xref: /linux/arch/x86/net/bpf_jit_comp.c (revision 9a0e89dcc9be8e0ba20aeb81c330a6352261667e)
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/bitfield.h>
12 #include <linux/bpf.h>
13 #include <linux/memory.h>
14 #include <linux/sort.h>
15 #include <asm/extable.h>
16 #include <asm/ftrace.h>
17 #include <asm/set_memory.h>
18 #include <asm/nospec-branch.h>
19 #include <asm/text-patching.h>
20 #include <asm/unwind.h>
21 #include <asm/cfi.h>
22 
23 static bool all_callee_regs_used[4] = {true, true, true, true};
24 
25 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
26 {
27 	if (len == 1)
28 		*ptr = bytes;
29 	else if (len == 2)
30 		*(u16 *)ptr = bytes;
31 	else {
32 		*(u32 *)ptr = bytes;
33 		barrier();
34 	}
35 	return ptr + len;
36 }
37 
38 #define EMIT(bytes, len) \
39 	do { prog = emit_code(prog, bytes, len); } while (0)
40 
41 #define EMIT1(b1)		EMIT(b1, 1)
42 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
43 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
44 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
45 #define EMIT5(b1, b2, b3, b4, b5) \
46 	do { EMIT1(b1); EMIT4(b2, b3, b4, b5); } while (0)
47 
48 #define EMIT1_off32(b1, off) \
49 	do { EMIT1(b1); EMIT(off, 4); } while (0)
50 #define EMIT2_off32(b1, b2, off) \
51 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
52 #define EMIT3_off32(b1, b2, b3, off) \
53 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
54 #define EMIT4_off32(b1, b2, b3, b4, off) \
55 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
56 
57 #ifdef CONFIG_X86_KERNEL_IBT
58 #define EMIT_ENDBR()		EMIT(gen_endbr(), 4)
59 #define EMIT_ENDBR_POISON()	EMIT(gen_endbr_poison(), 4)
60 #else
61 #define EMIT_ENDBR()		do { } while (0)
62 #define EMIT_ENDBR_POISON()	do { } while (0)
63 #endif
64 
65 static bool is_imm8(int value)
66 {
67 	return value <= 127 && value >= -128;
68 }
69 
70 /*
71  * Let us limit the positive offset to be <= 123.
72  * This is to ensure eventual jit convergence For the following patterns:
73  * ...
74  * pass4, final_proglen=4391:
75  *   ...
76  *   20e:    48 85 ff                test   rdi,rdi
77  *   211:    74 7d                   je     0x290
78  *   213:    48 8b 77 00             mov    rsi,QWORD PTR [rdi+0x0]
79  *   ...
80  *   289:    48 85 ff                test   rdi,rdi
81  *   28c:    74 17                   je     0x2a5
82  *   28e:    e9 7f ff ff ff          jmp    0x212
83  *   293:    bf 03 00 00 00          mov    edi,0x3
84  * Note that insn at 0x211 is 2-byte cond jump insn for offset 0x7d (-125)
85  * and insn at 0x28e is 5-byte jmp insn with offset -129.
86  *
87  * pass5, final_proglen=4392:
88  *   ...
89  *   20e:    48 85 ff                test   rdi,rdi
90  *   211:    0f 84 80 00 00 00       je     0x297
91  *   217:    48 8b 77 00             mov    rsi,QWORD PTR [rdi+0x0]
92  *   ...
93  *   28d:    48 85 ff                test   rdi,rdi
94  *   290:    74 1a                   je     0x2ac
95  *   292:    eb 84                   jmp    0x218
96  *   294:    bf 03 00 00 00          mov    edi,0x3
97  * Note that insn at 0x211 is 6-byte cond jump insn now since its offset
98  * becomes 0x80 based on previous round (0x293 - 0x213 = 0x80).
99  * At the same time, insn at 0x292 is a 2-byte insn since its offset is
100  * -124.
101  *
102  * pass6 will repeat the same code as in pass4 and this will prevent
103  * eventual convergence.
104  *
105  * To fix this issue, we need to break je (2->6 bytes) <-> jmp (5->2 bytes)
106  * cycle in the above. In the above example je offset <= 0x7c should work.
107  *
108  * For other cases, je <-> je needs offset <= 0x7b to avoid no convergence
109  * issue. For jmp <-> je and jmp <-> jmp cases, jmp offset <= 0x7c should
110  * avoid no convergence issue.
111  *
112  * Overall, let us limit the positive offset for 8bit cond/uncond jmp insn
113  * to maximum 123 (0x7b). This way, the jit pass can eventually converge.
114  */
115 static bool is_imm8_jmp_offset(int value)
116 {
117 	return value <= 123 && value >= -128;
118 }
119 
120 static bool is_simm32(s64 value)
121 {
122 	return value == (s64)(s32)value;
123 }
124 
125 static bool is_uimm32(u64 value)
126 {
127 	return value == (u64)(u32)value;
128 }
129 
130 /* mov dst, src */
131 #define EMIT_mov(DST, SRC)								 \
132 	do {										 \
133 		if (DST != SRC)								 \
134 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
135 	} while (0)
136 
137 static int bpf_size_to_x86_bytes(int bpf_size)
138 {
139 	if (bpf_size == BPF_W)
140 		return 4;
141 	else if (bpf_size == BPF_H)
142 		return 2;
143 	else if (bpf_size == BPF_B)
144 		return 1;
145 	else if (bpf_size == BPF_DW)
146 		return 4; /* imm32 */
147 	else
148 		return 0;
149 }
150 
151 /*
152  * List of x86 cond jumps opcodes (. + s8)
153  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
154  */
155 #define X86_JB  0x72
156 #define X86_JAE 0x73
157 #define X86_JE  0x74
158 #define X86_JNE 0x75
159 #define X86_JBE 0x76
160 #define X86_JA  0x77
161 #define X86_JL  0x7C
162 #define X86_JGE 0x7D
163 #define X86_JLE 0x7E
164 #define X86_JG  0x7F
165 
166 /* Pick a register outside of BPF range for JIT internal work */
167 #define AUX_REG (MAX_BPF_JIT_REG + 1)
168 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
169 #define X86_REG_R12 (MAX_BPF_JIT_REG + 3)
170 
171 /*
172  * The following table maps BPF registers to x86-64 registers.
173  *
174  * x86-64 register R12 is unused, since if used as base address
175  * register in load/store instructions, it always needs an
176  * extra byte of encoding and is callee saved.
177  *
178  * x86-64 register R9 is not used by BPF programs, but can be used by BPF
179  * trampoline. x86-64 register R10 is used for blinding (if enabled).
180  */
181 static const int reg2hex[] = {
182 	[BPF_REG_0] = 0,  /* RAX */
183 	[BPF_REG_1] = 7,  /* RDI */
184 	[BPF_REG_2] = 6,  /* RSI */
185 	[BPF_REG_3] = 2,  /* RDX */
186 	[BPF_REG_4] = 1,  /* RCX */
187 	[BPF_REG_5] = 0,  /* R8  */
188 	[BPF_REG_6] = 3,  /* RBX callee saved */
189 	[BPF_REG_7] = 5,  /* R13 callee saved */
190 	[BPF_REG_8] = 6,  /* R14 callee saved */
191 	[BPF_REG_9] = 7,  /* R15 callee saved */
192 	[BPF_REG_FP] = 5, /* RBP readonly */
193 	[BPF_REG_AX] = 2, /* R10 temp register */
194 	[AUX_REG] = 3,    /* R11 temp register */
195 	[X86_REG_R9] = 1, /* R9 register, 6th function argument */
196 	[X86_REG_R12] = 4, /* R12 callee saved */
197 };
198 
199 static const int reg2pt_regs[] = {
200 	[BPF_REG_0] = offsetof(struct pt_regs, ax),
201 	[BPF_REG_1] = offsetof(struct pt_regs, di),
202 	[BPF_REG_2] = offsetof(struct pt_regs, si),
203 	[BPF_REG_3] = offsetof(struct pt_regs, dx),
204 	[BPF_REG_4] = offsetof(struct pt_regs, cx),
205 	[BPF_REG_5] = offsetof(struct pt_regs, r8),
206 	[BPF_REG_6] = offsetof(struct pt_regs, bx),
207 	[BPF_REG_7] = offsetof(struct pt_regs, r13),
208 	[BPF_REG_8] = offsetof(struct pt_regs, r14),
209 	[BPF_REG_9] = offsetof(struct pt_regs, r15),
210 };
211 
212 /*
213  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
214  * which need extra byte of encoding.
215  * rax,rcx,...,rbp have simpler encoding
216  */
217 static bool is_ereg(u32 reg)
218 {
219 	return (1 << reg) & (BIT(BPF_REG_5) |
220 			     BIT(AUX_REG) |
221 			     BIT(BPF_REG_7) |
222 			     BIT(BPF_REG_8) |
223 			     BIT(BPF_REG_9) |
224 			     BIT(X86_REG_R9) |
225 			     BIT(X86_REG_R12) |
226 			     BIT(BPF_REG_AX));
227 }
228 
229 /*
230  * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
231  * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
232  * of encoding. al,cl,dl,bl have simpler encoding.
233  */
234 static bool is_ereg_8l(u32 reg)
235 {
236 	return is_ereg(reg) ||
237 	    (1 << reg) & (BIT(BPF_REG_1) |
238 			  BIT(BPF_REG_2) |
239 			  BIT(BPF_REG_FP));
240 }
241 
242 static bool is_axreg(u32 reg)
243 {
244 	return reg == BPF_REG_0;
245 }
246 
247 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
248 static u8 add_1mod(u8 byte, u32 reg)
249 {
250 	if (is_ereg(reg))
251 		byte |= 1;
252 	return byte;
253 }
254 
255 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
256 {
257 	if (is_ereg(r1))
258 		byte |= 1;
259 	if (is_ereg(r2))
260 		byte |= 4;
261 	return byte;
262 }
263 
264 static u8 add_3mod(u8 byte, u32 r1, u32 r2, u32 index)
265 {
266 	if (is_ereg(r1))
267 		byte |= 1;
268 	if (is_ereg(index))
269 		byte |= 2;
270 	if (is_ereg(r2))
271 		byte |= 4;
272 	return byte;
273 }
274 
275 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
276 static u8 add_1reg(u8 byte, u32 dst_reg)
277 {
278 	return byte + reg2hex[dst_reg];
279 }
280 
281 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
282 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
283 {
284 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
285 }
286 
287 /* Some 1-byte opcodes for binary ALU operations */
288 static u8 simple_alu_opcodes[] = {
289 	[BPF_ADD] = 0x01,
290 	[BPF_SUB] = 0x29,
291 	[BPF_AND] = 0x21,
292 	[BPF_OR] = 0x09,
293 	[BPF_XOR] = 0x31,
294 	[BPF_LSH] = 0xE0,
295 	[BPF_RSH] = 0xE8,
296 	[BPF_ARSH] = 0xF8,
297 };
298 
299 static void jit_fill_hole(void *area, unsigned int size)
300 {
301 	/* Fill whole space with INT3 instructions */
302 	memset(area, 0xcc, size);
303 }
304 
305 int bpf_arch_text_invalidate(void *dst, size_t len)
306 {
307 	return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
308 }
309 
310 struct jit_context {
311 	int cleanup_addr; /* Epilogue code offset */
312 
313 	/*
314 	 * Program specific offsets of labels in the code; these rely on the
315 	 * JIT doing at least 2 passes, recording the position on the first
316 	 * pass, only to generate the correct offset on the second pass.
317 	 */
318 	int tail_call_direct_label;
319 	int tail_call_indirect_label;
320 };
321 
322 /* Maximum number of bytes emitted while JITing one eBPF insn */
323 #define BPF_MAX_INSN_SIZE	128
324 #define BPF_INSN_SAFETY		64
325 
326 /* Number of bytes emit_patch() needs to generate instructions */
327 #define X86_PATCH_SIZE		5
328 /* Number of bytes that will be skipped on tailcall */
329 #define X86_TAIL_CALL_OFFSET	(12 + ENDBR_INSN_SIZE)
330 
331 static void push_r9(u8 **pprog)
332 {
333 	u8 *prog = *pprog;
334 
335 	EMIT2(0x41, 0x51);   /* push r9 */
336 	*pprog = prog;
337 }
338 
339 static void pop_r9(u8 **pprog)
340 {
341 	u8 *prog = *pprog;
342 
343 	EMIT2(0x41, 0x59);   /* pop r9 */
344 	*pprog = prog;
345 }
346 
347 static void push_r12(u8 **pprog)
348 {
349 	u8 *prog = *pprog;
350 
351 	EMIT2(0x41, 0x54);   /* push r12 */
352 	*pprog = prog;
353 }
354 
355 static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
356 {
357 	u8 *prog = *pprog;
358 
359 	if (callee_regs_used[0])
360 		EMIT1(0x53);         /* push rbx */
361 	if (callee_regs_used[1])
362 		EMIT2(0x41, 0x55);   /* push r13 */
363 	if (callee_regs_used[2])
364 		EMIT2(0x41, 0x56);   /* push r14 */
365 	if (callee_regs_used[3])
366 		EMIT2(0x41, 0x57);   /* push r15 */
367 	*pprog = prog;
368 }
369 
370 static void pop_r12(u8 **pprog)
371 {
372 	u8 *prog = *pprog;
373 
374 	EMIT2(0x41, 0x5C);   /* pop r12 */
375 	*pprog = prog;
376 }
377 
378 static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
379 {
380 	u8 *prog = *pprog;
381 
382 	if (callee_regs_used[3])
383 		EMIT2(0x41, 0x5F);   /* pop r15 */
384 	if (callee_regs_used[2])
385 		EMIT2(0x41, 0x5E);   /* pop r14 */
386 	if (callee_regs_used[1])
387 		EMIT2(0x41, 0x5D);   /* pop r13 */
388 	if (callee_regs_used[0])
389 		EMIT1(0x5B);         /* pop rbx */
390 	*pprog = prog;
391 }
392 
393 static void emit_nops(u8 **pprog, int len)
394 {
395 	u8 *prog = *pprog;
396 	int i, noplen;
397 
398 	while (len > 0) {
399 		noplen = len;
400 
401 		if (noplen > ASM_NOP_MAX)
402 			noplen = ASM_NOP_MAX;
403 
404 		for (i = 0; i < noplen; i++)
405 			EMIT1(x86_nops[noplen][i]);
406 		len -= noplen;
407 	}
408 
409 	*pprog = prog;
410 }
411 
412 /*
413  * Emit the various CFI preambles, see asm/cfi.h and the comments about FineIBT
414  * in arch/x86/kernel/alternative.c
415  */
416 static int emit_call(u8 **prog, void *func, void *ip);
417 
418 static void emit_fineibt(u8 **pprog, u8 *ip, u32 hash, int arity)
419 {
420 	u8 *prog = *pprog;
421 
422 	EMIT_ENDBR();
423 	EMIT1_off32(0x2d, hash);			/* subl $hash, %eax	*/
424 	if (cfi_bhi) {
425 		EMIT2(0x2e, 0x2e);			/* cs cs */
426 		emit_call(&prog, __bhi_args[arity], ip + 11);
427 	} else {
428 		EMIT3_off32(0x2e, 0x0f, 0x85, 3);	/* jne.d32,pn 3		*/
429 	}
430 	EMIT_ENDBR_POISON();
431 
432 	*pprog = prog;
433 }
434 
435 static void emit_kcfi(u8 **pprog, u32 hash)
436 {
437 	u8 *prog = *pprog;
438 
439 	EMIT1_off32(0xb8, hash);			/* movl $hash, %eax	*/
440 #ifdef CONFIG_CALL_PADDING
441 	for (int i = 0; i < CONFIG_FUNCTION_PADDING_CFI; i++)
442 		EMIT1(0x90);
443 #endif
444 	EMIT_ENDBR();
445 
446 	*pprog = prog;
447 }
448 
449 static void emit_cfi(u8 **pprog, u8 *ip, u32 hash, int arity)
450 {
451 	u8 *prog = *pprog;
452 
453 	switch (cfi_mode) {
454 	case CFI_FINEIBT:
455 		emit_fineibt(&prog, ip, hash, arity);
456 		break;
457 
458 	case CFI_KCFI:
459 		emit_kcfi(&prog, hash);
460 		break;
461 
462 	default:
463 		EMIT_ENDBR();
464 		break;
465 	}
466 
467 	*pprog = prog;
468 }
469 
470 static void emit_prologue_tail_call(u8 **pprog, bool is_subprog)
471 {
472 	u8 *prog = *pprog;
473 
474 	if (!is_subprog) {
475 		/* cmp rax, MAX_TAIL_CALL_CNT */
476 		EMIT4(0x48, 0x83, 0xF8, MAX_TAIL_CALL_CNT);
477 		EMIT2(X86_JA, 6);        /* ja 6 */
478 		/* rax is tail_call_cnt if <= MAX_TAIL_CALL_CNT.
479 		 * case1: entry of main prog.
480 		 * case2: tail callee of main prog.
481 		 */
482 		EMIT1(0x50);             /* push rax */
483 		/* Make rax as tail_call_cnt_ptr. */
484 		EMIT3(0x48, 0x89, 0xE0); /* mov rax, rsp */
485 		EMIT2(0xEB, 1);          /* jmp 1 */
486 		/* rax is tail_call_cnt_ptr if > MAX_TAIL_CALL_CNT.
487 		 * case: tail callee of subprog.
488 		 */
489 		EMIT1(0x50);             /* push rax */
490 		/* push tail_call_cnt_ptr */
491 		EMIT1(0x50);             /* push rax */
492 	} else { /* is_subprog */
493 		/* rax is tail_call_cnt_ptr. */
494 		EMIT1(0x50);             /* push rax */
495 		EMIT1(0x50);             /* push rax */
496 	}
497 
498 	*pprog = prog;
499 }
500 
501 /*
502  * Emit x86-64 prologue code for BPF program.
503  * bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
504  * while jumping to another program
505  */
506 static void emit_prologue(u8 **pprog, u8 *ip, u32 stack_depth, bool ebpf_from_cbpf,
507 			  bool tail_call_reachable, bool is_subprog,
508 			  bool is_exception_cb)
509 {
510 	u8 *prog = *pprog;
511 
512 	if (is_subprog) {
513 		emit_cfi(&prog, ip, cfi_bpf_subprog_hash, 5);
514 	} else {
515 		emit_cfi(&prog, ip, cfi_bpf_hash, 1);
516 	}
517 	/* BPF trampoline can be made to work without these nops,
518 	 * but let's waste 5 bytes for now and optimize later
519 	 */
520 	emit_nops(&prog, X86_PATCH_SIZE);
521 	if (!ebpf_from_cbpf) {
522 		if (tail_call_reachable && !is_subprog)
523 			/* When it's the entry of the whole tailcall context,
524 			 * zeroing rax means initialising tail_call_cnt.
525 			 */
526 			EMIT3(0x48, 0x31, 0xC0); /* xor rax, rax */
527 		else
528 			/* Keep the same instruction layout. */
529 			emit_nops(&prog, 3);     /* nop3 */
530 	}
531 	/* Exception callback receives FP as third parameter */
532 	if (is_exception_cb) {
533 		EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */
534 		EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */
535 		/* The main frame must have exception_boundary as true, so we
536 		 * first restore those callee-saved regs from stack, before
537 		 * reusing the stack frame.
