xref: /linux/arch/x86/kernel/alternative.c (revision 778e73d2411abc8f3a2d60dbf038acaec218792e)
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 s32 __smp_locks[], __smp_locks_end[];
164 void text_poke_early(void *addr, const void *opcode, size_t len);
165 
166 /*
167  * Matches NOP and NOPL, not any of the other possible NOPs.
168  */
169 static bool insn_is_nop(struct insn *insn)
170 {
171 	/* Anything NOP, but no REP NOP */
172 	if (insn->opcode.bytes[0] == 0x90 &&
173 	    (!insn->prefixes.nbytes || insn->prefixes.bytes[0] != 0xF3))
174 		return true;
175 
176 	/* NOPL */
177 	if (insn->opcode.bytes[0] == 0x0F && insn->opcode.bytes[1] == 0x1F)
178 		return true;
179 
180 	/* TODO: more nops */
181 
182 	return false;
183 }
184 
185 /*
186  * Find the offset of the first non-NOP instruction starting at @offset
187  * but no further than @len.
188  */
189 static int skip_nops(u8 *instr, int offset, int len)
190 {
191 	struct insn insn;
192 
193 	for (; offset < len; offset += insn.length) {
194 		if (insn_decode_kernel(&insn, &instr[offset]))
195 			break;
196 
197 		if (!insn_is_nop(&insn))
198 			break;
199 	}
200 
201 	return offset;
202 }
203 
204 /*
205  * Optimize a sequence of NOPs, possibly preceded by an unconditional jump
206  * to the end of the NOP sequence into a single NOP.
207  */
208 static bool __init_or_module
209 __optimize_nops(u8 *instr, size_t len, struct insn *insn, int *next, int *prev, int *target)
210 {
211 	int i = *next - insn->length;
212 
213 	switch (insn->opcode.bytes[0]) {
214 	case JMP8_INSN_OPCODE:
215 	case JMP32_INSN_OPCODE:
216 		*prev = i;
217 		*target = *next + insn->immediate.value;
218 		return false;
219 	}
220 
221 	if (insn_is_nop(insn)) {
222 		int nop = i;
223 
224 		*next = skip_nops(instr, *next, len);
225 		if (*target && *next == *target)
226 			nop = *prev;
227 
228 		add_nop(instr + nop, *next - nop);
229 		DUMP_BYTES(ALT, instr, len, "%px: [%d:%d) optimized NOPs: ", instr, nop, *next);
230 		return true;
231 	}
232 
233 	*target = 0;
234 	return false;
235 }
236 
237 /*
238  * "noinline" to cause control flow change and thus invalidate I$ and
239  * cause refetch after modification.
240  */
241 static void __init_or_module noinline optimize_nops(u8 *instr, size_t len)
242 {
243 	int prev, target = 0;
244 
245 	for (int next, i = 0; i < len; i = next) {
246 		struct insn insn;
247 
248 		if (insn_decode_kernel(&insn, &instr[i]))
249 			return;
250 
251 		next = i + insn.length;
252 
253 		__optimize_nops(instr, len, &insn, &next, &prev, &target);
254 	}
255 }
256 
257 static void __init_or_module noinline optimize_nops_inplace(u8 *instr, size_t len)
258 {
259 	unsigned long flags;
260 
261 	local_irq_save(flags);
262 	optimize_nops(instr, len);
263 	sync_core();
264 	local_irq_restore(flags);
265 }
266 
267 /*
268  * In this context, "source" is where the instructions are placed in the
269  * section .altinstr_replacement, for example during kernel build by the
270  * toolchain.
271  * "Destination" is where the instructions are being patched in by this
272  * machinery.
273  *
274  * The source offset is:
275  *
276  *   src_imm = target - src_next_ip                  (1)
277  *
278  * and the target offset is:
279  *
280  *   dst_imm = target - dst_next_ip                  (2)
281  *
282  * so rework (1) as an expression for target like:
283  *
284  *   target = src_imm + src_next_ip                  (1a)
285  *
286  * and substitute in (2) to get:
287  *
288  *   dst_imm = (src_imm + src_next_ip) - dst_next_ip (3)
289  *
290  * Now, since the instruction stream is 'identical' at src and dst (it
291  * is being copied after all) it can be stated that:
292  *
293  *   src_next_ip = src + ip_offset
294  *   dst_next_ip = dst + ip_offset                   (4)
295  *
296  * Substitute (4) in (3) and observe ip_offset being cancelled out to
297  * obtain:
298  *
299  *   dst_imm = src_imm + (src + ip_offset) - (dst + ip_offset)
300  *           = src_imm + src - dst + ip_offset - ip_offset
301  *           = src_imm + src - dst                   (5)
302  *
303  * IOW, only the relative displacement of the code block matters.
304  */
305 
306 #define apply_reloc_n(n_, p_, d_)				\
307 	do {							\
308 		s32 v = *(s##n_ *)(p_);				\
309 		v += (d_);					\
310 		BUG_ON((v >> 31) != (v >> (n_-1)));		\
311 		*(s##n_ *)(p_) = (s##n_)v;			\
312 	} while (0)
313 
314 
315 static __always_inline
316 void apply_reloc(int n, void *ptr, uintptr_t diff)
317 {
318 	switch (n) {
319 	case 1: apply_reloc_n(8, ptr, diff); break;
320 	case 2: apply_reloc_n(16, ptr, diff); break;
321 	case 4: apply_reloc_n(32, ptr, diff); break;
322 	default: BUG();
323 	}
324 }
325 
326 static __always_inline
327 bool need_reloc(unsigned long offset, u8 *src, size_t src_len)
328 {
329 	u8 *target = src + offset;
330 	/*
331 	 * If the target is inside the patched block, it's relative to the
332 	 * block itself and does not need relocation.
333 	 */
334 	return (target < src || target > src + src_len);
335 }
336 
337 static void __init_or_module noinline
338 apply_relocation(u8 *buf, size_t len, u8 *dest, u8 *src, size_t src_len)
339 {
340 	int prev, target = 0;
341 
342 	for (int next, i = 0; i < len; i = next) {
343 		struct insn insn;
344 
345 		if (WARN_ON_ONCE(insn_decode_kernel(&insn, &buf[i])))
346 			return;
347 
348 		next = i + insn.length;
349 
350 		if (__optimize_nops(buf, len, &insn, &next, &prev, &target))
351 			continue;
352 
353 		switch (insn.opcode.bytes[0]) {
354 		case 0x0f:
355 			if (insn.opcode.bytes[1] < 0x80 ||
356 			    insn.opcode.bytes[1] > 0x8f)
357 				break;
358 
359 			fallthrough;	/* Jcc.d32 */
360 		case 0x70 ... 0x7f:	/* Jcc.d8 */
361 		case JMP8_INSN_OPCODE:
362 		case JMP32_INSN_OPCODE:
363 		case CALL_INSN_OPCODE:
364 			if (need_reloc(next + insn.immediate.value, src, src_len)) {
365 				apply_reloc(insn.immediate.nbytes,
366 					    buf + i + insn_offset_immediate(&insn),
367 					    src - dest);
368 			}
369 
370 			/*
371 			 * Where possible, convert JMP.d32 into JMP.d8.
372 			 */
373 			if (insn.opcode.bytes[0] == JMP32_INSN_OPCODE) {
374 				s32 imm = insn.immediate.value;
375 				imm += src - dest;
376 				imm += JMP32_INSN_SIZE - JMP8_INSN_SIZE;
377 				if ((imm >> 31) == (imm >> 7)) {
378 					buf[i+0] = JMP8_INSN_OPCODE;
379 					buf[i+1] = (s8)imm;
380 
381 					memset(&buf[i+2], INT3_INSN_OPCODE, insn.length - 2);
382 				}
383 			}
384 			break;
385 		}
386 
387 		if (insn_rip_relative(&insn)) {
388 			if (need_reloc(next + insn.displacement.value, src, src_len)) {
389 				apply_reloc(insn.displacement.nbytes,
390 					    buf + i + insn_offset_displacement(&insn),
391 					    src - dest);
392 			}
393 		}
394 	}
395 }
396 
397 /* Low-level backend functions usable from alternative code replacements. */
398 DEFINE_ASM_FUNC(nop_func, "", .entry.text);
399 EXPORT_SYMBOL_GPL(nop_func);
400 
401 noinstr void BUG_func(void)
402 {
403 	BUG();
404 }
405 EXPORT_SYMBOL_GPL(BUG_func);
406 
407 #define CALL_RIP_REL_OPCODE	0xff
408 #define CALL_RIP_REL_MODRM	0x15
409 
410 /*
411  * Rewrite the "call BUG_func" replacement to point to the target of the
412  * indirect pv_ops call "call *disp(%ip)".
413  */
414 static int alt_replace_call(u8 *instr, u8 *insn_buff, struct alt_instr *a)
415 {
416 	void *target, *bug = &BUG_func;
417 	s32 disp;
418 
419 	if (a->replacementlen != 5 || insn_buff[0] != CALL_INSN_OPCODE) {
420 		pr_err("ALT_FLAG_DIRECT_CALL set for a non-call replacement instruction\n");
421 		BUG();
422 	}
423 
424 	if (a->instrlen != 6 ||
425 	    instr[0] != CALL_RIP_REL_OPCODE ||
426 	    instr[1] != CALL_RIP_REL_MODRM) {
427 		pr_err("ALT_FLAG_DIRECT_CALL set for unrecognized indirect call\n");
428 		BUG();
429 	}
430 
431 	/* Skip CALL_RIP_REL_OPCODE and CALL_RIP_REL_MODRM */
432 	disp = *(s32 *)(instr + 2);
433 #ifdef CONFIG_X86_64
434 	/* ff 15 00 00 00 00   call   *0x0(%rip) */
435 	/* target address is stored at "next instruction + disp". */
436 	target = *(void **)(instr + a->instrlen + disp);
437 #else
438 	/* ff 15 00 00 00 00   call   *0x0 */
439 	/* target address is stored at disp. */
440 	target = *(void **)disp;
441 #endif
442 	if (!target)
443 		target = bug;
444 
445 	/* (BUG_func - .) + (target - BUG_func) := target - . */
446 	*(s32 *)(insn_buff + 1) += target - bug;
447 
448 	if (target == &nop_func)
449 		return 0;
450 
451 	return 5;
452 }
453 
454 /*
455  * Replace instructions with better alternatives for this CPU type. This runs
456  * before SMP is initialized to avoid SMP problems with self modifying code.
