xref: /openbmc/linux/arch/x86/kernel/alternative.c (revision dff03381)
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 static int __initdata_or_module debug_alternative;
41 
42 static int __init debug_alt(char *str)
43 {
44 	debug_alternative = 1;
45 	return 1;
46 }
47 __setup("debug-alternative", debug_alt);
48 
49 static int noreplace_smp;
50 
51 static int __init setup_noreplace_smp(char *str)
52 {
53 	noreplace_smp = 1;
54 	return 1;
55 }
56 __setup("noreplace-smp", setup_noreplace_smp);
57 
58 #define DPRINTK(fmt, args...)						\
59 do {									\
60 	if (debug_alternative)						\
61 		printk(KERN_DEBUG pr_fmt(fmt) "\n", ##args);		\
62 } while (0)
63 
64 #define DUMP_BYTES(buf, len, fmt, args...)				\
65 do {									\
66 	if (unlikely(debug_alternative)) {				\
67 		int j;							\
68 									\
69 		if (!(len))						\
70 			break;						\
71 									\
72 		printk(KERN_DEBUG pr_fmt(fmt), ##args);			\
73 		for (j = 0; j < (len) - 1; j++)				\
74 			printk(KERN_CONT "%02hhx ", buf[j]);		\
75 		printk(KERN_CONT "%02hhx\n", buf[j]);			\
76 	}								\
77 } while (0)
78 
79 static const unsigned char x86nops[] =
80 {
81 	BYTES_NOP1,
82 	BYTES_NOP2,
83 	BYTES_NOP3,
84 	BYTES_NOP4,
85 	BYTES_NOP5,
86 	BYTES_NOP6,
87 	BYTES_NOP7,
88 	BYTES_NOP8,
89 };
90 
91 const unsigned char * const x86_nops[ASM_NOP_MAX+1] =
92 {
93 	NULL,
94 	x86nops,
95 	x86nops + 1,
96 	x86nops + 1 + 2,
97 	x86nops + 1 + 2 + 3,
98 	x86nops + 1 + 2 + 3 + 4,
99 	x86nops + 1 + 2 + 3 + 4 + 5,
100 	x86nops + 1 + 2 + 3 + 4 + 5 + 6,
101 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
102 };
103 
104 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
105 static void __init_or_module add_nops(void *insns, unsigned int len)
106 {
107 	while (len > 0) {
108 		unsigned int noplen = len;
109 		if (noplen > ASM_NOP_MAX)
110 			noplen = ASM_NOP_MAX;
111 		memcpy(insns, x86_nops[noplen], noplen);
112 		insns += noplen;
113 		len -= noplen;
114 	}
115 }
116 
117 extern s32 __retpoline_sites[], __retpoline_sites_end[];
118 extern s32 __return_sites[], __return_sites_end[];
119 extern s32 __ibt_endbr_seal[], __ibt_endbr_seal_end[];
120 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
121 extern s32 __smp_locks[], __smp_locks_end[];
122 void text_poke_early(void *addr, const void *opcode, size_t len);
123 
124 /*
125  * Are we looking at a near JMP with a 1 or 4-byte displacement.
126  */
127 static inline bool is_jmp(const u8 opcode)
128 {
129 	return opcode == 0xeb || opcode == 0xe9;
130 }
131 
132 static void __init_or_module
133 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insn_buff)
134 {
135 	u8 *next_rip, *tgt_rip;
136 	s32 n_dspl, o_dspl;
137 	int repl_len;
138 
139 	if (a->replacementlen != 5)
140 		return;
141 
142 	o_dspl = *(s32 *)(insn_buff + 1);
143 
144 	/* next_rip of the replacement JMP */
145 	next_rip = repl_insn + a->replacementlen;
146 	/* target rip of the replacement JMP */
147 	tgt_rip  = next_rip + o_dspl;
148 	n_dspl = tgt_rip - orig_insn;
149 
150 	DPRINTK("target RIP: %px, new_displ: 0x%x", tgt_rip, n_dspl);
151 
152 	if (tgt_rip - orig_insn >= 0) {
153 		if (n_dspl - 2 <= 127)
154 			goto two_byte_jmp;
155 		else
156 			goto five_byte_jmp;
157 	/* negative offset */
158 	} else {
159 		if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
160 			goto two_byte_jmp;
161 		else
162 			goto five_byte_jmp;
163 	}
164 
165 two_byte_jmp:
166 	n_dspl -= 2;
167 
168 	insn_buff[0] = 0xeb;
169 	insn_buff[1] = (s8)n_dspl;
170 	add_nops(insn_buff + 2, 3);
171 
172 	repl_len = 2;
173 	goto done;
174 
175 five_byte_jmp:
176 	n_dspl -= 5;
177 
178 	insn_buff[0] = 0xe9;
179 	*(s32 *)&insn_buff[1] = n_dspl;
180 
181 	repl_len = 5;
182 
183 done:
184 
185 	DPRINTK("final displ: 0x%08x, JMP 0x%lx",
186 		n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
187 }
188 
189 /*
190  * optimize_nops_range() - Optimize a sequence of single byte NOPs (0x90)
191  *
192  * @instr: instruction byte stream
193  * @instrlen: length of the above
194  * @off: offset within @instr where the first NOP has been detected
195  *
196  * Return: number of NOPs found (and replaced).
197  */
198 static __always_inline int optimize_nops_range(u8 *instr, u8 instrlen, int off)
199 {
200 	unsigned long flags;
201 	int i = off, nnops;
202 
203 	while (i < instrlen) {
204 		if (instr[i] != 0x90)
205 			break;
206 
207 		i++;
208 	}
209 
210 	nnops = i - off;
211 
212 	if (nnops <= 1)
213 		return nnops;
214 
215 	local_irq_save(flags);
216 	add_nops(instr + off, nnops);
217 	local_irq_restore(flags);
218 
219 	DUMP_BYTES(instr, instrlen, "%px: [%d:%d) optimized NOPs: ", instr, off, i);
220 
221 	return nnops;
222 }
223 
224 /*
225  * "noinline" to cause control flow change and thus invalidate I$ and
226  * cause refetch after modification.
227  */
228 static void __init_or_module noinline optimize_nops(u8 *instr, size_t len)
229 {
230 	struct insn insn;
231 	int i = 0;
232 
233 	/*
234 	 * Jump over the non-NOP insns and optimize single-byte NOPs into bigger
235 	 * ones.
236 	 */
237 	for (;;) {
238 		if (insn_decode_kernel(&insn, &instr[i]))
239 			return;
240 
241 		/*
242 		 * See if this and any potentially following NOPs can be
243 		 * optimized.
244 		 */
245 		if (insn.length == 1 && insn.opcode.bytes[0] == 0x90)
246 			i += optimize_nops_range(instr, len, i);
247 		else
248 			i += insn.length;
249 
250 		if (i >= len)
251 			return;
252 	}
253 }
254 
255 /*
256  * Replace instructions with better alternatives for this CPU type. This runs
257  * before SMP is initialized to avoid SMP problems with self modifying code.
