xref: /linux/arch/x86/kernel/alternative.c (revision 140eb5227767c6754742020a16d2691222b9c19b)
1 #define pr_fmt(fmt) "SMP alternatives: " fmt
2 
3 #include <linux/module.h>
4 #include <linux/sched.h>
5 #include <linux/mutex.h>
6 #include <linux/list.h>
7 #include <linux/stringify.h>
8 #include <linux/mm.h>
9 #include <linux/vmalloc.h>
10 #include <linux/memory.h>
11 #include <linux/stop_machine.h>
12 #include <linux/slab.h>
13 #include <linux/kdebug.h>
14 #include <asm/text-patching.h>
15 #include <asm/alternative.h>
16 #include <asm/sections.h>
17 #include <asm/pgtable.h>
18 #include <asm/mce.h>
19 #include <asm/nmi.h>
20 #include <asm/cacheflush.h>
21 #include <asm/tlbflush.h>
22 #include <asm/io.h>
23 #include <asm/fixmap.h>
24 
25 int __read_mostly alternatives_patched;
26 
27 EXPORT_SYMBOL_GPL(alternatives_patched);
28 
29 #define MAX_PATCH_LEN (255-1)
30 
31 static int __initdata_or_module debug_alternative;
32 
33 static int __init debug_alt(char *str)
34 {
35 	debug_alternative = 1;
36 	return 1;
37 }
38 __setup("debug-alternative", debug_alt);
39 
40 static int noreplace_smp;
41 
42 static int __init setup_noreplace_smp(char *str)
43 {
44 	noreplace_smp = 1;
45 	return 1;
46 }
47 __setup("noreplace-smp", setup_noreplace_smp);
48 
49 #ifdef CONFIG_PARAVIRT
50 static int __initdata_or_module noreplace_paravirt = 0;
51 
52 static int __init setup_noreplace_paravirt(char *str)
53 {
54 	noreplace_paravirt = 1;
55 	return 1;
56 }
57 __setup("noreplace-paravirt", setup_noreplace_paravirt);
58 #endif
59 
60 #define DPRINTK(fmt, args...)						\
61 do {									\
62 	if (debug_alternative)						\
63 		printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args);	\
64 } while (0)
65 
66 #define DUMP_BYTES(buf, len, fmt, args...)				\
67 do {									\
68 	if (unlikely(debug_alternative)) {				\
69 		int j;							\
70 									\
71 		if (!(len))						\
72 			break;						\
73 									\
74 		printk(KERN_DEBUG fmt, ##args);				\
75 		for (j = 0; j < (len) - 1; j++)				\
76 			printk(KERN_CONT "%02hhx ", buf[j]);		\
77 		printk(KERN_CONT "%02hhx\n", buf[j]);			\
78 	}								\
79 } while (0)
80 
81 /*
82  * Each GENERIC_NOPX is of X bytes, and defined as an array of bytes
83  * that correspond to that nop. Getting from one nop to the next, we
84  * add to the array the offset that is equal to the sum of all sizes of
85  * nops preceding the one we are after.
86  *
87  * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
88  * nice symmetry of sizes of the previous nops.
