xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision de4eda9d)
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/set_memory.h>
16 #include <asm/nospec-branch.h>
17 #include <asm/text-patching.h>
18 
19 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
20 {
21 	if (len == 1)
22 		*ptr = bytes;
23 	else if (len == 2)
24 		*(u16 *)ptr = bytes;
25 	else {
26 		*(u32 *)ptr = bytes;
27 		barrier();
28 	}
29 	return ptr + len;
30 }
31 
32 #define EMIT(bytes, len) \
33 	do { prog = emit_code(prog, bytes, len); } while (0)
34 
35 #define EMIT1(b1)		EMIT(b1, 1)
36 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
37 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
38 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
39 
40 #define EMIT1_off32(b1, off) \
41 	do { EMIT1(b1); EMIT(off, 4); } while (0)
42 #define EMIT2_off32(b1, b2, off) \
43 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
44 #define EMIT3_off32(b1, b2, b3, off) \
45 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
46 #define EMIT4_off32(b1, b2, b3, b4, off) \
47 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
48 
49 #ifdef CONFIG_X86_KERNEL_IBT
50 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
51 #else
52 #define EMIT_ENDBR()
53 #endif
54 
55 static bool is_imm8(int value)
56 {
57 	return value <= 127 && value >= -128;
58 }
59 
60 static bool is_simm32(s64 value)
61 {
62 	return value == (s64)(s32)value;
63 }
64 
65 static bool is_uimm32(u64 value)
66 {
67 	return value == (u64)(u32)value;
68 }
69 
70 /* mov dst, src */
71 #define EMIT_mov(DST, SRC)								 \
72 	do {										 \
73 		if (DST != SRC)								 \
74 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
75 	} while (0)
76 
77 static int bpf_size_to_x86_bytes(int bpf_size)
78 {
79 	if (bpf_size == BPF_W)
80 		return 4;
81 	else if (bpf_size == BPF_H)
82 		return 2;
83 	else if (bpf_size == BPF_B)
84 		return 1;
85 	else if (bpf_size == BPF_DW)
86 		return 4; /* imm32 */
87 	else
88 		return 0;
89 }
90 
91 /*
92  * List of x86 cond jumps opcodes (. + s8)
93  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
94  */
95 #define X86_JB  0x72
96 #define X86_JAE 0x73
97 #define X86_JE  0x74
98 #define X86_JNE 0x75
99 #define X86_JBE 0x76
100 #define X86_JA  0x77
101 #define X86_JL  0x7C
102 #define X86_JGE 0x7D
103 #define X86_JLE 0x7E
104 #define X86_JG  0x7F
105 
106 /* Pick a register outside of BPF range for JIT internal work */
107 #define AUX_REG (MAX_BPF_JIT_REG + 1)
108 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
109 
110 /*
111  * The following table maps BPF registers to x86-64 registers.
112  *
113  * x86-64 register R12 is unused, since if used as base address
114  * register in load/store instructions, it always needs an
115  * extra byte of encoding and is callee saved.
116  *
117  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
118  * trampoline. x86-64 register R10 is used for blinding (if enabled).
119  */
120 static const int reg2hex[] = {
121 	[BPF_REG_0] = 0,  /* RAX */
122 	[BPF_REG_1] = 7,  /* RDI */
123 	[BPF_REG_2] = 6,  /* RSI */
124 	[BPF_REG_3] = 2,  /* RDX */
125 	[BPF_REG_4] = 1,  /* RCX */
126 	[BPF_REG_5] = 0,  /* R8  */
127 	[BPF_REG_6] = 3,  /* RBX callee saved */
128 	[BPF_REG_7] = 5,  /* R13 callee saved */
129 	[BPF_REG_8] = 6,  /* R14 callee saved */
130 	[BPF_REG_9] = 7,  /* R15 callee saved */
131 	[BPF_REG_FP] = 5, /* RBP readonly */
132 	[BPF_REG_AX] = 2, /* R10 temp register */
133 	[AUX_REG] = 3,    /* R11 temp register */
134 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
135 };
136 
137 static const int reg2pt_regs[] = {
138 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
139 	[BPF_REG_1] = offsetof(struct pt_regs, di),
140 	[BPF_REG_2] = offsetof(struct pt_regs, si),
141 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
142 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
143 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
144 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
145 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
146 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
147 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
148 };
149 
150 /*
151  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
152  * which need extra byte of encoding.
153  * rax,rcx,...,rbp have simpler encoding
154  */
155 static bool is_ereg(u32 reg)
156 {
157 	return (1 << reg) & (BIT(BPF_REG_5) |
158 			     BIT(AUX_REG) |
159 			     BIT(BPF_REG_7) |
160 			     BIT(BPF_REG_8) |
161 			     BIT(BPF_REG_9) |
162 			     BIT(X86_REG_R9) |
163 			     BIT(BPF_REG_AX));
164 }
165 
166 /*
167  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
168  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
169  * of encoding. al,cl,dl,bl have simpler encoding.
170  */
171 static bool is_ereg_8l(u32 reg)
172 {
173 	return is_ereg(reg) ||
174 	    (1 << reg) & (BIT(BPF_REG_1) |
175 			  BIT(BPF_REG_2) |
176 			  BIT(BPF_REG_FP));
177 }
178 
179 static bool is_axreg(u32 reg)
180 {
181 	return reg == BPF_REG_0;
182 }
183 
184 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
185 static u8 add_1mod(u8 byte, u32 reg)
186 {
187 	if (is_ereg(reg))
188 		byte |= 1;
189 	return byte;
190 }
191 
192 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
193 {
194 	if (is_ereg(r1))
195 		byte |= 1;
196 	if (is_ereg(r2))
197 		byte |= 4;
198 	return byte;
199 }
200 
201 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
202 static u8 add_1reg(u8 byte, u32 dst_reg)
203 {
204 	return byte + reg2hex[dst_reg];
205 }
206 
207 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
208 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
209 {
210 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
211 }
212 
213 /* Some 1-byte opcodes for binary ALU operations */
214 static u8 simple_alu_opcodes[] = {
215 	[BPF_ADD] = 0x01,
216 	[BPF_SUB] = 0x29,
217 	[BPF_AND] = 0x21,
218 	[BPF_OR] = 0x09,
219 	[BPF_XOR] = 0x31,
220 	[BPF_LSH] = 0xE0,
221 	[BPF_RSH] = 0xE8,
222 	[BPF_ARSH] = 0xF8,
223 };
224 
225 static void jit_fill_hole(void *area, unsigned int size)
226 {
227 	/* Fill whole space with INT3 instructions */
228 	memset(area, 0xcc, size);
229 }
230 
231 int bpf_arch_text_invalidate(void *dst, size_t len)
232 {
233 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
234 }
235 
236 struct jit_context {
237 	int cleanup_addr; /* Epilogue code offset */
238 
239 	/*
240 	 * Program specific offsets of labels in the code; these rely on the
241 	 * JIT doing at least 2 passes, recording the position on the first
242 	 * pass, only to generate the correct offset on the second pass.
243 	 */
244 	int tail_call_direct_label;
245 	int tail_call_indirect_label;
246 };
247 
248 /* Maximum number of bytes emitted while JITing one eBPF insn */
249 #define BPF_MAX_INSN_SIZE	128
250 #define BPF_INSN_SAFETY		64
251 
252 /* Number of bytes emit_patch() needs to generate instructions */
253 #define X86_PATCH_SIZE		5
254 /* Number of bytes that will be skipped on tailcall */
255 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
256 
257 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
258 {
259 	u8 *prog = *pprog;
260 
261 	if (callee_regs_used[0])
262 		EMIT1(0x53);         /* push rbx */
263 	if (callee_regs_used[1])
264 		EMIT2(0x41, 0x55);   /* push r13 */
265 	if (callee_regs_used[2])
266 		EMIT2(0x41, 0x56);   /* push r14 */
267 	if (callee_regs_used[3])
268 		EMIT2(0x41, 0x57);   /* push r15 */
269 	*pprog = prog;
270 }
271 
272 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
273 {
274 	u8 *prog = *pprog;
275 
276 	if (callee_regs_used[3])
277 		EMIT2(0x41, 0x5F);   /* pop r15 */
278 	if (callee_regs_used[2])
279 		EMIT2(0x41, 0x5E);   /* pop r14 */
280 	if (callee_regs_used[1])
281 		EMIT2(0x41, 0x5D);   /* pop r13 */
282 	if (callee_regs_used[0])
283 		EMIT1(0x5B);         /* pop rbx */
284 	*pprog = prog;
285 }
286 
287 /*
288  * Emit x86-64 prologue code for BPF program.
289  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
290  * while jumping to another program
291  */
292 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
293 			  bool tail_call_reachable, bool is_subprog)
294 {
295 	u8 *prog = *pprog;
296 
297 	/* BPF trampoline can be made to work without these nops,
298 	 * but let's waste 5 bytes for now and optimize later
299 	 */
300 	EMIT_ENDBR();
301 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
302 	prog += X86_PATCH_SIZE;
303 	if (!ebpf_from_cbpf) {
304 		if (tail_call_reachable && !is_subprog)
305 			EMIT2(0x31, 0xC0); /* xor eax, eax */
306 		else
307 			EMIT2(0x66, 0x90); /* nop2 */
308 	}
309 	EMIT1(0x55);             /* push rbp */
310 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
311 
312 	/* X86_TAIL_CALL_OFFSET is here */
313 	EMIT_ENDBR();
314 
315 	/* sub rsp, rounded_stack_depth */
316 	if (stack_depth)
317 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
318 	if (tail_call_reachable)
319 		EMIT1(0x50);         /* push rax */
320 	*pprog = prog;
321 }
322 
323 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
324 {
325 	u8 *prog = *pprog;
326 	s64 offset;
327 
328 	offset = func - (ip + X86_PATCH_SIZE);
329 	if (!is_simm32(offset)) {
330 		pr_err("Target call %p is out of range\n", func);
331 		return -ERANGE;
332 	}
333 	EMIT1_off32(opcode, offset);
334 	*pprog = prog;
335 	return 0;
336 }
337 
338 static int emit_call(u8 **pprog, void *func, void *ip)
339 {
340 	return emit_patch(pprog, func, ip, 0xE8);
341 }
342 
343 static int emit_jump(u8 **pprog, void *func, void *ip)
344 {
345 	return emit_patch(pprog, func, ip, 0xE9);
346 }
347 
348 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
349 				void *old_addr, void *new_addr)
350 {
351 	const u8 *nop_insn = x86_nops[5];
352 	u8 old_insn[X86_PATCH_SIZE];
353 	u8 new_insn[X86_PATCH_SIZE];
354 	u8 *prog;
355 	int ret;
356 
357 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
358 	if (old_addr) {
359 		prog = old_insn;
360 		ret = t == BPF_MOD_CALL ?