457  * This implies that asymmetric systems where APs have less capabilities than
458  * the boot processor are not handled. Tough. Make sure you disable such
459  * features by hand.
460  *
461  * Marked "noinline" to cause control flow change and thus insn cache
462  * to refetch changed I$ lines.
463  */
464 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
465 						  struct alt_instr *end)
466 {
467 	struct alt_instr *a;
468 	u8 *instr, *replacement;
469 	u8 insn_buff[MAX_PATCH_LEN];
470 
471 	DPRINTK(ALT, "alt table %px, -> %px", start, end);
472 
473 	/*
474 	 * In the case CONFIG_X86_5LEVEL=y, KASAN_SHADOW_START is defined using
475 	 * cpu_feature_enabled(X86_FEATURE_LA57) and is therefore patched here.
476 	 * During the process, KASAN becomes confused seeing partial LA57
477 	 * conversion and triggers a false-positive out-of-bound report.
478 	 *
479 	 * Disable KASAN until the patching is complete.
480 	 */
481 	kasan_disable_current();
482 
483 	/*
484 	 * The scan order should be from start to end. A later scanned
485 	 * alternative code can overwrite previously scanned alternative code.
486 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
487 	 * patch code.
488 	 *
489 	 * So be careful if you want to change the scan order to any other
490 	 * order.
491 	 */
492 	for (a = start; a < end; a++) {
493 		int insn_buff_sz = 0;
494 
495 		instr = (u8 *)&a->instr_offset + a->instr_offset;
496 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
497 		BUG_ON(a->instrlen > sizeof(insn_buff));
498 		BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
499 
500 		/*
501 		 * Patch if either:
502 		 * - feature is present
503 		 * - feature not present but ALT_FLAG_NOT is set to mean,
504 		 *   patch if feature is *NOT* present.
505 		 */
506 		if (!boot_cpu_has(a->cpuid) == !(a->flags & ALT_FLAG_NOT)) {
507 			optimize_nops_inplace(instr, a->instrlen);
508 			continue;
509 		}
510 
511 		DPRINTK(ALT, "feat: %d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d) flags: 0x%x",
512 			a->cpuid >> 5,
513 			a->cpuid & 0x1f,
514 			instr, instr, a->instrlen,
515 			replacement, a->replacementlen, a->flags);
516 
517 		memcpy(insn_buff, replacement, a->replacementlen);
518 		insn_buff_sz = a->replacementlen;
519 
520 		if (a->flags & ALT_FLAG_DIRECT_CALL) {
521 			insn_buff_sz = alt_replace_call(instr, insn_buff, a);
522 			if (insn_buff_sz < 0)
523 				continue;
524 		}
525 
526 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
527 			insn_buff[insn_buff_sz] = 0x90;
528 
529 		apply_relocation(insn_buff, a->instrlen, instr, replacement, a->replacementlen);
530 
531 		DUMP_BYTES(ALT, instr, a->instrlen, "%px:   old_insn: ", instr);
532 		DUMP_BYTES(ALT, replacement, a->replacementlen, "%px:   rpl_insn: ", replacement);
533 		DUMP_BYTES(ALT, insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
534 
535 		text_poke_early(instr, insn_buff, insn_buff_sz);
536 	}
537 
538 	kasan_enable_current();
539 }
540 
541 static inline bool is_jcc32(struct insn *insn)
542 {
543 	/* Jcc.d32 second opcode byte is in the range: 0x80-0x8f */
544 	return insn->opcode.bytes[0] == 0x0f && (insn->opcode.bytes[1] & 0xf0) == 0x80;
545 }
546 
547 #if defined(CONFIG_RETPOLINE) && defined(CONFIG_OBJTOOL)
548 
549 /*
550  * CALL/JMP *%\reg
551  */
552 static int emit_indirect(int op, int reg, u8 *bytes)
553 {
554 	int i = 0;
555 	u8 modrm;
556 
557 	switch (op) {
558 	case CALL_INSN_OPCODE:
559 		modrm = 0x10; /* Reg = 2; CALL r/m */
560 		break;
561 
562 	case JMP32_INSN_OPCODE:
563 		modrm = 0x20; /* Reg = 4; JMP r/m */
564 		break;
565 
566 	default:
567 		WARN_ON_ONCE(1);
568 		return -1;
569 	}
570 
571 	if (reg >= 8) {
572 		bytes[i++] = 0x41; /* REX.B prefix */
573 		reg -= 8;
574 	}
575 
576 	modrm |= 0xc0; /* Mod = 3 */
577 	modrm += reg;
578 
579 	bytes[i++] = 0xff; /* opcode */
580 	bytes[i++] = modrm;
581 
582 	return i;
583 }
584 
585 static int emit_call_track_retpoline(void *addr, struct insn *insn, int reg, u8 *bytes)
586 {
587 	u8 op = insn->opcode.bytes[0];
588 	int i = 0;
589 
590 	/*
591 	 * Clang does 'weird' Jcc __x86_indirect_thunk_r11 conditional
592 	 * tail-calls. Deal with them.
593 	 */
594 	if (is_jcc32(insn)) {
595 		bytes[i++] = op;
596 		op = insn->opcode.bytes[1];
597 		goto clang_jcc;
598 	}
599 
600 	if (insn->length == 6)
601 		bytes[i++] = 0x2e; /* CS-prefix */
602 
603 	switch (op) {
604 	case CALL_INSN_OPCODE:
605 		__text_gen_insn(bytes+i, op, addr+i,
606 				__x86_indirect_call_thunk_array[reg],
607 				CALL_INSN_SIZE);
608 		i += CALL_INSN_SIZE;
609 		break;
610 
611 	case JMP32_INSN_OPCODE:
612 clang_jcc:
613 		__text_gen_insn(bytes+i, op, addr+i,
614 				__x86_indirect_jump_thunk_array[reg],
615 				JMP32_INSN_SIZE);
616 		i += JMP32_INSN_SIZE;
617 		break;
618 
619 	default:
620 		WARN(1, "%pS %px %*ph\n", addr, addr, 6, addr);
621 		return -1;
622 	}
623 
624 	WARN_ON_ONCE(i != insn->length);
625 
626 	return i;
627 }
628 
629 /*
630  * Rewrite the compiler generated retpoline thunk calls.
631  *
632  * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate
633  * indirect instructions, avoiding the extra indirection.
634  *
635  * For example, convert:
636  *
637  *   CALL __x86_indirect_thunk_\reg
638  *
639  * into:
640  *
641  *   CALL *%\reg
642  *
643  * It also tries to inline spectre_v2=retpoline,lfence when size permits.
644  */
645 static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes)
646 {
647 	retpoline_thunk_t *target;
648 	int reg, ret, i = 0;
649 	u8 op, cc;
650 
651 	target = addr + insn->length + insn->immediate.value;
652 	reg = target - __x86_indirect_thunk_array;
653 
654 	if (WARN_ON_ONCE(reg & ~0xf))
655 		return -1;
656 
657 	/* If anyone ever does: CALL/JMP *%rsp, we're in deep trouble. */
658 	BUG_ON(reg == 4);
659 
660 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE) &&
661 	    !cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
662 		if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
663 			return emit_call_track_retpoline(addr, insn, reg, bytes);
664 
665 		return -1;
666 	}
667 
668 	op = insn->opcode.bytes[0];
669 
670 	/*
671 	 * Convert:
672 	 *
673 	 *   Jcc.d32 __x86_indirect_thunk_\reg
674 	 *
675 	 * into:
676 	 *
677 	 *   Jncc.d8 1f
678 	 *   [ LFENCE ]
679 	 *   JMP *%\reg
680 	 *   [ NOP ]
681 	 * 1:
682 	 */
683 	if (is_jcc32(insn)) {
684 		cc = insn->opcode.bytes[1] & 0xf;
685 		cc ^= 1; /* invert condition */
686 
687 		bytes[i++] = 0x70 + cc;        /* Jcc.d8 */
688 		bytes[i++] = insn->length - 2; /* sizeof(Jcc.d8) == 2 */
689 
690 		/* Continue as if: JMP.d32 __x86_indirect_thunk_\reg */
691 		op = JMP32_INSN_OPCODE;
692 	}
693 
694 	/*
695 	 * For RETPOLINE_LFENCE: prepend the indirect CALL/JMP with an LFENCE.
696 	 */
697 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
698 		bytes[i++] = 0x0f;
699 		bytes[i++] = 0xae;
700 		bytes[i++] = 0xe8; /* LFENCE */
701 	}
702 
703 	ret = emit_indirect(op, reg, bytes + i);
704 	if (ret < 0)
705 		return ret;
706 	i += ret;
707 
708 	/*
709 	 * The compiler is supposed to EMIT an INT3 after every unconditional
710 	 * JMP instruction due to AMD BTC. However, if the compiler is too old
711 	 * or SLS isn't enabled, we still need an INT3 after indirect JMPs
712 	 * even on Intel.
713 	 */
714 	if (op == JMP32_INSN_OPCODE && i < insn->length)
715 		bytes[i++] = INT3_INSN_OPCODE;
716 
717 	for (; i < insn->length;)
718 		bytes[i++] = BYTES_NOP1;
719 
720 	return i;
721 }
722 
723 /*
724  * Generated by 'objtool --retpoline'.
