:
1607 		case BPF_ALU64 | BPF_MUL | BPF_X:
1608 			maybe_emit_mod(&prog, src_reg, dst_reg,
1609 				       BPF_CLASS(insn->code) == BPF_ALU64);
1610 
1611 			/* imul dst_reg, src_reg */
1612 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1613 			break;
1614 
1615 			/* Shifts */
1616 		case BPF_ALU | BPF_LSH | BPF_K:
1617 		case BPF_ALU | BPF_RSH | BPF_K:
1618 		case BPF_ALU | BPF_ARSH | BPF_K:
1619 		case BPF_ALU64 | BPF_LSH | BPF_K:
1620 		case BPF_ALU64 | BPF_RSH | BPF_K:
1621 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1622 			maybe_emit_1mod(&prog, dst_reg,
1623 					BPF_CLASS(insn->code) == BPF_ALU64);
1624 
1625 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1626 			if (imm32 == 1)
1627 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1628 			else
1629 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1630 			break;
1631 
1632 		case BPF_ALU | BPF_LSH | BPF_X:
1633 		case BPF_ALU | BPF_RSH | BPF_X:
1634 		case BPF_ALU | BPF_ARSH | BPF_X:
1635 		case BPF_ALU64 | BPF_LSH | BPF_X:
1636 		case BPF_ALU64 | BPF_RSH | BPF_X:
1637 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1638 			/* BMI2 shifts aren't better when shift count is already in rcx */
1639 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1640 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1641 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1642 				u8 op;
1643 
1644 				switch (BPF_OP(insn->code)) {
1645 				case BPF_LSH:
1646 					op = 1; /* prefix 0x66 */
1647 					break;
1648 				case BPF_RSH:
1649 					op = 3; /* prefix 0xf2 */
1650 					break;
1651 				case BPF_ARSH:
1652 					op = 2; /* prefix 0xf3 */
1653 					break;
1654 				}
1655 
1656 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1657 
1658 				break;
1659 			}
1660 
1661 			if (src_reg != BPF_REG_4) { /* common case */
1662 				/* Check for bad case when dst_reg == rcx */
1663 				if (dst_reg == BPF_REG_4) {
1664 					/* mov r11, dst_reg */
1665 					EMIT_mov(AUX_REG, dst_reg);
1666 					dst_reg = AUX_REG;
1667 				} else {
1668 					EMIT1(0x51); /* push rcx */
1669 				}
1670 				/* mov rcx, src_reg */
1671 				EMIT_mov(BPF_REG_4, src_reg);
1672 			}
1673 
1674 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1675 			maybe_emit_1mod(&prog, dst_reg,
1676 					BPF_CLASS(insn->code) == BPF_ALU64);
1677 
1678 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1679 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1680 
1681 			if (src_reg != BPF_REG_4) {
1682 				if (insn->dst_reg == BPF_REG_4)
1683 					/* mov dst_reg, r11 */
1684 					EMIT_mov(insn->dst_reg, AUX_REG);
1685 				else
1686 					EMIT1(0x59); /* pop rcx */
1687 			}
1688 
1689 			break;
1690 
1691 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1692 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1693 			switch (imm32) {
1694 			case 16:
1695 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1696 				EMIT1(0x66);
1697 				if (is_ereg(dst_reg))
1698 					EMIT1(0x41);
1699 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1700 
1701 				/* Emit 'movzwl eax, ax' */
1702 				if (is_ereg(dst_reg))
1703 					EMIT3(0x45, 0x0F, 0xB7);
1704 				else
1705 					EMIT2(0x0F, 0xB7);
1706 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1707 				break;
1708 			case 32:
1709 				/* Emit 'bswap eax' to swap lower 4 bytes */
1710 				if (is_ereg(dst_reg))
1711 					EMIT2(0x41, 0x0F);
1712 				else
1713 					EMIT1(0x0F);
1714 				EMIT1(add_1reg(0xC8, dst_reg));
1715 				break;
1716 			case 64:
1717 				/* Emit 'bswap rax' to swap 8 bytes */
1718 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1719 				      add_1reg(0xC8, dst_reg));
1720 				break;
1721 			}
1722 			break;
1723 
1724 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1725 			switch (imm32) {
1726 			case 16:
1727 				/*
1728 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1729 				 * into 64 bit
1730 				 */
1731 				if (is_ereg(dst_reg))
1732 					EMIT3(0x45, 0x0F, 0xB7);
1733 				else
1734 					EMIT2(0x0F, 0xB7);
1735 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1736 				break;
1737 			case 32:
1738 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1739 				if (is_ereg(dst_reg))
1740 					EMIT1(0x45);
1741 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1742 				break;
1743 			case 64:
1744 				/* nop */
1745 				break;
1746 			}
1747 			break;
1748 
1749 			/* speculation barrier */
1750 		case BPF_ST | BPF_NOSPEC:
1751 			EMIT_LFENCE();
1752 			break;
1753 
1754 			/* ST: *(u8*)(dst_reg + off) = imm */
1755 		case BPF_ST | BPF_MEM | BPF_B:
1756 			if (is_ereg(dst_reg))
1757 				EMIT2(0x41, 0xC6);
1758 			else
1759 				EMIT1(0xC6);
1760 			goto st;
1761 		case BPF_ST | BPF_MEM | BPF_H:
1762 			if (is_ereg(dst_reg))
1763 				EMIT3(0x66, 0x41, 0xC7);
1764 			else
1765 				EMIT2(0x66, 0xC7);
1766 			goto st;
1767 		case BPF_ST | BPF_MEM | BPF_W:
1768 			if (is_ereg(dst_reg))
1769 				EMIT2(0x41, 0xC7);
1770 			else
1771 				EMIT1(0xC7);
1772 			goto st;
1773 		case BPF_ST | BPF_MEM | BPF_DW:
1774 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1775 
1776 st:			if (is_imm8(insn->off))
1777 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1778 			else
1779 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1780 
1781 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1782 			break;
1783 
1784 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1785 		case BPF_STX | BPF_MEM | BPF_B:
1786 		case BPF_STX | BPF_MEM | BPF_H:
1787 		case BPF_STX | BPF_MEM | BPF_W:
1788 		case BPF_STX | BPF_MEM | BPF_DW:
1789 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1790 			break;
1791 
1792 		case BPF_ST | BPF_PROBE_MEM32 | BPF_B:
1793 		case BPF_ST | BPF_PROBE_MEM32 | BPF_H:
1794 		case BPF_ST | BPF_PROBE_MEM32 | BPF_W:
1795 		case BPF_ST | BPF_PROBE_MEM32 | BPF_DW:
1796 			start_of_ldx = prog;
1797 			emit_st_r12(&prog, BPF_SIZE(insn->code), dst_reg, insn->off, insn->imm);
1798 			goto populate_extable;
1799 
1800 			/* LDX: dst_reg = *(u8*)(src_reg + r12 + off) */
1801 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_B:
1802 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
1803 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
1804 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
1805 		case BPF_STX | BPF_PROBE_MEM32 | BPF_B:
1806 		case BPF_STX | BPF_PROBE_MEM32 | BPF_H:
1807 		case BPF_STX | BPF_PROBE_MEM32 | BPF_W:
1808 		case BPF_STX | BPF_PROBE_MEM32 | BPF_DW:
1809 			start_of_ldx = prog;
1810 			if (BPF_CLASS(insn->code) == BPF_LDX)
1811 				emit_ldx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1812 			else
1813 				emit_stx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1814 populate_extable:
1815 			{
1816 				struct exception_table_entry *ex;
