xref: /linux/arch/x86/net/bpf_jit_comp.c (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * BPF JIT compiler
4  *
5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6  * Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
7  */
8 #include <linux/netdevice.h>
9 #include <linux/filter.h>
10 #include <linux/if_vlan.h>
11 #include <linux/bpf.h>
12 #include <linux/memory.h>
13 #include <linux/sort.h>
14 #include <asm/extable.h>
15 #include <asm/ftrace.h>
16 #include <asm/set_memory.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/text-patching.h>
19 #include <asm/unwind.h>
20 
21 static bool all_callee_regs_used[4] = {true, true, true, true};
22 
23 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
24 {
25 	if (len == 1)
26 		*ptr = bytes;
27 	else if (len == 2)
28 		*(u16 *)ptr = bytes;
29 	else {
30 		*(u32 *)ptr = bytes;
31 		barrier();
32 	}
33 	return ptr + len;
34 }
35 
36 #define EMIT(bytes, len) \
37 	do { prog = emit_code(prog, bytes, len); } while (0)
38 
39 #define EMIT1(b1)		EMIT(b1, 1)
40 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
41 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
42 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
43 
44 #define EMIT1_off32(b1, off) \
45 	do { EMIT1(b1); EMIT(off, 4); } while (0)
46 #define EMIT2_off32(b1, b2, off) \
47 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
48 #define EMIT3_off32(b1, b2, b3, off) \
49 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
50 #define EMIT4_off32(b1, b2, b3, b4, off) \
51 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
52 
53 #ifdef CONFIG_X86_KERNEL_IBT
54 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
55 #else
56 #define EMIT_ENDBR()
57 #endif
58 
59 static bool is_imm8(int value)
60 {
61 	return value <= 127 && value >= -128;
62 }
63 
64 static bool is_simm32(s64 value)
65 {
66 	return value == (s64)(s32)value;
67 }
68 
69 static bool is_uimm32(u64 value)
70 {
71 	return value == (u64)(u32)value;
72 }
73 
74 /* mov dst, src */
75 #define EMIT_mov(DST, SRC)								 \
76 	do {										 \
77 		if (DST != SRC)								 \
78 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
79 	} while (0)
80 
81 static int bpf_size_to_x86_bytes(int bpf_size)
82 {
83 	if (bpf_size == BPF_W)
84 		return 4;
85 	else if (bpf_size == BPF_H)
86 		return 2;
87 	else if (bpf_size == BPF_B)
88 		return 1;
89 	else if (bpf_size == BPF_DW)
90 		return 4; /* imm32 */
91 	else
92 		return 0;
93 }
94 
95 /*
96  * List of x86 cond jumps opcodes (. + s8)
97  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
98  */
99 #define X86_JB  0x72
100 #define X86_JAE 0x73
101 #define X86_JE  0x74
102 #define X86_JNE 0x75
103 #define X86_JBE 0x76
104 #define X86_JA  0x77
105 #define X86_JL  0x7C
106 #define X86_JGE 0x7D
107 #define X86_JLE 0x7E
108 #define X86_JG  0x7F
109 
110 /* Pick a register outside of BPF range for JIT internal work */
111 #define AUX_REG (MAX_BPF_JIT_REG + 1)
112 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
113 
114 /*
115  * The following table maps BPF registers to x86-64 registers.
116  *
117  * x86-64 register R12 is unused, since if used as base address
118  * register in load/store instructions, it always needs an
119  * extra byte of encoding and is callee saved.
120  *
121  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
122  * trampoline. x86-64 register R10 is used for blinding (if enabled).
123  */
124 static const int reg2hex[] = {
125 	[BPF_REG_0] = 0,  /* RAX */
126 	[BPF_REG_1] = 7,  /* RDI */
127 	[BPF_REG_2] = 6,  /* RSI */
128 	[BPF_REG_3] = 2,  /* RDX */
129 	[BPF_REG_4] = 1,  /* RCX */
130 	[BPF_REG_5] = 0,  /* R8  */
131 	[BPF_REG_6] = 3,  /* RBX callee saved */
132 	[BPF_REG_7] = 5,  /* R13 callee saved */
133 	[BPF_REG_8] = 6,  /* R14 callee saved */
134 	[BPF_REG_9] = 7,  /* R15 callee saved */
135 	[BPF_REG_FP] = 5, /* RBP readonly */
136 	[BPF_REG_AX] = 2, /* R10 temp register */
137 	[AUX_REG] = 3,    /* R11 temp register */
138 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
139 };
140 
141 static const int reg2pt_regs[] = {
142 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
143 	[BPF_REG_1] = offsetof(struct pt_regs, di),
144 	[BPF_REG_2] = offsetof(struct pt_regs, si),
145 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
146 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
147 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
148 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
149 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
150 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
151 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
152 };
153 
154 /*
155  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
156  * which need extra byte of encoding.
157  * rax,rcx,...,rbp have simpler encoding
158  */
159 static bool is_ereg(u32 reg)
160 {
161 	return (1 << reg) & (BIT(BPF_REG_5) |
162 			     BIT(AUX_REG) |
163 			     BIT(BPF_REG_7) |
164 			     BIT(BPF_REG_8) |
165 			     BIT(BPF_REG_9) |
166 			     BIT(X86_REG_R9) |
167 			     BIT(BPF_REG_AX));
168 }
169 
170 /*
171  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
172  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
173  * of encoding. al,cl,dl,bl have simpler encoding.
174  */
175 static bool is_ereg_8l(u32 reg)
176 {
177 	return is_ereg(reg) ||
178 	    (1 << reg) & (BIT(BPF_REG_1) |
179 			  BIT(BPF_REG_2) |
180 			  BIT(BPF_REG_FP));
181 }
182 
183 static bool is_axreg(u32 reg)
184 {
185 	return reg == BPF_REG_0;
186 }
187 
188 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
189 static u8 add_1mod(u8 byte, u32 reg)
190 {
191 	if (is_ereg(reg))
192 		byte |= 1;
193 	return byte;
194 }
195 
196 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
197 {
198 	if (is_ereg(r1))
199 		byte |= 1;
200 	if (is_ereg(r2))
201 		byte |= 4;
202 	return byte;
203 }
204 
205 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
206 static u8 add_1reg(u8 byte, u32 dst_reg)
207 {
208 	return byte + reg2hex[dst_reg];
209 }
210 
211 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
212 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
213 {
214 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
215 }
216 
217 /* Some 1-byte opcodes for binary ALU operations */
218 static u8 simple_alu_opcodes[] = {
219 	[BPF_ADD] = 0x01,
220 	[BPF_SUB] = 0x29,
221 	[BPF_AND] = 0x21,
222 	[BPF_OR] = 0x09,
223 	[BPF_XOR] = 0x31,
224 	[BPF_LSH] = 0xE0,
225 	[BPF_RSH] = 0xE8,
226 	[BPF_ARSH] = 0xF8,
227 };
228 
229 static void jit_fill_hole(void *area, unsigned int size)
230 {
231 	/* Fill whole space with INT3 instructions */
232 	memset(area, 0xcc, size);
233 }
234 
235 int bpf_arch_text_invalidate(void *dst, size_t len)
236 {
237 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
238 }
239 
240 struct jit_context {
241 	int cleanup_addr; /* Epilogue code offset */
242 
243 	/*
244 	 * Program specific offsets of labels in the code; these rely on the
245 	 * JIT doing at least 2 passes, recording the position on the first
246 	 * pass, only to generate the correct offset on the second pass.
247 	 */
248 	int tail_call_direct_label;
249 	int tail_call_indirect_label;
250 };
251 
252 /* Maximum number of bytes emitted while JITing one eBPF insn */
253 #define BPF_MAX_INSN_SIZE	128
254 #define BPF_INSN_SAFETY		64
255 
256 /* Number of bytes emit_patch() needs to generate instructions */
257 #define X86_PATCH_SIZE		5
258 /* Number of bytes that will be skipped on tailcall */
259 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
260 
261 static void push_r12(u8 **pprog)
262 {
263 	u8 *prog = *pprog;
264 
265 	EMIT2(0x41, 0x54);   /* push r12 */
266 	*pprog = prog;
267 }
268 
269 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
270 {
271 	u8 *prog = *pprog;
272 
273 	if (callee_regs_used[0])
274 		EMIT1(0x53);         /* push rbx */
275 	if (callee_regs_used[1])
276 		EMIT2(0x41, 0x55);   /* push r13 */
277 	if (callee_regs_used[2])
278 		EMIT2(0x41, 0x56);   /* push r14 */
279 	if (callee_regs_used[3])
280 		EMIT2(0x41, 0x57);   /* push r15 */
281 	*pprog = prog;
282 }
283 
284 static void pop_r12(u8 **pprog)
285 {
286 	u8 *prog = *pprog;
287 
288 	EMIT2(0x41, 0x5C);   /* pop r12 */
289 	*pprog = prog;
290 }
291 
292 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
293 {
294 	u8 *prog = *pprog;
295 
296 	if (callee_regs_used[3])
297 		EMIT2(0x41, 0x5F);   /* pop r15 */
298 	if (callee_regs_used[2])
299 		EMIT2(0x41, 0x5E);   /* pop r14 */
300 	if (callee_regs_used[1])
301 		EMIT2(0x41, 0x5D);   /* pop r13 */
302 	if (callee_regs_used[0])
303 		EMIT1(0x5B);         /* pop rbx */
304 	*pprog = prog;
305 }
306 
307 /*
308  * Emit x86-64 prologue code for BPF program.
309  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
310  * while jumping to another program
311  */
312 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
313 			  bool tail_call_reachable, bool is_subprog,
314 			  bool is_exception_cb)
315 {
316 	u8 *prog = *pprog;
317 
318 	/* BPF trampoline can be made to work without these nops,
319 	 * but let's waste 5 bytes for now and optimize later
320 	 */
321 	EMIT_ENDBR();
322 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
323 	prog += X86_PATCH_SIZE;
324 	if (!ebpf_from_cbpf) {
325 		if (tail_call_reachable && !is_subprog)
326 			/* When it's the entry of the whole tailcall context,
327 			 * zeroing rax means initialising tail_call_cnt.
328 			 */
329 			EMIT2(0x31, 0xC0); /* xor eax, eax */
330 		else
331 			/* Keep the same instruction layout. */
332 			EMIT2(0x66, 0x90); /* nop2 */
333 	}
334 	/* Exception callback receives FP as third parameter */
335 	if (is_exception_cb) {
336 		EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */
337 		EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */
338 		/* The main frame must have exception_boundary as true, so we
339 		 * first restore those callee-saved regs from stack, before
340 		 * reusing the stack frame.
341 		 */
342 		pop_callee_regs(&prog, all_callee_regs_used);
343 		pop_r12(&prog);
344 		/* Reset the stack frame. */
345 		EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */
346 	} else {
347 		EMIT1(0x55);             /* push rbp */
348 		EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
349 	}
350 
351 	/* X86_TAIL_CALL_OFFSET is here */
352 	EMIT_ENDBR();
353 
354 	/* sub rsp, rounded_stack_depth */
355 	if (stack_depth)
356 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
357 	if (tail_call_reachable)
358 		EMIT1(0x50);         /* push rax */
359 	*pprog = prog;
360 }
361 
362 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
363 {
364 	u8 *prog = *pprog;
365 	s64 offset;
366 
367 	offset = func - (ip + X86_PATCH_SIZE);
368 	if (!is_simm32(offset)) {
369 		pr_err("Target call %p is out of range\n", func);
370 		return -ERANGE;
371 	}
372 	EMIT1_off32(opcode, offset);
373 	*pprog = prog;
374 	return 0;
375 }
376 
377 static int emit_call(u8 **pprog, void *func, void *ip)
378 {
379 	return emit_patch(pprog, func, ip, 0xE8);
380 }
381 
382 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
383 {
384 	OPTIMIZER_HIDE_VAR(func);
385 	x86_call_depth_emit_accounting(pprog, func);
386 	return emit_patch(pprog, func, ip, 0xE8);
387 }
388 
389 static int emit_jump(u8 **pprog, void *func, void *ip)
390 {
391 	return emit_patch(pprog, func, ip, 0xE9);
392 }
393 
394 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
395 				void *old_addr, void *new_addr)
396 {
397 	const u8 *nop_insn = x86_nops[5];
398 	u8 old_insn[X86_PATCH_SIZE];
399 	u8 new_insn[X86_PATCH_SIZE];
400 	u8 *prog;
401 	int ret;
402 
403 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
404 	if (old_addr) {
405 		prog = old_insn;
406 		ret = t == BPF_MOD_CALL ?
