xref: /linux/arch/x86/kernel/alternative.c (revision b17ef04bf3a4346d66404454d6a646343ddc9749)
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 	/*
408 	 * In the case CONFIG_X86_5LEVEL=y, KASAN_SHADOW_START is defined using
409 	 * cpu_feature_enabled(X86_FEATURE_LA57) and is therefore patched here.
410 	 * During the process, KASAN becomes confused seeing partial LA57
411 	 * conversion and triggers a false-positive out-of-bound report.
412 	 *
413 	 * Disable KASAN until the patching is complete.
414 	 */
415 	kasan_disable_current();
416 
417 	/*
418 	 * The scan order should be from start to end. A later scanned
419 	 * alternative code can overwrite previously scanned alternative code.
420 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
421 	 * patch code.
422 	 *
423 	 * So be careful if you want to change the scan order to any other
424 	 * order.
425 	 */
426 	for (a = start; a < end; a++) {
427 		int insn_buff_sz = 0;
428 
429 		instr = (u8 *)&a->instr_offset + a->instr_offset;
430 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
431 		BUG_ON(a->instrlen > sizeof(insn_buff));
432 		BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
433 
434 		/*
435 		 * Patch if either:
436 		 * - feature is present
437 		 * - feature not present but ALT_FLAG_NOT is set to mean,
438 		 *   patch if feature is *NOT* present.
439 		 */
440 		if (!boot_cpu_has(a->cpuid) == !(a->flags & ALT_FLAG_NOT)) {
441 			optimize_nops(instr, a->instrlen);
442 			continue;
443 		}
444 
445 		DPRINTK(ALT, "feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)",
446 			(a->flags & ALT_FLAG_NOT) ? "!" : "",
447 			a->cpuid >> 5,
448 			a->cpuid & 0x1f,
449 			instr, instr, a->instrlen,
450 			replacement, a->replacementlen);
451 
452 		memcpy(insn_buff, replacement, a->replacementlen);
453 		insn_buff_sz = a->replacementlen;
454 
455 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
456 			insn_buff[insn_buff_sz] = 0x90;
457 
458 		apply_relocation(insn_buff, a->instrlen, instr, replacement, a->replacementlen);
459 
460 		DUMP_BYTES(ALT, instr, a->instrlen, "%px:   old_insn: ", instr);
461 		DUMP_BYTES(ALT, replacement, a->replacementlen, "%px:   rpl_insn: ", replacement);
462 		DUMP_BYTES(ALT, insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
463 
464 		text_poke_early(instr, insn_buff, insn_buff_sz);
465 	}
466 
467 	kasan_enable_current();
468 }
469 
470 static inline bool is_jcc32(struct insn *insn)
471 {
472 	/* Jcc.d32 second opcode byte is in the range: 0x80-0x8f */
473 	return insn->opcode.bytes[0] == 0x0f && (insn->opcode.bytes[1] & 0xf0) == 0x80;
474 }
475 
476 #if defined(CONFIG_RETPOLINE) && defined(CONFIG_OBJTOOL)
477 
478 /*
479  * CALL/JMP *%\reg
480  */
481 static int emit_indirect(int op, int reg, u8 *bytes)
482 {
483 	int i = 0;
484 	u8 modrm;
485 
486 	switch (op) {
487 	case CALL_INSN_OPCODE:
488 		modrm = 0x10; /* Reg = 2; CALL r/m */
489 		break;
490 
491 	case JMP32_INSN_OPCODE:
492 		modrm = 0x20; /* Reg = 4; JMP r/m */
493 		break;
494 
495 	default:
496 		WARN_ON_ONCE(1);
497 		return -1;
498 	}
499 
500 	if (reg >= 8) {
501 		bytes[i++] = 0x41; /* REX.B prefix */
502 		reg -= 8;
503 	}
504 
505 	modrm |= 0xc0; /* Mod = 3 */
506 	modrm += reg;
507 
508 	bytes[i++] = 0xff; /* opcode */
509 	bytes[i++] = modrm;
510 
511 	return i;
512 }
513 
514 static int emit_call_track_retpoline(void *addr, struct insn *insn, int reg, u8 *bytes)
515 {
516 	u8 op = insn->opcode.bytes[0];
517 	int i = 0;
518 
519 	/*
520 	 * Clang does 'weird' Jcc __x86_indirect_thunk_r11 conditional
521 	 * tail-calls. Deal with them.
522 	 */
523 	if (is_jcc32(insn)) {
524 		bytes[i++] = op;
525 		op = insn->opcode.bytes[1];
526 		goto clang_jcc;
527 	}
528 
529 	if (insn->length == 6)
530 		bytes[i++] = 0x2e; /* CS-prefix */
531 
532 	switch (op) {
533 	case CALL_INSN_OPCODE:
534 		__text_gen_insn(bytes+i, op, addr+i,
535 				__x86_indirect_call_thunk_array[reg],
536 				CALL_INSN_SIZE);
537 		i += CALL_INSN_SIZE;
538 		break;
539 
540 	case JMP32_INSN_OPCODE:
541 clang_jcc:
542 		__text_gen_insn(bytes+i, op, addr+i,
543 				__x86_indirect_jump_thunk_array[reg],
544 				JMP32_INSN_SIZE);
545 		i += JMP32_INSN_SIZE;
546 		break;
547 
548 	default:
549 		WARN(1, "%pS %px %*ph\n", addr, addr, 6, addr);
550 		return -1;
551 	}
552 
553 	WARN_ON_ONCE(i != insn->length);
554 
555 	return i;
556 }
557 
558 /*
559  * Rewrite the compiler generated retpoline thunk calls.
560  *
561  * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate
562  * indirect instructions, avoiding the extra indirection.
563  *
564  * For example, convert:
565  *
566  *   CALL __x86_indirect_thunk_\reg
567  *
568  * into:
569  *
570  *   CALL *%\reg
571  *
572  * It also tries to inline spectre_v2=retpoline,lfence when size permits.
573  */
574 static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes)
575 {
576 	retpoline_thunk_t *target;
577 	int reg, ret, i = 0;
578 	u8 op, cc;
579 
580 	target = addr + insn->length + insn->immediate.value;
581 	reg = target - __x86_indirect_thunk_array;
582 
583 	if (WARN_ON_ONCE(reg & ~0xf))
584 		return -1;
585 
586 	/* If anyone ever does: CALL/JMP *%rsp, we're in deep trouble. */
587 	BUG_ON(reg == 4);
588 
589 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE) &&
590 	    !cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
591 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
592 			return emit_call_track_retpoline(addr, insn, reg, bytes);
593 
594 		return -1;
595 	}
596 
597 	op = insn->opcode.bytes[0];
598 
599 	/*
600 	 * Convert:
601 	 *
602 	 *   Jcc.d32 __x86_indirect_thunk_\reg
603 	 *
604 	 * into:
605 	 *
606 	 *   Jncc.d8 1f
607 	 *   [ LFENCE ]
608 	 *   JMP *%\reg
609 	 *   [ NOP ]
610 	 * 1:
611 	 */
612 	if (is_jcc32(insn)) {
613 		cc = insn->opcode.bytes[1] & 0xf;
614 		cc ^= 1; /* invert condition */
615 
616 		bytes[i++] = 0x70 + cc;        /* Jcc.d8 */
617 		bytes[i++] = insn->length - 2; /* sizeof(Jcc.d8) == 2 */
618 
619 		/* Continue as if: JMP.d32 __x86_indirect_thunk_\reg */
620 		op = JMP32_INSN_OPCODE;
621 	}
622 
623 	/*
624 	 * For RETPOLINE_LFENCE: prepend the indirect CALL/JMP with an LFENCE.
625 	 */
626 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
627 		bytes[i++] = 0x0f;
628 		bytes[i++] = 0xae;
629 		bytes[i++] = 0xe8; /* LFENCE */
630 	}
631 
632 	ret = emit_indirect(op, reg, bytes + i);
633 	if (ret < 0)
634 		return ret;
635 	i += ret;
636 
637 	/*
638 	 * The compiler is supposed to EMIT an INT3 after every unconditional
639 	 * JMP instruction due to AMD BTC. However, if the compiler is too old
640 	 * or SLS isn't enabled, we still need an INT3 after indirect JMPs
641 	 * even on Intel.
642 	 */
643 	if (op == JMP32_INSN_OPCODE && i < insn->length)
644 		bytes[i++] = INT3_INSN_OPCODE;
645 
646 	for (; i < insn->length;)
647 		bytes[i++] = BYTES_NOP1;
648 
649 	return i;
650 }
651 
652 /*
653  * Generated by 'objtool --retpoline'.
