xref: /linux/arch/x86/net/bpf_jit_comp.c (revision b8e85e6f3a09fc56b0ff574887798962ef8a8f80)
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 
117 /*
118  * The following table maps BPF registers to x86-64 registers.
119  *
120  * x86-64 register R12 is unused, since if used as base address
121  * register in load/store instructions, it always needs an
122  * extra byte of encoding and is callee saved.
123  *
124  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
125  * trampoline. x86-64 register R10 is used for blinding (if enabled).
126  */
127 static const int reg2hex[] = {
128 	[BPF_REG_0] = 0,  /* RAX */
129 	[BPF_REG_1] = 7,  /* RDI */
130 	[BPF_REG_2] = 6,  /* RSI */
131 	[BPF_REG_3] = 2,  /* RDX */
132 	[BPF_REG_4] = 1,  /* RCX */
133 	[BPF_REG_5] = 0,  /* R8  */
134 	[BPF_REG_6] = 3,  /* RBX callee saved */
135 	[BPF_REG_7] = 5,  /* R13 callee saved */
136 	[BPF_REG_8] = 6,  /* R14 callee saved */
137 	[BPF_REG_9] = 7,  /* R15 callee saved */
138 	[BPF_REG_FP] = 5, /* RBP readonly */
139 	[BPF_REG_AX] = 2, /* R10 temp register */
140 	[AUX_REG] = 3,    /* R11 temp register */
141 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
142 };
143 
144 static const int reg2pt_regs[] = {
145 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
146 	[BPF_REG_1] = offsetof(struct pt_regs, di),
147 	[BPF_REG_2] = offsetof(struct pt_regs, si),
148 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
149 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
150 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
151 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
152 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
153 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
154 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
155 };
156 
157 /*
158  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
159  * which need extra byte of encoding.
160  * rax,rcx,...,rbp have simpler encoding
161  */
162 static bool is_ereg(u32 reg)
163 {
164 	return (1 << reg) & (BIT(BPF_REG_5) |
165 			     BIT(AUX_REG) |
166 			     BIT(BPF_REG_7) |
167 			     BIT(BPF_REG_8) |
168 			     BIT(BPF_REG_9) |
169 			     BIT(X86_REG_R9) |
170 			     BIT(BPF_REG_AX));
171 }
172 
173 /*
174  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
175  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
176  * of encoding. al,cl,dl,bl have simpler encoding.
177  */
178 static bool is_ereg_8l(u32 reg)
179 {
180 	return is_ereg(reg) ||
181 	    (1 << reg) & (BIT(BPF_REG_1) |
182 			  BIT(BPF_REG_2) |
183 			  BIT(BPF_REG_FP));
184 }
185 
186 static bool is_axreg(u32 reg)
187 {
188 	return reg == BPF_REG_0;
189 }
190 
191 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
192 static u8 add_1mod(u8 byte, u32 reg)
193 {
194 	if (is_ereg(reg))
195 		byte |= 1;
196 	return byte;
197 }
198 
199 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
200 {
201 	if (is_ereg(r1))
202 		byte |= 1;
203 	if (is_ereg(r2))
204 		byte |= 4;
205 	return byte;
206 }
207 
208 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
209 static u8 add_1reg(u8 byte, u32 dst_reg)
210 {
211 	return byte + reg2hex[dst_reg];
212 }
213 
214 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
215 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
216 {
217 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
218 }
219 
220 /* Some 1-byte opcodes for binary ALU operations */
221 static u8 simple_alu_opcodes[] = {
222 	[BPF_ADD] = 0x01,
223 	[BPF_SUB] = 0x29,
224 	[BPF_AND] = 0x21,
225 	[BPF_OR] = 0x09,
226 	[BPF_XOR] = 0x31,
227 	[BPF_LSH] = 0xE0,
228 	[BPF_RSH] = 0xE8,
229 	[BPF_ARSH] = 0xF8,
230 };
231 
232 static void jit_fill_hole(void *area, unsigned int size)
233 {
234 	/* Fill whole space with INT3 instructions */
235 	memset(area, 0xcc, size);
236 }
237 
238 int bpf_arch_text_invalidate(void *dst, size_t len)
239 {
240 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
241 }
242 
243 struct jit_context {
244 	int cleanup_addr; /* Epilogue code offset */
245 
246 	/*
247 	 * Program specific offsets of labels in the code; these rely on the
248 	 * JIT doing at least 2 passes, recording the position on the first
249 	 * pass, only to generate the correct offset on the second pass.
250 	 */
251 	int tail_call_direct_label;
252 	int tail_call_indirect_label;
253 };
254 
255 /* Maximum number of bytes emitted while JITing one eBPF insn */
256 #define BPF_MAX_INSN_SIZE	128
257 #define BPF_INSN_SAFETY		64
258 
259 /* Number of bytes emit_patch() needs to generate instructions */
260 #define X86_PATCH_SIZE		5
261 /* Number of bytes that will be skipped on tailcall */
262 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
263 
264 static void push_r12(u8 **pprog)
265 {
266 	u8 *prog = *pprog;
267 
268 	EMIT2(0x41, 0x54);   /* push r12 */
269 	*pprog = prog;
270 }
271 
272 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
273 {
274 	u8 *prog = *pprog;
275 
276 	if (callee_regs_used[0])
277 		EMIT1(0x53);         /* push rbx */
278 	if (callee_regs_used[1])
279 		EMIT2(0x41, 0x55);   /* push r13 */
280 	if (callee_regs_used[2])
281 		EMIT2(0x41, 0x56);   /* push r14 */
282 	if (callee_regs_used[3])
283 		EMIT2(0x41, 0x57);   /* push r15 */
284 	*pprog = prog;
285 }
286 
287 static void pop_r12(u8 **pprog)
288 {
289 	u8 *prog = *pprog;
290 
291 	EMIT2(0x41, 0x5C);   /* pop r12 */
292 	*pprog = prog;
293 }
294 
295 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
296 {
297 	u8 *prog = *pprog;
298 
299 	if (callee_regs_used[3])
300 		EMIT2(0x41, 0x5F);   /* pop r15 */
301 	if (callee_regs_used[2])
302 		EMIT2(0x41, 0x5E);   /* pop r14 */
303 	if (callee_regs_used[1])
304 		EMIT2(0x41, 0x5D);   /* pop r13 */
305 	if (callee_regs_used[0])
306 		EMIT1(0x5B);         /* pop rbx */
307 	*pprog = prog;
308 }
309 
310 static void emit_nops(u8 **pprog, int len)
311 {
312 	u8 *prog = *pprog;
313 	int i, noplen;
314 
315 	while (len > 0) {
316 		noplen = len;
317 
318 		if (noplen > ASM_NOP_MAX)
319 			noplen = ASM_NOP_MAX;
320 
321 		for (i = 0; i < noplen; i++)
322 			EMIT1(x86_nops[noplen][i]);
323 		len -= noplen;
324 	}
325 
326 	*pprog = prog;
327 }
328 
329 /*
330  * Emit the various CFI preambles, see asm/cfi.h and the comments about FineIBT
331  * in arch/x86/kernel/alternative.c
332  */
333 
334 static void emit_fineibt(u8 **pprog, u32 hash)
335 {
336 	u8 *prog = *pprog;
337 
338 	EMIT_ENDBR();
339 	EMIT3_off32(0x41, 0x81, 0xea, hash);		/* subl $hash, %r10d	*/
340 	EMIT2(0x74, 0x07);				/* jz.d8 +7		*/
341 	EMIT2(0x0f, 0x0b);				/* ud2			*/
342 	EMIT1(0x90);					/* nop			*/
343 	EMIT_ENDBR_POISON();
344 
345 	*pprog = prog;
346 }
347 
348 static void emit_kcfi(u8 **pprog, u32 hash)
349 {
350 	u8 *prog = *pprog;
351 
352 	EMIT1_off32(0xb8, hash);			/* movl $hash, %eax	*/
353 #ifdef CONFIG_CALL_PADDING
354 	EMIT1(0x90);
355 	EMIT1(0x90);
356 	EMIT1(0x90);
357 	EMIT1(0x90);
358 	EMIT1(0x90);
359 	EMIT1(0x90);
360 	EMIT1(0x90);
361 	EMIT1(0x90);
362 	EMIT1(0x90);
363 	EMIT1(0x90);
364 	EMIT1(0x90);
365 #endif
366 	EMIT_ENDBR();
367 
368 	*pprog = prog;
369 }
370 
371 static void emit_cfi(u8 **pprog, u32 hash)
372 {
373 	u8 *prog = *pprog;
374 
375 	switch (cfi_mode) {
376 	case CFI_FINEIBT:
377 		emit_fineibt(&prog, hash);
378 		break;
379 
380 	case CFI_KCFI:
381 		emit_kcfi(&prog, hash);
382 		break;
383 
384 	default:
385 		EMIT_ENDBR();
386 		break;
387 	}
388 
389 	*pprog = prog;
390 }
391 
392 /*
393  * Emit x86-64 prologue code for BPF program.
394  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
395  * while jumping to another program
396  */
397 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
398 			  bool tail_call_reachable, bool is_subprog,
399 			  bool is_exception_cb)
400 {
401 	u8 *prog = *pprog;
402 
403 	emit_cfi(&prog, is_subprog ? cfi_bpf_subprog_hash : cfi_bpf_hash);
404 	/* BPF trampoline can be made to work without these nops,
405 	 * but let's waste 5 bytes for now and optimize later
406 	 */
407 	emit_nops(&prog, X86_PATCH_SIZE);
408 	if (!ebpf_from_cbpf) {
409 		if (tail_call_reachable && !is_subprog)
410 			/* When it's the entry of the whole tailcall context,
411 			 * zeroing rax means initialising tail_call_cnt.
412 			 */
413 			EMIT2(0x31, 0xC0); /* xor eax, eax */
414 		else
415 			/* Keep the same instruction layout. */
416 			EMIT2(0x66, 0x90); /* nop2 */
417 	}
418 	/* Exception callback receives FP as third parameter */
419 	if (is_exception_cb) {
420 		EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */
421 		EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */
422 		/* The main frame must have exception_boundary as true, so we
423 		 * first restore those callee-saved regs from stack, before
424 		 * reusing the stack frame.
425 		 */
426 		pop_callee_regs(&prog, all_callee_regs_used);
427 		pop_r12(&prog);
428 		/* Reset the stack frame. */
429 		EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */
430 	} else {
431 		EMIT1(0x55);             /* push rbp */
432 		EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
433 	}
434 
435 	/* X86_TAIL_CALL_OFFSET is here */
436 	EMIT_ENDBR();
437 
438 	/* sub rsp, rounded_stack_depth */
439 	if (stack_depth)
440 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
441 	if (tail_call_reachable)
442 		EMIT1(0x50);         /* push rax */
443 	*pprog = prog;
444 }
445 
446 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
447 {
448 	u8 *prog = *pprog;
449 	s64 offset;
450 
451 	offset = func - (ip + X86_PATCH_SIZE);
452 	if (!is_simm32(offset)) {
453 		pr_err("Target call %p is out of range\n", func);
454 		return -ERANGE;
455 	}
456 	EMIT1_off32(opcode, offset);
457 	*pprog = prog;
458 	return 0;
459 }
460 
461 static int emit_call(u8 **pprog, void *func, void *ip)
462 {
463 	return emit_patch(pprog, func, ip, 0xE8);
464 }
465 
466 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
467 {
468 	OPTIMIZER_HIDE_VAR(func);
469 	x86_call_depth_emit_accounting(pprog, func);
470 	return emit_patch(pprog, func, ip, 0xE8);
471 }
472 
473 static int emit_jump(u8 **pprog, void *func, void *ip)
474 {
475 	return emit_patch(pprog, func, ip, 0xE9);
476 }
477 
478 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
479 				void *old_addr, void *new_addr)
480 {
481 	const u8 *nop_insn = x86_nops[5];
482 	u8 old_insn[X86_PATCH_SIZE];
483 	u8 new_insn[X86_PATCH_SIZE];
484 	u8 *prog;
485 	int ret;
486 
487 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
488 	if (old_addr) {
489 		prog = old_insn;
490 		ret = t == BPF_MOD_CALL ?
491 		      emit_call(&prog, old_addr, ip) :
492 		      emit_jump(&prog, old_addr, ip);
493 		if (ret)
494 			return ret;
495 	}
496 
497 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
498 	if (new_addr) {
499 		prog = new_insn;
500 		ret = t == BPF_MOD_CALL ?
