xref: /openbmc/linux/arch/x86/kernel/alternative.c (revision 9134211f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #define pr_fmt(fmt) "SMP alternatives: " fmt
3 
4 #include <linux/module.h>
5 #include <linux/sched.h>
6 #include <linux/perf_event.h>
7 #include <linux/mutex.h>
8 #include <linux/list.h>
9 #include <linux/stringify.h>
10 #include <linux/highmem.h>
11 #include <linux/mm.h>
12 #include <linux/vmalloc.h>
13 #include <linux/memory.h>
14 #include <linux/stop_machine.h>
15 #include <linux/slab.h>
16 #include <linux/kdebug.h>
17 #include <linux/kprobes.h>
18 #include <linux/mmu_context.h>
19 #include <linux/bsearch.h>
20 #include <linux/sync_core.h>
21 #include <asm/text-patching.h>
22 #include <asm/alternative.h>
23 #include <asm/sections.h>
24 #include <asm/mce.h>
25 #include <asm/nmi.h>
26 #include <asm/cacheflush.h>
27 #include <asm/tlbflush.h>
28 #include <asm/insn.h>
29 #include <asm/io.h>
30 #include <asm/fixmap.h>
31 #include <asm/paravirt.h>
32 #include <asm/asm-prototypes.h>
33 
34 int __read_mostly alternatives_patched;
35 
36 EXPORT_SYMBOL_GPL(alternatives_patched);
37 
38 #define MAX_PATCH_LEN (255-1)
39 
40 #define DA_ALL		(~0)
41 #define DA_ALT		0x01
42 #define DA_RET		0x02
43 #define DA_RETPOLINE	0x04
44 #define DA_ENDBR	0x08
45 #define DA_SMP		0x10
46 
47 static unsigned int __initdata_or_module debug_alternative;
48 
49 static int __init debug_alt(char *str)
50 {
51 	if (str && *str == '=')
52 		str++;
53 
54 	if (!str || kstrtouint(str, 0, &debug_alternative))
55 		debug_alternative = DA_ALL;
56 
57 	return 1;
58 }
59 __setup("debug-alternative", debug_alt);
60 
61 static int noreplace_smp;
62 
63 static int __init setup_noreplace_smp(char *str)
64 {
65 	noreplace_smp = 1;
66 	return 1;
67 }
68 __setup("noreplace-smp", setup_noreplace_smp);
69 
70 #define DPRINTK(type, fmt, args...)					\
71 do {									\
72 	if (debug_alternative & DA_##type)				\
73 		printk(KERN_DEBUG pr_fmt(fmt) "\n", ##args);		\
74 } while (0)
75 
76 #define DUMP_BYTES(type, buf, len, fmt, args...)			\
77 do {									\
78 	if (unlikely(debug_alternative & DA_##type)) {			\
79 		int j;							\
80 									\
81 		if (!(len))						\
82 			break;						\
83 									\
84 		printk(KERN_DEBUG pr_fmt(fmt), ##args);			\
85 		for (j = 0; j < (len) - 1; j++)				\
86 			printk(KERN_CONT "%02hhx ", buf[j]);		\
87 		printk(KERN_CONT "%02hhx\n", buf[j]);			\
88 	}								\
89 } while (0)
90 
91 static const unsigned char x86nops[] =
92 {
93 	BYTES_NOP1,
94 	BYTES_NOP2,
95 	BYTES_NOP3,
96 	BYTES_NOP4,
97 	BYTES_NOP5,
98 	BYTES_NOP6,
99 	BYTES_NOP7,
100 	BYTES_NOP8,
101 #ifdef CONFIG_64BIT
102 	BYTES_NOP9,
103 	BYTES_NOP10,
104 	BYTES_NOP11,
105 #endif
106 };
107 
108 const unsigned char * const x86_nops[ASM_NOP_MAX+1] =
109 {
110 	NULL,
111 	x86nops,
112 	x86nops + 1,
113 	x86nops + 1 + 2,
114 	x86nops + 1 + 2 + 3,
115 	x86nops + 1 + 2 + 3 + 4,
116 	x86nops + 1 + 2 + 3 + 4 + 5,
117 	x86nops + 1 + 2 + 3 + 4 + 5 + 6,
118 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
119 #ifdef CONFIG_64BIT
120 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
121 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9,
122 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8 + 9 + 10,
123 #endif
124 };
125 
126 /*
127  * Fill the buffer with a single effective instruction of size @len.
128  *
129  * In order not to issue an ORC stack depth tracking CFI entry (Call Frame Info)
130  * for every single-byte NOP, try to generate the maximally available NOP of
131  * size <= ASM_NOP_MAX such that only a single CFI entry is generated (vs one for
132  * each single-byte NOPs). If @len to fill out is > ASM_NOP_MAX, pad with INT3 and
133  * *jump* over instead of executing long and daft NOPs.
134  */
135 static void __init_or_module add_nop(u8 *instr, unsigned int len)
136 {
137 	u8 *target = instr + len;
138 
139 	if (!len)
140 		return;
141 
142 	if (len <= ASM_NOP_MAX) {
143 		memcpy(instr, x86_nops[len], len);
144 		return;
145 	}
146 
147 	if (len < 128) {
148 		__text_gen_insn(instr, JMP8_INSN_OPCODE, instr, target, JMP8_INSN_SIZE);
149 		instr += JMP8_INSN_SIZE;
150 	} else {
151 		__text_gen_insn(instr, JMP32_INSN_OPCODE, instr, target, JMP32_INSN_SIZE);
152 		instr += JMP32_INSN_SIZE;
153 	}
154 
155 	for (;instr < target; instr++)
156 		*instr = INT3_INSN_OPCODE;
157 }
158 
159 extern s32 __retpoline_sites[], __retpoline_sites_end[];
160 extern s32 __return_sites[], __return_sites_end[];
161 extern s32 __cfi_sites[], __cfi_sites_end[];
162 extern s32 __ibt_endbr_seal[], __ibt_endbr_seal_end[];
163 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
164 extern s32 __smp_locks[], __smp_locks_end[];
165 void text_poke_early(void *addr, const void *opcode, size_t len);
166 
167 /*
168  * Matches NOP and NOPL, not any of the other possible NOPs.
169  */
170 static bool insn_is_nop(struct insn *insn)
171 {
172 	/* Anything NOP, but no REP NOP */
173 	if (insn->opcode.bytes[0] == 0x90 &&
174 	    (!insn->prefixes.nbytes || insn->prefixes.bytes[0] != 0xF3))
175 		return true;
176 
177 	/* NOPL */
178 	if (insn->opcode.bytes[0] == 0x0F && insn->opcode.bytes[1] == 0x1F)
179 		return true;
180 
181 	/* TODO: more nops */
182 
183 	return false;
184 }
185 
186 /*
187  * Find the offset of the first non-NOP instruction starting at @offset
188  * but no further than @len.
189  */
190 static int skip_nops(u8 *instr, int offset, int len)
191 {
192 	struct insn insn;
193 
194 	for (; offset < len; offset += insn.length) {
195 		if (insn_decode_kernel(&insn, &instr[offset]))
196 			break;
197 
198 		if (!insn_is_nop(&insn))
199 			break;
200 	}
201 
202 	return offset;
203 }
204 
205 /*
206  * Optimize a sequence of NOPs, possibly preceded by an unconditional jump
207  * to the end of the NOP sequence into a single NOP.
208  */
209 static bool __init_or_module
210 __optimize_nops(u8 *instr, size_t len, struct insn *insn, int *next, int *prev, int *target)
211 {
212 	int i = *next - insn->length;
213 
214 	switch (insn->opcode.bytes[0]) {
215 	case JMP8_INSN_OPCODE:
216 	case JMP32_INSN_OPCODE:
217 		*prev = i;
218 		*target = *next + insn->immediate.value;
219 		return false;
220 	}
221 
222 	if (insn_is_nop(insn)) {
223 		int nop = i;
224 
225 		*next = skip_nops(instr, *next, len);
226 		if (*target && *next == *target)
227 			nop = *prev;
228 
229 		add_nop(instr + nop, *next - nop);
230 		DUMP_BYTES(ALT, instr, len, "%px: [%d:%d) optimized NOPs: ", instr, nop, *next);
231 		return true;
232 	}
233 
234 	*target = 0;
235 	return false;
236 }
237 
238 /*
239  * "noinline" to cause control flow change and thus invalidate I$ and
240  * cause refetch after modification.
241  */
242 static void __init_or_module noinline optimize_nops(u8 *instr, size_t len)
243 {
244 	int prev, target = 0;
245 
246 	for (int next, i = 0; i < len; i = next) {
247 		struct insn insn;
248 
249 		if (insn_decode_kernel(&insn, &instr[i]))
250 			return;
251 
252 		next = i + insn.length;
253 
254 		__optimize_nops(instr, len, &insn, &next, &prev, &target);
255 	}
256 }
257 
258 /*
259  * In this context, "source" is where the instructions are placed in the
260  * section .altinstr_replacement, for example during kernel build by the
261  * toolchain.
262  * "Destination" is where the instructions are being patched in by this
263  * machinery.
264  *
265  * The source offset is:
266  *
267  *   src_imm = target - src_next_ip                  (1)
268  *
269  * and the target offset is:
270  *
271  *   dst_imm = target - dst_next_ip                  (2)
272  *
273  * so rework (1) as an expression for target like:
274  *
275  *   target = src_imm + src_next_ip                  (1a)
276  *
277  * and substitute in (2) to get:
278  *
279  *   dst_imm = (src_imm + src_next_ip) - dst_next_ip (3)
280  *
281  * Now, since the instruction stream is 'identical' at src and dst (it
282  * is being copied after all) it can be stated that:
283  *
284  *   src_next_ip = src + ip_offset
285  *   dst_next_ip = dst + ip_offset                   (4)
286  *
287  * Substitute (4) in (3) and observe ip_offset being cancelled out to
288  * obtain:
289  *
290  *   dst_imm = src_imm + (src + ip_offset) - (dst + ip_offset)
291  *           = src_imm + src - dst + ip_offset - ip_offset
292  *           = src_imm + src - dst                   (5)
293  *
294  * IOW, only the relative displacement of the code block matters.