725  */
726 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end)
727 {
728 	s32 *s;
729 
730 	for (s = start; s < end; s++) {
731 		void *addr = (void *)s + *s;
732 		struct insn insn;
733 		int len, ret;
734 		u8 bytes[16];
735 		u8 op1, op2;
736 
737 		ret = insn_decode_kernel(&insn, addr);
738 		if (WARN_ON_ONCE(ret < 0))
739 			continue;
740 
741 		op1 = insn.opcode.bytes[0];
742 		op2 = insn.opcode.bytes[1];
743 
744 		switch (op1) {
745 		case CALL_INSN_OPCODE:
746 		case JMP32_INSN_OPCODE:
747 			break;
748 
749 		case 0x0f: /* escape */
750 			if (op2 >= 0x80 && op2 <= 0x8f)
751 				break;
752 			fallthrough;
753 		default:
754 			WARN_ON_ONCE(1);
755 			continue;
756 		}
757 
758 		DPRINTK(RETPOLINE, "retpoline at: %pS (%px) len: %d to: %pS",
759 			addr, addr, insn.length,
760 			addr + insn.length + insn.immediate.value);
761 
762 		len = patch_retpoline(addr, &insn, bytes);
763 		if (len == insn.length) {
764 			optimize_nops(bytes, len);
765 			DUMP_BYTES(RETPOLINE, ((u8*)addr),  len, "%px: orig: ", addr);
766 			DUMP_BYTES(RETPOLINE, ((u8*)bytes), len, "%px: repl: ", addr);
767 			text_poke_early(addr, bytes, len);
768 		}
769 	}
770 }
771 
772 #ifdef CONFIG_RETHUNK
773 
774 /*
775  * Rewrite the compiler generated return thunk tail-calls.
776  *
777  * For example, convert:
778  *
779  *   JMP __x86_return_thunk
780  *
781  * into:
782  *
783  *   RET
784  */
785 static int patch_return(void *addr, struct insn *insn, u8 *bytes)
786 {
787 	int i = 0;
788 
789 	/* Patch the custom return thunks... */
790 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
791 		i = JMP32_INSN_SIZE;
792 		__text_gen_insn(bytes, JMP32_INSN_OPCODE, addr, x86_return_thunk, i);
793 	} else {
794 		/* ... or patch them out if not needed. */
795 		bytes[i++] = RET_INSN_OPCODE;
796 	}
797 
798 	for (; i < insn->length;)
799 		bytes[i++] = INT3_INSN_OPCODE;
800 	return i;
801 }
802 
803 void __init_or_module noinline apply_returns(s32 *start, s32 *end)
804 {
805 	s32 *s;
806 
807 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
808 		static_call_force_reinit();
809 
810 	for (s = start; s < end; s++) {
811 		void *dest = NULL, *addr = (void *)s + *s;
812 		struct insn insn;
813 		int len, ret;
814 		u8 bytes[16];
815 		u8 op;
816 
817 		ret = insn_decode_kernel(&insn, addr);
818 		if (WARN_ON_ONCE(ret < 0))
819 			continue;
820 
821 		op = insn.opcode.bytes[0];
822 		if (op == JMP32_INSN_OPCODE)
823 			dest = addr + insn.length + insn.immediate.value;
824 
825 		if (__static_call_fixup(addr, op, dest) ||
826 		    WARN_ONCE(dest != &__x86_return_thunk,
827 			      "missing return thunk: %pS-%pS: %*ph",
828 			      addr, dest, 5, addr))
829 			continue;
830 
831 		DPRINTK(RET, "return thunk at: %pS (%px) len: %d to: %pS",
832 			addr, addr, insn.length,
833 			addr + insn.length + insn.immediate.value);
834 
835 		len = patch_return(addr, &insn, bytes);
836 		if (len == insn.length) {
837 			DUMP_BYTES(RET, ((u8*)addr),  len, "%px: orig: ", addr);
838 			DUMP_BYTES(RET, ((u8*)bytes), len, "%px: repl: ", addr);
839 			text_poke_early(addr, bytes, len);
840 		}
841 	}
842 }
843 #else
844 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
845 #endif /* CONFIG_RETHUNK */
846 
847 #else /* !CONFIG_RETPOLINE || !CONFIG_OBJTOOL */
848 
849 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) { }
850 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
851 
852 #endif /* CONFIG_RETPOLINE && CONFIG_OBJTOOL */
853 
854 #ifdef CONFIG_X86_KERNEL_IBT
855 
856 static void poison_cfi(void *addr);
857 
858 static void __init_or_module poison_endbr(void *addr, bool warn)
859 {
860 	u32 endbr, poison = gen_endbr_poison();
861 
862 	if (WARN_ON_ONCE(get_kernel_nofault(endbr, addr)))
863 		return;
864 
865 	if (!is_endbr(endbr)) {
866 		WARN_ON_ONCE(warn);
867 		return;
868 	}
869 
870 	DPRINTK(ENDBR, "ENDBR at: %pS (%px)", addr, addr);
871 
872 	/*
873 	 * When we have IBT, the lack of ENDBR will trigger #CP
874 	 */
875 	DUMP_BYTES(ENDBR, ((u8*)addr), 4, "%px: orig: ", addr);
876 	DUMP_BYTES(ENDBR, ((u8*)&poison), 4, "%px: repl: ", addr);
877 	text_poke_early(addr, &poison, 4);
878 }
879 
880 /*
881  * Generated by: objtool --ibt
882  *
883  * Seal the functions for indirect calls by clobbering the ENDBR instructions
884  * and the kCFI hash value.
885  */
886 void __init_or_module noinline apply_seal_endbr(s32 *start, s32 *end)
887 {
888 	s32 *s;
889 
890 	for (s = start; s < end; s++) {
891 		void *addr = (void *)s + *s;
892 
893 		poison_endbr(addr, true);
894 		if (IS_ENABLED(CONFIG_FINEIBT))
895 			poison_cfi(addr - 16);
896 	}
897 }
898 
899 #else
900 
901 void __init_or_module apply_seal_endbr(s32 *start, s32 *end) { }
902 
903 #endif /* CONFIG_X86_KERNEL_IBT */
904 
905 #ifdef CONFIG_FINEIBT
906 
907 enum cfi_mode {
908 	CFI_DEFAULT,
909 	CFI_OFF,
910 	CFI_KCFI,
911 	CFI_FINEIBT,
912 };
913 
914 static enum cfi_mode cfi_mode __ro_after_init = CFI_DEFAULT;
915 static bool cfi_rand __ro_after_init = true;
916 static u32  cfi_seed __ro_after_init;
917 
918 /*
919  * Re-hash the CFI hash with a boot-time seed while making sure the result is
920  * not a valid ENDBR instruction.
921  */
922 static u32 cfi_rehash(u32 hash)
923 {
924 	hash ^= cfi_seed;
925 	while (unlikely(is_endbr(hash) || is_endbr(-hash))) {
926 		bool lsb = hash & 1;
927 		hash >>= 1;
928 		if (lsb)
929 			hash ^= 0x80200003;
930 	}
931 	return hash;
932 }
933 
934 static __init int cfi_parse_cmdline(char *str)
935 {
936 	if (!str)
937 		return -EINVAL;
938 
939 	while (str) {
940 		char *next = strchr(str, ',');
941 		if (next) {
942 			*next = 0;
943 			next++;
944 		}
945 
946 		if (!strcmp(str, "auto")) {
947 			cfi_mode = CFI_DEFAULT;
948 		} else if (!strcmp(str, "off")) {
949 			cfi_mode = CFI_OFF;
950 			cfi_rand = false;
951 		} else if (!strcmp(str, "kcfi")) {
952 			cfi_mode = CFI_KCFI;
953 		} else if (!strcmp(str, "fineibt")) {
954 			cfi_mode = CFI_FINEIBT;
955 		} else if (!strcmp(str, "norand")) {
956 			cfi_rand = false;
957 		} else {
958 			pr_err("Ignoring unknown cfi option (%s).", str);
959 		}
960 
961 		str = next;
962 	}
963 
964 	return 0;
965 }
966 early_param("cfi", cfi_parse_cmdline);
967 
968 /*
969  * kCFI						FineIBT
970  *
971  * __cfi_\func:					__cfi_\func:
972  *	movl   $0x12345678,%eax		// 5	     endbr64			// 4
973  *	nop					     subl   $0x12345678,%r10d   // 7
974  *	nop					     jz     1f			// 2
975  *	nop					     ud2			// 2
976  *	nop					1:   nop			// 1
977  *	nop
978  *	nop
979  *	nop
980  *	nop
981  *	nop
982  *	nop
983  *	nop
984  *
985  *
986  * caller:					caller:
987  *	movl	$(-0x12345678),%r10d	 // 6	     movl   $0x12345678,%r10d	// 6
988  *	addl	$-15(%r11),%r10d	 // 4	     sub    $16,%r11		// 4
989  *	je	1f			 // 2	     nop4			// 4
990  *	ud2				 // 2
991  * 1:	call	__x86_indirect_thunk_r11 // 5	     call   *%r11; nop2;	// 5
992  *
993  */
994 
995 asm(	".pushsection .rodata			\n"
996 	"fineibt_preamble_start:		\n"
997 	"	endbr64				\n"
998 	"	subl	$0x12345678, %r10d	\n"
999 	"	je	fineibt_preamble_end	\n"
1000 	"	ud2				\n"
1001 	"	nop				\n"
1002 	"fineibt_preamble_end:			\n"
1003 	".popsection\n"
1004 );
1005 
1006 extern u8 fineibt_preamble_start[];
1007 extern u8 fineibt_preamble_end[];
1008 
1009 #define fineibt_preamble_size (fineibt_preamble_end - fineibt_preamble_start)
1010 #define fineibt_preamble_hash 7
1011 
1012 asm(	".pushsection .rodata			\n"
1013 	"fineibt_caller_start:			\n"
1014 	"	movl	$0x12345678, %r10d	\n"
1015 	"	sub	$16, %r11		\n"
1016 	ASM_NOP4
1017 	"fineibt_caller_end:			\n"
1018 	".popsection				\n"
1019 );
1020 
1021 extern u8 fineibt_caller_start[];
1022 extern u8 fineibt_caller_end[];
1023 
1024 #define fineibt_caller_size (fineibt_caller_end - fineibt_caller_start)
1025 #define fineibt_caller_hash 2
1026 
1027 #define fineibt_caller_jmp (fineibt_caller_size - 2)
1028 
1029 static u32 decode_preamble_hash(void *addr)
1030 {
1031 	u8 *p = addr;
1032 
1033 	/* b8 78 56 34 12          mov    $0x12345678,%eax */
1034 	if (p[0] == 0xb8)
1035 		return *(u32 *)(addr + 1);
1036 
1037 	return 0; /* invalid hash value */
1038 }
1039 
1040 static u32 decode_caller_hash(void *addr)
1041 {
1042 	u8 *p = addr;
1043 
1044 	/* 41 ba 78 56 34 12       mov    $0x12345678,%r10d */
1045 	if (p[0] == 0x41 && p[1] == 0xba)
1046 		return -*(u32 *)(addr + 2);
1047 
1048 	/* e8 0c 78 56 34 12	   jmp.d8  +12 */
1049 	if (p[0] == JMP8_INSN_OPCODE && p[1] == fineibt_caller_jmp)
1050 		return -*(u32 *)(addr + 2);
1051 
1052 	return 0; /* invalid hash value */
1053 }
1054 
1055 /* .retpoline_sites */
1056 static int cfi_disable_callers(s32 *start, s32 *end)
1057 {
1058 	/*
1059 	 * Disable kCFI by patching in a JMP.d8, this leaves the hash immediate
1060 	 * in tact for later usage. Also see decode_caller_hash() and
1061 	 * cfi_rewrite_callers().