501 		      emit_call(&prog, new_addr, ip) :
502 		      emit_jump(&prog, new_addr, ip);
503 		if (ret)
504 			return ret;
505 	}
506 
507 	ret = -EBUSY;
508 	mutex_lock(&text_mutex);
509 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
510 		goto out;
511 	ret = 1;
512 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
513 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
514 		ret = 0;
515 	}
516 out:
517 	mutex_unlock(&text_mutex);
518 	return ret;
519 }
520 
521 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
522 		       void *old_addr, void *new_addr)
523 {
524 	if (!is_kernel_text((long)ip) &&
525 	    !is_bpf_text_address((long)ip))
526 		/* BPF poking in modules is not supported */
527 		return -EINVAL;
528 
529 	/*
530 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
531 	 * with an ENDBR instruction.
532 	 */
533 	if (is_endbr(*(u32 *)ip))
534 		ip += ENDBR_INSN_SIZE;
535 
536 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
537 }
538 
539 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
540 
541 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
542 {
543 	u8 *prog = *pprog;
544 
545 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
546 		EMIT_LFENCE();
547 		EMIT2(0xFF, 0xE0 + reg);
548 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
549 		OPTIMIZER_HIDE_VAR(reg);
550 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
551 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
552 		else
553 			emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
554 	} else {
555 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
556 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
557 			EMIT1(0xCC);		/* int3 */
558 	}
559 
560 	*pprog = prog;
561 }
562 
563 static void emit_return(u8 **pprog, u8 *ip)
564 {
565 	u8 *prog = *pprog;
566 
567 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
568 		emit_jump(&prog, x86_return_thunk, ip);
569 	} else {
570 		EMIT1(0xC3);		/* ret */
571 		if (IS_ENABLED(CONFIG_SLS))
572 			EMIT1(0xCC);	/* int3 */
573 	}
574 
575 	*pprog = prog;
576 }
577 
578 /*
579  * Generate the following code:
580  *
581  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
582  *   if (index >= array->map.max_entries)
583  *     goto out;
584  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
585  *     goto out;
586  *   prog = array->ptrs[index];
587  *   if (prog == NULL)
588  *     goto out;
589  *   goto *(prog->bpf_func + prologue_size);
590  * out:
591  */
592 static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog,
593 					u8 **pprog, bool *callee_regs_used,
594 					u32 stack_depth, u8 *ip,
595 					struct jit_context *ctx)
596 {
597 	int tcc_off = -4 - round_up(stack_depth, 8);
598 	u8 *prog = *pprog, *start = *pprog;
599 	int offset;
600 
601 	/*
602 	 * rdi - pointer to ctx
603 	 * rsi - pointer to bpf_array
604 	 * rdx - index in bpf_array
605 	 */
606 
607 	/*
608 	 * if (index >= array->map.max_entries)
609 	 *	goto out;
610 	 */
611 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
612 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
613 	      offsetof(struct bpf_array, map.max_entries));
614 
615 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
616 	EMIT2(X86_JBE, offset);                   /* jbe out */
617 
618 	/*
619 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
620 	 *	goto out;
621 	 */
622 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
623 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
624 
625 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
626 	EMIT2(X86_JAE, offset);                   /* jae out */
627 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
628 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
629 
630 	/* prog = array->ptrs[index]; */
631 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
632 		    offsetof(struct bpf_array, ptrs));
633 
634 	/*
635 	 * if (prog == NULL)
636 	 *	goto out;
637 	 */
638 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
639 
640 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
641 	EMIT2(X86_JE, offset);                    /* je out */
642 
643 	if (bpf_prog->aux->exception_boundary) {
644 		pop_callee_regs(&prog, all_callee_regs_used);
645 		pop_r12(&prog);
646 	} else {
647 		pop_callee_regs(&prog, callee_regs_used);
648 	}
649 
650 	EMIT1(0x58);                              /* pop rax */
651 	if (stack_depth)
652 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
653 			    round_up(stack_depth, 8));
654 
655 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
656 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
657 	      offsetof(struct bpf_prog, bpf_func));
658 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
659 	      X86_TAIL_CALL_OFFSET);
660 	/*
661 	 * Now we're ready to jump into next BPF program
662 	 * rdi == ctx (1st arg)
663 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
664 	 */
665 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
666 
667 	/* out: */
668 	ctx->tail_call_indirect_label = prog - start;
669 	*pprog = prog;
670 }
671 
672 static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog,
673 				      struct bpf_jit_poke_descriptor *poke,
674 				      u8 **pprog, u8 *ip,
675 				      bool *callee_regs_used, u32 stack_depth,
676 				      struct jit_context *ctx)
677 {
678 	int tcc_off = -4 - round_up(stack_depth, 8);
679 	u8 *prog = *pprog, *start = *pprog;
680 	int offset;
681 
682 	/*
683 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
684 	 *	goto out;
685 	 */
686 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
687 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
688 
689 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
690 	EMIT2(X86_JAE, offset);                       /* jae out */
691 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
692 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
693 
694 	poke->tailcall_bypass = ip + (prog - start);
695 	poke->adj_off = X86_TAIL_CALL_OFFSET;
696 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
697 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
698 
699 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
700 		  poke->tailcall_bypass);
701 
702 	if (bpf_prog->aux->exception_boundary) {
703 		pop_callee_regs(&prog, all_callee_regs_used);
704 		pop_r12(&prog);
705 	} else {
706 		pop_callee_regs(&prog, callee_regs_used);
707 	}
708 
709 	EMIT1(0x58);                                  /* pop rax */
710 	if (stack_depth)
711 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
712 
713 	emit_nops(&prog, X86_PATCH_SIZE);
714 
715 	/* out: */
716 	ctx->tail_call_direct_label = prog - start;
717 
718 	*pprog = prog;
719 }
720 
721 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
722 {
723 	struct bpf_jit_poke_descriptor *poke;
724 	struct bpf_array *array;
725 	struct bpf_prog *target;
726 	int i, ret;
727 
728 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
729 		poke = &prog->aux->poke_tab[i];
730 		if (poke->aux && poke->aux != prog->aux)
731 			continue;
732 
733 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
734 
735 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
736 			continue;
737 
738 		array = container_of(poke->tail_call.map, struct bpf_array, map);
739 		mutex_lock(&array->aux->poke_mutex);
740 		target = array->ptrs[poke->tail_call.key];
741 		if (target) {
742 			ret = __bpf_arch_text_poke(poke->tailcall_target,
743 						   BPF_MOD_JUMP, NULL,
744 						   (u8 *)target->bpf_func +
745 						   poke->adj_off);
746 			BUG_ON(ret < 0);
747 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
748 						   BPF_MOD_JUMP,
749 						   (u8 *)poke->tailcall_target +
750 						   X86_PATCH_SIZE, NULL);
751 			BUG_ON(ret < 0);
752 		}
753 		WRITE_ONCE(poke->tailcall_target_stable, true);
754 		mutex_unlock(&array->aux->poke_mutex);
755 	}
756 }
757 
758 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
759 			   u32 dst_reg, const u32 imm32)
760 {
761 	u8 *prog = *pprog;
762 	u8 b1, b2, b3;
763 
764 	/*
765 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
766 	 * (which zero-extends imm32) to save 2 bytes.
767 	 */
768 	if (sign_propagate && (s32)imm32 < 0) {
769 		/* 'mov %rax, imm32' sign extends imm32 */
770 		b1 = add_1mod(0x48, dst_reg);
771 		b2 = 0xC7;
772 		b3 = 0xC0;
773 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
774 		goto done;
775 	}
776 
777 	/*
778 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
779 	 * to save 3 bytes.
780 	 */
781 	if (imm32 == 0) {
782 		if (is_ereg(dst_reg))
783 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
784 		b2 = 0x31; /* xor */
785 		b3 = 0xC0;
786 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
787 		goto done;
788 	}
789 
790 	/* mov %eax, imm32 */
791 	if (is_ereg(dst_reg))
792 		EMIT1(add_1mod(0x40, dst_reg));
793 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
794 done:
795 	*pprog = prog;
796 }
797 
798 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
799 			   const u32 imm32_hi, const u32 imm32_lo)
800 {
801 	u8 *prog = *pprog;
802 
803 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
804 		/*
805 		 * For emitting plain u32, where sign bit must not be
806 		 * propagated LLVM tends to load imm64 over mov32
807 		 * directly, so save couple of bytes by just doing
808 		 * 'mov %eax, imm32' instead.
809 		 */
810 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
811 	} else {
812 		/* movabsq rax, imm64 */
813 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
814 		EMIT(imm32_lo, 4);
815 		EMIT(imm32_hi, 4);
816 	}
817 
818 	*pprog = prog;
819 }
820 
821 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
822 {
823 	u8 *prog = *pprog;
824 
825 	if (is64) {
826 		/* mov dst, src */
827 		EMIT_mov(dst_reg, src_reg);
828 	} else {
829 		/* mov32 dst, src */
830 		if (is_ereg(dst_reg) || is_ereg(src_reg))
831 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
832 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
833 	}
834 
835 	*pprog = prog;
836 }
837 
838 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
839 			   u32 src_reg)
840 {
841 	u8 *prog = *pprog;
842 
843 	if (is64) {
844 		/* movs[b,w,l]q dst, src */
845 		if (num_bits == 8)
846 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
847 			      add_2reg(0xC0, src_reg, dst_reg));
848 		else if (num_bits == 16)
849 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
850 			      add_2reg(0xC0, src_reg, dst_reg));
851 		else if (num_bits == 32)
852 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
853 			      add_2reg(0xC0, src_reg, dst_reg));
854 	} else {
855 		/* movs[b,w]l dst, src */
856 		if (num_bits == 8) {
857 			EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
858 			      add_2reg(0xC0, src_reg, dst_reg));
859 		} else if (num_bits == 16) {
860 			if (is_ereg(dst_reg) || is_ereg(src_reg))
861 				EMIT1(add_2mod(0x40, src_reg, dst_reg));
862 			EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
863 			      add_2reg(0xC0, src_reg, dst_reg));
864 		}
865 	}
866 
867 	*pprog = prog;
868 }
869 
870 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
871 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
872 {
873 	u8 *prog = *pprog;
874 
875 	if (is_imm8(off)) {
876 		/* 1-byte signed displacement.
