xref: /openbmc/linux/arch/x86/kernel/alternative.c (revision ed84ef1c)
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 
33 int __read_mostly alternatives_patched;
34 
35 EXPORT_SYMBOL_GPL(alternatives_patched);
36 
37 #define MAX_PATCH_LEN (255-1)
38 
39 static int __initdata_or_module debug_alternative;
40 
41 static int __init debug_alt(char *str)
42 {
43 	debug_alternative = 1;
44 	return 1;
45 }
46 __setup("debug-alternative", debug_alt);
47 
48 static int noreplace_smp;
49 
50 static int __init setup_noreplace_smp(char *str)
51 {
52 	noreplace_smp = 1;
53 	return 1;
54 }
55 __setup("noreplace-smp", setup_noreplace_smp);
56 
57 #define DPRINTK(fmt, args...)						\
58 do {									\
59 	if (debug_alternative)						\
60 		printk(KERN_DEBUG pr_fmt(fmt) "\n", ##args);		\
61 } while (0)
62 
63 #define DUMP_BYTES(buf, len, fmt, args...)				\
64 do {									\
65 	if (unlikely(debug_alternative)) {				\
66 		int j;							\
67 									\
68 		if (!(len))						\
69 			break;						\
70 									\
71 		printk(KERN_DEBUG pr_fmt(fmt), ##args);			\
72 		for (j = 0; j < (len) - 1; j++)				\
73 			printk(KERN_CONT "%02hhx ", buf[j]);		\
74 		printk(KERN_CONT "%02hhx\n", buf[j]);			\
75 	}								\
76 } while (0)
77 
78 static const unsigned char x86nops[] =
79 {
80 	BYTES_NOP1,
81 	BYTES_NOP2,
82 	BYTES_NOP3,
83 	BYTES_NOP4,
84 	BYTES_NOP5,
85 	BYTES_NOP6,
86 	BYTES_NOP7,
87 	BYTES_NOP8,
88 };
89 
90 const unsigned char * const x86_nops[ASM_NOP_MAX+1] =
91 {
92 	NULL,
93 	x86nops,
94 	x86nops + 1,
95 	x86nops + 1 + 2,
96 	x86nops + 1 + 2 + 3,
97 	x86nops + 1 + 2 + 3 + 4,
98 	x86nops + 1 + 2 + 3 + 4 + 5,
99 	x86nops + 1 + 2 + 3 + 4 + 5 + 6,
100 	x86nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
101 };
102 
103 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
104 static void __init_or_module add_nops(void *insns, unsigned int len)
105 {
106 	while (len > 0) {
107 		unsigned int noplen = len;
108 		if (noplen > ASM_NOP_MAX)
109 			noplen = ASM_NOP_MAX;
110 		memcpy(insns, x86_nops[noplen], noplen);
111 		insns += noplen;
112 		len -= noplen;
113 	}
114 }
115 
116 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
117 extern s32 __smp_locks[], __smp_locks_end[];
118 void text_poke_early(void *addr, const void *opcode, size_t len);
119 
120 /*
121  * Are we looking at a near JMP with a 1 or 4-byte displacement.
122  */
123 static inline bool is_jmp(const u8 opcode)
124 {
125 	return opcode == 0xeb || opcode == 0xe9;
126 }
127 
128 static void __init_or_module
129 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insn_buff)
130 {
131 	u8 *next_rip, *tgt_rip;
132 	s32 n_dspl, o_dspl;
133 	int repl_len;
134 
135 	if (a->replacementlen != 5)
136 		return;
137 
138 	o_dspl = *(s32 *)(insn_buff + 1);
139 
140 	/* next_rip of the replacement JMP */
141 	next_rip = repl_insn + a->replacementlen;
142 	/* target rip of the replacement JMP */
143 	tgt_rip  = next_rip + o_dspl;
144 	n_dspl = tgt_rip - orig_insn;
145 
146 	DPRINTK("target RIP: %px, new_displ: 0x%x", tgt_rip, n_dspl);
147 
148 	if (tgt_rip - orig_insn >= 0) {
149 		if (n_dspl - 2 <= 127)
150 			goto two_byte_jmp;
151 		else
152 			goto five_byte_jmp;
153 	/* negative offset */
154 	} else {
155 		if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
156 			goto two_byte_jmp;
157 		else
158 			goto five_byte_jmp;
159 	}
160 
161 two_byte_jmp:
162 	n_dspl -= 2;
163 
164 	insn_buff[0] = 0xeb;
165 	insn_buff[1] = (s8)n_dspl;
166 	add_nops(insn_buff + 2, 3);
167 
168 	repl_len = 2;
169 	goto done;
170 
171 five_byte_jmp:
172 	n_dspl -= 5;
173 
174 	insn_buff[0] = 0xe9;
175 	*(s32 *)&insn_buff[1] = n_dspl;
176 
177 	repl_len = 5;
178 
179 done:
180 
181 	DPRINTK("final displ: 0x%08x, JMP 0x%lx",
182 		n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
183 }
184 
185 /*
186  * optimize_nops_range() - Optimize a sequence of single byte NOPs (0x90)
187  *
188  * @instr: instruction byte stream
189  * @instrlen: length of the above
190  * @off: offset within @instr where the first NOP has been detected
191  *
192  * Return: number of NOPs found (and replaced).
