xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision caf83e49)
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 struct jit_context {
232 	int cleanup_addr; /* Epilogue code offset */
233 
234 	/*
235 	 * Program specific offsets of labels in the code; these rely on the
236 	 * JIT doing at least 2 passes, recording the position on the first
237 	 * pass, only to generate the correct offset on the second pass.
238 	 */
239 	int tail_call_direct_label;
240 	int tail_call_indirect_label;
241 };
242 
243 /* Maximum number of bytes emitted while JITing one eBPF insn */
244 #define BPF_MAX_INSN_SIZE	128
245 #define BPF_INSN_SAFETY		64
246 
247 /* Number of bytes emit_patch() needs to generate instructions */
248 #define X86_PATCH_SIZE		5
249 /* Number of bytes that will be skipped on tailcall */
250 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
251 
252 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
253 {
254 	u8 *prog = *pprog;
255 
256 	if (callee_regs_used[0])
257 		EMIT1(0x53);         /* push rbx */
258 	if (callee_regs_used[1])
259 		EMIT2(0x41, 0x55);   /* push r13 */
260 	if (callee_regs_used[2])
261 		EMIT2(0x41, 0x56);   /* push r14 */
262 	if (callee_regs_used[3])
263 		EMIT2(0x41, 0x57);   /* push r15 */
264 	*pprog = prog;
265 }
266 
267 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
268 {
269 	u8 *prog = *pprog;
270 
271 	if (callee_regs_used[3])
272 		EMIT2(0x41, 0x5F);   /* pop r15 */
273 	if (callee_regs_used[2])
274 		EMIT2(0x41, 0x5E);   /* pop r14 */
275 	if (callee_regs_used[1])
276 		EMIT2(0x41, 0x5D);   /* pop r13 */
277 	if (callee_regs_used[0])
278 		EMIT1(0x5B);         /* pop rbx */
279 	*pprog = prog;
280 }
281 
282 /*
283  * Emit x86-64 prologue code for BPF program.
284  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
285  * while jumping to another program
286  */
287 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
288 			  bool tail_call_reachable, bool is_subprog)
289 {
290 	u8 *prog = *pprog;
291 
292 	/* BPF trampoline can be made to work without these nops,
293 	 * but let's waste 5 bytes for now and optimize later
294 	 */
295 	EMIT_ENDBR();
296 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
297 	prog += X86_PATCH_SIZE;
298 	if (!ebpf_from_cbpf) {
299 		if (tail_call_reachable && !is_subprog)
300 			EMIT2(0x31, 0xC0); /* xor eax, eax */
301 		else
302 			EMIT2(0x66, 0x90); /* nop2 */
303 	}
304 	EMIT1(0x55);             /* push rbp */
305 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
306 
307 	/* X86_TAIL_CALL_OFFSET is here */
308 	EMIT_ENDBR();
309 
310 	/* sub rsp, rounded_stack_depth */
311 	if (stack_depth)
312 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
313 	if (tail_call_reachable)
314 		EMIT1(0x50);         /* push rax */
315 	*pprog = prog;
316 }
317 
318 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
319 {
320 	u8 *prog = *pprog;
321 	s64 offset;
322 
323 	offset = func - (ip + X86_PATCH_SIZE);
324 	if (!is_simm32(offset)) {
325 		pr_err("Target call %p is out of range\n", func);
326 		return -ERANGE;
327 	}
328 	EMIT1_off32(opcode, offset);
329 	*pprog = prog;
330 	return 0;
331 }
332 
333 static int emit_call(u8 **pprog, void *func, void *ip)
334 {
335 	return emit_patch(pprog, func, ip, 0xE8);
336 }
337 
338 static int emit_jump(u8 **pprog, void *func, void *ip)
339 {
340 	return emit_patch(pprog, func, ip, 0xE9);
341 }
342 
343 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
344 				void *old_addr, void *new_addr)
345 {
346 	const u8 *nop_insn = x86_nops[5];
347 	u8 old_insn[X86_PATCH_SIZE];
348 	u8 new_insn[X86_PATCH_SIZE];
349 	u8 *prog;
350 	int ret;
351 
352 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
353 	if (old_addr) {
354 		prog = old_insn;
355 		ret = t == BPF_MOD_CALL ?
356 		      emit_call(&prog, old_addr, ip) :
357 		      emit_jump(&prog, old_addr, ip);
358 		if (ret)
359 			return ret;
360 	}
361 
362 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
363 	if (new_addr) {
364 		prog = new_insn;
365 		ret = t == BPF_MOD_CALL ?
366 		      emit_call(&prog, new_addr, ip) :
367 		      emit_jump(&prog, new_addr, ip);
368 		if (ret)
369 			return ret;
370 	}
371 
372 	ret = -EBUSY;
373 	mutex_lock(&text_mutex);
374 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
375 		goto out;
376 	ret = 1;
377 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
378 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
379 		ret = 0;
380 	}
381 out:
382 	mutex_unlock(&text_mutex);
383 	return ret;
384 }
385 
386 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
387 		       void *old_addr, void *new_addr)
388 {
389 	if (!is_kernel_text((long)ip) &&
390 	    !is_bpf_text_address((long)ip))
391 		/* BPF poking in modules is not supported */
392 		return -EINVAL;
393 
394 	/*
395 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
396 	 * with an ENDBR instruction.
397 	 */
398 	if (is_endbr(*(u32 *)ip))
399 		ip += ENDBR_INSN_SIZE;
400 
401 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
402 }
403 
404 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
405 
406 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
407 {
408 	u8 *prog = *pprog;
409 
410 #ifdef CONFIG_RETPOLINE
411 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
412 		EMIT_LFENCE();
413 		EMIT2(0xFF, 0xE0 + reg);
414 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
415 		emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
416 	} else
417 #endif
418 	EMIT2(0xFF, 0xE0 + reg);
419 
420 	*pprog = prog;
421 }
422 
423 /*
424  * Generate the following code:
425  *
426  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
427  *   if (index >= array->map.max_entries)
428  *     goto out;
429  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
430  *     goto out;
431  *   prog = array->ptrs[index];
432  *   if (prog == NULL)
433  *     goto out;
434  *   goto *(prog->bpf_func + prologue_size);
435  * out:
436  */
437 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
438 					u32 stack_depth, u8 *ip,
439 					struct jit_context *ctx)
440 {
441 	int tcc_off = -4 - round_up(stack_depth, 8);
442 	u8 *prog = *pprog, *start = *pprog;
443 	int offset;
444 
445 	/*
446 	 * rdi - pointer to ctx
447 	 * rsi - pointer to bpf_array
448 	 * rdx - index in bpf_array
449 	 */
450 
451 	/*
452 	 * if (index >= array->map.max_entries)
453 	 *	goto out;
454 	 */
455 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
456 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
457 	      offsetof(struct bpf_array, map.max_entries));
458 
459 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
460 	EMIT2(X86_JBE, offset);                   /* jbe out */
461 
462 	/*
463 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
464 	 *	goto out;
465 	 */
466 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
467 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
468 
469 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
470 	EMIT2(X86_JAE, offset);                   /* jae out */
471 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
472 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
473 
474 	/* prog = array->ptrs[index]; */
475 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
476 		    offsetof(struct bpf_array, ptrs));
477 
478 	/*
479 	 * if (prog == NULL)
480 	 *	goto out;
481 	 */
482 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
483 
484 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
485 	EMIT2(X86_JE, offset);                    /* je out */
486 
487 	pop_callee_regs(&prog, callee_regs_used);
488 
489 	EMIT1(0x58);                              /* pop rax */
490 	if (stack_depth)
491 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
492 			    round_up(stack_depth, 8));
493 
494 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
495 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
496 	      offsetof(struct bpf_prog, bpf_func));
497 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
498 	      X86_TAIL_CALL_OFFSET);
499 	/*
500 	 * Now we're ready to jump into next BPF program
501 	 * rdi == ctx (1st arg)
502 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
503 	 */
504 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
505 
506 	/* out: */
507 	ctx->tail_call_indirect_label = prog - start;
508 	*pprog = prog;
509 }
510 
511 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
512 				      u8 **pprog, u8 *ip,
513 				      bool *callee_regs_used, u32 stack_depth,
514 				      struct jit_context *ctx)
515 {
516 	int tcc_off = -4 - round_up(stack_depth, 8);
517 	u8 *prog = *pprog, *start = *pprog;
518 	int offset;
519 
520 	/*
521 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
522 	 *	goto out;
523 	 */
524 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
525 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
526 
527 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
528 	EMIT2(X86_JAE, offset);                       /* jae out */
529 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
530 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
531 
532 	poke->tailcall_bypass = ip + (prog - start);
533 	poke->adj_off = X86_TAIL_CALL_OFFSET;
534 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
535 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
536 
537 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
538 		  poke->tailcall_bypass);
539 
540 	pop_callee_regs(&prog, callee_regs_used);
541 	EMIT1(0x58);                                  /* pop rax */
542 	if (stack_depth)
543 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
544 
545 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
546 	prog += X86_PATCH_SIZE;
547 
548 	/* out: */
549 	ctx->tail_call_direct_label = prog - start;
550 
551 	*pprog = prog;
552 }
553 
554 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
555 {
556 	struct bpf_jit_poke_descriptor *poke;
557 	struct bpf_array *array;
558 	struct bpf_prog *target;
559 	int i, ret;
560 
561 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
562 		poke = &prog->aux->poke_tab[i];
563 		if (poke->aux && poke->aux != prog->aux)
564 			continue;
565 
566 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
567 
568 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
569 			continue;
570 
571 		array = container_of(poke->tail_call.map, struct bpf_array, map);
572 		mutex_lock(&array->aux->poke_mutex);
573 		target = array->ptrs[poke->tail_call.key];
574 		if (target) {
575 			ret = __bpf_arch_text_poke(poke->tailcall_target,
576 						   BPF_MOD_JUMP, NULL,
577 						   (u8 *)target->bpf_func +
578 						   poke->adj_off);
579 			BUG_ON(ret < 0);
580 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
581 						   BPF_MOD_JUMP,
582 						   (u8 *)poke->tailcall_target +
583 						   X86_PATCH_SIZE, NULL);
584 			BUG_ON(ret < 0);
585 		}
586 		WRITE_ONCE(poke->tailcall_target_stable, true);
587 		mutex_unlock(&array->aux->poke_mutex);
588 	}
589 }
590 
591 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
592 			   u32 dst_reg, const u32 imm32)
593 {
594 	u8 *prog = *pprog;
595 	u8 b1, b2, b3;
596 
597 	/*
598 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
599 	 * (which zero-extends imm32) to save 2 bytes.
