xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision 7ae5c03a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * BPF JIT compiler
4  *
5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6  * Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
7  */
8 #include <linux/netdevice.h>
9 #include <linux/filter.h>
10 #include <linux/if_vlan.h>
11 #include <linux/bpf.h>
12 #include <linux/memory.h>
13 #include <linux/sort.h>
14 #include <asm/extable.h>
15 #include <asm/set_memory.h>
16 #include <asm/nospec-branch.h>
17 #include <asm/text-patching.h>
18 
19 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
20 {
21 	if (len == 1)
22 		*ptr = bytes;
23 	else if (len == 2)
24 		*(u16 *)ptr = bytes;
25 	else {
26 		*(u32 *)ptr = bytes;
27 		barrier();
28 	}
29 	return ptr + len;
30 }
31 
32 #define EMIT(bytes, len) \
33 	do { prog = emit_code(prog, bytes, len); } while (0)
34 
35 #define EMIT1(b1)		EMIT(b1, 1)
36 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
37 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
38 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
39 
40 #define EMIT1_off32(b1, off) \
41 	do { EMIT1(b1); EMIT(off, 4); } while (0)
42 #define EMIT2_off32(b1, b2, off) \
43 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
44 #define EMIT3_off32(b1, b2, b3, off) \
45 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
46 #define EMIT4_off32(b1, b2, b3, b4, off) \
47 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
48 
49 #ifdef CONFIG_X86_KERNEL_IBT
50 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
51 #else
52 #define EMIT_ENDBR()
53 #endif
54 
55 static bool is_imm8(int value)
56 {
57 	return value <= 127 && value >= -128;
58 }
59 
60 static bool is_simm32(s64 value)
61 {
62 	return value == (s64)(s32)value;
63 }
64 
65 static bool is_uimm32(u64 value)
66 {
67 	return value == (u64)(u32)value;
68 }
69 
70 /* mov dst, src */
71 #define EMIT_mov(DST, SRC)								 \
72 	do {										 \
73 		if (DST != SRC)								 \
74 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
75 	} while (0)
76 
77 static int bpf_size_to_x86_bytes(int bpf_size)
78 {
79 	if (bpf_size == BPF_W)
80 		return 4;
81 	else if (bpf_size == BPF_H)
82 		return 2;
83 	else if (bpf_size == BPF_B)
84 		return 1;
85 	else if (bpf_size == BPF_DW)
86 		return 4; /* imm32 */
87 	else
88 		return 0;
89 }
90 
91 /*
92  * List of x86 cond jumps opcodes (. + s8)
93  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
94  */
95 #define X86_JB  0x72
96 #define X86_JAE 0x73
97 #define X86_JE  0x74
98 #define X86_JNE 0x75
99 #define X86_JBE 0x76
100 #define X86_JA  0x77
101 #define X86_JL  0x7C
102 #define X86_JGE 0x7D
103 #define X86_JLE 0x7E
104 #define X86_JG  0x7F
105 
106 /* Pick a register outside of BPF range for JIT internal work */
107 #define AUX_REG (MAX_BPF_JIT_REG + 1)
108 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
109 
110 /*
111  * The following table maps BPF registers to x86-64 registers.
112  *
113  * x86-64 register R12 is unused, since if used as base address
114  * register in load/store instructions, it always needs an
115  * extra byte of encoding and is callee saved.
116  *
117  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
118  * trampoline. x86-64 register R10 is used for blinding (if enabled).
119  */
120 static const int reg2hex[] = {
121 	[BPF_REG_0] = 0,  /* RAX */
122 	[BPF_REG_1] = 7,  /* RDI */
123 	[BPF_REG_2] = 6,  /* RSI */
124 	[BPF_REG_3] = 2,  /* RDX */
125 	[BPF_REG_4] = 1,  /* RCX */
126 	[BPF_REG_5] = 0,  /* R8  */
127 	[BPF_REG_6] = 3,  /* RBX callee saved */
128 	[BPF_REG_7] = 5,  /* R13 callee saved */
129 	[BPF_REG_8] = 6,  /* R14 callee saved */
130 	[BPF_REG_9] = 7,  /* R15 callee saved */
131 	[BPF_REG_FP] = 5, /* RBP readonly */
132 	[BPF_REG_AX] = 2, /* R10 temp register */
133 	[AUX_REG] = 3,    /* R11 temp register */
134 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
135 };
136 
137 static const int reg2pt_regs[] = {
138 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
139 	[BPF_REG_1] = offsetof(struct pt_regs, di),
140 	[BPF_REG_2] = offsetof(struct pt_regs, si),
141 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
142 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
143 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
144 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
145 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
146 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
147 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
148 };
149 
150 /*
151  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
152  * which need extra byte of encoding.
153  * rax,rcx,...,rbp have simpler encoding
154  */
155 static bool is_ereg(u32 reg)
156 {
157 	return (1 << reg) & (BIT(BPF_REG_5) |
158 			     BIT(AUX_REG) |
159 			     BIT(BPF_REG_7) |
160 			     BIT(BPF_REG_8) |
161 			     BIT(BPF_REG_9) |
162 			     BIT(X86_REG_R9) |
163 			     BIT(BPF_REG_AX));
164 }
165 
166 /*
167  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
168  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
169  * of encoding. al,cl,dl,bl have simpler encoding.
170  */
171 static bool is_ereg_8l(u32 reg)
172 {
173 	return is_ereg(reg) ||
174 	    (1 << reg) & (BIT(BPF_REG_1) |
175 			  BIT(BPF_REG_2) |
176 			  BIT(BPF_REG_FP));
177 }
178 
179 static bool is_axreg(u32 reg)
180 {
181 	return reg == BPF_REG_0;
182 }
183 
184 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
185 static u8 add_1mod(u8 byte, u32 reg)
186 {
187 	if (is_ereg(reg))
188 		byte |= 1;
189 	return byte;
190 }
191 
192 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
193 {
194 	if (is_ereg(r1))
195 		byte |= 1;
196 	if (is_ereg(r2))
197 		byte |= 4;
198 	return byte;
199 }
200 
201 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
202 static u8 add_1reg(u8 byte, u32 dst_reg)
203 {
204 	return byte + reg2hex[dst_reg];
205 }
206 
207 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
208 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
209 {
210 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
211 }
212 
213 /* Some 1-byte opcodes for binary ALU operations */
214 static u8 simple_alu_opcodes[] = {
215 	[BPF_ADD] = 0x01,
216 	[BPF_SUB] = 0x29,
217 	[BPF_AND] = 0x21,
218 	[BPF_OR] = 0x09,
219 	[BPF_XOR] = 0x31,
220 	[BPF_LSH] = 0xE0,
221 	[BPF_RSH] = 0xE8,
222 	[BPF_ARSH] = 0xF8,
223 };
224 
225 static void jit_fill_hole(void *area, unsigned int size)
226 {
227 	/* Fill whole space with INT3 instructions */
228 	memset(area, 0xcc, size);
229 }
230 
231 int bpf_arch_text_invalidate(void *dst, size_t len)
232 {
233 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
234 }
235 
236 struct jit_context {
237 	int cleanup_addr; /* Epilogue code offset */
238 
239 	/*
240 	 * Program specific offsets of labels in the code; these rely on the
241 	 * JIT doing at least 2 passes, recording the position on the first
242 	 * pass, only to generate the correct offset on the second pass.
243 	 */
244 	int tail_call_direct_label;
245 	int tail_call_indirect_label;
246 };
247 
248 /* Maximum number of bytes emitted while JITing one eBPF insn */
249 #define BPF_MAX_INSN_SIZE	128
250 #define BPF_INSN_SAFETY		64
251 
252 /* Number of bytes emit_patch() needs to generate instructions */
253 #define X86_PATCH_SIZE		5
254 /* Number of bytes that will be skipped on tailcall */
255 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
256 
257 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
258 {
259 	u8 *prog = *pprog;
260 
261 	if (callee_regs_used[0])
262 		EMIT1(0x53);         /* push rbx */
263 	if (callee_regs_used[1])
264 		EMIT2(0x41, 0x55);   /* push r13 */
265 	if (callee_regs_used[2])
266 		EMIT2(0x41, 0x56);   /* push r14 */
267 	if (callee_regs_used[3])
268 		EMIT2(0x41, 0x57);   /* push r15 */
269 	*pprog = prog;
270 }
271 
272 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
273 {
274 	u8 *prog = *pprog;
275 
276 	if (callee_regs_used[3])
277 		EMIT2(0x41, 0x5F);   /* pop r15 */
278 	if (callee_regs_used[2])
279 		EMIT2(0x41, 0x5E);   /* pop r14 */
280 	if (callee_regs_used[1])
281 		EMIT2(0x41, 0x5D);   /* pop r13 */
282 	if (callee_regs_used[0])
283 		EMIT1(0x5B);         /* pop rbx */
284 	*pprog = prog;
285 }
286 
287 /*
288  * Emit x86-64 prologue code for BPF program.
289  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
290  * while jumping to another program
291  */
292 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
293 			  bool tail_call_reachable, bool is_subprog)
294 {
295 	u8 *prog = *pprog;
296 
297 	/* BPF trampoline can be made to work without these nops,
298 	 * but let's waste 5 bytes for now and optimize later
299 	 */
300 	EMIT_ENDBR();
301 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
302 	prog += X86_PATCH_SIZE;
303 	if (!ebpf_from_cbpf) {
304 		if (tail_call_reachable && !is_subprog)
305 			EMIT2(0x31, 0xC0); /* xor eax, eax */
306 		else
307 			EMIT2(0x66, 0x90); /* nop2 */
308 	}
309 	EMIT1(0x55);             /* push rbp */
310 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
311 
312 	/* X86_TAIL_CALL_OFFSET is here */
313 	EMIT_ENDBR();
314 
315 	/* sub rsp, rounded_stack_depth */
316 	if (stack_depth)
317 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
318 	if (tail_call_reachable)
319 		EMIT1(0x50);         /* push rax */
320 	*pprog = prog;
321 }
322 
323 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
324 {
325 	u8 *prog = *pprog;
326 	s64 offset;
327 
328 	offset = func - (ip + X86_PATCH_SIZE);
329 	if (!is_simm32(offset)) {
330 		pr_err("Target call %p is out of range\n", func);
331 		return -ERANGE;
332 	}
333 	EMIT1_off32(opcode, offset);
334 	*pprog = prog;
335 	return 0;
336 }
337 
338 static int emit_call(u8 **pprog, void *func, void *ip)
339 {
340 	return emit_patch(pprog, func, ip, 0xE8);
341 }
342 
343 static int emit_jump(u8 **pprog, void *func, void *ip)
344 {
345 	return emit_patch(pprog, func, ip, 0xE9);
346 }
347 
348 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
349 				void *old_addr, void *new_addr)
350 {
351 	const u8 *nop_insn = x86_nops[5];
352 	u8 old_insn[X86_PATCH_SIZE];
353 	u8 new_insn[X86_PATCH_SIZE];
354 	u8 *prog;
355 	int ret;
356 
357 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
358 	if (old_addr) {
359 		prog = old_insn;
360 		ret = t == BPF_MOD_CALL ?