295  */
296 
297 #define apply_reloc_n(n_, p_, d_)				\
298 	do {							\
299 		s32 v = *(s##n_ *)(p_);				\
300 		v += (d_);					\
301 		BUG_ON((v >> 31) != (v >> (n_-1)));		\
302 		*(s##n_ *)(p_) = (s##n_)v;			\
303 	} while (0)
304 
305 
306 static __always_inline
307 void apply_reloc(int n, void *ptr, uintptr_t diff)
308 {
309 	switch (n) {
310 	case 1: apply_reloc_n(8, ptr, diff); break;
311 	case 2: apply_reloc_n(16, ptr, diff); break;
312 	case 4: apply_reloc_n(32, ptr, diff); break;
313 	default: BUG();
314 	}
315 }
316 
317 static __always_inline
318 bool need_reloc(unsigned long offset, u8 *src, size_t src_len)
319 {
320 	u8 *target = src + offset;
321 	/*
322 	 * If the target is inside the patched block, it's relative to the
323 	 * block itself and does not need relocation.
324 	 */
325 	return (target < src || target > src + src_len);
326 }
327 
328 static void __init_or_module noinline
329 apply_relocation(u8 *buf, size_t len, u8 *dest, u8 *src, size_t src_len)
330 {
331 	int prev, target = 0;
332 
333 	for (int next, i = 0; i < len; i = next) {
334 		struct insn insn;
335 
336 		if (WARN_ON_ONCE(insn_decode_kernel(&insn, &buf[i])))
337 			return;
338 
339 		next = i + insn.length;
340 
341 		if (__optimize_nops(buf, len, &insn, &next, &prev, &target))
342 			continue;
343 
344 		switch (insn.opcode.bytes[0]) {
345 		case 0x0f:
346 			if (insn.opcode.bytes[1] < 0x80 ||
347 			    insn.opcode.bytes[1] > 0x8f)
348 				break;
349 
350 			fallthrough;	/* Jcc.d32 */
351 		case 0x70 ... 0x7f:	/* Jcc.d8 */
352 		case JMP8_INSN_OPCODE:
353 		case JMP32_INSN_OPCODE:
354 		case CALL_INSN_OPCODE:
355 			if (need_reloc(next + insn.immediate.value, src, src_len)) {
356 				apply_reloc(insn.immediate.nbytes,
357 					    buf + i + insn_offset_immediate(&insn),
358 					    src - dest);
359 			}
360 
361 			/*
362 			 * Where possible, convert JMP.d32 into JMP.d8.
363 			 */
364 			if (insn.opcode.bytes[0] == JMP32_INSN_OPCODE) {
365 				s32 imm = insn.immediate.value;
366 				imm += src - dest;
367 				imm += JMP32_INSN_SIZE - JMP8_INSN_SIZE;
368 				if ((imm >> 31) == (imm >> 7)) {
369 					buf[i+0] = JMP8_INSN_OPCODE;
370 					buf[i+1] = (s8)imm;
371 
372 					memset(&buf[i+2], INT3_INSN_OPCODE, insn.length - 2);
373 				}
374 			}
375 			break;
376 		}
377 
378 		if (insn_rip_relative(&insn)) {
379 			if (need_reloc(next + insn.displacement.value, src, src_len)) {
380 				apply_reloc(insn.displacement.nbytes,
381 					    buf + i + insn_offset_displacement(&insn),
382 					    src - dest);
383 			}
384 		}
385 	}
386 }
387 
388 /*
389  * Replace instructions with better alternatives for this CPU type. This runs
390  * before SMP is initialized to avoid SMP problems with self modifying code.
391  * This implies that asymmetric systems where APs have less capabilities than
392  * the boot processor are not handled. Tough. Make sure you disable such
393  * features by hand.
394  *
395  * Marked "noinline" to cause control flow change and thus insn cache
396  * to refetch changed I$ lines.
397  */
398 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
399 						  struct alt_instr *end)
400 {
401 	struct alt_instr *a;
402 	u8 *instr, *replacement;
403 	u8 insn_buff[MAX_PATCH_LEN];
404 
405 	DPRINTK(ALT, "alt table %px, -> %px", start, end);
406 	/*
407 	 * The scan order should be from start to end. A later scanned
408 	 * alternative code can overwrite previously scanned alternative code.
409 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
410 	 * patch code.
411 	 *
412 	 * So be careful if you want to change the scan order to any other
413 	 * order.
414 	 */
415 	for (a = start; a < end; a++) {
416 		int insn_buff_sz = 0;
417 
418 		instr = (u8 *)&a->instr_offset + a->instr_offset;
419 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
420 		BUG_ON(a->instrlen > sizeof(insn_buff));
421 		BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
422 
423 		/*
424 		 * Patch if either:
425 		 * - feature is present
426 		 * - feature not present but ALT_FLAG_NOT is set to mean,
427 		 *   patch if feature is *NOT* present.
428 		 */
429 		if (!boot_cpu_has(a->cpuid) == !(a->flags & ALT_FLAG_NOT)) {
430 			optimize_nops(instr, a->instrlen);
431 			continue;
432 		}
433 
434 		DPRINTK(ALT, "feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)",
435 			(a->flags & ALT_FLAG_NOT) ? "!" : "",
436 			a->cpuid >> 5,
437 			a->cpuid & 0x1f,
438 			instr, instr, a->instrlen,
439 			replacement, a->replacementlen);
440 
441 		memcpy(insn_buff, replacement, a->replacementlen);
442 		insn_buff_sz = a->replacementlen;
443 
444 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
445 			insn_buff[insn_buff_sz] = 0x90;
446 
447 		apply_relocation(insn_buff, a->instrlen, instr, replacement, a->replacementlen);
448 
449 		DUMP_BYTES(ALT, instr, a->instrlen, "%px:   old_insn: ", instr);
450 		DUMP_BYTES(ALT, replacement, a->replacementlen, "%px:   rpl_insn: ", replacement);
451 		DUMP_BYTES(ALT, insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
452 
453 		text_poke_early(instr, insn_buff, insn_buff_sz);
454 	}
455 }
456 
457 static inline bool is_jcc32(struct insn *insn)
458 {
459 	/* Jcc.d32 second opcode byte is in the range: 0x80-0x8f */
460 	return insn->opcode.bytes[0] == 0x0f && (insn->opcode.bytes[1] & 0xf0) == 0x80;
461 }
462 
463 #if defined(CONFIG_RETPOLINE) && defined(CONFIG_OBJTOOL)
464 
465 /*
466  * CALL/JMP *%\reg
467  */
468 static int emit_indirect(int op, int reg, u8 *bytes)
469 {
470 	int i = 0;
471 	u8 modrm;
472 
473 	switch (op) {
474 	case CALL_INSN_OPCODE:
475 		modrm = 0x10; /* Reg = 2; CALL r/m */
476 		break;
477 
478 	case JMP32_INSN_OPCODE:
479 		modrm = 0x20; /* Reg = 4; JMP r/m */
480 		break;
481 
482 	default:
483 		WARN_ON_ONCE(1);
484 		return -1;
485 	}
486 
487 	if (reg >= 8) {
488 		bytes[i++] = 0x41; /* REX.B prefix */
489 		reg -= 8;
490 	}
491 
492 	modrm |= 0xc0; /* Mod = 3 */
493 	modrm += reg;
494 
495 	bytes[i++] = 0xff; /* opcode */
496 	bytes[i++] = modrm;
497 
498 	return i;
499 }
500 
501 static int emit_call_track_retpoline(void *addr, struct insn *insn, int reg, u8 *bytes)
502 {
503 	u8 op = insn->opcode.bytes[0];
504 	int i = 0;
505 
506 	/*
507 	 * Clang does 'weird' Jcc __x86_indirect_thunk_r11 conditional
508 	 * tail-calls. Deal with them.
509 	 */
510 	if (is_jcc32(insn)) {
511 		bytes[i++] = op;
512 		op = insn->opcode.bytes[1];
513 		goto clang_jcc;
514 	}
515 
516 	if (insn->length == 6)
517 		bytes[i++] = 0x2e; /* CS-prefix */
518 
519 	switch (op) {
520 	case CALL_INSN_OPCODE:
521 		__text_gen_insn(bytes+i, op, addr+i,
522 				__x86_indirect_call_thunk_array[reg],
523 				CALL_INSN_SIZE);
524 		i += CALL_INSN_SIZE;
525 		break;
526 
527 	case JMP32_INSN_OPCODE:
528 clang_jcc:
529 		__text_gen_insn(bytes+i, op, addr+i,
530 				__x86_indirect_jump_thunk_array[reg],
531 				JMP32_INSN_SIZE);
532 		i += JMP32_INSN_SIZE;
533 		break;
534 
535 	default:
536 		WARN(1, "%pS %px %*ph\n", addr, addr, 6, addr);
537 		return -1;
538 	}
539 
540 	WARN_ON_ONCE(i != insn->length);
541 
542 	return i;
543 }
544 
545 /*
546  * Rewrite the compiler generated retpoline thunk calls.
547  *
548  * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate
549  * indirect instructions, avoiding the extra indirection.
550  *
551  * For example, convert:
552  *
553  *   CALL __x86_indirect_thunk_\reg
554  *
555  * into:
556  *
557  *   CALL *%\reg
558  *
559  * It also tries to inline spectre_v2=retpoline,lfence when size permits.
560  */
561 static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes)
562 {
563 	retpoline_thunk_t *target;
564 	int reg, ret, i = 0;
565 	u8 op, cc;
566 
567 	target = addr + insn->length + insn->immediate.value;
568 	reg = target - __x86_indirect_thunk_array;
569 
570 	if (WARN_ON_ONCE(reg & ~0xf))
571 		return -1;
572 
573 	/* If anyone ever does: CALL/JMP *%rsp, we're in deep trouble. */
574 	BUG_ON(reg == 4);
575 
576 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE) &&
577 	    !cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
578 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
579 			return emit_call_track_retpoline(addr, insn, reg, bytes);
580 
581 		return -1;
582 	}
583 
584 	op = insn->opcode.bytes[0];
585 
586 	/*
587 	 * Convert:
588 	 *
589 	 *   Jcc.d32 __x86_indirect_thunk_\reg
590 	 *
591 	 * into:
592 	 *
593 	 *   Jncc.d8 1f
594 	 *   [ LFENCE ]
595 	 *   JMP *%\reg
596 	 *   [ NOP ]
597 	 * 1:
598 	 */
599 	if (is_jcc32(insn)) {
600 		cc = insn->opcode.bytes[1] & 0xf;
601 		cc ^= 1; /* invert condition */
602 
603 		bytes[i++] = 0x70 + cc;        /* Jcc.d8 */
604 		bytes[i++] = insn->length - 2; /* sizeof(Jcc.d8) == 2 */
605 
606 		/* Continue as if: JMP.d32 __x86_indirect_thunk_\reg */
607 		op = JMP32_INSN_OPCODE;
608 	}
609 
610 	/*
611 	 * For RETPOLINE_LFENCE: prepend the indirect CALL/JMP with an LFENCE.