1817 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1818 				s64 delta;
1819 
1820 				if (!bpf_prog->aux->extable)
1821 					break;
1822 
1823 				if (excnt >= bpf_prog->aux->num_exentries) {
1824 					pr_err("mem32 extable bug\n");
1825 					return -EFAULT;
1826 				}
1827 				ex = &bpf_prog->aux->extable[excnt++];
1828 
1829 				delta = _insn - (u8 *)&ex->insn;
1830 				/* switch ex to rw buffer for writes */
1831 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1832 
1833 				ex->insn = delta;
1834 
1835 				ex->data = EX_TYPE_BPF;
1836 
1837 				ex->fixup = (prog - start_of_ldx) |
1838 					((BPF_CLASS(insn->code) == BPF_LDX ? reg2pt_regs[dst_reg] : DONT_CLEAR) << 8);
1839 			}
1840 			break;
1841 
1842 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1843 		case BPF_LDX | BPF_MEM | BPF_B:
1844 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1845 		case BPF_LDX | BPF_MEM | BPF_H:
1846 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1847 		case BPF_LDX | BPF_MEM | BPF_W:
1848 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1849 		case BPF_LDX | BPF_MEM | BPF_DW:
1850 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1851 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
1852 		case BPF_LDX | BPF_MEMSX | BPF_B:
1853 		case BPF_LDX | BPF_MEMSX | BPF_H:
1854 		case BPF_LDX | BPF_MEMSX | BPF_W:
1855 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1856 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1857 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1858 			insn_off = insn->off;
1859 
1860 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1861 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1862 				/* Conservatively check that src_reg + insn->off is a kernel address:
1863 				 *   src_reg + insn->off > TASK_SIZE_MAX + PAGE_SIZE
1864 				 *   and
1865 				 *   src_reg + insn->off < VSYSCALL_ADDR
1866 				 */
1867 
1868 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE - VSYSCALL_ADDR;
1869 				u8 *end_of_jmp;
1870 
1871 				/* movabsq r10, VSYSCALL_ADDR */
1872 				emit_mov_imm64(&prog, BPF_REG_AX, (long)VSYSCALL_ADDR >> 32,
1873 					       (u32)(long)VSYSCALL_ADDR);
1874 
1875 				/* mov src_reg, r11 */
1876 				EMIT_mov(AUX_REG, src_reg);
1877 
1878 				if (insn->off) {
1879 					/* add r11, insn->off */
1880 					maybe_emit_1mod(&prog, AUX_REG, true);
1881 					EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1882 				}
1883 
1884 				/* sub r11, r10 */
1885 				maybe_emit_mod(&prog, AUX_REG, BPF_REG_AX, true);
1886 				EMIT2(0x29, add_2reg(0xC0, AUX_REG, BPF_REG_AX));
1887 
1888 				/* movabsq r10, limit */
1889 				emit_mov_imm64(&prog, BPF_REG_AX, (long)limit >> 32,
1890 					       (u32)(long)limit);
1891 
1892 				/* cmp r10, r11 */
1893 				maybe_emit_mod(&prog, AUX_REG, BPF_REG_AX, true);
1894 				EMIT2(0x39, add_2reg(0xC0, AUX_REG, BPF_REG_AX));
1895 
1896 				/* if unsigned '>', goto load */
1897 				EMIT2(X86_JA, 0);
1898 				end_of_jmp = prog;
1899 
1900 				/* xor dst_reg, dst_reg */
1901 				emit_mov_imm32(&prog, false, dst_reg, 0);
1902 				/* jmp byte_after_ldx */
1903 				EMIT2(0xEB, 0);
1904 
1905 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1906 				start_of_ldx = prog;
1907 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1908 			}
1909 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
1910 			    BPF_MODE(insn->code) == BPF_MEMSX)
1911 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1912 			else
1913 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1914 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1915 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1916 				struct exception_table_entry *ex;
1917 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1918 				s64 delta;
1919 
1920 				/* populate jmp_offset for JMP above */
1921 				start_of_ldx[-1] = prog - start_of_ldx;
1922 
1923 				if (!bpf_prog->aux->extable)
1924 					break;
1925 
1926 				if (excnt >= bpf_prog->aux->num_exentries) {
1927 					pr_err("ex gen bug\n");
1928 					return -EFAULT;
1929 				}
1930 				ex = &bpf_prog->aux->extable[excnt++];
1931 
1932 				delta = _insn - (u8 *)&ex->insn;
1933 				if (!is_simm32(delta)) {
1934 					pr_err("extable->insn doesn't fit into 32-bit\n");
1935 					return -EFAULT;
1936 				}
1937 				/* switch ex to rw buffer for writes */
1938 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1939 
1940 				ex->insn = delta;
1941 
1942 				ex->data = EX_TYPE_BPF;
1943 
1944 				if (dst_reg > BPF_REG_9) {
1945 					pr_err("verifier error\n");
1946 					return -EFAULT;
1947 				}
1948 				/*
1949 				 * Compute size of x86 insn and its target dest x86 register.
1950 				 * ex_handler_bpf() will use lower 8 bits to adjust
1951 				 * pt_regs->ip to jump over this x86 instruction
1952 				 * and upper bits to figure out which pt_regs to zero out.
1953 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1954 				 * of 4 bytes will be ignored and rbx will be zero inited.
1955 				 */
1956 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1957 			}
1958 			break;
1959 
1960 		case BPF_STX | BPF_ATOMIC | BPF_W:
1961 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1962 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1963 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1964 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1965 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1966 				u32 real_src_reg = src_reg;
1967 				u32 real_dst_reg = dst_reg;
1968 				u8 *branch_target;
1969 
1970 				/*
1971 				 * Can't be implemented with a single x86 insn.
1972 				 * Need to do a CMPXCHG loop.
1973 				 */
1974 
1975 				/* Will need RAX as a CMPXCHG operand so save R0 */
1976 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1977 				if (src_reg == BPF_REG_0)
1978 					real_src_reg = BPF_REG_AX;
1979 				if (dst_reg == BPF_REG_0)
1980 					real_dst_reg = BPF_REG_AX;
1981 
1982 				branch_target = prog;
1983 				/* Load old value */
1984 				emit_ldx(&prog, BPF_SIZE(insn->code),
1985 					 BPF_REG_0, real_dst_reg, insn->off);
1986 				/*
1987 				 * Perform the (commutative) operation locally,
1988 				 * put the result in the AUX_REG.
1989 				 */
1990 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1991 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1992 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1993 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1994 				/* Attempt to swap in new value */
1995 				err = emit_atomic(&prog, BPF_CMPXCHG,
1996 						  real_dst_reg, AUX_REG,
1997 						  insn->off,
1998 						  BPF_SIZE(insn->code));
1999 				if (WARN_ON(err))
2000 					return err;
2001 				/*
2002 				 * ZF tells us whether we won the race. If it's
2003 				 * cleared we need to try again.