600 	 */
601 	if (sign_propagate && (s32)imm32 < 0) {
602 		/* 'mov %rax, imm32' sign extends imm32 */
603 		b1 = add_1mod(0x48, dst_reg);
604 		b2 = 0xC7;
605 		b3 = 0xC0;
606 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
607 		goto done;
608 	}
609 
610 	/*
611 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
612 	 * to save 3 bytes.
613 	 */
614 	if (imm32 == 0) {
615 		if (is_ereg(dst_reg))
616 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
617 		b2 = 0x31; /* xor */
618 		b3 = 0xC0;
619 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
620 		goto done;
621 	}
622 
623 	/* mov %eax, imm32 */
624 	if (is_ereg(dst_reg))
625 		EMIT1(add_1mod(0x40, dst_reg));
626 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
627 done:
628 	*pprog = prog;
629 }
630 
631 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
632 			   const u32 imm32_hi, const u32 imm32_lo)
633 {
634 	u8 *prog = *pprog;
635 
636 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
637 		/*
638 		 * For emitting plain u32, where sign bit must not be
639 		 * propagated LLVM tends to load imm64 over mov32
640 		 * directly, so save couple of bytes by just doing
641 		 * 'mov %eax, imm32' instead.
642 		 */
643 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
644 	} else {
645 		/* movabsq %rax, imm64 */
646 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
647 		EMIT(imm32_lo, 4);
648 		EMIT(imm32_hi, 4);
649 	}
650 
651 	*pprog = prog;
652 }
653 
654 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
655 {
656 	u8 *prog = *pprog;
657 
658 	if (is64) {
659 		/* mov dst, src */
660 		EMIT_mov(dst_reg, src_reg);
661 	} else {
662 		/* mov32 dst, src */
663 		if (is_ereg(dst_reg) || is_ereg(src_reg))
664 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
665 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
666 	}
667 
668 	*pprog = prog;
669 }
670 
671 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
672 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
673 {
674 	u8 *prog = *pprog;
675 
676 	if (is_imm8(off)) {
677 		/* 1-byte signed displacement.
678 		 *
679 		 * If off == 0 we could skip this and save one extra byte, but
680 		 * special case of x86 R13 which always needs an offset is not
681 		 * worth the hassle
682 		 */
683 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
684 	} else {
685 		/* 4-byte signed displacement */
686 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
687 	}
688 	*pprog = prog;
689 }
690 
691 /*
692  * Emit a REX byte if it will be necessary to address these registers
693  */
694 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
695 {
696 	u8 *prog = *pprog;
697 
698 	if (is64)
699 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
700 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
701 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
702 	*pprog = prog;
703 }
704 
705 /*
706  * Similar version of maybe_emit_mod() for a single register
707  */
708 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
709 {
710 	u8 *prog = *pprog;
711 
712 	if (is64)
713 		EMIT1(add_1mod(0x48, reg));
714 	else if (is_ereg(reg))
715 		EMIT1(add_1mod(0x40, reg));
716 	*pprog = prog;
717 }
718 
719 /* LDX: dst_reg = *(u8*)(src_reg + off) */
720 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
721 {
722 	u8 *prog = *pprog;
723 
724 	switch (size) {
725 	case BPF_B:
726 		/* Emit 'movzx rax, byte ptr [rax + off]' */
727 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
728 		break;
729 	case BPF_H:
730 		/* Emit 'movzx rax, word ptr [rax + off]' */
731 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
732 		break;
733 	case BPF_W:
734 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
735 		if (is_ereg(dst_reg) || is_ereg(src_reg))
736 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
737 		else
738 			EMIT1(0x8B);
739 		break;
740 	case BPF_DW:
741 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
742 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
743 		break;
744 	}
745 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
746 	*pprog = prog;
747 }
748 
749 /* STX: *(u8*)(dst_reg + off) = src_reg */
750 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
751 {
752 	u8 *prog = *pprog;
753 
754 	switch (size) {
755 	case BPF_B:
756 		/* Emit 'mov byte ptr [rax + off], al' */
757 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
758 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
759 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
760 		else
761 			EMIT1(0x88);
762 		break;
763 	case BPF_H:
764 		if (is_ereg(dst_reg) || is_ereg(src_reg))
765 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
766 		else
767 			EMIT2(0x66, 0x89);
768 		break;
769 	case BPF_W:
770 		if (is_ereg(dst_reg) || is_ereg(src_reg))
771 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
772 		else
773 			EMIT1(0x89);
774 		break;
775 	case BPF_DW:
776 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
777 		break;
778 	}
779 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
780 	*pprog = prog;
781 }
782 
783 static int emit_atomic(u8 **pprog, u8 atomic_op,
784 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
785 {
786 	u8 *prog = *pprog;
787 
788 	EMIT1(0xF0); /* lock prefix */
789 
790 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
791 
792 	/* emit opcode */
793 	switch (atomic_op) {
794 	case BPF_ADD:
795 	case BPF_AND:
796 	case BPF_OR:
797 	case BPF_XOR:
798 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
799 		EMIT1(simple_alu_opcodes[atomic_op]);
800 		break;
801 	case BPF_ADD | BPF_FETCH:
802 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
803 		EMIT2(0x0F, 0xC1);
804 		break;
805 	case BPF_XCHG:
806 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
807 		EMIT1(0x87);
808 		break;
809 	case BPF_CMPXCHG:
810 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
811 		EMIT2(0x0F, 0xB1);
812 		break;
813 	default:
814 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
815 		return -EFAULT;
816 	}
817 
818 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
819 
820 	*pprog = prog;
821 	return 0;
822 }
823 
824 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
825 {
826 	u32 reg = x->fixup >> 8;
827 
828 	/* jump over faulting load and clear dest register */
829 	*(unsigned long *)((void *)regs + reg) = 0;
830 	regs->ip += x->fixup & 0xff;
831 	return true;
832 }
833 
834 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
835 			     bool *regs_used, bool *tail_call_seen)
836 {
837 	int i;
838 
839 	for (i = 1; i <= insn_cnt; i++, insn++) {
840 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
841 			*tail_call_seen = true;
842 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
843 			regs_used[0] = true;
844 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
845 			regs_used[1] = true;
846 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
847 			regs_used[2] = true;
848 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
849 			regs_used[3] = true;
850 	}
851 }
852 
853 static void emit_nops(u8 **pprog, int len)
854 {
855 	u8 *prog = *pprog;
856 	int i, noplen;
857 
858 	while (len > 0) {
859 		noplen = len;
860 
861 		if (noplen > ASM_NOP_MAX)
862 			noplen = ASM_NOP_MAX;
863 
864 		for (i = 0; i < noplen; i++)
865 			EMIT1(x86_nops[noplen][i]);
866 		len -= noplen;
867 	}
868 
869 	*pprog = prog;
870 }
871 
872 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
873 
874 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
875 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
876 {
877 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
878 	struct bpf_insn *insn = bpf_prog->insnsi;
879 	bool callee_regs_used[4] = {};
880 	int insn_cnt = bpf_prog->len;
881 	bool tail_call_seen = false;
882 	bool seen_exit = false;
883 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
884 	int i, excnt = 0;
885 	int ilen, proglen = 0;
886 	u8 *prog = temp;
887 	int err;
888 
889 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