361 		      emit_call(&prog, old_addr, ip) :
362 		      emit_jump(&prog, old_addr, ip);
363 		if (ret)
364 			return ret;
365 	}
366 
367 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
368 	if (new_addr) {
369 		prog = new_insn;
370 		ret = t == BPF_MOD_CALL ?
371 		      emit_call(&prog, new_addr, ip) :
372 		      emit_jump(&prog, new_addr, ip);
373 		if (ret)
374 			return ret;
375 	}
376 
377 	ret = -EBUSY;
378 	mutex_lock(&text_mutex);
379 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
380 		goto out;
381 	ret = 1;
382 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
383 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
384 		ret = 0;
385 	}
386 out:
387 	mutex_unlock(&text_mutex);
388 	return ret;
389 }
390 
391 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
392 		       void *old_addr, void *new_addr)
393 {
394 	if (!is_kernel_text((long)ip) &&
395 	    !is_bpf_text_address((long)ip))
396 		/* BPF poking in modules is not supported */
397 		return -EINVAL;
398 
399 	/*
400 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
401 	 * with an ENDBR instruction.
402 	 */
403 	if (is_endbr(*(u32 *)ip))
404 		ip += ENDBR_INSN_SIZE;
405 
406 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
407 }
408 
409 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
410 
411 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
412 {
413 	u8 *prog = *pprog;
414 
415 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
416 		EMIT_LFENCE();
417 		EMIT2(0xFF, 0xE0 + reg);
418 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
419 		OPTIMIZER_HIDE_VAR(reg);
420 		emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
421 	} else {
422 		EMIT2(0xFF, 0xE0 + reg);
423 	}
424 
425 	*pprog = prog;
426 }
427 
428 static void emit_return(u8 **pprog, u8 *ip)
429 {
430 	u8 *prog = *pprog;
431 
432 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
433 		emit_jump(&prog, &__x86_return_thunk, ip);
434 	} else {
435 		EMIT1(0xC3);		/* ret */
436 		if (IS_ENABLED(CONFIG_SLS))
437 			EMIT1(0xCC);	/* int3 */
438 	}
439 
440 	*pprog = prog;
441 }
442 
443 /*
444  * Generate the following code:
445  *
446  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
447  *   if (index >= array->map.max_entries)
448  *     goto out;
449  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
450  *     goto out;
451  *   prog = array->ptrs[index];
452  *   if (prog == NULL)
453  *     goto out;
454  *   goto *(prog->bpf_func + prologue_size);
455  * out:
456  */
457 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
458 					u32 stack_depth, u8 *ip,
459 					struct jit_context *ctx)
460 {
461 	int tcc_off = -4 - round_up(stack_depth, 8);
462 	u8 *prog = *pprog, *start = *pprog;
463 	int offset;
464 
465 	/*
466 	 * rdi - pointer to ctx
467 	 * rsi - pointer to bpf_array
468 	 * rdx - index in bpf_array
469 	 */
470 
471 	/*
472 	 * if (index >= array->map.max_entries)
473 	 *	goto out;
474 	 */
475 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
476 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
477 	      offsetof(struct bpf_array, map.max_entries));
478 
479 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
480 	EMIT2(X86_JBE, offset);                   /* jbe out */
481 
482 	/*
483 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
484 	 *	goto out;
485 	 */
486 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
487 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
488 
489 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
490 	EMIT2(X86_JAE, offset);                   /* jae out */
491 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
492 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
493 
494 	/* prog = array->ptrs[index]; */
495 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
496 		    offsetof(struct bpf_array, ptrs));
497 
498 	/*
499 	 * if (prog == NULL)
500 	 *	goto out;
501 	 */
502 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
503 
504 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
505 	EMIT2(X86_JE, offset);                    /* je out */
506 
507 	pop_callee_regs(&prog, callee_regs_used);
508 
509 	EMIT1(0x58);                              /* pop rax */
510 	if (stack_depth)
511 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
512 			    round_up(stack_depth, 8));
513 
514 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
515 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
516 	      offsetof(struct bpf_prog, bpf_func));
517 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
518 	      X86_TAIL_CALL_OFFSET);
519 	/*
520 	 * Now we're ready to jump into next BPF program
521 	 * rdi == ctx (1st arg)
522 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
523 	 */
524 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
525 
526 	/* out: */
527 	ctx->tail_call_indirect_label = prog - start;
528 	*pprog = prog;
529 }
530 
531 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
532 				      u8 **pprog, u8 *ip,
533 				      bool *callee_regs_used, u32 stack_depth,
534 				      struct jit_context *ctx)
535 {
536 	int tcc_off = -4 - round_up(stack_depth, 8);
537 	u8 *prog = *pprog, *start = *pprog;
538 	int offset;
539 
540 	/*
541 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
542 	 *	goto out;
543 	 */
544 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
545 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
546 
547 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
548 	EMIT2(X86_JAE, offset);                       /* jae out */
549 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
550 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
551 
552 	poke->tailcall_bypass = ip + (prog - start);
553 	poke->adj_off = X86_TAIL_CALL_OFFSET;
554 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
555 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
556 
557 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
558 		  poke->tailcall_bypass);
559 
560 	pop_callee_regs(&prog, callee_regs_used);
561 	EMIT1(0x58);                                  /* pop rax */
562 	if (stack_depth)
563 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
564 
565 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
566 	prog += X86_PATCH_SIZE;
567 
568 	/* out: */
569 	ctx->tail_call_direct_label = prog - start;
570 
571 	*pprog = prog;
572 }
573 
574 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
575 {
576 	struct bpf_jit_poke_descriptor *poke;
577 	struct bpf_array *array;
578 	struct bpf_prog *target;
579 	int i, ret;
580 
581 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
582 		poke = &prog->aux->poke_tab[i];
583 		if (poke->aux && poke->aux != prog->aux)
584 			continue;
585 
586 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
587 
588 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
589 			continue;
590 
591 		array = container_of(poke->tail_call.map, struct bpf_array, map);
592 		mutex_lock(&array->aux->poke_mutex);
593 		target = array->ptrs[poke->tail_call.key];
594 		if (target) {
595 			ret = __bpf_arch_text_poke(poke->tailcall_target,
596 						   BPF_MOD_JUMP, NULL,
597 						   (u8 *)target->bpf_func +
598 						   poke->adj_off);
599 			BUG_ON(ret < 0);
600 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
601 						   BPF_MOD_JUMP,
602 						   (u8 *)poke->tailcall_target +
603 						   X86_PATCH_SIZE, NULL);
604 			BUG_ON(ret < 0);
605 		}
606 		WRITE_ONCE(poke->tailcall_target_stable, true);
607 		mutex_unlock(&array->aux->poke_mutex);
608 	}
609 }
610 
611 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
612 			   u32 dst_reg, const u32 imm32)
613 {
614 	u8 *prog = *pprog;
615 	u8 b1, b2, b3;
616 
617 	/*
618 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
619 	 * (which zero-extends imm32) to save 2 bytes.
620 	 */
621 	if (sign_propagate && (s32)imm32 < 0) {
622 		/* 'mov %rax, imm32' sign extends imm32 */
623 		b1 = add_1mod(0x48, dst_reg);
624 		b2 = 0xC7;
625 		b3 = 0xC0;
626 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
627 		goto done;
628 	}
629 
630 	/*
631 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
632 	 * to save 3 bytes.
633 	 */
634 	if (imm32 == 0) {
635 		if (is_ereg(dst_reg))
636 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
637 		b2 = 0x31; /* xor */
638 		b3 = 0xC0;
639 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
640 		goto done;
641 	}
642 
643 	/* mov %eax, imm32 */
644 	if (is_ereg(dst_reg))
645 		EMIT1(add_1mod(0x40, dst_reg));
646 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
647 done:
648 	*pprog = prog;
649 }
650 
651 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
652 			   const u32 imm32_hi, const u32 imm32_lo)
653 {
654 	u8 *prog = *pprog;
655 
656 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
657 		/*
658 		 * For emitting plain u32, where sign bit must not be
659 		 * propagated LLVM tends to load imm64 over mov32
660 		 * directly, so save couple of bytes by just doing
661 		 * 'mov %eax, imm32' instead.
662 		 */
663 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
664 	} else {
665 		/* movabsq %rax, imm64 */
666 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
667 		EMIT(imm32_lo, 4);
668 		EMIT(imm32_hi, 4);
669 	}
670 
671 	*pprog = prog;
672 }
673 
674 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
675 {
676 	u8 *prog = *pprog;
677 
678 	if (is64) {
679 		/* mov dst, src */
680 		EMIT_mov(dst_reg, src_reg);
681 	} else {
682 		/* mov32 dst, src */
683 		if (is_ereg(dst_reg) || is_ereg(src_reg))
684 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
685 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
686 	}
687 
688 	*pprog = prog;
689 }
690 
691 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
692 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
693 {
694 	u8 *prog = *pprog;
695 
696 	if (is_imm8(off)) {
697 		/* 1-byte signed displacement.