407 		      emit_call(&prog, old_addr, ip) :
408 		      emit_jump(&prog, old_addr, ip);
409 		if (ret)
410 			return ret;
411 	}
412 
413 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
414 	if (new_addr) {
415 		prog = new_insn;
416 		ret = t == BPF_MOD_CALL ?
417 		      emit_call(&prog, new_addr, ip) :
418 		      emit_jump(&prog, new_addr, ip);
419 		if (ret)
420 			return ret;
421 	}
422 
423 	ret = -EBUSY;
424 	mutex_lock(&text_mutex);
425 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
426 		goto out;
427 	ret = 1;
428 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
429 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
430 		ret = 0;
431 	}
432 out:
433 	mutex_unlock(&text_mutex);
434 	return ret;
435 }
436 
437 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
438 		       void *old_addr, void *new_addr)
439 {
440 	if (!is_kernel_text((long)ip) &&
441 	    !is_bpf_text_address((long)ip))
442 		/* BPF poking in modules is not supported */
443 		return -EINVAL;
444 
445 	/*
446 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
447 	 * with an ENDBR instruction.
448 	 */
449 	if (is_endbr(*(u32 *)ip))
450 		ip += ENDBR_INSN_SIZE;
451 
452 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
453 }
454 
455 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
456 
457 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
458 {
459 	u8 *prog = *pprog;
460 
461 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
462 		EMIT_LFENCE();
463 		EMIT2(0xFF, 0xE0 + reg);
464 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
465 		OPTIMIZER_HIDE_VAR(reg);
466 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
467 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
468 		else
469 			emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
470 	} else {
471 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
472 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
473 			EMIT1(0xCC);		/* int3 */
474 	}
475 
476 	*pprog = prog;
477 }
478 
479 static void emit_return(u8 **pprog, u8 *ip)
480 {
481 	u8 *prog = *pprog;
482 
483 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
484 		emit_jump(&prog, x86_return_thunk, ip);
485 	} else {
486 		EMIT1(0xC3);		/* ret */
487 		if (IS_ENABLED(CONFIG_SLS))
488 			EMIT1(0xCC);	/* int3 */
489 	}
490 
491 	*pprog = prog;
492 }
493 
494 /*
495  * Generate the following code:
496  *
497  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
498  *   if (index >= array->map.max_entries)
499  *     goto out;
500  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
501  *     goto out;
502  *   prog = array->ptrs[index];
503  *   if (prog == NULL)
504  *     goto out;
505  *   goto *(prog->bpf_func + prologue_size);
506  * out:
507  */
508 static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog,
509 					u8 **pprog, bool *callee_regs_used,
510 					u32 stack_depth, u8 *ip,
511 					struct jit_context *ctx)
512 {
513 	int tcc_off = -4 - round_up(stack_depth, 8);
514 	u8 *prog = *pprog, *start = *pprog;
515 	int offset;
516 
517 	/*
518 	 * rdi - pointer to ctx
519 	 * rsi - pointer to bpf_array
520 	 * rdx - index in bpf_array
521 	 */
522 
523 	/*
524 	 * if (index >= array->map.max_entries)
525 	 *	goto out;
526 	 */
527 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
528 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
529 	      offsetof(struct bpf_array, map.max_entries));
530 
531 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
532 	EMIT2(X86_JBE, offset);                   /* jbe out */
533 
534 	/*
535 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
536 	 *	goto out;
537 	 */
538 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
539 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
540 
541 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
542 	EMIT2(X86_JAE, offset);                   /* jae out */
543 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
544 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
545 
546 	/* prog = array->ptrs[index]; */
547 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
548 		    offsetof(struct bpf_array, ptrs));
549 
550 	/*
551 	 * if (prog == NULL)
552 	 *	goto out;
553 	 */
554 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
555 
556 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
557 	EMIT2(X86_JE, offset);                    /* je out */
558 
559 	if (bpf_prog->aux->exception_boundary) {
560 		pop_callee_regs(&prog, all_callee_regs_used);
561 		pop_r12(&prog);
562 	} else {
563 		pop_callee_regs(&prog, callee_regs_used);
564 	}
565 
566 	EMIT1(0x58);                              /* pop rax */
567 	if (stack_depth)
568 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
569 			    round_up(stack_depth, 8));
570 
571 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
572 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
573 	      offsetof(struct bpf_prog, bpf_func));
574 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
575 	      X86_TAIL_CALL_OFFSET);
576 	/*
577 	 * Now we're ready to jump into next BPF program
578 	 * rdi == ctx (1st arg)
579 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
580 	 */
581 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
582 
583 	/* out: */
584 	ctx->tail_call_indirect_label = prog - start;
585 	*pprog = prog;
586 }
587 
588 static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog,
589 				      struct bpf_jit_poke_descriptor *poke,
590 				      u8 **pprog, u8 *ip,
591 				      bool *callee_regs_used, u32 stack_depth,
592 				      struct jit_context *ctx)
593 {
594 	int tcc_off = -4 - round_up(stack_depth, 8);
595 	u8 *prog = *pprog, *start = *pprog;
596 	int offset;
597 
598 	/*
599 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
600 	 *	goto out;
601 	 */
602 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
603 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
604 
605 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
606 	EMIT2(X86_JAE, offset);                       /* jae out */
607 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
608 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
609 
610 	poke->tailcall_bypass = ip + (prog - start);
611 	poke->adj_off = X86_TAIL_CALL_OFFSET;
612 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
613 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
614 
615 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
616 		  poke->tailcall_bypass);
617 
618 	if (bpf_prog->aux->exception_boundary) {
619 		pop_callee_regs(&prog, all_callee_regs_used);
620 		pop_r12(&prog);
621 	} else {
622 		pop_callee_regs(&prog, callee_regs_used);
623 	}
624 
625 	EMIT1(0x58);                                  /* pop rax */
626 	if (stack_depth)
627 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
628 
629 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
630 	prog += X86_PATCH_SIZE;
631 
632 	/* out: */
633 	ctx->tail_call_direct_label = prog - start;
634 
635 	*pprog = prog;
636 }
637 
638 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
639 {
640 	struct bpf_jit_poke_descriptor *poke;
641 	struct bpf_array *array;
642 	struct bpf_prog *target;
643 	int i, ret;
644 
645 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
646 		poke = &prog->aux->poke_tab[i];
647 		if (poke->aux && poke->aux != prog->aux)
648 			continue;
649 
650 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
651 
652 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
653 			continue;
654 
655 		array = container_of(poke->tail_call.map, struct bpf_array, map);
656 		mutex_lock(&array->aux->poke_mutex);
657 		target = array->ptrs[poke->tail_call.key];
658 		if (target) {
659 			ret = __bpf_arch_text_poke(poke->tailcall_target,
660 						   BPF_MOD_JUMP, NULL,
661 						   (u8 *)target->bpf_func +
662 						   poke->adj_off);
663 			BUG_ON(ret < 0);
664 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
665 						   BPF_MOD_JUMP,
666 						   (u8 *)poke->tailcall_target +
667 						   X86_PATCH_SIZE, NULL);
668 			BUG_ON(ret < 0);
669 		}
670 		WRITE_ONCE(poke->tailcall_target_stable, true);
671 		mutex_unlock(&array->aux->poke_mutex);
672 	}
673 }
674 
675 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
676 			   u32 dst_reg, const u32 imm32)
677 {
678 	u8 *prog = *pprog;
679 	u8 b1, b2, b3;
680 
681 	/*
682 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
683 	 * (which zero-extends imm32) to save 2 bytes.
684 	 */
685 	if (sign_propagate && (s32)imm32 < 0) {
686 		/* 'mov %rax, imm32' sign extends imm32 */
687 		b1 = add_1mod(0x48, dst_reg);
688 		b2 = 0xC7;
689 		b3 = 0xC0;
690 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
691 		goto done;
692 	}
693 
694 	/*
695 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
696 	 * to save 3 bytes.
697 	 */
698 	if (imm32 == 0) {
699 		if (is_ereg(dst_reg))
700 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
701 		b2 = 0x31; /* xor */
702 		b3 = 0xC0;
703 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
704 		goto done;
705 	}
706 
707 	/* mov %eax, imm32 */
708 	if (is_ereg(dst_reg))
709 		EMIT1(add_1mod(0x40, dst_reg));
710 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
711 done:
712 	*pprog = prog;
713 }
714 
715 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
716 			   const u32 imm32_hi, const u32 imm32_lo)
717 {
718 	u8 *prog = *pprog;
719 
720 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
721 		/*
722 		 * For emitting plain u32, where sign bit must not be
723 		 * propagated LLVM tends to load imm64 over mov32
724 		 * directly, so save couple of bytes by just doing
725 		 * 'mov %eax, imm32' instead.
726 		 */
727 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
728 	} else {
729 		/* movabsq rax, imm64 */
730 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
731 		EMIT(imm32_lo, 4);
732 		EMIT(imm32_hi, 4);
733 	}
734 
735 	*pprog = prog;
736 }
737 
738 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
739 {
740 	u8 *prog = *pprog;
741 
742 	if (is64) {
743 		/* mov dst, src */
744 		EMIT_mov(dst_reg, src_reg);
745 	} else {
746 		/* mov32 dst, src */
747 		if (is_ereg(dst_reg) || is_ereg(src_reg))
748 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
749 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
750 	}
751 
752 	*pprog = prog;
753 }
754 
755 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
756 			   u32 src_reg)
757 {
758 	u8 *prog = *pprog;
759 
760 	if (is64) {
761 		/* movs[b,w,l]q dst, src */
762 		if (num_bits == 8)
763 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
764 			      add_2reg(0xC0, src_reg, dst_reg));
765 		else if (num_bits == 16)
766 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
767 			      add_2reg(0xC0, src_reg, dst_reg));
768 		else if (num_bits == 32)
769 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
770 			      add_2reg(0xC0, src_reg, dst_reg));
771 	} else {
772 		/* movs[b,w]l dst, src */
773 		if (num_bits == 8) {
774 			EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
775 			      add_2reg(0xC0, src_reg, dst_reg));
776 		} else if (num_bits == 16) {
777 			if (is_ereg(dst_reg) || is_ereg(src_reg))
778 				EMIT1(add_2mod(0x40, src_reg, dst_reg));
779 			EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
780 			      add_2reg(0xC0, src_reg, dst_reg));
781 		}
782 	}
783 
784 	*pprog = prog;
785 }
786 
787 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
788 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
789 {
790 	u8 *prog = *pprog;
791 
792 	if (is_imm8(off)) {
793 		/* 1-byte signed displacement.