193  */
194 static __always_inline int optimize_nops_range(u8 *instr, u8 instrlen, int off)
195 {
196 	unsigned long flags;
197 	int i = off, nnops;
198 
199 	while (i < instrlen) {
200 		if (instr[i] != 0x90)
201 			break;
202 
203 		i++;
204 	}
205 
206 	nnops = i - off;
207 
208 	if (nnops <= 1)
209 		return nnops;
210 
211 	local_irq_save(flags);
212 	add_nops(instr + off, nnops);
213 	local_irq_restore(flags);
214 
215 	DUMP_BYTES(instr, instrlen, "%px: [%d:%d) optimized NOPs: ", instr, off, i);
216 
217 	return nnops;
218 }
219 
220 /*
221  * "noinline" to cause control flow change and thus invalidate I$ and
222  * cause refetch after modification.
223  */
224 static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr)
225 {
226 	struct insn insn;
227 	int i = 0;
228 
229 	/*
230 	 * Jump over the non-NOP insns and optimize single-byte NOPs into bigger
231 	 * ones.
232 	 */
233 	for (;;) {
234 		if (insn_decode_kernel(&insn, &instr[i]))
235 			return;
236 
237 		/*
238 		 * See if this and any potentially following NOPs can be
239 		 * optimized.
240 		 */
241 		if (insn.length == 1 && insn.opcode.bytes[0] == 0x90)
242 			i += optimize_nops_range(instr, a->instrlen, i);
243 		else
244 			i += insn.length;
245 
246 		if (i >= a->instrlen)
247 			return;
248 	}
249 }
250 
251 /*
252  * Replace instructions with better alternatives for this CPU type. This runs
253  * before SMP is initialized to avoid SMP problems with self modifying code.
254  * This implies that asymmetric systems where APs have less capabilities than
255  * the boot processor are not handled. Tough. Make sure you disable such
256  * features by hand.
257  *
258  * Marked "noinline" to cause control flow change and thus insn cache
259  * to refetch changed I$ lines.
260  */
261 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
262 						  struct alt_instr *end)
263 {
264 	struct alt_instr *a;
265 	u8 *instr, *replacement;
266 	u8 insn_buff[MAX_PATCH_LEN];
267 
268 	DPRINTK("alt table %px, -> %px", start, end);
269 	/*
270 	 * The scan order should be from start to end. A later scanned
271 	 * alternative code can overwrite previously scanned alternative code.
272 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
273 	 * patch code.
274 	 *
275 	 * So be careful if you want to change the scan order to any other
276 	 * order.
277 	 */
278 	for (a = start; a < end; a++) {
279 		int insn_buff_sz = 0;
280 		/* Mask away "NOT" flag bit for feature to test. */
281 		u16 feature = a->cpuid & ~ALTINSTR_FLAG_INV;
282 
283 		instr = (u8 *)&a->instr_offset + a->instr_offset;
284 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
285 		BUG_ON(a->instrlen > sizeof(insn_buff));
286 		BUG_ON(feature >= (NCAPINTS + NBUGINTS) * 32);
287 
288 		/*
289 		 * Patch if either:
290 		 * - feature is present
291 		 * - feature not present but ALTINSTR_FLAG_INV is set to mean,
292 		 *   patch if feature is *NOT* present.
293 		 */
294 		if (!boot_cpu_has(feature) == !(a->cpuid & ALTINSTR_FLAG_INV))
295 			goto next;
296 
297 		DPRINTK("feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)",
298 			(a->cpuid & ALTINSTR_FLAG_INV) ? "!" : "",
299 			feature >> 5,
300 			feature & 0x1f,
301 			instr, instr, a->instrlen,
302 			replacement, a->replacementlen);
303 
304 		DUMP_BYTES(instr, a->instrlen, "%px:   old_insn: ", instr);
305 		DUMP_BYTES(replacement, a->replacementlen, "%px:   rpl_insn: ", replacement);
306 
307 		memcpy(insn_buff, replacement, a->replacementlen);
308 		insn_buff_sz = a->replacementlen;
309 
310 		/*
311 		 * 0xe8 is a relative jump; fix the offset.
312 		 *
313 		 * Instruction length is checked before the opcode to avoid
314 		 * accessing uninitialized bytes for zero-length replacements.