612 	 */
613 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
614 		bytes[i++] = 0x0f;
615 		bytes[i++] = 0xae;
616 		bytes[i++] = 0xe8; /* LFENCE */
617 	}
618 
619 	ret = emit_indirect(op, reg, bytes + i);
620 	if (ret < 0)
621 		return ret;
622 	i += ret;
623 
624 	/*
625 	 * The compiler is supposed to EMIT an INT3 after every unconditional
626 	 * JMP instruction due to AMD BTC. However, if the compiler is too old
627 	 * or SLS isn't enabled, we still need an INT3 after indirect JMPs
628 	 * even on Intel.
629 	 */
630 	if (op == JMP32_INSN_OPCODE && i < insn->length)
631 		bytes[i++] = INT3_INSN_OPCODE;
632 
633 	for (; i < insn->length;)
634 		bytes[i++] = BYTES_NOP1;
635 
636 	return i;
637 }
638 
639 /*
640  * Generated by 'objtool --retpoline'.
641  */
642 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end)
643 {
644 	s32 *s;
645 
646 	for (s = start; s < end; s++) {
647 		void *addr = (void *)s + *s;
648 		struct insn insn;
649 		int len, ret;
650 		u8 bytes[16];
651 		u8 op1, op2;
652 
653 		ret = insn_decode_kernel(&insn, addr);
654 		if (WARN_ON_ONCE(ret < 0))
655 			continue;
656 
657 		op1 = insn.opcode.bytes[0];
658 		op2 = insn.opcode.bytes[1];
659 
660 		switch (op1) {
661 		case CALL_INSN_OPCODE:
662 		case JMP32_INSN_OPCODE:
663 			break;
664 
665 		case 0x0f: /* escape */
666 			if (op2 >= 0x80 && op2 <= 0x8f)
667 				break;
668 			fallthrough;
669 		default:
670 			WARN_ON_ONCE(1);
671 			continue;
672 		}
673 
674 		DPRINTK(RETPOLINE, "retpoline at: %pS (%px) len: %d to: %pS",
675 			addr, addr, insn.length,
676 			addr + insn.length + insn.immediate.value);
677 
678 		len = patch_retpoline(addr, &insn, bytes);
679 		if (len == insn.length) {
680 			optimize_nops(bytes, len);
681 			DUMP_BYTES(RETPOLINE, ((u8*)addr),  len, "%px: orig: ", addr);
682 			DUMP_BYTES(RETPOLINE, ((u8*)bytes), len, "%px: repl: ", addr);
683 			text_poke_early(addr, bytes, len);
684 		}
685 	}
686 }
687 
688 #ifdef CONFIG_RETHUNK
689 
690 #ifdef CONFIG_CALL_THUNKS
691 void (*x86_return_thunk)(void) __ro_after_init = &__x86_return_thunk;
692 #endif
693 
694 /*
695  * Rewrite the compiler generated return thunk tail-calls.
696  *
697  * For example, convert:
698  *
699  *   JMP __x86_return_thunk
700  *
701  * into:
702  *
703  *   RET
704  */
705 static int patch_return(void *addr, struct insn *insn, u8 *bytes)
706 {
707 	int i = 0;
708 
709 	/* Patch the custom return thunks... */
710 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
711 		i = JMP32_INSN_SIZE;
712 		__text_gen_insn(bytes, JMP32_INSN_OPCODE, addr, x86_return_thunk, i);
713 	} else {
714 		/* ... or patch them out if not needed. */
715 		bytes[i++] = RET_INSN_OPCODE;
716 	}
717 
718 	for (; i < insn->length;)
719 		bytes[i++] = INT3_INSN_OPCODE;
720 	return i;
721 }
722 
723 void __init_or_module noinline apply_returns(s32 *start, s32 *end)
724 {
725 	s32 *s;
726 
727 	/*
728 	 * Do not patch out the default return thunks if those needed are the
729 	 * ones generated by the compiler.
730 	 */
731 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK) &&
732 	    (x86_return_thunk == __x86_return_thunk))
733 		return;
734 
735 	for (s = start; s < end; s++) {
736 		void *dest = NULL, *addr = (void *)s + *s;
737 		struct insn insn;
738 		int len, ret;
739 		u8 bytes[16];
740 		u8 op;
741 
742 		ret = insn_decode_kernel(&insn, addr);
743 		if (WARN_ON_ONCE(ret < 0))
744 			continue;
745 
746 		op = insn.opcode.bytes[0];
747 		if (op == JMP32_INSN_OPCODE)
748 			dest = addr + insn.length + insn.immediate.value;
749 
750 		if (__static_call_fixup(addr, op, dest) ||
751 		    WARN_ONCE(dest != &__x86_return_thunk,
752 			      "missing return thunk: %pS-%pS: %*ph",
753 			      addr, dest, 5, addr))
754 			continue;
755 
756 		DPRINTK(RET, "return thunk at: %pS (%px) len: %d to: %pS",
757 			addr, addr, insn.length,
758 			addr + insn.length + insn.immediate.value);
759 
760 		len = patch_return(addr, &insn, bytes);
761 		if (len == insn.length) {
762 			DUMP_BYTES(RET, ((u8*)addr),  len, "%px: orig: ", addr);
763 			DUMP_BYTES(RET, ((u8*)bytes), len, "%px: repl: ", addr);
764 			text_poke_early(addr, bytes, len);
765 		}
766 	}
767 }
768 #else
769 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
770 #endif /* CONFIG_RETHUNK */
771 
772 #else /* !CONFIG_RETPOLINE || !CONFIG_OBJTOOL */
773 
774 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) { }
775 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
776 
777 #endif /* CONFIG_RETPOLINE && CONFIG_OBJTOOL */
778 
779 #ifdef CONFIG_X86_KERNEL_IBT
780 
781 static void __init_or_module poison_endbr(void *addr, bool warn)
782 {
783 	u32 endbr, poison = gen_endbr_poison();
784 
785 	if (WARN_ON_ONCE(get_kernel_nofault(endbr, addr)))
786 		return;
787 
788 	if (!is_endbr(endbr)) {
789 		WARN_ON_ONCE(warn);
790 		return;
791 	}
792 
793 	DPRINTK(ENDBR, "ENDBR at: %pS (%px)", addr, addr);
794 
795 	/*
796 	 * When we have IBT, the lack of ENDBR will trigger #CP
797 	 */
798 	DUMP_BYTES(ENDBR, ((u8*)addr), 4, "%px: orig: ", addr);
799 	DUMP_BYTES(ENDBR, ((u8*)&poison), 4, "%px: repl: ", addr);
800 	text_poke_early(addr, &poison, 4);
801 }
802 
803 /*
804  * Generated by: objtool --ibt
805  */
806 void __init_or_module noinline apply_ibt_endbr(s32 *start, s32 *end)
807 {
808 	s32 *s;
809 
810 	for (s = start; s < end; s++) {
811 		void *addr = (void *)s + *s;
812 
813 		poison_endbr(addr, true);
814 		if (IS_ENABLED(CONFIG_FINEIBT))
815 			poison_endbr(addr - 16, false);
816 	}
817 }
818 
819 #else
820 
821 void __init_or_module apply_ibt_endbr(s32 *start, s32 *end) { }
822 
823 #endif /* CONFIG_X86_KERNEL_IBT */
824 
825 #ifdef CONFIG_FINEIBT
826 
827 enum cfi_mode {
828 	CFI_DEFAULT,
829 	CFI_OFF,
830 	CFI_KCFI,
831 	CFI_FINEIBT,
832 };
833 
834 static enum cfi_mode cfi_mode __ro_after_init = CFI_DEFAULT;
835 static bool cfi_rand __ro_after_init = true;
836 static u32  cfi_seed __ro_after_init;
837 
838 /*
839  * Re-hash the CFI hash with a boot-time seed while making sure the result is
840  * not a valid ENDBR instruction.