794 		 *
795 		 * If off == 0 we could skip this and save one extra byte, but
796 		 * special case of x86 R13 which always needs an offset is not
797 		 * worth the hassle
798 		 */
799 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
800 	} else {
801 		/* 4-byte signed displacement */
802 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
803 	}
804 	*pprog = prog;
805 }
806 
807 /*
808  * Emit a REX byte if it will be necessary to address these registers
809  */
810 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
811 {
812 	u8 *prog = *pprog;
813 
814 	if (is64)
815 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
816 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
817 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
818 	*pprog = prog;
819 }
820 
821 /*
822  * Similar version of maybe_emit_mod() for a single register
823  */
824 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
825 {
826 	u8 *prog = *pprog;
827 
828 	if (is64)
829 		EMIT1(add_1mod(0x48, reg));
830 	else if (is_ereg(reg))
831 		EMIT1(add_1mod(0x40, reg));
832 	*pprog = prog;
833 }
834 
835 /* LDX: dst_reg = *(u8*)(src_reg + off) */
836 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
837 {
838 	u8 *prog = *pprog;
839 
840 	switch (size) {
841 	case BPF_B:
842 		/* Emit 'movzx rax, byte ptr [rax + off]' */
843 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
844 		break;
845 	case BPF_H:
846 		/* Emit 'movzx rax, word ptr [rax + off]' */
847 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
848 		break;
849 	case BPF_W:
850 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
851 		if (is_ereg(dst_reg) || is_ereg(src_reg))
852 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
853 		else
854 			EMIT1(0x8B);
855 		break;
856 	case BPF_DW:
857 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
858 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
859 		break;
860 	}
861 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
862 	*pprog = prog;
863 }
864 
865 /* LDSX: dst_reg = *(s8*)(src_reg + off) */
866 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
867 {
868 	u8 *prog = *pprog;
869 
870 	switch (size) {
871 	case BPF_B:
872 		/* Emit 'movsx rax, byte ptr [rax + off]' */
873 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
874 		break;
875 	case BPF_H:
876 		/* Emit 'movsx rax, word ptr [rax + off]' */
877 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
878 		break;
879 	case BPF_W:
880 		/* Emit 'movsx rax, dword ptr [rax+0x14]' */
881 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
882 		break;
883 	}
884 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
885 	*pprog = prog;
886 }
887 
888 /* STX: *(u8*)(dst_reg + off) = src_reg */
889 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
890 {
891 	u8 *prog = *pprog;
892 
893 	switch (size) {
894 	case BPF_B:
895 		/* Emit 'mov byte ptr [rax + off], al' */
896 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
897 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
898 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
899 		else
900 			EMIT1(0x88);
901 		break;
902 	case BPF_H:
903 		if (is_ereg(dst_reg) || is_ereg(src_reg))
904 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
905 		else
906 			EMIT2(0x66, 0x89);
907 		break;
908 	case BPF_W:
909 		if (is_ereg(dst_reg) || is_ereg(src_reg))
910 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
911 		else
912 			EMIT1(0x89);
913 		break;
914 	case BPF_DW:
915 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
916 		break;
917 	}
918 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
919 	*pprog = prog;
920 }
921 
922 static int emit_atomic(u8 **pprog, u8 atomic_op,
923 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
924 {
925 	u8 *prog = *pprog;
926 
927 	EMIT1(0xF0); /* lock prefix */
928 
929 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
930 
931 	/* emit opcode */
932 	switch (atomic_op) {
933 	case BPF_ADD:
934 	case BPF_AND:
935 	case BPF_OR:
936 	case BPF_XOR:
937 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
938 		EMIT1(simple_alu_opcodes[atomic_op]);
939 		break;
940 	case BPF_ADD | BPF_FETCH:
941 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
942 		EMIT2(0x0F, 0xC1);
943 		break;
944 	case BPF_XCHG:
945 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
946 		EMIT1(0x87);
947 		break;
948 	case BPF_CMPXCHG:
949 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
950 		EMIT2(0x0F, 0xB1);
951 		break;
952 	default:
953 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
954 		return -EFAULT;
955 	}
956 
957 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
958 
959 	*pprog = prog;
960 	return 0;
961 }
962 
963 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
964 {
965 	u32 reg = x->fixup >> 8;
966 
967 	/* jump over faulting load and clear dest register */
968 	*(unsigned long *)((void *)regs + reg) = 0;
969 	regs->ip += x->fixup & 0xff;
970 	return true;
971 }
972 
973 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
974 			     bool *regs_used, bool *tail_call_seen)
975 {
976 	int i;
977 
978 	for (i = 1; i <= insn_cnt; i++, insn++) {
979 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
980 			*tail_call_seen = true;
981 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
982 			regs_used[0] = true;
983 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
984 			regs_used[1] = true;
985 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
986 			regs_used[2] = true;
987 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
988 			regs_used[3] = true;
989 	}
990 }
991 
992 static void emit_nops(u8 **pprog, int len)
993 {
994 	u8 *prog = *pprog;
995 	int i, noplen;
996 
997 	while (len > 0) {
998 		noplen = len;
999 
1000 		if (noplen > ASM_NOP_MAX)
1001 			noplen = ASM_NOP_MAX;
1002 
1003 		for (i = 0; i < noplen; i++)
1004 			EMIT1(x86_nops[noplen][i]);
1005 		len -= noplen;
1006 	}
1007 
1008 	*pprog = prog;
1009 }
1010 
1011 /* emit the 3-byte VEX prefix
1012  *
1013  * r: same as rex.r, extra bit for ModRM reg field
1014  * x: same as rex.x, extra bit for SIB index field
1015  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
1016  * m: opcode map select, encoding escape bytes e.g. 0x0f38
1017  * w: same as rex.w (32 bit or 64 bit) or opcode specific
1018  * src_reg2: additional source reg (encoded as BPF reg)
1019  * l: vector length (128 bit or 256 bit) or reserved
1020  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
1021  */
1022 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
1023 		      bool w, u8 src_reg2, bool l, u8 pp)
1024 {
1025 	u8 *prog = *pprog;
1026 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
1027 	u8 b1, b2;
1028 	u8 vvvv = reg2hex[src_reg2];
1029 
1030 	/* reg2hex gives only the lower 3 bit of vvvv */
1031 	if (is_ereg(src_reg2))
1032 		vvvv |= 1 << 3;
1033 
1034 	/*
1035 	 * 2nd byte of 3-byte VEX prefix
1036 	 * ~ means bit inverted encoding
1037 	 *
1038 	 *    7                           0
1039 	 *  +---+---+---+---+---+---+---+---+
1040 	 *  |~R |~X |~B |         m         |
1041 	 *  +---+---+---+---+---+---+---+---+
1042 	 */
1043 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
1044 	/*
1045 	 * 3rd byte of 3-byte VEX prefix
1046 	 *
1047 	 *    7                           0
1048 	 *  +---+---+---+---+---+---+---+---+
1049 	 *  | W |     ~vvvv     | L |   pp  |
1050 	 *  +---+---+---+---+---+---+---+---+
1051 	 */
1052 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1053 
1054 	EMIT3(b0, b1, b2);
1055 	*pprog = prog;
1056 }
1057 
1058 /* emit BMI2 shift instruction */
1059 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1060 {
1061 	u8 *prog = *pprog;
1062 	bool r = is_ereg(dst_reg);
1063 	u8 m = 2; /* escape code 0f38 */
1064 
1065 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1066 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1067 	*pprog = prog;
1068 }
1069 
1070 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1071 
1072 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1073 #define RESTORE_TAIL_CALL_CNT(stack)				\
1074 	EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
1075 
1076 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
1077 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
1078 {
1079 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1080 	struct bpf_insn *insn = bpf_prog->insnsi;
1081 	bool callee_regs_used[4] = {};
1082 	int insn_cnt = bpf_prog->len;
1083 	bool tail_call_seen = false;
1084 	bool seen_exit = false;
1085 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1086 	int i, excnt = 0;
1087 	int ilen, proglen = 0;
1088 	u8 *prog = temp;
1089 	int err;
1090 
1091 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
1092 			 &tail_call_seen);
1093 
1094 	/* tail call's presence in current prog implies it is reachable */
1095 	tail_call_reachable |= tail_call_seen;
1096 
1097 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
1098 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1099 		      bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
1100 	/* Exception callback will clobber callee regs for its own use, and
1101 	 * restore the original callee regs from main prog's stack frame.
1102 	 */
1103 	if (bpf_prog->aux->exception_boundary) {
1104 		/* We also need to save r12, which is not mapped to any BPF
1105 		 * register, as we throw after entry into the kernel, which may
1106 		 * overwrite r12.
1107 		 */
1108 		push_r12(&prog);
1109 		push_callee_regs(&prog, all_callee_regs_used);
1110 	} else {
1111 		push_callee_regs(&prog, callee_regs_used);
1112 	}
1113 
1114 	ilen = prog - temp;
1115 	if (rw_image)
1116 		memcpy(rw_image + proglen, temp, ilen);
1117 	proglen += ilen;
1118 	addrs[0] = proglen;
1119 	prog = temp;
1120 
1121 	for (i = 1; i <= insn_cnt; i++, insn++) {
1122 		const s32 imm32 = insn->imm;
1123 		u32 dst_reg = insn->dst_reg;
1124 		u32 src_reg = insn->src_reg;
1125 		u8 b2 = 0, b3 = 0;
1126 		u8 *start_of_ldx;
1127 		s64 jmp_offset;
1128 		s16 insn_off;
1129 		u8 jmp_cond;
1130 		u8 *func;
1131 		int nops;
1132 
1133 		switch (insn->code) {
1134 			/* ALU */
1135 		case BPF_ALU | BPF_ADD | BPF_X:
1136 		case BPF_ALU | BPF_SUB | BPF_X:
1137 		case BPF_ALU | BPF_AND | BPF_X:
1138 		case BPF_ALU | BPF_OR | BPF_X:
1139 		case BPF_ALU | BPF_XOR | BPF_X:
1140 		case BPF_ALU64 | BPF_ADD | BPF_X:
1141 		case BPF_ALU64 | BPF_SUB | BPF_X:
1142 		case BPF_ALU64 | BPF_AND | BPF_X:
1143 		case BPF_ALU64 | BPF_OR | BPF_X:
1144 		case BPF_ALU64 | BPF_XOR | BPF_X:
1145 			maybe_emit_mod(&prog, dst_reg, src_reg,
1146 				       BPF_CLASS(insn->code) == BPF_ALU64);
1147 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1148 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1149 			break;
1150 
1151 		case BPF_ALU64 | BPF_MOV | BPF_X:
1152 		case BPF_ALU | BPF_MOV | BPF_X:
1153 			if (insn->off == 0)
1154 				emit_mov_reg(&prog,
1155 					     BPF_CLASS(insn->code) == BPF_ALU64,
1156 					     dst_reg, src_reg);
1157 			else
1158 				emit_movsx_reg(&prog, insn->off,
1159 					       BPF_CLASS(insn->code) == BPF_ALU64,
1160 					       dst_reg, src_reg);
1161 			break;
1162 
1163 			/* neg dst */
1164 		case BPF_ALU | BPF_NEG:
1165 		case BPF_ALU64 | BPF_NEG:
1166 			maybe_emit_1mod(&prog, dst_reg,
1167 					BPF_CLASS(insn->code) == BPF_ALU64);
1168 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1169 			break;
1170 
1171 		case BPF_ALU | BPF_ADD | BPF_K:
1172 		case BPF_ALU | BPF_SUB | BPF_K:
1173 		case BPF_ALU | BPF_AND | BPF_K:
1174 		case BPF_ALU | BPF_OR | BPF_K:
1175 		case BPF_ALU | BPF_XOR | BPF_K:
1176 		case BPF_ALU64 | BPF_ADD | BPF_K:
1177 		case BPF_ALU64 | BPF_SUB | BPF_K:
1178 		case BPF_ALU64 | BPF_AND | BPF_K:
1179 		case BPF_ALU64 | BPF_OR | BPF_K:
1180 		case BPF_ALU64 | BPF_XOR | BPF_K:
1181 			maybe_emit_1mod(&prog, dst_reg,
1182 					BPF_CLASS(insn->code) == BPF_ALU64);
1183 
1184 			/*
1185 			 * b3 holds 'normal' opcode, b2 short form only valid
1186 			 * in case dst is eax/rax.
