xref: /linux/arch/x86/net/bpf_jit_comp.c (revision 32a4ca1361d7a51e5003d4af4dfbf570f1b5fd00)
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, bool *tail_call_seen)
1238 {
1239 	int i;
1240 
1241 	for (i = 1; i <= insn_cnt; i++, insn++) {
1242 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
1243 			*tail_call_seen = true;
1244 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
1245 			regs_used[0] = true;
1246 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
1247 			regs_used[1] = true;
1248 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
1249 			regs_used[2] = true;
1250 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
1251 			regs_used[3] = true;
1252 	}
1253 }
1254 
1255 /* emit the 3-byte VEX prefix
1256  *
1257  * r: same as rex.r, extra bit for ModRM reg field
1258  * x: same as rex.x, extra bit for SIB index field
1259  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
1260  * m: opcode map select, encoding escape bytes e.g. 0x0f38
1261  * w: same as rex.w (32 bit or 64 bit) or opcode specific
1262  * src_reg2: additional source reg (encoded as BPF reg)
1263  * l: vector length (128 bit or 256 bit) or reserved
1264  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
1265  */
1266 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
1267 		      bool w, u8 src_reg2, bool l, u8 pp)
1268 {
1269 	u8 *prog = *pprog;
1270 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
1271 	u8 b1, b2;
1272 	u8 vvvv = reg2hex[src_reg2];
1273 
1274 	/* reg2hex gives only the lower 3 bit of vvvv */
1275 	if (is_ereg(src_reg2))
1276 		vvvv |= 1 << 3;
1277 
1278 	/*
1279 	 * 2nd byte of 3-byte VEX prefix
1280 	 * ~ means bit inverted encoding
1281 	 *
1282 	 *    7                           0
1283 	 *  +---+---+---+---+---+---+---+---+
1284 	 *  |~R |~X |~B |         m         |
1285 	 *  +---+---+---+---+---+---+---+---+
1286 	 */
1287 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
1288 	/*
1289 	 * 3rd byte of 3-byte VEX prefix
1290 	 *
1291 	 *    7                           0
1292 	 *  +---+---+---+---+---+---+---+---+
1293 	 *  | W |     ~vvvv     | L |   pp  |
1294 	 *  +---+---+---+---+---+---+---+---+
1295 	 */
1296 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1297 
1298 	EMIT3(b0, b1, b2);
1299 	*pprog = prog;
1300 }
1301 
1302 /* emit BMI2 shift instruction */
1303 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1304 {
1305 	u8 *prog = *pprog;
1306 	bool r = is_ereg(dst_reg);
1307 	u8 m = 2; /* escape code 0f38 */
1308 
1309 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1310 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1311 	*pprog = prog;
1312 }
1313 
1314 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1315 
1316 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1317 #define RESTORE_TAIL_CALL_CNT(stack)				\
1318 	EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
1319 
1320 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
1321 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
1322 {
1323 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1324 	struct bpf_insn *insn = bpf_prog->insnsi;
1325 	bool callee_regs_used[4] = {};
1326 	int insn_cnt = bpf_prog->len;
1327 	bool tail_call_seen = false;
1328 	bool seen_exit = false;
1329 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1330 	u64 arena_vm_start, user_vm_start;
1331 	int i, excnt = 0;
1332 	int ilen, proglen = 0;
1333 	u8 *prog = temp;
1334 	int err;
1335 
1336 	arena_vm_start = bpf_arena_get_kern_vm_start(bpf_prog->aux->arena);
1337 	user_vm_start = bpf_arena_get_user_vm_start(bpf_prog->aux->arena);
1338 
1339 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
1340 			 &tail_call_seen);
1341 
1342 	/* tail call's presence in current prog implies it is reachable */
1343 	tail_call_reachable |= tail_call_seen;
1344 
1345 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
1346 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1347 		      bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
1348 	/* Exception callback will clobber callee regs for its own use, and
1349 	 * restore the original callee regs from main prog's stack frame.
1350 	 */
1351 	if (bpf_prog->aux->exception_boundary) {
1352 		/* We also need to save r12, which is not mapped to any BPF
1353 		 * register, as we throw after entry into the kernel, which may
1354 		 * overwrite r12.
1355 		 */
1356 		push_r12(&prog);
1357 		push_callee_regs(&prog, all_callee_regs_used);
1358 	} else {
1359 		if (arena_vm_start)
1360 			push_r12(&prog);
1361 		push_callee_regs(&prog, callee_regs_used);
1362 	}
1363 	if (arena_vm_start)
1364 		emit_mov_imm64(&prog, X86_REG_R12,
1365 			       arena_vm_start >> 32, (u32) arena_vm_start);
1366 
1367 	ilen = prog - temp;
1368 	if (rw_image)
1369 		memcpy(rw_image + proglen, temp, ilen);
1370 	proglen += ilen;
1371 	addrs[0] = proglen;
1372 	prog = temp;
1373 
1374 	for (i = 1; i <= insn_cnt; i++, insn++) {
1375 		const s32 imm32 = insn->imm;
1376 		u32 dst_reg = insn->dst_reg;
1377 		u32 src_reg = insn->src_reg;
1378 		u8 b2 = 0, b3 = 0;
1379 		u8 *start_of_ldx;
1380 		s64 jmp_offset;
1381 		s16 insn_off;
1382 		u8 jmp_cond;
1383 		u8 *func;
1384 		int nops;
1385 
1386 		switch (insn->code) {
1387 			/* ALU */
1388 		case BPF_ALU | BPF_ADD | BPF_X:
1389 		case BPF_ALU | BPF_SUB | BPF_X:
1390 		case BPF_ALU | BPF_AND | BPF_X:
1391 		case BPF_ALU | BPF_OR | BPF_X:
1392 		case BPF_ALU | BPF_XOR | BPF_X:
1393 		case BPF_ALU64 | BPF_ADD | BPF_X:
1394 		case BPF_ALU64 | BPF_SUB | BPF_X:
1395 		case BPF_ALU64 | BPF_AND | BPF_X:
1396 		case BPF_ALU64 | BPF_OR | BPF_X:
1397 		case BPF_ALU64 | BPF_XOR | BPF_X:
1398 			maybe_emit_mod(&prog, dst_reg, src_reg,
1399 				       BPF_CLASS(insn->code) == BPF_ALU64);
1400 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1401 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1402 			break;
1403 
1404 		case BPF_ALU64 | BPF_MOV | BPF_X:
1405 			if (insn_is_cast_user(insn)) {
1406 				if (dst_reg != src_reg)
1407 					/* 32-bit mov */
1408 					emit_mov_reg(&prog, false, dst_reg, src_reg);
1409 				/* shl dst_reg, 32 */
1410 				maybe_emit_1mod(&prog, dst_reg, true);
1411 				EMIT3(0xC1, add_1reg(0xE0, dst_reg), 32);
1412 
1413 				/* or dst_reg, user_vm_start */
1414 				maybe_emit_1mod(&prog, dst_reg, true);
1415 				if (is_axreg(dst_reg))
1416 					EMIT1_off32(0x0D,  user_vm_start >> 32);
1417 				else
1418 					EMIT2_off32(0x81, add_1reg(0xC8, dst_reg),  user_vm_start >> 32);
1419 
1420 				/* rol dst_reg, 32 */
1421 				maybe_emit_1mod(&prog, dst_reg, true);
1422 				EMIT3(0xC1, add_1reg(0xC0, dst_reg), 32);
1423 
1424 				/* xor r11, r11 */
1425 				EMIT3(0x4D, 0x31, 0xDB);
1426 
1427 				/* test dst_reg32, dst_reg32; check if lower 32-bit are zero */
1428 				maybe_emit_mod(&prog, dst_reg, dst_reg, false);
1429 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1430 
1431 				/* cmove r11, dst_reg; if so, set dst_reg to zero */
1432 				/* WARNING: Intel swapped src/dst register encoding in CMOVcc !!! */
1433 				maybe_emit_mod(&prog, AUX_REG, dst_reg, true);
1434 				EMIT3(0x0F, 0x44, add_2reg(0xC0, AUX_REG, dst_reg));
1435 				break;
1436 			} else if (insn_is_mov_percpu_addr(insn)) {
1437 				/* mov <dst>, <src> (if necessary) */
1438 				EMIT_mov(dst_reg, src_reg);
1439 #ifdef CONFIG_SMP
1440 				/* add <dst>, gs:[<off>] */
1441 				EMIT2(0x65, add_1mod(0x48, dst_reg));
1442 				EMIT3(0x03, add_2reg(0x04, 0, dst_reg), 0x25);
1443 				EMIT((u32)(unsigned long)&this_cpu_off, 4);
1444 #endif
1445 				break;
1446 			}
1447 			fallthrough;
1448 		case BPF_ALU | BPF_MOV | BPF_X:
1449 			if (insn->off == 0)
1450 				emit_mov_reg(&prog,
1451 					     BPF_CLASS(insn->code) == BPF_ALU64,
1452 					     dst_reg, src_reg);
1453 			else
1454 				emit_movsx_reg(&prog, insn->off,
1455 					       BPF_CLASS(insn->code) == BPF_ALU64,
1456 					       dst_reg, src_reg);
1457 			break;
1458 
1459 			/* neg dst */
1460 		case BPF_ALU | BPF_NEG:
1461 		case BPF_ALU64 | BPF_NEG:
1462 			maybe_emit_1mod(&prog, dst_reg,
1463 					BPF_CLASS(insn->code) == BPF_ALU64);
1464 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1465 			break;
1466 
1467 		case BPF_ALU | BPF_ADD | BPF_K:
1468 		case BPF_ALU | BPF_SUB | BPF_K:
1469 		case BPF_ALU | BPF_AND | BPF_K:
1470 		case BPF_ALU | BPF_OR | BPF_K:
1471 		case BPF_ALU | BPF_XOR | BPF_K:
1472 		case BPF_ALU64 | BPF_ADD | BPF_K:
1473 		case BPF_ALU64 | BPF_SUB | BPF_K:
1474 		case BPF_ALU64 | BPF_AND | BPF_K:
1475 		case BPF_ALU64 | BPF_OR | BPF_K:
1476 		case BPF_ALU64 | BPF_XOR | BPF_K:
1477 			maybe_emit_1mod(&prog, dst_reg,
1478 					BPF_CLASS(insn->code) == BPF_ALU64);
1479 
1480 			/*
1481 			 * b3 holds 'normal' opcode, b2 short form only valid
1482 			 * in case dst is eax/rax.