841  */
842 static u32 cfi_rehash(u32 hash)
843 {
844 	hash ^= cfi_seed;
845 	while (unlikely(is_endbr(hash) || is_endbr(-hash))) {
846 		bool lsb = hash & 1;
847 		hash >>= 1;
848 		if (lsb)
849 			hash ^= 0x80200003;
850 	}
851 	return hash;
852 }
853 
854 static __init int cfi_parse_cmdline(char *str)
855 {
856 	if (!str)
857 		return -EINVAL;
858 
859 	while (str) {
860 		char *next = strchr(str, ',');
861 		if (next) {
862 			*next = 0;
863 			next++;
864 		}
865 
866 		if (!strcmp(str, "auto")) {
867 			cfi_mode = CFI_DEFAULT;
868 		} else if (!strcmp(str, "off")) {
869 			cfi_mode = CFI_OFF;
870 			cfi_rand = false;
871 		} else if (!strcmp(str, "kcfi")) {
872 			cfi_mode = CFI_KCFI;
873 		} else if (!strcmp(str, "fineibt")) {
874 			cfi_mode = CFI_FINEIBT;
875 		} else if (!strcmp(str, "norand")) {
876 			cfi_rand = false;
877 		} else {
878 			pr_err("Ignoring unknown cfi option (%s).", str);
879 		}
880 
881 		str = next;
882 	}
883 
884 	return 0;
885 }
886 early_param("cfi", cfi_parse_cmdline);
887 
888 /*
889  * kCFI						FineIBT
890  *
891  * __cfi_\func:					__cfi_\func:
892  *	movl   $0x12345678,%eax		// 5	     endbr64			// 4
893  *	nop					     subl   $0x12345678,%r10d   // 7
894  *	nop					     jz     1f			// 2
895  *	nop					     ud2			// 2
896  *	nop					1:   nop			// 1
897  *	nop
898  *	nop
899  *	nop
900  *	nop
901  *	nop
902  *	nop
903  *	nop
904  *
905  *
906  * caller:					caller:
907  *	movl	$(-0x12345678),%r10d	 // 6	     movl   $0x12345678,%r10d	// 6
908  *	addl	$-15(%r11),%r10d	 // 4	     sub    $16,%r11		// 4
909  *	je	1f			 // 2	     nop4			// 4
910  *	ud2				 // 2
911  * 1:	call	__x86_indirect_thunk_r11 // 5	     call   *%r11; nop2;	// 5
912  *
913  */
914 
915 asm(	".pushsection .rodata			\n"
916 	"fineibt_preamble_start:		\n"
917 	"	endbr64				\n"
918 	"	subl	$0x12345678, %r10d	\n"
919 	"	je	fineibt_preamble_end	\n"
920 	"	ud2				\n"
921 	"	nop				\n"
922 	"fineibt_preamble_end:			\n"
923 	".popsection\n"
924 );
925 
926 extern u8 fineibt_preamble_start[];
927 extern u8 fineibt_preamble_end[];
928 
929 #define fineibt_preamble_size (fineibt_preamble_end - fineibt_preamble_start)
930 #define fineibt_preamble_hash 7
931 
932 asm(	".pushsection .rodata			\n"
933 	"fineibt_caller_start:			\n"
934 	"	movl	$0x12345678, %r10d	\n"
935 	"	sub	$16, %r11		\n"
936 	ASM_NOP4
937 	"fineibt_caller_end:			\n"
938 	".popsection				\n"
939 );
940 
941 extern u8 fineibt_caller_start[];
942 extern u8 fineibt_caller_end[];
943 
944 #define fineibt_caller_size (fineibt_caller_end - fineibt_caller_start)
945 #define fineibt_caller_hash 2
946 
947 #define fineibt_caller_jmp (fineibt_caller_size - 2)
948 
949 static u32 decode_preamble_hash(void *addr)
950 {
951 	u8 *p = addr;
952 
953 	/* b8 78 56 34 12          mov    $0x12345678,%eax */
954 	if (p[0] == 0xb8)
955 		return *(u32 *)(addr + 1);
956 
957 	return 0; /* invalid hash value */
958 }
959 
960 static u32 decode_caller_hash(void *addr)
961 {
962 	u8 *p = addr;
963 
964 	/* 41 ba 78 56 34 12       mov    $0x12345678,%r10d */
965 	if (p[0] == 0x41 && p[1] == 0xba)
966 		return -*(u32 *)(addr + 2);
967 
968 	/* e8 0c 78 56 34 12	   jmp.d8  +12 */
969 	if (p[0] == JMP8_INSN_OPCODE && p[1] == fineibt_caller_jmp)
970 		return -*(u32 *)(addr + 2);
971 
972 	return 0; /* invalid hash value */
973 }
974 
975 /* .retpoline_sites */
976 static int cfi_disable_callers(s32 *start, s32 *end)
977 {
978 	/*
979 	 * Disable kCFI by patching in a JMP.d8, this leaves the hash immediate
980 	 * in tact for later usage. Also see decode_caller_hash() and
981 	 * cfi_rewrite_callers().
982 	 */
983 	const u8 jmp[] = { JMP8_INSN_OPCODE, fineibt_caller_jmp };
984 	s32 *s;
985 
986 	for (s = start; s < end; s++) {
987 		void *addr = (void *)s + *s;
988 		u32 hash;
989 
990 		addr -= fineibt_caller_size;
991 		hash = decode_caller_hash(addr);
992 		if (!hash) /* nocfi callers */
993 			continue;
994 
995 		text_poke_early(addr, jmp, 2);
996 	}
997 
998 	return 0;
999 }
1000 
1001 static int cfi_enable_callers(s32 *start, s32 *end)
1002 {
1003 	/*
1004 	 * Re-enable kCFI, undo what cfi_disable_callers() did.
1005 	 */
1006 	const u8 mov[] = { 0x41, 0xba };
1007 	s32 *s;
1008 
1009 	for (s = start; s < end; s++) {
1010 		void *addr = (void *)s + *s;
1011 		u32 hash;
1012 
1013 		addr -= fineibt_caller_size;
1014 		hash = decode_caller_hash(addr);
1015 		if (!hash) /* nocfi callers */
1016 			continue;
1017 
1018 		text_poke_early(addr, mov, 2);
1019 	}
1020 
1021 	return 0;
1022 }
1023 
1024 /* .cfi_sites */
1025 static int cfi_rand_preamble(s32 *start, s32 *end)
1026 {
1027 	s32 *s;
1028 
1029 	for (s = start; s < end; s++) {
1030 		void *addr = (void *)s + *s;
1031 		u32 hash;
1032 
1033 		hash = decode_preamble_hash(addr);
1034 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1035 			 addr, addr, 5, addr))
1036 			return -EINVAL;
1037 
1038 		hash = cfi_rehash(hash);
1039 		text_poke_early(addr + 1, &hash, 4);
1040 	}
1041 
1042 	return 0;
1043 }
1044 
1045 static int cfi_rewrite_preamble(s32 *start, s32 *end)
1046 {
1047 	s32 *s;
1048 
1049 	for (s = start; s < end; s++) {
1050 		void *addr = (void *)s + *s;
1051 		u32 hash;
1052 
1053 		hash = decode_preamble_hash(addr);
1054 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1055 			 addr, addr, 5, addr))
1056 			return -EINVAL;
1057 
1058 		text_poke_early(addr, fineibt_preamble_start, fineibt_preamble_size);
1059 		WARN_ON(*(u32 *)(addr + fineibt_preamble_hash) != 0x12345678);
1060 		text_poke_early(addr + fineibt_preamble_hash, &hash, 4);
1061 	}
1062 
1063 	return 0;
1064 }
1065 
1066 /* .retpoline_sites */
1067 static int cfi_rand_callers(s32 *start, s32 *end)
1068 {
1069 	s32 *s;
1070 
1071 	for (s = start; s < end; s++) {
1072 		void *addr = (void *)s + *s;
1073 		u32 hash;
1074 
1075 		addr -= fineibt_caller_size;
1076 		hash = decode_caller_hash(addr);
1077 		if (hash) {
1078 			hash = -cfi_rehash(hash);
1079 			text_poke_early(addr + 2, &hash, 4);
1080 		}
1081 	}
1082 
1083 	return 0;
1084 }
1085 
1086 static int cfi_rewrite_callers(s32 *start, s32 *end)
1087 {
1088 	s32 *s;
1089 
1090 	for (s = start; s < end; s++) {
1091 		void *addr = (void *)s + *s;
1092 		u32 hash;
1093 
1094 		addr -= fineibt_caller_size;
1095 		hash = decode_caller_hash(addr);
1096 		if (hash) {
1097 			text_poke_early(addr, fineibt_caller_start, fineibt_caller_size);
1098 			WARN_ON(*(u32 *)(addr + fineibt_caller_hash) != 0x12345678);
1099 			text_poke_early(addr + fineibt_caller_hash, &hash, 4);
1100 		}
1101 		/* rely on apply_retpolines() */
1102 	}
1103 
1104 	return 0;
1105 }
1106 
1107 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1108 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1109 {
1110 	int ret;
1111 
1112 	if (WARN_ONCE(fineibt_preamble_size != 16,
1113 		      "FineIBT preamble wrong size: %ld", fineibt_preamble_size))
1114 		return;
1115 
1116 	if (cfi_mode == CFI_DEFAULT) {
1117 		cfi_mode = CFI_KCFI;
1118 		if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT))
1119 			cfi_mode = CFI_FINEIBT;
1120 	}
1121 
1122 	/*
1123 	 * Rewrite the callers to not use the __cfi_ stubs, such that we might
1124 	 * rewrite them. This disables all CFI. If this succeeds but any of the
1125 	 * later stages fails, we're without CFI.
