xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision 7f1005dd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * BPF JIT compiler
4  *
5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6  * Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
7  */
8 #include <linux/netdevice.h>
9 #include <linux/filter.h>
10 #include <linux/if_vlan.h>
11 #include <linux/bpf.h>
12 #include <linux/memory.h>
13 #include <linux/sort.h>
14 #include <asm/extable.h>
15 #include <asm/ftrace.h>
16 #include <asm/set_memory.h>
17 #include <asm/nospec-branch.h>
18 #include <asm/text-patching.h>
19 
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
21 {
22 	if (len == 1)
23 		*ptr = bytes;
24 	else if (len == 2)
25 		*(u16 *)ptr = bytes;
26 	else {
27 		*(u32 *)ptr = bytes;
28 		barrier();
29 	}
30 	return ptr + len;
31 }
32 
33 #define EMIT(bytes, len) \
34 	do { prog = emit_code(prog, bytes, len); } while (0)
35 
36 #define EMIT1(b1)		EMIT(b1, 1)
37 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
38 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
39 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
40 
41 #define EMIT1_off32(b1, off) \
42 	do { EMIT1(b1); EMIT(off, 4); } while (0)
43 #define EMIT2_off32(b1, b2, off) \
44 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
45 #define EMIT3_off32(b1, b2, b3, off) \
46 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
47 #define EMIT4_off32(b1, b2, b3, b4, off) \
48 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
49 
50 #ifdef CONFIG_X86_KERNEL_IBT
51 #define EMIT_ENDBR()	EMIT(gen_endbr(), 4)
52 #else
53 #define EMIT_ENDBR()
54 #endif
55 
56 static bool is_imm8(int value)
57 {
58 	return value <= 127 && value >= -128;
59 }
60 
61 static bool is_simm32(s64 value)
62 {
63 	return value == (s64)(s32)value;
64 }
65 
66 static bool is_uimm32(u64 value)
67 {
68 	return value == (u64)(u32)value;
69 }
70 
71 /* mov dst, src */
72 #define EMIT_mov(DST, SRC)								 \
73 	do {										 \
74 		if (DST != SRC)								 \
75 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
76 	} while (0)
77 
78 static int bpf_size_to_x86_bytes(int bpf_size)
79 {
80 	if (bpf_size == BPF_W)
81 		return 4;
82 	else if (bpf_size == BPF_H)
83 		return 2;
84 	else if (bpf_size == BPF_B)
85 		return 1;
86 	else if (bpf_size == BPF_DW)
87 		return 4; /* imm32 */
88 	else
89 		return 0;
90 }
91 
92 /*
93  * List of x86 cond jumps opcodes (. + s8)
94  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
95  */
96 #define X86_JB  0x72
97 #define X86_JAE 0x73
98 #define X86_JE  0x74
99 #define X86_JNE 0x75
100 #define X86_JBE 0x76
101 #define X86_JA  0x77
102 #define X86_JL  0x7C
103 #define X86_JGE 0x7D
104 #define X86_JLE 0x7E
105 #define X86_JG  0x7F
106 
107 /* Pick a register outside of BPF range for JIT internal work */
108 #define AUX_REG (MAX_BPF_JIT_REG + 1)
109 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
110 
111 /*
112  * The following table maps BPF registers to x86-64 registers.
113  *
114  * x86-64 register R12 is unused, since if used as base address
115  * register in load/store instructions, it always needs an
116  * extra byte of encoding and is callee saved.
117  *
118  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
119  * trampoline. x86-64 register R10 is used for blinding (if enabled).
120  */
121 static const int reg2hex[] = {
122 	[BPF_REG_0] = 0,  /* RAX */
123 	[BPF_REG_1] = 7,  /* RDI */
124 	[BPF_REG_2] = 6,  /* RSI */
125 	[BPF_REG_3] = 2,  /* RDX */
126 	[BPF_REG_4] = 1,  /* RCX */
127 	[BPF_REG_5] = 0,  /* R8  */
128 	[BPF_REG_6] = 3,  /* RBX callee saved */
129 	[BPF_REG_7] = 5,  /* R13 callee saved */
130 	[BPF_REG_8] = 6,  /* R14 callee saved */
131 	[BPF_REG_9] = 7,  /* R15 callee saved */
132 	[BPF_REG_FP] = 5, /* RBP readonly */
133 	[BPF_REG_AX] = 2, /* R10 temp register */
134 	[AUX_REG] = 3,    /* R11 temp register */
135 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
136 };
137 
138 static const int reg2pt_regs[] = {
139 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
140 	[BPF_REG_1] = offsetof(struct pt_regs, di),
141 	[BPF_REG_2] = offsetof(struct pt_regs, si),
142 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
143 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
144 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
145 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
146 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
147 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
148 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
149 };
150 
151 /*
152  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
153  * which need extra byte of encoding.
154  * rax,rcx,...,rbp have simpler encoding
155  */
156 static bool is_ereg(u32 reg)
157 {
158 	return (1 << reg) & (BIT(BPF_REG_5) |
159 			     BIT(AUX_REG) |
160 			     BIT(BPF_REG_7) |
161 			     BIT(BPF_REG_8) |
162 			     BIT(BPF_REG_9) |
163 			     BIT(X86_REG_R9) |
164 			     BIT(BPF_REG_AX));
165 }
166 
167 /*
168  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
169  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
170  * of encoding. al,cl,dl,bl have simpler encoding.
171  */
172 static bool is_ereg_8l(u32 reg)
173 {
174 	return is_ereg(reg) ||
175 	    (1 << reg) & (BIT(BPF_REG_1) |
176 			  BIT(BPF_REG_2) |
177 			  BIT(BPF_REG_FP));
178 }
179 
180 static bool is_axreg(u32 reg)
181 {
182 	return reg == BPF_REG_0;
183 }
184 
185 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
186 static u8 add_1mod(u8 byte, u32 reg)
187 {
188 	if (is_ereg(reg))
189 		byte |= 1;
190 	return byte;
191 }
192 
193 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
194 {
195 	if (is_ereg(r1))
196 		byte |= 1;
197 	if (is_ereg(r2))
198 		byte |= 4;
199 	return byte;
200 }
201 
202 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
203 static u8 add_1reg(u8 byte, u32 dst_reg)
204 {
205 	return byte + reg2hex[dst_reg];
206 }
207 
208 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
209 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
210 {
211 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
212 }
213 
214 /* Some 1-byte opcodes for binary ALU operations */
215 static u8 simple_alu_opcodes[] = {
216 	[BPF_ADD] = 0x01,
217 	[BPF_SUB] = 0x29,
218 	[BPF_AND] = 0x21,
219 	[BPF_OR] = 0x09,
220 	[BPF_XOR] = 0x31,
221 	[BPF_LSH] = 0xE0,
222 	[BPF_RSH] = 0xE8,
223 	[BPF_ARSH] = 0xF8,
224 };
225 
226 static void jit_fill_hole(void *area, unsigned int size)
227 {
228 	/* Fill whole space with INT3 instructions */
229 	memset(area, 0xcc, size);
230 }
231 
232 int bpf_arch_text_invalidate(void *dst, size_t len)
233 {
234 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
235 }
236 
237 struct jit_context {
238 	int cleanup_addr; /* Epilogue code offset */
239 
240 	/*
241 	 * Program specific offsets of labels in the code; these rely on the
242 	 * JIT doing at least 2 passes, recording the position on the first
243 	 * pass, only to generate the correct offset on the second pass.
244 	 */
245 	int tail_call_direct_label;
246 	int tail_call_indirect_label;
247 };
248 
249 /* Maximum number of bytes emitted while JITing one eBPF insn */
250 #define BPF_MAX_INSN_SIZE	128
251 #define BPF_INSN_SAFETY		64
252 
253 /* Number of bytes emit_patch() needs to generate instructions */
254 #define X86_PATCH_SIZE		5
255 /* Number of bytes that will be skipped on tailcall */
256 #define X86_TAIL_CALL_OFFSET	(11 + ENDBR_INSN_SIZE)
257 
258 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
259 {
260 	u8 *prog = *pprog;
261 
262 	if (callee_regs_used[0])
263 		EMIT1(0x53);         /* push rbx */
264 	if (callee_regs_used[1])
265 		EMIT2(0x41, 0x55);   /* push r13 */
266 	if (callee_regs_used[2])
267 		EMIT2(0x41, 0x56);   /* push r14 */
268 	if (callee_regs_used[3])
269 		EMIT2(0x41, 0x57);   /* push r15 */
270 	*pprog = prog;
271 }
272 
273 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
274 {
275 	u8 *prog = *pprog;
276 
277 	if (callee_regs_used[3])
278 		EMIT2(0x41, 0x5F);   /* pop r15 */
279 	if (callee_regs_used[2])
280 		EMIT2(0x41, 0x5E);   /* pop r14 */
281 	if (callee_regs_used[1])
282 		EMIT2(0x41, 0x5D);   /* pop r13 */
283 	if (callee_regs_used[0])
284 		EMIT1(0x5B);         /* pop rbx */
285 	*pprog = prog;
286 }
287 
288 /*
289  * Emit x86-64 prologue code for BPF program.
290  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
291  * while jumping to another program
292  */
293 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
294 			  bool tail_call_reachable, bool is_subprog)
295 {
296 	u8 *prog = *pprog;
297 
298 	/* BPF trampoline can be made to work without these nops,
299 	 * but let's waste 5 bytes for now and optimize later
300 	 */
301 	EMIT_ENDBR();
302 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
303 	prog += X86_PATCH_SIZE;
304 	if (!ebpf_from_cbpf) {
305 		if (tail_call_reachable && !is_subprog)
306 			EMIT2(0x31, 0xC0); /* xor eax, eax */
307 		else
308 			EMIT2(0x66, 0x90); /* nop2 */
309 	}
310 	EMIT1(0x55);             /* push rbp */
311 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
312 
313 	/* X86_TAIL_CALL_OFFSET is here */
314 	EMIT_ENDBR();
315 
316 	/* sub rsp, rounded_stack_depth */
317 	if (stack_depth)
318 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
319 	if (tail_call_reachable)
320 		EMIT1(0x50);         /* push rax */
321 	*pprog = prog;
322 }
323 
324 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
325 {
326 	u8 *prog = *pprog;
327 	s64 offset;
328 
329 	offset = func - (ip + X86_PATCH_SIZE);
330 	if (!is_simm32(offset)) {
331 		pr_err("Target call %p is out of range\n", func);
332 		return -ERANGE;
333 	}
334 	EMIT1_off32(opcode, offset);
335 	*pprog = prog;
336 	return 0;
337 }
338 
339 static int emit_call(u8 **pprog, void *func, void *ip)
340 {
341 	return emit_patch(pprog, func, ip, 0xE8);
342 }
343 
344 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
345 {
346 	OPTIMIZER_HIDE_VAR(func);
347 	x86_call_depth_emit_accounting(pprog, func);
348 	return emit_patch(pprog, func, ip, 0xE8);
349 }
350 
351 static int emit_jump(u8 **pprog, void *func, void *ip)
352 {
353 	return emit_patch(pprog, func, ip, 0xE9);
354 }
355 
356 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
357 				void *old_addr, void *new_addr)
358 {
359 	const u8 *nop_insn = x86_nops[5];
360 	u8 old_insn[X86_PATCH_SIZE];
361 	u8 new_insn[X86_PATCH_SIZE];
362 	u8 *prog;
363 	int ret;
364 
365 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
366 	if (old_addr) {
367 		prog = old_insn;
368 		ret = t == BPF_MOD_CALL ?
369 		      emit_call(&prog, old_addr, ip) :
370 		      emit_jump(&prog, old_addr, ip);
371 		if (ret)
372 			return ret;
373 	}
374 
375 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
376 	if (new_addr) {
377 		prog = new_insn;
378 		ret = t == BPF_MOD_CALL ?
379 		      emit_call(&prog, new_addr, ip) :
380 		      emit_jump(&prog, new_addr, ip);
381 		if (ret)
382 			return ret;
383 	}
384 
385 	ret = -EBUSY;
386 	mutex_lock(&text_mutex);
387 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
388 		goto out;
389 	ret = 1;
390 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
391 		text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
392 		ret = 0;
393 	}
394 out:
395 	mutex_unlock(&text_mutex);
396 	return ret;
397 }
398 
399 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
400 		       void *old_addr, void *new_addr)
401 {
402 	if (!is_kernel_text((long)ip) &&
403 	    !is_bpf_text_address((long)ip))
404 		/* BPF poking in modules is not supported */
405 		return -EINVAL;
406 
407 	/*
408 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
409 	 * with an ENDBR instruction.