1483 			 */
1484 			switch (BPF_OP(insn->code)) {
1485 			case BPF_ADD:
1486 				b3 = 0xC0;
1487 				b2 = 0x05;
1488 				break;
1489 			case BPF_SUB:
1490 				b3 = 0xE8;
1491 				b2 = 0x2D;
1492 				break;
1493 			case BPF_AND:
1494 				b3 = 0xE0;
1495 				b2 = 0x25;
1496 				break;
1497 			case BPF_OR:
1498 				b3 = 0xC8;
1499 				b2 = 0x0D;
1500 				break;
1501 			case BPF_XOR:
1502 				b3 = 0xF0;
1503 				b2 = 0x35;
1504 				break;
1505 			}
1506 
1507 			if (is_imm8(imm32))
1508 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1509 			else if (is_axreg(dst_reg))
1510 				EMIT1_off32(b2, imm32);
1511 			else
1512 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1513 			break;
1514 
1515 		case BPF_ALU64 | BPF_MOV | BPF_K:
1516 		case BPF_ALU | BPF_MOV | BPF_K:
1517 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1518 				       dst_reg, imm32);
1519 			break;
1520 
1521 		case BPF_LD | BPF_IMM | BPF_DW:
1522 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1523 			insn++;
1524 			i++;
1525 			break;
1526 
1527 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1528 		case BPF_ALU | BPF_MOD | BPF_X:
1529 		case BPF_ALU | BPF_DIV | BPF_X:
1530 		case BPF_ALU | BPF_MOD | BPF_K:
1531 		case BPF_ALU | BPF_DIV | BPF_K:
1532 		case BPF_ALU64 | BPF_MOD | BPF_X:
1533 		case BPF_ALU64 | BPF_DIV | BPF_X:
1534 		case BPF_ALU64 | BPF_MOD | BPF_K:
1535 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1536 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1537 
1538 			if (dst_reg != BPF_REG_0)
1539 				EMIT1(0x50); /* push rax */
1540 			if (dst_reg != BPF_REG_3)
1541 				EMIT1(0x52); /* push rdx */
1542 
1543 			if (BPF_SRC(insn->code) == BPF_X) {
1544 				if (src_reg == BPF_REG_0 ||
1545 				    src_reg == BPF_REG_3) {
1546 					/* mov r11, src_reg */
1547 					EMIT_mov(AUX_REG, src_reg);
1548 					src_reg = AUX_REG;
1549 				}
1550 			} else {
1551 				/* mov r11, imm32 */
1552 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1553 				src_reg = AUX_REG;
1554 			}
1555 
1556 			if (dst_reg != BPF_REG_0)
1557 				/* mov rax, dst_reg */
1558 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1559 
1560 			if (insn->off == 0) {
1561 				/*
1562 				 * xor edx, edx
1563 				 * equivalent to 'xor rdx, rdx', but one byte less
1564 				 */
1565 				EMIT2(0x31, 0xd2);
1566 
1567 				/* div src_reg */
1568 				maybe_emit_1mod(&prog, src_reg, is64);
1569 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1570 			} else {
1571 				if (BPF_CLASS(insn->code) == BPF_ALU)
1572 					EMIT1(0x99); /* cdq */
1573 				else
1574 					EMIT2(0x48, 0x99); /* cqo */
1575 
1576 				/* idiv src_reg */
1577 				maybe_emit_1mod(&prog, src_reg, is64);
1578 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1579 			}
1580 
1581 			if (BPF_OP(insn->code) == BPF_MOD &&
1582 			    dst_reg != BPF_REG_3)
1583 				/* mov dst_reg, rdx */
1584 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1585 			else if (BPF_OP(insn->code) == BPF_DIV &&
1586 				 dst_reg != BPF_REG_0)
1587 				/* mov dst_reg, rax */
1588 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1589 
1590 			if (dst_reg != BPF_REG_3)
1591 				EMIT1(0x5A); /* pop rdx */
1592 			if (dst_reg != BPF_REG_0)
1593 				EMIT1(0x58); /* pop rax */
1594 			break;
1595 		}
1596 
1597 		case BPF_ALU | BPF_MUL | BPF_K:
1598 		case BPF_ALU64 | BPF_MUL | BPF_K:
1599 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1600 				       BPF_CLASS(insn->code) == BPF_ALU64);
1601 
1602 			if (is_imm8(imm32))
1603 				/* imul dst_reg, dst_reg, imm8 */
1604 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1605 				      imm32);
1606 			else
1607 				/* imul dst_reg, dst_reg, imm32 */
1608 				EMIT2_off32(0x69,
1609 					    add_2reg(0xC0, dst_reg, dst_reg),
1610 					    imm32);
1611 			break;
1612 
1613 		case BPF_ALU | BPF_MUL | BPF_X:
1614 		case BPF_ALU64 | BPF_MUL | BPF_X:
1615 			maybe_emit_mod(&prog, src_reg, dst_reg,
1616 				       BPF_CLASS(insn->code) == BPF_ALU64);
1617 
1618 			/* imul dst_reg, src_reg */
1619 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1620 			break;
1621 
1622 			/* Shifts */
1623 		case BPF_ALU | BPF_LSH | BPF_K:
1624 		case BPF_ALU | BPF_RSH | BPF_K:
1625 		case BPF_ALU | BPF_ARSH | BPF_K:
1626 		case BPF_ALU64 | BPF_LSH | BPF_K:
1627 		case BPF_ALU64 | BPF_RSH | BPF_K:
1628 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1629 			maybe_emit_1mod(&prog, dst_reg,
1630 					BPF_CLASS(insn->code) == BPF_ALU64);
1631 
1632 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1633 			if (imm32 == 1)
1634 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1635 			else
1636 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1637 			break;
1638 
1639 		case BPF_ALU | BPF_LSH | BPF_X:
1640 		case BPF_ALU | BPF_RSH | BPF_X:
1641 		case BPF_ALU | BPF_ARSH | BPF_X:
1642 		case BPF_ALU64 | BPF_LSH | BPF_X:
1643 		case BPF_ALU64 | BPF_RSH | BPF_X:
1644 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1645 			/* BMI2 shifts aren't better when shift count is already in rcx */
1646 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1647 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1648 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1649 				u8 op;
1650 
1651 				switch (BPF_OP(insn->code)) {
1652 				case BPF_LSH:
1653 					op = 1; /* prefix 0x66 */
1654 					break;
1655 				case BPF_RSH:
1656 					op = 3; /* prefix 0xf2 */
1657 					break;
1658 				case BPF_ARSH:
1659 					op = 2; /* prefix 0xf3 */
1660 					break;
1661 				}
1662 
1663 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1664 
1665 				break;
1666 			}
1667 
1668 			if (src_reg != BPF_REG_4) { /* common case */
1669 				/* Check for bad case when dst_reg == rcx */
1670 				if (dst_reg == BPF_REG_4) {
1671 					/* mov r11, dst_reg */
1672 					EMIT_mov(AUX_REG, dst_reg);
1673 					dst_reg = AUX_REG;
1674 				} else {
1675 					EMIT1(0x51); /* push rcx */
1676 				}
1677 				/* mov rcx, src_reg */
1678 				EMIT_mov(BPF_REG_4, src_reg);
1679 			}
1680 
1681 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1682 			maybe_emit_1mod(&prog, dst_reg,
1683 					BPF_CLASS(insn->code) == BPF_ALU64);
1684 
1685 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1686 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1687 
1688 			if (src_reg != BPF_REG_4) {
1689 				if (insn->dst_reg == BPF_REG_4)
1690 					/* mov dst_reg, r11 */
1691 					EMIT_mov(insn->dst_reg, AUX_REG);
1692 				else
1693 					EMIT1(0x59); /* pop rcx */
1694 			}
1695 
1696 			break;
1697 
1698 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1699 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1700 			switch (imm32) {
1701 			case 16:
1702 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1703 				EMIT1(0x66);
1704 				if (is_ereg(dst_reg))
1705 					EMIT1(0x41);
1706 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1707 
1708 				/* Emit 'movzwl eax, ax' */
1709 				if (is_ereg(dst_reg))
1710 					EMIT3(0x45, 0x0F, 0xB7);
1711 				else
1712 					EMIT2(0x0F, 0xB7);
1713 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1714 				break;
1715 			case 32:
1716 				/* Emit 'bswap eax' to swap lower 4 bytes */
1717 				if (is_ereg(dst_reg))
1718 					EMIT2(0x41, 0x0F);
1719 				else
1720 					EMIT1(0x0F);
1721 				EMIT1(add_1reg(0xC8, dst_reg));
1722 				break;
1723 			case 64:
1724 				/* Emit 'bswap rax' to swap 8 bytes */
1725 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1726 				      add_1reg(0xC8, dst_reg));
1727 				break;
1728 			}
1729 			break;
1730 
1731 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1732 			switch (imm32) {
1733 			case 16:
1734 				/*
1735 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1736 				 * into 64 bit
1737 				 */
1738 				if (is_ereg(dst_reg))
1739 					EMIT3(0x45, 0x0F, 0xB7);
1740 				else
1741 					EMIT2(0x0F, 0xB7);
1742 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1743 				break;
1744 			case 32:
1745 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1746 				if (is_ereg(dst_reg))
1747 					EMIT1(0x45);
1748 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1749 				break;
1750 			case 64:
1751 				/* nop */
1752 				break;
1753 			}
1754 			break;
1755 
1756 			/* speculation barrier */
1757 		case BPF_ST | BPF_NOSPEC:
1758 			EMIT_LFENCE();
1759 			break;
1760 
1761 			/* ST: *(u8*)(dst_reg + off) = imm */
1762 		case BPF_ST | BPF_MEM | BPF_B:
1763 			if (is_ereg(dst_reg))
1764 				EMIT2(0x41, 0xC6);
1765 			else
1766 				EMIT1(0xC6);
1767 			goto st;
1768 		case BPF_ST | BPF_MEM | BPF_H:
1769 			if (is_ereg(dst_reg))
1770 				EMIT3(0x66, 0x41, 0xC7);
1771 			else
1772 				EMIT2(0x66, 0xC7);
1773 			goto st;
1774 		case BPF_ST | BPF_MEM | BPF_W:
1775 			if (is_ereg(dst_reg))
1776 				EMIT2(0x41, 0xC7);
1777 			else
1778 				EMIT1(0xC7);
1779 			goto st;
1780 		case BPF_ST | BPF_MEM | BPF_DW:
1781 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1782 
1783 st:			if (is_imm8(insn->off))
1784 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1785 			else
1786 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1787 
1788 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1789 			break;
1790 
1791 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1792 		case BPF_STX | BPF_MEM | BPF_B:
1793 		case BPF_STX | BPF_MEM | BPF_H:
1794 		case BPF_STX | BPF_MEM | BPF_W:
1795 		case BPF_STX | BPF_MEM | BPF_DW:
1796 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1797 			break;
1798 
1799 		case BPF_ST | BPF_PROBE_MEM32 | BPF_B:
1800 		case BPF_ST | BPF_PROBE_MEM32 | BPF_H:
1801 		case BPF_ST | BPF_PROBE_MEM32 | BPF_W:
1802 		case BPF_ST | BPF_PROBE_MEM32 | BPF_DW:
1803 			start_of_ldx = prog;
1804 			emit_st_r12(&prog, BPF_SIZE(insn->code), dst_reg, insn->off, insn->imm);
1805 			goto populate_extable;
1806 
1807 			/* LDX: dst_reg = *(u8*)(src_reg + r12 + off) */
1808 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_B:
1809 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
1810 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
1811 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
1812 		case BPF_STX | BPF_PROBE_MEM32 | BPF_B:
1813 		case BPF_STX | BPF_PROBE_MEM32 | BPF_H:
1814 		case BPF_STX | BPF_PROBE_MEM32 | BPF_W:
1815 		case BPF_STX | BPF_PROBE_MEM32 | BPF_DW:
1816 			start_of_ldx = prog;
1817 			if (BPF_CLASS(insn->code) == BPF_LDX)
1818 				emit_ldx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1819 			else
1820 				emit_stx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1821 populate_extable:
1822 			{
1823 				struct exception_table_entry *ex;
1824 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1825 				s64 delta;
1826 
1827 				if (!bpf_prog->aux->extable)
1828 					break;
1829 
1830 				if (excnt >= bpf_prog->aux->num_exentries) {
1831 					pr_err("mem32 extable bug\n");
1832 					return -EFAULT;
1833 				}
1834 				ex = &bpf_prog->aux->extable[excnt++];
1835 
1836 				delta = _insn - (u8 *)&ex->insn;
1837 				/* switch ex to rw buffer for writes */
1838 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1839 
1840 				ex->insn = delta;
1841 
1842 				ex->data = EX_TYPE_BPF;
1843 
1844 				ex->fixup = (prog - start_of_ldx) |
1845 					((BPF_CLASS(insn->code) == BPF_LDX ? reg2pt_regs[dst_reg] : DONT_CLEAR) << 8);
1846 			}
1847 			break;
1848 
1849 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1850 		case BPF_LDX | BPF_MEM | BPF_B:
1851 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1852 		case BPF_LDX | BPF_MEM | BPF_H:
1853 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1854 		case BPF_LDX | BPF_MEM | BPF_W:
1855 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1856 		case BPF_LDX | BPF_MEM | BPF_DW:
1857 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1858 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
1859 		case BPF_LDX | BPF_MEMSX | BPF_B:
1860 		case BPF_LDX | BPF_MEMSX | BPF_H:
1861 		case BPF_LDX | BPF_MEMSX | BPF_W:
1862 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1863 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1864 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1865 			insn_off = insn->off;
1866 
1867 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1868 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1869 				/* Conservatively check that src_reg + insn->off is a kernel address:
1870 				 *   src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
1871 				 * src_reg is used as scratch for src_reg += insn->off and restored
1872 				 * after emit_ldx if necessary
1873 				 */
1874 
1875 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
1876 				u8 *end_of_jmp;
1877 
1878 				/* At end of these emitted checks, insn->off will have been added
1879 				 * to src_reg, so no need to do relative load with insn->off offset
1880 				 */
1881 				insn_off = 0;
1882 
1883 				/* movabsq r11, limit */
1884 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1885 				EMIT((u32)limit, 4);
1886 				EMIT(limit >> 32, 4);
1887 
1888 				if (insn->off) {
1889 					/* add src_reg, insn->off */
1890 					maybe_emit_1mod(&prog, src_reg, true);
1891 					EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
1892 				}
1893 
1894 				/* cmp src_reg, r11 */
1895 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1896 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1897 
1898 				/* if unsigned '>=', goto load */
1899 				EMIT2(X86_JAE, 0);
1900 				end_of_jmp = prog;
1901 
1902 				/* xor dst_reg, dst_reg */
1903 				emit_mov_imm32(&prog, false, dst_reg, 0);
1904 				/* jmp byte_after_ldx */
1905 				EMIT2(0xEB, 0);
1906 
1907 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1908 				start_of_ldx = prog;
1909 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1910 			}
1911 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
1912 			    BPF_MODE(insn->code) == BPF_MEMSX)
1913 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1914 			else
1915 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1916 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1917 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1918 				struct exception_table_entry *ex;
1919 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1920 				s64 delta;
1921 
1922 				/* populate jmp_offset for JMP above */
1923 				start_of_ldx[-1] = prog - start_of_ldx;
1924 
1925 				if (insn->off && src_reg != dst_reg) {
1926 					/* sub src_reg, insn->off
1927 					 * Restore src_reg after "add src_reg, insn->off" in prev
1928 					 * if statement. But if src_reg == dst_reg, emit_ldx
1929 					 * above already clobbered src_reg, so no need to restore.
1930 					 * If add src_reg, insn->off was unnecessary, no need to
1931 					 * restore either.
1932 					 */
1933 					maybe_emit_1mod(&prog, src_reg, true);
1934 					EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
1935 				}
1936 
1937 				if (!bpf_prog->aux->extable)
1938 					break;
1939 
1940 				if (excnt >= bpf_prog->aux->num_exentries) {
1941 					pr_err("ex gen bug\n");
1942 					return -EFAULT;
1943 				}
1944 				ex = &bpf_prog->aux->extable[excnt++];
1945 
1946 				delta = _insn - (u8 *)&ex->insn;
1947 				if (!is_simm32(delta)) {
1948 					pr_err("extable->insn doesn't fit into 32-bit\n");
1949 					return -EFAULT;
1950 				}
1951 				/* switch ex to rw buffer for writes */
1952 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1953 
1954 				ex->insn = delta;
1955 
1956 				ex->data = EX_TYPE_BPF;
1957 
1958 				if (dst_reg > BPF_REG_9) {
1959 					pr_err("verifier error\n");
1960 					return -EFAULT;
1961 				}
1962 				/*
1963 				 * Compute size of x86 insn and its target dest x86 register.
1964 				 * ex_handler_bpf() will use lower 8 bits to adjust
1965 				 * pt_regs->ip to jump over this x86 instruction
1966 				 * and upper bits to figure out which pt_regs to zero out.
1967 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1968 				 * of 4 bytes will be ignored and rbx will be zero inited.
1969 				 */
1970 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1971 			}
1972 			break;
1973 
1974 		case BPF_STX | BPF_ATOMIC | BPF_W:
1975 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1976 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1977 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1978 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1979 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1980 				u32 real_src_reg = src_reg;
1981 				u32 real_dst_reg = dst_reg;
1982 				u8 *branch_target;
1983 
1984 				/*
1985 				 * Can't be implemented with a single x86 insn.
1986 				 * Need to do a CMPXCHG loop.
1987 				 */
1988 
1989 				/* Will need RAX as a CMPXCHG operand so save R0 */
1990 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1991 				if (src_reg == BPF_REG_0)
1992 					real_src_reg = BPF_REG_AX;
1993 				if (dst_reg == BPF_REG_0)
1994 					real_dst_reg = BPF_REG_AX;
1995 
1996 				branch_target = prog;
1997 				/* Load old value */
1998 				emit_ldx(&prog, BPF_SIZE(insn->code),
1999 					 BPF_REG_0, real_dst_reg, insn->off);
2000 				/*
2001 				 * Perform the (commutative) operation locally,
2002 				 * put the result in the AUX_REG.
2003 				 */
2004 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
2005 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
2006 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
2007 				      add_2reg(0xC0, AUX_REG, real_src_reg));
2008 				/* Attempt to swap in new value */
2009 				err = emit_atomic(&prog, BPF_CMPXCHG,
2010 						  real_dst_reg, AUX_REG,
2011 						  insn->off,
2012 						  BPF_SIZE(insn->code));
2013 				if (WARN_ON(err))
2014 					return err;
2015 				/*
2016 				 * ZF tells us whether we won the race. If it's
2017 				 * cleared we need to try again.