1126 	 */
1127 	ret = cfi_disable_callers(start_retpoline, end_retpoline);
1128 	if (ret)
1129 		goto err;
1130 
1131 	if (cfi_rand) {
1132 		if (builtin)
1133 			cfi_seed = get_random_u32();
1134 
1135 		ret = cfi_rand_preamble(start_cfi, end_cfi);
1136 		if (ret)
1137 			goto err;
1138 
1139 		ret = cfi_rand_callers(start_retpoline, end_retpoline);
1140 		if (ret)
1141 			goto err;
1142 	}
1143 
1144 	switch (cfi_mode) {
1145 	case CFI_OFF:
1146 		if (builtin)
1147 			pr_info("Disabling CFI\n");
1148 		return;
1149 
1150 	case CFI_KCFI:
1151 		ret = cfi_enable_callers(start_retpoline, end_retpoline);
1152 		if (ret)
1153 			goto err;
1154 
1155 		if (builtin)
1156 			pr_info("Using kCFI\n");
1157 		return;
1158 
1159 	case CFI_FINEIBT:
1160 		ret = cfi_rewrite_preamble(start_cfi, end_cfi);
1161 		if (ret)
1162 			goto err;
1163 
1164 		ret = cfi_rewrite_callers(start_retpoline, end_retpoline);
1165 		if (ret)
1166 			goto err;
1167 
1168 		if (builtin)
1169 			pr_info("Using FineIBT CFI\n");
1170 		return;
1171 
1172 	default:
1173 		break;
1174 	}
1175 
1176 err:
1177 	pr_err("Something went horribly wrong trying to rewrite the CFI implementation.\n");
1178 }
1179 
1180 #else
1181 
1182 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1183 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1184 {
1185 }
1186 
1187 #endif
1188 
1189 void apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1190 		   s32 *start_cfi, s32 *end_cfi)
1191 {
1192 	return __apply_fineibt(start_retpoline, end_retpoline,
1193 			       start_cfi, end_cfi,
1194 			       /* .builtin = */ false);
1195 }
1196 
1197 #ifdef CONFIG_SMP
1198 static void alternatives_smp_lock(const s32 *start, const s32 *end,
1199 				  u8 *text, u8 *text_end)
1200 {
1201 	const s32 *poff;
1202 
1203 	for (poff = start; poff < end; poff++) {
1204 		u8 *ptr = (u8 *)poff + *poff;
1205 
1206 		if (!*poff || ptr < text || ptr >= text_end)
1207 			continue;
1208 		/* turn DS segment override prefix into lock prefix */
1209 		if (*ptr == 0x3e)
1210 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
1211 	}
1212 }
1213 
1214 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
1215 				    u8 *text, u8 *text_end)
1216 {
1217 	const s32 *poff;
1218 
1219 	for (poff = start; poff < end; poff++) {
1220 		u8 *ptr = (u8 *)poff + *poff;
1221 
1222 		if (!*poff || ptr < text || ptr >= text_end)
1223 			continue;
1224 		/* turn lock prefix into DS segment override prefix */
1225 		if (*ptr == 0xf0)
1226 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
1227 	}
1228 }
1229 
1230 struct smp_alt_module {
1231 	/* what is this ??? */
1232 	struct module	*mod;
1233 	char		*name;
1234 
1235 	/* ptrs to lock prefixes */
1236 	const s32	*locks;
1237 	const s32	*locks_end;
1238 
1239 	/* .text segment, needed to avoid patching init code ;) */
1240 	u8		*text;
1241 	u8		*text_end;
1242 
1243 	struct list_head next;
1244 };
1245 static LIST_HEAD(smp_alt_modules);
1246 static bool uniproc_patched = false;	/* protected by text_mutex */
1247 
1248 void __init_or_module alternatives_smp_module_add(struct module *mod,
1249 						  char *name,
1250 						  void *locks, void *locks_end,
1251 						  void *text,  void *text_end)
1252 {
1253 	struct smp_alt_module *smp;
1254 
1255 	mutex_lock(&text_mutex);
1256 	if (!uniproc_patched)
1257 		goto unlock;
1258 
1259 	if (num_possible_cpus() == 1)
1260 		/* Don't bother remembering, we'll never have to undo it. */
1261 		goto smp_unlock;
1262 
1263 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
1264 	if (NULL == smp)
1265 		/* we'll run the (safe but slow) SMP code then ... */
1266 		goto unlock;
1267 
1268 	smp->mod	= mod;
1269 	smp->name	= name;
1270 	smp->locks	= locks;
1271 	smp->locks_end	= locks_end;
1272 	smp->text	= text;
1273 	smp->text_end	= text_end;
1274 	DPRINTK(SMP, "locks %p -> %p, text %p -> %p, name %s\n",
1275 		smp->locks, smp->locks_end,
1276 		smp->text, smp->text_end, smp->name);
1277 
1278 	list_add_tail(&smp->next, &smp_alt_modules);
1279 smp_unlock:
1280 	alternatives_smp_unlock(locks, locks_end, text, text_end);
1281 unlock:
1282 	mutex_unlock(&text_mutex);
1283 }
1284 
1285 void __init_or_module alternatives_smp_module_del(struct module *mod)
1286 {
1287 	struct smp_alt_module *item;
1288 
1289 	mutex_lock(&text_mutex);
1290 	list_for_each_entry(item, &smp_alt_modules, next) {
1291 		if (mod != item->mod)
1292 			continue;
1293 		list_del(&item->next);
1294 		kfree(item);
1295 		break;
1296 	}
1297 	mutex_unlock(&text_mutex);
1298 }
1299 
1300 void alternatives_enable_smp(void)
1301 {
1302 	struct smp_alt_module *mod;
1303 
1304 	/* Why bother if there are no other CPUs? */
1305 	BUG_ON(num_possible_cpus() == 1);
1306 
1307 	mutex_lock(&text_mutex);
1308 
1309 	if (uniproc_patched) {
1310 		pr_info("switching to SMP code\n");
1311 		BUG_ON(num_online_cpus() != 1);
1312 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
1313 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
1314 		list_for_each_entry(mod, &smp_alt_modules, next)
1315 			alternatives_smp_lock(mod->locks, mod->locks_end,
1316 					      mod->text, mod->text_end);
1317 		uniproc_patched = false;
1318 	}
1319 	mutex_unlock(&text_mutex);
1320 }
1321 
1322 /*
1323  * Return 1 if the address range is reserved for SMP-alternatives.
1324  * Must hold text_mutex.
1325  */
1326 int alternatives_text_reserved(void *start, void *end)
1327 {
1328 	struct smp_alt_module *mod;
1329 	const s32 *poff;
1330 	u8 *text_start = start;
1331 	u8 *text_end = end;
1332 
1333 	lockdep_assert_held(&text_mutex);
1334 
1335 	list_for_each_entry(mod, &smp_alt_modules, next) {
1336 		if (mod->text > text_end || mod->text_end < text_start)
1337 			continue;
1338 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
1339 			const u8 *ptr = (const u8 *)poff + *poff;
1340 
1341 			if (text_start <= ptr && text_end > ptr)
1342 				return 1;
1343 		}
1344 	}
1345 
1346 	return 0;
1347 }
1348 #endif /* CONFIG_SMP */
1349 
1350 #ifdef CONFIG_PARAVIRT
1351 
1352 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
1353 static void __init_or_module add_nops(void *insns, unsigned int len)
1354 {
1355 	while (len > 0) {
1356 		unsigned int noplen = len;
1357 		if (noplen > ASM_NOP_MAX)
1358 			noplen = ASM_NOP_MAX;
1359 		memcpy(insns, x86_nops[noplen], noplen);
1360 		insns += noplen;
1361 		len -= noplen;
1362 	}
1363 }
1364 
1365 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
1366 				     struct paravirt_patch_site *end)
1367 {
1368 	struct paravirt_patch_site *p;
1369 	char insn_buff[MAX_PATCH_LEN];
1370 
1371 	for (p = start; p < end; p++) {
1372 		unsigned int used;
1373 
1374 		BUG_ON(p->len > MAX_PATCH_LEN);
1375 		/* prep the buffer with the original instructions */
1376 		memcpy(insn_buff, p->instr, p->len);
1377 		used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
1378 
1379 		BUG_ON(used > p->len);
1380 
1381 		/* Pad the rest with nops */
1382 		add_nops(insn_buff + used, p->len - used);
1383 		text_poke_early(p->instr, insn_buff, p->len);
1384 	}
1385 }
1386 extern struct paravirt_patch_site __start_parainstructions[],
1387 	__stop_parainstructions[];
1388 #endif	/* CONFIG_PARAVIRT */
1389 
1390 /*
1391  * Self-test for the INT3 based CALL emulation code.
1392  *
1393  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
1394  * properly and that there is a stack gap between the INT3 frame and the
1395  * previous context. Without this gap doing a virtual PUSH on the interrupted
1396  * stack would corrupt the INT3 IRET frame.
1397  *
1398  * See entry_{32,64}.S for more details.
1399  */
1400 
1401 /*
1402  * We define the int3_magic() function in assembly to control the calling
1403  * convention such that we can 'call' it from assembly.
1404  */
1405 
1406 extern void int3_magic(unsigned int *ptr); /* defined in asm */
1407 
1408 asm (
1409 "	.pushsection	.init.text, \"ax\", @progbits\n"
1410 "	.type		int3_magic, @function\n"
1411 "int3_magic:\n"
1412 	ANNOTATE_NOENDBR
1413 "	movl	$1, (%" _ASM_ARG1 ")\n"
1414 	ASM_RET
1415 "	.size		int3_magic, .-int3_magic\n"
1416 "	.popsection\n"
1417 );
1418 
1419 extern void int3_selftest_ip(void); /* defined in asm below */
1420 
1421 static int __init
1422 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
1423 {
1424 	unsigned long selftest = (unsigned long)&int3_selftest_ip;
1425 	struct die_args *args = data;
1426 	struct pt_regs *regs = args->regs;
1427 
1428 	OPTIMIZER_HIDE_VAR(selftest);
1429 
1430 	if (!regs || user_mode(regs))
1431 		return NOTIFY_DONE;
1432 
1433 	if (val != DIE_INT3)
1434 		return NOTIFY_DONE;
1435 
1436 	if (regs->ip - INT3_INSN_SIZE != selftest)
1437 		return NOTIFY_DONE;
1438 
1439 	int3_emulate_call(regs, (unsigned long)&int3_magic);
1440 	return NOTIFY_STOP;
1441 }
1442 
1443 /* Must be noinline to ensure uniqueness of int3_selftest_ip. */
1444 static noinline void __init int3_selftest(void)
1445 {
1446 	static __initdata struct notifier_block int3_exception_nb = {
1447 		.notifier_call	= int3_exception_notify,
1448 		.priority	= INT_MAX-1, /* last */
1449 	};
1450 	unsigned int val = 0;
1451 
1452 	BUG_ON(register_die_notifier(&int3_exception_nb));
1453 
1454 	/*
1455 	 * Basically: int3_magic(&val); but really complicated :-)
1456 	 *
1457 	 * INT3 padded with NOP to CALL_INSN_SIZE. The int3_exception_nb
1458 	 * notifier above will emulate CALL for us.
1459 	 */
1460 	asm volatile ("int3_selftest_ip:\n\t"
1461 		      ANNOTATE_NOENDBR
1462 		      "    int3; nop; nop; nop; nop\n\t"
1463 		      : ASM_CALL_CONSTRAINT
1464 		      : __ASM_SEL_RAW(a, D) (&val)
1465 		      : "memory");
1466 
1467 	BUG_ON(val != 1);
1468 
1469 	unregister_die_notifier(&int3_exception_nb);
1470 }
1471 
1472 static __initdata int __alt_reloc_selftest_addr;
1473 
1474 __visible noinline void __init __alt_reloc_selftest(void *arg)
1475 {
1476 	WARN_ON(arg != &__alt_reloc_selftest_addr);
1477 }
1478 
1479 static noinline void __init alt_reloc_selftest(void)
1480 {
1481 	/*
1482 	 * Tests apply_relocation().