654  */
655 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end)
656 {
657 	s32 *s;
658 
659 	for (s = start; s < end; s++) {
660 		void *addr = (void *)s + *s;
661 		struct insn insn;
662 		int len, ret;
663 		u8 bytes[16];
664 		u8 op1, op2;
665 
666 		ret = insn_decode_kernel(&insn, addr);
667 		if (WARN_ON_ONCE(ret < 0))
668 			continue;
669 
670 		op1 = insn.opcode.bytes[0];
671 		op2 = insn.opcode.bytes[1];
672 
673 		switch (op1) {
674 		case CALL_INSN_OPCODE:
675 		case JMP32_INSN_OPCODE:
676 			break;
677 
678 		case 0x0f: /* escape */
679 			if (op2 >= 0x80 && op2 <= 0x8f)
680 				break;
681 			fallthrough;
682 		default:
683 			WARN_ON_ONCE(1);
684 			continue;
685 		}
686 
687 		DPRINTK(RETPOLINE, "retpoline at: %pS (%px) len: %d to: %pS",
688 			addr, addr, insn.length,
689 			addr + insn.length + insn.immediate.value);
690 
691 		len = patch_retpoline(addr, &insn, bytes);
692 		if (len == insn.length) {
693 			optimize_nops(bytes, len);
694 			DUMP_BYTES(RETPOLINE, ((u8*)addr),  len, "%px: orig: ", addr);
695 			DUMP_BYTES(RETPOLINE, ((u8*)bytes), len, "%px: repl: ", addr);
696 			text_poke_early(addr, bytes, len);
697 		}
698 	}
699 }
700 
701 #ifdef CONFIG_RETHUNK
702 
703 /*
704  * Rewrite the compiler generated return thunk tail-calls.
705  *
706  * For example, convert:
707  *
708  *   JMP __x86_return_thunk
709  *
710  * into:
711  *
712  *   RET
713  */
714 static int patch_return(void *addr, struct insn *insn, u8 *bytes)
715 {
716 	int i = 0;
717 
718 	/* Patch the custom return thunks... */
719 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
720 		i = JMP32_INSN_SIZE;
721 		__text_gen_insn(bytes, JMP32_INSN_OPCODE, addr, x86_return_thunk, i);
722 	} else {
723 		/* ... or patch them out if not needed. */
724 		bytes[i++] = RET_INSN_OPCODE;
725 	}
726 
727 	for (; i < insn->length;)
728 		bytes[i++] = INT3_INSN_OPCODE;
729 	return i;
730 }
731 
732 void __init_or_module noinline apply_returns(s32 *start, s32 *end)
733 {
734 	s32 *s;
735 
736 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
737 		static_call_force_reinit();
738 
739 	for (s = start; s < end; s++) {
740 		void *dest = NULL, *addr = (void *)s + *s;
741 		struct insn insn;
742 		int len, ret;
743 		u8 bytes[16];
744 		u8 op;
745 
746 		ret = insn_decode_kernel(&insn, addr);
747 		if (WARN_ON_ONCE(ret < 0))
748 			continue;
749 
750 		op = insn.opcode.bytes[0];
751 		if (op == JMP32_INSN_OPCODE)
752 			dest = addr + insn.length + insn.immediate.value;
753 
754 		if (__static_call_fixup(addr, op, dest) ||
755 		    WARN_ONCE(dest != &__x86_return_thunk,
756 			      "missing return thunk: %pS-%pS: %*ph",
757 			      addr, dest, 5, addr))
758 			continue;
759 
760 		DPRINTK(RET, "return thunk at: %pS (%px) len: %d to: %pS",
761 			addr, addr, insn.length,
762 			addr + insn.length + insn.immediate.value);
763 
764 		len = patch_return(addr, &insn, bytes);
765 		if (len == insn.length) {
766 			DUMP_BYTES(RET, ((u8*)addr),  len, "%px: orig: ", addr);
767 			DUMP_BYTES(RET, ((u8*)bytes), len, "%px: repl: ", addr);
768 			text_poke_early(addr, bytes, len);
769 		}
770 	}
771 }
772 #else
773 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
774 #endif /* CONFIG_RETHUNK */
775 
776 #else /* !CONFIG_RETPOLINE || !CONFIG_OBJTOOL */
777 
778 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) { }
779 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
780 
781 #endif /* CONFIG_RETPOLINE && CONFIG_OBJTOOL */
782 
783 #ifdef CONFIG_X86_KERNEL_IBT
784 
785 static void poison_cfi(void *addr);
786 
787 static void __init_or_module poison_endbr(void *addr, bool warn)
788 {
789 	u32 endbr, poison = gen_endbr_poison();
790 
791 	if (WARN_ON_ONCE(get_kernel_nofault(endbr, addr)))
792 		return;
793 
794 	if (!is_endbr(endbr)) {
795 		WARN_ON_ONCE(warn);
796 		return;
797 	}
798 
799 	DPRINTK(ENDBR, "ENDBR at: %pS (%px)", addr, addr);
800 
801 	/*
802 	 * When we have IBT, the lack of ENDBR will trigger #CP
803 	 */
804 	DUMP_BYTES(ENDBR, ((u8*)addr), 4, "%px: orig: ", addr);
805 	DUMP_BYTES(ENDBR, ((u8*)&poison), 4, "%px: repl: ", addr);
806 	text_poke_early(addr, &poison, 4);
807 }
808 
809 /*
810  * Generated by: objtool --ibt
811  *
812  * Seal the functions for indirect calls by clobbering the ENDBR instructions
813  * and the kCFI hash value.
814  */
815 void __init_or_module noinline apply_seal_endbr(s32 *start, s32 *end)
816 {
817 	s32 *s;
818 
819 	for (s = start; s < end; s++) {
820 		void *addr = (void *)s + *s;
821 
822 		poison_endbr(addr, true);
823 		if (IS_ENABLED(CONFIG_FINEIBT))
824 			poison_cfi(addr - 16);
825 	}
826 }
827 
828 #else
829 
830 void __init_or_module apply_seal_endbr(s32 *start, s32 *end) { }
831 
832 #endif /* CONFIG_X86_KERNEL_IBT */
833 
834 #ifdef CONFIG_FINEIBT
835 
836 enum cfi_mode {
837 	CFI_DEFAULT,
838 	CFI_OFF,
839 	CFI_KCFI,
840 	CFI_FINEIBT,
841 };
842 
843 static enum cfi_mode cfi_mode __ro_after_init = CFI_DEFAULT;
844 static bool cfi_rand __ro_after_init = true;
845 static u32  cfi_seed __ro_after_init;
846 
847 /*
848  * Re-hash the CFI hash with a boot-time seed while making sure the result is
849  * not a valid ENDBR instruction.