877 		 *
878 		 * If off == 0 we could skip this and save one extra byte, but
879 		 * special case of x86 R13 which always needs an offset is not
880 		 * worth the hassle
881 		 */
882 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
883 	} else {
884 		/* 4-byte signed displacement */
885 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
886 	}
887 	*pprog = prog;
888 }
889 
890 /*
891  * Emit a REX byte if it will be necessary to address these registers
892  */
893 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
894 {
895 	u8 *prog = *pprog;
896 
897 	if (is64)
898 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
899 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
900 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
901 	*pprog = prog;
902 }
903 
904 /*
905  * Similar version of maybe_emit_mod() for a single register
906  */
907 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
908 {
909 	u8 *prog = *pprog;
910 
911 	if (is64)
912 		EMIT1(add_1mod(0x48, reg));
913 	else if (is_ereg(reg))
914 		EMIT1(add_1mod(0x40, reg));
915 	*pprog = prog;
916 }
917 
918 /* LDX: dst_reg = *(u8*)(src_reg + off) */
919 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
920 {
921 	u8 *prog = *pprog;
922 
923 	switch (size) {
924 	case BPF_B:
925 		/* Emit 'movzx rax, byte ptr [rax + off]' */
926 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
927 		break;
928 	case BPF_H:
929 		/* Emit 'movzx rax, word ptr [rax + off]' */
930 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
931 		break;
932 	case BPF_W:
933 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
934 		if (is_ereg(dst_reg) || is_ereg(src_reg))
935 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
936 		else
937 			EMIT1(0x8B);
938 		break;
939 	case BPF_DW:
940 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
941 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
942 		break;
943 	}
944 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
945 	*pprog = prog;
946 }
947 
948 /* LDSX: dst_reg = *(s8*)(src_reg + off) */
949 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
950 {
951 	u8 *prog = *pprog;
952 
953 	switch (size) {
954 	case BPF_B:
955 		/* Emit 'movsx rax, byte ptr [rax + off]' */
956 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
957 		break;
958 	case BPF_H:
959 		/* Emit 'movsx rax, word ptr [rax + off]' */
960 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
961 		break;
962 	case BPF_W:
963 		/* Emit 'movsx rax, dword ptr [rax+0x14]' */
964 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
965 		break;
966 	}
967 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
968 	*pprog = prog;
969 }
970 
971 /* STX: *(u8*)(dst_reg + off) = src_reg */
972 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
973 {
974 	u8 *prog = *pprog;
975 
976 	switch (size) {
977 	case BPF_B:
978 		/* Emit 'mov byte ptr [rax + off], al' */
979 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
980 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
981 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
982 		else
983 			EMIT1(0x88);
984 		break;
985 	case BPF_H:
986 		if (is_ereg(dst_reg) || is_ereg(src_reg))
987 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
988 		else
989 			EMIT2(0x66, 0x89);
990 		break;
991 	case BPF_W:
992 		if (is_ereg(dst_reg) || is_ereg(src_reg))
993 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
994 		else
995 			EMIT1(0x89);
996 		break;
997 	case BPF_DW:
998 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
999 		break;
1000 	}
1001 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1002 	*pprog = prog;
1003 }
1004 
1005 static int emit_atomic(u8 **pprog, u8 atomic_op,
1006 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
1007 {
1008 	u8 *prog = *pprog;
1009 
1010 	EMIT1(0xF0); /* lock prefix */
1011 
1012 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
1013 
1014 	/* emit opcode */
1015 	switch (atomic_op) {
1016 	case BPF_ADD:
1017 	case BPF_AND:
1018 	case BPF_OR:
1019 	case BPF_XOR:
1020 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
1021 		EMIT1(simple_alu_opcodes[atomic_op]);
1022 		break;
1023 	case BPF_ADD | BPF_FETCH:
1024 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
1025 		EMIT2(0x0F, 0xC1);
1026 		break;
1027 	case BPF_XCHG:
1028 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
1029 		EMIT1(0x87);
1030 		break;
1031 	case BPF_CMPXCHG:
1032 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
1033 		EMIT2(0x0F, 0xB1);
1034 		break;
1035 	default:
1036 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
1037 		return -EFAULT;
1038 	}
1039 
1040 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1041 
1042 	*pprog = prog;
1043 	return 0;
1044 }
1045 
1046 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
1047 {
1048 	u32 reg = x->fixup >> 8;
1049 
1050 	/* jump over faulting load and clear dest register */
1051 	*(unsigned long *)((void *)regs + reg) = 0;
1052 	regs->ip += x->fixup & 0xff;
1053 	return true;
1054 }
1055 
1056 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
1057 			     bool *regs_used, bool *tail_call_seen)
1058 {
1059 	int i;
1060 
1061 	for (i = 1; i <= insn_cnt; i++, insn++) {
1062 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
1063 			*tail_call_seen = true;
1064 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
1065 			regs_used[0] = true;
1066 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
1067 			regs_used[1] = true;
1068 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
1069 			regs_used[2] = true;
1070 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
1071 			regs_used[3] = true;
1072 	}
1073 }
1074 
1075 /* emit the 3-byte VEX prefix
1076  *
1077  * r: same as rex.r, extra bit for ModRM reg field
1078  * x: same as rex.x, extra bit for SIB index field
1079  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
1080  * m: opcode map select, encoding escape bytes e.g. 0x0f38
1081  * w: same as rex.w (32 bit or 64 bit) or opcode specific
1082  * src_reg2: additional source reg (encoded as BPF reg)
1083  * l: vector length (128 bit or 256 bit) or reserved
1084  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
1085  */
1086 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
1087 		      bool w, u8 src_reg2, bool l, u8 pp)
1088 {
1089 	u8 *prog = *pprog;
1090 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
1091 	u8 b1, b2;
1092 	u8 vvvv = reg2hex[src_reg2];
1093 
1094 	/* reg2hex gives only the lower 3 bit of vvvv */
1095 	if (is_ereg(src_reg2))
1096 		vvvv |= 1 << 3;
1097 
1098 	/*
1099 	 * 2nd byte of 3-byte VEX prefix
1100 	 * ~ means bit inverted encoding
1101 	 *
1102 	 *    7                           0
1103 	 *  +---+---+---+---+---+---+---+---+
1104 	 *  |~R |~X |~B |         m         |
1105 	 *  +---+---+---+---+---+---+---+---+
1106 	 */
1107 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
1108 	/*
1109 	 * 3rd byte of 3-byte VEX prefix
1110 	 *
1111 	 *    7                           0
1112 	 *  +---+---+---+---+---+---+---+---+
1113 	 *  | W |     ~vvvv     | L |   pp  |
1114 	 *  +---+---+---+---+---+---+---+---+
1115 	 */
1116 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1117 
1118 	EMIT3(b0, b1, b2);
1119 	*pprog = prog;
1120 }
1121 
1122 /* emit BMI2 shift instruction */
1123 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1124 {
1125 	u8 *prog = *pprog;
1126 	bool r = is_ereg(dst_reg);
1127 	u8 m = 2; /* escape code 0f38 */
1128 
1129 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1130 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1131 	*pprog = prog;
1132 }
1133 
1134 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1135 
1136 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1137 #define RESTORE_TAIL_CALL_CNT(stack)				\
1138 	EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
1139 
1140 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
1141 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
1142 {
1143 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1144 	struct bpf_insn *insn = bpf_prog->insnsi;
1145 	bool callee_regs_used[4] = {};
1146 	int insn_cnt = bpf_prog->len;
1147 	bool tail_call_seen = false;
1148 	bool seen_exit = false;
1149 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1150 	int i, excnt = 0;
1151 	int ilen, proglen = 0;
1152 	u8 *prog = temp;
1153 	int err;
1154 
1155 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
1156 			 &tail_call_seen);
1157 
1158 	/* tail call's presence in current prog implies it is reachable */
1159 	tail_call_reachable |= tail_call_seen;
1160 
1161 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
1162 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1163 		      bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
1164 	/* Exception callback will clobber callee regs for its own use, and
1165 	 * restore the original callee regs from main prog's stack frame.
1166 	 */
1167 	if (bpf_prog->aux->exception_boundary) {
1168 		/* We also need to save r12, which is not mapped to any BPF
1169 		 * register, as we throw after entry into the kernel, which may
1170 		 * overwrite r12.
1171 		 */
1172 		push_r12(&prog);
1173 		push_callee_regs(&prog, all_callee_regs_used);
1174 	} else {
1175 		push_callee_regs(&prog, callee_regs_used);
1176 	}
1177 
1178 	ilen = prog - temp;
1179 	if (rw_image)
1180 		memcpy(rw_image + proglen, temp, ilen);
1181 	proglen += ilen;
1182 	addrs[0] = proglen;
1183 	prog = temp;
1184 
1185 	for (i = 1; i <= insn_cnt; i++, insn++) {
1186 		const s32 imm32 = insn->imm;
1187 		u32 dst_reg = insn->dst_reg;
1188 		u32 src_reg = insn->src_reg;
1189 		u8 b2 = 0, b3 = 0;
1190 		u8 *start_of_ldx;
1191 		s64 jmp_offset;
1192 		s16 insn_off;
1193 		u8 jmp_cond;
1194 		u8 *func;
1195 		int nops;
1196 
1197 		switch (insn->code) {
1198 			/* ALU */
1199 		case BPF_ALU | BPF_ADD | BPF_X:
1200 		case BPF_ALU | BPF_SUB | BPF_X:
1201 		case BPF_ALU | BPF_AND | BPF_X:
1202 		case BPF_ALU | BPF_OR | BPF_X:
1203 		case BPF_ALU | BPF_XOR | BPF_X:
1204 		case BPF_ALU64 | BPF_ADD | BPF_X:
1205 		case BPF_ALU64 | BPF_SUB | BPF_X:
1206 		case BPF_ALU64 | BPF_AND | BPF_X:
1207 		case BPF_ALU64 | BPF_OR | BPF_X:
1208 		case BPF_ALU64 | BPF_XOR | BPF_X:
1209 			maybe_emit_mod(&prog, dst_reg, src_reg,
1210 				       BPF_CLASS(insn->code) == BPF_ALU64);
1211 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1212 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1213 			break;
1214 
1215 		case BPF_ALU64 | BPF_MOV | BPF_X:
1216 		case BPF_ALU | BPF_MOV | BPF_X:
1217 			if (insn->off == 0)
1218 				emit_mov_reg(&prog,
1219 					     BPF_CLASS(insn->code) == BPF_ALU64,
1220 					     dst_reg, src_reg);
1221 			else
1222 				emit_movsx_reg(&prog, insn->off,
1223 					       BPF_CLASS(insn->code) == BPF_ALU64,
1224 					       dst_reg, src_reg);
1225 			break;
1226 
1227 			/* neg dst */
1228 		case BPF_ALU | BPF_NEG:
1229 		case BPF_ALU64 | BPF_NEG:
1230 			maybe_emit_1mod(&prog, dst_reg,
1231 					BPF_CLASS(insn->code) == BPF_ALU64);
1232 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1233 			break;
1234 
1235 		case BPF_ALU | BPF_ADD | BPF_K:
1236 		case BPF_ALU | BPF_SUB | BPF_K:
1237 		case BPF_ALU | BPF_AND | BPF_K:
1238 		case BPF_ALU | BPF_OR | BPF_K:
1239 		case BPF_ALU | BPF_XOR | BPF_K:
1240 		case BPF_ALU64 | BPF_ADD | BPF_K:
1241 		case BPF_ALU64 | BPF_SUB | BPF_K:
1242 		case BPF_ALU64 | BPF_AND | BPF_K:
1243 		case BPF_ALU64 | BPF_OR | BPF_K:
1244 		case BPF_ALU64 | BPF_XOR | BPF_K:
1245 			maybe_emit_1mod(&prog, dst_reg,
1246 					BPF_CLASS(insn->code) == BPF_ALU64);
1247 
1248 			/*
1249 			 * b3 holds 'normal' opcode, b2 short form only valid
1250 			 * in case dst is eax/rax.