315 		 */
316 		if (a->replacementlen == 5 && *insn_buff == 0xe8) {
317 			*(s32 *)(insn_buff + 1) += replacement - instr;
318 			DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
319 				*(s32 *)(insn_buff + 1),
320 				(unsigned long)instr + *(s32 *)(insn_buff + 1) + 5);
321 		}
322 
323 		if (a->replacementlen && is_jmp(replacement[0]))
324 			recompute_jump(a, instr, replacement, insn_buff);
325 
326 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
327 			insn_buff[insn_buff_sz] = 0x90;
328 
329 		DUMP_BYTES(insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
330 
331 		text_poke_early(instr, insn_buff, insn_buff_sz);
332 
333 next:
334 		optimize_nops(a, instr);
335 	}
336 }
337 
338 #ifdef CONFIG_SMP
339 static void alternatives_smp_lock(const s32 *start, const s32 *end,
340 				  u8 *text, u8 *text_end)
341 {
342 	const s32 *poff;
343 
344 	for (poff = start; poff < end; poff++) {
345 		u8 *ptr = (u8 *)poff + *poff;
346 
347 		if (!*poff || ptr < text || ptr >= text_end)
348 			continue;
349 		/* turn DS segment override prefix into lock prefix */
350 		if (*ptr == 0x3e)
351 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
352 	}
353 }
354 
355 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
356 				    u8 *text, u8 *text_end)
357 {
358 	const s32 *poff;
359 
360 	for (poff = start; poff < end; poff++) {
361 		u8 *ptr = (u8 *)poff + *poff;
362 
363 		if (!*poff || ptr < text || ptr >= text_end)
364 			continue;
365 		/* turn lock prefix into DS segment override prefix */
366 		if (*ptr == 0xf0)
367 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
368 	}
369 }
370 
371 struct smp_alt_module {
372 	/* what is this ??? */
373 	struct module	*mod;
374 	char		*name;
375 
376 	/* ptrs to lock prefixes */
377 	const s32	*locks;
378 	const s32	*locks_end;
379 
380 	/* .text segment, needed to avoid patching init code ;) */
381 	u8		*text;
382 	u8		*text_end;
383 
384 	struct list_head next;
385 };
386 static LIST_HEAD(smp_alt_modules);
387 static bool uniproc_patched = false;	/* protected by text_mutex */
388 
389 void __init_or_module alternatives_smp_module_add(struct module *mod,
390 						  char *name,
391 						  void *locks, void *locks_end,
392 						  void *text,  void *text_end)
393 {
394 	struct smp_alt_module *smp;
395 
396 	mutex_lock(&text_mutex);
397 	if (!uniproc_patched)
398 		goto unlock;
399 
400 	if (num_possible_cpus() == 1)
401 		/* Don't bother remembering, we'll never have to undo it. */
402 		goto smp_unlock;
403 
404 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
405 	if (NULL == smp)
406 		/* we'll run the (safe but slow) SMP code then ... */
407 		goto unlock;
408 
409 	smp->mod	= mod;
410 	smp->name	= name;
411 	smp->locks	= locks;
412 	smp->locks_end	= locks_end;
413 	smp->text	= text;
414 	smp->text_end	= text_end;
415 	DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
416 		smp->locks, smp->locks_end,
417 		smp->text, smp->text_end, smp->name);
418 
419 	list_add_tail(&smp->next, &smp_alt_modules);
420 smp_unlock:
421 	alternatives_smp_unlock(locks, locks_end, text, text_end);
422 unlock:
423 	mutex_unlock(&text_mutex);
424 }
425 
426 void __init_or_module alternatives_smp_module_del(struct module *mod)
427 {
428 	struct smp_alt_module *item;
429 
430 	mutex_lock(&text_mutex);
431 	list_for_each_entry(item, &smp_alt_modules, next) {
432 		if (mod != item->mod)
433 			continue;
434 		list_del(&item->next);
435 		kfree(item);
436 		break;
437 	}
438 	mutex_unlock(&text_mutex);
439 }
440 
441 void alternatives_enable_smp(void)
442 {
443 	struct smp_alt_module *mod;
444 
445 	/* Why bother if there are no other CPUs? */
446 	BUG_ON(num_possible_cpus() == 1);
447 
448 	mutex_lock(&text_mutex);
449 
450 	if (uniproc_patched) {
451 		pr_info("switching to SMP code\n");
452 		BUG_ON(num_online_cpus() != 1);
453 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
454 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
455 		list_for_each_entry(mod, &smp_alt_modules, next)
456 			alternatives_smp_lock(mod->locks, mod->locks_end,
457 					      mod->text, mod->text_end);
458 		uniproc_patched = false;
459 	}
460 	mutex_unlock(&text_mutex);
461 }
462 
463 /*
464  * Return 1 if the address range is reserved for SMP-alternatives.
465  * Must hold text_mutex.
466  */
467 int alternatives_text_reserved(void *start, void *end)
468 {
469 	struct smp_alt_module *mod;
470 	const s32 *poff;
471 	u8 *text_start = start;
472 	u8 *text_end = end;
473 
474 	lockdep_assert_held(&text_mutex);
475 
476 	list_for_each_entry(mod, &smp_alt_modules, next) {
477 		if (mod->text > text_end || mod->text_end < text_start)
478 			continue;
479 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
480 			const u8 *ptr = (const u8 *)poff + *poff;
481 
482 			if (text_start <= ptr && text_end > ptr)
483 				return 1;
484 		}
485 	}
486 
487 	return 0;
488 }
489 #endif /* CONFIG_SMP */
490 
491 #ifdef CONFIG_PARAVIRT
492 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
493 				     struct paravirt_patch_site *end)
494 {
495 	struct paravirt_patch_site *p;
496 	char insn_buff[MAX_PATCH_LEN];
497 
498 	for (p = start; p < end; p++) {
499 		unsigned int used;
500 
501 		BUG_ON(p->len > MAX_PATCH_LEN);
502 		/* prep the buffer with the original instructions */
503 		memcpy(insn_buff, p->instr, p->len);
504 		used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
505 
506 		BUG_ON(used > p->len);
507 
508 		/* Pad the rest with nops */
509 		add_nops(insn_buff + used, p->len - used);
510 		text_poke_early(p->instr, insn_buff, p->len);
511 	}
512 }
513 extern struct paravirt_patch_site __start_parainstructions[],
514 	__stop_parainstructions[];
515 #endif	/* CONFIG_PARAVIRT */
516 
517 /*
518  * Self-test for the INT3 based CALL emulation code.