2018 				 */
2019 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
2020 				/* Return the pre-modification value */
2021 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
2022 				/* Restore R0 after clobbering RAX */
2023 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
2024 				break;
2025 			}
2026 
2027 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
2028 					  insn->off, BPF_SIZE(insn->code));
2029 			if (err)
2030 				return err;
2031 			break;
2032 
2033 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_W:
2034 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_DW:
2035 			start_of_ldx = prog;
2036 			err = emit_atomic_index(&prog, insn->imm, BPF_SIZE(insn->code),
2037 						dst_reg, src_reg, X86_REG_R12, insn->off);
2038 			if (err)
2039 				return err;
2040 			goto populate_extable;
2041 
2042 			/* call */
2043 		case BPF_JMP | BPF_CALL: {
2044 			u8 *ip = image + addrs[i - 1];
2045 
2046 			func = (u8 *) __bpf_call_base + imm32;
2047 			if (tail_call_reachable) {
2048 				RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
2049 				ip += 7;
2050 			}
2051 			if (!imm32)
2052 				return -EINVAL;
2053 			ip += x86_call_depth_emit_accounting(&prog, func, ip);
2054 			if (emit_call(&prog, func, ip))
2055 				return -EINVAL;
2056 			break;
2057 		}
2058 
2059 		case BPF_JMP | BPF_TAIL_CALL:
2060 			if (imm32)
2061 				emit_bpf_tail_call_direct(bpf_prog,
2062 							  &bpf_prog->aux->poke_tab[imm32 - 1],
2063 							  &prog, image + addrs[i - 1],
2064 							  callee_regs_used,
2065 							  bpf_prog->aux->stack_depth,
2066 							  ctx);
2067 			else
2068 				emit_bpf_tail_call_indirect(bpf_prog,
2069 							    &prog,
2070 							    callee_regs_used,
2071 							    bpf_prog->aux->stack_depth,
2072 							    image + addrs[i - 1],
2073 							    ctx);
2074 			break;
2075 
2076 			/* cond jump */
2077 		case BPF_JMP | BPF_JEQ | BPF_X:
2078 		case BPF_JMP | BPF_JNE | BPF_X:
2079 		case BPF_JMP | BPF_JGT | BPF_X:
2080 		case BPF_JMP | BPF_JLT | BPF_X:
2081 		case BPF_JMP | BPF_JGE | BPF_X:
2082 		case BPF_JMP | BPF_JLE | BPF_X:
2083 		case BPF_JMP | BPF_JSGT | BPF_X:
2084 		case BPF_JMP | BPF_JSLT | BPF_X:
2085 		case BPF_JMP | BPF_JSGE | BPF_X:
2086 		case BPF_JMP | BPF_JSLE | BPF_X:
2087 		case BPF_JMP32 | BPF_JEQ | BPF_X:
2088 		case BPF_JMP32 | BPF_JNE | BPF_X:
2089 		case BPF_JMP32 | BPF_JGT | BPF_X:
2090 		case BPF_JMP32 | BPF_JLT | BPF_X:
2091 		case BPF_JMP32 | BPF_JGE | BPF_X:
2092 		case BPF_JMP32 | BPF_JLE | BPF_X:
2093 		case BPF_JMP32 | BPF_JSGT | BPF_X:
2094 		case BPF_JMP32 | BPF_JSLT | BPF_X:
2095 		case BPF_JMP32 | BPF_JSGE | BPF_X:
2096 		case BPF_JMP32 | BPF_JSLE | BPF_X:
2097 			/* cmp dst_reg, src_reg */
2098 			maybe_emit_mod(&prog, dst_reg, src_reg,
2099 				       BPF_CLASS(insn->code) == BPF_JMP);
2100 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
2101 			goto emit_cond_jmp;
2102 
2103 		case BPF_JMP | BPF_JSET | BPF_X:
2104 		case BPF_JMP32 | BPF_JSET | BPF_X:
2105 			/* test dst_reg, src_reg */
2106 			maybe_emit_mod(&prog, dst_reg, src_reg,
2107 				       BPF_CLASS(insn->code) == BPF_JMP);
2108 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
2109 			goto emit_cond_jmp;
2110 
2111 		case BPF_JMP | BPF_JSET | BPF_K:
2112 		case BPF_JMP32 | BPF_JSET | BPF_K:
2113 			/* test dst_reg, imm32 */
2114 			maybe_emit_1mod(&prog, dst_reg,
2115 					BPF_CLASS(insn->code) == BPF_JMP);
2116 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
2117 			goto emit_cond_jmp;
2118 
2119 		case BPF_JMP | BPF_JEQ | BPF_K:
2120 		case BPF_JMP | BPF_JNE | BPF_K:
2121 		case BPF_JMP | BPF_JGT | BPF_K:
2122 		case BPF_JMP | BPF_JLT | BPF_K:
2123 		case BPF_JMP | BPF_JGE | BPF_K:
2124 		case BPF_JMP | BPF_JLE | BPF_K:
2125 		case BPF_JMP | BPF_JSGT | BPF_K:
2126 		case BPF_JMP | BPF_JSLT | BPF_K:
2127 		case BPF_JMP | BPF_JSGE | BPF_K:
2128 		case BPF_JMP | BPF_JSLE | BPF_K:
2129 		case BPF_JMP32 | BPF_JEQ | BPF_K:
2130 		case BPF_JMP32 | BPF_JNE | BPF_K:
2131 		case BPF_JMP32 | BPF_JGT | BPF_K:
2132 		case BPF_JMP32 | BPF_JLT | BPF_K:
2133 		case BPF_JMP32 | BPF_JGE | BPF_K:
2134 		case BPF_JMP32 | BPF_JLE | BPF_K:
2135 		case BPF_JMP32 | BPF_JSGT | BPF_K:
2136 		case BPF_JMP32 | BPF_JSLT | BPF_K:
2137 		case BPF_JMP32 | BPF_JSGE | BPF_K:
2138 		case BPF_JMP32 | BPF_JSLE | BPF_K:
2139 			/* test dst_reg, dst_reg to save one extra byte */
2140 			if (imm32 == 0) {
2141 				maybe_emit_mod(&prog, dst_reg, dst_reg,
2142 					       BPF_CLASS(insn->code) == BPF_JMP);
2143 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
2144 				goto emit_cond_jmp;
2145 			}
2146 
2147 			/* cmp dst_reg, imm8/32 */
2148 			maybe_emit_1mod(&prog, dst_reg,
2149 					BPF_CLASS(insn->code) == BPF_JMP);
2150 
2151 			if (is_imm8(imm32))
2152 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
2153 			else
2154 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
2155 
2156 emit_cond_jmp:		/* Convert BPF opcode to x86 */
2157 			switch (BPF_OP(insn->code)) {
2158 			case BPF_JEQ:
2159 				jmp_cond = X86_JE;
2160 				break;
2161 			case BPF_JSET:
2162 			case BPF_JNE:
2163 				jmp_cond = X86_JNE;
2164 				break;
2165 			case BPF_JGT:
2166 				/* GT is unsigned '>', JA in x86 */
2167 				jmp_cond = X86_JA;
2168 				break;
2169 			case BPF_JLT:
2170 				/* LT is unsigned '<', JB in x86 */
2171 				jmp_cond = X86_JB;
2172 				break;
2173 			case BPF_JGE:
2174 				/* GE is unsigned '>=', JAE in x86 */
2175 				jmp_cond = X86_JAE;
2176 				break;
2177 			case BPF_JLE:
2178 				/* LE is unsigned '<=', JBE in x86 */
2179 				jmp_cond = X86_JBE;
2180 				break;
2181 			case BPF_JSGT:
2182 				/* Signed '>', GT in x86 */
2183 				jmp_cond = X86_JG;
2184 				break;
2185 			case BPF_JSLT:
2186 				/* Signed '<', LT in x86 */
2187 				jmp_cond = X86_JL;
2188 				break;
2189 			case BPF_JSGE:
2190 				/* Signed '>=', GE in x86 */
2191 				jmp_cond = X86_JGE;
2192 				break;
2193 			case BPF_JSLE:
2194 				/* Signed '<=', LE in x86 */
2195 				jmp_cond = X86_JLE;
2196 				break;
2197 			default: /* to silence GCC warning */
2198 				return -EFAULT;
2199 			}
2200 			jmp_offset = addrs[i + insn->off] - addrs[i];
2201 			if (is_imm8(jmp_offset)) {
2202 				if (jmp_padding) {
2203 					/* To keep the jmp_offset valid, the extra bytes are
2204 					 * padded before the jump insn, so we subtract the
2205 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
2206 					 *
2207 					 * If the previous pass already emits an imm8
2208 					 * jmp_cond, then this BPF insn won't shrink, so
2209 					 * "nops" is 0.
2210 					 *
2211 					 * On the other hand, if the previous pass emits an
2212 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
2213 					 * keep the image from shrinking further.
2214 					 *
2215 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
2216 					 *     is 2 bytes, so the size difference is 4 bytes.
2217 					 */
2218 					nops = INSN_SZ_DIFF - 2;
2219 					if (nops != 0 && nops != 4) {
2220 						pr_err("unexpected jmp_cond padding: %d bytes\n",
2221 						       nops);
2222 						return -EFAULT;
2223 					}
2224 					emit_nops(&prog, nops);
2225 				}
2226 				EMIT2(jmp_cond, jmp_offset);
2227 			} else if (is_simm32(jmp_offset)) {
2228 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
2229 			} else {
2230 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
2231 				return -EFAULT;
2232 			}
2233 
2234 			break;
2235 
2236 		case BPF_JMP | BPF_JA:
2237 		case BPF_JMP32 | BPF_JA:
2238 			if (BPF_CLASS(insn->code) == BPF_JMP) {
2239 				if (insn->off == -1)
2240 					/* -1 jmp instructions will always jump
2241 					 * backwards two bytes. Explicitly handling
2242 					 * this case avoids wasting too many passes
2243 					 * when there are long sequences of replaced
2244 					 * dead code.
2245 					 */
2246 					jmp_offset = -2;
2247 				else
2248 					jmp_offset = addrs[i + insn->off] - addrs[i];
2249 			} else {
2250 				if (insn->imm == -1)
2251 					jmp_offset = -2;
2252 				else
2253 					jmp_offset = addrs[i + insn->imm] - addrs[i];
2254 			}
2255 
2256 			if (!jmp_offset) {
2257 				/*
2258 				 * If jmp_padding is enabled, the extra nops will
2259 				 * be inserted. Otherwise, optimize out nop jumps.
2260 				 */
2261 				if (jmp_padding) {
2262 					/* There are 3 possible conditions.
2263 					 * (1) This BPF_JA is already optimized out in
2264 					 *     the previous run, so there is no need
2265 					 *     to pad any extra byte (0 byte).
2266 					 * (2) The previous pass emits an imm8 jmp,
2267 					 *     so we pad 2 bytes to match the previous
2268 					 *     insn size.
2269 					 * (3) Similarly, the previous pass emits an
2270 					 *     imm32 jmp, and 5 bytes is padded.
2271 					 */
2272 					nops = INSN_SZ_DIFF;
2273 					if (nops != 0 && nops != 2 && nops != 5) {
2274 						pr_err("unexpected nop jump padding: %d bytes\n",
2275 						       nops);
2276 						return -EFAULT;
2277 					}
2278 					emit_nops(&prog, nops);
2279 				}
2280 				break;
2281 			}
2282 emit_jmp:
2283 			if (is_imm8(jmp_offset)) {
2284 				if (jmp_padding) {
2285 					/* To avoid breaking jmp_offset, the extra bytes
2286 					 * are padded before the actual jmp insn, so
2287 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
2288 					 *
2289 					 * If the previous pass already emits an imm8
2290 					 * jmp, there is nothing to pad (0 byte).
