xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision 47010c04)
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 		OPTIMIZER_HIDE_VAR(reg);
416 		emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
417 	} else
418 #endif
419 	EMIT2(0xFF, 0xE0 + reg);
420 
421 	*pprog = prog;
422 }
423 
424 /*
425  * Generate the following code:
426  *
427  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
428  *   if (index >= array->map.max_entries)
429  *     goto out;
430  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
431  *     goto out;
432  *   prog = array->ptrs[index];
433  *   if (prog == NULL)
434  *     goto out;
435  *   goto *(prog->bpf_func + prologue_size);
436  * out:
437  */
438 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
439 					u32 stack_depth, u8 *ip,
440 					struct jit_context *ctx)
441 {
442 	int tcc_off = -4 - round_up(stack_depth, 8);
443 	u8 *prog = *pprog, *start = *pprog;
444 	int offset;
445 
446 	/*
447 	 * rdi - pointer to ctx
448 	 * rsi - pointer to bpf_array
449 	 * rdx - index in bpf_array
450 	 */
451 
452 	/*
453 	 * if (index >= array->map.max_entries)
454 	 *	goto out;
455 	 */
456 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
457 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
458 	      offsetof(struct bpf_array, map.max_entries));
459 
460 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
461 	EMIT2(X86_JBE, offset);                   /* jbe out */
462 
463 	/*
464 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
465 	 *	goto out;
466 	 */
467 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
468 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
469 
470 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
471 	EMIT2(X86_JAE, offset);                   /* jae out */
472 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
473 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
474 
475 	/* prog = array->ptrs[index]; */
476 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
477 		    offsetof(struct bpf_array, ptrs));
478 
479 	/*
480 	 * if (prog == NULL)
481 	 *	goto out;
482 	 */
483 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
484 
485 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
486 	EMIT2(X86_JE, offset);                    /* je out */
487 
488 	pop_callee_regs(&prog, callee_regs_used);
489 
490 	EMIT1(0x58);                              /* pop rax */
491 	if (stack_depth)
492 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
493 			    round_up(stack_depth, 8));
494 
495 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
496 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
497 	      offsetof(struct bpf_prog, bpf_func));
498 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
499 	      X86_TAIL_CALL_OFFSET);
500 	/*
501 	 * Now we're ready to jump into next BPF program
502 	 * rdi == ctx (1st arg)
503 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
504 	 */
505 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
506 
507 	/* out: */
508 	ctx->tail_call_indirect_label = prog - start;
509 	*pprog = prog;
510 }
511 
512 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
513 				      u8 **pprog, u8 *ip,
514 				      bool *callee_regs_used, u32 stack_depth,
515 				      struct jit_context *ctx)
516 {
517 	int tcc_off = -4 - round_up(stack_depth, 8);
518 	u8 *prog = *pprog, *start = *pprog;
519 	int offset;
520 
521 	/*
522 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
523 	 *	goto out;
524 	 */
525 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
526 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
527 
528 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
529 	EMIT2(X86_JAE, offset);                       /* jae out */
530 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
531 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
532 
533 	poke->tailcall_bypass = ip + (prog - start);
534 	poke->adj_off = X86_TAIL_CALL_OFFSET;
535 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
536 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
537 
538 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
539 		  poke->tailcall_bypass);
540 
541 	pop_callee_regs(&prog, callee_regs_used);
542 	EMIT1(0x58);                                  /* pop rax */
543 	if (stack_depth)
544 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
545 
546 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
547 	prog += X86_PATCH_SIZE;
548 
549 	/* out: */
550 	ctx->tail_call_direct_label = prog - start;
551 
552 	*pprog = prog;
553 }
554 
555 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
556 {
557 	struct bpf_jit_poke_descriptor *poke;
558 	struct bpf_array *array;
559 	struct bpf_prog *target;
560 	int i, ret;
561 
562 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
563 		poke = &prog->aux->poke_tab[i];
564 		if (poke->aux && poke->aux != prog->aux)
565 			continue;
566 
567 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
568 
569 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
570 			continue;
571 
572 		array = container_of(poke->tail_call.map, struct bpf_array, map);
573 		mutex_lock(&array->aux->poke_mutex);
574 		target = array->ptrs[poke->tail_call.key];
575 		if (target) {
576 			ret = __bpf_arch_text_poke(poke->tailcall_target,
577 						   BPF_MOD_JUMP, NULL,
578 						   (u8 *)target->bpf_func +
579 						   poke->adj_off);
580 			BUG_ON(ret < 0);
581 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
582 						   BPF_MOD_JUMP,
583 						   (u8 *)poke->tailcall_target +
584 						   X86_PATCH_SIZE, NULL);
585 			BUG_ON(ret < 0);
586 		}
587 		WRITE_ONCE(poke->tailcall_target_stable, true);
588 		mutex_unlock(&array->aux->poke_mutex);
589 	}
590 }
591 
592 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
593 			   u32 dst_reg, const u32 imm32)
594 {
595 	u8 *prog = *pprog;
596 	u8 b1, b2, b3;
597 
598 	/*
599 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
600 	 * (which zero-extends imm32) to save 2 bytes.
601 	 */
602 	if (sign_propagate && (s32)imm32 < 0) {
603 		/* 'mov %rax, imm32' sign extends imm32 */
604 		b1 = add_1mod(0x48, dst_reg);
605 		b2 = 0xC7;
606 		b3 = 0xC0;
607 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
608 		goto done;
609 	}
610 
611 	/*
612 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
613 	 * to save 3 bytes.
614 	 */
615 	if (imm32 == 0) {
616 		if (is_ereg(dst_reg))
617 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
618 		b2 = 0x31; /* xor */
619 		b3 = 0xC0;
620 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
621 		goto done;
622 	}
623 
624 	/* mov %eax, imm32 */
625 	if (is_ereg(dst_reg))
626 		EMIT1(add_1mod(0x40, dst_reg));
627 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
628 done:
629 	*pprog = prog;
630 }
631 
632 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
633 			   const u32 imm32_hi, const u32 imm32_lo)
634 {
635 	u8 *prog = *pprog;
636 
637 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
638 		/*
639 		 * For emitting plain u32, where sign bit must not be
640 		 * propagated LLVM tends to load imm64 over mov32
641 		 * directly, so save couple of bytes by just doing
642 		 * 'mov %eax, imm32' instead.
643 		 */
644 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
645 	} else {
646 		/* movabsq %rax, imm64 */
647 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
648 		EMIT(imm32_lo, 4);
649 		EMIT(imm32_hi, 4);
650 	}
651 
652 	*pprog = prog;
653 }
654 
655 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
656 {
657 	u8 *prog = *pprog;
658 
659 	if (is64) {
660 		/* mov dst, src */
661 		EMIT_mov(dst_reg, src_reg);
662 	} else {
663 		/* mov32 dst, src */
664 		if (is_ereg(dst_reg) || is_ereg(src_reg))
665 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
666 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
667 	}
668 
669 	*pprog = prog;
670 }
671 
672 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
673 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
674 {
675 	u8 *prog = *pprog;
676 
677 	if (is_imm8(off)) {
678 		/* 1-byte signed displacement.