361 		      emit_call(&prog, old_addr, ip) :
362 		      emit_jump(&prog, old_addr, ip);
363 		if (ret)
364 			return ret;
365 	}
366 
367 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
368 	if (new_addr) {
369 		prog = new_insn;
370 		ret = t == BPF_MOD_CALL ?
371 		      emit_call(&prog, new_addr, ip) :
372 		      emit_jump(&prog, new_addr, ip);
373 		if (ret)
374 			return ret;
375 	}
376 
377 	ret = -EBUSY;
378 	mutex_lock(&text_mutex);
379 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
380 		goto out;
381 	ret = 1;
382 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
383 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
384 		ret = 0;
385 	}
386 out:
387 	mutex_unlock(&text_mutex);
388 	return ret;
389 }
390 
391 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
392 		       void *old_addr, void *new_addr)
393 {
394 	if (!is_kernel_text((long)ip) &&
395 	    !is_bpf_text_address((long)ip))
396 		/* BPF poking in modules is not supported */
397 		return -EINVAL;
398 
399 	/*
400 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
401 	 * with an ENDBR instruction.
402 	 */
403 	if (is_endbr(*(u32 *)ip))
404 		ip += ENDBR_INSN_SIZE;
405 
406 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
407 }
408 
409 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
410 
411 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
412 {
413 	u8 *prog = *pprog;
414 
415 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
416 		EMIT_LFENCE();
417 		EMIT2(0xFF, 0xE0 + reg);
418 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
419 		OPTIMIZER_HIDE_VAR(reg);
420 		emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
421 	} else {
422 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
423 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
424 			EMIT1(0xCC);		/* int3 */
425 	}
426 
427 	*pprog = prog;
428 }
429 
430 static void emit_return(u8 **pprog, u8 *ip)
431 {
432 	u8 *prog = *pprog;
433 
434 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
435 		emit_jump(&prog, &__x86_return_thunk, ip);
436 	} else {
437 		EMIT1(0xC3);		/* ret */
438 		if (IS_ENABLED(CONFIG_SLS))
439 			EMIT1(0xCC);	/* int3 */
440 	}
441 
442 	*pprog = prog;
443 }
444 
445 /*
446  * Generate the following code:
447  *
448  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
449  *   if (index >= array->map.max_entries)
450  *     goto out;
451  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
452  *     goto out;
453  *   prog = array->ptrs[index];
454  *   if (prog == NULL)
455  *     goto out;
456  *   goto *(prog->bpf_func + prologue_size);
457  * out:
458  */
459 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
460 					u32 stack_depth, u8 *ip,
461 					struct jit_context *ctx)
462 {
463 	int tcc_off = -4 - round_up(stack_depth, 8);
464 	u8 *prog = *pprog, *start = *pprog;
465 	int offset;
466 
467 	/*
468 	 * rdi - pointer to ctx
469 	 * rsi - pointer to bpf_array
470 	 * rdx - index in bpf_array
471 	 */
472 
473 	/*
474 	 * if (index >= array->map.max_entries)
475 	 *	goto out;
476 	 */
477 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
478 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
479 	      offsetof(struct bpf_array, map.max_entries));
480 
481 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
482 	EMIT2(X86_JBE, offset);                   /* jbe out */
483 
484 	/*
485 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
486 	 *	goto out;
487 	 */
488 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
489 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
490 
491 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
492 	EMIT2(X86_JAE, offset);                   /* jae out */
493 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
494 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
495 
496 	/* prog = array->ptrs[index]; */
497 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
498 		    offsetof(struct bpf_array, ptrs));
499 
500 	/*
501 	 * if (prog == NULL)
502 	 *	goto out;
503 	 */
504 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
505 
506 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
507 	EMIT2(X86_JE, offset);                    /* je out */
508 
509 	pop_callee_regs(&prog, callee_regs_used);
510 
511 	EMIT1(0x58);                              /* pop rax */
512 	if (stack_depth)
513 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
514 			    round_up(stack_depth, 8));
515 
516 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
517 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
518 	      offsetof(struct bpf_prog, bpf_func));
519 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
520 	      X86_TAIL_CALL_OFFSET);
521 	/*
522 	 * Now we're ready to jump into next BPF program
523 	 * rdi == ctx (1st arg)
524 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
525 	 */
526 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
527 
528 	/* out: */
529 	ctx->tail_call_indirect_label = prog - start;
530 	*pprog = prog;
531 }
532 
533 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
534 				      u8 **pprog, u8 *ip,
535 				      bool *callee_regs_used, u32 stack_depth,
536 				      struct jit_context *ctx)
537 {
538 	int tcc_off = -4 - round_up(stack_depth, 8);
539 	u8 *prog = *pprog, *start = *pprog;
540 	int offset;
541 
542 	/*
543 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
544 	 *	goto out;
545 	 */
546 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
547 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
548 
549 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
550 	EMIT2(X86_JAE, offset);                       /* jae out */
551 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
552 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
553 
554 	poke->tailcall_bypass = ip + (prog - start);
555 	poke->adj_off = X86_TAIL_CALL_OFFSET;
556 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
557 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
558 
559 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
560 		  poke->tailcall_bypass);
561 
562 	pop_callee_regs(&prog, callee_regs_used);
563 	EMIT1(0x58);                                  /* pop rax */
564 	if (stack_depth)
565 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
566 
567 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
568 	prog += X86_PATCH_SIZE;
569 
570 	/* out: */
571 	ctx->tail_call_direct_label = prog - start;
572 
573 	*pprog = prog;
574 }
575 
576 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
577 {
578 	struct bpf_jit_poke_descriptor *poke;
579 	struct bpf_array *array;
580 	struct bpf_prog *target;
581 	int i, ret;
582 
583 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
584 		poke = &prog->aux->poke_tab[i];
585 		if (poke->aux && poke->aux != prog->aux)
586 			continue;
587 
588 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
589 
590 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
591 			continue;
592 
593 		array = container_of(poke->tail_call.map, struct bpf_array, map);
594 		mutex_lock(&array->aux->poke_mutex);
595 		target = array->ptrs[poke->tail_call.key];
596 		if (target) {
597 			ret = __bpf_arch_text_poke(poke->tailcall_target,
598 						   BPF_MOD_JUMP, NULL,
599 						   (u8 *)target->bpf_func +
600 						   poke->adj_off);
601 			BUG_ON(ret < 0);
602 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
603 						   BPF_MOD_JUMP,
604 						   (u8 *)poke->tailcall_target +
605 						   X86_PATCH_SIZE, NULL);
606 			BUG_ON(ret < 0);
607 		}
608 		WRITE_ONCE(poke->tailcall_target_stable, true);
609 		mutex_unlock(&array->aux->poke_mutex);
610 	}
611 }
612 
613 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
614 			   u32 dst_reg, const u32 imm32)
615 {
616 	u8 *prog = *pprog;
617 	u8 b1, b2, b3;
618 
619 	/*
620 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
621 	 * (which zero-extends imm32) to save 2 bytes.
622 	 */
623 	if (sign_propagate && (s32)imm32 < 0) {
624 		/* 'mov %rax, imm32' sign extends imm32 */
625 		b1 = add_1mod(0x48, dst_reg);
626 		b2 = 0xC7;
627 		b3 = 0xC0;
628 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
629 		goto done;
630 	}
631 
632 	/*
633 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
634 	 * to save 3 bytes.
635 	 */
636 	if (imm32 == 0) {
637 		if (is_ereg(dst_reg))
638 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
639 		b2 = 0x31; /* xor */
640 		b3 = 0xC0;
641 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
642 		goto done;
643 	}
644 
645 	/* mov %eax, imm32 */
646 	if (is_ereg(dst_reg))
647 		EMIT1(add_1mod(0x40, dst_reg));
648 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
649 done:
650 	*pprog = prog;
651 }
652 
653 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
654 			   const u32 imm32_hi, const u32 imm32_lo)
655 {
656 	u8 *prog = *pprog;
657 
658 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
659 		/*
660 		 * For emitting plain u32, where sign bit must not be
661 		 * propagated LLVM tends to load imm64 over mov32
662 		 * directly, so save couple of bytes by just doing
663 		 * 'mov %eax, imm32' instead.
664 		 */
665 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
666 	} else {
667 		/* movabsq rax, imm64 */
668 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
669 		EMIT(imm32_lo, 4);
670 		EMIT(imm32_hi, 4);
671 	}
672 
673 	*pprog = prog;
674 }
675 
676 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
677 {
678 	u8 *prog = *pprog;
679 
680 	if (is64) {
681 		/* mov dst, src */
682 		EMIT_mov(dst_reg, src_reg);
683 	} else {
684 		/* mov32 dst, src */
685 		if (is_ereg(dst_reg) || is_ereg(src_reg))
686 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
687 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
688 	}
689 
690 	*pprog = prog;
691 }
692 
693 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
694 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
695 {
696 	u8 *prog = *pprog;
697 
698 	if (is_imm8(off)) {
699 		/* 1-byte signed displacement.