410 	 */
411 	if (is_endbr(*(u32 *)ip))
412 		ip += ENDBR_INSN_SIZE;
413 
414 	return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
415 }
416 
417 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
418 
419 static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
420 {
421 	u8 *prog = *pprog;
422 
423 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
424 		EMIT_LFENCE();
425 		EMIT2(0xFF, 0xE0 + reg);
426 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
427 		OPTIMIZER_HIDE_VAR(reg);
428 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
429 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
430 		else
431 			emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
432 	} else {
433 		EMIT2(0xFF, 0xE0 + reg);	/* jmp *%\reg */
434 		if (IS_ENABLED(CONFIG_RETPOLINE) || IS_ENABLED(CONFIG_SLS))
435 			EMIT1(0xCC);		/* int3 */
436 	}
437 
438 	*pprog = prog;
439 }
440 
441 static void emit_return(u8 **pprog, u8 *ip)
442 {
443 	u8 *prog = *pprog;
444 
445 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
446 		emit_jump(&prog, x86_return_thunk, ip);
447 	} else {
448 		EMIT1(0xC3);		/* ret */
449 		if (IS_ENABLED(CONFIG_SLS))
450 			EMIT1(0xCC);	/* int3 */
451 	}
452 
453 	*pprog = prog;
454 }
455 
456 /*
457  * Generate the following code:
458  *
459  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
460  *   if (index >= array->map.max_entries)
461  *     goto out;
462  *   if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
463  *     goto out;
464  *   prog = array->ptrs[index];
465  *   if (prog == NULL)
466  *     goto out;
467  *   goto *(prog->bpf_func + prologue_size);
468  * out:
469  */
470 static void emit_bpf_tail_call_indirect(u8 **pprog, bool *callee_regs_used,
471 					u32 stack_depth, u8 *ip,
472 					struct jit_context *ctx)
473 {
474 	int tcc_off = -4 - round_up(stack_depth, 8);
475 	u8 *prog = *pprog, *start = *pprog;
476 	int offset;
477 
478 	/*
479 	 * rdi - pointer to ctx
480 	 * rsi - pointer to bpf_array
481 	 * rdx - index in bpf_array
482 	 */
483 
484 	/*
485 	 * if (index >= array->map.max_entries)
486 	 *	goto out;
487 	 */
488 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
489 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
490 	      offsetof(struct bpf_array, map.max_entries));
491 
492 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
493 	EMIT2(X86_JBE, offset);                   /* jbe out */
494 
495 	/*
496 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
497 	 *	goto out;
498 	 */
499 	EMIT2_off32(0x8B, 0x85, tcc_off);         /* mov eax, dword ptr [rbp - tcc_off] */
500 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
501 
502 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
503 	EMIT2(X86_JAE, offset);                   /* jae out */
504 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
505 	EMIT2_off32(0x89, 0x85, tcc_off);         /* mov dword ptr [rbp - tcc_off], eax */
506 
507 	/* prog = array->ptrs[index]; */
508 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
509 		    offsetof(struct bpf_array, ptrs));
510 
511 	/*
512 	 * if (prog == NULL)
513 	 *	goto out;
514 	 */
515 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
516 
517 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
518 	EMIT2(X86_JE, offset);                    /* je out */
519 
520 	pop_callee_regs(&prog, callee_regs_used);
521 
522 	EMIT1(0x58);                              /* pop rax */
523 	if (stack_depth)
524 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
525 			    round_up(stack_depth, 8));
526 
527 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
528 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
529 	      offsetof(struct bpf_prog, bpf_func));
530 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
531 	      X86_TAIL_CALL_OFFSET);
532 	/*
533 	 * Now we're ready to jump into next BPF program
534 	 * rdi == ctx (1st arg)
535 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
536 	 */
537 	emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
538 
539 	/* out: */
540 	ctx->tail_call_indirect_label = prog - start;
541 	*pprog = prog;
542 }
543 
544 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
545 				      u8 **pprog, u8 *ip,
546 				      bool *callee_regs_used, u32 stack_depth,
547 				      struct jit_context *ctx)
548 {
549 	int tcc_off = -4 - round_up(stack_depth, 8);
550 	u8 *prog = *pprog, *start = *pprog;
551 	int offset;
552 
553 	/*
554 	 * if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
555 	 *	goto out;
556 	 */
557 	EMIT2_off32(0x8B, 0x85, tcc_off);             /* mov eax, dword ptr [rbp - tcc_off] */
558 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);         /* cmp eax, MAX_TAIL_CALL_CNT */
559 
560 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
561 	EMIT2(X86_JAE, offset);                       /* jae out */
562 	EMIT3(0x83, 0xC0, 0x01);                      /* add eax, 1 */
563 	EMIT2_off32(0x89, 0x85, tcc_off);             /* mov dword ptr [rbp - tcc_off], eax */
564 
565 	poke->tailcall_bypass = ip + (prog - start);
566 	poke->adj_off = X86_TAIL_CALL_OFFSET;
567 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
568 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
569 
570 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
571 		  poke->tailcall_bypass);
572 
573 	pop_callee_regs(&prog, callee_regs_used);
574 	EMIT1(0x58);                                  /* pop rax */
575 	if (stack_depth)
576 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
577 
578 	memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
579 	prog += X86_PATCH_SIZE;
580 
581 	/* out: */
582 	ctx->tail_call_direct_label = prog - start;
583 
584 	*pprog = prog;
585 }
586 
587 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
588 {
589 	struct bpf_jit_poke_descriptor *poke;
590 	struct bpf_array *array;
591 	struct bpf_prog *target;
592 	int i, ret;
593 
594 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
595 		poke = &prog->aux->poke_tab[i];
596 		if (poke->aux && poke->aux != prog->aux)
597 			continue;
598 
599 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
600 
601 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
602 			continue;
603 
604 		array = container_of(poke->tail_call.map, struct bpf_array, map);
605 		mutex_lock(&array->aux->poke_mutex);
606 		target = array->ptrs[poke->tail_call.key];
607 		if (target) {
608 			ret = __bpf_arch_text_poke(poke->tailcall_target,
609 						   BPF_MOD_JUMP, NULL,
610 						   (u8 *)target->bpf_func +
611 						   poke->adj_off);
612 			BUG_ON(ret < 0);
613 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
614 						   BPF_MOD_JUMP,
615 						   (u8 *)poke->tailcall_target +
616 						   X86_PATCH_SIZE, NULL);
617 			BUG_ON(ret < 0);
618 		}
619 		WRITE_ONCE(poke->tailcall_target_stable, true);
620 		mutex_unlock(&array->aux->poke_mutex);
621 	}
622 }
623 
624 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
625 			   u32 dst_reg, const u32 imm32)
626 {
627 	u8 *prog = *pprog;
628 	u8 b1, b2, b3;
629 
630 	/*
631 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
632 	 * (which zero-extends imm32) to save 2 bytes.
633 	 */
634 	if (sign_propagate && (s32)imm32 < 0) {
635 		/* 'mov %rax, imm32' sign extends imm32 */
636 		b1 = add_1mod(0x48, dst_reg);
637 		b2 = 0xC7;
638 		b3 = 0xC0;
639 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
640 		goto done;
641 	}
642 
643 	/*
644 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
645 	 * to save 3 bytes.
646 	 */
647 	if (imm32 == 0) {
648 		if (is_ereg(dst_reg))
649 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
650 		b2 = 0x31; /* xor */
651 		b3 = 0xC0;
652 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
653 		goto done;
654 	}
655 
656 	/* mov %eax, imm32 */
657 	if (is_ereg(dst_reg))
658 		EMIT1(add_1mod(0x40, dst_reg));
659 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
660 done:
661 	*pprog = prog;
662 }
663 
664 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
665 			   const u32 imm32_hi, const u32 imm32_lo)
666 {
667 	u8 *prog = *pprog;
668 
669 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
670 		/*
671 		 * For emitting plain u32, where sign bit must not be
672 		 * propagated LLVM tends to load imm64 over mov32
673 		 * directly, so save couple of bytes by just doing
674 		 * 'mov %eax, imm32' instead.
675 		 */
676 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
677 	} else {
678 		/* movabsq rax, imm64 */
679 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
680 		EMIT(imm32_lo, 4);
681 		EMIT(imm32_hi, 4);
682 	}
683 
684 	*pprog = prog;
685 }
686 
687 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
688 {
689 	u8 *prog = *pprog;
690 
691 	if (is64) {
692 		/* mov dst, src */
693 		EMIT_mov(dst_reg, src_reg);
694 	} else {
695 		/* mov32 dst, src */
696 		if (is_ereg(dst_reg) || is_ereg(src_reg))
697 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
698 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
699 	}
700 
701 	*pprog = prog;
702 }
703 
704 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
705 			   u32 src_reg)
706 {
707 	u8 *prog = *pprog;
708 
709 	if (is64) {
710 		/* movs[b,w,l]q dst, src */
711 		if (num_bits == 8)
712 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
713 			      add_2reg(0xC0, src_reg, dst_reg));
714 		else if (num_bits == 16)
715 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
716 			      add_2reg(0xC0, src_reg, dst_reg));
717 		else if (num_bits == 32)
718 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
719 			      add_2reg(0xC0, src_reg, dst_reg));
720 	} else {
721 		/* movs[b,w]l dst, src */
722 		if (num_bits == 8) {
723 			EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
724 			      add_2reg(0xC0, src_reg, dst_reg));
725 		} else if (num_bits == 16) {
726 			if (is_ereg(dst_reg) || is_ereg(src_reg))
727 				EMIT1(add_2mod(0x40, src_reg, dst_reg));
728 			EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
729 			      add_2reg(0xC0, src_reg, dst_reg));
730 		}
731 	}
732 
733 	*pprog = prog;
734 }
735 
736 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
737 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
738 {
739 	u8 *prog = *pprog;
740 
741 	if (is_imm8(off)) {
742 		/* 1-byte signed displacement.