1483 	 *
1484 	 * This has a relative immediate (CALL) in a place other than the first
1485 	 * instruction and additionally on x86_64 we get a RIP-relative LEA:
1486 	 *
1487 	 *   lea    0x0(%rip),%rdi  # 5d0: R_X86_64_PC32    .init.data+0x5566c
1488 	 *   call   +0              # 5d5: R_X86_64_PLT32   __alt_reloc_selftest-0x4
1489 	 *
1490 	 * Getting this wrong will either crash and burn or tickle the WARN
1491 	 * above.
1492 	 */
1493 	asm_inline volatile (
1494 		ALTERNATIVE("", "lea %[mem], %%" _ASM_ARG1 "; call __alt_reloc_selftest;", X86_FEATURE_ALWAYS)
1495 		: /* output */
1496 		: [mem] "m" (__alt_reloc_selftest_addr)
1497 		: _ASM_ARG1
1498 	);
1499 }
1500 
1501 void __init alternative_instructions(void)
1502 {
1503 	int3_selftest();
1504 
1505 	/*
1506 	 * The patching is not fully atomic, so try to avoid local
1507 	 * interruptions that might execute the to be patched code.
1508 	 * Other CPUs are not running.
1509 	 */
1510 	stop_nmi();
1511 
1512 	/*
1513 	 * Don't stop machine check exceptions while patching.
1514 	 * MCEs only happen when something got corrupted and in this
1515 	 * case we must do something about the corruption.
1516 	 * Ignoring it is worse than an unlikely patching race.
1517 	 * Also machine checks tend to be broadcast and if one CPU
1518 	 * goes into machine check the others follow quickly, so we don't
1519 	 * expect a machine check to cause undue problems during to code
1520 	 * patching.
1521 	 */
1522 
1523 	/*
1524 	 * Paravirt patching and alternative patching can be combined to
1525 	 * replace a function call with a short direct code sequence (e.g.
1526 	 * by setting a constant return value instead of doing that in an
1527 	 * external function).
1528 	 * In order to make this work the following sequence is required:
1529 	 * 1. set (artificial) features depending on used paravirt
1530 	 *    functions which can later influence alternative patching
1531 	 * 2. apply paravirt patching (generally replacing an indirect
1532 	 *    function call with a direct one)
1533 	 * 3. apply alternative patching (e.g. replacing a direct function
1534 	 *    call with a custom code sequence)
1535 	 * Doing paravirt patching after alternative patching would clobber
1536 	 * the optimization of the custom code with a function call again.
1537 	 */
1538 	paravirt_set_cap();
1539 
1540 	/*
1541 	 * First patch paravirt functions, such that we overwrite the indirect
1542 	 * call with the direct call.
1543 	 */
1544 	apply_paravirt(__parainstructions, __parainstructions_end);
1545 
1546 	__apply_fineibt(__retpoline_sites, __retpoline_sites_end,
1547 			__cfi_sites, __cfi_sites_end, true);
1548 
1549 	/*
1550 	 * Rewrite the retpolines, must be done before alternatives since
1551 	 * those can rewrite the retpoline thunks.
1552 	 */
1553 	apply_retpolines(__retpoline_sites, __retpoline_sites_end);
1554 	apply_returns(__return_sites, __return_sites_end);
1555 
1556 	/*
1557 	 * Then patch alternatives, such that those paravirt calls that are in
1558 	 * alternatives can be overwritten by their immediate fragments.
1559 	 */
1560 	apply_alternatives(__alt_instructions, __alt_instructions_end);
1561 
1562 	/*
1563 	 * Now all calls are established. Apply the call thunks if
1564 	 * required.
1565 	 */
1566 	callthunks_patch_builtin_calls();
1567 
1568 	apply_ibt_endbr(__ibt_endbr_seal, __ibt_endbr_seal_end);
1569 
1570 #ifdef CONFIG_SMP
1571 	/* Patch to UP if other cpus not imminent. */
1572 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
1573 		uniproc_patched = true;
1574 		alternatives_smp_module_add(NULL, "core kernel",
1575 					    __smp_locks, __smp_locks_end,
1576 					    _text, _etext);
1577 	}
1578 
1579 	if (!uniproc_patched || num_possible_cpus() == 1) {
1580 		free_init_pages("SMP alternatives",
1581 				(unsigned long)__smp_locks,
1582 				(unsigned long)__smp_locks_end);
1583 	}
1584 #endif
1585 
1586 	restart_nmi();
1587 	alternatives_patched = 1;
1588 
1589 	alt_reloc_selftest();
1590 }
1591 
1592 /**
1593  * text_poke_early - Update instructions on a live kernel at boot time
1594  * @addr: address to modify
1595  * @opcode: source of the copy
1596  * @len: length to copy
1597  *
1598  * When you use this code to patch more than one byte of an instruction
1599  * you need to make sure that other CPUs cannot execute this code in parallel.
1600  * Also no thread must be currently preempted in the middle of these
1601  * instructions. And on the local CPU you need to be protected against NMI or
1602  * MCE handlers seeing an inconsistent instruction while you patch.
1603  */
1604 void __init_or_module text_poke_early(void *addr, const void *opcode,
1605 				      size_t len)
1606 {
1607 	unsigned long flags;
1608 
1609 	if (boot_cpu_has(X86_FEATURE_NX) &&
1610 	    is_module_text_address((unsigned long)addr)) {
1611 		/*
1612 		 * Modules text is marked initially as non-executable, so the
1613 		 * code cannot be running and speculative code-fetches are
1614 		 * prevented. Just change the code.
1615 		 */
1616 		memcpy(addr, opcode, len);
1617 	} else {
1618 		local_irq_save(flags);
1619 		memcpy(addr, opcode, len);
1620 		local_irq_restore(flags);
1621 		sync_core();
1622 
1623 		/*
1624 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
1625 		 * that causes hangs on some VIA CPUs.
1626 		 */
1627 	}
1628 }
1629 
1630 typedef struct {
1631 	struct mm_struct *mm;
1632 } temp_mm_state_t;
1633 
1634 /*
1635  * Using a temporary mm allows to set temporary mappings that are not accessible
1636  * by other CPUs. Such mappings are needed to perform sensitive memory writes
1637  * that override the kernel memory protections (e.g., W^X), without exposing the
1638  * temporary page-table mappings that are required for these write operations to
1639  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
1640  * mapping is torn down.
1641  *
1642  * Context: The temporary mm needs to be used exclusively by a single core. To
1643  *          harden security IRQs must be disabled while the temporary mm is
1644  *          loaded, thereby preventing interrupt handler bugs from overriding
1645  *          the kernel memory protection.
1646  */
1647 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
1648 {
1649 	temp_mm_state_t temp_state;
1650 
1651 	lockdep_assert_irqs_disabled();
1652 
1653 	/*
1654 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
1655 	 * with a stale address space WITHOUT being in lazy mode after
1656 	 * restoring the previous mm.
1657 	 */
1658 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
1659 		leave_mm(smp_processor_id());
1660 
1661 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1662 	switch_mm_irqs_off(NULL, mm, current);
1663 
1664 	/*
1665 	 * If breakpoints are enabled, disable them while the temporary mm is
1666 	 * used. Userspace might set up watchpoints on addresses that are used
1667 	 * in the temporary mm, which would lead to wrong signals being sent or
1668 	 * crashes.
1669 	 *
1670 	 * Note that breakpoints are not disabled selectively, which also causes
1671 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
1672 	 * undesirable, but still seems reasonable as the code that runs in the
1673 	 * temporary mm should be short.
1674 	 */
1675 	if (hw_breakpoint_active())
1676 		hw_breakpoint_disable();
1677 
1678 	return temp_state;
1679 }
1680 
1681 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
1682 {
1683 	lockdep_assert_irqs_disabled();
1684 	switch_mm_irqs_off(NULL, prev_state.mm, current);
1685 
1686 	/*
1687 	 * Restore the breakpoints if they were disabled before the temporary mm
1688 	 * was loaded.
1689 	 */
1690 	if (hw_breakpoint_active())
1691 		hw_breakpoint_restore();
1692 }
1693 
1694 __ro_after_init struct mm_struct *poking_mm;
1695 __ro_after_init unsigned long poking_addr;
1696 
1697 static void text_poke_memcpy(void *dst, const void *src, size_t len)
1698 {
1699 	memcpy(dst, src, len);
1700 }
1701 
1702 static void text_poke_memset(void *dst, const void *src, size_t len)
1703 {
1704 	int c = *(const int *)src;
1705 
1706 	memset(dst, c, len);
1707 }
1708 
1709 typedef void text_poke_f(void *dst, const void *src, size_t len);
1710 
1711 static void *__text_poke(text_poke_f func, void *addr, const void *src, size_t len)
1712 {
1713 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
1714 	struct page *pages[2] = {NULL};
1715 	temp_mm_state_t prev;
1716 	unsigned long flags;
1717 	pte_t pte, *ptep;
1718 	spinlock_t *ptl;
1719 	pgprot_t pgprot;
1720 
1721 	/*
1722 	 * While boot memory allocator is running we cannot use struct pages as
1723 	 * they are not yet initialized. There is no way to recover.
1724 	 */
1725 	BUG_ON(!after_bootmem);
1726 
1727 	if (!core_kernel_text((unsigned long)addr)) {
1728 		pages[0] = vmalloc_to_page(addr);
1729 		if (cross_page_boundary)
1730 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
1731 	} else {
1732 		pages[0] = virt_to_page(addr);
1733 		WARN_ON(!PageReserved(pages[0]));
1734 		if (cross_page_boundary)
1735 			pages[1] = virt_to_page(addr + PAGE_SIZE);
1736 	}
1737 	/*
1738 	 * If something went wrong, crash and burn since recovery paths are not
1739 	 * implemented.
1740 	 */
1741 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
1742 
1743 	/*
1744 	 * Map the page without the global bit, as TLB flushing is done with
1745 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
1746 	 */
1747 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
1748 
1749 	/*
1750 	 * The lock is not really needed, but this allows to avoid open-coding.
1751 	 */
1752 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
1753 
1754 	/*
1755 	 * This must not fail; preallocated in poking_init().