698 		 *
699 		 * If off == 0 we could skip this and save one extra byte, but
700 		 * special case of x86 R13 which always needs an offset is not
701 		 * worth the hassle
702 		 */
703 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
704 	} else {
705 		/* 4-byte signed displacement */
706 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
707 	}
708 	*pprog = prog;
709 }
710 
711 /*
712  * Emit a REX byte if it will be necessary to address these registers
713  */
714 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
715 {
716 	u8 *prog = *pprog;
717 
718 	if (is64)
719 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
720 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
721 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
722 	*pprog = prog;
723 }
724 
725 /*
726  * Similar version of maybe_emit_mod() for a single register
727  */
728 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
729 {
730 	u8 *prog = *pprog;
731 
732 	if (is64)
733 		EMIT1(add_1mod(0x48, reg));
734 	else if (is_ereg(reg))
735 		EMIT1(add_1mod(0x40, reg));
736 	*pprog = prog;
737 }
738 
739 /* LDX: dst_reg = *(u8*)(src_reg + off) */
740 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
741 {
742 	u8 *prog = *pprog;
743 
744 	switch (size) {
745 	case BPF_B:
746 		/* Emit 'movzx rax, byte ptr [rax + off]' */
747 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
748 		break;
749 	case BPF_H:
750 		/* Emit 'movzx rax, word ptr [rax + off]' */
751 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
752 		break;
753 	case BPF_W:
754 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
755 		if (is_ereg(dst_reg) || is_ereg(src_reg))
756 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
757 		else
758 			EMIT1(0x8B);
759 		break;
760 	case BPF_DW:
761 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
762 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
763 		break;
764 	}
765 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
766 	*pprog = prog;
767 }
768 
769 /* STX: *(u8*)(dst_reg + off) = src_reg */
770 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
771 {
772 	u8 *prog = *pprog;
773 
774 	switch (size) {
775 	case BPF_B:
776 		/* Emit 'mov byte ptr [rax + off], al' */
777 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
778 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
779 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
780 		else
781 			EMIT1(0x88);
782 		break;
783 	case BPF_H:
784 		if (is_ereg(dst_reg) || is_ereg(src_reg))
785 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
786 		else
787 			EMIT2(0x66, 0x89);
788 		break;
789 	case BPF_W:
790 		if (is_ereg(dst_reg) || is_ereg(src_reg))
791 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
792 		else
793 			EMIT1(0x89);
794 		break;
795 	case BPF_DW:
796 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
797 		break;
798 	}
799 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
800 	*pprog = prog;
801 }
802 
803 static int emit_atomic(u8 **pprog, u8 atomic_op,
804 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
805 {
806 	u8 *prog = *pprog;
807 
808 	EMIT1(0xF0); /* lock prefix */
809 
810 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
811 
812 	/* emit opcode */
813 	switch (atomic_op) {
814 	case BPF_ADD:
815 	case BPF_AND:
816 	case BPF_OR:
817 	case BPF_XOR:
818 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
819 		EMIT1(simple_alu_opcodes[atomic_op]);
820 		break;
821 	case BPF_ADD | BPF_FETCH:
822 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
823 		EMIT2(0x0F, 0xC1);
824 		break;
825 	case BPF_XCHG:
826 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
827 		EMIT1(0x87);
828 		break;
829 	case BPF_CMPXCHG:
830 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
831 		EMIT2(0x0F, 0xB1);
832 		break;
833 	default:
834 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
835 		return -EFAULT;
836 	}
837 
838 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
839 
840 	*pprog = prog;
841 	return 0;
842 }
843 
844 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
845 {
846 	u32 reg = x->fixup >> 8;
847 
848 	/* jump over faulting load and clear dest register */
849 	*(unsigned long *)((void *)regs + reg) = 0;
850 	regs->ip += x->fixup & 0xff;
851 	return true;
852 }
853 
854 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
855 			     bool *regs_used, bool *tail_call_seen)
856 {
857 	int i;
858 
859 	for (i = 1; i <= insn_cnt; i++, insn++) {
860 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
861 			*tail_call_seen = true;
862 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
863 			regs_used[0] = true;
864 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
865 			regs_used[1] = true;
866 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
867 			regs_used[2] = true;
868 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
869 			regs_used[3] = true;
870 	}
871 }
872 
873 static void emit_nops(u8 **pprog, int len)
874 {
875 	u8 *prog = *pprog;
876 	int i, noplen;
877 
878 	while (len > 0) {
879 		noplen = len;
880 
881 		if (noplen > ASM_NOP_MAX)
882 			noplen = ASM_NOP_MAX;
883 
884 		for (i = 0; i < noplen; i++)
885 			EMIT1(x86_nops[noplen][i]);
886 		len -= noplen;
887 	}
888 
889 	*pprog = prog;
890 }
891 
892 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
893 
894 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
895 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
896 {
897 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
898 	struct bpf_insn *insn = bpf_prog->insnsi;
899 	bool callee_regs_used[4] = {};
900 	int insn_cnt = bpf_prog->len;
901 	bool tail_call_seen = false;
902 	bool seen_exit = false;
903 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
904 	int i, excnt = 0;
905 	int ilen, proglen = 0;
906 	u8 *prog = temp;
907 	int err;
908 
909 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
910 			 &tail_call_seen);
911 
912 	/* tail call's presence in current prog implies it is reachable */
913 	tail_call_reachable |= tail_call_seen;
914 
915 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
916 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
917 		      bpf_prog->aux->func_idx != 0);
918 	push_callee_regs(&prog, callee_regs_used);
919 
920 	ilen = prog - temp;
921 	if (rw_image)
922 		memcpy(rw_image + proglen, temp, ilen);
923 	proglen += ilen;
924 	addrs[0] = proglen;
925 	prog = temp;
926 
927 	for (i = 1; i <= insn_cnt; i++, insn++) {
928 		const s32 imm32 = insn->imm;
929 		u32 dst_reg = insn->dst_reg;
930 		u32 src_reg = insn->src_reg;
931 		u8 b2 = 0, b3 = 0;
932 		u8 *start_of_ldx;
933 		s64 jmp_offset;
934 		u8 jmp_cond;
935 		u8 *func;
936 		int nops;
937 
938 		switch (insn->code) {
939 			/* ALU */
940 		case BPF_ALU | BPF_ADD | BPF_X:
941 		case BPF_ALU | BPF_SUB | BPF_X:
942 		case BPF_ALU | BPF_AND | BPF_X:
943 		case BPF_ALU | BPF_OR | BPF_X:
944 		case BPF_ALU | BPF_XOR | BPF_X:
945 		case BPF_ALU64 | BPF_ADD | BPF_X:
946 		case BPF_ALU64 | BPF_SUB | BPF_X:
947 		case BPF_ALU64 | BPF_AND | BPF_X:
948 		case BPF_ALU64 | BPF_OR | BPF_X:
949 		case BPF_ALU64 | BPF_XOR | BPF_X:
950 			maybe_emit_mod(&prog, dst_reg, src_reg,
951 				       BPF_CLASS(insn->code) == BPF_ALU64);
952 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
953 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
954 			break;
955 
956 		case BPF_ALU64 | BPF_MOV | BPF_X:
957 		case BPF_ALU | BPF_MOV | BPF_X:
958 			emit_mov_reg(&prog,
959 				     BPF_CLASS(insn->code) == BPF_ALU64,
960 				     dst_reg, src_reg);
961 			break;
962 
963 			/* neg dst */
964 		case BPF_ALU | BPF_NEG:
965 		case BPF_ALU64 | BPF_NEG:
966 			maybe_emit_1mod(&prog, dst_reg,
967 					BPF_CLASS(insn->code) == BPF_ALU64);
968 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
969 			break;
970 
971 		case BPF_ALU | BPF_ADD | BPF_K:
972 		case BPF_ALU | BPF_SUB | BPF_K:
973 		case BPF_ALU | BPF_AND | BPF_K:
974 		case BPF_ALU | BPF_OR | BPF_K:
975 		case BPF_ALU | BPF_XOR | BPF_K:
976 		case BPF_ALU64 | BPF_ADD | BPF_K:
977 		case BPF_ALU64 | BPF_SUB | BPF_K:
978 		case BPF_ALU64 | BPF_AND | BPF_K:
979 		case BPF_ALU64 | BPF_OR | BPF_K:
980 		case BPF_ALU64 | BPF_XOR | BPF_K:
981 			maybe_emit_1mod(&prog, dst_reg,
982 					BPF_CLASS(insn->code) == BPF_ALU64);
983 
984 			/*
985 			 * b3 holds 'normal' opcode, b2 short form only valid
986 			 * in case dst is eax/rax.