538 		 */
539 		pop_callee_regs(&prog, all_callee_regs_used);
540 		pop_r12(&prog);
541 		/* Reset the stack frame. */
542 		EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */
543 	} else {
544 		EMIT1(0x55);             /* push rbp */
545 		EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
546 	}
547 
548 	/* X86_TAIL_CALL_OFFSET is here */
549 	EMIT_ENDBR();
550 
551 	/* sub rsp, rounded_stack_depth */
552 	if (stack_depth)
553 		EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
554 	if (tail_call_reachable)
555 		emit_prologue_tail_call(&prog, is_subprog);
556 	*pprog = prog;
557 }
558 
559 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
560 {
561 	u8 *prog = *pprog;
562 	s64 offset;
563 
564 	offset = func - (ip + X86_PATCH_SIZE);
565 	if (!is_simm32(offset)) {
566 		pr_err("Target call %p is out of range\n", func);
567 		return -ERANGE;
568 	}
569 	EMIT1_off32(opcode, offset);
570 	*pprog = prog;
571 	return 0;
572 }
573 
574 static int emit_call(u8 **pprog, void *func, void *ip)
575 {
576 	return emit_patch(pprog, func, ip, 0xE8);
577 }
578 
579 static int emit_rsb_call(u8 **pprog, void *func, void *ip)
580 {
581 	OPTIMIZER_HIDE_VAR(func);
582 	ip += x86_call_depth_emit_accounting(pprog, func, ip);
583 	return emit_patch(pprog, func, ip, 0xE8);
584 }
585 
586 static int emit_jump(u8 **pprog, void *func, void *ip)
587 {
588 	return emit_patch(pprog, func, ip, 0xE9);
589 }
590 
591 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
592 				enum bpf_text_poke_type new_t,
593 				void *old_addr, void *new_addr)
594 {
595 	const u8 *nop_insn = x86_nops[5];
596 	u8 old_insn[X86_PATCH_SIZE];
597 	u8 new_insn[X86_PATCH_SIZE];
598 	u8 *prog;
599 	int ret;
600 
601 	memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
602 	if (old_t != BPF_MOD_NOP && old_addr) {
603 		prog = old_insn;
604 		ret = old_t == BPF_MOD_CALL ?
605 		      emit_call(&prog, old_addr, ip) :
606 		      emit_jump(&prog, old_addr, ip);
607 		if (ret)
608 			return ret;
609 	}
610 
611 	memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
612 	if (new_t != BPF_MOD_NOP && new_addr) {
613 		prog = new_insn;
614 		ret = new_t == BPF_MOD_CALL ?
615 		      emit_call(&prog, new_addr, ip) :
616 		      emit_jump(&prog, new_addr, ip);
617 		if (ret)
618 			return ret;
619 	}
620 
621 	ret = -EBUSY;
622 	mutex_lock(&text_mutex);
623 	if (memcmp(ip, old_insn, X86_PATCH_SIZE))
624 		goto out;
625 	ret = 1;
626 	if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
627 		smp_text_poke_single(ip, new_insn, X86_PATCH_SIZE, NULL);
628 		ret = 0;
629 	}
630 out:
631 	mutex_unlock(&text_mutex);
632 	return ret;
633 }
634 
635 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type old_t,
636 		       enum bpf_text_poke_type new_t, void *old_addr,
637 		       void *new_addr)
638 {
639 	if (!is_kernel_text((long)ip) &&
640 	    !is_bpf_text_address((long)ip))
641 		/* BPF poking in modules is not supported */
642 		return -EINVAL;
643 
644 	/*
645 	 * See emit_prologue(), for IBT builds the trampoline hook is preceded
646 	 * with an ENDBR instruction.
647 	 */
648 	if (is_endbr(ip))
649 		ip += ENDBR_INSN_SIZE;
650 
651 	return __bpf_arch_text_poke(ip, old_t, new_t, old_addr, new_addr);
652 }
653 
654 #define EMIT_LFENCE()	EMIT3(0x0F, 0xAE, 0xE8)
655 
656 static void __emit_indirect_jump(u8 **pprog, int reg, bool ereg)
657 {
658 	u8 *prog = *pprog;
659 
660 	if (ereg)
661 		EMIT1(0x41);
662 
663 	EMIT2(0xFF, 0xE0 + reg);
664 
665 	*pprog = prog;
666 }
667 
668 static void emit_indirect_jump(u8 **pprog, int bpf_reg, u8 *ip)
669 {
670 	u8 *prog = *pprog;
671 	int reg = reg2hex[bpf_reg];
672 	bool ereg = is_ereg(bpf_reg);
673 
674 	if (cpu_feature_enabled(X86_FEATURE_INDIRECT_THUNK_ITS)) {
675 		OPTIMIZER_HIDE_VAR(reg);
676 		emit_jump(&prog, its_static_thunk(reg + 8*ereg), ip);
677 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
678 		EMIT_LFENCE();
679 		__emit_indirect_jump(&prog, reg, ereg);
680 	} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
681 		OPTIMIZER_HIDE_VAR(reg);
682 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
683 			emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg + 8*ereg], ip);
684 		else
685 			emit_jump(&prog, &__x86_indirect_thunk_array[reg + 8*ereg], ip);
686 	} else {
687 		__emit_indirect_jump(&prog, reg, ereg);
688 		if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) || IS_ENABLED(CONFIG_MITIGATION_SLS))
689 			EMIT1(0xCC);		/* int3 */
690 	}
691 
692 	*pprog = prog;
693 }
694 
695 static void emit_return(u8 **pprog, u8 *ip)
696 {
697 	u8 *prog = *pprog;
698 
699 	if (cpu_wants_rethunk()) {
700 		emit_jump(&prog, x86_return_thunk, ip);
701 	} else {
702 		EMIT1(0xC3);		/* ret */
703 		if (IS_ENABLED(CONFIG_MITIGATION_SLS))
704 			EMIT1(0xCC);	/* int3 */
705 	}
706 
707 	*pprog = prog;
708 }
709 
710 #define BPF_TAIL_CALL_CNT_PTR_STACK_OFF(stack)	(-16 - round_up(stack, 8))
711 
712 /*
713  * Generate the following code:
714  *
715  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
716  *   if (index >= array->map.max_entries)
717  *     goto out;
718  *   if ((*tcc_ptr)++ >= MAX_TAIL_CALL_CNT)
719  *     goto out;
720  *   prog = array->ptrs[index];
721  *   if (prog == NULL)
722  *     goto out;
723  *   goto *(prog->bpf_func + prologue_size);
724  * out:
725  */
726 static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog,
727 					u8 **pprog, bool *callee_regs_used,
728 					u32 stack_depth, u8 *ip,
729 					struct jit_context *ctx)
730 {
731 	int tcc_ptr_off = BPF_TAIL_CALL_CNT_PTR_STACK_OFF(stack_depth);
732 	u8 *prog = *pprog, *start = *pprog;
733 	int offset;
734 
735 	/*
736 	 * rdi - pointer to ctx
737 	 * rsi - pointer to bpf_array
738 	 * rdx - index in bpf_array
739 	 */
740 
741 	/*
742 	 * if (index >= array->map.max_entries)
743 	 *	goto out;
744 	 */
745 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
746 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
747 	      offsetof(struct bpf_array, map.max_entries));
748 
749 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
750 	EMIT2(X86_JBE, offset);                   /* jbe out */
751 
752 	/*
753 	 * if ((*tcc_ptr)++ >= MAX_TAIL_CALL_CNT)
754 	 *	goto out;
755 	 */
756 	EMIT3_off32(0x48, 0x8B, 0x85, tcc_ptr_off); /* mov rax, qword ptr [rbp - tcc_ptr_off] */
757 	EMIT4(0x48, 0x83, 0x38, MAX_TAIL_CALL_CNT); /* cmp qword ptr [rax], MAX_TAIL_CALL_CNT */
758 
759 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
760 	EMIT2(X86_JAE, offset);                   /* jae out */
761 
762 	/* prog = array->ptrs[index]; */
763 	EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6,       /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
764 		    offsetof(struct bpf_array, ptrs));
765 
766 	/*
767 	 * if (prog == NULL)
768 	 *	goto out;
769 	 */
770 	EMIT3(0x48, 0x85, 0xC9);                  /* test rcx,rcx */
771 
772 	offset = ctx->tail_call_indirect_label - (prog + 2 - start);
773 	EMIT2(X86_JE, offset);                    /* je out */
774 
775 	/* Inc tail_call_cnt if the slot is populated. */
776 	EMIT4(0x48, 0x83, 0x00, 0x01);            /* add qword ptr [rax], 1 */
777 
778 	if (bpf_prog->aux->exception_boundary) {
779 		pop_callee_regs(&prog, all_callee_regs_used);
780 		pop_r12(&prog);
781 	} else {
782 		pop_callee_regs(&prog, callee_regs_used);
783 		if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
784 			pop_r12(&prog);
785 	}
786 
787 	/* Pop tail_call_cnt_ptr. */
788 	EMIT1(0x58);                              /* pop rax */
789 	/* Pop tail_call_cnt, if it's main prog.
790 	 * Pop tail_call_cnt_ptr, if it's subprog.
791 	 */
792 	EMIT1(0x58);                              /* pop rax */
793 	if (stack_depth)
794 		EMIT3_off32(0x48, 0x81, 0xC4,     /* add rsp, sd */
795 			    round_up(stack_depth, 8));
796 
797 	/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
798 	EMIT4(0x48, 0x8B, 0x49,                   /* mov rcx, qword ptr [rcx + 32] */
799 	      offsetof(struct bpf_prog, bpf_func));
800 	EMIT4(0x48, 0x83, 0xC1,                   /* add rcx, X86_TAIL_CALL_OFFSET */
801 	      X86_TAIL_CALL_OFFSET);
802 	/*
803 	 * Now we're ready to jump into next BPF program
804 	 * rdi == ctx (1st arg)
805 	 * rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
806 	 */
807 	emit_indirect_jump(&prog, BPF_REG_4 /* R4 -> rcx */, ip + (prog - start));
808 
809 	/* out: */
810 	ctx->tail_call_indirect_label = prog - start;
811 	*pprog = prog;
812 }
813 
814 static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog,
815 				      struct bpf_jit_poke_descriptor *poke,
816 				      u8 **pprog, u8 *ip,
817 				      bool *callee_regs_used, u32 stack_depth,
818 				      struct jit_context *ctx)
819 {
820 	int tcc_ptr_off = BPF_TAIL_CALL_CNT_PTR_STACK_OFF(stack_depth);
821 	u8 *prog = *pprog, *start = *pprog;
822 	int offset;
823 
824 	/*
825 	 * if ((*tcc_ptr)++ >= MAX_TAIL_CALL_CNT)
826 	 *	goto out;
827 	 */
828 	EMIT3_off32(0x48, 0x8B, 0x85, tcc_ptr_off);   /* mov rax, qword ptr [rbp - tcc_ptr_off] */
829 	EMIT4(0x48, 0x83, 0x38, MAX_TAIL_CALL_CNT);   /* cmp qword ptr [rax], MAX_TAIL_CALL_CNT */
830 
831 	offset = ctx->tail_call_direct_label - (prog + 2 - start);
832 	EMIT2(X86_JAE, offset);                       /* jae out */
833 
834 	poke->tailcall_bypass = ip + (prog - start);
835 	poke->adj_off = X86_TAIL_CALL_OFFSET;
836 	poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
837 	poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
838 
839 	emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
840 		  poke->tailcall_bypass);
841 
842 	/* Inc tail_call_cnt if the slot is populated. */
843 	EMIT4(0x48, 0x83, 0x00, 0x01);                /* add qword ptr [rax], 1 */
844 
845 	if (bpf_prog->aux->exception_boundary) {
846 		pop_callee_regs(&prog, all_callee_regs_used);
847 		pop_r12(&prog);
848 	} else {
849 		pop_callee_regs(&prog, callee_regs_used);
850 		if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
851 			pop_r12(&prog);
852 	}
853 
854 	/* Pop tail_call_cnt_ptr. */
855 	EMIT1(0x58);                                  /* pop rax */
856 	/* Pop tail_call_cnt, if it's main prog.
857 	 * Pop tail_call_cnt_ptr, if it's subprog.
858 	 */
859 	EMIT1(0x58);                                  /* pop rax */
860 	if (stack_depth)
861 		EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
862 
863 	emit_nops(&prog, X86_PATCH_SIZE);
864 
865 	/* out: */
866 	ctx->tail_call_direct_label = prog - start;
867 
868 	*pprog = prog;
869 }
870 
871 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
872 {
873 	struct bpf_jit_poke_descriptor *poke;
874 	struct bpf_array *array;
875 	struct bpf_prog *target;
876 	int i, ret;
877 
878 	for (i = 0; i < prog->aux->size_poke_tab; i++) {
879 		poke = &prog->aux->poke_tab[i];
880 		if (poke->aux && poke->aux != prog->aux)
881 			continue;
882 
883 		WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
884 
885 		if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
886 			continue;
887 
888 		array = container_of(poke->tail_call.map, struct bpf_array, map);
889 		mutex_lock(&array->aux->poke_mutex);
890 		target = array->ptrs[poke->tail_call.key];
891 		if (target) {
892 			ret = __bpf_arch_text_poke(poke->tailcall_target,
893 						   BPF_MOD_NOP, BPF_MOD_JUMP,
894 						   NULL,
895 						   (u8 *)target->bpf_func +
896 						   poke->adj_off);
897 			BUG_ON(ret < 0);
898 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
899 						   BPF_MOD_JUMP, BPF_MOD_NOP,
900 						   (u8 *)poke->tailcall_target +
901 						   X86_PATCH_SIZE, NULL);
902 			BUG_ON(ret < 0);
903 		}
904 		WRITE_ONCE(poke->tailcall_target_stable, true);
905 		mutex_unlock(&array->aux->poke_mutex);
906 	}
907 }
908 
909 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
910 			   u32 dst_reg, const u32 imm32)
911 {
912 	u8 *prog = *pprog;
913 	u8 b1, b2, b3;
914 
915 	/*
916 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
917 	 * (which zero-extends imm32) to save 2 bytes.
918 	 */
919 	if (sign_propagate && (s32)imm32 < 0) {
920 		/* 'mov %rax, imm32' sign extends imm32 */
921 		b1 = add_1mod(0x48, dst_reg);
922 		b2 = 0xC7;
923 		b3 = 0xC0;
924 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
925 		goto done;
926 	}
927 
928 	/*
929 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
930 	 * to save 3 bytes.
931 	 */
932 	if (imm32 == 0) {
933 		if (is_ereg(dst_reg))
934 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
935 		b2 = 0x31; /* xor */
936 		b3 = 0xC0;
937 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
938 		goto done;
939 	}
940 
941 	/* mov %eax, imm32 */
942 	if (is_ereg(dst_reg))
943 		EMIT1(add_1mod(0x40, dst_reg));
944 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
945 done:
946 	*pprog = prog;
947 }
948 
949 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
950 			   const u32 imm32_hi, const u32 imm32_lo)
951 {
952 	u64 imm64 = ((u64)imm32_hi << 32) | (u32)imm32_lo;
953 	u8 *prog = *pprog;
954 
955 	if (is_uimm32(imm64)) {
956 		/*
957 		 * For emitting plain u32, where sign bit must not be
958 		 * propagated LLVM tends to load imm64 over mov32
959 		 * directly, so save couple of bytes by just doing
960 		 * 'mov %eax, imm32' instead.
961 		 */
962 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
963 	} else if (is_simm32(imm64)) {
964 		emit_mov_imm32(&prog, true, dst_reg, imm32_lo);
965 	} else {
966 		/* movabsq rax, imm64 */
967 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
968 		EMIT(imm32_lo, 4);
969 		EMIT(imm32_hi, 4);
970 	}
971 
972 	*pprog = prog;
973 }
974 
975 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
976 {
977 	u8 *prog = *pprog;
978 
979 	if (is64) {
980 		/* mov dst, src */
981 		EMIT_mov(dst_reg, src_reg);
982 	} else {
983 		/* mov32 dst, src */
984 		if (is_ereg(dst_reg) || is_ereg(src_reg))
985 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
986 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
987 	}
988 
989 	*pprog = prog;
990 }
991 
992 static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
993 			   u32 src_reg)
994 {
995 	u8 *prog = *pprog;
996 
997 	if (is64) {
998 		/* movs[b,w,l]q dst, src */
999 		if (num_bits == 8)
1000 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
1001 			      add_2reg(0xC0, src_reg, dst_reg));
1002 		else if (num_bits == 16)
1003 			EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
1004 			      add_2reg(0xC0, src_reg, dst_reg));
1005 		else if (num_bits == 32)
1006 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
1007 			      add_2reg(0xC0, src_reg, dst_reg));
1008 	} else {
1009 		/* movs[b,w]l dst, src */
1010 		if (num_bits == 8) {
1011 			EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
1012 			      add_2reg(0xC0, src_reg, dst_reg));
1013 		} else if (num_bits == 16) {
1014 			if (is_ereg(dst_reg) || is_ereg(src_reg))
1015 				EMIT1(add_2mod(0x40, src_reg, dst_reg));
1016 			EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
1017 			      add_2reg(0xC0, src_reg, dst_reg));
1018 		}
1019 	}
1020 
1021 	*pprog = prog;
1022 }
1023 
1024 /* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
1025 static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
1026 {
1027 	u8 *prog = *pprog;
1028 
1029 	if (is_imm8(off)) {
1030 		/* 1-byte signed displacement.