519  *
520  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
521  * properly and that there is a stack gap between the INT3 frame and the
522  * previous context. Without this gap doing a virtual PUSH on the interrupted
523  * stack would corrupt the INT3 IRET frame.
524  *
525  * See entry_{32,64}.S for more details.
526  */
527 
528 /*
529  * We define the int3_magic() function in assembly to control the calling
530  * convention such that we can 'call' it from assembly.
531  */
532 
533 extern void int3_magic(unsigned int *ptr); /* defined in asm */
534 
535 asm (
536 "	.pushsection	.init.text, \"ax\", @progbits\n"
537 "	.type		int3_magic, @function\n"
538 "int3_magic:\n"
539 "	movl	$1, (%" _ASM_ARG1 ")\n"
540 "	ret\n"
541 "	.size		int3_magic, .-int3_magic\n"
542 "	.popsection\n"
543 );
544 
545 extern __initdata unsigned long int3_selftest_ip; /* defined in asm below */
546 
547 static int __init
548 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
549 {
550 	struct die_args *args = data;
551 	struct pt_regs *regs = args->regs;
552 
553 	if (!regs || user_mode(regs))
554 		return NOTIFY_DONE;
555 
556 	if (val != DIE_INT3)
557 		return NOTIFY_DONE;
558 
559 	if (regs->ip - INT3_INSN_SIZE != int3_selftest_ip)
560 		return NOTIFY_DONE;
561 
562 	int3_emulate_call(regs, (unsigned long)&int3_magic);
563 	return NOTIFY_STOP;
564 }
565 
566 static void __init int3_selftest(void)
567 {
568 	static __initdata struct notifier_block int3_exception_nb = {
569 		.notifier_call	= int3_exception_notify,
570 		.priority	= INT_MAX-1, /* last */
571 	};
572 	unsigned int val = 0;
573 
574 	BUG_ON(register_die_notifier(&int3_exception_nb));
575 
576 	/*
577 	 * Basically: int3_magic(&val); but really complicated :-)
578 	 *
579 	 * Stick the address of the INT3 instruction into int3_selftest_ip,
580 	 * then trigger the INT3, padded with NOPs to match a CALL instruction
581 	 * length.
582 	 */
583 	asm volatile ("1: int3; nop; nop; nop; nop\n\t"
584 		      ".pushsection .init.data,\"aw\"\n\t"
585 		      ".align " __ASM_SEL(4, 8) "\n\t"
586 		      ".type int3_selftest_ip, @object\n\t"
587 		      ".size int3_selftest_ip, " __ASM_SEL(4, 8) "\n\t"
588 		      "int3_selftest_ip:\n\t"
589 		      __ASM_SEL(.long, .quad) " 1b\n\t"
590 		      ".popsection\n\t"
591 		      : ASM_CALL_CONSTRAINT
592 		      : __ASM_SEL_RAW(a, D) (&val)
593 		      : "memory");
594 
595 	BUG_ON(val != 1);
596 
597 	unregister_die_notifier(&int3_exception_nb);
598 }
599 
600 void __init alternative_instructions(void)
601 {
602 	int3_selftest();
603 
604 	/*
605 	 * The patching is not fully atomic, so try to avoid local
606 	 * interruptions that might execute the to be patched code.
607 	 * Other CPUs are not running.
608 	 */
609 	stop_nmi();
610 
611 	/*
612 	 * Don't stop machine check exceptions while patching.
613 	 * MCEs only happen when something got corrupted and in this
614 	 * case we must do something about the corruption.
615 	 * Ignoring it is worse than an unlikely patching race.
616 	 * Also machine checks tend to be broadcast and if one CPU
617 	 * goes into machine check the others follow quickly, so we don't
618 	 * expect a machine check to cause undue problems during to code
619 	 * patching.
620 	 */
621 
622 	/*
623 	 * Paravirt patching and alternative patching can be combined to
624 	 * replace a function call with a short direct code sequence (e.g.
625 	 * by setting a constant return value instead of doing that in an
626 	 * external function).
627 	 * In order to make this work the following sequence is required:
628 	 * 1. set (artificial) features depending on used paravirt
629 	 *    functions which can later influence alternative patching
630 	 * 2. apply paravirt patching (generally replacing an indirect
631 	 *    function call with a direct one)
632 	 * 3. apply alternative patching (e.g. replacing a direct function
633 	 *    call with a custom code sequence)
634 	 * Doing paravirt patching after alternative patching would clobber
635 	 * the optimization of the custom code with a function call again.
636 	 */
637 	paravirt_set_cap();
638 
639 	/*
640 	 * First patch paravirt functions, such that we overwrite the indirect
641 	 * call with the direct call.
642 	 */
643 	apply_paravirt(__parainstructions, __parainstructions_end);
644 
645 	/*
646 	 * Then patch alternatives, such that those paravirt calls that are in
647 	 * alternatives can be overwritten by their immediate fragments.