987 			 */
988 			switch (BPF_OP(insn->code)) {
989 			case BPF_ADD:
990 				b3 = 0xC0;
991 				b2 = 0x05;
992 				break;
993 			case BPF_SUB:
994 				b3 = 0xE8;
995 				b2 = 0x2D;
996 				break;
997 			case BPF_AND:
998 				b3 = 0xE0;
999 				b2 = 0x25;
1000 				break;
1001 			case BPF_OR:
1002 				b3 = 0xC8;
1003 				b2 = 0x0D;
1004 				break;
1005 			case BPF_XOR:
1006 				b3 = 0xF0;
1007 				b2 = 0x35;
1008 				break;
1009 			}
1010 
1011 			if (is_imm8(imm32))
1012 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1013 			else if (is_axreg(dst_reg))
1014 				EMIT1_off32(b2, imm32);
1015 			else
1016 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1017 			break;
1018 
1019 		case BPF_ALU64 | BPF_MOV | BPF_K:
1020 		case BPF_ALU | BPF_MOV | BPF_K:
1021 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1022 				       dst_reg, imm32);
1023 			break;
1024 
1025 		case BPF_LD | BPF_IMM | BPF_DW:
1026 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1027 			insn++;
1028 			i++;
1029 			break;
1030 
1031 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1032 		case BPF_ALU | BPF_MOD | BPF_X:
1033 		case BPF_ALU | BPF_DIV | BPF_X:
1034 		case BPF_ALU | BPF_MOD | BPF_K:
1035 		case BPF_ALU | BPF_DIV | BPF_K:
1036 		case BPF_ALU64 | BPF_MOD | BPF_X:
1037 		case BPF_ALU64 | BPF_DIV | BPF_X:
1038 		case BPF_ALU64 | BPF_MOD | BPF_K:
1039 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1040 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1041 
1042 			if (dst_reg != BPF_REG_0)
1043 				EMIT1(0x50); /* push rax */
1044 			if (dst_reg != BPF_REG_3)
1045 				EMIT1(0x52); /* push rdx */
1046 
1047 			if (BPF_SRC(insn->code) == BPF_X) {
1048 				if (src_reg == BPF_REG_0 ||
1049 				    src_reg == BPF_REG_3) {
1050 					/* mov r11, src_reg */
1051 					EMIT_mov(AUX_REG, src_reg);
1052 					src_reg = AUX_REG;
1053 				}
1054 			} else {
1055 				/* mov r11, imm32 */
1056 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1057 				src_reg = AUX_REG;
1058 			}
1059 
1060 			if (dst_reg != BPF_REG_0)
1061 				/* mov rax, dst_reg */
1062 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1063 
1064 			/*
1065 			 * xor edx, edx
1066 			 * equivalent to 'xor rdx, rdx', but one byte less
1067 			 */
1068 			EMIT2(0x31, 0xd2);
1069 
1070 			/* div src_reg */
1071 			maybe_emit_1mod(&prog, src_reg, is64);
1072 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1073 
1074 			if (BPF_OP(insn->code) == BPF_MOD &&
1075 			    dst_reg != BPF_REG_3)
1076 				/* mov dst_reg, rdx */
1077 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1078 			else if (BPF_OP(insn->code) == BPF_DIV &&
1079 				 dst_reg != BPF_REG_0)
1080 				/* mov dst_reg, rax */
1081 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1082 
1083 			if (dst_reg != BPF_REG_3)
1084 				EMIT1(0x5A); /* pop rdx */
1085 			if (dst_reg != BPF_REG_0)
1086 				EMIT1(0x58); /* pop rax */
1087 			break;
1088 		}
1089 
1090 		case BPF_ALU | BPF_MUL | BPF_K:
1091 		case BPF_ALU64 | BPF_MUL | BPF_K:
1092 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1093 				       BPF_CLASS(insn->code) == BPF_ALU64);
1094 
1095 			if (is_imm8(imm32))
1096 				/* imul dst_reg, dst_reg, imm8 */
1097 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1098 				      imm32);
1099 			else
1100 				/* imul dst_reg, dst_reg, imm32 */
1101 				EMIT2_off32(0x69,
1102 					    add_2reg(0xC0, dst_reg, dst_reg),
1103 					    imm32);
1104 			break;
1105 
1106 		case BPF_ALU | BPF_MUL | BPF_X:
1107 		case BPF_ALU64 | BPF_MUL | BPF_X:
1108 			maybe_emit_mod(&prog, src_reg, dst_reg,
1109 				       BPF_CLASS(insn->code) == BPF_ALU64);
1110 
1111 			/* imul dst_reg, src_reg */
1112 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1113 			break;
1114 
1115 			/* Shifts */
1116 		case BPF_ALU | BPF_LSH | BPF_K:
1117 		case BPF_ALU | BPF_RSH | BPF_K:
1118 		case BPF_ALU | BPF_ARSH | BPF_K:
1119 		case BPF_ALU64 | BPF_LSH | BPF_K:
1120 		case BPF_ALU64 | BPF_RSH | BPF_K:
1121 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1122 			maybe_emit_1mod(&prog, dst_reg,
1123 					BPF_CLASS(insn->code) == BPF_ALU64);
1124 
1125 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1126 			if (imm32 == 1)
1127 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1128 			else
1129 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1130 			break;
1131 
1132 		case BPF_ALU | BPF_LSH | BPF_X:
1133 		case BPF_ALU | BPF_RSH | BPF_X:
1134 		case BPF_ALU | BPF_ARSH | BPF_X:
1135 		case BPF_ALU64 | BPF_LSH | BPF_X:
1136 		case BPF_ALU64 | BPF_RSH | BPF_X:
1137 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1138 
1139 			/* Check for bad case when dst_reg == rcx */
1140 			if (dst_reg == BPF_REG_4) {
1141 				/* mov r11, dst_reg */
1142 				EMIT_mov(AUX_REG, dst_reg);
1143 				dst_reg = AUX_REG;
1144 			}
1145 
1146 			if (src_reg != BPF_REG_4) { /* common case */
1147 				EMIT1(0x51); /* push rcx */
1148 
1149 				/* mov rcx, src_reg */
1150 				EMIT_mov(BPF_REG_4, src_reg);
1151 			}
1152 
1153 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1154 			maybe_emit_1mod(&prog, dst_reg,
1155 					BPF_CLASS(insn->code) == BPF_ALU64);
1156 
1157 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1158 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1159 
1160 			if (src_reg != BPF_REG_4)
1161 				EMIT1(0x59); /* pop rcx */
1162 
1163 			if (insn->dst_reg == BPF_REG_4)
1164 				/* mov dst_reg, r11 */
1165 				EMIT_mov(insn->dst_reg, AUX_REG);
1166 			break;
1167 
1168 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1169 			switch (imm32) {
1170 			case 16:
1171 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1172 				EMIT1(0x66);
1173 				if (is_ereg(dst_reg))
1174 					EMIT1(0x41);
1175 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1176 
1177 				/* Emit 'movzwl eax, ax' */
1178 				if (is_ereg(dst_reg))
1179 					EMIT3(0x45, 0x0F, 0xB7);
1180 				else
1181 					EMIT2(0x0F, 0xB7);
1182 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1183 				break;
1184 			case 32:
1185 				/* Emit 'bswap eax' to swap lower 4 bytes */
1186 				if (is_ereg(dst_reg))
1187 					EMIT2(0x41, 0x0F);
1188 				else
1189 					EMIT1(0x0F);
1190 				EMIT1(add_1reg(0xC8, dst_reg));
1191 				break;
1192 			case 64:
1193 				/* Emit 'bswap rax' to swap 8 bytes */
1194 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1195 				      add_1reg(0xC8, dst_reg));
1196 				break;
1197 			}
1198 			break;
1199 
1200 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1201 			switch (imm32) {
1202 			case 16:
1203 				/*
1204 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1205 				 * into 64 bit
1206 				 */
1207 				if (is_ereg(dst_reg))
1208 					EMIT3(0x45, 0x0F, 0xB7);
1209 				else
1210 					EMIT2(0x0F, 0xB7);
1211 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1212 				break;
1213 			case 32:
1214 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1215 				if (is_ereg(dst_reg))
1216 					EMIT1(0x45);
1217 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1218 				break;
1219 			case 64:
1220 				/* nop */
1221 				break;
1222 			}
1223 			break;
1224 
1225 			/* speculation barrier */
1226 		case BPF_ST | BPF_NOSPEC:
1227 			if (boot_cpu_has(X86_FEATURE_XMM2))
1228 				EMIT_LFENCE();
1229 			break;
1230 
1231 			/* ST: *(u8*)(dst_reg + off) = imm */
1232 		case BPF_ST | BPF_MEM | BPF_B:
1233 			if (is_ereg(dst_reg))
1234 				EMIT2(0x41, 0xC6);
1235 			else
1236 				EMIT1(0xC6);
1237 			goto st;
1238 		case BPF_ST | BPF_MEM | BPF_H:
1239 			if (is_ereg(dst_reg))
1240 				EMIT3(0x66, 0x41, 0xC7);
1241 			else
1242 				EMIT2(0x66, 0xC7);
1243 			goto st;
1244 		case BPF_ST | BPF_MEM | BPF_W:
1245 			if (is_ereg(dst_reg))
1246 				EMIT2(0x41, 0xC7);
1247 			else
1248 				EMIT1(0xC7);
1249 			goto st;
1250 		case BPF_ST | BPF_MEM | BPF_DW:
1251 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1252 
1253 st:			if (is_imm8(insn->off))
1254 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1255 			else
1256 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1257 
1258 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1259 			break;
1260 
1261 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1262 		case BPF_STX | BPF_MEM | BPF_B:
1263 		case BPF_STX | BPF_MEM | BPF_H:
1264 		case BPF_STX | BPF_MEM | BPF_W:
1265 		case BPF_STX | BPF_MEM | BPF_DW:
1266 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1267 			break;
1268 
1269 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1270 		case BPF_LDX | BPF_MEM | BPF_B:
1271 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1272 		case BPF_LDX | BPF_MEM | BPF_H:
1273 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1274 		case BPF_LDX | BPF_MEM | BPF_W:
1275 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1276 		case BPF_LDX | BPF_MEM | BPF_DW:
1277 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1278 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1279 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1280 				 * add abs(insn->off) to the limit to make sure that negative
1281 				 * offset won't be an issue.
1282 				 * insn->off is s16, so it won't affect valid pointers.
1283 				 */
1284 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1285 				u8 *end_of_jmp1, *end_of_jmp2;
1286 
1287 				/* Conservatively check that src_reg + insn->off is a kernel address:
1288 				 * 1. src_reg + insn->off >= limit
1289 				 * 2. src_reg + insn->off doesn't become small positive.
1290 				 * Cannot do src_reg + insn->off >= limit in one branch,
1291 				 * since it needs two spare registers, but JIT has only one.
1292 				 */
1293 
1294 				/* movabsq r11, limit */
1295 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1296 				EMIT((u32)limit, 4);
1297 				EMIT(limit >> 32, 4);
1298 				/* cmp src_reg, r11 */
1299 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1300 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1301 				/* if unsigned '<' goto end_of_jmp2 */
1302 				EMIT2(X86_JB, 0);
1303 				end_of_jmp1 = prog;
1304 
1305 				/* mov r11, src_reg */
1306 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1307 				/* add r11, insn->off */
1308 				maybe_emit_1mod(&prog, AUX_REG, true);
1309 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1310 				/* jmp if not carry to start_of_ldx
1311 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1312 				 * that has to be rejected.