1031 		 *
1032 		 * If off == 0 we could skip this and save one extra byte, but
1033 		 * special case of x86 R13 which always needs an offset is not
1034 		 * worth the hassle
1035 		 */
1036 		EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
1037 	} else {
1038 		/* 4-byte signed displacement */
1039 		EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
1040 	}
1041 	*pprog = prog;
1042 }
1043 
1044 static void emit_insn_suffix_SIB(u8 **pprog, u32 ptr_reg, u32 val_reg, u32 index_reg, int off)
1045 {
1046 	u8 *prog = *pprog;
1047 
1048 	if (is_imm8(off)) {
1049 		EMIT3(add_2reg(0x44, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
1050 	} else {
1051 		EMIT2_off32(add_2reg(0x84, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
1052 	}
1053 	*pprog = prog;
1054 }
1055 
1056 /*
1057  * Emit a REX byte if it will be necessary to address these registers
1058  */
1059 static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
1060 {
1061 	u8 *prog = *pprog;
1062 
1063 	if (is64)
1064 		EMIT1(add_2mod(0x48, dst_reg, src_reg));
1065 	else if (is_ereg(dst_reg) || is_ereg(src_reg))
1066 		EMIT1(add_2mod(0x40, dst_reg, src_reg));
1067 	*pprog = prog;
1068 }
1069 
1070 /*
1071  * Similar version of maybe_emit_mod() for a single register
1072  */
1073 static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
1074 {
1075 	u8 *prog = *pprog;
1076 
1077 	if (is64)
1078 		EMIT1(add_1mod(0x48, reg));
1079 	else if (is_ereg(reg))
1080 		EMIT1(add_1mod(0x40, reg));
1081 	*pprog = prog;
1082 }
1083 
1084 /* LDX: dst_reg = *(u8*)(src_reg + off) */
1085 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1086 {
1087 	u8 *prog = *pprog;
1088 
1089 	switch (size) {
1090 	case BPF_B:
1091 		/* Emit 'movzx rax, byte ptr [rax + off]' */
1092 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
1093 		break;
1094 	case BPF_H:
1095 		/* Emit 'movzx rax, word ptr [rax + off]' */
1096 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
1097 		break;
1098 	case BPF_W:
1099 		/* Emit 'mov eax, dword ptr [rax+0x14]' */
1100 		if (is_ereg(dst_reg) || is_ereg(src_reg))
1101 			EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
1102 		else
1103 			EMIT1(0x8B);
1104 		break;
1105 	case BPF_DW:
1106 		/* Emit 'mov rax, qword ptr [rax+0x14]' */
1107 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
1108 		break;
1109 	}
1110 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
1111 	*pprog = prog;
1112 }
1113 
1114 /* LDSX: dst_reg = *(s8*)(src_reg + off) */
1115 static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1116 {
1117 	u8 *prog = *pprog;
1118 
1119 	switch (size) {
1120 	case BPF_B:
1121 		/* Emit 'movsx rax, byte ptr [rax + off]' */
1122 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
1123 		break;
1124 	case BPF_H:
1125 		/* Emit 'movsx rax, word ptr [rax + off]' */
1126 		EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
1127 		break;
1128 	case BPF_W:
1129 		/* Emit 'movsx rax, dword ptr [rax+0x14]' */
1130 		EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
1131 		break;
1132 	}
1133 	emit_insn_suffix(&prog, src_reg, dst_reg, off);
1134 	*pprog = prog;
1135 }
1136 
1137 static void emit_ldx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1138 {
1139 	u8 *prog = *pprog;
1140 
1141 	switch (size) {
1142 	case BPF_B:
1143 		/* movzx rax, byte ptr [rax + r12 + off] */
1144 		EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB6);
1145 		break;
1146 	case BPF_H:
1147 		/* movzx rax, word ptr [rax + r12 + off] */
1148 		EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB7);
1149 		break;
1150 	case BPF_W:
1151 		/* mov eax, dword ptr [rax + r12 + off] */
1152 		EMIT2(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x8B);
1153 		break;
1154 	case BPF_DW:
1155 		/* mov rax, qword ptr [rax + r12 + off] */
1156 		EMIT2(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x8B);
1157 		break;
1158 	}
1159 	emit_insn_suffix_SIB(&prog, src_reg, dst_reg, index_reg, off);
1160 	*pprog = prog;
1161 }
1162 
1163 static void emit_ldsx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1164 {
1165 	u8 *prog = *pprog;
1166 
1167 	switch (size) {
1168 	case BPF_B:
1169 		/* movsx rax, byte ptr [rax + r12 + off] */
1170 		EMIT3(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x0F, 0xBE);
1171 		break;
1172 	case BPF_H:
1173 		/* movsx rax, word ptr [rax + r12 + off] */
1174 		EMIT3(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x0F, 0xBF);
1175 		break;
1176 	case BPF_W:
1177 		/* movsx rax, dword ptr [rax + r12 + off] */
1178 		EMIT2(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x63);
1179 		break;
1180 	}
1181 	emit_insn_suffix_SIB(&prog, src_reg, dst_reg, index_reg, off);
1182 	*pprog = prog;
1183 }
1184 
1185 static void emit_ldx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1186 {
1187 	emit_ldx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
1188 }
1189 
1190 static void emit_ldsx_r12(u8 **prog, u32 size, u32 dst_reg, u32 src_reg, int off)
1191 {
1192 	emit_ldsx_index(prog, size, dst_reg, src_reg, X86_REG_R12, off);
1193 }
1194 
1195 /* STX: *(u8*)(dst_reg + off) = src_reg */
1196 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1197 {
1198 	u8 *prog = *pprog;
1199 
1200 	switch (size) {
1201 	case BPF_B:
1202 		/* Emit 'mov byte ptr [rax + off], al' */
1203 		if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
1204 			/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
1205 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
1206 		else
1207 			EMIT1(0x88);
1208 		break;
1209 	case BPF_H:
1210 		if (is_ereg(dst_reg) || is_ereg(src_reg))
1211 			EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
1212 		else
1213 			EMIT2(0x66, 0x89);
1214 		break;
1215 	case BPF_W:
1216 		if (is_ereg(dst_reg) || is_ereg(src_reg))
1217 			EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
1218 		else
1219 			EMIT1(0x89);
1220 		break;
1221 	case BPF_DW:
1222 		EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
1223 		break;
1224 	}
1225 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1226 	*pprog = prog;
1227 }
1228 
1229 /* STX: *(u8*)(dst_reg + index_reg + off) = src_reg */
1230 static void emit_stx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
1231 {
1232 	u8 *prog = *pprog;
1233 
1234 	switch (size) {
1235 	case BPF_B:
1236 		/* mov byte ptr [rax + r12 + off], al */
1237 		EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x88);
1238 		break;
1239 	case BPF_H:
1240 		/* mov word ptr [rax + r12 + off], ax */
1241 		EMIT3(0x66, add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
1242 		break;
1243 	case BPF_W:
1244 		/* mov dword ptr [rax + r12 + 1], eax */
1245 		EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
1246 		break;
1247 	case BPF_DW:
1248 		/* mov qword ptr [rax + r12 + 1], rax */
1249 		EMIT2(add_3mod(0x48, dst_reg, src_reg, index_reg), 0x89);
1250 		break;
1251 	}
1252 	emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off);
1253 	*pprog = prog;
1254 }
1255 
1256 static void emit_stx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
1257 {
1258 	emit_stx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
1259 }
1260 
1261 /* ST: *(u8*)(dst_reg + index_reg + off) = imm32 */
1262 static void emit_st_index(u8 **pprog, u32 size, u32 dst_reg, u32 index_reg, int off, int imm)
1263 {
1264 	u8 *prog = *pprog;
1265 
1266 	switch (size) {
1267 	case BPF_B:
1268 		/* mov byte ptr [rax + r12 + off], imm8 */
1269 		EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC6);
1270 		break;
1271 	case BPF_H:
1272 		/* mov word ptr [rax + r12 + off], imm16 */
1273 		EMIT3(0x66, add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
1274 		break;
1275 	case BPF_W:
1276 		/* mov dword ptr [rax + r12 + 1], imm32 */
1277 		EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
1278 		break;
1279 	case BPF_DW:
1280 		/* mov qword ptr [rax + r12 + 1], imm32 */
1281 		EMIT2(add_3mod(0x48, dst_reg, 0, index_reg), 0xC7);
1282 		break;
1283 	}
1284 	emit_insn_suffix_SIB(&prog, dst_reg, 0, index_reg, off);
1285 	EMIT(imm, bpf_size_to_x86_bytes(size));
1286 	*pprog = prog;
1287 }
1288 
1289 static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
1290 {
1291 	emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
1292 }
1293 
1294 static void emit_store_stack_imm64(u8 **pprog, int reg, int stack_off, u64 imm64)
1295 {
1296 	/*
1297 	 * mov reg, imm64
1298 	 * mov QWORD PTR [rbp + stack_off], reg
1299 	 */
1300 	emit_mov_imm64(pprog, reg, imm64 >> 32, (u32) imm64);
1301 	emit_stx(pprog, BPF_DW, BPF_REG_FP, reg, stack_off);
1302 }
1303 
1304 static int emit_atomic_rmw(u8 **pprog, u32 atomic_op,
1305 			   u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
1306 {
1307 	u8 *prog = *pprog;
1308 
1309 	if (atomic_op != BPF_XCHG)
1310 		EMIT1(0xF0); /* lock prefix */
1311 
1312 	maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
1313 
1314 	/* emit opcode */
1315 	switch (atomic_op) {
1316 	case BPF_ADD:
1317 	case BPF_AND:
1318 	case BPF_OR:
1319 	case BPF_XOR:
1320 		/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
1321 		EMIT1(simple_alu_opcodes[atomic_op]);
1322 		break;
1323 	case BPF_ADD | BPF_FETCH:
1324 		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
1325 		EMIT2(0x0F, 0xC1);
1326 		break;
1327 	case BPF_XCHG:
1328 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
1329 		EMIT1(0x87);
1330 		break;
1331 	case BPF_CMPXCHG:
1332 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
1333 		EMIT2(0x0F, 0xB1);
1334 		break;
1335 	default:
1336 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
1337 		return -EFAULT;
1338 	}
1339 
1340 	emit_insn_suffix(&prog, dst_reg, src_reg, off);
1341 
1342 	*pprog = prog;
1343 	return 0;
1344 }
1345 
1346 static int emit_atomic_rmw_index(u8 **pprog, u32 atomic_op, u32 size,
1347 				 u32 dst_reg, u32 src_reg, u32 index_reg,
1348 				 int off)
1349 {
1350 	u8 *prog = *pprog;
1351 
1352 	if (atomic_op != BPF_XCHG)
1353 		EMIT1(0xF0); /* lock prefix */
1354 
1355 	switch (size) {
1356 	case BPF_W:
1357 		EMIT1(add_3mod(0x40, dst_reg, src_reg, index_reg));
1358 		break;
1359 	case BPF_DW:
1360 		EMIT1(add_3mod(0x48, dst_reg, src_reg, index_reg));
1361 		break;
1362 	default:
1363 		pr_err("bpf_jit: 1- and 2-byte RMW atomics are not supported\n");
1364 		return -EFAULT;
1365 	}
1366 
1367 	/* emit opcode */
1368 	switch (atomic_op) {
1369 	case BPF_ADD:
1370 	case BPF_AND:
1371 	case BPF_OR:
1372 	case BPF_XOR:
1373 		/* lock *(u32/u64*)(dst_reg + idx_reg + off) <op>= src_reg */
1374 		EMIT1(simple_alu_opcodes[atomic_op]);
1375 		break;
1376 	case BPF_ADD | BPF_FETCH:
1377 		/* src_reg = atomic_fetch_add(dst_reg + idx_reg + off, src_reg); */
1378 		EMIT2(0x0F, 0xC1);
1379 		break;
1380 	case BPF_XCHG:
1381 		/* src_reg = atomic_xchg(dst_reg + idx_reg + off, src_reg); */
1382 		EMIT1(0x87);
1383 		break;
1384 	case BPF_CMPXCHG:
1385 		/* r0 = atomic_cmpxchg(dst_reg + idx_reg + off, r0, src_reg); */
1386 		EMIT2(0x0F, 0xB1);
1387 		break;
1388 	default:
1389 		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
1390 		return -EFAULT;
1391 	}
1392 	emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off);
1393 	*pprog = prog;
1394 	return 0;
1395 }
1396 
1397 static int emit_atomic_ld_st(u8 **pprog, u32 atomic_op, u32 dst_reg,
1398 			     u32 src_reg, s16 off, u8 bpf_size)
1399 {
1400 	switch (atomic_op) {
1401 	case BPF_LOAD_ACQ:
1402 		/* dst_reg = smp_load_acquire(src_reg + off16) */
1403 		emit_ldx(pprog, bpf_size, dst_reg, src_reg, off);
1404 		break;
1405 	case BPF_STORE_REL:
1406 		/* smp_store_release(dst_reg + off16, src_reg) */
1407 		emit_stx(pprog, bpf_size, dst_reg, src_reg, off);
1408 		break;
1409 	default:
1410 		pr_err("bpf_jit: unknown atomic load/store opcode %02x\n",
1411 		       atomic_op);
1412 		return -EFAULT;
1413 	}
1414 
1415 	return 0;
1416 }
1417 
1418 static int emit_atomic_ld_st_index(u8 **pprog, u32 atomic_op, u32 size,
1419 				   u32 dst_reg, u32 src_reg, u32 index_reg,
1420 				   int off)
1421 {
1422 	switch (atomic_op) {
1423 	case BPF_LOAD_ACQ:
1424 		/* dst_reg = smp_load_acquire(src_reg + idx_reg + off16) */
1425 		emit_ldx_index(pprog, size, dst_reg, src_reg, index_reg, off);
1426 		break;
1427 	case BPF_STORE_REL:
1428 		/* smp_store_release(dst_reg + idx_reg + off16, src_reg) */
1429 		emit_stx_index(pprog, size, dst_reg, src_reg, index_reg, off);
1430 		break;
1431 	default:
1432 		pr_err("bpf_jit: unknown atomic load/store opcode %02x\n",
1433 		       atomic_op);
1434 		return -EFAULT;
1435 	}
1436 
1437 	return 0;
1438 }
1439 
1440 /*
1441  * Metadata encoding for exception handling in JITed code.
1442  *
1443  * Format of `fixup` and `data` fields in `struct exception_table_entry`:
1444  *
1445  * Bit layout of `fixup` (32-bit):
1446  *
1447  * +-----------+--------+-----------+---------+----------+
1448  * | 31        | 30-24  |   23-16   |   15-8  |    7-0   |
1449  * |           |        |           |         |          |
1450  * | ARENA_ACC | Unused | ARENA_REG | DST_REG | INSN_LEN |
1451  * +-----------+--------+-----------+---------+----------+
1452  *
1453  * - INSN_LEN (8 bits): Length of faulting insn (max x86 insn = 15 bytes (fits in 8 bits)).
1454  * - DST_REG  (8 bits): Offset of dst_reg from reg2pt_regs[] (max offset = 112 (fits in 8 bits)).
1455  *                      This is set to DONT_CLEAR if the insn is a store.
1456  * - ARENA_REG (8 bits): Offset of the register that is used to calculate the
1457  *                       address for load/store when accessing the arena region.
1458  * - ARENA_ACCESS (1 bit): This bit is set when the faulting instruction accessed the arena region.
1459  *
1460  * Bit layout of `data` (32-bit):
1461  *
1462  * +--------------+--------+--------------+
1463  * |	31-16	  |  15-8  |     7-0      |
1464  * |              |	   |              |
1465  * | ARENA_OFFSET | Unused |  EX_TYPE_BPF |
1466  * +--------------+--------+--------------+
1467  *
1468  * - ARENA_OFFSET (16 bits): Offset used to calculate the address for load/store when
1469  *                           accessing the arena region.
1470  */
1471 
1472 #define DONT_CLEAR 1
1473 #define FIXUP_INSN_LEN_MASK	GENMASK(7, 0)
1474 #define FIXUP_REG_MASK		GENMASK(15, 8)
1475 #define FIXUP_ARENA_REG_MASK	GENMASK(23, 16)
1476 #define FIXUP_ARENA_ACCESS	BIT(31)
1477 #define DATA_ARENA_OFFSET_MASK	GENMASK(31, 16)
1478 
1479 bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
1480 {
1481 	u32 reg = FIELD_GET(FIXUP_REG_MASK, x->fixup);
1482 	u32 insn_len = FIELD_GET(FIXUP_INSN_LEN_MASK, x->fixup);
1483 	bool is_arena = !!(x->fixup & FIXUP_ARENA_ACCESS);
1484 	bool is_write = (reg == DONT_CLEAR);
1485 	unsigned long addr;
1486 	s16 off;
1487 	u32 arena_reg;
1488 
1489 	if (is_arena) {
1490 		arena_reg = FIELD_GET(FIXUP_ARENA_REG_MASK, x->fixup);
1491 		off = FIELD_GET(DATA_ARENA_OFFSET_MASK, x->data);
1492 		addr = *(unsigned long *)((void *)regs + arena_reg) + off;
1493 		bpf_prog_report_arena_violation(is_write, addr, regs->ip);
1494 	}
1495 
1496 	/* jump over faulting load and clear dest register */
1497 	if (reg != DONT_CLEAR)
1498 		*(unsigned long *)((void *)regs + reg) = 0;
1499 	regs->ip += insn_len;
1500 
1501 	return true;
1502 }
1503 
1504 static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
1505 			     bool *regs_used)
1506 {
1507 	int i;
1508 
1509 	for (i = 1; i <= insn_cnt; i++, insn++) {
1510 		if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
1511 			regs_used[0] = true;
1512 		if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
1513 			regs_used[1] = true;
1514 		if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
1515 			regs_used[2] = true;
1516 		if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
1517 			regs_used[3] = true;
1518 	}
1519 }
1520 
1521 /* emit the 3-byte VEX prefix
1522  *
1523  * r: same as rex.r, extra bit for ModRM reg field
1524  * x: same as rex.x, extra bit for SIB index field
1525  * b: same as rex.b, extra bit for ModRM r/m, or SIB base
1526  * m: opcode map select, encoding escape bytes e.g. 0x0f38
1527  * w: same as rex.w (32 bit or 64 bit) or opcode specific
1528  * src_reg2: additional source reg (encoded as BPF reg)
1529  * l: vector length (128 bit or 256 bit) or reserved
1530  * pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
1531  */
1532 static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
1533 		      bool w, u8 src_reg2, bool l, u8 pp)
1534 {
1535 	u8 *prog = *pprog;
1536 	const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
1537 	u8 b1, b2;
1538 	u8 vvvv = reg2hex[src_reg2];
1539 
1540 	/* reg2hex gives only the lower 3 bit of vvvv */
1541 	if (is_ereg(src_reg2))
1542 		vvvv |= 1 << 3;
1543 
1544 	/*
1545 	 * 2nd byte of 3-byte VEX prefix
1546 	 * ~ means bit inverted encoding
1547 	 *
1548 	 *    7                           0
1549 	 *  +---+---+---+---+---+---+---+---+
1550 	 *  |~R |~X |~B |         m         |
1551 	 *  +---+---+---+---+---+---+---+---+
1552 	 */
1553 	b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
1554 	/*
1555 	 * 3rd byte of 3-byte VEX prefix
1556 	 *
1557 	 *    7                           0
1558 	 *  +---+---+---+---+---+---+---+---+
1559 	 *  | W |     ~vvvv     | L |   pp  |
1560 	 *  +---+---+---+---+---+---+---+---+
1561 	 */
1562 	b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
1563 
1564 	EMIT3(b0, b1, b2);
1565 	*pprog = prog;
1566 }
1567 
1568 /* emit BMI2 shift instruction */
1569 static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
1570 {
1571 	u8 *prog = *pprog;
1572 	bool r = is_ereg(dst_reg);
1573 	u8 m = 2; /* escape code 0f38 */
1574 
1575 	emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
1576 	EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
1577 	*pprog = prog;
1578 }
1579 
1580 static void emit_priv_frame_ptr(u8 **pprog, void __percpu *priv_frame_ptr)
1581 {
1582 	u8 *prog = *pprog;
1583 
1584 	/* movabs r9, priv_frame_ptr */
1585 	emit_mov_imm64(&prog, X86_REG_R9, (__force long) priv_frame_ptr >> 32,
1586 		       (u32) (__force long) priv_frame_ptr);
1587 
1588 #ifdef CONFIG_SMP
1589 	/* add <r9>, gs:[<off>] */
1590 	EMIT2(0x65, 0x4c);
1591 	EMIT3(0x03, 0x0c, 0x25);
1592 	EMIT((u32)(unsigned long)&this_cpu_off, 4);
1593 #endif
1594 
1595 	*pprog = prog;
1596 }
1597 
1598 #define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
1599 
1600 #define __LOAD_TCC_PTR(off)			\
1601 	EMIT3_off32(0x48, 0x8B, 0x85, off)
1602 /* mov rax, qword ptr [rbp - rounded_stack_depth - 16] */
1603 #define LOAD_TAIL_CALL_CNT_PTR(stack)				\
1604 	__LOAD_TCC_PTR(BPF_TAIL_CALL_CNT_PTR_STACK_OFF(stack))
1605 
1606 /* Memory size/value to protect private stack overflow/underflow */
1607 #define PRIV_STACK_GUARD_SZ    8
1608 #define PRIV_STACK_GUARD_VAL   0xEB9F12345678eb9fULL
1609 
1610 static int emit_spectre_bhb_barrier(u8 **pprog, u8 *ip,
1611 				    struct bpf_prog *bpf_prog)
1612 {
1613 	u8 *prog = *pprog;
1614 	u8 *func;
1615 
1616 	if (cpu_feature_enabled(X86_FEATURE_CLEAR_BHB_LOOP)) {
1617 		/* The clearing sequence clobbers eax and ecx. */
1618 		EMIT1(0x50); /* push rax */
1619 		EMIT1(0x51); /* push rcx */
1620 		ip += 2;
1621 
1622 		func = (u8 *)clear_bhb_loop;
1623 		ip += x86_call_depth_emit_accounting(&prog, func, ip);
1624 
1625 		if (emit_call(&prog, func, ip))
1626 			return -EINVAL;
1627 		EMIT1(0x59); /* pop rcx */
1628 		EMIT1(0x58); /* pop rax */
1629 	}
1630 	/* Insert IBHF instruction */
1631 	if ((cpu_feature_enabled(X86_FEATURE_CLEAR_BHB_LOOP) &&
1632 	     cpu_feature_enabled(X86_FEATURE_HYPERVISOR)) ||
1633 	    cpu_feature_enabled(X86_FEATURE_CLEAR_BHB_HW)) {
1634 		/*
1635 		 * Add an Indirect Branch History Fence (IBHF). IBHF acts as a
1636 		 * fence preventing branch history from before the fence from
1637 		 * affecting indirect branches after the fence. This is
1638 		 * specifically used in cBPF jitted code to prevent Intra-mode
1639 		 * BHI attacks. The IBHF instruction is designed to be a NOP on
1640 		 * hardware that doesn't need or support it.  The REP and REX.W
1641 		 * prefixes are required by the microcode, and they also ensure
1642 		 * that the NOP is unlikely to be used in existing code.