679 		 *
680 		 * If off == 0 we could skip this and save one extra byte, but
681 		 * special case of x86 R13 which always needs an offset is not
682 		 * worth the hassle
683 		 */
684 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
685 	} else {
686 		/* 4-byte signed displacement */
687 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
688 	}
689 	*pprog = prog;
690 }
691 
692 /*
693  * Emit a REX byte if it will be necessary to address these registers
694  */
695 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
696 {
697 	u8 *prog = *pprog;
698 
699 	if (is64)
700 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
701 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
702 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
703 	*pprog = prog;
704 }
705 
706 /*
707  * Similar version of maybe_emit_mod() for a single register
708  */
709 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
710 {
711 	u8 *prog = *pprog;
712 
713 	if (is64)
714 		EMIT1(add_1mod(0x48, reg));
715 	else if (is_ereg(reg))
716 		EMIT1(add_1mod(0x40, reg));
717 	*pprog = prog;
718 }
719 
720 /* LDX: dst_reg = *(u8*)(src_reg + off) */
721 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
722 {
723 	u8 *prog = *pprog;
724 
725 	switch (size) {
726 	case BPF_B:
727 		/* Emit 'movzx rax, byte ptr [rax + off]' */
728 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
729 		break;
730 	case BPF_H:
731 		/* Emit 'movzx rax, word ptr [rax + off]' */
732 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
733 		break;
734 	case BPF_W:
735 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
736 		if (is_ereg(dst_reg) || is_ereg(src_reg))
737 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
738 		else
739 			EMIT1(0x8B);
740 		break;
741 	case BPF_DW:
742 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
743 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
744 		break;
745 	}
746 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
747 	*pprog = prog;
748 }
749 
750 /* STX: *(u8*)(dst_reg + off) = src_reg */
751 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
752 {
753 	u8 *prog = *pprog;
754 
755 	switch (size) {
756 	case BPF_B:
757 		/* Emit 'mov byte ptr [rax + off], al' */
758 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
759 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
760 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
761 		else
762 			EMIT1(0x88);
763 		break;
764 	case BPF_H:
765 		if (is_ereg(dst_reg) || is_ereg(src_reg))
766 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
767 		else
768 			EMIT2(0x66, 0x89);
769 		break;
770 	case BPF_W:
771 		if (is_ereg(dst_reg) || is_ereg(src_reg))
772 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
773 		else
774 			EMIT1(0x89);
775 		break;
776 	case BPF_DW:
777 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
778 		break;
779 	}
780 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
781 	*pprog = prog;
782 }
783 
784 static int emit_atomic(u8 **pprog, u8 atomic_op,
785 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
786 {
787 	u8 *prog = *pprog;
788 
789 	EMIT1(0xF0); /* lock prefix */
790 
791 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
792 
793 	/* emit opcode */
794 	switch (atomic_op) {
795 	case BPF_ADD:
796 	case BPF_AND:
797 	case BPF_OR:
798 	case BPF_XOR:
799 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
800 		EMIT1(simple_alu_opcodes[atomic_op]);
801 		break;
802 	case BPF_ADD | BPF_FETCH:
803 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
804 		EMIT2(0x0F, 0xC1);
805 		break;
806 	case BPF_XCHG:
807 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
808 		EMIT1(0x87);
809 		break;
810 	case BPF_CMPXCHG:
811 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
812 		EMIT2(0x0F, 0xB1);
813 		break;
814 	default:
815 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
816 		return -EFAULT;
817 	}
818 
819 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
820 
821 	*pprog = prog;
822 	return 0;
823 }
824 
825 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
826 {
827 	u32 reg = x->fixup >> 8;
828 
829 	/* jump over faulting load and clear dest register */
830 	*(unsigned long *)((void *)regs + reg) = 0;
831 	regs->ip += x->fixup & 0xff;
832 	return true;
833 }
834 
835 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
836 			     bool *regs_used, bool *tail_call_seen)
837 {
838 	int i;
839 
840 	for (i = 1; i <= insn_cnt; i++, insn++) {
841 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
842 			*tail_call_seen = true;
843 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
844 			regs_used[0] = true;
845 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
846 			regs_used[1] = true;
847 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
848 			regs_used[2] = true;
849 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
850 			regs_used[3] = true;
851 	}
852 }
853 
854 static void emit_nops(u8 **pprog, int len)
855 {
856 	u8 *prog = *pprog;
857 	int i, noplen;
858 
859 	while (len > 0) {
860 		noplen = len;
861 
862 		if (noplen > ASM_NOP_MAX)
863 			noplen = ASM_NOP_MAX;
864 
865 		for (i = 0; i < noplen; i++)
866 			EMIT1(x86_nops[noplen][i]);
867 		len -= noplen;
868 	}
869 
870 	*pprog = prog;
871 }
872 
873 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
874 
875 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
876 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
877 {
878 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
879 	struct bpf_insn *insn = bpf_prog->insnsi;
880 	bool callee_regs_used[4] = {};
881 	int insn_cnt = bpf_prog->len;
882 	bool tail_call_seen = false;
883 	bool seen_exit = false;
884 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
885 	int i, excnt = 0;
886 	int ilen, proglen = 0;
887 	u8 *prog = temp;
888 	int err;
889 
890 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
891 			 &tail_call_seen);
892 
893 	/* tail call's presence in current prog implies it is reachable */
894 	tail_call_reachable |= tail_call_seen;
895 
896 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
897 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
898 		      bpf_prog->aux->func_idx != 0);
899 	push_callee_regs(&prog, callee_regs_used);
900 
901 	ilen = prog - temp;
902 	if (rw_image)
903 		memcpy(rw_image + proglen, temp, ilen);
904 	proglen += ilen;
905 	addrs[0] = proglen;
906 	prog = temp;
907 
908 	for (i = 1; i <= insn_cnt; i++, insn++) {
909 		const s32 imm32 = insn->imm;
910 		u32 dst_reg = insn->dst_reg;
911 		u32 src_reg = insn->src_reg;
912 		u8 b2 = 0, b3 = 0;
913 		u8 *start_of_ldx;
914 		s64 jmp_offset;
915 		u8 jmp_cond;
916 		u8 *func;
917 		int nops;
918 
919 		switch (insn->code) {
920 			/* ALU */
921 		case BPF_ALU | BPF_ADD | BPF_X:
922 		case BPF_ALU | BPF_SUB | BPF_X:
923 		case BPF_ALU | BPF_AND | BPF_X:
924 		case BPF_ALU | BPF_OR | BPF_X:
925 		case BPF_ALU | BPF_XOR | BPF_X:
926 		case BPF_ALU64 | BPF_ADD | BPF_X:
927 		case BPF_ALU64 | BPF_SUB | BPF_X:
928 		case BPF_ALU64 | BPF_AND | BPF_X:
929 		case BPF_ALU64 | BPF_OR | BPF_X:
930 		case BPF_ALU64 | BPF_XOR | BPF_X:
931 			maybe_emit_mod(&prog, dst_reg, src_reg,
932 				       BPF_CLASS(insn->code) == BPF_ALU64);
933 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
934 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
935 			break;
936 
937 		case BPF_ALU64 | BPF_MOV | BPF_X:
938 		case BPF_ALU | BPF_MOV | BPF_X:
939 			emit_mov_reg(&prog,
940 				     BPF_CLASS(insn->code) == BPF_ALU64,
941 				     dst_reg, src_reg);
942 			break;
943 
944 			/* neg dst */
945 		case BPF_ALU | BPF_NEG:
946 		case BPF_ALU64 | BPF_NEG:
947 			maybe_emit_1mod(&prog, dst_reg,
948 					BPF_CLASS(insn->code) == BPF_ALU64);
949 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
950 			break;
951 
952 		case BPF_ALU | BPF_ADD | BPF_K:
953 		case BPF_ALU | BPF_SUB | BPF_K:
954 		case BPF_ALU | BPF_AND | BPF_K:
955 		case BPF_ALU | BPF_OR | BPF_K:
956 		case BPF_ALU | BPF_XOR | BPF_K:
957 		case BPF_ALU64 | BPF_ADD | BPF_K:
958 		case BPF_ALU64 | BPF_SUB | BPF_K:
959 		case BPF_ALU64 | BPF_AND | BPF_K:
960 		case BPF_ALU64 | BPF_OR | BPF_K:
961 		case BPF_ALU64 | BPF_XOR | BPF_K:
962 			maybe_emit_1mod(&prog, dst_reg,
963 					BPF_CLASS(insn->code) == BPF_ALU64);
964 
965 			/*
966 			 * b3 holds 'normal' opcode, b2 short form only valid
967 			 * in case dst is eax/rax.
