xref: /linux/arch/x86/kernel/alternative.c (revision 87c9c16317882dd6dbbc07e349bc3223e14f3244)
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 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  * "noinline" to cause control flow change and thus invalidate I$ and
187  * cause refetch after modification.
188  */
189 static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr)
190 {
191 	unsigned long flags;
192 	struct insn insn;
193 	int nop, i = 0;
194 
195 	/*
196 	 * Jump over the non-NOP insns, the remaining bytes must be single-byte
197 	 * NOPs, optimize them.
198 	 */
199 	for (;;) {
200 		if (insn_decode_kernel(&insn, &instr[i]))
201 			return;
202 
203 		if (insn.length == 1 && insn.opcode.bytes[0] == 0x90)
204 			break;
205 
206 		if ((i += insn.length) >= a->instrlen)
207 			return;
208 	}
209 
210 	for (nop = i; i < a->instrlen; i++) {
211 		if (WARN_ONCE(instr[i] != 0x90, "Not a NOP at 0x%px\n", &instr[i]))
212 			return;
213 	}
214 
215 	local_irq_save(flags);
216 	add_nops(instr + nop, i - nop);
217 	local_irq_restore(flags);
218 
219 	DUMP_BYTES(instr, a->instrlen, "%px: [%d:%d) optimized NOPs: ",
220 		   instr, nop, a->instrlen);
221 }
222 
223 /*
224  * Replace instructions with better alternatives for this CPU type. This runs
225  * before SMP is initialized to avoid SMP problems with self modifying code.
226  * This implies that asymmetric systems where APs have less capabilities than
227  * the boot processor are not handled. Tough. Make sure you disable such
228  * features by hand.
229  *
230  * Marked "noinline" to cause control flow change and thus insn cache
231  * to refetch changed I$ lines.
232  */
233 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
234 						  struct alt_instr *end)
235 {
236 	struct alt_instr *a;
237 	u8 *instr, *replacement;
238 	u8 insn_buff[MAX_PATCH_LEN];
239 
240 	DPRINTK("alt table %px, -> %px", start, end);
241 	/*
242 	 * The scan order should be from start to end. A later scanned
243 	 * alternative code can overwrite previously scanned alternative code.
244 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
245 	 * patch code.
246 	 *
247 	 * So be careful if you want to change the scan order to any other
248 	 * order.
249 	 */
250 	for (a = start; a < end; a++) {
251 		int insn_buff_sz = 0;
252 		/* Mask away "NOT" flag bit for feature to test. */
253 		u16 feature = a->cpuid & ~ALTINSTR_FLAG_INV;
254 
255 		instr = (u8 *)&a->instr_offset + a->instr_offset;
256 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
257 		BUG_ON(a->instrlen > sizeof(insn_buff));
258 		BUG_ON(feature >= (NCAPINTS + NBUGINTS) * 32);
259 
260 		/*
261 		 * Patch if either:
262 		 * - feature is present
263 		 * - feature not present but ALTINSTR_FLAG_INV is set to mean,
264 		 *   patch if feature is *NOT* present.
265 		 */
266 		if (!boot_cpu_has(feature) == !(a->cpuid & ALTINSTR_FLAG_INV))
267 			goto next;
268 
269 		DPRINTK("feat: %s%d*32+%d, old: (%pS (%px) len: %d), repl: (%px, len: %d)",
270 			(a->cpuid & ALTINSTR_FLAG_INV) ? "!" : "",
271 			feature >> 5,
272 			feature & 0x1f,
273 			instr, instr, a->instrlen,
274 			replacement, a->replacementlen);
275 
276 		DUMP_BYTES(instr, a->instrlen, "%px: old_insn: ", instr);
277 		DUMP_BYTES(replacement, a->replacementlen, "%px: rpl_insn: ", replacement);
278 
279 		memcpy(insn_buff, replacement, a->replacementlen);
280 		insn_buff_sz = a->replacementlen;
281 
282 		/*
283 		 * 0xe8 is a relative jump; fix the offset.
284 		 *
285 		 * Instruction length is checked before the opcode to avoid
286 		 * accessing uninitialized bytes for zero-length replacements.
