xref: /linux/arch/x86/net/bpf_jit_comp.c (revision 7c7e33b799ac169e5fab8abfc6819fce8b26d53d)
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 
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
21 {
22 	if (len == 1)
23 		*ptr = bytes;
24 	else if (len == 2)
25 		*(u16 *)ptr = bytes;
26 	else {
27 		*(u32 *)ptr = bytes;
28 		barrier();
29 	}
30 	return ptr + len;
31 }
32 
33 #define EMIT(bytes, len) \
34 	do { prog = emit_code(prog, bytes, len); } while (0)
35 
36 #define EMIT1(b1)		EMIT(b1, 1)
37 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
38 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
39 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
40 
41 #define EMIT1_off32(b1, off) \
42 	do { EMIT1(b1); EMIT(off, 4); } while (0)
43 #define EMIT2_off32(b1, b2, off) \
44 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
45 #define EMIT3_off32(b1, b2, b3, off) \
46 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
47 #define EMIT4_off32(b1, b2, b3, b4, off) \
48 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
49 
50 #ifdef CONFIG_X86_KERNEL_IBT
51 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
52 #else
53 #define EMIT_ENDBR()
54 #endif
55 
56 static bool is_imm8(int value)
57 {
58 	return value <= 127 && value >= -128;
59 }
60 
61 static bool is_simm32(s64 value)
62 {
63 	return value == (s64)(s32)value;
64 }
65 
66 static bool is_uimm32(u64 value)
67 {
68 	return value == (u64)(u32)value;
69 }
70 
71 /* mov dst, src */
72 #define EMIT_mov(DST, SRC)								 \
73 	do {										 \
74 		if (DST != SRC)								 \
75 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
76 	} while (0)
77 
78 static int bpf_size_to_x86_bytes(int bpf_size)
79 {
80 	if (bpf_size == BPF_W)
81 		return 4;
82 	else if (bpf_size == BPF_H)
83 		return 2;
84 	else if (bpf_size == BPF_B)
85 		return 1;
86 	else if (bpf_size == BPF_DW)
87 		return 4; /* imm32 */
88 	else
89 		return 0;
90 }
91 
92 /*
93  * List of x86 cond jumps opcodes (. + s8)
94  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
95  */
96 #define X86_JB  0x72
97 #define X86_JAE 0x73
98 #define X86_JE  0x74
99 #define X86_JNE 0x75
100 #define X86_JBE 0x76
101 #define X86_JA  0x77
102 #define X86_JL  0x7C
103 #define X86_JGE 0x7D
104 #define X86_JLE 0x7E
105 #define X86_JG  0x7F
106 
107 /* Pick a register outside of BPF range for JIT internal work */
108 #define AUX_REG (MAX_BPF_JIT_REG + 1)
109 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
110 
111 /*
112  * The following table maps BPF registers to x86-64 registers.
113  *
114  * x86-64 register R12 is unused, since if used as base address
115  * register in load/store instructions, it always needs an
116  * extra byte of encoding and is callee saved.
117  *
118  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
119  * trampoline. x86-64 register R10 is used for blinding (if enabled).
120  */
121 static const int reg2hex[] = {
122 	[BPF_REG_0] = 0,  /* RAX */
123 	[BPF_REG_1] = 7,  /* RDI */
124 	[BPF_REG_2] = 6,  /* RSI */
125 	[BPF_REG_3] = 2,  /* RDX */
126 	[BPF_REG_4] = 1,  /* RCX */
127 	[BPF_REG_5] = 0,  /* R8  */
128 	[BPF_REG_6] = 3,  /* RBX callee saved */
129 	[BPF_REG_7] = 5,  /* R13 callee saved */
130 	[BPF_REG_8] = 6,  /* R14 callee saved */
131 	[BPF_REG_9] = 7,  /* R15 callee saved */
132 	[BPF_REG_FP] = 5, /* RBP readonly */
133 	[BPF_REG_AX] = 2, /* R10 temp register */
134 	[AUX_REG] = 3,    /* R11 temp register */
135 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
136 };
137 
138 static const int reg2pt_regs[] = {
139 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
140 	[BPF_REG_1] = offsetof(struct pt_regs, di),
141 	[BPF_REG_2] = offsetof(struct pt_regs, si),
142 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
143 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
144 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
145 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
146 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
147 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
148 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
149 };
150 
151 /*
152  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
153  * which need extra byte of encoding.
154  * rax,rcx,...,rbp have simpler encoding
155  */
156 static bool is_ereg(u32 reg)
157 {
158 	return (1 << reg) & (BIT(BPF_REG_5) |
159 			     BIT(AUX_REG) |
160 			     BIT(BPF_REG_7) |
161 			     BIT(BPF_REG_8) |
162 			     BIT(BPF_REG_9) |
163 			     BIT(X86_REG_R9) |
164 			     BIT(BPF_REG_AX));
165 }
166 
167 /*
168  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
169  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
170  * of encoding. al,cl,dl,bl have simpler encoding.
171  */
172 static bool is_ereg_8l(u32 reg)
173 {
174 	return is_ereg(reg) ||
175 	    (1 << reg) & (BIT(BPF_REG_1) |
176 			  BIT(BPF_REG_2) |
177 			  BIT(BPF_REG_FP));
178 }
179 
180 static bool is_axreg(u32 reg)
181 {
182 	return reg == BPF_REG_0;
183 }
184 
185 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
186 static u8 add_1mod(u8 byte, u32 reg)
187 {
188 	if (is_ereg(reg))
189 		byte |= 1;
190 	return byte;
191 }
192 
193 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
194 {
195 	if (is_ereg(r1))
196 		byte |= 1;
197 	if (is_ereg(r2))
198 		byte |= 4;
199 	return byte;
200 }
201 
202 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
203 static u8 add_1reg(u8 byte, u32 dst_reg)
204 {
205 	return byte + reg2hex[dst_reg];
206 }
207 
208 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
209 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
210 {
211 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
212 }
213 
214 /* Some 1-byte opcodes for binary ALU operations */
215 static u8 simple_alu_opcodes[] = {
216 	[BPF_ADD] = 0x01,
217 	[BPF_SUB] = 0x29,
218 	[BPF_AND] = 0x21,
219 	[BPF_OR] = 0x09,
220 	[BPF_XOR] = 0x31,
221 	[BPF_LSH] = 0xE0,
222 	[BPF_RSH] = 0xE8,
223 	[BPF_ARSH] = 0xF8,
224 };
225 
226 static void jit_fill_hole(void *area, unsigned int size)
227 {
228 	/* Fill whole space with INT3 instructions */
229 	memset(area, 0xcc, size);
230 }
231 
232 int bpf_arch_text_invalidate(void *dst, size_t len)
233 {
234 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
235 }
236 
237 struct jit_context {
238 	int cleanup_addr; /* Epilogue code offset */
239 
240 	/*
241 	 * Program specific offsets of labels in the code; these rely on the
242 	 * JIT doing at least 2 passes, recording the position on the first
243 	 * pass, only to generate the correct offset on the second pass.
244 	 */
245 	int tail_call_direct_label;
246 	int tail_call_indirect_label;
247 };
248 
249 /* Maximum number of bytes emitted while JITing one eBPF insn */
250 #define BPF_MAX_INSN_SIZE	128
251 #define BPF_INSN_SAFETY		64
252 
253 /* Number of bytes emit_patch() needs to generate instructions */
254 #define X86_PATCH_SIZE		5
255 /* Number of bytes that will be skipped on tailcall */
256 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
257 
258 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
259 {
260 	u8 *prog = *pprog;
261 
262 	if (callee_regs_used[0])
263 		EMIT1(0x53);         /* push rbx */
264 	if (callee_regs_used[1])
265 		EMIT2(0x41, 0x55);   /* push r13 */
266 	if (callee_regs_used[2])
267 		EMIT2(0x41, 0x56);   /* push r14 */
268 	if (callee_regs_used[3])
269 		EMIT2(0x41, 0x57);   /* push r15 */
270 	*pprog = prog;
271 }
272 
273 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
274 {
275 	u8 *prog = *pprog;
276 
277 	if (callee_regs_used[3])
278 		EMIT2(0x41, 0x5F);   /* pop r15 */
279 	if (callee_regs_used[2])
280 		EMIT2(0x41, 0x5E);   /* pop r14 */
281 	if (callee_regs_used[1])
282 		EMIT2(0x41, 0x5D);   /* pop r13 */
283 	if (callee_regs_used[0])
284 		EMIT1(0x5B);         /* pop rbx */
285 	*pprog = prog;
286 }
287 
288 /*
289  * Emit x86-64 prologue code for BPF program.
290  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
291  * while jumping to another program
292  */
293 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
294 			  bool tail_call_reachable, bool is_subprog)
295 {
296 	u8 *prog = *pprog;
297 
298 	/* BPF trampoline can be made to work without these nops,
299 	 * but let's waste 5 bytes for now and optimize later
300 	 */
301 	EMIT_ENDBR();
302 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
303 	prog += X86_PATCH_SIZE;
304 	if (!ebpf_from_cbpf) {
305 		if (tail_call_reachable && !is_subprog)
306 			EMIT2(0x31, 0xC0); /* xor eax, eax */
307 		else
308 			EMIT2(0x66, 0x90); /* nop2 */
309 	}
310 	EMIT1(0x55);             /* push rbp */
311 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
312 
313 	/* X86_TAIL_CALL_OFFSET is here */
314 	EMIT_ENDBR();
315 
316 	/* sub rsp, rounded_stack_depth */
317 	if (stack_depth)
318 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
319 	if (tail_call_reachable)
320 		EMIT1(0x50);         /* push rax */
321 	*pprog = prog;
322 }
323 
324 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
325 {
326 	u8 *prog = *pprog;
327 	s64 offset;
328 
329 	offset = func - (ip + X86_PATCH_SIZE);
330 	if (!is_simm32(offset)) {
331 		pr_err("Target call %p is out of range\n", func);
332 		return -ERANGE;
333 	}
334 	EMIT1_off32(opcode, offset);
335 	*pprog = prog;
336 	return 0;
337 }
338 
339 static int emit_call(u8 **pprog, void *func, void *ip)
340 {
341 	return emit_patch(pprog, func, ip, 0xE8);
342 }
343 
344 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
345 {
346 	OPTIMIZER_HIDE_VAR(func);
347 	x86_call_depth_emit_accounting(pprog, func);
348 	return emit_patch(pprog, func, ip, 0xE8);
349 }
350 
351 static int emit_jump(u8 **pprog, void *func, void *ip)
352 {
353 	return emit_patch(pprog, func, ip, 0xE9);
354 }
355 
356 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
357 				void *old_addr, void *new_addr)
358 {
359 	const u8 *nop_insn = x86_nops[5];
360 	u8 old_insn[X86_PATCH_SIZE];
361 	u8 new_insn[X86_PATCH_SIZE];
362 	u8 *prog;
363 	int ret;
364 
365 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
366 	if (old_addr) {
367 		prog = old_insn;
368 		ret = t == BPF_MOD_CALL ?