700 		 *
701 		 * If off == 0 we could skip this and save one extra byte, but
702 		 * special case of x86 R13 which always needs an offset is not
703 		 * worth the hassle
704 		 */
705 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
706 	} else {
707 		/* 4-byte signed displacement */
708 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
709 	}
710 	*pprog = prog;
711 }
712 
713 /*
714  * Emit a REX byte if it will be necessary to address these registers
715  */
716 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
717 {
718 	u8 *prog = *pprog;
719 
720 	if (is64)
721 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
722 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
723 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
724 	*pprog = prog;
725 }
726 
727 /*
728  * Similar version of maybe_emit_mod() for a single register
729  */
730 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
731 {
732 	u8 *prog = *pprog;
733 
734 	if (is64)
735 		EMIT1(add_1mod(0x48, reg));
736 	else if (is_ereg(reg))
737 		EMIT1(add_1mod(0x40, reg));
738 	*pprog = prog;
739 }
740 
741 /* LDX: dst_reg = *(u8*)(src_reg + off) */
742 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
743 {
744 	u8 *prog = *pprog;
745 
746 	switch (size) {
747 	case BPF_B:
748 		/* Emit 'movzx rax, byte ptr [rax + off]' */
749 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
750 		break;
751 	case BPF_H:
752 		/* Emit 'movzx rax, word ptr [rax + off]' */
753 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
754 		break;
755 	case BPF_W:
756 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
757 		if (is_ereg(dst_reg) || is_ereg(src_reg))
758 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
759 		else
760 			EMIT1(0x8B);
761 		break;
762 	case BPF_DW:
763 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
764 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
765 		break;
766 	}
767 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
768 	*pprog = prog;
769 }
770 
771 /* STX: *(u8*)(dst_reg + off) = src_reg */
772 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
773 {
774 	u8 *prog = *pprog;
775 
776 	switch (size) {
777 	case BPF_B:
778 		/* Emit 'mov byte ptr [rax + off], al' */
779 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
780 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
781 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
782 		else
783 			EMIT1(0x88);
784 		break;
785 	case BPF_H:
786 		if (is_ereg(dst_reg) || is_ereg(src_reg))
787 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
788 		else
789 			EMIT2(0x66, 0x89);
790 		break;
791 	case BPF_W:
792 		if (is_ereg(dst_reg) || is_ereg(src_reg))
793 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
794 		else
795 			EMIT1(0x89);
796 		break;
797 	case BPF_DW:
798 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
799 		break;
800 	}
801 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
802 	*pprog = prog;
803 }
804 
805 static int emit_atomic(u8 **pprog, u8 atomic_op,
806 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
807 {
808 	u8 *prog = *pprog;
809 
810 	EMIT1(0xF0); /* lock prefix */
811 
812 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
813 
814 	/* emit opcode */
815 	switch (atomic_op) {
816 	case BPF_ADD:
817 	case BPF_AND:
818 	case BPF_OR:
819 	case BPF_XOR:
820 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
821 		EMIT1(simple_alu_opcodes[atomic_op]);
822 		break;
823 	case BPF_ADD | BPF_FETCH:
824 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
825 		EMIT2(0x0F, 0xC1);
826 		break;
827 	case BPF_XCHG:
828 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
829 		EMIT1(0x87);
830 		break;
831 	case BPF_CMPXCHG:
832 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
833 		EMIT2(0x0F, 0xB1);
834 		break;
835 	default:
836 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
837 		return -EFAULT;
838 	}
839 
840 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
841 
842 	*pprog = prog;
843 	return 0;
844 }
845 
846 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
847 {
848 	u32 reg = x->fixup >> 8;
849 
850 	/* jump over faulting load and clear dest register */
851 	*(unsigned long *)((void *)regs + reg) = 0;
852 	regs->ip += x->fixup & 0xff;
853 	return true;
854 }
855 
856 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
857 			     bool *regs_used, bool *tail_call_seen)
858 {
859 	int i;
860 
861 	for (i = 1; i <= insn_cnt; i++, insn++) {
862 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
863 			*tail_call_seen = true;
864 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
865 			regs_used[0] = true;
866 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
867 			regs_used[1] = true;
868 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
869 			regs_used[2] = true;
870 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
871 			regs_used[3] = true;
872 	}
873 }
874 
875 static void emit_nops(u8 **pprog, int len)
876 {
877 	u8 *prog = *pprog;
878 	int i, noplen;
879 
880 	while (len > 0) {
881 		noplen = len;
882 
883 		if (noplen > ASM_NOP_MAX)
884 			noplen = ASM_NOP_MAX;
885 
886 		for (i = 0; i < noplen; i++)
887 			EMIT1(x86_nops[noplen][i]);
888 		len -= noplen;
889 	}
890 
891 	*pprog = prog;
892 }
893 
894 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
895 
896 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
897 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
898 {
899 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
900 	struct bpf_insn *insn = bpf_prog->insnsi;
901 	bool callee_regs_used[4] = {};
902 	int insn_cnt = bpf_prog->len;
903 	bool tail_call_seen = false;
904 	bool seen_exit = false;
905 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
906 	int i, excnt = 0;
907 	int ilen, proglen = 0;
908 	u8 *prog = temp;
909 	int err;
910 
911 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
912 			 &tail_call_seen);
913 
914 	/* tail call's presence in current prog implies it is reachable */
915 	tail_call_reachable |= tail_call_seen;
916 
917 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
918 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
919 		      bpf_prog->aux->func_idx != 0);
920 	push_callee_regs(&prog, callee_regs_used);
921 
922 	ilen = prog - temp;
923 	if (rw_image)
924 		memcpy(rw_image + proglen, temp, ilen);
925 	proglen += ilen;
926 	addrs[0] = proglen;
927 	prog = temp;
928 
929 	for (i = 1; i <= insn_cnt; i++, insn++) {
930 		const s32 imm32 = insn->imm;
931 		u32 dst_reg = insn->dst_reg;
932 		u32 src_reg = insn->src_reg;
933 		u8 b2 = 0, b3 = 0;
934 		u8 *start_of_ldx;
935 		s64 jmp_offset;
936 		u8 jmp_cond;
937 		u8 *func;
938 		int nops;
939 
940 		switch (insn->code) {
941 			/* ALU */
942 		case BPF_ALU | BPF_ADD | BPF_X:
943 		case BPF_ALU | BPF_SUB | BPF_X:
944 		case BPF_ALU | BPF_AND | BPF_X:
945 		case BPF_ALU | BPF_OR | BPF_X:
946 		case BPF_ALU | BPF_XOR | BPF_X:
947 		case BPF_ALU64 | BPF_ADD | BPF_X:
948 		case BPF_ALU64 | BPF_SUB | BPF_X:
949 		case BPF_ALU64 | BPF_AND | BPF_X:
950 		case BPF_ALU64 | BPF_OR | BPF_X:
951 		case BPF_ALU64 | BPF_XOR | BPF_X:
952 			maybe_emit_mod(&prog, dst_reg, src_reg,
953 				       BPF_CLASS(insn->code) == BPF_ALU64);
954 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
955 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
956 			break;
957 
958 		case BPF_ALU64 | BPF_MOV | BPF_X:
959 		case BPF_ALU | BPF_MOV | BPF_X:
960 			emit_mov_reg(&prog,
961 				     BPF_CLASS(insn->code) == BPF_ALU64,
962 				     dst_reg, src_reg);
963 			break;
964 
965 			/* neg dst */
966 		case BPF_ALU | BPF_NEG:
967 		case BPF_ALU64 | BPF_NEG:
968 			maybe_emit_1mod(&prog, dst_reg,
969 					BPF_CLASS(insn->code) == BPF_ALU64);
970 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
971 			break;
972 
973 		case BPF_ALU | BPF_ADD | BPF_K:
974 		case BPF_ALU | BPF_SUB | BPF_K:
975 		case BPF_ALU | BPF_AND | BPF_K:
976 		case BPF_ALU | BPF_OR | BPF_K:
977 		case BPF_ALU | BPF_XOR | BPF_K:
978 		case BPF_ALU64 | BPF_ADD | BPF_K:
979 		case BPF_ALU64 | BPF_SUB | BPF_K:
980 		case BPF_ALU64 | BPF_AND | BPF_K:
981 		case BPF_ALU64 | BPF_OR | BPF_K:
982 		case BPF_ALU64 | BPF_XOR | BPF_K:
983 			maybe_emit_1mod(&prog, dst_reg,
984 					BPF_CLASS(insn->code) == BPF_ALU64);
985 
986 			/*
987 			 * b3 holds 'normal' opcode, b2 short form only valid
988 			 * in case dst is eax/rax.