258  * This implies that asymmetric systems where APs have less capabilities than
259  * the boot processor are not handled. Tough. Make sure you disable such
260  * features by hand.
261  *
262  * Marked "noinline" to cause control flow change and thus insn cache
263  * to refetch changed I$ lines.
264  */
265 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
266 						  struct alt_instr *end)
267 {
268 	struct alt_instr *a;
269 	u8 *instr, *replacement;
270 	u8 insn_buff[MAX_PATCH_LEN];
271 
272 	DPRINTK("alt table %px, -> %px", start, end);
273 	/*
274 	 * The scan order should be from start to end. A later scanned
275 	 * alternative code can overwrite previously scanned alternative code.
276 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
277 	 * patch code.
278 	 *
279 	 * So be careful if you want to change the scan order to any other
280 	 * order.
281 	 */
282 	for (a = start; a < end; a++) {
283 		int insn_buff_sz = 0;
284 		/* Mask away "NOT" flag bit for feature to test. */
285 		u16 feature = a->cpuid & ~ALTINSTR_FLAG_INV;
286 
287 		instr = (u8 *)&a->instr_offset + a->instr_offset;
288 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
289 		BUG_ON(a->instrlen > sizeof(insn_buff));
290 		BUG_ON(feature >= (NCAPINTS + NBUGINTS) * 32);
291 
292 		/*
293 		 * Patch if either:
294 		 * - feature is present
295 		 * - feature not present but ALTINSTR_FLAG_INV is set to mean,
296 		 *   patch if feature is *NOT* present.
297 		 */
298 		if (!boot_cpu_has(feature) == !(a->cpuid & ALTINSTR_FLAG_INV))
299 			goto next;
300 
301 		DPRINTK("feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)",
302 			(a->cpuid & ALTINSTR_FLAG_INV) ? "!" : "",
303 			feature >> 5,
304 			feature & 0x1f,
305 			instr, instr, a->instrlen,
306 			replacement, a->replacementlen);
307 
308 		DUMP_BYTES(instr, a->instrlen, "%px:   old_insn: ", instr);
309 		DUMP_BYTES(replacement, a->replacementlen, "%px:   rpl_insn: ", replacement);
310 
311 		memcpy(insn_buff, replacement, a->replacementlen);
312 		insn_buff_sz = a->replacementlen;
313 
314 		/*
315 		 * 0xe8 is a relative jump; fix the offset.
316 		 *
317 		 * Instruction length is checked before the opcode to avoid
318 		 * accessing uninitialized bytes for zero-length replacements.
319 		 */
320 		if (a->replacementlen == 5 && *insn_buff == 0xe8) {
321 			*(s32 *)(insn_buff + 1) += replacement - instr;
322 			DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
323 				*(s32 *)(insn_buff + 1),
324 				(unsigned long)instr + *(s32 *)(insn_buff + 1) + 5);
325 		}
326 
327 		if (a->replacementlen && is_jmp(replacement[0]))
328 			recompute_jump(a, instr, replacement, insn_buff);
329 
330 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
331 			insn_buff[insn_buff_sz] = 0x90;
332 
333 		DUMP_BYTES(insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
334 
335 		text_poke_early(instr, insn_buff, insn_buff_sz);
336 
337 next:
338 		optimize_nops(instr, a->instrlen);
339 	}
340 }
341 
342 #if defined(CONFIG_RETPOLINE) && defined(CONFIG_OBJTOOL)
343 
344 /*
345  * CALL/JMP *%\reg
346  */
347 static int emit_indirect(int op, int reg, u8 *bytes)
348 {
349 	int i = 0;
350 	u8 modrm;
351 
352 	switch (op) {
353 	case CALL_INSN_OPCODE:
354 		modrm = 0x10; /* Reg = 2; CALL r/m */
355 		break;
356 
357 	case JMP32_INSN_OPCODE:
358 		modrm = 0x20; /* Reg = 4; JMP r/m */
359 		break;
360 
361 	default:
362 		WARN_ON_ONCE(1);
363 		return -1;
364 	}
365 
366 	if (reg >= 8) {
367 		bytes[i++] = 0x41; /* REX.B prefix */
368 		reg -= 8;
369 	}
370 
371 	modrm |= 0xc0; /* Mod = 3 */
372 	modrm += reg;
373 
374 	bytes[i++] = 0xff; /* opcode */
375 	bytes[i++] = modrm;
376 
377 	return i;
378 }
379 
380 /*
381  * Rewrite the compiler generated retpoline thunk calls.
382  *
383  * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate
384  * indirect instructions, avoiding the extra indirection.
385  *
386  * For example, convert:
387  *
388  *   CALL __x86_indirect_thunk_\reg
389  *
390  * into:
391  *
392  *   CALL *%\reg
393  *
394  * It also tries to inline spectre_v2=retpoline,lfence when size permits.
395  */
396 static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes)
397 {
398 	retpoline_thunk_t *target;
399 	int reg, ret, i = 0;
400 	u8 op, cc;
401 
402 	target = addr + insn->length + insn->immediate.value;
403 	reg = target - __x86_indirect_thunk_array;
404 
405 	if (WARN_ON_ONCE(reg & ~0xf))
406 		return -1;
407 
408 	/* If anyone ever does: CALL/JMP *%rsp, we're in deep trouble. */
409 	BUG_ON(reg == 4);
410 
411 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE) &&
412 	    !cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE))
413 		return -1;
414 
415 	op = insn->opcode.bytes[0];
416 
417 	/*
418 	 * Convert:
419 	 *
420 	 *   Jcc.d32 __x86_indirect_thunk_\reg
421 	 *
422 	 * into:
423 	 *
424 	 *   Jncc.d8 1f
425 	 *   [ LFENCE ]
426 	 *   JMP *%\reg
427 	 *   [ NOP ]
428 	 * 1:
429 	 */
430 	/* Jcc.d32 second opcode byte is in the range: 0x80-0x8f */
431 	if (op == 0x0f && (insn->opcode.bytes[1] & 0xf0) == 0x80) {
432 		cc = insn->opcode.bytes[1] & 0xf;
433 		cc ^= 1; /* invert condition */
434 
435 		bytes[i++] = 0x70 + cc;        /* Jcc.d8 */
436 		bytes[i++] = insn->length - 2; /* sizeof(Jcc.d8) == 2 */
437 
438 		/* Continue as if: JMP.d32 __x86_indirect_thunk_\reg */
439 		op = JMP32_INSN_OPCODE;
440 	}
441 
442 	/*
443 	 * For RETPOLINE_LFENCE: prepend the indirect CALL/JMP with an LFENCE.
444 	 */
445 	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
446 		bytes[i++] = 0x0f;
447 		bytes[i++] = 0xae;
448 		bytes[i++] = 0xe8; /* LFENCE */
449 	}
450 
451 	ret = emit_indirect(op, reg, bytes + i);
452 	if (ret < 0)
453 		return ret;
454 	i += ret;
455 
456 	for (; i < insn->length;)
457 		bytes[i++] = BYTES_NOP1;
458 
459 	return i;
460 }
461 
462 /*
463  * Generated by 'objtool --retpoline'.