1187 			 */
1188 			switch (BPF_OP(insn->code)) {
1189 			case BPF_ADD:
1190 				b3 = 0xC0;
1191 				b2 = 0x05;
1192 				break;
1193 			case BPF_SUB:
1194 				b3 = 0xE8;
1195 				b2 = 0x2D;
1196 				break;
1197 			case BPF_AND:
1198 				b3 = 0xE0;
1199 				b2 = 0x25;
1200 				break;
1201 			case BPF_OR:
1202 				b3 = 0xC8;
1203 				b2 = 0x0D;
1204 				break;
1205 			case BPF_XOR:
1206 				b3 = 0xF0;
1207 				b2 = 0x35;
1208 				break;
1209 			}
1210 
1211 			if (is_imm8(imm32))
1212 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1213 			else if (is_axreg(dst_reg))
1214 				EMIT1_off32(b2, imm32);
1215 			else
1216 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1217 			break;
1218 
1219 		case BPF_ALU64 | BPF_MOV | BPF_K:
1220 		case BPF_ALU | BPF_MOV | BPF_K:
1221 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1222 				       dst_reg, imm32);
1223 			break;
1224 
1225 		case BPF_LD | BPF_IMM | BPF_DW:
1226 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1227 			insn++;
1228 			i++;
1229 			break;
1230 
1231 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1232 		case BPF_ALU | BPF_MOD | BPF_X:
1233 		case BPF_ALU | BPF_DIV | BPF_X:
1234 		case BPF_ALU | BPF_MOD | BPF_K:
1235 		case BPF_ALU | BPF_DIV | BPF_K:
1236 		case BPF_ALU64 | BPF_MOD | BPF_X:
1237 		case BPF_ALU64 | BPF_DIV | BPF_X:
1238 		case BPF_ALU64 | BPF_MOD | BPF_K:
1239 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1240 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1241 
1242 			if (dst_reg != BPF_REG_0)
1243 				EMIT1(0x50); /* push rax */
1244 			if (dst_reg != BPF_REG_3)
1245 				EMIT1(0x52); /* push rdx */
1246 
1247 			if (BPF_SRC(insn->code) == BPF_X) {
1248 				if (src_reg == BPF_REG_0 ||
1249 				    src_reg == BPF_REG_3) {
1250 					/* mov r11, src_reg */
1251 					EMIT_mov(AUX_REG, src_reg);
1252 					src_reg = AUX_REG;
1253 				}
1254 			} else {
1255 				/* mov r11, imm32 */
1256 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1257 				src_reg = AUX_REG;
1258 			}
1259 
1260 			if (dst_reg != BPF_REG_0)
1261 				/* mov rax, dst_reg */
1262 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1263 
1264 			if (insn->off == 0) {
1265 				/*
1266 				 * xor edx, edx
1267 				 * equivalent to 'xor rdx, rdx', but one byte less
1268 				 */
1269 				EMIT2(0x31, 0xd2);
1270 
1271 				/* div src_reg */
1272 				maybe_emit_1mod(&prog, src_reg, is64);
1273 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1274 			} else {
1275 				if (BPF_CLASS(insn->code) == BPF_ALU)
1276 					EMIT1(0x99); /* cdq */
1277 				else
1278 					EMIT2(0x48, 0x99); /* cqo */
1279 
1280 				/* idiv src_reg */
1281 				maybe_emit_1mod(&prog, src_reg, is64);
1282 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1283 			}
1284 
1285 			if (BPF_OP(insn->code) == BPF_MOD &&
1286 			    dst_reg != BPF_REG_3)
1287 				/* mov dst_reg, rdx */
1288 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1289 			else if (BPF_OP(insn->code) == BPF_DIV &&
1290 				 dst_reg != BPF_REG_0)
1291 				/* mov dst_reg, rax */
1292 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1293 
1294 			if (dst_reg != BPF_REG_3)
1295 				EMIT1(0x5A); /* pop rdx */
1296 			if (dst_reg != BPF_REG_0)
1297 				EMIT1(0x58); /* pop rax */
1298 			break;
1299 		}
1300 
1301 		case BPF_ALU | BPF_MUL | BPF_K:
1302 		case BPF_ALU64 | BPF_MUL | BPF_K:
1303 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1304 				       BPF_CLASS(insn->code) == BPF_ALU64);
1305 
1306 			if (is_imm8(imm32))
1307 				/* imul dst_reg, dst_reg, imm8 */
1308 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1309 				      imm32);
1310 			else
1311 				/* imul dst_reg, dst_reg, imm32 */
1312 				EMIT2_off32(0x69,
1313 					    add_2reg(0xC0, dst_reg, dst_reg),
1314 					    imm32);
1315 			break;
1316 
1317 		case BPF_ALU | BPF_MUL | BPF_X:
1318 		case BPF_ALU64 | BPF_MUL | BPF_X:
1319 			maybe_emit_mod(&prog, src_reg, dst_reg,
1320 				       BPF_CLASS(insn->code) == BPF_ALU64);
1321 
1322 			/* imul dst_reg, src_reg */
1323 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1324 			break;
1325 
1326 			/* Shifts */
1327 		case BPF_ALU | BPF_LSH | BPF_K:
1328 		case BPF_ALU | BPF_RSH | BPF_K:
1329 		case BPF_ALU | BPF_ARSH | BPF_K:
1330 		case BPF_ALU64 | BPF_LSH | BPF_K:
1331 		case BPF_ALU64 | BPF_RSH | BPF_K:
1332 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1333 			maybe_emit_1mod(&prog, dst_reg,
1334 					BPF_CLASS(insn->code) == BPF_ALU64);
1335 
1336 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1337 			if (imm32 == 1)
1338 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1339 			else
1340 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1341 			break;
1342 
1343 		case BPF_ALU | BPF_LSH | BPF_X:
1344 		case BPF_ALU | BPF_RSH | BPF_X:
1345 		case BPF_ALU | BPF_ARSH | BPF_X:
1346 		case BPF_ALU64 | BPF_LSH | BPF_X:
1347 		case BPF_ALU64 | BPF_RSH | BPF_X:
1348 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1349 			/* BMI2 shifts aren't better when shift count is already in rcx */
1350 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1351 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1352 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1353 				u8 op;
1354 
1355 				switch (BPF_OP(insn->code)) {
1356 				case BPF_LSH:
1357 					op = 1; /* prefix 0x66 */
1358 					break;
1359 				case BPF_RSH:
1360 					op = 3; /* prefix 0xf2 */
1361 					break;
1362 				case BPF_ARSH:
1363 					op = 2; /* prefix 0xf3 */
1364 					break;
1365 				}
1366 
1367 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1368 
1369 				break;
1370 			}
1371 
1372 			if (src_reg != BPF_REG_4) { /* common case */
1373 				/* Check for bad case when dst_reg == rcx */
1374 				if (dst_reg == BPF_REG_4) {
1375 					/* mov r11, dst_reg */
1376 					EMIT_mov(AUX_REG, dst_reg);
1377 					dst_reg = AUX_REG;
1378 				} else {
1379 					EMIT1(0x51); /* push rcx */
1380 				}
1381 				/* mov rcx, src_reg */
1382 				EMIT_mov(BPF_REG_4, src_reg);
1383 			}
1384 
1385 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1386 			maybe_emit_1mod(&prog, dst_reg,
1387 					BPF_CLASS(insn->code) == BPF_ALU64);
1388 
1389 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1390 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1391 
1392 			if (src_reg != BPF_REG_4) {
1393 				if (insn->dst_reg == BPF_REG_4)
1394 					/* mov dst_reg, r11 */
1395 					EMIT_mov(insn->dst_reg, AUX_REG);
1396 				else
1397 					EMIT1(0x59); /* pop rcx */
1398 			}
1399 
1400 			break;
1401 
1402 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1403 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1404 			switch (imm32) {
1405 			case 16:
1406 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1407 				EMIT1(0x66);
1408 				if (is_ereg(dst_reg))
1409 					EMIT1(0x41);
1410 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1411 
1412 				/* Emit 'movzwl eax, ax' */
1413 				if (is_ereg(dst_reg))
1414 					EMIT3(0x45, 0x0F, 0xB7);
1415 				else
1416 					EMIT2(0x0F, 0xB7);
1417 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1418 				break;
1419 			case 32:
1420 				/* Emit 'bswap eax' to swap lower 4 bytes */
1421 				if (is_ereg(dst_reg))
1422 					EMIT2(0x41, 0x0F);
1423 				else
1424 					EMIT1(0x0F);
1425 				EMIT1(add_1reg(0xC8, dst_reg));
1426 				break;
1427 			case 64:
1428 				/* Emit 'bswap rax' to swap 8 bytes */
1429 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1430 				      add_1reg(0xC8, dst_reg));
1431 				break;
1432 			}
1433 			break;
1434 
1435 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1436 			switch (imm32) {
1437 			case 16:
1438 				/*
1439 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1440 				 * into 64 bit
1441 				 */
1442 				if (is_ereg(dst_reg))
1443 					EMIT3(0x45, 0x0F, 0xB7);
1444 				else
1445 					EMIT2(0x0F, 0xB7);
1446 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1447 				break;
1448 			case 32:
1449 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1450 				if (is_ereg(dst_reg))
1451 					EMIT1(0x45);
1452 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1453 				break;
1454 			case 64:
1455 				/* nop */
1456 				break;
1457 			}
1458 			break;
1459 
1460 			/* speculation barrier */
1461 		case BPF_ST | BPF_NOSPEC:
1462 			EMIT_LFENCE();
1463 			break;
1464 
1465 			/* ST: *(u8*)(dst_reg + off) = imm */
1466 		case BPF_ST | BPF_MEM | BPF_B:
1467 			if (is_ereg(dst_reg))
1468 				EMIT2(0x41, 0xC6);
1469 			else
1470 				EMIT1(0xC6);
1471 			goto st;
1472 		case BPF_ST | BPF_MEM | BPF_H:
1473 			if (is_ereg(dst_reg))
1474 				EMIT3(0x66, 0x41, 0xC7);
1475 			else
1476 				EMIT2(0x66, 0xC7);
1477 			goto st;
1478 		case BPF_ST | BPF_MEM | BPF_W:
1479 			if (is_ereg(dst_reg))
1480 				EMIT2(0x41, 0xC7);
1481 			else
1482 				EMIT1(0xC7);
1483 			goto st;
1484 		case BPF_ST | BPF_MEM | BPF_DW:
1485 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1486 
1487 st:			if (is_imm8(insn->off))
1488 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1489 			else
1490 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1491 
1492 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1493 			break;
1494 
1495 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1496 		case BPF_STX | BPF_MEM | BPF_B:
1497 		case BPF_STX | BPF_MEM | BPF_H:
1498 		case BPF_STX | BPF_MEM | BPF_W:
1499 		case BPF_STX | BPF_MEM | BPF_DW:
1500 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1501 			break;
1502 
1503 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1504 		case BPF_LDX | BPF_MEM | BPF_B:
1505 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1506 		case BPF_LDX | BPF_MEM | BPF_H:
1507 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1508 		case BPF_LDX | BPF_MEM | BPF_W:
1509 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1510 		case BPF_LDX | BPF_MEM | BPF_DW:
1511 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1512 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
1513 		case BPF_LDX | BPF_MEMSX | BPF_B:
1514 		case BPF_LDX | BPF_MEMSX | BPF_H:
1515 		case BPF_LDX | BPF_MEMSX | BPF_W:
1516 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1517 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1518 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1519 			insn_off = insn->off;
1520 
1521 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1522 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1523 				/* Conservatively check that src_reg + insn->off is a kernel address:
1524 				 *   src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
1525 				 * src_reg is used as scratch for src_reg += insn->off and restored
1526 				 * after emit_ldx if necessary
1527 				 */
1528 
1529 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
1530 				u8 *end_of_jmp;
1531 
1532 				/* At end of these emitted checks, insn->off will have been added
1533 				 * to src_reg, so no need to do relative load with insn->off offset
1534 				 */
1535 				insn_off = 0;
1536 
1537 				/* movabsq r11, limit */
1538 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1539 				EMIT((u32)limit, 4);
1540 				EMIT(limit >> 32, 4);
1541 
1542 				if (insn->off) {
1543 					/* add src_reg, insn->off */
1544 					maybe_emit_1mod(&prog, src_reg, true);
1545 					EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
1546 				}
1547 
1548 				/* cmp src_reg, r11 */
1549 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1550 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1551 
1552 				/* if unsigned '>=', goto load */
1553 				EMIT2(X86_JAE, 0);
1554 				end_of_jmp = prog;
1555 
1556 				/* xor dst_reg, dst_reg */
1557 				emit_mov_imm32(&prog, false, dst_reg, 0);
1558 				/* jmp byte_after_ldx */
1559 				EMIT2(0xEB, 0);
1560 
1561 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1562 				start_of_ldx = prog;
1563 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1564 			}
1565 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
1566 			    BPF_MODE(insn->code) == BPF_MEMSX)
1567 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1568 			else
1569 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1570 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1571 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1572 				struct exception_table_entry *ex;
1573 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1574 				s64 delta;
1575 
1576 				/* populate jmp_offset for JMP above */
1577 				start_of_ldx[-1] = prog - start_of_ldx;
1578 
1579 				if (insn->off && src_reg != dst_reg) {
1580 					/* sub src_reg, insn->off
1581 					 * Restore src_reg after "add src_reg, insn->off" in prev
1582 					 * if statement. But if src_reg == dst_reg, emit_ldx
1583 					 * above already clobbered src_reg, so no need to restore.
1584 					 * If add src_reg, insn->off was unnecessary, no need to
1585 					 * restore either.