1313 				 */
1314 				EMIT2(0x73 /* JNC */, 0);
1315 				end_of_jmp2 = prog;
1316 
1317 				/* xor dst_reg, dst_reg */
1318 				emit_mov_imm32(&prog, false, dst_reg, 0);
1319 				/* jmp byte_after_ldx */
1320 				EMIT2(0xEB, 0);
1321 
1322 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1323 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1324 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1325 				start_of_ldx = prog;
1326 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1327 			}
1328 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1329 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1330 				struct exception_table_entry *ex;
1331 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1332 				s64 delta;
1333 
1334 				/* populate jmp_offset for JMP above */
1335 				start_of_ldx[-1] = prog - start_of_ldx;
1336 
1337 				if (!bpf_prog->aux->extable)
1338 					break;
1339 
1340 				if (excnt >= bpf_prog->aux->num_exentries) {
1341 					pr_err("ex gen bug\n");
1342 					return -EFAULT;
1343 				}
1344 				ex = &bpf_prog->aux->extable[excnt++];
1345 
1346 				delta = _insn - (u8 *)&ex->insn;
1347 				if (!is_simm32(delta)) {
1348 					pr_err("extable->insn doesn't fit into 32-bit\n");
1349 					return -EFAULT;
1350 				}
1351 				/* switch ex to rw buffer for writes */
1352 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1353 
1354 				ex->insn = delta;
1355 
1356 				ex->data = EX_TYPE_BPF;
1357 
1358 				if (dst_reg > BPF_REG_9) {
1359 					pr_err("verifier error\n");
1360 					return -EFAULT;
1361 				}
1362 				/*
1363 				 * Compute size of x86 insn and its target dest x86 register.
1364 				 * ex_handler_bpf() will use lower 8 bits to adjust
1365 				 * pt_regs->ip to jump over this x86 instruction
1366 				 * and upper bits to figure out which pt_regs to zero out.
1367 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1368 				 * of 4 bytes will be ignored and rbx will be zero inited.
1369 				 */
1370 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1371 			}
1372 			break;
1373 
1374 		case BPF_STX | BPF_ATOMIC | BPF_W:
1375 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1376 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1377 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1378 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1379 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1380 				u32 real_src_reg = src_reg;
1381 				u32 real_dst_reg = dst_reg;
1382 				u8 *branch_target;
1383 
1384 				/*
1385 				 * Can't be implemented with a single x86 insn.
1386 				 * Need to do a CMPXCHG loop.
1387 				 */
1388 
1389 				/* Will need RAX as a CMPXCHG operand so save R0 */
1390 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1391 				if (src_reg == BPF_REG_0)
1392 					real_src_reg = BPF_REG_AX;
1393 				if (dst_reg == BPF_REG_0)
1394 					real_dst_reg = BPF_REG_AX;
1395 
1396 				branch_target = prog;
1397 				/* Load old value */
1398 				emit_ldx(&prog, BPF_SIZE(insn->code),
1399 					 BPF_REG_0, real_dst_reg, insn->off);
1400 				/*
1401 				 * Perform the (commutative) operation locally,
1402 				 * put the result in the AUX_REG.
1403 				 */
1404 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1405 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1406 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1407 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1408 				/* Attempt to swap in new value */
1409 				err = emit_atomic(&prog, BPF_CMPXCHG,
1410 						  real_dst_reg, AUX_REG,
1411 						  insn->off,
1412 						  BPF_SIZE(insn->code));
1413 				if (WARN_ON(err))
1414 					return err;
1415 				/*
1416 				 * ZF tells us whether we won the race. If it's
1417 				 * cleared we need to try again.
1418 				 */
1419 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1420 				/* Return the pre-modification value */
1421 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1422 				/* Restore R0 after clobbering RAX */
1423 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1424 				break;
1425 			}
1426 
1427 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1428 					  insn->off, BPF_SIZE(insn->code));
1429 			if (err)
1430 				return err;
1431 			break;
1432 
1433 			/* call */
1434 		case BPF_JMP | BPF_CALL:
1435 			func = (u8 *) __bpf_call_base + imm32;
1436 			if (tail_call_reachable) {
1437 				/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1438 				EMIT3_off32(0x48, 0x8B, 0x85,
1439 					    -round_up(bpf_prog->aux->stack_depth, 8) - 8);
1440 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1441 					return -EINVAL;
1442 			} else {
1443 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1444 					return -EINVAL;
1445 			}
1446 			break;
1447 
1448 		case BPF_JMP | BPF_TAIL_CALL:
1449 			if (imm32)
1450 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1451 							  &prog, image + addrs[i - 1],
1452 							  callee_regs_used,
1453 							  bpf_prog->aux->stack_depth,
1454 							  ctx);
1455 			else
1456 				emit_bpf_tail_call_indirect(&prog,
1457 							    callee_regs_used,
1458 							    bpf_prog->aux->stack_depth,
1459 							    image + addrs[i - 1],
1460 							    ctx);
1461 			break;
1462 
1463 			/* cond jump */
1464 		case BPF_JMP | BPF_JEQ | BPF_X:
1465 		case BPF_JMP | BPF_JNE | BPF_X:
1466 		case BPF_JMP | BPF_JGT | BPF_X:
1467 		case BPF_JMP | BPF_JLT | BPF_X:
1468 		case BPF_JMP | BPF_JGE | BPF_X:
1469 		case BPF_JMP | BPF_JLE | BPF_X:
1470 		case BPF_JMP | BPF_JSGT | BPF_X:
1471 		case BPF_JMP | BPF_JSLT | BPF_X:
1472 		case BPF_JMP | BPF_JSGE | BPF_X:
1473 		case BPF_JMP | BPF_JSLE | BPF_X:
1474 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1475 		case BPF_JMP32 | BPF_JNE | BPF_X:
1476 		case BPF_JMP32 | BPF_JGT | BPF_X:
1477 		case BPF_JMP32 | BPF_JLT | BPF_X:
1478 		case BPF_JMP32 | BPF_JGE | BPF_X:
1479 		case BPF_JMP32 | BPF_JLE | BPF_X:
1480 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1481 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1482 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1483 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1484 			/* cmp dst_reg, src_reg */
1485 			maybe_emit_mod(&prog, dst_reg, src_reg,
1486 				       BPF_CLASS(insn->code) == BPF_JMP);
1487 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1488 			goto emit_cond_jmp;
1489 
1490 		case BPF_JMP | BPF_JSET | BPF_X:
1491 		case BPF_JMP32 | BPF_JSET | BPF_X:
1492 			/* test dst_reg, src_reg */
1493 			maybe_emit_mod(&prog, dst_reg, src_reg,
1494 				       BPF_CLASS(insn->code) == BPF_JMP);
1495 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1496 			goto emit_cond_jmp;
1497 
1498 		case BPF_JMP | BPF_JSET | BPF_K:
1499 		case BPF_JMP32 | BPF_JSET | BPF_K:
1500 			/* test dst_reg, imm32 */
1501 			maybe_emit_1mod(&prog, dst_reg,
1502 					BPF_CLASS(insn->code) == BPF_JMP);
1503 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1504 			goto emit_cond_jmp;
1505 
1506 		case BPF_JMP | BPF_JEQ | BPF_K:
1507 		case BPF_JMP | BPF_JNE | BPF_K:
1508 		case BPF_JMP | BPF_JGT | BPF_K:
1509 		case BPF_JMP | BPF_JLT | BPF_K:
1510 		case BPF_JMP | BPF_JGE | BPF_K:
1511 		case BPF_JMP | BPF_JLE | BPF_K:
1512 		case BPF_JMP | BPF_JSGT | BPF_K:
1513 		case BPF_JMP | BPF_JSLT | BPF_K:
1514 		case BPF_JMP | BPF_JSGE | BPF_K:
1515 		case BPF_JMP | BPF_JSLE | BPF_K:
1516 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1517 		case BPF_JMP32 | BPF_JNE | BPF_K:
1518 		case BPF_JMP32 | BPF_JGT | BPF_K:
1519 		case BPF_JMP32 | BPF_JLT | BPF_K:
1520 		case BPF_JMP32 | BPF_JGE | BPF_K:
1521 		case BPF_JMP32 | BPF_JLE | BPF_K:
1522 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1523 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1524 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1525 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1526 			/* test dst_reg, dst_reg to save one extra byte */
1527 			if (imm32 == 0) {
1528 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1529 					       BPF_CLASS(insn->code) == BPF_JMP);
1530 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1531 				goto emit_cond_jmp;
1532 			}
1533 
1534 			/* cmp dst_reg, imm8/32 */
1535 			maybe_emit_1mod(&prog, dst_reg,
1536 					BPF_CLASS(insn->code) == BPF_JMP);
1537 
1538 			if (is_imm8(imm32))
1539 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1540 			else
1541 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1542 
1543 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1544 			switch (BPF_OP(insn->code)) {
1545 			case BPF_JEQ:
1546 				jmp_cond = X86_JE;
1547 				break;
1548 			case BPF_JSET:
1549 			case BPF_JNE:
1550 				jmp_cond = X86_JNE;
1551 				break;
1552 			case BPF_JGT:
1553 				/* GT is unsigned '>', JA in x86 */
1554 				jmp_cond = X86_JA;
1555 				break;
1556 			case BPF_JLT:
1557 				/* LT is unsigned '<', JB in x86 */
1558 				jmp_cond = X86_JB;
1559 				break;
1560 			case BPF_JGE:
1561 				/* GE is unsigned '>=', JAE in x86 */
1562 				jmp_cond = X86_JAE;
1563 				break;
1564 			case BPF_JLE:
1565 				/* LE is unsigned '<=', JBE in x86 */
1566 				jmp_cond = X86_JBE;
1567 				break;
1568 			case BPF_JSGT:
1569 				/* Signed '>', GT in x86 */
1570 				jmp_cond = X86_JG;
1571 				break;
1572 			case BPF_JSLT:
1573 				/* Signed '<', LT in x86 */
1574 				jmp_cond = X86_JL;
1575 				break;
1576 			case BPF_JSGE:
1577 				/* Signed '>=', GE in x86 */
1578 				jmp_cond = X86_JGE;
1579 				break;
1580 			case BPF_JSLE:
1581 				/* Signed '<=', LE in x86 */
1582 				jmp_cond = X86_JLE;
1583 				break;
1584 			default: /* to silence GCC warning */
1585 				return -EFAULT;
1586 			}
1587 			jmp_offset = addrs[i + insn->off] - addrs[i];
1588 			if (is_imm8(jmp_offset)) {
1589 				if (jmp_padding) {
1590 					/* To keep the jmp_offset valid, the extra bytes are
1591 					 * padded before the jump insn, so we subtract the
1592 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1593 					 *
1594 					 * If the previous pass already emits an imm8
1595 					 * jmp_cond, then this BPF insn won't shrink, so
1596 					 * "nops" is 0.