464  */
465 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end)
466 {
467 	s32 *s;
468 
469 	for (s = start; s < end; s++) {
470 		void *addr = (void *)s + *s;
471 		struct insn insn;
472 		int len, ret;
473 		u8 bytes[16];
474 		u8 op1, op2;
475 
476 		ret = insn_decode_kernel(&insn, addr);
477 		if (WARN_ON_ONCE(ret < 0))
478 			continue;
479 
480 		op1 = insn.opcode.bytes[0];
481 		op2 = insn.opcode.bytes[1];
482 
483 		switch (op1) {
484 		case CALL_INSN_OPCODE:
485 		case JMP32_INSN_OPCODE:
486 			break;
487 
488 		case 0x0f: /* escape */
489 			if (op2 >= 0x80 && op2 <= 0x8f)
490 				break;
491 			fallthrough;
492 		default:
493 			WARN_ON_ONCE(1);
494 			continue;
495 		}
496 
497 		DPRINTK("retpoline at: %pS (%px) len: %d to: %pS",
498 			addr, addr, insn.length,
499 			addr + insn.length + insn.immediate.value);
500 
501 		len = patch_retpoline(addr, &insn, bytes);
502 		if (len == insn.length) {
503 			optimize_nops(bytes, len);
504 			DUMP_BYTES(((u8*)addr),  len, "%px: orig: ", addr);
505 			DUMP_BYTES(((u8*)bytes), len, "%px: repl: ", addr);
506 			text_poke_early(addr, bytes, len);
507 		}
508 	}
509 }
510 
511 #ifdef CONFIG_RETHUNK
512 /*
513  * Rewrite the compiler generated return thunk tail-calls.
514  *
515  * For example, convert:
516  *
517  *   JMP __x86_return_thunk
518  *
519  * into:
520  *
521  *   RET
522  */
523 static int patch_return(void *addr, struct insn *insn, u8 *bytes)
524 {
525 	int i = 0;
526 
527 	if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
528 		return -1;
529 
530 	bytes[i++] = RET_INSN_OPCODE;
531 
532 	for (; i < insn->length;)
533 		bytes[i++] = INT3_INSN_OPCODE;
534 
535 	return i;
536 }
537 
538 void __init_or_module noinline apply_returns(s32 *start, s32 *end)
539 {
540 	s32 *s;
541 
542 	for (s = start; s < end; s++) {
543 		void *dest = NULL, *addr = (void *)s + *s;
544 		struct insn insn;
545 		int len, ret;
546 		u8 bytes[16];
547 		u8 op;
548 
549 		ret = insn_decode_kernel(&insn, addr);
550 		if (WARN_ON_ONCE(ret < 0))
551 			continue;
552 
553 		op = insn.opcode.bytes[0];
554 		if (op == JMP32_INSN_OPCODE)
555 			dest = addr + insn.length + insn.immediate.value;
556 
557 		if (__static_call_fixup(addr, op, dest) ||
558 		    WARN_ONCE(dest != &__x86_return_thunk,
559 			      "missing return thunk: %pS-%pS: %*ph",
560 			      addr, dest, 5, addr))
561 			continue;
562 
563 		DPRINTK("return thunk at: %pS (%px) len: %d to: %pS",
564 			addr, addr, insn.length,
565 			addr + insn.length + insn.immediate.value);
566 
567 		len = patch_return(addr, &insn, bytes);
568 		if (len == insn.length) {
569 			DUMP_BYTES(((u8*)addr),  len, "%px: orig: ", addr);
570 			DUMP_BYTES(((u8*)bytes), len, "%px: repl: ", addr);
571 			text_poke_early(addr, bytes, len);
572 		}
573 	}
574 }
575 #else
576 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
577 #endif /* CONFIG_RETHUNK */
578 
579 #else /* !CONFIG_RETPOLINE || !CONFIG_OBJTOOL */
580 
581 void __init_or_module noinline apply_retpolines(s32 *start, s32 *end) { }
582 void __init_or_module noinline apply_returns(s32 *start, s32 *end) { }
583 
584 #endif /* CONFIG_RETPOLINE && CONFIG_OBJTOOL */
585 
586 #ifdef CONFIG_X86_KERNEL_IBT
587 
588 /*
589  * Generated by: objtool --ibt
590  */
591 void __init_or_module noinline apply_ibt_endbr(s32 *start, s32 *end)
592 {
593 	s32 *s;
594 
595 	for (s = start; s < end; s++) {
596 		u32 endbr, poison = gen_endbr_poison();
597 		void *addr = (void *)s + *s;
598 
599 		if (WARN_ON_ONCE(get_kernel_nofault(endbr, addr)))
600 			continue;
601 
602 		if (WARN_ON_ONCE(!is_endbr(endbr)))
603 			continue;
604 
605 		DPRINTK("ENDBR at: %pS (%px)", addr, addr);
606 
607 		/*
608 		 * When we have IBT, the lack of ENDBR will trigger #CP
609 		 */
610 		DUMP_BYTES(((u8*)addr), 4, "%px: orig: ", addr);
611 		DUMP_BYTES(((u8*)&poison), 4, "%px: repl: ", addr);
612 		text_poke_early(addr, &poison, 4);
613 	}
614 }
615 
616 #else
617 
618 void __init_or_module noinline apply_ibt_endbr(s32 *start, s32 *end) { }
619 
620 #endif /* CONFIG_X86_KERNEL_IBT */
621 
622 #ifdef CONFIG_SMP
623 static void alternatives_smp_lock(const s32 *start, const s32 *end,
624 				  u8 *text, u8 *text_end)
625 {
626 	const s32 *poff;
627 
628 	for (poff = start; poff < end; poff++) {
629 		u8 *ptr = (u8 *)poff + *poff;
630 
631 		if (!*poff || ptr < text || ptr >= text_end)
632 			continue;
633 		/* turn DS segment override prefix into lock prefix */
634 		if (*ptr == 0x3e)
635 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
636 	}
637 }
638 
639 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
640 				    u8 *text, u8 *text_end)
641 {
642 	const s32 *poff;
643 
644 	for (poff = start; poff < end; poff++) {
645 		u8 *ptr = (u8 *)poff + *poff;
646 
647 		if (!*poff || ptr < text || ptr >= text_end)
648 			continue;
649 		/* turn lock prefix into DS segment override prefix */
650 		if (*ptr == 0xf0)
651 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
652 	}
653 }
654 
655 struct smp_alt_module {
656 	/* what is this ??? */
657 	struct module	*mod;
658 	char		*name;
659 
660 	/* ptrs to lock prefixes */
661 	const s32	*locks;
662 	const s32	*locks_end;
663 
664 	/* .text segment, needed to avoid patching init code ;) */
665 	u8		*text;
666 	u8		*text_end;
667 
668 	struct list_head next;
669 };
670 static LIST_HEAD(smp_alt_modules);
671 static bool uniproc_patched = false;	/* protected by text_mutex */
672 
673 void __init_or_module alternatives_smp_module_add(struct module *mod,
674 						  char *name,
675 						  void *locks, void *locks_end,
676 						  void *text,  void *text_end)