1586 					 */
1587 					maybe_emit_1mod(&prog, src_reg, true);
1588 					EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
1589 				}
1590 
1591 				if (!bpf_prog->aux->extable)
1592 					break;
1593 
1594 				if (excnt >= bpf_prog->aux->num_exentries) {
1595 					pr_err("ex gen bug\n");
1596 					return -EFAULT;
1597 				}
1598 				ex = &bpf_prog->aux->extable[excnt++];
1599 
1600 				delta = _insn - (u8 *)&ex->insn;
1601 				if (!is_simm32(delta)) {
1602 					pr_err("extable->insn doesn't fit into 32-bit\n");
1603 					return -EFAULT;
1604 				}
1605 				/* switch ex to rw buffer for writes */
1606 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1607 
1608 				ex->insn = delta;
1609 
1610 				ex->data = EX_TYPE_BPF;
1611 
1612 				if (dst_reg > BPF_REG_9) {
1613 					pr_err("verifier error\n");
1614 					return -EFAULT;
1615 				}
1616 				/*
1617 				 * Compute size of x86 insn and its target dest x86 register.
1618 				 * ex_handler_bpf() will use lower 8 bits to adjust
1619 				 * pt_regs->ip to jump over this x86 instruction
1620 				 * and upper bits to figure out which pt_regs to zero out.
1621 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1622 				 * of 4 bytes will be ignored and rbx will be zero inited.
1623 				 */
1624 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1625 			}
1626 			break;
1627 
1628 		case BPF_STX | BPF_ATOMIC | BPF_W:
1629 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1630 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1631 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1632 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1633 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1634 				u32 real_src_reg = src_reg;
1635 				u32 real_dst_reg = dst_reg;
1636 				u8 *branch_target;
1637 
1638 				/*
1639 				 * Can't be implemented with a single x86 insn.
1640 				 * Need to do a CMPXCHG loop.
1641 				 */
1642 
1643 				/* Will need RAX as a CMPXCHG operand so save R0 */
1644 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1645 				if (src_reg == BPF_REG_0)
1646 					real_src_reg = BPF_REG_AX;
1647 				if (dst_reg == BPF_REG_0)
1648 					real_dst_reg = BPF_REG_AX;
1649 
1650 				branch_target = prog;
1651 				/* Load old value */
1652 				emit_ldx(&prog, BPF_SIZE(insn->code),
1653 					 BPF_REG_0, real_dst_reg, insn->off);
1654 				/*
1655 				 * Perform the (commutative) operation locally,
1656 				 * put the result in the AUX_REG.
1657 				 */
1658 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1659 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1660 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1661 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1662 				/* Attempt to swap in new value */
1663 				err = emit_atomic(&prog, BPF_CMPXCHG,
1664 						  real_dst_reg, AUX_REG,
1665 						  insn->off,
1666 						  BPF_SIZE(insn->code));
1667 				if (WARN_ON(err))
1668 					return err;
1669 				/*
1670 				 * ZF tells us whether we won the race. If it's
1671 				 * cleared we need to try again.
1672 				 */
1673 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1674 				/* Return the pre-modification value */
1675 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1676 				/* Restore R0 after clobbering RAX */
1677 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1678 				break;
1679 			}
1680 
1681 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1682 					  insn->off, BPF_SIZE(insn->code));
1683 			if (err)
1684 				return err;
1685 			break;
1686 
1687 			/* call */
1688 		case BPF_JMP | BPF_CALL: {
1689 			int offs;
1690 
1691 			func = (u8 *) __bpf_call_base + imm32;
1692 			if (tail_call_reachable) {
1693 				RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
1694 				if (!imm32)
1695 					return -EINVAL;
1696 				offs = 7 + x86_call_depth_emit_accounting(&prog, func);
1697 			} else {
1698 				if (!imm32)
1699 					return -EINVAL;
1700 				offs = x86_call_depth_emit_accounting(&prog, func);
1701 			}
1702 			if (emit_call(&prog, func, image + addrs[i - 1] + offs))
1703 				return -EINVAL;
1704 			break;
1705 		}
1706 
1707 		case BPF_JMP | BPF_TAIL_CALL:
1708 			if (imm32)
1709 				emit_bpf_tail_call_direct(bpf_prog,
1710 							  &bpf_prog->aux->poke_tab[imm32 - 1],
1711 							  &prog, image + addrs[i - 1],
1712 							  callee_regs_used,
1713 							  bpf_prog->aux->stack_depth,
1714 							  ctx);
1715 			else
1716 				emit_bpf_tail_call_indirect(bpf_prog,
1717 							    &prog,
1718 							    callee_regs_used,
1719 							    bpf_prog->aux->stack_depth,
1720 							    image + addrs[i - 1],
1721 							    ctx);
1722 			break;
1723 
1724 			/* cond jump */
1725 		case BPF_JMP | BPF_JEQ | BPF_X:
1726 		case BPF_JMP | BPF_JNE | BPF_X:
1727 		case BPF_JMP | BPF_JGT | BPF_X:
1728 		case BPF_JMP | BPF_JLT | BPF_X:
1729 		case BPF_JMP | BPF_JGE | BPF_X:
1730 		case BPF_JMP | BPF_JLE | BPF_X:
1731 		case BPF_JMP | BPF_JSGT | BPF_X:
1732 		case BPF_JMP | BPF_JSLT | BPF_X:
1733 		case BPF_JMP | BPF_JSGE | BPF_X:
1734 		case BPF_JMP | BPF_JSLE | BPF_X:
1735 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1736 		case BPF_JMP32 | BPF_JNE | BPF_X:
1737 		case BPF_JMP32 | BPF_JGT | BPF_X:
1738 		case BPF_JMP32 | BPF_JLT | BPF_X:
1739 		case BPF_JMP32 | BPF_JGE | BPF_X:
1740 		case BPF_JMP32 | BPF_JLE | BPF_X:
1741 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1742 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1743 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1744 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1745 			/* cmp dst_reg, src_reg */
1746 			maybe_emit_mod(&prog, dst_reg, src_reg,
1747 				       BPF_CLASS(insn->code) == BPF_JMP);
1748 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1749 			goto emit_cond_jmp;
1750 
1751 		case BPF_JMP | BPF_JSET | BPF_X:
1752 		case BPF_JMP32 | BPF_JSET | BPF_X:
1753 			/* test dst_reg, src_reg */
1754 			maybe_emit_mod(&prog, dst_reg, src_reg,
1755 				       BPF_CLASS(insn->code) == BPF_JMP);
1756 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1757 			goto emit_cond_jmp;
1758 
1759 		case BPF_JMP | BPF_JSET | BPF_K:
1760 		case BPF_JMP32 | BPF_JSET | BPF_K:
1761 			/* test dst_reg, imm32 */
1762 			maybe_emit_1mod(&prog, dst_reg,
1763 					BPF_CLASS(insn->code) == BPF_JMP);
1764 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1765 			goto emit_cond_jmp;
1766 
1767 		case BPF_JMP | BPF_JEQ | BPF_K:
1768 		case BPF_JMP | BPF_JNE | BPF_K:
1769 		case BPF_JMP | BPF_JGT | BPF_K:
1770 		case BPF_JMP | BPF_JLT | BPF_K:
1771 		case BPF_JMP | BPF_JGE | BPF_K:
1772 		case BPF_JMP | BPF_JLE | BPF_K:
1773 		case BPF_JMP | BPF_JSGT | BPF_K:
1774 		case BPF_JMP | BPF_JSLT | BPF_K:
1775 		case BPF_JMP | BPF_JSGE | BPF_K:
1776 		case BPF_JMP | BPF_JSLE | BPF_K:
1777 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1778 		case BPF_JMP32 | BPF_JNE | BPF_K:
1779 		case BPF_JMP32 | BPF_JGT | BPF_K:
1780 		case BPF_JMP32 | BPF_JLT | BPF_K:
1781 		case BPF_JMP32 | BPF_JGE | BPF_K:
1782 		case BPF_JMP32 | BPF_JLE | BPF_K:
1783 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1784 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1785 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1786 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1787 			/* test dst_reg, dst_reg to save one extra byte */
1788 			if (imm32 == 0) {
1789 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1790 					       BPF_CLASS(insn->code) == BPF_JMP);
1791 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1792 				goto emit_cond_jmp;
1793 			}
1794 
1795 			/* cmp dst_reg, imm8/32 */
1796 			maybe_emit_1mod(&prog, dst_reg,
1797 					BPF_CLASS(insn->code) == BPF_JMP);
1798 
1799 			if (is_imm8(imm32))
1800 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1801 			else
1802 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1803 
1804 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1805 			switch (BPF_OP(insn->code)) {
1806 			case BPF_JEQ:
1807 				jmp_cond = X86_JE;
1808 				break;
1809 			case BPF_JSET:
1810 			case BPF_JNE:
1811 				jmp_cond = X86_JNE;
1812 				break;
1813 			case BPF_JGT:
1814 				/* GT is unsigned '>', JA in x86 */
1815 				jmp_cond = X86_JA;
1816 				break;
1817 			case BPF_JLT:
1818 				/* LT is unsigned '<', JB in x86 */
1819 				jmp_cond = X86_JB;
1820 				break;
1821 			case BPF_JGE:
1822 				/* GE is unsigned '>=', JAE in x86 */
1823 				jmp_cond = X86_JAE;
1824 				break;
1825 			case BPF_JLE:
1826 				/* LE is unsigned '<=', JBE in x86 */
1827 				jmp_cond = X86_JBE;
1828 				break;
1829 			case BPF_JSGT:
1830 				/* Signed '>', GT in x86 */
1831 				jmp_cond = X86_JG;
1832 				break;
1833 			case BPF_JSLT:
1834 				/* Signed '<', LT in x86 */
1835 				jmp_cond = X86_JL;
1836 				break;
1837 			case BPF_JSGE:
1838 				/* Signed '>=', GE in x86 */
1839 				jmp_cond = X86_JGE;
1840 				break;
1841 			case BPF_JSLE:
1842 				/* Signed '<=', LE in x86 */
1843 				jmp_cond = X86_JLE;
1844 				break;
1845 			default: /* to silence GCC warning */
1846 				return -EFAULT;
1847 			}
1848 			jmp_offset = addrs[i + insn->off] - addrs[i];
1849 			if (is_imm8(jmp_offset)) {
1850 				if (jmp_padding) {
1851 					/* To keep the jmp_offset valid, the extra bytes are
1852 					 * padded before the jump insn, so we subtract the
1853 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1854 					 *
1855 					 * If the previous pass already emits an imm8
1856 					 * jmp_cond, then this BPF insn won't shrink, so
1857 					 * "nops" is 0.
1858 					 *
1859 					 * On the other hand, if the previous pass emits an
1860 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1861 					 * keep the image from shrinking further.
1862 					 *
1863 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1864 					 *     is 2 bytes, so the size difference is 4 bytes.
1865 					 */
1866 					nops = INSN_SZ_DIFF - 2;
1867 					if (nops != 0 && nops != 4) {
1868 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1869 						       nops);
1870 						return -EFAULT;
1871 					}
1872 					emit_nops(&prog, nops);
1873 				}
1874 				EMIT2(jmp_cond, jmp_offset);
1875 			} else if (is_simm32(jmp_offset)) {
1876 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1877 			} else {
1878 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1879 				return -EFAULT;
1880 			}
1881 
1882 			break;
1883 
1884 		case BPF_JMP | BPF_JA:
1885 		case BPF_JMP32 | BPF_JA:
1886 			if (BPF_CLASS(insn->code) == BPF_JMP) {
1887 				if (insn->off == -1)
1888 					/* -1 jmp instructions will always jump
1889 					 * backwards two bytes. Explicitly handling
1890 					 * this case avoids wasting too many passes
1891 					 * when there are long sequences of replaced
1892 					 * dead code.
1893 					 */
1894 					jmp_offset = -2;
1895 				else
1896 					jmp_offset = addrs[i + insn->off] - addrs[i];
1897 			} else {
1898 				if (insn->imm == -1)
1899 					jmp_offset = -2;
1900 				else
1901 					jmp_offset = addrs[i + insn->imm] - addrs[i];
1902 			}
1903 
1904 			if (!jmp_offset) {
1905 				/*
1906 				 * If jmp_padding is enabled, the extra nops will
1907 				 * be inserted. Otherwise, optimize out nop jumps.
1908 				 */
1909 				if (jmp_padding) {
1910 					/* There are 3 possible conditions.
1911 					 * (1) This BPF_JA is already optimized out in
1912 					 *     the previous run, so there is no need
1913 					 *     to pad any extra byte (0 byte).