369 		      emit_call(&prog, old_addr, ip) :
370 		      emit_jump(&prog, old_addr, ip);
371 		if (ret)
372 			return ret;
373 	}
374 
375 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
376 	if (new_addr) {
377 		prog = new_insn;
378 		ret = t == BPF_MOD_CALL ?
379 		      emit_call(&prog, new_addr, ip) :
380 		      emit_jump(&prog, new_addr, ip);
381 		if (ret)
382 			return ret;
383 	}
384 
385 	ret = -EBUSY;
386 	mutex_lock(&text_mutex);
387 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
388 		goto out;
389 	ret = 1;
390 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
391 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
392 		ret = 0;
393 	}
394 out:
395 	mutex_unlock(&text_mutex);
396 	return ret;
397 }
398 
399 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
400 		       void *old_addr, void *new_addr)
401 {
402 	if (!is_kernel_text((long)ip) &&
403 	    !is_bpf_text_address((long)ip))
404 		/* BPF poking in modules is not supported */
405 		return -EINVAL;
406 
407 	/*
408 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
409 	 * with an ENDBR instruction.
410 	 */
411 	if (is_endbr(*(u32 *)ip))
412 		ip += ENDBR_INSN_SIZE;
413 
414 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
415 }
416 
417 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
418 
419 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
420 {
421 	u8 *prog = *pprog;
422 
423 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
424 		EMIT_LFENCE();
425 		EMIT2(0xFF, 0xE0 + reg);
426 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
427 		OPTIMIZER_HIDE_VAR(reg);
428 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
429 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
430 		else
431 			emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
432 	} else {
433 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
434 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
435 			EMIT1(0xCC);		/* int3 */
436 	}
437 
438 	*pprog = prog;
439 }
440 
441 static void emit_return(u8 **pprog, u8 *ip)
442 {
443 	u8 *prog = *pprog;
444 
445 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
446 		emit_jump(&prog, x86_return_thunk, ip);
447 	} else {
448 		EMIT1(0xC3);		/* ret */
449 		if (IS_ENABLED(CONFIG_SLS))
450 			EMIT1(0xCC);	/* int3 */
451 	}
452 
453 	*pprog = prog;
454 }
455 
456 /*
457  * Generate the following code:
458  *
459  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
460  *   if (index >= array->map.max_entries)
461  *     goto out;
462  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
463  *     goto out;
464  *   prog = array->ptrs[index];
465  *   if (prog == NULL)
466  *     goto out;
467  *   goto *(prog->bpf_func + prologue_size);
468  * out:
469  */
470 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
471 					u32 stack_depth, u8 *ip,
472 					struct jit_context *ctx)
473 {
474 	int tcc_off = -4 - round_up(stack_depth, 8);
475 	u8 *prog = *pprog, *start = *pprog;
476 	int offset;
477 
478 	/*
479 	 * rdi - pointer to ctx
480 	 * rsi - pointer to bpf_array
481 	 * rdx - index in bpf_array
482 	 */
483 
484 	/*
485 	 * if (index >= array->map.max_entries)
486 	 *	goto out;
487 	 */
488 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
489 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
490 	      offsetof(struct bpf_array, map.max_entries));
491 
492 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
493 	EMIT2(X86_JBE, offset);                   /* jbe out */
494 
495 	/*
496 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
497 	 *	goto out;
498 	 */
499 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
500 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
501 
502 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
503 	EMIT2(X86_JAE, offset);                   /* jae out */
504 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
505 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
506 
507 	/* prog = array->ptrs[index]; */
508 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
509 		    offsetof(struct bpf_array, ptrs));
510 
511 	/*
512 	 * if (prog == NULL)
513 	 *	goto out;
514 	 */
515 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
516 
517 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
518 	EMIT2(X86_JE, offset);                    /* je out */
519 
520 	pop_callee_regs(&prog, callee_regs_used);
521 
522 	EMIT1(0x58);                              /* pop rax */
523 	if (stack_depth)
524 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
525 			    round_up(stack_depth, 8));
526 
527 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
528 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
529 	      offsetof(struct bpf_prog, bpf_func));
530 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
531 	      X86_TAIL_CALL_OFFSET);
532 	/*
533 	 * Now we're ready to jump into next BPF program
534 	 * rdi == ctx (1st arg)
535 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
536 	 */
537 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
538 
539 	/* out: */
540 	ctx->tail_call_indirect_label = prog - start;
541 	*pprog = prog;
542 }
543 
544 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
545 				      u8 **pprog, u8 *ip,
546 				      bool *callee_regs_used, u32 stack_depth,
547 				      struct jit_context *ctx)
548 {
549 	int tcc_off = -4 - round_up(stack_depth, 8);
550 	u8 *prog = *pprog, *start = *pprog;
551 	int offset;
552 
553 	/*
554 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
555 	 *	goto out;
556 	 */
557 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
558 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
559 
560 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
561 	EMIT2(X86_JAE, offset);                       /* jae out */
562 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
563 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
564 
565 	poke->tailcall_bypass = ip + (prog - start);
566 	poke->adj_off = X86_TAIL_CALL_OFFSET;
567 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
568 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
569 
570 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
571 		  poke->tailcall_bypass);
572 
573 	pop_callee_regs(&prog, callee_regs_used);
574 	EMIT1(0x58);                                  /* pop rax */
575 	if (stack_depth)
576 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
577 
578 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
579 	prog += X86_PATCH_SIZE;
580 
581 	/* out: */
582 	ctx->tail_call_direct_label = prog - start;
583 
584 	*pprog = prog;
585 }
586 
587 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
588 {
589 	struct bpf_jit_poke_descriptor *poke;
590 	struct bpf_array *array;
591 	struct bpf_prog *target;
592 	int i, ret;
593 
594 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
595 		poke = &prog->aux->poke_tab[i];
596 		if (poke->aux && poke->aux != prog->aux)
597 			continue;
598 
599 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
600 
601 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
602 			continue;
603 
604 		array = container_of(poke->tail_call.map, struct bpf_array, map);
605 		mutex_lock(&array->aux->poke_mutex);
606 		target = array->ptrs[poke->tail_call.key];
607 		if (target) {
608 			ret = __bpf_arch_text_poke(poke->tailcall_target,
609 						   BPF_MOD_JUMP, NULL,
610 						   (u8 *)target->bpf_func +
611 						   poke->adj_off);
612 			BUG_ON(ret < 0);
613 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
614 						   BPF_MOD_JUMP,
615 						   (u8 *)poke->tailcall_target +
616 						   X86_PATCH_SIZE, NULL);
617 			BUG_ON(ret < 0);
618 		}
619 		WRITE_ONCE(poke->tailcall_target_stable, true);
620 		mutex_unlock(&array->aux->poke_mutex);
621 	}
622 }
623 
624 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
625 			   u32 dst_reg, const u32 imm32)
626 {
627 	u8 *prog = *pprog;
628 	u8 b1, b2, b3;
629 
630 	/*
631 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
632 	 * (which zero-extends imm32) to save 2 bytes.
633 	 */
634 	if (sign_propagate && (s32)imm32 < 0) {
635 		/* 'mov %rax, imm32' sign extends imm32 */
636 		b1 = add_1mod(0x48, dst_reg);
637 		b2 = 0xC7;
638 		b3 = 0xC0;
639 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
640 		goto done;
641 	}
642 
643 	/*
644 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
645 	 * to save 3 bytes.
646 	 */
647 	if (imm32 == 0) {
648 		if (is_ereg(dst_reg))
649 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
650 		b2 = 0x31; /* xor */
651 		b3 = 0xC0;
652 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
653 		goto done;
654 	}
655 
656 	/* mov %eax, imm32 */
657 	if (is_ereg(dst_reg))
658 		EMIT1(add_1mod(0x40, dst_reg));
659 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
660 done:
661 	*pprog = prog;
662 }
663 
664 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
665 			   const u32 imm32_hi, const u32 imm32_lo)
666 {
667 	u8 *prog = *pprog;
668 
669 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
670 		/*
671 		 * For emitting plain u32, where sign bit must not be
672 		 * propagated LLVM tends to load imm64 over mov32
673 		 * directly, so save couple of bytes by just doing
674 		 * 'mov %eax, imm32' instead.
675 		 */
676 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
677 	} else {
678 		/* movabsq rax, imm64 */
679 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
680 		EMIT(imm32_lo, 4);
681 		EMIT(imm32_hi, 4);
682 	}
683 
684 	*pprog = prog;
685 }
686 
687 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
688 {
689 	u8 *prog = *pprog;
690 
691 	if (is64) {
692 		/* mov dst, src */
693 		EMIT_mov(dst_reg, src_reg);
694 	} else {
695 		/* mov32 dst, src */
696 		if (is_ereg(dst_reg) || is_ereg(src_reg))
697 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
698 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
699 	}
700 
701 	*pprog = prog;
702 }
703 
704 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
705 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
706 {
707 	u8 *prog = *pprog;
708 
709 	if (is_imm8(off)) {
710 		/* 1-byte signed displacement.