677 {
678 	struct smp_alt_module *smp;
679 
680 	mutex_lock(&text_mutex);
681 	if (!uniproc_patched)
682 		goto unlock;
683 
684 	if (num_possible_cpus() == 1)
685 		/* Don't bother remembering, we'll never have to undo it. */
686 		goto smp_unlock;
687 
688 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
689 	if (NULL == smp)
690 		/* we'll run the (safe but slow) SMP code then ... */
691 		goto unlock;
692 
693 	smp->mod	= mod;
694 	smp->name	= name;
695 	smp->locks	= locks;
696 	smp->locks_end	= locks_end;
697 	smp->text	= text;
698 	smp->text_end	= text_end;
699 	DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
700 		smp->locks, smp->locks_end,
701 		smp->text, smp->text_end, smp->name);
702 
703 	list_add_tail(&smp->next, &smp_alt_modules);
704 smp_unlock:
705 	alternatives_smp_unlock(locks, locks_end, text, text_end);
706 unlock:
707 	mutex_unlock(&text_mutex);
708 }
709 
710 void __init_or_module alternatives_smp_module_del(struct module *mod)
711 {
712 	struct smp_alt_module *item;
713 
714 	mutex_lock(&text_mutex);
715 	list_for_each_entry(item, &smp_alt_modules, next) {
716 		if (mod != item->mod)
717 			continue;
718 		list_del(&item->next);
719 		kfree(item);
720 		break;
721 	}
722 	mutex_unlock(&text_mutex);
723 }
724 
725 void alternatives_enable_smp(void)
726 {
727 	struct smp_alt_module *mod;
728 
729 	/* Why bother if there are no other CPUs? */
730 	BUG_ON(num_possible_cpus() == 1);
731 
732 	mutex_lock(&text_mutex);
733 
734 	if (uniproc_patched) {
735 		pr_info("switching to SMP code\n");
736 		BUG_ON(num_online_cpus() != 1);
737 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
738 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
739 		list_for_each_entry(mod, &smp_alt_modules, next)
740 			alternatives_smp_lock(mod->locks, mod->locks_end,
741 					      mod->text, mod->text_end);
742 		uniproc_patched = false;
743 	}
744 	mutex_unlock(&text_mutex);
745 }
746 
747 /*
748  * Return 1 if the address range is reserved for SMP-alternatives.
749  * Must hold text_mutex.
750  */
751 int alternatives_text_reserved(void *start, void *end)
752 {
753 	struct smp_alt_module *mod;
754 	const s32 *poff;
755 	u8 *text_start = start;
756 	u8 *text_end = end;
757 
758 	lockdep_assert_held(&text_mutex);
759 
760 	list_for_each_entry(mod, &smp_alt_modules, next) {
761 		if (mod->text > text_end || mod->text_end < text_start)
762 			continue;
763 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
764 			const u8 *ptr = (const u8 *)poff + *poff;
765 
766 			if (text_start <= ptr && text_end > ptr)
767 				return 1;
768 		}
769 	}
770 
771 	return 0;
772 }
773 #endif /* CONFIG_SMP */
774 
775 #ifdef CONFIG_PARAVIRT
776 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
777 				     struct paravirt_patch_site *end)
778 {
779 	struct paravirt_patch_site *p;
780 	char insn_buff[MAX_PATCH_LEN];
781 
782 	for (p = start; p < end; p++) {
783 		unsigned int used;
784 
785 		BUG_ON(p->len > MAX_PATCH_LEN);
786 		/* prep the buffer with the original instructions */
787 		memcpy(insn_buff, p->instr, p->len);
788 		used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
789 
790 		BUG_ON(used > p->len);
791 
792 		/* Pad the rest with nops */
793 		add_nops(insn_buff + used, p->len - used);
794 		text_poke_early(p->instr, insn_buff, p->len);
795 	}
796 }
797 extern struct paravirt_patch_site __start_parainstructions[],
798 	__stop_parainstructions[];
799 #endif	/* CONFIG_PARAVIRT */
800 
801 /*
802  * Self-test for the INT3 based CALL emulation code.
803  *
804  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
805  * properly and that there is a stack gap between the INT3 frame and the
806  * previous context. Without this gap doing a virtual PUSH on the interrupted
807  * stack would corrupt the INT3 IRET frame.
808  *
809  * See entry_{32,64}.S for more details.
810  */
811 
812 /*
813  * We define the int3_magic() function in assembly to control the calling
814  * convention such that we can 'call' it from assembly.
815  */
816 
817 extern void int3_magic(unsigned int *ptr); /* defined in asm */
818 
819 asm (
820 "	.pushsection	.init.text, \"ax\", @progbits\n"
821 "	.type		int3_magic, @function\n"
822 "int3_magic:\n"
823 	ANNOTATE_NOENDBR
824 "	movl	$1, (%" _ASM_ARG1 ")\n"
825 	ASM_RET
826 "	.size		int3_magic, .-int3_magic\n"
827 "	.popsection\n"
828 );
829 
830 extern void int3_selftest_ip(void); /* defined in asm below */
831 
832 static int __init
833 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
834 {
835 	unsigned long selftest = (unsigned long)&int3_selftest_ip;
836 	struct die_args *args = data;
837 	struct pt_regs *regs = args->regs;
838 
839 	OPTIMIZER_HIDE_VAR(selftest);
840 
841 	if (!regs || user_mode(regs))
842 		return NOTIFY_DONE;
843 
844 	if (val != DIE_INT3)
845 		return NOTIFY_DONE;
846 
847 	if (regs->ip - INT3_INSN_SIZE != selftest)
848 		return NOTIFY_DONE;
849 
850 	int3_emulate_call(regs, (unsigned long)&int3_magic);
851 	return NOTIFY_STOP;
852 }
853 
854 /* Must be noinline to ensure uniqueness of int3_selftest_ip. */
855 static noinline void __init int3_selftest(void)
856 {
857 	static __initdata struct notifier_block int3_exception_nb = {
858 		.notifier_call	= int3_exception_notify,
859 		.priority	= INT_MAX-1, /* last */
860 	};
861 	unsigned int val = 0;
862 
863 	BUG_ON(register_die_notifier(&int3_exception_nb));
864 
865 	/*
866 	 * Basically: int3_magic(&val); but really complicated :-)
867 	 *
868 	 * INT3 padded with NOP to CALL_INSN_SIZE. The int3_exception_nb
869 	 * notifier above will emulate CALL for us.
