xref: /linux/arch/x86/net/bpf_jit_comp.c (revision 0da908c291070d89482f6211dbe81d4d43c3f7cb)
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 		u8 jmp_cond;
1007 		u8 *func;
1008 		int nops;
1009 
1010 		switch (insn->code) {
1011 			/* ALU */
1012 		case BPF_ALU | BPF_ADD | BPF_X:
1013 		case BPF_ALU | BPF_SUB | BPF_X:
1014 		case BPF_ALU | BPF_AND | BPF_X:
1015 		case BPF_ALU | BPF_OR | BPF_X:
1016 		case BPF_ALU | BPF_XOR | BPF_X:
1017 		case BPF_ALU64 | BPF_ADD | BPF_X:
1018 		case BPF_ALU64 | BPF_SUB | BPF_X:
1019 		case BPF_ALU64 | BPF_AND | BPF_X:
1020 		case BPF_ALU64 | BPF_OR | BPF_X:
1021 		case BPF_ALU64 | BPF_XOR | BPF_X:
1022 			maybe_emit_mod(&prog, dst_reg, src_reg,
1023 				       BPF_CLASS(insn->code) == BPF_ALU64);
1024 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1025 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1026 			break;
1027 
1028 		case BPF_ALU64 | BPF_MOV | BPF_X:
1029 		case BPF_ALU | BPF_MOV | BPF_X:
1030 			emit_mov_reg(&prog,
1031 				     BPF_CLASS(insn->code) == BPF_ALU64,
1032 				     dst_reg, src_reg);
1033 			break;
1034 
1035 			/* neg dst */
1036 		case BPF_ALU | BPF_NEG:
1037 		case BPF_ALU64 | BPF_NEG:
1038 			maybe_emit_1mod(&prog, dst_reg,
1039 					BPF_CLASS(insn->code) == BPF_ALU64);
1040 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1041 			break;
1042 
1043 		case BPF_ALU | BPF_ADD | BPF_K:
1044 		case BPF_ALU | BPF_SUB | BPF_K:
1045 		case BPF_ALU | BPF_AND | BPF_K:
1046 		case BPF_ALU | BPF_OR | BPF_K:
1047 		case BPF_ALU | BPF_XOR | BPF_K:
1048 		case BPF_ALU64 | BPF_ADD | BPF_K:
1049 		case BPF_ALU64 | BPF_SUB | BPF_K:
1050 		case BPF_ALU64 | BPF_AND | BPF_K:
1051 		case BPF_ALU64 | BPF_OR | BPF_K:
1052 		case BPF_ALU64 | BPF_XOR | BPF_K:
1053 			maybe_emit_1mod(&prog, dst_reg,
1054 					BPF_CLASS(insn->code) == BPF_ALU64);
1055 
1056 			/*
1057 			 * b3 holds 'normal' opcode, b2 short form only valid
1058 			 * in case dst is eax/rax.
1059 			 */
1060 			switch (BPF_OP(insn->code)) {
1061 			case BPF_ADD:
1062 				b3 = 0xC0;
1063 				b2 = 0x05;
1064 				break;
1065 			case BPF_SUB:
1066 				b3 = 0xE8;
1067 				b2 = 0x2D;
1068 				break;
1069 			case BPF_AND:
1070 				b3 = 0xE0;
1071 				b2 = 0x25;
1072 				break;
1073 			case BPF_OR:
1074 				b3 = 0xC8;
1075 				b2 = 0x0D;
1076 				break;
1077 			case BPF_XOR:
1078 				b3 = 0xF0;
1079 				b2 = 0x35;
1080 				break;
1081 			}
1082 
1083 			if (is_imm8(imm32))
1084 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1085 			else if (is_axreg(dst_reg))
1086 				EMIT1_off32(b2, imm32);
1087 			else
1088 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1089 			break;
1090 
1091 		case BPF_ALU64 | BPF_MOV | BPF_K:
1092 		case BPF_ALU | BPF_MOV | BPF_K:
1093 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1094 				       dst_reg, imm32);
1095 			break;
1096 
1097 		case BPF_LD | BPF_IMM | BPF_DW:
1098 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1099 			insn++;
1100 			i++;
1101 			break;
1102 
1103 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1104 		case BPF_ALU | BPF_MOD | BPF_X:
1105 		case BPF_ALU | BPF_DIV | BPF_X:
1106 		case BPF_ALU | BPF_MOD | BPF_K:
1107 		case BPF_ALU | BPF_DIV | BPF_K:
1108 		case BPF_ALU64 | BPF_MOD | BPF_X:
1109 		case BPF_ALU64 | BPF_DIV | BPF_X:
1110 		case BPF_ALU64 | BPF_MOD | BPF_K:
1111 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1112 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1113 
1114 			if (dst_reg != BPF_REG_0)
1115 				EMIT1(0x50); /* push rax */
1116 			if (dst_reg != BPF_REG_3)
1117 				EMIT1(0x52); /* push rdx */
1118 
1119 			if (BPF_SRC(insn->code) == BPF_X) {
1120 				if (src_reg == BPF_REG_0 ||
1121 				    src_reg == BPF_REG_3) {
1122 					/* mov r11, src_reg */
1123 					EMIT_mov(AUX_REG, src_reg);
1124 					src_reg = AUX_REG;
1125 				}
1126 			} else {
1127 				/* mov r11, imm32 */
1128 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1129 				src_reg = AUX_REG;
1130 			}
1131 
1132 			if (dst_reg != BPF_REG_0)
1133 				/* mov rax, dst_reg */
1134 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1135 
1136 			/*
1137 			 * xor edx, edx
1138 			 * equivalent to 'xor rdx, rdx', but one byte less
1139 			 */
1140 			EMIT2(0x31, 0xd2);
1141 
1142 			/* div src_reg */
1143 			maybe_emit_1mod(&prog, src_reg, is64);
1144 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1145 
1146 			if (BPF_OP(insn->code) == BPF_MOD &&
1147 			    dst_reg != BPF_REG_3)
1148 				/* mov dst_reg, rdx */
1149 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1150 			else if (BPF_OP(insn->code) == BPF_DIV &&
1151 				 dst_reg != BPF_REG_0)
1152 				/* mov dst_reg, rax */
1153 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1154 
1155 			if (dst_reg != BPF_REG_3)
1156 				EMIT1(0x5A); /* pop rdx */
1157 			if (dst_reg != BPF_REG_0)
1158 				EMIT1(0x58); /* pop rax */
1159 			break;
1160 		}
1161 
1162 		case BPF_ALU | BPF_MUL | BPF_K:
1163 		case BPF_ALU64 | BPF_MUL | BPF_K:
1164 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1165 				       BPF_CLASS(insn->code) == BPF_ALU64);
1166 
1167 			if (is_imm8(imm32))
1168 				/* imul dst_reg, dst_reg, imm8 */
1169 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1170 				      imm32);
1171 			else
1172 				/* imul dst_reg, dst_reg, imm32 */
1173 				EMIT2_off32(0x69,
1174 					    add_2reg(0xC0, dst_reg, dst_reg),
1175 					    imm32);
1176 			break;
1177 
1178 		case BPF_ALU | BPF_MUL | BPF_X:
1179 		case BPF_ALU64 | BPF_MUL | BPF_X:
1180 			maybe_emit_mod(&prog, src_reg, dst_reg,
1181 				       BPF_CLASS(insn->code) == BPF_ALU64);
1182 
1183 			/* imul dst_reg, src_reg */
1184 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1185 			break;
1186 
1187 			/* Shifts */
1188 		case BPF_ALU | BPF_LSH | BPF_K:
1189 		case BPF_ALU | BPF_RSH | BPF_K:
1190 		case BPF_ALU | BPF_ARSH | BPF_K:
1191 		case BPF_ALU64 | BPF_LSH | BPF_K:
1192 		case BPF_ALU64 | BPF_RSH | BPF_K:
1193 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1194 			maybe_emit_1mod(&prog, dst_reg,