989 			 */
990 			switch (BPF_OP(insn->code)) {
991 			case BPF_ADD:
992 				b3 = 0xC0;
993 				b2 = 0x05;
994 				break;
995 			case BPF_SUB:
996 				b3 = 0xE8;
997 				b2 = 0x2D;
998 				break;
999 			case BPF_AND:
1000 				b3 = 0xE0;
1001 				b2 = 0x25;
1002 				break;
1003 			case BPF_OR:
1004 				b3 = 0xC8;
1005 				b2 = 0x0D;
1006 				break;
1007 			case BPF_XOR:
1008 				b3 = 0xF0;
1009 				b2 = 0x35;
1010 				break;
1011 			}
1012 
1013 			if (is_imm8(imm32))
1014 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1015 			else if (is_axreg(dst_reg))
1016 				EMIT1_off32(b2, imm32);
1017 			else
1018 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1019 			break;
1020 
1021 		case BPF_ALU64 | BPF_MOV | BPF_K:
1022 		case BPF_ALU | BPF_MOV | BPF_K:
1023 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1024 				       dst_reg, imm32);
1025 			break;
1026 
1027 		case BPF_LD | BPF_IMM | BPF_DW:
1028 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1029 			insn++;
1030 			i++;
1031 			break;
1032 
1033 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1034 		case BPF_ALU | BPF_MOD | BPF_X:
1035 		case BPF_ALU | BPF_DIV | BPF_X:
1036 		case BPF_ALU | BPF_MOD | BPF_K:
1037 		case BPF_ALU | BPF_DIV | BPF_K:
1038 		case BPF_ALU64 | BPF_MOD | BPF_X:
1039 		case BPF_ALU64 | BPF_DIV | BPF_X:
1040 		case BPF_ALU64 | BPF_MOD | BPF_K:
1041 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1042 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1043 
1044 			if (dst_reg != BPF_REG_0)
1045 				EMIT1(0x50); /* push rax */
1046 			if (dst_reg != BPF_REG_3)
1047 				EMIT1(0x52); /* push rdx */
1048 
1049 			if (BPF_SRC(insn->code) == BPF_X) {
1050 				if (src_reg == BPF_REG_0 ||
1051 				    src_reg == BPF_REG_3) {
1052 					/* mov r11, src_reg */
1053 					EMIT_mov(AUX_REG, src_reg);
1054 					src_reg = AUX_REG;
1055 				}
1056 			} else {
1057 				/* mov r11, imm32 */
1058 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1059 				src_reg = AUX_REG;
1060 			}
1061 
1062 			if (dst_reg != BPF_REG_0)
1063 				/* mov rax, dst_reg */
1064 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1065 
1066 			/*
1067 			 * xor edx, edx
1068 			 * equivalent to 'xor rdx, rdx', but one byte less
1069 			 */
1070 			EMIT2(0x31, 0xd2);
1071 
1072 			/* div src_reg */
1073 			maybe_emit_1mod(&prog, src_reg, is64);
1074 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1075 
1076 			if (BPF_OP(insn->code) == BPF_MOD &&
1077 			    dst_reg != BPF_REG_3)
1078 				/* mov dst_reg, rdx */
1079 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1080 			else if (BPF_OP(insn->code) == BPF_DIV &&
1081 				 dst_reg != BPF_REG_0)
1082 				/* mov dst_reg, rax */
1083 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1084 
1085 			if (dst_reg != BPF_REG_3)
1086 				EMIT1(0x5A); /* pop rdx */
1087 			if (dst_reg != BPF_REG_0)
1088 				EMIT1(0x58); /* pop rax */
1089 			break;
1090 		}
1091 
1092 		case BPF_ALU | BPF_MUL | BPF_K:
1093 		case BPF_ALU64 | BPF_MUL | BPF_K:
1094 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1095 				       BPF_CLASS(insn->code) == BPF_ALU64);
1096 
1097 			if (is_imm8(imm32))
1098 				/* imul dst_reg, dst_reg, imm8 */
1099 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1100 				      imm32);
1101 			else
1102 				/* imul dst_reg, dst_reg, imm32 */
1103 				EMIT2_off32(0x69,
1104 					    add_2reg(0xC0, dst_reg, dst_reg),
1105 					    imm32);
1106 			break;
1107 
1108 		case BPF_ALU | BPF_MUL | BPF_X:
1109 		case BPF_ALU64 | BPF_MUL | BPF_X:
1110 			maybe_emit_mod(&prog, src_reg, dst_reg,
1111 				       BPF_CLASS(insn->code) == BPF_ALU64);
1112 
1113 			/* imul dst_reg, src_reg */
1114 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1115 			break;
1116 
1117 			/* Shifts */
1118 		case BPF_ALU | BPF_LSH | BPF_K:
1119 		case BPF_ALU | BPF_RSH | BPF_K:
1120 		case BPF_ALU | BPF_ARSH | BPF_K:
1121 		case BPF_ALU64 | BPF_LSH | BPF_K:
1122 		case BPF_ALU64 | BPF_RSH | BPF_K:
1123 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1124 			maybe_emit_1mod(&prog, dst_reg,
1125 					BPF_CLASS(insn->code) == BPF_ALU64);
1126 
1127 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1128 			if (imm32 == 1)
1129 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1130 			else
1131 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1132 			break;
1133 
1134 		case BPF_ALU | BPF_LSH | BPF_X:
1135 		case BPF_ALU | BPF_RSH | BPF_X:
1136 		case BPF_ALU | BPF_ARSH | BPF_X:
1137 		case BPF_ALU64 | BPF_LSH | BPF_X:
1138 		case BPF_ALU64 | BPF_RSH | BPF_X:
1139 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1140 
1141 			/* Check for bad case when dst_reg == rcx */
1142 			if (dst_reg == BPF_REG_4) {
1143 				/* mov r11, dst_reg */
1144 				EMIT_mov(AUX_REG, dst_reg);
1145 				dst_reg = AUX_REG;
1146 			}
1147 
1148 			if (src_reg != BPF_REG_4) { /* common case */
1149 				EMIT1(0x51); /* push rcx */
1150 
1151 				/* mov rcx, src_reg */
1152 				EMIT_mov(BPF_REG_4, src_reg);
1153 			}
1154 
1155 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1156 			maybe_emit_1mod(&prog, dst_reg,
1157 					BPF_CLASS(insn->code) == BPF_ALU64);
1158 
1159 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1160 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1161 
1162 			if (src_reg != BPF_REG_4)
1163 				EMIT1(0x59); /* pop rcx */
1164 
1165 			if (insn->dst_reg == BPF_REG_4)
1166 				/* mov dst_reg, r11 */
1167 				EMIT_mov(insn->dst_reg, AUX_REG);
1168 			break;
1169 
1170 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1171 			switch (imm32) {
1172 			case 16:
1173 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1174 				EMIT1(0x66);
1175 				if (is_ereg(dst_reg))
1176 					EMIT1(0x41);
1177 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1178 
1179 				/* Emit 'movzwl eax, ax' */
1180 				if (is_ereg(dst_reg))
1181 					EMIT3(0x45, 0x0F, 0xB7);
1182 				else
1183 					EMIT2(0x0F, 0xB7);
1184 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1185 				break;
1186 			case 32:
1187 				/* Emit 'bswap eax' to swap lower 4 bytes */
1188 				if (is_ereg(dst_reg))
1189 					EMIT2(0x41, 0x0F);
1190 				else
1191 					EMIT1(0x0F);
1192 				EMIT1(add_1reg(0xC8, dst_reg));
1193 				break;
1194 			case 64:
1195 				/* Emit 'bswap rax' to swap 8 bytes */
1196 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1197 				      add_1reg(0xC8, dst_reg));
1198 				break;
1199 			}
1200 			break;
1201 
1202 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1203 			switch (imm32) {
1204 			case 16:
1205 				/*
1206 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1207 				 * into 64 bit
1208 				 */
1209 				if (is_ereg(dst_reg))
1210 					EMIT3(0x45, 0x0F, 0xB7);
1211 				else
1212 					EMIT2(0x0F, 0xB7);
1213 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1214 				break;
1215 			case 32:
1216 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1217 				if (is_ereg(dst_reg))
1218 					EMIT1(0x45);
1219 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1220 				break;
1221 			case 64:
1222 				/* nop */
1223 				break;
1224 			}
1225 			break;
1226 
1227 			/* speculation barrier */
1228 		case BPF_ST | BPF_NOSPEC:
1229 			if (boot_cpu_has(X86_FEATURE_XMM2))
1230 				EMIT_LFENCE();
1231 			break;
1232 
1233 			/* ST: *(u8*)(dst_reg + off) = imm */
1234 		case BPF_ST | BPF_MEM | BPF_B:
1235 			if (is_ereg(dst_reg))
1236 				EMIT2(0x41, 0xC6);
1237 			else
1238 				EMIT1(0xC6);
1239 			goto st;
1240 		case BPF_ST | BPF_MEM | BPF_H:
1241 			if (is_ereg(dst_reg))
1242 				EMIT3(0x66, 0x41, 0xC7);
1243 			else
1244 				EMIT2(0x66, 0xC7);
1245 			goto st;
1246 		case BPF_ST | BPF_MEM | BPF_W:
1247 			if (is_ereg(dst_reg))
1248 				EMIT2(0x41, 0xC7);
1249 			else
1250 				EMIT1(0xC7);
1251 			goto st;
1252 		case BPF_ST | BPF_MEM | BPF_DW:
1253 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1254 
1255 st:			if (is_imm8(insn->off))
1256 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1257 			else
1258 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1259 
1260 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1261 			break;
1262 
1263 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1264 		case BPF_STX | BPF_MEM | BPF_B:
1265 		case BPF_STX | BPF_MEM | BPF_H:
1266 		case BPF_STX | BPF_MEM | BPF_W:
1267 		case BPF_STX | BPF_MEM | BPF_DW:
1268 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1269 			break;
1270 
1271 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1272 		case BPF_LDX | BPF_MEM | BPF_B:
1273 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1274 		case BPF_LDX | BPF_MEM | BPF_H:
1275 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1276 		case BPF_LDX | BPF_MEM | BPF_W:
1277 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1278 		case BPF_LDX | BPF_MEM | BPF_DW:
1279 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1280 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1281 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1282 				 * add abs(insn->off) to the limit to make sure that negative
1283 				 * offset won't be an issue.
1284 				 * insn->off is s16, so it won't affect valid pointers.
1285 				 */
1286 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1287 				u8 *end_of_jmp1, *end_of_jmp2;
1288 
1289 				/* Conservatively check that src_reg + insn->off is a kernel address:
1290 				 * 1. src_reg + insn->off >= limit
1291 				 * 2. src_reg + insn->off doesn't become small positive.
1292 				 * Cannot do src_reg + insn->off >= limit in one branch,
1293 				 * since it needs two spare registers, but JIT has only one.
1294 				 */
1295 
1296 				/* movabsq r11, limit */
1297 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1298 				EMIT((u32)limit, 4);
1299 				EMIT(limit >> 32, 4);
1300 				/* cmp src_reg, r11 */
1301 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1302 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1303 				/* if unsigned '<' goto end_of_jmp2 */
1304 				EMIT2(X86_JB, 0);
1305 				end_of_jmp1 = prog;
1306 
1307 				/* mov r11, src_reg */
1308 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1309 				/* add r11, insn->off */
1310 				maybe_emit_1mod(&prog, AUX_REG, true);
1311 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1312 				/* jmp if not carry to start_of_ldx
1313 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1314 				 * that has to be rejected.
1315 				 */
1316 				EMIT2(0x73 /* JNC */, 0);
1317 				end_of_jmp2 = prog;
1318 
1319 				/* xor dst_reg, dst_reg */
1320 				emit_mov_imm32(&prog, false, dst_reg, 0);
1321 				/* jmp byte_after_ldx */
1322 				EMIT2(0xEB, 0);
1323 
1324 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1325 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1326 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1327 				start_of_ldx = prog;
1328 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1329 			}
1330 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1331 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1332 				struct exception_table_entry *ex;
1333 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1334 				s64 delta;
1335 
1336 				/* populate jmp_offset for JMP above */
1337 				start_of_ldx[-1] = prog - start_of_ldx;
1338 
1339 				if (!bpf_prog->aux->extable)
1340 					break;
1341 
1342 				if (excnt >= bpf_prog->aux->num_exentries) {
1343 					pr_err("ex gen bug\n");
1344 					return -EFAULT;
1345 				}
1346 				ex = &bpf_prog->aux->extable[excnt++];
1347 
1348 				delta = _insn - (u8 *)&ex->insn;
1349 				if (!is_simm32(delta)) {
1350 					pr_err("extable->insn doesn't fit into 32-bit\n");
1351 					return -EFAULT;
1352 				}
1353 				/* switch ex to rw buffer for writes */
1354 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1355 
1356 				ex->insn = delta;
1357 
1358 				ex->data = EX_TYPE_BPF;
1359 
1360 				if (dst_reg > BPF_REG_9) {
1361 					pr_err("verifier error\n");
1362 					return -EFAULT;
1363 				}
1364 				/*
1365 				 * Compute size of x86 insn and its target dest x86 register.