648 	 */
649 	apply_alternatives(__alt_instructions, __alt_instructions_end);
650 
651 #ifdef CONFIG_SMP
652 	/* Patch to UP if other cpus not imminent. */
653 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
654 		uniproc_patched = true;
655 		alternatives_smp_module_add(NULL, "core kernel",
656 					    __smp_locks, __smp_locks_end,
657 					    _text, _etext);
658 	}
659 
660 	if (!uniproc_patched || num_possible_cpus() == 1) {
661 		free_init_pages("SMP alternatives",
662 				(unsigned long)__smp_locks,
663 				(unsigned long)__smp_locks_end);
664 	}
665 #endif
666 
667 	restart_nmi();
668 	alternatives_patched = 1;
669 }
670 
671 /**
672  * text_poke_early - Update instructions on a live kernel at boot time
673  * @addr: address to modify
674  * @opcode: source of the copy
675  * @len: length to copy
676  *
677  * When you use this code to patch more than one byte of an instruction
678  * you need to make sure that other CPUs cannot execute this code in parallel.
679  * Also no thread must be currently preempted in the middle of these
680  * instructions. And on the local CPU you need to be protected against NMI or
681  * MCE handlers seeing an inconsistent instruction while you patch.
682  */
683 void __init_or_module text_poke_early(void *addr, const void *opcode,
684 				      size_t len)
685 {
686 	unsigned long flags;
687 
688 	if (boot_cpu_has(X86_FEATURE_NX) &&
689 	    is_module_text_address((unsigned long)addr)) {
690 		/*
691 		 * Modules text is marked initially as non-executable, so the
692 		 * code cannot be running and speculative code-fetches are
693 		 * prevented. Just change the code.
694 		 */
695 		memcpy(addr, opcode, len);
696 	} else {
697 		local_irq_save(flags);
698 		memcpy(addr, opcode, len);
699 		local_irq_restore(flags);
700 		sync_core();
701 
702 		/*
703 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
704 		 * that causes hangs on some VIA CPUs.
705 		 */
706 	}
707 }
708 
709 typedef struct {
710 	struct mm_struct *mm;
711 } temp_mm_state_t;
712 
713 /*
714  * Using a temporary mm allows to set temporary mappings that are not accessible
715  * by other CPUs. Such mappings are needed to perform sensitive memory writes
716  * that override the kernel memory protections (e.g., W^X), without exposing the
717  * temporary page-table mappings that are required for these write operations to
718  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
719  * mapping is torn down.
720  *
721  * Context: The temporary mm needs to be used exclusively by a single core. To
722  *          harden security IRQs must be disabled while the temporary mm is
723  *          loaded, thereby preventing interrupt handler bugs from overriding
724  *          the kernel memory protection.
725  */
726 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
727 {
728 	temp_mm_state_t temp_state;
729 
730 	lockdep_assert_irqs_disabled();
731 
732 	/*
733 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
734 	 * with a stale address space WITHOUT being in lazy mode after
735 	 * restoring the previous mm.
736 	 */
737 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
738 		leave_mm(smp_processor_id());
739 
740 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
741 	switch_mm_irqs_off(NULL, mm, current);
742 
743 	/*
744 	 * If breakpoints are enabled, disable them while the temporary mm is
745 	 * used. Userspace might set up watchpoints on addresses that are used
746 	 * in the temporary mm, which would lead to wrong signals being sent or
747 	 * crashes.
748 	 *
749 	 * Note that breakpoints are not disabled selectively, which also causes
750 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
751 	 * undesirable, but still seems reasonable as the code that runs in the
752 	 * temporary mm should be short.
753 	 */
754 	if (hw_breakpoint_active())
755 		hw_breakpoint_disable();
756 
757 	return temp_state;
758 }
759 
760 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
761 {
762 	lockdep_assert_irqs_disabled();
763 	switch_mm_irqs_off(NULL, prev_state.mm, current);
764 
765 	/*
766 	 * Restore the breakpoints if they were disabled before the temporary mm
767 	 * was loaded.
768 	 */
769 	if (hw_breakpoint_active())
770 		hw_breakpoint_restore();
771 }
772 
773 __ro_after_init struct mm_struct *poking_mm;
774 __ro_after_init unsigned long poking_addr;
775 
776 static void *__text_poke(void *addr, const void *opcode, size_t len)
777 {
778 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
779 	struct page *pages[2] = {NULL};
780 	temp_mm_state_t prev;
781 	unsigned long flags;
782 	pte_t pte, *ptep;
783 	spinlock_t *ptl;
784 	pgprot_t pgprot;
785 
786 	/*
787 	 * While boot memory allocator is running we cannot use struct pages as
788 	 * they are not yet initialized. There is no way to recover.
789 	 */
790 	BUG_ON(!after_bootmem);
791 
792 	if (!core_kernel_text((unsigned long)addr)) {
793 		pages[0] = vmalloc_to_page(addr);
794 		if (cross_page_boundary)
795 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
796 	} else {
797 		pages[0] = virt_to_page(addr);
798 		WARN_ON(!PageReserved(pages[0]));
799 		if (cross_page_boundary)
800 			pages[1] = virt_to_page(addr + PAGE_SIZE);
801 	}
802 	/*
803 	 * If something went wrong, crash and burn since recovery paths are not
804 	 * implemented.
805 	 */
806 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
807 
808 	/*
809 	 * Map the page without the global bit, as TLB flushing is done with
810 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
811 	 */
812 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
813 
814 	/*
815 	 * The lock is not really needed, but this allows to avoid open-coding.