711 		 *
712 		 * If off == 0 we could skip this and save one extra byte, but
713 		 * special case of x86 R13 which always needs an offset is not
714 		 * worth the hassle
715 		 */
716 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
717 	} else {
718 		/* 4-byte signed displacement */
719 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
720 	}
721 	*pprog = prog;
722 }
723 
724 /*
725  * Emit a REX byte if it will be necessary to address these registers
726  */
727 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
728 {
729 	u8 *prog = *pprog;
730 
731 	if (is64)
732 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
733 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
734 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
735 	*pprog = prog;
736 }
737 
738 /*
739  * Similar version of maybe_emit_mod() for a single register
740  */
741 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
742 {
743 	u8 *prog = *pprog;
744 
745 	if (is64)
746 		EMIT1(add_1mod(0x48, reg));
747 	else if (is_ereg(reg))
748 		EMIT1(add_1mod(0x40, reg));
749 	*pprog = prog;
750 }
751 
752 /* LDX: dst_reg = *(u8*)(src_reg + off) */
753 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
754 {
755 	u8 *prog = *pprog;
756 
757 	switch (size) {
758 	case BPF_B:
759 		/* Emit 'movzx rax, byte ptr [rax + off]' */
760 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
761 		break;
762 	case BPF_H:
763 		/* Emit 'movzx rax, word ptr [rax + off]' */
764 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
765 		break;
766 	case BPF_W:
767 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
768 		if (is_ereg(dst_reg) || is_ereg(src_reg))
769 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
770 		else
771 			EMIT1(0x8B);
772 		break;
773 	case BPF_DW:
774 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
775 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
776 		break;
777 	}
778 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
779 	*pprog = prog;
780 }
781 
782 /* STX: *(u8*)(dst_reg + off) = src_reg */
783 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
784 {
785 	u8 *prog = *pprog;
786 
787 	switch (size) {
788 	case BPF_B:
789 		/* Emit 'mov byte ptr [rax + off], al' */
790 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
791 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
792 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
793 		else
794 			EMIT1(0x88);
795 		break;
796 	case BPF_H:
797 		if (is_ereg(dst_reg) || is_ereg(src_reg))
798 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
799 		else
800 			EMIT2(0x66, 0x89);
801 		break;
802 	case BPF_W:
803 		if (is_ereg(dst_reg) || is_ereg(src_reg))
804 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
805 		else
806 			EMIT1(0x89);
807 		break;
808 	case BPF_DW:
809 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
810 		break;
811 	}
812 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
813 	*pprog = prog;
814 }
815 
816 static int emit_atomic(u8 **pprog, u8 atomic_op,
817 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
818 {
819 	u8 *prog = *pprog;
820 
821 	EMIT1(0xF0); /* lock prefix */
822 
823 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
824 
825 	/* emit opcode */
826 	switch (atomic_op) {
827 	case BPF_ADD:
828 	case BPF_AND:
829 	case BPF_OR:
830 	case BPF_XOR:
831 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
832 		EMIT1(simple_alu_opcodes[atomic_op]);
833 		break;
834 	case BPF_ADD | BPF_FETCH:
835 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
836 		EMIT2(0x0F, 0xC1);
837 		break;
838 	case BPF_XCHG:
839 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
840 		EMIT1(0x87);
841 		break;
842 	case BPF_CMPXCHG:
843 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
844 		EMIT2(0x0F, 0xB1);
845 		break;
846 	default:
847 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
848 		return -EFAULT;
849 	}
850 
851 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
852 
853 	*pprog = prog;
854 	return 0;
855 }
856 
857 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
858 {
859 	u32 reg = x->fixup >> 8;
860 
861 	/* jump over faulting load and clear dest register */
862 	*(unsigned long *)((void *)regs + reg) = 0;
863 	regs->ip += x->fixup & 0xff;
864 	return true;
865 }
866 
867 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
868 			     bool *regs_used, bool *tail_call_seen)
869 {
870 	int i;
871 
872 	for (i = 1; i <= insn_cnt; i++, insn++) {
873 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
874 			*tail_call_seen = true;
875 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
876 			regs_used[0] = true;
877 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
878 			regs_used[1] = true;
879 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
880 			regs_used[2] = true;
881 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
882 			regs_used[3] = true;
883 	}
884 }
885 
886 static void emit_nops(u8 **pprog, int len)
887 {
888 	u8 *prog = *pprog;
889 	int i, noplen;
890 
891 	while (len > 0) {
892 		noplen = len;
893 
894 		if (noplen > ASM_NOP_MAX)
895 			noplen = ASM_NOP_MAX;
896 
897 		for (i = 0; i < noplen; i++)
898 			EMIT1(x86_nops[noplen][i]);
899 		len -= noplen;
900 	}
901 
902 	*pprog = prog;
903 }
904 
905 /* emit the 3-byte VEX prefix
906  *
907  * r: same as rex.r, extra bit for ModRM reg field
908  * x: same as rex.x, extra bit for SIB index field
909  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
910  * m: opcode map select, encoding escape bytes e.g. 0x0f38
911  * w: same as rex.w (32 bit or 64 bit) or opcode specific
912  * src_reg2: additional source reg (encoded as BPF reg)
913  * l: vector length (128 bit or 256 bit) or reserved
914  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
915  */
916 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
917 		      bool w, u8 src_reg2, bool l, u8 pp)
918 {
919 	u8 *prog = *pprog;
920 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
921 	u8 b1, b2;
922 	u8 vvvv = reg2hex[src_reg2];
923 
924 	/* reg2hex gives only the lower 3 bit of vvvv */
925 	if (is_ereg(src_reg2))
926 		vvvv |= 1 << 3;
927 
928 	/*
929 	 * 2nd byte of 3-byte VEX prefix
930 	 * ~ means bit inverted encoding
931 	 *
932 	 *    7                           0
933 	 *  +---+---+---+---+---+---+---+---+
934 	 *  |~R |~X |~B |         m         |
935 	 *  +---+---+---+---+---+---+---+---+
936 	 */
937 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
938 	/*
939 	 * 3rd byte of 3-byte VEX prefix
940 	 *
941 	 *    7                           0
942 	 *  +---+---+---+---+---+---+---+---+
943 	 *  | W |     ~vvvv     | L |   pp  |
944 	 *  +---+---+---+---+---+---+---+---+
945 	 */
946 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
947 
948 	EMIT3(b0, b1, b2);
949 	*pprog = prog;
950 }
951 
952 /* emit BMI2 shift instruction */
953 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
954 {
955 	u8 *prog = *pprog;
956 	bool r = is_ereg(dst_reg);
957 	u8 m = 2; /* escape code 0f38 */
958 
959 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
960 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
961 	*pprog = prog;
962 }
963 
964 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
965 
966 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
967 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
968 {
969 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
970 	struct bpf_insn *insn = bpf_prog->insnsi;
971 	bool callee_regs_used[4] = {};
972 	int insn_cnt = bpf_prog->len;
973 	bool tail_call_seen = false;
974 	bool seen_exit = false;
975 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
976 	int i, excnt = 0;
977 	int ilen, proglen = 0;
978 	u8 *prog = temp;
979 	int err;
980 
981 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
982 			 &tail_call_seen);
983 
984 	/* tail call's presence in current prog implies it is reachable */
985 	tail_call_reachable |= tail_call_seen;
986 
987 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
988 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
989 		      bpf_prog->aux->func_idx != 0);
990 	push_callee_regs(&prog, callee_regs_used);
991 
992 	ilen = prog - temp;
993 	if (rw_image)
994 		memcpy(rw_image + proglen, temp, ilen);
995 	proglen += ilen;
996 	addrs[0] = proglen;
997 	prog = temp;
998 
999 	for (i = 1; i <= insn_cnt; i++, insn++) {
1000 		const s32 imm32 = insn->imm;
1001 		u32 dst_reg = insn->dst_reg;
1002 		u32 src_reg = insn->src_reg;
1003 		u8 b2 = 0, b3 = 0;
1004 		u8 *start_of_ldx;
1005 		s64 jmp_offset;
1006 		s16 insn_off;
1007 		u8 jmp_cond;
1008 		u8 *func;
1009 		int nops;
1010 
1011 		switch (insn->code) {
1012 			/* ALU */
1013 		case BPF_ALU | BPF_ADD | BPF_X:
1014 		case BPF_ALU | BPF_SUB | BPF_X:
1015 		case BPF_ALU | BPF_AND | BPF_X:
1016 		case BPF_ALU | BPF_OR | BPF_X:
1017 		case BPF_ALU | BPF_XOR | BPF_X:
1018 		case BPF_ALU64 | BPF_ADD | BPF_X:
1019 		case BPF_ALU64 | BPF_SUB | BPF_X:
1020 		case BPF_ALU64 | BPF_AND | BPF_X:
1021 		case BPF_ALU64 | BPF_OR | BPF_X:
1022 		case BPF_ALU64 | BPF_XOR | BPF_X:
1023 			maybe_emit_mod(&prog, dst_reg, src_reg,
1024 				       BPF_CLASS(insn->code) == BPF_ALU64);
1025 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1026 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1027 			break;
1028 
1029 		case BPF_ALU64 | BPF_MOV | BPF_X:
1030 		case BPF_ALU | BPF_MOV | BPF_X:
1031 			emit_mov_reg(&prog,
1032 				     BPF_CLASS(insn->code) == BPF_ALU64,
1033 				     dst_reg, src_reg);
1034 			break;
1035 
1036 			/* neg dst */
1037 		case BPF_ALU | BPF_NEG:
1038 		case BPF_ALU64 | BPF_NEG:
1039 			maybe_emit_1mod(&prog, dst_reg,
1040 					BPF_CLASS(insn->code) == BPF_ALU64);
1041 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1042 			break;
1043 
1044 		case BPF_ALU | BPF_ADD | BPF_K:
1045 		case BPF_ALU | BPF_SUB | BPF_K:
1046 		case BPF_ALU | BPF_AND | BPF_K:
1047 		case BPF_ALU | BPF_OR | BPF_K:
1048 		case BPF_ALU | BPF_XOR | BPF_K:
1049 		case BPF_ALU64 | BPF_ADD | BPF_K:
1050 		case BPF_ALU64 | BPF_SUB | BPF_K:
1051 		case BPF_ALU64 | BPF_AND | BPF_K:
1052 		case BPF_ALU64 | BPF_OR | BPF_K:
1053 		case BPF_ALU64 | BPF_XOR | BPF_K:
1054 			maybe_emit_1mod(&prog, dst_reg,
1055 					BPF_CLASS(insn->code) == BPF_ALU64);
1056 
1057 			/*
1058 			 * b3 holds 'normal' opcode, b2 short form only valid
1059 			 * in case dst is eax/rax.