1195 					BPF_CLASS(insn->code) == BPF_ALU64);
1196 
1197 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1198 			if (imm32 == 1)
1199 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1200 			else
1201 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1202 			break;
1203 
1204 		case BPF_ALU | BPF_LSH | BPF_X:
1205 		case BPF_ALU | BPF_RSH | BPF_X:
1206 		case BPF_ALU | BPF_ARSH | BPF_X:
1207 		case BPF_ALU64 | BPF_LSH | BPF_X:
1208 		case BPF_ALU64 | BPF_RSH | BPF_X:
1209 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1210 			/* BMI2 shifts aren't better when shift count is already in rcx */
1211 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1212 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1213 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1214 				u8 op;
1215 
1216 				switch (BPF_OP(insn->code)) {
1217 				case BPF_LSH:
1218 					op = 1; /* prefix 0x66 */
1219 					break;
1220 				case BPF_RSH:
1221 					op = 3; /* prefix 0xf2 */
1222 					break;
1223 				case BPF_ARSH:
1224 					op = 2; /* prefix 0xf3 */
1225 					break;
1226 				}
1227 
1228 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1229 
1230 				break;
1231 			}
1232 
1233 			if (src_reg != BPF_REG_4) { /* common case */
1234 				/* Check for bad case when dst_reg == rcx */
1235 				if (dst_reg == BPF_REG_4) {
1236 					/* mov r11, dst_reg */
1237 					EMIT_mov(AUX_REG, dst_reg);
1238 					dst_reg = AUX_REG;
1239 				} else {
1240 					EMIT1(0x51); /* push rcx */
1241 				}
1242 				/* mov rcx, src_reg */
1243 				EMIT_mov(BPF_REG_4, src_reg);
1244 			}
1245 
1246 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1247 			maybe_emit_1mod(&prog, dst_reg,
1248 					BPF_CLASS(insn->code) == BPF_ALU64);
1249 
1250 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1251 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1252 
1253 			if (src_reg != BPF_REG_4) {
1254 				if (insn->dst_reg == BPF_REG_4)
1255 					/* mov dst_reg, r11 */
1256 					EMIT_mov(insn->dst_reg, AUX_REG);
1257 				else
1258 					EMIT1(0x59); /* pop rcx */
1259 			}
1260 
1261 			break;
1262 
1263 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1264 			switch (imm32) {
1265 			case 16:
1266 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1267 				EMIT1(0x66);
1268 				if (is_ereg(dst_reg))
1269 					EMIT1(0x41);
1270 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1271 
1272 				/* Emit 'movzwl eax, ax' */
1273 				if (is_ereg(dst_reg))
1274 					EMIT3(0x45, 0x0F, 0xB7);
1275 				else
1276 					EMIT2(0x0F, 0xB7);
1277 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1278 				break;
1279 			case 32:
1280 				/* Emit 'bswap eax' to swap lower 4 bytes */
1281 				if (is_ereg(dst_reg))
1282 					EMIT2(0x41, 0x0F);
1283 				else
1284 					EMIT1(0x0F);
1285 				EMIT1(add_1reg(0xC8, dst_reg));
1286 				break;
1287 			case 64:
1288 				/* Emit 'bswap rax' to swap 8 bytes */
1289 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1290 				      add_1reg(0xC8, dst_reg));
1291 				break;
1292 			}
1293 			break;
1294 
1295 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1296 			switch (imm32) {
1297 			case 16:
1298 				/*
1299 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1300 				 * into 64 bit
1301 				 */
1302 				if (is_ereg(dst_reg))
1303 					EMIT3(0x45, 0x0F, 0xB7);
1304 				else
1305 					EMIT2(0x0F, 0xB7);
1306 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1307 				break;
1308 			case 32:
1309 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1310 				if (is_ereg(dst_reg))
1311 					EMIT1(0x45);
1312 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1313 				break;
1314 			case 64:
1315 				/* nop */
1316 				break;
1317 			}
1318 			break;
1319 
1320 			/* speculation barrier */
1321 		case BPF_ST | BPF_NOSPEC:
1322 			EMIT_LFENCE();
1323 			break;
1324 
1325 			/* ST: *(u8*)(dst_reg + off) = imm */
1326 		case BPF_ST | BPF_MEM | BPF_B:
1327 			if (is_ereg(dst_reg))
1328 				EMIT2(0x41, 0xC6);
1329 			else
1330 				EMIT1(0xC6);
1331 			goto st;
1332 		case BPF_ST | BPF_MEM | BPF_H:
1333 			if (is_ereg(dst_reg))
1334 				EMIT3(0x66, 0x41, 0xC7);
1335 			else
1336 				EMIT2(0x66, 0xC7);
1337 			goto st;
1338 		case BPF_ST | BPF_MEM | BPF_W:
1339 			if (is_ereg(dst_reg))
1340 				EMIT2(0x41, 0xC7);
1341 			else
1342 				EMIT1(0xC7);
1343 			goto st;
1344 		case BPF_ST | BPF_MEM | BPF_DW:
1345 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1346 
1347 st:			if (is_imm8(insn->off))
1348 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1349 			else
1350 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1351 
1352 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1353 			break;
1354 
1355 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1356 		case BPF_STX | BPF_MEM | BPF_B:
1357 		case BPF_STX | BPF_MEM | BPF_H:
1358 		case BPF_STX | BPF_MEM | BPF_W:
1359 		case BPF_STX | BPF_MEM | BPF_DW:
1360 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1361 			break;
1362 
1363 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1364 		case BPF_LDX | BPF_MEM | BPF_B:
1365 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1366 		case BPF_LDX | BPF_MEM | BPF_H:
1367 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1368 		case BPF_LDX | BPF_MEM | BPF_W:
1369 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1370 		case BPF_LDX | BPF_MEM | BPF_DW:
1371 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1372 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1373 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1374 				 * add abs(insn->off) to the limit to make sure that negative
1375 				 * offset won't be an issue.
1376 				 * insn->off is s16, so it won't affect valid pointers.
1377 				 */
1378 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1379 				u8 *end_of_jmp1, *end_of_jmp2;
1380 
1381 				/* Conservatively check that src_reg + insn->off is a kernel address:
1382 				 * 1. src_reg + insn->off >= limit
1383 				 * 2. src_reg + insn->off doesn't become small positive.