1914 					 * (2) The previous pass emits an imm8 jmp,
1915 					 *     so we pad 2 bytes to match the previous
1916 					 *     insn size.
1917 					 * (3) Similarly, the previous pass emits an
1918 					 *     imm32 jmp, and 5 bytes is padded.
1919 					 */
1920 					nops = INSN_SZ_DIFF;
1921 					if (nops != 0 && nops != 2 && nops != 5) {
1922 						pr_err("unexpected nop jump padding: %d bytes\n",
1923 						       nops);
1924 						return -EFAULT;
1925 					}
1926 					emit_nops(&prog, nops);
1927 				}
1928 				break;
1929 			}
1930 emit_jmp:
1931 			if (is_imm8(jmp_offset)) {
1932 				if (jmp_padding) {
1933 					/* To avoid breaking jmp_offset, the extra bytes
1934 					 * are padded before the actual jmp insn, so
1935 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1936 					 *
1937 					 * If the previous pass already emits an imm8
1938 					 * jmp, there is nothing to pad (0 byte).
1939 					 *
1940 					 * If it emits an imm32 jmp (5 bytes) previously
1941 					 * and now an imm8 jmp (2 bytes), then we pad
1942 					 * (5 - 2 = 3) bytes to stop the image from
1943 					 * shrinking further.
1944 					 */
1945 					nops = INSN_SZ_DIFF - 2;
1946 					if (nops != 0 && nops != 3) {
1947 						pr_err("unexpected jump padding: %d bytes\n",
1948 						       nops);
1949 						return -EFAULT;
1950 					}
1951 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1952 				}
1953 				EMIT2(0xEB, jmp_offset);
1954 			} else if (is_simm32(jmp_offset)) {
1955 				EMIT1_off32(0xE9, jmp_offset);
1956 			} else {
1957 				pr_err("jmp gen bug %llx\n", jmp_offset);
1958 				return -EFAULT;
1959 			}
1960 			break;
1961 
1962 		case BPF_JMP | BPF_EXIT:
1963 			if (seen_exit) {
1964 				jmp_offset = ctx->cleanup_addr - addrs[i];
1965 				goto emit_jmp;
1966 			}
1967 			seen_exit = true;
1968 			/* Update cleanup_addr */
1969 			ctx->cleanup_addr = proglen;
1970 			if (bpf_prog->aux->exception_boundary) {
1971 				pop_callee_regs(&prog, all_callee_regs_used);
1972 				pop_r12(&prog);
1973 			} else {
1974 				pop_callee_regs(&prog, callee_regs_used);
1975 			}
1976 			EMIT1(0xC9);         /* leave */
1977 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1978 			break;
1979 
1980 		default:
1981 			/*
1982 			 * By design x86-64 JIT should support all BPF instructions.
1983 			 * This error will be seen if new instruction was added
1984 			 * to the interpreter, but not to the JIT, or if there is
1985 			 * junk in bpf_prog.
1986 			 */
1987 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1988 			return -EINVAL;
1989 		}
1990 
1991 		ilen = prog - temp;
1992 		if (ilen > BPF_MAX_INSN_SIZE) {
1993 			pr_err("bpf_jit: fatal insn size error\n");
1994 			return -EFAULT;
1995 		}
1996 
1997 		if (image) {
1998 			/*
1999 			 * When populating the image, assert that:
2000 			 *
2001 			 *  i) We do not write beyond the allocated space, and
2002 			 * ii) addrs[i] did not change from the prior run, in order
2003 			 *     to validate assumptions made for computing branch
2004 			 *     displacements.
2005 			 */
2006 			if (unlikely(proglen + ilen > oldproglen ||
2007 				     proglen + ilen != addrs[i])) {
2008 				pr_err("bpf_jit: fatal error\n");
2009 				return -EFAULT;
2010 			}
2011 			memcpy(rw_image + proglen, temp, ilen);
2012 		}
2013 		proglen += ilen;
2014 		addrs[i] = proglen;
2015 		prog = temp;
2016 	}
2017 
2018 	if (image && excnt != bpf_prog->aux->num_exentries) {
2019 		pr_err("extable is not populated\n");
2020 		return -EFAULT;
2021 	}
2022 	return proglen;
2023 }
2024 
2025 static void clean_stack_garbage(const struct btf_func_model *m,
2026 				u8 **pprog, int nr_stack_slots,
2027 				int stack_size)
2028 {
2029 	int arg_size, off;
2030 	u8 *prog;
2031 
2032 	/* Generally speaking, the compiler will pass the arguments
2033 	 * on-stack with "push" instruction, which will take 8-byte
2034 	 * on the stack. In this case, there won't be garbage values
2035 	 * while we copy the arguments from origin stack frame to current
2036 	 * in BPF_DW.
2037 	 *
2038 	 * However, sometimes the compiler will only allocate 4-byte on
2039 	 * the stack for the arguments. For now, this case will only
2040 	 * happen if there is only one argument on-stack and its size
2041 	 * not more than 4 byte. In this case, there will be garbage
2042 	 * values on the upper 4-byte where we store the argument on
2043 	 * current stack frame.
2044 	 *
2045 	 * arguments on origin stack:
2046 	 *
2047 	 * stack_arg_1(4-byte) xxx(4-byte)
2048 	 *
2049 	 * what we copy:
2050 	 *
2051 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
2052 	 *
2053 	 * and the xxx is the garbage values which we should clean here.
2054 	 */
2055 	if (nr_stack_slots != 1)
2056 		return;
2057 
2058 	/* the size of the last argument */
2059 	arg_size = m->arg_size[m->nr_args - 1];
2060 	if (arg_size <= 4) {
2061 		off = -(stack_size - 4);
2062 		prog = *pprog;
2063 		/* mov DWORD PTR [rbp + off], 0 */
2064 		if (!is_imm8(off))
2065 			EMIT2_off32(0xC7, 0x85, off);
2066 		else
2067 			EMIT3(0xC7, 0x45, off);
2068 		EMIT(0, 4);
2069 		*pprog = prog;
2070 	}
2071 }
2072 
2073 /* get the count of the regs that are used to pass arguments */
2074 static int get_nr_used_regs(const struct btf_func_model *m)
2075 {
2076 	int i, arg_regs, nr_used_regs = 0;
2077 
2078 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2079 		arg_regs = (m->arg_size[i] + 7) / 8;
2080 		if (nr_used_regs + arg_regs <= 6)
2081 			nr_used_regs += arg_regs;
2082 
2083 		if (nr_used_regs >= 6)
2084 			break;
2085 	}
2086 
2087 	return nr_used_regs;
2088 }
2089 
2090 static void save_args(const struct btf_func_model *m, u8 **prog,
2091 		      int stack_size, bool for_call_origin)
2092 {
2093 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2094 	int i, j;
2095 
2096 	/* Store function arguments to stack.
2097 	 * For a function that accepts two pointers the sequence will be:
2098 	 * mov QWORD PTR [rbp-0x10],rdi
2099 	 * mov QWORD PTR [rbp-0x8],rsi
2100 	 */
2101 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2102 		arg_regs = (m->arg_size[i] + 7) / 8;
2103 
2104 		/* According to the research of Yonghong, struct members
2105 		 * should be all in register or all on the stack.
2106 		 * Meanwhile, the compiler will pass the argument on regs
2107 		 * if the remaining regs can hold the argument.
2108 		 *
2109 		 * Disorder of the args can happen. For example:
2110 		 *
2111 		 * struct foo_struct {
2112 		 *     long a;
2113 		 *     int b;
2114 		 * };
2115 		 * int foo(char, char, char, char, char, struct foo_struct,
2116 		 *         char);
2117 		 *
2118 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2119 		 * by stack.
2120 		 */
2121 		if (nr_regs + arg_regs > 6) {
2122 			/* copy function arguments from origin stack frame
2123 			 * into current stack frame.
2124 			 *
2125 			 * The starting address of the arguments on-stack
2126 			 * is:
2127 			 *   rbp + 8(push rbp) +
2128 			 *   8(return addr of origin call) +
2129 			 *   8(return addr of the caller)
2130 			 * which means: rbp + 24
2131 			 */
2132 			for (j = 0; j < arg_regs; j++) {
2133 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2134 					 nr_stack_slots * 8 + 0x18);
2135 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2136 					 -stack_size);
2137 
2138 				if (!nr_stack_slots)
2139 					first_off = stack_size;
2140 				stack_size -= 8;
2141 				nr_stack_slots++;
2142 			}
2143 		} else {
2144 			/* Only copy the arguments on-stack to current
2145 			 * 'stack_size' and ignore the regs, used to
2146 			 * prepare the arguments on-stack for orign call.
2147 			 */
2148 			if (for_call_origin) {
2149 				nr_regs += arg_regs;
2150 				continue;
2151 			}
2152 
2153 			/* copy the arguments from regs into stack */
2154 			for (j = 0; j < arg_regs; j++) {
2155 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2156 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2157 					 -stack_size);
2158 				stack_size -= 8;
2159 				nr_regs++;
2160 			}
2161 		}
2162 	}
2163 
2164 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2165 }
2166 
2167 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2168 			 int stack_size)
2169 {
2170 	int i, j, arg_regs, nr_regs = 0;
2171 
2172 	/* Restore function arguments from stack.
2173 	 * For a function that accepts two pointers the sequence will be:
2174 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2175 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2176 	 *
2177 	 * The logic here is similar to what we do in save_args()
2178 	 */
2179 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2180 		arg_regs = (m->arg_size[i] + 7) / 8;
2181 		if (nr_regs + arg_regs <= 6) {
2182 			for (j = 0; j < arg_regs; j++) {
2183 				emit_ldx(prog, BPF_DW,
2184 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2185 					 BPF_REG_FP,
2186 					 -stack_size);
2187 				stack_size -= 8;
2188 				nr_regs++;
2189 			}
2190 		} else {
2191 			stack_size -= 8 * arg_regs;
2192 		}
2193 
2194 		if (nr_regs >= 6)
2195 			break;
2196 	}
2197 }
2198 
2199 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2200 			   struct bpf_tramp_link *l, int stack_size,
2201 			   int run_ctx_off, bool save_ret)
2202 {
2203 	u8 *prog = *pprog;
2204 	u8 *jmp_insn;
2205 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2206 	struct bpf_prog *p = l->link.prog;
2207 	u64 cookie = l->cookie;
2208 
2209 	/* mov rdi, cookie */
2210 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2211 
2212 	/* Prepare struct bpf_tramp_run_ctx.
2213 	 *
2214 	 * bpf_tramp_run_ctx is already preserved by
2215 	 * arch_prepare_bpf_trampoline().
2216 	 *
2217 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2218 	 */
2219 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2220 
2221 	/* arg1: mov rdi, progs[i] */
2222 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2223 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2224 	if (!is_imm8(-run_ctx_off))
2225 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
2226 	else
2227 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
2228 
2229 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), prog))
2230 		return -EINVAL;
2231 	/* remember prog start time returned by __bpf_prog_enter */
2232 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
2233 
2234 	/* if (__bpf_prog_enter*(prog) == 0)
2235 	 *	goto skip_exec_of_prog;
2236 	 */
2237 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
2238 	/* emit 2 nops that will be replaced with JE insn */
2239 	jmp_insn = prog;
2240 	emit_nops(&prog, 2);
2241 
2242 	/* arg1: lea rdi, [rbp - stack_size] */
2243 	if (!is_imm8(-stack_size))
2244 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
2245 	else
2246 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
2247 	/* arg2: progs[i]->insnsi for interpreter */
2248 	if (!p->jited)
2249 		emit_mov_imm64(&prog, BPF_REG_2,
2250 			       (long) p->insnsi >> 32,
2251 			       (u32) (long) p->insnsi);
2252 	/* call JITed bpf program or interpreter */
2253 	if (emit_rsb_call(&prog, p->bpf_func, prog))
2254 		return -EINVAL;
2255 
2256 	/*
2257 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
2258 	 * of the previous call which is then passed on the stack to
2259 	 * the next BPF program.