743 		 *
744 		 * If off == 0 we could skip this and save one extra byte, but
745 		 * special case of x86 R13 which always needs an offset is not
746 		 * worth the hassle
747 		 */
748 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
749 	} else {
750 		/* 4-byte signed displacement */
751 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
752 	}
753 	*pprog = prog;
754 }
755 
756 /*
757  * Emit a REX byte if it will be necessary to address these registers
758  */
759 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
760 {
761 	u8 *prog = *pprog;
762 
763 	if (is64)
764 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
765 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
766 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
767 	*pprog = prog;
768 }
769 
770 /*
771  * Similar version of maybe_emit_mod() for a single register
772  */
773 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
774 {
775 	u8 *prog = *pprog;
776 
777 	if (is64)
778 		EMIT1(add_1mod(0x48, reg));
779 	else if (is_ereg(reg))
780 		EMIT1(add_1mod(0x40, reg));
781 	*pprog = prog;
782 }
783 
784 /* LDX: dst_reg = *(u8*)(src_reg + off) */
785 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
786 {
787 	u8 *prog = *pprog;
788 
789 	switch (size) {
790 	case BPF_B:
791 		/* Emit 'movzx rax, byte ptr [rax + off]' */
792 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
793 		break;
794 	case BPF_H:
795 		/* Emit 'movzx rax, word ptr [rax + off]' */
796 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
797 		break;
798 	case BPF_W:
799 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
800 		if (is_ereg(dst_reg) || is_ereg(src_reg))
801 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
802 		else
803 			EMIT1(0x8B);
804 		break;
805 	case BPF_DW:
806 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
807 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
808 		break;
809 	}
810 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
811 	*pprog = prog;
812 }
813 
814 /* LDSX: dst_reg = *(s8*)(src_reg + off) */
815 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
816 {
817 	u8 *prog = *pprog;
818 
819 	switch (size) {
820 	case BPF_B:
821 		/* Emit 'movsx rax, byte ptr [rax + off]' */
822 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
823 		break;
824 	case BPF_H:
825 		/* Emit 'movsx rax, word ptr [rax + off]' */
826 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
827 		break;
828 	case BPF_W:
829 		/* Emit 'movsx rax, dword ptr [rax+0x14]' */
830 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
831 		break;
832 	}
833 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
834 	*pprog = prog;
835 }
836 
837 /* STX: *(u8*)(dst_reg + off) = src_reg */
838 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
839 {
840 	u8 *prog = *pprog;
841 
842 	switch (size) {
843 	case BPF_B:
844 		/* Emit 'mov byte ptr [rax + off], al' */
845 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
846 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
847 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
848 		else
849 			EMIT1(0x88);
850 		break;
851 	case BPF_H:
852 		if (is_ereg(dst_reg) || is_ereg(src_reg))
853 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
854 		else
855 			EMIT2(0x66, 0x89);
856 		break;
857 	case BPF_W:
858 		if (is_ereg(dst_reg) || is_ereg(src_reg))
859 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
860 		else
861 			EMIT1(0x89);
862 		break;
863 	case BPF_DW:
864 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
865 		break;
866 	}
867 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
868 	*pprog = prog;
869 }
870 
871 static int emit_atomic(u8 **pprog, u8 atomic_op,
872 		       u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
873 {
874 	u8 *prog = *pprog;
875 
876 	EMIT1(0xF0); /* lock prefix */
877 
878 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
879 
880 	/* emit opcode */
881 	switch (atomic_op) {
882 	case BPF_ADD:
883 	case BPF_AND:
884 	case BPF_OR:
885 	case BPF_XOR:
886 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
887 		EMIT1(simple_alu_opcodes[atomic_op]);
888 		break;
889 	case BPF_ADD | BPF_FETCH:
890 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
891 		EMIT2(0x0F, 0xC1);
892 		break;
893 	case BPF_XCHG:
894 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
895 		EMIT1(0x87);
896 		break;
897 	case BPF_CMPXCHG:
898 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
899 		EMIT2(0x0F, 0xB1);
900 		break;
901 	default:
902 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
903 		return -EFAULT;
904 	}
905 
906 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
907 
908 	*pprog = prog;
909 	return 0;
910 }
911 
912 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
913 {
914 	u32 reg = x->fixup >> 8;
915 
916 	/* jump over faulting load and clear dest register */
917 	*(unsigned long *)((void *)regs + reg) = 0;
918 	regs->ip += x->fixup & 0xff;
919 	return true;
920 }
921 
922 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
923 			     bool *regs_used, bool *tail_call_seen)
924 {
925 	int i;
926 
927 	for (i = 1; i <= insn_cnt; i++, insn++) {
928 		if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
929 			*tail_call_seen = true;
930 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
931 			regs_used[0] = true;
932 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
933 			regs_used[1] = true;
934 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
935 			regs_used[2] = true;
936 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
937 			regs_used[3] = true;
938 	}
939 }
940 
941 static void emit_nops(u8 **pprog, int len)
942 {
943 	u8 *prog = *pprog;
944 	int i, noplen;
945 
946 	while (len > 0) {
947 		noplen = len;
948 
949 		if (noplen > ASM_NOP_MAX)
950 			noplen = ASM_NOP_MAX;
951 
952 		for (i = 0; i < noplen; i++)
953 			EMIT1(x86_nops[noplen][i]);
954 		len -= noplen;
955 	}
956 
957 	*pprog = prog;
958 }
959 
960 /* emit the 3-byte VEX prefix
961  *
962  * r: same as rex.r, extra bit for ModRM reg field
963  * x: same as rex.x, extra bit for SIB index field
964  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
965  * m: opcode map select, encoding escape bytes e.g. 0x0f38
966  * w: same as rex.w (32 bit or 64 bit) or opcode specific
967  * src_reg2: additional source reg (encoded as BPF reg)
968  * l: vector length (128 bit or 256 bit) or reserved
969  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
970  */
971 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
972 		      bool w, u8 src_reg2, bool l, u8 pp)
973 {
974 	u8 *prog = *pprog;
975 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
976 	u8 b1, b2;
977 	u8 vvvv = reg2hex[src_reg2];
978 
979 	/* reg2hex gives only the lower 3 bit of vvvv */
980 	if (is_ereg(src_reg2))
981 		vvvv |= 1 << 3;
982 
983 	/*
984 	 * 2nd byte of 3-byte VEX prefix
985 	 * ~ means bit inverted encoding
986 	 *
987 	 *    7                           0
988 	 *  +---+---+---+---+---+---+---+---+
989 	 *  |~R |~X |~B |         m         |
990 	 *  +---+---+---+---+---+---+---+---+
991 	 */
992 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
993 	/*
994 	 * 3rd byte of 3-byte VEX prefix
995 	 *
996 	 *    7                           0
997 	 *  +---+---+---+---+---+---+---+---+
998 	 *  | W |     ~vvvv     | L |   pp  |
999 	 *  +---+---+---+---+---+---+---+---+
1000 	 */
1001 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1002 
1003 	EMIT3(b0, b1, b2);
1004 	*pprog = prog;
1005 }
1006 
1007 /* emit BMI2 shift instruction */
1008 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1009 {
1010 	u8 *prog = *pprog;
1011 	bool r = is_ereg(dst_reg);
1012 	u8 m = 2; /* escape code 0f38 */
1013 
1014 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1015 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1016 	*pprog = prog;
1017 }
1018 
1019 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1020 
1021 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
1022 #define RESTORE_TAIL_CALL_CNT(stack)				\
1023 	EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
1024 
1025 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
1026 		  int oldproglen, struct jit_context *ctx, bool jmp_padding)
1027 {
1028 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1029 	struct bpf_insn *insn = bpf_prog->insnsi;
1030 	bool callee_regs_used[4] = {};
1031 	int insn_cnt = bpf_prog->len;
1032 	bool tail_call_seen = false;
1033 	bool seen_exit = false;
1034 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1035 	int i, excnt = 0;
1036 	int ilen, proglen = 0;
1037 	u8 *prog = temp;
1038 	int err;
1039 
1040 	detect_reg_usage(insn, insn_cnt, callee_regs_used,
1041 			 &tail_call_seen);
1042 
1043 	/* tail call's presence in current prog implies it is reachable */
1044 	tail_call_reachable |= tail_call_seen;
1045 
1046 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
1047 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1048 		      bpf_prog->aux->func_idx != 0);
1049 	push_callee_regs(&prog, callee_regs_used);
1050 
1051 	ilen = prog - temp;
1052 	if (rw_image)
1053 		memcpy(rw_image + proglen, temp, ilen);
1054 	proglen += ilen;
1055 	addrs[0] = proglen;
1056 	prog = temp;
1057 
1058 	for (i = 1; i <= insn_cnt; i++, insn++) {
1059 		const s32 imm32 = insn->imm;
1060 		u32 dst_reg = insn->dst_reg;
1061 		u32 src_reg = insn->src_reg;
1062 		u8 b2 = 0, b3 = 0;
1063 		u8 *start_of_ldx;
1064 		s64 jmp_offset;
1065 		s16 insn_off;
1066 		u8 jmp_cond;
1067 		u8 *func;
1068 		int nops;
1069 
1070 		switch (insn->code) {
1071 			/* ALU */
1072 		case BPF_ALU | BPF_ADD | BPF_X:
1073 		case BPF_ALU | BPF_SUB | BPF_X:
1074 		case BPF_ALU | BPF_AND | BPF_X:
1075 		case BPF_ALU | BPF_OR | BPF_X:
1076 		case BPF_ALU | BPF_XOR | BPF_X:
1077 		case BPF_ALU64 | BPF_ADD | BPF_X:
1078 		case BPF_ALU64 | BPF_SUB | BPF_X:
1079 		case BPF_ALU64 | BPF_AND | BPF_X:
1080 		case BPF_ALU64 | BPF_OR | BPF_X:
1081 		case BPF_ALU64 | BPF_XOR | BPF_X:
1082 			maybe_emit_mod(&prog, dst_reg, src_reg,
1083 				       BPF_CLASS(insn->code) == BPF_ALU64);
1084 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1085 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1086 			break;
1087 
1088 		case BPF_ALU64 | BPF_MOV | BPF_X:
1089 		case BPF_ALU | BPF_MOV | BPF_X:
1090 			if (insn->off == 0)
1091 				emit_mov_reg(&prog,
1092 					     BPF_CLASS(insn->code) == BPF_ALU64,
1093 					     dst_reg, src_reg);
1094 			else
1095 				emit_movsx_reg(&prog, insn->off,
1096 					       BPF_CLASS(insn->code) == BPF_ALU64,
1097 					       dst_reg, src_reg);
1098 			break;
1099 
1100 			/* neg dst */
1101 		case BPF_ALU | BPF_NEG:
1102 		case BPF_ALU64 | BPF_NEG:
1103 			maybe_emit_1mod(&prog, dst_reg,
1104 					BPF_CLASS(insn->code) == BPF_ALU64);
1105 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1106 			break;
1107 
1108 		case BPF_ALU | BPF_ADD | BPF_K:
1109 		case BPF_ALU | BPF_SUB | BPF_K:
1110 		case BPF_ALU | BPF_AND | BPF_K:
1111 		case BPF_ALU | BPF_OR | BPF_K:
1112 		case BPF_ALU | BPF_XOR | BPF_K:
1113 		case BPF_ALU64 | BPF_ADD | BPF_K:
1114 		case BPF_ALU64 | BPF_SUB | BPF_K:
1115 		case BPF_ALU64 | BPF_AND | BPF_K:
1116 		case BPF_ALU64 | BPF_OR | BPF_K:
1117 		case BPF_ALU64 | BPF_XOR | BPF_K:
1118 			maybe_emit_1mod(&prog, dst_reg,
1119 					BPF_CLASS(insn->code) == BPF_ALU64);
1120 
1121 			/*
1122 			 * b3 holds 'normal' opcode, b2 short form only valid
1123 			 * in case dst is eax/rax.