1060 			 */
1061 			switch (BPF_OP(insn->code)) {
1062 			case BPF_ADD:
1063 				b3 = 0xC0;
1064 				b2 = 0x05;
1065 				break;
1066 			case BPF_SUB:
1067 				b3 = 0xE8;
1068 				b2 = 0x2D;
1069 				break;
1070 			case BPF_AND:
1071 				b3 = 0xE0;
1072 				b2 = 0x25;
1073 				break;
1074 			case BPF_OR:
1075 				b3 = 0xC8;
1076 				b2 = 0x0D;
1077 				break;
1078 			case BPF_XOR:
1079 				b3 = 0xF0;
1080 				b2 = 0x35;
1081 				break;
1082 			}
1083 
1084 			if (is_imm8(imm32))
1085 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1086 			else if (is_axreg(dst_reg))
1087 				EMIT1_off32(b2, imm32);
1088 			else
1089 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1090 			break;
1091 
1092 		case BPF_ALU64 | BPF_MOV | BPF_K:
1093 		case BPF_ALU | BPF_MOV | BPF_K:
1094 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1095 				       dst_reg, imm32);
1096 			break;
1097 
1098 		case BPF_LD | BPF_IMM | BPF_DW:
1099 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1100 			insn++;
1101 			i++;
1102 			break;
1103 
1104 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1105 		case BPF_ALU | BPF_MOD | BPF_X:
1106 		case BPF_ALU | BPF_DIV | BPF_X:
1107 		case BPF_ALU | BPF_MOD | BPF_K:
1108 		case BPF_ALU | BPF_DIV | BPF_K:
1109 		case BPF_ALU64 | BPF_MOD | BPF_X:
1110 		case BPF_ALU64 | BPF_DIV | BPF_X:
1111 		case BPF_ALU64 | BPF_MOD | BPF_K:
1112 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1113 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1114 
1115 			if (dst_reg != BPF_REG_0)
1116 				EMIT1(0x50); /* push rax */
1117 			if (dst_reg != BPF_REG_3)
1118 				EMIT1(0x52); /* push rdx */
1119 
1120 			if (BPF_SRC(insn->code) == BPF_X) {
1121 				if (src_reg == BPF_REG_0 ||
1122 				    src_reg == BPF_REG_3) {
1123 					/* mov r11, src_reg */
1124 					EMIT_mov(AUX_REG, src_reg);
1125 					src_reg = AUX_REG;
1126 				}
1127 			} else {
1128 				/* mov r11, imm32 */
1129 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1130 				src_reg = AUX_REG;
1131 			}
1132 
1133 			if (dst_reg != BPF_REG_0)
1134 				/* mov rax, dst_reg */
1135 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1136 
1137 			/*
1138 			 * xor edx, edx
1139 			 * equivalent to 'xor rdx, rdx', but one byte less
1140 			 */
1141 			EMIT2(0x31, 0xd2);
1142 
1143 			/* div src_reg */
1144 			maybe_emit_1mod(&prog, src_reg, is64);
1145 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1146 
1147 			if (BPF_OP(insn->code) == BPF_MOD &&
1148 			    dst_reg != BPF_REG_3)
1149 				/* mov dst_reg, rdx */
1150 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1151 			else if (BPF_OP(insn->code) == BPF_DIV &&
1152 				 dst_reg != BPF_REG_0)
1153 				/* mov dst_reg, rax */
1154 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1155 
1156 			if (dst_reg != BPF_REG_3)
1157 				EMIT1(0x5A); /* pop rdx */
1158 			if (dst_reg != BPF_REG_0)
1159 				EMIT1(0x58); /* pop rax */
1160 			break;
1161 		}
1162 
1163 		case BPF_ALU | BPF_MUL | BPF_K:
1164 		case BPF_ALU64 | BPF_MUL | BPF_K:
1165 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1166 				       BPF_CLASS(insn->code) == BPF_ALU64);
1167 
1168 			if (is_imm8(imm32))
1169 				/* imul dst_reg, dst_reg, imm8 */
1170 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1171 				      imm32);
1172 			else
1173 				/* imul dst_reg, dst_reg, imm32 */
1174 				EMIT2_off32(0x69,
1175 					    add_2reg(0xC0, dst_reg, dst_reg),
1176 					    imm32);
1177 			break;
1178 
1179 		case BPF_ALU | BPF_MUL | BPF_X:
1180 		case BPF_ALU64 | BPF_MUL | BPF_X:
1181 			maybe_emit_mod(&prog, src_reg, dst_reg,
1182 				       BPF_CLASS(insn->code) == BPF_ALU64);
1183 
1184 			/* imul dst_reg, src_reg */
1185 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1186 			break;
1187 
1188 			/* Shifts */
1189 		case BPF_ALU | BPF_LSH | BPF_K:
1190 		case BPF_ALU | BPF_RSH | BPF_K:
1191 		case BPF_ALU | BPF_ARSH | BPF_K:
1192 		case BPF_ALU64 | BPF_LSH | BPF_K:
1193 		case BPF_ALU64 | BPF_RSH | BPF_K:
1194 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1195 			maybe_emit_1mod(&prog, dst_reg,
1196 					BPF_CLASS(insn->code) == BPF_ALU64);
1197 
1198 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1199 			if (imm32 == 1)
1200 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1201 			else
1202 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1203 			break;
1204 
1205 		case BPF_ALU | BPF_LSH | BPF_X:
1206 		case BPF_ALU | BPF_RSH | BPF_X:
1207 		case BPF_ALU | BPF_ARSH | BPF_X:
1208 		case BPF_ALU64 | BPF_LSH | BPF_X:
1209 		case BPF_ALU64 | BPF_RSH | BPF_X:
1210 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1211 			/* BMI2 shifts aren't better when shift count is already in rcx */
1212 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1213 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1214 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1215 				u8 op;
1216 
1217 				switch (BPF_OP(insn->code)) {
1218 				case BPF_LSH:
1219 					op = 1; /* prefix 0x66 */
1220 					break;
1221 				case BPF_RSH:
1222 					op = 3; /* prefix 0xf2 */
1223 					break;
1224 				case BPF_ARSH:
1225 					op = 2; /* prefix 0xf3 */
1226 					break;
1227 				}
1228 
1229 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1230 
1231 				break;
1232 			}
1233 
1234 			if (src_reg != BPF_REG_4) { /* common case */
1235 				/* Check for bad case when dst_reg == rcx */
1236 				if (dst_reg == BPF_REG_4) {
1237 					/* mov r11, dst_reg */
1238 					EMIT_mov(AUX_REG, dst_reg);
1239 					dst_reg = AUX_REG;
1240 				} else {
1241 					EMIT1(0x51); /* push rcx */
1242 				}
1243 				/* mov rcx, src_reg */
1244 				EMIT_mov(BPF_REG_4, src_reg);
1245 			}
1246 
1247 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1248 			maybe_emit_1mod(&prog, dst_reg,
1249 					BPF_CLASS(insn->code) == BPF_ALU64);
1250 
1251 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1252 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1253 
1254 			if (src_reg != BPF_REG_4) {
1255 				if (insn->dst_reg == BPF_REG_4)
1256 					/* mov dst_reg, r11 */
1257 					EMIT_mov(insn->dst_reg, AUX_REG);
1258 				else
1259 					EMIT1(0x59); /* pop rcx */
1260 			}
1261 
1262 			break;
1263 
1264 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1265 			switch (imm32) {
1266 			case 16:
1267 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1268 				EMIT1(0x66);
1269 				if (is_ereg(dst_reg))
1270 					EMIT1(0x41);
1271 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1272 
1273 				/* Emit 'movzwl eax, ax' */
1274 				if (is_ereg(dst_reg))
1275 					EMIT3(0x45, 0x0F, 0xB7);
1276 				else
1277 					EMIT2(0x0F, 0xB7);
1278 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1279 				break;
1280 			case 32:
1281 				/* Emit 'bswap eax' to swap lower 4 bytes */
1282 				if (is_ereg(dst_reg))
1283 					EMIT2(0x41, 0x0F);
1284 				else
1285 					EMIT1(0x0F);
1286 				EMIT1(add_1reg(0xC8, dst_reg));
1287 				break;
1288 			case 64:
1289 				/* Emit 'bswap rax' to swap 8 bytes */
1290 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1291 				      add_1reg(0xC8, dst_reg));
1292 				break;
1293 			}
1294 			break;
1295 
1296 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1297 			switch (imm32) {
1298 			case 16:
1299 				/*
1300 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1301 				 * into 64 bit
1302 				 */
1303 				if (is_ereg(dst_reg))
1304 					EMIT3(0x45, 0x0F, 0xB7);
1305 				else
1306 					EMIT2(0x0F, 0xB7);
1307 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1308 				break;
1309 			case 32:
1310 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1311 				if (is_ereg(dst_reg))
1312 					EMIT1(0x45);
1313 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1314 				break;
1315 			case 64:
1316 				/* nop */
1317 				break;
1318 			}
1319 			break;
1320 
1321 			/* speculation barrier */
1322 		case BPF_ST | BPF_NOSPEC:
1323 			EMIT_LFENCE();
1324 			break;
1325 
1326 			/* ST: *(u8*)(dst_reg + off) = imm */
1327 		case BPF_ST | BPF_MEM | BPF_B:
1328 			if (is_ereg(dst_reg))
1329 				EMIT2(0x41, 0xC6);
1330 			else
1331 				EMIT1(0xC6);
1332 			goto st;
1333 		case BPF_ST | BPF_MEM | BPF_H:
1334 			if (is_ereg(dst_reg))
1335 				EMIT3(0x66, 0x41, 0xC7);
1336 			else
1337 				EMIT2(0x66, 0xC7);
1338 			goto st;
1339 		case BPF_ST | BPF_MEM | BPF_W:
1340 			if (is_ereg(dst_reg))
1341 				EMIT2(0x41, 0xC7);
1342 			else
1343 				EMIT1(0xC7);
1344 			goto st;
1345 		case BPF_ST | BPF_MEM | BPF_DW:
1346 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1347 
1348 st:			if (is_imm8(insn->off))
1349 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1350 			else
1351 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1352 
1353 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1354 			break;
1355 
1356 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1357 		case BPF_STX | BPF_MEM | BPF_B:
1358 		case BPF_STX | BPF_MEM | BPF_H:
1359 		case BPF_STX | BPF_MEM | BPF_W:
1360 		case BPF_STX | BPF_MEM | BPF_DW:
1361 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1362 			break;
1363 
1364 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1365 		case BPF_LDX | BPF_MEM | BPF_B:
1366 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1367 		case BPF_LDX | BPF_MEM | BPF_H:
1368 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1369 		case BPF_LDX | BPF_MEM | BPF_W:
1370 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1371 		case BPF_LDX | BPF_MEM | BPF_DW:
1372 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1373 			insn_off = insn->off;
1374 
1375 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1376 				/* Conservatively check that src_reg + insn->off is a kernel address:
1377 				 *   src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
1378 				 * src_reg is used as scratch for src_reg += insn->off and restored
1379 				 * after emit_ldx if necessary
1380 				 */
1381 
1382 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
1383 				u8 *end_of_jmp;
1384 
1385 				/* At end of these emitted checks, insn->off will have been added
1386 				 * to src_reg, so no need to do relative load with insn->off offset
1387 				 */
1388 				insn_off = 0;
1389 
1390 				/* movabsq r11, limit */
1391 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1392 				EMIT((u32)limit, 4);
1393 				EMIT(limit >> 32, 4);
1394 
1395 				if (insn->off) {
1396 					/* add src_reg, insn->off */
1397 					maybe_emit_1mod(&prog, src_reg, true);
1398 					EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
1399 				}
1400 
1401 				/* cmp src_reg, r11 */
1402 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1403 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1404 
1405 				/* if unsigned '>=', goto load */
1406 				EMIT2(X86_JAE, 0);
1407 				end_of_jmp = prog;
1408 
1409 				/* xor dst_reg, dst_reg */
1410 				emit_mov_imm32(&prog, false, dst_reg, 0);
1411 				/* jmp byte_after_ldx */
1412 				EMIT2(0xEB, 0);
1413 
1414 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1415 				start_of_ldx = prog;
1416 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1417 			}
1418 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1419 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1420 				struct exception_table_entry *ex;
1421 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1422 				s64 delta;
1423 
1424 				/* populate jmp_offset for JMP above */
1425 				start_of_ldx[-1] = prog - start_of_ldx;
1426 
1427 				if (insn->off && src_reg != dst_reg) {
1428 					/* sub src_reg, insn->off
1429 					 * Restore src_reg after "add src_reg, insn->off" in prev
1430 					 * if statement. But if src_reg == dst_reg, emit_ldx
1431 					 * above already clobbered src_reg, so no need to restore.