816 	 */
817 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
818 
819 	/*
820 	 * This must not fail; preallocated in poking_init().
821 	 */
822 	VM_BUG_ON(!ptep);
823 
824 	local_irq_save(flags);
825 
826 	pte = mk_pte(pages[0], pgprot);
827 	set_pte_at(poking_mm, poking_addr, ptep, pte);
828 
829 	if (cross_page_boundary) {
830 		pte = mk_pte(pages[1], pgprot);
831 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
832 	}
833 
834 	/*
835 	 * Loading the temporary mm behaves as a compiler barrier, which
836 	 * guarantees that the PTE will be set at the time memcpy() is done.
837 	 */
838 	prev = use_temporary_mm(poking_mm);
839 
840 	kasan_disable_current();
841 	memcpy((u8 *)poking_addr + offset_in_page(addr), opcode, len);
842 	kasan_enable_current();
843 
844 	/*
845 	 * Ensure that the PTE is only cleared after the instructions of memcpy
846 	 * were issued by using a compiler barrier.
847 	 */
848 	barrier();
849 
850 	pte_clear(poking_mm, poking_addr, ptep);
851 	if (cross_page_boundary)
852 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
853 
854 	/*
855 	 * Loading the previous page-table hierarchy requires a serializing
856 	 * instruction that already allows the core to see the updated version.
857 	 * Xen-PV is assumed to serialize execution in a similar manner.
858 	 */
859 	unuse_temporary_mm(prev);
860 
861 	/*
862 	 * Flushing the TLB might involve IPIs, which would require enabled
863 	 * IRQs, but not if the mm is not used, as it is in this point.
864 	 */
865 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
866 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
867 			   PAGE_SHIFT, false);
868 
869 	/*
870 	 * If the text does not match what we just wrote then something is
871 	 * fundamentally screwy; there's nothing we can really do about that.
872 	 */
873 	BUG_ON(memcmp(addr, opcode, len));
874 
875 	local_irq_restore(flags);
876 	pte_unmap_unlock(ptep, ptl);
877 	return addr;
878 }
879 
880 /**
881  * text_poke - Update instructions on a live kernel
882  * @addr: address to modify
883  * @opcode: source of the copy
884  * @len: length to copy
885  *
886  * Only atomic text poke/set should be allowed when not doing early patching.
887  * It means the size must be writable atomically and the address must be aligned
888  * in a way that permits an atomic write. It also makes sure we fit on a single
889  * page.
890  *
891  * Note that the caller must ensure that if the modified code is part of a
892  * module, the module would not be removed during poking. This can be achieved
893  * by registering a module notifier, and ordering module removal and patching
894  * trough a mutex.
895  */
896 void *text_poke(void *addr, const void *opcode, size_t len)
897 {
898 	lockdep_assert_held(&text_mutex);
899 
900 	return __text_poke(addr, opcode, len);
901 }
902 
903 /**
904  * text_poke_kgdb - Update instructions on a live kernel by kgdb
905  * @addr: address to modify
906  * @opcode: source of the copy
907  * @len: length to copy
908  *
909  * Only atomic text poke/set should be allowed when not doing early patching.
910  * It means the size must be writable atomically and the address must be aligned
911  * in a way that permits an atomic write. It also makes sure we fit on a single
912  * page.
913  *
914  * Context: should only be used by kgdb, which ensures no other core is running,
915  *	    despite the fact it does not hold the text_mutex.
916  */
917 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
918 {
919 	return __text_poke(addr, opcode, len);
920 }
921 
922 static void do_sync_core(void *info)
923 {
924 	sync_core();
925 }
926 
927 void text_poke_sync(void)
928 {
929 	on_each_cpu(do_sync_core, NULL, 1);
930 }
931 
932 struct text_poke_loc {
933 	s32 rel_addr; /* addr := _stext + rel_addr */
934 	s32 rel32;
935 	u8 opcode;
936 	const u8 text[POKE_MAX_OPCODE_SIZE];
937 	u8 old;
938 };
939 
940 struct bp_patching_desc {
941 	struct text_poke_loc *vec;
942 	int nr_entries;
943 	atomic_t refs;
944 };
945 
946 static struct bp_patching_desc *bp_desc;
947 
948 static __always_inline
949 struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
950 {
951 	struct bp_patching_desc *desc = __READ_ONCE(*descp); /* rcu_dereference */
952 
953 	if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
954 		return NULL;
955 
956 	return desc;
957 }
958 
959 static __always_inline void put_desc(struct bp_patching_desc *desc)
960 {
961 	smp_mb__before_atomic();
962 	arch_atomic_dec(&desc->refs);
963 }
964 
965 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
966 {
967 	return _stext + tp->rel_addr;
968 }
969 
970 static __always_inline int patch_cmp(const void *key, const void *elt)
971 {
972 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
973 
974 	if (key < text_poke_addr(tp))
975 		return -1;
976 	if (key > text_poke_addr(tp))
977 		return 1;
978 	return 0;
979 }
980 
981 noinstr int poke_int3_handler(struct pt_regs *regs)
982 {
983 	struct bp_patching_desc *desc;
984 	struct text_poke_loc *tp;
985 	int len, ret = 0;
986 	void *ip;
987 
988 	if (user_mode(regs))
989 		return 0;
990 
991 	/*
992 	 * Having observed our INT3 instruction, we now must observe
993 	 * bp_desc:
994 	 *
995 	 *	bp_desc = desc			INT3
996 	 *	WMB				RMB
997 	 *	write INT3			if (desc)
998 	 */
999 	smp_rmb();
1000 
1001 	desc = try_get_desc(&bp_desc);
1002 	if (!desc)
1003 		return 0;
1004 
1005 	/*
1006 	 * Discount the INT3. See text_poke_bp_batch().