1124 			 */
1125 			switch (BPF_OP(insn->code)) {
1126 			case BPF_ADD:
1127 				b3 = 0xC0;
1128 				b2 = 0x05;
1129 				break;
1130 			case BPF_SUB:
1131 				b3 = 0xE8;
1132 				b2 = 0x2D;
1133 				break;
1134 			case BPF_AND:
1135 				b3 = 0xE0;
1136 				b2 = 0x25;
1137 				break;
1138 			case BPF_OR:
1139 				b3 = 0xC8;
1140 				b2 = 0x0D;
1141 				break;
1142 			case BPF_XOR:
1143 				b3 = 0xF0;
1144 				b2 = 0x35;
1145 				break;
1146 			}
1147 
1148 			if (is_imm8(imm32))
1149 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1150 			else if (is_axreg(dst_reg))
1151 				EMIT1_off32(b2, imm32);
1152 			else
1153 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1154 			break;
1155 
1156 		case BPF_ALU64 | BPF_MOV | BPF_K:
1157 		case BPF_ALU | BPF_MOV | BPF_K:
1158 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1159 				       dst_reg, imm32);
1160 			break;
1161 
1162 		case BPF_LD | BPF_IMM | BPF_DW:
1163 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1164 			insn++;
1165 			i++;
1166 			break;
1167 
1168 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1169 		case BPF_ALU | BPF_MOD | BPF_X:
1170 		case BPF_ALU | BPF_DIV | BPF_X:
1171 		case BPF_ALU | BPF_MOD | BPF_K:
1172 		case BPF_ALU | BPF_DIV | BPF_K:
1173 		case BPF_ALU64 | BPF_MOD | BPF_X:
1174 		case BPF_ALU64 | BPF_DIV | BPF_X:
1175 		case BPF_ALU64 | BPF_MOD | BPF_K:
1176 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1177 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1178 
1179 			if (dst_reg != BPF_REG_0)
1180 				EMIT1(0x50); /* push rax */
1181 			if (dst_reg != BPF_REG_3)
1182 				EMIT1(0x52); /* push rdx */
1183 
1184 			if (BPF_SRC(insn->code) == BPF_X) {
1185 				if (src_reg == BPF_REG_0 ||
1186 				    src_reg == BPF_REG_3) {
1187 					/* mov r11, src_reg */
1188 					EMIT_mov(AUX_REG, src_reg);
1189 					src_reg = AUX_REG;
1190 				}
1191 			} else {
1192 				/* mov r11, imm32 */
1193 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1194 				src_reg = AUX_REG;
1195 			}
1196 
1197 			if (dst_reg != BPF_REG_0)
1198 				/* mov rax, dst_reg */
1199 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1200 
1201 			if (insn->off == 0) {
1202 				/*
1203 				 * xor edx, edx
1204 				 * equivalent to 'xor rdx, rdx', but one byte less
1205 				 */
1206 				EMIT2(0x31, 0xd2);
1207 
1208 				/* div src_reg */
1209 				maybe_emit_1mod(&prog, src_reg, is64);
1210 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1211 			} else {
1212 				if (BPF_CLASS(insn->code) == BPF_ALU)
1213 					EMIT1(0x99); /* cdq */
1214 				else
1215 					EMIT2(0x48, 0x99); /* cqo */
1216 
1217 				/* idiv src_reg */
1218 				maybe_emit_1mod(&prog, src_reg, is64);
1219 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1220 			}
1221 
1222 			if (BPF_OP(insn->code) == BPF_MOD &&
1223 			    dst_reg != BPF_REG_3)
1224 				/* mov dst_reg, rdx */
1225 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1226 			else if (BPF_OP(insn->code) == BPF_DIV &&
1227 				 dst_reg != BPF_REG_0)
1228 				/* mov dst_reg, rax */
1229 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1230 
1231 			if (dst_reg != BPF_REG_3)
1232 				EMIT1(0x5A); /* pop rdx */
1233 			if (dst_reg != BPF_REG_0)
1234 				EMIT1(0x58); /* pop rax */
1235 			break;
1236 		}
1237 
1238 		case BPF_ALU | BPF_MUL | BPF_K:
1239 		case BPF_ALU64 | BPF_MUL | BPF_K:
1240 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1241 				       BPF_CLASS(insn->code) == BPF_ALU64);
1242 
1243 			if (is_imm8(imm32))
1244 				/* imul dst_reg, dst_reg, imm8 */
1245 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1246 				      imm32);
1247 			else
1248 				/* imul dst_reg, dst_reg, imm32 */
1249 				EMIT2_off32(0x69,
1250 					    add_2reg(0xC0, dst_reg, dst_reg),
1251 					    imm32);
1252 			break;
1253 
1254 		case BPF_ALU | BPF_MUL | BPF_X:
1255 		case BPF_ALU64 | BPF_MUL | BPF_X:
1256 			maybe_emit_mod(&prog, src_reg, dst_reg,
1257 				       BPF_CLASS(insn->code) == BPF_ALU64);
1258 
1259 			/* imul dst_reg, src_reg */
1260 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1261 			break;
1262 
1263 			/* Shifts */
1264 		case BPF_ALU | BPF_LSH | BPF_K:
1265 		case BPF_ALU | BPF_RSH | BPF_K:
1266 		case BPF_ALU | BPF_ARSH | BPF_K:
1267 		case BPF_ALU64 | BPF_LSH | BPF_K:
1268 		case BPF_ALU64 | BPF_RSH | BPF_K:
1269 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1270 			maybe_emit_1mod(&prog, dst_reg,
1271 					BPF_CLASS(insn->code) == BPF_ALU64);
1272 
1273 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1274 			if (imm32 == 1)
1275 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1276 			else
1277 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1278 			break;
1279 
1280 		case BPF_ALU | BPF_LSH | BPF_X:
1281 		case BPF_ALU | BPF_RSH | BPF_X:
1282 		case BPF_ALU | BPF_ARSH | BPF_X:
1283 		case BPF_ALU64 | BPF_LSH | BPF_X:
1284 		case BPF_ALU64 | BPF_RSH | BPF_X:
1285 		case BPF_ALU64 | BPF_ARSH | BPF_X:
1286 			/* BMI2 shifts aren't better when shift count is already in rcx */
1287 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
1288 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
1289 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
1290 				u8 op;
1291 
1292 				switch (BPF_OP(insn->code)) {
1293 				case BPF_LSH:
1294 					op = 1; /* prefix 0x66 */
1295 					break;
1296 				case BPF_RSH:
1297 					op = 3; /* prefix 0xf2 */
1298 					break;
1299 				case BPF_ARSH:
1300 					op = 2; /* prefix 0xf3 */
1301 					break;
1302 				}
1303 
1304 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
1305 
1306 				break;
1307 			}
1308 
1309 			if (src_reg != BPF_REG_4) { /* common case */
1310 				/* Check for bad case when dst_reg == rcx */
1311 				if (dst_reg == BPF_REG_4) {
1312 					/* mov r11, dst_reg */
1313 					EMIT_mov(AUX_REG, dst_reg);
1314 					dst_reg = AUX_REG;
1315 				} else {
1316 					EMIT1(0x51); /* push rcx */
1317 				}
1318 				/* mov rcx, src_reg */
1319 				EMIT_mov(BPF_REG_4, src_reg);
1320 			}
1321 
1322 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
1323 			maybe_emit_1mod(&prog, dst_reg,
1324 					BPF_CLASS(insn->code) == BPF_ALU64);
1325 
1326 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1327 			EMIT2(0xD3, add_1reg(b3, dst_reg));
1328 
1329 			if (src_reg != BPF_REG_4) {
1330 				if (insn->dst_reg == BPF_REG_4)
1331 					/* mov dst_reg, r11 */
1332 					EMIT_mov(insn->dst_reg, AUX_REG);
1333 				else
1334 					EMIT1(0x59); /* pop rcx */
1335 			}
1336 
1337 			break;
1338 
1339 		case BPF_ALU | BPF_END | BPF_FROM_BE:
1340 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
1341 			switch (imm32) {
1342 			case 16:
1343 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
1344 				EMIT1(0x66);
1345 				if (is_ereg(dst_reg))
1346 					EMIT1(0x41);
1347 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
1348 
1349 				/* Emit 'movzwl eax, ax' */
1350 				if (is_ereg(dst_reg))
1351 					EMIT3(0x45, 0x0F, 0xB7);
1352 				else
1353 					EMIT2(0x0F, 0xB7);
1354 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1355 				break;
1356 			case 32:
1357 				/* Emit 'bswap eax' to swap lower 4 bytes */
1358 				if (is_ereg(dst_reg))
1359 					EMIT2(0x41, 0x0F);
1360 				else
1361 					EMIT1(0x0F);
1362 				EMIT1(add_1reg(0xC8, dst_reg));
1363 				break;
1364 			case 64:
1365 				/* Emit 'bswap rax' to swap 8 bytes */
1366 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
1367 				      add_1reg(0xC8, dst_reg));
1368 				break;
1369 			}
1370 			break;
1371 
1372 		case BPF_ALU | BPF_END | BPF_FROM_LE:
1373 			switch (imm32) {
1374 			case 16:
1375 				/*
1376 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
1377 				 * into 64 bit
1378 				 */
1379 				if (is_ereg(dst_reg))
1380 					EMIT3(0x45, 0x0F, 0xB7);
1381 				else
1382 					EMIT2(0x0F, 0xB7);
1383 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
1384 				break;
1385 			case 32:
1386 				/* Emit 'mov eax, eax' to clear upper 32-bits */
1387 				if (is_ereg(dst_reg))
1388 					EMIT1(0x45);
1389 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
1390 				break;
1391 			case 64:
1392 				/* nop */
1393 				break;
1394 			}
1395 			break;
1396 
1397 			/* speculation barrier */
1398 		case BPF_ST | BPF_NOSPEC:
1399 			EMIT_LFENCE();
1400 			break;
1401 
1402 			/* ST: *(u8*)(dst_reg + off) = imm */
1403 		case BPF_ST | BPF_MEM | BPF_B:
1404 			if (is_ereg(dst_reg))
1405 				EMIT2(0x41, 0xC6);
1406 			else
1407 				EMIT1(0xC6);
1408 			goto st;
1409 		case BPF_ST | BPF_MEM | BPF_H:
1410 			if (is_ereg(dst_reg))
1411 				EMIT3(0x66, 0x41, 0xC7);
1412 			else
1413 				EMIT2(0x66, 0xC7);
1414 			goto st;
1415 		case BPF_ST | BPF_MEM | BPF_W:
1416 			if (is_ereg(dst_reg))
1417 				EMIT2(0x41, 0xC7);
1418 			else
1419 				EMIT1(0xC7);
1420 			goto st;
1421 		case BPF_ST | BPF_MEM | BPF_DW:
1422 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1423 
1424 st:			if (is_imm8(insn->off))
1425 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
1426 			else
1427 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1428 
1429 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1430 			break;
1431 
1432 			/* STX: *(u8*)(dst_reg + off) = src_reg */
1433 		case BPF_STX | BPF_MEM | BPF_B:
1434 		case BPF_STX | BPF_MEM | BPF_H:
1435 		case BPF_STX | BPF_MEM | BPF_W:
1436 		case BPF_STX | BPF_MEM | BPF_DW:
1437 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1438 			break;
1439 
1440 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
1441 		case BPF_LDX | BPF_MEM | BPF_B:
1442 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1443 		case BPF_LDX | BPF_MEM | BPF_H:
1444 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1445 		case BPF_LDX | BPF_MEM | BPF_W:
1446 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1447 		case BPF_LDX | BPF_MEM | BPF_DW:
1448 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1449 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
1450 		case BPF_LDX | BPF_MEMSX | BPF_B:
1451 		case BPF_LDX | BPF_MEMSX | BPF_H:
1452 		case BPF_LDX | BPF_MEMSX | BPF_W:
1453 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1454 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1455 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1456 			insn_off = insn->off;
1457 
1458 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1459 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1460 				/* Conservatively check that src_reg + insn->off is a kernel address:
1461 				 *   src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
1462 				 * src_reg is used as scratch for src_reg += insn->off and restored
1463 				 * after emit_ldx if necessary
1464 				 */
1465 
1466 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
1467 				u8 *end_of_jmp;
1468 
1469 				/* At end of these emitted checks, insn->off will have been added
1470 				 * to src_reg, so no need to do relative load with insn->off offset
1471 				 */
1472 				insn_off = 0;
1473 
1474 				/* movabsq r11, limit */
1475 				EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
1476 				EMIT((u32)limit, 4);
1477 				EMIT(limit >> 32, 4);
1478 
1479 				if (insn->off) {
1480 					/* add src_reg, insn->off */
1481 					maybe_emit_1mod(&prog, src_reg, true);
1482 					EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
1483 				}
1484 
1485 				/* cmp src_reg, r11 */
1486 				maybe_emit_mod(&prog, src_reg, AUX_REG, true);
1487 				EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
1488 
1489 				/* if unsigned '>=', goto load */
1490 				EMIT2(X86_JAE, 0);
1491 				end_of_jmp = prog;
1492 
1493 				/* xor dst_reg, dst_reg */
1494 				emit_mov_imm32(&prog, false, dst_reg, 0);
1495 				/* jmp byte_after_ldx */
1496 				EMIT2(0xEB, 0);
1497 
1498 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
1499 				start_of_ldx = prog;
1500 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
1501 			}
1502 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
1503 			    BPF_MODE(insn->code) == BPF_MEMSX)
1504 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1505 			else
1506 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
1507 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
1508 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
1509 				struct exception_table_entry *ex;
1510 				u8 *_insn = image + proglen + (start_of_ldx - temp);
1511 				s64 delta;
1512 
1513 				/* populate jmp_offset for JMP above */
1514 				start_of_ldx[-1] = prog - start_of_ldx;
1515 
1516 				if (insn->off && src_reg != dst_reg) {
1517 					/* sub src_reg, insn->off
1518 					 * Restore src_reg after "add src_reg, insn->off" in prev
1519 					 * if statement. But if src_reg == dst_reg, emit_ldx
1520 					 * above already clobbered src_reg, so no need to restore.
1521 					 * If add src_reg, insn->off was unnecessary, no need to
1522 					 * restore either.
1523 					 */
1524 					maybe_emit_1mod(&prog, src_reg, true);
1525 					EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
1526 				}
1527 
1528 				if (!bpf_prog->aux->extable)
1529 					break;
1530 
1531 				if (excnt >= bpf_prog->aux->num_exentries) {
1532 					pr_err("ex gen bug\n");
1533 					return -EFAULT;
1534 				}
1535 				ex = &bpf_prog->aux->extable[excnt++];
1536 
1537 				delta = _insn - (u8 *)&ex->insn;
1538 				if (!is_simm32(delta)) {
1539 					pr_err("extable->insn doesn't fit into 32-bit\n");
1540 					return -EFAULT;
1541 				}
1542 				/* switch ex to rw buffer for writes */
1543 				ex = (void *)rw_image + ((void *)ex - (void *)image);
1544 
1545 				ex->insn = delta;
1546 
1547 				ex->data = EX_TYPE_BPF;
1548 
1549 				if (dst_reg > BPF_REG_9) {
1550 					pr_err("verifier error\n");
1551 					return -EFAULT;
1552 				}
1553 				/*
1554 				 * Compute size of x86 insn and its target dest x86 register.