1384 				 * Cannot do src_reg + insn->off >= limit in one branch,
1385 				 * since it needs two spare registers, but JIT has only one.
1386 				 */
1387 
1388 				/* movabsq r11, limit */
1389 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1390 				EMIT((u32)limit, 4);
1391 				EMIT(limit >> 32, 4);
1392 				/* cmp src_reg, r11 */
1393 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1394 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1395 				/* if unsigned '<' goto end_of_jmp2 */
1396 				EMIT2(X86_JB, 0);
1397 				end_of_jmp1 = prog;
1398 
1399 				/* mov r11, src_reg */
1400 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1401 				/* add r11, insn->off */
1402 				maybe_emit_1mod(&prog, AUX_REG, true);
1403 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1404 				/* jmp if not carry to start_of_ldx
1405 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1406 				 * that has to be rejected.
1407 				 */
1408 				EMIT2(0x73 /* JNC */, 0);
1409 				end_of_jmp2 = prog;
1410 
1411 				/* xor dst_reg, dst_reg */
1412 				emit_mov_imm32(&prog, false, dst_reg, 0);
1413 				/* jmp byte_after_ldx */
1414 				EMIT2(0xEB, 0);
1415 
1416 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1417 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1418 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1419 				start_of_ldx = prog;
1420 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1421 			}
1422 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1423 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1424 				struct exception_table_entry *ex;
1425 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1426 				s64 delta;
1427 
1428 				/* populate jmp_offset for JMP above */
1429 				start_of_ldx[-1] = prog - start_of_ldx;
1430 
1431 				if (!bpf_prog->aux->extable)
1432 					break;
1433 
1434 				if (excnt >= bpf_prog->aux->num_exentries) {
1435 					pr_err("ex gen bug\n");
1436 					return -EFAULT;
1437 				}
1438 				ex = &bpf_prog->aux->extable[excnt++];
1439 
1440 				delta = _insn - (u8 *)&ex->insn;
1441 				if (!is_simm32(delta)) {
1442 					pr_err("extable->insn doesn't fit into 32-bit\n");
1443 					return -EFAULT;
1444 				}
1445 				/* switch ex to rw buffer for writes */
1446 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1447 
1448 				ex->insn = delta;
1449 
1450 				ex->data = EX_TYPE_BPF;
1451 
1452 				if (dst_reg > BPF_REG_9) {
1453 					pr_err("verifier error\n");
1454 					return -EFAULT;
1455 				}
1456 				/*
1457 				 * Compute size of x86 insn and its target dest x86 register.
1458 				 * ex_handler_bpf() will use lower 8 bits to adjust
1459 				 * pt_regs->ip to jump over this x86 instruction
1460 				 * and upper bits to figure out which pt_regs to zero out.
1461 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1462 				 * of 4 bytes will be ignored and rbx will be zero inited.
1463 				 */
1464 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1465 			}
1466 			break;
1467 
1468 		case BPF_STX | BPF_ATOMIC | BPF_W:
1469 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1470 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1471 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1472 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1473 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1474 				u32 real_src_reg = src_reg;
1475 				u32 real_dst_reg = dst_reg;
1476 				u8 *branch_target;
1477 
1478 				/*
1479 				 * Can't be implemented with a single x86 insn.
1480 				 * Need to do a CMPXCHG loop.
1481 				 */
1482 
1483 				/* Will need RAX as a CMPXCHG operand so save R0 */
1484 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1485 				if (src_reg == BPF_REG_0)
1486 					real_src_reg = BPF_REG_AX;
1487 				if (dst_reg == BPF_REG_0)
1488 					real_dst_reg = BPF_REG_AX;
1489 
1490 				branch_target = prog;
1491 				/* Load old value */
1492 				emit_ldx(&prog, BPF_SIZE(insn->code),
1493 					 BPF_REG_0, real_dst_reg, insn->off);
1494 				/*
1495 				 * Perform the (commutative) operation locally,
1496 				 * put the result in the AUX_REG.
1497 				 */
1498 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1499 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1500 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1501 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1502 				/* Attempt to swap in new value */
1503 				err = emit_atomic(&prog, BPF_CMPXCHG,
1504 						  real_dst_reg, AUX_REG,
1505 						  insn->off,
1506 						  BPF_SIZE(insn->code));
1507 				if (WARN_ON(err))
1508 					return err;
1509 				/*
1510 				 * ZF tells us whether we won the race. If it's
1511 				 * cleared we need to try again.
1512 				 */
1513 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1514 				/* Return the pre-modification value */
1515 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1516 				/* Restore R0 after clobbering RAX */
1517 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1518 				break;
1519 			}
1520 
1521 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1522 					  insn->off, BPF_SIZE(insn->code));
1523 			if (err)
1524 				return err;
1525 			break;
1526 
1527 			/* call */
1528 		case BPF_JMP | BPF_CALL: {
1529 			int offs;
1530 
1531 			func = (u8 *) __bpf_call_base + imm32;
1532 			if (tail_call_reachable) {
1533 				/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1534 				EMIT3_off32(0x48, 0x8B, 0x85,
1535 					    -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1536 				if (!imm32)
1537 					return -EINVAL;
1538 				offs = 7 + x86_call_depth_emit_accounting(&prog, func);
1539 			} else {
1540 				if (!imm32)
1541 					return -EINVAL;
1542 				offs = x86_call_depth_emit_accounting(&prog, func);
1543 			}
1544 			if (emit_call(&prog, func, image + addrs[i - 1] + offs))
1545 				return -EINVAL;
1546 			break;
1547 		}
1548 
1549 		case BPF_JMP | BPF_TAIL_CALL:
1550 			if (imm32)
1551 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1552 							  &prog, image + addrs[i - 1],
1553 							  callee_regs_used,
1554 							  bpf_prog->aux->stack_depth,
1555 							  ctx);
1556 			else
1557 				emit_bpf_tail_call_indirect(&prog,
1558 							    callee_regs_used,
1559 							    bpf_prog->aux->stack_depth,
1560 							    image + addrs[i - 1],
1561 							    ctx);
1562 			break;
1563 
1564 			/* cond jump */
1565 		case BPF_JMP | BPF_JEQ | BPF_X:
1566 		case BPF_JMP | BPF_JNE | BPF_X:
1567 		case BPF_JMP | BPF_JGT | BPF_X:
1568 		case BPF_JMP | BPF_JLT | BPF_X:
1569 		case BPF_JMP | BPF_JGE | BPF_X:
1570 		case BPF_JMP | BPF_JLE | BPF_X:
1571 		case BPF_JMP | BPF_JSGT | BPF_X:
1572 		case BPF_JMP | BPF_JSLT | BPF_X:
1573 		case BPF_JMP | BPF_JSGE | BPF_X:
1574 		case BPF_JMP | BPF_JSLE | BPF_X:
1575 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1576 		case BPF_JMP32 | BPF_JNE | BPF_X:
1577 		case BPF_JMP32 | BPF_JGT | BPF_X:
1578 		case BPF_JMP32 | BPF_JLT | BPF_X:
1579 		case BPF_JMP32 | BPF_JGE | BPF_X:
1580 		case BPF_JMP32 | BPF_JLE | BPF_X:
1581 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1582 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1583 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1584 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1585 			/* cmp dst_reg, src_reg */
1586 			maybe_emit_mod(&prog, dst_reg, src_reg,
1587 				       BPF_CLASS(insn->code) == BPF_JMP);
1588 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1589 			goto emit_cond_jmp;
1590 
1591 		case BPF_JMP | BPF_JSET | BPF_X:
1592 		case BPF_JMP32 | BPF_JSET | BPF_X:
1593 			/* test dst_reg, src_reg */
1594 			maybe_emit_mod(&prog, dst_reg, src_reg,
1595 				       BPF_CLASS(insn->code) == BPF_JMP);
1596 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1597 			goto emit_cond_jmp;
1598 
1599 		case BPF_JMP | BPF_JSET | BPF_K:
1600 		case BPF_JMP32 | BPF_JSET | BPF_K:
1601 			/* test dst_reg, imm32 */
1602 			maybe_emit_1mod(&prog, dst_reg,
1603 					BPF_CLASS(insn->code) == BPF_JMP);
1604 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1605 			goto emit_cond_jmp;
1606 
1607 		case BPF_JMP | BPF_JEQ | BPF_K:
1608 		case BPF_JMP | BPF_JNE | BPF_K:
1609 		case BPF_JMP | BPF_JGT | BPF_K:
1610 		case BPF_JMP | BPF_JLT | BPF_K:
1611 		case BPF_JMP | BPF_JGE | BPF_K:
1612 		case BPF_JMP | BPF_JLE | BPF_K:
1613 		case BPF_JMP | BPF_JSGT | BPF_K:
1614 		case BPF_JMP | BPF_JSLT | BPF_K:
1615 		case BPF_JMP | BPF_JSGE | BPF_K:
1616 		case BPF_JMP | BPF_JSLE | BPF_K:
1617 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1618 		case BPF_JMP32 | BPF_JNE | BPF_K:
1619 		case BPF_JMP32 | BPF_JGT | BPF_K:
1620 		case BPF_JMP32 | BPF_JLT | BPF_K:
1621 		case BPF_JMP32 | BPF_JGE | BPF_K:
1622 		case BPF_JMP32 | BPF_JLE | BPF_K:
1623 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1624 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1625 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1626 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1627 			/* test dst_reg, dst_reg to save one extra byte */
1628 			if (imm32 == 0) {
1629 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1630 					       BPF_CLASS(insn->code) == BPF_JMP);
1631 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1632 				goto emit_cond_jmp;
1633 			}
1634 
1635 			/* cmp dst_reg, imm8/32 */
1636 			maybe_emit_1mod(&prog, dst_reg,
1637 					BPF_CLASS(insn->code) == BPF_JMP);
1638 
1639 			if (is_imm8(imm32))
1640 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1641 			else
1642 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1643 
1644 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1645 			switch (BPF_OP(insn->code)) {
1646 			case BPF_JEQ:
1647 				jmp_cond = X86_JE;
1648 				break;
1649 			case BPF_JSET:
1650 			case BPF_JNE:
1651 				jmp_cond = X86_JNE;
1652 				break;
1653 			case BPF_JGT:
1654 				/* GT is unsigned '>', JA in x86 */
1655 				jmp_cond = X86_JA;
1656 				break;
1657 			case BPF_JLT:
1658 				/* LT is unsigned '<', JB in x86 */
1659 				jmp_cond = X86_JB;
1660 				break;
1661 			case BPF_JGE:
1662 				/* GE is unsigned '>=', JAE in x86 */
1663 				jmp_cond = X86_JAE;
1664 				break;
1665 			case BPF_JLE:
1666 				/* LE is unsigned '<=', JBE in x86 */
1667 				jmp_cond = X86_JBE;
1668 				break;
1669 			case BPF_JSGT:
1670 				/* Signed '>', GT in x86 */
1671 				jmp_cond = X86_JG;
1672 				break;
1673 			case BPF_JSLT:
1674 				/* Signed '<', LT in x86 */
1675 				jmp_cond = X86_JL;
1676 				break;
1677 			case BPF_JSGE:
1678 				/* Signed '>=', GE in x86 */
1679 				jmp_cond = X86_JGE;
1680 				break;
1681 			case BPF_JSLE:
1682 				/* Signed '<=', LE in x86 */
1683 				jmp_cond = X86_JLE;
1684 				break;
1685 			default: /* to silence GCC warning */
1686 				return -EFAULT;
1687 			}
1688 			jmp_offset = addrs[i + insn->off] - addrs[i];
1689 			if (is_imm8(jmp_offset)) {
1690 				if (jmp_padding) {
1691 					/* To keep the jmp_offset valid, the extra bytes are
1692 					 * padded before the jump insn, so we subtract the
1693 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1694 					 *
1695 					 * If the previous pass already emits an imm8
1696 					 * jmp_cond, then this BPF insn won't shrink, so
1697 					 * "nops" is 0.
1698 					 *
1699 					 * On the other hand, if the previous pass emits an
1700 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1701 					 * keep the image from shrinking further.
1702 					 *
1703 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1704 					 *     is 2 bytes, so the size difference is 4 bytes.
1705 					 */
1706 					nops = INSN_SZ_DIFF - 2;
1707 					if (nops != 0 && nops != 4) {
1708 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1709 						       nops);
1710 						return -EFAULT;
1711 					}
1712 					emit_nops(&prog, nops);
1713 				}
1714 				EMIT2(jmp_cond, jmp_offset);
1715 			} else if (is_simm32(jmp_offset)) {
1716 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1717 			} else {
1718 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1719 				return -EFAULT;
1720 			}
1721 
1722 			break;
1723 
1724 		case BPF_JMP | BPF_JA:
1725 			if (insn->off == -1)
1726 				/* -1 jmp instructions will always jump
1727 				 * backwards two bytes. Explicitly handling
1728 				 * this case avoids wasting too many passes
1729 				 * when there are long sequences of replaced
1730 				 * dead code.
1731 				 */
1732 				jmp_offset = -2;
1733 			else
1734 				jmp_offset = addrs[i + insn->off] - addrs[i];
1735 
1736 			if (!jmp_offset) {
1737 				/*
1738 				 * If jmp_padding is enabled, the extra nops will
1739 				 * be inserted. Otherwise, optimize out nop jumps.
1740 				 */
1741 				if (jmp_padding) {
1742 					/* There are 3 possible conditions.
1743 					 * (1) This BPF_JA is already optimized out in
1744 					 *     the previous run, so there is no need
1745 					 *     to pad any extra byte (0 byte).
1746 					 * (2) The previous pass emits an imm8 jmp,
1747 					 *     so we pad 2 bytes to match the previous
1748 					 *     insn size.
1749 					 * (3) Similarly, the previous pass emits an
1750 					 *     imm32 jmp, and 5 bytes is padded.