89  */
90 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
91 static const unsigned char intelnops[] =
92 {
93 	GENERIC_NOP1,
94 	GENERIC_NOP2,
95 	GENERIC_NOP3,
96 	GENERIC_NOP4,
97 	GENERIC_NOP5,
98 	GENERIC_NOP6,
99 	GENERIC_NOP7,
100 	GENERIC_NOP8,
101 	GENERIC_NOP5_ATOMIC
102 };
103 static const unsigned char * const intel_nops[ASM_NOP_MAX+2] =
104 {
105 	NULL,
106 	intelnops,
107 	intelnops + 1,
108 	intelnops + 1 + 2,
109 	intelnops + 1 + 2 + 3,
110 	intelnops + 1 + 2 + 3 + 4,
111 	intelnops + 1 + 2 + 3 + 4 + 5,
112 	intelnops + 1 + 2 + 3 + 4 + 5 + 6,
113 	intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
114 	intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
115 };
116 #endif
117 
118 #ifdef K8_NOP1
119 static const unsigned char k8nops[] =
120 {
121 	K8_NOP1,
122 	K8_NOP2,
123 	K8_NOP3,
124 	K8_NOP4,
125 	K8_NOP5,
126 	K8_NOP6,
127 	K8_NOP7,
128 	K8_NOP8,
129 	K8_NOP5_ATOMIC
130 };
131 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
132 {
133 	NULL,
134 	k8nops,
135 	k8nops + 1,
136 	k8nops + 1 + 2,
137 	k8nops + 1 + 2 + 3,
138 	k8nops + 1 + 2 + 3 + 4,
139 	k8nops + 1 + 2 + 3 + 4 + 5,
140 	k8nops + 1 + 2 + 3 + 4 + 5 + 6,
141 	k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
142 	k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
143 };
144 #endif
145 
146 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
147 static const unsigned char k7nops[] =
148 {
149 	K7_NOP1,
150 	K7_NOP2,
151 	K7_NOP3,
152 	K7_NOP4,
153 	K7_NOP5,
154 	K7_NOP6,
155 	K7_NOP7,
156 	K7_NOP8,
157 	K7_NOP5_ATOMIC
158 };
159 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
160 {
161 	NULL,
162 	k7nops,
163 	k7nops + 1,
164 	k7nops + 1 + 2,
165 	k7nops + 1 + 2 + 3,
166 	k7nops + 1 + 2 + 3 + 4,
167 	k7nops + 1 + 2 + 3 + 4 + 5,
168 	k7nops + 1 + 2 + 3 + 4 + 5 + 6,
169 	k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
170 	k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
171 };
172 #endif
173 
174 #ifdef P6_NOP1
175 static const unsigned char p6nops[] =
176 {
177 	P6_NOP1,
178 	P6_NOP2,
179 	P6_NOP3,
180 	P6_NOP4,
181 	P6_NOP5,
182 	P6_NOP6,
183 	P6_NOP7,
184 	P6_NOP8,
185 	P6_NOP5_ATOMIC
186 };
187 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
188 {
189 	NULL,
190 	p6nops,
191 	p6nops + 1,
192 	p6nops + 1 + 2,
193 	p6nops + 1 + 2 + 3,
194 	p6nops + 1 + 2 + 3 + 4,
195 	p6nops + 1 + 2 + 3 + 4 + 5,
196 	p6nops + 1 + 2 + 3 + 4 + 5 + 6,
197 	p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
198 	p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
199 };
200 #endif
201 
202 /* Initialize these to a safe default */
203 #ifdef CONFIG_X86_64
204 const unsigned char * const *ideal_nops = p6_nops;
205 #else
206 const unsigned char * const *ideal_nops = intel_nops;
207 #endif
208 
209 void __init arch_init_ideal_nops(void)
210 {
211 	switch (boot_cpu_data.x86_vendor) {
212 	case X86_VENDOR_INTEL:
213 		/*
214 		 * Due to a decoder implementation quirk, some
215 		 * specific Intel CPUs actually perform better with
216 		 * the "k8_nops" than with the SDM-recommended NOPs.