968 			 */
969 			switch (BPF_OP(insn->code)) {
970 			case BPF_ADD:
971 				b3 = 0xC0;
972 				b2 = 0x05;
973 				break;
974 			case BPF_SUB:
975 				b3 = 0xE8;
976 				b2 = 0x2D;
977 				break;
978 			case BPF_AND:
979 				b3 = 0xE0;
980 				b2 = 0x25;
981 				break;
982 			case BPF_OR:
983 				b3 = 0xC8;
984 				b2 = 0x0D;
985 				break;
986 			case BPF_XOR:
987 				b3 = 0xF0;
988 				b2 = 0x35;
989 				break;
990 			}
991 
992 			if (is_imm8(imm32))
993 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
994 			else if (is_axreg(dst_reg))
995 				EMIT1_off32(b2, imm32);
996 			else
997 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
998 			break;
999 
1000 		case BPF_ALU64 | BPF_MOV | BPF_K:
1001 		case BPF_ALU | BPF_MOV | BPF_K:
1002 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1003 				       dst_reg, imm32);
1004 			break;
1005 
1006 		case BPF_LD | BPF_IMM | BPF_DW:
1007 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1008 			insn++;
1009 			i++;
1010 			break;
1011 
1012 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1013 		case BPF_ALU | BPF_MOD | BPF_X:
1014 		case BPF_ALU | BPF_DIV | BPF_X:
1015 		case BPF_ALU | BPF_MOD | BPF_K:
1016 		case BPF_ALU | BPF_DIV | BPF_K:
1017 		case BPF_ALU64 | BPF_MOD | BPF_X:
1018 		case BPF_ALU64 | BPF_DIV | BPF_X:
1019 		case BPF_ALU64 | BPF_MOD | BPF_K:
1020 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1021 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1022 
1023 			if (dst_reg != BPF_REG_0)
1024 				EMIT1(0x50); /* push rax */
1025 			if (dst_reg != BPF_REG_3)
1026 				EMIT1(0x52); /* push rdx */
1027 
1028 			if (BPF_SRC(insn->code) == BPF_X) {
1029 				if (src_reg == BPF_REG_0 ||
1030 				    src_reg == BPF_REG_3) {
1031 					/* mov r11, src_reg */
1032 					EMIT_mov(AUX_REG, src_reg);
1033 					src_reg = AUX_REG;
1034 				}
1035 			} else {
1036 				/* mov r11, imm32 */
1037 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1038 				src_reg = AUX_REG;
1039 			}
1040 
1041 			if (dst_reg != BPF_REG_0)
1042 				/* mov rax, dst_reg */
1043 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1044 
1045 			/*
1046 			 * xor edx, edx
1047 			 * equivalent to 'xor rdx, rdx', but one byte less
1048 			 */
1049 			EMIT2(0x31, 0xd2);
1050 
1051 			/* div src_reg */
1052 			maybe_emit_1mod(&prog, src_reg, is64);
1053 			EMIT2(0xF7, add_1reg(0xF0, src_reg));
1054 
1055 			if (BPF_OP(insn->code) == BPF_MOD &&
1056 			    dst_reg != BPF_REG_3)
1057 				/* mov dst_reg, rdx */
1058 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1059 			else if (BPF_OP(insn->code) == BPF_DIV &&
1060 				 dst_reg != BPF_REG_0)
1061 				/* mov dst_reg, rax */
1062 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1063 
1064 			if (dst_reg != BPF_REG_3)
1065 				EMIT1(0x5A); /* pop rdx */
1066 			if (dst_reg != BPF_REG_0)
1067 				EMIT1(0x58); /* pop rax */
1068 			break;
1069 		}
1070 
1071 		case BPF_ALU | BPF_MUL | BPF_K:
1072 		case BPF_ALU64 | BPF_MUL | BPF_K:
1073 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1074 				       BPF_CLASS(insn->code) == BPF_ALU64);
1075 
1076 			if (is_imm8(imm32))
1077 				/* imul dst_reg, dst_reg, imm8 */
1078 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1079 				      imm32);
1080 			else
1081 				/* imul dst_reg, dst_reg, imm32 */
1082 				EMIT2_off32(0x69,
1083 					    add_2reg(0xC0, dst_reg, dst_reg),
1084 					    imm32);
1085 			break;
1086 
1087 		case BPF_ALU | BPF_MUL | BPF_X:
1088 		case BPF_ALU64 | BPF_MUL | BPF_X:
1089 			maybe_emit_mod(&prog, src_reg, dst_reg,
1090 				       BPF_CLASS(insn->code) == BPF_ALU64);
1091 
1092 			/* imul dst_reg, src_reg */
1093 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1094 			break;
1095 
1096 			/* Shifts */
1097 		case BPF_ALU | BPF_LSH | BPF_K:
1098 		case BPF_ALU | BPF_RSH | BPF_K:
1099 		case BPF_ALU | BPF_ARSH | BPF_K:
1100 		case BPF_ALU64 | BPF_LSH | BPF_K:
1101 		case BPF_ALU64 | BPF_RSH | BPF_K:
1102 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1103 			maybe_emit_1mod(&prog, dst_reg,
1104 					BPF_CLASS(insn->code) == BPF_ALU64);
1105 
1106 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1107 			if (imm32 == 1)
1108 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1109 			else
1110 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1111 			break;
1112 
1113 		case BPF_ALU | BPF_LSH | BPF_X:
1114 		case BPF_ALU | BPF_RSH | BPF_X:
1115 		case BPF_ALU | BPF_ARSH | BPF_X:
1116 		case BPF_ALU64 | BPF_LSH | BPF_X:
1117 		case BPF_ALU64 | BPF_RSH | BPF_X:
1118 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1119 
1120 			/* Check for bad case when dst_reg == rcx */
1121 			if (dst_reg == BPF_REG_4) {
1122 				/* mov r11, dst_reg */
1123 				EMIT_mov(AUX_REG, dst_reg);
1124 				dst_reg = AUX_REG;
1125 			}
1126 
1127 			if (src_reg != BPF_REG_4) { /* common case */
1128 				EMIT1(0x51); /* push rcx */
1129 
1130 				/* mov rcx, src_reg */
1131 				EMIT_mov(BPF_REG_4, src_reg);
1132 			}
1133 
1134 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1135 			maybe_emit_1mod(&prog, dst_reg,
1136 					BPF_CLASS(insn->code) == BPF_ALU64);
1137 
1138 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1139 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1140 
1141 			if (src_reg != BPF_REG_4)
1142 				EMIT1(0x59); /* pop rcx */
1143 
1144 			if (insn->dst_reg == BPF_REG_4)
1145 				/* mov dst_reg, r11 */
1146 				EMIT_mov(insn->dst_reg, AUX_REG);
1147 			break;
1148 
1149 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1150 			switch (imm32) {
1151 			case 16:
1152 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1153 				EMIT1(0x66);
1154 				if (is_ereg(dst_reg))
1155 					EMIT1(0x41);
1156 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1157 
1158 				/* Emit 'movzwl eax, ax' */
1159 				if (is_ereg(dst_reg))
1160 					EMIT3(0x45, 0x0F, 0xB7);
1161 				else
1162 					EMIT2(0x0F, 0xB7);
1163 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1164 				break;
1165 			case 32:
1166 				/* Emit 'bswap eax' to swap lower 4 bytes */
1167 				if (is_ereg(dst_reg))
1168 					EMIT2(0x41, 0x0F);
1169 				else
1170 					EMIT1(0x0F);
1171 				EMIT1(add_1reg(0xC8, dst_reg));
1172 				break;
1173 			case 64:
1174 				/* Emit 'bswap rax' to swap 8 bytes */
1175 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1176 				      add_1reg(0xC8, dst_reg));
1177 				break;
1178 			}
1179 			break;
1180 
1181 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1182 			switch (imm32) {
1183 			case 16:
1184 				/*
1185 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1186 				 * into 64 bit
1187 				 */
1188 				if (is_ereg(dst_reg))
1189 					EMIT3(0x45, 0x0F, 0xB7);
1190 				else
1191 					EMIT2(0x0F, 0xB7);
1192 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1193 				break;
1194 			case 32:
1195 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1196 				if (is_ereg(dst_reg))
1197 					EMIT1(0x45);
1198 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1199 				break;
1200 			case 64:
1201 				/* nop */
1202 				break;
1203 			}
1204 			break;
1205 
1206 			/* speculation barrier */
1207 		case BPF_ST | BPF_NOSPEC:
1208 			if (boot_cpu_has(X86_FEATURE_XMM2))
1209 				EMIT_LFENCE();
1210 			break;
1211 
1212 			/* ST: *(u8*)(dst_reg + off) = imm */
1213 		case BPF_ST | BPF_MEM | BPF_B:
1214 			if (is_ereg(dst_reg))
1215 				EMIT2(0x41, 0xC6);
1216 			else
1217 				EMIT1(0xC6);
1218 			goto st;
1219 		case BPF_ST | BPF_MEM | BPF_H:
1220 			if (is_ereg(dst_reg))
1221 				EMIT3(0x66, 0x41, 0xC7);
1222 			else
1223 				EMIT2(0x66, 0xC7);
1224 			goto st;
1225 		case BPF_ST | BPF_MEM | BPF_W:
1226 			if (is_ereg(dst_reg))
1227 				EMIT2(0x41, 0xC7);
1228 			else
1229 				EMIT1(0xC7);
1230 			goto st;
1231 		case BPF_ST | BPF_MEM | BPF_DW:
1232 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1233 
1234 st:			if (is_imm8(insn->off))
1235 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1236 			else
1237 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1238 
1239 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1240 			break;
1241 
1242 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1243 		case BPF_STX | BPF_MEM | BPF_B:
1244 		case BPF_STX | BPF_MEM | BPF_H:
1245 		case BPF_STX | BPF_MEM | BPF_W:
1246 		case BPF_STX | BPF_MEM | BPF_DW:
1247 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1248 			break;
1249 
1250 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1251 		case BPF_LDX | BPF_MEM | BPF_B:
1252 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1253 		case BPF_LDX | BPF_MEM | BPF_H:
1254 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1255 		case BPF_LDX | BPF_MEM | BPF_W:
1256 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1257 		case BPF_LDX | BPF_MEM | BPF_DW:
1258 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1259 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1260 				/* Though the verifier prevents negative insn->off in BPF_PROBE_MEM
1261 				 * add abs(insn->off) to the limit to make sure that negative
1262 				 * offset won't be an issue.