1062 	 */
1063 	const u8 jmp[] = { JMP8_INSN_OPCODE, fineibt_caller_jmp };
1064 	s32 *s;
1065 
1066 	for (s = start; s < end; s++) {
1067 		void *addr = (void *)s + *s;
1068 		u32 hash;
1069 
1070 		addr -= fineibt_caller_size;
1071 		hash = decode_caller_hash(addr);
1072 		if (!hash) /* nocfi callers */
1073 			continue;
1074 
1075 		text_poke_early(addr, jmp, 2);
1076 	}
1077 
1078 	return 0;
1079 }
1080 
1081 static int cfi_enable_callers(s32 *start, s32 *end)
1082 {
1083 	/*
1084 	 * Re-enable kCFI, undo what cfi_disable_callers() did.
1085 	 */
1086 	const u8 mov[] = { 0x41, 0xba };
1087 	s32 *s;
1088 
1089 	for (s = start; s < end; s++) {
1090 		void *addr = (void *)s + *s;
1091 		u32 hash;
1092 
1093 		addr -= fineibt_caller_size;
1094 		hash = decode_caller_hash(addr);
1095 		if (!hash) /* nocfi callers */
1096 			continue;
1097 
1098 		text_poke_early(addr, mov, 2);
1099 	}
1100 
1101 	return 0;
1102 }
1103 
1104 /* .cfi_sites */
1105 static int cfi_rand_preamble(s32 *start, s32 *end)
1106 {
1107 	s32 *s;
1108 
1109 	for (s = start; s < end; s++) {
1110 		void *addr = (void *)s + *s;
1111 		u32 hash;
1112 
1113 		hash = decode_preamble_hash(addr);
1114 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1115 			 addr, addr, 5, addr))
1116 			return -EINVAL;
1117 
1118 		hash = cfi_rehash(hash);
1119 		text_poke_early(addr + 1, &hash, 4);
1120 	}
1121 
1122 	return 0;
1123 }
1124 
1125 static int cfi_rewrite_preamble(s32 *start, s32 *end)
1126 {
1127 	s32 *s;
1128 
1129 	for (s = start; s < end; s++) {
1130 		void *addr = (void *)s + *s;
1131 		u32 hash;
1132 
1133 		hash = decode_preamble_hash(addr);
1134 		if (WARN(!hash, "no CFI hash found at: %pS %px %*ph\n",
1135 			 addr, addr, 5, addr))
1136 			return -EINVAL;
1137 
1138 		text_poke_early(addr, fineibt_preamble_start, fineibt_preamble_size);
1139 		WARN_ON(*(u32 *)(addr + fineibt_preamble_hash) != 0x12345678);
1140 		text_poke_early(addr + fineibt_preamble_hash, &hash, 4);
1141 	}
1142 
1143 	return 0;
1144 }
1145 
1146 static void cfi_rewrite_endbr(s32 *start, s32 *end)
1147 {
1148 	s32 *s;
1149 
1150 	for (s = start; s < end; s++) {
1151 		void *addr = (void *)s + *s;
1152 
1153 		poison_endbr(addr+16, false);
1154 	}
1155 }
1156 
1157 /* .retpoline_sites */
1158 static int cfi_rand_callers(s32 *start, s32 *end)
1159 {
1160 	s32 *s;
1161 
1162 	for (s = start; s < end; s++) {
1163 		void *addr = (void *)s + *s;
1164 		u32 hash;
1165 
1166 		addr -= fineibt_caller_size;
1167 		hash = decode_caller_hash(addr);
1168 		if (hash) {
1169 			hash = -cfi_rehash(hash);
1170 			text_poke_early(addr + 2, &hash, 4);
1171 		}
1172 	}
1173 
1174 	return 0;
1175 }
1176 
1177 static int cfi_rewrite_callers(s32 *start, s32 *end)
1178 {
1179 	s32 *s;
1180 
1181 	for (s = start; s < end; s++) {
1182 		void *addr = (void *)s + *s;
1183 		u32 hash;
1184 
1185 		addr -= fineibt_caller_size;
1186 		hash = decode_caller_hash(addr);
1187 		if (hash) {
1188 			text_poke_early(addr, fineibt_caller_start, fineibt_caller_size);
1189 			WARN_ON(*(u32 *)(addr + fineibt_caller_hash) != 0x12345678);
1190 			text_poke_early(addr + fineibt_caller_hash, &hash, 4);
1191 		}
1192 		/* rely on apply_retpolines() */
1193 	}
1194 
1195 	return 0;
1196 }
1197 
1198 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1199 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1200 {
1201 	int ret;
1202 
1203 	if (WARN_ONCE(fineibt_preamble_size != 16,
1204 		      "FineIBT preamble wrong size: %ld", fineibt_preamble_size))
1205 		return;
1206 
1207 	if (cfi_mode == CFI_DEFAULT) {
1208 		cfi_mode = CFI_KCFI;
1209 		if (HAS_KERNEL_IBT && cpu_feature_enabled(X86_FEATURE_IBT))
1210 			cfi_mode = CFI_FINEIBT;
1211 	}
1212 
1213 	/*
1214 	 * Rewrite the callers to not use the __cfi_ stubs, such that we might
1215 	 * rewrite them. This disables all CFI. If this succeeds but any of the
1216 	 * later stages fails, we're without CFI.