1555 				 * ex_handler_bpf() will use lower 8 bits to adjust
1556 				 * pt_regs->ip to jump over this x86 instruction
1557 				 * and upper bits to figure out which pt_regs to zero out.
1558 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1559 				 * of 4 bytes will be ignored and rbx will be zero inited.
1560 				 */
1561 				ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
1562 			}
1563 			break;
1564 
1565 		case BPF_STX | BPF_ATOMIC | BPF_W:
1566 		case BPF_STX | BPF_ATOMIC | BPF_DW:
1567 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
1568 			    insn->imm == (BPF_OR | BPF_FETCH) ||
1569 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
1570 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
1571 				u32 real_src_reg = src_reg;
1572 				u32 real_dst_reg = dst_reg;
1573 				u8 *branch_target;
1574 
1575 				/*
1576 				 * Can't be implemented with a single x86 insn.
1577 				 * Need to do a CMPXCHG loop.
1578 				 */
1579 
1580 				/* Will need RAX as a CMPXCHG operand so save R0 */
1581 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
1582 				if (src_reg == BPF_REG_0)
1583 					real_src_reg = BPF_REG_AX;
1584 				if (dst_reg == BPF_REG_0)
1585 					real_dst_reg = BPF_REG_AX;
1586 
1587 				branch_target = prog;
1588 				/* Load old value */
1589 				emit_ldx(&prog, BPF_SIZE(insn->code),
1590 					 BPF_REG_0, real_dst_reg, insn->off);
1591 				/*
1592 				 * Perform the (commutative) operation locally,
1593 				 * put the result in the AUX_REG.
1594 				 */
1595 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
1596 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
1597 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
1598 				      add_2reg(0xC0, AUX_REG, real_src_reg));
1599 				/* Attempt to swap in new value */
1600 				err = emit_atomic(&prog, BPF_CMPXCHG,
1601 						  real_dst_reg, AUX_REG,
1602 						  insn->off,
1603 						  BPF_SIZE(insn->code));
1604 				if (WARN_ON(err))
1605 					return err;
1606 				/*
1607 				 * ZF tells us whether we won the race. If it's
1608 				 * cleared we need to try again.
1609 				 */
1610 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
1611 				/* Return the pre-modification value */
1612 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
1613 				/* Restore R0 after clobbering RAX */
1614 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
1615 				break;
1616 			}
1617 
1618 			err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
1619 					  insn->off, BPF_SIZE(insn->code));
1620 			if (err)
1621 				return err;
1622 			break;
1623 
1624 			/* call */
1625 		case BPF_JMP | BPF_CALL: {
1626 			int offs;
1627 
1628 			func = (u8 *) __bpf_call_base + imm32;
1629 			if (tail_call_reachable) {
1630 				RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
1631 				if (!imm32)
1632 					return -EINVAL;
1633 				offs = 7 + x86_call_depth_emit_accounting(&prog, func);
1634 			} else {
1635 				if (!imm32)
1636 					return -EINVAL;
1637 				offs = x86_call_depth_emit_accounting(&prog, func);
1638 			}
1639 			if (emit_call(&prog, func, image + addrs[i - 1] + offs))
1640 				return -EINVAL;
1641 			break;
1642 		}
1643 
1644 		case BPF_JMP | BPF_TAIL_CALL:
1645 			if (imm32)
1646 				emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1647 							  &prog, image + addrs[i - 1],
1648 							  callee_regs_used,
1649 							  bpf_prog->aux->stack_depth,
1650 							  ctx);
1651 			else
1652 				emit_bpf_tail_call_indirect(&prog,
1653 							    callee_regs_used,
1654 							    bpf_prog->aux->stack_depth,
1655 							    image + addrs[i - 1],
1656 							    ctx);
1657 			break;
1658 
1659 			/* cond jump */
1660 		case BPF_JMP | BPF_JEQ | BPF_X:
1661 		case BPF_JMP | BPF_JNE | BPF_X:
1662 		case BPF_JMP | BPF_JGT | BPF_X:
1663 		case BPF_JMP | BPF_JLT | BPF_X:
1664 		case BPF_JMP | BPF_JGE | BPF_X:
1665 		case BPF_JMP | BPF_JLE | BPF_X:
1666 		case BPF_JMP | BPF_JSGT | BPF_X:
1667 		case BPF_JMP | BPF_JSLT | BPF_X:
1668 		case BPF_JMP | BPF_JSGE | BPF_X:
1669 		case BPF_JMP | BPF_JSLE | BPF_X:
1670 		case BPF_JMP32 | BPF_JEQ | BPF_X:
1671 		case BPF_JMP32 | BPF_JNE | BPF_X:
1672 		case BPF_JMP32 | BPF_JGT | BPF_X:
1673 		case BPF_JMP32 | BPF_JLT | BPF_X:
1674 		case BPF_JMP32 | BPF_JGE | BPF_X:
1675 		case BPF_JMP32 | BPF_JLE | BPF_X:
1676 		case BPF_JMP32 | BPF_JSGT | BPF_X:
1677 		case BPF_JMP32 | BPF_JSLT | BPF_X:
1678 		case BPF_JMP32 | BPF_JSGE | BPF_X:
1679 		case BPF_JMP32 | BPF_JSLE | BPF_X:
1680 			/* cmp dst_reg, src_reg */
1681 			maybe_emit_mod(&prog, dst_reg, src_reg,
1682 				       BPF_CLASS(insn->code) == BPF_JMP);
1683 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1684 			goto emit_cond_jmp;
1685 
1686 		case BPF_JMP | BPF_JSET | BPF_X:
1687 		case BPF_JMP32 | BPF_JSET | BPF_X:
1688 			/* test dst_reg, src_reg */
1689 			maybe_emit_mod(&prog, dst_reg, src_reg,
1690 				       BPF_CLASS(insn->code) == BPF_JMP);
1691 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1692 			goto emit_cond_jmp;
1693 
1694 		case BPF_JMP | BPF_JSET | BPF_K:
1695 		case BPF_JMP32 | BPF_JSET | BPF_K:
1696 			/* test dst_reg, imm32 */
1697 			maybe_emit_1mod(&prog, dst_reg,
1698 					BPF_CLASS(insn->code) == BPF_JMP);
1699 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1700 			goto emit_cond_jmp;
1701 
1702 		case BPF_JMP | BPF_JEQ | BPF_K:
1703 		case BPF_JMP | BPF_JNE | BPF_K:
1704 		case BPF_JMP | BPF_JGT | BPF_K:
1705 		case BPF_JMP | BPF_JLT | BPF_K:
1706 		case BPF_JMP | BPF_JGE | BPF_K:
1707 		case BPF_JMP | BPF_JLE | BPF_K:
1708 		case BPF_JMP | BPF_JSGT | BPF_K:
1709 		case BPF_JMP | BPF_JSLT | BPF_K:
1710 		case BPF_JMP | BPF_JSGE | BPF_K:
1711 		case BPF_JMP | BPF_JSLE | BPF_K:
1712 		case BPF_JMP32 | BPF_JEQ | BPF_K:
1713 		case BPF_JMP32 | BPF_JNE | BPF_K:
1714 		case BPF_JMP32 | BPF_JGT | BPF_K:
1715 		case BPF_JMP32 | BPF_JLT | BPF_K:
1716 		case BPF_JMP32 | BPF_JGE | BPF_K:
1717 		case BPF_JMP32 | BPF_JLE | BPF_K:
1718 		case BPF_JMP32 | BPF_JSGT | BPF_K:
1719 		case BPF_JMP32 | BPF_JSLT | BPF_K:
1720 		case BPF_JMP32 | BPF_JSGE | BPF_K:
1721 		case BPF_JMP32 | BPF_JSLE | BPF_K:
1722 			/* test dst_reg, dst_reg to save one extra byte */
1723 			if (imm32 == 0) {
1724 				maybe_emit_mod(&prog, dst_reg, dst_reg,
1725 					       BPF_CLASS(insn->code) == BPF_JMP);
1726 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1727 				goto emit_cond_jmp;
1728 			}
1729 
1730 			/* cmp dst_reg, imm8/32 */
1731 			maybe_emit_1mod(&prog, dst_reg,
1732 					BPF_CLASS(insn->code) == BPF_JMP);
1733 
1734 			if (is_imm8(imm32))
1735 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1736 			else
1737 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1738 
1739 emit_cond_jmp:		/* Convert BPF opcode to x86 */
1740 			switch (BPF_OP(insn->code)) {
1741 			case BPF_JEQ:
1742 				jmp_cond = X86_JE;
1743 				break;
1744 			case BPF_JSET:
1745 			case BPF_JNE:
1746 				jmp_cond = X86_JNE;
1747 				break;
1748 			case BPF_JGT:
1749 				/* GT is unsigned '>', JA in x86 */
1750 				jmp_cond = X86_JA;
1751 				break;
1752 			case BPF_JLT:
1753 				/* LT is unsigned '<', JB in x86 */
1754 				jmp_cond = X86_JB;
1755 				break;
1756 			case BPF_JGE:
1757 				/* GE is unsigned '>=', JAE in x86 */
1758 				jmp_cond = X86_JAE;
1759 				break;
1760 			case BPF_JLE:
1761 				/* LE is unsigned '<=', JBE in x86 */
1762 				jmp_cond = X86_JBE;
1763 				break;
1764 			case BPF_JSGT:
1765 				/* Signed '>', GT in x86 */
1766 				jmp_cond = X86_JG;
1767 				break;
1768 			case BPF_JSLT:
1769 				/* Signed '<', LT in x86 */
1770 				jmp_cond = X86_JL;
1771 				break;
1772 			case BPF_JSGE:
1773 				/* Signed '>=', GE in x86 */
1774 				jmp_cond = X86_JGE;
1775 				break;
1776 			case BPF_JSLE:
1777 				/* Signed '<=', LE in x86 */
1778 				jmp_cond = X86_JLE;
1779 				break;
1780 			default: /* to silence GCC warning */
1781 				return -EFAULT;
1782 			}
1783 			jmp_offset = addrs[i + insn->off] - addrs[i];
1784 			if (is_imm8(jmp_offset)) {
1785 				if (jmp_padding) {
1786 					/* To keep the jmp_offset valid, the extra bytes are
1787 					 * padded before the jump insn, so we subtract the
1788 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
1789 					 *
1790 					 * If the previous pass already emits an imm8
1791 					 * jmp_cond, then this BPF insn won't shrink, so
1792 					 * "nops" is 0.
1793 					 *
1794 					 * On the other hand, if the previous pass emits an
1795 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
1796 					 * keep the image from shrinking further.
1797 					 *
1798 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
1799 					 *     is 2 bytes, so the size difference is 4 bytes.
1800 					 */
1801 					nops = INSN_SZ_DIFF - 2;
1802 					if (nops != 0 && nops != 4) {
1803 						pr_err("unexpected jmp_cond padding: %d bytes\n",
1804 						       nops);
1805 						return -EFAULT;
1806 					}
1807 					emit_nops(&prog, nops);
1808 				}
1809 				EMIT2(jmp_cond, jmp_offset);
1810 			} else if (is_simm32(jmp_offset)) {
1811 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1812 			} else {
1813 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1814 				return -EFAULT;
1815 			}
1816 
1817 			break;
1818 
1819 		case BPF_JMP | BPF_JA:
1820 		case BPF_JMP32 | BPF_JA:
1821 			if (BPF_CLASS(insn->code) == BPF_JMP) {
1822 				if (insn->off == -1)
1823 					/* -1 jmp instructions will always jump
1824 					 * backwards two bytes. Explicitly handling
1825 					 * this case avoids wasting too many passes
1826 					 * when there are long sequences of replaced
1827 					 * dead code.
1828 					 */
1829 					jmp_offset = -2;
1830 				else
1831 					jmp_offset = addrs[i + insn->off] - addrs[i];
1832 			} else {
1833 				if (insn->imm == -1)
1834 					jmp_offset = -2;
1835 				else
1836 					jmp_offset = addrs[i + insn->imm] - addrs[i];
1837 			}
1838 
1839 			if (!jmp_offset) {
1840 				/*
1841 				 * If jmp_padding is enabled, the extra nops will
1842 				 * be inserted. Otherwise, optimize out nop jumps.