890 			 &tail_call_seen);
891 
892 	/* tail call's presence in current prog implies it is reachable */
893 	tail_call_reachable |= tail_call_seen;
894 
895 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
896 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
897 		      bpf_prog->aux->func_idx != 0);
898 	push_callee_regs(&prog, callee_regs_used);
899 
900 	ilen = prog - temp;
901 	if (rw_image)
902 		memcpy(rw_image + proglen, temp, ilen);
903 	proglen += ilen;
904 	addrs[0] = proglen;
905 	prog = temp;
906 
907 	for (i = 1; i <= insn_cnt; i++, insn++) {
908 		const s32 imm32 = insn->imm;
909 		u32 dst_reg = insn->dst_reg;
910 		u32 src_reg = insn->src_reg;
911 		u8 b2 = 0, b3 = 0;
912 		u8 *start_of_ldx;
913 		s64 jmp_offset;
914 		u8 jmp_cond;
915 		u8 *func;
916 		int nops;
917 
918 		switch (insn->code) {
919 			/* ALU */
920 		case BPF_ALU | BPF_ADD | BPF_X:
921 		case BPF_ALU | BPF_SUB | BPF_X:
922 		case BPF_ALU | BPF_AND | BPF_X:
923 		case BPF_ALU | BPF_OR | BPF_X:
924 		case BPF_ALU | BPF_XOR | BPF_X:
925 		case BPF_ALU64 | BPF_ADD | BPF_X:
926 		case BPF_ALU64 | BPF_SUB | BPF_X:
927 		case BPF_ALU64 | BPF_AND | BPF_X:
928 		case BPF_ALU64 | BPF_OR | BPF_X:
929 		case BPF_ALU64 | BPF_XOR | BPF_X:
930 			maybe_emit_mod(&prog, dst_reg, src_reg,
931 				       BPF_CLASS(insn->code) == BPF_ALU64);
932 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
933 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
934 			break;
935 
936 		case BPF_ALU64 | BPF_MOV | BPF_X:
937 		case BPF_ALU | BPF_MOV | BPF_X:
938 			emit_mov_reg(&prog,
939 				     BPF_CLASS(insn->code) == BPF_ALU64,
940 				     dst_reg, src_reg);
941 			break;
942 
943 			/* neg dst */
944 		case BPF_ALU | BPF_NEG:
945 		case BPF_ALU64 | BPF_NEG:
946 			maybe_emit_1mod(&prog, dst_reg,
947 					BPF_CLASS(insn->code) == BPF_ALU64);
948 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
949 			break;
950 
951 		case BPF_ALU | BPF_ADD | BPF_K:
952 		case BPF_ALU | BPF_SUB | BPF_K:
953 		case BPF_ALU | BPF_AND | BPF_K:
954 		case BPF_ALU | BPF_OR | BPF_K:
955 		case BPF_ALU | BPF_XOR | BPF_K:
956 		case BPF_ALU64 | BPF_ADD | BPF_K:
957 		case BPF_ALU64 | BPF_SUB | BPF_K:
958 		case BPF_ALU64 | BPF_AND | BPF_K:
959 		case BPF_ALU64 | BPF_OR | BPF_K:
960 		case BPF_ALU64 | BPF_XOR | BPF_K:
961 			maybe_emit_1mod(&prog, dst_reg,
962 					BPF_CLASS(insn->code) == BPF_ALU64);
963 
964 			/*
965 			 * b3 holds 'normal' opcode, b2 short form only valid
966 			 * in case dst is eax/rax.
967 			 */
968 			switch (BPF_OP(insn->code)) {
969 			case BPF_ADD:
970 				b3 = 0xC0;
971 				b2 = 0x05;
972 				break;
973 			case BPF_SUB:
974 				b3 = 0xE8;
975 				b2 = 0x2D;
976 				break;
977 			case BPF_AND:
978 				b3 = 0xE0;
979 				b2 = 0x25;
980 				break;
981 			case BPF_OR:
982 				b3 = 0xC8;
983 				b2 = 0x0D;
984 				break;
985 			case BPF_XOR:
986 				b3 = 0xF0;
987 				b2 = 0x35;
988 				break;
989 			}
990 
991 			if (is_imm8(imm32))
992 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
993 			else if (is_axreg(dst_reg))
994 				EMIT1_off32(b2, imm32);
995 			else
996 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
997 			break;
998 
999 		case BPF_ALU64 | BPF_MOV | BPF_K:
1000 		case BPF_ALU | BPF_MOV | BPF_K:
1001 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1002 				       dst_reg, imm32);
1003 			break;
1004 
1005 		case BPF_LD | BPF_IMM | BPF_DW:
1006 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1007 			insn++;
1008 			i++;
1009 			break;
1010 
1011 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1012 		case BPF_ALU | BPF_MOD | BPF_X:
1013 		case BPF_ALU | BPF_DIV | BPF_X:
1014 		case BPF_ALU | BPF_MOD | BPF_K:
1015 		case BPF_ALU | BPF_DIV | BPF_K:
1016 		case BPF_ALU64 | BPF_MOD | BPF_X:
1017 		case BPF_ALU64 | BPF_DIV | BPF_X:
1018 		case BPF_ALU64 | BPF_MOD | BPF_K:
1019 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1020 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1021 
1022 			if (dst_reg != BPF_REG_0)
1023 				EMIT1(0x50); /* push rax */
1024 			if (dst_reg != BPF_REG_3)
1025 				EMIT1(0x52); /* push rdx */
1026 
1027 			if (BPF_SRC(insn->code) == BPF_X) {
1028 				if (src_reg == BPF_REG_0 ||
1029 				    src_reg == BPF_REG_3) {
1030 					/* mov r11, src_reg */
1031 					EMIT_mov(AUX_REG, src_reg);
1032 					src_reg = AUX_REG;
1033 				}
1034 			} else {
1035 				/* mov r11, imm32 */
1036 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1037 				src_reg = AUX_REG;
1038 			}
1039 
1040 			if (dst_reg != BPF_REG_0)
1041 				/* mov rax, dst_reg */
1042 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1043 
1044 			/*
1045 			 * xor edx, edx
1046 			 * equivalent to 'xor rdx, rdx', but one byte less
1047 			 */
1048 			EMIT2(0x31, 0xd2);
1049 
1050 			/* div src_reg */
1051 			maybe_emit_1mod(&prog, src_reg, is64);
1052 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1053 
1054 			if (BPF_OP(insn->code) == BPF_MOD &&
1055 			    dst_reg != BPF_REG_3)
1056 				/* mov dst_reg, rdx */
1057 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1058 			else if (BPF_OP(insn->code) == BPF_DIV &&
1059 				 dst_reg != BPF_REG_0)
1060 				/* mov dst_reg, rax */
1061 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1062 
1063 			if (dst_reg != BPF_REG_3)
1064 				EMIT1(0x5A); /* pop rdx */
1065 			if (dst_reg != BPF_REG_0)
1066 				EMIT1(0x58); /* pop rax */
1067 			break;
1068 		}
1069 
1070 		case BPF_ALU | BPF_MUL | BPF_K:
1071 		case BPF_ALU64 | BPF_MUL | BPF_K:
1072 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1073 				       BPF_CLASS(insn->code) == BPF_ALU64);
1074 
1075 			if (is_imm8(imm32))
1076 				/* imul dst_reg, dst_reg, imm8 */
1077 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1078 				      imm32);
1079 			else
1080 				/* imul dst_reg, dst_reg, imm32 */
1081 				EMIT2_off32(0x69,
1082 					    add_2reg(0xC0, dst_reg, dst_reg),
1083 					    imm32);
1084 			break;
1085 
1086 		case BPF_ALU | BPF_MUL | BPF_X:
1087 		case BPF_ALU64 | BPF_MUL | BPF_X:
1088 			maybe_emit_mod(&prog, src_reg, dst_reg,
1089 				       BPF_CLASS(insn->code) == BPF_ALU64);
1090 
1091 			/* imul dst_reg, src_reg */
1092 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1093 			break;
1094 
1095 			/* Shifts */
1096 		case BPF_ALU | BPF_LSH | BPF_K:
1097 		case BPF_ALU | BPF_RSH | BPF_K:
1098 		case BPF_ALU | BPF_ARSH | BPF_K:
1099 		case BPF_ALU64 | BPF_LSH | BPF_K:
1100 		case BPF_ALU64 | BPF_RSH | BPF_K:
1101 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1102 			maybe_emit_1mod(&prog, dst_reg,
1103 					BPF_CLASS(insn->code) == BPF_ALU64);
1104 
1105 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1106 			if (imm32 == 1)
1107 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1108 			else
1109 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1110 			break;
1111 
1112 		case BPF_ALU | BPF_LSH | BPF_X:
1113 		case BPF_ALU | BPF_RSH | BPF_X:
1114 		case BPF_ALU | BPF_ARSH | BPF_X:
1115 		case BPF_ALU64 | BPF_LSH | BPF_X:
1116 		case BPF_ALU64 | BPF_RSH | BPF_X:
1117 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1118 
1119 			/* Check for bad case when dst_reg == rcx */
1120 			if (dst_reg == BPF_REG_4) {
1121 				/* mov r11, dst_reg */
1122 				EMIT_mov(AUX_REG, dst_reg);
1123 				dst_reg = AUX_REG;
1124 			}
1125 
1126 			if (src_reg != BPF_REG_4) { /* common case */
1127 				EMIT1(0x51); /* push rcx */
1128 
1129 				/* mov rcx, src_reg */
1130 				EMIT_mov(BPF_REG_4, src_reg);
1131 			}
1132 
1133 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1134 			maybe_emit_1mod(&prog, dst_reg,
1135 					BPF_CLASS(insn->code) == BPF_ALU64);
1136 
1137 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1138 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1139 
1140 			if (src_reg != BPF_REG_4)
1141 				EMIT1(0x59); /* pop rcx */
1142 
1143 			if (insn->dst_reg == BPF_REG_4)
1144 				/* mov dst_reg, r11 */
1145 				EMIT_mov(insn->dst_reg, AUX_REG);
1146 			break;
1147 
1148 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1149 			switch (imm32) {