287 		 */
288 		if (a->replacementlen == 5 && *insn_buff == 0xe8) {
289 			*(s32 *)(insn_buff + 1) += replacement - instr;
290 			DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
291 				*(s32 *)(insn_buff + 1),
292 				(unsigned long)instr + *(s32 *)(insn_buff + 1) + 5);
293 		}
294 
295 		if (a->replacementlen && is_jmp(replacement[0]))
296 			recompute_jump(a, instr, replacement, insn_buff);
297 
298 		for (; insn_buff_sz < a->instrlen; insn_buff_sz++)
299 			insn_buff[insn_buff_sz] = 0x90;
300 
301 		DUMP_BYTES(insn_buff, insn_buff_sz, "%px: final_insn: ", instr);
302 
303 		text_poke_early(instr, insn_buff, insn_buff_sz);
304 
305 next:
306 		optimize_nops(a, instr);
307 	}
308 }
309 
310 #ifdef CONFIG_SMP
311 static void alternatives_smp_lock(const s32 *start, const s32 *end,
312 				  u8 *text, u8 *text_end)
313 {
314 	const s32 *poff;
315 
316 	for (poff = start; poff < end; poff++) {
317 		u8 *ptr = (u8 *)poff + *poff;
318 
319 		if (!*poff || ptr < text || ptr >= text_end)
320 			continue;
321 		/* turn DS segment override prefix into lock prefix */
322 		if (*ptr == 0x3e)
323 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
324 	}
325 }
326 
327 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
328 				    u8 *text, u8 *text_end)
329 {
330 	const s32 *poff;
331 
332 	for (poff = start; poff < end; poff++) {
333 		u8 *ptr = (u8 *)poff + *poff;
334 
335 		if (!*poff || ptr < text || ptr >= text_end)
336 			continue;
337 		/* turn lock prefix into DS segment override prefix */
338 		if (*ptr == 0xf0)
339 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
340 	}
341 }
342 
343 struct smp_alt_module {
344 	/* what is this ??? */
345 	struct module	*mod;
346 	char		*name;
347 
348 	/* ptrs to lock prefixes */
349 	const s32	*locks;
350 	const s32	*locks_end;
351 
352 	/* .text segment, needed to avoid patching init code ;) */
353 	u8		*text;
354 	u8		*text_end;
355 
356 	struct list_head next;
357 };
358 static LIST_HEAD(smp_alt_modules);
359 static bool uniproc_patched = false;	/* protected by text_mutex */
360 
361 void __init_or_module alternatives_smp_module_add(struct module *mod,
362 						  char *name,
363 						  void *locks, void *locks_end,
364 						  void *text,  void *text_end)
365 {
366 	struct smp_alt_module *smp;
367 
368 	mutex_lock(&text_mutex);
369 	if (!uniproc_patched)
370 		goto unlock;
371 
372 	if (num_possible_cpus() == 1)
373 		/* Don't bother remembering, we'll never have to undo it. */
374 		goto smp_unlock;
375 
376 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
377 	if (NULL == smp)
378 		/* we'll run the (safe but slow) SMP code then ... */
379 		goto unlock;
380 
381 	smp->mod	= mod;
382 	smp->name	= name;
383 	smp->locks	= locks;
384 	smp->locks_end	= locks_end;
385 	smp->text	= text;
386 	smp->text_end	= text_end;
387 	DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
388 		smp->locks, smp->locks_end,
389 		smp->text, smp->text_end, smp->name);
390 
391 	list_add_tail(&smp->next, &smp_alt_modules);
392 smp_unlock:
393 	alternatives_smp_unlock(locks, locks_end, text, text_end);
394 unlock:
395 	mutex_unlock(&text_mutex);
396 }
397 
398 void __init_or_module alternatives_smp_module_del(struct module *mod)
399 {
400 	struct smp_alt_module *item;
401 
402 	mutex_lock(&text_mutex);
403 	list_for_each_entry(item, &smp_alt_modules, next) {
404 		if (mod != item->mod)
405 			continue;
406 		list_del(&item->next);
407 		kfree(item);
408 		break;
409 	}
410 	mutex_unlock(&text_mutex);
411 }
412 
413 void alternatives_enable_smp(void)
414 {
415 	struct smp_alt_module *mod;
416 
417 	/* Why bother if there are no other CPUs? */
418 	BUG_ON(num_possible_cpus() == 1);
419 
420 	mutex_lock(&text_mutex);
421 
422 	if (uniproc_patched) {
423 		pr_info("switching to SMP code\n");
424 		BUG_ON(num_online_cpus() != 1);
425 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
426 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
427 		list_for_each_entry(mod, &smp_alt_modules, next)
428 			alternatives_smp_lock(mod->locks, mod->locks_end,
429 					      mod->text, mod->text_end);
430 		uniproc_patched = false;
431 	}
432 	mutex_unlock(&text_mutex);
433 }
434 
435 /*
436  * Return 1 if the address range is reserved for SMP-alternatives.
437  * Must hold text_mutex.
438  */
439 int alternatives_text_reserved(void *start, void *end)
440 {
441 	struct smp_alt_module *mod;
442 	const s32 *poff;
443 	u8 *text_start = start;
444 	u8 *text_end = end;
445 
446 	lockdep_assert_held(&text_mutex);
447 
448 	list_for_each_entry(mod, &smp_alt_modules, next) {
449 		if (mod->text > text_end || mod->text_end < text_start)
450 			continue;
451 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
452 			const u8 *ptr = (const u8 *)poff + *poff;
453 
454 			if (text_start <= ptr && text_end > ptr)
455 				return 1;
456 		}
457 	}
458 
459 	return 0;
460 }
461 #endif /* CONFIG_SMP */
462 
463 #ifdef CONFIG_PARAVIRT
464 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
465 				     struct paravirt_patch_site *end)
466 {
467 	struct paravirt_patch_site *p;
468 	char insn_buff[MAX_PATCH_LEN];
469 
470 	for (p = start; p < end; p++) {
471 		unsigned int used;
472 
473 		BUG_ON(p->len > MAX_PATCH_LEN);
474 		/* prep the buffer with the original instructions */
475 		memcpy(insn_buff, p->instr, p->len);
476 		used = paravirt_patch(p->type, insn_buff, (unsigned long)p->instr, p->len);
477 
478 		BUG_ON(used > p->len);
479 
480 		/* Pad the rest with nops */
481 		add_nops(insn_buff + used, p->len - used);
482 		text_poke_early(p->instr, insn_buff, p->len);
483 	}
484 }
485 extern struct paravirt_patch_site __start_parainstructions[],
486 	__stop_parainstructions[];
487 #endif	/* CONFIG_PARAVIRT */
488 
489 /*
490  * Self-test for the INT3 based CALL emulation code.