1843 				 */
1844 				if (jmp_padding) {
1845 					/* There are 3 possible conditions.
1846 					 * (1) This BPF_JA is already optimized out in
1847 					 *     the previous run, so there is no need
1848 					 *     to pad any extra byte (0 byte).
1849 					 * (2) The previous pass emits an imm8 jmp,
1850 					 *     so we pad 2 bytes to match the previous
1851 					 *     insn size.
1852 					 * (3) Similarly, the previous pass emits an
1853 					 *     imm32 jmp, and 5 bytes is padded.
1854 					 */
1855 					nops = INSN_SZ_DIFF;
1856 					if (nops != 0 && nops != 2 && nops != 5) {
1857 						pr_err("unexpected nop jump padding: %d bytes\n",
1858 						       nops);
1859 						return -EFAULT;
1860 					}
1861 					emit_nops(&prog, nops);
1862 				}
1863 				break;
1864 			}
1865 emit_jmp:
1866 			if (is_imm8(jmp_offset)) {
1867 				if (jmp_padding) {
1868 					/* To avoid breaking jmp_offset, the extra bytes
1869 					 * are padded before the actual jmp insn, so
1870 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
1871 					 *
1872 					 * If the previous pass already emits an imm8
1873 					 * jmp, there is nothing to pad (0 byte).
1874 					 *
1875 					 * If it emits an imm32 jmp (5 bytes) previously
1876 					 * and now an imm8 jmp (2 bytes), then we pad
1877 					 * (5 - 2 = 3) bytes to stop the image from
1878 					 * shrinking further.
1879 					 */
1880 					nops = INSN_SZ_DIFF - 2;
1881 					if (nops != 0 && nops != 3) {
1882 						pr_err("unexpected jump padding: %d bytes\n",
1883 						       nops);
1884 						return -EFAULT;
1885 					}
1886 					emit_nops(&prog, INSN_SZ_DIFF - 2);
1887 				}
1888 				EMIT2(0xEB, jmp_offset);
1889 			} else if (is_simm32(jmp_offset)) {
1890 				EMIT1_off32(0xE9, jmp_offset);
1891 			} else {
1892 				pr_err("jmp gen bug %llx\n", jmp_offset);
1893 				return -EFAULT;
1894 			}
1895 			break;
1896 
1897 		case BPF_JMP | BPF_EXIT:
1898 			if (seen_exit) {
1899 				jmp_offset = ctx->cleanup_addr - addrs[i];
1900 				goto emit_jmp;
1901 			}
1902 			seen_exit = true;
1903 			/* Update cleanup_addr */
1904 			ctx->cleanup_addr = proglen;
1905 			pop_callee_regs(&prog, callee_regs_used);
1906 			EMIT1(0xC9);         /* leave */
1907 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
1908 			break;
1909 
1910 		default:
1911 			/*
1912 			 * By design x86-64 JIT should support all BPF instructions.
1913 			 * This error will be seen if new instruction was added
1914 			 * to the interpreter, but not to the JIT, or if there is
1915 			 * junk in bpf_prog.
1916 			 */
1917 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1918 			return -EINVAL;
1919 		}
1920 
1921 		ilen = prog - temp;
1922 		if (ilen > BPF_MAX_INSN_SIZE) {
1923 			pr_err("bpf_jit: fatal insn size error\n");
1924 			return -EFAULT;
1925 		}
1926 
1927 		if (image) {
1928 			/*
1929 			 * When populating the image, assert that:
1930 			 *
1931 			 *  i) We do not write beyond the allocated space, and
1932 			 * ii) addrs[i] did not change from the prior run, in order
1933 			 *     to validate assumptions made for computing branch
1934 			 *     displacements.
1935 			 */
1936 			if (unlikely(proglen + ilen > oldproglen ||
1937 				     proglen + ilen != addrs[i])) {
1938 				pr_err("bpf_jit: fatal error\n");
1939 				return -EFAULT;
1940 			}
1941 			memcpy(rw_image + proglen, temp, ilen);
1942 		}
1943 		proglen += ilen;
1944 		addrs[i] = proglen;
1945 		prog = temp;
1946 	}
1947 
1948 	if (image && excnt != bpf_prog->aux->num_exentries) {
1949 		pr_err("extable is not populated\n");
1950 		return -EFAULT;
1951 	}
1952 	return proglen;
1953 }
1954 
1955 static void clean_stack_garbage(const struct btf_func_model *m,
1956 				u8 **pprog, int nr_stack_slots,
1957 				int stack_size)
1958 {
1959 	int arg_size, off;
1960 	u8 *prog;
1961 
1962 	/* Generally speaking, the compiler will pass the arguments
1963 	 * on-stack with "push" instruction, which will take 8-byte
1964 	 * on the stack. In this case, there won't be garbage values
1965 	 * while we copy the arguments from origin stack frame to current
1966 	 * in BPF_DW.
1967 	 *
1968 	 * However, sometimes the compiler will only allocate 4-byte on
1969 	 * the stack for the arguments. For now, this case will only
1970 	 * happen if there is only one argument on-stack and its size
1971 	 * not more than 4 byte. In this case, there will be garbage
1972 	 * values on the upper 4-byte where we store the argument on
1973 	 * current stack frame.
1974 	 *
1975 	 * arguments on origin stack:
1976 	 *
1977 	 * stack_arg_1(4-byte) xxx(4-byte)
1978 	 *
1979 	 * what we copy:
1980 	 *
1981 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
1982 	 *
1983 	 * and the xxx is the garbage values which we should clean here.
1984 	 */
1985 	if (nr_stack_slots != 1)
1986 		return;
1987 
1988 	/* the size of the last argument */
1989 	arg_size = m->arg_size[m->nr_args - 1];
1990 	if (arg_size <= 4) {
1991 		off = -(stack_size - 4);
1992 		prog = *pprog;
1993 		/* mov DWORD PTR [rbp + off], 0 */
1994 		if (!is_imm8(off))
1995 			EMIT2_off32(0xC7, 0x85, off);
1996 		else
1997 			EMIT3(0xC7, 0x45, off);
1998 		EMIT(0, 4);
1999 		*pprog = prog;
2000 	}
2001 }
2002 
2003 /* get the count of the regs that are used to pass arguments */
2004 static int get_nr_used_regs(const struct btf_func_model *m)
2005 {
2006 	int i, arg_regs, nr_used_regs = 0;
2007 
2008 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2009 		arg_regs = (m->arg_size[i] + 7) / 8;
2010 		if (nr_used_regs + arg_regs <= 6)
2011 			nr_used_regs += arg_regs;
2012 
2013 		if (nr_used_regs >= 6)
2014 			break;
2015 	}
2016 
2017 	return nr_used_regs;
2018 }
2019 
2020 static void save_args(const struct btf_func_model *m, u8 **prog,
2021 		      int stack_size, bool for_call_origin)
2022 {
2023 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2024 	int i, j;
2025 
2026 	/* Store function arguments to stack.
2027 	 * For a function that accepts two pointers the sequence will be:
2028 	 * mov QWORD PTR [rbp-0x10],rdi
2029 	 * mov QWORD PTR [rbp-0x8],rsi
2030 	 */
2031 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2032 		arg_regs = (m->arg_size[i] + 7) / 8;
2033 
2034 		/* According to the research of Yonghong, struct members
2035 		 * should be all in register or all on the stack.
2036 		 * Meanwhile, the compiler will pass the argument on regs
2037 		 * if the remaining regs can hold the argument.
2038 		 *
2039 		 * Disorder of the args can happen. For example:
2040 		 *
2041 		 * struct foo_struct {
2042 		 *     long a;
2043 		 *     int b;
2044 		 * };
2045 		 * int foo(char, char, char, char, char, struct foo_struct,
2046 		 *         char);
2047 		 *
2048 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2049 		 * by stack.
2050 		 */
2051 		if (nr_regs + arg_regs > 6) {
2052 			/* copy function arguments from origin stack frame
2053 			 * into current stack frame.
2054 			 *
2055 			 * The starting address of the arguments on-stack
2056 			 * is:
2057 			 *   rbp + 8(push rbp) +
2058 			 *   8(return addr of origin call) +
2059 			 *   8(return addr of the caller)
2060 			 * which means: rbp + 24
2061 			 */
2062 			for (j = 0; j < arg_regs; j++) {
2063 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2064 					 nr_stack_slots * 8 + 0x18);
2065 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2066 					 -stack_size);
2067 
2068 				if (!nr_stack_slots)
2069 					first_off = stack_size;
2070 				stack_size -= 8;
2071 				nr_stack_slots++;
2072 			}
2073 		} else {
2074 			/* Only copy the arguments on-stack to current
2075 			 * 'stack_size' and ignore the regs, used to
2076 			 * prepare the arguments on-stack for orign call.
2077 			 */
2078 			if (for_call_origin) {
2079 				nr_regs += arg_regs;
2080 				continue;
2081 			}
2082 
2083 			/* copy the arguments from regs into stack */
2084 			for (j = 0; j < arg_regs; j++) {
2085 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2086 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2087 					 -stack_size);
2088 				stack_size -= 8;
2089 				nr_regs++;
2090 			}
2091 		}
2092 	}
2093 
2094 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2095 }
2096 
2097 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2098 			 int stack_size)
2099 {
2100 	int i, j, arg_regs, nr_regs = 0;
2101 
2102 	/* Restore function arguments from stack.
2103 	 * For a function that accepts two pointers the sequence will be:
2104 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2105 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2106 	 *
2107 	 * The logic here is similar to what we do in save_args()
2108 	 */
2109 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2110 		arg_regs = (m->arg_size[i] + 7) / 8;
2111 		if (nr_regs + arg_regs <= 6) {
2112 			for (j = 0; j < arg_regs; j++) {
2113 				emit_ldx(prog, BPF_DW,
2114 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2115 					 BPF_REG_FP,
2116 					 -stack_size);
2117 				stack_size -= 8;
2118 				nr_regs++;
2119 			}
2120 		} else {
2121 			stack_size -= 8 * arg_regs;
2122 		}
2123 
2124 		if (nr_regs >= 6)
2125 			break;
2126 	}
2127 }
2128 
2129 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2130 			   struct bpf_tramp_link *l, int stack_size,
2131 			   int run_ctx_off, bool save_ret)
2132 {
2133 	u8 *prog = *pprog;
2134 	u8 *jmp_insn;
2135 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2136 	struct bpf_prog *p = l->link.prog;
2137 	u64 cookie = l->cookie;
2138 
2139 	/* mov rdi, cookie */
2140 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2141 
2142 	/* Prepare struct bpf_tramp_run_ctx.
2143 	 *
2144 	 * bpf_tramp_run_ctx is already preserved by
2145 	 * arch_prepare_bpf_trampoline().
2146 	 *
2147 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2148 	 */
2149 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2150 
2151 	/* arg1: mov rdi, progs[i] */
2152 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2153 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2154 	if (!is_imm8(-run_ctx_off))
2155 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
2156 	else
2157 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
2158 
2159 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), prog))
2160 		return -EINVAL;
2161 	/* remember prog start time returned by __bpf_prog_enter */
2162 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
2163 
2164 	/* if (__bpf_prog_enter*(prog) == 0)
2165 	 *	goto skip_exec_of_prog;
2166 	 */
2167 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
2168 	/* emit 2 nops that will be replaced with JE insn */
2169 	jmp_insn = prog;
2170 	emit_nops(&prog, 2);
2171 
2172 	/* arg1: lea rdi, [rbp - stack_size] */
2173 	if (!is_imm8(-stack_size))
2174 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
2175 	else
2176 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
2177 	/* arg2: progs[i]->insnsi for interpreter */
2178 	if (!p->jited)
2179 		emit_mov_imm64(&prog, BPF_REG_2,
2180 			       (long) p->insnsi >> 32,
2181 			       (u32) (long) p->insnsi);
2182 	/* call JITed bpf program or interpreter */
2183 	if (emit_rsb_call(&prog, p->bpf_func, prog))
2184 		return -EINVAL;
2185 
2186 	/*
2187 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
2188 	 * of the previous call which is then passed on the stack to
2189 	 * the next BPF program.