1432 					 * If add src_reg, insn->off was unnecessary, no need to
1433 					 * restore either.
1434 					 */
1435 					maybe_emit_1mod(&prog, src_reg, true);
1436 					EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
1437 				}
1438 
1439 				if (!bpf_prog->aux->extable)
1440 					break;
1441 
1442 				if (excnt >= bpf_prog->aux->num_exentries) {
1443 					pr_err("ex gen bug\n");
1444 					return -EFAULT;
1445 				}
1446 				ex = &bpf_prog->aux->extable[excnt++];
1447 
1448 				delta = _insn - (u8 *)&ex->insn;
1449 				if (!is_simm32(delta)) {
1450 					pr_err("extable->insn doesn't fit into 32-bit\n");
1451 					return -EFAULT;
1452 				}
1453 				/* switch ex to rw buffer for writes */
1454 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1455 
1456 				ex->insn = delta;
1457 
1458 				ex->data = EX_TYPE_BPF;
1459 
1460 				if (dst_reg > BPF_REG_9) {
1461 					pr_err("verifier error\n");
1462 					return -EFAULT;
1463 				}
1464 				/*
1465 				 * Compute size of x86 insn and its target dest x86 register.
1466 				 * ex_handler_bpf() will use lower 8 bits to adjust
1467 				 * pt_regs->ip to jump over this x86 instruction
1468 				 * and upper bits to figure out which pt_regs to zero out.
1469 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1470 				 * of 4 bytes will be ignored and rbx will be zero inited.
1471 				 */
1472 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1473 			}
1474 			break;
1475 
1476 		case BPF_STX | BPF_ATOMIC | BPF_W:
1477 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1478 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1479 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1480 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1481 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1482 				u32 real_src_reg = src_reg;
1483 				u32 real_dst_reg = dst_reg;
1484 				u8 *branch_target;
1485 
1486 				/*
1487 				 * Can't be implemented with a single x86 insn.
1488 				 * Need to do a CMPXCHG loop.
1489 				 */
1490 
1491 				/* Will need RAX as a CMPXCHG operand so save R0 */
1492 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1493 				if (src_reg == BPF_REG_0)
1494 					real_src_reg = BPF_REG_AX;
1495 				if (dst_reg == BPF_REG_0)
1496 					real_dst_reg = BPF_REG_AX;
1497 
1498 				branch_target = prog;
1499 				/* Load old value */
1500 				emit_ldx(&prog, BPF_SIZE(insn->code),
1501 					 BPF_REG_0, real_dst_reg, insn->off);
1502 				/*
1503 				 * Perform the (commutative) operation locally,
1504 				 * put the result in the AUX_REG.
1505 				 */
1506 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1507 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1508 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1509 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1510 				/* Attempt to swap in new value */
1511 				err = emit_atomic(&prog, BPF_CMPXCHG,
1512 						  real_dst_reg, AUX_REG,
1513 						  insn->off,
1514 						  BPF_SIZE(insn->code));
1515 				if (WARN_ON(err))
1516 					return err;
1517 				/*
1518 				 * ZF tells us whether we won the race. If it's
1519 				 * cleared we need to try again.
1520 				 */
1521 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1522 				/* Return the pre-modification value */
1523 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1524 				/* Restore R0 after clobbering RAX */
1525 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1526 				break;
1527 			}
1528 
1529 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1530 					  insn->off, BPF_SIZE(insn->code));
1531 			if (err)
1532 				return err;
1533 			break;
1534 
1535 			/* call */
1536 		case BPF_JMP | BPF_CALL: {
1537 			int offs;
1538 
1539 			func = (u8 *) __bpf_call_base + imm32;
1540 			if (tail_call_reachable) {
1541 				/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1542 				EMIT3_off32(0x48, 0x8B, 0x85,
1543 					    -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1544 				if (!imm32)
1545 					return -EINVAL;
1546 				offs = 7 + x86_call_depth_emit_accounting(&prog, func);
1547 			} else {
1548 				if (!imm32)
1549 					return -EINVAL;
1550 				offs = x86_call_depth_emit_accounting(&prog, func);
1551 			}
1552 			if (emit_call(&prog, func, image + addrs[i - 1] + offs))
1553 				return -EINVAL;
1554 			break;
1555 		}
1556 
1557 		case BPF_JMP | BPF_TAIL_CALL:
1558 			if (imm32)
1559 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1560 							  &prog, image + addrs[i - 1],
1561 							  callee_regs_used,
1562 							  bpf_prog->aux->stack_depth,
1563 							  ctx);
1564 			else
1565 				emit_bpf_tail_call_indirect(&prog,
1566 							    callee_regs_used,
1567 							    bpf_prog->aux->stack_depth,
1568 							    image + addrs[i - 1],
1569 							    ctx);
1570 			break;
1571 
1572 			/* cond jump */
1573 		case BPF_JMP | BPF_JEQ | BPF_X:
1574 		case BPF_JMP | BPF_JNE | BPF_X:
1575 		case BPF_JMP | BPF_JGT | BPF_X:
1576 		case BPF_JMP | BPF_JLT | BPF_X:
1577 		case BPF_JMP | BPF_JGE | BPF_X:
1578 		case BPF_JMP | BPF_JLE | BPF_X:
1579 		case BPF_JMP | BPF_JSGT | BPF_X:
1580 		case BPF_JMP | BPF_JSLT | BPF_X:
1581 		case BPF_JMP | BPF_JSGE | BPF_X:
1582 		case BPF_JMP | BPF_JSLE | BPF_X:
1583 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1584 		case BPF_JMP32 | BPF_JNE | BPF_X:
1585 		case BPF_JMP32 | BPF_JGT | BPF_X:
1586 		case BPF_JMP32 | BPF_JLT | BPF_X:
1587 		case BPF_JMP32 | BPF_JGE | BPF_X:
1588 		case BPF_JMP32 | BPF_JLE | BPF_X:
1589 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1590 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1591 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1592 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1593 			/* cmp dst_reg, src_reg */
1594 			maybe_emit_mod(&prog, dst_reg, src_reg,
1595 				       BPF_CLASS(insn->code) == BPF_JMP);
1596 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1597 			goto emit_cond_jmp;
1598 
1599 		case BPF_JMP | BPF_JSET | BPF_X:
1600 		case BPF_JMP32 | BPF_JSET | BPF_X:
1601 			/* test dst_reg, src_reg */
1602 			maybe_emit_mod(&prog, dst_reg, src_reg,
1603 				       BPF_CLASS(insn->code) == BPF_JMP);
1604 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1605 			goto emit_cond_jmp;
1606 
1607 		case BPF_JMP | BPF_JSET | BPF_K:
1608 		case BPF_JMP32 | BPF_JSET | BPF_K:
1609 			/* test dst_reg, imm32 */
1610 			maybe_emit_1mod(&prog, dst_reg,
1611 					BPF_CLASS(insn->code) == BPF_JMP);
1612 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1613 			goto emit_cond_jmp;
1614 
1615 		case BPF_JMP | BPF_JEQ | BPF_K:
1616 		case BPF_JMP | BPF_JNE | BPF_K:
1617 		case BPF_JMP | BPF_JGT | BPF_K:
1618 		case BPF_JMP | BPF_JLT | BPF_K:
1619 		case BPF_JMP | BPF_JGE | BPF_K:
1620 		case BPF_JMP | BPF_JLE | BPF_K:
1621 		case BPF_JMP | BPF_JSGT | BPF_K:
1622 		case BPF_JMP | BPF_JSLT | BPF_K:
1623 		case BPF_JMP | BPF_JSGE | BPF_K:
1624 		case BPF_JMP | BPF_JSLE | BPF_K:
1625 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1626 		case BPF_JMP32 | BPF_JNE | BPF_K:
1627 		case BPF_JMP32 | BPF_JGT | BPF_K:
1628 		case BPF_JMP32 | BPF_JLT | BPF_K:
1629 		case BPF_JMP32 | BPF_JGE | BPF_K:
1630 		case BPF_JMP32 | BPF_JLE | BPF_K:
1631 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1632 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1633 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1634 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1635 			/* test dst_reg, dst_reg to save one extra byte */
1636 			if (imm32 == 0) {
1637 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1638 					       BPF_CLASS(insn->code) == BPF_JMP);
1639 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1640 				goto emit_cond_jmp;
1641 			}
1642 
1643 			/* cmp dst_reg, imm8/32 */
1644 			maybe_emit_1mod(&prog, dst_reg,
1645 					BPF_CLASS(insn->code) == BPF_JMP);
1646 
1647 			if (is_imm8(imm32))
1648 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1649 			else
1650 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1651 
1652 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1653 			switch (BPF_OP(insn->code)) {
1654 			case BPF_JEQ:
1655 				jmp_cond = X86_JE;
1656 				break;
1657 			case BPF_JSET:
1658 			case BPF_JNE:
1659 				jmp_cond = X86_JNE;
1660 				break;
1661 			case BPF_JGT:
1662 				/* GT is unsigned '>', JA in x86 */
1663 				jmp_cond = X86_JA;
1664 				break;
1665 			case BPF_JLT:
1666 				/* LT is unsigned '<', JB in x86 */
1667 				jmp_cond = X86_JB;
1668 				break;
1669 			case BPF_JGE:
1670 				/* GE is unsigned '>=', JAE in x86 */
1671 				jmp_cond = X86_JAE;
1672 				break;
1673 			case BPF_JLE:
1674 				/* LE is unsigned '<=', JBE in x86 */
1675 				jmp_cond = X86_JBE;
1676 				break;
1677 			case BPF_JSGT:
1678 				/* Signed '>', GT in x86 */
1679 				jmp_cond = X86_JG;
1680 				break;
1681 			case BPF_JSLT:
1682 				/* Signed '<', LT in x86 */
1683 				jmp_cond = X86_JL;
1684 				break;
1685 			case BPF_JSGE:
1686 				/* Signed '>=', GE in x86 */
1687 				jmp_cond = X86_JGE;
1688 				break;
1689 			case BPF_JSLE:
1690 				/* Signed '<=', LE in x86 */
1691 				jmp_cond = X86_JLE;
1692 				break;
1693 			default: /* to silence GCC warning */
1694 				return -EFAULT;
1695 			}
1696 			jmp_offset = addrs[i + insn->off] - addrs[i];
1697 			if (is_imm8(jmp_offset)) {
1698 				if (jmp_padding) {
1699 					/* To keep the jmp_offset valid, the extra bytes are
1700 					 * padded before the jump insn, so we subtract the
1701 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1702 					 *
1703 					 * If the previous pass already emits an imm8
1704 					 * jmp_cond, then this BPF insn won't shrink, so
1705 					 * "nops" is 0.