2004 				 */
2005 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
2006 				/* Return the pre-modification value */
2007 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
2008 				/* Restore R0 after clobbering RAX */
2009 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
2010 				break;
2011 			}
2012 
2013 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
2014 					  insn->off, BPF_SIZE(insn->code));
2015 			if (err)
2016 				return err;
2017 			break;
2018 
2019 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_W:
2020 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_DW:
2021 			start_of_ldx = prog;
2022 			err = emit_atomic_index(&prog, insn->imm, BPF_SIZE(insn->code),
2023 						dst_reg, src_reg, X86_REG_R12, insn->off);
2024 			if (err)
2025 				return err;
2026 			goto populate_extable;
2027 
2028 			/* call */
2029 		case BPF_JMP | BPF_CALL: {
2030 			u8 *ip = image + addrs[i - 1];
2031 
2032 			func = (u8 *) __bpf_call_base + imm32;
2033 			if (tail_call_reachable) {
2034 				RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
2035 				ip += 7;
2036 			}
2037 			if (!imm32)
2038 				return -EINVAL;
2039 			ip += x86_call_depth_emit_accounting(&prog, func, ip);
2040 			if (emit_call(&prog, func, ip))
2041 				return -EINVAL;
2042 			break;
2043 		}
2044 
2045 		case BPF_JMP | BPF_TAIL_CALL:
2046 			if (imm32)
2047 				emit_bpf_tail_call_direct(bpf_prog,
2048 							  &bpf_prog->aux->poke_tab[imm32 - 1],
2049 							  &prog, image + addrs[i - 1],
2050 							  callee_regs_used,
2051 							  bpf_prog->aux->stack_depth,
2052 							  ctx);
2053 			else
2054 				emit_bpf_tail_call_indirect(bpf_prog,
2055 							    &prog,
2056 							    callee_regs_used,
2057 							    bpf_prog->aux->stack_depth,
2058 							    image + addrs[i - 1],
2059 							    ctx);
2060 			break;
2061 
2062 			/* cond jump */
2063 		case BPF_JMP | BPF_JEQ | BPF_X:
2064 		case BPF_JMP | BPF_JNE | BPF_X:
2065 		case BPF_JMP | BPF_JGT | BPF_X:
2066 		case BPF_JMP | BPF_JLT | BPF_X:
2067 		case BPF_JMP | BPF_JGE | BPF_X:
2068 		case BPF_JMP | BPF_JLE | BPF_X:
2069 		case BPF_JMP | BPF_JSGT | BPF_X:
2070 		case BPF_JMP | BPF_JSLT | BPF_X:
2071 		case BPF_JMP | BPF_JSGE | BPF_X:
2072 		case BPF_JMP | BPF_JSLE | BPF_X:
2073 		case BPF_JMP32 | BPF_JEQ | BPF_X:
2074 		case BPF_JMP32 | BPF_JNE | BPF_X:
2075 		case BPF_JMP32 | BPF_JGT | BPF_X:
2076 		case BPF_JMP32 | BPF_JLT | BPF_X:
2077 		case BPF_JMP32 | BPF_JGE | BPF_X:
2078 		case BPF_JMP32 | BPF_JLE | BPF_X:
2079 		case BPF_JMP32 | BPF_JSGT | BPF_X:
2080 		case BPF_JMP32 | BPF_JSLT | BPF_X:
2081 		case BPF_JMP32 | BPF_JSGE | BPF_X:
2082 		case BPF_JMP32 | BPF_JSLE | BPF_X:
2083 			/* cmp dst_reg, src_reg */
2084 			maybe_emit_mod(&prog, dst_reg, src_reg,
2085 				       BPF_CLASS(insn->code) == BPF_JMP);
2086 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
2087 			goto emit_cond_jmp;
2088 
2089 		case BPF_JMP | BPF_JSET | BPF_X:
2090 		case BPF_JMP32 | BPF_JSET | BPF_X:
2091 			/* test dst_reg, src_reg */
2092 			maybe_emit_mod(&prog, dst_reg, src_reg,
2093 				       BPF_CLASS(insn->code) == BPF_JMP);
2094 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
2095 			goto emit_cond_jmp;
2096 
2097 		case BPF_JMP | BPF_JSET | BPF_K:
2098 		case BPF_JMP32 | BPF_JSET | BPF_K:
2099 			/* test dst_reg, imm32 */
2100 			maybe_emit_1mod(&prog, dst_reg,
2101 					BPF_CLASS(insn->code) == BPF_JMP);
2102 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
2103 			goto emit_cond_jmp;
2104 
2105 		case BPF_JMP | BPF_JEQ | BPF_K:
2106 		case BPF_JMP | BPF_JNE | BPF_K:
2107 		case BPF_JMP | BPF_JGT | BPF_K:
2108 		case BPF_JMP | BPF_JLT | BPF_K:
2109 		case BPF_JMP | BPF_JGE | BPF_K:
2110 		case BPF_JMP | BPF_JLE | BPF_K:
2111 		case BPF_JMP | BPF_JSGT | BPF_K:
2112 		case BPF_JMP | BPF_JSLT | BPF_K:
2113 		case BPF_JMP | BPF_JSGE | BPF_K:
2114 		case BPF_JMP | BPF_JSLE | BPF_K:
2115 		case BPF_JMP32 | BPF_JEQ | BPF_K:
2116 		case BPF_JMP32 | BPF_JNE | BPF_K:
2117 		case BPF_JMP32 | BPF_JGT | BPF_K:
2118 		case BPF_JMP32 | BPF_JLT | BPF_K:
2119 		case BPF_JMP32 | BPF_JGE | BPF_K:
2120 		case BPF_JMP32 | BPF_JLE | BPF_K:
2121 		case BPF_JMP32 | BPF_JSGT | BPF_K:
2122 		case BPF_JMP32 | BPF_JSLT | BPF_K:
2123 		case BPF_JMP32 | BPF_JSGE | BPF_K:
2124 		case BPF_JMP32 | BPF_JSLE | BPF_K:
2125 			/* test dst_reg, dst_reg to save one extra byte */
2126 			if (imm32 == 0) {
2127 				maybe_emit_mod(&prog, dst_reg, dst_reg,
2128 					       BPF_CLASS(insn->code) == BPF_JMP);
2129 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
2130 				goto emit_cond_jmp;
2131 			}
2132 
2133 			/* cmp dst_reg, imm8/32 */
2134 			maybe_emit_1mod(&prog, dst_reg,
2135 					BPF_CLASS(insn->code) == BPF_JMP);
2136 
2137 			if (is_imm8(imm32))
2138 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
2139 			else
2140 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
2141 
2142 emit_cond_jmp:		/* Convert BPF opcode to x86 */
2143 			switch (BPF_OP(insn->code)) {
2144 			case BPF_JEQ:
2145 				jmp_cond = X86_JE;
2146 				break;
2147 			case BPF_JSET:
2148 			case BPF_JNE:
2149 				jmp_cond = X86_JNE;
2150 				break;
2151 			case BPF_JGT:
2152 				/* GT is unsigned '>', JA in x86 */
2153 				jmp_cond = X86_JA;
2154 				break;
2155 			case BPF_JLT:
2156 				/* LT is unsigned '<', JB in x86 */
2157 				jmp_cond = X86_JB;
2158 				break;
2159 			case BPF_JGE:
2160 				/* GE is unsigned '>=', JAE in x86 */
2161 				jmp_cond = X86_JAE;
2162 				break;
2163 			case BPF_JLE:
2164 				/* LE is unsigned '<=', JBE in x86 */
2165 				jmp_cond = X86_JBE;
2166 				break;
2167 			case BPF_JSGT:
2168 				/* Signed '>', GT in x86 */
2169 				jmp_cond = X86_JG;
2170 				break;
2171 			case BPF_JSLT:
2172 				/* Signed '<', LT in x86 */
2173 				jmp_cond = X86_JL;
2174 				break;
2175 			case BPF_JSGE:
2176 				/* Signed '>=', GE in x86 */
2177 				jmp_cond = X86_JGE;
2178 				break;
2179 			case BPF_JSLE:
2180 				/* Signed '<=', LE in x86 */
2181 				jmp_cond = X86_JLE;
2182 				break;
2183 			default: /* to silence GCC warning */
2184 				return -EFAULT;
2185 			}
2186 			jmp_offset = addrs[i + insn->off] - addrs[i];
2187 			if (is_imm8(jmp_offset)) {
2188 				if (jmp_padding) {
2189 					/* To keep the jmp_offset valid, the extra bytes are
2190 					 * padded before the jump insn, so we subtract the
2191 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
2192 					 *
2193 					 * If the previous pass already emits an imm8
2194 					 * jmp_cond, then this BPF insn won't shrink, so
2195 					 * "nops" is 0.
2196 					 *
2197 					 * On the other hand, if the previous pass emits an
2198 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
2199 					 * keep the image from shrinking further.
2200 					 *
2201 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
2202 					 *     is 2 bytes, so the size difference is 4 bytes.
2203 					 */
2204 					nops = INSN_SZ_DIFF - 2;
2205 					if (nops != 0 && nops != 4) {
2206 						pr_err("unexpected jmp_cond padding: %d bytes\n",
2207 						       nops);
2208 						return -EFAULT;
2209 					}
2210 					emit_nops(&prog, nops);
2211 				}
2212 				EMIT2(jmp_cond, jmp_offset);
2213 			} else if (is_simm32(jmp_offset)) {
2214 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
2215 			} else {
2216 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
2217 				return -EFAULT;
2218 			}
2219 
2220 			break;
2221 
2222 		case BPF_JMP | BPF_JA:
2223 		case BPF_JMP32 | BPF_JA:
2224 			if (BPF_CLASS(insn->code) == BPF_JMP) {
2225 				if (insn->off == -1)
2226 					/* -1 jmp instructions will always jump
2227 					 * backwards two bytes. Explicitly handling
2228 					 * this case avoids wasting too many passes
2229 					 * when there are long sequences of replaced
2230 					 * dead code.
2231 					 */
2232 					jmp_offset = -2;
2233 				else
2234 					jmp_offset = addrs[i + insn->off] - addrs[i];
2235 			} else {
2236 				if (insn->imm == -1)
2237 					jmp_offset = -2;
2238 				else
2239 					jmp_offset = addrs[i + insn->imm] - addrs[i];
2240 			}
2241 
2242 			if (!jmp_offset) {
2243 				/*
2244 				 * If jmp_padding is enabled, the extra nops will
2245 				 * be inserted. Otherwise, optimize out nop jumps.
2246 				 */
2247 				if (jmp_padding) {
2248 					/* There are 3 possible conditions.
2249 					 * (1) This BPF_JA is already optimized out in
2250 					 *     the previous run, so there is no need
2251 					 *     to pad any extra byte (0 byte).
2252 					 * (2) The previous pass emits an imm8 jmp,
2253 					 *     so we pad 2 bytes to match the previous
2254 					 *     insn size.