1366 				 * ex_handler_bpf() will use lower 8 bits to adjust
1367 				 * pt_regs->ip to jump over this x86 instruction
1368 				 * and upper bits to figure out which pt_regs to zero out.
1369 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1370 				 * of 4 bytes will be ignored and rbx will be zero inited.
1371 				 */
1372 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1373 			}
1374 			break;
1375 
1376 		case BPF_STX | BPF_ATOMIC | BPF_W:
1377 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1378 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1379 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1380 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1381 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1382 				u32 real_src_reg = src_reg;
1383 				u32 real_dst_reg = dst_reg;
1384 				u8 *branch_target;
1385 
1386 				/*
1387 				 * Can't be implemented with a single x86 insn.
1388 				 * Need to do a CMPXCHG loop.
1389 				 */
1390 
1391 				/* Will need RAX as a CMPXCHG operand so save R0 */
1392 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1393 				if (src_reg == BPF_REG_0)
1394 					real_src_reg = BPF_REG_AX;
1395 				if (dst_reg == BPF_REG_0)
1396 					real_dst_reg = BPF_REG_AX;
1397 
1398 				branch_target = prog;
1399 				/* Load old value */
1400 				emit_ldx(&prog, BPF_SIZE(insn->code),
1401 					 BPF_REG_0, real_dst_reg, insn->off);
1402 				/*
1403 				 * Perform the (commutative) operation locally,
1404 				 * put the result in the AUX_REG.
1405 				 */
1406 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1407 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1408 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1409 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1410 				/* Attempt to swap in new value */
1411 				err = emit_atomic(&prog, BPF_CMPXCHG,
1412 						  real_dst_reg, AUX_REG,
1413 						  insn->off,
1414 						  BPF_SIZE(insn->code));
1415 				if (WARN_ON(err))
1416 					return err;
1417 				/*
1418 				 * ZF tells us whether we won the race. If it's
1419 				 * cleared we need to try again.
1420 				 */
1421 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1422 				/* Return the pre-modification value */
1423 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1424 				/* Restore R0 after clobbering RAX */
1425 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1426 				break;
1427 			}
1428 
1429 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1430 					  insn->off, BPF_SIZE(insn->code));
1431 			if (err)
1432 				return err;
1433 			break;
1434 
1435 			/* call */
1436 		case BPF_JMP | BPF_CALL:
1437 			func = (u8 *) __bpf_call_base + imm32;
1438 			if (tail_call_reachable) {
1439 				/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1440 				EMIT3_off32(0x48, 0x8B, 0x85,
1441 					    -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1442 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1443 					return -EINVAL;
1444 			} else {
1445 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1446 					return -EINVAL;
1447 			}
1448 			break;
1449 
1450 		case BPF_JMP | BPF_TAIL_CALL:
1451 			if (imm32)
1452 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1453 							  &prog, image + addrs[i - 1],
1454 							  callee_regs_used,
1455 							  bpf_prog->aux->stack_depth,
1456 							  ctx);
1457 			else
1458 				emit_bpf_tail_call_indirect(&prog,
1459 							    callee_regs_used,
1460 							    bpf_prog->aux->stack_depth,
1461 							    image + addrs[i - 1],
1462 							    ctx);
1463 			break;
1464 
1465 			/* cond jump */
1466 		case BPF_JMP | BPF_JEQ | BPF_X:
1467 		case BPF_JMP | BPF_JNE | BPF_X:
1468 		case BPF_JMP | BPF_JGT | BPF_X:
1469 		case BPF_JMP | BPF_JLT | BPF_X:
1470 		case BPF_JMP | BPF_JGE | BPF_X:
1471 		case BPF_JMP | BPF_JLE | BPF_X:
1472 		case BPF_JMP | BPF_JSGT | BPF_X:
1473 		case BPF_JMP | BPF_JSLT | BPF_X:
1474 		case BPF_JMP | BPF_JSGE | BPF_X:
1475 		case BPF_JMP | BPF_JSLE | BPF_X:
1476 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1477 		case BPF_JMP32 | BPF_JNE | BPF_X:
1478 		case BPF_JMP32 | BPF_JGT | BPF_X:
1479 		case BPF_JMP32 | BPF_JLT | BPF_X:
1480 		case BPF_JMP32 | BPF_JGE | BPF_X:
1481 		case BPF_JMP32 | BPF_JLE | BPF_X:
1482 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1483 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1484 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1485 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1486 			/* cmp dst_reg, src_reg */
1487 			maybe_emit_mod(&prog, dst_reg, src_reg,
1488 				       BPF_CLASS(insn->code) == BPF_JMP);
1489 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1490 			goto emit_cond_jmp;
1491 
1492 		case BPF_JMP | BPF_JSET | BPF_X:
1493 		case BPF_JMP32 | BPF_JSET | BPF_X:
1494 			/* test dst_reg, src_reg */
1495 			maybe_emit_mod(&prog, dst_reg, src_reg,
1496 				       BPF_CLASS(insn->code) == BPF_JMP);
1497 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1498 			goto emit_cond_jmp;
1499 
1500 		case BPF_JMP | BPF_JSET | BPF_K:
1501 		case BPF_JMP32 | BPF_JSET | BPF_K:
1502 			/* test dst_reg, imm32 */
1503 			maybe_emit_1mod(&prog, dst_reg,
1504 					BPF_CLASS(insn->code) == BPF_JMP);
1505 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1506 			goto emit_cond_jmp;
1507 
1508 		case BPF_JMP | BPF_JEQ | BPF_K:
1509 		case BPF_JMP | BPF_JNE | BPF_K:
1510 		case BPF_JMP | BPF_JGT | BPF_K:
1511 		case BPF_JMP | BPF_JLT | BPF_K:
1512 		case BPF_JMP | BPF_JGE | BPF_K:
1513 		case BPF_JMP | BPF_JLE | BPF_K:
1514 		case BPF_JMP | BPF_JSGT | BPF_K:
1515 		case BPF_JMP | BPF_JSLT | BPF_K:
1516 		case BPF_JMP | BPF_JSGE | BPF_K:
1517 		case BPF_JMP | BPF_JSLE | BPF_K:
1518 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1519 		case BPF_JMP32 | BPF_JNE | BPF_K:
1520 		case BPF_JMP32 | BPF_JGT | BPF_K:
1521 		case BPF_JMP32 | BPF_JLT | BPF_K:
1522 		case BPF_JMP32 | BPF_JGE | BPF_K:
1523 		case BPF_JMP32 | BPF_JLE | BPF_K:
1524 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1525 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1526 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1527 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1528 			/* test dst_reg, dst_reg to save one extra byte */
1529 			if (imm32 == 0) {
1530 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1531 					       BPF_CLASS(insn->code) == BPF_JMP);
1532 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1533 				goto emit_cond_jmp;
1534 			}
1535 
1536 			/* cmp dst_reg, imm8/32 */
1537 			maybe_emit_1mod(&prog, dst_reg,
1538 					BPF_CLASS(insn->code) == BPF_JMP);
1539 
1540 			if (is_imm8(imm32))
1541 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1542 			else
1543 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1544 
1545 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1546 			switch (BPF_OP(insn->code)) {
1547 			case BPF_JEQ:
1548 				jmp_cond = X86_JE;
1549 				break;
1550 			case BPF_JSET:
1551 			case BPF_JNE:
1552 				jmp_cond = X86_JNE;
1553 				break;
1554 			case BPF_JGT:
1555 				/* GT is unsigned '>', JA in x86 */
1556 				jmp_cond = X86_JA;
1557 				break;
1558 			case BPF_JLT:
1559 				/* LT is unsigned '<', JB in x86 */
1560 				jmp_cond = X86_JB;
1561 				break;
1562 			case BPF_JGE:
1563 				/* GE is unsigned '>=', JAE in x86 */
1564 				jmp_cond = X86_JAE;
1565 				break;
1566 			case BPF_JLE:
1567 				/* LE is unsigned '<=', JBE in x86 */
1568 				jmp_cond = X86_JBE;
1569 				break;
1570 			case BPF_JSGT:
1571 				/* Signed '>', GT in x86 */
1572 				jmp_cond = X86_JG;
1573 				break;
1574 			case BPF_JSLT:
1575 				/* Signed '<', LT in x86 */
1576 				jmp_cond = X86_JL;
1577 				break;
1578 			case BPF_JSGE:
1579 				/* Signed '>=', GE in x86 */
1580 				jmp_cond = X86_JGE;
1581 				break;
1582 			case BPF_JSLE:
1583 				/* Signed '<=', LE in x86 */
1584 				jmp_cond = X86_JLE;
1585 				break;
1586 			default: /* to silence GCC warning */
1587 				return -EFAULT;
1588 			}
1589 			jmp_offset = addrs[i + insn->off] - addrs[i];
1590 			if (is_imm8(jmp_offset)) {
1591 				if (jmp_padding) {
1592 					/* To keep the jmp_offset valid, the extra bytes are
1593 					 * padded before the jump insn, so we subtract the
1594 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1595 					 *
1596 					 * If the previous pass already emits an imm8
1597 					 * jmp_cond, then this BPF insn won't shrink, so
1598 					 * "nops" is 0.
1599 					 *
1600 					 * On the other hand, if the previous pass emits an
1601 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1602 					 * keep the image from shrinking further.
1603 					 *
1604 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1605 					 *     is 2 bytes, so the size difference is 4 bytes.