1751 					 */
1752 					nops = INSN_SZ_DIFF;
1753 					if (nops != 0 && nops != 2 && nops != 5) {
1754 						pr_err("unexpected nop jump padding: %d bytes\n",
1755 						       nops);
1756 						return -EFAULT;
1757 					}
1758 					emit_nops(&prog, nops);
1759 				}
1760 				break;
1761 			}
1762 emit_jmp:
1763 			if (is_imm8(jmp_offset)) {
1764 				if (jmp_padding) {
1765 					/* To avoid breaking jmp_offset, the extra bytes
1766 					 * are padded before the actual jmp insn, so
1767 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1768 					 *
1769 					 * If the previous pass already emits an imm8
1770 					 * jmp, there is nothing to pad (0 byte).
1771 					 *
1772 					 * If it emits an imm32 jmp (5 bytes) previously
1773 					 * and now an imm8 jmp (2 bytes), then we pad
1774 					 * (5 - 2 = 3) bytes to stop the image from
1775 					 * shrinking further.
1776 					 */
1777 					nops = INSN_SZ_DIFF - 2;
1778 					if (nops != 0 && nops != 3) {
1779 						pr_err("unexpected jump padding: %d bytes\n",
1780 						       nops);
1781 						return -EFAULT;
1782 					}
1783 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1784 				}
1785 				EMIT2(0xEB, jmp_offset);
1786 			} else if (is_simm32(jmp_offset)) {
1787 				EMIT1_off32(0xE9, jmp_offset);
1788 			} else {
1789 				pr_err("jmp gen bug %llx\n", jmp_offset);
1790 				return -EFAULT;
1791 			}
1792 			break;
1793 
1794 		case BPF_JMP | BPF_EXIT:
1795 			if (seen_exit) {
1796 				jmp_offset = ctx->cleanup_addr - addrs[i];
1797 				goto emit_jmp;
1798 			}
1799 			seen_exit = true;
1800 			/* Update cleanup_addr */
1801 			ctx->cleanup_addr = proglen;
1802 			pop_callee_regs(&prog, callee_regs_used);
1803 			EMIT1(0xC9);         /* leave */
1804 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1805 			break;
1806 
1807 		default:
1808 			/*
1809 			 * By design x86-64 JIT should support all BPF instructions.
1810 			 * This error will be seen if new instruction was added
1811 			 * to the interpreter, but not to the JIT, or if there is
1812 			 * junk in bpf_prog.
1813 			 */
1814 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1815 			return -EINVAL;
1816 		}
1817 
1818 		ilen = prog - temp;
1819 		if (ilen > BPF_MAX_INSN_SIZE) {
1820 			pr_err("bpf_jit: fatal insn size error\n");
1821 			return -EFAULT;
1822 		}
1823 
1824 		if (image) {
1825 			/*
1826 			 * When populating the image, assert that:
1827 			 *
1828 			 *  i) We do not write beyond the allocated space, and
1829 			 * ii) addrs[i] did not change from the prior run, in order
1830 			 *     to validate assumptions made for computing branch
1831 			 *     displacements.
1832 			 */
1833 			if (unlikely(proglen + ilen > oldproglen ||
1834 				     proglen + ilen != addrs[i])) {
1835 				pr_err("bpf_jit: fatal error\n");
1836 				return -EFAULT;
1837 			}
1838 			memcpy(rw_image + proglen, temp, ilen);
1839 		}
1840 		proglen += ilen;
1841 		addrs[i] = proglen;
1842 		prog = temp;
1843 	}
1844 
1845 	if (image && excnt != bpf_prog->aux->num_exentries) {
1846 		pr_err("extable is not populated\n");
1847 		return -EFAULT;
1848 	}
1849 	return proglen;
1850 }
1851 
1852 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1853 		      int stack_size)
1854 {
1855 	int i, j, arg_size, nr_regs;
1856 	/* Store function arguments to stack.
1857 	 * For a function that accepts two pointers the sequence will be:
1858 	 * mov QWORD PTR [rbp-0x10],rdi
1859 	 * mov QWORD PTR [rbp-0x8],rsi
1860 	 */
1861 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1862 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1863 			nr_regs = (m->arg_size[i] + 7) / 8;
1864 			arg_size = 8;
1865 		} else {
1866 			nr_regs = 1;
1867 			arg_size = m->arg_size[i];
1868 		}
1869 
1870 		while (nr_regs) {
1871 			emit_stx(prog, bytes_to_bpf_size(arg_size),
1872 				 BPF_REG_FP,
1873 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1874 				 -(stack_size - j * 8));
1875 			nr_regs--;
1876 			j++;
1877 		}
1878 	}
1879 }
1880 
1881 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1882 			 int stack_size)
1883 {
1884 	int i, j, arg_size, nr_regs;
1885 
1886 	/* Restore function arguments from stack.
1887 	 * For a function that accepts two pointers the sequence will be:
1888 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1889 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1890 	 */
1891 	for (i = 0, j = 0; i < min(nr_args, 6); i++) {
1892 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG) {
1893 			nr_regs = (m->arg_size[i] + 7) / 8;
1894 			arg_size = 8;
1895 		} else {
1896 			nr_regs = 1;
1897 			arg_size = m->arg_size[i];
1898 		}
1899 
1900 		while (nr_regs) {
1901 			emit_ldx(prog, bytes_to_bpf_size(arg_size),
1902 				 j == 5 ? X86_REG_R9 : BPF_REG_1 + j,
1903 				 BPF_REG_FP,
1904 				 -(stack_size - j * 8));
1905 			nr_regs--;
1906 			j++;
1907 		}
1908 	}
1909 }
1910 
1911 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1912 			   struct bpf_tramp_link *l, int stack_size,
1913 			   int run_ctx_off, bool save_ret)
1914 {
1915 	u8 *prog = *pprog;
1916 	u8 *jmp_insn;
1917 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1918 	struct bpf_prog *p = l->link.prog;
1919 	u64 cookie = l->cookie;
1920 
1921 	/* mov rdi, cookie */
1922 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
1923 
1924 	/* Prepare struct bpf_tramp_run_ctx.
1925 	 *
1926 	 * bpf_tramp_run_ctx is already preserved by
1927 	 * arch_prepare_bpf_trampoline().
1928 	 *
1929 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
1930 	 */
1931 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
1932 
1933 	/* arg1: mov rdi, progs[i] */
1934 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1935 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
1936 	EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
1937 
1938 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), prog))
1939 		return -EINVAL;
1940 	/* remember prog start time returned by __bpf_prog_enter */
1941 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1942 
1943 	/* if (__bpf_prog_enter*(prog) == 0)
1944 	 *	goto skip_exec_of_prog;
1945 	 */
1946 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1947 	/* emit 2 nops that will be replaced with JE insn */
1948 	jmp_insn = prog;
1949 	emit_nops(&prog, 2);
1950 
1951 	/* arg1: lea rdi, [rbp - stack_size] */
1952 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1953 	/* arg2: progs[i]->insnsi for interpreter */
1954 	if (!p->jited)
1955 		emit_mov_imm64(&prog, BPF_REG_2,
1956 			       (long) p->insnsi >> 32,
1957 			       (u32) (long) p->insnsi);
1958 	/* call JITed bpf program or interpreter */
1959 	if (emit_rsb_call(&prog, p->bpf_func, prog))
1960 		return -EINVAL;
1961 
1962 	/*
1963 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1964 	 * of the previous call which is then passed on the stack to
1965 	 * the next BPF program.