1251 			 */
1252 			switch (BPF_OP(insn->code)) {
1253 			case BPF_ADD:
1254 				b3 = 0xC0;
1255 				b2 = 0x05;
1256 				break;
1257 			case BPF_SUB:
1258 				b3 = 0xE8;
1259 				b2 = 0x2D;
1260 				break;
1261 			case BPF_AND:
1262 				b3 = 0xE0;
1263 				b2 = 0x25;
1264 				break;
1265 			case BPF_OR:
1266 				b3 = 0xC8;
1267 				b2 = 0x0D;
1268 				break;
1269 			case BPF_XOR:
1270 				b3 = 0xF0;
1271 				b2 = 0x35;
1272 				break;
1273 			}
1274 
1275 			if (is_imm8(imm32))
1276 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1277 			else if (is_axreg(dst_reg))
1278 				EMIT1_off32(b2, imm32);
1279 			else
1280 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1281 			break;
1282 
1283 		case BPF_ALU64 | BPF_MOV | BPF_K:
1284 		case BPF_ALU | BPF_MOV | BPF_K:
1285 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1286 				       dst_reg, imm32);
1287 			break;
1288 
1289 		case BPF_LD | BPF_IMM | BPF_DW:
1290 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1291 			insn++;
1292 			i++;
1293 			break;
1294 
1295 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1296 		case BPF_ALU | BPF_MOD | BPF_X:
1297 		case BPF_ALU | BPF_DIV | BPF_X:
1298 		case BPF_ALU | BPF_MOD | BPF_K:
1299 		case BPF_ALU | BPF_DIV | BPF_K:
1300 		case BPF_ALU64 | BPF_MOD | BPF_X:
1301 		case BPF_ALU64 | BPF_DIV | BPF_X:
1302 		case BPF_ALU64 | BPF_MOD | BPF_K:
1303 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1304 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1305 
1306 			if (dst_reg != BPF_REG_0)
1307 				EMIT1(0x50); /* push rax */
1308 			if (dst_reg != BPF_REG_3)
1309 				EMIT1(0x52); /* push rdx */
1310 
1311 			if (BPF_SRC(insn->code) == BPF_X) {
1312 				if (src_reg == BPF_REG_0 ||
1313 				    src_reg == BPF_REG_3) {
1314 					/* mov r11, src_reg */
1315 					EMIT_mov(AUX_REG, src_reg);
1316 					src_reg = AUX_REG;
1317 				}
1318 			} else {
1319 				/* mov r11, imm32 */
1320 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1321 				src_reg = AUX_REG;
1322 			}
1323 
1324 			if (dst_reg != BPF_REG_0)
1325 				/* mov rax, dst_reg */
1326 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1327 
1328 			if (insn->off == 0) {
1329 				/*
1330 				 * xor edx, edx
1331 				 * equivalent to 'xor rdx, rdx', but one byte less
1332 				 */
1333 				EMIT2(0x31, 0xd2);
1334 
1335 				/* div src_reg */
1336 				maybe_emit_1mod(&prog, src_reg, is64);
1337 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1338 			} else {
1339 				if (BPF_CLASS(insn->code) == BPF_ALU)
1340 					EMIT1(0x99); /* cdq */
1341 				else
1342 					EMIT2(0x48, 0x99); /* cqo */
1343 
1344 				/* idiv src_reg */
1345 				maybe_emit_1mod(&prog, src_reg, is64);
1346 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1347 			}
1348 
1349 			if (BPF_OP(insn->code) == BPF_MOD &&
1350 			    dst_reg != BPF_REG_3)
1351 				/* mov dst_reg, rdx */
1352 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1353 			else if (BPF_OP(insn->code) == BPF_DIV &&
1354 				 dst_reg != BPF_REG_0)
1355 				/* mov dst_reg, rax */
1356 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1357 
1358 			if (dst_reg != BPF_REG_3)
1359 				EMIT1(0x5A); /* pop rdx */
1360 			if (dst_reg != BPF_REG_0)
1361 				EMIT1(0x58); /* pop rax */
1362 			break;
1363 		}
1364 
1365 		case BPF_ALU | BPF_MUL | BPF_K:
1366 		case BPF_ALU64 | BPF_MUL | BPF_K:
1367 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1368 				       BPF_CLASS(insn->code) == BPF_ALU64);
1369 
1370 			if (is_imm8(imm32))
1371 				/* imul dst_reg, dst_reg, imm8 */
1372 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1373 				      imm32);
1374 			else
1375 				/* imul dst_reg, dst_reg, imm32 */
1376 				EMIT2_off32(0x69,
1377 					    add_2reg(0xC0, dst_reg, dst_reg),
1378 					    imm32);
1379 			break;
1380 
1381 		case BPF_ALU | BPF_MUL | BPF_X:
1382 		case BPF_ALU64 | BPF_MUL | BPF_X:
1383 			maybe_emit_mod(&prog, src_reg, dst_reg,
1384 				       BPF_CLASS(insn->code) == BPF_ALU64);
1385 
1386 			/* imul dst_reg, src_reg */
1387 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1388 			break;
1389 
1390 			/* Shifts */
1391 		case BPF_ALU | BPF_LSH | BPF_K:
1392 		case BPF_ALU | BPF_RSH | BPF_K:
1393 		case BPF_ALU | BPF_ARSH | BPF_K:
1394 		case BPF_ALU64 | BPF_LSH | BPF_K:
1395 		case BPF_ALU64 | BPF_RSH | BPF_K:
1396 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1397 			maybe_emit_1mod(&prog, dst_reg,
1398 					BPF_CLASS(insn->code) == BPF_ALU64);
1399 
1400 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1401 			if (imm32 == 1)
1402 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1403 			else
1404 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1405 			break;
1406 
1407 		case BPF_ALU | BPF_LSH | BPF_X:
1408 		case BPF_ALU | BPF_RSH | BPF_X:
1409 		case BPF_ALU | BPF_ARSH | BPF_X:
1410 		case BPF_ALU64 | BPF_LSH | BPF_X:
1411 		case BPF_ALU64 | BPF_RSH | BPF_X:
1412 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1413 			/* BMI2 shifts aren't better when shift count is already in rcx */
1414 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1415 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1416 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1417 				u8 op;
1418 
1419 				switch (BPF_OP(insn->code)) {
1420 				case BPF_LSH:
1421 					op = 1; /* prefix 0x66 */
1422 					break;
1423 				case BPF_RSH:
1424 					op = 3; /* prefix 0xf2 */
1425 					break;
1426 				case BPF_ARSH:
1427 					op = 2; /* prefix 0xf3 */
1428 					break;
1429 				}
1430 
1431 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1432 
1433 				break;
1434 			}
1435 
1436 			if (src_reg != BPF_REG_4) { /* common case */
1437 				/* Check for bad case when dst_reg == rcx */
1438 				if (dst_reg == BPF_REG_4) {
1439 					/* mov r11, dst_reg */
1440 					EMIT_mov(AUX_REG, dst_reg);
1441 					dst_reg = AUX_REG;
1442 				} else {
1443 					EMIT1(0x51); /* push rcx */
1444 				}
1445 				/* mov rcx, src_reg */
1446 				EMIT_mov(BPF_REG_4, src_reg);
1447 			}
1448 
1449 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1450 			maybe_emit_1mod(&prog, dst_reg,
1451 					BPF_CLASS(insn->code) == BPF_ALU64);
1452 
1453 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1454 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1455 
1456 			if (src_reg != BPF_REG_4) {
1457 				if (insn->dst_reg == BPF_REG_4)
1458 					/* mov dst_reg, r11 */
1459 					EMIT_mov(insn->dst_reg, AUX_REG);
1460 				else
1461 					EMIT1(0x59); /* pop rcx */
1462 			}
1463 
1464 			break;
1465 
1466 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1467 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1468 			switch (imm32) {
1469 			case 16:
1470 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1471 				EMIT1(0x66);
1472 				if (is_ereg(dst_reg))
1473 					EMIT1(0x41);
1474 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1475 
1476 				/* Emit 'movzwl eax, ax' */
1477 				if (is_ereg(dst_reg))
1478 					EMIT3(0x45, 0x0F, 0xB7);
1479 				else
1480 					EMIT2(0x0F, 0xB7);
1481 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1482 				break;
1483 			case 32:
1484 				/* Emit 'bswap eax' to swap lower 4 bytes */
1485 				if (is_ereg(dst_reg))
1486 					EMIT2(0x41, 0x0F);
1487 				else
1488 					EMIT1(0x0F);
1489 				EMIT1(add_1reg(0xC8, dst_reg));
1490 				break;
1491 			case 64:
1492 				/* Emit 'bswap rax' to swap 8 bytes */
1493 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1494 				      add_1reg(0xC8, dst_reg));
1495 				break;
1496 			}
1497 			break;
1498 
1499 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1500 			switch (imm32) {
1501 			case 16:
1502 				/*
1503 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1504 				 * into 64 bit
1505 				 */
1506 				if (is_ereg(dst_reg))
1507 					EMIT3(0x45, 0x0F, 0xB7);
1508 				else
1509 					EMIT2(0x0F, 0xB7);
1510 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1511 				break;
1512 			case 32:
1513 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1514 				if (is_ereg(dst_reg))
1515 					EMIT1(0x45);
1516 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1517 				break;
1518 			case 64:
1519 				/* nop */
1520 				break;
1521 			}
1522 			break;
1523 
1524 			/* speculation barrier */
1525 		case BPF_ST | BPF_NOSPEC:
1526 			EMIT_LFENCE();
1527 			break;
1528 
1529 			/* ST: *(u8*)(dst_reg + off) = imm */
1530 		case BPF_ST | BPF_MEM | BPF_B:
1531 			if (is_ereg(dst_reg))
1532 				EMIT2(0x41, 0xC6);
1533 			else
1534 				EMIT1(0xC6);
1535 			goto st;
1536 		case BPF_ST | BPF_MEM | BPF_H:
1537 			if (is_ereg(dst_reg))
1538 				EMIT3(0x66, 0x41, 0xC7);
1539 			else
1540 				EMIT2(0x66, 0xC7);
1541 			goto st;
1542 		case BPF_ST | BPF_MEM | BPF_W:
1543 			if (is_ereg(dst_reg))
1544 				EMIT2(0x41, 0xC7);
1545 			else
1546 				EMIT1(0xC7);
1547 			goto st;
1548 		case BPF_ST | BPF_MEM | BPF_DW:
1549 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1550 
1551 st:			if (is_imm8(insn->off))
1552 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1553 			else
1554 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1555 
1556 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1557 			break;
1558 
1559 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1560 		case BPF_STX | BPF_MEM | BPF_B:
1561 		case BPF_STX | BPF_MEM | BPF_H:
1562 		case BPF_STX | BPF_MEM | BPF_W:
1563 		case BPF_STX | BPF_MEM | BPF_DW:
1564 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1565 			break;
1566 
1567 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1568 		case BPF_LDX | BPF_MEM | BPF_B:
1569 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1570 		case BPF_LDX | BPF_MEM | BPF_H:
1571 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1572 		case BPF_LDX | BPF_MEM | BPF_W:
1573 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1574 		case BPF_LDX | BPF_MEM | BPF_DW:
1575 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1576 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
1577 		case BPF_LDX | BPF_MEMSX | BPF_B:
1578 		case BPF_LDX | BPF_MEMSX | BPF_H:
1579 		case BPF_LDX | BPF_MEMSX | BPF_W:
1580 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1581 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1582 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1583 			insn_off = insn->off;
1584 
1585 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1586 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1587 				/* Conservatively check that src_reg + insn->off is a kernel address:
1588 				 *   src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
1589 				 * src_reg is used as scratch for src_reg += insn->off and restored
1590 				 * after emit_ldx if necessary
1591 				 */
1592 
1593 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
1594 				u8 *end_of_jmp;
1595 
1596 				/* At end of these emitted checks, insn->off will have been added
1597 				 * to src_reg, so no need to do relative load with insn->off offset
1598 				 */
1599 				insn_off = 0;
1600 
1601 				/* movabsq r11, limit */
1602 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1603 				EMIT((u32)limit, 4);
1604 				EMIT(limit >> 32, 4);
1605 
1606 				if (insn->off) {
1607 					/* add src_reg, insn->off */
1608 					maybe_emit_1mod(&prog, src_reg, true);
1609 					EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
1610 				}
1611 
1612 				/* cmp src_reg, r11 */
1613 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1614 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1615 
1616 				/* if unsigned '>=', goto load */
1617 				EMIT2(X86_JAE, 0);
1618 				end_of_jmp = prog;
1619 
1620 				/* xor dst_reg, dst_reg */
1621 				emit_mov_imm32(&prog, false, dst_reg, 0);
1622 				/* jmp byte_after_ldx */
1623 				EMIT2(0xEB, 0);
1624 
1625 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1626 				start_of_ldx = prog;
1627 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1628 			}
1629 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
1630 			    BPF_MODE(insn->code) == BPF_MEMSX)
1631 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1632 			else
1633 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1634 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1635 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1636 				struct exception_table_entry *ex;
1637 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1638 				s64 delta;
1639 
1640 				/* populate jmp_offset for JMP above */
1641 				start_of_ldx[-1] = prog - start_of_ldx;
1642 
1643 				if (insn->off && src_reg != dst_reg) {
1644 					/* sub src_reg, insn->off
1645 					 * Restore src_reg after "add src_reg, insn->off" in prev
1646 					 * if statement. But if src_reg == dst_reg, emit_ldx
1647 					 * above already clobbered src_reg, so no need to restore.
1648 					 * If add src_reg, insn->off was unnecessary, no need to
1649 					 * restore either.
1650 					 */
1651 					maybe_emit_1mod(&prog, src_reg, true);
1652 					EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
1653 				}
1654 
1655 				if (!bpf_prog->aux->extable)
1656 					break;
1657 
1658 				if (excnt >= bpf_prog->aux->num_exentries) {
1659 					pr_err("ex gen bug\n");
1660 					return -EFAULT;
1661 				}
1662 				ex = &bpf_prog->aux->extable[excnt++];
1663 
1664 				delta = _insn - (u8 *)&ex->insn;
1665 				if (!is_simm32(delta)) {
1666 					pr_err("extable->insn doesn't fit into 32-bit\n");
1667 					return -EFAULT;
1668 				}
1669 				/* switch ex to rw buffer for writes */
1670 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1671 
1672 				ex->insn = delta;
1673 
1674 				ex->data = EX_TYPE_BPF;
1675 
1676 				if (dst_reg > BPF_REG_9) {
1677 					pr_err("verifier error\n");
1678 					return -EFAULT;
1679 				}
1680 				/*
1681 				 * Compute size of x86 insn and its target dest x86 register.
1682 				 * ex_handler_bpf() will use lower 8 bits to adjust
1683 				 * pt_regs->ip to jump over this x86 instruction
1684 				 * and upper bits to figure out which pt_regs to zero out.
1685 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1686 				 * of 4 bytes will be ignored and rbx will be zero inited.
1687 				 */
1688 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1689 			}
1690 			break;
1691 
1692 		case BPF_STX | BPF_ATOMIC | BPF_W:
1693 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1694 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1695 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1696 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1697 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1698 				u32 real_src_reg = src_reg;
1699 				u32 real_dst_reg = dst_reg;
1700 				u8 *branch_target;
1701 
1702 				/*
1703 				 * Can't be implemented with a single x86 insn.
1704 				 * Need to do a CMPXCHG loop.
1705 				 */
1706 
1707 				/* Will need RAX as a CMPXCHG operand so save R0 */
1708 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1709 				if (src_reg == BPF_REG_0)
1710 					real_src_reg = BPF_REG_AX;
1711 				if (dst_reg == BPF_REG_0)
1712 					real_dst_reg = BPF_REG_AX;
1713 
1714 				branch_target = prog;
1715 				/* Load old value */
1716 				emit_ldx(&prog, BPF_SIZE(insn->code),
1717 					 BPF_REG_0, real_dst_reg, insn->off);
1718 				/*
1719 				 * Perform the (commutative) operation locally,
1720 				 * put the result in the AUX_REG.