2260 	 *
2261 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
2262 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
2263 	 */
2264 	if (save_ret)
2265 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2266 
2267 	/* replace 2 nops with JE insn, since jmp target is known */
2268 	jmp_insn[0] = X86_JE;
2269 	jmp_insn[1] = prog - jmp_insn - 2;
2270 
2271 	/* arg1: mov rdi, progs[i] */
2272 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2273 	/* arg2: mov rsi, rbx <- start time in nsec */
2274 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
2275 	/* arg3: lea rdx, [rbp - run_ctx_off] */
2276 	if (!is_imm8(-run_ctx_off))
2277 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
2278 	else
2279 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
2280 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), prog))
2281 		return -EINVAL;
2282 
2283 	*pprog = prog;
2284 	return 0;
2285 }
2286 
2287 static void emit_align(u8 **pprog, u32 align)
2288 {
2289 	u8 *target, *prog = *pprog;
2290 
2291 	target = PTR_ALIGN(prog, align);
2292 	if (target != prog)
2293 		emit_nops(&prog, target - prog);
2294 
2295 	*pprog = prog;
2296 }
2297 
2298 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2299 {
2300 	u8 *prog = *pprog;
2301 	s64 offset;
2302 
2303 	offset = func - (ip + 2 + 4);
2304 	if (!is_simm32(offset)) {
2305 		pr_err("Target %p is out of range\n", func);
2306 		return -EINVAL;
2307 	}
2308 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2309 	*pprog = prog;
2310 	return 0;
2311 }
2312 
2313 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2314 		      struct bpf_tramp_links *tl, int stack_size,
2315 		      int run_ctx_off, bool save_ret)
2316 {
2317 	int i;
2318 	u8 *prog = *pprog;
2319 
2320 	for (i = 0; i < tl->nr_links; i++) {
2321 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2322 				    run_ctx_off, save_ret))
2323 			return -EINVAL;
2324 	}
2325 	*pprog = prog;
2326 	return 0;
2327 }
2328 
2329 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2330 			      struct bpf_tramp_links *tl, int stack_size,
2331 			      int run_ctx_off, u8 **branches)
2332 {
2333 	u8 *prog = *pprog;
2334 	int i;
2335 
2336 	/* The first fmod_ret program will receive a garbage return value.
2337 	 * Set this to 0 to avoid confusing the program.
2338 	 */
2339 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2340 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2341 	for (i = 0; i < tl->nr_links; i++) {
2342 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
2343 			return -EINVAL;
2344 
2345 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2346 		 * if (*(u64 *)(rbp - 8) !=  0)
2347 		 *	goto do_fexit;
2348 		 */
2349 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2350 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2351 
2352 		/* Save the location of the branch and Generate 6 nops
2353 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2354 		 * are replaced with a conditional jump once do_fexit (i.e. the
2355 		 * start of the fexit invocation) is finalized.
2356 		 */
2357 		branches[i] = prog;
2358 		emit_nops(&prog, 4 + 2);
2359 	}
2360 
2361 	*pprog = prog;
2362 	return 0;
2363 }
2364 
2365 /* Example:
2366  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2367  * its 'struct btf_func_model' will be nr_args=2
2368  * The assembly code when eth_type_trans is executing after trampoline:
2369  *
2370  * push rbp
2371  * mov rbp, rsp
2372  * sub rsp, 16                     // space for skb and dev
2373  * push rbx                        // temp regs to pass start time
2374  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2375  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2376  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2377  * mov rbx, rax                    // remember start time in bpf stats are enabled
2378  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2379  * call addr_of_jited_FENTRY_prog
2380  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2381  * mov rsi, rbx                    // prog start time
2382  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2383  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2384  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2385  * pop rbx
2386  * leave
2387  * ret
2388  *
2389  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2390  * replaced with 'call generated_bpf_trampoline'. When it returns
2391  * eth_type_trans will continue executing with original skb and dev pointers.
2392  *
2393  * The assembly code when eth_type_trans is called from trampoline:
2394  *
2395  * push rbp
2396  * mov rbp, rsp
2397  * sub rsp, 24                     // space for skb, dev, return value
2398  * push rbx                        // temp regs to pass start time
2399  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2400  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2401  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2402  * mov rbx, rax                    // remember start time if bpf stats are enabled
2403  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2404  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2405  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2406  * mov rsi, rbx                    // prog start time
2407  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2408  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2409  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2410  * call eth_type_trans+5           // execute body of eth_type_trans
2411  * mov qword ptr [rbp - 8], rax    // save return value
2412  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2413  * mov rbx, rax                    // remember start time in bpf stats are enabled
2414  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2415  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2416  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2417  * mov rsi, rbx                    // prog start time
2418  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2419  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2420  * pop rbx
2421  * leave
2422  * add rsp, 8                      // skip eth_type_trans's frame
2423  * ret                             // return to its caller
2424  */
2425 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2426 				const struct btf_func_model *m, u32 flags,
2427 				struct bpf_tramp_links *tlinks,
2428 				void *func_addr)
2429 {
2430 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2431 	int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
2432 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2433 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2434 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2435 	void *orig_call = func_addr;
2436 	u8 **branches = NULL;
2437 	u8 *prog;
2438 	bool save_ret;
2439 
2440 	/* extra registers for struct arguments */
2441 	for (i = 0; i < m->nr_args; i++)
2442 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2443 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2444 
2445 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
2446 	 * are passed through regs, the remains are through stack.
2447 	 */
2448 	if (nr_regs > MAX_BPF_FUNC_ARGS)
2449 		return -ENOTSUPP;
2450 
2451 	/* Generated trampoline stack layout:
2452 	 *
2453 	 * RBP + 8         [ return address  ]
2454 	 * RBP + 0         [ RBP             ]
2455 	 *
2456 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2457 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2458 	 *
2459 	 *                 [ reg_argN        ]  always
2460 	 *                 [ ...             ]
2461 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2462 	 *
2463 	 * RBP - nregs_off [ regs count	     ]  always
2464 	 *
2465 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2466 	 *
2467 	 * RBP - rbx_off   [ rbx value       ]  always
2468 	 *
2469 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2470 	 *
2471 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
2472 	 *                     [ ...        ]
2473 	 *                     [ stack_arg2 ]
2474 	 * RBP - arg_stack_off [ stack_arg1 ]
2475 	 * RSP                 [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
2476 	 */
2477 
2478 	/* room for return value of orig_call or fentry prog */
2479 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2480 	if (save_ret)
2481 		stack_size += 8;
2482 
2483 	stack_size += nr_regs * 8;
2484 	regs_off = stack_size;
2485 
2486 	/* regs count  */
2487 	stack_size += 8;
2488 	nregs_off = stack_size;
2489 
2490 	if (flags & BPF_TRAMP_F_IP_ARG)
2491 		stack_size += 8; /* room for IP address argument */
2492 
2493 	ip_off = stack_size;
2494 
2495 	stack_size += 8;
2496 	rbx_off = stack_size;
2497 
2498 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2499 	run_ctx_off = stack_size;
2500 
2501 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
2502 		/* the space that used to pass arguments on-stack */
2503 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
2504 		/* make sure the stack pointer is 16-byte aligned if we
2505 		 * need pass arguments on stack, which means
2506 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
2507 		 * should be 16-byte aligned. Following code depend on
2508 		 * that stack_size is already 8-byte aligned.
2509 		 */
2510 		stack_size += (stack_size % 16) ? 0 : 8;
2511 	}
2512 
2513 	arg_stack_off = stack_size;
2514 
2515 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2516 		/* skip patched call instruction and point orig_call to actual
2517 		 * body of the kernel function.
2518 		 */
2519 		if (is_endbr(*(u32 *)orig_call))
2520 			orig_call += ENDBR_INSN_SIZE;
2521 		orig_call += X86_PATCH_SIZE;
2522 	}
2523 
2524 	prog = image;
2525 
2526 	EMIT_ENDBR();
2527 	/*
2528 	 * This is the direct-call trampoline, as such it needs accounting
2529 	 * for the __fentry__ call.
2530 	 */
2531 	x86_call_depth_emit_accounting(&prog, NULL);
2532 	EMIT1(0x55);		 /* push rbp */
2533 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2534 	if (!is_imm8(stack_size))
2535 		/* sub rsp, stack_size */
2536 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
2537 	else
2538 		/* sub rsp, stack_size */
2539 		EMIT4(0x48, 0x83, 0xEC, stack_size);
2540 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2541 		EMIT1(0x50);		/* push rax */
2542 	/* mov QWORD PTR [rbp - rbx_off], rbx */
2543 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
2544 
2545 	/* Store number of argument registers of the traced function:
2546 	 *   mov rax, nr_regs
2547 	 *   mov QWORD PTR [rbp - nregs_off], rax
2548 	 */
2549 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2550 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2551 
2552 	if (flags & BPF_TRAMP_F_IP_ARG) {
2553 		/* Store IP address of the traced function:
2554 		 * movabsq rax, func_addr
2555 		 * mov QWORD PTR [rbp - ip_off], rax
2556 		 */
2557 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2558 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2559 	}
2560 
2561 	save_args(m, &prog, regs_off, false);
2562 
2563 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2564 		/* arg1: mov rdi, im */
2565 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2566 		if (emit_rsb_call(&prog, __bpf_tramp_enter, prog)) {
2567 			ret = -EINVAL;
2568 			goto cleanup;
2569 		}
2570 	}
2571 
2572 	if (fentry->nr_links)
2573 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2574 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2575 			return -EINVAL;
2576 
2577 	if (fmod_ret->nr_links) {
2578 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2579 				   GFP_KERNEL);
2580 		if (!branches)
2581 			return -ENOMEM;
2582 
2583 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2584 				       run_ctx_off, branches)) {
2585 			ret = -EINVAL;
2586 			goto cleanup;
2587 		}
2588 	}
2589 
2590 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2591 		restore_regs(m, &prog, regs_off);
2592 		save_args(m, &prog, arg_stack_off, true);
2593 
2594 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2595 			/* Before calling the original function, restore the
2596 			 * tail_call_cnt from stack to rax.
2597 			 */
2598 			RESTORE_TAIL_CALL_CNT(stack_size);
2599 
2600 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2601 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
2602 			EMIT2(0xff, 0xd3); /* call *rbx */
2603 		} else {
2604 			/* call original function */
2605 			if (emit_rsb_call(&prog, orig_call, prog)) {
2606 				ret = -EINVAL;
2607 				goto cleanup;
2608 			}
2609 		}
2610 		/* remember return value in a stack for bpf prog to access */
2611 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2612 		im->ip_after_call = prog;
2613 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2614 		prog += X86_PATCH_SIZE;
2615 	}
2616 
2617 	if (fmod_ret->nr_links) {
2618 		/* From Intel 64 and IA-32 Architectures Optimization
2619 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2620 		 * Coding Rule 11: All branch targets should be 16-byte
2621 		 * aligned.
2622 		 */
2623 		emit_align(&prog, 16);
2624 		/* Update the branches saved in invoke_bpf_mod_ret with the
2625 		 * aligned address of do_fexit.
2626 		 */
2627 		for (i = 0; i < fmod_ret->nr_links; i++)
2628 			emit_cond_near_jump(&branches[i], prog, branches[i],
2629 					    X86_JNE);
2630 	}
2631 
2632 	if (fexit->nr_links)
2633 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2634 			ret = -EINVAL;
2635 			goto cleanup;
2636 		}
2637 
2638 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2639 		restore_regs(m, &prog, regs_off);
2640 
2641 	/* This needs to be done regardless. If there were fmod_ret programs,
2642 	 * the return value is only updated on the stack and still needs to be
2643 	 * restored to R0.
2644 	 */
2645 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2646 		im->ip_epilogue = prog;
2647 		/* arg1: mov rdi, im */
2648 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2649 		if (emit_rsb_call(&prog, __bpf_tramp_exit, prog)) {
2650 			ret = -EINVAL;
2651 			goto cleanup;
2652 		}
2653 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2654 		/* Before running the original function, restore the
2655 		 * tail_call_cnt from stack to rax.
2656 		 */
2657 		RESTORE_TAIL_CALL_CNT(stack_size);
2658 
2659 	/* restore return value of orig_call or fentry prog back into RAX */
2660 	if (save_ret)
2661 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2662 
2663 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
2664 	EMIT1(0xC9); /* leave */
2665 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2666 		/* skip our return address and return to parent */
2667 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2668 	emit_return(&prog, prog);
2669 	/* Make sure the trampoline generation logic doesn't overflow */
2670 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2671 		ret = -EFAULT;
2672 		goto cleanup;
2673 	}
2674 	ret = prog - (u8 *)image;
2675 
2676 cleanup:
2677 	kfree(branches);
2678 	return ret;
2679 }
2680 
2681 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2682 {
2683 	u8 *jg_reloc, *prog = *pprog;
2684 	int pivot, err, jg_bytes = 1;
2685 	s64 jg_offset;
2686 
2687 	if (a == b) {
2688 		/* Leaf node of recursion, i.e. not a range of indices
2689 		 * anymore.