491  *
492  * This exercises int3_emulate_call() to make sure INT3 pt_regs are set up
493  * properly and that there is a stack gap between the INT3 frame and the
494  * previous context. Without this gap doing a virtual PUSH on the interrupted
495  * stack would corrupt the INT3 IRET frame.
496  *
497  * See entry_{32,64}.S for more details.
498  */
499 
500 /*
501  * We define the int3_magic() function in assembly to control the calling
502  * convention such that we can 'call' it from assembly.
503  */
504 
505 extern void int3_magic(unsigned int *ptr); /* defined in asm */
506 
507 asm (
508 "	.pushsection	.init.text, \"ax\", @progbits\n"
509 "	.type		int3_magic, @function\n"
510 "int3_magic:\n"
511 "	movl	$1, (%" _ASM_ARG1 ")\n"
512 "	ret\n"
513 "	.size		int3_magic, .-int3_magic\n"
514 "	.popsection\n"
515 );
516 
517 extern __initdata unsigned long int3_selftest_ip; /* defined in asm below */
518 
519 static int __init
520 int3_exception_notify(struct notifier_block *self, unsigned long val, void *data)
521 {
522 	struct die_args *args = data;
523 	struct pt_regs *regs = args->regs;
524 
525 	if (!regs || user_mode(regs))
526 		return NOTIFY_DONE;
527 
528 	if (val != DIE_INT3)
529 		return NOTIFY_DONE;
530 
531 	if (regs->ip - INT3_INSN_SIZE != int3_selftest_ip)
532 		return NOTIFY_DONE;
533 
534 	int3_emulate_call(regs, (unsigned long)&int3_magic);
535 	return NOTIFY_STOP;
536 }
537 
538 static void __init int3_selftest(void)
539 {
540 	static __initdata struct notifier_block int3_exception_nb = {
541 		.notifier_call	= int3_exception_notify,
542 		.priority	= INT_MAX-1, /* last */
543 	};
544 	unsigned int val = 0;
545 
546 	BUG_ON(register_die_notifier(&int3_exception_nb));
547 
548 	/*
549 	 * Basically: int3_magic(&val); but really complicated :-)
550 	 *
551 	 * Stick the address of the INT3 instruction into int3_selftest_ip,
552 	 * then trigger the INT3, padded with NOPs to match a CALL instruction
553 	 * length.
554 	 */
555 	asm volatile ("1: int3; nop; nop; nop; nop\n\t"
556 		      ".pushsection .init.data,\"aw\"\n\t"
557 		      ".align " __ASM_SEL(4, 8) "\n\t"
558 		      ".type int3_selftest_ip, @object\n\t"
559 		      ".size int3_selftest_ip, " __ASM_SEL(4, 8) "\n\t"
560 		      "int3_selftest_ip:\n\t"
561 		      __ASM_SEL(.long, .quad) " 1b\n\t"
562 		      ".popsection\n\t"
563 		      : ASM_CALL_CONSTRAINT
564 		      : __ASM_SEL_RAW(a, D) (&val)
565 		      : "memory");
566 
567 	BUG_ON(val != 1);
568 
569 	unregister_die_notifier(&int3_exception_nb);
570 }
571 
572 void __init alternative_instructions(void)
573 {
574 	int3_selftest();
575 
576 	/*
577 	 * The patching is not fully atomic, so try to avoid local
578 	 * interruptions that might execute the to be patched code.
579 	 * Other CPUs are not running.
580 	 */
581 	stop_nmi();
582 
583 	/*
584 	 * Don't stop machine check exceptions while patching.
585 	 * MCEs only happen when something got corrupted and in this
586 	 * case we must do something about the corruption.
587 	 * Ignoring it is worse than an unlikely patching race.
588 	 * Also machine checks tend to be broadcast and if one CPU
589 	 * goes into machine check the others follow quickly, so we don't
590 	 * expect a machine check to cause undue problems during to code
591 	 * patching.
592 	 */
593 
594 	/*
595 	 * Paravirt patching and alternative patching can be combined to
596 	 * replace a function call with a short direct code sequence (e.g.
597 	 * by setting a constant return value instead of doing that in an
598 	 * external function).