870 	 */
871 	asm volatile ("int3_selftest_ip:\n\t"
872 		      ANNOTATE_NOENDBR
873 		      "    int3; nop; nop; nop; nop\n\t"
874 		      : ASM_CALL_CONSTRAINT
875 		      : __ASM_SEL_RAW(a, D) (&val)
876 		      : "memory");
877 
878 	BUG_ON(val != 1);
879 
880 	unregister_die_notifier(&int3_exception_nb);
881 }
882 
883 void __init alternative_instructions(void)
884 {
885 	int3_selftest();
886 
887 	/*
888 	 * The patching is not fully atomic, so try to avoid local
889 	 * interruptions that might execute the to be patched code.
890 	 * Other CPUs are not running.
891 	 */
892 	stop_nmi();
893 
894 	/*
895 	 * Don't stop machine check exceptions while patching.
896 	 * MCEs only happen when something got corrupted and in this
897 	 * case we must do something about the corruption.
898 	 * Ignoring it is worse than an unlikely patching race.
899 	 * Also machine checks tend to be broadcast and if one CPU
900 	 * goes into machine check the others follow quickly, so we don't
901 	 * expect a machine check to cause undue problems during to code
902 	 * patching.
903 	 */
904 
905 	/*
906 	 * Paravirt patching and alternative patching can be combined to
907 	 * replace a function call with a short direct code sequence (e.g.
908 	 * by setting a constant return value instead of doing that in an
909 	 * external function).
910 	 * In order to make this work the following sequence is required:
911 	 * 1. set (artificial) features depending on used paravirt
912 	 *    functions which can later influence alternative patching
913 	 * 2. apply paravirt patching (generally replacing an indirect
914 	 *    function call with a direct one)
915 	 * 3. apply alternative patching (e.g. replacing a direct function
916 	 *    call with a custom code sequence)
917 	 * Doing paravirt patching after alternative patching would clobber
918 	 * the optimization of the custom code with a function call again.
919 	 */
920 	paravirt_set_cap();
921 
922 	/*
923 	 * First patch paravirt functions, such that we overwrite the indirect
924 	 * call with the direct call.
925 	 */
926 	apply_paravirt(__parainstructions, __parainstructions_end);
927 
928 	/*
929 	 * Rewrite the retpolines, must be done before alternatives since
930 	 * those can rewrite the retpoline thunks.
931 	 */
932 	apply_retpolines(__retpoline_sites, __retpoline_sites_end);
933 	apply_returns(__return_sites, __return_sites_end);
934 
935 	/*
936 	 * Then patch alternatives, such that those paravirt calls that are in
937 	 * alternatives can be overwritten by their immediate fragments.
938 	 */
939 	apply_alternatives(__alt_instructions, __alt_instructions_end);
940 
941 	apply_ibt_endbr(__ibt_endbr_seal, __ibt_endbr_seal_end);
942 
943 #ifdef CONFIG_SMP
944 	/* Patch to UP if other cpus not imminent. */
945 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
946 		uniproc_patched = true;
947 		alternatives_smp_module_add(NULL, "core kernel",
948 					    __smp_locks, __smp_locks_end,
949 					    _text, _etext);
950 	}
951 
952 	if (!uniproc_patched || num_possible_cpus() == 1) {
953 		free_init_pages("SMP alternatives",
954 				(unsigned long)__smp_locks,
955 				(unsigned long)__smp_locks_end);
956 	}
957 #endif
958 
959 	restart_nmi();
960 	alternatives_patched = 1;
961 }
962 
963 /**
964  * text_poke_early - Update instructions on a live kernel at boot time
965  * @addr: address to modify
966  * @opcode: source of the copy
967  * @len: length to copy
968  *
969  * When you use this code to patch more than one byte of an instruction
970  * you need to make sure that other CPUs cannot execute this code in parallel.
971  * Also no thread must be currently preempted in the middle of these
972  * instructions. And on the local CPU you need to be protected against NMI or
973  * MCE handlers seeing an inconsistent instruction while you patch.
974  */
975 void __init_or_module text_poke_early(void *addr, const void *opcode,
976 				      size_t len)
977 {
978 	unsigned long flags;
979 
980 	if (boot_cpu_has(X86_FEATURE_NX) &&
981 	    is_module_text_address((unsigned long)addr)) {
982 		/*
983 		 * Modules text is marked initially as non-executable, so the
984 		 * code cannot be running and speculative code-fetches are
985 		 * prevented. Just change the code.
986 		 */
987 		memcpy(addr, opcode, len);
988 	} else {
989 		local_irq_save(flags);
990 		memcpy(addr, opcode, len);
991 		local_irq_restore(flags);
992 		sync_core();
993 
994 		/*
995 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
996 		 * that causes hangs on some VIA CPUs.
997 		 */
998 	}
999 }
1000 
1001 typedef struct {
1002 	struct mm_struct *mm;
1003 } temp_mm_state_t;
1004 
1005 /*
1006  * Using a temporary mm allows to set temporary mappings that are not accessible
1007  * by other CPUs. Such mappings are needed to perform sensitive memory writes
1008  * that override the kernel memory protections (e.g., W^X), without exposing the
1009  * temporary page-table mappings that are required for these write operations to
1010  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
1011  * mapping is torn down.
1012  *
1013  * Context: The temporary mm needs to be used exclusively by a single core. To
1014  *          harden security IRQs must be disabled while the temporary mm is
1015  *          loaded, thereby preventing interrupt handler bugs from overriding
1016  *          the kernel memory protection.
1017  */
1018 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
1019 {
1020 	temp_mm_state_t temp_state;
1021 
1022 	lockdep_assert_irqs_disabled();
1023 
1024 	/*
1025 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
1026 	 * with a stale address space WITHOUT being in lazy mode after
1027 	 * restoring the previous mm.
1028 	 */
1029 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
1030 		leave_mm(smp_processor_id());
1031 
1032 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
1033 	switch_mm_irqs_off(NULL, mm, current);
1034 
1035 	/*
1036 	 * If breakpoints are enabled, disable them while the temporary mm is
1037 	 * used. Userspace might set up watchpoints on addresses that are used
1038 	 * in the temporary mm, which would lead to wrong signals being sent or
1039 	 * crashes.
1040 	 *
1041 	 * Note that breakpoints are not disabled selectively, which also causes
1042 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
1043 	 * undesirable, but still seems reasonable as the code that runs in the
1044 	 * temporary mm should be short.
1045 	 */
1046 	if (hw_breakpoint_active())
1047 		hw_breakpoint_disable();
1048 
1049 	return temp_state;
1050 }
1051 
1052 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
1053 {
1054 	lockdep_assert_irqs_disabled();
1055 	switch_mm_irqs_off(NULL, prev_state.mm, current);
1056 
1057 	/*
1058 	 * Restore the breakpoints if they were disabled before the temporary mm
1059 	 * was loaded.