217 		 */
218 		if (boot_cpu_data.x86 == 6 &&
219 		    boot_cpu_data.x86_model >= 0x0f &&
220 		    boot_cpu_data.x86_model != 0x1c &&
221 		    boot_cpu_data.x86_model != 0x26 &&
222 		    boot_cpu_data.x86_model != 0x27 &&
223 		    boot_cpu_data.x86_model < 0x30) {
224 			ideal_nops = k8_nops;
225 		} else if (boot_cpu_has(X86_FEATURE_NOPL)) {
226 			   ideal_nops = p6_nops;
227 		} else {
228 #ifdef CONFIG_X86_64
229 			ideal_nops = k8_nops;
230 #else
231 			ideal_nops = intel_nops;
232 #endif
233 		}
234 		break;
235 
236 	case X86_VENDOR_AMD:
237 		if (boot_cpu_data.x86 > 0xf) {
238 			ideal_nops = p6_nops;
239 			return;
240 		}
241 
242 		/* fall through */
243 
244 	default:
245 #ifdef CONFIG_X86_64
246 		ideal_nops = k8_nops;
247 #else
248 		if (boot_cpu_has(X86_FEATURE_K8))
249 			ideal_nops = k8_nops;
250 		else if (boot_cpu_has(X86_FEATURE_K7))
251 			ideal_nops = k7_nops;
252 		else
253 			ideal_nops = intel_nops;
254 #endif
255 	}
256 }
257 
258 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
259 static void __init_or_module add_nops(void *insns, unsigned int len)
260 {
261 	while (len > 0) {
262 		unsigned int noplen = len;
263 		if (noplen > ASM_NOP_MAX)
264 			noplen = ASM_NOP_MAX;
265 		memcpy(insns, ideal_nops[noplen], noplen);
266 		insns += noplen;
267 		len -= noplen;
268 	}
269 }
270 
271 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
272 extern s32 __smp_locks[], __smp_locks_end[];
273 void *text_poke_early(void *addr, const void *opcode, size_t len);
274 
275 /*
276  * Are we looking at a near JMP with a 1 or 4-byte displacement.
277  */
278 static inline bool is_jmp(const u8 opcode)
279 {
280 	return opcode == 0xeb || opcode == 0xe9;
281 }
282 
283 static void __init_or_module
284 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf)
285 {
286 	u8 *next_rip, *tgt_rip;
287 	s32 n_dspl, o_dspl;
288 	int repl_len;
289 
290 	if (a->replacementlen != 5)
291 		return;
292 
293 	o_dspl = *(s32 *)(insnbuf + 1);
294 
295 	/* next_rip of the replacement JMP */
296 	next_rip = repl_insn + a->replacementlen;
297 	/* target rip of the replacement JMP */
298 	tgt_rip  = next_rip + o_dspl;
299 	n_dspl = tgt_rip - orig_insn;
300 
301 	DPRINTK("target RIP: %p, new_displ: 0x%x", tgt_rip, n_dspl);
302 
303 	if (tgt_rip - orig_insn >= 0) {
304 		if (n_dspl - 2 <= 127)
305 			goto two_byte_jmp;
306 		else
307 			goto five_byte_jmp;
308 	/* negative offset */
309 	} else {
310 		if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
311 			goto two_byte_jmp;
312 		else
313 			goto five_byte_jmp;
314 	}
315 
316 two_byte_jmp:
317 	n_dspl -= 2;
318 
319 	insnbuf[0] = 0xeb;
320 	insnbuf[1] = (s8)n_dspl;
321 	add_nops(insnbuf + 2, 3);
322 
323 	repl_len = 2;
324 	goto done;
325 
326 five_byte_jmp:
327 	n_dspl -= 5;
328 
329 	insnbuf[0] = 0xe9;
330 	*(s32 *)&insnbuf[1] = n_dspl;
331 
332 	repl_len = 5;
333 
334 done:
335 
336 	DPRINTK("final displ: 0x%08x, JMP 0x%lx",
337 		n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
338 }
339 
340 /*
341  * "noinline" to cause control flow change and thus invalidate I$ and
342  * cause refetch after modification.
343  */
344 static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr)
345 {
346 	unsigned long flags;
347 	int i;
348 
349 	for (i = 0; i < a->padlen; i++) {
350 		if (instr[i] != 0x90)
351 			return;
352 	}
353 
354 	local_irq_save(flags);
355 	add_nops(instr + (a->instrlen - a->padlen), a->padlen);
356 	local_irq_restore(flags);
357 
358 	DUMP_BYTES(instr, a->instrlen, "%p: [%d:%d) optimized NOPs: ",
359 		   instr, a->instrlen - a->padlen, a->padlen);
360 }
361 
362 /*
363  * Replace instructions with better alternatives for this CPU type. This runs
364  * before SMP is initialized to avoid SMP problems with self modifying code.
365  * This implies that asymmetric systems where APs have less capabilities than
366  * the boot processor are not handled. Tough. Make sure you disable such
367  * features by hand.