2255 					 * (3) Similarly, the previous pass emits an
2256 					 *     imm32 jmp, and 5 bytes is padded.
2257 					 */
2258 					nops = INSN_SZ_DIFF;
2259 					if (nops != 0 && nops != 2 && nops != 5) {
2260 						pr_err("unexpected nop jump padding: %d bytes\n",
2261 						       nops);
2262 						return -EFAULT;
2263 					}
2264 					emit_nops(&prog, nops);
2265 				}
2266 				break;
2267 			}
2268 emit_jmp:
2269 			if (is_imm8(jmp_offset)) {
2270 				if (jmp_padding) {
2271 					/* To avoid breaking jmp_offset, the extra bytes
2272 					 * are padded before the actual jmp insn, so
2273 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
2274 					 *
2275 					 * If the previous pass already emits an imm8
2276 					 * jmp, there is nothing to pad (0 byte).
2277 					 *
2278 					 * If it emits an imm32 jmp (5 bytes) previously
2279 					 * and now an imm8 jmp (2 bytes), then we pad
2280 					 * (5 - 2 = 3) bytes to stop the image from
2281 					 * shrinking further.
2282 					 */
2283 					nops = INSN_SZ_DIFF - 2;
2284 					if (nops != 0 && nops != 3) {
2285 						pr_err("unexpected jump padding: %d bytes\n",
2286 						       nops);
2287 						return -EFAULT;
2288 					}
2289 					emit_nops(&prog, INSN_SZ_DIFF - 2);
2290 				}
2291 				EMIT2(0xEB, jmp_offset);
2292 			} else if (is_simm32(jmp_offset)) {
2293 				EMIT1_off32(0xE9, jmp_offset);
2294 			} else {
2295 				pr_err("jmp gen bug %llx\n", jmp_offset);
2296 				return -EFAULT;
2297 			}
2298 			break;
2299 
2300 		case BPF_JMP | BPF_EXIT:
2301 			if (seen_exit) {
2302 				jmp_offset = ctx->cleanup_addr - addrs[i];
2303 				goto emit_jmp;
2304 			}
2305 			seen_exit = true;
2306 			/* Update cleanup_addr */
2307 			ctx->cleanup_addr = proglen;
2308 			if (bpf_prog->aux->exception_boundary) {
2309 				pop_callee_regs(&prog, all_callee_regs_used);
2310 				pop_r12(&prog);
2311 			} else {
2312 				pop_callee_regs(&prog, callee_regs_used);
2313 				if (arena_vm_start)
2314 					pop_r12(&prog);
2315 			}
2316 			EMIT1(0xC9);         /* leave */
2317 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
2318 			break;
2319 
2320 		default:
2321 			/*
2322 			 * By design x86-64 JIT should support all BPF instructions.
2323 			 * This error will be seen if new instruction was added
2324 			 * to the interpreter, but not to the JIT, or if there is
2325 			 * junk in bpf_prog.
2326 			 */
2327 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
2328 			return -EINVAL;
2329 		}
2330 
2331 		ilen = prog - temp;
2332 		if (ilen > BPF_MAX_INSN_SIZE) {
2333 			pr_err("bpf_jit: fatal insn size error\n");
2334 			return -EFAULT;
2335 		}
2336 
2337 		if (image) {
2338 			/*
2339 			 * When populating the image, assert that:
2340 			 *
2341 			 *  i) We do not write beyond the allocated space, and
2342 			 * ii) addrs[i] did not change from the prior run, in order
2343 			 *     to validate assumptions made for computing branch
2344 			 *     displacements.
2345 			 */
2346 			if (unlikely(proglen + ilen > oldproglen ||
2347 				     proglen + ilen != addrs[i])) {
2348 				pr_err("bpf_jit: fatal error\n");
2349 				return -EFAULT;
2350 			}
2351 			memcpy(rw_image + proglen, temp, ilen);
2352 		}
2353 		proglen += ilen;
2354 		addrs[i] = proglen;
2355 		prog = temp;
2356 	}
2357 
2358 	if (image && excnt != bpf_prog->aux->num_exentries) {
2359 		pr_err("extable is not populated\n");
2360 		return -EFAULT;
2361 	}
2362 	return proglen;
2363 }
2364 
2365 static void clean_stack_garbage(const struct btf_func_model *m,
2366 				u8 **pprog, int nr_stack_slots,
2367 				int stack_size)
2368 {
2369 	int arg_size, off;
2370 	u8 *prog;
2371 
2372 	/* Generally speaking, the compiler will pass the arguments
2373 	 * on-stack with "push" instruction, which will take 8-byte
2374 	 * on the stack. In this case, there won't be garbage values
2375 	 * while we copy the arguments from origin stack frame to current
2376 	 * in BPF_DW.
2377 	 *
2378 	 * However, sometimes the compiler will only allocate 4-byte on
2379 	 * the stack for the arguments. For now, this case will only
2380 	 * happen if there is only one argument on-stack and its size
2381 	 * not more than 4 byte. In this case, there will be garbage
2382 	 * values on the upper 4-byte where we store the argument on
2383 	 * current stack frame.
2384 	 *
2385 	 * arguments on origin stack:
2386 	 *
2387 	 * stack_arg_1(4-byte) xxx(4-byte)
2388 	 *
2389 	 * what we copy:
2390 	 *
2391 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
2392 	 *
2393 	 * and the xxx is the garbage values which we should clean here.
2394 	 */
2395 	if (nr_stack_slots != 1)
2396 		return;
2397 
2398 	/* the size of the last argument */
2399 	arg_size = m->arg_size[m->nr_args - 1];
2400 	if (arg_size <= 4) {
2401 		off = -(stack_size - 4);
2402 		prog = *pprog;
2403 		/* mov DWORD PTR [rbp + off], 0 */
2404 		if (!is_imm8(off))
2405 			EMIT2_off32(0xC7, 0x85, off);
2406 		else
2407 			EMIT3(0xC7, 0x45, off);
2408 		EMIT(0, 4);
2409 		*pprog = prog;
2410 	}
2411 }
2412 
2413 /* get the count of the regs that are used to pass arguments */
2414 static int get_nr_used_regs(const struct btf_func_model *m)
2415 {
2416 	int i, arg_regs, nr_used_regs = 0;
2417 
2418 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2419 		arg_regs = (m->arg_size[i] + 7) / 8;
2420 		if (nr_used_regs + arg_regs <= 6)
2421 			nr_used_regs += arg_regs;
2422 
2423 		if (nr_used_regs >= 6)
2424 			break;
2425 	}
2426 
2427 	return nr_used_regs;
2428 }
2429 
2430 static void save_args(const struct btf_func_model *m, u8 **prog,
2431 		      int stack_size, bool for_call_origin)
2432 {
2433 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2434 	int i, j;
2435 
2436 	/* Store function arguments to stack.
2437 	 * For a function that accepts two pointers the sequence will be:
2438 	 * mov QWORD PTR [rbp-0x10],rdi
2439 	 * mov QWORD PTR [rbp-0x8],rsi
2440 	 */
2441 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2442 		arg_regs = (m->arg_size[i] + 7) / 8;
2443 
2444 		/* According to the research of Yonghong, struct members
2445 		 * should be all in register or all on the stack.
2446 		 * Meanwhile, the compiler will pass the argument on regs
2447 		 * if the remaining regs can hold the argument.
2448 		 *
2449 		 * Disorder of the args can happen. For example:
2450 		 *
2451 		 * struct foo_struct {
2452 		 *     long a;
2453 		 *     int b;
2454 		 * };
2455 		 * int foo(char, char, char, char, char, struct foo_struct,
2456 		 *         char);
2457 		 *
2458 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2459 		 * by stack.
2460 		 */
2461 		if (nr_regs + arg_regs > 6) {
2462 			/* copy function arguments from origin stack frame
2463 			 * into current stack frame.
2464 			 *
2465 			 * The starting address of the arguments on-stack
2466 			 * is:
2467 			 *   rbp + 8(push rbp) +
2468 			 *   8(return addr of origin call) +
2469 			 *   8(return addr of the caller)
2470 			 * which means: rbp + 24
2471 			 */
2472 			for (j = 0; j < arg_regs; j++) {
2473 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2474 					 nr_stack_slots * 8 + 0x18);
2475 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2476 					 -stack_size);
2477 
2478 				if (!nr_stack_slots)
2479 					first_off = stack_size;
2480 				stack_size -= 8;
2481 				nr_stack_slots++;
2482 			}
2483 		} else {
2484 			/* Only copy the arguments on-stack to current
2485 			 * 'stack_size' and ignore the regs, used to
2486 			 * prepare the arguments on-stack for origin call.