1060 	 */
1061 	if (hw_breakpoint_active())
1062 		hw_breakpoint_restore();
1063 }
1064 
1065 __ro_after_init struct mm_struct *poking_mm;
1066 __ro_after_init unsigned long poking_addr;
1067 
1068 static void text_poke_memcpy(void *dst, const void *src, size_t len)
1069 {
1070 	memcpy(dst, src, len);
1071 }
1072 
1073 static void text_poke_memset(void *dst, const void *src, size_t len)
1074 {
1075 	int c = *(const int *)src;
1076 
1077 	memset(dst, c, len);
1078 }
1079 
1080 typedef void text_poke_f(void *dst, const void *src, size_t len);
1081 
1082 static void *__text_poke(text_poke_f func, void *addr, const void *src, size_t len)
1083 {
1084 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
1085 	struct page *pages[2] = {NULL};
1086 	temp_mm_state_t prev;
1087 	unsigned long flags;
1088 	pte_t pte, *ptep;
1089 	spinlock_t *ptl;
1090 	pgprot_t pgprot;
1091 
1092 	/*
1093 	 * While boot memory allocator is running we cannot use struct pages as
1094 	 * they are not yet initialized. There is no way to recover.
1095 	 */
1096 	BUG_ON(!after_bootmem);
1097 
1098 	if (!core_kernel_text((unsigned long)addr)) {
1099 		pages[0] = vmalloc_to_page(addr);
1100 		if (cross_page_boundary)
1101 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
1102 	} else {
1103 		pages[0] = virt_to_page(addr);
1104 		WARN_ON(!PageReserved(pages[0]));
1105 		if (cross_page_boundary)
1106 			pages[1] = virt_to_page(addr + PAGE_SIZE);
1107 	}
1108 	/*
1109 	 * If something went wrong, crash and burn since recovery paths are not
1110 	 * implemented.
1111 	 */
1112 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
1113 
1114 	/*
1115 	 * Map the page without the global bit, as TLB flushing is done with
1116 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
1117 	 */
1118 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
1119 
1120 	/*
1121 	 * The lock is not really needed, but this allows to avoid open-coding.
1122 	 */
1123 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
1124 
1125 	/*
1126 	 * This must not fail; preallocated in poking_init().
1127 	 */
1128 	VM_BUG_ON(!ptep);
1129 
1130 	local_irq_save(flags);
1131 
1132 	pte = mk_pte(pages[0], pgprot);
1133 	set_pte_at(poking_mm, poking_addr, ptep, pte);
1134 
1135 	if (cross_page_boundary) {
1136 		pte = mk_pte(pages[1], pgprot);
1137 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
1138 	}
1139 
1140 	/*
1141 	 * Loading the temporary mm behaves as a compiler barrier, which
1142 	 * guarantees that the PTE will be set at the time memcpy() is done.
1143 	 */
1144 	prev = use_temporary_mm(poking_mm);
1145 
1146 	kasan_disable_current();
1147 	func((u8 *)poking_addr + offset_in_page(addr), src, len);
1148 	kasan_enable_current();
1149 
1150 	/*
1151 	 * Ensure that the PTE is only cleared after the instructions of memcpy
1152 	 * were issued by using a compiler barrier.
1153 	 */
1154 	barrier();
1155 
1156 	pte_clear(poking_mm, poking_addr, ptep);
1157 	if (cross_page_boundary)
1158 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
1159 
1160 	/*
1161 	 * Loading the previous page-table hierarchy requires a serializing
1162 	 * instruction that already allows the core to see the updated version.
1163 	 * Xen-PV is assumed to serialize execution in a similar manner.
1164 	 */
1165 	unuse_temporary_mm(prev);
1166 
1167 	/*
1168 	 * Flushing the TLB might involve IPIs, which would require enabled
1169 	 * IRQs, but not if the mm is not used, as it is in this point.
1170 	 */
1171 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
1172 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
1173 			   PAGE_SHIFT, false);
1174 
1175 	if (func == text_poke_memcpy) {
1176 		/*
1177 		 * If the text does not match what we just wrote then something is
1178 		 * fundamentally screwy; there's nothing we can really do about that.
1179 		 */
1180 		BUG_ON(memcmp(addr, src, len));
1181 	}
1182 
1183 	local_irq_restore(flags);
1184 	pte_unmap_unlock(ptep, ptl);
1185 	return addr;
1186 }
1187 
1188 /**
1189  * text_poke - Update instructions on a live kernel
1190  * @addr: address to modify
1191  * @opcode: source of the copy
1192  * @len: length to copy
1193  *
1194  * Only atomic text poke/set should be allowed when not doing early patching.
1195  * It means the size must be writable atomically and the address must be aligned
1196  * in a way that permits an atomic write. It also makes sure we fit on a single
1197  * page.
1198  *
1199  * Note that the caller must ensure that if the modified code is part of a
1200  * module, the module would not be removed during poking. This can be achieved
1201  * by registering a module notifier, and ordering module removal and patching
1202  * trough a mutex.
1203  */
1204 void *text_poke(void *addr, const void *opcode, size_t len)
1205 {
1206 	lockdep_assert_held(&text_mutex);
1207 
1208 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1209 }
1210 
1211 /**
1212  * text_poke_kgdb - Update instructions on a live kernel by kgdb
1213  * @addr: address to modify
1214  * @opcode: source of the copy
1215  * @len: length to copy
1216  *
1217  * Only atomic text poke/set should be allowed when not doing early patching.
1218  * It means the size must be writable atomically and the address must be aligned
1219  * in a way that permits an atomic write. It also makes sure we fit on a single
1220  * page.
1221  *
1222  * Context: should only be used by kgdb, which ensures no other core is running,
1223  *	    despite the fact it does not hold the text_mutex.
1224  */
1225 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
1226 {
1227 	return __text_poke(text_poke_memcpy, addr, opcode, len);
1228 }
1229 
1230 /**
1231  * text_poke_copy - Copy instructions into (an unused part of) RX memory
1232  * @addr: address to modify
1233  * @opcode: source of the copy
1234  * @len: length to copy, could be more than 2x PAGE_SIZE
1235  *
1236  * Not safe against concurrent execution; useful for JITs to dump
1237  * new code blocks into unused regions of RX memory. Can be used in
1238  * conjunction with synchronize_rcu_tasks() to wait for existing
1239  * execution to quiesce after having made sure no existing functions
1240  * pointers are live.
1241  */
1242 void *text_poke_copy(void *addr, const void *opcode, size_t len)
1243 {
1244 	unsigned long start = (unsigned long)addr;
1245 	size_t patched = 0;
1246 
1247 	if (WARN_ON_ONCE(core_kernel_text(start)))
1248 		return NULL;
1249 
1250 	mutex_lock(&text_mutex);
1251 	while (patched < len) {
1252 		unsigned long ptr = start + patched;
1253 		size_t s;
1254 
1255 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1256 
1257 		__text_poke(text_poke_memcpy, (void *)ptr, opcode + patched, s);
1258 		patched += s;
1259 	}
1260 	mutex_unlock(&text_mutex);
1261 	return addr;
1262 }
1263 
1264 /**
1265  * text_poke_set - memset into (an unused part of) RX memory
1266  * @addr: address to modify
1267  * @c: the byte to fill the area with
1268  * @len: length to copy, could be more than 2x PAGE_SIZE
1269  *
1270  * This is useful to overwrite unused regions of RX memory with illegal
1271  * instructions.