850  */
851 static u32 cfi_rehash(u32 hash)
852 {
853 	hash ^= cfi_seed;
854 	while (unlikely(is_endbr(hash) || is_endbr(-hash))) {
855 		bool lsb = hash & 1;
856 		hash >>= 1;
857 		if (lsb)
858 			hash ^= 0x80200003;
859 	}
860 	return hash;
861 }
862 
863 static __init int cfi_parse_cmdline(char *str)
864 {
865 	if (!str)
866 		return -EINVAL;
867 
868 	while (str) {
869 		char *next = strchr(str, ',');
870 		if (next) {
871 			*next = 0;
872 			next++;
873 		}
874 
875 		if (!strcmp(str, "auto")) {
876 			cfi_mode = CFI_DEFAULT;
877 		} else if (!strcmp(str, "off")) {
878 			cfi_mode = CFI_OFF;
879 			cfi_rand = false;
880 		} else if (!strcmp(str, "kcfi")) {
881 			cfi_mode = CFI_KCFI;
882 		} else if (!strcmp(str, "fineibt")) {
883 			cfi_mode = CFI_FINEIBT;
884 		} else if (!strcmp(str, "norand")) {
885 			cfi_rand = false;
886 		} else {
887 			pr_err("Ignoring unknown cfi option (%s).", str);
888 		}
889 
890 		str = next;
891 	}
892 
893 	return 0;
894 }
895 early_param("cfi", cfi_parse_cmdline);
896 
897 /*
898  * kCFI						FineIBT
899  *
900  * __cfi_\func:					__cfi_\func:
901  *	movl   $0x12345678,%eax		// 5	     endbr64			// 4
902  *	nop					     subl   $0x12345678,%r10d   // 7
903  *	nop					     jz     1f			// 2
904  *	nop					     ud2			// 2
905  *	nop					1:   nop			// 1
906  *	nop
907  *	nop
908  *	nop
909  *	nop
910  *	nop
911  *	nop
912  *	nop
913  *
914  *
915  * caller:					caller:
916  *	movl	$(-0x12345678),%r10d	 // 6	     movl   $0x12345678,%r10d	// 6
917  *	addl	$-15(%r11),%r10d	 // 4	     sub    $16,%r11		// 4
918  *	je	1f			 // 2	     nop4			// 4
919  *	ud2				 // 2
920  * 1:	call	__x86_indirect_thunk_r11 // 5	     call   *%r11; nop2;	// 5
921  *
922  */
923 
924 asm(	".pushsection .rodata			\n"
925 	"fineibt_preamble_start:		\n"
926 	"	endbr64				\n"
927 	"	subl	$0x12345678, %r10d	\n"
928 	"	je	fineibt_preamble_end	\n"
929 	"	ud2				\n"
930 	"	nop				\n"
931 	"fineibt_preamble_end:			\n"
932 	".popsection\n"
933 );
934 
935 extern u8 fineibt_preamble_start[];
936 extern u8 fineibt_preamble_end[];
937 
938 #define fineibt_preamble_size (fineibt_preamble_end - fineibt_preamble_start)
939 #define fineibt_preamble_hash 7
940 
941 asm(	".pushsection .rodata			\n"
942 	"fineibt_caller_start:			\n"
943 	"	movl	$0x12345678, %r10d	\n"
944 	"	sub	$16, %r11		\n"
945 	ASM_NOP4
946 	"fineibt_caller_end:			\n"
947 	".popsection				\n"
948 );
949 
950 extern u8 fineibt_caller_start[];
951 extern u8 fineibt_caller_end[];
952 
953 #define fineibt_caller_size (fineibt_caller_end - fineibt_caller_start)
954 #define fineibt_caller_hash 2
955 
956 #define fineibt_caller_jmp (fineibt_caller_size - 2)
957 
958 static u32 decode_preamble_hash(void *addr)
959 {
960 	u8 *p = addr;
961 
962 	/* b8 78 56 34 12          mov    $0x12345678,%eax */
963 	if (p[0] == 0xb8)
964 		return *(u32 *)(addr + 1);
965 
966 	return 0; /* invalid hash value */
967 }
968 
969 static u32 decode_caller_hash(void *addr)
970 {
971 	u8 *p = addr;
972 
973 	/* 41 ba 78 56 34 12       mov    $0x12345678,%r10d */
974 	if (p[0] == 0x41 && p[1] == 0xba)
975 		return -*(u32 *)(addr + 2);
976 
977 	/* e8 0c 78 56 34 12	   jmp.d8  +12 */
978 	if (p[0] == JMP8_INSN_OPCODE && p[1] == fineibt_caller_jmp)
979 		return -*(u32 *)(addr + 2);
980 
981 	return 0; /* invalid hash value */
982 }
983 
984 /* .retpoline_sites */
985 static int cfi_disable_callers(s32 *start, s32 *end)
986 {
987 	/*
988 	 * Disable kCFI by patching in a JMP.d8, this leaves the hash immediate
989 	 * in tact for later usage. Also see decode_caller_hash() and
990 	 * cfi_rewrite_callers().
991 	 */
992 	const u8 jmp[] = { JMP8_INSN_OPCODE, fineibt_caller_jmp };
993 	s32 *s;
994 
995 	for (s = start; s < end; s++) {
996 		void *addr = (void *)s + *s;
997 		u32 hash;
998 
999 		addr -= fineibt_caller_size;
1000 		hash = decode_caller_hash(addr);
1001 		if (!hash) /* nocfi callers */
1002 			continue;
1003 
1004 		text_poke_early(addr, jmp, 2);
1005 	}
1006 
1007 	return 0;
1008 }
1009 
1010 static int cfi_enable_callers(s32 *start, s32 *end)
1011 {
1012 	/*
1013 	 * Re-enable kCFI, undo what cfi_disable_callers() did.
1014 	 */
1015 	const u8 mov[] = { 0x41, 0xba };
1016 	s32 *s;
1017 
1018 	for (s = start; s < end; s++) {
1019 		void *addr = (void *)s + *s;
1020 		u32 hash;
1021 
1022 		addr -= fineibt_caller_size;
1023 		hash = decode_caller_hash(addr);
1024 		if (!hash) /* nocfi callers */
1025 			continue;
1026 
1027 		text_poke_early(addr, mov, 2);
1028 	}
1029 
1030 	return 0;
1031 }
1032 
1033 /* .cfi_sites */
1034 static int cfi_rand_preamble(s32 *start, s32 *end)
1035 {
1036 	s32 *s;
1037 
1038 	for (s = start; s < end; s++) {
1039 		void *addr = (void *)s + *s;
1040 		u32 hash;
1041 
1042 		hash = decode_preamble_hash(addr);
1043 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1044 			 addr, addr, 5, addr))
1045 			return -EINVAL;
1046 
1047 		hash = cfi_rehash(hash);
1048 		text_poke_early(addr + 1, &hash, 4);
1049 	}
1050 
1051 	return 0;
1052 }
1053 
1054 static int cfi_rewrite_preamble(s32 *start, s32 *end)
1055 {
1056 	s32 *s;
1057 
1058 	for (s = start; s < end; s++) {
1059 		void *addr = (void *)s + *s;
1060 		u32 hash;
1061 
1062 		hash = decode_preamble_hash(addr);
1063 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1064 			 addr, addr, 5, addr))
1065 			return -EINVAL;
1066 
1067 		text_poke_early(addr, fineibt_preamble_start, fineibt_preamble_size);
1068 		WARN_ON(*(u32 *)(addr + fineibt_preamble_hash) != 0x12345678);
1069 		text_poke_early(addr + fineibt_preamble_hash, &hash, 4);
1070 	}
1071 
1072 	return 0;
1073 }
1074 
1075 static void cfi_rewrite_endbr(s32 *start, s32 *end)
1076 {
1077 	s32 *s;
1078 
1079 	for (s = start; s < end; s++) {
1080 		void *addr = (void *)s + *s;
1081 
1082 		poison_endbr(addr+16, false);
1083 	}
1084 }
1085 
1086 /* .retpoline_sites */
1087 static int cfi_rand_callers(s32 *start, s32 *end)
1088 {
1089 	s32 *s;
1090 
1091 	for (s = start; s < end; s++) {
1092 		void *addr = (void *)s + *s;
1093 		u32 hash;
1094 
1095 		addr -= fineibt_caller_size;
1096 		hash = decode_caller_hash(addr);
1097 		if (hash) {
1098 			hash = -cfi_rehash(hash);
1099 			text_poke_early(addr + 2, &hash, 4);
1100 		}
1101 	}
1102 
1103 	return 0;
1104 }
1105 
1106 static int cfi_rewrite_callers(s32 *start, s32 *end)
1107 {
1108 	s32 *s;
1109 
1110 	for (s = start; s < end; s++) {
1111 		void *addr = (void *)s + *s;
1112 		u32 hash;
1113 
1114 		addr -= fineibt_caller_size;
1115 		hash = decode_caller_hash(addr);
1116 		if (hash) {
1117 			text_poke_early(addr, fineibt_caller_start, fineibt_caller_size);
1118 			WARN_ON(*(u32 *)(addr + fineibt_caller_hash) != 0x12345678);
1119 			text_poke_early(addr + fineibt_caller_hash, &hash, 4);
1120 		}
1121 		/* rely on apply_retpolines() */
1122 	}
1123 
1124 	return 0;
1125 }
1126 
1127 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1128 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1129 {
1130 	int ret;
1131 
1132 	if (WARN_ONCE(fineibt_preamble_size != 16,
1133 		      "FineIBT preamble wrong size: %ld", fineibt_preamble_size))
1134 		return;
1135 
1136 	if (cfi_mode == CFI_DEFAULT) {
1137 		cfi_mode = CFI_KCFI;
1138 		if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT))
1139 			cfi_mode = CFI_FINEIBT;
1140 	}
1141 
1142 	/*
1143 	 * Rewrite the callers to not use the __cfi_ stubs, such that we might
1144 	 * rewrite them. This disables all CFI. If this succeeds but any of the
1145 	 * later stages fails, we're without CFI.