1217 	 */
1218 	ret = cfi_disable_callers(start_retpoline, end_retpoline);
1219 	if (ret)
1220 		goto err;
1221 
1222 	if (cfi_rand) {
1223 		if (builtin)
1224 			cfi_seed = get_random_u32();
1225 
1226 		ret = cfi_rand_preamble(start_cfi, end_cfi);
1227 		if (ret)
1228 			goto err;
1229 
1230 		ret = cfi_rand_callers(start_retpoline, end_retpoline);
1231 		if (ret)
1232 			goto err;
1233 	}
1234 
1235 	switch (cfi_mode) {
1236 	case CFI_OFF:
1237 		if (builtin)
1238 			pr_info("Disabling CFI\n");
1239 		return;
1240 
1241 	case CFI_KCFI:
1242 		ret = cfi_enable_callers(start_retpoline, end_retpoline);
1243 		if (ret)
1244 			goto err;
1245 
1246 		if (builtin)
1247 			pr_info("Using kCFI\n");
1248 		return;
1249 
1250 	case CFI_FINEIBT:
1251 		/* place the FineIBT preamble at func()-16 */
1252 		ret = cfi_rewrite_preamble(start_cfi, end_cfi);
1253 		if (ret)
1254 			goto err;
1255 
1256 		/* rewrite the callers to target func()-16 */
1257 		ret = cfi_rewrite_callers(start_retpoline, end_retpoline);
1258 		if (ret)
1259 			goto err;
1260 
1261 		/* now that nobody targets func()+0, remove ENDBR there */
1262 		cfi_rewrite_endbr(start_cfi, end_cfi);
1263 
1264 		if (builtin)
1265 			pr_info("Using FineIBT CFI\n");
1266 		return;
1267 
1268 	default:
1269 		break;
1270 	}
1271 
1272 err:
1273 	pr_err("Something went horribly wrong trying to rewrite the CFI implementation.\n");
1274 }
1275 
1276 static inline void poison_hash(void *addr)
1277 {
1278 	*(u32 *)addr = 0;
1279 }
1280 
1281 static void poison_cfi(void *addr)
1282 {
1283 	switch (cfi_mode) {
1284 	case CFI_FINEIBT:
1285 		/*
1286 		 * __cfi_\func:
1287 		 *	osp nopl (%rax)
1288 		 *	subl	$0, %r10d
1289 		 *	jz	1f
1290 		 *	ud2
1291 		 * 1:	nop
1292 		 */
1293 		poison_endbr(addr, false);
1294 		poison_hash(addr + fineibt_preamble_hash);
1295 		break;
1296 
1297 	case CFI_KCFI:
1298 		/*
1299 		 * __cfi_\func:
1300 		 *	movl	$0, %eax
1301 		 *	.skip	11, 0x90
1302 		 */
1303 		poison_hash(addr + 1);
1304 		break;
1305 
1306 	default:
1307 		break;
1308 	}
1309 }
1310 
1311 #else
1312 
1313 static void __apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1314 			    s32 *start_cfi, s32 *end_cfi, bool builtin)
1315 {
1316 }
1317 
1318 #ifdef CONFIG_X86_KERNEL_IBT
1319 static void poison_cfi(void *addr) { }
1320 #endif
1321 
1322 #endif
1323 
1324 void apply_fineibt(s32 *start_retpoline, s32 *end_retpoline,
1325 		   s32 *start_cfi, s32 *end_cfi)
1326 {
1327 	return __apply_fineibt(start_retpoline, end_retpoline,
1328 			       start_cfi, end_cfi,
1329 			       /* .builtin = */ false);
1330 }
1331 
1332 #ifdef CONFIG_SMP
1333 static void alternatives_smp_lock(const s32 *start, const s32 *end,
1334 				  u8 *text, u8 *text_end)
1335 {
1336 	const s32 *poff;
1337 
1338 	for (poff = start; poff < end; poff++) {
1339 		u8 *ptr = (u8 *)poff + *poff;
1340 
1341 		if (!*poff || ptr < text || ptr >= text_end)
1342 			continue;
1343 		/* turn DS segment override prefix into lock prefix */
1344 		if (*ptr == 0x3e)
1345 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
1346 	}
1347 }
1348 
1349 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
1350 				    u8 *text, u8 *text_end)
1351 {
1352 	const s32 *poff;
1353 
1354 	for (poff = start; poff < end; poff++) {
1355 		u8 *ptr = (u8 *)poff + *poff;
1356 
1357 		if (!*poff || ptr < text || ptr >= text_end)
1358 			continue;
1359 		/* turn lock prefix into DS segment override prefix */
1360 		if (*ptr == 0xf0)
1361 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
1362 	}
1363 }
1364 
1365 struct smp_alt_module {
1366 	/* what is this ??? */
1367 	struct module	*mod;
1368 	char		*name;
1369 
1370 	/* ptrs to lock prefixes */
1371 	const s32	*locks;
1372 	const s32	*locks_end;
1373 
1374 	/* .text segment, needed to avoid patching init code ;) */
1375 	u8		*text;
1376 	u8		*text_end;
1377 
1378 	struct list_head next;
1379 };
1380 static LIST_HEAD(smp_alt_modules);
1381 static bool uniproc_patched = false;	/* protected by text_mutex */
1382 
1383 void __init_or_module alternatives_smp_module_add(struct module *mod,
1384 						  char *name,
1385 						  void *locks, void *locks_end,
1386 						  void *text,  void *text_end)
1387 {
1388 	struct smp_alt_module *smp;
1389 
1390 	mutex_lock(&text_mutex);
1391 	if (!uniproc_patched)
1392 		goto unlock;
1393 
1394 	if (num_possible_cpus() == 1)
1395 		/* Don't bother remembering, we'll never have to undo it. */
1396 		goto smp_unlock;
1397 
1398 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
1399 	if (NULL == smp)
1400 		/* we'll run the (safe but slow) SMP code then ... */
1401 		goto unlock;
1402 
1403 	smp->mod	= mod;
1404 	smp->name	= name;
1405 	smp->locks	= locks;
1406 	smp->locks_end	= locks_end;
1407 	smp->text	= text;
1408 	smp->text_end	= text_end;
1409 	DPRINTK(SMP, "locks %p -> %p, text %p -> %p, name %s\n",
1410 		smp->locks, smp->locks_end,
1411 		smp->text, smp->text_end, smp->name);
1412 
1413 	list_add_tail(&smp->next, &smp_alt_modules);
1414 smp_unlock:
1415 	alternatives_smp_unlock(locks, locks_end, text, text_end);
1416 unlock:
1417 	mutex_unlock(&text_mutex);
1418 }
1419 
1420 void __init_or_module alternatives_smp_module_del(struct module *mod)
1421 {
1422 	struct smp_alt_module *item;
1423 
1424 	mutex_lock(&text_mutex);
1425 	list_for_each_entry(item, &smp_alt_modules, next) {
1426 		if (mod != item->mod)
1427 			continue;
1428 		list_del(&item->next);
1429 		kfree(item);
1430 		break;
1431 	}
1432 	mutex_unlock(&text_mutex);
1433 }
1434 
1435 void alternatives_enable_smp(void)
1436 {
1437 	struct smp_alt_module *mod;
1438 
1439 	/* Why bother if there are no other CPUs? */
1440 	BUG_ON(num_possible_cpus() == 1);
1441 
1442 	mutex_lock(&text_mutex);
1443 
1444 	if (uniproc_patched) {
1445 		pr_info("switching to SMP code\n");
1446 		BUG_ON(num_online_cpus() != 1);
1447 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
1448 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
1449 		list_for_each_entry(mod, &smp_alt_modules, next)
1450 			alternatives_smp_lock(mod->locks, mod->locks_end,
1451 					      mod->text, mod->text_end);
1452 		uniproc_patched = false;
1453 	}
1454 	mutex_unlock(&text_mutex);
1455 }
1456 
1457 /*
1458  * Return 1 if the address range is reserved for SMP-alternatives.
1459  * Must hold text_mutex.
1460  */
1461 int alternatives_text_reserved(void *start, void *end)
1462 {
1463 	struct smp_alt_module *mod;
1464 	const s32 *poff;
1465 	u8 *text_start = start;
1466 	u8 *text_end = end;
1467 
1468 	lockdep_assert_held(&text_mutex);
1469 
1470 	list_for_each_entry(mod, &smp_alt_modules, next) {
1471 		if (mod->text > text_end || mod->text_end < text_start)
1472 			continue;
1473 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
1474 			const u8 *ptr = (const u8 *)poff + *poff;
1475 
1476 			if (text_start <= ptr && text_end > ptr)
1477 				return 1;
1478 		}
1479 	}
1480 
1481 	return 0;
1482 }
1483 #endif /* CONFIG_SMP */
1484 
1485 /*
1486  * Self-test for the INT3 based CALL emulation code.
1487  *
1488  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
1489  * properly and that there is a stack gap between the INT3 frame and the
1490  * previous context. Without this gap doing a virtual PUSH on the interrupted
1491  * stack would corrupt the INT3 IRET frame.
1492  *
1493  * See entry_{32,64}.S for more details.
1494  */
1495 
1496 /*
1497  * We define the int3_magic() function in assembly to control the calling
1498  * convention such that we can 'call' it from assembly.
1499  */
1500 
1501 extern void int3_magic(unsigned int *ptr); /* defined in asm */
1502 
1503 asm (
1504 "	.pushsection	.init.text, \"ax\", @progbits\n"
1505 "	.type		int3_magic, @function\n"
1506 "int3_magic:\n"
1507 	ANNOTATE_NOENDBR
1508 "	movl	$1, (%" _ASM_ARG1 ")\n"
1509 	ASM_RET
1510 "	.size		int3_magic, .-int3_magic\n"
1511 "	.popsection\n"
1512 );
1513 
1514 extern void int3_selftest_ip(void); /* defined in asm below */
1515 
1516 static int __init
1517 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
1518 {
1519 	unsigned long selftest = (unsigned long)&int3_selftest_ip;
1520 	struct die_args *args = data;
1521 	struct pt_regs *regs = args->regs;
1522 
1523 	OPTIMIZER_HIDE_VAR(selftest);
1524 
1525 	if (!regs || user_mode(regs))
1526 		return NOTIFY_DONE;
1527 
1528 	if (val != DIE_INT3)
1529 		return NOTIFY_DONE;
1530 
1531 	if (regs->ip - INT3_INSN_SIZE != selftest)
1532 		return NOTIFY_DONE;
1533 
1534 	int3_emulate_call(regs, (unsigned long)&int3_magic);
1535 	return NOTIFY_STOP;
1536 }
1537 
1538 /* Must be noinline to ensure uniqueness of int3_selftest_ip. */
1539 static noinline void __init int3_selftest(void)
1540 {
1541 	static __initdata struct notifier_block int3_exception_nb = {
1542 		.notifier_call	= int3_exception_notify,
1543 		.priority	= INT_MAX-1, /* last */
1544 	};
1545 	unsigned int val = 0;
1546 
1547 	BUG_ON(register_die_notifier(&int3_exception_nb));
1548 
1549 	/*
1550 	 * Basically: int3_magic(&val); but really complicated :-)
1551 	 *
1552 	 * INT3 padded with NOP to CALL_INSN_SIZE. The int3_exception_nb
1553 	 * notifier above will emulate CALL for us.
1554 	 */
1555 	asm volatile ("int3_selftest_ip:\n\t"
1556 		      ANNOTATE_NOENDBR
1557 		      "    int3; nop; nop; nop; nop\n\t"
1558 		      : ASM_CALL_CONSTRAINT
1559 		      : __ASM_SEL_RAW(a, D) (&val)
1560 		      : "memory");
1561 
1562 	BUG_ON(val != 1);
1563 
1564 	unregister_die_notifier(&int3_exception_nb);
1565 }
1566 
1567 static __initdata int __alt_reloc_selftest_addr;
1568 
1569 extern void __init __alt_reloc_selftest(void *arg);
1570 __visible noinline void __init __alt_reloc_selftest(void *arg)
1571 {
1572 	WARN_ON(arg != &__alt_reloc_selftest_addr);
1573 }
1574 
1575 static noinline void __init alt_reloc_selftest(void)
1576 {
1577 	/*
1578 	 * Tests apply_relocation().
1579 	 *
1580 	 * This has a relative immediate (CALL) in a place other than the first
1581 	 * instruction and additionally on x86_64 we get a RIP-relative LEA:
1582 	 *
1583 	 *   lea    0x0(%rip),%rdi  # 5d0: R_X86_64_PC32    .init.data+0x5566c
1584 	 *   call   +0              # 5d5: R_X86_64_PLT32   __alt_reloc_selftest-0x4
1585 	 *
1586 	 * Getting this wrong will either crash and burn or tickle the WARN
1587 	 * above.