2487 			 */
2488 			if (for_call_origin) {
2489 				nr_regs += arg_regs;
2490 				continue;
2491 			}
2492 
2493 			/* copy the arguments from regs into stack */
2494 			for (j = 0; j < arg_regs; j++) {
2495 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2496 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2497 					 -stack_size);
2498 				stack_size -= 8;
2499 				nr_regs++;
2500 			}
2501 		}
2502 	}
2503 
2504 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2505 }
2506 
2507 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2508 			 int stack_size)
2509 {
2510 	int i, j, arg_regs, nr_regs = 0;
2511 
2512 	/* Restore function arguments from stack.
2513 	 * For a function that accepts two pointers the sequence will be:
2514 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2515 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2516 	 *
2517 	 * The logic here is similar to what we do in save_args()
2518 	 */
2519 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2520 		arg_regs = (m->arg_size[i] + 7) / 8;
2521 		if (nr_regs + arg_regs <= 6) {
2522 			for (j = 0; j < arg_regs; j++) {
2523 				emit_ldx(prog, BPF_DW,
2524 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2525 					 BPF_REG_FP,
2526 					 -stack_size);
2527 				stack_size -= 8;
2528 				nr_regs++;
2529 			}
2530 		} else {
2531 			stack_size -= 8 * arg_regs;
2532 		}
2533 
2534 		if (nr_regs >= 6)
2535 			break;
2536 	}
2537 }
2538 
2539 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2540 			   struct bpf_tramp_link *l, int stack_size,
2541 			   int run_ctx_off, bool save_ret,
2542 			   void *image, void *rw_image)
2543 {
2544 	u8 *prog = *pprog;
2545 	u8 *jmp_insn;
2546 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2547 	struct bpf_prog *p = l->link.prog;
2548 	u64 cookie = l->cookie;
2549 
2550 	/* mov rdi, cookie */
2551 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2552 
2553 	/* Prepare struct bpf_tramp_run_ctx.
2554 	 *
2555 	 * bpf_tramp_run_ctx is already preserved by
2556 	 * arch_prepare_bpf_trampoline().
2557 	 *
2558 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2559 	 */
2560 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2561 
2562 	/* arg1: mov rdi, progs[i] */
2563 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2564 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2565 	if (!is_imm8(-run_ctx_off))
2566 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
2567 	else
2568 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
2569 
2570 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image)))
2571 		return -EINVAL;
2572 	/* remember prog start time returned by __bpf_prog_enter */
2573 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
2574 
2575 	/* if (__bpf_prog_enter*(prog) == 0)
2576 	 *	goto skip_exec_of_prog;
2577 	 */
2578 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
2579 	/* emit 2 nops that will be replaced with JE insn */
2580 	jmp_insn = prog;
2581 	emit_nops(&prog, 2);
2582 
2583 	/* arg1: lea rdi, [rbp - stack_size] */
2584 	if (!is_imm8(-stack_size))
2585 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
2586 	else
2587 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
2588 	/* arg2: progs[i]->insnsi for interpreter */
2589 	if (!p->jited)
2590 		emit_mov_imm64(&prog, BPF_REG_2,
2591 			       (long) p->insnsi >> 32,
2592 			       (u32) (long) p->insnsi);
2593 	/* call JITed bpf program or interpreter */
2594 	if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image)))
2595 		return -EINVAL;
2596 
2597 	/*
2598 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
2599 	 * of the previous call which is then passed on the stack to
2600 	 * the next BPF program.
2601 	 *
2602 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
2603 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
2604 	 */
2605 	if (save_ret)
2606 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2607 
2608 	/* replace 2 nops with JE insn, since jmp target is known */
2609 	jmp_insn[0] = X86_JE;
2610 	jmp_insn[1] = prog - jmp_insn - 2;
2611 
2612 	/* arg1: mov rdi, progs[i] */
2613 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2614 	/* arg2: mov rsi, rbx <- start time in nsec */
2615 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
2616 	/* arg3: lea rdx, [rbp - run_ctx_off] */
2617 	if (!is_imm8(-run_ctx_off))
2618 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
2619 	else
2620 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
2621 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
2622 		return -EINVAL;
2623 
2624 	*pprog = prog;
2625 	return 0;
2626 }
2627 
2628 static void emit_align(u8 **pprog, u32 align)
2629 {
2630 	u8 *target, *prog = *pprog;
2631 
2632 	target = PTR_ALIGN(prog, align);
2633 	if (target != prog)
2634 		emit_nops(&prog, target - prog);
2635 
2636 	*pprog = prog;
2637 }
2638 
2639 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2640 {
2641 	u8 *prog = *pprog;
2642 	s64 offset;
2643 
2644 	offset = func - (ip + 2 + 4);
2645 	if (!is_simm32(offset)) {
2646 		pr_err("Target %p is out of range\n", func);
2647 		return -EINVAL;
2648 	}
2649 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2650 	*pprog = prog;
2651 	return 0;
2652 }
2653 
2654 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2655 		      struct bpf_tramp_links *tl, int stack_size,
2656 		      int run_ctx_off, bool save_ret,
2657 		      void *image, void *rw_image)
2658 {
2659 	int i;
2660 	u8 *prog = *pprog;
2661 
2662 	for (i = 0; i < tl->nr_links; i++) {
2663 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2664 				    run_ctx_off, save_ret, image, rw_image))
2665 			return -EINVAL;
2666 	}
2667 	*pprog = prog;
2668 	return 0;
2669 }
2670 
2671 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2672 			      struct bpf_tramp_links *tl, int stack_size,
2673 			      int run_ctx_off, u8 **branches,
2674 			      void *image, void *rw_image)
2675 {
2676 	u8 *prog = *pprog;
2677 	int i;
2678 
2679 	/* The first fmod_ret program will receive a garbage return value.
2680 	 * Set this to 0 to avoid confusing the program.
2681 	 */
2682 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2683 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2684 	for (i = 0; i < tl->nr_links; i++) {
2685 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true,
2686 				    image, rw_image))
2687 			return -EINVAL;
2688 
2689 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2690 		 * if (*(u64 *)(rbp - 8) !=  0)
2691 		 *	goto do_fexit;
2692 		 */
2693 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2694 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2695 
2696 		/* Save the location of the branch and Generate 6 nops
2697 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2698 		 * are replaced with a conditional jump once do_fexit (i.e. the
2699 		 * start of the fexit invocation) is finalized.
2700 		 */
2701 		branches[i] = prog;
2702 		emit_nops(&prog, 4 + 2);
2703 	}
2704 
2705 	*pprog = prog;
2706 	return 0;
2707 }
2708 
2709 /* Example:
2710  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2711  * its 'struct btf_func_model' will be nr_args=2
2712  * The assembly code when eth_type_trans is executing after trampoline:
2713  *
2714  * push rbp
2715  * mov rbp, rsp
2716  * sub rsp, 16                     // space for skb and dev
2717  * push rbx                        // temp regs to pass start time
2718  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2719  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2720  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2721  * mov rbx, rax                    // remember start time in bpf stats are enabled
2722  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2723  * call addr_of_jited_FENTRY_prog
2724  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2725  * mov rsi, rbx                    // prog start time
2726  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2727  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2728  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2729  * pop rbx
2730  * leave
2731  * ret
2732  *
2733  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2734  * replaced with 'call generated_bpf_trampoline'. When it returns
2735  * eth_type_trans will continue executing with original skb and dev pointers.
2736  *
2737  * The assembly code when eth_type_trans is called from trampoline:
2738  *
2739  * push rbp
2740  * mov rbp, rsp
2741  * sub rsp, 24                     // space for skb, dev, return value
2742  * push rbx                        // temp regs to pass start time
2743  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2744  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2745  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2746  * mov rbx, rax                    // remember start time if bpf stats are enabled
2747  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2748  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2749  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2750  * mov rsi, rbx                    // prog start time
2751  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2752  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2753  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2754  * call eth_type_trans+5           // execute body of eth_type_trans
2755  * mov qword ptr [rbp - 8], rax    // save return value
2756  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2757  * mov rbx, rax                    // remember start time in bpf stats are enabled
2758  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2759  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2760  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2761  * mov rsi, rbx                    // prog start time
2762  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2763  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2764  * pop rbx
2765  * leave
2766  * add rsp, 8                      // skip eth_type_trans's frame
2767  * ret                             // return to its caller
2768  */
2769 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
2770 					 void *rw_image_end, void *image,
2771 					 const struct btf_func_model *m, u32 flags,
2772 					 struct bpf_tramp_links *tlinks,
2773 					 void *func_addr)
2774 {
2775 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2776 	int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
2777 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2778 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2779 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2780 	void *orig_call = func_addr;
2781 	u8 **branches = NULL;
2782 	u8 *prog;
2783 	bool save_ret;
2784 
2785 	/*
2786 	 * F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is
2787 	 * explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG
2788 	 * because @func_addr.