1146 	 */
1147 	ret = cfi_disable_callers(start_retpoline, end_retpoline);
1148 	if (ret)
1149 		goto err;
1150 
1151 	if (cfi_rand) {
1152 		if (builtin)
1153 			cfi_seed = get_random_u32();
1154 
1155 		ret = cfi_rand_preamble(start_cfi, end_cfi);
1156 		if (ret)
1157 			goto err;
1158 
1159 		ret = cfi_rand_callers(start_retpoline, end_retpoline);
1160 		if (ret)
1161 			goto err;
1162 	}
1163 
1164 	switch (cfi_mode) {
1165 	case CFI_OFF:
1166 		if (builtin)
1167 			pr_info("Disabling CFI\n");
1168 		return;
1169 
1170 	case CFI_KCFI:
1171 		ret = cfi_enable_callers(start_retpoline, end_retpoline);
1172 		if (ret)
1173 			goto err;
1174 
1175 		if (builtin)
1176 			pr_info("Using kCFI\n");
1177 		return;
1178 
1179 	case CFI_FINEIBT:
1180 		/* place the FineIBT preamble at func()-16 */
1181 		ret = cfi_rewrite_preamble(start_cfi, end_cfi);
1182 		if (ret)
1183 			goto err;
1184 
1185 		/* rewrite the callers to target func()-16 */
1186 		ret = cfi_rewrite_callers(start_retpoline, end_retpoline);
1187 		if (ret)
1188 			goto err;
1189 
1190 		/* now that nobody targets func()+0, remove ENDBR there */
1191 		cfi_rewrite_endbr(start_cfi, end_cfi);
1192 
1193 		if (builtin)
1194 			pr_info("Using FineIBT CFI\n");
1195 		return;
1196 
1197 	default:
1198 		break;
1199 	}
1200 
1201 err:
1202 	pr_err("Something went horribly wrong trying to rewrite the CFI implementation.\n");
1203 }
1204 
1205 static inline void poison_hash(void *addr)
1206 {
1207 	*(u32 *)addr = 0;
1208 }
1209 
1210 static void poison_cfi(void *addr)
1211 {
1212 	switch (cfi_mode) {
1213 	case CFI_FINEIBT:
1214 		/*
1215 		 * __cfi_\func:
1216 		 *	osp nopl (%rax)
1217 		 *	subl	$0, %r10d
1218 		 *	jz	1f
1219 		 *	ud2
1220 		 * 1:	nop
1221 		 */
1222 		poison_endbr(addr, false);
1223 		poison_hash(addr + fineibt_preamble_hash);
1224 		break;
1225 
1226 	case CFI_KCFI:
1227 		/*
1228 		 * __cfi_\func:
1229 		 *	movl	$0, %eax
1230 		 *	.skip	11, 0x90
1231 		 */
1232 		poison_hash(addr + 1);
1233 		break;
1234 
1235 	default:
1236 		break;
1237 	}
1238 }
1239 
1240 #else
1241 
1242 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1243 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1244 {
1245 }
1246 
1247 #ifdef CONFIG_X86_KERNEL_IBT
1248 static void poison_cfi(void *addr) { }
1249 #endif
1250 
1251 #endif
1252 
1253 void apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1254 		   s32 *start_cfi, s32 *end_cfi)
1255 {
1256 	return __apply_fineibt(start_retpoline, end_retpoline,
1257 			       start_cfi, end_cfi,
1258 			       /* .builtin = */ false);
1259 }
1260 
1261 #ifdef CONFIG_SMP
1262 static void alternatives_smp_lock(const s32 *start, const s32 *end,
1263 				  u8 *text, u8 *text_end)
1264 {
1265 	const s32 *poff;
1266 
1267 	for (poff = start; poff < end; poff++) {
1268 		u8 *ptr = (u8 *)poff + *poff;
1269 
1270 		if (!*poff || ptr < text || ptr >= text_end)
1271 			continue;
1272 		/* turn DS segment override prefix into lock prefix */
1273 		if (*ptr == 0x3e)
1274 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
1275 	}
1276 }
1277 
1278 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
1279 				    u8 *text, u8 *text_end)
1280 {
1281 	const s32 *poff;
1282 
1283 	for (poff = start; poff < end; poff++) {
1284 		u8 *ptr = (u8 *)poff + *poff;
1285 
1286 		if (!*poff || ptr < text || ptr >= text_end)
1287 			continue;
1288 		/* turn lock prefix into DS segment override prefix */
1289 		if (*ptr == 0xf0)
1290 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
1291 	}
1292 }
1293 
1294 struct smp_alt_module {
1295 	/* what is this ??? */
1296 	struct module	*mod;
1297 	char		*name;
1298 
1299 	/* ptrs to lock prefixes */
1300 	const s32	*locks;
1301 	const s32	*locks_end;
1302 
1303 	/* .text segment, needed to avoid patching init code ;) */
1304 	u8		*text;
1305 	u8		*text_end;
1306 
1307 	struct list_head next;
1308 };
1309 static LIST_HEAD(smp_alt_modules);
1310 static bool uniproc_patched = false;	/* protected by text_mutex */
1311 
1312 void __init_or_module alternatives_smp_module_add(struct module *mod,
1313 						  char *name,
1314 						  void *locks, void *locks_end,
1315 						  void *text,  void *text_end)
1316 {
1317 	struct smp_alt_module *smp;
1318 
1319 	mutex_lock(&text_mutex);
1320 	if (!uniproc_patched)
1321 		goto unlock;
1322 
1323 	if (num_possible_cpus() == 1)
1324 		/* Don't bother remembering, we'll never have to undo it. */
1325 		goto smp_unlock;
1326 
1327 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
1328 	if (NULL == smp)
1329 		/* we'll run the (safe but slow) SMP code then ... */
1330 		goto unlock;
1331 
1332 	smp->mod	= mod;
1333 	smp->name	= name;
1334 	smp->locks	= locks;
1335 	smp->locks_end	= locks_end;
1336 	smp->text	= text;
1337 	smp->text_end	= text_end;
1338 	DPRINTK(SMP, "locks %p -> %p, text %p -> %p, name %s\n",
1339 		smp->locks, smp->locks_end,
1340 		smp->text, smp->text_end, smp->name);
1341 
1342 	list_add_tail(&smp->next, &smp_alt_modules);
1343 smp_unlock:
1344 	alternatives_smp_unlock(locks, locks_end, text, text_end);
1345 unlock:
1346 	mutex_unlock(&text_mutex);
1347 }
1348 
1349 void __init_or_module alternatives_smp_module_del(struct module *mod)
1350 {
1351 	struct smp_alt_module *item;
1352 
1353 	mutex_lock(&text_mutex);
1354 	list_for_each_entry(item, &smp_alt_modules, next) {
1355 		if (mod != item->mod)
1356 			continue;
1357 		list_del(&item->next);
1358 		kfree(item);
1359 		break;
1360 	}
1361 	mutex_unlock(&text_mutex);
1362 }
1363 
1364 void alternatives_enable_smp(void)
1365 {
1366 	struct smp_alt_module *mod;
1367 
1368 	/* Why bother if there are no other CPUs? */
1369 	BUG_ON(num_possible_cpus() == 1);
1370 
1371 	mutex_lock(&text_mutex);
1372 
1373 	if (uniproc_patched) {
1374 		pr_info("switching to SMP code\n");
1375 		BUG_ON(num_online_cpus() != 1);
1376 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
1377 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
1378 		list_for_each_entry(mod, &smp_alt_modules, next)
1379 			alternatives_smp_lock(mod->locks, mod->locks_end,
1380 					      mod->text, mod->text_end);
1381 		uniproc_patched = false;
1382 	}
1383 	mutex_unlock(&text_mutex);
1384 }
1385 
1386 /*
1387  * Return 1 if the address range is reserved for SMP-alternatives.
1388  * Must hold text_mutex.
1389  */
1390 int alternatives_text_reserved(void *start, void *end)
1391 {
1392 	struct smp_alt_module *mod;
1393 	const s32 *poff;
1394 	u8 *text_start = start;
1395 	u8 *text_end = end;
1396 
1397 	lockdep_assert_held(&text_mutex);
1398 
1399 	list_for_each_entry(mod, &smp_alt_modules, next) {
1400 		if (mod->text > text_end || mod->text_end < text_start)
1401 			continue;
1402 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
1403 			const u8 *ptr = (const u8 *)poff + *poff;
1404 
1405 			if (text_start <= ptr && text_end > ptr)
1406 				return 1;
1407 		}
1408 	}
1409 
1410 	return 0;
1411 }
1412 #endif /* CONFIG_SMP */
1413 
1414 #ifdef CONFIG_PARAVIRT
1415 
1416 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
1417 static void __init_or_module add_nops(void *insns, unsigned int len)
1418 {
1419 	while (len > 0) {
1420 		unsigned int noplen = len;
1421 		if (noplen > ASM_NOP_MAX)
1422 			noplen = ASM_NOP_MAX;
1423 		memcpy(insns, x86_nops[noplen], noplen);
1424 		insns += noplen;
1425 		len -= noplen;
1426 	}
1427 }
1428 
1429 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
1430 				     struct paravirt_patch_site *end)
1431 {
1432 	struct paravirt_patch_site *p;
1433 	char insn_buff[MAX_PATCH_LEN];
1434 
1435 	for (p = start; p < end; p++) {
1436 		unsigned int used;
1437 
1438 		BUG_ON(p->len > MAX_PATCH_LEN);
1439 		/* prep the buffer with the original instructions */
1440 		memcpy(insn_buff, p->instr, p->len);
1441 		used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
1442 
1443 		BUG_ON(used > p->len);
1444 
1445 		/* Pad the rest with nops */
1446 		add_nops(insn_buff + used, p->len - used);
1447 		text_poke_early(p->instr, insn_buff, p->len);
1448 	}
1449 }
1450 extern struct paravirt_patch_site __start_parainstructions[],
1451 	__stop_parainstructions[];
1452 #endif	/* CONFIG_PARAVIRT */
1453 
1454 /*
1455  * Self-test for the INT3 based CALL emulation code.