1150 			case 16:
1151 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1152 				EMIT1(0x66);
1153 				if (is_ereg(dst_reg))
1154 					EMIT1(0x41);
1155 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1156 
1157 				/* Emit 'movzwl eax, ax' */
1158 				if (is_ereg(dst_reg))
1159 					EMIT3(0x45, 0x0F, 0xB7);
1160 				else
1161 					EMIT2(0x0F, 0xB7);
1162 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1163 				break;
1164 			case 32:
1165 				/* Emit 'bswap eax' to swap lower 4 bytes */
1166 				if (is_ereg(dst_reg))
1167 					EMIT2(0x41, 0x0F);
1168 				else
1169 					EMIT1(0x0F);
1170 				EMIT1(add_1reg(0xC8, dst_reg));
1171 				break;
1172 			case 64:
1173 				/* Emit 'bswap rax' to swap 8 bytes */
1174 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1175 				      add_1reg(0xC8, dst_reg));
1176 				break;
1177 			}
1178 			break;
1179 
1180 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1181 			switch (imm32) {
1182 			case 16:
1183 				/*
1184 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1185 				 * into 64 bit
1186 				 */
1187 				if (is_ereg(dst_reg))
1188 					EMIT3(0x45, 0x0F, 0xB7);
1189 				else
1190 					EMIT2(0x0F, 0xB7);
1191 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1192 				break;
1193 			case 32:
1194 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1195 				if (is_ereg(dst_reg))
1196 					EMIT1(0x45);
1197 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1198 				break;
1199 			case 64:
1200 				/* nop */
1201 				break;
1202 			}
1203 			break;
1204 
1205 			/* speculation barrier */
1206 		case BPF_ST | BPF_NOSPEC:
1207 			if (boot_cpu_has(X86_FEATURE_XMM2))
1208 				EMIT_LFENCE();
1209 			break;
1210 
1211 			/* ST: *(u8*)(dst_reg + off) = imm */
1212 		case BPF_ST | BPF_MEM | BPF_B:
1213 			if (is_ereg(dst_reg))
1214 				EMIT2(0x41, 0xC6);
1215 			else
1216 				EMIT1(0xC6);
1217 			goto st;
1218 		case BPF_ST | BPF_MEM | BPF_H:
1219 			if (is_ereg(dst_reg))
1220 				EMIT3(0x66, 0x41, 0xC7);
1221 			else
1222 				EMIT2(0x66, 0xC7);
1223 			goto st;
1224 		case BPF_ST | BPF_MEM | BPF_W:
1225 			if (is_ereg(dst_reg))
1226 				EMIT2(0x41, 0xC7);
1227 			else
1228 				EMIT1(0xC7);
1229 			goto st;
1230 		case BPF_ST | BPF_MEM | BPF_DW:
1231 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1232 
1233 st:			if (is_imm8(insn->off))
1234 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1235 			else
1236 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1237 
1238 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1239 			break;
1240 
1241 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1242 		case BPF_STX | BPF_MEM | BPF_B:
1243 		case BPF_STX | BPF_MEM | BPF_H:
1244 		case BPF_STX | BPF_MEM | BPF_W:
1245 		case BPF_STX | BPF_MEM | BPF_DW:
1246 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1247 			break;
1248 
1249 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1250 		case BPF_LDX | BPF_MEM | BPF_B:
1251 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1252 		case BPF_LDX | BPF_MEM | BPF_H:
1253 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1254 		case BPF_LDX | BPF_MEM | BPF_W:
1255 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1256 		case BPF_LDX | BPF_MEM | BPF_DW:
1257 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1258 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1259 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1260 				 * add abs(insn->off) to the limit to make sure that negative
1261 				 * offset won't be an issue.
1262 				 * insn->off is s16, so it won't affect valid pointers.
1263 				 */
1264 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1265 				u8 *end_of_jmp1, *end_of_jmp2;
1266 
1267 				/* Conservatively check that src_reg + insn->off is a kernel address:
1268 				 * 1. src_reg + insn->off >= limit
1269 				 * 2. src_reg + insn->off doesn't become small positive.
1270 				 * Cannot do src_reg + insn->off >= limit in one branch,
1271 				 * since it needs two spare registers, but JIT has only one.
1272 				 */
1273 
1274 				/* movabsq r11, limit */
1275 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1276 				EMIT((u32)limit, 4);
1277 				EMIT(limit >> 32, 4);
1278 				/* cmp src_reg, r11 */
1279 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1280 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1281 				/* if unsigned '<' goto end_of_jmp2 */
1282 				EMIT2(X86_JB, 0);
1283 				end_of_jmp1 = prog;
1284 
1285 				/* mov r11, src_reg */
1286 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1287 				/* add r11, insn->off */
1288 				maybe_emit_1mod(&prog, AUX_REG, true);
1289 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1290 				/* jmp if not carry to start_of_ldx
1291 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1292 				 * that has to be rejected.
1293 				 */
1294 				EMIT2(0x73 /* JNC */, 0);
1295 				end_of_jmp2 = prog;
1296 
1297 				/* xor dst_reg, dst_reg */
1298 				emit_mov_imm32(&prog, false, dst_reg, 0);
1299 				/* jmp byte_after_ldx */
1300 				EMIT2(0xEB, 0);
1301 
1302 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1303 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1304 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1305 				start_of_ldx = prog;
1306 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1307 			}
1308 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1309 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1310 				struct exception_table_entry *ex;
1311 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1312 				s64 delta;
1313 
1314 				/* populate jmp_offset for JMP above */
1315 				start_of_ldx[-1] = prog - start_of_ldx;
1316 
1317 				if (!bpf_prog->aux->extable)
1318 					break;
1319 
1320 				if (excnt >= bpf_prog->aux->num_exentries) {
1321 					pr_err("ex gen bug\n");
1322 					return -EFAULT;
1323 				}
1324 				ex = &bpf_prog->aux->extable[excnt++];
1325 
1326 				delta = _insn - (u8 *)&ex->insn;
1327 				if (!is_simm32(delta)) {
1328 					pr_err("extable->insn doesn't fit into 32-bit\n");
1329 					return -EFAULT;
1330 				}
1331 				/* switch ex to rw buffer for writes */
1332 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1333 
1334 				ex->insn = delta;
1335 
1336 				ex->data = EX_TYPE_BPF;
1337 
1338 				if (dst_reg > BPF_REG_9) {
1339 					pr_err("verifier error\n");
1340 					return -EFAULT;
1341 				}
1342 				/*
1343 				 * Compute size of x86 insn and its target dest x86 register.
1344 				 * ex_handler_bpf() will use lower 8 bits to adjust
1345 				 * pt_regs->ip to jump over this x86 instruction
1346 				 * and upper bits to figure out which pt_regs to zero out.
1347 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1348 				 * of 4 bytes will be ignored and rbx will be zero inited.
1349 				 */
1350 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1351 			}
1352 			break;
1353 
1354 		case BPF_STX | BPF_ATOMIC | BPF_W:
1355 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1356 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1357 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1358 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1359 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1360 				u32 real_src_reg = src_reg;
1361 				u32 real_dst_reg = dst_reg;
1362 				u8 *branch_target;
1363 
1364 				/*
1365 				 * Can't be implemented with a single x86 insn.
1366 				 * Need to do a CMPXCHG loop.
1367 				 */
1368 
1369 				/* Will need RAX as a CMPXCHG operand so save R0 */
1370 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1371 				if (src_reg == BPF_REG_0)
1372 					real_src_reg = BPF_REG_AX;
1373 				if (dst_reg == BPF_REG_0)
1374 					real_dst_reg = BPF_REG_AX;
1375 
1376 				branch_target = prog;
1377 				/* Load old value */
1378 				emit_ldx(&prog, BPF_SIZE(insn->code),
1379 					 BPF_REG_0, real_dst_reg, insn->off);
1380 				/*
1381 				 * Perform the (commutative) operation locally,
1382 				 * put the result in the AUX_REG.