368  *
369  * Marked "noinline" to cause control flow change and thus insn cache
370  * to refetch changed I$ lines.
371  */
372 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
373 						  struct alt_instr *end)
374 {
375 	struct alt_instr *a;
376 	u8 *instr, *replacement;
377 	u8 insnbuf[MAX_PATCH_LEN];
378 
379 	DPRINTK("alt table %p -> %p", start, end);
380 	/*
381 	 * The scan order should be from start to end. A later scanned
382 	 * alternative code can overwrite previously scanned alternative code.
383 	 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
384 	 * patch code.
385 	 *
386 	 * So be careful if you want to change the scan order to any other
387 	 * order.
388 	 */
389 	for (a = start; a < end; a++) {
390 		int insnbuf_sz = 0;
391 
392 		instr = (u8 *)&a->instr_offset + a->instr_offset;
393 		replacement = (u8 *)&a->repl_offset + a->repl_offset;
394 		BUG_ON(a->instrlen > sizeof(insnbuf));
395 		BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
396 		if (!boot_cpu_has(a->cpuid)) {
397 			if (a->padlen > 1)
398 				optimize_nops(a, instr);
399 
400 			continue;
401 		}
402 
403 		DPRINTK("feat: %d*32+%d, old: (%p, len: %d), repl: (%p, len: %d), pad: %d",
404 			a->cpuid >> 5,
405 			a->cpuid & 0x1f,
406 			instr, a->instrlen,
407 			replacement, a->replacementlen, a->padlen);
408 
409 		DUMP_BYTES(instr, a->instrlen, "%p: old_insn: ", instr);
410 		DUMP_BYTES(replacement, a->replacementlen, "%p: rpl_insn: ", replacement);
411 
412 		memcpy(insnbuf, replacement, a->replacementlen);
413 		insnbuf_sz = a->replacementlen;
414 
415 		/*
416 		 * 0xe8 is a relative jump; fix the offset.
417 		 *
418 		 * Instruction length is checked before the opcode to avoid
419 		 * accessing uninitialized bytes for zero-length replacements.
420 		 */
421 		if (a->replacementlen == 5 && *insnbuf == 0xe8) {
422 			*(s32 *)(insnbuf + 1) += replacement - instr;
423 			DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
424 				*(s32 *)(insnbuf + 1),
425 				(unsigned long)instr + *(s32 *)(insnbuf + 1) + 5);
426 		}
427 
428 		if (a->replacementlen && is_jmp(replacement[0]))
429 			recompute_jump(a, instr, replacement, insnbuf);
430 
431 		if (a->instrlen > a->replacementlen) {
432 			add_nops(insnbuf + a->replacementlen,
433 				 a->instrlen - a->replacementlen);
434 			insnbuf_sz += a->instrlen - a->replacementlen;
435 		}
436 		DUMP_BYTES(insnbuf, insnbuf_sz, "%p: final_insn: ", instr);
437 
438 		text_poke_early(instr, insnbuf, insnbuf_sz);
439 	}
440 }
441 
442 #ifdef CONFIG_SMP
443 static void alternatives_smp_lock(const s32 *start, const s32 *end,
444 				  u8 *text, u8 *text_end)
445 {
446 	const s32 *poff;
447 
448 	for (poff = start; poff < end; poff++) {
449 		u8 *ptr = (u8 *)poff + *poff;
450 
451 		if (!*poff || ptr < text || ptr >= text_end)
452 			continue;
453 		/* turn DS segment override prefix into lock prefix */
454 		if (*ptr == 0x3e)
455 			text_poke(ptr, ((unsigned char []){0xf0}), 1);
456 	}
457 }
458 
459 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