1643 		 *
1644 		 * IBHF is not a valid instruction in 32-bit mode.
1645 		 */
1646 		EMIT5(0xF3, 0x48, 0x0F, 0x1E, 0xF8); /* ibhf */
1647 	}
1648 	*pprog = prog;
1649 	return 0;
1650 }
1651 
1652 static int do_jit(struct bpf_verifier_env *env, struct bpf_prog *bpf_prog, int *addrs, u8 *image,
1653 		  u8 *rw_image, int oldproglen, struct jit_context *ctx, bool jmp_padding)
1654 {
1655 	bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
1656 	struct bpf_insn *insn = bpf_prog->insnsi;
1657 	bool callee_regs_used[4] = {};
1658 	int insn_cnt = bpf_prog->len;
1659 	bool seen_exit = false;
1660 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
1661 	void __percpu *priv_frame_ptr = NULL;
1662 	u64 arena_vm_start, user_vm_start;
1663 	void __percpu *priv_stack_ptr;
1664 	int i, excnt = 0;
1665 	int ilen, proglen = 0;
1666 	u8 *ip, *prog = temp;
1667 	u32 stack_depth;
1668 	int err;
1669 
1670 	stack_depth = bpf_prog->aux->stack_depth;
1671 	priv_stack_ptr = bpf_prog->aux->priv_stack_ptr;
1672 	if (priv_stack_ptr) {
1673 		priv_frame_ptr = priv_stack_ptr + PRIV_STACK_GUARD_SZ + round_up(stack_depth, 8);
1674 		stack_depth = 0;
1675 	}
1676 
1677 	arena_vm_start = bpf_arena_get_kern_vm_start(bpf_prog->aux->arena);
1678 	user_vm_start = bpf_arena_get_user_vm_start(bpf_prog->aux->arena);
1679 
1680 	detect_reg_usage(insn, insn_cnt, callee_regs_used);
1681 
1682 	emit_prologue(&prog, image, stack_depth,
1683 		      bpf_prog_was_classic(bpf_prog), tail_call_reachable,
1684 		      bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
1685 
1686 	bpf_prog->aux->ksym.fp_start = prog - temp;
1687 
1688 	/* Exception callback will clobber callee regs for its own use, and
1689 	 * restore the original callee regs from main prog's stack frame.
1690 	 */
1691 	if (bpf_prog->aux->exception_boundary) {
1692 		/* We also need to save r12, which is not mapped to any BPF
1693 		 * register, as we throw after entry into the kernel, which may
1694 		 * overwrite r12.
1695 		 */
1696 		push_r12(&prog);
1697 		push_callee_regs(&prog, all_callee_regs_used);
1698 	} else {
1699 		if (arena_vm_start)
1700 			push_r12(&prog);
1701 		push_callee_regs(&prog, callee_regs_used);
1702 	}
1703 	if (arena_vm_start)
1704 		emit_mov_imm64(&prog, X86_REG_R12,
1705 			       arena_vm_start >> 32, (u32) arena_vm_start);
1706 
1707 	if (priv_frame_ptr)
1708 		emit_priv_frame_ptr(&prog, priv_frame_ptr);
1709 
1710 	ilen = prog - temp;
1711 	if (rw_image)
1712 		memcpy(rw_image + proglen, temp, ilen);
1713 	proglen += ilen;
1714 	addrs[0] = proglen;
1715 	prog = temp;
1716 
1717 	for (i = 1; i <= insn_cnt; i++, insn++) {
1718 		const s32 imm32 = insn->imm;
1719 		u32 dst_reg = insn->dst_reg;
1720 		u32 src_reg = insn->src_reg;
1721 		u8 b2 = 0, b3 = 0;
1722 		u8 *start_of_ldx;
1723 		s64 jmp_offset;
1724 		s16 insn_off;
1725 		u8 jmp_cond;
1726 		u8 *func;
1727 		int nops;
1728 
1729 		if (priv_frame_ptr) {
1730 			if (src_reg == BPF_REG_FP)
1731 				src_reg = X86_REG_R9;
1732 
1733 			if (dst_reg == BPF_REG_FP)
1734 				dst_reg = X86_REG_R9;
1735 		}
1736 
1737 		if (bpf_insn_is_indirect_target(env, bpf_prog, i - 1))
1738 			EMIT_ENDBR();
1739 
1740 		ip = image + addrs[i - 1] + (prog - temp);
1741 
1742 		switch (insn->code) {
1743 			/* ALU */
1744 		case BPF_ALU | BPF_ADD | BPF_X:
1745 		case BPF_ALU | BPF_SUB | BPF_X:
1746 		case BPF_ALU | BPF_AND | BPF_X:
1747 		case BPF_ALU | BPF_OR | BPF_X:
1748 		case BPF_ALU | BPF_XOR | BPF_X:
1749 		case BPF_ALU64 | BPF_ADD | BPF_X:
1750 		case BPF_ALU64 | BPF_SUB | BPF_X:
1751 		case BPF_ALU64 | BPF_AND | BPF_X:
1752 		case BPF_ALU64 | BPF_OR | BPF_X:
1753 		case BPF_ALU64 | BPF_XOR | BPF_X:
1754 			maybe_emit_mod(&prog, dst_reg, src_reg,
1755 				       BPF_CLASS(insn->code) == BPF_ALU64);
1756 			b2 = simple_alu_opcodes[BPF_OP(insn->code)];
1757 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
1758 			break;
1759 
1760 		case BPF_ALU64 | BPF_MOV | BPF_X:
1761 			if (insn_is_cast_user(insn)) {
1762 				if (dst_reg != src_reg)
1763 					/* 32-bit mov */
1764 					emit_mov_reg(&prog, false, dst_reg, src_reg);
1765 				/* shl dst_reg, 32 */
1766 				maybe_emit_1mod(&prog, dst_reg, true);
1767 				EMIT3(0xC1, add_1reg(0xE0, dst_reg), 32);
1768 
1769 				/* or dst_reg, user_vm_start */
1770 				maybe_emit_1mod(&prog, dst_reg, true);
1771 				if (is_axreg(dst_reg))
1772 					EMIT1_off32(0x0D,  user_vm_start >> 32);
1773 				else
1774 					EMIT2_off32(0x81, add_1reg(0xC8, dst_reg),  user_vm_start >> 32);
1775 
1776 				/* rol dst_reg, 32 */
1777 				maybe_emit_1mod(&prog, dst_reg, true);
1778 				EMIT3(0xC1, add_1reg(0xC0, dst_reg), 32);
1779 
1780 				/* xor r11, r11 */
1781 				EMIT3(0x4D, 0x31, 0xDB);
1782 
1783 				/* test dst_reg32, dst_reg32; check if lower 32-bit are zero */
1784 				maybe_emit_mod(&prog, dst_reg, dst_reg, false);
1785 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1786 
1787 				/* cmove r11, dst_reg; if so, set dst_reg to zero */
1788 				/* WARNING: Intel swapped src/dst register encoding in CMOVcc !!! */
1789 				maybe_emit_mod(&prog, AUX_REG, dst_reg, true);
1790 				EMIT3(0x0F, 0x44, add_2reg(0xC0, AUX_REG, dst_reg));
1791 				break;
1792 			} else if (insn_is_mov_percpu_addr(insn)) {
1793 				/* mov <dst>, <src> (if necessary) */
1794 				EMIT_mov(dst_reg, src_reg);
1795 #ifdef CONFIG_SMP
1796 				/* add <dst>, gs:[<off>] */
1797 				EMIT2(0x65, add_1mod(0x48, dst_reg));
1798 				EMIT3(0x03, add_2reg(0x04, 0, dst_reg), 0x25);
1799 				EMIT((u32)(unsigned long)&this_cpu_off, 4);
1800 #endif
1801 				break;
1802 			}
1803 			fallthrough;
1804 		case BPF_ALU | BPF_MOV | BPF_X:
1805 			if (insn->off == 0)
1806 				emit_mov_reg(&prog,
1807 					     BPF_CLASS(insn->code) == BPF_ALU64,
1808 					     dst_reg, src_reg);
1809 			else
1810 				emit_movsx_reg(&prog, insn->off,
1811 					       BPF_CLASS(insn->code) == BPF_ALU64,
1812 					       dst_reg, src_reg);
1813 			break;
1814 
1815 			/* neg dst */
1816 		case BPF_ALU | BPF_NEG:
1817 		case BPF_ALU64 | BPF_NEG:
1818 			maybe_emit_1mod(&prog, dst_reg,
1819 					BPF_CLASS(insn->code) == BPF_ALU64);
1820 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
1821 			break;
1822 
1823 		case BPF_ALU | BPF_ADD | BPF_K:
1824 		case BPF_ALU | BPF_SUB | BPF_K:
1825 		case BPF_ALU | BPF_AND | BPF_K:
1826 		case BPF_ALU | BPF_OR | BPF_K:
1827 		case BPF_ALU | BPF_XOR | BPF_K:
1828 		case BPF_ALU64 | BPF_ADD | BPF_K:
1829 		case BPF_ALU64 | BPF_SUB | BPF_K:
1830 		case BPF_ALU64 | BPF_AND | BPF_K:
1831 		case BPF_ALU64 | BPF_OR | BPF_K:
1832 		case BPF_ALU64 | BPF_XOR | BPF_K:
1833 			maybe_emit_1mod(&prog, dst_reg,
1834 					BPF_CLASS(insn->code) == BPF_ALU64);
1835 
1836 			/*
1837 			 * b3 holds 'normal' opcode, b2 short form only valid
1838 			 * in case dst is eax/rax.
1839 			 */
1840 			switch (BPF_OP(insn->code)) {
1841 			case BPF_ADD:
1842 				b3 = 0xC0;
1843 				b2 = 0x05;
1844 				break;
1845 			case BPF_SUB:
1846 				b3 = 0xE8;
1847 				b2 = 0x2D;
1848 				break;
1849 			case BPF_AND:
1850 				b3 = 0xE0;
1851 				b2 = 0x25;
1852 				break;
1853 			case BPF_OR:
1854 				b3 = 0xC8;
1855 				b2 = 0x0D;
1856 				break;
1857 			case BPF_XOR:
1858 				b3 = 0xF0;
1859 				b2 = 0x35;
1860 				break;
1861 			}
1862 
1863 			if (is_imm8(imm32))
1864 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
1865 			else if (is_axreg(dst_reg))
1866 				EMIT1_off32(b2, imm32);
1867 			else
1868 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
1869 			break;
1870 
1871 		case BPF_ALU64 | BPF_MOV | BPF_K:
1872 		case BPF_ALU | BPF_MOV | BPF_K:
1873 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
1874 				       dst_reg, imm32);
1875 			break;
1876 
1877 		case BPF_LD | BPF_IMM | BPF_DW:
1878 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
1879 			insn++;
1880 			i++;
1881 			break;
1882 
1883 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
1884 		case BPF_ALU | BPF_MOD | BPF_X:
1885 		case BPF_ALU | BPF_DIV | BPF_X:
1886 		case BPF_ALU | BPF_MOD | BPF_K:
1887 		case BPF_ALU | BPF_DIV | BPF_K:
1888 		case BPF_ALU64 | BPF_MOD | BPF_X:
1889 		case BPF_ALU64 | BPF_DIV | BPF_X:
1890 		case BPF_ALU64 | BPF_MOD | BPF_K:
1891 		case BPF_ALU64 | BPF_DIV | BPF_K: {
1892 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
1893 
1894 			if (dst_reg != BPF_REG_0)
1895 				EMIT1(0x50); /* push rax */
1896 			if (dst_reg != BPF_REG_3)
1897 				EMIT1(0x52); /* push rdx */
1898 
1899 			if (BPF_SRC(insn->code) == BPF_X) {
1900 				if (src_reg == BPF_REG_0 ||
1901 				    src_reg == BPF_REG_3) {
1902 					/* mov r11, src_reg */
1903 					EMIT_mov(AUX_REG, src_reg);
1904 					src_reg = AUX_REG;
1905 				}
1906 			} else {
1907 				/* mov r11, imm32 */
1908 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
1909 				src_reg = AUX_REG;
1910 			}
1911 
1912 			if (dst_reg != BPF_REG_0)
1913 				/* mov rax, dst_reg */
1914 				emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
1915 
1916 			if (insn->off == 0) {
1917 				/*
1918 				 * xor edx, edx
1919 				 * equivalent to 'xor rdx, rdx', but one byte less
1920 				 */
1921 				EMIT2(0x31, 0xd2);
1922 
1923 				/* div src_reg */
1924 				maybe_emit_1mod(&prog, src_reg, is64);
1925 				EMIT2(0xF7, add_1reg(0xF0, src_reg));
1926 			} else {
1927 				if (BPF_CLASS(insn->code) == BPF_ALU)
1928 					EMIT1(0x99); /* cdq */
1929 				else
1930 					EMIT2(0x48, 0x99); /* cqo */
1931 
1932 				/* idiv src_reg */
1933 				maybe_emit_1mod(&prog, src_reg, is64);
1934 				EMIT2(0xF7, add_1reg(0xF8, src_reg));
1935 			}
1936 
1937 			if (BPF_OP(insn->code) == BPF_MOD &&
1938 			    dst_reg != BPF_REG_3)
1939 				/* mov dst_reg, rdx */
1940 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
1941 			else if (BPF_OP(insn->code) == BPF_DIV &&
1942 				 dst_reg != BPF_REG_0)
1943 				/* mov dst_reg, rax */
1944 				emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
1945 
1946 			if (dst_reg != BPF_REG_3)
1947 				EMIT1(0x5A); /* pop rdx */
1948 			if (dst_reg != BPF_REG_0)
1949 				EMIT1(0x58); /* pop rax */
1950 			break;
1951 		}
1952 
1953 		case BPF_ALU | BPF_MUL | BPF_K:
1954 		case BPF_ALU64 | BPF_MUL | BPF_K:
1955 			maybe_emit_mod(&prog, dst_reg, dst_reg,
1956 				       BPF_CLASS(insn->code) == BPF_ALU64);
1957 
1958 			if (is_imm8(imm32))
1959 				/* imul dst_reg, dst_reg, imm8 */
1960 				EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
1961 				      imm32);
1962 			else
1963 				/* imul dst_reg, dst_reg, imm32 */
1964 				EMIT2_off32(0x69,
1965 					    add_2reg(0xC0, dst_reg, dst_reg),
1966 					    imm32);
1967 			break;
1968 
1969 		case BPF_ALU | BPF_MUL | BPF_X:
1970 		case BPF_ALU64 | BPF_MUL | BPF_X:
1971 			maybe_emit_mod(&prog, src_reg, dst_reg,
1972 				       BPF_CLASS(insn->code) == BPF_ALU64);
1973 
1974 			/* imul dst_reg, src_reg */
1975 			EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
1976 			break;
1977 
1978 			/* Shifts */
1979 		case BPF_ALU | BPF_LSH | BPF_K:
1980 		case BPF_ALU | BPF_RSH | BPF_K:
1981 		case BPF_ALU | BPF_ARSH | BPF_K:
1982 		case BPF_ALU64 | BPF_LSH | BPF_K:
1983 		case BPF_ALU64 | BPF_RSH | BPF_K:
1984 		case BPF_ALU64 | BPF_ARSH | BPF_K:
1985 			maybe_emit_1mod(&prog, dst_reg,
1986 					BPF_CLASS(insn->code) == BPF_ALU64);
1987 
1988 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
1989 			if (imm32 == 1)
1990 				EMIT2(0xD1, add_1reg(b3, dst_reg));
1991 			else
1992 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
1993 			break;
1994 
1995 		case BPF_ALU | BPF_LSH | BPF_X:
1996 		case BPF_ALU | BPF_RSH | BPF_X:
1997 		case BPF_ALU | BPF_ARSH | BPF_X:
1998 		case BPF_ALU64 | BPF_LSH | BPF_X:
1999 		case BPF_ALU64 | BPF_RSH | BPF_X:
2000 		case BPF_ALU64 | BPF_ARSH | BPF_X:
2001 			/* BMI2 shifts aren't better when shift count is already in rcx */
2002 			if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
2003 				/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
2004 				bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
2005 				u8 op;
2006 
2007 				switch (BPF_OP(insn->code)) {
2008 				case BPF_LSH:
2009 					op = 1; /* prefix 0x66 */
2010 					break;
2011 				case BPF_RSH:
2012 					op = 3; /* prefix 0xf2 */
2013 					break;
2014 				case BPF_ARSH:
2015 					op = 2; /* prefix 0xf3 */
2016 					break;
2017 				}
2018 
2019 				emit_shiftx(&prog, dst_reg, src_reg, w, op);
2020 
2021 				break;
2022 			}
2023 
2024 			if (src_reg != BPF_REG_4) { /* common case */
2025 				/* Check for bad case when dst_reg == rcx */
2026 				if (dst_reg == BPF_REG_4) {
2027 					/* mov r11, dst_reg */
2028 					EMIT_mov(AUX_REG, dst_reg);
2029 					dst_reg = AUX_REG;
2030 				} else {
2031 					EMIT1(0x51); /* push rcx */
2032 				}
2033 				/* mov rcx, src_reg */
2034 				EMIT_mov(BPF_REG_4, src_reg);
2035 			}
2036 
2037 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
2038 			maybe_emit_1mod(&prog, dst_reg,
2039 					BPF_CLASS(insn->code) == BPF_ALU64);
2040 
2041 			b3 = simple_alu_opcodes[BPF_OP(insn->code)];
2042 			EMIT2(0xD3, add_1reg(b3, dst_reg));
2043 
2044 			if (src_reg != BPF_REG_4) {
2045 				if (insn->dst_reg == BPF_REG_4)
2046 					/* mov dst_reg, r11 */
2047 					EMIT_mov(insn->dst_reg, AUX_REG);
2048 				else
2049 					EMIT1(0x59); /* pop rcx */
2050 			}
2051 
2052 			break;
2053 
2054 		case BPF_ALU | BPF_END | BPF_FROM_BE:
2055 		case BPF_ALU64 | BPF_END | BPF_FROM_LE:
2056 			switch (imm32) {
2057 			case 16:
2058 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
2059 				EMIT1(0x66);
2060 				if (is_ereg(dst_reg))
2061 					EMIT1(0x41);
2062 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
2063 
2064 				/* Emit 'movzwl eax, ax' */
2065 				if (is_ereg(dst_reg))
2066 					EMIT3(0x45, 0x0F, 0xB7);
2067 				else
2068 					EMIT2(0x0F, 0xB7);
2069 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
2070 				break;
2071 			case 32:
2072 				/* Emit 'bswap eax' to swap lower 4 bytes */
2073 				if (is_ereg(dst_reg))
2074 					EMIT2(0x41, 0x0F);
2075 				else
2076 					EMIT1(0x0F);
2077 				EMIT1(add_1reg(0xC8, dst_reg));
2078 				break;
2079 			case 64:
2080 				/* Emit 'bswap rax' to swap 8 bytes */
2081 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
2082 				      add_1reg(0xC8, dst_reg));
2083 				break;
2084 			}
2085 			break;
2086 
2087 		case BPF_ALU | BPF_END | BPF_FROM_LE:
2088 			switch (imm32) {
2089 			case 16:
2090 				/*
2091 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
2092 				 * into 64 bit
2093 				 */
2094 				if (is_ereg(dst_reg))
2095 					EMIT3(0x45, 0x0F, 0xB7);
2096 				else
2097 					EMIT2(0x0F, 0xB7);
2098 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
2099 				break;
2100 			case 32:
2101 				/* Emit 'mov eax, eax' to clear upper 32-bits */
2102 				if (is_ereg(dst_reg))
2103 					EMIT1(0x45);
2104 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
2105 				break;
2106 			case 64:
2107 				/* nop */
2108 				break;
2109 			}
2110 			break;
2111 
2112 			/* speculation barrier */
2113 		case BPF_ST | BPF_NOSPEC:
2114 			EMIT_LFENCE();
2115 			break;
2116 
2117 			/* ST: *(u8*)(dst_reg + off) = imm */
2118 		case BPF_ST | BPF_MEM | BPF_B:
2119 			if (is_ereg(dst_reg))
2120 				EMIT2(0x41, 0xC6);
2121 			else
2122 				EMIT1(0xC6);
2123 			goto st;
2124 		case BPF_ST | BPF_MEM | BPF_H:
2125 			if (is_ereg(dst_reg))
2126 				EMIT3(0x66, 0x41, 0xC7);
2127 			else
2128 				EMIT2(0x66, 0xC7);
2129 			goto st;
2130 		case BPF_ST | BPF_MEM | BPF_W:
2131 			if (is_ereg(dst_reg))
2132 				EMIT2(0x41, 0xC7);
2133 			else
2134 				EMIT1(0xC7);
2135 			goto st;
2136 		case BPF_ST | BPF_MEM | BPF_DW:
2137 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
2138 
2139 st:			if (is_imm8(insn->off))
2140 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
2141 			else
2142 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
2143 
2144 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
2145 			break;
2146 
2147 			/* STX: *(u8*)(dst_reg + off) = src_reg */
2148 		case BPF_STX | BPF_MEM | BPF_B:
2149 		case BPF_STX | BPF_MEM | BPF_H:
2150 		case BPF_STX | BPF_MEM | BPF_W:
2151 		case BPF_STX | BPF_MEM | BPF_DW:
2152 			emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
2153 			break;
2154 
2155 		case BPF_ST | BPF_PROBE_MEM32 | BPF_B:
2156 		case BPF_ST | BPF_PROBE_MEM32 | BPF_H:
2157 		case BPF_ST | BPF_PROBE_MEM32 | BPF_W:
2158 		case BPF_ST | BPF_PROBE_MEM32 | BPF_DW:
2159 			start_of_ldx = prog;
2160 			emit_st_r12(&prog, BPF_SIZE(insn->code), dst_reg, insn->off, insn->imm);
2161 			goto populate_extable;
2162 
2163 			/* LDX: dst_reg = *(u8*)(src_reg + r12 + off) */
2164 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_B:
2165 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
2166 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
2167 		case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
2168 		case BPF_LDX | BPF_PROBE_MEM32SX | BPF_B:
2169 		case BPF_LDX | BPF_PROBE_MEM32SX | BPF_H:
2170 		case BPF_LDX | BPF_PROBE_MEM32SX | BPF_W:
2171 		case BPF_STX | BPF_PROBE_MEM32 | BPF_B:
2172 		case BPF_STX | BPF_PROBE_MEM32 | BPF_H:
2173 		case BPF_STX | BPF_PROBE_MEM32 | BPF_W:
2174 		case BPF_STX | BPF_PROBE_MEM32 | BPF_DW:
2175 			start_of_ldx = prog;
2176 			if (BPF_CLASS(insn->code) == BPF_LDX) {
2177 				if (BPF_MODE(insn->code) == BPF_PROBE_MEM32SX)
2178 					emit_ldsx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
2179 				else
2180 					emit_ldx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
2181 			} else {
2182 				emit_stx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
2183 			}
2184 populate_extable:
2185 			{
2186 				struct exception_table_entry *ex;
2187 				u8 *_insn = image + proglen + (start_of_ldx - temp);
2188 				u32 arena_reg, fixup_reg;
2189 				s64 delta;
2190 
2191 				if (!bpf_prog->aux->extable)
2192 					break;
2193 
2194 				if (excnt >= bpf_prog->aux->num_exentries) {
2195 					pr_err("mem32 extable bug\n");
2196 					return -EFAULT;
2197 				}
2198 				ex = &bpf_prog->aux->extable[excnt++];
2199 
2200 				delta = _insn - (u8 *)&ex->insn;
2201 				/* switch ex to rw buffer for writes */
2202 				ex = (void *)rw_image + ((void *)ex - (void *)image);
2203 
2204 				ex->insn = delta;
2205 
2206 				ex->data = EX_TYPE_BPF;
2207 
2208 				/*
2209 				 * src_reg/dst_reg holds the address in the arena region with upper
2210 				 * 32-bits being zero because of a preceding addr_space_cast(r<n>,
2211 				 * 0x0, 0x1) instruction. This address is adjusted with the addition
2212 				 * of arena_vm_start (see the implementation of BPF_PROBE_MEM32 and
2213 				 * BPF_PROBE_ATOMIC) before being used for the memory access. Pass
2214 				 * the reg holding the unmodified 32-bit address to
2215 				 * ex_handler_bpf().