1721 				 */
1722 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1723 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1724 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1725 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1726 				/* Attempt to swap in new value */
1727 				err = emit_atomic(&prog, BPF_CMPXCHG,
1728 						  real_dst_reg, AUX_REG,
1729 						  insn->off,
1730 						  BPF_SIZE(insn->code));
1731 				if (WARN_ON(err))
1732 					return err;
1733 				/*
1734 				 * ZF tells us whether we won the race. If it's
1735 				 * cleared we need to try again.
1736 				 */
1737 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1738 				/* Return the pre-modification value */
1739 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1740 				/* Restore R0 after clobbering RAX */
1741 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1742 				break;
1743 			}
1744 
1745 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1746 					  insn->off, BPF_SIZE(insn->code));
1747 			if (err)
1748 				return err;
1749 			break;
1750 
1751 			/* call */
1752 		case BPF_JMP | BPF_CALL: {
1753 			int offs;
1754 
1755 			func = (u8 *) __bpf_call_base + imm32;
1756 			if (tail_call_reachable) {
1757 				RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
1758 				if (!imm32)
1759 					return -EINVAL;
1760 				offs = 7 + x86_call_depth_emit_accounting(&prog, func);
1761 			} else {
1762 				if (!imm32)
1763 					return -EINVAL;
1764 				offs = x86_call_depth_emit_accounting(&prog, func);
1765 			}
1766 			if (emit_call(&prog, func, image + addrs[i - 1] + offs))
1767 				return -EINVAL;
1768 			break;
1769 		}
1770 
1771 		case BPF_JMP | BPF_TAIL_CALL:
1772 			if (imm32)
1773 				emit_bpf_tail_call_direct(bpf_prog,
1774 							  &bpf_prog->aux->poke_tab[imm32 - 1],
1775 							  &prog, image + addrs[i - 1],
1776 							  callee_regs_used,
1777 							  bpf_prog->aux->stack_depth,
1778 							  ctx);
1779 			else
1780 				emit_bpf_tail_call_indirect(bpf_prog,
1781 							    &prog,
1782 							    callee_regs_used,
1783 							    bpf_prog->aux->stack_depth,
1784 							    image + addrs[i - 1],
1785 							    ctx);
1786 			break;
1787 
1788 			/* cond jump */
1789 		case BPF_JMP | BPF_JEQ | BPF_X:
1790 		case BPF_JMP | BPF_JNE | BPF_X:
1791 		case BPF_JMP | BPF_JGT | BPF_X:
1792 		case BPF_JMP | BPF_JLT | BPF_X:
1793 		case BPF_JMP | BPF_JGE | BPF_X:
1794 		case BPF_JMP | BPF_JLE | BPF_X:
1795 		case BPF_JMP | BPF_JSGT | BPF_X:
1796 		case BPF_JMP | BPF_JSLT | BPF_X:
1797 		case BPF_JMP | BPF_JSGE | BPF_X:
1798 		case BPF_JMP | BPF_JSLE | BPF_X:
1799 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1800 		case BPF_JMP32 | BPF_JNE | BPF_X:
1801 		case BPF_JMP32 | BPF_JGT | BPF_X:
1802 		case BPF_JMP32 | BPF_JLT | BPF_X:
1803 		case BPF_JMP32 | BPF_JGE | BPF_X:
1804 		case BPF_JMP32 | BPF_JLE | BPF_X:
1805 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1806 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1807 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1808 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1809 			/* cmp dst_reg, src_reg */
1810 			maybe_emit_mod(&prog, dst_reg, src_reg,
1811 				       BPF_CLASS(insn->code) == BPF_JMP);
1812 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1813 			goto emit_cond_jmp;
1814 
1815 		case BPF_JMP | BPF_JSET | BPF_X:
1816 		case BPF_JMP32 | BPF_JSET | BPF_X:
1817 			/* test dst_reg, src_reg */
1818 			maybe_emit_mod(&prog, dst_reg, src_reg,
1819 				       BPF_CLASS(insn->code) == BPF_JMP);
1820 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1821 			goto emit_cond_jmp;
1822 
1823 		case BPF_JMP | BPF_JSET | BPF_K:
1824 		case BPF_JMP32 | BPF_JSET | BPF_K:
1825 			/* test dst_reg, imm32 */
1826 			maybe_emit_1mod(&prog, dst_reg,
1827 					BPF_CLASS(insn->code) == BPF_JMP);
1828 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1829 			goto emit_cond_jmp;
1830 
1831 		case BPF_JMP | BPF_JEQ | BPF_K:
1832 		case BPF_JMP | BPF_JNE | BPF_K:
1833 		case BPF_JMP | BPF_JGT | BPF_K:
1834 		case BPF_JMP | BPF_JLT | BPF_K:
1835 		case BPF_JMP | BPF_JGE | BPF_K:
1836 		case BPF_JMP | BPF_JLE | BPF_K:
1837 		case BPF_JMP | BPF_JSGT | BPF_K:
1838 		case BPF_JMP | BPF_JSLT | BPF_K:
1839 		case BPF_JMP | BPF_JSGE | BPF_K:
1840 		case BPF_JMP | BPF_JSLE | BPF_K:
1841 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1842 		case BPF_JMP32 | BPF_JNE | BPF_K:
1843 		case BPF_JMP32 | BPF_JGT | BPF_K:
1844 		case BPF_JMP32 | BPF_JLT | BPF_K:
1845 		case BPF_JMP32 | BPF_JGE | BPF_K:
1846 		case BPF_JMP32 | BPF_JLE | BPF_K:
1847 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1848 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1849 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1850 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1851 			/* test dst_reg, dst_reg to save one extra byte */
1852 			if (imm32 == 0) {
1853 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1854 					       BPF_CLASS(insn->code) == BPF_JMP);
1855 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1856 				goto emit_cond_jmp;
1857 			}
1858 
1859 			/* cmp dst_reg, imm8/32 */
1860 			maybe_emit_1mod(&prog, dst_reg,
1861 					BPF_CLASS(insn->code) == BPF_JMP);
1862 
1863 			if (is_imm8(imm32))
1864 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1865 			else
1866 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1867 
1868 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1869 			switch (BPF_OP(insn->code)) {
1870 			case BPF_JEQ:
1871 				jmp_cond = X86_JE;
1872 				break;
1873 			case BPF_JSET:
1874 			case BPF_JNE:
1875 				jmp_cond = X86_JNE;
1876 				break;
1877 			case BPF_JGT:
1878 				/* GT is unsigned '>', JA in x86 */
1879 				jmp_cond = X86_JA;
1880 				break;
1881 			case BPF_JLT:
1882 				/* LT is unsigned '<', JB in x86 */
1883 				jmp_cond = X86_JB;
1884 				break;
1885 			case BPF_JGE:
1886 				/* GE is unsigned '>=', JAE in x86 */
1887 				jmp_cond = X86_JAE;
1888 				break;
1889 			case BPF_JLE:
1890 				/* LE is unsigned '<=', JBE in x86 */
1891 				jmp_cond = X86_JBE;
1892 				break;
1893 			case BPF_JSGT:
1894 				/* Signed '>', GT in x86 */
1895 				jmp_cond = X86_JG;
1896 				break;
1897 			case BPF_JSLT:
1898 				/* Signed '<', LT in x86 */
1899 				jmp_cond = X86_JL;
1900 				break;
1901 			case BPF_JSGE:
1902 				/* Signed '>=', GE in x86 */
1903 				jmp_cond = X86_JGE;
1904 				break;
1905 			case BPF_JSLE:
1906 				/* Signed '<=', LE in x86 */
1907 				jmp_cond = X86_JLE;
1908 				break;
1909 			default: /* to silence GCC warning */
1910 				return -EFAULT;
1911 			}
1912 			jmp_offset = addrs[i + insn->off] - addrs[i];
1913 			if (is_imm8(jmp_offset)) {
1914 				if (jmp_padding) {
1915 					/* To keep the jmp_offset valid, the extra bytes are
1916 					 * padded before the jump insn, so we subtract the
1917 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1918 					 *
1919 					 * If the previous pass already emits an imm8
1920 					 * jmp_cond, then this BPF insn won't shrink, so
1921 					 * "nops" is 0.
1922 					 *
1923 					 * On the other hand, if the previous pass emits an
1924 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1925 					 * keep the image from shrinking further.
1926 					 *
1927 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1928 					 *     is 2 bytes, so the size difference is 4 bytes.
1929 					 */
1930 					nops = INSN_SZ_DIFF - 2;
1931 					if (nops != 0 && nops != 4) {
1932 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1933 						       nops);
1934 						return -EFAULT;
1935 					}
1936 					emit_nops(&prog, nops);
1937 				}
1938 				EMIT2(jmp_cond, jmp_offset);
1939 			} else if (is_simm32(jmp_offset)) {
1940 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1941 			} else {
1942 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1943 				return -EFAULT;
1944 			}
1945 
1946 			break;
1947 
1948 		case BPF_JMP | BPF_JA:
1949 		case BPF_JMP32 | BPF_JA:
1950 			if (BPF_CLASS(insn->code) == BPF_JMP) {
1951 				if (insn->off == -1)
1952 					/* -1 jmp instructions will always jump
1953 					 * backwards two bytes. Explicitly handling
1954 					 * this case avoids wasting too many passes
1955 					 * when there are long sequences of replaced
1956 					 * dead code.
1957 					 */
1958 					jmp_offset = -2;
1959 				else
1960 					jmp_offset = addrs[i + insn->off] - addrs[i];
1961 			} else {
1962 				if (insn->imm == -1)
1963 					jmp_offset = -2;
1964 				else
1965 					jmp_offset = addrs[i + insn->imm] - addrs[i];
1966 			}
1967 
1968 			if (!jmp_offset) {
1969 				/*
1970 				 * If jmp_padding is enabled, the extra nops will
1971 				 * be inserted. Otherwise, optimize out nop jumps.
1972 				 */
1973 				if (jmp_padding) {
1974 					/* There are 3 possible conditions.
1975 					 * (1) This BPF_JA is already optimized out in
1976 					 *     the previous run, so there is no need
1977 					 *     to pad any extra byte (0 byte).
1978 					 * (2) The previous pass emits an imm8 jmp,
1979 					 *     so we pad 2 bytes to match the previous
1980 					 *     insn size.
1981 					 * (3) Similarly, the previous pass emits an
1982 					 *     imm32 jmp, and 5 bytes is padded.
1983 					 */
1984 					nops = INSN_SZ_DIFF;
1985 					if (nops != 0 && nops != 2 && nops != 5) {
1986 						pr_err("unexpected nop jump padding: %d bytes\n",
1987 						       nops);
1988 						return -EFAULT;
1989 					}
1990 					emit_nops(&prog, nops);
1991 				}
1992 				break;
1993 			}
1994 emit_jmp:
1995 			if (is_imm8(jmp_offset)) {
1996 				if (jmp_padding) {
1997 					/* To avoid breaking jmp_offset, the extra bytes
1998 					 * are padded before the actual jmp insn, so
1999 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
2000 					 *
2001 					 * If the previous pass already emits an imm8
2002 					 * jmp, there is nothing to pad (0 byte).
2003 					 *
2004 					 * If it emits an imm32 jmp (5 bytes) previously
2005 					 * and now an imm8 jmp (2 bytes), then we pad
2006 					 * (5 - 2 = 3) bytes to stop the image from
2007 					 * shrinking further.
2008 					 */
2009 					nops = INSN_SZ_DIFF - 2;
2010 					if (nops != 0 && nops != 3) {
2011 						pr_err("unexpected jump padding: %d bytes\n",
2012 						       nops);
2013 						return -EFAULT;
2014 					}
2015 					emit_nops(&prog, INSN_SZ_DIFF - 2);
2016 				}
2017 				EMIT2(0xEB, jmp_offset);
2018 			} else if (is_simm32(jmp_offset)) {
2019 				EMIT1_off32(0xE9, jmp_offset);
2020 			} else {
2021 				pr_err("jmp gen bug %llx\n", jmp_offset);
2022 				return -EFAULT;
2023 			}
2024 			break;
2025 
2026 		case BPF_JMP | BPF_EXIT:
2027 			if (seen_exit) {
2028 				jmp_offset = ctx->cleanup_addr - addrs[i];
2029 				goto emit_jmp;
2030 			}
2031 			seen_exit = true;
2032 			/* Update cleanup_addr */
2033 			ctx->cleanup_addr = proglen;
2034 			if (bpf_prog->aux->exception_boundary) {
2035 				pop_callee_regs(&prog, all_callee_regs_used);
2036 				pop_r12(&prog);
2037 			} else {
2038 				pop_callee_regs(&prog, callee_regs_used);
2039 			}
2040 			EMIT1(0xC9);         /* leave */
2041 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
2042 			break;
2043 
2044 		default:
2045 			/*
2046 			 * By design x86-64 JIT should support all BPF instructions.