1597 					 *
1598 					 * On the other hand, if the previous pass emits an
1599 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1600 					 * keep the image from shrinking further.
1601 					 *
1602 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1603 					 *     is 2 bytes, so the size difference is 4 bytes.
1604 					 */
1605 					nops = INSN_SZ_DIFF - 2;
1606 					if (nops != 0 && nops != 4) {
1607 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1608 						       nops);
1609 						return -EFAULT;
1610 					}
1611 					emit_nops(&prog, nops);
1612 				}
1613 				EMIT2(jmp_cond, jmp_offset);
1614 			} else if (is_simm32(jmp_offset)) {
1615 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1616 			} else {
1617 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1618 				return -EFAULT;
1619 			}
1620 
1621 			break;
1622 
1623 		case BPF_JMP | BPF_JA:
1624 			if (insn->off == -1)
1625 				/* -1 jmp instructions will always jump
1626 				 * backwards two bytes. Explicitly handling
1627 				 * this case avoids wasting too many passes
1628 				 * when there are long sequences of replaced
1629 				 * dead code.
1630 				 */
1631 				jmp_offset = -2;
1632 			else
1633 				jmp_offset = addrs[i + insn->off] - addrs[i];
1634 
1635 			if (!jmp_offset) {
1636 				/*
1637 				 * If jmp_padding is enabled, the extra nops will
1638 				 * be inserted. Otherwise, optimize out nop jumps.
1639 				 */
1640 				if (jmp_padding) {
1641 					/* There are 3 possible conditions.
1642 					 * (1) This BPF_JA is already optimized out in
1643 					 *     the previous run, so there is no need
1644 					 *     to pad any extra byte (0 byte).
1645 					 * (2) The previous pass emits an imm8 jmp,
1646 					 *     so we pad 2 bytes to match the previous
1647 					 *     insn size.
1648 					 * (3) Similarly, the previous pass emits an
1649 					 *     imm32 jmp, and 5 bytes is padded.
1650 					 */
1651 					nops = INSN_SZ_DIFF;
1652 					if (nops != 0 && nops != 2 && nops != 5) {
1653 						pr_err("unexpected nop jump padding: %d bytes\n",
1654 						       nops);
1655 						return -EFAULT;
1656 					}
1657 					emit_nops(&prog, nops);
1658 				}
1659 				break;
1660 			}
1661 emit_jmp:
1662 			if (is_imm8(jmp_offset)) {
1663 				if (jmp_padding) {
1664 					/* To avoid breaking jmp_offset, the extra bytes
1665 					 * are padded before the actual jmp insn, so
1666 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1667 					 *
1668 					 * If the previous pass already emits an imm8
1669 					 * jmp, there is nothing to pad (0 byte).
1670 					 *
1671 					 * If it emits an imm32 jmp (5 bytes) previously
1672 					 * and now an imm8 jmp (2 bytes), then we pad
1673 					 * (5 - 2 = 3) bytes to stop the image from
1674 					 * shrinking further.
1675 					 */
1676 					nops = INSN_SZ_DIFF - 2;
1677 					if (nops != 0 && nops != 3) {
1678 						pr_err("unexpected jump padding: %d bytes\n",
1679 						       nops);
1680 						return -EFAULT;
1681 					}
1682 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1683 				}
1684 				EMIT2(0xEB, jmp_offset);
1685 			} else if (is_simm32(jmp_offset)) {
1686 				EMIT1_off32(0xE9, jmp_offset);
1687 			} else {
1688 				pr_err("jmp gen bug %llx\n", jmp_offset);
1689 				return -EFAULT;
1690 			}
1691 			break;
1692 
1693 		case BPF_JMP | BPF_EXIT:
1694 			if (seen_exit) {
1695 				jmp_offset = ctx->cleanup_addr - addrs[i];
1696 				goto emit_jmp;
1697 			}
1698 			seen_exit = true;
1699 			/* Update cleanup_addr */
1700 			ctx->cleanup_addr = proglen;
1701 			pop_callee_regs(&prog, callee_regs_used);
1702 			EMIT1(0xC9);         /* leave */
1703 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1704 			break;
1705 
1706 		default:
1707 			/*
1708 			 * By design x86-64 JIT should support all BPF instructions.
1709 			 * This error will be seen if new instruction was added
1710 			 * to the interpreter, but not to the JIT, or if there is
1711 			 * junk in bpf_prog.
1712 			 */
1713 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1714 			return -EINVAL;
1715 		}
1716 
1717 		ilen = prog - temp;
1718 		if (ilen > BPF_MAX_INSN_SIZE) {
1719 			pr_err("bpf_jit: fatal insn size error\n");
1720 			return -EFAULT;
1721 		}
1722 
1723 		if (image) {
1724 			/*
1725 			 * When populating the image, assert that:
1726 			 *
1727 			 *  i) We do not write beyond the allocated space, and
1728 			 * ii) addrs[i] did not change from the prior run, in order
1729 			 *     to validate assumptions made for computing branch
1730 			 *     displacements.
1731 			 */
1732 			if (unlikely(proglen + ilen > oldproglen ||
1733 				     proglen + ilen != addrs[i])) {
1734 				pr_err("bpf_jit: fatal error\n");
1735 				return -EFAULT;
1736 			}
1737 			memcpy(rw_image + proglen, temp, ilen);
1738 		}
1739 		proglen += ilen;
1740 		addrs[i] = proglen;
1741 		prog = temp;
1742 	}
1743 
1744 	if (image && excnt != bpf_prog->aux->num_exentries) {
1745 		pr_err("extable is not populated\n");
1746 		return -EFAULT;
1747 	}
1748 	return proglen;
1749 }
1750 
1751 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1752 		      int stack_size)
1753 {
1754 	int i;
1755 	/* Store function arguments to stack.
1756 	 * For a function that accepts two pointers the sequence will be:
1757 	 * mov QWORD PTR [rbp-0x10],rdi
1758 	 * mov QWORD PTR [rbp-0x8],rsi
1759 	 */
1760 	for (i = 0; i < min(nr_args, 6); i++)
1761 		emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1762 			 BPF_REG_FP,
1763 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1764 			 -(stack_size - i * 8));
1765 }
1766 
1767 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1768 			 int stack_size)
1769 {
1770 	int i;
1771 
1772 	/* Restore function arguments from stack.
1773 	 * For a function that accepts two pointers the sequence will be:
1774 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1775 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1776 	 */
1777 	for (i = 0; i < min(nr_args, 6); i++)
1778 		emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1779 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1780 			 BPF_REG_FP,
1781 			 -(stack_size - i * 8));
1782 }
1783 
1784 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1785 			   struct bpf_tramp_link *l, int stack_size,
1786 			   int run_ctx_off, bool save_ret)
1787 {
1788 	void (*exit)(struct bpf_prog *prog, u64 start,
1789 		     struct bpf_tramp_run_ctx *run_ctx) = __bpf_prog_exit;
1790 	u64 (*enter)(struct bpf_prog *prog,
1791 		     struct bpf_tramp_run_ctx *run_ctx) = __bpf_prog_enter;
1792 	u8 *prog = *pprog;
1793 	u8 *jmp_insn;
1794 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1795 	struct bpf_prog *p = l->link.prog;
1796 	u64 cookie = l->cookie;
1797 
1798 	/* mov rdi, cookie */
1799 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
1800 
1801 	/* Prepare struct bpf_tramp_run_ctx.
1802 	 *
1803 	 * bpf_tramp_run_ctx is already preserved by
1804 	 * arch_prepare_bpf_trampoline().
1805 	 *
1806 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
1807 	 */
1808 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
1809 
1810 	if (p->aux->sleepable) {
1811 		enter = __bpf_prog_enter_sleepable;
1812 		exit = __bpf_prog_exit_sleepable;
1813 	} else if (p->expected_attach_type == BPF_LSM_CGROUP) {
1814 		enter = __bpf_prog_enter_lsm_cgroup;
1815 		exit = __bpf_prog_exit_lsm_cgroup;
1816 	}
1817 
1818 	/* arg1: mov rdi, progs[i] */
1819 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1820 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
1821 	EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
1822 
1823 	if (emit_call(&prog, enter, prog))
1824 		return -EINVAL;
1825 	/* remember prog start time returned by __bpf_prog_enter */
1826 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1827 
1828 	/* if (__bpf_prog_enter*(prog) == 0)
1829 	 *	goto skip_exec_of_prog;
1830 	 */
1831 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1832 	/* emit 2 nops that will be replaced with JE insn */
1833 	jmp_insn = prog;
1834 	emit_nops(&prog, 2);
1835 
1836 	/* arg1: lea rdi, [rbp - stack_size] */
1837 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1838 	/* arg2: progs[i]->insnsi for interpreter */
1839 	if (!p->jited)
1840 		emit_mov_imm64(&prog, BPF_REG_2,
1841 			       (long) p->insnsi >> 32,
1842 			       (u32) (long) p->insnsi);
1843 	/* call JITed bpf program or interpreter */
1844 	if (emit_call(&prog, p->bpf_func, prog))
1845 		return -EINVAL;
1846 
1847 	/*
1848 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1849 	 * of the previous call which is then passed on the stack to
1850 	 * the next BPF program.