1456  *
1457  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
1458  * properly and that there is a stack gap between the INT3 frame and the
1459  * previous context. Without this gap doing a virtual PUSH on the interrupted
1460  * stack would corrupt the INT3 IRET frame.
1461  *
1462  * See entry_{32,64}.S for more details.
1463  */
1464 
1465 /*
1466  * We define the int3_magic() function in assembly to control the calling
1467  * convention such that we can 'call' it from assembly.
1468  */
1469 
1470 extern void int3_magic(unsigned int *ptr); /* defined in asm */
1471 
1472 asm (
1473 "	.pushsection	.init.text, \"ax\", @progbits\n"
1474 "	.type		int3_magic, @function\n"
1475 "int3_magic:\n"
1476 	ANNOTATE_NOENDBR
1477 "	movl	$1, (%" _ASM_ARG1 ")\n"
1478 	ASM_RET
1479 "	.size		int3_magic, .-int3_magic\n"
1480 "	.popsection\n"
1481 );
1482 
1483 extern void int3_selftest_ip(void); /* defined in asm below */
1484 
1485 static int __init
1486 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
1487 {
1488 	unsigned long selftest = (unsigned long)&int3_selftest_ip;
1489 	struct die_args *args = data;
1490 	struct pt_regs *regs = args->regs;
1491 
1492 	OPTIMIZER_HIDE_VAR(selftest);
1493 
1494 	if (!regs || user_mode(regs))
1495 		return NOTIFY_DONE;
1496 
1497 	if (val != DIE_INT3)
1498 		return NOTIFY_DONE;
1499 
1500 	if (regs->ip - INT3_INSN_SIZE != selftest)
1501 		return NOTIFY_DONE;
1502 
1503 	int3_emulate_call(regs, (unsigned long)&int3_magic);
1504 	return NOTIFY_STOP;
1505 }
1506 
1507 /* Must be noinline to ensure uniqueness of int3_selftest_ip. */
1508 static noinline void __init int3_selftest(void)
1509 {
1510 	static __initdata struct notifier_block int3_exception_nb = {
1511 		.notifier_call	= int3_exception_notify,
1512 		.priority	= INT_MAX-1, /* last */
1513 	};
1514 	unsigned int val = 0;
1515 
1516 	BUG_ON(register_die_notifier(&int3_exception_nb));
1517 
1518 	/*
1519 	 * Basically: int3_magic(&val); but really complicated :-)
1520 	 *
1521 	 * INT3 padded with NOP to CALL_INSN_SIZE. The int3_exception_nb
1522 	 * notifier above will emulate CALL for us.
1523 	 */
1524 	asm volatile ("int3_selftest_ip:\n\t"
1525 		      ANNOTATE_NOENDBR
1526 		      "    int3; nop; nop; nop; nop\n\t"
1527 		      : ASM_CALL_CONSTRAINT
1528 		      : __ASM_SEL_RAW(a, D) (&val)
1529 		      : "memory");
1530 
1531 	BUG_ON(val != 1);
1532 
1533 	unregister_die_notifier(&int3_exception_nb);
1534 }
1535 
1536 static __initdata int __alt_reloc_selftest_addr;
1537 
1538 extern void __init __alt_reloc_selftest(void *arg);
1539 __visible noinline void __init __alt_reloc_selftest(void *arg)
1540 {
1541 	WARN_ON(arg != &__alt_reloc_selftest_addr);
1542 }
1543 
1544 static noinline void __init alt_reloc_selftest(void)
1545 {
1546 	/*
1547 	 * Tests apply_relocation().
1548 	 *
1549 	 * This has a relative immediate (CALL) in a place other than the first
1550 	 * instruction and additionally on x86_64 we get a RIP-relative LEA:
1551 	 *
1552 	 *   lea    0x0(%rip),%rdi  # 5d0: R_X86_64_PC32    .init.data+0x5566c
1553 	 *   call   +0              # 5d5: R_X86_64_PLT32   __alt_reloc_selftest-0x4
1554 	 *
1555 	 * Getting this wrong will either crash and burn or tickle the WARN
1556 	 * above.
1557 	 */
1558 	asm_inline volatile (
1559 		ALTERNATIVE("", "lea %[mem], %%" _ASM_ARG1 "; call __alt_reloc_selftest;", X86_FEATURE_ALWAYS)
1560 		: /* output */
1561 		: [mem] "m" (__alt_reloc_selftest_addr)
1562 		: _ASM_ARG1
1563 	);
1564 }
1565 
1566 void __init alternative_instructions(void)
1567 {
1568 	int3_selftest();
1569 
1570 	/*
1571 	 * The patching is not fully atomic, so try to avoid local
1572 	 * interruptions that might execute the to be patched code.
1573 	 * Other CPUs are not running.
1574 	 */
1575 	stop_nmi();
1576 
1577 	/*
1578 	 * Don't stop machine check exceptions while patching.
1579 	 * MCEs only happen when something got corrupted and in this
1580 	 * case we must do something about the corruption.
1581 	 * Ignoring it is worse than an unlikely patching race.
1582 	 * Also machine checks tend to be broadcast and if one CPU
1583 	 * goes into machine check the others follow quickly, so we don't
1584 	 * expect a machine check to cause undue problems during to code
1585 	 * patching.
1586 	 */
1587 
1588 	/*
1589 	 * Paravirt patching and alternative patching can be combined to
1590 	 * replace a function call with a short direct code sequence (e.g.
1591 	 * by setting a constant return value instead of doing that in an
1592 	 * external function).
1593 	 * In order to make this work the following sequence is required:
1594 	 * 1. set (artificial) features depending on used paravirt
1595 	 *    functions which can later influence alternative patching
1596 	 * 2. apply paravirt patching (generally replacing an indirect
1597 	 *    function call with a direct one)
1598 	 * 3. apply alternative patching (e.g. replacing a direct function
1599 	 *    call with a custom code sequence)
1600 	 * Doing paravirt patching after alternative patching would clobber
1601 	 * the optimization of the custom code with a function call again.
1602 	 */
1603 	paravirt_set_cap();
1604 
1605 	/*
1606 	 * First patch paravirt functions, such that we overwrite the indirect
1607 	 * call with the direct call.
1608 	 */
1609 	apply_paravirt(__parainstructions, __parainstructions_end);
1610 
1611 	__apply_fineibt(__retpoline_sites, __retpoline_sites_end,
1612 			__cfi_sites, __cfi_sites_end, true);
1613 
1614 	/*
1615 	 * Rewrite the retpolines, must be done before alternatives since
1616 	 * those can rewrite the retpoline thunks.
1617 	 */
1618 	apply_retpolines(__retpoline_sites, __retpoline_sites_end);
1619 	apply_returns(__return_sites, __return_sites_end);
1620 
1621 	/*
1622 	 * Then patch alternatives, such that those paravirt calls that are in
1623 	 * alternatives can be overwritten by their immediate fragments.
1624 	 */
1625 	apply_alternatives(__alt_instructions, __alt_instructions_end);
1626 
1627 	/*
1628 	 * Now all calls are established. Apply the call thunks if
1629 	 * required.
1630 	 */
1631 	callthunks_patch_builtin_calls();
1632 
1633 	/*
1634 	 * Seal all functions that do not have their address taken.
1635 	 */
1636 	apply_seal_endbr(__ibt_endbr_seal, __ibt_endbr_seal_end);
1637 
1638 #ifdef CONFIG_SMP
1639 	/* Patch to UP if other cpus not imminent. */
1640 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
1641 		uniproc_patched = true;
1642 		alternatives_smp_module_add(NULL, "core kernel",
1643 					    __smp_locks, __smp_locks_end,
1644 					    _text, _etext);
1645 	}
1646 
1647 	if (!uniproc_patched || num_possible_cpus() == 1) {
1648 		free_init_pages("SMP alternatives",
1649 				(unsigned long)__smp_locks,
1650 				(unsigned long)__smp_locks_end);
1651 	}
1652 #endif
1653 
1654 	restart_nmi();
1655 	alternatives_patched = 1;
1656 
1657 	alt_reloc_selftest();
1658 }
1659 
1660 /**
1661  * text_poke_early - Update instructions on a live kernel at boot time
1662  * @addr: address to modify
1663  * @opcode: source of the copy
1664  * @len: length to copy
1665  *
1666  * When you use this code to patch more than one byte of an instruction
1667  * you need to make sure that other CPUs cannot execute this code in parallel.