2047 			 * This error will be seen if new instruction was added
2048 			 * to the interpreter, but not to the JIT, or if there is
2049 			 * junk in bpf_prog.
2050 			 */
2051 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
2052 			return -EINVAL;
2053 		}
2054 
2055 		ilen = prog - temp;
2056 		if (ilen > BPF_MAX_INSN_SIZE) {
2057 			pr_err("bpf_jit: fatal insn size error\n");
2058 			return -EFAULT;
2059 		}
2060 
2061 		if (image) {
2062 			/*
2063 			 * When populating the image, assert that:
2064 			 *
2065 			 *  i) We do not write beyond the allocated space, and
2066 			 * ii) addrs[i] did not change from the prior run, in order
2067 			 *     to validate assumptions made for computing branch
2068 			 *     displacements.
2069 			 */
2070 			if (unlikely(proglen + ilen > oldproglen ||
2071 				     proglen + ilen != addrs[i])) {
2072 				pr_err("bpf_jit: fatal error\n");
2073 				return -EFAULT;
2074 			}
2075 			memcpy(rw_image + proglen, temp, ilen);
2076 		}
2077 		proglen += ilen;
2078 		addrs[i] = proglen;
2079 		prog = temp;
2080 	}
2081 
2082 	if (image && excnt != bpf_prog->aux->num_exentries) {
2083 		pr_err("extable is not populated\n");
2084 		return -EFAULT;
2085 	}
2086 	return proglen;
2087 }
2088 
2089 static void clean_stack_garbage(const struct btf_func_model *m,
2090 				u8 **pprog, int nr_stack_slots,
2091 				int stack_size)
2092 {
2093 	int arg_size, off;
2094 	u8 *prog;
2095 
2096 	/* Generally speaking, the compiler will pass the arguments
2097 	 * on-stack with "push" instruction, which will take 8-byte
2098 	 * on the stack. In this case, there won't be garbage values
2099 	 * while we copy the arguments from origin stack frame to current
2100 	 * in BPF_DW.
2101 	 *
2102 	 * However, sometimes the compiler will only allocate 4-byte on
2103 	 * the stack for the arguments. For now, this case will only
2104 	 * happen if there is only one argument on-stack and its size
2105 	 * not more than 4 byte. In this case, there will be garbage
2106 	 * values on the upper 4-byte where we store the argument on
2107 	 * current stack frame.
2108 	 *
2109 	 * arguments on origin stack:
2110 	 *
2111 	 * stack_arg_1(4-byte) xxx(4-byte)
2112 	 *
2113 	 * what we copy:
2114 	 *
2115 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
2116 	 *
2117 	 * and the xxx is the garbage values which we should clean here.
2118 	 */
2119 	if (nr_stack_slots != 1)
2120 		return;
2121 
2122 	/* the size of the last argument */
2123 	arg_size = m->arg_size[m->nr_args - 1];
2124 	if (arg_size <= 4) {
2125 		off = -(stack_size - 4);
2126 		prog = *pprog;
2127 		/* mov DWORD PTR [rbp + off], 0 */
2128 		if (!is_imm8(off))
2129 			EMIT2_off32(0xC7, 0x85, off);
2130 		else
2131 			EMIT3(0xC7, 0x45, off);
2132 		EMIT(0, 4);
2133 		*pprog = prog;
2134 	}
2135 }
2136 
2137 /* get the count of the regs that are used to pass arguments */
2138 static int get_nr_used_regs(const struct btf_func_model *m)
2139 {
2140 	int i, arg_regs, nr_used_regs = 0;
2141 
2142 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2143 		arg_regs = (m->arg_size[i] + 7) / 8;
2144 		if (nr_used_regs + arg_regs <= 6)
2145 			nr_used_regs += arg_regs;
2146 
2147 		if (nr_used_regs >= 6)
2148 			break;
2149 	}
2150 
2151 	return nr_used_regs;
2152 }
2153 
2154 static void save_args(const struct btf_func_model *m, u8 **prog,
2155 		      int stack_size, bool for_call_origin)
2156 {
2157 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2158 	int i, j;
2159 
2160 	/* Store function arguments to stack.
2161 	 * For a function that accepts two pointers the sequence will be:
2162 	 * mov QWORD PTR [rbp-0x10],rdi
2163 	 * mov QWORD PTR [rbp-0x8],rsi
2164 	 */
2165 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2166 		arg_regs = (m->arg_size[i] + 7) / 8;
2167 
2168 		/* According to the research of Yonghong, struct members
2169 		 * should be all in register or all on the stack.
2170 		 * Meanwhile, the compiler will pass the argument on regs
2171 		 * if the remaining regs can hold the argument.
2172 		 *
2173 		 * Disorder of the args can happen. For example:
2174 		 *
2175 		 * struct foo_struct {
2176 		 *     long a;
2177 		 *     int b;
2178 		 * };
2179 		 * int foo(char, char, char, char, char, struct foo_struct,
2180 		 *         char);
2181 		 *
2182 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2183 		 * by stack.
2184 		 */
2185 		if (nr_regs + arg_regs > 6) {
2186 			/* copy function arguments from origin stack frame
2187 			 * into current stack frame.
2188 			 *
2189 			 * The starting address of the arguments on-stack
2190 			 * is:
2191 			 *   rbp + 8(push rbp) +
2192 			 *   8(return addr of origin call) +
2193 			 *   8(return addr of the caller)
2194 			 * which means: rbp + 24
2195 			 */
2196 			for (j = 0; j < arg_regs; j++) {
2197 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2198 					 nr_stack_slots * 8 + 0x18);
2199 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2200 					 -stack_size);
2201 
2202 				if (!nr_stack_slots)
2203 					first_off = stack_size;
2204 				stack_size -= 8;
2205 				nr_stack_slots++;
2206 			}
2207 		} else {
2208 			/* Only copy the arguments on-stack to current
2209 			 * 'stack_size' and ignore the regs, used to
2210 			 * prepare the arguments on-stack for origin call.
2211 			 */
2212 			if (for_call_origin) {
2213 				nr_regs += arg_regs;
2214 				continue;
2215 			}
2216 
2217 			/* copy the arguments from regs into stack */
2218 			for (j = 0; j < arg_regs; j++) {
2219 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2220 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2221 					 -stack_size);
2222 				stack_size -= 8;
2223 				nr_regs++;
2224 			}
2225 		}
2226 	}
2227 
2228 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2229 }
2230 
2231 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2232 			 int stack_size)
2233 {
2234 	int i, j, arg_regs, nr_regs = 0;
2235 
2236 	/* Restore function arguments from stack.
2237 	 * For a function that accepts two pointers the sequence will be:
2238 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2239 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2240 	 *
2241 	 * The logic here is similar to what we do in save_args()
2242 	 */
2243 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2244 		arg_regs = (m->arg_size[i] + 7) / 8;
2245 		if (nr_regs + arg_regs <= 6) {
2246 			for (j = 0; j < arg_regs; j++) {
2247 				emit_ldx(prog, BPF_DW,
2248 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2249 					 BPF_REG_FP,
2250 					 -stack_size);
2251 				stack_size -= 8;
2252 				nr_regs++;
2253 			}
2254 		} else {
2255 			stack_size -= 8 * arg_regs;
2256 		}
2257 
2258 		if (nr_regs >= 6)
2259 			break;
2260 	}
2261 }
2262 
2263 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2264 			   struct bpf_tramp_link *l, int stack_size,
2265 			   int run_ctx_off, bool save_ret,
2266 			   void *image, void *rw_image)
2267 {
2268 	u8 *prog = *pprog;
2269 	u8 *jmp_insn;
2270 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2271 	struct bpf_prog *p = l->link.prog;
2272 	u64 cookie = l->cookie;
2273 
2274 	/* mov rdi, cookie */
2275 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2276 
2277 	/* Prepare struct bpf_tramp_run_ctx.
2278 	 *
2279 	 * bpf_tramp_run_ctx is already preserved by
2280 	 * arch_prepare_bpf_trampoline().
2281 	 *
2282 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2283 	 */
2284 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2285 
2286 	/* arg1: mov rdi, progs[i] */
2287 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2288 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2289 	if (!is_imm8(-run_ctx_off))
2290 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
2291 	else
2292 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
2293 
2294 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image)))
2295 		return -EINVAL;
2296 	/* remember prog start time returned by __bpf_prog_enter */
2297 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
2298 
2299 	/* if (__bpf_prog_enter*(prog) == 0)
2300 	 *	goto skip_exec_of_prog;
2301 	 */
2302 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
2303 	/* emit 2 nops that will be replaced with JE insn */
2304 	jmp_insn = prog;
2305 	emit_nops(&prog, 2);
2306 
2307 	/* arg1: lea rdi, [rbp - stack_size] */
2308 	if (!is_imm8(-stack_size))
2309 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
2310 	else
2311 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
2312 	/* arg2: progs[i]->insnsi for interpreter */
2313 	if (!p->jited)
2314 		emit_mov_imm64(&prog, BPF_REG_2,
2315 			       (long) p->insnsi >> 32,
2316 			       (u32) (long) p->insnsi);
2317 	/* call JITed bpf program or interpreter */
2318 	if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image)))
2319 		return -EINVAL;
2320 
2321 	/*
2322 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
2323 	 * of the previous call which is then passed on the stack to
2324 	 * the next BPF program.
2325 	 *
2326 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
2327 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
2328 	 */
2329 	if (save_ret)
2330 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2331 
2332 	/* replace 2 nops with JE insn, since jmp target is known */
2333 	jmp_insn[0] = X86_JE;
2334 	jmp_insn[1] = prog - jmp_insn - 2;
2335 
2336 	/* arg1: mov rdi, progs[i] */
2337 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2338 	/* arg2: mov rsi, rbx <- start time in nsec */
2339 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
2340 	/* arg3: lea rdx, [rbp - run_ctx_off] */
2341 	if (!is_imm8(-run_ctx_off))
2342 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
2343 	else
2344 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
2345 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
2346 		return -EINVAL;
2347 
2348 	*pprog = prog;
2349 	return 0;
2350 }
2351 
2352 static void emit_align(u8 **pprog, u32 align)
2353 {
2354 	u8 *target, *prog = *pprog;
2355 
2356 	target = PTR_ALIGN(prog, align);
2357 	if (target != prog)
2358 		emit_nops(&prog, target - prog);
2359 
2360 	*pprog = prog;
2361 }
2362 
2363 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2364 {
2365 	u8 *prog = *pprog;
2366 	s64 offset;
2367 
2368 	offset = func - (ip + 2 + 4);
2369 	if (!is_simm32(offset)) {
2370 		pr_err("Target %p is out of range\n", func);
2371 		return -EINVAL;
2372 	}
2373 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2374 	*pprog = prog;
2375 	return 0;
2376 }
2377 
2378 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2379 		      struct bpf_tramp_links *tl, int stack_size,
2380 		      int run_ctx_off, bool save_ret,
2381 		      void *image, void *rw_image)
2382 {
2383 	int i;
2384 	u8 *prog = *pprog;
2385 
2386 	for (i = 0; i < tl->nr_links; i++) {
2387 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2388 				    run_ctx_off, save_ret, image, rw_image))
2389 			return -EINVAL;
2390 	}
2391 	*pprog = prog;
2392 	return 0;
2393 }
2394 
2395 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2396 			      struct bpf_tramp_links *tl, int stack_size,
2397 			      int run_ctx_off, u8 **branches,
2398 			      void *image, void *rw_image)
2399 {
2400 	u8 *prog = *pprog;
2401 	int i;
2402 
2403 	/* The first fmod_ret program will receive a garbage return value.
2404 	 * Set this to 0 to avoid confusing the program.
2405 	 */
2406 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2407 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2408 	for (i = 0; i < tl->nr_links; i++) {
2409 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true,
2410 				    image, rw_image))
2411 			return -EINVAL;
2412 
2413 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2414 		 * if (*(u64 *)(rbp - 8) !=  0)
2415 		 *	goto do_fexit;
2416 		 */
2417 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2418 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2419 
2420 		/* Save the location of the branch and Generate 6 nops
2421 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2422 		 * are replaced with a conditional jump once do_fexit (i.e. the
2423 		 * start of the fexit invocation) is finalized.