1263 				 * insn->off is s16, so it won't affect valid pointers.
1264 				 */
1265 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE + abs(insn->off);
1266 				u8 *end_of_jmp1, *end_of_jmp2;
1267 
1268 				/* Conservatively check that src_reg + insn->off is a kernel address:
1269 				 * 1. src_reg + insn->off >= limit
1270 				 * 2. src_reg + insn->off doesn't become small positive.
1271 				 * Cannot do src_reg + insn->off >= limit in one branch,
1272 				 * since it needs two spare registers, but JIT has only one.
1273 				 */
1274 
1275 				/* movabsq r11, limit */
1276 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1277 				EMIT((u32)limit, 4);
1278 				EMIT(limit >> 32, 4);
1279 				/* cmp src_reg, r11 */
1280 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1281 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1282 				/* if unsigned '<' goto end_of_jmp2 */
1283 				EMIT2(X86_JB, 0);
1284 				end_of_jmp1 = prog;
1285 
1286 				/* mov r11, src_reg */
1287 				emit_mov_reg(&prog, true, AUX_REG, src_reg);
1288 				/* add r11, insn->off */
1289 				maybe_emit_1mod(&prog, AUX_REG, true);
1290 				EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
1291 				/* jmp if not carry to start_of_ldx
1292 				 * Otherwise ERR_PTR(-EINVAL) + 128 will be the user addr
1293 				 * that has to be rejected.
1294 				 */
1295 				EMIT2(0x73 /* JNC */, 0);
1296 				end_of_jmp2 = prog;
1297 
1298 				/* xor dst_reg, dst_reg */
1299 				emit_mov_imm32(&prog, false, dst_reg, 0);
1300 				/* jmp byte_after_ldx */
1301 				EMIT2(0xEB, 0);
1302 
1303 				/* populate jmp_offset for JB above to jump to xor dst_reg */
1304 				end_of_jmp1[-1] = end_of_jmp2 - end_of_jmp1;
1305 				/* populate jmp_offset for JNC above to jump to start_of_ldx */
1306 				start_of_ldx = prog;
1307 				end_of_jmp2[-1] = start_of_ldx - end_of_jmp2;
1308 			}
1309 			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1310 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1311 				struct exception_table_entry *ex;
1312 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1313 				s64 delta;
1314 
1315 				/* populate jmp_offset for JMP above */
1316 				start_of_ldx[-1] = prog - start_of_ldx;
1317 
1318 				if (!bpf_prog->aux->extable)
1319 					break;
1320 
1321 				if (excnt >= bpf_prog->aux->num_exentries) {
1322 					pr_err("ex gen bug\n");
1323 					return -EFAULT;
1324 				}
1325 				ex = &bpf_prog->aux->extable[excnt++];
1326 
1327 				delta = _insn - (u8 *)&ex->insn;
1328 				if (!is_simm32(delta)) {
1329 					pr_err("extable->insn doesn't fit into 32-bit\n");
1330 					return -EFAULT;
1331 				}
1332 				/* switch ex to rw buffer for writes */
1333 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1334 
1335 				ex->insn = delta;
1336 
1337 				ex->data = EX_TYPE_BPF;
1338 
1339 				if (dst_reg > BPF_REG_9) {
1340 					pr_err("verifier error\n");
1341 					return -EFAULT;
1342 				}
1343 				/*
1344 				 * Compute size of x86 insn and its target dest x86 register.
1345 				 * ex_handler_bpf() will use lower 8 bits to adjust
1346 				 * pt_regs->ip to jump over this x86 instruction
1347 				 * and upper bits to figure out which pt_regs to zero out.
1348 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1349 				 * of 4 bytes will be ignored and rbx will be zero inited.
1350 				 */
1351 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1352 			}
1353 			break;
1354 
1355 		case BPF_STX | BPF_ATOMIC | BPF_W:
1356 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1357 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1358 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1359 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1360 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1361 				u32 real_src_reg = src_reg;
1362 				u32 real_dst_reg = dst_reg;
1363 				u8 *branch_target;
1364 
1365 				/*
1366 				 * Can't be implemented with a single x86 insn.
1367 				 * Need to do a CMPXCHG loop.
1368 				 */
1369 
1370 				/* Will need RAX as a CMPXCHG operand so save R0 */
1371 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1372 				if (src_reg == BPF_REG_0)
1373 					real_src_reg = BPF_REG_AX;
1374 				if (dst_reg == BPF_REG_0)
1375 					real_dst_reg = BPF_REG_AX;
1376 
1377 				branch_target = prog;
1378 				/* Load old value */
1379 				emit_ldx(&prog, BPF_SIZE(insn->code),
1380 					 BPF_REG_0, real_dst_reg, insn->off);
1381 				/*
1382 				 * Perform the (commutative) operation locally,
1383 				 * put the result in the AUX_REG.
1384 				 */
1385 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1386 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1387 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1388 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1389 				/* Attempt to swap in new value */
1390 				err = emit_atomic(&prog, BPF_CMPXCHG,
1391 						  real_dst_reg, AUX_REG,
1392 						  insn->off,
1393 						  BPF_SIZE(insn->code));
1394 				if (WARN_ON(err))
1395 					return err;
1396 				/*
1397 				 * ZF tells us whether we won the race. If it's
1398 				 * cleared we need to try again.