2291 					 *
2292 					 * If it emits an imm32 jmp (5 bytes) previously
2293 					 * and now an imm8 jmp (2 bytes), then we pad
2294 					 * (5 - 2 = 3) bytes to stop the image from
2295 					 * shrinking further.
2296 					 */
2297 					nops = INSN_SZ_DIFF - 2;
2298 					if (nops != 0 && nops != 3) {
2299 						pr_err("unexpected jump padding: %d bytes\n",
2300 						       nops);
2301 						return -EFAULT;
2302 					}
2303 					emit_nops(&prog, INSN_SZ_DIFF - 2);
2304 				}
2305 				EMIT2(0xEB, jmp_offset);
2306 			} else if (is_simm32(jmp_offset)) {
2307 				EMIT1_off32(0xE9, jmp_offset);
2308 			} else {
2309 				pr_err("jmp gen bug %llx\n", jmp_offset);
2310 				return -EFAULT;
2311 			}
2312 			break;
2313 
2314 		case BPF_JMP | BPF_EXIT:
2315 			if (seen_exit) {
2316 				jmp_offset = ctx->cleanup_addr - addrs[i];
2317 				goto emit_jmp;
2318 			}
2319 			seen_exit = true;
2320 			/* Update cleanup_addr */
2321 			ctx->cleanup_addr = proglen;
2322 			if (bpf_prog->aux->exception_boundary) {
2323 				pop_callee_regs(&prog, all_callee_regs_used);
2324 				pop_r12(&prog);
2325 			} else {
2326 				pop_callee_regs(&prog, callee_regs_used);
2327 				if (arena_vm_start)
2328 					pop_r12(&prog);
2329 			}
2330 			EMIT1(0xC9);         /* leave */
2331 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
2332 			break;
2333 
2334 		default:
2335 			/*
2336 			 * By design x86-64 JIT should support all BPF instructions.
2337 			 * This error will be seen if new instruction was added
2338 			 * to the interpreter, but not to the JIT, or if there is
2339 			 * junk in bpf_prog.
2340 			 */
2341 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
2342 			return -EINVAL;
2343 		}
2344 
2345 		ilen = prog - temp;
2346 		if (ilen > BPF_MAX_INSN_SIZE) {
2347 			pr_err("bpf_jit: fatal insn size error\n");
2348 			return -EFAULT;
2349 		}
2350 
2351 		if (image) {
2352 			/*
2353 			 * When populating the image, assert that:
2354 			 *
2355 			 *  i) We do not write beyond the allocated space, and
2356 			 * ii) addrs[i] did not change from the prior run, in order
2357 			 *     to validate assumptions made for computing branch
2358 			 *     displacements.
2359 			 */
2360 			if (unlikely(proglen + ilen > oldproglen ||
2361 				     proglen + ilen != addrs[i])) {
2362 				pr_err("bpf_jit: fatal error\n");
2363 				return -EFAULT;
2364 			}
2365 			memcpy(rw_image + proglen, temp, ilen);
2366 		}
2367 		proglen += ilen;
2368 		addrs[i] = proglen;
2369 		prog = temp;
2370 	}
2371 
2372 	if (image && excnt != bpf_prog->aux->num_exentries) {
2373 		pr_err("extable is not populated\n");
2374 		return -EFAULT;
2375 	}
2376 	return proglen;
2377 }
2378 
2379 static void clean_stack_garbage(const struct btf_func_model *m,
2380 				u8 **pprog, int nr_stack_slots,
2381 				int stack_size)
2382 {
2383 	int arg_size, off;
2384 	u8 *prog;
2385 
2386 	/* Generally speaking, the compiler will pass the arguments
2387 	 * on-stack with "push" instruction, which will take 8-byte
2388 	 * on the stack. In this case, there won't be garbage values
2389 	 * while we copy the arguments from origin stack frame to current
2390 	 * in BPF_DW.
2391 	 *
2392 	 * However, sometimes the compiler will only allocate 4-byte on
2393 	 * the stack for the arguments. For now, this case will only
2394 	 * happen if there is only one argument on-stack and its size
2395 	 * not more than 4 byte. In this case, there will be garbage
2396 	 * values on the upper 4-byte where we store the argument on
2397 	 * current stack frame.
2398 	 *
2399 	 * arguments on origin stack:
2400 	 *
2401 	 * stack_arg_1(4-byte) xxx(4-byte)
2402 	 *
2403 	 * what we copy:
2404 	 *
2405 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
2406 	 *
2407 	 * and the xxx is the garbage values which we should clean here.
2408 	 */
2409 	if (nr_stack_slots != 1)
2410 		return;
2411 
2412 	/* the size of the last argument */
2413 	arg_size = m->arg_size[m->nr_args - 1];
2414 	if (arg_size <= 4) {
2415 		off = -(stack_size - 4);
2416 		prog = *pprog;
2417 		/* mov DWORD PTR [rbp + off], 0 */
2418 		if (!is_imm8(off))
2419 			EMIT2_off32(0xC7, 0x85, off);
2420 		else
2421 			EMIT3(0xC7, 0x45, off);
2422 		EMIT(0, 4);
2423 		*pprog = prog;
2424 	}
2425 }
2426 
2427 /* get the count of the regs that are used to pass arguments */
2428 static int get_nr_used_regs(const struct btf_func_model *m)
2429 {
2430 	int i, arg_regs, nr_used_regs = 0;
2431 
2432 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2433 		arg_regs = (m->arg_size[i] + 7) / 8;
2434 		if (nr_used_regs + arg_regs <= 6)
2435 			nr_used_regs += arg_regs;
2436 
2437 		if (nr_used_regs >= 6)
2438 			break;
2439 	}
2440 
2441 	return nr_used_regs;
2442 }
2443 
2444 static void save_args(const struct btf_func_model *m, u8 **prog,
2445 		      int stack_size, bool for_call_origin)
2446 {
2447 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2448 	int i, j;
2449 
2450 	/* Store function arguments to stack.
2451 	 * For a function that accepts two pointers the sequence will be:
2452 	 * mov QWORD PTR [rbp-0x10],rdi
2453 	 * mov QWORD PTR [rbp-0x8],rsi
2454 	 */
2455 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2456 		arg_regs = (m->arg_size[i] + 7) / 8;
2457 
2458 		/* According to the research of Yonghong, struct members
2459 		 * should be all in register or all on the stack.
2460 		 * Meanwhile, the compiler will pass the argument on regs
2461 		 * if the remaining regs can hold the argument.
2462 		 *
2463 		 * Disorder of the args can happen. For example:
2464 		 *
2465 		 * struct foo_struct {
2466 		 *     long a;
2467 		 *     int b;
2468 		 * };
2469 		 * int foo(char, char, char, char, char, struct foo_struct,
2470 		 *         char);
2471 		 *
2472 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2473 		 * by stack.
2474 		 */
2475 		if (nr_regs + arg_regs > 6) {
2476 			/* copy function arguments from origin stack frame
2477 			 * into current stack frame.
2478 			 *
2479 			 * The starting address of the arguments on-stack
2480 			 * is:
2481 			 *   rbp + 8(push rbp) +
2482 			 *   8(return addr of origin call) +
2483 			 *   8(return addr of the caller)
2484 			 * which means: rbp + 24
2485 			 */
2486 			for (j = 0; j < arg_regs; j++) {
2487 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2488 					 nr_stack_slots * 8 + 0x18);
2489 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2490 					 -stack_size);
2491 
2492 				if (!nr_stack_slots)
2493 					first_off = stack_size;
2494 				stack_size -= 8;
2495 				nr_stack_slots++;
2496 			}
2497 		} else {
2498 			/* Only copy the arguments on-stack to current
2499 			 * 'stack_size' and ignore the regs, used to
2500 			 * prepare the arguments on-stack for origin call.
2501 			 */
2502 			if (for_call_origin) {
2503 				nr_regs += arg_regs;
2504 				continue;
2505 			}
2506 
2507 			/* copy the arguments from regs into stack */
2508 			for (j = 0; j < arg_regs; j++) {
2509 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2510 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2511 					 -stack_size);
2512 				stack_size -= 8;
2513 				nr_regs++;
2514 			}
2515 		}
2516 	}
2517 
2518 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2519 }
2520 
2521 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2522 			 int stack_size)
2523 {
2524 	int i, j, arg_regs, nr_regs = 0;
2525 
2526 	/* Restore function arguments from stack.
2527 	 * For a function that accepts two pointers the sequence will be:
2528 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2529 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2530 	 *
2531 	 * The logic here is similar to what we do in save_args()
2532 	 */
2533 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2534 		arg_regs = (m->arg_size[i] + 7) / 8;
2535 		if (nr_regs + arg_regs <= 6) {
2536 			for (j = 0; j < arg_regs; j++) {
2537 				emit_ldx(prog, BPF_DW,
2538 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2539 					 BPF_REG_FP,
2540 					 -stack_size);
2541 				stack_size -= 8;
2542 				nr_regs++;
2543 			}
2544 		} else {
2545 			stack_size -= 8 * arg_regs;
2546 		}
2547 
2548 		if (nr_regs >= 6)
2549 			break;
2550 	}
2551 }
2552 
2553 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2554 			   struct bpf_tramp_link *l, int stack_size,
2555 			   int run_ctx_off, bool save_ret,
2556 			   void *image, void *rw_image)
2557 {
2558 	u8 *prog = *pprog;
2559 	u8 *jmp_insn;
2560 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2561 	struct bpf_prog *p = l->link.prog;
2562 	u64 cookie = l->cookie;
2563 
2564 	/* mov rdi, cookie */
2565 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2566 
2567 	/* Prepare struct bpf_tramp_run_ctx.
2568 	 *
2569 	 * bpf_tramp_run_ctx is already preserved by
2570 	 * arch_prepare_bpf_trampoline().