1966 	 *
1967 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1968 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1969 	 */
1970 	if (save_ret)
1971 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1972 
1973 	/* replace 2 nops with JE insn, since jmp target is known */
1974 	jmp_insn[0] = X86_JE;
1975 	jmp_insn[1] = prog - jmp_insn - 2;
1976 
1977 	/* arg1: mov rdi, progs[i] */
1978 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1979 	/* arg2: mov rsi, rbx <- start time in nsec */
1980 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1981 	/* arg3: lea rdx, [rbp - run_ctx_off] */
1982 	EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
1983 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), prog))
1984 		return -EINVAL;
1985 
1986 	*pprog = prog;
1987 	return 0;
1988 }
1989 
1990 static void emit_align(u8 **pprog, u32 align)
1991 {
1992 	u8 *target, *prog = *pprog;
1993 
1994 	target = PTR_ALIGN(prog, align);
1995 	if (target != prog)
1996 		emit_nops(&prog, target - prog);
1997 
1998 	*pprog = prog;
1999 }
2000 
2001 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2002 {
2003 	u8 *prog = *pprog;
2004 	s64 offset;
2005 
2006 	offset = func - (ip + 2 + 4);
2007 	if (!is_simm32(offset)) {
2008 		pr_err("Target %p is out of range\n", func);
2009 		return -EINVAL;
2010 	}
2011 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2012 	*pprog = prog;
2013 	return 0;
2014 }
2015 
2016 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2017 		      struct bpf_tramp_links *tl, int stack_size,
2018 		      int run_ctx_off, bool save_ret)
2019 {
2020 	int i;
2021 	u8 *prog = *pprog;
2022 
2023 	for (i = 0; i < tl->nr_links; i++) {
2024 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2025 				    run_ctx_off, save_ret))
2026 			return -EINVAL;
2027 	}
2028 	*pprog = prog;
2029 	return 0;
2030 }
2031 
2032 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2033 			      struct bpf_tramp_links *tl, int stack_size,
2034 			      int run_ctx_off, u8 **branches)
2035 {
2036 	u8 *prog = *pprog;
2037 	int i;
2038 
2039 	/* The first fmod_ret program will receive a garbage return value.
2040 	 * Set this to 0 to avoid confusing the program.
2041 	 */
2042 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2043 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2044 	for (i = 0; i < tl->nr_links; i++) {
2045 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
2046 			return -EINVAL;
2047 
2048 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2049 		 * if (*(u64 *)(rbp - 8) !=  0)
2050 		 *	goto do_fexit;
2051 		 */
2052 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2053 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2054 
2055 		/* Save the location of the branch and Generate 6 nops
2056 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2057 		 * are replaced with a conditional jump once do_fexit (i.e. the
2058 		 * start of the fexit invocation) is finalized.
2059 		 */
2060 		branches[i] = prog;
2061 		emit_nops(&prog, 4 + 2);
2062 	}
2063 
2064 	*pprog = prog;
2065 	return 0;
2066 }
2067 
2068 /* Example:
2069  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2070  * its 'struct btf_func_model' will be nr_args=2
2071  * The assembly code when eth_type_trans is executing after trampoline:
2072  *
2073  * push rbp
2074  * mov rbp, rsp
2075  * sub rsp, 16                     // space for skb and dev
2076  * push rbx                        // temp regs to pass start time
2077  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2078  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2079  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2080  * mov rbx, rax                    // remember start time in bpf stats are enabled
2081  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2082  * call addr_of_jited_FENTRY_prog
2083  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2084  * mov rsi, rbx                    // prog start time
2085  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2086  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2087  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2088  * pop rbx
2089  * leave
2090  * ret
2091  *
2092  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2093  * replaced with 'call generated_bpf_trampoline'. When it returns
2094  * eth_type_trans will continue executing with original skb and dev pointers.
2095  *
2096  * The assembly code when eth_type_trans is called from trampoline:
2097  *
2098  * push rbp
2099  * mov rbp, rsp
2100  * sub rsp, 24                     // space for skb, dev, return value
2101  * push rbx                        // temp regs to pass start time
2102  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2103  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2104  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2105  * mov rbx, rax                    // remember start time if bpf stats are enabled
2106  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2107  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2108  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2109  * mov rsi, rbx                    // prog start time
2110  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2111  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2112  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2113  * call eth_type_trans+5           // execute body of eth_type_trans
2114  * mov qword ptr [rbp - 8], rax    // save return value
2115  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2116  * mov rbx, rax                    // remember start time in bpf stats are enabled
2117  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2118  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2119  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2120  * mov rsi, rbx                    // prog start time
2121  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2122  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2123  * pop rbx
2124  * leave
2125  * add rsp, 8                      // skip eth_type_trans's frame
2126  * ret                             // return to its caller
2127  */
2128 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2129 				const struct btf_func_model *m, u32 flags,
2130 				struct bpf_tramp_links *tlinks,
2131 				void *func_addr)
2132 {
2133 	int ret, i, nr_args = m->nr_args, extra_nregs = 0;
2134 	int regs_off, ip_off, args_off, stack_size = nr_args * 8, run_ctx_off;
2135 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2136 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2137 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2138 	void *orig_call = func_addr;
2139 	u8 **branches = NULL;
2140 	u8 *prog;
2141 	bool save_ret;
2142 
2143 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2144 	if (nr_args > 6)
2145 		return -ENOTSUPP;
2146 
2147 	for (i = 0; i < MAX_BPF_FUNC_ARGS; i++) {
2148 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2149 			extra_nregs += (m->arg_size[i] + 7) / 8 - 1;
2150 	}
2151 	if (nr_args + extra_nregs > 6)
2152 		return -ENOTSUPP;
2153 	stack_size += extra_nregs * 8;
2154 
2155 	/* Generated trampoline stack layout:
2156 	 *
2157 	 * RBP + 8         [ return address  ]
2158 	 * RBP + 0         [ RBP             ]
2159 	 *
2160 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2161 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2162 	 *
2163 	 *                 [ reg_argN        ]  always
2164 	 *                 [ ...             ]
2165 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2166 	 *
2167 	 * RBP - args_off  [ arg regs count  ]  always
2168 	 *
2169 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2170 	 *
2171 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2172 	 */
2173 
2174 	/* room for return value of orig_call or fentry prog */
2175 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2176 	if (save_ret)
2177 		stack_size += 8;
2178 
2179 	regs_off = stack_size;
2180 
2181 	/* args count  */
2182 	stack_size += 8;
2183 	args_off = stack_size;
2184 
2185 	if (flags & BPF_TRAMP_F_IP_ARG)
2186 		stack_size += 8; /* room for IP address argument */
2187 
2188 	ip_off = stack_size;
2189 
2190 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2191 	run_ctx_off = stack_size;
2192 
2193 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2194 		/* skip patched call instruction and point orig_call to actual
2195 		 * body of the kernel function.
2196 		 */
2197 		if (is_endbr(*(u32 *)orig_call))
2198 			orig_call += ENDBR_INSN_SIZE;
2199 		orig_call += X86_PATCH_SIZE;
2200 	}
2201 
2202 	prog = image;
2203 
2204 	EMIT_ENDBR();
2205 	/*
2206 	 * This is the direct-call trampoline, as such it needs accounting
2207 	 * for the __fentry__ call.