1588 	 */
1589 	asm_inline volatile (
1590 		ALTERNATIVE("", "lea %[mem], %%" _ASM_ARG1 "; call __alt_reloc_selftest;", X86_FEATURE_ALWAYS)
1591 		: /* output */
1592 		: [mem] "m" (__alt_reloc_selftest_addr)
1593 		: _ASM_ARG1
1594 	);
1595 }
1596 
1597 void __init alternative_instructions(void)
1598 {
1599 	int3_selftest();
1600 
1601 	/*
1602 	 * The patching is not fully atomic, so try to avoid local
1603 	 * interruptions that might execute the to be patched code.
1604 	 * Other CPUs are not running.
1605 	 */
1606 	stop_nmi();
1607 
1608 	/*
1609 	 * Don't stop machine check exceptions while patching.
1610 	 * MCEs only happen when something got corrupted and in this
1611 	 * case we must do something about the corruption.
1612 	 * Ignoring it is worse than an unlikely patching race.
1613 	 * Also machine checks tend to be broadcast and if one CPU
1614 	 * goes into machine check the others follow quickly, so we don't
1615 	 * expect a machine check to cause undue problems during to code
1616 	 * patching.
1617 	 */
1618 
1619 	/*
1620 	 * Make sure to set (artificial) features depending on used paravirt
1621 	 * functions which can later influence alternative patching.
1622 	 */
1623 	paravirt_set_cap();
1624 
1625 	__apply_fineibt(__retpoline_sites, __retpoline_sites_end,
1626 			__cfi_sites, __cfi_sites_end, true);
1627 
1628 	/*
1629 	 * Rewrite the retpolines, must be done before alternatives since
1630 	 * those can rewrite the retpoline thunks.
1631 	 */
1632 	apply_retpolines(__retpoline_sites, __retpoline_sites_end);
1633 	apply_returns(__return_sites, __return_sites_end);
1634 
1635 	apply_alternatives(__alt_instructions, __alt_instructions_end);
1636 
1637 	/*
1638 	 * Now all calls are established. Apply the call thunks if
1639 	 * required.
1640 	 */
1641 	callthunks_patch_builtin_calls();
1642 
1643 	/*
1644 	 * Seal all functions that do not have their address taken.
1645 	 */
1646 	apply_seal_endbr(__ibt_endbr_seal, __ibt_endbr_seal_end);
1647 
1648 #ifdef CONFIG_SMP
1649 	/* Patch to UP if other cpus not imminent. */
1650 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
1651 		uniproc_patched = true;
1652 		alternatives_smp_module_add(NULL, "core kernel",
1653 					    __smp_locks, __smp_locks_end,
1654 					    _text, _etext);
1655 	}
1656 
1657 	if (!uniproc_patched || num_possible_cpus() == 1) {
1658 		free_init_pages("SMP alternatives",
1659 				(unsigned long)__smp_locks,
1660 				(unsigned long)__smp_locks_end);
1661 	}
1662 #endif
1663 
1664 	restart_nmi();
1665 	alternatives_patched = 1;
1666 
1667 	alt_reloc_selftest();
1668 }
1669 
1670 /**
1671  * text_poke_early - Update instructions on a live kernel at boot time
1672  * @addr: address to modify
1673  * @opcode: source of the copy
1674  * @len: length to copy
1675  *
1676  * When you use this code to patch more than one byte of an instruction
1677  * you need to make sure that other CPUs cannot execute this code in parallel.
1678  * Also no thread must be currently preempted in the middle of these
1679  * instructions. And on the local CPU you need to be protected against NMI or
1680  * MCE handlers seeing an inconsistent instruction while you patch.
1681  */
1682 void __init_or_module text_poke_early(void *addr, const void *opcode,
1683 				      size_t len)
1684 {
1685 	unsigned long flags;
1686 
1687 	if (boot_cpu_has(X86_FEATURE_NX) &&
1688 	    is_module_text_address((unsigned long)addr)) {
1689 		/*
1690 		 * Modules text is marked initially as non-executable, so the
1691 		 * code cannot be running and speculative code-fetches are
1692 		 * prevented. Just change the code.
1693 		 */
1694 		memcpy(addr, opcode, len);
1695 	} else {
1696 		local_irq_save(flags);
1697 		memcpy(addr, opcode, len);
1698 		sync_core();
1699 		local_irq_restore(flags);
1700 
1701 		/*
1702 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
1703 		 * that causes hangs on some VIA CPUs.
1704 		 */
1705 	}
1706 }
1707 
1708 typedef struct {
1709 	struct mm_struct *mm;
1710 } temp_mm_state_t;
1711 
1712 /*
1713  * Using a temporary mm allows to set temporary mappings that are not accessible
1714  * by other CPUs. Such mappings are needed to perform sensitive memory writes
1715  * that override the kernel memory protections (e.g., W^X), without exposing the
1716  * temporary page-table mappings that are required for these write operations to
1717  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
1718  * mapping is torn down.
1719  *
1720  * Context: The temporary mm needs to be used exclusively by a single core. To
1721  *          harden security IRQs must be disabled while the temporary mm is
1722  *          loaded, thereby preventing interrupt handler bugs from overriding
1723  *          the kernel memory protection.
1724  */
1725 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
1726 {
1727 	temp_mm_state_t temp_state;
1728 
1729 	lockdep_assert_irqs_disabled();
1730 
1731 	/*
1732 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
1733 	 * with a stale address space WITHOUT being in lazy mode after
1734 	 * restoring the previous mm.
1735 	 */
1736 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
1737 		leave_mm(smp_processor_id());
1738 
1739 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1740 	switch_mm_irqs_off(NULL, mm, current);
1741 
1742 	/*
1743 	 * If breakpoints are enabled, disable them while the temporary mm is
1744 	 * used. Userspace might set up watchpoints on addresses that are used
1745 	 * in the temporary mm, which would lead to wrong signals being sent or
1746 	 * crashes.
1747 	 *
1748 	 * Note that breakpoints are not disabled selectively, which also causes
1749 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
1750 	 * undesirable, but still seems reasonable as the code that runs in the
1751 	 * temporary mm should be short.
1752 	 */
1753 	if (hw_breakpoint_active())
1754 		hw_breakpoint_disable();
1755 
1756 	return temp_state;
1757 }
1758 
1759 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
1760 {
1761 	lockdep_assert_irqs_disabled();
1762 	switch_mm_irqs_off(NULL, prev_state.mm, current);
1763 
1764 	/*
1765 	 * Restore the breakpoints if they were disabled before the temporary mm
1766 	 * was loaded.
1767 	 */
1768 	if (hw_breakpoint_active())
1769 		hw_breakpoint_restore();
1770 }
1771 
1772 __ro_after_init struct mm_struct *poking_mm;
1773 __ro_after_init unsigned long poking_addr;
1774 
1775 static void text_poke_memcpy(void *dst, const void *src, size_t len)
1776 {
1777 	memcpy(dst, src, len);
1778 }
1779 
1780 static void text_poke_memset(void *dst, const void *src, size_t len)
1781 {
1782 	int c = *(const int *)src;
1783 
1784 	memset(dst, c, len);
1785 }
1786 
1787 typedef void text_poke_f(void *dst, const void *src, size_t len);
1788 
1789 static void *__text_poke(text_poke_f func, void *addr, const void *src, size_t len)
1790 {
1791 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
1792 	struct page *pages[2] = {NULL};
1793 	temp_mm_state_t prev;
1794 	unsigned long flags;
1795 	pte_t pte, *ptep;
1796 	spinlock_t *ptl;
1797 	pgprot_t pgprot;
1798 
1799 	/*
1800 	 * While boot memory allocator is running we cannot use struct pages as
1801 	 * they are not yet initialized. There is no way to recover.
1802 	 */
1803 	BUG_ON(!after_bootmem);
1804 
1805 	if (!core_kernel_text((unsigned long)addr)) {
1806 		pages[0] = vmalloc_to_page(addr);
1807 		if (cross_page_boundary)
1808 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
1809 	} else {
1810 		pages[0] = virt_to_page(addr);
1811 		WARN_ON(!PageReserved(pages[0]));
1812 		if (cross_page_boundary)
1813 			pages[1] = virt_to_page(addr + PAGE_SIZE);
1814 	}
1815 	/*
1816 	 * If something went wrong, crash and burn since recovery paths are not
1817 	 * implemented.
1818 	 */
1819 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
1820 
1821 	/*
1822 	 * Map the page without the global bit, as TLB flushing is done with
1823 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
1824 	 */
1825 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
1826 
1827 	/*
1828 	 * The lock is not really needed, but this allows to avoid open-coding.
1829 	 */
1830 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
1831 
1832 	/*
1833 	 * This must not fail; preallocated in poking_init().
1834 	 */
1835 	VM_BUG_ON(!ptep);
1836 
1837 	local_irq_save(flags);
1838 
1839 	pte = mk_pte(pages[0], pgprot);
1840 	set_pte_at(poking_mm, poking_addr, ptep, pte);
1841 
1842 	if (cross_page_boundary) {
1843 		pte = mk_pte(pages[1], pgprot);
1844 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
1845 	}
1846 
1847 	/*
1848 	 * Loading the temporary mm behaves as a compiler barrier, which
1849 	 * guarantees that the PTE will be set at the time memcpy() is done.
1850 	 */
1851 	prev = use_temporary_mm(poking_mm);
1852 
1853 	kasan_disable_current();
1854 	func((u8 *)poking_addr + offset_in_page(addr), src, len);
1855 	kasan_enable_current();
1856 
1857 	/*
1858 	 * Ensure that the PTE is only cleared after the instructions of memcpy
1859 	 * were issued by using a compiler barrier.