1272  */
1273 void *text_poke_set(void *addr, int c, size_t len)
1274 {
1275 	unsigned long start = (unsigned long)addr;
1276 	size_t patched = 0;
1277 
1278 	if (WARN_ON_ONCE(core_kernel_text(start)))
1279 		return NULL;
1280 
1281 	mutex_lock(&text_mutex);
1282 	while (patched < len) {
1283 		unsigned long ptr = start + patched;
1284 		size_t s;
1285 
1286 		s = min_t(size_t, PAGE_SIZE * 2 - offset_in_page(ptr), len - patched);
1287 
1288 		__text_poke(text_poke_memset, (void *)ptr, (void *)&c, s);
1289 		patched += s;
1290 	}
1291 	mutex_unlock(&text_mutex);
1292 	return addr;
1293 }
1294 
1295 static void do_sync_core(void *info)
1296 {
1297 	sync_core();
1298 }
1299 
1300 void text_poke_sync(void)
1301 {
1302 	on_each_cpu(do_sync_core, NULL, 1);
1303 }
1304 
1305 struct text_poke_loc {
1306 	/* addr := _stext + rel_addr */
1307 	s32 rel_addr;
1308 	s32 disp;
1309 	u8 len;
1310 	u8 opcode;
1311 	const u8 text[POKE_MAX_OPCODE_SIZE];
1312 	/* see text_poke_bp_batch() */
1313 	u8 old;
1314 };
1315 
1316 struct bp_patching_desc {
1317 	struct text_poke_loc *vec;
1318 	int nr_entries;
1319 	atomic_t refs;
1320 };
1321 
1322 static struct bp_patching_desc *bp_desc;
1323 
1324 static __always_inline
1325 struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
1326 {
1327 	/* rcu_dereference */
1328 	struct bp_patching_desc *desc = __READ_ONCE(*descp);
1329 
1330 	if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
1331 		return NULL;
1332 
1333 	return desc;
1334 }
1335 
1336 static __always_inline void put_desc(struct bp_patching_desc *desc)
1337 {
1338 	smp_mb__before_atomic();
1339 	arch_atomic_dec(&desc->refs);
1340 }
1341 
1342 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
1343 {
1344 	return _stext + tp->rel_addr;
1345 }
1346 
1347 static __always_inline int patch_cmp(const void *key, const void *elt)
1348 {
1349 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
1350 
1351 	if (key < text_poke_addr(tp))
1352 		return -1;
1353 	if (key > text_poke_addr(tp))
1354 		return 1;
1355 	return 0;
1356 }
1357 
1358 noinstr int poke_int3_handler(struct pt_regs *regs)
1359 {
1360 	struct bp_patching_desc *desc;
1361 	struct text_poke_loc *tp;
1362 	int ret = 0;
1363 	void *ip;
1364 
1365 	if (user_mode(regs))
1366 		return 0;
1367 
1368 	/*
1369 	 * Having observed our INT3 instruction, we now must observe
1370 	 * bp_desc:
1371 	 *
1372 	 *	bp_desc = desc			INT3
1373 	 *	WMB				RMB
1374 	 *	write INT3			if (desc)
1375 	 */
1376 	smp_rmb();
1377 
1378 	desc = try_get_desc(&bp_desc);
1379 	if (!desc)
1380 		return 0;
1381 
1382 	/*
1383 	 * Discount the INT3. See text_poke_bp_batch().
1384 	 */
1385 	ip = (void *) regs->ip - INT3_INSN_SIZE;
1386 
1387 	/*
1388 	 * Skip the binary search if there is a single member in the vector.
1389 	 */
1390 	if (unlikely(desc->nr_entries > 1)) {
1391 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
1392 				      sizeof(struct text_poke_loc),
1393 				      patch_cmp);
1394 		if (!tp)
1395 			goto out_put;
1396 	} else {
1397 		tp = desc->vec;
1398 		if (text_poke_addr(tp) != ip)
1399 			goto out_put;
1400 	}
1401 
1402 	ip += tp->len;
1403 
1404 	switch (tp->opcode) {
1405 	case INT3_INSN_OPCODE:
1406 		/*
1407 		 * Someone poked an explicit INT3, they'll want to handle it,
1408 		 * do not consume.
1409 		 */
1410 		goto out_put;
1411 
1412 	case RET_INSN_OPCODE:
1413 		int3_emulate_ret(regs);
1414 		break;
1415 
1416 	case CALL_INSN_OPCODE:
1417 		int3_emulate_call(regs, (long)ip + tp->disp);
1418 		break;
1419 
1420 	case JMP32_INSN_OPCODE:
1421 	case JMP8_INSN_OPCODE:
1422 		int3_emulate_jmp(regs, (long)ip + tp->disp);
1423 		break;
1424 
1425 	default:
1426 		BUG();
1427 	}
1428 
1429 	ret = 1;
1430 
1431 out_put:
1432 	put_desc(desc);
1433 	return ret;
1434 }
1435 
1436 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
1437 static struct text_poke_loc tp_vec[TP_VEC_MAX];
1438 static int tp_vec_nr;
1439 
1440 /**
1441  * text_poke_bp_batch() -- update instructions on live kernel on SMP
1442  * @tp:			vector of instructions to patch
1443  * @nr_entries:		number of entries in the vector
1444  *
1445  * Modify multi-byte instruction by using int3 breakpoint on SMP.
1446  * We completely avoid stop_machine() here, and achieve the
1447  * synchronization using int3 breakpoint.
1448  *
1449  * The way it is done:
1450  *	- For each entry in the vector:
1451  *		- add a int3 trap to the address that will be patched
1452  *	- sync cores
1453  *	- For each entry in the vector:
1454  *		- update all but the first byte of the patched range
1455  *	- sync cores
1456  *	- For each entry in the vector:
1457  *		- replace the first byte (int3) by the first byte of
1458  *		  replacing opcode
1459  *	- sync cores
1460  */
1461 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
1462 {
1463 	struct bp_patching_desc desc = {
1464 		.vec = tp,
1465 		.nr_entries = nr_entries,
1466 		.refs = ATOMIC_INIT(1),
1467 	};
1468 	unsigned char int3 = INT3_INSN_OPCODE;
1469 	unsigned int i;
1470 	int do_sync;
1471 
1472 	lockdep_assert_held(&text_mutex);
1473 
1474 	smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
1475 
1476 	/*
1477 	 * Corresponding read barrier in int3 notifier for making sure the
1478 	 * nr_entries and handler are correctly ordered wrt. patching.
1479 	 */
1480 	smp_wmb();
1481 
1482 	/*
1483 	 * First step: add a int3 trap to the address that will be patched.
1484 	 */
1485 	for (i = 0; i < nr_entries; i++) {
1486 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
1487 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
1488 	}
1489 
1490 	text_poke_sync();
1491 
1492 	/*
1493 	 * Second step: update all but the first byte of the patched range.