2789 	 */
2790 	WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) &&
2791 		     (flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET)));
2792 
2793 	/* extra registers for struct arguments */
2794 	for (i = 0; i < m->nr_args; i++) {
2795 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2796 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2797 	}
2798 
2799 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
2800 	 * are passed through regs, the remains are through stack.
2801 	 */
2802 	if (nr_regs > MAX_BPF_FUNC_ARGS)
2803 		return -ENOTSUPP;
2804 
2805 	/* Generated trampoline stack layout:
2806 	 *
2807 	 * RBP + 8         [ return address  ]
2808 	 * RBP + 0         [ RBP             ]
2809 	 *
2810 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2811 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2812 	 *
2813 	 *                 [ reg_argN        ]  always
2814 	 *                 [ ...             ]
2815 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2816 	 *
2817 	 * RBP - nregs_off [ regs count	     ]  always
2818 	 *
2819 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2820 	 *
2821 	 * RBP - rbx_off   [ rbx value       ]  always
2822 	 *
2823 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2824 	 *
2825 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
2826 	 *                     [ ...        ]
2827 	 *                     [ stack_arg2 ]
2828 	 * RBP - arg_stack_off [ stack_arg1 ]
2829 	 * RSP                 [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
2830 	 */
2831 
2832 	/* room for return value of orig_call or fentry prog */
2833 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2834 	if (save_ret)
2835 		stack_size += 8;
2836 
2837 	stack_size += nr_regs * 8;
2838 	regs_off = stack_size;
2839 
2840 	/* regs count  */
2841 	stack_size += 8;
2842 	nregs_off = stack_size;
2843 
2844 	if (flags & BPF_TRAMP_F_IP_ARG)
2845 		stack_size += 8; /* room for IP address argument */
2846 
2847 	ip_off = stack_size;
2848 
2849 	stack_size += 8;
2850 	rbx_off = stack_size;
2851 
2852 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2853 	run_ctx_off = stack_size;
2854 
2855 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
2856 		/* the space that used to pass arguments on-stack */
2857 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
2858 		/* make sure the stack pointer is 16-byte aligned if we
2859 		 * need pass arguments on stack, which means
2860 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
2861 		 * should be 16-byte aligned. Following code depend on
2862 		 * that stack_size is already 8-byte aligned.
2863 		 */
2864 		stack_size += (stack_size % 16) ? 0 : 8;
2865 	}
2866 
2867 	arg_stack_off = stack_size;
2868 
2869 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2870 		/* skip patched call instruction and point orig_call to actual
2871 		 * body of the kernel function.
2872 		 */
2873 		if (is_endbr(*(u32 *)orig_call))
2874 			orig_call += ENDBR_INSN_SIZE;
2875 		orig_call += X86_PATCH_SIZE;
2876 	}
2877 
2878 	prog = rw_image;
2879 
2880 	if (flags & BPF_TRAMP_F_INDIRECT) {
2881 		/*
2882 		 * Indirect call for bpf_struct_ops
2883 		 */
2884 		emit_cfi(&prog, cfi_get_func_hash(func_addr));
2885 	} else {
2886 		/*
2887 		 * Direct-call fentry stub, as such it needs accounting for the
2888 		 * __fentry__ call.
2889 		 */
2890 		x86_call_depth_emit_accounting(&prog, NULL, image);
2891 	}
2892 	EMIT1(0x55);		 /* push rbp */
2893 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2894 	if (!is_imm8(stack_size)) {
2895 		/* sub rsp, stack_size */
2896 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
2897 	} else {
2898 		/* sub rsp, stack_size */
2899 		EMIT4(0x48, 0x83, 0xEC, stack_size);
2900 	}
2901 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2902 		EMIT1(0x50);		/* push rax */
2903 	/* mov QWORD PTR [rbp - rbx_off], rbx */
2904 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
2905 
2906 	/* Store number of argument registers of the traced function:
2907 	 *   mov rax, nr_regs
2908 	 *   mov QWORD PTR [rbp - nregs_off], rax
2909 	 */
2910 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2911 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2912 
2913 	if (flags & BPF_TRAMP_F_IP_ARG) {
2914 		/* Store IP address of the traced function:
2915 		 * movabsq rax, func_addr
2916 		 * mov QWORD PTR [rbp - ip_off], rax
2917 		 */
2918 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2919 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2920 	}
2921 
2922 	save_args(m, &prog, regs_off, false);
2923 
2924 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2925 		/* arg1: mov rdi, im */
2926 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2927 		if (emit_rsb_call(&prog, __bpf_tramp_enter,
2928 				  image + (prog - (u8 *)rw_image))) {
2929 			ret = -EINVAL;
2930 			goto cleanup;
2931 		}
2932 	}
2933 
2934 	if (fentry->nr_links) {
2935 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2936 			       flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image))
2937 			return -EINVAL;
2938 	}
2939 
2940 	if (fmod_ret->nr_links) {
2941 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2942 				   GFP_KERNEL);
2943 		if (!branches)
2944 			return -ENOMEM;
2945 
2946 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2947 				       run_ctx_off, branches, image, rw_image)) {
2948 			ret = -EINVAL;
2949 			goto cleanup;
2950 		}
2951 	}
2952 
2953 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2954 		restore_regs(m, &prog, regs_off);
2955 		save_args(m, &prog, arg_stack_off, true);
2956 
2957 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
2958 			/* Before calling the original function, restore the
2959 			 * tail_call_cnt from stack to rax.
2960 			 */
2961 			RESTORE_TAIL_CALL_CNT(stack_size);
2962 		}
2963 
2964 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2965 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
2966 			EMIT2(0xff, 0xd3); /* call *rbx */
2967 		} else {
2968 			/* call original function */
2969 			if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) {
2970 				ret = -EINVAL;
2971 				goto cleanup;
2972 			}
2973 		}
2974 		/* remember return value in a stack for bpf prog to access */
2975 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2976 		im->ip_after_call = image + (prog - (u8 *)rw_image);
2977 		emit_nops(&prog, X86_PATCH_SIZE);
2978 	}
2979 
2980 	if (fmod_ret->nr_links) {
2981 		/* From Intel 64 and IA-32 Architectures Optimization
2982 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2983 		 * Coding Rule 11: All branch targets should be 16-byte
2984 		 * aligned.
2985 		 */
2986 		emit_align(&prog, 16);
2987 		/* Update the branches saved in invoke_bpf_mod_ret with the
2988 		 * aligned address of do_fexit.
2989 		 */
2990 		for (i = 0; i < fmod_ret->nr_links; i++) {
2991 			emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image),
2992 					    image + (branches[i] - (u8 *)rw_image), X86_JNE);
2993 		}
2994 	}
2995 
2996 	if (fexit->nr_links) {
2997 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off,
2998 			       false, image, rw_image)) {
2999 			ret = -EINVAL;
3000 			goto cleanup;
3001 		}
3002 	}
3003 
3004 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
3005 		restore_regs(m, &prog, regs_off);
3006 
3007 	/* This needs to be done regardless. If there were fmod_ret programs,
3008 	 * the return value is only updated on the stack and still needs to be
3009 	 * restored to R0.
3010 	 */
3011 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3012 		im->ip_epilogue = image + (prog - (u8 *)rw_image);
3013 		/* arg1: mov rdi, im */
3014 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
3015 		if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) {
3016 			ret = -EINVAL;
3017 			goto cleanup;
3018 		}
3019 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
3020 		/* Before running the original function, restore the
3021 		 * tail_call_cnt from stack to rax.