1851 	 *
1852 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1853 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1854 	 */
1855 	if (save_ret)
1856 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1857 
1858 	/* replace 2 nops with JE insn, since jmp target is known */
1859 	jmp_insn[0] = X86_JE;
1860 	jmp_insn[1] = prog - jmp_insn - 2;
1861 
1862 	/* arg1: mov rdi, progs[i] */
1863 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1864 	/* arg2: mov rsi, rbx <- start time in nsec */
1865 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1866 	/* arg3: lea rdx, [rbp - run_ctx_off] */
1867 	EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
1868 	if (emit_call(&prog, exit, prog))
1869 		return -EINVAL;
1870 
1871 	*pprog = prog;
1872 	return 0;
1873 }
1874 
1875 static void emit_align(u8 **pprog, u32 align)
1876 {
1877 	u8 *target, *prog = *pprog;
1878 
1879 	target = PTR_ALIGN(prog, align);
1880 	if (target != prog)
1881 		emit_nops(&prog, target - prog);
1882 
1883 	*pprog = prog;
1884 }
1885 
1886 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1887 {
1888 	u8 *prog = *pprog;
1889 	s64 offset;
1890 
1891 	offset = func - (ip + 2 + 4);
1892 	if (!is_simm32(offset)) {
1893 		pr_err("Target %p is out of range\n", func);
1894 		return -EINVAL;
1895 	}
1896 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1897 	*pprog = prog;
1898 	return 0;
1899 }
1900 
1901 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1902 		      struct bpf_tramp_links *tl, int stack_size,
1903 		      int run_ctx_off, bool save_ret)
1904 {
1905 	int i;
1906 	u8 *prog = *pprog;
1907 
1908 	for (i = 0; i < tl->nr_links; i++) {
1909 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
1910 				    run_ctx_off, save_ret))
1911 			return -EINVAL;
1912 	}
1913 	*pprog = prog;
1914 	return 0;
1915 }
1916 
1917 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1918 			      struct bpf_tramp_links *tl, int stack_size,
1919 			      int run_ctx_off, u8 **branches)
1920 {
1921 	u8 *prog = *pprog;
1922 	int i;
1923 
1924 	/* The first fmod_ret program will receive a garbage return value.
1925 	 * Set this to 0 to avoid confusing the program.
1926 	 */
1927 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1928 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1929 	for (i = 0; i < tl->nr_links; i++) {
1930 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
1931 			return -EINVAL;
1932 
1933 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
1934 		 * if (*(u64 *)(rbp - 8) !=  0)
1935 		 *	goto do_fexit;
1936 		 */
1937 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
1938 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1939 
1940 		/* Save the location of the branch and Generate 6 nops
1941 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
1942 		 * are replaced with a conditional jump once do_fexit (i.e. the
1943 		 * start of the fexit invocation) is finalized.
1944 		 */
1945 		branches[i] = prog;
1946 		emit_nops(&prog, 4 + 2);
1947 	}
1948 
1949 	*pprog = prog;
1950 	return 0;
1951 }
1952 
1953 /* Example:
1954  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1955  * its 'struct btf_func_model' will be nr_args=2
1956  * The assembly code when eth_type_trans is executing after trampoline:
1957  *
1958  * push rbp
1959  * mov rbp, rsp
1960  * sub rsp, 16                     // space for skb and dev
1961  * push rbx                        // temp regs to pass start time
1962  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1963  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1964  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1965  * mov rbx, rax                    // remember start time in bpf stats are enabled
1966  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1967  * call addr_of_jited_FENTRY_prog
1968  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1969  * mov rsi, rbx                    // prog start time
1970  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1971  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
1972  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
1973  * pop rbx
1974  * leave
1975  * ret
1976  *
1977  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1978  * replaced with 'call generated_bpf_trampoline'. When it returns
1979  * eth_type_trans will continue executing with original skb and dev pointers.
1980  *
1981  * The assembly code when eth_type_trans is called from trampoline:
1982  *
1983  * push rbp
1984  * mov rbp, rsp
1985  * sub rsp, 24                     // space for skb, dev, return value
1986  * push rbx                        // temp regs to pass start time
1987  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
1988  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
1989  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1990  * mov rbx, rax                    // remember start time if bpf stats are enabled
1991  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1992  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
1993  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1994  * mov rsi, rbx                    // prog start time
1995  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1996  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
1997  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
1998  * call eth_type_trans+5           // execute body of eth_type_trans
1999  * mov qword ptr [rbp - 8], rax    // save return value
2000  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2001  * mov rbx, rax                    // remember start time in bpf stats are enabled
2002  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2003  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2004  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2005  * mov rsi, rbx                    // prog start time
2006  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2007  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2008  * pop rbx
2009  * leave
2010  * add rsp, 8                      // skip eth_type_trans's frame
2011  * ret                             // return to its caller
2012  */
2013 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2014 				const struct btf_func_model *m, u32 flags,
2015 				struct bpf_tramp_links *tlinks,
2016 				void *orig_call)
2017 {
2018 	int ret, i, nr_args = m->nr_args;
2019 	int regs_off, ip_off, args_off, stack_size = nr_args * 8, run_ctx_off;
2020 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2021 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2022 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2023 	u8 **branches = NULL;
2024 	u8 *prog;
2025 	bool save_ret;
2026 
2027 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
2028 	if (nr_args > 6)
2029 		return -ENOTSUPP;
2030 
2031 	/* Generated trampoline stack layout:
2032 	 *
2033 	 * RBP + 8         [ return address  ]
2034 	 * RBP + 0         [ RBP             ]
2035 	 *
2036 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2037 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2038 	 *
2039 	 *                 [ reg_argN        ]  always
2040 	 *                 [ ...             ]
2041 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2042 	 *
2043 	 * RBP - args_off  [ args count      ]  always
2044 	 *
2045 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2046 	 *
2047 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2048 	 */
2049 
2050 	/* room for return value of orig_call or fentry prog */
2051 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2052 	if (save_ret)
2053 		stack_size += 8;
2054 
2055 	regs_off = stack_size;
2056 
2057 	/* args count  */
2058 	stack_size += 8;
2059 	args_off = stack_size;
2060 
2061 	if (flags & BPF_TRAMP_F_IP_ARG)
2062 		stack_size += 8; /* room for IP address argument */
2063 
2064 	ip_off = stack_size;
2065 
2066 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2067 	run_ctx_off = stack_size;
2068 
2069 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2070 		/* skip patched call instruction and point orig_call to actual
2071 		 * body of the kernel function.
2072 		 */
2073 		if (is_endbr(*(u32 *)orig_call))
2074 			orig_call += ENDBR_INSN_SIZE;
2075 		orig_call += X86_PATCH_SIZE;
2076 	}
2077 
2078 	prog = image;
2079 
2080 	EMIT_ENDBR();
2081 	EMIT1(0x55);		 /* push rbp */
2082 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2083 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2084 	EMIT1(0x53);		 /* push rbx */
2085 
2086 	/* Store number of arguments of the traced function:
2087 	 *   mov rax, nr_args
2088 	 *   mov QWORD PTR [rbp - args_off], rax
2089 	 */
2090 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args);
2091 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2092 
2093 	if (flags & BPF_TRAMP_F_IP_ARG) {
2094 		/* Store IP address of the traced function:
2095 		 * mov rax, QWORD PTR [rbp + 8]
2096 		 * sub rax, X86_PATCH_SIZE
2097 		 * mov QWORD PTR [rbp - ip_off], rax
2098 		 */
2099 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2100 		EMIT4(0x48, 0x83, 0xe8, X86_PATCH_SIZE);
2101 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2102 	}
2103 
2104 	save_regs(m, &prog, nr_args, regs_off);
2105 
2106 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2107 		/* arg1: mov rdi, im */
2108 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2109 		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2110 			ret = -EINVAL;
2111 			goto cleanup;
2112 		}
2113 	}
2114 
2115 	if (fentry->nr_links)
2116 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2117 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2118 			return -EINVAL;
2119 
2120 	if (fmod_ret->nr_links) {
2121 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2122 				   GFP_KERNEL);
2123 		if (!branches)
2124 			return -ENOMEM;
2125 
2126 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2127 				       run_ctx_off, branches)) {
2128 			ret = -EINVAL;
2129 			goto cleanup;
2130 		}
2131 	}
2132 
2133 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2134 		restore_regs(m, &prog, nr_args, regs_off);
2135 
2136 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2137 			emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2138 			EMIT2(0xff, 0xd0); /* call *rax */
2139 		} else {
2140 			/* call original function */
2141 			if (emit_call(&prog, orig_call, prog)) {
2142 				ret = -EINVAL;
2143 				goto cleanup;
2144 			}
2145 		}
2146 		/* remember return value in a stack for bpf prog to access */
2147 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2148 		im->ip_after_call = prog;
2149 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2150 		prog += X86_PATCH_SIZE;
2151 	}
2152 
2153 	if (fmod_ret->nr_links) {
2154 		/* From Intel 64 and IA-32 Architectures Optimization
2155 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2156 		 * Coding Rule 11: All branch targets should be 16-byte
2157 		 * aligned.
2158 		 */
2159 		emit_align(&prog, 16);
2160 		/* Update the branches saved in invoke_bpf_mod_ret with the
2161 		 * aligned address of do_fexit.
2162 		 */
2163 		for (i = 0; i < fmod_ret->nr_links; i++)
2164 			emit_cond_near_jump(&branches[i], prog, branches[i],
2165 					    X86_JNE);
2166 	}
2167 
2168 	if (fexit->nr_links)
2169 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2170 			ret = -EINVAL;
2171 			goto cleanup;
2172 		}
2173 
2174 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2175 		restore_regs(m, &prog, nr_args, regs_off);
2176 
2177 	/* This needs to be done regardless. If there were fmod_ret programs,
2178 	 * the return value is only updated on the stack and still needs to be
2179 	 * restored to R0.
2180 	 */
2181 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2182 		im->ip_epilogue = prog;
2183 		/* arg1: mov rdi, im */
2184 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2185 		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2186 			ret = -EINVAL;
2187 			goto cleanup;
2188 		}
2189 	}
2190 	/* restore return value of orig_call or fentry prog back into RAX */
2191 	if (save_ret)
2192 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2193 
2194 	EMIT1(0x5B); /* pop rbx */
2195 	EMIT1(0xC9); /* leave */
2196 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2197 		/* skip our return address and return to parent */
2198 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2199 	emit_return(&prog, prog);
2200 	/* Make sure the trampoline generation logic doesn't overflow */
2201 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2202 		ret = -EFAULT;
2203 		goto cleanup;
2204 	}
2205 	ret = prog - (u8 *)image;
2206 
2207 cleanup:
2208 	kfree(branches);
2209 	return ret;
2210 }
2211 
2212 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2213 {
2214 	u8 *jg_reloc, *prog = *pprog;
2215 	int pivot, err, jg_bytes = 1;
2216 	s64 jg_offset;
2217 
2218 	if (a == b) {
2219 		/* Leaf node of recursion, i.e. not a range of indices
2220 		 * anymore.