1668  * Also no thread must be currently preempted in the middle of these
1669  * instructions. And on the local CPU you need to be protected against NMI or
1670  * MCE handlers seeing an inconsistent instruction while you patch.
1671  */
1672 void __init_or_module text_poke_early(void *addr, const void *opcode,
1673 				      size_t len)
1674 {
1675 	unsigned long flags;
1676 
1677 	if (boot_cpu_has(X86_FEATURE_NX) &&
1678 	    is_module_text_address((unsigned long)addr)) {
1679 		/*
1680 		 * Modules text is marked initially as non-executable, so the
1681 		 * code cannot be running and speculative code-fetches are
1682 		 * prevented. Just change the code.
1683 		 */
1684 		memcpy(addr, opcode, len);
1685 	} else {
1686 		local_irq_save(flags);
1687 		memcpy(addr, opcode, len);
1688 		local_irq_restore(flags);
1689 		sync_core();
1690 
1691 		/*
1692 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
1693 		 * that causes hangs on some VIA CPUs.
1694 		 */
1695 	}
1696 }
1697 
1698 typedef struct {
1699 	struct mm_struct *mm;
1700 } temp_mm_state_t;
1701 
1702 /*
1703  * Using a temporary mm allows to set temporary mappings that are not accessible
1704  * by other CPUs. Such mappings are needed to perform sensitive memory writes
1705  * that override the kernel memory protections (e.g., W^X), without exposing the
1706  * temporary page-table mappings that are required for these write operations to
1707  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
1708  * mapping is torn down.
1709  *
1710  * Context: The temporary mm needs to be used exclusively by a single core. To
1711  *          harden security IRQs must be disabled while the temporary mm is
1712  *          loaded, thereby preventing interrupt handler bugs from overriding
1713  *          the kernel memory protection.
1714  */
1715 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
1716 {
1717 	temp_mm_state_t temp_state;
1718 
1719 	lockdep_assert_irqs_disabled();
1720 
1721 	/*
1722 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
1723 	 * with a stale address space WITHOUT being in lazy mode after
1724 	 * restoring the previous mm.
1725 	 */
1726 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
1727 		leave_mm(smp_processor_id());
1728 
1729 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1730 	switch_mm_irqs_off(NULL, mm, current);
1731 
1732 	/*
1733 	 * If breakpoints are enabled, disable them while the temporary mm is
1734 	 * used. Userspace might set up watchpoints on addresses that are used
1735 	 * in the temporary mm, which would lead to wrong signals being sent or
1736 	 * crashes.
1737 	 *
1738 	 * Note that breakpoints are not disabled selectively, which also causes
1739 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
1740 	 * undesirable, but still seems reasonable as the code that runs in the
1741 	 * temporary mm should be short.
1742 	 */
1743 	if (hw_breakpoint_active())
1744 		hw_breakpoint_disable();
1745 
1746 	return temp_state;
1747 }
1748 
1749 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
1750 {
1751 	lockdep_assert_irqs_disabled();
1752 	switch_mm_irqs_off(NULL, prev_state.mm, current);
1753 
1754 	/*
1755 	 * Restore the breakpoints if they were disabled before the temporary mm
1756 	 * was loaded.
1757 	 */
1758 	if (hw_breakpoint_active())
1759 		hw_breakpoint_restore();
1760 }
1761 
1762 __ro_after_init struct mm_struct *poking_mm;
1763 __ro_after_init unsigned long poking_addr;
1764 
1765 static void text_poke_memcpy(void *dst, const void *src, size_t len)
1766 {
1767 	memcpy(dst, src, len);
1768 }
1769 
1770 static void text_poke_memset(void *dst, const void *src, size_t len)
1771 {
1772 	int c = *(const int *)src;
1773 
1774 	memset(dst, c, len);
1775 }
1776 
1777 typedef void text_poke_f(void *dst, const void *src, size_t len);
1778 
1779 static void *__text_poke(text_poke_f func, void *addr, const void *src, size_t len)
1780 {
1781 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
1782 	struct page *pages[2] = {NULL};
1783 	temp_mm_state_t prev;
1784 	unsigned long flags;
1785 	pte_t pte, *ptep;
1786 	spinlock_t *ptl;
1787 	pgprot_t pgprot;
1788 
1789 	/*
1790 	 * While boot memory allocator is running we cannot use struct pages as
1791 	 * they are not yet initialized. There is no way to recover.
1792 	 */
1793 	BUG_ON(!after_bootmem);
1794 
1795 	if (!core_kernel_text((unsigned long)addr)) {
1796 		pages[0] = vmalloc_to_page(addr);
1797 		if (cross_page_boundary)
1798 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
1799 	} else {
1800 		pages[0] = virt_to_page(addr);
1801 		WARN_ON(!PageReserved(pages[0]));
1802 		if (cross_page_boundary)
1803 			pages[1] = virt_to_page(addr + PAGE_SIZE);
1804 	}
1805 	/*
1806 	 * If something went wrong, crash and burn since recovery paths are not
1807 	 * implemented.
1808 	 */
1809 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
1810 
1811 	/*
1812 	 * Map the page without the global bit, as TLB flushing is done with
1813 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
1814 	 */
1815 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
1816 
1817 	/*
1818 	 * The lock is not really needed, but this allows to avoid open-coding.
1819 	 */
1820 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
1821 
1822 	/*
1823 	 * This must not fail; preallocated in poking_init().
1824 	 */
1825 	VM_BUG_ON(!ptep);
1826 
1827 	local_irq_save(flags);
1828 
1829 	pte = mk_pte(pages[0], pgprot);
1830 	set_pte_at(poking_mm, poking_addr, ptep, pte);
1831 
1832 	if (cross_page_boundary) {
1833 		pte = mk_pte(pages[1], pgprot);
1834 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
1835 	}
1836 
1837 	/*
1838 	 * Loading the temporary mm behaves as a compiler barrier, which
1839 	 * guarantees that the PTE will be set at the time memcpy() is done.
1840 	 */
1841 	prev = use_temporary_mm(poking_mm);
1842 
1843 	kasan_disable_current();
1844 	func((u8 *)poking_addr + offset_in_page(addr), src, len);
1845 	kasan_enable_current();
1846 
1847 	/*
1848 	 * Ensure that the PTE is only cleared after the instructions of memcpy
1849 	 * were issued by using a compiler barrier.
1850 	 */
1851 	barrier();
1852 
1853 	pte_clear(poking_mm, poking_addr, ptep);
1854 	if (cross_page_boundary)
1855 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
1856 
1857 	/*
1858 	 * Loading the previous page-table hierarchy requires a serializing
1859 	 * instruction that already allows the core to see the updated version.
1860 	 * Xen-PV is assumed to serialize execution in a similar manner.
1861 	 */
1862 	unuse_temporary_mm(prev);
1863 
1864 	/*
1865 	 * Flushing the TLB might involve IPIs, which would require enabled
1866 	 * IRQs, but not if the mm is not used, as it is in this point.
1867 	 */
1868 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
1869 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
1870 			   PAGE_SHIFT, false);
1871 
1872 	if (func == text_poke_memcpy) {
1873 		/*
1874 		 * If the text does not match what we just wrote then something is
1875 		 * fundamentally screwy; there's nothing we can really do about that.
1876 		 */
1877 		BUG_ON(memcmp(addr, src, len));
1878 	}
1879 
1880 	local_irq_restore(flags);
1881 	pte_unmap_unlock(ptep, ptl);
1882 	return addr;
1883 }
1884 
1885 /**
1886  * text_poke - Update instructions on a live kernel
1887  * @addr: address to modify
1888  * @opcode: source of the copy
1889  * @len: length to copy
1890  *
1891  * Only atomic text poke/set should be allowed when not doing early patching.
1892  * It means the size must be writable atomically and the address must be aligned
1893  * in a way that permits an atomic write. It also makes sure we fit on a single
1894  * page.
1895  *
1896  * Note that the caller must ensure that if the modified code is part of a
1897  * module, the module would not be removed during poking. This can be achieved
1898  * by registering a module notifier, and ordering module removal and patching
1899  * trough a mutex.
1900  */
1901 void *text_poke(void *addr, const void *opcode, size_t len)
1902 {
1903 	lockdep_assert_held(&text_mutex);
1904 
1905 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1906 }
1907 
1908 /**
1909  * text_poke_kgdb - Update instructions on a live kernel by kgdb
1910  * @addr: address to modify
1911  * @opcode: source of the copy
1912  * @len: length to copy
1913  *
1914  * Only atomic text poke/set should be allowed when not doing early patching.