1399 				 */
1400 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1401 				/* Return the pre-modification value */
1402 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1403 				/* Restore R0 after clobbering RAX */
1404 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1405 				break;
1406 			}
1407 
1408 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1409 					  insn->off, BPF_SIZE(insn->code));
1410 			if (err)
1411 				return err;
1412 			break;
1413 
1414 			/* call */
1415 		case BPF_JMP | BPF_CALL:
1416 			func = (u8 *) __bpf_call_base + imm32;
1417 			if (tail_call_reachable) {
1418 				EMIT3_off32(0x48, 0x8B, 0x85,
1419 					    -(bpf_prog->aux->stack_depth + 8));
1420 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1] + 7))
1421 					return -EINVAL;
1422 			} else {
1423 				if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1424 					return -EINVAL;
1425 			}
1426 			break;
1427 
1428 		case BPF_JMP | BPF_TAIL_CALL:
1429 			if (imm32)
1430 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1431 							  &prog, image + addrs[i - 1],
1432 							  callee_regs_used,
1433 							  bpf_prog->aux->stack_depth,
1434 							  ctx);
1435 			else
1436 				emit_bpf_tail_call_indirect(&prog,
1437 							    callee_regs_used,
1438 							    bpf_prog->aux->stack_depth,
1439 							    image + addrs[i - 1],
1440 							    ctx);
1441 			break;
1442 
1443 			/* cond jump */
1444 		case BPF_JMP | BPF_JEQ | BPF_X:
1445 		case BPF_JMP | BPF_JNE | BPF_X:
1446 		case BPF_JMP | BPF_JGT | BPF_X:
1447 		case BPF_JMP | BPF_JLT | BPF_X:
1448 		case BPF_JMP | BPF_JGE | BPF_X:
1449 		case BPF_JMP | BPF_JLE | BPF_X:
1450 		case BPF_JMP | BPF_JSGT | BPF_X:
1451 		case BPF_JMP | BPF_JSLT | BPF_X:
1452 		case BPF_JMP | BPF_JSGE | BPF_X:
1453 		case BPF_JMP | BPF_JSLE | BPF_X:
1454 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1455 		case BPF_JMP32 | BPF_JNE | BPF_X:
1456 		case BPF_JMP32 | BPF_JGT | BPF_X:
1457 		case BPF_JMP32 | BPF_JLT | BPF_X:
1458 		case BPF_JMP32 | BPF_JGE | BPF_X:
1459 		case BPF_JMP32 | BPF_JLE | BPF_X:
1460 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1461 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1462 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1463 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1464 			/* cmp dst_reg, src_reg */
1465 			maybe_emit_mod(&prog, dst_reg, src_reg,
1466 				       BPF_CLASS(insn->code) == BPF_JMP);
1467 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1468 			goto emit_cond_jmp;
1469 
1470 		case BPF_JMP | BPF_JSET | BPF_X:
1471 		case BPF_JMP32 | BPF_JSET | BPF_X:
1472 			/* test dst_reg, src_reg */
1473 			maybe_emit_mod(&prog, dst_reg, src_reg,
1474 				       BPF_CLASS(insn->code) == BPF_JMP);
1475 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1476 			goto emit_cond_jmp;
1477 
1478 		case BPF_JMP | BPF_JSET | BPF_K:
1479 		case BPF_JMP32 | BPF_JSET | BPF_K:
1480 			/* test dst_reg, imm32 */
1481 			maybe_emit_1mod(&prog, dst_reg,
1482 					BPF_CLASS(insn->code) == BPF_JMP);
1483 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1484 			goto emit_cond_jmp;
1485 
1486 		case BPF_JMP | BPF_JEQ | BPF_K:
1487 		case BPF_JMP | BPF_JNE | BPF_K:
1488 		case BPF_JMP | BPF_JGT | BPF_K:
1489 		case BPF_JMP | BPF_JLT | BPF_K:
1490 		case BPF_JMP | BPF_JGE | BPF_K:
1491 		case BPF_JMP | BPF_JLE | BPF_K:
1492 		case BPF_JMP | BPF_JSGT | BPF_K:
1493 		case BPF_JMP | BPF_JSLT | BPF_K:
1494 		case BPF_JMP | BPF_JSGE | BPF_K:
1495 		case BPF_JMP | BPF_JSLE | BPF_K:
1496 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1497 		case BPF_JMP32 | BPF_JNE | BPF_K:
1498 		case BPF_JMP32 | BPF_JGT | BPF_K:
1499 		case BPF_JMP32 | BPF_JLT | BPF_K:
1500 		case BPF_JMP32 | BPF_JGE | BPF_K:
1501 		case BPF_JMP32 | BPF_JLE | BPF_K:
1502 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1503 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1504 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1505 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1506 			/* test dst_reg, dst_reg to save one extra byte */
1507 			if (imm32 == 0) {
1508 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1509 					       BPF_CLASS(insn->code) == BPF_JMP);
1510 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1511 				goto emit_cond_jmp;
1512 			}
1513 
1514 			/* cmp dst_reg, imm8/32 */
1515 			maybe_emit_1mod(&prog, dst_reg,
1516 					BPF_CLASS(insn->code) == BPF_JMP);
1517 
1518 			if (is_imm8(imm32))
1519 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1520 			else
1521 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1522 
1523 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1524 			switch (BPF_OP(insn->code)) {
1525 			case BPF_JEQ:
1526 				jmp_cond = X86_JE;
1527 				break;
1528 			case BPF_JSET:
1529 			case BPF_JNE:
1530 				jmp_cond = X86_JNE;
1531 				break;
1532 			case BPF_JGT:
1533 				/* GT is unsigned '>', JA in x86 */
1534 				jmp_cond = X86_JA;
1535 				break;
1536 			case BPF_JLT:
1537 				/* LT is unsigned '<', JB in x86 */
1538 				jmp_cond = X86_JB;
1539 				break;
1540 			case BPF_JGE:
1541 				/* GE is unsigned '>=', JAE in x86 */
1542 				jmp_cond = X86_JAE;
1543 				break;
1544 			case BPF_JLE:
1545 				/* LE is unsigned '<=', JBE in x86 */
1546 				jmp_cond = X86_JBE;
1547 				break;
1548 			case BPF_JSGT:
1549 				/* Signed '>', GT in x86 */
1550 				jmp_cond = X86_JG;
1551 				break;
1552 			case BPF_JSLT:
1553 				/* Signed '<', LT in x86 */
1554 				jmp_cond = X86_JL;
1555 				break;
1556 			case BPF_JSGE:
1557 				/* Signed '>=', GE in x86 */
1558 				jmp_cond = X86_JGE;
1559 				break;
1560 			case BPF_JSLE:
1561 				/* Signed '<=', LE in x86 */
1562 				jmp_cond = X86_JLE;
1563 				break;
1564 			default: /* to silence GCC warning */
1565 				return -EFAULT;
1566 			}
1567 			jmp_offset = addrs[i + insn->off] - addrs[i];
1568 			if (is_imm8(jmp_offset)) {
1569 				if (jmp_padding) {
1570 					/* To keep the jmp_offset valid, the extra bytes are
1571 					 * padded before the jump insn, so we subtract the
1572 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1573 					 *
1574 					 * If the previous pass already emits an imm8
1575 					 * jmp_cond, then this BPF insn won't shrink, so
1576 					 * "nops" is 0.
1577 					 *
1578 					 * On the other hand, if the previous pass emits an
1579 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1580 					 * keep the image from shrinking further.
1581 					 *
1582 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1583 					 *     is 2 bytes, so the size difference is 4 bytes.
1584 					 */
1585 					nops = INSN_SZ_DIFF - 2;
1586 					if (nops != 0 && nops != 4) {
1587 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1588 						       nops);
1589 						return -EFAULT;
1590 					}
1591 					emit_nops(&prog, nops);
1592 				}
1593 				EMIT2(jmp_cond, jmp_offset);
1594 			} else if (is_simm32(jmp_offset)) {
1595 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1596 			} else {
1597 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1598 				return -EFAULT;
1599 			}
1600 
1601 			break;
1602 
1603 		case BPF_JMP | BPF_JA:
1604 			if (insn->off == -1)
1605 				/* -1 jmp instructions will always jump
1606 				 * backwards two bytes. Explicitly handling
1607 				 * this case avoids wasting too many passes
1608 				 * when there are long sequences of replaced
1609 				 * dead code.
1610 				 */
1611 				jmp_offset = -2;
1612 			else
1613 				jmp_offset = addrs[i + insn->off] - addrs[i];
1614 
1615 			if (!jmp_offset) {
1616 				/*
1617 				 * If jmp_padding is enabled, the extra nops will
1618 				 * be inserted. Otherwise, optimize out nop jumps.
1619 				 */
1620 				if (jmp_padding) {
1621 					/* There are 3 possible conditions.
1622 					 * (1) This BPF_JA is already optimized out in
1623 					 *     the previous run, so there is no need
1624 					 *     to pad any extra byte (0 byte).
1625 					 * (2) The previous pass emits an imm8 jmp,
1626 					 *     so we pad 2 bytes to match the previous
1627 					 *     insn size.
1628 					 * (3) Similarly, the previous pass emits an
1629 					 *     imm32 jmp, and 5 bytes is padded.
1630 					 */
1631 					nops = INSN_SZ_DIFF;
1632 					if (nops != 0 && nops != 2 && nops != 5) {
1633 						pr_err("unexpected nop jump padding: %d bytes\n",
1634 						       nops);
1635 						return -EFAULT;
1636 					}
1637 					emit_nops(&prog, nops);
1638 				}
1639 				break;
1640 			}
1641 emit_jmp:
1642 			if (is_imm8(jmp_offset)) {
1643 				if (jmp_padding) {
1644 					/* To avoid breaking jmp_offset, the extra bytes
1645 					 * are padded before the actual jmp insn, so
1646 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1647 					 *
1648 					 * If the previous pass already emits an imm8
1649 					 * jmp, there is nothing to pad (0 byte).
1650 					 *
1651 					 * If it emits an imm32 jmp (5 bytes) previously
1652 					 * and now an imm8 jmp (2 bytes), then we pad
1653 					 * (5 - 2 = 3) bytes to stop the image from
1654 					 * shrinking further.