1706 					 *
1707 					 * On the other hand, if the previous pass emits an
1708 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1709 					 * keep the image from shrinking further.
1710 					 *
1711 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1712 					 *     is 2 bytes, so the size difference is 4 bytes.
1713 					 */
1714 					nops = INSN_SZ_DIFF - 2;
1715 					if (nops != 0 && nops != 4) {
1716 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1717 						       nops);
1718 						return -EFAULT;
1719 					}
1720 					emit_nops(&prog, nops);
1721 				}
1722 				EMIT2(jmp_cond, jmp_offset);
1723 			} else if (is_simm32(jmp_offset)) {
1724 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1725 			} else {
1726 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1727 				return -EFAULT;
1728 			}
1729 
1730 			break;
1731 
1732 		case BPF_JMP | BPF_JA:
1733 			if (insn->off == -1)
1734 				/* -1 jmp instructions will always jump
1735 				 * backwards two bytes. Explicitly handling
1736 				 * this case avoids wasting too many passes
1737 				 * when there are long sequences of replaced
1738 				 * dead code.
1739 				 */
1740 				jmp_offset = -2;
1741 			else
1742 				jmp_offset = addrs[i + insn->off] - addrs[i];
1743 
1744 			if (!jmp_offset) {
1745 				/*
1746 				 * If jmp_padding is enabled, the extra nops will
1747 				 * be inserted. Otherwise, optimize out nop jumps.
1748 				 */
1749 				if (jmp_padding) {
1750 					/* There are 3 possible conditions.
1751 					 * (1) This BPF_JA is already optimized out in
1752 					 *     the previous run, so there is no need
1753 					 *     to pad any extra byte (0 byte).
1754 					 * (2) The previous pass emits an imm8 jmp,
1755 					 *     so we pad 2 bytes to match the previous
1756 					 *     insn size.
1757 					 * (3) Similarly, the previous pass emits an
1758 					 *     imm32 jmp, and 5 bytes is padded.
1759 					 */
1760 					nops = INSN_SZ_DIFF;
1761 					if (nops != 0 && nops != 2 && nops != 5) {
1762 						pr_err("unexpected nop jump padding: %d bytes\n",
1763 						       nops);
1764 						return -EFAULT;
1765 					}
1766 					emit_nops(&prog, nops);
1767 				}
1768 				break;
1769 			}
1770 emit_jmp:
1771 			if (is_imm8(jmp_offset)) {
1772 				if (jmp_padding) {
1773 					/* To avoid breaking jmp_offset, the extra bytes
1774 					 * are padded before the actual jmp insn, so
1775 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1776 					 *
1777 					 * If the previous pass already emits an imm8
1778 					 * jmp, there is nothing to pad (0 byte).
1779 					 *
1780 					 * If it emits an imm32 jmp (5 bytes) previously
1781 					 * and now an imm8 jmp (2 bytes), then we pad
1782 					 * (5 - 2 = 3) bytes to stop the image from
1783 					 * shrinking further.
1784 					 */
1785 					nops = INSN_SZ_DIFF - 2;
1786 					if (nops != 0 && nops != 3) {
1787 						pr_err("unexpected jump padding: %d bytes\n",
1788 						       nops);
1789 						return -EFAULT;
1790 					}
1791 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1792 				}
1793 				EMIT2(0xEB, jmp_offset);
1794 			} else if (is_simm32(jmp_offset)) {
1795 				EMIT1_off32(0xE9, jmp_offset);
1796 			} else {
1797 				pr_err("jmp gen bug %llx\n", jmp_offset);
1798 				return -EFAULT;
1799 			}
1800 			break;
1801 
1802 		case BPF_JMP | BPF_EXIT:
1803 			if (seen_exit) {
1804 				jmp_offset = ctx->cleanup_addr - addrs[i];
1805 				goto emit_jmp;
1806 			}
1807 			seen_exit = true;
1808 			/* Update cleanup_addr */
1809 			ctx->cleanup_addr = proglen;
1810 			pop_callee_regs(&prog, callee_regs_used);
1811 			EMIT1(0xC9);         /* leave */
1812 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1813 			break;
1814 
1815 		default:
1816 			/*
1817 			 * By design x86-64 JIT should support all BPF instructions.
1818 			 * This error will be seen if new instruction was added
1819 			 * to the interpreter, but not to the JIT, or if there is
1820 			 * junk in bpf_prog.
1821 			 */
1822 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1823 			return -EINVAL;
1824 		}
1825 
1826 		ilen = prog - temp;
1827 		if (ilen > BPF_MAX_INSN_SIZE) {
1828 			pr_err("bpf_jit: fatal insn size error\n");
1829 			return -EFAULT;
1830 		}
1831 
1832 		if (image) {
1833 			/*
1834 			 * When populating the image, assert that:
1835 			 *
1836 			 *  i) We do not write beyond the allocated space, and
1837 			 * ii) addrs[i] did not change from the prior run, in order
1838 			 *     to validate assumptions made for computing branch
1839 			 *     displacements.
1840 			 */
1841 			if (unlikely(proglen + ilen > oldproglen ||
1842 				     proglen + ilen != addrs[i])) {
1843 				pr_err("bpf_jit: fatal error\n");
1844 				return -EFAULT;
1845 			}
1846 			memcpy(rw_image + proglen, temp, ilen);
1847 		}
1848 		proglen += ilen;
1849 		addrs[i] = proglen;
1850 		prog = temp;
1851 	}
1852 
1853 	if (image && excnt != bpf_prog->aux->num_exentries) {
1854 		pr_err("extable is not populated\n");
1855 		return -EFAULT;
1856 	}
1857 	return proglen;
1858 }
1859 
1860 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_regs,
1861 		      int stack_size)
1862 {
1863 	int i, j, arg_size;
1864 	bool next_same_struct = false;
1865 
1866 	/* Store function arguments to stack.
1867 	 * For a function that accepts two pointers the sequence will be:
1868 	 * mov QWORD PTR [rbp-0x10],rdi
1869 	 * mov QWORD PTR [rbp-0x8],rsi
1870 	 */
1871 	for (i = 0, j = 0; i < min(nr_regs, 6); i++) {
1872 		/* The arg_size is at most 16 bytes, enforced by the verifier. */
1873 		arg_size = m->arg_size[j];
1874 		if (arg_size > 8) {
1875 			arg_size = 8;
1876 			next_same_struct = !next_same_struct;
1877 		}
1878 
1879 		emit_stx(prog, bytes_to_bpf_size(arg_size),
1880 			 BPF_REG_FP,
1881 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1882 			 -(stack_size - i * 8));
1883 
1884 		j = next_same_struct ? j : j + 1;
1885 	}
1886 }
1887 
1888 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_regs,
1889 			 int stack_size)
1890 {
1891 	int i, j, arg_size;
1892 	bool next_same_struct = false;
1893 
1894 	/* Restore function arguments from stack.
1895 	 * For a function that accepts two pointers the sequence will be:
1896 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1897 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1898 	 */
1899 	for (i = 0, j = 0; i < min(nr_regs, 6); i++) {
1900 		/* The arg_size is at most 16 bytes, enforced by the verifier. */
1901 		arg_size = m->arg_size[j];
1902 		if (arg_size > 8) {
1903 			arg_size = 8;
1904 			next_same_struct = !next_same_struct;
1905 		}
1906 
1907 		emit_ldx(prog, bytes_to_bpf_size(arg_size),
1908 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1909 			 BPF_REG_FP,
1910 			 -(stack_size - i * 8));
1911 
1912 		j = next_same_struct ? j : j + 1;
1913 	}
1914 }
1915 
1916 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1917 			   struct bpf_tramp_link *l, int stack_size,
1918 			   int run_ctx_off, bool save_ret)
1919 {
1920 	u8 *prog = *pprog;
1921 	u8 *jmp_insn;
1922 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1923 	struct bpf_prog *p = l->link.prog;
1924 	u64 cookie = l->cookie;
1925 
1926 	/* mov rdi, cookie */
1927 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
1928 
1929 	/* Prepare struct bpf_tramp_run_ctx.
1930 	 *
1931 	 * bpf_tramp_run_ctx is already preserved by
1932 	 * arch_prepare_bpf_trampoline().
1933 	 *
1934 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
1935 	 */
1936 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
1937 
1938 	/* arg1: mov rdi, progs[i] */
1939 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1940 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
1941 	EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
1942 
1943 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), prog))
1944 		return -EINVAL;
1945 	/* remember prog start time returned by __bpf_prog_enter */
1946 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1947 
1948 	/* if (__bpf_prog_enter*(prog) == 0)
1949 	 *	goto skip_exec_of_prog;
1950 	 */
1951 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1952 	/* emit 2 nops that will be replaced with JE insn */
1953 	jmp_insn = prog;
1954 	emit_nops(&prog, 2);
1955 
1956 	/* arg1: lea rdi, [rbp - stack_size] */
1957 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1958 	/* arg2: progs[i]->insnsi for interpreter */
1959 	if (!p->jited)
1960 		emit_mov_imm64(&prog, BPF_REG_2,
1961 			       (long) p->insnsi >> 32,
1962 			       (u32) (long) p->insnsi);
1963 	/* call JITed bpf program or interpreter */
1964 	if (emit_rsb_call(&prog, p->bpf_func, prog))
1965 		return -EINVAL;
1966 
1967 	/*
1968 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1969 	 * of the previous call which is then passed on the stack to
1970 	 * the next BPF program.