1606 					 */
1607 					nops = INSN_SZ_DIFF - 2;
1608 					if (nops != 0 && nops != 4) {
1609 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1610 						       nops);
1611 						return -EFAULT;
1612 					}
1613 					emit_nops(&prog, nops);
1614 				}
1615 				EMIT2(jmp_cond, jmp_offset);
1616 			} else if (is_simm32(jmp_offset)) {
1617 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1618 			} else {
1619 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1620 				return -EFAULT;
1621 			}
1622 
1623 			break;
1624 
1625 		case BPF_JMP | BPF_JA:
1626 			if (insn->off == -1)
1627 				/* -1 jmp instructions will always jump
1628 				 * backwards two bytes. Explicitly handling
1629 				 * this case avoids wasting too many passes
1630 				 * when there are long sequences of replaced
1631 				 * dead code.
1632 				 */
1633 				jmp_offset = -2;
1634 			else
1635 				jmp_offset = addrs[i + insn->off] - addrs[i];
1636 
1637 			if (!jmp_offset) {
1638 				/*
1639 				 * If jmp_padding is enabled, the extra nops will
1640 				 * be inserted. Otherwise, optimize out nop jumps.
1641 				 */
1642 				if (jmp_padding) {
1643 					/* There are 3 possible conditions.
1644 					 * (1) This BPF_JA is already optimized out in
1645 					 *     the previous run, so there is no need
1646 					 *     to pad any extra byte (0 byte).
1647 					 * (2) The previous pass emits an imm8 jmp,
1648 					 *     so we pad 2 bytes to match the previous
1649 					 *     insn size.
1650 					 * (3) Similarly, the previous pass emits an
1651 					 *     imm32 jmp, and 5 bytes is padded.
1652 					 */
1653 					nops = INSN_SZ_DIFF;
1654 					if (nops != 0 && nops != 2 && nops != 5) {
1655 						pr_err("unexpected nop jump padding: %d bytes\n",
1656 						       nops);
1657 						return -EFAULT;
1658 					}
1659 					emit_nops(&prog, nops);
1660 				}
1661 				break;
1662 			}
1663 emit_jmp:
1664 			if (is_imm8(jmp_offset)) {
1665 				if (jmp_padding) {
1666 					/* To avoid breaking jmp_offset, the extra bytes
1667 					 * are padded before the actual jmp insn, so
1668 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1669 					 *
1670 					 * If the previous pass already emits an imm8
1671 					 * jmp, there is nothing to pad (0 byte).
1672 					 *
1673 					 * If it emits an imm32 jmp (5 bytes) previously
1674 					 * and now an imm8 jmp (2 bytes), then we pad
1675 					 * (5 - 2 = 3) bytes to stop the image from
1676 					 * shrinking further.
1677 					 */
1678 					nops = INSN_SZ_DIFF - 2;
1679 					if (nops != 0 && nops != 3) {
1680 						pr_err("unexpected jump padding: %d bytes\n",
1681 						       nops);
1682 						return -EFAULT;
1683 					}
1684 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1685 				}
1686 				EMIT2(0xEB, jmp_offset);
1687 			} else if (is_simm32(jmp_offset)) {
1688 				EMIT1_off32(0xE9, jmp_offset);
1689 			} else {
1690 				pr_err("jmp gen bug %llx\n", jmp_offset);
1691 				return -EFAULT;
1692 			}
1693 			break;
1694 
1695 		case BPF_JMP | BPF_EXIT:
1696 			if (seen_exit) {
1697 				jmp_offset = ctx->cleanup_addr - addrs[i];
1698 				goto emit_jmp;
1699 			}
1700 			seen_exit = true;
1701 			/* Update cleanup_addr */
1702 			ctx->cleanup_addr = proglen;
1703 			pop_callee_regs(&prog, callee_regs_used);
1704 			EMIT1(0xC9);         /* leave */
1705 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1706 			break;
1707 
1708 		default:
1709 			/*
1710 			 * By design x86-64 JIT should support all BPF instructions.
1711 			 * This error will be seen if new instruction was added
1712 			 * to the interpreter, but not to the JIT, or if there is
1713 			 * junk in bpf_prog.
1714 			 */
1715 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1716 			return -EINVAL;
1717 		}
1718 
1719 		ilen = prog - temp;
1720 		if (ilen > BPF_MAX_INSN_SIZE) {
1721 			pr_err("bpf_jit: fatal insn size error\n");
1722 			return -EFAULT;
1723 		}
1724 
1725 		if (image) {
1726 			/*
1727 			 * When populating the image, assert that:
1728 			 *
1729 			 *  i) We do not write beyond the allocated space, and
1730 			 * ii) addrs[i] did not change from the prior run, in order
1731 			 *     to validate assumptions made for computing branch
1732 			 *     displacements.
1733 			 */
1734 			if (unlikely(proglen + ilen > oldproglen ||
1735 				     proglen + ilen != addrs[i])) {
1736 				pr_err("bpf_jit: fatal error\n");
1737 				return -EFAULT;
1738 			}
1739 			memcpy(rw_image + proglen, temp, ilen);
1740 		}
1741 		proglen += ilen;
1742 		addrs[i] = proglen;
1743 		prog = temp;
1744 	}
1745 
1746 	if (image && excnt != bpf_prog->aux->num_exentries) {
1747 		pr_err("extable is not populated\n");
1748 		return -EFAULT;
1749 	}
1750 	return proglen;
1751 }
1752 
1753 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1754 		      int stack_size)
1755 {
1756 	int i, j, arg_size, nr_regs;
1757 	/* Store function arguments to stack.
1758 	 * For a function that accepts two pointers the sequence will be:
1759 	 * mov QWORD PTR [rbp-0x10],rdi
1760 	 * mov QWORD PTR [rbp-0x8],rsi
1761 	 */
1762 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1763 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1764 			nr_regs = (m->arg_size[i] + 7) / 8;
1765 			arg_size = 8;
1766 		} else {
1767 			nr_regs = 1;
1768 			arg_size = m->arg_size[i];
1769 		}
1770 
1771 		while (nr_regs) {
1772 			emit_stx(prog, bytes_to_bpf_size(arg_size),
1773 				 BPF_REG_FP,
1774 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1775 				 -(stack_size - j * 8));
1776 			nr_regs--;
1777 			j++;
1778 		}
1779 	}
1780 }
1781 
1782 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1783 			 int stack_size)
1784 {
1785 	int i, j, arg_size, nr_regs;
1786 
1787 	/* Restore function arguments from stack.
1788 	 * For a function that accepts two pointers the sequence will be:
1789 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1790 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1791 	 */
1792 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1793 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1794 			nr_regs = (m->arg_size[i] + 7) / 8;
1795 			arg_size = 8;
1796 		} else {
1797 			nr_regs = 1;
1798 			arg_size = m->arg_size[i];
1799 		}
1800 
1801 		while (nr_regs) {
1802 			emit_ldx(prog, bytes_to_bpf_size(arg_size),
1803 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1804 				 BPF_REG_FP,
1805 				 -(stack_size - j * 8));
1806 			nr_regs--;
1807 			j++;
1808 		}
1809 	}
1810 }
1811 
1812 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1813 			   struct bpf_tramp_link *l, int stack_size,
1814 			   int run_ctx_off, bool save_ret)
1815 {
1816 	void (*exit)(struct bpf_prog *prog, u64 start,
1817 		     struct bpf_tramp_run_ctx *run_ctx) = __bpf_prog_exit;
1818 	u64 (*enter)(struct bpf_prog *prog,
1819 		     struct bpf_tramp_run_ctx *run_ctx) = __bpf_prog_enter;
1820 	u8 *prog = *pprog;
1821 	u8 *jmp_insn;
1822 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1823 	struct bpf_prog *p = l->link.prog;
1824 	u64 cookie = l->cookie;
1825 
1826 	/* mov rdi, cookie */
1827 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
1828 
1829 	/* Prepare struct bpf_tramp_run_ctx.
1830 	 *
1831 	 * bpf_tramp_run_ctx is already preserved by
1832 	 * arch_prepare_bpf_trampoline().
1833 	 *
1834 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
1835 	 */
1836 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
1837 
1838 	if (p->aux->sleepable) {
1839 		enter = __bpf_prog_enter_sleepable;
1840 		exit = __bpf_prog_exit_sleepable;
1841 	} else if (p->type == BPF_PROG_TYPE_STRUCT_OPS) {
1842 		enter = __bpf_prog_enter_struct_ops;
1843 		exit = __bpf_prog_exit_struct_ops;
1844 	} else if (p->expected_attach_type == BPF_LSM_CGROUP) {
1845 		enter = __bpf_prog_enter_lsm_cgroup;
1846 		exit = __bpf_prog_exit_lsm_cgroup;
1847 	}
1848 
1849 	/* arg1: mov rdi, progs[i] */
1850 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1851 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
1852 	EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
1853 
1854 	if (emit_call(&prog, enter, prog))
1855 		return -EINVAL;
1856 	/* remember prog start time returned by __bpf_prog_enter */
1857 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1858 
1859 	/* if (__bpf_prog_enter*(prog) == 0)
1860 	 *	goto skip_exec_of_prog;
1861 	 */
1862 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1863 	/* emit 2 nops that will be replaced with JE insn */
1864 	jmp_insn = prog;
1865 	emit_nops(&prog, 2);
1866 
1867 	/* arg1: lea rdi, [rbp - stack_size] */
1868 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1869 	/* arg2: progs[i]->insnsi for interpreter */
1870 	if (!p->jited)
1871 		emit_mov_imm64(&prog, BPF_REG_2,
1872 			       (long) p->insnsi >> 32,
1873 			       (u32) (long) p->insnsi);
1874 	/* call JITed bpf program or interpreter */
1875 	if (emit_call(&prog, p->bpf_func, prog))
1876 		return -EINVAL;
1877 
1878 	/*
1879 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1880 	 * of the previous call which is then passed on the stack to
1881 	 * the next BPF program.