1655 					 */
1656 					nops = INSN_SZ_DIFF - 2;
1657 					if (nops != 0 && nops != 3) {
1658 						pr_err("unexpected jump padding: %d bytes\n",
1659 						       nops);
1660 						return -EFAULT;
1661 					}
1662 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1663 				}
1664 				EMIT2(0xEB, jmp_offset);
1665 			} else if (is_simm32(jmp_offset)) {
1666 				EMIT1_off32(0xE9, jmp_offset);
1667 			} else {
1668 				pr_err("jmp gen bug %llx\n", jmp_offset);
1669 				return -EFAULT;
1670 			}
1671 			break;
1672 
1673 		case BPF_JMP | BPF_EXIT:
1674 			if (seen_exit) {
1675 				jmp_offset = ctx->cleanup_addr - addrs[i];
1676 				goto emit_jmp;
1677 			}
1678 			seen_exit = true;
1679 			/* Update cleanup_addr */
1680 			ctx->cleanup_addr = proglen;
1681 			pop_callee_regs(&prog, callee_regs_used);
1682 			EMIT1(0xC9);         /* leave */
1683 			EMIT1(0xC3);         /* ret */
1684 			break;
1685 
1686 		default:
1687 			/*
1688 			 * By design x86-64 JIT should support all BPF instructions.
1689 			 * This error will be seen if new instruction was added
1690 			 * to the interpreter, but not to the JIT, or if there is
1691 			 * junk in bpf_prog.
1692 			 */
1693 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1694 			return -EINVAL;
1695 		}
1696 
1697 		ilen = prog - temp;
1698 		if (ilen > BPF_MAX_INSN_SIZE) {
1699 			pr_err("bpf_jit: fatal insn size error\n");
1700 			return -EFAULT;
1701 		}
1702 
1703 		if (image) {
1704 			/*
1705 			 * When populating the image, assert that:
1706 			 *
1707 			 *  i) We do not write beyond the allocated space, and
1708 			 * ii) addrs[i] did not change from the prior run, in order
1709 			 *     to validate assumptions made for computing branch
1710 			 *     displacements.
1711 			 */
1712 			if (unlikely(proglen + ilen > oldproglen ||
1713 				     proglen + ilen != addrs[i])) {
1714 				pr_err("bpf_jit: fatal error\n");
1715 				return -EFAULT;
1716 			}
1717 			memcpy(rw_image + proglen, temp, ilen);
1718 		}
1719 		proglen += ilen;
1720 		addrs[i] = proglen;
1721 		prog = temp;
1722 	}
1723 
1724 	if (image && excnt != bpf_prog->aux->num_exentries) {
1725 		pr_err("extable is not populated\n");
1726 		return -EFAULT;
1727 	}
1728 	return proglen;
1729 }
1730 
1731 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1732 		      int stack_size)
1733 {
1734 	int i;
1735 	/* Store function arguments to stack.
1736 	 * For a function that accepts two pointers the sequence will be:
1737 	 * mov QWORD PTR [rbp-0x10],rdi
1738 	 * mov QWORD PTR [rbp-0x8],rsi
1739 	 */
1740 	for (i = 0; i < min(nr_args, 6); i++)
1741 		emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1742 			 BPF_REG_FP,
1743 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1744 			 -(stack_size - i * 8));
1745 }
1746 
1747 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1748 			 int stack_size)
1749 {
1750 	int i;
1751 
1752 	/* Restore function arguments from stack.
1753 	 * For a function that accepts two pointers the sequence will be:
1754 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1755 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1756 	 */
1757 	for (i = 0; i < min(nr_args, 6); i++)
1758 		emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1759 			 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1760 			 BPF_REG_FP,
1761 			 -(stack_size - i * 8));
1762 }
1763 
1764 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1765 			   struct bpf_prog *p, int stack_size, bool save_ret)
1766 {
1767 	u8 *prog = *pprog;
1768 	u8 *jmp_insn;
1769 
1770 	/* arg1: mov rdi, progs[i] */
1771 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1772 	if (emit_call(&prog,
1773 		      p->aux->sleepable ? __bpf_prog_enter_sleepable :
1774 		      __bpf_prog_enter, prog))
1775 			return -EINVAL;
1776 	/* remember prog start time returned by __bpf_prog_enter */
1777 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1778 
1779 	/* if (__bpf_prog_enter*(prog) == 0)
1780 	 *	goto skip_exec_of_prog;
1781 	 */
1782 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
1783 	/* emit 2 nops that will be replaced with JE insn */
1784 	jmp_insn = prog;
1785 	emit_nops(&prog, 2);
1786 
1787 	/* arg1: lea rdi, [rbp - stack_size] */
1788 	EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1789 	/* arg2: progs[i]->insnsi for interpreter */
1790 	if (!p->jited)
1791 		emit_mov_imm64(&prog, BPF_REG_2,
1792 			       (long) p->insnsi >> 32,
1793 			       (u32) (long) p->insnsi);
1794 	/* call JITed bpf program or interpreter */
1795 	if (emit_call(&prog, p->bpf_func, prog))
1796 		return -EINVAL;
1797 
1798 	/*
1799 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1800 	 * of the previous call which is then passed on the stack to
1801 	 * the next BPF program.
1802 	 *
1803 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
1804 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
1805 	 */
1806 	if (save_ret)
1807 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1808 
1809 	/* replace 2 nops with JE insn, since jmp target is known */
1810 	jmp_insn[0] = X86_JE;
1811 	jmp_insn[1] = prog - jmp_insn - 2;
1812 
1813 	/* arg1: mov rdi, progs[i] */
1814 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
1815 	/* arg2: mov rsi, rbx <- start time in nsec */
1816 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1817 	if (emit_call(&prog,
1818 		      p->aux->sleepable ? __bpf_prog_exit_sleepable :
1819 		      __bpf_prog_exit, prog))
1820 			return -EINVAL;
1821 
1822 	*pprog = prog;
1823 	return 0;
1824 }
1825 
1826 static void emit_align(u8 **pprog, u32 align)
1827 {
1828 	u8 *target, *prog = *pprog;
1829 
1830 	target = PTR_ALIGN(prog, align);
1831 	if (target != prog)
1832 		emit_nops(&prog, target - prog);
1833 
1834 	*pprog = prog;
1835 }
1836 
1837 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1838 {
1839 	u8 *prog = *pprog;
1840 	s64 offset;
1841 
1842 	offset = func - (ip + 2 + 4);
1843 	if (!is_simm32(offset)) {
1844 		pr_err("Target %p is out of range\n", func);
1845 		return -EINVAL;
1846 	}
1847 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1848 	*pprog = prog;
1849 	return 0;
1850 }
1851 
1852 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1853 		      struct bpf_tramp_progs *tp, int stack_size,
1854 		      bool save_ret)
1855 {
1856 	int i;
1857 	u8 *prog = *pprog;
1858 
1859 	for (i = 0; i < tp->nr_progs; i++) {
1860 		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size,
1861 				    save_ret))
1862 			return -EINVAL;
1863 	}
1864 	*pprog = prog;
1865 	return 0;
1866 }
1867 
1868 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1869 			      struct bpf_tramp_progs *tp, int stack_size,
1870 			      u8 **branches)
1871 {
1872 	u8 *prog = *pprog;
1873 	int i;
1874 
1875 	/* The first fmod_ret program will receive a garbage return value.
1876 	 * Set this to 0 to avoid confusing the program.
1877 	 */
1878 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1879 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1880 	for (i = 0; i < tp->nr_progs; i++) {
1881 		if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1882 			return -EINVAL;
1883 
1884 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
1885 		 * if (*(u64 *)(rbp - 8) !=  0)
1886 		 *	goto do_fexit;
1887 		 */
1888 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
1889 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1890 
1891 		/* Save the location of the branch and Generate 6 nops
1892 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
1893 		 * are replaced with a conditional jump once do_fexit (i.e. the
1894 		 * start of the fexit invocation) is finalized.
1895 		 */
1896 		branches[i] = prog;
1897 		emit_nops(&prog, 4 + 2);
1898 	}
1899 
1900 	*pprog = prog;
1901 	return 0;
1902 }
1903 
1904 static bool is_valid_bpf_tramp_flags(unsigned int flags)
1905 {
1906 	if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1907 	    (flags & BPF_TRAMP_F_SKIP_FRAME))
1908 		return false;
1909 
1910 	/*
1911 	 * BPF_TRAMP_F_RET_FENTRY_RET is only used by bpf_struct_ops,
1912 	 * and it must be used alone.