2216 				 */
2217 				if (BPF_CLASS(insn->code) == BPF_LDX) {
2218 					arena_reg = reg2pt_regs[src_reg];
2219 					fixup_reg = reg2pt_regs[dst_reg];
2220 				} else {
2221 					arena_reg = reg2pt_regs[dst_reg];
2222 					fixup_reg = DONT_CLEAR;
2223 				}
2224 
2225 				ex->fixup = FIELD_PREP(FIXUP_INSN_LEN_MASK, prog - start_of_ldx) |
2226 					    FIELD_PREP(FIXUP_ARENA_REG_MASK, arena_reg) |
2227 					    FIELD_PREP(FIXUP_REG_MASK, fixup_reg);
2228 				ex->fixup |= FIXUP_ARENA_ACCESS;
2229 
2230 				ex->data |= FIELD_PREP(DATA_ARENA_OFFSET_MASK, insn->off);
2231 			}
2232 			break;
2233 
2234 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
2235 		case BPF_LDX | BPF_MEM | BPF_B:
2236 		case BPF_LDX | BPF_PROBE_MEM | BPF_B:
2237 		case BPF_LDX | BPF_MEM | BPF_H:
2238 		case BPF_LDX | BPF_PROBE_MEM | BPF_H:
2239 		case BPF_LDX | BPF_MEM | BPF_W:
2240 		case BPF_LDX | BPF_PROBE_MEM | BPF_W:
2241 		case BPF_LDX | BPF_MEM | BPF_DW:
2242 		case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
2243 			/* LDXS: dst_reg = *(s8*)(src_reg + off) */
2244 		case BPF_LDX | BPF_MEMSX | BPF_B:
2245 		case BPF_LDX | BPF_MEMSX | BPF_H:
2246 		case BPF_LDX | BPF_MEMSX | BPF_W:
2247 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
2248 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
2249 		case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
2250 			insn_off = insn->off;
2251 
2252 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
2253 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
2254 				/* Conservatively check that src_reg + insn->off is a kernel address:
2255 				 *   src_reg + insn->off > TASK_SIZE_MAX + PAGE_SIZE
2256 				 *   and
2257 				 *   src_reg + insn->off < VSYSCALL_ADDR
2258 				 */
2259 
2260 				u64 limit = TASK_SIZE_MAX + PAGE_SIZE - VSYSCALL_ADDR;
2261 				u8 *end_of_jmp;
2262 
2263 				/* movabsq r10, VSYSCALL_ADDR */
2264 				emit_mov_imm64(&prog, BPF_REG_AX, (long)VSYSCALL_ADDR >> 32,
2265 					       (u32)(long)VSYSCALL_ADDR);
2266 
2267 				/* mov src_reg, r11 */
2268 				EMIT_mov(AUX_REG, src_reg);
2269 
2270 				if (insn->off) {
2271 					/* add r11, insn->off */
2272 					maybe_emit_1mod(&prog, AUX_REG, true);
2273 					EMIT2_off32(0x81, add_1reg(0xC0, AUX_REG), insn->off);
2274 				}
2275 
2276 				/* sub r11, r10 */
2277 				maybe_emit_mod(&prog, AUX_REG, BPF_REG_AX, true);
2278 				EMIT2(0x29, add_2reg(0xC0, AUX_REG, BPF_REG_AX));
2279 
2280 				/* movabsq r10, limit */
2281 				emit_mov_imm64(&prog, BPF_REG_AX, (long)limit >> 32,
2282 					       (u32)(long)limit);
2283 
2284 				/* cmp r10, r11 */
2285 				maybe_emit_mod(&prog, AUX_REG, BPF_REG_AX, true);
2286 				EMIT2(0x39, add_2reg(0xC0, AUX_REG, BPF_REG_AX));
2287 
2288 				/* if unsigned '>', goto load */
2289 				EMIT2(X86_JA, 0);
2290 				end_of_jmp = prog;
2291 
2292 				/* xor dst_reg, dst_reg */
2293 				emit_mov_imm32(&prog, false, dst_reg, 0);
2294 				/* jmp byte_after_ldx */
2295 				EMIT2(0xEB, 0);
2296 
2297 				/* populate jmp_offset for JAE above to jump to start_of_ldx */
2298 				start_of_ldx = prog;
2299 				end_of_jmp[-1] = start_of_ldx - end_of_jmp;
2300 			}
2301 			if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
2302 			    BPF_MODE(insn->code) == BPF_MEMSX)
2303 				emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
2304 			else
2305 				emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
2306 			if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
2307 			    BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
2308 				struct exception_table_entry *ex;
2309 				u8 *_insn = image + proglen + (start_of_ldx - temp);
2310 				s64 delta;
2311 
2312 				/* populate jmp_offset for JMP above */
2313 				start_of_ldx[-1] = prog - start_of_ldx;
2314 
2315 				if (!bpf_prog->aux->extable)
2316 					break;
2317 
2318 				if (excnt >= bpf_prog->aux->num_exentries) {
2319 					pr_err("ex gen bug\n");
2320 					return -EFAULT;
2321 				}
2322 				ex = &bpf_prog->aux->extable[excnt++];
2323 
2324 				delta = _insn - (u8 *)&ex->insn;
2325 				if (!is_simm32(delta)) {
2326 					pr_err("extable->insn doesn't fit into 32-bit\n");
2327 					return -EFAULT;
2328 				}
2329 				/* switch ex to rw buffer for writes */
2330 				ex = (void *)rw_image + ((void *)ex - (void *)image);
2331 
2332 				ex->insn = delta;
2333 
2334 				ex->data = EX_TYPE_BPF;
2335 
2336 				if (dst_reg > BPF_REG_9) {
2337 					pr_err("verifier error\n");
2338 					return -EFAULT;
2339 				}
2340 				/*
2341 				 * Compute size of x86 insn and its target dest x86 register.
2342 				 * ex_handler_bpf() will use lower 8 bits to adjust
2343 				 * pt_regs->ip to jump over this x86 instruction
2344 				 * and upper bits to figure out which pt_regs to zero out.
2345 				 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
2346 				 * of 4 bytes will be ignored and rbx will be zero inited.
2347 				 */
2348 				ex->fixup = FIELD_PREP(FIXUP_INSN_LEN_MASK, prog - start_of_ldx) |
2349 					    FIELD_PREP(FIXUP_REG_MASK, reg2pt_regs[dst_reg]);
2350 			}
2351 			break;
2352 
2353 		case BPF_STX | BPF_ATOMIC | BPF_B:
2354 		case BPF_STX | BPF_ATOMIC | BPF_H:
2355 			if (!bpf_atomic_is_load_store(insn)) {
2356 				pr_err("bpf_jit: 1- and 2-byte RMW atomics are not supported\n");
2357 				return -EFAULT;
2358 			}
2359 			fallthrough;
2360 		case BPF_STX | BPF_ATOMIC | BPF_W:
2361 		case BPF_STX | BPF_ATOMIC | BPF_DW:
2362 			if (insn->imm == (BPF_AND | BPF_FETCH) ||
2363 			    insn->imm == (BPF_OR | BPF_FETCH) ||
2364 			    insn->imm == (BPF_XOR | BPF_FETCH)) {
2365 				bool is64 = BPF_SIZE(insn->code) == BPF_DW;
2366 				u32 real_src_reg = src_reg;
2367 				u32 real_dst_reg = dst_reg;
2368 				u8 *branch_target;
2369 
2370 				/*
2371 				 * Can't be implemented with a single x86 insn.
2372 				 * Need to do a CMPXCHG loop.
2373 				 */
2374 
2375 				/* Will need RAX as a CMPXCHG operand so save R0 */
2376 				emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
2377 				if (src_reg == BPF_REG_0)
2378 					real_src_reg = BPF_REG_AX;
2379 				if (dst_reg == BPF_REG_0)
2380 					real_dst_reg = BPF_REG_AX;
2381 
2382 				branch_target = prog;
2383 				/* Load old value */
2384 				emit_ldx(&prog, BPF_SIZE(insn->code),
2385 					 BPF_REG_0, real_dst_reg, insn->off);
2386 				/*
2387 				 * Perform the (commutative) operation locally,
2388 				 * put the result in the AUX_REG.
2389 				 */
2390 				emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
2391 				maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
2392 				EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
2393 				      add_2reg(0xC0, AUX_REG, real_src_reg));
2394 				/* Attempt to swap in new value */
2395 				err = emit_atomic_rmw(&prog, BPF_CMPXCHG,
2396 						      real_dst_reg, AUX_REG,
2397 						      insn->off,
2398 						      BPF_SIZE(insn->code));
2399 				if (WARN_ON(err))
2400 					return err;
2401 				/*
2402 				 * ZF tells us whether we won the race. If it's
2403 				 * cleared we need to try again.
2404 				 */
2405 				EMIT2(X86_JNE, -(prog - branch_target) - 2);
2406 				/* Return the pre-modification value */
2407 				emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
2408 				/* Restore R0 after clobbering RAX */
2409 				emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
2410 				break;
2411 			}
2412 
2413 			if (bpf_atomic_is_load_store(insn))
2414 				err = emit_atomic_ld_st(&prog, insn->imm, dst_reg, src_reg,
2415 							insn->off, BPF_SIZE(insn->code));
2416 			else
2417 				err = emit_atomic_rmw(&prog, insn->imm, dst_reg, src_reg,
2418 						      insn->off, BPF_SIZE(insn->code));
2419 			if (err)
2420 				return err;
2421 			break;
2422 
2423 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_B:
2424 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_H:
2425 			if (!bpf_atomic_is_load_store(insn)) {
2426 				pr_err("bpf_jit: 1- and 2-byte RMW atomics are not supported\n");
2427 				return -EFAULT;
2428 			}
2429 			fallthrough;
2430 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_W:
2431 		case BPF_STX | BPF_PROBE_ATOMIC | BPF_DW:
2432 			start_of_ldx = prog;
2433 
2434 			if (bpf_atomic_is_load_store(insn))
2435 				err = emit_atomic_ld_st_index(&prog, insn->imm,
2436 							      BPF_SIZE(insn->code), dst_reg,
2437 							      src_reg, X86_REG_R12, insn->off);
2438 			else
2439 				err = emit_atomic_rmw_index(&prog, insn->imm, BPF_SIZE(insn->code),
2440 							    dst_reg, src_reg, X86_REG_R12,
2441 							    insn->off);
2442 			if (err)
2443 				return err;
2444 			goto populate_extable;
2445 
2446 			/* call */
2447 		case BPF_JMP | BPF_CALL: {
2448 			func = (u8 *) __bpf_call_base + imm32;
2449 			if (src_reg == BPF_PSEUDO_CALL && tail_call_reachable) {
2450 				LOAD_TAIL_CALL_CNT_PTR(stack_depth);
2451 				ip += 7;
2452 			}
2453 			if (!imm32)
2454 				return -EINVAL;
2455 			if (priv_frame_ptr) {
2456 				push_r9(&prog);
2457 				ip += 2;
2458 			}
2459 			ip += x86_call_depth_emit_accounting(&prog, func, ip);
2460 			if (emit_call(&prog, func, ip))
2461 				return -EINVAL;
2462 			if (priv_frame_ptr)
2463 				pop_r9(&prog);
2464 			break;
2465 		}
2466 
2467 		case BPF_JMP | BPF_TAIL_CALL:
2468 			if (imm32)
2469 				emit_bpf_tail_call_direct(bpf_prog,
2470 							  &bpf_prog->aux->poke_tab[imm32 - 1],
2471 							  &prog,
2472 							  ip,
2473 							  callee_regs_used,
2474 							  stack_depth,
2475 							  ctx);
2476 			else
2477 				emit_bpf_tail_call_indirect(bpf_prog,
2478 							    &prog,
2479 							    callee_regs_used,
2480 							    stack_depth,
2481 							    ip,
2482 							    ctx);
2483 			break;
2484 
2485 			/* cond jump */
2486 		case BPF_JMP | BPF_JEQ | BPF_X:
2487 		case BPF_JMP | BPF_JNE | BPF_X:
2488 		case BPF_JMP | BPF_JGT | BPF_X:
2489 		case BPF_JMP | BPF_JLT | BPF_X:
2490 		case BPF_JMP | BPF_JGE | BPF_X:
2491 		case BPF_JMP | BPF_JLE | BPF_X:
2492 		case BPF_JMP | BPF_JSGT | BPF_X:
2493 		case BPF_JMP | BPF_JSLT | BPF_X:
2494 		case BPF_JMP | BPF_JSGE | BPF_X:
2495 		case BPF_JMP | BPF_JSLE | BPF_X:
2496 		case BPF_JMP32 | BPF_JEQ | BPF_X:
2497 		case BPF_JMP32 | BPF_JNE | BPF_X:
2498 		case BPF_JMP32 | BPF_JGT | BPF_X:
2499 		case BPF_JMP32 | BPF_JLT | BPF_X:
2500 		case BPF_JMP32 | BPF_JGE | BPF_X:
2501 		case BPF_JMP32 | BPF_JLE | BPF_X:
2502 		case BPF_JMP32 | BPF_JSGT | BPF_X:
2503 		case BPF_JMP32 | BPF_JSLT | BPF_X:
2504 		case BPF_JMP32 | BPF_JSGE | BPF_X:
2505 		case BPF_JMP32 | BPF_JSLE | BPF_X:
2506 			/* cmp dst_reg, src_reg */
2507 			maybe_emit_mod(&prog, dst_reg, src_reg,
2508 				       BPF_CLASS(insn->code) == BPF_JMP);
2509 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
2510 			goto emit_cond_jmp;
2511 
2512 		case BPF_JMP | BPF_JSET | BPF_X:
2513 		case BPF_JMP32 | BPF_JSET | BPF_X:
2514 			/* test dst_reg, src_reg */
2515 			maybe_emit_mod(&prog, dst_reg, src_reg,
2516 				       BPF_CLASS(insn->code) == BPF_JMP);
2517 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
2518 			goto emit_cond_jmp;
2519 
2520 		case BPF_JMP | BPF_JSET | BPF_K:
2521 		case BPF_JMP32 | BPF_JSET | BPF_K:
2522 			/* test dst_reg, imm32 */
2523 			maybe_emit_1mod(&prog, dst_reg,
2524 					BPF_CLASS(insn->code) == BPF_JMP);
2525 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
2526 			goto emit_cond_jmp;
2527 
2528 		case BPF_JMP | BPF_JEQ | BPF_K:
2529 		case BPF_JMP | BPF_JNE | BPF_K:
2530 		case BPF_JMP | BPF_JGT | BPF_K:
2531 		case BPF_JMP | BPF_JLT | BPF_K:
2532 		case BPF_JMP | BPF_JGE | BPF_K:
2533 		case BPF_JMP | BPF_JLE | BPF_K:
2534 		case BPF_JMP | BPF_JSGT | BPF_K:
2535 		case BPF_JMP | BPF_JSLT | BPF_K:
2536 		case BPF_JMP | BPF_JSGE | BPF_K:
2537 		case BPF_JMP | BPF_JSLE | BPF_K:
2538 		case BPF_JMP32 | BPF_JEQ | BPF_K:
2539 		case BPF_JMP32 | BPF_JNE | BPF_K:
2540 		case BPF_JMP32 | BPF_JGT | BPF_K:
2541 		case BPF_JMP32 | BPF_JLT | BPF_K:
2542 		case BPF_JMP32 | BPF_JGE | BPF_K:
2543 		case BPF_JMP32 | BPF_JLE | BPF_K:
2544 		case BPF_JMP32 | BPF_JSGT | BPF_K:
2545 		case BPF_JMP32 | BPF_JSLT | BPF_K:
2546 		case BPF_JMP32 | BPF_JSGE | BPF_K:
2547 		case BPF_JMP32 | BPF_JSLE | BPF_K:
2548 			/* test dst_reg, dst_reg to save one extra byte */
2549 			if (imm32 == 0) {
2550 				maybe_emit_mod(&prog, dst_reg, dst_reg,
2551 					       BPF_CLASS(insn->code) == BPF_JMP);
2552 				EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
2553 				goto emit_cond_jmp;
2554 			}
2555 
2556 			/* cmp dst_reg, imm8/32 */
2557 			maybe_emit_1mod(&prog, dst_reg,
2558 					BPF_CLASS(insn->code) == BPF_JMP);
2559 
2560 			if (is_imm8(imm32))
2561 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
2562 			else
2563 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
2564 
2565 emit_cond_jmp:		/* Convert BPF opcode to x86 */
2566 			switch (BPF_OP(insn->code)) {
2567 			case BPF_JEQ:
2568 				jmp_cond = X86_JE;
2569 				break;
2570 			case BPF_JSET:
2571 			case BPF_JNE:
2572 				jmp_cond = X86_JNE;
2573 				break;
2574 			case BPF_JGT:
2575 				/* GT is unsigned '>', JA in x86 */
2576 				jmp_cond = X86_JA;
2577 				break;
2578 			case BPF_JLT:
2579 				/* LT is unsigned '<', JB in x86 */
2580 				jmp_cond = X86_JB;
2581 				break;
2582 			case BPF_JGE:
2583 				/* GE is unsigned '>=', JAE in x86 */
2584 				jmp_cond = X86_JAE;
2585 				break;
2586 			case BPF_JLE:
2587 				/* LE is unsigned '<=', JBE in x86 */
2588 				jmp_cond = X86_JBE;
2589 				break;
2590 			case BPF_JSGT:
2591 				/* Signed '>', GT in x86 */
2592 				jmp_cond = X86_JG;
2593 				break;
2594 			case BPF_JSLT:
2595 				/* Signed '<', LT in x86 */
2596 				jmp_cond = X86_JL;
2597 				break;
2598 			case BPF_JSGE:
2599 				/* Signed '>=', GE in x86 */
2600 				jmp_cond = X86_JGE;
2601 				break;
2602 			case BPF_JSLE:
2603 				/* Signed '<=', LE in x86 */
2604 				jmp_cond = X86_JLE;
2605 				break;
2606 			default: /* to silence GCC warning */
2607 				return -EFAULT;
2608 			}
2609 			jmp_offset = addrs[i + insn->off] - addrs[i];
2610 			if (is_imm8_jmp_offset(jmp_offset)) {
2611 				if (jmp_padding) {
2612 					/* To keep the jmp_offset valid, the extra bytes are
2613 					 * padded before the jump insn, so we subtract the
2614 					 * 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
2615 					 *
2616 					 * If the previous pass already emits an imm8
2617 					 * jmp_cond, then this BPF insn won't shrink, so
2618 					 * "nops" is 0.