1383 				 */
1384 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1385 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1386 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1387 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1388 				/* Attempt to swap in new value */
1389 				err = emit_atomic(&prog, BPF_CMPXCHG,
1390 						  real_dst_reg, AUX_REG,
1391 						  insn->off,
1392 						  BPF_SIZE(insn->code));
1393 				if (WARN_ON(err))
1394 					return err;
1395 				/*
1396 				 * ZF tells us whether we won the race. If it's
1397 				 * cleared we need to try again.
1398 				 */
1399 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1400 				/* Return the pre-modification value */
1401 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1402 				/* Restore R0 after clobbering RAX */
1403 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1404 				break;
1405 			}
1406 
1407 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1408 					  insn->off, BPF_SIZE(insn->code));
1409 			if (err)
1410 				return err;
1411 			break;
1412 
1413 			/* call */
1414 		case BPF_JMP | BPF_CALL:
1415 			func = (u8 *) __bpf_call_base + imm32;
1416 			if (tail_call_reachable) {
1417 				EMIT3_off32(0x48, 0x8B, 0x85,
1418 					    -(bpf_prog->aux->stack_depth + 8));
1419 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1420 					return -EINVAL;
1421 			} else {
1422 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1423 					return -EINVAL;
1424 			}
1425 			break;
1426 
1427 		case BPF_JMP | BPF_TAIL_CALL:
1428 			if (imm32)
1429 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1430 							  &prog, image + addrs[i - 1],
1431 							  callee_regs_used,
1432 							  bpf_prog->aux->stack_depth,
1433 							  ctx);
1434 			else
1435 				emit_bpf_tail_call_indirect(&prog,
1436 							    callee_regs_used,
1437 							    bpf_prog->aux->stack_depth,
1438 							    image + addrs[i - 1],
1439 							    ctx);
1440 			break;
1441 
1442 			/* cond jump */
1443 		case BPF_JMP | BPF_JEQ | BPF_X:
1444 		case BPF_JMP | BPF_JNE | BPF_X:
1445 		case BPF_JMP | BPF_JGT | BPF_X:
1446 		case BPF_JMP | BPF_JLT | BPF_X:
1447 		case BPF_JMP | BPF_JGE | BPF_X:
1448 		case BPF_JMP | BPF_JLE | BPF_X:
1449 		case BPF_JMP | BPF_JSGT | BPF_X:
1450 		case BPF_JMP | BPF_JSLT | BPF_X:
1451 		case BPF_JMP | BPF_JSGE | BPF_X:
1452 		case BPF_JMP | BPF_JSLE | BPF_X:
1453 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1454 		case BPF_JMP32 | BPF_JNE | BPF_X:
1455 		case BPF_JMP32 | BPF_JGT | BPF_X:
1456 		case BPF_JMP32 | BPF_JLT | BPF_X:
1457 		case BPF_JMP32 | BPF_JGE | BPF_X:
1458 		case BPF_JMP32 | BPF_JLE | BPF_X:
1459 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1460 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1461 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1462 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1463 			/* cmp dst_reg, src_reg */
1464 			maybe_emit_mod(&prog, dst_reg, src_reg,
1465 				       BPF_CLASS(insn->code) == BPF_JMP);
1466 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1467 			goto emit_cond_jmp;
1468 
1469 		case BPF_JMP | BPF_JSET | BPF_X:
1470 		case BPF_JMP32 | BPF_JSET | BPF_X:
1471 			/* test dst_reg, src_reg */
1472 			maybe_emit_mod(&prog, dst_reg, src_reg,
1473 				       BPF_CLASS(insn->code) == BPF_JMP);
1474 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1475 			goto emit_cond_jmp;
1476 
1477 		case BPF_JMP | BPF_JSET | BPF_K:
1478 		case BPF_JMP32 | BPF_JSET | BPF_K:
1479 			/* test dst_reg, imm32 */
1480 			maybe_emit_1mod(&prog, dst_reg,
1481 					BPF_CLASS(insn->code) == BPF_JMP);
1482 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1483 			goto emit_cond_jmp;
1484 
1485 		case BPF_JMP | BPF_JEQ | BPF_K:
1486 		case BPF_JMP | BPF_JNE | BPF_K:
1487 		case BPF_JMP | BPF_JGT | BPF_K:
1488 		case BPF_JMP | BPF_JLT | BPF_K:
1489 		case BPF_JMP | BPF_JGE | BPF_K:
1490 		case BPF_JMP | BPF_JLE | BPF_K:
1491 		case BPF_JMP | BPF_JSGT | BPF_K:
1492 		case BPF_JMP | BPF_JSLT | BPF_K:
1493 		case BPF_JMP | BPF_JSGE | BPF_K:
1494 		case BPF_JMP | BPF_JSLE | BPF_K:
1495 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1496 		case BPF_JMP32 | BPF_JNE | BPF_K:
1497 		case BPF_JMP32 | BPF_JGT | BPF_K:
1498 		case BPF_JMP32 | BPF_JLT | BPF_K:
1499 		case BPF_JMP32 | BPF_JGE | BPF_K:
1500 		case BPF_JMP32 | BPF_JLE | BPF_K:
1501 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1502 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1503 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1504 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1505 			/* test dst_reg, dst_reg to save one extra byte */
1506 			if (imm32 == 0) {
1507 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1508 					       BPF_CLASS(insn->code) == BPF_JMP);
1509 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1510 				goto emit_cond_jmp;
1511 			}
1512 
1513 			/* cmp dst_reg, imm8/32 */
1514 			maybe_emit_1mod(&prog, dst_reg,
1515 					BPF_CLASS(insn->code) == BPF_JMP);
1516 
1517 			if (is_imm8(imm32))
1518 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1519 			else
1520 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1521 
1522 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1523 			switch (BPF_OP(insn->code)) {
1524 			case BPF_JEQ:
1525 				jmp_cond = X86_JE;
1526 				break;
1527 			case BPF_JSET:
1528 			case BPF_JNE:
1529 				jmp_cond = X86_JNE;
1530 				break;
1531 			case BPF_JGT:
1532 				/* GT is unsigned '>', JA in x86 */
1533 				jmp_cond = X86_JA;
1534 				break;
1535 			case BPF_JLT:
1536 				/* LT is unsigned '<', JB in x86 */
1537 				jmp_cond = X86_JB;
1538 				break;
1539 			case BPF_JGE:
1540 				/* GE is unsigned '>=', JAE in x86 */
1541 				jmp_cond = X86_JAE;
1542 				break;
1543 			case BPF_JLE:
1544 				/* LE is unsigned '<=', JBE in x86 */
1545 				jmp_cond = X86_JBE;
1546 				break;
1547 			case BPF_JSGT:
1548 				/* Signed '>', GT in x86 */
1549 				jmp_cond = X86_JG;
1550 				break;
1551 			case BPF_JSLT:
1552 				/* Signed '<', LT in x86 */
1553 				jmp_cond = X86_JL;
1554 				break;
1555 			case BPF_JSGE:
1556 				/* Signed '>=', GE in x86 */
1557 				jmp_cond = X86_JGE;
1558 				break;
1559 			case BPF_JSLE:
1560 				/* Signed '<=', LE in x86 */
1561 				jmp_cond = X86_JLE;
1562 				break;
1563 			default: /* to silence GCC warning */
1564 				return -EFAULT;
1565 			}
1566 			jmp_offset = addrs[i + insn->off] - addrs[i];
1567 			if (is_imm8(jmp_offset)) {
1568 				if (jmp_padding) {
1569 					/* To keep the jmp_offset valid, the extra bytes are
1570 					 * padded before the jump insn, so we subtract the
1571 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1572 					 *
1573 					 * If the previous pass already emits an imm8
1574 					 * jmp_cond, then this BPF insn won't shrink, so
1575 					 * "nops" is 0.
1576 					 *
1577 					 * On the other hand, if the previous pass emits an
1578 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1579 					 * keep the image from shrinking further.
1580 					 *
1581 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1582 					 *     is 2 bytes, so the size difference is 4 bytes.
1583 					 */
1584 					nops = INSN_SZ_DIFF - 2;
1585 					if (nops != 0 && nops != 4) {
1586 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1587 						       nops);
1588 						return -EFAULT;
1589 					}
1590 					emit_nops(&prog, nops);
1591 				}
1592 				EMIT2(jmp_cond, jmp_offset);
1593 			} else if (is_simm32(jmp_offset)) {
1594 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1595 			} else {
1596 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1597 				return -EFAULT;
1598 			}
1599 
1600 			break;
1601 
1602 		case BPF_JMP | BPF_JA:
1603 			if (insn->off == -1)
1604 				/* -1 jmp instructions will always jump
1605 				 * backwards two bytes. Explicitly handling
1606 				 * this case avoids wasting too many passes
1607 				 * when there are long sequences of replaced
1608 				 * dead code.
1609 				 */
1610 				jmp_offset = -2;
1611 			else
1612 				jmp_offset = addrs[i + insn->off] - addrs[i];
1613 
1614 			if (!jmp_offset) {
1615 				/*
1616 				 * If jmp_padding is enabled, the extra nops will
1617 				 * be inserted. Otherwise, optimize out nop jumps.