2571 	 *
2572 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2573 	 */
2574 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2575 
2576 	/* arg1: mov rdi, progs[i] */
2577 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2578 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2579 	if (!is_imm8(-run_ctx_off))
2580 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
2581 	else
2582 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
2583 
2584 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image)))
2585 		return -EINVAL;
2586 	/* remember prog start time returned by __bpf_prog_enter */
2587 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
2588 
2589 	/* if (__bpf_prog_enter*(prog) == 0)
2590 	 *	goto skip_exec_of_prog;
2591 	 */
2592 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
2593 	/* emit 2 nops that will be replaced with JE insn */
2594 	jmp_insn = prog;
2595 	emit_nops(&prog, 2);
2596 
2597 	/* arg1: lea rdi, [rbp - stack_size] */
2598 	if (!is_imm8(-stack_size))
2599 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
2600 	else
2601 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
2602 	/* arg2: progs[i]->insnsi for interpreter */
2603 	if (!p->jited)
2604 		emit_mov_imm64(&prog, BPF_REG_2,
2605 			       (long) p->insnsi >> 32,
2606 			       (u32) (long) p->insnsi);
2607 	/* call JITed bpf program or interpreter */
2608 	if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image)))
2609 		return -EINVAL;
2610 
2611 	/*
2612 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
2613 	 * of the previous call which is then passed on the stack to
2614 	 * the next BPF program.
2615 	 *
2616 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
2617 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
2618 	 */
2619 	if (save_ret)
2620 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2621 
2622 	/* replace 2 nops with JE insn, since jmp target is known */
2623 	jmp_insn[0] = X86_JE;
2624 	jmp_insn[1] = prog - jmp_insn - 2;
2625 
2626 	/* arg1: mov rdi, progs[i] */
2627 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2628 	/* arg2: mov rsi, rbx <- start time in nsec */
2629 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
2630 	/* arg3: lea rdx, [rbp - run_ctx_off] */
2631 	if (!is_imm8(-run_ctx_off))
2632 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
2633 	else
2634 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
2635 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
2636 		return -EINVAL;
2637 
2638 	*pprog = prog;
2639 	return 0;
2640 }
2641 
2642 static void emit_align(u8 **pprog, u32 align)
2643 {
2644 	u8 *target, *prog = *pprog;
2645 
2646 	target = PTR_ALIGN(prog, align);
2647 	if (target != prog)
2648 		emit_nops(&prog, target - prog);
2649 
2650 	*pprog = prog;
2651 }
2652 
2653 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2654 {
2655 	u8 *prog = *pprog;
2656 	s64 offset;
2657 
2658 	offset = func - (ip + 2 + 4);
2659 	if (!is_simm32(offset)) {
2660 		pr_err("Target %p is out of range\n", func);
2661 		return -EINVAL;
2662 	}
2663 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2664 	*pprog = prog;
2665 	return 0;
2666 }
2667 
2668 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2669 		      struct bpf_tramp_links *tl, int stack_size,
2670 		      int run_ctx_off, bool save_ret,
2671 		      void *image, void *rw_image)
2672 {
2673 	int i;
2674 	u8 *prog = *pprog;
2675 
2676 	for (i = 0; i < tl->nr_links; i++) {
2677 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2678 				    run_ctx_off, save_ret, image, rw_image))
2679 			return -EINVAL;
2680 	}
2681 	*pprog = prog;
2682 	return 0;
2683 }
2684 
2685 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2686 			      struct bpf_tramp_links *tl, int stack_size,
2687 			      int run_ctx_off, u8 **branches,
2688 			      void *image, void *rw_image)
2689 {
2690 	u8 *prog = *pprog;
2691 	int i;
2692 
2693 	/* The first fmod_ret program will receive a garbage return value.
2694 	 * Set this to 0 to avoid confusing the program.
2695 	 */
2696 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2697 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2698 	for (i = 0; i < tl->nr_links; i++) {
2699 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true,
2700 				    image, rw_image))
2701 			return -EINVAL;
2702 
2703 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2704 		 * if (*(u64 *)(rbp - 8) !=  0)
2705 		 *	goto do_fexit;
2706 		 */
2707 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2708 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2709 
2710 		/* Save the location of the branch and Generate 6 nops
2711 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2712 		 * are replaced with a conditional jump once do_fexit (i.e. the
2713 		 * start of the fexit invocation) is finalized.
2714 		 */
2715 		branches[i] = prog;
2716 		emit_nops(&prog, 4 + 2);
2717 	}
2718 
2719 	*pprog = prog;
2720 	return 0;
2721 }
2722 
2723 /* Example:
2724  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2725  * its 'struct btf_func_model' will be nr_args=2
2726  * The assembly code when eth_type_trans is executing after trampoline:
2727  *
2728  * push rbp
2729  * mov rbp, rsp
2730  * sub rsp, 16                     // space for skb and dev
2731  * push rbx                        // temp regs to pass start time
2732  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2733  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2734  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2735  * mov rbx, rax                    // remember start time in bpf stats are enabled
2736  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2737  * call addr_of_jited_FENTRY_prog
2738  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2739  * mov rsi, rbx                    // prog start time
2740  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2741  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2742  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2743  * pop rbx
2744  * leave
2745  * ret
2746  *
2747  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2748  * replaced with 'call generated_bpf_trampoline'. When it returns
2749  * eth_type_trans will continue executing with original skb and dev pointers.
2750  *
2751  * The assembly code when eth_type_trans is called from trampoline:
2752  *
2753  * push rbp
2754  * mov rbp, rsp
2755  * sub rsp, 24                     // space for skb, dev, return value
2756  * push rbx                        // temp regs to pass start time
2757  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2758  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2759  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2760  * mov rbx, rax                    // remember start time if bpf stats are enabled
2761  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2762  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2763  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2764  * mov rsi, rbx                    // prog start time
2765  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2766  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2767  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2768  * call eth_type_trans+5           // execute body of eth_type_trans
2769  * mov qword ptr [rbp - 8], rax    // save return value
2770  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2771  * mov rbx, rax                    // remember start time in bpf stats are enabled
2772  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2773  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2774  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2775  * mov rsi, rbx                    // prog start time
2776  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2777  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2778  * pop rbx
2779  * leave
2780  * add rsp, 8                      // skip eth_type_trans's frame
2781  * ret                             // return to its caller
2782  */
2783 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
2784 					 void *rw_image_end, void *image,
2785 					 const struct btf_func_model *m, u32 flags,
2786 					 struct bpf_tramp_links *tlinks,
2787 					 void *func_addr)
2788 {
2789 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2790 	int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
2791 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2792 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2793 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2794 	void *orig_call = func_addr;
2795 	u8 **branches = NULL;
2796 	u8 *prog;
2797 	bool save_ret;
2798 
2799 	/*
2800 	 * F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is
2801 	 * explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG
2802 	 * because @func_addr.
2803 	 */
2804 	WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) &&
2805 		     (flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET)));
2806 
2807 	/* extra registers for struct arguments */
2808 	for (i = 0; i < m->nr_args; i++) {
2809 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2810 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2811 	}
2812 
2813 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
2814 	 * are passed through regs, the remains are through stack.
2815 	 */
2816 	if (nr_regs > MAX_BPF_FUNC_ARGS)
2817 		return -ENOTSUPP;
2818 
2819 	/* Generated trampoline stack layout:
2820 	 *
2821 	 * RBP + 8         [ return address  ]
2822 	 * RBP + 0         [ RBP             ]
2823 	 *
2824 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2825 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2826 	 *
2827 	 *                 [ reg_argN        ]  always
2828 	 *                 [ ...             ]
2829 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2830 	 *
2831 	 * RBP - nregs_off [ regs count	     ]  always
2832 	 *
2833 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2834 	 *
2835 	 * RBP - rbx_off   [ rbx value       ]  always
2836 	 *
2837 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2838 	 *
2839 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
2840 	 *                     [ ...        ]
2841 	 *                     [ stack_arg2 ]
2842 	 * RBP - arg_stack_off [ stack_arg1 ]
2843 	 * RSP                 [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
2844 	 */
2845 
2846 	/* room for return value of orig_call or fentry prog */
2847 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2848 	if (save_ret)
2849 		stack_size += 8;
2850 
2851 	stack_size += nr_regs * 8;
2852 	regs_off = stack_size;
2853 
2854 	/* regs count  */
2855 	stack_size += 8;
2856 	nregs_off = stack_size;
2857 
2858 	if (flags & BPF_TRAMP_F_IP_ARG)
2859 		stack_size += 8; /* room for IP address argument */
2860 
2861 	ip_off = stack_size;
2862 
2863 	stack_size += 8;
2864 	rbx_off = stack_size;
2865 
2866 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2867 	run_ctx_off = stack_size;
2868 
2869 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
2870 		/* the space that used to pass arguments on-stack */
2871 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
2872 		/* make sure the stack pointer is 16-byte aligned if we
2873 		 * need pass arguments on stack, which means
2874 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
2875 		 * should be 16-byte aligned. Following code depend on
2876 		 * that stack_size is already 8-byte aligned.
2877 		 */
2878 		stack_size += (stack_size % 16) ? 0 : 8;
2879 	}
2880 
2881 	arg_stack_off = stack_size;
2882 
2883 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2884 		/* skip patched call instruction and point orig_call to actual
2885 		 * body of the kernel function.
2886 		 */
2887 		if (is_endbr(*(u32 *)orig_call))
2888 			orig_call += ENDBR_INSN_SIZE;
2889 		orig_call += X86_PATCH_SIZE;
2890 	}
2891 
2892 	prog = rw_image;
2893 
2894 	if (flags & BPF_TRAMP_F_INDIRECT) {
2895 		/*
2896 		 * Indirect call for bpf_struct_ops
2897 		 */
2898 		emit_cfi(&prog, cfi_get_func_hash(func_addr));
2899 	} else {
2900 		/*
2901 		 * Direct-call fentry stub, as such it needs accounting for the
2902 		 * __fentry__ call.
