xref: /linux/arch/x86/kernel/alternative.c (revision a115bc070b1fc57ab23f3972401425927b5b465c)
1 #include <linux/module.h>
2 #include <linux/sched.h>
3 #include <linux/mutex.h>
4 #include <linux/list.h>
5 #include <linux/stringify.h>
6 #include <linux/kprobes.h>
7 #include <linux/mm.h>
8 #include <linux/vmalloc.h>
9 #include <linux/memory.h>
10 #include <linux/stop_machine.h>
11 #include <asm/alternative.h>
12 #include <asm/sections.h>
13 #include <asm/pgtable.h>
14 #include <asm/mce.h>
15 #include <asm/nmi.h>
16 #include <asm/vsyscall.h>
17 #include <asm/cacheflush.h>
18 #include <asm/tlbflush.h>
19 #include <asm/io.h>
20 #include <asm/fixmap.h>
21 
22 #define MAX_PATCH_LEN (255-1)
23 
24 #ifdef CONFIG_HOTPLUG_CPU
25 static int smp_alt_once;
26 
27 static int __init bootonly(char *str)
28 {
29 	smp_alt_once = 1;
30 	return 1;
31 }
32 __setup("smp-alt-boot", bootonly);
33 #else
34 #define smp_alt_once 1
35 #endif
36 
37 static int __initdata_or_module debug_alternative;
38 
39 static int __init debug_alt(char *str)
40 {
41 	debug_alternative = 1;
42 	return 1;
43 }
44 __setup("debug-alternative", debug_alt);
45 
46 static int noreplace_smp;
47 
48 static int __init setup_noreplace_smp(char *str)
49 {
50 	noreplace_smp = 1;
51 	return 1;
52 }
53 __setup("noreplace-smp", setup_noreplace_smp);
54 
55 #ifdef CONFIG_PARAVIRT
56 static int __initdata_or_module noreplace_paravirt = 0;
57 
58 static int __init setup_noreplace_paravirt(char *str)
59 {
60 	noreplace_paravirt = 1;
61 	return 1;
62 }
63 __setup("noreplace-paravirt", setup_noreplace_paravirt);
64 #endif
65 
66 #define DPRINTK(fmt, args...) if (debug_alternative) \
67 	printk(KERN_DEBUG fmt, args)
68 
69 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
70 /* Use inline assembly to define this because the nops are defined
71    as inline assembly strings in the include files and we cannot
72    get them easily into strings. */
73 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nintelnops: "
74 	GENERIC_NOP1 GENERIC_NOP2 GENERIC_NOP3 GENERIC_NOP4 GENERIC_NOP5 GENERIC_NOP6
75 	GENERIC_NOP7 GENERIC_NOP8
76     "\t.previous");
77 extern const unsigned char intelnops[];
78 static const unsigned char *const __initconst_or_module
79 intel_nops[ASM_NOP_MAX+1] = {
80 	NULL,
81 	intelnops,
82 	intelnops + 1,
83 	intelnops + 1 + 2,
84 	intelnops + 1 + 2 + 3,
85 	intelnops + 1 + 2 + 3 + 4,
86 	intelnops + 1 + 2 + 3 + 4 + 5,
87 	intelnops + 1 + 2 + 3 + 4 + 5 + 6,
88 	intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
89 };
90 #endif
91 
92 #ifdef K8_NOP1
93 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nk8nops: "
94 	K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
95 	K8_NOP7 K8_NOP8
96     "\t.previous");
97 extern const unsigned char k8nops[];
98 static const unsigned char *const __initconst_or_module
99 k8_nops[ASM_NOP_MAX+1] = {
100 	NULL,
101 	k8nops,
102 	k8nops + 1,
103 	k8nops + 1 + 2,
104 	k8nops + 1 + 2 + 3,
105 	k8nops + 1 + 2 + 3 + 4,
106 	k8nops + 1 + 2 + 3 + 4 + 5,
107 	k8nops + 1 + 2 + 3 + 4 + 5 + 6,
108 	k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
109 };
110 #endif
111 
112 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
113 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nk7nops: "
114 	K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
115 	K7_NOP7 K7_NOP8
116     "\t.previous");
117 extern const unsigned char k7nops[];