1971 	 *
1972 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1973 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1974 	 */
1975 	if (save_ret)
1976 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1977 
1978 	/* replace 2 nops with JE insn, since jmp target is known */
1979 	jmp_insn[0] = X86_JE;
1980 	jmp_insn[1] = prog - jmp_insn - 2;
1981 
1982 	/* arg1: mov rdi, progs[i] */
1983 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1984 	/* arg2: mov rsi, rbx <- start time in nsec */
1985 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1986 	/* arg3: lea rdx, [rbp - run_ctx_off] */
1987 	EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
1988 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), prog))
1989 		return -EINVAL;
1990 
1991 	*pprog = prog;
1992 	return 0;
1993 }
1994 
1995 static void emit_align(u8 **pprog, u32 align)
1996 {
1997 	u8 *target, *prog = *pprog;
1998 
1999 	target = PTR_ALIGN(prog, align);
2000 	if (target != prog)
2001 		emit_nops(&prog, target - prog);
2002 
2003 	*pprog = prog;
2004 }
2005 
2006 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2007 {
2008 	u8 *prog = *pprog;
2009 	s64 offset;
2010 
2011 	offset = func - (ip + 2 + 4);
2012 	if (!is_simm32(offset)) {
2013 		pr_err("Target %p is out of range\n", func);
2014 		return -EINVAL;
2015 	}
2016 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2017 	*pprog = prog;
2018 	return 0;
2019 }
2020 
2021 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2022 		      struct bpf_tramp_links *tl, int stack_size,
2023 		      int run_ctx_off, bool save_ret)
2024 {
2025 	int i;
2026 	u8 *prog = *pprog;
2027 
2028 	for (i = 0; i < tl->nr_links; i++) {
2029 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2030 				    run_ctx_off, save_ret))
2031 			return -EINVAL;
2032 	}
2033 	*pprog = prog;
2034 	return 0;
2035 }
2036 
2037 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2038 			      struct bpf_tramp_links *tl, int stack_size,
2039 			      int run_ctx_off, u8 **branches)
2040 {
2041 	u8 *prog = *pprog;
2042 	int i;
2043 
2044 	/* The first fmod_ret program will receive a garbage return value.
2045 	 * Set this to 0 to avoid confusing the program.
2046 	 */
2047 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2048 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2049 	for (i = 0; i < tl->nr_links; i++) {
2050 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
2051 			return -EINVAL;
2052 
2053 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2054 		 * if (*(u64 *)(rbp - 8) !=  0)
2055 		 *	goto do_fexit;
2056 		 */
2057 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2058 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2059 
2060 		/* Save the location of the branch and Generate 6 nops
2061 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2062 		 * are replaced with a conditional jump once do_fexit (i.e. the
2063 		 * start of the fexit invocation) is finalized.
2064 		 */
2065 		branches[i] = prog;
2066 		emit_nops(&prog, 4 + 2);
2067 	}
2068 
2069 	*pprog = prog;
2070 	return 0;
2071 }
2072 
2073 /* Example:
2074  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2075  * its 'struct btf_func_model' will be nr_args=2
2076  * The assembly code when eth_type_trans is executing after trampoline:
2077  *
2078  * push rbp
2079  * mov rbp, rsp
2080  * sub rsp, 16                     // space for skb and dev
2081  * push rbx                        // temp regs to pass start time
2082  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2083  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2084  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2085  * mov rbx, rax                    // remember start time in bpf stats are enabled
2086  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2087  * call addr_of_jited_FENTRY_prog
2088  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2089  * mov rsi, rbx                    // prog start time
2090  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2091  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2092  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2093  * pop rbx
2094  * leave
2095  * ret
2096  *
2097  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2098  * replaced with 'call generated_bpf_trampoline'. When it returns
2099  * eth_type_trans will continue executing with original skb and dev pointers.
2100  *
2101  * The assembly code when eth_type_trans is called from trampoline:
2102  *
2103  * push rbp
2104  * mov rbp, rsp
2105  * sub rsp, 24                     // space for skb, dev, return value
2106  * push rbx                        // temp regs to pass start time
2107  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2108  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2109  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2110  * mov rbx, rax                    // remember start time if bpf stats are enabled
2111  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2112  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2113  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2114  * mov rsi, rbx                    // prog start time
2115  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2116  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2117  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2118  * call eth_type_trans+5           // execute body of eth_type_trans
2119  * mov qword ptr [rbp - 8], rax    // save return value
2120  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2121  * mov rbx, rax                    // remember start time in bpf stats are enabled
2122  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2123  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2124  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2125  * mov rsi, rbx                    // prog start time
2126  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2127  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2128  * pop rbx
2129  * leave
2130  * add rsp, 8                      // skip eth_type_trans's frame
2131  * ret                             // return to its caller
2132  */
2133 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2134 				const struct btf_func_model *m, u32 flags,
2135 				struct bpf_tramp_links *tlinks,
2136 				void *func_addr)
2137 {
2138 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2139 	int regs_off, nregs_off, ip_off, run_ctx_off;
2140 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2141 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2142 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2143 	void *orig_call = func_addr;
2144 	u8 **branches = NULL;
2145 	u8 *prog;
2146 	bool save_ret;
2147 
2148 	/* extra registers for struct arguments */
2149 	for (i = 0; i < m->nr_args; i++)
2150 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2151 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2152 
2153 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2154 	if (nr_regs > 6)
2155 		return -ENOTSUPP;
2156 
2157 	/* Generated trampoline stack layout:
2158 	 *
2159 	 * RBP + 8         [ return address  ]
2160 	 * RBP + 0         [ RBP             ]
2161 	 *
2162 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2163 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2164 	 *
2165 	 *                 [ reg_argN        ]  always
2166 	 *                 [ ...             ]
2167 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2168 	 *
2169 	 * RBP - nregs_off [ regs count	     ]  always
2170 	 *
2171 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2172 	 *
2173 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2174 	 */
2175 
2176 	/* room for return value of orig_call or fentry prog */
2177 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2178 	if (save_ret)
2179 		stack_size += 8;
2180 
2181 	stack_size += nr_regs * 8;
2182 	regs_off = stack_size;
2183 
2184 	/* regs count  */
2185 	stack_size += 8;
2186 	nregs_off = stack_size;
2187 
2188 	if (flags & BPF_TRAMP_F_IP_ARG)
2189 		stack_size += 8; /* room for IP address argument */
2190 
2191 	ip_off = stack_size;
2192 
2193 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2194 	run_ctx_off = stack_size;
2195 
2196 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2197 		/* skip patched call instruction and point orig_call to actual
2198 		 * body of the kernel function.
2199 		 */
2200 		if (is_endbr(*(u32 *)orig_call))
2201 			orig_call += ENDBR_INSN_SIZE;
2202 		orig_call += X86_PATCH_SIZE;
2203 	}
2204 
2205 	prog = image;
2206 
2207 	EMIT_ENDBR();
2208 	/*
2209 	 * This is the direct-call trampoline, as such it needs accounting
2210 	 * for the __fentry__ call.
2211 	 */
2212 	x86_call_depth_emit_accounting(&prog, NULL);
2213 	EMIT1(0x55);		 /* push rbp */
2214 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2215 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2216 	EMIT1(0x53);		 /* push rbx */
2217 
2218 	/* Store number of argument registers of the traced function:
2219 	 *   mov rax, nr_regs
2220 	 *   mov QWORD PTR [rbp - nregs_off], rax
2221 	 */
2222 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2223 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2224 
2225 	if (flags & BPF_TRAMP_F_IP_ARG) {
2226 		/* Store IP address of the traced function:
2227 		 * movabsq rax, func_addr
2228 		 * mov QWORD PTR [rbp - ip_off], rax
2229 		 */
2230 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2231 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2232 	}
2233 
2234 	save_regs(m, &prog, nr_regs, regs_off);
2235 
2236 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2237 		/* arg1: mov rdi, im */
2238 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2239 		if (emit_rsb_call(&prog, __bpf_tramp_enter, prog)) {
2240 			ret = -EINVAL;
2241 			goto cleanup;
2242 		}
2243 	}
2244 
2245 	if (fentry->nr_links)
2246 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2247 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2248 			return -EINVAL;
2249 
2250 	if (fmod_ret->nr_links) {
2251 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2252 				   GFP_KERNEL);
2253 		if (!branches)
2254 			return -ENOMEM;
2255 
2256 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2257 				       run_ctx_off, branches)) {
2258 			ret = -EINVAL;
2259 			goto cleanup;
2260 		}
2261 	}
2262 
2263 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2264 		restore_regs(m, &prog, nr_regs, regs_off);
2265 
2266 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2267 			emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2268 			EMIT2(0xff, 0xd0); /* call *rax */
2269 		} else {
2270 			/* call original function */
2271 			if (emit_rsb_call(&prog, orig_call, prog)) {
2272 				ret = -EINVAL;
2273 				goto cleanup;
2274 			}
2275 		}
2276 		/* remember return value in a stack for bpf prog to access */
2277 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2278 		im->ip_after_call = prog;
2279 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2280 		prog += X86_PATCH_SIZE;
2281 	}
2282 
2283 	if (fmod_ret->nr_links) {
2284 		/* From Intel 64 and IA-32 Architectures Optimization
2285 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2286 		 * Coding Rule 11: All branch targets should be 16-byte
2287 		 * aligned.
2288 		 */
2289 		emit_align(&prog, 16);
2290 		/* Update the branches saved in invoke_bpf_mod_ret with the
2291 		 * aligned address of do_fexit.
2292 		 */
2293 		for (i = 0; i < fmod_ret->nr_links; i++)
2294 			emit_cond_near_jump(&branches[i], prog, branches[i],
2295 					    X86_JNE);
2296 	}
2297 
2298 	if (fexit->nr_links)
2299 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2300 			ret = -EINVAL;
2301 			goto cleanup;
2302 		}
2303 
2304 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2305 		restore_regs(m, &prog, nr_regs, regs_off);
2306 
2307 	/* This needs to be done regardless. If there were fmod_ret programs,
2308 	 * the return value is only updated on the stack and still needs to be
2309 	 * restored to R0.