1913 	 */
1914 	if ((flags & BPF_TRAMP_F_RET_FENTRY_RET) &&
1915 	    (flags & ~BPF_TRAMP_F_RET_FENTRY_RET))
1916 		return false;
1917 
1918 	return true;
1919 }
1920 
1921 /* Example:
1922  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1923  * its 'struct btf_func_model' will be nr_args=2
1924  * The assembly code when eth_type_trans is executing after trampoline:
1925  *
1926  * push rbp
1927  * mov rbp, rsp
1928  * sub rsp, 16                     // space for skb and dev
1929  * push rbx                        // temp regs to pass start time
1930  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
1931  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
1932  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1933  * mov rbx, rax                    // remember start time in bpf stats are enabled
1934  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
1935  * call addr_of_jited_FENTRY_prog
1936  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1937  * mov rsi, rbx                    // prog start time
1938  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1939  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
1940  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
1941  * pop rbx
1942  * leave
1943  * ret
1944  *
1945  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1946  * replaced with 'call generated_bpf_trampoline'. When it returns
1947  * eth_type_trans will continue executing with original skb and dev pointers.
1948  *
1949  * The assembly code when eth_type_trans is called from trampoline:
1950  *
1951  * push rbp
1952  * mov rbp, rsp
1953  * sub rsp, 24                     // space for skb, dev, return value
1954  * push rbx                        // temp regs to pass start time
1955  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
1956  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
1957  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1958  * mov rbx, rax                    // remember start time if bpf stats are enabled
1959  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1960  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
1961  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1962  * mov rsi, rbx                    // prog start time
1963  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1964  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
1965  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
1966  * call eth_type_trans+5           // execute body of eth_type_trans
1967  * mov qword ptr [rbp - 8], rax    // save return value
1968  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
1969  * mov rbx, rax                    // remember start time in bpf stats are enabled
1970  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
1971  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
1972  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
1973  * mov rsi, rbx                    // prog start time
1974  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
1975  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
1976  * pop rbx
1977  * leave
1978  * add rsp, 8                      // skip eth_type_trans's frame
1979  * ret                             // return to its caller
1980  */
1981 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
1982 				const struct btf_func_model *m, u32 flags,
1983 				struct bpf_tramp_progs *tprogs,
1984 				void *orig_call)
1985 {
1986 	int ret, i, nr_args = m->nr_args;
1987 	int regs_off, ip_off, args_off, stack_size = nr_args * 8;
1988 	struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
1989 	struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
1990 	struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
1991 	u8 **branches = NULL;
1992 	u8 *prog;
1993 	bool save_ret;
1994 
1995 	/* x86-64 supports up to 6 arguments. 7+ can be added in the future */
1996 	if (nr_args > 6)
1997 		return -ENOTSUPP;
1998 
1999 	if (!is_valid_bpf_tramp_flags(flags))
2000 		return -EINVAL;
2001 
2002 	/* Generated trampoline stack layout:
2003 	 *
2004 	 * RBP + 8         [ return address  ]
2005 	 * RBP + 0         [ RBP             ]
2006 	 *
2007 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2008 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2009 	 *
2010 	 *                 [ reg_argN        ]  always
2011 	 *                 [ ...             ]
2012 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2013 	 *
2014 	 * RBP - args_off  [ args count      ]  always
2015 	 *
2016 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2017 	 */
2018 
2019 	/* room for return value of orig_call or fentry prog */
2020 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2021 	if (save_ret)
2022 		stack_size += 8;
2023 
2024 	regs_off = stack_size;
2025 
2026 	/* args count  */
2027 	stack_size += 8;
2028 	args_off = stack_size;
2029 
2030 	if (flags & BPF_TRAMP_F_IP_ARG)
2031 		stack_size += 8; /* room for IP address argument */
2032 
2033 	ip_off = stack_size;
2034 
2035 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2036 		/* skip patched call instruction and point orig_call to actual
2037 		 * body of the kernel function.
2038 		 */
2039 		if (is_endbr(*(u32 *)orig_call))
2040 			orig_call += ENDBR_INSN_SIZE;
2041 		orig_call += X86_PATCH_SIZE;
2042 	}
2043 
2044 	prog = image;
2045 
2046 	EMIT_ENDBR();
2047 	EMIT1(0x55);		 /* push rbp */
2048 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2049 	EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
2050 	EMIT1(0x53);		 /* push rbx */
2051 
2052 	/* Store number of arguments of the traced function:
2053 	 *   mov rax, nr_args
2054 	 *   mov QWORD PTR [rbp - args_off], rax
2055 	 */
2056 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_args);
2057 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -args_off);
2058 
2059 	if (flags & BPF_TRAMP_F_IP_ARG) {
2060 		/* Store IP address of the traced function:
2061 		 * mov rax, QWORD PTR [rbp + 8]
2062 		 * sub rax, X86_PATCH_SIZE
2063 		 * mov QWORD PTR [rbp - ip_off], rax
2064 		 */
2065 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, 8);
2066 		EMIT4(0x48, 0x83, 0xe8, X86_PATCH_SIZE);
2067 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2068 	}
2069 
2070 	save_regs(m, &prog, nr_args, regs_off);
2071 
2072 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2073 		/* arg1: mov rdi, im */
2074 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2075 		if (emit_call(&prog, __bpf_tramp_enter, prog)) {
2076 			ret = -EINVAL;
2077 			goto cleanup;
2078 		}
2079 	}
2080 
2081 	if (fentry->nr_progs)
2082 		if (invoke_bpf(m, &prog, fentry, regs_off,
2083 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2084 			return -EINVAL;
2085 
2086 	if (fmod_ret->nr_progs) {
2087 		branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
2088 				   GFP_KERNEL);
2089 		if (!branches)
2090 			return -ENOMEM;
2091 
2092 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2093 				       branches)) {
2094 			ret = -EINVAL;
2095 			goto cleanup;
2096 		}
2097 	}
2098 
2099 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2100 		restore_regs(m, &prog, nr_args, regs_off);
2101 
2102 		/* call original function */
2103 		if (emit_call(&prog, orig_call, prog)) {
2104 			ret = -EINVAL;
2105 			goto cleanup;
2106 		}
2107 		/* remember return value in a stack for bpf prog to access */
2108 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2109 		im->ip_after_call = prog;
2110 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2111 		prog += X86_PATCH_SIZE;
2112 	}
2113 
2114 	if (fmod_ret->nr_progs) {
2115 		/* From Intel 64 and IA-32 Architectures Optimization
2116 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2117 		 * Coding Rule 11: All branch targets should be 16-byte
2118 		 * aligned.
2119 		 */
2120 		emit_align(&prog, 16);
2121 		/* Update the branches saved in invoke_bpf_mod_ret with the
2122 		 * aligned address of do_fexit.
2123 		 */
2124 		for (i = 0; i < fmod_ret->nr_progs; i++)
2125 			emit_cond_near_jump(&branches[i], prog, branches[i],
2126 					    X86_JNE);
2127 	}
2128 
2129 	if (fexit->nr_progs)
2130 		if (invoke_bpf(m, &prog, fexit, regs_off, false)) {
2131 			ret = -EINVAL;
2132 			goto cleanup;
2133 		}
2134 
2135 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2136 		restore_regs(m, &prog, nr_args, regs_off);
2137 
2138 	/* This needs to be done regardless. If there were fmod_ret programs,
2139 	 * the return value is only updated on the stack and still needs to be
2140 	 * restored to R0.
2141 	 */
2142 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2143 		im->ip_epilogue = prog;
2144 		/* arg1: mov rdi, im */
2145 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2146 		if (emit_call(&prog, __bpf_tramp_exit, prog)) {
2147 			ret = -EINVAL;
2148 			goto cleanup;
2149 		}
2150 	}
2151 	/* restore return value of orig_call or fentry prog back into RAX */
2152 	if (save_ret)
2153 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2154 
2155 	EMIT1(0x5B); /* pop rbx */
2156 	EMIT1(0xC9); /* leave */
2157 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2158 		/* skip our return address and return to parent */
2159 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2160 	EMIT1(0xC3); /* ret */
2161 	/* Make sure the trampoline generation logic doesn't overflow */
2162 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2163 		ret = -EFAULT;
2164 		goto cleanup;
2165 	}
2166 	ret = prog - (u8 *)image;
2167 
2168 cleanup:
2169 	kfree(branches);
2170 	return ret;
2171 }
2172 
2173 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
2174 {
2175 	u8 *jg_reloc, *prog = *pprog;
2176 	int pivot, err, jg_bytes = 1;
2177 	s64 jg_offset;
2178 
2179 	if (a == b) {
2180 		/* Leaf node of recursion, i.e. not a range of indices
2181 		 * anymore.