3022 		 */
3023 		RESTORE_TAIL_CALL_CNT(stack_size);
3024 	}
3025 
3026 	/* restore return value of orig_call or fentry prog back into RAX */
3027 	if (save_ret)
3028 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
3029 
3030 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
3031 	EMIT1(0xC9); /* leave */
3032 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
3033 		/* skip our return address and return to parent */
3034 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
3035 	}
3036 	emit_return(&prog, image + (prog - (u8 *)rw_image));
3037 	/* Make sure the trampoline generation logic doesn't overflow */
3038 	if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) {
3039 		ret = -EFAULT;
3040 		goto cleanup;
3041 	}
3042 	ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY;
3043 
3044 cleanup:
3045 	kfree(branches);
3046 	return ret;
3047 }
3048 
3049 void *arch_alloc_bpf_trampoline(unsigned int size)
3050 {
3051 	return bpf_prog_pack_alloc(size, jit_fill_hole);
3052 }
3053 
3054 void arch_free_bpf_trampoline(void *image, unsigned int size)
3055 {
3056 	bpf_prog_pack_free(image, size);
3057 }
3058 
3059 int arch_protect_bpf_trampoline(void *image, unsigned int size)
3060 {
3061 	return 0;
3062 }
3063 
3064 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
3065 				const struct btf_func_model *m, u32 flags,
3066 				struct bpf_tramp_links *tlinks,
3067 				void *func_addr)
3068 {
3069 	void *rw_image, *tmp;
3070 	int ret;
3071 	u32 size = image_end - image;
3072 
3073 	/* rw_image doesn't need to be in module memory range, so we can
3074 	 * use kvmalloc.
3075 	 */
3076 	rw_image = kvmalloc(size, GFP_KERNEL);
3077 	if (!rw_image)
3078 		return -ENOMEM;
3079 
3080 	ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
3081 					    flags, tlinks, func_addr);
3082 	if (ret < 0)
3083 		goto out;
3084 
3085 	tmp = bpf_arch_text_copy(image, rw_image, size);
3086 	if (IS_ERR(tmp))
3087 		ret = PTR_ERR(tmp);
3088 out:
3089 	kvfree(rw_image);
3090 	return ret;
3091 }
3092 
3093 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
3094 			     struct bpf_tramp_links *tlinks, void *func_addr)
3095 {
3096 	struct bpf_tramp_image im;
3097 	void *image;
3098 	int ret;
3099 
3100 	/* Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
3101 	 * This will NOT cause fragmentation in direct map, as we do not
3102 	 * call set_memory_*() on this buffer.
3103 	 *
3104 	 * We cannot use kvmalloc here, because we need image to be in
3105 	 * module memory range.
3106 	 */
3107 	image = bpf_jit_alloc_exec(PAGE_SIZE);
3108 	if (!image)
3109 		return -ENOMEM;
3110 
3111 	ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
3112 					    m, flags, tlinks, func_addr);
3113 	bpf_jit_free_exec(image);
3114 	return ret;
3115 }
3116 
3117 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
3118 {
3119 	u8 *jg_reloc, *prog = *pprog;
3120 	int pivot, err, jg_bytes = 1;
3121 	s64 jg_offset;
3122 
3123 	if (a == b) {
3124 		/* Leaf node of recursion, i.e. not a range of indices
3125 		 * anymore.
3126 		 */
3127 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
3128 		if (!is_simm32(progs[a]))
3129 			return -1;
3130 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
3131 			    progs[a]);
3132 		err = emit_cond_near_jump(&prog,	/* je func */
3133 					  (void *)progs[a], image + (prog - buf),
3134 					  X86_JE);
3135 		if (err)
3136 			return err;
3137 
3138 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
3139 
3140 		*pprog = prog;
3141 		return 0;
3142 	}
3143 
3144 	/* Not a leaf node, so we pivot, and recursively descend into
3145 	 * the lower and upper ranges.
3146 	 */
3147 	pivot = (b - a) / 2;
3148 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
3149 	if (!is_simm32(progs[a + pivot]))
3150 		return -1;
3151 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
3152 
3153 	if (pivot > 2) {				/* jg upper_part */
3154 		/* Require near jump. */
3155 		jg_bytes = 4;
3156 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
3157 	} else {
3158 		EMIT2(X86_JG, 0);
3159 	}
3160 	jg_reloc = prog;
3161 
3162 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
3163 				  progs, image, buf);
3164 	if (err)
3165 		return err;
3166 
3167 	/* From Intel 64 and IA-32 Architectures Optimization
3168 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
3169 	 * Coding Rule 11: All branch targets should be 16-byte
3170 	 * aligned.
3171 	 */
3172 	emit_align(&prog, 16);
3173 	jg_offset = prog - jg_reloc;
3174 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
3175 
3176 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
3177 				  b, progs, image, buf);
3178 	if (err)
3179 		return err;
3180 
3181 	*pprog = prog;
3182 	return 0;
3183 }
3184 
3185 static int cmp_ips(const void *a, const void *b)
3186 {
3187 	const s64 *ipa = a;
3188 	const s64 *ipb = b;
3189 
3190 	if (*ipa > *ipb)
3191 		return 1;
3192 	if (*ipa < *ipb)
3193 		return -1;
3194 	return 0;
3195 }
3196 
3197 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
3198 {
3199 	u8 *prog = buf;
3200 
3201 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
3202 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
3203 }
3204 
3205 struct x64_jit_data {
3206 	struct bpf_binary_header *rw_header;
3207 	struct bpf_binary_header *header;
3208 	int *addrs;
3209 	u8 *image;
3210 	int proglen;
3211 	struct jit_context ctx;
3212 };
3213 
3214 #define MAX_PASSES 20
3215 #define PADDING_PASSES (MAX_PASSES - 5)
3216 
3217 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
3218 {
3219 	struct bpf_binary_header *rw_header = NULL;
3220 	struct bpf_binary_header *header = NULL;
3221 	struct bpf_prog *tmp, *orig_prog = prog;
3222 	struct x64_jit_data *jit_data;
3223 	int proglen, oldproglen = 0;
3224 	struct jit_context ctx = {};
3225 	bool tmp_blinded = false;
3226 	bool extra_pass = false;
3227 	bool padding = false;
3228 	u8 *rw_image = NULL;
3229 	u8 *image = NULL;
3230 	int *addrs;
3231 	int pass;
3232 	int i;
3233 
3234 	if (!prog->jit_requested)
3235 		return orig_prog;
3236 
3237 	tmp = bpf_jit_blind_constants(prog);
3238 	/*
3239 	 * If blinding was requested and we failed during blinding,
3240 	 * we must fall back to the interpreter.
3241 	 */
3242 	if (IS_ERR(tmp))
3243 		return orig_prog;
3244 	if (tmp != prog) {
3245 		tmp_blinded = true;
3246 		prog = tmp;
3247 	}
3248 
3249 	jit_data = prog->aux->jit_data;
3250 	if (!jit_data) {
3251 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
3252 		if (!jit_data) {
3253 			prog = orig_prog;
3254 			goto out;
3255 		}
3256 		prog->aux->jit_data = jit_data;
3257 	}
3258 	addrs = jit_data->addrs;
3259 	if (addrs) {
3260 		ctx = jit_data->ctx;
3261 		oldproglen = jit_data->proglen;
3262 		image = jit_data->image;
3263 		header = jit_data->header;
3264 		rw_header = jit_data->rw_header;
3265 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
3266 		extra_pass = true;
3267 		padding = true;
3268 		goto skip_init_addrs;
3269 	}
3270 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
3271 	if (!addrs) {
3272 		prog = orig_prog;
3273 		goto out_addrs;
3274 	}
3275 
3276 	/*
3277 	 * Before first pass, make a rough estimation of addrs[]
3278 	 * each BPF instruction is translated to less than 64 bytes
3279 	 */
3280 	for (proglen = 0, i = 0; i <= prog->len; i++) {
3281 		proglen += 64;
3282 		addrs[i] = proglen;
3283 	}
3284 	ctx.cleanup_addr = proglen;
3285 skip_init_addrs:
3286 
3287 	/*
3288 	 * JITed image shrinks with every pass and the loop iterates
3289 	 * until the image stops shrinking. Very large BPF programs
3290 	 * may converge on the last pass. In such case do one more
3291 	 * pass to emit the final image.