2903 		 */
2904 		x86_call_depth_emit_accounting(&prog, NULL, image);
2905 	}
2906 	EMIT1(0x55);		 /* push rbp */
2907 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2908 	if (!is_imm8(stack_size)) {
2909 		/* sub rsp, stack_size */
2910 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
2911 	} else {
2912 		/* sub rsp, stack_size */
2913 		EMIT4(0x48, 0x83, 0xEC, stack_size);
2914 	}
2915 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2916 		EMIT1(0x50);		/* push rax */
2917 	/* mov QWORD PTR [rbp - rbx_off], rbx */
2918 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
2919 
2920 	/* Store number of argument registers of the traced function:
2921 	 *   mov rax, nr_regs
2922 	 *   mov QWORD PTR [rbp - nregs_off], rax
2923 	 */
2924 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2925 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2926 
2927 	if (flags & BPF_TRAMP_F_IP_ARG) {
2928 		/* Store IP address of the traced function:
2929 		 * movabsq rax, func_addr
2930 		 * mov QWORD PTR [rbp - ip_off], rax
2931 		 */
2932 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2933 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2934 	}
2935 
2936 	save_args(m, &prog, regs_off, false);
2937 
2938 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2939 		/* arg1: mov rdi, im */
2940 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2941 		if (emit_rsb_call(&prog, __bpf_tramp_enter,
2942 				  image + (prog - (u8 *)rw_image))) {
2943 			ret = -EINVAL;
2944 			goto cleanup;
2945 		}
2946 	}
2947 
2948 	if (fentry->nr_links) {
2949 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2950 			       flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image))
2951 			return -EINVAL;
2952 	}
2953 
2954 	if (fmod_ret->nr_links) {
2955 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2956 				   GFP_KERNEL);
2957 		if (!branches)
2958 			return -ENOMEM;
2959 
2960 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2961 				       run_ctx_off, branches, image, rw_image)) {
2962 			ret = -EINVAL;
2963 			goto cleanup;
2964 		}
2965 	}
2966 
2967 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2968 		restore_regs(m, &prog, regs_off);
2969 		save_args(m, &prog, arg_stack_off, true);
2970 
2971 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
2972 			/* Before calling the original function, restore the
2973 			 * tail_call_cnt from stack to rax.
2974 			 */
2975 			RESTORE_TAIL_CALL_CNT(stack_size);
2976 		}
2977 
2978 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2979 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
2980 			EMIT2(0xff, 0xd3); /* call *rbx */
2981 		} else {
2982 			/* call original function */
2983 			if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) {
2984 				ret = -EINVAL;
2985 				goto cleanup;
2986 			}
2987 		}
2988 		/* remember return value in a stack for bpf prog to access */
2989 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2990 		im->ip_after_call = image + (prog - (u8 *)rw_image);
2991 		emit_nops(&prog, X86_PATCH_SIZE);
2992 	}
2993 
2994 	if (fmod_ret->nr_links) {
2995 		/* From Intel 64 and IA-32 Architectures Optimization
2996 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2997 		 * Coding Rule 11: All branch targets should be 16-byte
2998 		 * aligned.
2999 		 */
3000 		emit_align(&prog, 16);
3001 		/* Update the branches saved in invoke_bpf_mod_ret with the
3002 		 * aligned address of do_fexit.
3003 		 */
3004 		for (i = 0; i < fmod_ret->nr_links; i++) {
3005 			emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image),
3006 					    image + (branches[i] - (u8 *)rw_image), X86_JNE);
3007 		}
3008 	}
3009 
3010 	if (fexit->nr_links) {
3011 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off,
3012 			       false, image, rw_image)) {
3013 			ret = -EINVAL;
3014 			goto cleanup;
3015 		}
3016 	}
3017 
3018 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
3019 		restore_regs(m, &prog, regs_off);
3020 
3021 	/* This needs to be done regardless. If there were fmod_ret programs,
3022 	 * the return value is only updated on the stack and still needs to be
3023 	 * restored to R0.
3024 	 */
3025 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3026 		im->ip_epilogue = image + (prog - (u8 *)rw_image);
3027 		/* arg1: mov rdi, im */
3028 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
3029 		if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) {
3030 			ret = -EINVAL;
3031 			goto cleanup;
3032 		}
3033 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
3034 		/* Before running the original function, restore the
3035 		 * tail_call_cnt from stack to rax.
3036 		 */
3037 		RESTORE_TAIL_CALL_CNT(stack_size);
3038 	}
3039 
3040 	/* restore return value of orig_call or fentry prog back into RAX */
3041 	if (save_ret)
3042 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
3043 
3044 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
3045 	EMIT1(0xC9); /* leave */
3046 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
3047 		/* skip our return address and return to parent */
3048 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
3049 	}
3050 	emit_return(&prog, image + (prog - (u8 *)rw_image));
3051 	/* Make sure the trampoline generation logic doesn't overflow */
3052 	if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) {
3053 		ret = -EFAULT;
3054 		goto cleanup;
3055 	}
3056 	ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY;
3057 
3058 cleanup:
3059 	kfree(branches);
3060 	return ret;
3061 }
3062 
3063 void *arch_alloc_bpf_trampoline(unsigned int size)
3064 {
3065 	return bpf_prog_pack_alloc(size, jit_fill_hole);
3066 }
3067 
3068 void arch_free_bpf_trampoline(void *image, unsigned int size)
3069 {
3070 	bpf_prog_pack_free(image, size);
3071 }
3072 
3073 int arch_protect_bpf_trampoline(void *image, unsigned int size)
3074 {
3075 	return 0;
3076 }
3077 
3078 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
3079 				const struct btf_func_model *m, u32 flags,
3080 				struct bpf_tramp_links *tlinks,
3081 				void *func_addr)
3082 {
3083 	void *rw_image, *tmp;
3084 	int ret;
3085 	u32 size = image_end - image;
3086 
3087 	/* rw_image doesn't need to be in module memory range, so we can
3088 	 * use kvmalloc.
3089 	 */
3090 	rw_image = kvmalloc(size, GFP_KERNEL);
3091 	if (!rw_image)
3092 		return -ENOMEM;
3093 
3094 	ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
3095 					    flags, tlinks, func_addr);
3096 	if (ret < 0)
3097 		goto out;
3098 
3099 	tmp = bpf_arch_text_copy(image, rw_image, size);
3100 	if (IS_ERR(tmp))
3101 		ret = PTR_ERR(tmp);
3102 out:
3103 	kvfree(rw_image);
3104 	return ret;
3105 }
3106 
3107 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
3108 			     struct bpf_tramp_links *tlinks, void *func_addr)
3109 {
3110 	struct bpf_tramp_image im;
3111 	void *image;
3112 	int ret;
3113 
3114 	/* Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
3115 	 * This will NOT cause fragmentation in direct map, as we do not
3116 	 * call set_memory_*() on this buffer.
3117 	 *
3118 	 * We cannot use kvmalloc here, because we need image to be in
3119 	 * module memory range.
3120 	 */
3121 	image = bpf_jit_alloc_exec(PAGE_SIZE);
3122 	if (!image)
3123 		return -ENOMEM;
3124 
3125 	ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
3126 					    m, flags, tlinks, func_addr);
3127 	bpf_jit_free_exec(image);
3128 	return ret;
3129 }
3130 
3131 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
3132 {
3133 	u8 *jg_reloc, *prog = *pprog;
3134 	int pivot, err, jg_bytes = 1;
3135 	s64 jg_offset;
3136 
3137 	if (a == b) {
3138 		/* Leaf node of recursion, i.e. not a range of indices
3139 		 * anymore.
3140 		 */
3141 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
3142 		if (!is_simm32(progs[a]))
3143 			return -1;
3144 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
3145 			    progs[a]);
3146 		err = emit_cond_near_jump(&prog,	/* je func */
3147 					  (void *)progs[a], image + (prog - buf),
3148 					  X86_JE);
3149 		if (err)
3150 			return err;
3151 
3152 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
3153 
3154 		*pprog = prog;
3155 		return 0;
3156 	}
3157 
3158 	/* Not a leaf node, so we pivot, and recursively descend into
3159 	 * the lower and upper ranges.
3160 	 */
3161 	pivot = (b - a) / 2;
3162 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
3163 	if (!is_simm32(progs[a + pivot]))
3164 		return -1;
3165 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
3166 
3167 	if (pivot > 2) {				/* jg upper_part */
3168 		/* Require near jump. */
3169 		jg_bytes = 4;
3170 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
3171 	} else {
3172 		EMIT2(X86_JG, 0);
3173 	}
3174 	jg_reloc = prog;
3175 
3176 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
3177 				  progs, image, buf);
3178 	if (err)
3179 		return err;
3180 
3181 	/* From Intel 64 and IA-32 Architectures Optimization
3182 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
3183 	 * Coding Rule 11: All branch targets should be 16-byte
3184 	 * aligned.
3185 	 */
3186 	emit_align(&prog, 16);
3187 	jg_offset = prog - jg_reloc;
3188 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
3189 
3190 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
3191 				  b, progs, image, buf);
3192 	if (err)
3193 		return err;
3194 
3195 	*pprog = prog;
3196 	return 0;
3197 }
3198 
3199 static int cmp_ips(const void *a, const void *b)
3200 {
3201 	const s64 *ipa = a;
3202 	const s64 *ipb = b;
3203 
3204 	if (*ipa > *ipb)
3205 		return 1;
3206 	if (*ipa < *ipb)
3207 		return -1;
3208 	return 0;
3209 }
3210 
3211 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
3212 {
3213 	u8 *prog = buf;
3214 
3215 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
3216 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
3217 }
3218 
3219 struct x64_jit_data {
3220 	struct bpf_binary_header *rw_header;
3221 	struct bpf_binary_header *header;
3222 	int *addrs;
3223 	u8 *image;
3224 	int proglen;
3225 	struct jit_context ctx;
3226 };
3227 
3228 #define MAX_PASSES 20
3229 #define PADDING_PASSES (MAX_PASSES - 5)
3230 
3231 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
3232 {
3233 	struct bpf_binary_header *rw_header = NULL;
3234 	struct bpf_binary_header *header = NULL;
3235 	struct bpf_prog *tmp, *orig_prog = prog;
3236 	struct x64_jit_data *jit_data;
3237 	int proglen, oldproglen = 0;
3238 	struct jit_context ctx = {};
3239 	bool tmp_blinded = false;
3240 	bool extra_pass = false;
3241 	bool padding = false;
3242 	u8 *rw_image = NULL;
3243 	u8 *image = NULL;
3244 	int *addrs;
3245 	int pass;
3246 	int i;
3247 
3248 	if (!prog->jit_requested)
3249 		return orig_prog;
3250 
3251 	tmp = bpf_jit_blind_constants(prog);
3252 	/*
3253 	 * If blinding was requested and we failed during blinding,
3254 	 * we must fall back to the interpreter.