599 	 * In order to make this work the following sequence is required:
600 	 * 1. set (artificial) features depending on used paravirt
601 	 *    functions which can later influence alternative patching
602 	 * 2. apply paravirt patching (generally replacing an indirect
603 	 *    function call with a direct one)
604 	 * 3. apply alternative patching (e.g. replacing a direct function
605 	 *    call with a custom code sequence)
606 	 * Doing paravirt patching after alternative patching would clobber
607 	 * the optimization of the custom code with a function call again.
608 	 */
609 	paravirt_set_cap();
610 
611 	/*
612 	 * First patch paravirt functions, such that we overwrite the indirect
613 	 * call with the direct call.
614 	 */
615 	apply_paravirt(__parainstructions, __parainstructions_end);
616 
617 	/*
618 	 * Then patch alternatives, such that those paravirt calls that are in
619 	 * alternatives can be overwritten by their immediate fragments.
620 	 */
621 	apply_alternatives(__alt_instructions, __alt_instructions_end);
622 
623 #ifdef CONFIG_SMP
624 	/* Patch to UP if other cpus not imminent. */
625 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
626 		uniproc_patched = true;
627 		alternatives_smp_module_add(NULL, "core kernel",
628 					    __smp_locks, __smp_locks_end,
629 					    _text, _etext);
630 	}
631 
632 	if (!uniproc_patched || num_possible_cpus() == 1) {
633 		free_init_pages("SMP alternatives",
634 				(unsigned long)__smp_locks,
635 				(unsigned long)__smp_locks_end);
636 	}
637 #endif
638 
639 	restart_nmi();
640 	alternatives_patched = 1;
641 }
642 
643 /**
644  * text_poke_early - Update instructions on a live kernel at boot time
645  * @addr: address to modify
646  * @opcode: source of the copy
647  * @len: length to copy
648  *
649  * When you use this code to patch more than one byte of an instruction
650  * you need to make sure that other CPUs cannot execute this code in parallel.
651  * Also no thread must be currently preempted in the middle of these
652  * instructions. And on the local CPU you need to be protected against NMI or
653  * MCE handlers seeing an inconsistent instruction while you patch.
654  */
655 void __init_or_module text_poke_early(void *addr, const void *opcode,
656 				      size_t len)
657 {
658 	unsigned long flags;
659 
660 	if (boot_cpu_has(X86_FEATURE_NX) &&
661 	    is_module_text_address((unsigned long)addr)) {
662 		/*
663 		 * Modules text is marked initially as non-executable, so the
664 		 * code cannot be running and speculative code-fetches are
665 		 * prevented. Just change the code.
666 		 */
667 		memcpy(addr, opcode, len);
668 	} else {
669 		local_irq_save(flags);
670 		memcpy(addr, opcode, len);
671 		local_irq_restore(flags);
672 		sync_core();
673 
674 		/*
675 		 * Could also do a CLFLUSH here to speed up CPU recovery; but
676 		 * that causes hangs on some VIA CPUs.
677 		 */
678 	}
679 }
680 
681 typedef struct {
682 	struct mm_struct *mm;
683 } temp_mm_state_t;
684 
685 /*
686  * Using a temporary mm allows to set temporary mappings that are not accessible
687  * by other CPUs. Such mappings are needed to perform sensitive memory writes
688  * that override the kernel memory protections (e.g., W^X), without exposing the
689  * temporary page-table mappings that are required for these write operations to
690  * other CPUs. Using a temporary mm also allows to avoid TLB shootdowns when the
691  * mapping is torn down.
692  *
693  * Context: The temporary mm needs to be used exclusively by a single core. To
694  *          harden security IRQs must be disabled while the temporary mm is
695  *          loaded, thereby preventing interrupt handler bugs from overriding
696  *          the kernel memory protection.
697  */
698 static inline temp_mm_state_t use_temporary_mm(struct mm_struct *mm)
699 {
700 	temp_mm_state_t temp_state;
701 
702 	lockdep_assert_irqs_disabled();
703 
704 	/*
705 	 * Make sure not to be in TLB lazy mode, as otherwise we'll end up
706 	 * with a stale address space WITHOUT being in lazy mode after
707 	 * restoring the previous mm.
708 	 */
709 	if (this_cpu_read(cpu_tlbstate_shared.is_lazy))
710 		leave_mm(smp_processor_id());
711 
712 	temp_state.mm = this_cpu_read(cpu_tlbstate.loaded_mm);
713 	switch_mm_irqs_off(NULL, mm, current);
714 
715 	/*
716 	 * If breakpoints are enabled, disable them while the temporary mm is
717 	 * used. Userspace might set up watchpoints on addresses that are used
718 	 * in the temporary mm, which would lead to wrong signals being sent or
719 	 * crashes.
720 	 *
721 	 * Note that breakpoints are not disabled selectively, which also causes
722 	 * kernel breakpoints (e.g., perf's) to be disabled. This might be
723 	 * undesirable, but still seems reasonable as the code that runs in the
724 	 * temporary mm should be short.