2424 		 */
2425 		branches[i] = prog;
2426 		emit_nops(&prog, 4 + 2);
2427 	}
2428 
2429 	*pprog = prog;
2430 	return 0;
2431 }
2432 
2433 /* Example:
2434  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2435  * its 'struct btf_func_model' will be nr_args=2
2436  * The assembly code when eth_type_trans is executing after trampoline:
2437  *
2438  * push rbp
2439  * mov rbp, rsp
2440  * sub rsp, 16                     // space for skb and dev
2441  * push rbx                        // temp regs to pass start time
2442  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2443  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2444  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2445  * mov rbx, rax                    // remember start time in bpf stats are enabled
2446  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2447  * call addr_of_jited_FENTRY_prog
2448  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2449  * mov rsi, rbx                    // prog start time
2450  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2451  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2452  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2453  * pop rbx
2454  * leave
2455  * ret
2456  *
2457  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2458  * replaced with 'call generated_bpf_trampoline'. When it returns
2459  * eth_type_trans will continue executing with original skb and dev pointers.
2460  *
2461  * The assembly code when eth_type_trans is called from trampoline:
2462  *
2463  * push rbp
2464  * mov rbp, rsp
2465  * sub rsp, 24                     // space for skb, dev, return value
2466  * push rbx                        // temp regs to pass start time
2467  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2468  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2469  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2470  * mov rbx, rax                    // remember start time if bpf stats are enabled
2471  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2472  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2473  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2474  * mov rsi, rbx                    // prog start time
2475  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2476  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2477  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2478  * call eth_type_trans+5           // execute body of eth_type_trans
2479  * mov qword ptr [rbp - 8], rax    // save return value
2480  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2481  * mov rbx, rax                    // remember start time in bpf stats are enabled
2482  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2483  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2484  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2485  * mov rsi, rbx                    // prog start time
2486  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2487  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2488  * pop rbx
2489  * leave
2490  * add rsp, 8                      // skip eth_type_trans's frame
2491  * ret                             // return to its caller
2492  */
2493 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
2494 					 void *rw_image_end, void *image,
2495 					 const struct btf_func_model *m, u32 flags,
2496 					 struct bpf_tramp_links *tlinks,
2497 					 void *func_addr)
2498 {
2499 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2500 	int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
2501 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2502 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2503 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2504 	void *orig_call = func_addr;
2505 	u8 **branches = NULL;
2506 	u8 *prog;
2507 	bool save_ret;
2508 
2509 	/*
2510 	 * F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is
2511 	 * explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG
2512 	 * because @func_addr.
2513 	 */
2514 	WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) &&
2515 		     (flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET)));
2516 
2517 	/* extra registers for struct arguments */
2518 	for (i = 0; i < m->nr_args; i++) {
2519 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2520 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2521 	}
2522 
2523 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
2524 	 * are passed through regs, the remains are through stack.
2525 	 */
2526 	if (nr_regs > MAX_BPF_FUNC_ARGS)
2527 		return -ENOTSUPP;
2528 
2529 	/* Generated trampoline stack layout:
2530 	 *
2531 	 * RBP + 8         [ return address  ]
2532 	 * RBP + 0         [ RBP             ]
2533 	 *
2534 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2535 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2536 	 *
2537 	 *                 [ reg_argN        ]  always
2538 	 *                 [ ...             ]
2539 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2540 	 *
2541 	 * RBP - nregs_off [ regs count	     ]  always
2542 	 *
2543 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2544 	 *
2545 	 * RBP - rbx_off   [ rbx value       ]  always
2546 	 *
2547 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2548 	 *
2549 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
2550 	 *                     [ ...        ]
2551 	 *                     [ stack_arg2 ]
2552 	 * RBP - arg_stack_off [ stack_arg1 ]
2553 	 * RSP                 [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
2554 	 */
2555 
2556 	/* room for return value of orig_call or fentry prog */
2557 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2558 	if (save_ret)
2559 		stack_size += 8;
2560 
2561 	stack_size += nr_regs * 8;
2562 	regs_off = stack_size;
2563 
2564 	/* regs count  */
2565 	stack_size += 8;
2566 	nregs_off = stack_size;
2567 
2568 	if (flags & BPF_TRAMP_F_IP_ARG)
2569 		stack_size += 8; /* room for IP address argument */
2570 
2571 	ip_off = stack_size;
2572 
2573 	stack_size += 8;
2574 	rbx_off = stack_size;
2575 
2576 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2577 	run_ctx_off = stack_size;
2578 
2579 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
2580 		/* the space that used to pass arguments on-stack */
2581 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
2582 		/* make sure the stack pointer is 16-byte aligned if we
2583 		 * need pass arguments on stack, which means
2584 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
2585 		 * should be 16-byte aligned. Following code depend on
2586 		 * that stack_size is already 8-byte aligned.
2587 		 */
2588 		stack_size += (stack_size % 16) ? 0 : 8;
2589 	}
2590 
2591 	arg_stack_off = stack_size;
2592 
2593 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2594 		/* skip patched call instruction and point orig_call to actual
2595 		 * body of the kernel function.
2596 		 */
2597 		if (is_endbr(*(u32 *)orig_call))
2598 			orig_call += ENDBR_INSN_SIZE;
2599 		orig_call += X86_PATCH_SIZE;
2600 	}
2601 
2602 	prog = rw_image;
2603 
2604 	if (flags & BPF_TRAMP_F_INDIRECT) {
2605 		/*
2606 		 * Indirect call for bpf_struct_ops
2607 		 */
2608 		emit_cfi(&prog, cfi_get_func_hash(func_addr));
2609 	} else {
2610 		/*
2611 		 * Direct-call fentry stub, as such it needs accounting for the
2612 		 * __fentry__ call.
2613 		 */
2614 		x86_call_depth_emit_accounting(&prog, NULL);
2615 	}
2616 	EMIT1(0x55);		 /* push rbp */
2617 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2618 	if (!is_imm8(stack_size)) {
2619 		/* sub rsp, stack_size */
2620 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
2621 	} else {
2622 		/* sub rsp, stack_size */
2623 		EMIT4(0x48, 0x83, 0xEC, stack_size);
2624 	}
2625 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2626 		EMIT1(0x50);		/* push rax */
2627 	/* mov QWORD PTR [rbp - rbx_off], rbx */
2628 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
2629 
2630 	/* Store number of argument registers of the traced function:
2631 	 *   mov rax, nr_regs
2632 	 *   mov QWORD PTR [rbp - nregs_off], rax
2633 	 */
2634 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2635 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2636 
2637 	if (flags & BPF_TRAMP_F_IP_ARG) {
2638 		/* Store IP address of the traced function:
2639 		 * movabsq rax, func_addr
2640 		 * mov QWORD PTR [rbp - ip_off], rax
2641 		 */
2642 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2643 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2644 	}
2645 
2646 	save_args(m, &prog, regs_off, false);
2647 
2648 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2649 		/* arg1: mov rdi, im */
2650 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2651 		if (emit_rsb_call(&prog, __bpf_tramp_enter,
2652 				  image + (prog - (u8 *)rw_image))) {
2653 			ret = -EINVAL;
2654 			goto cleanup;
2655 		}
2656 	}
2657 
2658 	if (fentry->nr_links) {
2659 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2660 			       flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image))
2661 			return -EINVAL;
2662 	}
2663 
2664 	if (fmod_ret->nr_links) {
2665 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2666 				   GFP_KERNEL);
2667 		if (!branches)
2668 			return -ENOMEM;
2669 
2670 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2671 				       run_ctx_off, branches, image, rw_image)) {
2672 			ret = -EINVAL;
2673 			goto cleanup;
2674 		}
2675 	}
2676 
2677 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2678 		restore_regs(m, &prog, regs_off);
2679 		save_args(m, &prog, arg_stack_off, true);
2680 
2681 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
2682 			/* Before calling the original function, restore the
2683 			 * tail_call_cnt from stack to rax.
2684 			 */
2685 			RESTORE_TAIL_CALL_CNT(stack_size);
2686 		}
2687 
2688 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2689 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
2690 			EMIT2(0xff, 0xd3); /* call *rbx */
2691 		} else {
2692 			/* call original function */
2693 			if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) {
2694 				ret = -EINVAL;
2695 				goto cleanup;
2696 			}
2697 		}
2698 		/* remember return value in a stack for bpf prog to access */
2699 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2700 		im->ip_after_call = image + (prog - (u8 *)rw_image);
2701 		emit_nops(&prog, X86_PATCH_SIZE);
2702 	}
2703 
2704 	if (fmod_ret->nr_links) {
2705 		/* From Intel 64 and IA-32 Architectures Optimization
2706 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2707 		 * Coding Rule 11: All branch targets should be 16-byte
2708 		 * aligned.
2709 		 */
2710 		emit_align(&prog, 16);
2711 		/* Update the branches saved in invoke_bpf_mod_ret with the
2712 		 * aligned address of do_fexit.
2713 		 */
2714 		for (i = 0; i < fmod_ret->nr_links; i++) {
2715 			emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image),
2716 					    image + (branches[i] - (u8 *)rw_image), X86_JNE);
2717 		}
2718 	}
2719 
2720 	if (fexit->nr_links) {
2721 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off,
2722 			       false, image, rw_image)) {
2723 			ret = -EINVAL;
2724 			goto cleanup;
2725 		}
2726 	}
2727 
2728 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2729 		restore_regs(m, &prog, regs_off);
2730 
2731 	/* This needs to be done regardless. If there were fmod_ret programs,
2732 	 * the return value is only updated on the stack and still needs to be
2733 	 * restored to R0.
2734 	 */
2735 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2736 		im->ip_epilogue = image + (prog - (u8 *)rw_image);
2737 		/* arg1: mov rdi, im */
2738 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2739 		if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) {
2740 			ret = -EINVAL;
2741 			goto cleanup;
2742 		}
2743 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
2744 		/* Before running the original function, restore the
2745 		 * tail_call_cnt from stack to rax.
2746 		 */
2747 		RESTORE_TAIL_CALL_CNT(stack_size);
2748 	}
2749 
2750 	/* restore return value of orig_call or fentry prog back into RAX */
2751 	if (save_ret)
2752 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2753 
2754 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
2755 	EMIT1(0xC9); /* leave */
2756 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2757 		/* skip our return address and return to parent */
2758 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2759 	}
2760 	emit_return(&prog, image + (prog - (u8 *)rw_image));
2761 	/* Make sure the trampoline generation logic doesn't overflow */
2762 	if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) {
2763 		ret = -EFAULT;
2764 		goto cleanup;
2765 	}
2766 	ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY;
2767 
2768 cleanup:
2769 	kfree(branches);
2770 	return ret;
2771 }
2772 
2773 void *arch_alloc_bpf_trampoline(unsigned int size)
2774 {
2775 	return bpf_prog_pack_alloc(size, jit_fill_hole);
2776 }
2777 
2778 void arch_free_bpf_trampoline(void *image, unsigned int size)
2779 {
2780 	bpf_prog_pack_free(image, size);
2781 }
2782 
2783 void arch_protect_bpf_trampoline(void *image, unsigned int size)
2784 {
2785 }
2786 
2787 void arch_unprotect_bpf_trampoline(void *image, unsigned int size)
2788 {
2789 }
2790 
2791 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2792 				const struct btf_func_model *m, u32 flags,
2793 				struct bpf_tramp_links *tlinks,
2794 				void *func_addr)
2795 {
2796 	void *rw_image, *tmp;
2797 	int ret;
2798 	u32 size = image_end - image;
2799 
2800 	/* rw_image doesn't need to be in module memory range, so we can
2801 	 * use kvmalloc.
2802 	 */
2803 	rw_image = kvmalloc(size, GFP_KERNEL);
2804 	if (!rw_image)
2805 		return -ENOMEM;
2806 
2807 	ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
2808 					    flags, tlinks, func_addr);
2809 	if (ret < 0)
2810 		goto out;
2811 
2812 	tmp = bpf_arch_text_copy(image, rw_image, size);
2813 	if (IS_ERR(tmp))
2814 		ret = PTR_ERR(tmp);
2815 out:
2816 	kvfree(rw_image);
2817 	return ret;
2818 }
2819 
2820 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
2821 			     struct bpf_tramp_links *tlinks, void *func_addr)
2822 {
2823 	struct bpf_tramp_image im;
2824 	void *image;
2825 	int ret;
2826 
2827 	/* Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
2828 	 * This will NOT cause fragmentation in direct map, as we do not
2829 	 * call set_memory_*() on this buffer.
2830 	 *
2831 	 * We cannot use kvmalloc here, because we need image to be in
2832 	 * module memory range.
2833 	 */
2834 	image = bpf_jit_alloc_exec(PAGE_SIZE);
2835 	if (!image)
2836 		return -ENOMEM;
2837 
2838 	ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
2839 					    m, flags, tlinks, func_addr);
2840 	bpf_jit_free_exec(image);
2841 	return ret;
2842 }
2843 
2844 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2845 {
2846 	u8 *jg_reloc, *prog = *pprog;
2847 	int pivot, err, jg_bytes = 1;
2848 	s64 jg_offset;
2849 
2850 	if (a == b) {
2851 		/* Leaf node of recursion, i.e. not a range of indices
2852 		 * anymore.