1007 	 */
1008 	ip = (void *) regs->ip - INT3_INSN_SIZE;
1009 
1010 	/*
1011 	 * Skip the binary search if there is a single member in the vector.
1012 	 */
1013 	if (unlikely(desc->nr_entries > 1)) {
1014 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
1015 				      sizeof(struct text_poke_loc),
1016 				      patch_cmp);
1017 		if (!tp)
1018 			goto out_put;
1019 	} else {
1020 		tp = desc->vec;
1021 		if (text_poke_addr(tp) != ip)
1022 			goto out_put;
1023 	}
1024 
1025 	len = text_opcode_size(tp->opcode);
1026 	ip += len;
1027 
1028 	switch (tp->opcode) {
1029 	case INT3_INSN_OPCODE:
1030 		/*
1031 		 * Someone poked an explicit INT3, they'll want to handle it,
1032 		 * do not consume.
1033 		 */
1034 		goto out_put;
1035 
1036 	case RET_INSN_OPCODE:
1037 		int3_emulate_ret(regs);
1038 		break;
1039 
1040 	case CALL_INSN_OPCODE:
1041 		int3_emulate_call(regs, (long)ip + tp->rel32);
1042 		break;
1043 
1044 	case JMP32_INSN_OPCODE:
1045 	case JMP8_INSN_OPCODE:
1046 		int3_emulate_jmp(regs, (long)ip + tp->rel32);
1047 		break;
1048 
1049 	default:
1050 		BUG();
1051 	}
1052 
1053 	ret = 1;
1054 
1055 out_put:
1056 	put_desc(desc);
1057 	return ret;
1058 }
1059 
1060 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
1061 static struct text_poke_loc tp_vec[TP_VEC_MAX];
1062 static int tp_vec_nr;
1063 
1064 /**
1065  * text_poke_bp_batch() -- update instructions on live kernel on SMP
1066  * @tp:			vector of instructions to patch
1067  * @nr_entries:		number of entries in the vector
1068  *
1069  * Modify multi-byte instruction by using int3 breakpoint on SMP.
1070  * We completely avoid stop_machine() here, and achieve the
1071  * synchronization using int3 breakpoint.
1072  *
1073  * The way it is done:
1074  *	- For each entry in the vector:
1075  *		- add a int3 trap to the address that will be patched
1076  *	- sync cores
1077  *	- For each entry in the vector:
1078  *		- update all but the first byte of the patched range
1079  *	- sync cores
1080  *	- For each entry in the vector:
1081  *		- replace the first byte (int3) by the first byte of
1082  *		  replacing opcode
1083  *	- sync cores
1084  */
1085 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
1086 {
1087 	struct bp_patching_desc desc = {
1088 		.vec = tp,
1089 		.nr_entries = nr_entries,
1090 		.refs = ATOMIC_INIT(1),
1091 	};
1092 	unsigned char int3 = INT3_INSN_OPCODE;
1093 	unsigned int i;
1094 	int do_sync;
1095 
1096 	lockdep_assert_held(&text_mutex);
1097 
1098 	smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
1099 
1100 	/*
1101 	 * Corresponding read barrier in int3 notifier for making sure the
1102 	 * nr_entries and handler are correctly ordered wrt. patching.
1103 	 */
1104 	smp_wmb();
1105 
1106 	/*
1107 	 * First step: add a int3 trap to the address that will be patched.
1108 	 */
1109 	for (i = 0; i < nr_entries; i++) {
1110 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
1111 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
1112 	}
1113 
1114 	text_poke_sync();
1115 
1116 	/*
1117 	 * Second step: update all but the first byte of the patched range.
1118 	 */
1119 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1120 		u8 old[POKE_MAX_OPCODE_SIZE] = { tp[i].old, };
1121 		int len = text_opcode_size(tp[i].opcode);
1122 
1123 		if (len - INT3_INSN_SIZE > 0) {
1124 			memcpy(old + INT3_INSN_SIZE,
1125 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1126 			       len - INT3_INSN_SIZE);
1127 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1128 				  (const char *)tp[i].text + INT3_INSN_SIZE,
1129 				  len - INT3_INSN_SIZE);
1130 			do_sync++;
1131 		}
1132 
1133 		/*
1134 		 * Emit a perf event to record the text poke, primarily to
1135 		 * support Intel PT decoding which must walk the executable code
1136 		 * to reconstruct the trace. The flow up to here is:
1137 		 *   - write INT3 byte
1138 		 *   - IPI-SYNC
1139 		 *   - write instruction tail
1140 		 * At this point the actual control flow will be through the
1141 		 * INT3 and handler and not hit the old or new instruction.