725 	 */
726 	if (hw_breakpoint_active())
727 		hw_breakpoint_disable();
728 
729 	return temp_state;
730 }
731 
732 static inline void unuse_temporary_mm(temp_mm_state_t prev_state)
733 {
734 	lockdep_assert_irqs_disabled();
735 	switch_mm_irqs_off(NULL, prev_state.mm, current);
736 
737 	/*
738 	 * Restore the breakpoints if they were disabled before the temporary mm
739 	 * was loaded.
740 	 */
741 	if (hw_breakpoint_active())
742 		hw_breakpoint_restore();
743 }
744 
745 __ro_after_init struct mm_struct *poking_mm;
746 __ro_after_init unsigned long poking_addr;
747 
748 static void *__text_poke(void *addr, const void *opcode, size_t len)
749 {
750 	bool cross_page_boundary = offset_in_page(addr) + len > PAGE_SIZE;
751 	struct page *pages[2] = {NULL};
752 	temp_mm_state_t prev;
753 	unsigned long flags;
754 	pte_t pte, *ptep;
755 	spinlock_t *ptl;
756 	pgprot_t pgprot;
757 
758 	/*
759 	 * While boot memory allocator is running we cannot use struct pages as
760 	 * they are not yet initialized. There is no way to recover.
761 	 */
762 	BUG_ON(!after_bootmem);
763 
764 	if (!core_kernel_text((unsigned long)addr)) {
765 		pages[0] = vmalloc_to_page(addr);
766 		if (cross_page_boundary)
767 			pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
768 	} else {
769 		pages[0] = virt_to_page(addr);
770 		WARN_ON(!PageReserved(pages[0]));
771 		if (cross_page_boundary)
772 			pages[1] = virt_to_page(addr + PAGE_SIZE);
773 	}
774 	/*
775 	 * If something went wrong, crash and burn since recovery paths are not
776 	 * implemented.
777 	 */
778 	BUG_ON(!pages[0] || (cross_page_boundary && !pages[1]));
779 
780 	/*
781 	 * Map the page without the global bit, as TLB flushing is done with
782 	 * flush_tlb_mm_range(), which is intended for non-global PTEs.
783 	 */
784 	pgprot = __pgprot(pgprot_val(PAGE_KERNEL) & ~_PAGE_GLOBAL);
785 
786 	/*
787 	 * The lock is not really needed, but this allows to avoid open-coding.
788 	 */
789 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
790 
791 	/*
792 	 * This must not fail; preallocated in poking_init().
793 	 */
794 	VM_BUG_ON(!ptep);
795 
796 	local_irq_save(flags);
797 
798 	pte = mk_pte(pages[0], pgprot);
799 	set_pte_at(poking_mm, poking_addr, ptep, pte);
800 
801 	if (cross_page_boundary) {
802 		pte = mk_pte(pages[1], pgprot);
803 		set_pte_at(poking_mm, poking_addr + PAGE_SIZE, ptep + 1, pte);
804 	}
805 
806 	/*
807 	 * Loading the temporary mm behaves as a compiler barrier, which
808 	 * guarantees that the PTE will be set at the time memcpy() is done.
809 	 */
810 	prev = use_temporary_mm(poking_mm);
811 
812 	kasan_disable_current();
813 	memcpy((u8 *)poking_addr + offset_in_page(addr), opcode, len);
814 	kasan_enable_current();
815 
816 	/*
817 	 * Ensure that the PTE is only cleared after the instructions of memcpy
818 	 * were issued by using a compiler barrier.
819 	 */
820 	barrier();
821 
822 	pte_clear(poking_mm, poking_addr, ptep);
823 	if (cross_page_boundary)
824 		pte_clear(poking_mm, poking_addr + PAGE_SIZE, ptep + 1);
825 
826 	/*
827 	 * Loading the previous page-table hierarchy requires a serializing
828 	 * instruction that already allows the core to see the updated version.
829 	 * Xen-PV is assumed to serialize execution in a similar manner.
830 	 */
831 	unuse_temporary_mm(prev);
832 
833 	/*
834 	 * Flushing the TLB might involve IPIs, which would require enabled
835 	 * IRQs, but not if the mm is not used, as it is in this point.
836 	 */
837 	flush_tlb_mm_range(poking_mm, poking_addr, poking_addr +
838 			   (cross_page_boundary ? 2 : 1) * PAGE_SIZE,
839 			   PAGE_SHIFT, false);
840 
841 	/*
842 	 * If the text does not match what we just wrote then something is
843 	 * fundamentally screwy; there's nothing we can really do about that.
844 	 */
845 	BUG_ON(memcmp(addr, opcode, len));
846 
847 	local_irq_restore(flags);
848 	pte_unmap_unlock(ptep, ptl);
849 	return addr;
850 }
851 
852 /**
853  * text_poke - Update instructions on a live kernel
854  * @addr: address to modify
855  * @opcode: source of the copy
856  * @len: length to copy
857  *
858  * Only atomic text poke/set should be allowed when not doing early patching.
859  * It means the size must be writable atomically and the address must be aligned
860  * in a way that permits an atomic write. It also makes sure we fit on a single
861  * page.