2853 		 */
2854 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2855 		if (!is_simm32(progs[a]))
2856 			return -1;
2857 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2858 			    progs[a]);
2859 		err = emit_cond_near_jump(&prog,	/* je func */
2860 					  (void *)progs[a], image + (prog - buf),
2861 					  X86_JE);
2862 		if (err)
2863 			return err;
2864 
2865 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2866 
2867 		*pprog = prog;
2868 		return 0;
2869 	}
2870 
2871 	/* Not a leaf node, so we pivot, and recursively descend into
2872 	 * the lower and upper ranges.
2873 	 */
2874 	pivot = (b - a) / 2;
2875 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2876 	if (!is_simm32(progs[a + pivot]))
2877 		return -1;
2878 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2879 
2880 	if (pivot > 2) {				/* jg upper_part */
2881 		/* Require near jump. */
2882 		jg_bytes = 4;
2883 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2884 	} else {
2885 		EMIT2(X86_JG, 0);
2886 	}
2887 	jg_reloc = prog;
2888 
2889 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2890 				  progs, image, buf);
2891 	if (err)
2892 		return err;
2893 
2894 	/* From Intel 64 and IA-32 Architectures Optimization
2895 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2896 	 * Coding Rule 11: All branch targets should be 16-byte
2897 	 * aligned.
2898 	 */
2899 	emit_align(&prog, 16);
2900 	jg_offset = prog - jg_reloc;
2901 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2902 
2903 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2904 				  b, progs, image, buf);
2905 	if (err)
2906 		return err;
2907 
2908 	*pprog = prog;
2909 	return 0;
2910 }
2911 
2912 static int cmp_ips(const void *a, const void *b)
2913 {
2914 	const s64 *ipa = a;
2915 	const s64 *ipb = b;
2916 
2917 	if (*ipa > *ipb)
2918 		return 1;
2919 	if (*ipa < *ipb)
2920 		return -1;
2921 	return 0;
2922 }
2923 
2924 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2925 {
2926 	u8 *prog = buf;
2927 
2928 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2929 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2930 }
2931 
2932 struct x64_jit_data {
2933 	struct bpf_binary_header *rw_header;
2934 	struct bpf_binary_header *header;
2935 	int *addrs;
2936 	u8 *image;
2937 	int proglen;
2938 	struct jit_context ctx;
2939 };
2940 
2941 #define MAX_PASSES 20
2942 #define PADDING_PASSES (MAX_PASSES - 5)
2943 
2944 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2945 {
2946 	struct bpf_binary_header *rw_header = NULL;
2947 	struct bpf_binary_header *header = NULL;
2948 	struct bpf_prog *tmp, *orig_prog = prog;
2949 	struct x64_jit_data *jit_data;
2950 	int proglen, oldproglen = 0;
2951 	struct jit_context ctx = {};
2952 	bool tmp_blinded = false;
2953 	bool extra_pass = false;
2954 	bool padding = false;
2955 	u8 *rw_image = NULL;
2956 	u8 *image = NULL;
2957 	int *addrs;
2958 	int pass;
2959 	int i;
2960 
2961 	if (!prog->jit_requested)
2962 		return orig_prog;
2963 
2964 	tmp = bpf_jit_blind_constants(prog);
2965 	/*
2966 	 * If blinding was requested and we failed during blinding,
2967 	 * we must fall back to the interpreter.
2968 	 */
2969 	if (IS_ERR(tmp))
2970 		return orig_prog;
2971 	if (tmp != prog) {
2972 		tmp_blinded = true;
2973 		prog = tmp;
2974 	}
2975 
2976 	jit_data = prog->aux->jit_data;
2977 	if (!jit_data) {
2978 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2979 		if (!jit_data) {
2980 			prog = orig_prog;
2981 			goto out;
2982 		}
2983 		prog->aux->jit_data = jit_data;
2984 	}
2985 	addrs = jit_data->addrs;
2986 	if (addrs) {
2987 		ctx = jit_data->ctx;
2988 		oldproglen = jit_data->proglen;
2989 		image = jit_data->image;
2990 		header = jit_data->header;
2991 		rw_header = jit_data->rw_header;
2992 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2993 		extra_pass = true;
2994 		padding = true;
2995 		goto skip_init_addrs;
2996 	}
2997 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2998 	if (!addrs) {
2999 		prog = orig_prog;
3000 		goto out_addrs;
3001 	}
3002 
3003 	/*
3004 	 * Before first pass, make a rough estimation of addrs[]
3005 	 * each BPF instruction is translated to less than 64 bytes
3006 	 */
3007 	for (proglen = 0, i = 0; i <= prog->len; i++) {
3008 		proglen += 64;
3009 		addrs[i] = proglen;
3010 	}
3011 	ctx.cleanup_addr = proglen;
3012 skip_init_addrs:
3013 
3014 	/*
3015 	 * JITed image shrinks with every pass and the loop iterates
3016 	 * until the image stops shrinking. Very large BPF programs
3017 	 * may converge on the last pass. In such case do one more
3018 	 * pass to emit the final image.
3019 	 */
3020 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
3021 		if (!padding && pass >= PADDING_PASSES)
3022 			padding = true;
3023 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
3024 		if (proglen <= 0) {
3025 out_image:
3026 			image = NULL;
3027 			if (header) {
3028 				bpf_arch_text_copy(&header->size, &rw_header->size,
3029 						   sizeof(rw_header->size));
3030 				bpf_jit_binary_pack_free(header, rw_header);
3031 			}
3032 			/* Fall back to interpreter mode */
3033 			prog = orig_prog;
3034 			if (extra_pass) {
3035 				prog->bpf_func = NULL;
3036 				prog->jited = 0;
3037 				prog->jited_len = 0;
3038 			}
3039 			goto out_addrs;
3040 		}
3041 		if (image) {
3042 			if (proglen != oldproglen) {
3043 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
3044 				       proglen, oldproglen);
3045 				goto out_image;
3046 			}
3047 			break;
3048 		}
3049 		if (proglen == oldproglen) {
3050 			/*
3051 			 * The number of entries in extable is the number of BPF_LDX
3052 			 * insns that access kernel memory via "pointer to BTF type".
3053 			 * The verifier changed their opcode from LDX|MEM|size
3054 			 * to LDX|PROBE_MEM|size to make JITing easier.
3055 			 */
3056 			u32 align = __alignof__(struct exception_table_entry);
3057 			u32 extable_size = prog->aux->num_exentries *
3058 				sizeof(struct exception_table_entry);
3059 
3060 			/* allocate module memory for x86 insns and extable */
3061 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
3062 							   &image, align, &rw_header, &rw_image,
3063 							   jit_fill_hole);
3064 			if (!header) {
3065 				prog = orig_prog;
3066 				goto out_addrs;
3067 			}
3068 			prog->aux->extable = (void *) image + roundup(proglen, align);
3069 		}
3070 		oldproglen = proglen;
3071 		cond_resched();
3072 	}
3073 
3074 	if (bpf_jit_enable > 1)
3075 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
3076 
3077 	if (image) {
3078 		if (!prog->is_func || extra_pass) {
3079 			/*
3080 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
3081 			 *   1) header is not pointing to proper module memory;
3082 			 *   2) the arch doesn't support bpf_arch_text_copy().
3083 			 *
3084 			 * Both cases are serious bugs and justify WARN_ON.
3085 			 */
3086 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
3087 				/* header has been freed */
3088 				header = NULL;
3089 				goto out_image;
3090 			}
3091 
3092 			bpf_tail_call_direct_fixup(prog);
3093 		} else {
3094 			jit_data->addrs = addrs;
3095 			jit_data->ctx = ctx;
3096 			jit_data->proglen = proglen;
3097 			jit_data->image = image;
3098 			jit_data->header = header;
3099 			jit_data->rw_header = rw_header;
3100 		}
3101 		/*
3102 		 * ctx.prog_offset is used when CFI preambles put code *before*
3103 		 * the function. See emit_cfi(). For FineIBT specifically this code
3104 		 * can also be executed and bpf_prog_kallsyms_add() will
3105 		 * generate an additional symbol to cover this, hence also
3106 		 * decrement proglen.
3107 		 */
3108 		prog->bpf_func = (void *)image + cfi_get_offset();
3109 		prog->jited = 1;
3110 		prog->jited_len = proglen - cfi_get_offset();
3111 	} else {
3112 		prog = orig_prog;
3113 	}
3114 
3115 	if (!image || !prog->is_func || extra_pass) {
3116 		if (image)
3117 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
3118 out_addrs:
3119 		kvfree(addrs);
3120 		kfree(jit_data);
3121 		prog->aux->jit_data = NULL;
3122 	}
3123 out:
3124 	if (tmp_blinded)
3125 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
3126 					   tmp : orig_prog);
3127 	return prog;
3128 }
3129 
3130 bool bpf_jit_supports_kfunc_call(void)
3131 {
3132 	return true;
3133 }
3134 
3135 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
3136 {
3137 	if (text_poke_copy(dst, src, len) == NULL)
3138 		return ERR_PTR(-EINVAL);
3139 	return dst;
3140 }
3141 
3142 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
3143 bool bpf_jit_supports_subprog_tailcalls(void)
3144 {
3145 	return true;
3146 }
3147 
3148 void bpf_jit_free(struct bpf_prog *prog)
3149 {
3150 	if (prog->jited) {
3151 		struct x64_jit_data *jit_data = prog->aux->jit_data;
3152 		struct bpf_binary_header *hdr;
3153 
3154 		/*
3155 		 * If we fail the final pass of JIT (from jit_subprogs),
3156 		 * the program may not be finalized yet. Call finalize here
3157 		 * before freeing it.
3158 		 */
3159 		if (jit_data) {
3160 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
3161 						     jit_data->rw_header);
3162 			kvfree(jit_data->addrs);
3163 			kfree(jit_data);
3164 		}
3165 		prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset();
3166 		hdr = bpf_jit_binary_pack_hdr(prog);
3167 		bpf_jit_binary_pack_free(hdr, NULL);
3168 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
3169 	}
3170 
3171 	bpf_prog_unlock_free(prog);
3172 }
3173 
3174 bool bpf_jit_supports_exceptions(void)
3175 {
3176 	/* We unwind through both kernel frames (starting from within bpf_throw
3177 	 * call) and BPF frames. Therefore we require ORC unwinder to be enabled
3178 	 * to walk kernel frames and reach BPF frames in the stack trace.
3179 	 */
3180 	return IS_ENABLED(CONFIG_UNWINDER_ORC);
3181 }
3182 
3183 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3184 {
3185 #if defined(CONFIG_UNWINDER_ORC)
3186 	struct unwind_state state;
3187 	unsigned long addr;
3188 
3189 	for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
3190 	     unwind_next_frame(&state)) {
3191 		addr = unwind_get_return_address(&state);
3192 		if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
3193 			break;
3194 	}
3195 	return;
3196 #endif
3197 	WARN(1, "verification of programs using bpf_throw should have failed\n");
3198 }
3199 
3200 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
3201 			       struct bpf_prog *new, struct bpf_prog *old)
3202 {
3203 	u8 *old_addr, *new_addr, *old_bypass_addr;
3204 	int ret;
3205 
3206 	old_bypass_addr = old ? NULL : poke->bypass_addr;
3207 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
3208 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
3209 
3210 	/*
3211 	 * On program loading or teardown, the program's kallsym entry
3212 	 * might not be in place, so we use __bpf_arch_text_poke to skip
3213 	 * the kallsyms check.
3214 	 */
3215 	if (new) {
3216 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3217 					   BPF_MOD_JUMP,
3218 					   old_addr, new_addr);
3219 		BUG_ON(ret < 0);
3220 		if (!old) {
3221 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3222 						   BPF_MOD_JUMP,
3223 						   poke->bypass_addr,
3224 						   NULL);
3225 			BUG_ON(ret < 0);
3226 		}
3227 	} else {
3228 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3229 					   BPF_MOD_JUMP,
3230 					   old_bypass_addr,
3231 					   poke->bypass_addr);
3232 		BUG_ON(ret < 0);
3233 		/* let other CPUs finish the execution of program
3234 		 * so that it will not possible to expose them
3235 		 * to invalid nop, stack unwind, nop state
3236 		 */
3237 		if (!ret)
3238 			synchronize_rcu();
3239 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3240 					   BPF_MOD_JUMP,
3241 					   old_addr, NULL);
3242 		BUG_ON(ret < 0);
3243 	}
3244 }
3245