2310 	 */
2311 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2312 		im->ip_epilogue = prog;
2313 		/* arg1: mov rdi, im */
2314 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2315 		if (emit_rsb_call(&prog, __bpf_tramp_exit, prog)) {
2316 			ret = -EINVAL;
2317 			goto cleanup;
2318 		}
2319 	}
2320 	/* restore return value of orig_call or fentry prog back into RAX */
2321 	if (save_ret)
2322 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2323 
2324 	EMIT1(0x5B); /* pop rbx */
2325 	EMIT1(0xC9); /* leave */
2326 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2327 		/* skip our return address and return to parent */
2328 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2329 	emit_return(&prog, prog);
2330 	/* Make sure the trampoline generation logic doesn't overflow */
2331 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2332 		ret = -EFAULT;
2333 		goto cleanup;
2334 	}
2335 	ret = prog - (u8 *)image;
2336 
2337 cleanup:
2338 	kfree(branches);
2339 	return ret;
2340 }
2341 
2342 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2343 {
2344 	u8 *jg_reloc, *prog = *pprog;
2345 	int pivot, err, jg_bytes = 1;
2346 	s64 jg_offset;
2347 
2348 	if (a == b) {
2349 		/* Leaf node of recursion, i.e. not a range of indices
2350 		 * anymore.
2351 		 */
2352 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2353 		if (!is_simm32(progs[a]))
2354 			return -1;
2355 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2356 			    progs[a]);
2357 		err = emit_cond_near_jump(&prog,	/* je func */
2358 					  (void *)progs[a], image + (prog - buf),
2359 					  X86_JE);
2360 		if (err)
2361 			return err;
2362 
2363 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2364 
2365 		*pprog = prog;
2366 		return 0;
2367 	}
2368 
2369 	/* Not a leaf node, so we pivot, and recursively descend into
2370 	 * the lower and upper ranges.
2371 	 */
2372 	pivot = (b - a) / 2;
2373 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2374 	if (!is_simm32(progs[a + pivot]))
2375 		return -1;
2376 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2377 
2378 	if (pivot > 2) {				/* jg upper_part */
2379 		/* Require near jump. */
2380 		jg_bytes = 4;
2381 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2382 	} else {
2383 		EMIT2(X86_JG, 0);
2384 	}
2385 	jg_reloc = prog;
2386 
2387 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2388 				  progs, image, buf);
2389 	if (err)
2390 		return err;
2391 
2392 	/* From Intel 64 and IA-32 Architectures Optimization
2393 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2394 	 * Coding Rule 11: All branch targets should be 16-byte
2395 	 * aligned.
2396 	 */
2397 	emit_align(&prog, 16);
2398 	jg_offset = prog - jg_reloc;
2399 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2400 
2401 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2402 				  b, progs, image, buf);
2403 	if (err)
2404 		return err;
2405 
2406 	*pprog = prog;
2407 	return 0;
2408 }
2409 
2410 static int cmp_ips(const void *a, const void *b)
2411 {
2412 	const s64 *ipa = a;
2413 	const s64 *ipb = b;
2414 
2415 	if (*ipa > *ipb)
2416 		return 1;
2417 	if (*ipa < *ipb)
2418 		return -1;
2419 	return 0;
2420 }
2421 
2422 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2423 {
2424 	u8 *prog = buf;
2425 
2426 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2427 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2428 }
2429 
2430 struct x64_jit_data {
2431 	struct bpf_binary_header *rw_header;
2432 	struct bpf_binary_header *header;
2433 	int *addrs;
2434 	u8 *image;
2435 	int proglen;
2436 	struct jit_context ctx;
2437 };
2438 
2439 #define MAX_PASSES 20
2440 #define PADDING_PASSES (MAX_PASSES - 5)
2441 
2442 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2443 {
2444 	struct bpf_binary_header *rw_header = NULL;
2445 	struct bpf_binary_header *header = NULL;
2446 	struct bpf_prog *tmp, *orig_prog = prog;
2447 	struct x64_jit_data *jit_data;
2448 	int proglen, oldproglen = 0;
2449 	struct jit_context ctx = {};
2450 	bool tmp_blinded = false;
2451 	bool extra_pass = false;
2452 	bool padding = false;
2453 	u8 *rw_image = NULL;
2454 	u8 *image = NULL;
2455 	int *addrs;
2456 	int pass;
2457 	int i;
2458 
2459 	if (!prog->jit_requested)
2460 		return orig_prog;
2461 
2462 	tmp = bpf_jit_blind_constants(prog);
2463 	/*
2464 	 * If blinding was requested and we failed during blinding,
2465 	 * we must fall back to the interpreter.
2466 	 */
2467 	if (IS_ERR(tmp))
2468 		return orig_prog;
2469 	if (tmp != prog) {
2470 		tmp_blinded = true;
2471 		prog = tmp;
2472 	}
2473 
2474 	jit_data = prog->aux->jit_data;
2475 	if (!jit_data) {
2476 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2477 		if (!jit_data) {
2478 			prog = orig_prog;
2479 			goto out;
2480 		}
2481 		prog->aux->jit_data = jit_data;
2482 	}
2483 	addrs = jit_data->addrs;
2484 	if (addrs) {
2485 		ctx = jit_data->ctx;
2486 		oldproglen = jit_data->proglen;
2487 		image = jit_data->image;
2488 		header = jit_data->header;
2489 		rw_header = jit_data->rw_header;
2490 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2491 		extra_pass = true;
2492 		padding = true;
2493 		goto skip_init_addrs;
2494 	}
2495 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2496 	if (!addrs) {
2497 		prog = orig_prog;
2498 		goto out_addrs;
2499 	}
2500 
2501 	/*
2502 	 * Before first pass, make a rough estimation of addrs[]
2503 	 * each BPF instruction is translated to less than 64 bytes
2504 	 */
2505 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2506 		proglen += 64;
2507 		addrs[i] = proglen;
2508 	}
2509 	ctx.cleanup_addr = proglen;
2510 skip_init_addrs:
2511 
2512 	/*
2513 	 * JITed image shrinks with every pass and the loop iterates
2514 	 * until the image stops shrinking. Very large BPF programs
2515 	 * may converge on the last pass. In such case do one more
2516 	 * pass to emit the final image.
2517 	 */
2518 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2519 		if (!padding && pass >= PADDING_PASSES)
2520 			padding = true;
2521 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2522 		if (proglen <= 0) {
2523 out_image:
2524 			image = NULL;
2525 			if (header) {
2526 				bpf_arch_text_copy(&header->size, &rw_header->size,
2527 						   sizeof(rw_header->size));
2528 				bpf_jit_binary_pack_free(header, rw_header);
2529 			}
2530 			/* Fall back to interpreter mode */
2531 			prog = orig_prog;
2532 			if (extra_pass) {
2533 				prog->bpf_func = NULL;
2534 				prog->jited = 0;
2535 				prog->jited_len = 0;
2536 			}
2537 			goto out_addrs;
2538 		}
2539 		if (image) {
2540 			if (proglen != oldproglen) {
2541 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2542 				       proglen, oldproglen);
2543 				goto out_image;
2544 			}
2545 			break;
2546 		}
2547 		if (proglen == oldproglen) {
2548 			/*
2549 			 * The number of entries in extable is the number of BPF_LDX
2550 			 * insns that access kernel memory via "pointer to BTF type".
2551 			 * The verifier changed their opcode from LDX|MEM|size
2552 			 * to LDX|PROBE_MEM|size to make JITing easier.
2553 			 */
2554 			u32 align = __alignof__(struct exception_table_entry);
2555 			u32 extable_size = prog->aux->num_exentries *
2556 				sizeof(struct exception_table_entry);
2557 
2558 			/* allocate module memory for x86 insns and extable */
2559 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2560 							   &image, align, &rw_header, &rw_image,
2561 							   jit_fill_hole);
2562 			if (!header) {
2563 				prog = orig_prog;
2564 				goto out_addrs;
2565 			}
2566 			prog->aux->extable = (void *) image + roundup(proglen, align);
2567 		}
2568 		oldproglen = proglen;
2569 		cond_resched();
2570 	}
2571 
2572 	if (bpf_jit_enable > 1)
2573 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
2574 
2575 	if (image) {
2576 		if (!prog->is_func || extra_pass) {
2577 			/*
2578 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2579 			 *   1) header is not pointing to proper module memory;
2580 			 *   2) the arch doesn't support bpf_arch_text_copy().
2581 			 *
2582 			 * Both cases are serious bugs and justify WARN_ON.
2583 			 */
2584 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2585 				/* header has been freed */
2586 				header = NULL;
2587 				goto out_image;
2588 			}
2589 
2590 			bpf_tail_call_direct_fixup(prog);
2591 		} else {
2592 			jit_data->addrs = addrs;
2593 			jit_data->ctx = ctx;
2594 			jit_data->proglen = proglen;
2595 			jit_data->image = image;
2596 			jit_data->header = header;
2597 			jit_data->rw_header = rw_header;
2598 		}
2599 		prog->bpf_func = (void *)image;
2600 		prog->jited = 1;
2601 		prog->jited_len = proglen;
2602 	} else {
2603 		prog = orig_prog;
2604 	}
2605 
2606 	if (!image || !prog->is_func || extra_pass) {
2607 		if (image)
2608 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2609 out_addrs:
2610 		kvfree(addrs);
2611 		kfree(jit_data);
2612 		prog->aux->jit_data = NULL;
2613 	}
2614 out:
2615 	if (tmp_blinded)
2616 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2617 					   tmp : orig_prog);
2618 	return prog;
2619 }
2620 
2621 bool bpf_jit_supports_kfunc_call(void)
2622 {
2623 	return true;
2624 }
2625 
2626 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2627 {
2628 	if (text_poke_copy(dst, src, len) == NULL)
2629 		return ERR_PTR(-EINVAL);
2630 	return dst;
2631 }
2632 
2633 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2634 bool bpf_jit_supports_subprog_tailcalls(void)
2635 {
2636 	return true;
2637 }
2638 
2639 void bpf_jit_free(struct bpf_prog *prog)
2640 {
2641 	if (prog->jited) {
2642 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2643 		struct bpf_binary_header *hdr;
2644 
2645 		/*
2646 		 * If we fail the final pass of JIT (from jit_subprogs),
2647 		 * the program may not be finalized yet. Call finalize here
2648 		 * before freeing it.
2649 		 */
2650 		if (jit_data) {
2651 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2652 						     jit_data->rw_header);
2653 			kvfree(jit_data->addrs);
2654 			kfree(jit_data);
2655 		}
2656 		hdr = bpf_jit_binary_pack_hdr(prog);
2657 		bpf_jit_binary_pack_free(hdr, NULL);
2658 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2659 	}
2660 
2661 	bpf_prog_unlock_free(prog);
2662 }
2663