862  *
863  * Note that the caller must ensure that if the modified code is part of a
864  * module, the module would not be removed during poking. This can be achieved
865  * by registering a module notifier, and ordering module removal and patching
866  * trough a mutex.
867  */
868 void *text_poke(void *addr, const void *opcode, size_t len)
869 {
870 	lockdep_assert_held(&text_mutex);
871 
872 	return __text_poke(addr, opcode, len);
873 }
874 
875 /**
876  * text_poke_kgdb - Update instructions on a live kernel by kgdb
877  * @addr: address to modify
878  * @opcode: source of the copy
879  * @len: length to copy
880  *
881  * Only atomic text poke/set should be allowed when not doing early patching.
882  * It means the size must be writable atomically and the address must be aligned
883  * in a way that permits an atomic write. It also makes sure we fit on a single
884  * page.
885  *
886  * Context: should only be used by kgdb, which ensures no other core is running,
887  *	    despite the fact it does not hold the text_mutex.
888  */
889 void *text_poke_kgdb(void *addr, const void *opcode, size_t len)
890 {
891 	return __text_poke(addr, opcode, len);
892 }
893 
894 static void do_sync_core(void *info)
895 {
896 	sync_core();
897 }
898 
899 void text_poke_sync(void)
900 {
901 	on_each_cpu(do_sync_core, NULL, 1);
902 }
903 
904 struct text_poke_loc {
905 	s32 rel_addr; /* addr := _stext + rel_addr */
906 	s32 rel32;
907 	u8 opcode;
908 	const u8 text[POKE_MAX_OPCODE_SIZE];
909 	u8 old;
910 };
911 
912 struct bp_patching_desc {
913 	struct text_poke_loc *vec;
914 	int nr_entries;
915 	atomic_t refs;
916 };
917 
918 static struct bp_patching_desc *bp_desc;
919 
920 static __always_inline
921 struct bp_patching_desc *try_get_desc(struct bp_patching_desc **descp)
922 {
923 	struct bp_patching_desc *desc = __READ_ONCE(*descp); /* rcu_dereference */
924 
925 	if (!desc || !arch_atomic_inc_not_zero(&desc->refs))
926 		return NULL;
927 
928 	return desc;
929 }
930 
931 static __always_inline void put_desc(struct bp_patching_desc *desc)
932 {
933 	smp_mb__before_atomic();
934 	arch_atomic_dec(&desc->refs);
935 }
936 
937 static __always_inline void *text_poke_addr(struct text_poke_loc *tp)
938 {
939 	return _stext + tp->rel_addr;
940 }
941 
942 static __always_inline int patch_cmp(const void *key, const void *elt)
943 {
944 	struct text_poke_loc *tp = (struct text_poke_loc *) elt;
945 
946 	if (key < text_poke_addr(tp))
947 		return -1;
948 	if (key > text_poke_addr(tp))
949 		return 1;
950 	return 0;
951 }
952 
953 noinstr int poke_int3_handler(struct pt_regs *regs)
954 {
955 	struct bp_patching_desc *desc;
956 	struct text_poke_loc *tp;
957 	int len, ret = 0;
958 	void *ip;
959 
960 	if (user_mode(regs))
961 		return 0;
962 
963 	/*
964 	 * Having observed our INT3 instruction, we now must observe
965 	 * bp_desc:
966 	 *
967 	 *	bp_desc = desc			INT3
968 	 *	WMB				RMB
969 	 *	write INT3			if (desc)
970 	 */
971 	smp_rmb();
972 
973 	desc = try_get_desc(&bp_desc);
974 	if (!desc)
975 		return 0;
976 
977 	/*
978 	 * Discount the INT3. See text_poke_bp_batch().
979 	 */
980 	ip = (void *) regs->ip - INT3_INSN_SIZE;
981 
982 	/*
983 	 * Skip the binary search if there is a single member in the vector.
984 	 */
985 	if (unlikely(desc->nr_entries > 1)) {
986 		tp = __inline_bsearch(ip, desc->vec, desc->nr_entries,
987 				      sizeof(struct text_poke_loc),
988 				      patch_cmp);
989 		if (!tp)
990 			goto out_put;
991 	} else {
992 		tp = desc->vec;
993 		if (text_poke_addr(tp) != ip)
994 			goto out_put;
995 	}
996 
997 	len = text_opcode_size(tp->opcode);
998 	ip += len;
999 
1000 	switch (tp->opcode) {
1001 	case INT3_INSN_OPCODE:
1002 		/*
1003 		 * Someone poked an explicit INT3, they'll want to handle it,
1004 		 * do not consume.