2208 	 */
2209 	x86_call_depth_emit_accounting(&prog, NULL);
2210 	EMIT1(0x55);		 /* push rbp */
2211 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2212 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2213 	EMIT1(0x53);		 /* push rbx */
2214 
2215 	/* Store number of argument registers of the traced function:
2216 	 *   mov rax, nr_args + extra_nregs
2217 	 *   mov QWORD PTR [rbp - args_off], rax
2218 	 */
2219 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args + extra_nregs);
2220 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2221 
2222 	if (flags & BPF_TRAMP_F_IP_ARG) {
2223 		/* Store IP address of the traced function:
2224 		 * movabsq rax, func_addr
2225 		 * mov QWORD PTR [rbp - ip_off], rax
2226 		 */
2227 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2228 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2229 	}
2230 
2231 	save_regs(m, &prog, nr_args, regs_off);
2232 
2233 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2234 		/* arg1: mov rdi, im */
2235 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2236 		if (emit_rsb_call(&prog, __bpf_tramp_enter, prog)) {
2237 			ret = -EINVAL;
2238 			goto cleanup;
2239 		}
2240 	}
2241 
2242 	if (fentry->nr_links)
2243 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2244 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2245 			return -EINVAL;
2246 
2247 	if (fmod_ret->nr_links) {
2248 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2249 				   GFP_KERNEL);
2250 		if (!branches)
2251 			return -ENOMEM;
2252 
2253 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2254 				       run_ctx_off, branches)) {
2255 			ret = -EINVAL;
2256 			goto cleanup;
2257 		}
2258 	}
2259 
2260 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2261 		restore_regs(m, &prog, nr_args, regs_off);
2262 
2263 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2264 			emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2265 			EMIT2(0xff, 0xd0); /* call *rax */
2266 		} else {
2267 			/* call original function */
2268 			if (emit_rsb_call(&prog, orig_call, prog)) {
2269 				ret = -EINVAL;
2270 				goto cleanup;
2271 			}
2272 		}
2273 		/* remember return value in a stack for bpf prog to access */
2274 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2275 		im->ip_after_call = prog;
2276 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2277 		prog += X86_PATCH_SIZE;
2278 	}
2279 
2280 	if (fmod_ret->nr_links) {
2281 		/* From Intel 64 and IA-32 Architectures Optimization
2282 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2283 		 * Coding Rule 11: All branch targets should be 16-byte
2284 		 * aligned.
2285 		 */
2286 		emit_align(&prog, 16);
2287 		/* Update the branches saved in invoke_bpf_mod_ret with the
2288 		 * aligned address of do_fexit.
2289 		 */
2290 		for (i = 0; i < fmod_ret->nr_links; i++)
2291 			emit_cond_near_jump(&branches[i], prog, branches[i],
2292 					    X86_JNE);
2293 	}
2294 
2295 	if (fexit->nr_links)
2296 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2297 			ret = -EINVAL;
2298 			goto cleanup;
2299 		}
2300 
2301 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2302 		restore_regs(m, &prog, nr_args, regs_off);
2303 
2304 	/* This needs to be done regardless. If there were fmod_ret programs,
2305 	 * the return value is only updated on the stack and still needs to be
2306 	 * restored to R0.
2307 	 */
2308 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2309 		im->ip_epilogue = prog;
2310 		/* arg1: mov rdi, im */
2311 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2312 		if (emit_rsb_call(&prog, __bpf_tramp_exit, prog)) {
2313 			ret = -EINVAL;
2314 			goto cleanup;
2315 		}
2316 	}
2317 	/* restore return value of orig_call or fentry prog back into RAX */
2318 	if (save_ret)
2319 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2320 
2321 	EMIT1(0x5B); /* pop rbx */
2322 	EMIT1(0xC9); /* leave */
2323 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2324 		/* skip our return address and return to parent */
2325 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2326 	emit_return(&prog, prog);
2327 	/* Make sure the trampoline generation logic doesn't overflow */
2328 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2329 		ret = -EFAULT;
2330 		goto cleanup;
2331 	}
2332 	ret = prog - (u8 *)image;
2333 
2334 cleanup:
2335 	kfree(branches);
2336 	return ret;
2337 }
2338 
2339 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2340 {
2341 	u8 *jg_reloc, *prog = *pprog;
2342 	int pivot, err, jg_bytes = 1;
2343 	s64 jg_offset;
2344 
2345 	if (a == b) {
2346 		/* Leaf node of recursion, i.e. not a range of indices
2347 		 * anymore.
2348 		 */
2349 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2350 		if (!is_simm32(progs[a]))
2351 			return -1;
2352 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2353 			    progs[a]);
2354 		err = emit_cond_near_jump(&prog,	/* je func */
2355 					  (void *)progs[a], image + (prog - buf),
2356 					  X86_JE);
2357 		if (err)
2358 			return err;
2359 
2360 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2361 
2362 		*pprog = prog;
2363 		return 0;
2364 	}
2365 
2366 	/* Not a leaf node, so we pivot, and recursively descend into
2367 	 * the lower and upper ranges.
2368 	 */
2369 	pivot = (b - a) / 2;
2370 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2371 	if (!is_simm32(progs[a + pivot]))
2372 		return -1;
2373 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2374 
2375 	if (pivot > 2) {				/* jg upper_part */
2376 		/* Require near jump. */
2377 		jg_bytes = 4;
2378 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2379 	} else {
2380 		EMIT2(X86_JG, 0);
2381 	}
2382 	jg_reloc = prog;
2383 
2384 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2385 				  progs, image, buf);
2386 	if (err)
2387 		return err;
2388 
2389 	/* From Intel 64 and IA-32 Architectures Optimization
2390 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2391 	 * Coding Rule 11: All branch targets should be 16-byte
2392 	 * aligned.
2393 	 */
2394 	emit_align(&prog, 16);
2395 	jg_offset = prog - jg_reloc;
2396 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2397 
2398 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2399 				  b, progs, image, buf);
2400 	if (err)
2401 		return err;
2402 
2403 	*pprog = prog;
2404 	return 0;
2405 }
2406 
2407 static int cmp_ips(const void *a, const void *b)
2408 {
2409 	const s64 *ipa = a;
2410 	const s64 *ipb = b;
2411 
2412 	if (*ipa > *ipb)
2413 		return 1;
2414 	if (*ipa < *ipb)
2415 		return -1;
2416 	return 0;
2417 }
2418 
2419 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2420 {
2421 	u8 *prog = buf;
2422 
2423 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2424 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2425 }
2426 
2427 struct x64_jit_data {
2428 	struct bpf_binary_header *rw_header;
2429 	struct bpf_binary_header *header;
2430 	int *addrs;
2431 	u8 *image;
2432 	int proglen;
2433 	struct jit_context ctx;
2434 };
2435 
2436 #define MAX_PASSES 20
2437 #define PADDING_PASSES (MAX_PASSES - 5)
2438 
2439 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2440 {
2441 	struct bpf_binary_header *rw_header = NULL;
2442 	struct bpf_binary_header *header = NULL;
2443 	struct bpf_prog *tmp, *orig_prog = prog;
2444 	struct x64_jit_data *jit_data;
2445 	int proglen, oldproglen = 0;
2446 	struct jit_context ctx = {};
2447 	bool tmp_blinded = false;
2448 	bool extra_pass = false;
2449 	bool padding = false;
2450 	u8 *rw_image = NULL;
2451 	u8 *image = NULL;
2452 	int *addrs;
2453 	int pass;
2454 	int i;
2455 
2456 	if (!prog->jit_requested)
2457 		return orig_prog;
2458 
2459 	tmp = bpf_jit_blind_constants(prog);
2460 	/*
2461 	 * If blinding was requested and we failed during blinding,
2462 	 * we must fall back to the interpreter.