1618 				 */
1619 				if (jmp_padding) {
1620 					/* There are 3 possible conditions.
1621 					 * (1) This BPF_JA is already optimized out in
1622 					 *     the previous run, so there is no need
1623 					 *     to pad any extra byte (0 byte).
1624 					 * (2) The previous pass emits an imm8 jmp,
1625 					 *     so we pad 2 bytes to match the previous
1626 					 *     insn size.
1627 					 * (3) Similarly, the previous pass emits an
1628 					 *     imm32 jmp, and 5 bytes is padded.
1629 					 */
1630 					nops = INSN_SZ_DIFF;
1631 					if (nops != 0 && nops != 2 && nops != 5) {
1632 						pr_err("unexpected nop jump padding: %d bytes\n",
1633 						       nops);
1634 						return -EFAULT;
1635 					}
1636 					emit_nops(&prog, nops);
1637 				}
1638 				break;
1639 			}
1640 emit_jmp:
1641 			if (is_imm8(jmp_offset)) {
1642 				if (jmp_padding) {
1643 					/* To avoid breaking jmp_offset, the extra bytes
1644 					 * are padded before the actual jmp insn, so
1645 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1646 					 *
1647 					 * If the previous pass already emits an imm8
1648 					 * jmp, there is nothing to pad (0 byte).
1649 					 *
1650 					 * If it emits an imm32 jmp (5 bytes) previously
1651 					 * and now an imm8 jmp (2 bytes), then we pad
1652 					 * (5 - 2 = 3) bytes to stop the image from
1653 					 * shrinking further.
1654 					 */
1655 					nops = INSN_SZ_DIFF - 2;
1656 					if (nops != 0 && nops != 3) {
1657 						pr_err("unexpected jump padding: %d bytes\n",
1658 						       nops);
1659 						return -EFAULT;
1660 					}
1661 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1662 				}
1663 				EMIT2(0xEB, jmp_offset);
1664 			} else if (is_simm32(jmp_offset)) {
1665 				EMIT1_off32(0xE9, jmp_offset);
1666 			} else {
1667 				pr_err("jmp gen bug %llx\n", jmp_offset);
1668 				return -EFAULT;
1669 			}
1670 			break;
1671 
1672 		case BPF_JMP | BPF_EXIT:
1673 			if (seen_exit) {
1674 				jmp_offset = ctx->cleanup_addr - addrs[i];
1675 				goto emit_jmp;
1676 			}
1677 			seen_exit = true;
1678 			/* Update cleanup_addr */
1679 			ctx->cleanup_addr = proglen;
1680 			pop_callee_regs(&prog, callee_regs_used);
1681 			EMIT1(0xC9);         /* leave */
1682 			EMIT1(0xC3);         /* ret */
1683 			break;
1684 
1685 		default:
1686 			/*
1687 			 * By design x86-64 JIT should support all BPF instructions.
1688 			 * This error will be seen if new instruction was added
1689 			 * to the interpreter, but not to the JIT, or if there is
1690 			 * junk in bpf_prog.
1691 			 */
1692 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1693 			return -EINVAL;
1694 		}
1695 
1696 		ilen = prog - temp;
1697 		if (ilen > BPF_MAX_INSN_SIZE) {
1698 			pr_err("bpf_jit: fatal insn size error\n");
1699 			return -EFAULT;
1700 		}
1701 
1702 		if (image) {
1703 			/*
1704 			 * When populating the image, assert that:
1705 			 *
1706 			 *  i) We do not write beyond the allocated space, and
1707 			 * ii) addrs[i] did not change from the prior run, in order
1708 			 *     to validate assumptions made for computing branch
1709 			 *     displacements.
1710 			 */
1711 			if (unlikely(proglen + ilen > oldproglen ||
1712 				     proglen + ilen != addrs[i])) {
1713 				pr_err("bpf_jit: fatal error\n");
1714 				return -EFAULT;
1715 			}
1716 			memcpy(rw_image + proglen, temp, ilen);
1717 		}
1718 		proglen += ilen;
1719 		addrs[i] = proglen;
1720 		prog = temp;
1721 	}
1722 
1723 	if (image && excnt != bpf_prog->aux->num_exentries) {
1724 		pr_err("extable is not populated\n");
1725 		return -EFAULT;
1726 	}
1727 	return proglen;
1728 }
1729 
1730 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1731 		      int stack_size)
1732 {
1733 	int i;
1734 	/* Store function arguments to stack.
1735 	 * For a function that accepts two pointers the sequence will be:
1736 	 * mov QWORD PTR [rbp-0x10],rdi
1737 	 * mov QWORD PTR [rbp-0x8],rsi
1738 	 */
1739 	for (i = 0; i < min(nr_args, 6); i++)
1740 		emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1741 			 BPF_REG_FP,
1742 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1743 			 -(stack_size - i * 8));
1744 }
1745 
1746 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1747 			 int stack_size)
1748 {
1749 	int i;
1750 
1751 	/* Restore function arguments from stack.
1752 	 * For a function that accepts two pointers the sequence will be:
1753 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1754 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1755 	 */
1756 	for (i = 0; i < min(nr_args, 6); i++)
1757 		emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1758 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1759 			 BPF_REG_FP,
1760 			 -(stack_size - i * 8));
1761 }
1762 
1763 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1764 			   struct bpf_prog *p, int stack_size, bool save_ret)
1765 {
1766 	u8 *prog = *pprog;
1767 	u8 *jmp_insn;
1768 
1769 	/* arg1: mov rdi, progs[i] */
1770 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1771 	if (emit_call(&prog,
1772 		      p->aux->sleepable ? __bpf_prog_enter_sleepable :
1773 		      __bpf_prog_enter, prog))
1774 			return -EINVAL;
1775 	/* remember prog start time returned by __bpf_prog_enter */
1776 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1777 
1778 	/* if (__bpf_prog_enter*(prog) == 0)
1779 	 *	goto skip_exec_of_prog;
1780 	 */
1781 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1782 	/* emit 2 nops that will be replaced with JE insn */
1783 	jmp_insn = prog;
1784 	emit_nops(&prog, 2);
1785 
1786 	/* arg1: lea rdi, [rbp - stack_size] */
1787 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1788 	/* arg2: progs[i]->insnsi for interpreter */
1789 	if (!p->jited)
1790 		emit_mov_imm64(&prog, BPF_REG_2,
1791 			       (long) p->insnsi >> 32,
1792 			       (u32) (long) p->insnsi);
1793 	/* call JITed bpf program or interpreter */
1794 	if (emit_call(&prog, p->bpf_func, prog))
1795 		return -EINVAL;
1796 
1797 	/*
1798 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1799 	 * of the previous call which is then passed on the stack to
1800 	 * the next BPF program.
1801 	 *
1802 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1803 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1804 	 */
1805 	if (save_ret)
1806 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1807 
1808 	/* replace 2 nops with JE insn, since jmp target is known */
1809 	jmp_insn[0] = X86_JE;
1810 	jmp_insn[1] = prog - jmp_insn - 2;
1811 
1812 	/* arg1: mov rdi, progs[i] */
1813 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1814 	/* arg2: mov rsi, rbx <- start time in nsec */
1815 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1816 	if (emit_call(&prog,
1817 		      p->aux->sleepable ? __bpf_prog_exit_sleepable :
1818 		      __bpf_prog_exit, prog))
1819 			return -EINVAL;
1820 
1821 	*pprog = prog;
1822 	return 0;
1823 }
1824 
1825 static void emit_align(u8 **pprog, u32 align)
1826 {
1827 	u8 *target, *prog = *pprog;
1828 
1829 	target = PTR_ALIGN(prog, align);
1830 	if (target != prog)
1831 		emit_nops(&prog, target - prog);
1832 
1833 	*pprog = prog;
1834 }
1835 
1836 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1837 {
1838 	u8 *prog = *pprog;
1839 	s64 offset;
1840 
1841 	offset = func - (ip + 2 + 4);
1842 	if (!is_simm32(offset)) {
1843 		pr_err("Target %p is out of range\n", func);
1844 		return -EINVAL;
1845 	}
1846 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1847 	*pprog = prog;
1848 	return 0;
1849 }
1850 
1851 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1852 		      struct bpf_tramp_progs *tp, int stack_size,
1853 		      bool save_ret)
1854 {
1855 	int i;
1856 	u8 *prog = *pprog;
1857 
1858 	for (i = 0; i < tp->nr_progs; i++) {
1859 		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size,
1860 				    save_ret))
1861 			return -EINVAL;
1862 	}
1863 	*pprog = prog;
1864 	return 0;
1865 }
1866 
1867 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1868 			      struct bpf_tramp_progs *tp, int stack_size,
1869 			      u8 **branches)
1870 {
1871 	u8 *prog = *pprog;
1872 	int i;
1873 
1874 	/* The first fmod_ret program will receive a garbage return value.
1875 	 * Set this to 0 to avoid confusing the program.