2190 	 *
2191 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
2192 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
2193 	 */
2194 	if (save_ret)
2195 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2196 
2197 	/* replace 2 nops with JE insn, since jmp target is known */
2198 	jmp_insn[0] = X86_JE;
2199 	jmp_insn[1] = prog - jmp_insn - 2;
2200 
2201 	/* arg1: mov rdi, progs[i] */
2202 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2203 	/* arg2: mov rsi, rbx <- start time in nsec */
2204 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
2205 	/* arg3: lea rdx, [rbp - run_ctx_off] */
2206 	if (!is_imm8(-run_ctx_off))
2207 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
2208 	else
2209 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
2210 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), prog))
2211 		return -EINVAL;
2212 
2213 	*pprog = prog;
2214 	return 0;
2215 }
2216 
2217 static void emit_align(u8 **pprog, u32 align)
2218 {
2219 	u8 *target, *prog = *pprog;
2220 
2221 	target = PTR_ALIGN(prog, align);
2222 	if (target != prog)
2223 		emit_nops(&prog, target - prog);
2224 
2225 	*pprog = prog;
2226 }
2227 
2228 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
2229 {
2230 	u8 *prog = *pprog;
2231 	s64 offset;
2232 
2233 	offset = func - (ip + 2 + 4);
2234 	if (!is_simm32(offset)) {
2235 		pr_err("Target %p is out of range\n", func);
2236 		return -EINVAL;
2237 	}
2238 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
2239 	*pprog = prog;
2240 	return 0;
2241 }
2242 
2243 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
2244 		      struct bpf_tramp_links *tl, int stack_size,
2245 		      int run_ctx_off, bool save_ret)
2246 {
2247 	int i;
2248 	u8 *prog = *pprog;
2249 
2250 	for (i = 0; i < tl->nr_links; i++) {
2251 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
2252 				    run_ctx_off, save_ret))
2253 			return -EINVAL;
2254 	}
2255 	*pprog = prog;
2256 	return 0;
2257 }
2258 
2259 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
2260 			      struct bpf_tramp_links *tl, int stack_size,
2261 			      int run_ctx_off, u8 **branches)
2262 {
2263 	u8 *prog = *pprog;
2264 	int i;
2265 
2266 	/* The first fmod_ret program will receive a garbage return value.
2267 	 * Set this to 0 to avoid confusing the program.
2268 	 */
2269 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
2270 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2271 	for (i = 0; i < tl->nr_links; i++) {
2272 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true))
2273 			return -EINVAL;
2274 
2275 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
2276 		 * if (*(u64 *)(rbp - 8) !=  0)
2277 		 *	goto do_fexit;
2278 		 */
2279 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
2280 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
2281 
2282 		/* Save the location of the branch and Generate 6 nops
2283 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
2284 		 * are replaced with a conditional jump once do_fexit (i.e. the
2285 		 * start of the fexit invocation) is finalized.
2286 		 */
2287 		branches[i] = prog;
2288 		emit_nops(&prog, 4 + 2);
2289 	}
2290 
2291 	*pprog = prog;
2292 	return 0;
2293 }
2294 
2295 /* Example:
2296  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
2297  * its 'struct btf_func_model' will be nr_args=2
2298  * The assembly code when eth_type_trans is executing after trampoline:
2299  *
2300  * push rbp
2301  * mov rbp, rsp
2302  * sub rsp, 16                     // space for skb and dev
2303  * push rbx                        // temp regs to pass start time
2304  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
2305  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
2306  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2307  * mov rbx, rax                    // remember start time in bpf stats are enabled
2308  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
2309  * call addr_of_jited_FENTRY_prog
2310  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2311  * mov rsi, rbx                    // prog start time
2312  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2313  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
2314  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
2315  * pop rbx
2316  * leave
2317  * ret
2318  *
2319  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
2320  * replaced with 'call generated_bpf_trampoline'. When it returns
2321  * eth_type_trans will continue executing with original skb and dev pointers.
2322  *
2323  * The assembly code when eth_type_trans is called from trampoline:
2324  *
2325  * push rbp
2326  * mov rbp, rsp
2327  * sub rsp, 24                     // space for skb, dev, return value
2328  * push rbx                        // temp regs to pass start time
2329  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
2330  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
2331  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2332  * mov rbx, rax                    // remember start time if bpf stats are enabled
2333  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2334  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
2335  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2336  * mov rsi, rbx                    // prog start time
2337  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2338  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
2339  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
2340  * call eth_type_trans+5           // execute body of eth_type_trans
2341  * mov qword ptr [rbp - 8], rax    // save return value
2342  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
2343  * mov rbx, rax                    // remember start time in bpf stats are enabled
2344  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
2345  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
2346  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
2347  * mov rsi, rbx                    // prog start time
2348  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
2349  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
2350  * pop rbx
2351  * leave
2352  * add rsp, 8                      // skip eth_type_trans's frame
2353  * ret                             // return to its caller
2354  */
2355 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
2356 				const struct btf_func_model *m, u32 flags,
2357 				struct bpf_tramp_links *tlinks,
2358 				void *func_addr)
2359 {
2360 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
2361 	int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
2362 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
2363 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
2364 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
2365 	void *orig_call = func_addr;
2366 	u8 **branches = NULL;
2367 	u8 *prog;
2368 	bool save_ret;
2369 
2370 	/* extra registers for struct arguments */
2371 	for (i = 0; i < m->nr_args; i++)
2372 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2373 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
2374 
2375 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
2376 	 * are passed through regs, the remains are through stack.
2377 	 */
2378 	if (nr_regs > MAX_BPF_FUNC_ARGS)
2379 		return -ENOTSUPP;
2380 
2381 	/* Generated trampoline stack layout:
2382 	 *
2383 	 * RBP + 8         [ return address  ]
2384 	 * RBP + 0         [ RBP             ]
2385 	 *
2386 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
2387 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
2388 	 *
2389 	 *                 [ reg_argN        ]  always
2390 	 *                 [ ...             ]
2391 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
2392 	 *
2393 	 * RBP - nregs_off [ regs count	     ]  always
2394 	 *
2395 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
2396 	 *
2397 	 * RBP - rbx_off   [ rbx value       ]  always
2398 	 *
2399 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
2400 	 *
2401 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
2402 	 *                     [ ...        ]
2403 	 *                     [ stack_arg2 ]
2404 	 * RBP - arg_stack_off [ stack_arg1 ]
2405 	 * RSP                 [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
2406 	 */
2407 
2408 	/* room for return value of orig_call or fentry prog */
2409 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2410 	if (save_ret)
2411 		stack_size += 8;
2412 
2413 	stack_size += nr_regs * 8;
2414 	regs_off = stack_size;
2415 
2416 	/* regs count  */
2417 	stack_size += 8;
2418 	nregs_off = stack_size;
2419 
2420 	if (flags & BPF_TRAMP_F_IP_ARG)
2421 		stack_size += 8; /* room for IP address argument */
2422 
2423 	ip_off = stack_size;
2424 
2425 	stack_size += 8;
2426 	rbx_off = stack_size;
2427 
2428 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
2429 	run_ctx_off = stack_size;
2430 
2431 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
2432 		/* the space that used to pass arguments on-stack */
2433 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
2434 		/* make sure the stack pointer is 16-byte aligned if we
2435 		 * need pass arguments on stack, which means
2436 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
2437 		 * should be 16-byte aligned. Following code depend on
2438 		 * that stack_size is already 8-byte aligned.
2439 		 */
2440 		stack_size += (stack_size % 16) ? 0 : 8;
2441 	}
2442 
2443 	arg_stack_off = stack_size;
2444 
2445 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2446 		/* skip patched call instruction and point orig_call to actual
2447 		 * body of the kernel function.
2448 		 */
2449 		if (is_endbr(*(u32 *)orig_call))
2450 			orig_call += ENDBR_INSN_SIZE;
2451 		orig_call += X86_PATCH_SIZE;
2452 	}
2453 
2454 	prog = image;
2455 
2456 	EMIT_ENDBR();
2457 	/*
2458 	 * This is the direct-call trampoline, as such it needs accounting
2459 	 * for the __fentry__ call.
2460 	 */
2461 	x86_call_depth_emit_accounting(&prog, NULL);
2462 	EMIT1(0x55);		 /* push rbp */
2463 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
2464 	if (!is_imm8(stack_size))
2465 		/* sub rsp, stack_size */
2466 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
2467 	else
2468 		/* sub rsp, stack_size */
2469 		EMIT4(0x48, 0x83, 0xEC, stack_size);
2470 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2471 		EMIT1(0x50);		/* push rax */
2472 	/* mov QWORD PTR [rbp - rbx_off], rbx */
2473 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
2474 
2475 	/* Store number of argument registers of the traced function:
2476 	 *   mov rax, nr_regs
2477 	 *   mov QWORD PTR [rbp - nregs_off], rax
2478 	 */
2479 	emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
2480 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
2481 
2482 	if (flags & BPF_TRAMP_F_IP_ARG) {
2483 		/* Store IP address of the traced function:
2484 		 * movabsq rax, func_addr
2485 		 * mov QWORD PTR [rbp - ip_off], rax
2486 		 */
2487 		emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
2488 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
2489 	}
2490 
2491 	save_args(m, &prog, regs_off, false);
2492 
2493 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2494 		/* arg1: mov rdi, im */
2495 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2496 		if (emit_rsb_call(&prog, __bpf_tramp_enter, prog)) {
2497 			ret = -EINVAL;
2498 			goto cleanup;
2499 		}
2500 	}
2501 
2502 	if (fentry->nr_links)
2503 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
2504 			       flags & BPF_TRAMP_F_RET_FENTRY_RET))
2505 			return -EINVAL;
2506 
2507 	if (fmod_ret->nr_links) {
2508 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
2509 				   GFP_KERNEL);
2510 		if (!branches)
2511 			return -ENOMEM;
2512 
2513 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
2514 				       run_ctx_off, branches)) {
2515 			ret = -EINVAL;
2516 			goto cleanup;
2517 		}
2518 	}
2519 
2520 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2521 		restore_regs(m, &prog, regs_off);
2522 		save_args(m, &prog, arg_stack_off, true);
2523 
2524 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2525 			/* Before calling the original function, restore the
2526 			 * tail_call_cnt from stack to rax.
2527 			 */
2528 			RESTORE_TAIL_CALL_CNT(stack_size);
2529 
2530 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
2531 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
2532 			EMIT2(0xff, 0xd3); /* call *rbx */
2533 		} else {
2534 			/* call original function */
2535 			if (emit_rsb_call(&prog, orig_call, prog)) {
2536 				ret = -EINVAL;
2537 				goto cleanup;
2538 			}
2539 		}
2540 		/* remember return value in a stack for bpf prog to access */
2541 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
2542 		im->ip_after_call = prog;
2543 		memcpy(prog, x86_nops[5], X86_PATCH_SIZE);
2544 		prog += X86_PATCH_SIZE;
2545 	}
2546 
2547 	if (fmod_ret->nr_links) {
2548 		/* From Intel 64 and IA-32 Architectures Optimization
2549 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2550 		 * Coding Rule 11: All branch targets should be 16-byte
2551 		 * aligned.
2552 		 */
2553 		emit_align(&prog, 16);
2554 		/* Update the branches saved in invoke_bpf_mod_ret with the
2555 		 * aligned address of do_fexit.
2556 		 */
2557 		for (i = 0; i < fmod_ret->nr_links; i++)
2558 			emit_cond_near_jump(&branches[i], prog, branches[i],
2559 					    X86_JNE);
2560 	}
2561 
2562 	if (fexit->nr_links)
2563 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, false)) {
2564 			ret = -EINVAL;
2565 			goto cleanup;
2566 		}
2567 
2568 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2569 		restore_regs(m, &prog, regs_off);
2570 
2571 	/* This needs to be done regardless. If there were fmod_ret programs,
2572 	 * the return value is only updated on the stack and still needs to be
2573 	 * restored to R0.
2574 	 */
2575 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2576 		im->ip_epilogue = prog;
2577 		/* arg1: mov rdi, im */
2578 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
2579 		if (emit_rsb_call(&prog, __bpf_tramp_exit, prog)) {
2580 			ret = -EINVAL;
2581 			goto cleanup;
2582 		}
2583 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
2584 		/* Before running the original function, restore the
2585 		 * tail_call_cnt from stack to rax.