1882 	 *
1883 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1884 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1885 	 */
1886 	if (save_ret)
1887 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1888 
1889 	/* replace 2 nops with JE insn, since jmp target is known */
1890 	jmp_insn[0] = X86_JE;
1891 	jmp_insn[1] = prog - jmp_insn - 2;
1892 
1893 	/* arg1: mov rdi, progs[i] */
1894 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1895 	/* arg2: mov rsi, rbx <- start time in nsec */
1896 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1897 	/* arg3: lea rdx, [rbp - run_ctx_off] */
1898 	EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
1899 	if (emit_call(&prog, exit, prog))
1900 		return -EINVAL;
1901 
1902 	*pprog = prog;
1903 	return 0;
1904 }
1905 
1906 static void emit_align(u8 **pprog, u32 align)
1907 {
1908 	u8 *target, *prog = *pprog;
1909 
1910 	target = PTR_ALIGN(prog, align);
1911 	if (target != prog)
1912 		emit_nops(&prog, target - prog);
1913 
1914 	*pprog = prog;
1915 }
1916 
1917 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1918 {
1919 	u8 *prog = *pprog;
1920 	s64 offset;
1921 
1922 	offset = func - (ip + 2 + 4);
1923 	if (!is_simm32(offset)) {
1924 		pr_err("Target %p is out of range\n", func);
1925 		return -EINVAL;
1926 	}
1927 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1928 	*pprog = prog;
1929 	return 0;
1930 }
1931 
1932 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1933 		      struct bpf_tramp_links *tl, int stack_size,
1934 		      int run_ctx_off, bool save_ret)
1935 {
1936 	int i;
1937 	u8 *prog = *pprog;
1938 
1939 	for (i = 0; i < tl->nr_links; i++) {
1940 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
1941 				    run_ctx_off, save_ret))
1942 			return -EINVAL;
1943 	}
1944 	*pprog = prog;
1945 	return 0;
1946 }
1947 
1948 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1949 			      struct bpf_tramp_links *tl, int stack_size,
1950 			      int run_ctx_off, u8 **branches)
1951 {
1952 	u8 *prog = *pprog;
1953 	int i;
1954 
1955 	/* The first fmod_ret program will receive a garbage return value.
1956 	 * Set this to 0 to avoid confusing the program.
1957 	 */
1958 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1959 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1960 	for (i = 0; i < tl->nr_links; i++) {
1961 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
1962 			return -EINVAL;
1963 
1964 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
1965 		 * if (*(u64 *)(rbp - 8) !=  0)
1966 		 *	goto do_fexit;
1967 		 */
1968 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
1969 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1970 
1971 		/* Save the location of the branch and Generate 6 nops
1972 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
1973 		 * are replaced with a conditional jump once do_fexit (i.e. the
1974 		 * start of the fexit invocation) is finalized.
1975 		 */
1976 		branches[i] = prog;
1977 		emit_nops(&prog, 4 + 2);
1978 	}
1979 
1980 	*pprog = prog;
1981 	return 0;
1982 }
1983 
1984 /* Example:
1985  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1986  * its 'struct btf_func_model' will be nr_args=2
1987  * The assembly code when eth_type_trans is executing after trampoline:
1988  *
1989  * push rbp
1990  * mov rbp, rsp
1991  * sub rsp, 16                     // space for skb and dev
1992  * push rbx                        // temp regs to pass start time
1993  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1994  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1995  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1996  * mov rbx, rax                    // remember start time in bpf stats are enabled
1997  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1998  * call addr_of_jited_FENTRY_prog
1999  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2000  * mov rsi, rbx                    // prog start time
2001  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2002  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2003  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2004  * pop rbx
2005  * leave
2006  * ret
2007  *
2008  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2009  * replaced with 'call generated_bpf_trampoline'. When it returns
2010  * eth_type_trans will continue executing with original skb and dev pointers.
2011  *
2012  * The assembly code when eth_type_trans is called from trampoline:
2013  *
2014  * push rbp
2015  * mov rbp, rsp
2016  * sub rsp, 24                     // space for skb, dev, return value
2017  * push rbx                        // temp regs to pass start time
2018  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2019  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2020  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2021  * mov rbx, rax                    // remember start time if bpf stats are enabled
2022  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2023  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2024  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2025  * mov rsi, rbx                    // prog start time
2026  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2027  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2028  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2029  * call eth_type_trans+5           // execute body of eth_type_trans
2030  * mov qword ptr [rbp - 8], rax    // save return value
2031  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2032  * mov rbx, rax                    // remember start time in bpf stats are enabled
2033  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2034  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2035  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2036  * mov rsi, rbx                    // prog start time
2037  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2038  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2039  * pop rbx
2040  * leave
2041  * add rsp, 8                      // skip eth_type_trans's frame
2042  * ret                             // return to its caller
2043  */
2044 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2045 				const struct btf_func_model *m, u32 flags,
2046 				struct bpf_tramp_links *tlinks,
2047 				void *func_addr)
2048 {
2049 	int ret, i, nr_args = m->nr_args, extra_nregs = 0;
2050 	int regs_off, ip_off, args_off, stack_size = nr_args * 8, run_ctx_off;
2051 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2052 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2053 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2054 	void *orig_call = func_addr;
2055 	u8 **branches = NULL;
2056 	u8 *prog;
2057 	bool save_ret;
2058 
2059 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2060 	if (nr_args > 6)
2061 		return -ENOTSUPP;
2062 
2063 	for (i = 0; i < MAX_BPF_FUNC_ARGS; i++) {
2064 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2065 			extra_nregs += (m->arg_size[i] + 7) / 8 - 1;
2066 	}
2067 	if (nr_args + extra_nregs > 6)
2068 		return -ENOTSUPP;
2069 	stack_size += extra_nregs * 8;
2070 
2071 	/* Generated trampoline stack layout:
2072 	 *
2073 	 * RBP + 8         [ return address  ]
2074 	 * RBP + 0         [ RBP             ]
2075 	 *
2076 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2077 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2078 	 *
2079 	 *                 [ reg_argN        ]  always
2080 	 *                 [ ...             ]
2081 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2082 	 *
2083 	 * RBP - args_off  [ arg regs count  ]  always
2084 	 *
2085 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2086 	 *
2087 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2088 	 */
2089 
2090 	/* room for return value of orig_call or fentry prog */
2091 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2092 	if (save_ret)
2093 		stack_size += 8;
2094 
2095 	regs_off = stack_size;
2096 
2097 	/* args count  */
2098 	stack_size += 8;
2099 	args_off = stack_size;
2100 
2101 	if (flags & BPF_TRAMP_F_IP_ARG)
2102 		stack_size += 8; /* room for IP address argument */
2103 
2104 	ip_off = stack_size;
2105 
2106 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2107 	run_ctx_off = stack_size;
2108 
2109 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2110 		/* skip patched call instruction and point orig_call to actual
2111 		 * body of the kernel function.
2112 		 */
2113 		if (is_endbr(*(u32 *)orig_call))
2114 			orig_call += ENDBR_INSN_SIZE;
2115 		orig_call += X86_PATCH_SIZE;
2116 	}
2117 
2118 	prog = image;
2119 
2120 	EMIT_ENDBR();
2121 	EMIT1(0x55);		 /* push rbp */
2122 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2123 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2124 	EMIT1(0x53);		 /* push rbx */
2125 
2126 	/* Store number of argument registers of the traced function:
2127 	 *   mov rax, nr_args + extra_nregs
2128 	 *   mov QWORD PTR [rbp - args_off], rax
2129 	 */
2130 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args + extra_nregs);
2131 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2132 
2133 	if (flags & BPF_TRAMP_F_IP_ARG) {
2134 		/* Store IP address of the traced function:
2135 		 * movabsq rax, func_addr
2136 		 * mov QWORD PTR [rbp - ip_off], rax
2137 		 */
2138 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2139 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2140 	}
2141 
2142 	save_regs(m, &prog, nr_args, regs_off);
2143 
2144 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2145 		/* arg1: mov rdi, im */
2146 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2147 		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2148 			ret = -EINVAL;
2149 			goto cleanup;
2150 		}
2151 	}
2152 
2153 	if (fentry->nr_links)
2154 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2155 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2156 			return -EINVAL;
2157 
2158 	if (fmod_ret->nr_links) {
2159 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2160 				   GFP_KERNEL);
2161 		if (!branches)
2162 			return -ENOMEM;
2163 
2164 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2165 				       run_ctx_off, branches)) {
2166 			ret = -EINVAL;
2167 			goto cleanup;
2168 		}
2169 	}
2170 
2171 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2172 		restore_regs(m, &prog, nr_args, regs_off);
2173 
2174 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2175 			emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2176 			EMIT2(0xff, 0xd0); /* call *rax */
2177 		} else {
2178 			/* call original function */
2179 			if (emit_call(&prog, orig_call, prog)) {
2180 				ret = -EINVAL;
2181 				goto cleanup;
2182 			}
2183 		}
2184 		/* remember return value in a stack for bpf prog to access */
2185 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2186 		im->ip_after_call = prog;
2187 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2188 		prog += X86_PATCH_SIZE;
2189 	}
2190 
2191 	if (fmod_ret->nr_links) {
2192 		/* From Intel 64 and IA-32 Architectures Optimization
2193 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2194 		 * Coding Rule 11: All branch targets should be 16-byte
2195 		 * aligned.
2196 		 */
2197 		emit_align(&prog, 16);
2198 		/* Update the branches saved in invoke_bpf_mod_ret with the
2199 		 * aligned address of do_fexit.
2200 		 */
2201 		for (i = 0; i < fmod_ret->nr_links; i++)
2202 			emit_cond_near_jump(&branches[i], prog, branches[i],
2203 					    X86_JNE);
2204 	}
2205 
2206 	if (fexit->nr_links)
2207 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2208 			ret = -EINVAL;
2209 			goto cleanup;
2210 		}
2211 
2212 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2213 		restore_regs(m, &prog, nr_args, regs_off);
2214 
2215 	/* This needs to be done regardless. If there were fmod_ret programs,
2216 	 * the return value is only updated on the stack and still needs to be
2217 	 * restored to R0.