1915  * It means the size must be writable atomically and the address must be aligned
1916  * in a way that permits an atomic write. It also makes sure we fit on a single
1917  * page.
1918  *
1919  * Context: should only be used by kgdb, which ensures no other core is running,
1920  *	    despite the fact it does not hold the text_mutex.
1921  */
1922 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
1923 {
1924 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1925 }
1926 
1927 void *text_poke_copy_locked(void *addr, const void *opcode, size_t len,
1928 			    bool core_ok)
1929 {
1930 	unsigned long start = (unsigned long)addr;
1931 	size_t patched = 0;
1932 
1933 	if (WARN_ON_ONCE(!core_ok && core_kernel_text(start)))
1934 		return NULL;
1935 
1936 	while (patched < len) {
1937 		unsigned long ptr = start + patched;
1938 		size_t s;
1939 
1940 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1941 
1942 		__text_poke(text_poke_memcpy, (void *)ptr, opcode + patched, s);
1943 		patched += s;
1944 	}
1945 	return addr;
1946 }
1947 
1948 /**
1949  * text_poke_copy - Copy instructions into (an unused part of) RX memory
1950  * @addr: address to modify
1951  * @opcode: source of the copy
1952  * @len: length to copy, could be more than 2x PAGE_SIZE
1953  *
1954  * Not safe against concurrent execution; useful for JITs to dump
1955  * new code blocks into unused regions of RX memory. Can be used in
1956  * conjunction with synchronize_rcu_tasks() to wait for existing
1957  * execution to quiesce after having made sure no existing functions
1958  * pointers are live.
1959  */
1960 void *text_poke_copy(void *addr, const void *opcode, size_t len)
1961 {
1962 	mutex_lock(&text_mutex);
1963 	addr = text_poke_copy_locked(addr, opcode, len, false);
1964 	mutex_unlock(&text_mutex);
1965 	return addr;
1966 }
1967 
1968 /**
1969  * text_poke_set - memset into (an unused part of) RX memory
1970  * @addr: address to modify
1971  * @c: the byte to fill the area with
1972  * @len: length to copy, could be more than 2x PAGE_SIZE
1973  *
1974  * This is useful to overwrite unused regions of RX memory with illegal
1975  * instructions.
1976  */
1977 void *text_poke_set(void *addr, int c, size_t len)
1978 {
1979 	unsigned long start = (unsigned long)addr;
1980 	size_t patched = 0;
1981 
1982 	if (WARN_ON_ONCE(core_kernel_text(start)))
1983 		return NULL;
1984 
1985 	mutex_lock(&text_mutex);
1986 	while (patched < len) {
1987 		unsigned long ptr = start + patched;
1988 		size_t s;
1989 
1990 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1991 
1992 		__text_poke(text_poke_memset, (void *)ptr, (void *)&c, s);
1993 		patched += s;
1994 	}
1995 	mutex_unlock(&text_mutex);
1996 	return addr;
1997 }
1998 
1999 static void do_sync_core(void *info)
2000 {
2001 	sync_core();
2002 }
2003 
2004 void text_poke_sync(void)
2005 {
2006 	on_each_cpu(do_sync_core, NULL, 1);
2007 }
2008 
2009 /*
2010  * NOTE: crazy scheme to allow patching Jcc.d32 but not increase the size of
2011  * this thing. When len == 6 everything is prefixed with 0x0f and we map
2012  * opcode to Jcc.d8, using len to distinguish.
2013  */
2014 struct text_poke_loc {
2015 	/* addr := _stext + rel_addr */
2016 	s32 rel_addr;
2017 	s32 disp;
2018 	u8 len;
2019 	u8 opcode;
2020 	const u8 text[POKE_MAX_OPCODE_SIZE];
2021 	/* see text_poke_bp_batch() */
2022 	u8 old;
2023 };
2024 
2025 struct bp_patching_desc {
2026 	struct text_poke_loc *vec;
2027 	int nr_entries;
2028 	atomic_t refs;
2029 };
2030 
2031 static struct bp_patching_desc bp_desc;
2032 
2033 static __always_inline
2034 struct bp_patching_desc *try_get_desc(void)
2035 {
2036 	struct bp_patching_desc *desc = &bp_desc;
2037 
2038 	if (!raw_atomic_inc_not_zero(&desc->refs))
2039 		return NULL;
2040 
2041 	return desc;
2042 }
2043 
2044 static __always_inline void put_desc(void)
2045 {
2046 	struct bp_patching_desc *desc = &bp_desc;
2047 
2048 	smp_mb__before_atomic();
2049 	raw_atomic_dec(&desc->refs);
2050 }
2051 
2052 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
2053 {
2054 	return _stext + tp->rel_addr;
2055 }
2056 
2057 static __always_inline int patch_cmp(const void *key, const void *elt)
2058 {
2059 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
2060 
2061 	if (key < text_poke_addr(tp))
2062 		return -1;
2063 	if (key > text_poke_addr(tp))
2064 		return 1;
2065 	return 0;
2066 }
2067 
2068 noinstr int poke_int3_handler(struct pt_regs *regs)
2069 {
2070 	struct bp_patching_desc *desc;
2071 	struct text_poke_loc *tp;
2072 	int ret = 0;
2073 	void *ip;
2074 
2075 	if (user_mode(regs))
2076 		return 0;
2077 
2078 	/*
2079 	 * Having observed our INT3 instruction, we now must observe
2080 	 * bp_desc with non-zero refcount:
2081 	 *
2082 	 *	bp_desc.refs = 1		INT3
2083 	 *	WMB				RMB
2084 	 *	write INT3			if (bp_desc.refs != 0)
2085 	 */
2086 	smp_rmb();
2087 
2088 	desc = try_get_desc();
2089 	if (!desc)
2090 		return 0;
2091 
2092 	/*
2093 	 * Discount the INT3. See text_poke_bp_batch().
2094 	 */
2095 	ip = (void *) regs->ip - INT3_INSN_SIZE;
2096 
2097 	/*
2098 	 * Skip the binary search if there is a single member in the vector.
2099 	 */
2100 	if (unlikely(desc->nr_entries > 1)) {
2101 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
2102 				      sizeof(struct text_poke_loc),
2103 				      patch_cmp);
2104 		if (!tp)
2105 			goto out_put;
2106 	} else {
2107 		tp = desc->vec;
2108 		if (text_poke_addr(tp) != ip)
2109 			goto out_put;
2110 	}
2111 
2112 	ip += tp->len;
2113 
2114 	switch (tp->opcode) {
2115 	case INT3_INSN_OPCODE:
2116 		/*
2117 		 * Someone poked an explicit INT3, they'll want to handle it,
2118 		 * do not consume.
2119 		 */
2120 		goto out_put;
2121 
2122 	case RET_INSN_OPCODE:
2123 		int3_emulate_ret(regs);
2124 		break;
2125 
2126 	case CALL_INSN_OPCODE:
2127 		int3_emulate_call(regs, (long)ip + tp->disp);
2128 		break;
2129 
2130 	case JMP32_INSN_OPCODE:
2131 	case JMP8_INSN_OPCODE:
2132 		int3_emulate_jmp(regs, (long)ip + tp->disp);
2133 		break;
2134 
2135 	case 0x70 ... 0x7f: /* Jcc */
2136 		int3_emulate_jcc(regs, tp->opcode & 0xf, (long)ip, tp->disp);
2137 		break;
2138 
2139 	default:
2140 		BUG();
2141 	}
2142 
2143 	ret = 1;
2144 
2145 out_put:
2146 	put_desc();
2147 	return ret;
2148 }
2149 
2150 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
2151 static struct text_poke_loc tp_vec[TP_VEC_MAX];
2152 static int tp_vec_nr;
2153 
2154 /**
2155  * text_poke_bp_batch() -- update instructions on live kernel on SMP
2156  * @tp:			vector of instructions to patch
2157  * @nr_entries:		number of entries in the vector
2158  *
2159  * Modify multi-byte instruction by using int3 breakpoint on SMP.
2160  * We completely avoid stop_machine() here, and achieve the
2161  * synchronization using int3 breakpoint.
2162  *
2163  * The way it is done:
2164  *	- For each entry in the vector:
2165  *		- add a int3 trap to the address that will be patched
2166  *	- sync cores
2167  *	- For each entry in the vector:
2168  *		- update all but the first byte of the patched range
2169  *	- sync cores
2170  *	- For each entry in the vector:
2171  *		- replace the first byte (int3) by the first byte of
2172  *		  replacing opcode
2173  *	- sync cores
2174  */
2175 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
2176 {
2177 	unsigned char int3 = INT3_INSN_OPCODE;
2178 	unsigned int i;
2179 	int do_sync;
2180 
2181 	lockdep_assert_held(&text_mutex);
2182 
2183 	bp_desc.vec = tp;
2184 	bp_desc.nr_entries = nr_entries;
2185 
2186 	/*
2187 	 * Corresponds to the implicit memory barrier in try_get_desc() to
2188 	 * ensure reading a non-zero refcount provides up to date bp_desc data.