1005 		 */
1006 		goto out_put;
1007 
1008 	case RET_INSN_OPCODE:
1009 		int3_emulate_ret(regs);
1010 		break;
1011 
1012 	case CALL_INSN_OPCODE:
1013 		int3_emulate_call(regs, (long)ip + tp->rel32);
1014 		break;
1015 
1016 	case JMP32_INSN_OPCODE:
1017 	case JMP8_INSN_OPCODE:
1018 		int3_emulate_jmp(regs, (long)ip + tp->rel32);
1019 		break;
1020 
1021 	default:
1022 		BUG();
1023 	}
1024 
1025 	ret = 1;
1026 
1027 out_put:
1028 	put_desc(desc);
1029 	return ret;
1030 }
1031 
1032 #define TP_VEC_MAX (PAGE_SIZE / sizeof(struct text_poke_loc))
1033 static struct text_poke_loc tp_vec[TP_VEC_MAX];
1034 static int tp_vec_nr;
1035 
1036 /**
1037  * text_poke_bp_batch() -- update instructions on live kernel on SMP
1038  * @tp:			vector of instructions to patch
1039  * @nr_entries:		number of entries in the vector
1040  *
1041  * Modify multi-byte instruction by using int3 breakpoint on SMP.
1042  * We completely avoid stop_machine() here, and achieve the
1043  * synchronization using int3 breakpoint.
1044  *
1045  * The way it is done:
1046  *	- For each entry in the vector:
1047  *		- add a int3 trap to the address that will be patched
1048  *	- sync cores
1049  *	- For each entry in the vector:
1050  *		- update all but the first byte of the patched range
1051  *	- sync cores
1052  *	- For each entry in the vector:
1053  *		- replace the first byte (int3) by the first byte of
1054  *		  replacing opcode
1055  *	- sync cores
1056  */
1057 static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries)
1058 {
1059 	struct bp_patching_desc desc = {
1060 		.vec = tp,
1061 		.nr_entries = nr_entries,
1062 		.refs = ATOMIC_INIT(1),
1063 	};
1064 	unsigned char int3 = INT3_INSN_OPCODE;
1065 	unsigned int i;
1066 	int do_sync;
1067 
1068 	lockdep_assert_held(&text_mutex);
1069 
1070 	smp_store_release(&bp_desc, &desc); /* rcu_assign_pointer */
1071 
1072 	/*
1073 	 * Corresponding read barrier in int3 notifier for making sure the
1074 	 * nr_entries and handler are correctly ordered wrt. patching.
1075 	 */
1076 	smp_wmb();
1077 
1078 	/*
1079 	 * First step: add a int3 trap to the address that will be patched.
1080 	 */
1081 	for (i = 0; i < nr_entries; i++) {
1082 		tp[i].old = *(u8 *)text_poke_addr(&tp[i]);
1083 		text_poke(text_poke_addr(&tp[i]), &int3, INT3_INSN_SIZE);
1084 	}
1085 
1086 	text_poke_sync();
1087 
1088 	/*
1089 	 * Second step: update all but the first byte of the patched range.
1090 	 */
1091 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1092 		u8 old[POKE_MAX_OPCODE_SIZE] = { tp[i].old, };
1093 		int len = text_opcode_size(tp[i].opcode);
1094 
1095 		if (len - INT3_INSN_SIZE > 0) {
1096 			memcpy(old + INT3_INSN_SIZE,
1097 			       text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1098 			       len - INT3_INSN_SIZE);
1099 			text_poke(text_poke_addr(&tp[i]) + INT3_INSN_SIZE,
1100 				  (const char *)tp[i].text + INT3_INSN_SIZE,
1101 				  len - INT3_INSN_SIZE);
1102 			do_sync++;
1103 		}
1104 
1105 		/*
1106 		 * Emit a perf event to record the text poke, primarily to
1107 		 * support Intel PT decoding which must walk the executable code
1108 		 * to reconstruct the trace. The flow up to here is:
1109 		 *   - write INT3 byte
1110 		 *   - IPI-SYNC
1111 		 *   - write instruction tail
1112 		 * At this point the actual control flow will be through the
1113 		 * INT3 and handler and not hit the old or new instruction.
1114 		 * Intel PT outputs FUP/TIP packets for the INT3, so the flow
1115 		 * can still be decoded. Subsequently:
1116 		 *   - emit RECORD_TEXT_POKE with the new instruction
1117 		 *   - IPI-SYNC
1118 		 *   - write first byte
1119 		 *   - IPI-SYNC
1120 		 * So before the text poke event timestamp, the decoder will see
1121 		 * either the old instruction flow or FUP/TIP of INT3. After the
1122 		 * text poke event timestamp, the decoder will see either the
1123 		 * new instruction flow or FUP/TIP of INT3. Thus decoders can
1124 		 * use the timestamp as the point at which to modify the
1125 		 * executable code.
1126 		 * The old instruction is recorded so that the event can be
1127 		 * processed forwards or backwards.
1128 		 */
1129 		perf_event_text_poke(text_poke_addr(&tp[i]), old, len,
1130 				     tp[i].text, len);
1131 	}
1132 
1133 	if (do_sync) {
1134 		/*
1135 		 * According to Intel, this core syncing is very likely
1136 		 * not necessary and we'd be safe even without it. But
1137 		 * better safe than sorry (plus there's not only Intel).
1138 		 */
1139 		text_poke_sync();
1140 	}
1141 
1142 	/*
1143 	 * Third step: replace the first byte (int3) by the first byte of
1144 	 * replacing opcode.