2690 		 */
2691 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2692 		if (!is_simm32(progs[a]))
2693 			return -1;
2694 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2695 			    progs[a]);
2696 		err = emit_cond_near_jump(&prog,	/* je func */
2697 					  (void *)progs[a], image + (prog - buf),
2698 					  X86_JE);
2699 		if (err)
2700 			return err;
2701 
2702 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2703 
2704 		*pprog = prog;
2705 		return 0;
2706 	}
2707 
2708 	/* Not a leaf node, so we pivot, and recursively descend into
2709 	 * the lower and upper ranges.
2710 	 */
2711 	pivot = (b - a) / 2;
2712 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2713 	if (!is_simm32(progs[a + pivot]))
2714 		return -1;
2715 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2716 
2717 	if (pivot > 2) {				/* jg upper_part */
2718 		/* Require near jump. */
2719 		jg_bytes = 4;
2720 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2721 	} else {
2722 		EMIT2(X86_JG, 0);
2723 	}
2724 	jg_reloc = prog;
2725 
2726 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2727 				  progs, image, buf);
2728 	if (err)
2729 		return err;
2730 
2731 	/* From Intel 64 and IA-32 Architectures Optimization
2732 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2733 	 * Coding Rule 11: All branch targets should be 16-byte
2734 	 * aligned.
2735 	 */
2736 	emit_align(&prog, 16);
2737 	jg_offset = prog - jg_reloc;
2738 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2739 
2740 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2741 				  b, progs, image, buf);
2742 	if (err)
2743 		return err;
2744 
2745 	*pprog = prog;
2746 	return 0;
2747 }
2748 
2749 static int cmp_ips(const void *a, const void *b)
2750 {
2751 	const s64 *ipa = a;
2752 	const s64 *ipb = b;
2753 
2754 	if (*ipa > *ipb)
2755 		return 1;
2756 	if (*ipa < *ipb)
2757 		return -1;
2758 	return 0;
2759 }
2760 
2761 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2762 {
2763 	u8 *prog = buf;
2764 
2765 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2766 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2767 }
2768 
2769 struct x64_jit_data {
2770 	struct bpf_binary_header *rw_header;
2771 	struct bpf_binary_header *header;
2772 	int *addrs;
2773 	u8 *image;
2774 	int proglen;
2775 	struct jit_context ctx;
2776 };
2777 
2778 #define MAX_PASSES 20
2779 #define PADDING_PASSES (MAX_PASSES - 5)
2780 
2781 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2782 {
2783 	struct bpf_binary_header *rw_header = NULL;
2784 	struct bpf_binary_header *header = NULL;
2785 	struct bpf_prog *tmp, *orig_prog = prog;
2786 	struct x64_jit_data *jit_data;
2787 	int proglen, oldproglen = 0;
2788 	struct jit_context ctx = {};
2789 	bool tmp_blinded = false;
2790 	bool extra_pass = false;
2791 	bool padding = false;
2792 	u8 *rw_image = NULL;
2793 	u8 *image = NULL;
2794 	int *addrs;
2795 	int pass;
2796 	int i;
2797 
2798 	if (!prog->jit_requested)
2799 		return orig_prog;
2800 
2801 	tmp = bpf_jit_blind_constants(prog);
2802 	/*
2803 	 * If blinding was requested and we failed during blinding,
2804 	 * we must fall back to the interpreter.
2805 	 */
2806 	if (IS_ERR(tmp))
2807 		return orig_prog;
2808 	if (tmp != prog) {
2809 		tmp_blinded = true;
2810 		prog = tmp;
2811 	}
2812 
2813 	jit_data = prog->aux->jit_data;
2814 	if (!jit_data) {
2815 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2816 		if (!jit_data) {
2817 			prog = orig_prog;
2818 			goto out;
2819 		}
2820 		prog->aux->jit_data = jit_data;
2821 	}
2822 	addrs = jit_data->addrs;
2823 	if (addrs) {
2824 		ctx = jit_data->ctx;
2825 		oldproglen = jit_data->proglen;
2826 		image = jit_data->image;
2827 		header = jit_data->header;
2828 		rw_header = jit_data->rw_header;
2829 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2830 		extra_pass = true;
2831 		padding = true;
2832 		goto skip_init_addrs;
2833 	}
2834 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2835 	if (!addrs) {
2836 		prog = orig_prog;
2837 		goto out_addrs;
2838 	}
2839 
2840 	/*
2841 	 * Before first pass, make a rough estimation of addrs[]
2842 	 * each BPF instruction is translated to less than 64 bytes
2843 	 */
2844 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2845 		proglen += 64;
2846 		addrs[i] = proglen;
2847 	}
2848 	ctx.cleanup_addr = proglen;
2849 skip_init_addrs:
2850 
2851 	/*
2852 	 * JITed image shrinks with every pass and the loop iterates
2853 	 * until the image stops shrinking. Very large BPF programs
2854 	 * may converge on the last pass. In such case do one more
2855 	 * pass to emit the final image.
2856 	 */
2857 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2858 		if (!padding && pass >= PADDING_PASSES)
2859 			padding = true;
2860 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2861 		if (proglen <= 0) {
2862 out_image:
2863 			image = NULL;
2864 			if (header) {
2865 				bpf_arch_text_copy(&header->size, &rw_header->size,
2866 						   sizeof(rw_header->size));
2867 				bpf_jit_binary_pack_free(header, rw_header);
2868 			}
2869 			/* Fall back to interpreter mode */
2870 			prog = orig_prog;
2871 			if (extra_pass) {
2872 				prog->bpf_func = NULL;
2873 				prog->jited = 0;
2874 				prog->jited_len = 0;
2875 			}
2876 			goto out_addrs;
2877 		}
2878 		if (image) {
2879 			if (proglen != oldproglen) {
2880 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2881 				       proglen, oldproglen);
2882 				goto out_image;
2883 			}
2884 			break;
2885 		}
2886 		if (proglen == oldproglen) {
2887 			/*
2888 			 * The number of entries in extable is the number of BPF_LDX
2889 			 * insns that access kernel memory via "pointer to BTF type".
2890 			 * The verifier changed their opcode from LDX|MEM|size
2891 			 * to LDX|PROBE_MEM|size to make JITing easier.
2892 			 */
2893 			u32 align = __alignof__(struct exception_table_entry);
2894 			u32 extable_size = prog->aux->num_exentries *
2895 				sizeof(struct exception_table_entry);
2896 
2897 			/* allocate module memory for x86 insns and extable */
2898 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2899 							   &image, align, &rw_header, &rw_image,
2900 							   jit_fill_hole);
2901 			if (!header) {
2902 				prog = orig_prog;
2903 				goto out_addrs;
2904 			}
2905 			prog->aux->extable = (void *) image + roundup(proglen, align);
2906 		}
2907 		oldproglen = proglen;
2908 		cond_resched();
2909 	}
2910 
2911 	if (bpf_jit_enable > 1)
2912 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
2913 
2914 	if (image) {
2915 		if (!prog->is_func || extra_pass) {
2916 			/*
2917 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2918 			 *   1) header is not pointing to proper module memory;
2919 			 *   2) the arch doesn't support bpf_arch_text_copy().
2920 			 *
2921 			 * Both cases are serious bugs and justify WARN_ON.
2922 			 */
2923 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2924 				/* header has been freed */
2925 				header = NULL;
2926 				goto out_image;
2927 			}
2928 
2929 			bpf_tail_call_direct_fixup(prog);
2930 		} else {
2931 			jit_data->addrs = addrs;
2932 			jit_data->ctx = ctx;
2933 			jit_data->proglen = proglen;
2934 			jit_data->image = image;
2935 			jit_data->header = header;
2936 			jit_data->rw_header = rw_header;
2937 		}
2938 		prog->bpf_func = (void *)image;
2939 		prog->jited = 1;
2940 		prog->jited_len = proglen;
2941 	} else {
2942 		prog = orig_prog;
2943 	}
2944 
2945 	if (!image || !prog->is_func || extra_pass) {
2946 		if (image)
2947 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2948 out_addrs:
2949 		kvfree(addrs);
2950 		kfree(jit_data);
2951 		prog->aux->jit_data = NULL;
2952 	}
2953 out:
2954 	if (tmp_blinded)
2955 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2956 					   tmp : orig_prog);
2957 	return prog;
2958 }
2959 
2960 bool bpf_jit_supports_kfunc_call(void)
2961 {
2962 	return true;
2963 }
2964 
2965 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2966 {
2967 	if (text_poke_copy(dst, src, len) == NULL)
2968 		return ERR_PTR(-EINVAL);
2969 	return dst;
2970 }
2971 
2972 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2973 bool bpf_jit_supports_subprog_tailcalls(void)
2974 {
2975 	return true;
2976 }
2977 
2978 void bpf_jit_free(struct bpf_prog *prog)
2979 {
2980 	if (prog->jited) {
2981 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2982 		struct bpf_binary_header *hdr;
2983 
2984 		/*
2985 		 * If we fail the final pass of JIT (from jit_subprogs),
2986 		 * the program may not be finalized yet. Call finalize here
2987 		 * before freeing it.
2988 		 */
2989 		if (jit_data) {
2990 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2991 						     jit_data->rw_header);
2992 			kvfree(jit_data->addrs);
2993 			kfree(jit_data);
2994 		}
2995 		hdr = bpf_jit_binary_pack_hdr(prog);
2996 		bpf_jit_binary_pack_free(hdr, NULL);
2997 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2998 	}
2999 
3000 	bpf_prog_unlock_free(prog);
3001 }
3002 
3003 bool bpf_jit_supports_exceptions(void)
3004 {
3005 	/* We unwind through both kernel frames (starting from within bpf_throw
3006 	 * call) and BPF frames. Therefore we require ORC unwinder to be enabled
3007 	 * to walk kernel frames and reach BPF frames in the stack trace.
3008 	 */
3009 	return IS_ENABLED(CONFIG_UNWINDER_ORC);
3010 }
3011 
3012 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3013 {
3014 #if defined(CONFIG_UNWINDER_ORC)
3015 	struct unwind_state state;
3016 	unsigned long addr;
3017 
3018 	for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
3019 	     unwind_next_frame(&state)) {
3020 		addr = unwind_get_return_address(&state);
3021 		if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
3022 			break;
3023 	}
3024 	return;
3025 #endif
3026 	WARN(1, "verification of programs using bpf_throw should have failed\n");
3027 }
3028 
3029 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
3030 			       struct bpf_prog *new, struct bpf_prog *old)
3031 {
3032 	u8 *old_addr, *new_addr, *old_bypass_addr;
3033 	int ret;
3034 
3035 	old_bypass_addr = old ? NULL : poke->bypass_addr;
3036 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
3037 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
3038 
3039 	/*
3040 	 * On program loading or teardown, the program's kallsym entry
3041 	 * might not be in place, so we use __bpf_arch_text_poke to skip
3042 	 * the kallsyms check.
3043 	 */
3044 	if (new) {
3045 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3046 					   BPF_MOD_JUMP,
3047 					   old_addr, new_addr);
3048 		BUG_ON(ret < 0);
3049 		if (!old) {
3050 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3051 						   BPF_MOD_JUMP,
3052 						   poke->bypass_addr,
3053 						   NULL);
3054 			BUG_ON(ret < 0);
3055 		}
3056 	} else {
3057 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
3058 					   BPF_MOD_JUMP,
3059 					   old_bypass_addr,
3060 					   poke->bypass_addr);
3061 		BUG_ON(ret < 0);
3062 		/* let other CPUs finish the execution of program
3063 		 * so that it will not possible to expose them
3064 		 * to invalid nop, stack unwind, nop state
3065 		 */
3066 		if (!ret)
3067 			synchronize_rcu();
3068 		ret = __bpf_arch_text_poke(poke->tailcall_target,
3069 					   BPF_MOD_JUMP,
3070 					   old_addr, NULL);
3071 		BUG_ON(ret < 0);
3072 	}
3073 }
3074