460 				    u8 *text, u8 *text_end)
461 {
462 	const s32 *poff;
463 
464 	for (poff = start; poff < end; poff++) {
465 		u8 *ptr = (u8 *)poff + *poff;
466 
467 		if (!*poff || ptr < text || ptr >= text_end)
468 			continue;
469 		/* turn lock prefix into DS segment override prefix */
470 		if (*ptr == 0xf0)
471 			text_poke(ptr, ((unsigned char []){0x3E}), 1);
472 	}
473 }
474 
475 struct smp_alt_module {
476 	/* what is this ??? */
477 	struct module	*mod;
478 	char		*name;
479 
480 	/* ptrs to lock prefixes */
481 	const s32	*locks;
482 	const s32	*locks_end;
483 
484 	/* .text segment, needed to avoid patching init code ;) */
485 	u8		*text;
486 	u8		*text_end;
487 
488 	struct list_head next;
489 };
490 static LIST_HEAD(smp_alt_modules);
491 static bool uniproc_patched = false;	/* protected by text_mutex */
492 
493 void __init_or_module alternatives_smp_module_add(struct module *mod,
494 						  char *name,
495 						  void *locks, void *locks_end,
496 						  void *text,  void *text_end)
497 {
498 	struct smp_alt_module *smp;
499 
500 	mutex_lock(&text_mutex);
501 	if (!uniproc_patched)
502 		goto unlock;
503 
504 	if (num_possible_cpus() == 1)
505 		/* Don't bother remembering, we'll never have to undo it. */
506 		goto smp_unlock;
507 
508 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
509 	if (NULL == smp)
510 		/* we'll run the (safe but slow) SMP code then ... */
511 		goto unlock;
512 
513 	smp->mod	= mod;
514 	smp->name	= name;
515 	smp->locks	= locks;
516 	smp->locks_end	= locks_end;
517 	smp->text	= text;
518 	smp->text_end	= text_end;
519 	DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
520 		smp->locks, smp->locks_end,
521 		smp->text, smp->text_end, smp->name);
522 
523 	list_add_tail(&smp->next, &smp_alt_modules);
524 smp_unlock:
525 	alternatives_smp_unlock(locks, locks_end, text, text_end);
526 unlock:
527 	mutex_unlock(&text_mutex);
528 }
529 
530 void __init_or_module alternatives_smp_module_del(struct module *mod)
531 {
532 	struct smp_alt_module *item;
533 
534 	mutex_lock(&text_mutex);
535 	list_for_each_entry(item, &smp_alt_modules, next) {
536 		if (mod != item->mod)
537 			continue;
538 		list_del(&item->next);
539 		kfree(item);
540 		break;
541 	}
542 	mutex_unlock(&text_mutex);
543 }
544 
545 void alternatives_enable_smp(void)
546 {
547 	struct smp_alt_module *mod;
548 
549 	/* Why bother if there are no other CPUs? */
550 	BUG_ON(num_possible_cpus() == 1);
551 
552 	mutex_lock(&text_mutex);
553 
554 	if (uniproc_patched) {
555 		pr_info("switching to SMP code\n");
556 		BUG_ON(num_online_cpus() != 1);
557 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
558 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
559 		list_for_each_entry(mod, &smp_alt_modules, next)
560 			alternatives_smp_lock(mod->locks, mod->locks_end,
561 					      mod->text, mod->text_end);
562 		uniproc_patched = false;
563 	}
564 	mutex_unlock(&text_mutex);
565 }
566 
567 /*
568  * Return 1 if the address range is reserved for SMP-alternatives.
569  * Must hold text_mutex.