2619 					 *
2620 					 * On the other hand, if the previous pass emits an
2621 					 * imm32 jmp_cond, the extra 4 bytes(*) is padded to
2622 					 * keep the image from shrinking further.
2623 					 *
2624 					 * (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
2625 					 *     is 2 bytes, so the size difference is 4 bytes.
2626 					 */
2627 					nops = INSN_SZ_DIFF - 2;
2628 					if (nops != 0 && nops != 4) {
2629 						pr_err("unexpected jmp_cond padding: %d bytes\n",
2630 						       nops);
2631 						return -EFAULT;
2632 					}
2633 					emit_nops(&prog, nops);
2634 				}
2635 				EMIT2(jmp_cond, jmp_offset);
2636 			} else if (is_simm32(jmp_offset)) {
2637 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
2638 			} else {
2639 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
2640 				return -EFAULT;
2641 			}
2642 
2643 			break;
2644 
2645 		case BPF_JMP | BPF_JA | BPF_X:
2646 			emit_indirect_jump(&prog, insn->dst_reg, ip);
2647 			break;
2648 		case BPF_JMP | BPF_JA:
2649 		case BPF_JMP32 | BPF_JA:
2650 			if (BPF_CLASS(insn->code) == BPF_JMP) {
2651 				if (insn->off == -1)
2652 					/* -1 jmp instructions will always jump
2653 					 * backwards two bytes. Explicitly handling
2654 					 * this case avoids wasting too many passes
2655 					 * when there are long sequences of replaced
2656 					 * dead code.
2657 					 */
2658 					jmp_offset = -2;
2659 				else
2660 					jmp_offset = addrs[i + insn->off] - addrs[i];
2661 			} else {
2662 				if (insn->imm == -1)
2663 					jmp_offset = -2;
2664 				else
2665 					jmp_offset = addrs[i + insn->imm] - addrs[i];
2666 			}
2667 
2668 			if (!jmp_offset) {
2669 				/*
2670 				 * If jmp_padding is enabled, the extra nops will
2671 				 * be inserted. Otherwise, optimize out nop jumps.
2672 				 */
2673 				if (jmp_padding) {
2674 					/* There are 3 possible conditions.
2675 					 * (1) This BPF_JA is already optimized out in
2676 					 *     the previous run, so there is no need
2677 					 *     to pad any extra byte (0 byte).
2678 					 * (2) The previous pass emits an imm8 jmp,
2679 					 *     so we pad 2 bytes to match the previous
2680 					 *     insn size.
2681 					 * (3) Similarly, the previous pass emits an
2682 					 *     imm32 jmp, and 5 bytes is padded.
2683 					 */
2684 					nops = INSN_SZ_DIFF;
2685 					if (nops != 0 && nops != 2 && nops != 5) {
2686 						pr_err("unexpected nop jump padding: %d bytes\n",
2687 						       nops);
2688 						return -EFAULT;
2689 					}
2690 					emit_nops(&prog, nops);
2691 				}
2692 				break;
2693 			}
2694 emit_jmp:
2695 			if (is_imm8_jmp_offset(jmp_offset)) {
2696 				if (jmp_padding) {
2697 					/* To avoid breaking jmp_offset, the extra bytes
2698 					 * are padded before the actual jmp insn, so
2699 					 * 2 bytes is subtracted from INSN_SZ_DIFF.
2700 					 *
2701 					 * If the previous pass already emits an imm8
2702 					 * jmp, there is nothing to pad (0 byte).
2703 					 *
2704 					 * If it emits an imm32 jmp (5 bytes) previously
2705 					 * and now an imm8 jmp (2 bytes), then we pad
2706 					 * (5 - 2 = 3) bytes to stop the image from
2707 					 * shrinking further.
2708 					 */
2709 					nops = INSN_SZ_DIFF - 2;
2710 					if (nops != 0 && nops != 3) {
2711 						pr_err("unexpected jump padding: %d bytes\n",
2712 						       nops);
2713 						return -EFAULT;
2714 					}
2715 					emit_nops(&prog, INSN_SZ_DIFF - 2);
2716 				}
2717 				EMIT2(0xEB, jmp_offset);
2718 			} else if (is_simm32(jmp_offset)) {
2719 				EMIT1_off32(0xE9, jmp_offset);
2720 			} else {
2721 				pr_err("jmp gen bug %llx\n", jmp_offset);
2722 				return -EFAULT;
2723 			}
2724 			break;
2725 
2726 		case BPF_JMP | BPF_EXIT:
2727 			if (seen_exit) {
2728 				jmp_offset = ctx->cleanup_addr - addrs[i];
2729 				goto emit_jmp;
2730 			}
2731 			seen_exit = true;
2732 			/* Update cleanup_addr */
2733 			ctx->cleanup_addr = proglen;
2734 			if (bpf_prog_was_classic(bpf_prog) &&
2735 			    !ns_capable_noaudit(&init_user_ns, CAP_SYS_ADMIN)) {
2736 				if (emit_spectre_bhb_barrier(&prog, ip, bpf_prog))
2737 					return -EINVAL;
2738 			}
2739 			if (bpf_prog->aux->exception_boundary) {
2740 				pop_callee_regs(&prog, all_callee_regs_used);
2741 				pop_r12(&prog);
2742 			} else {
2743 				pop_callee_regs(&prog, callee_regs_used);
2744 				if (arena_vm_start)
2745 					pop_r12(&prog);
2746 			}
2747 			EMIT1(0xC9);         /* leave */
2748 			bpf_prog->aux->ksym.fp_end = prog - temp;
2749 
2750 			emit_return(&prog, image + addrs[i - 1] + (prog - temp));
2751 			break;
2752 
2753 		default:
2754 			/*
2755 			 * By design x86-64 JIT should support all BPF instructions.
2756 			 * This error will be seen if new instruction was added
2757 			 * to the interpreter, but not to the JIT, or if there is
2758 			 * junk in bpf_prog.
2759 			 */
2760 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
2761 			return -EINVAL;
2762 		}
2763 
2764 		ilen = prog - temp;
2765 		if (ilen > BPF_MAX_INSN_SIZE) {
2766 			pr_err("bpf_jit: fatal insn size error\n");
2767 			return -EFAULT;
2768 		}
2769 
2770 		if (image) {
2771 			/*
2772 			 * When populating the image, assert that:
2773 			 *
2774 			 *  i) We do not write beyond the allocated space, and
2775 			 * ii) addrs[i] did not change from the prior run, in order
2776 			 *     to validate assumptions made for computing branch
2777 			 *     displacements.
2778 			 */
2779 			if (unlikely(proglen + ilen > oldproglen ||
2780 				     proglen + ilen != addrs[i])) {
2781 				pr_err("bpf_jit: fatal error\n");
2782 				return -EFAULT;
2783 			}
2784 			memcpy(rw_image + proglen, temp, ilen);
2785 		}
2786 		proglen += ilen;
2787 		addrs[i] = proglen;
2788 		prog = temp;
2789 	}
2790 
2791 	if (image && excnt != bpf_prog->aux->num_exentries) {
2792 		pr_err("extable is not populated\n");
2793 		return -EFAULT;
2794 	}
2795 	return proglen;
2796 }
2797 
2798 static void clean_stack_garbage(const struct btf_func_model *m,
2799 				u8 **pprog, int nr_stack_slots,
2800 				int stack_size)
2801 {
2802 	int arg_size, off;
2803 	u8 *prog;
2804 
2805 	/* Generally speaking, the compiler will pass the arguments
2806 	 * on-stack with "push" instruction, which will take 8-byte
2807 	 * on the stack. In this case, there won't be garbage values
2808 	 * while we copy the arguments from origin stack frame to current
2809 	 * in BPF_DW.
2810 	 *
2811 	 * However, sometimes the compiler will only allocate 4-byte on
2812 	 * the stack for the arguments. For now, this case will only
2813 	 * happen if there is only one argument on-stack and its size
2814 	 * not more than 4 byte. In this case, there will be garbage
2815 	 * values on the upper 4-byte where we store the argument on
2816 	 * current stack frame.
2817 	 *
2818 	 * arguments on origin stack:
2819 	 *
2820 	 * stack_arg_1(4-byte) xxx(4-byte)
2821 	 *
2822 	 * what we copy:
2823 	 *
2824 	 * stack_arg_1(8-byte): stack_arg_1(origin) xxx
2825 	 *
2826 	 * and the xxx is the garbage values which we should clean here.
2827 	 */
2828 	if (nr_stack_slots != 1)
2829 		return;
2830 
2831 	/* the size of the last argument */
2832 	arg_size = m->arg_size[m->nr_args - 1];
2833 	if (arg_size <= 4) {
2834 		off = -(stack_size - 4);
2835 		prog = *pprog;
2836 		/* mov DWORD PTR [rbp + off], 0 */
2837 		if (!is_imm8(off))
2838 			EMIT2_off32(0xC7, 0x85, off);
2839 		else
2840 			EMIT3(0xC7, 0x45, off);
2841 		EMIT(0, 4);
2842 		*pprog = prog;
2843 	}
2844 }
2845 
2846 /* get the count of the regs that are used to pass arguments */
2847 static int get_nr_used_regs(const struct btf_func_model *m)
2848 {
2849 	int i, arg_regs, nr_used_regs = 0;
2850 
2851 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2852 		arg_regs = (m->arg_size[i] + 7) / 8;
2853 		if (nr_used_regs + arg_regs <= 6)
2854 			nr_used_regs += arg_regs;
2855 
2856 		if (nr_used_regs >= 6)
2857 			break;
2858 	}
2859 
2860 	return nr_used_regs;
2861 }
2862 
2863 static void save_args(const struct btf_func_model *m, u8 **prog,
2864 		      int stack_size, bool for_call_origin, u32 flags)
2865 {
2866 	int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
2867 	bool use_jmp = bpf_trampoline_use_jmp(flags);
2868 	int i, j;
2869 
2870 	/* Store function arguments to stack.
2871 	 * For a function that accepts two pointers the sequence will be:
2872 	 * mov QWORD PTR [rbp-0x10],rdi
2873 	 * mov QWORD PTR [rbp-0x8],rsi
2874 	 */
2875 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2876 		arg_regs = (m->arg_size[i] + 7) / 8;
2877 
2878 		/* According to the research of Yonghong, struct members
2879 		 * should be all in register or all on the stack.
2880 		 * Meanwhile, the compiler will pass the argument on regs
2881 		 * if the remaining regs can hold the argument.
2882 		 *
2883 		 * Disorder of the args can happen. For example:
2884 		 *
2885 		 * struct foo_struct {
2886 		 *     long a;
2887 		 *     int b;
2888 		 * };
2889 		 * int foo(char, char, char, char, char, struct foo_struct,
2890 		 *         char);
2891 		 *
2892 		 * the arg1-5,arg7 will be passed by regs, and arg6 will
2893 		 * by stack.
2894 		 */
2895 		if (nr_regs + arg_regs > 6) {
2896 			/* copy function arguments from origin stack frame
2897 			 * into current stack frame.
2898 			 *
2899 			 * The starting address of the arguments on-stack
2900 			 * is:
2901 			 *   rbp + 8(push rbp) +
2902 			 *   8(return addr of origin call) +
2903 			 *   8(return addr of the caller)
2904 			 * which means: rbp + 24
2905 			 */
2906 			for (j = 0; j < arg_regs; j++) {
2907 				emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
2908 					 nr_stack_slots * 8 + 16 + (!use_jmp) * 8);
2909 				emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
2910 					 -stack_size);
2911 
2912 				if (!nr_stack_slots)
2913 					first_off = stack_size;
2914 				stack_size -= 8;
2915 				nr_stack_slots++;
2916 			}
2917 		} else {
2918 			/* Only copy the arguments on-stack to current
2919 			 * 'stack_size' and ignore the regs, used to
2920 			 * prepare the arguments on-stack for origin call.
2921 			 */
2922 			if (for_call_origin) {
2923 				nr_regs += arg_regs;
2924 				continue;
2925 			}
2926 
2927 			/* copy the arguments from regs into stack */
2928 			for (j = 0; j < arg_regs; j++) {
2929 				emit_stx(prog, BPF_DW, BPF_REG_FP,
2930 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2931 					 -stack_size);
2932 				stack_size -= 8;
2933 				nr_regs++;
2934 			}
2935 		}
2936 	}
2937 
2938 	clean_stack_garbage(m, prog, nr_stack_slots, first_off);
2939 }
2940 
2941 static void restore_regs(const struct btf_func_model *m, u8 **prog,
2942 			 int stack_size)
2943 {
2944 	int i, j, arg_regs, nr_regs = 0;
2945 
2946 	/* Restore function arguments from stack.
2947 	 * For a function that accepts two pointers the sequence will be:
2948 	 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
2949 	 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
2950 	 *
2951 	 * The logic here is similar to what we do in save_args()
2952 	 */
2953 	for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
2954 		arg_regs = (m->arg_size[i] + 7) / 8;
2955 		if (nr_regs + arg_regs <= 6) {
2956 			for (j = 0; j < arg_regs; j++) {
2957 				emit_ldx(prog, BPF_DW,
2958 					 nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
2959 					 BPF_REG_FP,
2960 					 -stack_size);
2961 				stack_size -= 8;
2962 				nr_regs++;
2963 			}
2964 		} else {
2965 			stack_size -= 8 * arg_regs;
2966 		}
2967 
2968 		if (nr_regs >= 6)
2969 			break;
2970 	}
2971 }
2972 
2973 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
2974 			   struct bpf_tramp_link *l, int stack_size,
2975 			   int run_ctx_off, bool save_ret,
2976 			   void *image, void *rw_image)
2977 {
2978 	u8 *prog = *pprog;
2979 	u8 *jmp_insn;
2980 	int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
2981 	struct bpf_prog *p = l->link.prog;
2982 	u64 cookie = l->cookie;
2983 
2984 	/* mov rdi, cookie */
2985 	emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
2986 
2987 	/* Prepare struct bpf_tramp_run_ctx.
2988 	 *
2989 	 * bpf_tramp_run_ctx is already preserved by
2990 	 * arch_prepare_bpf_trampoline().
2991 	 *
2992 	 * mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
2993 	 */
2994 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
2995 
2996 	/* arg1: mov rdi, progs[i] */
2997 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
2998 	/* arg2: lea rsi, [rbp - ctx_cookie_off] */
2999 	if (!is_imm8(-run_ctx_off))
3000 		EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
3001 	else
3002 		EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
3003 
3004 	if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image)))
3005 		return -EINVAL;
3006 	/* remember prog start time returned by __bpf_prog_enter */
3007 	emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
3008 
3009 	/* if (__bpf_prog_enter*(prog) == 0)
3010 	 *	goto skip_exec_of_prog;
3011 	 */
3012 	EMIT3(0x48, 0x85, 0xC0);  /* test rax,rax */
3013 	/* emit 2 nops that will be replaced with JE insn */
3014 	jmp_insn = prog;
3015 	emit_nops(&prog, 2);
3016 
3017 	/* arg1: lea rdi, [rbp - stack_size] */
3018 	if (!is_imm8(-stack_size))
3019 		EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
3020 	else
3021 		EMIT4(0x48, 0x8D, 0x7D, -stack_size);
3022 	/* arg2: progs[i]->insnsi for interpreter */
3023 	if (!p->jited)
3024 		emit_mov_imm64(&prog, BPF_REG_2,
3025 			       (long) p->insnsi >> 32,
3026 			       (u32) (long) p->insnsi);
3027 	/* call JITed bpf program or interpreter */
3028 	if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image)))
3029 		return -EINVAL;
3030 
3031 	/*
3032 	 * BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
3033 	 * of the previous call which is then passed on the stack to
3034 	 * the next BPF program.