1756 	 */
1757 	VM_BUG_ON(!ptep);
1758 
1759 	local_irq_save(flags);
1760 
1761 	pte = mk_pte(pages[0], pgprot);
1762 	set_pte_at(poking_mm, poking_addr, ptep, pte);
1763 
1764 	if (cross_page_boundary) {
1765 		pte = mk_pte(pages[1], pgprot);
1766 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
1767 	}
1768 
1769 	/*
1770 	 * Loading the temporary mm behaves as a compiler barrier, which
1771 	 * guarantees that the PTE will be set at the time memcpy() is done.
1772 	 */
1773 	prev = use_temporary_mm(poking_mm);
1774 
1775 	kasan_disable_current();
1776 	func((u8 *)poking_addr + offset_in_page(addr), src, len);
1777 	kasan_enable_current();
1778 
1779 	/*
1780 	 * Ensure that the PTE is only cleared after the instructions of memcpy
1781 	 * were issued by using a compiler barrier.
1782 	 */
1783 	barrier();
1784 
1785 	pte_clear(poking_mm, poking_addr, ptep);
1786 	if (cross_page_boundary)
1787 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
1788 
1789 	/*
1790 	 * Loading the previous page-table hierarchy requires a serializing
1791 	 * instruction that already allows the core to see the updated version.
1792 	 * Xen-PV is assumed to serialize execution in a similar manner.
1793 	 */
1794 	unuse_temporary_mm(prev);
1795 
1796 	/*
1797 	 * Flushing the TLB might involve IPIs, which would require enabled
1798 	 * IRQs, but not if the mm is not used, as it is in this point.
1799 	 */
1800 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
1801 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
1802 			   PAGE_SHIFT, false);
1803 
1804 	if (func == text_poke_memcpy) {
1805 		/*
1806 		 * If the text does not match what we just wrote then something is
1807 		 * fundamentally screwy; there's nothing we can really do about that.
1808 		 */
1809 		BUG_ON(memcmp(addr, src, len));
1810 	}
1811 
1812 	local_irq_restore(flags);
1813 	pte_unmap_unlock(ptep, ptl);
1814 	return addr;
1815 }
1816 
1817 /**
1818  * text_poke - Update instructions on a live kernel
1819  * @addr: address to modify
1820  * @opcode: source of the copy
1821  * @len: length to copy
1822  *
1823  * Only atomic text poke/set should be allowed when not doing early patching.
1824  * It means the size must be writable atomically and the address must be aligned
1825  * in a way that permits an atomic write. It also makes sure we fit on a single
1826  * page.
1827  *
1828  * Note that the caller must ensure that if the modified code is part of a
1829  * module, the module would not be removed during poking. This can be achieved
1830  * by registering a module notifier, and ordering module removal and patching
1831  * trough a mutex.
1832  */
1833 void *text_poke(void *addr, const void *opcode, size_t len)
1834 {
1835 	lockdep_assert_held(&text_mutex);
1836 
1837 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1838 }
1839 
1840 /**
1841  * text_poke_kgdb - Update instructions on a live kernel by kgdb
1842  * @addr: address to modify
1843  * @opcode: source of the copy
1844  * @len: length to copy
1845  *
1846  * Only atomic text poke/set should be allowed when not doing early patching.
1847  * It means the size must be writable atomically and the address must be aligned
1848  * in a way that permits an atomic write. It also makes sure we fit on a single
1849  * page.
1850  *
1851  * Context: should only be used by kgdb, which ensures no other core is running,
1852  *	    despite the fact it does not hold the text_mutex.
1853  */
1854 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
1855 {
1856 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1857 }
1858 
1859 void *text_poke_copy_locked(void *addr, const void *opcode, size_t len,
1860 			    bool core_ok)
1861 {
1862 	unsigned long start = (unsigned long)addr;
1863 	size_t patched = 0;
1864 
1865 	if (WARN_ON_ONCE(!core_ok && core_kernel_text(start)))
1866 		return NULL;
1867 
1868 	while (patched < len) {
1869 		unsigned long ptr = start + patched;
1870 		size_t s;
1871 
1872 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1873 
1874 		__text_poke(text_poke_memcpy, (void *)ptr, opcode + patched, s);
1875 		patched += s;
1876 	}
1877 	return addr;
1878 }
1879 
1880 /**
1881  * text_poke_copy - Copy instructions into (an unused part of) RX memory
1882  * @addr: address to modify
1883  * @opcode: source of the copy
1884  * @len: length to copy, could be more than 2x PAGE_SIZE
1885  *
1886  * Not safe against concurrent execution; useful for JITs to dump
1887  * new code blocks into unused regions of RX memory. Can be used in
1888  * conjunction with synchronize_rcu_tasks() to wait for existing
1889  * execution to quiesce after having made sure no existing functions
1890  * pointers are live.
1891  */
1892 void *text_poke_copy(void *addr, const void *opcode, size_t len)
1893 {
1894 	mutex_lock(&text_mutex);
1895 	addr = text_poke_copy_locked(addr, opcode, len, false);
1896 	mutex_unlock(&text_mutex);
1897 	return addr;
1898 }
1899 
1900 /**
1901  * text_poke_set - memset into (an unused part of) RX memory
1902  * @addr: address to modify
1903  * @c: the byte to fill the area with
1904  * @len: length to copy, could be more than 2x PAGE_SIZE
1905  *
1906  * This is useful to overwrite unused regions of RX memory with illegal
1907  * instructions.
1908  */
1909 void *text_poke_set(void *addr, int c, size_t len)
1910 {
1911 	unsigned long start = (unsigned long)addr;
1912 	size_t patched = 0;
1913 
1914 	if (WARN_ON_ONCE(core_kernel_text(start)))
1915 		return NULL;
1916 
1917 	mutex_lock(&text_mutex);
1918 	while (patched < len) {
1919 		unsigned long ptr = start + patched;
1920 		size_t s;
1921 
1922 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1923 
1924 		__text_poke(text_poke_memset, (void *)ptr, (void *)&c, s);
1925 		patched += s;
1926 	}
1927 	mutex_unlock(&text_mutex);
1928 	return addr;
1929 }
1930 
1931 static void do_sync_core(void *info)
1932 {
1933 	sync_core();
1934 }
1935 
1936 void text_poke_sync(void)
1937 {
1938 	on_each_cpu(do_sync_core, NULL, 1);
1939 }
1940 
1941 /*
1942  * NOTE: crazy scheme to allow patching Jcc.d32 but not increase the size of
1943  * this thing. When len == 6 everything is prefixed with 0x0f and we map
1944  * opcode to Jcc.d8, using len to distinguish.
1945  */
1946 struct text_poke_loc {
1947 	/* addr := _stext + rel_addr */
1948 	s32 rel_addr;
1949 	s32 disp;
1950 	u8 len;
1951 	u8 opcode;
1952 	const u8 text[POKE_MAX_OPCODE_SIZE];
1953 	/* see text_poke_bp_batch() */
1954 	u8 old;
1955 };
1956 
1957 struct bp_patching_desc {
1958 	struct text_poke_loc *vec;
1959 	int nr_entries;
1960 	atomic_t refs;
1961 };
1962 
1963 static struct bp_patching_desc bp_desc;
1964 
1965 static __always_inline
1966 struct bp_patching_desc *try_get_desc(void)
1967 {
1968 	struct bp_patching_desc *desc = &bp_desc;
1969 
1970 	if (!raw_atomic_inc_not_zero(&desc->refs))
1971 		return NULL;
1972 
1973 	return desc;
1974 }
1975 
1976 static __always_inline void put_desc(void)
1977 {
1978 	struct bp_patching_desc *desc = &bp_desc;
1979 
1980 	smp_mb__before_atomic();
1981 	raw_atomic_dec(&desc->refs);
1982 }
1983 
1984 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
1985 {
1986 	return _stext + tp->rel_addr;
1987 }
1988 
1989 static __always_inline int patch_cmp(const void *key, const void *elt)
1990 {
1991 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
1992 
1993 	if (key < text_poke_addr(tp))
1994 		return -1;
1995 	if (key > text_poke_addr(tp))
1996 		return 1;
1997 	return 0;
1998 }
1999 
2000 noinstr int poke_int3_handler(struct pt_regs *regs)
2001 {
2002 	struct bp_patching_desc *desc;
2003 	struct text_poke_loc *tp;
2004 	int ret = 0;
2005 	void *ip;
2006 
2007 	if (user_mode(regs))
2008 		return 0;
2009 
2010 	/*
2011 	 * Having observed our INT3 instruction, we now must observe
2012 	 * bp_desc with non-zero refcount:
2013 	 *
2014 	 *	bp_desc.refs = 1		INT3
2015 	 *	WMB				RMB
2016 	 *	write INT3			if (bp_desc.refs != 0)
2017 	 */
2018 	smp_rmb();
2019 
2020 	desc = try_get_desc();
2021 	if (!desc)
2022 		return 0;
2023 
2024 	/*
2025 	 * Discount the INT3. See text_poke_bp_batch().
2026 	 */
2027 	ip = (void *) regs->ip - INT3_INSN_SIZE;
2028 
2029 	/*
2030 	 * Skip the binary search if there is a single member in the vector.
2031 	 */
2032 	if (unlikely(desc->nr_entries > 1)) {
2033 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
2034 				      sizeof(struct text_poke_loc),
2035 				      patch_cmp);
2036 		if (!tp)
2037 			goto out_put;
2038 	} else {
2039 		tp = desc->vec;
2040 		if (text_poke_addr(tp) != ip)
2041 			goto out_put;
2042 	}
2043 
2044 	ip += tp->len;
2045 
2046 	switch (tp->opcode) {
2047 	case INT3_INSN_OPCODE:
2048 		/*
2049 		 * Someone poked an explicit INT3, they'll want to handle it,
2050 		 * do not consume.