2189 	 */
2190 	atomic_set_release(&bp_desc.refs, 1);
2191 
2192 	/*
2193 	 * Function tracing can enable thousands of places that need to be
2194 	 * updated. This can take quite some time, and with full kernel debugging
2195 	 * enabled, this could cause the softlockup watchdog to trigger.
2196 	 * This function gets called every 256 entries added to be patched.
2197 	 * Call cond_resched() here to make sure that other tasks can get scheduled
2198 	 * while processing all the functions being patched.
2199 	 */
2200 	cond_resched();
2201 
2202 	/*
2203 	 * Corresponding read barrier in int3 notifier for making sure the
2204 	 * nr_entries and handler are correctly ordered wrt. patching.
2205 	 */
2206 	smp_wmb();
2207 
2208 	/*
2209 	 * First step: add a int3 trap to the address that will be patched.
2210 	 */
2211 	for (i = 0; i < nr_entries; i++) {
2212 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
2213 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
2214 	}
2215 
2216 	text_poke_sync();
2217 
2218 	/*
2219 	 * Second step: update all but the first byte of the patched range.
2220 	 */
2221 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2222 		u8 old[POKE_MAX_OPCODE_SIZE+1] = { tp[i].old, };
2223 		u8 _new[POKE_MAX_OPCODE_SIZE+1];
2224 		const u8 *new = tp[i].text;
2225 		int len = tp[i].len;
2226 
2227 		if (len - INT3_INSN_SIZE > 0) {
2228 			memcpy(old + INT3_INSN_SIZE,
2229 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2230 			       len - INT3_INSN_SIZE);
2231 
2232 			if (len == 6) {
2233 				_new[0] = 0x0f;
2234 				memcpy(_new + 1, new, 5);
2235 				new = _new;
2236 			}
2237 
2238 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2239 				  new + INT3_INSN_SIZE,
2240 				  len - INT3_INSN_SIZE);
2241 
2242 			do_sync++;
2243 		}
2244 
2245 		/*
2246 		 * Emit a perf event to record the text poke, primarily to
2247 		 * support Intel PT decoding which must walk the executable code
2248 		 * to reconstruct the trace. The flow up to here is:
2249 		 *   - write INT3 byte
2250 		 *   - IPI-SYNC
2251 		 *   - write instruction tail
2252 		 * At this point the actual control flow will be through the
2253 		 * INT3 and handler and not hit the old or new instruction.
2254 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
2255 		 * can still be decoded. Subsequently:
2256 		 *   - emit RECORD_TEXT_POKE with the new instruction
2257 		 *   - IPI-SYNC
2258 		 *   - write first byte
2259 		 *   - IPI-SYNC
2260 		 * So before the text poke event timestamp, the decoder will see
2261 		 * either the old instruction flow or FUP/TIP of INT3. After the
2262 		 * text poke event timestamp, the decoder will see either the
2263 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
2264 		 * use the timestamp as the point at which to modify the
2265 		 * executable code.
2266 		 * The old instruction is recorded so that the event can be
2267 		 * processed forwards or backwards.
2268 		 */
2269 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len, new, len);
2270 	}
2271 
2272 	if (do_sync) {
2273 		/*
2274 		 * According to Intel, this core syncing is very likely
2275 		 * not necessary and we'd be safe even without it. But
2276 		 * better safe than sorry (plus there's not only Intel).
2277 		 */
2278 		text_poke_sync();
2279 	}
2280 
2281 	/*
2282 	 * Third step: replace the first byte (int3) by the first byte of
2283 	 * replacing opcode.
2284 	 */
2285 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2286 		u8 byte = tp[i].text[0];
2287 
2288 		if (tp[i].len == 6)
2289 			byte = 0x0f;
2290 
2291 		if (byte == INT3_INSN_OPCODE)
2292 			continue;
2293 
2294 		text_poke(text_poke_addr(&tp[i]), &byte, INT3_INSN_SIZE);
2295 		do_sync++;
2296 	}
2297 
2298 	if (do_sync)
2299 		text_poke_sync();
2300 
2301 	/*
2302 	 * Remove and wait for refs to be zero.
2303 	 */
2304 	if (!atomic_dec_and_test(&bp_desc.refs))
2305 		atomic_cond_read_acquire(&bp_desc.refs, !VAL);
2306 }
2307 
2308 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
2309 			       const void *opcode, size_t len, const void *emulate)
2310 {
2311 	struct insn insn;
2312 	int ret, i = 0;
2313 
2314 	if (len == 6)
2315 		i = 1;
2316 	memcpy((void *)tp->text, opcode+i, len-i);
2317 	if (!emulate)
2318 		emulate = opcode;
2319 
2320 	ret = insn_decode_kernel(&insn, emulate);
2321 	BUG_ON(ret < 0);
2322 
2323 	tp->rel_addr = addr - (void *)_stext;
2324 	tp->len = len;
2325 	tp->opcode = insn.opcode.bytes[0];
2326 
2327 	if (is_jcc32(&insn)) {
2328 		/*
2329 		 * Map Jcc.d32 onto Jcc.d8 and use len to distinguish.
2330 		 */
2331 		tp->opcode = insn.opcode.bytes[1] - 0x10;
2332 	}
2333 
2334 	switch (tp->opcode) {
2335 	case RET_INSN_OPCODE:
2336 	case JMP32_INSN_OPCODE:
2337 	case JMP8_INSN_OPCODE:
2338 		/*
2339 		 * Control flow instructions without implied execution of the
2340 		 * next instruction can be padded with INT3.
2341 		 */
2342 		for (i = insn.length; i < len; i++)
2343 			BUG_ON(tp->text[i] != INT3_INSN_OPCODE);
2344 		break;
2345 
2346 	default:
2347 		BUG_ON(len != insn.length);
2348 	}
2349 
2350 	switch (tp->opcode) {
2351 	case INT3_INSN_OPCODE:
2352 	case RET_INSN_OPCODE:
2353 		break;
2354 
2355 	case CALL_INSN_OPCODE:
2356 	case JMP32_INSN_OPCODE:
2357 	case JMP8_INSN_OPCODE:
2358 	case 0x70 ... 0x7f: /* Jcc */
2359 		tp->disp = insn.immediate.value;
2360 		break;
2361 
2362 	default: /* assume NOP */
2363 		switch (len) {
2364 		case 2: /* NOP2 -- emulate as JMP8+0 */
2365 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2366 			tp->opcode = JMP8_INSN_OPCODE;
2367 			tp->disp = 0;
2368 			break;
2369 
2370 		case 5: /* NOP5 -- emulate as JMP32+0 */
2371 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2372 			tp->opcode = JMP32_INSN_OPCODE;
2373 			tp->disp = 0;
2374 			break;
2375 
2376 		default: /* unknown instruction */
2377 			BUG();
2378 		}
2379 		break;
2380 	}
2381 }
2382 
2383 /*
2384  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
2385  * early if needed.
2386  */
2387 static bool tp_order_fail(void *addr)
2388 {
2389 	struct text_poke_loc *tp;
2390 
2391 	if (!tp_vec_nr)
2392 		return false;
2393 
2394 	if (!addr) /* force */
2395 		return true;
2396 
2397 	tp = &tp_vec[tp_vec_nr - 1];
2398 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
2399 		return true;
2400 
2401 	return false;
2402 }
2403 
2404 static void text_poke_flush(void *addr)
2405 {
2406 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
2407 		text_poke_bp_batch(tp_vec, tp_vec_nr);
2408 		tp_vec_nr = 0;
2409 	}
2410 }
2411 
2412 void text_poke_finish(void)
2413 {
2414 	text_poke_flush(NULL);
2415 }
2416 
2417 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
2418 {
2419 	struct text_poke_loc *tp;
2420 
2421 	text_poke_flush(addr);
2422 
2423 	tp = &tp_vec[tp_vec_nr++];
2424 	text_poke_loc_init(tp, addr, opcode, len, emulate);
2425 }
2426 
2427 /**
2428  * text_poke_bp() -- update instructions on live kernel on SMP
2429  * @addr:	address to patch
2430  * @opcode:	opcode of new instruction
2431  * @len:	length to copy
2432  * @emulate:	instruction to be emulated
2433  *
2434  * Update a single instruction with the vector in the stack, avoiding
2435  * dynamically allocated memory. This function should be used when it is
2436  * not possible to allocate memory.
2437  */
2438 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
2439 {
2440 	struct text_poke_loc tp;
2441 
2442 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
2443 	text_poke_bp_batch(&tp, 1);
2444 }
2445