3035 	 *
3036 	 * BPF_TRAMP_FENTRY trampoline may need to return the return
3037 	 * value of BPF_PROG_TYPE_STRUCT_OPS prog.
3038 	 */
3039 	if (save_ret)
3040 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
3041 
3042 	/* replace 2 nops with JE insn, since jmp target is known */
3043 	jmp_insn[0] = X86_JE;
3044 	jmp_insn[1] = prog - jmp_insn - 2;
3045 
3046 	/* arg1: mov rdi, progs[i] */
3047 	emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
3048 	/* arg2: mov rsi, rbx <- start time in nsec */
3049 	emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
3050 	/* arg3: lea rdx, [rbp - run_ctx_off] */
3051 	if (!is_imm8(-run_ctx_off))
3052 		EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
3053 	else
3054 		EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
3055 	if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
3056 		return -EINVAL;
3057 
3058 	*pprog = prog;
3059 	return 0;
3060 }
3061 
3062 static void emit_align(u8 **pprog, u32 align)
3063 {
3064 	u8 *target, *prog = *pprog;
3065 
3066 	target = PTR_ALIGN(prog, align);
3067 	if (target != prog)
3068 		emit_nops(&prog, target - prog);
3069 
3070 	*pprog = prog;
3071 }
3072 
3073 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
3074 {
3075 	u8 *prog = *pprog;
3076 	s64 offset;
3077 
3078 	offset = func - (ip + 2 + 4);
3079 	if (!is_simm32(offset)) {
3080 		pr_err("Target %p is out of range\n", func);
3081 		return -EINVAL;
3082 	}
3083 	EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
3084 	*pprog = prog;
3085 	return 0;
3086 }
3087 
3088 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
3089 		      struct bpf_tramp_links *tl, int stack_size,
3090 		      int run_ctx_off, int func_meta_off, bool save_ret,
3091 		      void *image, void *rw_image, u64 func_meta,
3092 		      int cookie_off)
3093 {
3094 	int i, cur_cookie = (cookie_off - stack_size) / 8;
3095 	u8 *prog = *pprog;
3096 
3097 	for (i = 0; i < tl->nr_links; i++) {
3098 		if (tl->links[i]->link.prog->call_session_cookie) {
3099 			emit_store_stack_imm64(&prog, BPF_REG_0, -func_meta_off,
3100 				func_meta | (cur_cookie << BPF_TRAMP_COOKIE_INDEX_SHIFT));
3101 			cur_cookie--;
3102 		}
3103 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
3104 				    run_ctx_off, save_ret, image, rw_image))
3105 			return -EINVAL;
3106 	}
3107 	*pprog = prog;
3108 	return 0;
3109 }
3110 
3111 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
3112 			      struct bpf_tramp_links *tl, int stack_size,
3113 			      int run_ctx_off, u8 **branches,
3114 			      void *image, void *rw_image)
3115 {
3116 	u8 *prog = *pprog;
3117 	int i;
3118 
3119 	/* The first fmod_ret program will receive a garbage return value.
3120 	 * Set this to 0 to avoid confusing the program.
3121 	 */
3122 	emit_mov_imm32(&prog, false, BPF_REG_0, 0);
3123 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
3124 	for (i = 0; i < tl->nr_links; i++) {
3125 		if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true,
3126 				    image, rw_image))
3127 			return -EINVAL;
3128 
3129 		/* mod_ret prog stored return value into [rbp - 8]. Emit:
3130 		 * if (*(u64 *)(rbp - 8) !=  0)
3131 		 *	goto do_fexit;
3132 		 */
3133 		/* cmp QWORD PTR [rbp - 0x8], 0x0 */
3134 		EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
3135 
3136 		/* Save the location of the branch and Generate 6 nops
3137 		 * (4 bytes for an offset and 2 bytes for the jump) These nops
3138 		 * are replaced with a conditional jump once do_fexit (i.e. the
3139 		 * start of the fexit invocation) is finalized.
3140 		 */
3141 		branches[i] = prog;
3142 		emit_nops(&prog, 4 + 2);
3143 	}
3144 
3145 	*pprog = prog;
3146 	return 0;
3147 }
3148 
3149 /* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
3150 #define LOAD_TRAMP_TAIL_CALL_CNT_PTR(stack)	\
3151 	__LOAD_TCC_PTR(-round_up(stack, 8) - 8)
3152 
3153 /* Example:
3154  * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
3155  * its 'struct btf_func_model' will be nr_args=2
3156  * The assembly code when eth_type_trans is executing after trampoline:
3157  *
3158  * push rbp
3159  * mov rbp, rsp
3160  * sub rsp, 16                     // space for skb and dev
3161  * push rbx                        // temp regs to pass start time
3162  * mov qword ptr [rbp - 16], rdi   // save skb pointer to stack
3163  * mov qword ptr [rbp - 8], rsi    // save dev pointer to stack
3164  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
3165  * mov rbx, rax                    // remember start time in bpf stats are enabled
3166  * lea rdi, [rbp - 16]             // R1==ctx of bpf prog
3167  * call addr_of_jited_FENTRY_prog
3168  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
3169  * mov rsi, rbx                    // prog start time
3170  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
3171  * mov rdi, qword ptr [rbp - 16]   // restore skb pointer from stack
3172  * mov rsi, qword ptr [rbp - 8]    // restore dev pointer from stack
3173  * pop rbx
3174  * leave
3175  * ret
3176  *
3177  * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
3178  * replaced with 'call generated_bpf_trampoline'. When it returns
3179  * eth_type_trans will continue executing with original skb and dev pointers.
3180  *
3181  * The assembly code when eth_type_trans is called from trampoline:
3182  *
3183  * push rbp
3184  * mov rbp, rsp
3185  * sub rsp, 24                     // space for skb, dev, return value
3186  * push rbx                        // temp regs to pass start time
3187  * mov qword ptr [rbp - 24], rdi   // save skb pointer to stack
3188  * mov qword ptr [rbp - 16], rsi   // save dev pointer to stack
3189  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
3190  * mov rbx, rax                    // remember start time if bpf stats are enabled
3191  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
3192  * call addr_of_jited_FENTRY_prog  // bpf prog can access skb and dev
3193  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
3194  * mov rsi, rbx                    // prog start time
3195  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
3196  * mov rdi, qword ptr [rbp - 24]   // restore skb pointer from stack
3197  * mov rsi, qword ptr [rbp - 16]   // restore dev pointer from stack
3198  * call eth_type_trans+5           // execute body of eth_type_trans
3199  * mov qword ptr [rbp - 8], rax    // save return value
3200  * call __bpf_prog_enter           // rcu_read_lock and preempt_disable
3201  * mov rbx, rax                    // remember start time in bpf stats are enabled
3202  * lea rdi, [rbp - 24]             // R1==ctx of bpf prog
3203  * call addr_of_jited_FEXIT_prog   // bpf prog can access skb, dev, return value
3204  * movabsq rdi, 64bit_addr_of_struct_bpf_prog  // unused if bpf stats are off
3205  * mov rsi, rbx                    // prog start time
3206  * call __bpf_prog_exit            // rcu_read_unlock, preempt_enable and stats math
3207  * mov rax, qword ptr [rbp - 8]    // restore eth_type_trans's return value
3208  * pop rbx
3209  * leave
3210  * add rsp, 8                      // skip eth_type_trans's frame
3211  * ret                             // return to its caller
3212  */
3213 static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
3214 					 void *rw_image_end, void *image,
3215 					 const struct btf_func_model *m, u32 flags,
3216 					 struct bpf_tramp_links *tlinks,
3217 					 void *func_addr)
3218 {
3219 	int i, ret, nr_regs = m->nr_args, stack_size = 0;
3220 	int regs_off, func_meta_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
3221 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
3222 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
3223 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
3224 	void *orig_call = func_addr;
3225 	int cookie_off, cookie_cnt;
3226 	u8 **branches = NULL;
3227 	u64 func_meta;
3228 	u8 *prog;
3229 	bool save_ret;
3230 
3231 	/*
3232 	 * F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is
3233 	 * explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG
3234 	 * because @func_addr.
3235 	 */
3236 	WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) &&
3237 		     (flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET)));
3238 
3239 	/* extra registers for struct arguments */
3240 	for (i = 0; i < m->nr_args; i++) {
3241 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
3242 			nr_regs += (m->arg_size[i] + 7) / 8 - 1;
3243 	}
3244 
3245 	/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
3246 	 * are passed through regs, the remains are through stack.
3247 	 */
3248 	if (nr_regs > MAX_BPF_FUNC_ARGS)
3249 		return -ENOTSUPP;
3250 
3251 	/* Generated trampoline stack layout:
3252 	 *
3253 	 * RBP + 8         [ return address  ]
3254 	 * RBP + 0         [ RBP             ]
3255 	 *
3256 	 * RBP - 8         [ return value    ]  BPF_TRAMP_F_CALL_ORIG or
3257 	 *                                      BPF_TRAMP_F_RET_FENTRY_RET flags
3258 	 *
3259 	 *                 [ reg_argN        ]  always
3260 	 *                 [ ...             ]
3261 	 * RBP - regs_off  [ reg_arg1        ]  program's ctx pointer
3262 	 *
3263 	 * RBP - func_meta_off [ regs count, etc ]  always
3264 	 *
3265 	 * RBP - ip_off    [ traced function ]  BPF_TRAMP_F_IP_ARG flag
3266 	 *
3267 	 * RBP - rbx_off   [ rbx value       ]  always
3268 	 *
3269 	 * RBP - run_ctx_off [ bpf_tramp_run_ctx ]
3270 	 *
3271 	 *                     [ stack_argN ]  BPF_TRAMP_F_CALL_ORIG
3272 	 *                     [ ...        ]
3273 	 *                     [ stack_arg2 ]
3274 	 * RBP - arg_stack_off [ stack_arg1 ]
3275 	 * RSP                 [ tail_call_cnt_ptr ] BPF_TRAMP_F_TAIL_CALL_CTX
3276 	 */
3277 
3278 	/* room for return value of orig_call or fentry prog */
3279 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
3280 	if (save_ret)
3281 		stack_size += 8;
3282 
3283 	stack_size += nr_regs * 8;
3284 	regs_off = stack_size;
3285 
3286 	/* function matedata, such as regs count  */
3287 	stack_size += 8;
3288 	func_meta_off = stack_size;
3289 
3290 	if (flags & BPF_TRAMP_F_IP_ARG)
3291 		stack_size += 8; /* room for IP address argument */
3292 
3293 	ip_off = stack_size;
3294 
3295 	cookie_cnt = bpf_fsession_cookie_cnt(tlinks);
3296 	/* room for session cookies */
3297 	stack_size += cookie_cnt * 8;
3298 	cookie_off = stack_size;
3299 
3300 	stack_size += 8;
3301 	rbx_off = stack_size;
3302 
3303 	stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
3304 	run_ctx_off = stack_size;
3305 
3306 	if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
3307 		/* the space that used to pass arguments on-stack */
3308 		stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
3309 		/* make sure the stack pointer is 16-byte aligned if we
3310 		 * need pass arguments on stack, which means
3311 		 *  [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
3312 		 * should be 16-byte aligned. Following code depend on
3313 		 * that stack_size is already 8-byte aligned.
3314 		 */
3315 		if (bpf_trampoline_use_jmp(flags)) {
3316 			/* no rip in the "jmp" case */
3317 			stack_size += (stack_size % 16) ? 8 : 0;
3318 		} else {
3319 			stack_size += (stack_size % 16) ? 0 : 8;
3320 		}
3321 	}
3322 
3323 	arg_stack_off = stack_size;
3324 
3325 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3326 		/* skip patched call instruction and point orig_call to actual
3327 		 * body of the kernel function.
3328 		 */
3329 		if (is_endbr(orig_call))
3330 			orig_call += ENDBR_INSN_SIZE;
3331 		orig_call += X86_PATCH_SIZE;
3332 	}
3333 
3334 	prog = rw_image;
3335 
3336 	if (flags & BPF_TRAMP_F_INDIRECT) {
3337 		/*
3338 		 * Indirect call for bpf_struct_ops
3339 		 */
3340 		emit_cfi(&prog, image,
3341 			 cfi_get_func_hash(func_addr),
3342 			 cfi_get_func_arity(func_addr));
3343 	} else {
3344 		/*
3345 		 * Direct-call fentry stub, as such it needs accounting for the
3346 		 * __fentry__ call.
3347 		 */
3348 		x86_call_depth_emit_accounting(&prog, NULL, image);
3349 	}
3350 	EMIT1(0x55);		 /* push rbp */
3351 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
3352 	if (im)
3353 		im->ksym.fp_start = prog - (u8 *)rw_image;
3354 
3355 	if (!is_imm8(stack_size)) {
3356 		/* sub rsp, stack_size */
3357 		EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
3358 	} else {
3359 		/* sub rsp, stack_size */
3360 		EMIT4(0x48, 0x83, 0xEC, stack_size);
3361 	}
3362 	if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
3363 		EMIT1(0x50);		/* push rax */
3364 	/* mov QWORD PTR [rbp - rbx_off], rbx */
3365 	emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
3366 
3367 	func_meta = nr_regs;
3368 	/* Store number of argument registers of the traced function */
3369 	emit_store_stack_imm64(&prog, BPF_REG_0, -func_meta_off, func_meta);
3370 
3371 	if (flags & BPF_TRAMP_F_IP_ARG) {
3372 		/* Store IP address of the traced function */
3373 		emit_store_stack_imm64(&prog, BPF_REG_0, -ip_off, (long)func_addr);
3374 	}
3375 
3376 	save_args(m, &prog, regs_off, false, flags);
3377 
3378 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3379 		/* arg1: mov rdi, im */
3380 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
3381 		if (emit_rsb_call(&prog, __bpf_tramp_enter,
3382 				  image + (prog - (u8 *)rw_image))) {
3383 			ret = -EINVAL;
3384 			goto cleanup;
3385 		}
3386 	}
3387 
3388 	if (bpf_fsession_cnt(tlinks)) {
3389 		/* clear all the session cookies' value */
3390 		for (int i = 0; i < cookie_cnt; i++)
3391 			emit_store_stack_imm64(&prog, BPF_REG_0, -cookie_off + 8 * i, 0);
3392 		/* clear the return value to make sure fentry always get 0 */
3393 		emit_store_stack_imm64(&prog, BPF_REG_0, -8, 0);
3394 	}
3395 
3396 	if (fentry->nr_links) {
3397 		if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off, func_meta_off,
3398 			       flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image,
3399 			       func_meta, cookie_off))
3400 			return -EINVAL;
3401 	}
3402 
3403 	if (fmod_ret->nr_links) {
3404 		branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
3405 				   GFP_KERNEL);
3406 		if (!branches)
3407 			return -ENOMEM;
3408 
3409 		if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
3410 				       run_ctx_off, branches, image, rw_image)) {
3411 			ret = -EINVAL;
3412 			goto cleanup;
3413 		}
3414 	}
3415 
3416 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3417 		restore_regs(m, &prog, regs_off);
3418 		save_args(m, &prog, arg_stack_off, true, flags);
3419 
3420 		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
3421 			/* Before calling the original function, load the
3422 			 * tail_call_cnt_ptr from stack to rax.
3423 			 */
3424 			LOAD_TRAMP_TAIL_CALL_CNT_PTR(stack_size);
3425 		}
3426 
3427 		if (flags & BPF_TRAMP_F_ORIG_STACK) {
3428 			emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
3429 			EMIT2(0xff, 0xd3); /* call *rbx */
3430 		} else {
3431 			/* call original function */
3432 			if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) {
3433 				ret = -EINVAL;
3434 				goto cleanup;
3435 			}
3436 		}
3437 		/* remember return value in a stack for bpf prog to access */
3438 		emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
3439 		im->ip_after_call = image + (prog - (u8 *)rw_image);
3440 		emit_nops(&prog, X86_PATCH_SIZE);
3441 	}
3442 
3443 	if (fmod_ret->nr_links) {
3444 		/* From Intel 64 and IA-32 Architectures Optimization
3445 		 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
3446 		 * Coding Rule 11: All branch targets should be 16-byte
3447 		 * aligned.
3448 		 */
3449 		emit_align(&prog, 16);
3450 		/* Update the branches saved in invoke_bpf_mod_ret with the
3451 		 * aligned address of do_fexit.
3452 		 */
3453 		for (i = 0; i < fmod_ret->nr_links; i++) {
3454 			emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image),
3455 					    image + (branches[i] - (u8 *)rw_image), X86_JNE);
3456 		}
3457 	}
3458 
3459 	/* set the "is_return" flag for fsession */
3460 	func_meta |= (1ULL << BPF_TRAMP_IS_RETURN_SHIFT);
3461 	if (bpf_fsession_cnt(tlinks))
3462 		emit_store_stack_imm64(&prog, BPF_REG_0, -func_meta_off, func_meta);
3463 
3464 	if (fexit->nr_links) {
3465 		if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off, func_meta_off,
3466 			       false, image, rw_image, func_meta, cookie_off)) {
3467 			ret = -EINVAL;
3468 			goto cleanup;
3469 		}
3470 	}
3471 
3472 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
3473 		restore_regs(m, &prog, regs_off);
3474 
3475 	/* This needs to be done regardless. If there were fmod_ret programs,
3476 	 * the return value is only updated on the stack and still needs to be
3477 	 * restored to R0.
3478 	 */
3479 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
3480 		im->ip_epilogue = image + (prog - (u8 *)rw_image);
3481 		/* arg1: mov rdi, im */
3482 		emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
3483 		if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) {
3484 			ret = -EINVAL;
3485 			goto cleanup;
3486 		}
3487 	} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
3488 		/* Before running the original function, load the
3489 		 * tail_call_cnt_ptr from stack to rax.
3490 		 */
3491 		LOAD_TRAMP_TAIL_CALL_CNT_PTR(stack_size);
3492 	}
3493 
3494 	/* restore return value of orig_call or fentry prog back into RAX */
3495 	if (save_ret)
3496 		emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
3497 
3498 	emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
3499 
3500 	EMIT1(0xC9); /* leave */
3501 	if (im)
3502 		im->ksym.fp_end = prog - (u8 *)rw_image;
3503 
3504 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
3505 		/* skip our return address and return to parent */
3506 		EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
3507 	}
3508 	emit_return(&prog, image + (prog - (u8 *)rw_image));
3509 	/* Make sure the trampoline generation logic doesn't overflow */
3510 	if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) {
3511 		ret = -EFAULT;
3512 		goto cleanup;
3513 	}
3514 	ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY;
3515 
3516 cleanup:
3517 	kfree(branches);
3518 	return ret;
3519 }
3520 
3521 void *arch_alloc_bpf_trampoline(unsigned int size)
3522 {
3523 	return bpf_prog_pack_alloc(size, jit_fill_hole);
3524 }
3525 
3526 void arch_free_bpf_trampoline(void *image, unsigned int size)
3527 {
3528 	bpf_prog_pack_free(image, size);
3529 }
3530 
3531 int arch_protect_bpf_trampoline(void *image, unsigned int size)
3532 {
3533 	return 0;
3534 }
3535 
3536 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
3537 				const struct btf_func_model *m, u32 flags,
3538 				struct bpf_tramp_links *tlinks,
3539 				void *func_addr)
3540 {
3541 	void *rw_image, *tmp;
3542 	int ret;
3543 	u32 size = image_end - image;
3544 
3545 	/* rw_image doesn't need to be in module memory range, so we can
3546 	 * use kvmalloc.