2586 		 */
2587 		RESTORE_TAIL_CALL_CNT(stack_size);
2588 
2589 	/* restore return value of orig_call or fentry prog back into RAX */
2590 	if (save_ret)
2591 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
2592 
2593 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
2594 	EMIT1(0xC9); /* leave */
2595 	if (flags & BPF_TRAMP_F_SKIP_FRAME)
2596 		/* skip our return address and return to parent */
2597 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
2598 	emit_return(&prog, prog);
2599 	/* Make sure the trampoline generation logic doesn't overflow */
2600 	if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
2601 		ret = -EFAULT;
2602 		goto cleanup;
2603 	}
2604 	ret = prog - (u8 *)image;
2605 
2606 cleanup:
2607 	kfree(branches);
2608 	return ret;
2609 }
2610 
2611 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
2612 {
2613 	u8 *jg_reloc, *prog = *pprog;
2614 	int pivot, err, jg_bytes = 1;
2615 	s64 jg_offset;
2616 
2617 	if (a == b) {
2618 		/* Leaf node of recursion, i.e. not a range of indices
2619 		 * anymore.
2620 		 */
2621 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
2622 		if (!is_simm32(progs[a]))
2623 			return -1;
2624 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
2625 			    progs[a]);
2626 		err = emit_cond_near_jump(&prog,	/* je func */
2627 					  (void *)progs[a], image + (prog - buf),
2628 					  X86_JE);
2629 		if (err)
2630 			return err;
2631 
2632 		emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
2633 
2634 		*pprog = prog;
2635 		return 0;
2636 	}
2637 
2638 	/* Not a leaf node, so we pivot, and recursively descend into
2639 	 * the lower and upper ranges.
2640 	 */
2641 	pivot = (b - a) / 2;
2642 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
2643 	if (!is_simm32(progs[a + pivot]))
2644 		return -1;
2645 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
2646 
2647 	if (pivot > 2) {				/* jg upper_part */
2648 		/* Require near jump. */
2649 		jg_bytes = 4;
2650 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
2651 	} else {
2652 		EMIT2(X86_JG, 0);
2653 	}
2654 	jg_reloc = prog;
2655 
2656 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
2657 				  progs, image, buf);
2658 	if (err)
2659 		return err;
2660 
2661 	/* From Intel 64 and IA-32 Architectures Optimization
2662 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
2663 	 * Coding Rule 11: All branch targets should be 16-byte
2664 	 * aligned.
2665 	 */
2666 	emit_align(&prog, 16);
2667 	jg_offset = prog - jg_reloc;
2668 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
2669 
2670 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
2671 				  b, progs, image, buf);
2672 	if (err)
2673 		return err;
2674 
2675 	*pprog = prog;
2676 	return 0;
2677 }
2678 
2679 static int cmp_ips(const void *a, const void *b)
2680 {
2681 	const s64 *ipa = a;
2682 	const s64 *ipb = b;
2683 
2684 	if (*ipa > *ipb)
2685 		return 1;
2686 	if (*ipa < *ipb)
2687 		return -1;
2688 	return 0;
2689 }
2690 
2691 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
2692 {
2693 	u8 *prog = buf;
2694 
2695 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
2696 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
2697 }
2698 
2699 struct x64_jit_data {
2700 	struct bpf_binary_header *rw_header;
2701 	struct bpf_binary_header *header;
2702 	int *addrs;
2703 	u8 *image;
2704 	int proglen;
2705 	struct jit_context ctx;
2706 };
2707 
2708 #define MAX_PASSES 20
2709 #define PADDING_PASSES (MAX_PASSES - 5)
2710 
2711 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
2712 {
2713 	struct bpf_binary_header *rw_header = NULL;
2714 	struct bpf_binary_header *header = NULL;
2715 	struct bpf_prog *tmp, *orig_prog = prog;
2716 	struct x64_jit_data *jit_data;
2717 	int proglen, oldproglen = 0;
2718 	struct jit_context ctx = {};
2719 	bool tmp_blinded = false;
2720 	bool extra_pass = false;
2721 	bool padding = false;
2722 	u8 *rw_image = NULL;
2723 	u8 *image = NULL;
2724 	int *addrs;
2725 	int pass;
2726 	int i;
2727 
2728 	if (!prog->jit_requested)
2729 		return orig_prog;
2730 
2731 	tmp = bpf_jit_blind_constants(prog);
2732 	/*
2733 	 * If blinding was requested and we failed during blinding,
2734 	 * we must fall back to the interpreter.
2735 	 */
2736 	if (IS_ERR(tmp))
2737 		return orig_prog;
2738 	if (tmp != prog) {
2739 		tmp_blinded = true;
2740 		prog = tmp;
2741 	}
2742 
2743 	jit_data = prog->aux->jit_data;
2744 	if (!jit_data) {
2745 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
2746 		if (!jit_data) {
2747 			prog = orig_prog;
2748 			goto out;
2749 		}
2750 		prog->aux->jit_data = jit_data;
2751 	}
2752 	addrs = jit_data->addrs;
2753 	if (addrs) {
2754 		ctx = jit_data->ctx;
2755 		oldproglen = jit_data->proglen;
2756 		image = jit_data->image;
2757 		header = jit_data->header;
2758 		rw_header = jit_data->rw_header;
2759 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
2760 		extra_pass = true;
2761 		padding = true;
2762 		goto skip_init_addrs;
2763 	}
2764 	addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
2765 	if (!addrs) {
2766 		prog = orig_prog;
2767 		goto out_addrs;
2768 	}
2769 
2770 	/*
2771 	 * Before first pass, make a rough estimation of addrs[]
2772 	 * each BPF instruction is translated to less than 64 bytes
2773 	 */
2774 	for (proglen = 0, i = 0; i <= prog->len; i++) {
2775 		proglen += 64;
2776 		addrs[i] = proglen;
2777 	}
2778 	ctx.cleanup_addr = proglen;
2779 skip_init_addrs:
2780 
2781 	/*
2782 	 * JITed image shrinks with every pass and the loop iterates
2783 	 * until the image stops shrinking. Very large BPF programs
2784 	 * may converge on the last pass. In such case do one more
2785 	 * pass to emit the final image.
2786 	 */
2787 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
2788 		if (!padding && pass >= PADDING_PASSES)
2789 			padding = true;
2790 		proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
2791 		if (proglen <= 0) {
2792 out_image:
2793 			image = NULL;
2794 			if (header) {
2795 				bpf_arch_text_copy(&header->size, &rw_header->size,
2796 						   sizeof(rw_header->size));
2797 				bpf_jit_binary_pack_free(header, rw_header);
2798 			}
2799 			/* Fall back to interpreter mode */
2800 			prog = orig_prog;
2801 			if (extra_pass) {
2802 				prog->bpf_func = NULL;
2803 				prog->jited = 0;
2804 				prog->jited_len = 0;
2805 			}
2806 			goto out_addrs;
2807 		}
2808 		if (image) {
2809 			if (proglen != oldproglen) {
2810 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
2811 				       proglen, oldproglen);
2812 				goto out_image;
2813 			}
2814 			break;
2815 		}
2816 		if (proglen == oldproglen) {
2817 			/*
2818 			 * The number of entries in extable is the number of BPF_LDX
2819 			 * insns that access kernel memory via "pointer to BTF type".
2820 			 * The verifier changed their opcode from LDX|MEM|size
2821 			 * to LDX|PROBE_MEM|size to make JITing easier.
2822 			 */
2823 			u32 align = __alignof__(struct exception_table_entry);
2824 			u32 extable_size = prog->aux->num_exentries *
2825 				sizeof(struct exception_table_entry);
2826 
2827 			/* allocate module memory for x86 insns and extable */
2828 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
2829 							   &image, align, &rw_header, &rw_image,
2830 							   jit_fill_hole);
2831 			if (!header) {
2832 				prog = orig_prog;
2833 				goto out_addrs;
2834 			}
2835 			prog->aux->extable = (void *) image + roundup(proglen, align);
2836 		}
2837 		oldproglen = proglen;
2838 		cond_resched();
2839 	}
2840 
2841 	if (bpf_jit_enable > 1)
2842 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
2843 
2844 	if (image) {
2845 		if (!prog->is_func || extra_pass) {
2846 			/*
2847 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
2848 			 *   1) header is not pointing to proper module memory;
2849 			 *   2) the arch doesn't support bpf_arch_text_copy().
2850 			 *
2851 			 * Both cases are serious bugs and justify WARN_ON.
2852 			 */
2853 			if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
2854 				/* header has been freed */
2855 				header = NULL;
2856 				goto out_image;
2857 			}
2858 
2859 			bpf_tail_call_direct_fixup(prog);
2860 		} else {
2861 			jit_data->addrs = addrs;
2862 			jit_data->ctx = ctx;
2863 			jit_data->proglen = proglen;
2864 			jit_data->image = image;
2865 			jit_data->header = header;
2866 			jit_data->rw_header = rw_header;
2867 		}
2868 		prog->bpf_func = (void *)image;
2869 		prog->jited = 1;
2870 		prog->jited_len = proglen;
2871 	} else {
2872 		prog = orig_prog;
2873 	}
2874 
2875 	if (!image || !prog->is_func || extra_pass) {
2876 		if (image)
2877 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
2878 out_addrs:
2879 		kvfree(addrs);
2880 		kfree(jit_data);
2881 		prog->aux->jit_data = NULL;
2882 	}
2883 out:
2884 	if (tmp_blinded)
2885 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2886 					   tmp : orig_prog);
2887 	return prog;
2888 }
2889 
2890 bool bpf_jit_supports_kfunc_call(void)
2891 {
2892 	return true;
2893 }
2894 
2895 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
2896 {
2897 	if (text_poke_copy(dst, src, len) == NULL)
2898 		return ERR_PTR(-EINVAL);
2899 	return dst;
2900 }
2901 
2902 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
2903 bool bpf_jit_supports_subprog_tailcalls(void)
2904 {
2905 	return true;
2906 }
2907 
2908 void bpf_jit_free(struct bpf_prog *prog)
2909 {
2910 	if (prog->jited) {
2911 		struct x64_jit_data *jit_data = prog->aux->jit_data;
2912 		struct bpf_binary_header *hdr;
2913 
2914 		/*
2915 		 * If we fail the final pass of JIT (from jit_subprogs),
2916 		 * the program may not be finalized yet. Call finalize here
2917 		 * before freeing it.
2918 		 */
2919 		if (jit_data) {
2920 			bpf_jit_binary_pack_finalize(prog, jit_data->header,
2921 						     jit_data->rw_header);
2922 			kvfree(jit_data->addrs);
2923 			kfree(jit_data);
2924 		}
2925 		hdr = bpf_jit_binary_pack_hdr(prog);
2926 		bpf_jit_binary_pack_free(hdr, NULL);
2927 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2928 	}
2929 
2930 	bpf_prog_unlock_free(prog);
2931 }
2932 
2933 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
2934 			       struct bpf_prog *new, struct bpf_prog *old)
2935 {
2936 	u8 *old_addr, *new_addr, *old_bypass_addr;
2937 	int ret;
2938 
2939 	old_bypass_addr = old ? NULL : poke->bypass_addr;
2940 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
2941 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
2942 
2943 	/*
2944 	 * On program loading or teardown, the program's kallsym entry
2945 	 * might not be in place, so we use __bpf_arch_text_poke to skip
2946 	 * the kallsyms check.
2947 	 */
2948 	if (new) {
2949 		ret = __bpf_arch_text_poke(poke->tailcall_target,
2950 					   BPF_MOD_JUMP,
2951 					   old_addr, new_addr);
2952 		BUG_ON(ret < 0);
2953 		if (!old) {
2954 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
2955 						   BPF_MOD_JUMP,
2956 						   poke->bypass_addr,
2957 						   NULL);
2958 			BUG_ON(ret < 0);
2959 		}
2960 	} else {
2961 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
2962 					   BPF_MOD_JUMP,
2963 					   old_bypass_addr,
2964 					   poke->bypass_addr);
2965 		BUG_ON(ret < 0);
2966 		/* let other CPUs finish the execution of program
2967 		 * so that it will not possible to expose them
2968 		 * to invalid nop, stack unwind, nop state
2969 		 */
2970 		if (!ret)
2971 			synchronize_rcu();
2972 		ret = __bpf_arch_text_poke(poke->tailcall_target,
2973 					   BPF_MOD_JUMP,
2974 					   old_addr, NULL);
2975 		BUG_ON(ret < 0);
2976 	}
2977 }
2978