1145 	 */
1146 	for (do_sync = 0, i = 0; i < nr_entries; i++) {
1147 		if (tp[i].text[0] == INT3_INSN_OPCODE)
1148 			continue;
1149 
1150 		text_poke(text_poke_addr(&tp[i]), tp[i].text, INT3_INSN_SIZE);
1151 		do_sync++;
1152 	}
1153 
1154 	if (do_sync)
1155 		text_poke_sync();
1156 
1157 	/*
1158 	 * Remove and synchronize_rcu(), except we have a very primitive
1159 	 * refcount based completion.
1160 	 */
1161 	WRITE_ONCE(bp_desc, NULL); /* RCU_INIT_POINTER */
1162 	if (!atomic_dec_and_test(&desc.refs))
1163 		atomic_cond_read_acquire(&desc.refs, !VAL);
1164 }
1165 
1166 static void text_poke_loc_init(struct text_poke_loc *tp, void *addr,
1167 			       const void *opcode, size_t len, const void *emulate)
1168 {
1169 	struct insn insn;
1170 	int ret;
1171 
1172 	memcpy((void *)tp->text, opcode, len);
1173 	if (!emulate)
1174 		emulate = opcode;
1175 
1176 	ret = insn_decode_kernel(&insn, emulate);
1177 
1178 	BUG_ON(ret < 0);
1179 	BUG_ON(len != insn.length);
1180 
1181 	tp->rel_addr = addr - (void *)_stext;
1182 	tp->opcode = insn.opcode.bytes[0];
1183 
1184 	switch (tp->opcode) {
1185 	case INT3_INSN_OPCODE:
1186 	case RET_INSN_OPCODE:
1187 		break;
1188 
1189 	case CALL_INSN_OPCODE:
1190 	case JMP32_INSN_OPCODE:
1191 	case JMP8_INSN_OPCODE:
1192 		tp->rel32 = insn.immediate.value;
1193 		break;
1194 
1195 	default: /* assume NOP */
1196 		switch (len) {
1197 		case 2: /* NOP2 -- emulate as JMP8+0 */
1198 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1199 			tp->opcode = JMP8_INSN_OPCODE;
1200 			tp->rel32 = 0;
1201 			break;
1202 
1203 		case 5: /* NOP5 -- emulate as JMP32+0 */
1204 			BUG_ON(memcmp(emulate, x86_nops[len], len));
1205 			tp->opcode = JMP32_INSN_OPCODE;
1206 			tp->rel32 = 0;
1207 			break;
1208 
1209 		default: /* unknown instruction */
1210 			BUG();
1211 		}
1212 		break;
1213 	}
1214 }
1215 
1216 /*
1217  * We hard rely on the tp_vec being ordered; ensure this is so by flushing
1218  * early if needed.
1219  */
1220 static bool tp_order_fail(void *addr)
1221 {
1222 	struct text_poke_loc *tp;
1223 
1224 	if (!tp_vec_nr)
1225 		return false;
1226 
1227 	if (!addr) /* force */
1228 		return true;
1229 
1230 	tp = &tp_vec[tp_vec_nr - 1];
1231 	if ((unsigned long)text_poke_addr(tp) > (unsigned long)addr)
1232 		return true;
1233 
1234 	return false;
1235 }
1236 
1237 static void text_poke_flush(void *addr)
1238 {
1239 	if (tp_vec_nr == TP_VEC_MAX || tp_order_fail(addr)) {
1240 		text_poke_bp_batch(tp_vec, tp_vec_nr);
1241 		tp_vec_nr = 0;
1242 	}
1243 }
1244 
1245 void text_poke_finish(void)
1246 {
1247 	text_poke_flush(NULL);
1248 }
1249 
1250 void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const void *emulate)
1251 {
1252 	struct text_poke_loc *tp;
1253 
1254 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1255 		text_poke_early(addr, opcode, len);
1256 		return;
1257 	}
1258 
1259 	text_poke_flush(addr);
1260 
1261 	tp = &tp_vec[tp_vec_nr++];
1262 	text_poke_loc_init(tp, addr, opcode, len, emulate);
1263 }
1264 
1265 /**
1266  * text_poke_bp() -- update instructions on live kernel on SMP
1267  * @addr:	address to patch
1268  * @opcode:	opcode of new instruction
1269  * @len:	length to copy
1270  * @emulate:	instruction to be emulated
1271  *
1272  * Update a single instruction with the vector in the stack, avoiding
1273  * dynamically allocated memory. This function should be used when it is
1274  * not possible to allocate memory.
1275  */
1276 void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *emulate)
1277 {
1278 	struct text_poke_loc tp;
1279 
1280 	if (unlikely(system_state == SYSTEM_BOOTING)) {
1281 		text_poke_early(addr, opcode, len);
1282 		return;
1283 	}
1284 
1285 	text_poke_loc_init(&tp, addr, opcode, len, emulate);
1286 	text_poke_bp_batch(&tp, 1);
1287 }
1288