2182 		 */
2183 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2184 		if (!is_simm32(progs[a]))
2185 			return -1;
2186 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2187 			    progs[a]);
2188 		err = emit_cond_near_jump(&prog,	/* je func */
2189 					  (void *)progs[a], prog,
2190 					  X86_JE);
2191 		if (err)
2192 			return err;
2193 
2194 		emit_indirect_jump(&prog, 2 /* rdx */, prog);
2195 
2196 		*pprog = prog;
2197 		return 0;
2198 	}
2199 
2200 	/* Not a leaf node, so we pivot, and recursively descend into
2201 	 * the lower and upper ranges.
2202 	 */
2203 	pivot = (b - a) / 2;
2204 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2205 	if (!is_simm32(progs[a + pivot]))
2206 		return -1;
2207 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2208 
2209 	if (pivot > 2) {				/* jg upper_part */
2210 		/* Require near jump. */
2211 		jg_bytes = 4;
2212 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2213 	} else {
2214 		EMIT2(X86_JG, 0);
2215 	}
2216 	jg_reloc = prog;
2217 
2218 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2219 				  progs);
2220 	if (err)
2221 		return err;
2222 
2223 	/* From Intel 64 and IA-32 Architectures Optimization
2224 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2225 	 * Coding Rule 11: All branch targets should be 16-byte
2226 	 * aligned.
2227 	 */
2228 	emit_align(&prog, 16);
2229 	jg_offset = prog - jg_reloc;
2230 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2231 
2232 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2233 				  b, progs);
2234 	if (err)
2235 		return err;
2236 
2237 	*pprog = prog;
2238 	return 0;
2239 }
2240 
2241 static int cmp_ips(const void *a, const void *b)
2242 {
2243 	const s64 *ipa = a;
2244 	const s64 *ipb = b;
2245 
2246 	if (*ipa > *ipb)
2247 		return 1;
2248 	if (*ipa < *ipb)
2249 		return -1;
2250 	return 0;
2251 }
2252 
2253 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
2254 {
2255 	u8 *prog = image;
2256 
2257 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2258 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
2259 }
2260 
2261 struct x64_jit_data {
2262 	struct bpf_binary_header *rw_header;
2263 	struct bpf_binary_header *header;
2264 	int *addrs;
2265 	u8 *image;
2266 	int proglen;
2267 	struct jit_context ctx;
2268 };
2269 
2270 #define MAX_PASSES 20
2271 #define PADDING_PASSES (MAX_PASSES - 5)
2272 
2273 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2274 {
2275 	struct bpf_binary_header *rw_header = NULL;
2276 	struct bpf_binary_header *header = NULL;
2277 	struct bpf_prog *tmp, *orig_prog = prog;
2278 	struct x64_jit_data *jit_data;
2279 	int proglen, oldproglen = 0;
2280 	struct jit_context ctx = {};
2281 	bool tmp_blinded = false;
2282 	bool extra_pass = false;
2283 	bool padding = false;
2284 	u8 *rw_image = NULL;
2285 	u8 *image = NULL;
2286 	int *addrs;
2287 	int pass;
2288 	int i;
2289 
2290 	if (!prog->jit_requested)
2291 		return orig_prog;
2292 
2293 	tmp = bpf_jit_blind_constants(prog);
2294 	/*
2295 	 * If blinding was requested and we failed during blinding,
2296 	 * we must fall back to the interpreter.
2297 	 */
2298 	if (IS_ERR(tmp))
2299 		return orig_prog;
2300 	if (tmp != prog) {
2301 		tmp_blinded = true;
2302 		prog = tmp;
2303 	}
2304 
2305 	jit_data = prog->aux->jit_data;
2306 	if (!jit_data) {
2307 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2308 		if (!jit_data) {
2309 			prog = orig_prog;
2310 			goto out;
2311 		}
2312 		prog->aux->jit_data = jit_data;
2313 	}
2314 	addrs = jit_data->addrs;
2315 	if (addrs) {
2316 		ctx = jit_data->ctx;
2317 		oldproglen = jit_data->proglen;
2318 		image = jit_data->image;
2319 		header = jit_data->header;
2320 		rw_header = jit_data->rw_header;
2321 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2322 		extra_pass = true;
2323 		padding = true;
2324 		goto skip_init_addrs;
2325 	}
2326 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2327 	if (!addrs) {
2328 		prog = orig_prog;
2329 		goto out_addrs;
2330 	}
2331 
2332 	/*
2333 	 * Before first pass, make a rough estimation of addrs[]
2334 	 * each BPF instruction is translated to less than 64 bytes
2335 	 */
2336 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2337 		proglen += 64;
2338 		addrs[i] = proglen;
2339 	}
2340 	ctx.cleanup_addr = proglen;
2341 skip_init_addrs:
2342 
2343 	/*
2344 	 * JITed image shrinks with every pass and the loop iterates
2345 	 * until the image stops shrinking. Very large BPF programs
2346 	 * may converge on the last pass. In such case do one more
2347 	 * pass to emit the final image.
2348 	 */
2349 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2350 		if (!padding && pass >= PADDING_PASSES)
2351 			padding = true;
2352 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2353 		if (proglen <= 0) {
2354 out_image:
2355 			image = NULL;
2356 			if (header) {
2357 				bpf_arch_text_copy(&header->size, &rw_header->size,
2358 						   sizeof(rw_header->size));
2359 				bpf_jit_binary_pack_free(header, rw_header);
2360 			}
2361 			/* Fall back to interpreter mode */
2362 			prog = orig_prog;
2363 			if (extra_pass) {
2364 				prog->bpf_func = NULL;
2365 				prog->jited = 0;
2366 				prog->jited_len = 0;
2367 			}
2368 			goto out_addrs;
2369 		}
2370 		if (image) {
2371 			if (proglen != oldproglen) {
2372 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2373 				       proglen, oldproglen);
2374 				goto out_image;
2375 			}
2376 			break;
2377 		}
2378 		if (proglen == oldproglen) {
2379 			/*
2380 			 * The number of entries in extable is the number of BPF_LDX
2381 			 * insns that access kernel memory via "pointer to BTF type".
2382 			 * The verifier changed their opcode from LDX|MEM|size
2383 			 * to LDX|PROBE_MEM|size to make JITing easier.
2384 			 */
2385 			u32 align = __alignof__(struct exception_table_entry);
2386 			u32 extable_size = prog->aux->num_exentries *
2387 				sizeof(struct exception_table_entry);
2388 
2389 			/* allocate module memory for x86 insns and extable */
2390 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2391 							   &image, align, &rw_header, &rw_image,
2392 							   jit_fill_hole);
2393 			if (!header) {
2394 				prog = orig_prog;
2395 				goto out_addrs;
2396 			}
2397 			prog->aux->extable = (void *) image + roundup(proglen, align);
2398 		}
2399 		oldproglen = proglen;
2400 		cond_resched();
2401 	}
2402 
2403 	if (bpf_jit_enable > 1)
2404 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
2405 
2406 	if (image) {
2407 		if (!prog->is_func || extra_pass) {
2408 			/*
2409 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2410 			 *   1) header is not pointing to proper module memory;
2411 			 *   2) the arch doesn't support bpf_arch_text_copy().
2412 			 *
2413 			 * Both cases are serious bugs and justify WARN_ON.
2414 			 */
2415 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2416 				/* header has been freed */
2417 				header = NULL;
2418 				goto out_image;
2419 			}
2420 
2421 			bpf_tail_call_direct_fixup(prog);
2422 		} else {
2423 			jit_data->addrs = addrs;
2424 			jit_data->ctx = ctx;
2425 			jit_data->proglen = proglen;
2426 			jit_data->image = image;
2427 			jit_data->header = header;
2428 			jit_data->rw_header = rw_header;
2429 		}
2430 		prog->bpf_func = (void *)image;
2431 		prog->jited = 1;
2432 		prog->jited_len = proglen;
2433 	} else {
2434 		prog = orig_prog;
2435 	}
2436 
2437 	if (!image || !prog->is_func || extra_pass) {
2438 		if (image)
2439 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2440 out_addrs:
2441 		kvfree(addrs);
2442 		kfree(jit_data);
2443 		prog->aux->jit_data = NULL;
2444 	}
2445 out:
2446 	if (tmp_blinded)
2447 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2448 					   tmp : orig_prog);
2449 	return prog;
2450 }
2451 
2452 bool bpf_jit_supports_kfunc_call(void)
2453 {
2454 	return true;
2455 }
2456 
2457 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2458 {
2459 	if (text_poke_copy(dst, src, len) == NULL)
2460 		return ERR_PTR(-EINVAL);
2461 	return dst;
2462 }
2463