3547 	 */
3548 	rw_image = kvmalloc(size, GFP_KERNEL);
3549 	if (!rw_image)
3550 		return -ENOMEM;
3551 
3552 	ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
3553 					    flags, tlinks, func_addr);
3554 	if (ret < 0)
3555 		goto out;
3556 
3557 	tmp = bpf_arch_text_copy(image, rw_image, size);
3558 	if (IS_ERR(tmp))
3559 		ret = PTR_ERR(tmp);
3560 out:
3561 	kvfree(rw_image);
3562 	return ret;
3563 }
3564 
3565 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
3566 			     struct bpf_tramp_links *tlinks, void *func_addr)
3567 {
3568 	struct bpf_tramp_image im;
3569 	void *image;
3570 	int ret;
3571 
3572 	/* Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
3573 	 * This will NOT cause fragmentation in direct map, as we do not
3574 	 * call set_memory_*() on this buffer.
3575 	 *
3576 	 * We cannot use kvmalloc here, because we need image to be in
3577 	 * module memory range.
3578 	 */
3579 	image = bpf_jit_alloc_exec(PAGE_SIZE);
3580 	if (!image)
3581 		return -ENOMEM;
3582 
3583 	ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
3584 					    m, flags, tlinks, func_addr);
3585 	bpf_jit_free_exec(image);
3586 	return ret;
3587 }
3588 
3589 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
3590 {
3591 	u8 *jg_reloc, *prog = *pprog;
3592 	int pivot, err, jg_bytes = 1;
3593 	s64 jg_offset;
3594 
3595 	if (a == b) {
3596 		/* Leaf node of recursion, i.e. not a range of indices
3597 		 * anymore.
3598 		 */
3599 		EMIT1(add_1mod(0x48, BPF_REG_3));	/* cmp rdx,func */
3600 		if (!is_simm32(progs[a]))
3601 			return -1;
3602 		EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
3603 			    progs[a]);
3604 		err = emit_cond_near_jump(&prog,	/* je func */
3605 					  (void *)progs[a], image + (prog - buf),
3606 					  X86_JE);
3607 		if (err)
3608 			return err;
3609 
3610 		emit_indirect_jump(&prog, BPF_REG_3 /* R3 -> rdx */, image + (prog - buf));
3611 
3612 		*pprog = prog;
3613 		return 0;
3614 	}
3615 
3616 	/* Not a leaf node, so we pivot, and recursively descend into
3617 	 * the lower and upper ranges.
3618 	 */
3619 	pivot = (b - a) / 2;
3620 	EMIT1(add_1mod(0x48, BPF_REG_3));		/* cmp rdx,func */
3621 	if (!is_simm32(progs[a + pivot]))
3622 		return -1;
3623 	EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
3624 
3625 	if (pivot > 2) {				/* jg upper_part */
3626 		/* Require near jump. */
3627 		jg_bytes = 4;
3628 		EMIT2_off32(0x0F, X86_JG + 0x10, 0);
3629 	} else {
3630 		EMIT2(X86_JG, 0);
3631 	}
3632 	jg_reloc = prog;
3633 
3634 	err = emit_bpf_dispatcher(&prog, a, a + pivot,	/* emit lower_part */
3635 				  progs, image, buf);
3636 	if (err)
3637 		return err;
3638 
3639 	/* From Intel 64 and IA-32 Architectures Optimization
3640 	 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
3641 	 * Coding Rule 11: All branch targets should be 16-byte
3642 	 * aligned.
3643 	 */
3644 	emit_align(&prog, 16);
3645 	jg_offset = prog - jg_reloc;
3646 	emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
3647 
3648 	err = emit_bpf_dispatcher(&prog, a + pivot + 1,	/* emit upper_part */
3649 				  b, progs, image, buf);
3650 	if (err)
3651 		return err;
3652 
3653 	*pprog = prog;
3654 	return 0;
3655 }
3656 
3657 static int cmp_ips(const void *a, const void *b)
3658 {
3659 	const s64 *ipa = a;
3660 	const s64 *ipb = b;
3661 
3662 	if (*ipa > *ipb)
3663 		return 1;
3664 	if (*ipa < *ipb)
3665 		return -1;
3666 	return 0;
3667 }
3668 
3669 int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
3670 {
3671 	u8 *prog = buf;
3672 
3673 	sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
3674 	return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
3675 }
3676 
3677 static void priv_stack_init_guard(void __percpu *priv_stack_ptr, int alloc_size)
3678 {
3679 	int cpu, underflow_idx = (alloc_size - PRIV_STACK_GUARD_SZ) >> 3;
3680 	u64 *stack_ptr;
3681 
3682 	for_each_possible_cpu(cpu) {
3683 		stack_ptr = per_cpu_ptr(priv_stack_ptr, cpu);
3684 		stack_ptr[0] = PRIV_STACK_GUARD_VAL;
3685 		stack_ptr[underflow_idx] = PRIV_STACK_GUARD_VAL;
3686 	}
3687 }
3688 
3689 static void priv_stack_check_guard(void __percpu *priv_stack_ptr, int alloc_size,
3690 				   struct bpf_prog *prog)
3691 {
3692 	int cpu, underflow_idx = (alloc_size - PRIV_STACK_GUARD_SZ) >> 3;
3693 	u64 *stack_ptr;
3694 
3695 	for_each_possible_cpu(cpu) {
3696 		stack_ptr = per_cpu_ptr(priv_stack_ptr, cpu);
3697 		if (stack_ptr[0] != PRIV_STACK_GUARD_VAL ||
3698 		    stack_ptr[underflow_idx] != PRIV_STACK_GUARD_VAL) {
3699 			pr_err("BPF private stack overflow/underflow detected for prog %sx\n",
3700 			       bpf_jit_get_prog_name(prog));
3701 			break;
3702 		}
3703 	}
3704 }
3705 
3706 struct x64_jit_data {
3707 	struct bpf_binary_header *rw_header;
3708 	struct bpf_binary_header *header;
3709 	int *addrs;
3710 	u8 *image;
3711 	int proglen;
3712 	struct jit_context ctx;
3713 };
3714 
3715 #define MAX_PASSES 20
3716 #define PADDING_PASSES (MAX_PASSES - 5)
3717 
3718 struct bpf_prog *bpf_int_jit_compile(struct bpf_verifier_env *env, struct bpf_prog *prog)
3719 {
3720 	struct bpf_binary_header *rw_header = NULL;
3721 	struct bpf_binary_header *header = NULL;
3722 	void __percpu *priv_stack_ptr = NULL;
3723 	struct x64_jit_data *jit_data;
3724 	int priv_stack_alloc_sz;
3725 	int proglen, oldproglen = 0;
3726 	struct jit_context ctx = {};
3727 	bool extra_pass = false;
3728 	bool padding = false;
3729 	u8 *rw_image = NULL;
3730 	u8 *image = NULL;
3731 	int *addrs;
3732 	int pass;
3733 	int i;
3734 
3735 	if (!prog->jit_requested)
3736 		return prog;
3737 
3738 	jit_data = prog->aux->jit_data;
3739 	if (!jit_data) {
3740 		jit_data = kzalloc_obj(*jit_data);
3741 		if (!jit_data)
3742 			return prog;
3743 		prog->aux->jit_data = jit_data;
3744 	}
3745 	priv_stack_ptr = prog->aux->priv_stack_ptr;
3746 	if (!priv_stack_ptr && prog->aux->jits_use_priv_stack) {
3747 		/* Allocate actual private stack size with verifier-calculated
3748 		 * stack size plus two memory guards to protect overflow and
3749 		 * underflow.
3750 		 */
3751 		priv_stack_alloc_sz = round_up(prog->aux->stack_depth, 8) +
3752 				      2 * PRIV_STACK_GUARD_SZ;
3753 		priv_stack_ptr = __alloc_percpu_gfp(priv_stack_alloc_sz, 8, GFP_KERNEL);
3754 		if (!priv_stack_ptr)
3755 			goto out_priv_stack;
3756 
3757 		priv_stack_init_guard(priv_stack_ptr, priv_stack_alloc_sz);
3758 		prog->aux->priv_stack_ptr = priv_stack_ptr;
3759 	}
3760 	addrs = jit_data->addrs;
3761 	if (addrs) {
3762 		ctx = jit_data->ctx;
3763 		oldproglen = jit_data->proglen;
3764 		image = jit_data->image;
3765 		header = jit_data->header;
3766 		rw_header = jit_data->rw_header;
3767 		rw_image = (void *)rw_header + ((void *)image - (void *)header);
3768 		extra_pass = true;
3769 		padding = true;
3770 		goto skip_init_addrs;
3771 	}
3772 	addrs = kvmalloc_objs(*addrs, prog->len + 1);
3773 	if (!addrs)
3774 		goto out_addrs;
3775 
3776 	/*
3777 	 * Before first pass, make a rough estimation of addrs[]
3778 	 * each BPF instruction is translated to less than 64 bytes
3779 	 */
3780 	for (proglen = 0, i = 0; i <= prog->len; i++) {
3781 		proglen += 64;
3782 		addrs[i] = proglen;
3783 	}
3784 	ctx.cleanup_addr = proglen;
3785 skip_init_addrs:
3786 
3787 	/*
3788 	 * JITed image shrinks with every pass and the loop iterates
3789 	 * until the image stops shrinking. Very large BPF programs
3790 	 * may converge on the last pass. In such case do one more
3791 	 * pass to emit the final image.
3792 	 */
3793 	for (pass = 0; pass < MAX_PASSES || image; pass++) {
3794 		if (!padding && pass >= PADDING_PASSES)
3795 			padding = true;
3796 		proglen = do_jit(env, prog, addrs, image, rw_image, oldproglen, &ctx, padding);
3797 		if (proglen <= 0) {
3798 out_image:
3799 			image = NULL;
3800 			if (header) {
3801 				bpf_arch_text_copy(&header->size, &rw_header->size,
3802 						   sizeof(rw_header->size));
3803 				bpf_jit_binary_pack_free(header, rw_header);
3804 			}
3805 			if (extra_pass) {
3806 				prog->bpf_func = NULL;
3807 				prog->jited = 0;
3808 				prog->jited_len = 0;
3809 			}
3810 			goto out_addrs;
3811 		}
3812 		if (image) {
3813 			if (proglen != oldproglen) {
3814 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
3815 				       proglen, oldproglen);
3816 				goto out_image;
3817 			}
3818 			break;
3819 		}
3820 		if (proglen == oldproglen) {
3821 			/*
3822 			 * The number of entries in extable is the number of BPF_LDX
3823 			 * insns that access kernel memory via "pointer to BTF type".
3824 			 * The verifier changed their opcode from LDX|MEM|size
3825 			 * to LDX|PROBE_MEM|size to make JITing easier.
3826 			 */
3827 			u32 align = __alignof__(struct exception_table_entry);
3828 			u32 extable_size = prog->aux->num_exentries *
3829 				sizeof(struct exception_table_entry);
3830 
3831 			/* allocate module memory for x86 insns and extable */
3832 			header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
3833 							   &image, align, &rw_header, &rw_image,
3834 							   jit_fill_hole);
3835 			if (!header)
3836 				goto out_addrs;
3837 			prog->aux->extable = (void *) image + roundup(proglen, align);
3838 		}
3839 		oldproglen = proglen;
3840 		cond_resched();
3841 	}
3842 
3843 	if (bpf_jit_enable > 1)
3844 		bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
3845 
3846 	if (image) {
3847 		if (!prog->is_func || extra_pass) {
3848 			/*
3849 			 * bpf_jit_binary_pack_finalize fails in two scenarios:
3850 			 *   1) header is not pointing to proper module memory;
3851 			 *   2) the arch doesn't support bpf_arch_text_copy().
3852 			 *
3853 			 * Both cases are serious bugs and justify WARN_ON.
3854 			 */
3855 			if (WARN_ON(bpf_jit_binary_pack_finalize(header, rw_header))) {
3856 				/* header has been freed */
3857 				header = NULL;
3858 				goto out_image;
3859 			}
3860 
3861 			bpf_tail_call_direct_fixup(prog);
3862 		} else {
3863 			jit_data->addrs = addrs;
3864 			jit_data->ctx = ctx;
3865 			jit_data->proglen = proglen;
3866 			jit_data->image = image;
3867 			jit_data->header = header;
3868 			jit_data->rw_header = rw_header;
3869 		}
3870 
3871 		/*
3872 		 * The bpf_prog_update_insn_ptrs function expects addrs to
3873 		 * point to the first byte of the jitted instruction (unlike
3874 		 * the bpf_prog_fill_jited_linfo below, which, for historical
3875 		 * reasons, expects to point to the next instruction)
3876 		 */
3877 		bpf_prog_update_insn_ptrs(prog, addrs, image);
3878 
3879 		/*
3880 		 * ctx.prog_offset is used when CFI preambles put code *before*
3881 		 * the function. See emit_cfi(). For FineIBT specifically this code
3882 		 * can also be executed and bpf_prog_kallsyms_add() will
3883 		 * generate an additional symbol to cover this, hence also
3884 		 * decrement proglen.
3885 		 */
3886 		prog->bpf_func = (void *)image + cfi_get_offset();
3887 		prog->jited = 1;
3888 		prog->jited_len = proglen - cfi_get_offset();
3889 	}
3890 
3891 	if (!image || !prog->is_func || extra_pass) {
3892 		if (image)
3893 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
3894 out_addrs:
3895 		kvfree(addrs);
3896 		if (!image && priv_stack_ptr) {
3897 			free_percpu(priv_stack_ptr);
3898 			prog->aux->priv_stack_ptr = NULL;
3899 		}
3900 out_priv_stack:
3901 		kfree(jit_data);
3902 		prog->aux->jit_data = NULL;
3903 	}
3904 
3905 	return prog;
3906 }
3907 
3908 bool bpf_jit_supports_kfunc_call(void)
3909 {
3910 	return true;
3911 }
3912 
3913 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
3914 {
3915 	if (text_poke_copy(dst, src, len) == NULL)
3916 		return ERR_PTR(-EINVAL);
3917 	return dst;
3918 }
3919 
3920 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
3921 bool bpf_jit_supports_subprog_tailcalls(void)
3922 {
3923 	return true;
3924 }
3925 
3926 bool bpf_jit_supports_percpu_insn(void)
3927 {
3928 	return true;
3929 }
3930 
3931 void bpf_jit_free(struct bpf_prog *prog)
3932 {
3933 	if (prog->jited) {
3934 		struct x64_jit_data *jit_data = prog->aux->jit_data;
3935 		struct bpf_binary_header *hdr;
3936 		void __percpu *priv_stack_ptr;
3937 		int priv_stack_alloc_sz;
3938 
3939 		/*
3940 		 * If we fail the final pass of JIT (from jit_subprogs),
3941 		 * the program may not be finalized yet. Call finalize here
3942 		 * before freeing it.
3943 		 */
3944 		if (jit_data) {
3945 			bpf_jit_binary_pack_finalize(jit_data->header,
3946 						     jit_data->rw_header);
3947 			kvfree(jit_data->addrs);
3948 			kfree(jit_data);
3949 		}
3950 		prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset();
3951 		hdr = bpf_jit_binary_pack_hdr(prog);
3952 		bpf_jit_binary_pack_free(hdr, NULL);
3953 		priv_stack_ptr = prog->aux->priv_stack_ptr;
3954 		if (priv_stack_ptr) {
3955 			priv_stack_alloc_sz = round_up(prog->aux->stack_depth, 8) +
3956 					      2 * PRIV_STACK_GUARD_SZ;
3957 			priv_stack_check_guard(priv_stack_ptr, priv_stack_alloc_sz, prog);
3958 			free_percpu(prog->aux->priv_stack_ptr);
3959 		}
3960 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
3961 	}
3962 
3963 	bpf_prog_unlock_free(prog);
3964 }
3965 
3966 bool bpf_jit_supports_exceptions(void)
3967 {
3968 	/* We unwind through both kernel frames (starting from within bpf_throw
3969 	 * call) and BPF frames. Therefore we require ORC unwinder to be enabled
3970 	 * to walk kernel frames and reach BPF frames in the stack trace.
3971 	 */
3972 	return IS_ENABLED(CONFIG_UNWINDER_ORC);
3973 }
3974 
3975 bool bpf_jit_supports_private_stack(void)
3976 {
3977 	return true;
3978 }
3979 
3980 void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
3981 {
3982 #if defined(CONFIG_UNWINDER_ORC)
3983 	struct unwind_state state;
3984 	unsigned long addr;
3985 
3986 	for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
3987 	     unwind_next_frame(&state)) {
3988 		addr = unwind_get_return_address(&state);
3989 		if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
3990 			break;
3991 	}
3992 	return;
3993 #endif
3994 }
3995 
3996 void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
3997 			       struct bpf_prog *new, struct bpf_prog *old)
3998 {
3999 	u8 *old_addr, *new_addr, *old_bypass_addr;
4000 	enum bpf_text_poke_type t;
4001 	int ret;
4002 
4003 	old_bypass_addr = old ? NULL : poke->bypass_addr;
4004 	old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
4005 	new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
4006 
4007 	/*
4008 	 * On program loading or teardown, the program's kallsym entry
4009 	 * might not be in place, so we use __bpf_arch_text_poke to skip
4010 	 * the kallsyms check.
4011 	 */
4012 	if (new) {
4013 		t = old_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
4014 		ret = __bpf_arch_text_poke(poke->tailcall_target,
4015 					   t, BPF_MOD_JUMP,
4016 					   old_addr, new_addr);
4017 		BUG_ON(ret < 0);
4018 		if (!old) {
4019 			ret = __bpf_arch_text_poke(poke->tailcall_bypass,
4020 						   BPF_MOD_JUMP, BPF_MOD_NOP,
4021 						   poke->bypass_addr,
4022 						   NULL);
4023 			BUG_ON(ret < 0);
4024 		}
4025 	} else {
4026 		t = old_bypass_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
4027 		ret = __bpf_arch_text_poke(poke->tailcall_bypass,
4028 					   t, BPF_MOD_JUMP, old_bypass_addr,
4029 					   poke->bypass_addr);
4030 		BUG_ON(ret < 0);
4031 		/* let other CPUs finish the execution of program
4032 		 * so that it will not possible to expose them
4033 		 * to invalid nop, stack unwind, nop state
4034 		 */
4035 		if (!ret)
4036 			synchronize_rcu();
4037 		t = old_addr ? BPF_MOD_JUMP : BPF_MOD_NOP;
4038 		ret = __bpf_arch_text_poke(poke->tailcall_target,
4039 					   t, BPF_MOD_NOP, old_addr, NULL);
4040 		BUG_ON(ret < 0);
4041 	}
4042 }
4043 
4044 bool bpf_jit_supports_arena(void)
4045 {
4046 	return true;
4047 }
4048 
4049 bool bpf_jit_supports_insn(struct bpf_insn *insn, bool in_arena)
4050 {
4051 	if (!in_arena)
4052 		return true;
4053 	switch (insn->code) {
4054 	case BPF_STX | BPF_ATOMIC | BPF_W:
4055 	case BPF_STX | BPF_ATOMIC | BPF_DW:
4056 		if (insn->imm == (BPF_AND | BPF_FETCH) ||
4057 		    insn->imm == (BPF_OR | BPF_FETCH) ||
4058 		    insn->imm == (BPF_XOR | BPF_FETCH))
4059 			return false;
4060 	}
4061 	return true;
4062 }
4063 
4064 bool bpf_jit_supports_ptr_xchg(void)
4065 {
4066 	return true;
4067 }
4068 
4069 /* x86-64 JIT emits its own code to filter user addresses so return 0 here */
4070 u64 bpf_arch_uaddress_limit(void)
4071 {
4072 	return 0;
4073 }
4074 
4075 bool bpf_jit_supports_timed_may_goto(void)
4076 {
4077 	return true;
4078 }
4079 
4080 bool bpf_jit_supports_fsession(void)
4081 {
4082 	return true;
4083 }
4084