2218 	 */
2219 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2220 		im->ip_epilogue = prog;
2221 		/* arg1: mov rdi, im */
2222 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2223 		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2224 			ret = -EINVAL;
2225 			goto cleanup;
2226 		}
2227 	}
2228 	/* restore return value of orig_call or fentry prog back into RAX */
2229 	if (save_ret)
2230 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2231 
2232 	EMIT1(0x5B); /* pop rbx */
2233 	EMIT1(0xC9); /* leave */
2234 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2235 		/* skip our return address and return to parent */
2236 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2237 	emit_return(&prog, prog);
2238 	/* Make sure the trampoline generation logic doesn't overflow */
2239 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2240 		ret = -EFAULT;
2241 		goto cleanup;
2242 	}
2243 	ret = prog - (u8 *)image;
2244 
2245 cleanup:
2246 	kfree(branches);
2247 	return ret;
2248 }
2249 
2250 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2251 {
2252 	u8 *jg_reloc, *prog = *pprog;
2253 	int pivot, err, jg_bytes = 1;
2254 	s64 jg_offset;
2255 
2256 	if (a == b) {
2257 		/* Leaf node of recursion, i.e. not a range of indices
2258 		 * anymore.
2259 		 */
2260 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2261 		if (!is_simm32(progs[a]))
2262 			return -1;
2263 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2264 			    progs[a]);
2265 		err = emit_cond_near_jump(&prog,	/* je func */
2266 					  (void *)progs[a], image + (prog - buf),
2267 					  X86_JE);
2268 		if (err)
2269 			return err;
2270 
2271 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2272 
2273 		*pprog = prog;
2274 		return 0;
2275 	}
2276 
2277 	/* Not a leaf node, so we pivot, and recursively descend into
2278 	 * the lower and upper ranges.
2279 	 */
2280 	pivot = (b - a) / 2;
2281 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2282 	if (!is_simm32(progs[a + pivot]))
2283 		return -1;
2284 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2285 
2286 	if (pivot > 2) {				/* jg upper_part */
2287 		/* Require near jump. */
2288 		jg_bytes = 4;
2289 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2290 	} else {
2291 		EMIT2(X86_JG, 0);
2292 	}
2293 	jg_reloc = prog;
2294 
2295 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2296 				  progs, image, buf);
2297 	if (err)
2298 		return err;
2299 
2300 	/* From Intel 64 and IA-32 Architectures Optimization
2301 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2302 	 * Coding Rule 11: All branch targets should be 16-byte
2303 	 * aligned.
2304 	 */
2305 	emit_align(&prog, 16);
2306 	jg_offset = prog - jg_reloc;
2307 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2308 
2309 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2310 				  b, progs, image, buf);
2311 	if (err)
2312 		return err;
2313 
2314 	*pprog = prog;
2315 	return 0;
2316 }
2317 
2318 static int cmp_ips(const void *a, const void *b)
2319 {
2320 	const s64 *ipa = a;
2321 	const s64 *ipb = b;
2322 
2323 	if (*ipa > *ipb)
2324 		return 1;
2325 	if (*ipa < *ipb)
2326 		return -1;
2327 	return 0;
2328 }
2329 
2330 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2331 {
2332 	u8 *prog = buf;
2333 
2334 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2335 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2336 }
2337 
2338 struct x64_jit_data {
2339 	struct bpf_binary_header *rw_header;
2340 	struct bpf_binary_header *header;
2341 	int *addrs;
2342 	u8 *image;
2343 	int proglen;
2344 	struct jit_context ctx;
2345 };
2346 
2347 #define MAX_PASSES 20
2348 #define PADDING_PASSES (MAX_PASSES - 5)
2349 
2350 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2351 {
2352 	struct bpf_binary_header *rw_header = NULL;
2353 	struct bpf_binary_header *header = NULL;
2354 	struct bpf_prog *tmp, *orig_prog = prog;
2355 	struct x64_jit_data *jit_data;
2356 	int proglen, oldproglen = 0;
2357 	struct jit_context ctx = {};
2358 	bool tmp_blinded = false;
2359 	bool extra_pass = false;
2360 	bool padding = false;
2361 	u8 *rw_image = NULL;
2362 	u8 *image = NULL;
2363 	int *addrs;
2364 	int pass;
2365 	int i;
2366 
2367 	if (!prog->jit_requested)
2368 		return orig_prog;
2369 
2370 	tmp = bpf_jit_blind_constants(prog);
2371 	/*
2372 	 * If blinding was requested and we failed during blinding,
2373 	 * we must fall back to the interpreter.
2374 	 */
2375 	if (IS_ERR(tmp))
2376 		return orig_prog;
2377 	if (tmp != prog) {
2378 		tmp_blinded = true;
2379 		prog = tmp;
2380 	}
2381 
2382 	jit_data = prog->aux->jit_data;
2383 	if (!jit_data) {
2384 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2385 		if (!jit_data) {
2386 			prog = orig_prog;
2387 			goto out;
2388 		}
2389 		prog->aux->jit_data = jit_data;
2390 	}
2391 	addrs = jit_data->addrs;
2392 	if (addrs) {
2393 		ctx = jit_data->ctx;
2394 		oldproglen = jit_data->proglen;
2395 		image = jit_data->image;
2396 		header = jit_data->header;
2397 		rw_header = jit_data->rw_header;
2398 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2399 		extra_pass = true;
2400 		padding = true;
2401 		goto skip_init_addrs;
2402 	}
2403 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2404 	if (!addrs) {
2405 		prog = orig_prog;
2406 		goto out_addrs;
2407 	}
2408 
2409 	/*
2410 	 * Before first pass, make a rough estimation of addrs[]
2411 	 * each BPF instruction is translated to less than 64 bytes
2412 	 */
2413 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2414 		proglen += 64;
2415 		addrs[i] = proglen;
2416 	}
2417 	ctx.cleanup_addr = proglen;
2418 skip_init_addrs:
2419 
2420 	/*
2421 	 * JITed image shrinks with every pass and the loop iterates
2422 	 * until the image stops shrinking. Very large BPF programs
2423 	 * may converge on the last pass. In such case do one more
2424 	 * pass to emit the final image.
2425 	 */
2426 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2427 		if (!padding && pass >= PADDING_PASSES)
2428 			padding = true;
2429 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2430 		if (proglen <= 0) {
2431 out_image:
2432 			image = NULL;
2433 			if (header) {
2434 				bpf_arch_text_copy(&header->size, &rw_header->size,
2435 						   sizeof(rw_header->size));
2436 				bpf_jit_binary_pack_free(header, rw_header);
2437 			}
2438 			/* Fall back to interpreter mode */
2439 			prog = orig_prog;
2440 			if (extra_pass) {
2441 				prog->bpf_func = NULL;
2442 				prog->jited = 0;
2443 				prog->jited_len = 0;
2444 			}
2445 			goto out_addrs;
2446 		}
2447 		if (image) {
2448 			if (proglen != oldproglen) {
2449 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2450 				       proglen, oldproglen);
2451 				goto out_image;
2452 			}
2453 			break;
2454 		}
2455 		if (proglen == oldproglen) {
2456 			/*
2457 			 * The number of entries in extable is the number of BPF_LDX
2458 			 * insns that access kernel memory via "pointer to BTF type".
2459 			 * The verifier changed their opcode from LDX|MEM|size
2460 			 * to LDX|PROBE_MEM|size to make JITing easier.
2461 			 */
2462 			u32 align = __alignof__(struct exception_table_entry);
2463 			u32 extable_size = prog->aux->num_exentries *
2464 				sizeof(struct exception_table_entry);
2465 
2466 			/* allocate module memory for x86 insns and extable */
2467 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2468 							   &image, align, &rw_header, &rw_image,
2469 							   jit_fill_hole);
2470 			if (!header) {
2471 				prog = orig_prog;
2472 				goto out_addrs;
2473 			}
2474 			prog->aux->extable = (void *) image + roundup(proglen, align);
2475 		}
2476 		oldproglen = proglen;
2477 		cond_resched();
2478 	}
2479 
2480 	if (bpf_jit_enable > 1)
2481 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2482 
2483 	if (image) {
2484 		if (!prog->is_func || extra_pass) {
2485 			/*
2486 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2487 			 *   1) header is not pointing to proper module memory;
2488 			 *   2) the arch doesn't support bpf_arch_text_copy().
2489 			 *
2490 			 * Both cases are serious bugs and justify WARN_ON.
2491 			 */
2492 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2493 				/* header has been freed */
2494 				header = NULL;
2495 				goto out_image;
2496 			}
2497 
2498 			bpf_tail_call_direct_fixup(prog);
2499 		} else {
2500 			jit_data->addrs = addrs;
2501 			jit_data->ctx = ctx;
2502 			jit_data->proglen = proglen;
2503 			jit_data->image = image;
2504 			jit_data->header = header;
2505 			jit_data->rw_header = rw_header;
2506 		}
2507 		prog->bpf_func = (void *)image;
2508 		prog->jited = 1;
2509 		prog->jited_len = proglen;
2510 	} else {
2511 		prog = orig_prog;
2512 	}
2513 
2514 	if (!image || !prog->is_func || extra_pass) {
2515 		if (image)
2516 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2517 out_addrs:
2518 		kvfree(addrs);
2519 		kfree(jit_data);
2520 		prog->aux->jit_data = NULL;
2521 	}
2522 out:
2523 	if (tmp_blinded)
2524 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2525 					   tmp : orig_prog);
2526 	return prog;
2527 }
2528 
2529 bool bpf_jit_supports_kfunc_call(void)
2530 {
2531 	return true;
2532 }
2533 
2534 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2535 {
2536 	if (text_poke_copy(dst, src, len) == NULL)
2537 		return ERR_PTR(-EINVAL);
2538 	return dst;
2539 }
2540 
2541 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2542 bool bpf_jit_supports_subprog_tailcalls(void)
2543 {
2544 	return true;
2545 }
2546 
2547 void bpf_jit_free(struct bpf_prog *prog)
2548 {
2549 	if (prog->jited) {
2550 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2551 		struct bpf_binary_header *hdr;
2552 
2553 		/*
2554 		 * If we fail the final pass of JIT (from jit_subprogs),
2555 		 * the program may not be finalized yet. Call finalize here
2556 		 * before freeing it.
2557 		 */
2558 		if (jit_data) {
2559 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2560 						     jit_data->rw_header);
2561 			kvfree(jit_data->addrs);
2562 			kfree(jit_data);
2563 		}
2564 		hdr = bpf_jit_binary_pack_hdr(prog);
2565 		bpf_jit_binary_pack_free(hdr, NULL);
2566 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2567 	}
2568 
2569 	bpf_prog_unlock_free(prog);
2570 }
2571