1876 	 */
1877 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1878 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1879 	for (i = 0; i < tp->nr_progs; i++) {
1880 		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1881 			return -EINVAL;
1882 
1883 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
1884 		 * if (*(u64 *)(rbp - 8) !=  0)
1885 		 *	goto do_fexit;
1886 		 */
1887 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
1888 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1889 
1890 		/* Save the location of the branch and Generate 6 nops
1891 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
1892 		 * are replaced with a conditional jump once do_fexit (i.e. the
1893 		 * start of the fexit invocation) is finalized.
1894 		 */
1895 		branches[i] = prog;
1896 		emit_nops(&prog, 4 + 2);
1897 	}
1898 
1899 	*pprog = prog;
1900 	return 0;
1901 }
1902 
1903 static bool is_valid_bpf_tramp_flags(unsigned int flags)
1904 {
1905 	if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1906 	    (flags & BPF_TRAMP_F_SKIP_FRAME))
1907 		return false;
1908 
1909 	/*
1910 	 * BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
1911 	 * and it must be used alone.
1912 	 */
1913 	if ((flags & BPF_TRAMP_F_RET_FENTRY_RET) &&
1914 	    (flags & ~BPF_TRAMP_F_RET_FENTRY_RET))
1915 		return false;
1916 
1917 	return true;
1918 }
1919 
1920 /* Example:
1921  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1922  * its 'struct btf_func_model' will be nr_args=2
1923  * The assembly code when eth_type_trans is executing after trampoline:
1924  *
1925  * push rbp
1926  * mov rbp, rsp
1927  * sub rsp, 16                     // space for skb and dev
1928  * push rbx                        // temp regs to pass start time
1929  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1930  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1931  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1932  * mov rbx, rax                    // remember start time in bpf stats are enabled
1933  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1934  * call addr_of_jited_FENTRY_prog
1935  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1936  * mov rsi, rbx                    // prog start time
1937  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1938  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
1939  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
1940  * pop rbx
1941  * leave
1942  * ret
1943  *
1944  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1945  * replaced with 'call generated_bpf_trampoline'. When it returns
1946  * eth_type_trans will continue executing with original skb and dev pointers.
1947  *
1948  * The assembly code when eth_type_trans is called from trampoline:
1949  *
1950  * push rbp
1951  * mov rbp, rsp
1952  * sub rsp, 24                     // space for skb, dev, return value
1953  * push rbx                        // temp regs to pass start time
1954  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
1955  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
1956  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1957  * mov rbx, rax                    // remember start time if bpf stats are enabled
1958  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1959  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
1960  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1961  * mov rsi, rbx                    // prog start time
1962  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1963  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
1964  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
1965  * call eth_type_trans+5           // execute body of eth_type_trans
1966  * mov qword ptr [rbp - 8], rax    // save return value
1967  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1968  * mov rbx, rax                    // remember start time in bpf stats are enabled
1969  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1970  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
1971  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1972  * mov rsi, rbx                    // prog start time
1973  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1974  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
1975  * pop rbx
1976  * leave
1977  * add rsp, 8                      // skip eth_type_trans's frame
1978  * ret                             // return to its caller
1979  */
1980 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1981 				const struct btf_func_model *m, u32 flags,
1982 				struct bpf_tramp_progs *tprogs,
1983 				void *orig_call)
1984 {
1985 	int ret, i, nr_args = m->nr_args;
1986 	int regs_off, ip_off, args_off, stack_size = nr_args * 8;
1987 	struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
1988 	struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
1989 	struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
1990 	u8 **branches = NULL;
1991 	u8 *prog;
1992 	bool save_ret;
1993 
1994 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
1995 	if (nr_args > 6)
1996 		return -ENOTSUPP;
1997 
1998 	if (!is_valid_bpf_tramp_flags(flags))
1999 		return -EINVAL;
2000 
2001 	/* Generated trampoline stack layout:
2002 	 *
2003 	 * RBP + 8         [ return address  ]
2004 	 * RBP + 0         [ RBP             ]
2005 	 *
2006 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2007 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2008 	 *
2009 	 *                 [ reg_argN        ]  always
2010 	 *                 [ ...             ]
2011 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2012 	 *
2013 	 * RBP - args_off  [ args count      ]  always
2014 	 *
2015 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2016 	 */
2017 
2018 	/* room for return value of orig_call or fentry prog */
2019 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2020 	if (save_ret)
2021 		stack_size += 8;
2022 
2023 	regs_off = stack_size;
2024 
2025 	/* args count  */
2026 	stack_size += 8;
2027 	args_off = stack_size;
2028 
2029 	if (flags & BPF_TRAMP_F_IP_ARG)
2030 		stack_size += 8; /* room for IP address argument */
2031 
2032 	ip_off = stack_size;
2033 
2034 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2035 		/* skip patched call instruction and point orig_call to actual
2036 		 * body of the kernel function.
2037 		 */
2038 		if (is_endbr(*(u32 *)orig_call))
2039 			orig_call += ENDBR_INSN_SIZE;
2040 		orig_call += X86_PATCH_SIZE;
2041 	}
2042 
2043 	prog = image;
2044 
2045 	EMIT_ENDBR();
2046 	EMIT1(0x55);		 /* push rbp */
2047 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2048 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2049 	EMIT1(0x53);		 /* push rbx */
2050 
2051 	/* Store number of arguments of the traced function:
2052 	 *   mov rax, nr_args
2053 	 *   mov QWORD PTR [rbp - args_off], rax
2054 	 */
2055 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args);
2056 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2057 
2058 	if (flags & BPF_TRAMP_F_IP_ARG) {
2059 		/* Store IP address of the traced function:
2060 		 * mov rax, QWORD PTR [rbp + 8]
2061 		 * sub rax, X86_PATCH_SIZE
2062 		 * mov QWORD PTR [rbp - ip_off], rax
2063 		 */
2064 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2065 		EMIT4(0x48, 0x83, 0xe8, X86_PATCH_SIZE);
2066 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2067 	}
2068 
2069 	save_regs(m, &prog, nr_args, regs_off);
2070 
2071 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2072 		/* arg1: mov rdi, im */
2073 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2074 		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2075 			ret = -EINVAL;
2076 			goto cleanup;
2077 		}
2078 	}
2079 
2080 	if (fentry->nr_progs)
2081 		if (invoke_bpf(m, &prog, fentry, regs_off,
2082 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2083 			return -EINVAL;
2084 
2085 	if (fmod_ret->nr_progs) {
2086 		branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
2087 				   GFP_KERNEL);
2088 		if (!branches)
2089 			return -ENOMEM;
2090 
2091 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2092 				       branches)) {
2093 			ret = -EINVAL;
2094 			goto cleanup;
2095 		}
2096 	}
2097 
2098 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2099 		restore_regs(m, &prog, nr_args, regs_off);
2100 
2101 		/* call original function */
2102 		if (emit_call(&prog, orig_call, prog)) {
2103 			ret = -EINVAL;
2104 			goto cleanup;
2105 		}
2106 		/* remember return value in a stack for bpf prog to access */
2107 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2108 		im->ip_after_call = prog;
2109 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2110 		prog += X86_PATCH_SIZE;
2111 	}
2112 
2113 	if (fmod_ret->nr_progs) {
2114 		/* From Intel 64 and IA-32 Architectures Optimization
2115 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2116 		 * Coding Rule 11: All branch targets should be 16-byte
2117 		 * aligned.
2118 		 */
2119 		emit_align(&prog, 16);
2120 		/* Update the branches saved in invoke_bpf_mod_ret with the
2121 		 * aligned address of do_fexit.
2122 		 */
2123 		for (i = 0; i < fmod_ret->nr_progs; i++)
2124 			emit_cond_near_jump(&branches[i], prog, branches[i],
2125 					    X86_JNE);
2126 	}
2127 
2128 	if (fexit->nr_progs)
2129 		if (invoke_bpf(m, &prog, fexit, regs_off, false)) {
2130 			ret = -EINVAL;
2131 			goto cleanup;
2132 		}
2133 
2134 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2135 		restore_regs(m, &prog, nr_args, regs_off);
2136 
2137 	/* This needs to be done regardless. If there were fmod_ret programs,
2138 	 * the return value is only updated on the stack and still needs to be
2139 	 * restored to R0.
2140 	 */
2141 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2142 		im->ip_epilogue = prog;
2143 		/* arg1: mov rdi, im */
2144 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2145 		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2146 			ret = -EINVAL;
2147 			goto cleanup;
2148 		}
2149 	}
2150 	/* restore return value of orig_call or fentry prog back into RAX */
2151 	if (save_ret)
2152 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2153 
2154 	EMIT1(0x5B); /* pop rbx */
2155 	EMIT1(0xC9); /* leave */
2156 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2157 		/* skip our return address and return to parent */
2158 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2159 	EMIT1(0xC3); /* ret */
2160 	/* Make sure the trampoline generation logic doesn't overflow */
2161 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2162 		ret = -EFAULT;
2163 		goto cleanup;
2164 	}
2165 	ret = prog - (u8 *)image;
2166 
2167 cleanup:
2168 	kfree(branches);
2169 	return ret;
2170 }
2171 
2172 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2173 {
2174 	u8 *jg_reloc, *prog = *pprog;
2175 	int pivot, err, jg_bytes = 1;
2176 	s64 jg_offset;
2177 
2178 	if (a == b) {
2179 		/* Leaf node of recursion, i.e. not a range of indices
2180 		 * anymore.