1142 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
1143 		 * can still be decoded. Subsequently:
1144 		 *   - emit RECORD_TEXT_POKE with the new instruction
1145 		 *   - IPI-SYNC
1146 		 *   - write first byte
1147 		 *   - IPI-SYNC
1148 		 * So before the text poke event timestamp, the decoder will see
1149 		 * either the old instruction flow or FUP/TIP of INT3. After the
1150 		 * text poke event timestamp, the decoder will see either the
1151 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
1152 		 * use the timestamp as the point at which to modify the
1153 		 * executable code.
1154 		 * The old instruction is recorded so that the event can be
1155 		 * processed forwards or backwards.
1156 		 */
1157 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len,
1158 				     tp[i].text, len);
1159 	}
1160 
1161 	if (do_sync) {
1162 		/*
1163 		 * According to Intel, this core syncing is very likely
1164 		 * not necessary and we'd be safe even without it. But
1165 		 * better safe than sorry (plus there's not only Intel).
1166 		 */
1167 		text_poke_sync();
1168 	}
1169 
1170 	/*
1171 	 * Third step: replace the first byte (int3) by the first byte of
1172 	 * replacing opcode.
1173 	 */
1174 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1175 		if (tp[i].text[0] == INT3_INSN_OPCODE)
1176 			continue;
1177 
1178 		text_poke(text_poke_addr(&tp[i]), tp[i].text, INT3_INSN_SIZE);
1179 		do_sync++;
1180 	}
1181 
1182 	if (do_sync)
1183 		text_poke_sync();
1184 
1185 	/*
1186 	 * Remove and synchronize_rcu(), except we have a very primitive
1187 	 * refcount based completion.
1188 	 */
1189 	WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
1190 	if (!atomic_dec_and_test(&desc.refs))
1191 		atomic_cond_read_acquire(&desc.refs, !VAL);
1192 }
1193 
1194 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
1195 			       const void *opcode, size_t len, const void *emulate)
1196 {
1197 	struct insn insn;
1198 	int ret;
1199 
1200 	memcpy((void *)tp->text, opcode, len);
1201 	if (!emulate)
1202 		emulate = opcode;
1203 
1204 	ret = insn_decode_kernel(&insn, emulate);
1205 
1206 	BUG_ON(ret < 0);
1207 	BUG_ON(len != insn.length);
1208 
1209 	tp->rel_addr = addr - (void *)_stext;
1210 	tp->opcode = insn.opcode.bytes[0];
1211 
1212 	switch (tp->opcode) {
1213 	case INT3_INSN_OPCODE:
1214 	case RET_INSN_OPCODE:
1215 		break;
1216 
1217 	case CALL_INSN_OPCODE:
1218 	case JMP32_INSN_OPCODE:
1219 	case JMP8_INSN_OPCODE:
1220 		tp->rel32 = insn.immediate.value;
1221 		break;
1222 
1223 	default: /* assume NOP */
1224 		switch (len) {
1225 		case 2: /* NOP2 -- emulate as JMP8+0 */
1226 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1227 			tp->opcode = JMP8_INSN_OPCODE;
1228 			tp->rel32 = 0;
1229 			break;
1230 
1231 		case 5: /* NOP5 -- emulate as JMP32+0 */
1232 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1233 			tp->opcode = JMP32_INSN_OPCODE;
1234 			tp->rel32 = 0;
1235 			break;
1236 
1237 		default: /* unknown instruction */
1238 			BUG();
1239 		}
1240 		break;
1241 	}
1242 }
1243 
1244 /*
1245  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
1246  * early if needed.
1247  */
1248 static bool tp_order_fail(void *addr)
1249 {
1250 	struct text_poke_loc *tp;
1251 
1252 	if (!tp_vec_nr)
1253 		return false;
1254 
1255 	if (!addr) /* force */
1256 		return true;
1257 
1258 	tp = &tp_vec[tp_vec_nr - 1];
1259 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
1260 		return true;
1261 
1262 	return false;
1263 }
1264 
1265 static void text_poke_flush(void *addr)
1266 {
1267 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
1268 		text_poke_bp_batch(tp_vec, tp_vec_nr);
1269 		tp_vec_nr = 0;
1270 	}
1271 }
1272 
1273 void text_poke_finish(void)
1274 {
1275 	text_poke_flush(NULL);
1276 }
1277 
1278 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
1279 {
1280 	struct text_poke_loc *tp;
1281 
1282 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1283 		text_poke_early(addr, opcode, len);
1284 		return;
1285 	}
1286 
1287 	text_poke_flush(addr);
1288 
1289 	tp = &tp_vec[tp_vec_nr++];
1290 	text_poke_loc_init(tp, addr, opcode, len, emulate);
1291 }
1292 
1293 /**
1294  * text_poke_bp() -- update instructions on live kernel on SMP
1295  * @addr:	address to patch
1296  * @opcode:	opcode of new instruction
1297  * @len:	length to copy
1298  * @emulate:	instruction to be emulated
1299  *
1300  * Update a single instruction with the vector in the stack, avoiding
1301  * dynamically allocated memory. This function should be used when it is
1302  * not possible to allocate memory.
1303  */
1304 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
1305 {
1306 	struct text_poke_loc tp;
1307 
1308 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1309 		text_poke_early(addr, opcode, len);
1310 		return;
1311 	}
1312 
1313 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
1314 	text_poke_bp_batch(&tp, 1);
1315 }
1316