1860 	 */
1861 	barrier();
1862 
1863 	pte_clear(poking_mm, poking_addr, ptep);
1864 	if (cross_page_boundary)
1865 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
1866 
1867 	/*
1868 	 * Loading the previous page-table hierarchy requires a serializing
1869 	 * instruction that already allows the core to see the updated version.
1870 	 * Xen-PV is assumed to serialize execution in a similar manner.
1871 	 */
1872 	unuse_temporary_mm(prev);
1873 
1874 	/*
1875 	 * Flushing the TLB might involve IPIs, which would require enabled
1876 	 * IRQs, but not if the mm is not used, as it is in this point.
1877 	 */
1878 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
1879 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
1880 			   PAGE_SHIFT, false);
1881 
1882 	if (func == text_poke_memcpy) {
1883 		/*
1884 		 * If the text does not match what we just wrote then something is
1885 		 * fundamentally screwy; there's nothing we can really do about that.
1886 		 */
1887 		BUG_ON(memcmp(addr, src, len));
1888 	}
1889 
1890 	local_irq_restore(flags);
1891 	pte_unmap_unlock(ptep, ptl);
1892 	return addr;
1893 }
1894 
1895 /**
1896  * text_poke - Update instructions on a live kernel
1897  * @addr: address to modify
1898  * @opcode: source of the copy
1899  * @len: length to copy
1900  *
1901  * Only atomic text poke/set should be allowed when not doing early patching.
1902  * It means the size must be writable atomically and the address must be aligned
1903  * in a way that permits an atomic write. It also makes sure we fit on a single
1904  * page.
1905  *
1906  * Note that the caller must ensure that if the modified code is part of a
1907  * module, the module would not be removed during poking. This can be achieved
1908  * by registering a module notifier, and ordering module removal and patching
1909  * through a mutex.
1910  */
1911 void *text_poke(void *addr, const void *opcode, size_t len)
1912 {
1913 	lockdep_assert_held(&text_mutex);
1914 
1915 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1916 }
1917 
1918 /**
1919  * text_poke_kgdb - Update instructions on a live kernel by kgdb
1920  * @addr: address to modify
1921  * @opcode: source of the copy
1922  * @len: length to copy
1923  *
1924  * Only atomic text poke/set should be allowed when not doing early patching.
1925  * It means the size must be writable atomically and the address must be aligned
1926  * in a way that permits an atomic write. It also makes sure we fit on a single
1927  * page.
1928  *
1929  * Context: should only be used by kgdb, which ensures no other core is running,
1930  *	    despite the fact it does not hold the text_mutex.
1931  */
1932 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
1933 {
1934 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1935 }
1936 
1937 void *text_poke_copy_locked(void *addr, const void *opcode, size_t len,
1938 			    bool core_ok)
1939 {
1940 	unsigned long start = (unsigned long)addr;
1941 	size_t patched = 0;
1942 
1943 	if (WARN_ON_ONCE(!core_ok && core_kernel_text(start)))
1944 		return NULL;
1945 
1946 	while (patched < len) {
1947 		unsigned long ptr = start + patched;
1948 		size_t s;
1949 
1950 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1951 
1952 		__text_poke(text_poke_memcpy, (void *)ptr, opcode + patched, s);
1953 		patched += s;
1954 	}
1955 	return addr;
1956 }
1957 
1958 /**
1959  * text_poke_copy - Copy instructions into (an unused part of) RX memory
1960  * @addr: address to modify
1961  * @opcode: source of the copy
1962  * @len: length to copy, could be more than 2x PAGE_SIZE
1963  *
1964  * Not safe against concurrent execution; useful for JITs to dump
1965  * new code blocks into unused regions of RX memory. Can be used in
1966  * conjunction with synchronize_rcu_tasks() to wait for existing
1967  * execution to quiesce after having made sure no existing functions
1968  * pointers are live.
1969  */
1970 void *text_poke_copy(void *addr, const void *opcode, size_t len)
1971 {
1972 	mutex_lock(&text_mutex);
1973 	addr = text_poke_copy_locked(addr, opcode, len, false);
1974 	mutex_unlock(&text_mutex);
1975 	return addr;
1976 }
1977 
1978 /**
1979  * text_poke_set - memset into (an unused part of) RX memory
1980  * @addr: address to modify
1981  * @c: the byte to fill the area with
1982  * @len: length to copy, could be more than 2x PAGE_SIZE
1983  *
1984  * This is useful to overwrite unused regions of RX memory with illegal
1985  * instructions.
1986  */
1987 void *text_poke_set(void *addr, int c, size_t len)
1988 {
1989 	unsigned long start = (unsigned long)addr;
1990 	size_t patched = 0;
1991 
1992 	if (WARN_ON_ONCE(core_kernel_text(start)))
1993 		return NULL;
1994 
1995 	mutex_lock(&text_mutex);
1996 	while (patched < len) {
1997 		unsigned long ptr = start + patched;
1998 		size_t s;
1999 
2000 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
2001 
2002 		__text_poke(text_poke_memset, (void *)ptr, (void *)&c, s);
2003 		patched += s;
2004 	}
2005 	mutex_unlock(&text_mutex);
2006 	return addr;
2007 }
2008 
2009 static void do_sync_core(void *info)
2010 {
2011 	sync_core();
2012 }
2013 
2014 void text_poke_sync(void)
2015 {
2016 	on_each_cpu(do_sync_core, NULL, 1);
2017 }
2018 
2019 /*
2020  * NOTE: crazy scheme to allow patching Jcc.d32 but not increase the size of
2021  * this thing. When len == 6 everything is prefixed with 0x0f and we map
2022  * opcode to Jcc.d8, using len to distinguish.
2023  */
2024 struct text_poke_loc {
2025 	/* addr := _stext + rel_addr */
2026 	s32 rel_addr;
2027 	s32 disp;
2028 	u8 len;
2029 	u8 opcode;
2030 	const u8 text[POKE_MAX_OPCODE_SIZE];
2031 	/* see text_poke_bp_batch() */
2032 	u8 old;
2033 };
2034 
2035 struct bp_patching_desc {
2036 	struct text_poke_loc *vec;
2037 	int nr_entries;
2038 	atomic_t refs;
2039 };
2040 
2041 static struct bp_patching_desc bp_desc;
2042 
2043 static __always_inline
2044 struct bp_patching_desc *try_get_desc(void)
2045 {
2046 	struct bp_patching_desc *desc = &bp_desc;
2047 
2048 	if (!raw_atomic_inc_not_zero(&desc->refs))
2049 		return NULL;
2050 
2051 	return desc;
2052 }
2053 
2054 static __always_inline void put_desc(void)
2055 {
2056 	struct bp_patching_desc *desc = &bp_desc;
2057 
2058 	smp_mb__before_atomic();
2059 	raw_atomic_dec(&desc->refs);
2060 }
2061 
2062 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
2063 {
2064 	return _stext + tp->rel_addr;
2065 }
2066 
2067 static __always_inline int patch_cmp(const void *key, const void *elt)
2068 {
2069 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
2070 
2071 	if (key < text_poke_addr(tp))
2072 		return -1;
2073 	if (key > text_poke_addr(tp))
2074 		return 1;
2075 	return 0;
2076 }
2077 
2078 noinstr int poke_int3_handler(struct pt_regs *regs)
2079 {
2080 	struct bp_patching_desc *desc;
2081 	struct text_poke_loc *tp;
2082 	int ret = 0;
2083 	void *ip;
2084 
2085 	if (user_mode(regs))
2086 		return 0;
2087 
2088 	/*
2089 	 * Having observed our INT3 instruction, we now must observe
2090 	 * bp_desc with non-zero refcount:
2091 	 *
2092 	 *	bp_desc.refs = 1		INT3
2093 	 *	WMB				RMB
2094 	 *	write INT3			if (bp_desc.refs != 0)
2095 	 */
2096 	smp_rmb();
2097 
2098 	desc = try_get_desc();
2099 	if (!desc)
2100 		return 0;
2101 
2102 	/*
2103 	 * Discount the INT3. See text_poke_bp_batch().
2104 	 */
2105 	ip = (void *) regs->ip - INT3_INSN_SIZE;
2106 
2107 	/*
2108 	 * Skip the binary search if there is a single member in the vector.
2109 	 */
2110 	if (unlikely(desc->nr_entries > 1)) {
2111 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
2112 				      sizeof(struct text_poke_loc),
2113 				      patch_cmp);
2114 		if (!tp)
2115 			goto out_put;
2116 	} else {
2117 		tp = desc->vec;
2118 		if (text_poke_addr(tp) != ip)
2119 			goto out_put;
2120 	}
2121 
2122 	ip += tp->len;
2123 
2124 	switch (tp->opcode) {
2125 	case INT3_INSN_OPCODE:
2126 		/*
2127 		 * Someone poked an explicit INT3, they'll want to handle it,
2128 		 * do not consume.
2129 		 */
2130 		goto out_put;
2131 
2132 	case RET_INSN_OPCODE:
2133 		int3_emulate_ret(regs);
2134 		break;
2135 
2136 	case CALL_INSN_OPCODE:
2137 		int3_emulate_call(regs, (long)ip + tp->disp);
2138 		break;
2139 
2140 	case JMP32_INSN_OPCODE:
2141 	case JMP8_INSN_OPCODE:
2142 		int3_emulate_jmp(regs, (long)ip + tp->disp);
2143 		break;
2144 
2145 	case 0x70 ... 0x7f: /* Jcc */
2146 		int3_emulate_jcc(regs, tp->opcode & 0xf, (long)ip, tp->disp);
2147 		break;
2148 
2149 	default:
2150 		BUG();
2151 	}
2152 
2153 	ret = 1;
2154 
2155 out_put:
2156 	put_desc();
2157 	return ret;
2158 }
2159 
2160 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
2161 static struct text_poke_loc tp_vec[TP_VEC_MAX];
2162 static int tp_vec_nr;
2163 
2164 /**
2165  * text_poke_bp_batch() -- update instructions on live kernel on SMP
2166  * @tp:			vector of instructions to patch
2167  * @nr_entries:		number of entries in the vector
2168  *
2169  * Modify multi-byte instruction by using int3 breakpoint on SMP.