2181 		 */
2182 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2183 		if (!is_simm32(progs[a]))
2184 			return -1;
2185 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2186 			    progs[a]);
2187 		err = emit_cond_near_jump(&prog,	/* je func */
2188 					  (void *)progs[a], prog,
2189 					  X86_JE);
2190 		if (err)
2191 			return err;
2192 
2193 		emit_indirect_jump(&prog, 2 /* rdx */, prog);
2194 
2195 		*pprog = prog;
2196 		return 0;
2197 	}
2198 
2199 	/* Not a leaf node, so we pivot, and recursively descend into
2200 	 * the lower and upper ranges.
2201 	 */
2202 	pivot = (b - a) / 2;
2203 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2204 	if (!is_simm32(progs[a + pivot]))
2205 		return -1;
2206 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2207 
2208 	if (pivot > 2) {				/* jg upper_part */
2209 		/* Require near jump. */
2210 		jg_bytes = 4;
2211 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2212 	} else {
2213 		EMIT2(X86_JG, 0);
2214 	}
2215 	jg_reloc = prog;
2216 
2217 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2218 				  progs);
2219 	if (err)
2220 		return err;
2221 
2222 	/* From Intel 64 and IA-32 Architectures Optimization
2223 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2224 	 * Coding Rule 11: All branch targets should be 16-byte
2225 	 * aligned.
2226 	 */
2227 	emit_align(&prog, 16);
2228 	jg_offset = prog - jg_reloc;
2229 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2230 
2231 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2232 				  b, progs);
2233 	if (err)
2234 		return err;
2235 
2236 	*pprog = prog;
2237 	return 0;
2238 }
2239 
2240 static int cmp_ips(const void *a, const void *b)
2241 {
2242 	const s64 *ipa = a;
2243 	const s64 *ipb = b;
2244 
2245 	if (*ipa > *ipb)
2246 		return 1;
2247 	if (*ipa < *ipb)
2248 		return -1;
2249 	return 0;
2250 }
2251 
2252 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2253 {
2254 	u8 *prog = image;
2255 
2256 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2257 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2258 }
2259 
2260 struct x64_jit_data {
2261 	struct bpf_binary_header *rw_header;
2262 	struct bpf_binary_header *header;
2263 	int *addrs;
2264 	u8 *image;
2265 	int proglen;
2266 	struct jit_context ctx;
2267 };
2268 
2269 #define MAX_PASSES 20
2270 #define PADDING_PASSES (MAX_PASSES - 5)
2271 
2272 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2273 {
2274 	struct bpf_binary_header *rw_header = NULL;
2275 	struct bpf_binary_header *header = NULL;
2276 	struct bpf_prog *tmp, *orig_prog = prog;
2277 	struct x64_jit_data *jit_data;
2278 	int proglen, oldproglen = 0;
2279 	struct jit_context ctx = {};
2280 	bool tmp_blinded = false;
2281 	bool extra_pass = false;
2282 	bool padding = false;
2283 	u8 *rw_image = NULL;
2284 	u8 *image = NULL;
2285 	int *addrs;
2286 	int pass;
2287 	int i;
2288 
2289 	if (!prog->jit_requested)
2290 		return orig_prog;
2291 
2292 	tmp = bpf_jit_blind_constants(prog);
2293 	/*
2294 	 * If blinding was requested and we failed during blinding,
2295 	 * we must fall back to the interpreter.
2296 	 */
2297 	if (IS_ERR(tmp))
2298 		return orig_prog;
2299 	if (tmp != prog) {
2300 		tmp_blinded = true;
2301 		prog = tmp;
2302 	}
2303 
2304 	jit_data = prog->aux->jit_data;
2305 	if (!jit_data) {
2306 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2307 		if (!jit_data) {
2308 			prog = orig_prog;
2309 			goto out;
2310 		}
2311 		prog->aux->jit_data = jit_data;
2312 	}
2313 	addrs = jit_data->addrs;
2314 	if (addrs) {
2315 		ctx = jit_data->ctx;
2316 		oldproglen = jit_data->proglen;
2317 		image = jit_data->image;
2318 		header = jit_data->header;
2319 		rw_header = jit_data->rw_header;
2320 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2321 		extra_pass = true;
2322 		padding = true;
2323 		goto skip_init_addrs;
2324 	}
2325 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2326 	if (!addrs) {
2327 		prog = orig_prog;
2328 		goto out_addrs;
2329 	}
2330 
2331 	/*
2332 	 * Before first pass, make a rough estimation of addrs[]
2333 	 * each BPF instruction is translated to less than 64 bytes
2334 	 */
2335 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2336 		proglen += 64;
2337 		addrs[i] = proglen;
2338 	}
2339 	ctx.cleanup_addr = proglen;
2340 skip_init_addrs:
2341 
2342 	/*
2343 	 * JITed image shrinks with every pass and the loop iterates
2344 	 * until the image stops shrinking. Very large BPF programs
2345 	 * may converge on the last pass. In such case do one more
2346 	 * pass to emit the final image.
2347 	 */
2348 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2349 		if (!padding && pass >= PADDING_PASSES)
2350 			padding = true;
2351 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2352 		if (proglen <= 0) {
2353 out_image:
2354 			image = NULL;
2355 			if (header) {
2356 				bpf_arch_text_copy(&header->size, &rw_header->size,
2357 						   sizeof(rw_header->size));
2358 				bpf_jit_binary_pack_free(header, rw_header);
2359 			}
2360 			/* Fall back to interpreter mode */
2361 			prog = orig_prog;
2362 			if (extra_pass) {
2363 				prog->bpf_func = NULL;
2364 				prog->jited = 0;
2365 				prog->jited_len = 0;
2366 			}
2367 			goto out_addrs;
2368 		}
2369 		if (image) {
2370 			if (proglen != oldproglen) {
2371 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2372 				       proglen, oldproglen);
2373 				goto out_image;
2374 			}
2375 			break;
2376 		}
2377 		if (proglen == oldproglen) {
2378 			/*
2379 			 * The number of entries in extable is the number of BPF_LDX
2380 			 * insns that access kernel memory via "pointer to BTF type".
2381 			 * The verifier changed their opcode from LDX|MEM|size
2382 			 * to LDX|PROBE_MEM|size to make JITing easier.
2383 			 */
2384 			u32 align = __alignof__(struct exception_table_entry);
2385 			u32 extable_size = prog->aux->num_exentries *
2386 				sizeof(struct exception_table_entry);
2387 
2388 			/* allocate module memory for x86 insns and extable */
2389 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2390 							   &image, align, &rw_header, &rw_image,
2391 							   jit_fill_hole);
2392 			if (!header) {
2393 				prog = orig_prog;
2394 				goto out_addrs;
2395 			}
2396 			prog->aux->extable = (void *) image + roundup(proglen, align);
2397 		}
2398 		oldproglen = proglen;
2399 		cond_resched();
2400 	}
2401 
2402 	if (bpf_jit_enable > 1)
2403 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2404 
2405 	if (image) {
2406 		if (!prog->is_func || extra_pass) {
2407 			/*
2408 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2409 			 *   1) header is not pointing to proper module memory;
2410 			 *   2) the arch doesn't support bpf_arch_text_copy().
2411 			 *
2412 			 * Both cases are serious bugs and justify WARN_ON.
2413 			 */
2414 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2415 				/* header has been freed */
2416 				header = NULL;
2417 				goto out_image;
2418 			}
2419 
2420 			bpf_tail_call_direct_fixup(prog);
2421 		} else {
2422 			jit_data->addrs = addrs;
2423 			jit_data->ctx = ctx;
2424 			jit_data->proglen = proglen;
2425 			jit_data->image = image;
2426 			jit_data->header = header;
2427 			jit_data->rw_header = rw_header;
2428 		}
2429 		prog->bpf_func = (void *)image;
2430 		prog->jited = 1;
2431 		prog->jited_len = proglen;
2432 	} else {
2433 		prog = orig_prog;
2434 	}
2435 
2436 	if (!image || !prog->is_func || extra_pass) {
2437 		if (image)
2438 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2439 out_addrs:
2440 		kvfree(addrs);
2441 		kfree(jit_data);
2442 		prog->aux->jit_data = NULL;
2443 	}
2444 out:
2445 	if (tmp_blinded)
2446 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2447 					   tmp : orig_prog);
2448 	return prog;
2449 }
2450 
2451 bool bpf_jit_supports_kfunc_call(void)
2452 {
2453 	return true;
2454 }
2455 
2456 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2457 {
2458 	if (text_poke_copy(dst, src, len) == NULL)
2459 		return ERR_PTR(-EINVAL);
2460 	return dst;
2461 }
2462