2221 		 */
2222 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2223 		if (!is_simm32(progs[a]))
2224 			return -1;
2225 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2226 			    progs[a]);
2227 		err = emit_cond_near_jump(&prog,	/* je func */
2228 					  (void *)progs[a], prog,
2229 					  X86_JE);
2230 		if (err)
2231 			return err;
2232 
2233 		emit_indirect_jump(&prog, 2 /* rdx */, prog);
2234 
2235 		*pprog = prog;
2236 		return 0;
2237 	}
2238 
2239 	/* Not a leaf node, so we pivot, and recursively descend into
2240 	 * the lower and upper ranges.
2241 	 */
2242 	pivot = (b - a) / 2;
2243 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2244 	if (!is_simm32(progs[a + pivot]))
2245 		return -1;
2246 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2247 
2248 	if (pivot > 2) {				/* jg upper_part */
2249 		/* Require near jump. */
2250 		jg_bytes = 4;
2251 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2252 	} else {
2253 		EMIT2(X86_JG, 0);
2254 	}
2255 	jg_reloc = prog;
2256 
2257 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2258 				  progs);
2259 	if (err)
2260 		return err;
2261 
2262 	/* From Intel 64 and IA-32 Architectures Optimization
2263 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2264 	 * Coding Rule 11: All branch targets should be 16-byte
2265 	 * aligned.
2266 	 */
2267 	emit_align(&prog, 16);
2268 	jg_offset = prog - jg_reloc;
2269 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2270 
2271 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2272 				  b, progs);
2273 	if (err)
2274 		return err;
2275 
2276 	*pprog = prog;
2277 	return 0;
2278 }
2279 
2280 static int cmp_ips(const void *a, const void *b)
2281 {
2282 	const s64 *ipa = a;
2283 	const s64 *ipb = b;
2284 
2285 	if (*ipa > *ipb)
2286 		return 1;
2287 	if (*ipa < *ipb)
2288 		return -1;
2289 	return 0;
2290 }
2291 
2292 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2293 {
2294 	u8 *prog = image;
2295 
2296 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2297 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2298 }
2299 
2300 struct x64_jit_data {
2301 	struct bpf_binary_header *rw_header;
2302 	struct bpf_binary_header *header;
2303 	int *addrs;
2304 	u8 *image;
2305 	int proglen;
2306 	struct jit_context ctx;
2307 };
2308 
2309 #define MAX_PASSES 20
2310 #define PADDING_PASSES (MAX_PASSES - 5)
2311 
2312 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2313 {
2314 	struct bpf_binary_header *rw_header = NULL;
2315 	struct bpf_binary_header *header = NULL;
2316 	struct bpf_prog *tmp, *orig_prog = prog;
2317 	struct x64_jit_data *jit_data;
2318 	int proglen, oldproglen = 0;
2319 	struct jit_context ctx = {};
2320 	bool tmp_blinded = false;
2321 	bool extra_pass = false;
2322 	bool padding = false;
2323 	u8 *rw_image = NULL;
2324 	u8 *image = NULL;
2325 	int *addrs;
2326 	int pass;
2327 	int i;
2328 
2329 	if (!prog->jit_requested)
2330 		return orig_prog;
2331 
2332 	tmp = bpf_jit_blind_constants(prog);
2333 	/*
2334 	 * If blinding was requested and we failed during blinding,
2335 	 * we must fall back to the interpreter.
2336 	 */
2337 	if (IS_ERR(tmp))
2338 		return orig_prog;
2339 	if (tmp != prog) {
2340 		tmp_blinded = true;
2341 		prog = tmp;
2342 	}
2343 
2344 	jit_data = prog->aux->jit_data;
2345 	if (!jit_data) {
2346 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2347 		if (!jit_data) {
2348 			prog = orig_prog;
2349 			goto out;
2350 		}
2351 		prog->aux->jit_data = jit_data;
2352 	}
2353 	addrs = jit_data->addrs;
2354 	if (addrs) {
2355 		ctx = jit_data->ctx;
2356 		oldproglen = jit_data->proglen;
2357 		image = jit_data->image;
2358 		header = jit_data->header;
2359 		rw_header = jit_data->rw_header;
2360 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2361 		extra_pass = true;
2362 		padding = true;
2363 		goto skip_init_addrs;
2364 	}
2365 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2366 	if (!addrs) {
2367 		prog = orig_prog;
2368 		goto out_addrs;
2369 	}
2370 
2371 	/*
2372 	 * Before first pass, make a rough estimation of addrs[]
2373 	 * each BPF instruction is translated to less than 64 bytes
2374 	 */
2375 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2376 		proglen += 64;
2377 		addrs[i] = proglen;
2378 	}
2379 	ctx.cleanup_addr = proglen;
2380 skip_init_addrs:
2381 
2382 	/*
2383 	 * JITed image shrinks with every pass and the loop iterates
2384 	 * until the image stops shrinking. Very large BPF programs
2385 	 * may converge on the last pass. In such case do one more
2386 	 * pass to emit the final image.
2387 	 */
2388 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2389 		if (!padding && pass >= PADDING_PASSES)
2390 			padding = true;
2391 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2392 		if (proglen <= 0) {
2393 out_image:
2394 			image = NULL;
2395 			if (header) {
2396 				bpf_arch_text_copy(&header->size, &rw_header->size,
2397 						   sizeof(rw_header->size));
2398 				bpf_jit_binary_pack_free(header, rw_header);
2399 			}
2400 			/* Fall back to interpreter mode */
2401 			prog = orig_prog;
2402 			if (extra_pass) {
2403 				prog->bpf_func = NULL;
2404 				prog->jited = 0;
2405 				prog->jited_len = 0;
2406 			}
2407 			goto out_addrs;
2408 		}
2409 		if (image) {
2410 			if (proglen != oldproglen) {
2411 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2412 				       proglen, oldproglen);
2413 				goto out_image;
2414 			}
2415 			break;
2416 		}
2417 		if (proglen == oldproglen) {
2418 			/*
2419 			 * The number of entries in extable is the number of BPF_LDX
2420 			 * insns that access kernel memory via "pointer to BTF type".
2421 			 * The verifier changed their opcode from LDX|MEM|size
2422 			 * to LDX|PROBE_MEM|size to make JITing easier.
2423 			 */
2424 			u32 align = __alignof__(struct exception_table_entry);
2425 			u32 extable_size = prog->aux->num_exentries *
2426 				sizeof(struct exception_table_entry);
2427 
2428 			/* allocate module memory for x86 insns and extable */
2429 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2430 							   &image, align, &rw_header, &rw_image,
2431 							   jit_fill_hole);
2432 			if (!header) {
2433 				prog = orig_prog;
2434 				goto out_addrs;
2435 			}
2436 			prog->aux->extable = (void *) image + roundup(proglen, align);
2437 		}
2438 		oldproglen = proglen;
2439 		cond_resched();
2440 	}
2441 
2442 	if (bpf_jit_enable > 1)
2443 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2444 
2445 	if (image) {
2446 		if (!prog->is_func || extra_pass) {
2447 			/*
2448 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2449 			 *   1) header is not pointing to proper module memory;
2450 			 *   2) the arch doesn't support bpf_arch_text_copy().
2451 			 *
2452 			 * Both cases are serious bugs and justify WARN_ON.
2453 			 */
2454 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2455 				/* header has been freed */
2456 				header = NULL;
2457 				goto out_image;
2458 			}
2459 
2460 			bpf_tail_call_direct_fixup(prog);
2461 		} else {
2462 			jit_data->addrs = addrs;
2463 			jit_data->ctx = ctx;
2464 			jit_data->proglen = proglen;
2465 			jit_data->image = image;
2466 			jit_data->header = header;
2467 			jit_data->rw_header = rw_header;
2468 		}
2469 		prog->bpf_func = (void *)image;
2470 		prog->jited = 1;
2471 		prog->jited_len = proglen;
2472 	} else {
2473 		prog = orig_prog;
2474 	}
2475 
2476 	if (!image || !prog->is_func || extra_pass) {
2477 		if (image)
2478 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2479 out_addrs:
2480 		kvfree(addrs);
2481 		kfree(jit_data);
2482 		prog->aux->jit_data = NULL;
2483 	}
2484 out:
2485 	if (tmp_blinded)
2486 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2487 					   tmp : orig_prog);
2488 	return prog;
2489 }
2490 
2491 bool bpf_jit_supports_kfunc_call(void)
2492 {
2493 	return true;
2494 }
2495 
2496 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2497 {
2498 	if (text_poke_copy(dst, src, len) == NULL)
2499 		return ERR_PTR(-EINVAL);
2500 	return dst;
2501 }
2502 
2503 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2504 bool bpf_jit_supports_subprog_tailcalls(void)
2505 {
2506 	return true;
2507 }
2508 
2509 void bpf_jit_free(struct bpf_prog *prog)
2510 {
2511 	if (prog->jited) {
2512 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2513 		struct bpf_binary_header *hdr;
2514 
2515 		/*
2516 		 * If we fail the final pass of JIT (from jit_subprogs),
2517 		 * the program may not be finalized yet. Call finalize here
2518 		 * before freeing it.
2519 		 */
2520 		if (jit_data) {
2521 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2522 						     jit_data->rw_header);
2523 			kvfree(jit_data->addrs);
2524 			kfree(jit_data);
2525 		}
2526 		hdr = bpf_jit_binary_pack_hdr(prog);
2527 		bpf_jit_binary_pack_free(hdr, NULL);
2528 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2529 	}
2530 
2531 	bpf_prog_unlock_free(prog);
2532 }
2533