1494 	 */
1495 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1496 		u8 old[POKE_MAX_OPCODE_SIZE] = { tp[i].old, };
1497 		int len = tp[i].len;
1498 
1499 		if (len - INT3_INSN_SIZE > 0) {
1500 			memcpy(old + INT3_INSN_SIZE,
1501 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1502 			       len - INT3_INSN_SIZE);
1503 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1504 				  (const char *)tp[i].text + INT3_INSN_SIZE,
1505 				  len - INT3_INSN_SIZE);
1506 			do_sync++;
1507 		}
1508 
1509 		/*
1510 		 * Emit a perf event to record the text poke, primarily to
1511 		 * support Intel PT decoding which must walk the executable code
1512 		 * to reconstruct the trace. The flow up to here is:
1513 		 *   - write INT3 byte
1514 		 *   - IPI-SYNC
1515 		 *   - write instruction tail
1516 		 * At this point the actual control flow will be through the
1517 		 * INT3 and handler and not hit the old or new instruction.
1518 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
1519 		 * can still be decoded. Subsequently:
1520 		 *   - emit RECORD_TEXT_POKE with the new instruction
1521 		 *   - IPI-SYNC
1522 		 *   - write first byte
1523 		 *   - IPI-SYNC
1524 		 * So before the text poke event timestamp, the decoder will see
1525 		 * either the old instruction flow or FUP/TIP of INT3. After the
1526 		 * text poke event timestamp, the decoder will see either the
1527 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
1528 		 * use the timestamp as the point at which to modify the
1529 		 * executable code.
1530 		 * The old instruction is recorded so that the event can be
1531 		 * processed forwards or backwards.
1532 		 */
1533 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len,
1534 				     tp[i].text, len);
1535 	}
1536 
1537 	if (do_sync) {
1538 		/*
1539 		 * According to Intel, this core syncing is very likely
1540 		 * not necessary and we'd be safe even without it. But
1541 		 * better safe than sorry (plus there's not only Intel).
1542 		 */
1543 		text_poke_sync();
1544 	}
1545 
1546 	/*
1547 	 * Third step: replace the first byte (int3) by the first byte of
1548 	 * replacing opcode.
1549 	 */
1550 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1551 		if (tp[i].text[0] == INT3_INSN_OPCODE)
1552 			continue;
1553 
1554 		text_poke(text_poke_addr(&tp[i]), tp[i].text, INT3_INSN_SIZE);
1555 		do_sync++;
1556 	}
1557 
1558 	if (do_sync)
1559 		text_poke_sync();
1560 
1561 	/*
1562 	 * Remove and synchronize_rcu(), except we have a very primitive
1563 	 * refcount based completion.
1564 	 */
1565 	WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
1566 	if (!atomic_dec_and_test(&desc.refs))
1567 		atomic_cond_read_acquire(&desc.refs, !VAL);
1568 }
1569 
1570 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
1571 			       const void *opcode, size_t len, const void *emulate)
1572 {
1573 	struct insn insn;
1574 	int ret, i;
1575 
1576 	memcpy((void *)tp->text, opcode, len);
1577 	if (!emulate)
1578 		emulate = opcode;
1579 
1580 	ret = insn_decode_kernel(&insn, emulate);
1581 	BUG_ON(ret < 0);
1582 
1583 	tp->rel_addr = addr - (void *)_stext;
1584 	tp->len = len;
1585 	tp->opcode = insn.opcode.bytes[0];
1586 
1587 	switch (tp->opcode) {
1588 	case RET_INSN_OPCODE:
1589 	case JMP32_INSN_OPCODE:
1590 	case JMP8_INSN_OPCODE:
1591 		/*
1592 		 * Control flow instructions without implied execution of the
1593 		 * next instruction can be padded with INT3.
1594 		 */
1595 		for (i = insn.length; i < len; i++)
1596 			BUG_ON(tp->text[i] != INT3_INSN_OPCODE);
1597 		break;
1598 
1599 	default:
1600 		BUG_ON(len != insn.length);
1601 	};
1602 
1603 
1604 	switch (tp->opcode) {
1605 	case INT3_INSN_OPCODE:
1606 	case RET_INSN_OPCODE:
1607 		break;
1608 
1609 	case CALL_INSN_OPCODE:
1610 	case JMP32_INSN_OPCODE:
1611 	case JMP8_INSN_OPCODE:
1612 		tp->disp = insn.immediate.value;
1613 		break;
1614 
1615 	default: /* assume NOP */
1616 		switch (len) {
1617 		case 2: /* NOP2 -- emulate as JMP8+0 */
1618 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1619 			tp->opcode = JMP8_INSN_OPCODE;
1620 			tp->disp = 0;
1621 			break;
1622 
1623 		case 5: /* NOP5 -- emulate as JMP32+0 */
1624 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1625 			tp->opcode = JMP32_INSN_OPCODE;
1626 			tp->disp = 0;
1627 			break;
1628 
1629 		default: /* unknown instruction */
1630 			BUG();
1631 		}
1632 		break;
1633 	}
1634 }
1635 
1636 /*
1637  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
1638  * early if needed.
1639  */
1640 static bool tp_order_fail(void *addr)
1641 {
1642 	struct text_poke_loc *tp;
1643 
1644 	if (!tp_vec_nr)
1645 		return false;
1646 
1647 	if (!addr) /* force */
1648 		return true;
1649 
1650 	tp = &tp_vec[tp_vec_nr - 1];
1651 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
1652 		return true;
1653 
1654 	return false;
1655 }
1656 
1657 static void text_poke_flush(void *addr)
1658 {
1659 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
1660 		text_poke_bp_batch(tp_vec, tp_vec_nr);
1661 		tp_vec_nr = 0;
1662 	}
1663 }
1664 
1665 void text_poke_finish(void)
1666 {
1667 	text_poke_flush(NULL);
1668 }
1669 
1670 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
1671 {
1672 	struct text_poke_loc *tp;
1673 
1674 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1675 		text_poke_early(addr, opcode, len);
1676 		return;
1677 	}
1678 
1679 	text_poke_flush(addr);
1680 
1681 	tp = &tp_vec[tp_vec_nr++];
1682 	text_poke_loc_init(tp, addr, opcode, len, emulate);
1683 }
1684 
1685 /**
1686  * text_poke_bp() -- update instructions on live kernel on SMP
1687  * @addr:	address to patch
1688  * @opcode:	opcode of new instruction
1689  * @len:	length to copy
1690  * @emulate:	instruction to be emulated
1691  *
1692  * Update a single instruction with the vector in the stack, avoiding
1693  * dynamically allocated memory. This function should be used when it is
1694  * not possible to allocate memory.
1695  */
1696 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
1697 {
1698 	struct text_poke_loc tp;
1699 
1700 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1701 		text_poke_early(addr, opcode, len);
1702 		return;
1703 	}
1704 
1705 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
1706 	text_poke_bp_batch(&tp, 1);
1707 }
1708