3292 	 */
3293 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
3294 		if (!padding && pass >= PADDING_PASSES)
3295 			padding = true;
3296 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
3297 		if (proglen <= 0) {
3298 out_image:
3299 			image = NULL;
3300 			if (header) {
3301 				bpf_arch_text_copy(&header->size, &rw_header->size,
3302 						   sizeof(rw_header->size));
3303 				bpf_jit_binary_pack_free(header, rw_header);
3304 			}
3305 			/* Fall back to interpreter mode */
3306 			prog = orig_prog;
3307 			if (extra_pass) {
3308 				prog->bpf_func = NULL;
3309 				prog->jited = 0;
3310 				prog->jited_len = 0;
3311 			}
3312 			goto out_addrs;
3313 		}
3314 		if (image) {
3315 			if (proglen != oldproglen) {
3316 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
3317 				       proglen, oldproglen);
3318 				goto out_image;
3319 			}
3320 			break;
3321 		}
3322 		if (proglen == oldproglen) {
3323 			/*
3324 			 * The number of entries in extable is the number of BPF_LDX
3325 			 * insns that access kernel memory via "pointer to BTF type".
3326 			 * The verifier changed their opcode from LDX|MEM|size
3327 			 * to LDX|PROBE_MEM|size to make JITing easier.
3328 			 */
3329 			u32 align = __alignof__(struct exception_table_entry);
3330 			u32 extable_size = prog->aux->num_exentries *
3331 				sizeof(struct exception_table_entry);
3332 
3333 			/* allocate module memory for x86 insns and extable */
3334 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
3335 							   &image, align, &rw_header, &rw_image,
3336 							   jit_fill_hole);
3337 			if (!header) {
3338 				prog = orig_prog;
3339 				goto out_addrs;
3340 			}
3341 			prog->aux->extable = (void *) image + roundup(proglen, align);
3342 		}
3343 		oldproglen = proglen;
3344 		cond_resched();
3345 	}
3346 
3347 	if (bpf_jit_enable > 1)
3348 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
3349 
3350 	if (image) {
3351 		if (!prog->is_func || extra_pass) {
3352 			/*
3353 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
3354 			 *   1) header is not pointing to proper module memory;
3355 			 *   2) the arch doesn't support bpf_arch_text_copy().
3356 			 *
3357 			 * Both cases are serious bugs and justify WARN_ON.
3358 			 */
3359 			if (WARN_ON(bpf_jit_binary_pack_finalize(header, rw_header))) {
3360 				/* header has been freed */
3361 				header = NULL;
3362 				goto out_image;
3363 			}
3364 
3365 			bpf_tail_call_direct_fixup(prog);
3366 		} else {
3367 			jit_data->addrs = addrs;
3368 			jit_data->ctx = ctx;
3369 			jit_data->proglen = proglen;
3370 			jit_data->image = image;
3371 			jit_data->header = header;
3372 			jit_data->rw_header = rw_header;
3373 		}
3374 		/*
3375 		 * ctx.prog_offset is used when CFI preambles put code *before*
3376 		 * the function. See emit_cfi(). For FineIBT specifically this code
3377 		 * can also be executed and bpf_prog_kallsyms_add() will
3378 		 * generate an additional symbol to cover this, hence also
3379 		 * decrement proglen.
3380 		 */
3381 		prog->bpf_func = (void *)image + cfi_get_offset();
3382 		prog->jited = 1;
3383 		prog->jited_len = proglen - cfi_get_offset();
3384 	} else {
3385 		prog = orig_prog;
3386 	}
3387 
3388 	if (!image || !prog->is_func || extra_pass) {
3389 		if (image)
3390 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
3391 out_addrs:
3392 		kvfree(addrs);
3393 		kfree(jit_data);
3394 		prog->aux->jit_data = NULL;
3395 	}
3396 out:
3397 	if (tmp_blinded)
3398 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
3399 					   tmp : orig_prog);
3400 	return prog;
3401 }
3402 
3403 bool bpf_jit_supports_kfunc_call(void)
3404 {
3405 	return true;
3406 }
3407 
3408 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
3409 {
3410 	if (text_poke_copy(dst, src, len) == NULL)
3411 		return ERR_PTR(-EINVAL);
3412 	return dst;
3413 }
3414 
3415 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
3416 bool bpf_jit_supports_subprog_tailcalls(void)
3417 {
3418 	return true;
3419 }
3420 
3421 bool bpf_jit_supports_percpu_insn(void)
3422 {
3423 	return true;
3424 }
3425 
3426 void bpf_jit_free(struct bpf_prog *prog)
3427 {
3428 	if (prog->jited) {
3429 		struct x64_jit_data *jit_data = prog->aux->jit_data;
3430 		struct bpf_binary_header *hdr;
3431 
3432 		/*
3433 		 * If we fail the final pass of JIT (from jit_subprogs),
3434 		 * the program may not be finalized yet. Call finalize here
3435 		 * before freeing it.
3436 		 */
3437 		if (jit_data) {
3438 			bpf_jit_binary_pack_finalize(jit_data->header,
3439 						     jit_data->rw_header);
3440 			kvfree(jit_data->addrs);
3441 			kfree(jit_data);
3442 		}
3443 		prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset();
3444 		hdr = bpf_jit_binary_pack_hdr(prog);
3445 		bpf_jit_binary_pack_free(hdr, NULL);
3446 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
3447 	}
3448 
3449 	bpf_prog_unlock_free(prog);
3450 }
3451 
3452 bool bpf_jit_supports_exceptions(void)
3453 {
3454 	/* We unwind through both kernel frames (starting from within bpf_throw
3455 	 * call) and BPF frames. Therefore we require ORC unwinder to be enabled
3456 	 * to walk kernel frames and reach BPF frames in the stack trace.
3457 	 */
3458 	return IS_ENABLED(CONFIG_UNWINDER_ORC);
3459 }
3460 
3461 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3462 {
3463 #if defined(CONFIG_UNWINDER_ORC)
3464 	struct unwind_state state;
3465 	unsigned long addr;
3466 
3467 	for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
3468 	     unwind_next_frame(&state)) {
3469 		addr = unwind_get_return_address(&state);
3470 		if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
3471 			break;
3472 	}
3473 	return;
3474 #endif
3475 	WARN(1, "verification of programs using bpf_throw should have failed\n");
3476 }
3477 
3478 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
3479 			       struct bpf_prog *new, struct bpf_prog *old)
3480 {
3481 	u8 *old_addr, *new_addr, *old_bypass_addr;
3482 	int ret;
3483 
3484 	old_bypass_addr = old ? NULL : poke->bypass_addr;
3485 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
3486 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
3487 
3488 	/*
3489 	 * On program loading or teardown, the program's kallsym entry
3490 	 * might not be in place, so we use __bpf_arch_text_poke to skip
3491 	 * the kallsyms check.
3492 	 */
3493 	if (new) {
3494 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3495 					   BPF_MOD_JUMP,
3496 					   old_addr, new_addr);
3497 		BUG_ON(ret < 0);
3498 		if (!old) {
3499 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3500 						   BPF_MOD_JUMP,
3501 						   poke->bypass_addr,
3502 						   NULL);
3503 			BUG_ON(ret < 0);
3504 		}
3505 	} else {
3506 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3507 					   BPF_MOD_JUMP,
3508 					   old_bypass_addr,
3509 					   poke->bypass_addr);
3510 		BUG_ON(ret < 0);
3511 		/* let other CPUs finish the execution of program
3512 		 * so that it will not possible to expose them
3513 		 * to invalid nop, stack unwind, nop state
3514 		 */
3515 		if (!ret)
3516 			synchronize_rcu();
3517 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3518 					   BPF_MOD_JUMP,
3519 					   old_addr, NULL);
3520 		BUG_ON(ret < 0);
3521 	}
3522 }
3523 
3524 bool bpf_jit_supports_arena(void)
3525 {
3526 	return true;
3527 }
3528 
3529 bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
3530 {
3531 	if (!in_arena)
3532 		return true;
3533 	switch (insn->code) {
3534 	case BPF_STX | BPF_ATOMIC | BPF_W:
3535 	case BPF_STX | BPF_ATOMIC | BPF_DW:
3536 		if (insn->imm == (BPF_AND | BPF_FETCH) ||
3537 		    insn->imm == (BPF_OR | BPF_FETCH) ||
3538 		    insn->imm == (BPF_XOR | BPF_FETCH))
3539 			return false;
3540 	}
3541 	return true;
3542 }
3543 
3544 bool bpf_jit_supports_ptr_xchg(void)
3545 {
3546 	return true;
3547 }
3548 
3549 /* x86-64 JIT emits its own code to filter user addresses so return 0 here */
3550 u64 bpf_arch_uaddress_limit(void)
3551 {
3552 	return 0;
3553 }
3554