2051 		 */
2052 		goto out_put;
2053 
2054 	case RET_INSN_OPCODE:
2055 		int3_emulate_ret(regs);
2056 		break;
2057 
2058 	case CALL_INSN_OPCODE:
2059 		int3_emulate_call(regs, (long)ip + tp->disp);
2060 		break;
2061 
2062 	case JMP32_INSN_OPCODE:
2063 	case JMP8_INSN_OPCODE:
2064 		int3_emulate_jmp(regs, (long)ip + tp->disp);
2065 		break;
2066 
2067 	case 0x70 ... 0x7f: /* Jcc */
2068 		int3_emulate_jcc(regs, tp->opcode & 0xf, (long)ip, tp->disp);
2069 		break;
2070 
2071 	default:
2072 		BUG();
2073 	}
2074 
2075 	ret = 1;
2076 
2077 out_put:
2078 	put_desc();
2079 	return ret;
2080 }
2081 
2082 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
2083 static struct text_poke_loc tp_vec[TP_VEC_MAX];
2084 static int tp_vec_nr;
2085 
2086 /**
2087  * text_poke_bp_batch() -- update instructions on live kernel on SMP
2088  * @tp:			vector of instructions to patch
2089  * @nr_entries:		number of entries in the vector
2090  *
2091  * Modify multi-byte instruction by using int3 breakpoint on SMP.
2092  * We completely avoid stop_machine() here, and achieve the
2093  * synchronization using int3 breakpoint.
2094  *
2095  * The way it is done:
2096  *	- For each entry in the vector:
2097  *		- add a int3 trap to the address that will be patched
2098  *	- sync cores
2099  *	- For each entry in the vector:
2100  *		- update all but the first byte of the patched range
2101  *	- sync cores
2102  *	- For each entry in the vector:
2103  *		- replace the first byte (int3) by the first byte of
2104  *		  replacing opcode
2105  *	- sync cores
2106  */
2107 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
2108 {
2109 	unsigned char int3 = INT3_INSN_OPCODE;
2110 	unsigned int i;
2111 	int do_sync;
2112 
2113 	lockdep_assert_held(&text_mutex);
2114 
2115 	bp_desc.vec = tp;
2116 	bp_desc.nr_entries = nr_entries;
2117 
2118 	/*
2119 	 * Corresponds to the implicit memory barrier in try_get_desc() to
2120 	 * ensure reading a non-zero refcount provides up to date bp_desc data.
2121 	 */
2122 	atomic_set_release(&bp_desc.refs, 1);
2123 
2124 	/*
2125 	 * Function tracing can enable thousands of places that need to be
2126 	 * updated. This can take quite some time, and with full kernel debugging
2127 	 * enabled, this could cause the softlockup watchdog to trigger.
2128 	 * This function gets called every 256 entries added to be patched.
2129 	 * Call cond_resched() here to make sure that other tasks can get scheduled
2130 	 * while processing all the functions being patched.
2131 	 */
2132 	cond_resched();
2133 
2134 	/*
2135 	 * Corresponding read barrier in int3 notifier for making sure the
2136 	 * nr_entries and handler are correctly ordered wrt. patching.
2137 	 */
2138 	smp_wmb();
2139 
2140 	/*
2141 	 * First step: add a int3 trap to the address that will be patched.
2142 	 */
2143 	for (i = 0; i < nr_entries; i++) {
2144 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
2145 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
2146 	}
2147 
2148 	text_poke_sync();
2149 
2150 	/*
2151 	 * Second step: update all but the first byte of the patched range.
2152 	 */
2153 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2154 		u8 old[POKE_MAX_OPCODE_SIZE+1] = { tp[i].old, };
2155 		u8 _new[POKE_MAX_OPCODE_SIZE+1];
2156 		const u8 *new = tp[i].text;
2157 		int len = tp[i].len;
2158 
2159 		if (len - INT3_INSN_SIZE > 0) {
2160 			memcpy(old + INT3_INSN_SIZE,
2161 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2162 			       len - INT3_INSN_SIZE);
2163 
2164 			if (len == 6) {
2165 				_new[0] = 0x0f;
2166 				memcpy(_new + 1, new, 5);
2167 				new = _new;
2168 			}
2169 
2170 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2171 				  new + INT3_INSN_SIZE,
2172 				  len - INT3_INSN_SIZE);
2173 
2174 			do_sync++;
2175 		}
2176 
2177 		/*
2178 		 * Emit a perf event to record the text poke, primarily to
2179 		 * support Intel PT decoding which must walk the executable code
2180 		 * to reconstruct the trace. The flow up to here is:
2181 		 *   - write INT3 byte
2182 		 *   - IPI-SYNC
2183 		 *   - write instruction tail
2184 		 * At this point the actual control flow will be through the
2185 		 * INT3 and handler and not hit the old or new instruction.
2186 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
2187 		 * can still be decoded. Subsequently:
2188 		 *   - emit RECORD_TEXT_POKE with the new instruction
2189 		 *   - IPI-SYNC
2190 		 *   - write first byte
2191 		 *   - IPI-SYNC
2192 		 * So before the text poke event timestamp, the decoder will see
2193 		 * either the old instruction flow or FUP/TIP of INT3. After the
2194 		 * text poke event timestamp, the decoder will see either the
2195 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
2196 		 * use the timestamp as the point at which to modify the
2197 		 * executable code.
2198 		 * The old instruction is recorded so that the event can be
2199 		 * processed forwards or backwards.
2200 		 */
2201 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len, new, len);
2202 	}
2203 
2204 	if (do_sync) {
2205 		/*
2206 		 * According to Intel, this core syncing is very likely
2207 		 * not necessary and we'd be safe even without it. But
2208 		 * better safe than sorry (plus there's not only Intel).
2209 		 */
2210 		text_poke_sync();
2211 	}
2212 
2213 	/*
2214 	 * Third step: replace the first byte (int3) by the first byte of
2215 	 * replacing opcode.
2216 	 */
2217 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2218 		u8 byte = tp[i].text[0];
2219 
2220 		if (tp[i].len == 6)
2221 			byte = 0x0f;
2222 
2223 		if (byte == INT3_INSN_OPCODE)
2224 			continue;
2225 
2226 		text_poke(text_poke_addr(&tp[i]), &byte, INT3_INSN_SIZE);
2227 		do_sync++;
2228 	}
2229 
2230 	if (do_sync)
2231 		text_poke_sync();
2232 
2233 	/*
2234 	 * Remove and wait for refs to be zero.
2235 	 */
2236 	if (!atomic_dec_and_test(&bp_desc.refs))
2237 		atomic_cond_read_acquire(&bp_desc.refs, !VAL);
2238 }
2239 
2240 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
2241 			       const void *opcode, size_t len, const void *emulate)
2242 {
2243 	struct insn insn;
2244 	int ret, i = 0;
2245 
2246 	if (len == 6)
2247 		i = 1;
2248 	memcpy((void *)tp->text, opcode+i, len-i);
2249 	if (!emulate)
2250 		emulate = opcode;
2251 
2252 	ret = insn_decode_kernel(&insn, emulate);
2253 	BUG_ON(ret < 0);
2254 
2255 	tp->rel_addr = addr - (void *)_stext;
2256 	tp->len = len;
2257 	tp->opcode = insn.opcode.bytes[0];
2258 
2259 	if (is_jcc32(&insn)) {
2260 		/*
2261 		 * Map Jcc.d32 onto Jcc.d8 and use len to distinguish.
2262 		 */
2263 		tp->opcode = insn.opcode.bytes[1] - 0x10;
2264 	}
2265 
2266 	switch (tp->opcode) {
2267 	case RET_INSN_OPCODE:
2268 	case JMP32_INSN_OPCODE:
2269 	case JMP8_INSN_OPCODE:
2270 		/*
2271 		 * Control flow instructions without implied execution of the
2272 		 * next instruction can be padded with INT3.
2273 		 */
2274 		for (i = insn.length; i < len; i++)
2275 			BUG_ON(tp->text[i] != INT3_INSN_OPCODE);
2276 		break;
2277 
2278 	default:
2279 		BUG_ON(len != insn.length);
2280 	}
2281 
2282 	switch (tp->opcode) {
2283 	case INT3_INSN_OPCODE:
2284 	case RET_INSN_OPCODE:
2285 		break;
2286 
2287 	case CALL_INSN_OPCODE:
2288 	case JMP32_INSN_OPCODE:
2289 	case JMP8_INSN_OPCODE:
2290 	case 0x70 ... 0x7f: /* Jcc */
2291 		tp->disp = insn.immediate.value;
2292 		break;
2293 
2294 	default: /* assume NOP */
2295 		switch (len) {
2296 		case 2: /* NOP2 -- emulate as JMP8+0 */
2297 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2298 			tp->opcode = JMP8_INSN_OPCODE;
2299 			tp->disp = 0;
2300 			break;
2301 
2302 		case 5: /* NOP5 -- emulate as JMP32+0 */
2303 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2304 			tp->opcode = JMP32_INSN_OPCODE;
2305 			tp->disp = 0;
2306 			break;
2307 
2308 		default: /* unknown instruction */
2309 			BUG();
2310 		}
2311 		break;
2312 	}
2313 }
2314 
2315 /*
2316  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
2317  * early if needed.
2318  */
2319 static bool tp_order_fail(void *addr)
2320 {
2321 	struct text_poke_loc *tp;
2322 
2323 	if (!tp_vec_nr)
2324 		return false;
2325 
2326 	if (!addr) /* force */
2327 		return true;
2328 
2329 	tp = &tp_vec[tp_vec_nr - 1];
2330 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
2331 		return true;
2332 
2333 	return false;
2334 }
2335 
2336 static void text_poke_flush(void *addr)
2337 {
2338 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
2339 		text_poke_bp_batch(tp_vec, tp_vec_nr);
2340 		tp_vec_nr = 0;
2341 	}
2342 }
2343 
2344 void text_poke_finish(void)
2345 {
2346 	text_poke_flush(NULL);
2347 }
2348 
2349 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
2350 {
2351 	struct text_poke_loc *tp;
2352 
2353 	text_poke_flush(addr);
2354 
2355 	tp = &tp_vec[tp_vec_nr++];
2356 	text_poke_loc_init(tp, addr, opcode, len, emulate);
2357 }
2358 
2359 /**
2360  * text_poke_bp() -- update instructions on live kernel on SMP
2361  * @addr:	address to patch
2362  * @opcode:	opcode of new instruction
2363  * @len:	length to copy
2364  * @emulate:	instruction to be emulated
2365  *
2366  * Update a single instruction with the vector in the stack, avoiding
2367  * dynamically allocated memory. This function should be used when it is
2368  * not possible to allocate memory.
2369  */
2370 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
2371 {
2372 	struct text_poke_loc tp;
2373 
2374 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
2375 	text_poke_bp_batch(&tp, 1);
2376 }
2377