2463 	 */
2464 	if (IS_ERR(tmp))
2465 		return orig_prog;
2466 	if (tmp != prog) {
2467 		tmp_blinded = true;
2468 		prog = tmp;
2469 	}
2470 
2471 	jit_data = prog->aux->jit_data;
2472 	if (!jit_data) {
2473 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2474 		if (!jit_data) {
2475 			prog = orig_prog;
2476 			goto out;
2477 		}
2478 		prog->aux->jit_data = jit_data;
2479 	}
2480 	addrs = jit_data->addrs;
2481 	if (addrs) {
2482 		ctx = jit_data->ctx;
2483 		oldproglen = jit_data->proglen;
2484 		image = jit_data->image;
2485 		header = jit_data->header;
2486 		rw_header = jit_data->rw_header;
2487 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2488 		extra_pass = true;
2489 		padding = true;
2490 		goto skip_init_addrs;
2491 	}
2492 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2493 	if (!addrs) {
2494 		prog = orig_prog;
2495 		goto out_addrs;
2496 	}
2497 
2498 	/*
2499 	 * Before first pass, make a rough estimation of addrs[]
2500 	 * each BPF instruction is translated to less than 64 bytes
2501 	 */
2502 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2503 		proglen += 64;
2504 		addrs[i] = proglen;
2505 	}
2506 	ctx.cleanup_addr = proglen;
2507 skip_init_addrs:
2508 
2509 	/*
2510 	 * JITed image shrinks with every pass and the loop iterates
2511 	 * until the image stops shrinking. Very large BPF programs
2512 	 * may converge on the last pass. In such case do one more
2513 	 * pass to emit the final image.
2514 	 */
2515 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2516 		if (!padding && pass >= PADDING_PASSES)
2517 			padding = true;
2518 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2519 		if (proglen <= 0) {
2520 out_image:
2521 			image = NULL;
2522 			if (header) {
2523 				bpf_arch_text_copy(&header->size, &rw_header->size,
2524 						   sizeof(rw_header->size));
2525 				bpf_jit_binary_pack_free(header, rw_header);
2526 			}
2527 			/* Fall back to interpreter mode */
2528 			prog = orig_prog;
2529 			if (extra_pass) {
2530 				prog->bpf_func = NULL;
2531 				prog->jited = 0;
2532 				prog->jited_len = 0;
2533 			}
2534 			goto out_addrs;
2535 		}
2536 		if (image) {
2537 			if (proglen != oldproglen) {
2538 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2539 				       proglen, oldproglen);
2540 				goto out_image;
2541 			}
2542 			break;
2543 		}
2544 		if (proglen == oldproglen) {
2545 			/*
2546 			 * The number of entries in extable is the number of BPF_LDX
2547 			 * insns that access kernel memory via "pointer to BTF type".
2548 			 * The verifier changed their opcode from LDX|MEM|size
2549 			 * to LDX|PROBE_MEM|size to make JITing easier.
2550 			 */
2551 			u32 align = __alignof__(struct exception_table_entry);
2552 			u32 extable_size = prog->aux->num_exentries *
2553 				sizeof(struct exception_table_entry);
2554 
2555 			/* allocate module memory for x86 insns and extable */
2556 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2557 							   &image, align, &rw_header, &rw_image,
2558 							   jit_fill_hole);
2559 			if (!header) {
2560 				prog = orig_prog;
2561 				goto out_addrs;
2562 			}
2563 			prog->aux->extable = (void *) image + roundup(proglen, align);
2564 		}
2565 		oldproglen = proglen;
2566 		cond_resched();
2567 	}
2568 
2569 	if (bpf_jit_enable > 1)
2570 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2571 
2572 	if (image) {
2573 		if (!prog->is_func || extra_pass) {
2574 			/*
2575 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2576 			 *   1) header is not pointing to proper module memory;
2577 			 *   2) the arch doesn't support bpf_arch_text_copy().
2578 			 *
2579 			 * Both cases are serious bugs and justify WARN_ON.
2580 			 */
2581 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2582 				/* header has been freed */
2583 				header = NULL;
2584 				goto out_image;
2585 			}
2586 
2587 			bpf_tail_call_direct_fixup(prog);
2588 		} else {
2589 			jit_data->addrs = addrs;
2590 			jit_data->ctx = ctx;
2591 			jit_data->proglen = proglen;
2592 			jit_data->image = image;
2593 			jit_data->header = header;
2594 			jit_data->rw_header = rw_header;
2595 		}
2596 		prog->bpf_func = (void *)image;
2597 		prog->jited = 1;
2598 		prog->jited_len = proglen;
2599 	} else {
2600 		prog = orig_prog;
2601 	}
2602 
2603 	if (!image || !prog->is_func || extra_pass) {
2604 		if (image)
2605 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2606 out_addrs:
2607 		kvfree(addrs);
2608 		kfree(jit_data);
2609 		prog->aux->jit_data = NULL;
2610 	}
2611 out:
2612 	if (tmp_blinded)
2613 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2614 					   tmp : orig_prog);
2615 	return prog;
2616 }
2617 
2618 bool bpf_jit_supports_kfunc_call(void)
2619 {
2620 	return true;
2621 }
2622 
2623 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2624 {
2625 	if (text_poke_copy(dst, src, len) == NULL)
2626 		return ERR_PTR(-EINVAL);
2627 	return dst;
2628 }
2629 
2630 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2631 bool bpf_jit_supports_subprog_tailcalls(void)
2632 {
2633 	return true;
2634 }
2635 
2636 void bpf_jit_free(struct bpf_prog *prog)
2637 {
2638 	if (prog->jited) {
2639 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2640 		struct bpf_binary_header *hdr;
2641 
2642 		/*
2643 		 * If we fail the final pass of JIT (from jit_subprogs),
2644 		 * the program may not be finalized yet. Call finalize here
2645 		 * before freeing it.
2646 		 */
2647 		if (jit_data) {
2648 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2649 						     jit_data->rw_header);
2650 			kvfree(jit_data->addrs);
2651 			kfree(jit_data);
2652 		}
2653 		hdr = bpf_jit_binary_pack_hdr(prog);
2654 		bpf_jit_binary_pack_free(hdr, NULL);
2655 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2656 	}
2657 
2658 	bpf_prog_unlock_free(prog);
2659 }
2660