2170  * We completely avoid stop_machine() here, and achieve the
2171  * synchronization using int3 breakpoint.
2172  *
2173  * The way it is done:
2174  *	- For each entry in the vector:
2175  *		- add a int3 trap to the address that will be patched
2176  *	- sync cores
2177  *	- For each entry in the vector:
2178  *		- update all but the first byte of the patched range
2179  *	- sync cores
2180  *	- For each entry in the vector:
2181  *		- replace the first byte (int3) by the first byte of
2182  *		  replacing opcode
2183  *	- sync cores
2184  */
2185 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
2186 {
2187 	unsigned char int3 = INT3_INSN_OPCODE;
2188 	unsigned int i;
2189 	int do_sync;
2190 
2191 	lockdep_assert_held(&text_mutex);
2192 
2193 	bp_desc.vec = tp;
2194 	bp_desc.nr_entries = nr_entries;
2195 
2196 	/*
2197 	 * Corresponds to the implicit memory barrier in try_get_desc() to
2198 	 * ensure reading a non-zero refcount provides up to date bp_desc data.
2199 	 */
2200 	atomic_set_release(&bp_desc.refs, 1);
2201 
2202 	/*
2203 	 * Function tracing can enable thousands of places that need to be
2204 	 * updated. This can take quite some time, and with full kernel debugging
2205 	 * enabled, this could cause the softlockup watchdog to trigger.
2206 	 * This function gets called every 256 entries added to be patched.
2207 	 * Call cond_resched() here to make sure that other tasks can get scheduled
2208 	 * while processing all the functions being patched.
2209 	 */
2210 	cond_resched();
2211 
2212 	/*
2213 	 * Corresponding read barrier in int3 notifier for making sure the
2214 	 * nr_entries and handler are correctly ordered wrt. patching.
2215 	 */
2216 	smp_wmb();
2217 
2218 	/*
2219 	 * First step: add a int3 trap to the address that will be patched.
2220 	 */
2221 	for (i = 0; i < nr_entries; i++) {
2222 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
2223 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
2224 	}
2225 
2226 	text_poke_sync();
2227 
2228 	/*
2229 	 * Second step: update all but the first byte of the patched range.
2230 	 */
2231 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2232 		u8 old[POKE_MAX_OPCODE_SIZE+1] = { tp[i].old, };
2233 		u8 _new[POKE_MAX_OPCODE_SIZE+1];
2234 		const u8 *new = tp[i].text;
2235 		int len = tp[i].len;
2236 
2237 		if (len - INT3_INSN_SIZE > 0) {
2238 			memcpy(old + INT3_INSN_SIZE,
2239 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2240 			       len - INT3_INSN_SIZE);
2241 
2242 			if (len == 6) {
2243 				_new[0] = 0x0f;
2244 				memcpy(_new + 1, new, 5);
2245 				new = _new;
2246 			}
2247 
2248 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
2249 				  new + INT3_INSN_SIZE,
2250 				  len - INT3_INSN_SIZE);
2251 
2252 			do_sync++;
2253 		}
2254 
2255 		/*
2256 		 * Emit a perf event to record the text poke, primarily to
2257 		 * support Intel PT decoding which must walk the executable code
2258 		 * to reconstruct the trace. The flow up to here is:
2259 		 *   - write INT3 byte
2260 		 *   - IPI-SYNC
2261 		 *   - write instruction tail
2262 		 * At this point the actual control flow will be through the
2263 		 * INT3 and handler and not hit the old or new instruction.
2264 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
2265 		 * can still be decoded. Subsequently:
2266 		 *   - emit RECORD_TEXT_POKE with the new instruction
2267 		 *   - IPI-SYNC
2268 		 *   - write first byte
2269 		 *   - IPI-SYNC
2270 		 * So before the text poke event timestamp, the decoder will see
2271 		 * either the old instruction flow or FUP/TIP of INT3. After the
2272 		 * text poke event timestamp, the decoder will see either the
2273 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
2274 		 * use the timestamp as the point at which to modify the
2275 		 * executable code.
2276 		 * The old instruction is recorded so that the event can be
2277 		 * processed forwards or backwards.
2278 		 */
2279 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len, new, len);
2280 	}
2281 
2282 	if (do_sync) {
2283 		/*
2284 		 * According to Intel, this core syncing is very likely
2285 		 * not necessary and we'd be safe even without it. But
2286 		 * better safe than sorry (plus there's not only Intel).
2287 		 */
2288 		text_poke_sync();
2289 	}
2290 
2291 	/*
2292 	 * Third step: replace the first byte (int3) by the first byte of
2293 	 * replacing opcode.
2294 	 */
2295 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
2296 		u8 byte = tp[i].text[0];
2297 
2298 		if (tp[i].len == 6)
2299 			byte = 0x0f;
2300 
2301 		if (byte == INT3_INSN_OPCODE)
2302 			continue;
2303 
2304 		text_poke(text_poke_addr(&tp[i]), &byte, INT3_INSN_SIZE);
2305 		do_sync++;
2306 	}
2307 
2308 	if (do_sync)
2309 		text_poke_sync();
2310 
2311 	/*
2312 	 * Remove and wait for refs to be zero.
2313 	 */
2314 	if (!atomic_dec_and_test(&bp_desc.refs))
2315 		atomic_cond_read_acquire(&bp_desc.refs, !VAL);
2316 }
2317 
2318 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
2319 			       const void *opcode, size_t len, const void *emulate)
2320 {
2321 	struct insn insn;
2322 	int ret, i = 0;
2323 
2324 	if (len == 6)
2325 		i = 1;
2326 	memcpy((void *)tp->text, opcode+i, len-i);
2327 	if (!emulate)
2328 		emulate = opcode;
2329 
2330 	ret = insn_decode_kernel(&insn, emulate);
2331 	BUG_ON(ret < 0);
2332 
2333 	tp->rel_addr = addr - (void *)_stext;
2334 	tp->len = len;
2335 	tp->opcode = insn.opcode.bytes[0];
2336 
2337 	if (is_jcc32(&insn)) {
2338 		/*
2339 		 * Map Jcc.d32 onto Jcc.d8 and use len to distinguish.
2340 		 */
2341 		tp->opcode = insn.opcode.bytes[1] - 0x10;
2342 	}
2343 
2344 	switch (tp->opcode) {
2345 	case RET_INSN_OPCODE:
2346 	case JMP32_INSN_OPCODE:
2347 	case JMP8_INSN_OPCODE:
2348 		/*
2349 		 * Control flow instructions without implied execution of the
2350 		 * next instruction can be padded with INT3.
2351 		 */
2352 		for (i = insn.length; i < len; i++)
2353 			BUG_ON(tp->text[i] != INT3_INSN_OPCODE);
2354 		break;
2355 
2356 	default:
2357 		BUG_ON(len != insn.length);
2358 	}
2359 
2360 	switch (tp->opcode) {
2361 	case INT3_INSN_OPCODE:
2362 	case RET_INSN_OPCODE:
2363 		break;
2364 
2365 	case CALL_INSN_OPCODE:
2366 	case JMP32_INSN_OPCODE:
2367 	case JMP8_INSN_OPCODE:
2368 	case 0x70 ... 0x7f: /* Jcc */
2369 		tp->disp = insn.immediate.value;
2370 		break;
2371 
2372 	default: /* assume NOP */
2373 		switch (len) {
2374 		case 2: /* NOP2 -- emulate as JMP8+0 */
2375 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2376 			tp->opcode = JMP8_INSN_OPCODE;
2377 			tp->disp = 0;
2378 			break;
2379 
2380 		case 5: /* NOP5 -- emulate as JMP32+0 */
2381 			BUG_ON(memcmp(emulate, x86_nops[len], len));
2382 			tp->opcode = JMP32_INSN_OPCODE;
2383 			tp->disp = 0;
2384 			break;
2385 
2386 		default: /* unknown instruction */
2387 			BUG();
2388 		}
2389 		break;
2390 	}
2391 }
2392 
2393 /*
2394  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
2395  * early if needed.
2396  */
2397 static bool tp_order_fail(void *addr)
2398 {
2399 	struct text_poke_loc *tp;
2400 
2401 	if (!tp_vec_nr)
2402 		return false;
2403 
2404 	if (!addr) /* force */
2405 		return true;
2406 
2407 	tp = &tp_vec[tp_vec_nr - 1];
2408 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
2409 		return true;
2410 
2411 	return false;
2412 }
2413 
2414 static void text_poke_flush(void *addr)
2415 {
2416 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
2417 		text_poke_bp_batch(tp_vec, tp_vec_nr);
2418 		tp_vec_nr = 0;
2419 	}
2420 }
2421 
2422 void text_poke_finish(void)
2423 {
2424 	text_poke_flush(NULL);
2425 }
2426 
2427 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
2428 {
2429 	struct text_poke_loc *tp;
2430 
2431 	text_poke_flush(addr);
2432 
2433 	tp = &tp_vec[tp_vec_nr++];
2434 	text_poke_loc_init(tp, addr, opcode, len, emulate);
2435 }
2436 
2437 /**
2438  * text_poke_bp() -- update instructions on live kernel on SMP
2439  * @addr:	address to patch
2440  * @opcode:	opcode of new instruction
2441  * @len:	length to copy
2442  * @emulate:	instruction to be emulated
2443  *
2444  * Update a single instruction with the vector in the stack, avoiding
2445  * dynamically allocated memory. This function should be used when it is
2446  * not possible to allocate memory.
2447  */
2448 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
2449 {
2450 	struct text_poke_loc tp;
2451 
2452 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
2453 	text_poke_bp_batch(&tp, 1);
2454 }
2455