570  */
571 int alternatives_text_reserved(void *start, void *end)
572 {
573 	struct smp_alt_module *mod;
574 	const s32 *poff;
575 	u8 *text_start = start;
576 	u8 *text_end = end;
577 
578 	lockdep_assert_held(&text_mutex);
579 
580 	list_for_each_entry(mod, &smp_alt_modules, next) {
581 		if (mod->text > text_end || mod->text_end < text_start)
582 			continue;
583 		for (poff = mod->locks; poff < mod->locks_end; poff++) {
584 			const u8 *ptr = (const u8 *)poff + *poff;
585 
586 			if (text_start <= ptr && text_end > ptr)
587 				return 1;
588 		}
589 	}
590 
591 	return 0;
592 }
593 #endif /* CONFIG_SMP */
594 
595 #ifdef CONFIG_PARAVIRT
596 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
597 				     struct paravirt_patch_site *end)
598 {
599 	struct paravirt_patch_site *p;
600 	char insnbuf[MAX_PATCH_LEN];
601 
602 	if (noreplace_paravirt)
603 		return;
604 
605 	for (p = start; p < end; p++) {
606 		unsigned int used;
607 
608 		BUG_ON(p->len > MAX_PATCH_LEN);
609 		/* prep the buffer with the original instructions */
610 		memcpy(insnbuf, p->instr, p->len);
611 		used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
612 					 (unsigned long)p->instr, p->len);
613 
614 		BUG_ON(used > p->len);
615 
616 		/* Pad the rest with nops */
617 		add_nops(insnbuf + used, p->len - used);
618 		text_poke_early(p->instr, insnbuf, p->len);
619 	}
620 }
621 extern struct paravirt_patch_site __start_parainstructions[],
622 	__stop_parainstructions[];
623 #endif	/* CONFIG_PARAVIRT */
624 
625 void __init alternative_instructions(void)
626 {
627 	/* The patching is not fully atomic, so try to avoid local interruptions
628 	   that might execute the to be patched code.
629 	   Other CPUs are not running. */
630 	stop_nmi();
631 
632 	/*
633 	 * Don't stop machine check exceptions while patching.
634 	 * MCEs only happen when something got corrupted and in this
635 	 * case we must do something about the corruption.
636 	 * Ignoring it is worse than a unlikely patching race.
637 	 * Also machine checks tend to be broadcast and if one CPU
638 	 * goes into machine check the others follow quickly, so we don't
639 	 * expect a machine check to cause undue problems during to code
640 	 * patching.
641 	 */
642 
643 	apply_alternatives(__alt_instructions, __alt_instructions_end);
644 
645 #ifdef CONFIG_SMP
646 	/* Patch to UP if other cpus not imminent. */
647 	if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
648 		uniproc_patched = true;
649 		alternatives_smp_module_add(NULL, "core kernel",
650 					    __smp_locks, __smp_locks_end,
651 					    _text, _etext);
652 	}
653 
654 	if (!uniproc_patched || num_possible_cpus() == 1)
655 		free_init_pages("SMP alternatives",
656 				(unsigned long)__smp_locks,
657 				(unsigned long)__smp_locks_end);
658 #endif
659 
660 	apply_paravirt(__parainstructions, __parainstructions_end);
661 
662 	restart_nmi();
663 	alternatives_patched = 1;
664 }
665 
666 /**
667  * text_poke_early - Update instructions on a live kernel at boot time
668  * @addr: address to modify
669  * @opcode: source of the copy
670  * @len: length to copy
671  *
672  * When you use this code to patch more than one byte of an instruction
673  * you need to make sure that other CPUs cannot execute this code in parallel.
674  * Also no thread must be currently preempted in the middle of these
675  * instructions. And on the local CPU you need to be protected again NMI or MCE
676  * handlers seeing an inconsistent instruction while you patch.
677  */
678 void *__init_or_module text_poke_early(void *addr, const void *opcode,
679 					      size_t len)
680 {
681 	unsigned long flags;
682 	local_irq_save(flags);
683 	memcpy(addr, opcode, len);
684 	local_irq_restore(flags);
685 	/* Could also do a CLFLUSH here to speed up CPU recovery; but
686 	   that causes hangs on some VIA CPUs. */
687 	return addr;
688 }
689 
690 /**
691  * text_poke - Update instructions on a live kernel
692  * @addr: address to modify
693  * @opcode: source of the copy
694  * @len: length to copy
695  *
696  * Only atomic text poke/set should be allowed when not doing early patching.
697  * It means the size must be writable atomically and the address must be aligned
698  * in a way that permits an atomic write. It also makes sure we fit on a single
699  * page.
700  *
701  * Note: Must be called under text_mutex.