3255 	 */
3256 	if (IS_ERR(tmp))
3257 		return orig_prog;
3258 	if (tmp != prog) {
3259 		tmp_blinded = true;
3260 		prog = tmp;
3261 	}
3262 
3263 	jit_data = prog->aux->jit_data;
3264 	if (!jit_data) {
3265 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
3266 		if (!jit_data) {
3267 			prog = orig_prog;
3268 			goto out;
3269 		}
3270 		prog->aux->jit_data = jit_data;
3271 	}
3272 	addrs = jit_data->addrs;
3273 	if (addrs) {
3274 		ctx = jit_data->ctx;
3275 		oldproglen = jit_data->proglen;
3276 		image = jit_data->image;
3277 		header = jit_data->header;
3278 		rw_header = jit_data->rw_header;
3279 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
3280 		extra_pass = true;
3281 		padding = true;
3282 		goto skip_init_addrs;
3283 	}
3284 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
3285 	if (!addrs) {
3286 		prog = orig_prog;
3287 		goto out_addrs;
3288 	}
3289 
3290 	/*
3291 	 * Before first pass, make a rough estimation of addrs[]
3292 	 * each BPF instruction is translated to less than 64 bytes
3293 	 */
3294 	for (proglen = 0, i = 0; i <= prog->len; i++) {
3295 		proglen += 64;
3296 		addrs[i] = proglen;
3297 	}
3298 	ctx.cleanup_addr = proglen;
3299 skip_init_addrs:
3300 
3301 	/*
3302 	 * JITed image shrinks with every pass and the loop iterates
3303 	 * until the image stops shrinking. Very large BPF programs
3304 	 * may converge on the last pass. In such case do one more
3305 	 * pass to emit the final image.
3306 	 */
3307 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
3308 		if (!padding && pass >= PADDING_PASSES)
3309 			padding = true;
3310 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
3311 		if (proglen <= 0) {
3312 out_image:
3313 			image = NULL;
3314 			if (header) {
3315 				bpf_arch_text_copy(&header->size, &rw_header->size,
3316 						   sizeof(rw_header->size));
3317 				bpf_jit_binary_pack_free(header, rw_header);
3318 			}
3319 			/* Fall back to interpreter mode */
3320 			prog = orig_prog;
3321 			if (extra_pass) {
3322 				prog->bpf_func = NULL;
3323 				prog->jited = 0;
3324 				prog->jited_len = 0;
3325 			}
3326 			goto out_addrs;
3327 		}
3328 		if (image) {
3329 			if (proglen != oldproglen) {
3330 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
3331 				       proglen, oldproglen);
3332 				goto out_image;
3333 			}
3334 			break;
3335 		}
3336 		if (proglen == oldproglen) {
3337 			/*
3338 			 * The number of entries in extable is the number of BPF_LDX
3339 			 * insns that access kernel memory via "pointer to BTF type".
3340 			 * The verifier changed their opcode from LDX|MEM|size
3341 			 * to LDX|PROBE_MEM|size to make JITing easier.
3342 			 */
3343 			u32 align = __alignof__(struct exception_table_entry);
3344 			u32 extable_size = prog->aux->num_exentries *
3345 				sizeof(struct exception_table_entry);
3346 
3347 			/* allocate module memory for x86 insns and extable */
3348 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
3349 							   &image, align, &rw_header, &rw_image,
3350 							   jit_fill_hole);
3351 			if (!header) {
3352 				prog = orig_prog;
3353 				goto out_addrs;
3354 			}
3355 			prog->aux->extable = (void *) image + roundup(proglen, align);
3356 		}
3357 		oldproglen = proglen;
3358 		cond_resched();
3359 	}
3360 
3361 	if (bpf_jit_enable > 1)
3362 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
3363 
3364 	if (image) {
3365 		if (!prog->is_func || extra_pass) {
3366 			/*
3367 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
3368 			 *   1) header is not pointing to proper module memory;
3369 			 *   2) the arch doesn't support bpf_arch_text_copy().
3370 			 *
3371 			 * Both cases are serious bugs and justify WARN_ON.
3372 			 */
3373 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
3374 				/* header has been freed */
3375 				header = NULL;
3376 				goto out_image;
3377 			}
3378 
3379 			bpf_tail_call_direct_fixup(prog);
3380 		} else {
3381 			jit_data->addrs = addrs;
3382 			jit_data->ctx = ctx;
3383 			jit_data->proglen = proglen;
3384 			jit_data->image = image;
3385 			jit_data->header = header;
3386 			jit_data->rw_header = rw_header;
3387 		}
3388 		/*
3389 		 * ctx.prog_offset is used when CFI preambles put code *before*
3390 		 * the function. See emit_cfi(). For FineIBT specifically this code
3391 		 * can also be executed and bpf_prog_kallsyms_add() will
3392 		 * generate an additional symbol to cover this, hence also
3393 		 * decrement proglen.
3394 		 */
3395 		prog->bpf_func = (void *)image + cfi_get_offset();
3396 		prog->jited = 1;
3397 		prog->jited_len = proglen - cfi_get_offset();
3398 	} else {
3399 		prog = orig_prog;
3400 	}
3401 
3402 	if (!image || !prog->is_func || extra_pass) {
3403 		if (image)
3404 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
3405 out_addrs:
3406 		kvfree(addrs);
3407 		kfree(jit_data);
3408 		prog->aux->jit_data = NULL;
3409 	}
3410 out:
3411 	if (tmp_blinded)
3412 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
3413 					   tmp : orig_prog);
3414 	return prog;
3415 }
3416 
3417 bool bpf_jit_supports_kfunc_call(void)
3418 {
3419 	return true;
3420 }
3421 
3422 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
3423 {
3424 	if (text_poke_copy(dst, src, len) == NULL)
3425 		return ERR_PTR(-EINVAL);
3426 	return dst;
3427 }
3428 
3429 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
3430 bool bpf_jit_supports_subprog_tailcalls(void)
3431 {
3432 	return true;
3433 }
3434 
3435 bool bpf_jit_supports_percpu_insn(void)
3436 {
3437 	return true;
3438 }
3439 
3440 void bpf_jit_free(struct bpf_prog *prog)
3441 {
3442 	if (prog->jited) {
3443 		struct x64_jit_data *jit_data = prog->aux->jit_data;
3444 		struct bpf_binary_header *hdr;
3445 
3446 		/*
3447 		 * If we fail the final pass of JIT (from jit_subprogs),
3448 		 * the program may not be finalized yet. Call finalize here
3449 		 * before freeing it.
3450 		 */
3451 		if (jit_data) {
3452 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
3453 						     jit_data->rw_header);
3454 			kvfree(jit_data->addrs);
3455 			kfree(jit_data);
3456 		}
3457 		prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset();
3458 		hdr = bpf_jit_binary_pack_hdr(prog);
3459 		bpf_jit_binary_pack_free(hdr, NULL);
3460 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
3461 	}
3462 
3463 	bpf_prog_unlock_free(prog);
3464 }
3465 
3466 bool bpf_jit_supports_exceptions(void)
3467 {
3468 	/* We unwind through both kernel frames (starting from within bpf_throw
3469 	 * call) and BPF frames. Therefore we require ORC unwinder to be enabled
3470 	 * to walk kernel frames and reach BPF frames in the stack trace.
3471 	 */
3472 	return IS_ENABLED(CONFIG_UNWINDER_ORC);
3473 }
3474 
3475 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3476 {
3477 #if defined(CONFIG_UNWINDER_ORC)
3478 	struct unwind_state state;
3479 	unsigned long addr;
3480 
3481 	for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
3482 	     unwind_next_frame(&state)) {
3483 		addr = unwind_get_return_address(&state);
3484 		if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
3485 			break;
3486 	}
3487 	return;
3488 #endif
3489 	WARN(1, "verification of programs using bpf_throw should have failed\n");
3490 }
3491 
3492 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
3493 			       struct bpf_prog *new, struct bpf_prog *old)
3494 {
3495 	u8 *old_addr, *new_addr, *old_bypass_addr;
3496 	int ret;
3497 
3498 	old_bypass_addr = old ? NULL : poke->bypass_addr;
3499 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
3500 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
3501 
3502 	/*
3503 	 * On program loading or teardown, the program's kallsym entry
3504 	 * might not be in place, so we use __bpf_arch_text_poke to skip
3505 	 * the kallsyms check.
3506 	 */
3507 	if (new) {
3508 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3509 					   BPF_MOD_JUMP,
3510 					   old_addr, new_addr);
3511 		BUG_ON(ret < 0);
3512 		if (!old) {
3513 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3514 						   BPF_MOD_JUMP,
3515 						   poke->bypass_addr,
3516 						   NULL);
3517 			BUG_ON(ret < 0);
3518 		}
3519 	} else {
3520 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3521 					   BPF_MOD_JUMP,
3522 					   old_bypass_addr,
3523 					   poke->bypass_addr);
3524 		BUG_ON(ret < 0);
3525 		/* let other CPUs finish the execution of program
3526 		 * so that it will not possible to expose them
3527 		 * to invalid nop, stack unwind, nop state
3528 		 */
3529 		if (!ret)
3530 			synchronize_rcu();
3531 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3532 					   BPF_MOD_JUMP,
3533 					   old_addr, NULL);
3534 		BUG_ON(ret < 0);
3535 	}
3536 }
3537 
3538 bool bpf_jit_supports_arena(void)
3539 {
3540 	return true;
3541 }
3542 
3543 bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
3544 {
3545 	if (!in_arena)
3546 		return true;
3547 	switch (insn->code) {
3548 	case BPF_STX | BPF_ATOMIC | BPF_W:
3549 	case BPF_STX | BPF_ATOMIC | BPF_DW:
3550 		if (insn->imm == (BPF_AND | BPF_FETCH) ||
3551 		    insn->imm == (BPF_OR | BPF_FETCH) ||
3552 		    insn->imm == (BPF_XOR | BPF_FETCH))
3553 			return false;
3554 	}
3555 	return true;
3556 }
3557 
3558 bool bpf_jit_supports_ptr_xchg(void)
3559 {
3560 	return true;
3561 }
3562