702  */
703 void *text_poke(void *addr, const void *opcode, size_t len)
704 {
705 	unsigned long flags;
706 	char *vaddr;
707 	struct page *pages[2];
708 	int i;
709 
710 	if (!core_kernel_text((unsigned long)addr)) {
711 		pages[0] = vmalloc_to_page(addr);
712 		pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
713 	} else {
714 		pages[0] = virt_to_page(addr);
715 		WARN_ON(!PageReserved(pages[0]));
716 		pages[1] = virt_to_page(addr + PAGE_SIZE);
717 	}
718 	BUG_ON(!pages[0]);
719 	local_irq_save(flags);
720 	set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
721 	if (pages[1])
722 		set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
723 	vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
724 	memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
725 	clear_fixmap(FIX_TEXT_POKE0);
726 	if (pages[1])
727 		clear_fixmap(FIX_TEXT_POKE1);
728 	local_flush_tlb();
729 	sync_core();
730 	/* Could also do a CLFLUSH here to speed up CPU recovery; but
731 	   that causes hangs on some VIA CPUs. */
732 	for (i = 0; i < len; i++)
733 		BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
734 	local_irq_restore(flags);
735 	return addr;
736 }
737 
738 static void do_sync_core(void *info)
739 {
740 	sync_core();
741 }
742 
743 static bool bp_patching_in_progress;
744 static void *bp_int3_handler, *bp_int3_addr;
745 
746 int poke_int3_handler(struct pt_regs *regs)
747 {
748 	/*
749 	 * Having observed our INT3 instruction, we now must observe
750 	 * bp_patching_in_progress.
751 	 *
752 	 * 	in_progress = TRUE		INT3
753 	 * 	WMB				RMB
754 	 * 	write INT3			if (in_progress)
755 	 *
756 	 * Idem for bp_int3_handler.
757 	 */
758 	smp_rmb();
759 
760 	if (likely(!bp_patching_in_progress))
761 		return 0;
762 
763 	if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr)
764 		return 0;
765 
766 	/* set up the specified breakpoint handler */
767 	regs->ip = (unsigned long) bp_int3_handler;
768 
769 	return 1;
770 
771 }
772 
773 /**
774  * text_poke_bp() -- update instructions on live kernel on SMP
775  * @addr:	address to patch
776  * @opcode:	opcode of new instruction
777  * @len:	length to copy
778  * @handler:	address to jump to when the temporary breakpoint is hit
779  *
780  * Modify multi-byte instruction by using int3 breakpoint on SMP.
781  * We completely avoid stop_machine() here, and achieve the
782  * synchronization using int3 breakpoint.
783  *
784  * The way it is done:
785  *	- add a int3 trap to the address that will be patched
786  *	- sync cores
787  *	- update all but the first byte of the patched range
788  *	- sync cores
789  *	- replace the first byte (int3) by the first byte of
790  *	  replacing opcode
791  *	- sync cores
792  *
793  * Note: must be called under text_mutex.
794  */
795 void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler)
796 {
797 	unsigned char int3 = 0xcc;
798 
799 	bp_int3_handler = handler;
800 	bp_int3_addr = (u8 *)addr + sizeof(int3);
801 	bp_patching_in_progress = true;
802 	/*
803 	 * Corresponding read barrier in int3 notifier for making sure the
804 	 * in_progress and handler are correctly ordered wrt. patching.
805 	 */
806 	smp_wmb();
807 
808 	text_poke(addr, &int3, sizeof(int3));
809 
810 	on_each_cpu(do_sync_core, NULL, 1);
811 
812 	if (len - sizeof(int3) > 0) {
813 		/* patch all but the first byte */
814 		text_poke((char *)addr + sizeof(int3),
815 			  (const char *) opcode + sizeof(int3),
816 			  len - sizeof(int3));
817 		/*
818 		 * According to Intel, this core syncing is very likely
819 		 * not necessary and we'd be safe even without it. But
820 		 * better safe than sorry (plus there's not only Intel).
821 		 */
822 		on_each_cpu(do_sync_core, NULL, 1);
823 	}
824 
825 	/* patch the first byte */
826 	text_poke(addr, opcode, sizeof(int3));
827 
828 	on_each_cpu(do_sync_core, NULL, 1);
829 	/*
830 	 * sync_core() implies an smp_mb() and orders this store against
831 	 * the writing of the new instruction.
832 	 */
833 	bp_patching_in_progress = false;
834 
835 	return addr;
836 }
837 
838