118 static const unsigned char *const __initconst_or_module
119 k7_nops[ASM_NOP_MAX+1] = {
120 	NULL,
121 	k7nops,
122 	k7nops + 1,
123 	k7nops + 1 + 2,
124 	k7nops + 1 + 2 + 3,
125 	k7nops + 1 + 2 + 3 + 4,
126 	k7nops + 1 + 2 + 3 + 4 + 5,
127 	k7nops + 1 + 2 + 3 + 4 + 5 + 6,
128 	k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
129 };
130 #endif
131 
132 #ifdef P6_NOP1
133 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\np6nops: "
134 	P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
135 	P6_NOP7 P6_NOP8
136     "\t.previous");
137 extern const unsigned char p6nops[];
138 static const unsigned char *const __initconst_or_module
139 p6_nops[ASM_NOP_MAX+1] = {
140 	NULL,
141 	p6nops,
142 	p6nops + 1,
143 	p6nops + 1 + 2,
144 	p6nops + 1 + 2 + 3,
145 	p6nops + 1 + 2 + 3 + 4,
146 	p6nops + 1 + 2 + 3 + 4 + 5,
147 	p6nops + 1 + 2 + 3 + 4 + 5 + 6,
148 	p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
149 };
150 #endif
151 
152 #ifdef CONFIG_X86_64
153 
154 extern char __vsyscall_0;
155 static const unsigned char *const *__init_or_module find_nop_table(void)
156 {
157 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
158 	    boot_cpu_has(X86_FEATURE_NOPL))
159 		return p6_nops;
160 	else
161 		return k8_nops;
162 }
163 
164 #else /* CONFIG_X86_64 */
165 
166 static const unsigned char *const *__init_or_module find_nop_table(void)
167 {
168 	if (boot_cpu_has(X86_FEATURE_K8))
169 		return k8_nops;
170 	else if (boot_cpu_has(X86_FEATURE_K7))
171 		return k7_nops;
172 	else if (boot_cpu_has(X86_FEATURE_NOPL))
173 		return p6_nops;
174 	else
175 		return intel_nops;
176 }
177 
178 #endif /* CONFIG_X86_64 */
179 
180 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
181 static void __init_or_module add_nops(void *insns, unsigned int len)
182 {
183 	const unsigned char *const *noptable = find_nop_table();
184 
185 	while (len > 0) {
186 		unsigned int noplen = len;
187 		if (noplen > ASM_NOP_MAX)
188 			noplen = ASM_NOP_MAX;
189 		memcpy(insns, noptable[noplen], noplen);
190 		insns += noplen;
191 		len -= noplen;
192 	}
193 }
194 
195 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
196 extern u8 *__smp_locks[], *__smp_locks_end[];
197 static void *text_poke_early(void *addr, const void *opcode, size_t len);
198 
199 /* Replace instructions with better alternatives for this CPU type.
200    This runs before SMP is initialized to avoid SMP problems with
201    self modifying code. This implies that assymetric systems where
202    APs have less capabilities than the boot processor are not handled.
203    Tough. Make sure you disable such features by hand. */
204 
205 void __init_or_module apply_alternatives(struct alt_instr *start,
206 					 struct alt_instr *end)
207 {
208 	struct alt_instr *a;
209 	u8 insnbuf[MAX_PATCH_LEN];
210 
211 	DPRINTK("%s: alt table %p -> %p\n", __func__, start, end);
212 	for (a = start; a < end; a++) {
213 		u8 *instr = a->instr;
214 		BUG_ON(a->replacementlen > a->instrlen);
215 		BUG_ON(a->instrlen > sizeof(insnbuf));
216 		if (!boot_cpu_has(a->cpuid))
217 			continue;
218 #ifdef CONFIG_X86_64
219 		/* vsyscall code is not mapped yet. resolve it manually. */
220 		if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
221 			instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
222 			DPRINTK("%s: vsyscall fixup: %p => %p\n",
223 				__func__, a->instr, instr);
224 		}
225 #endif
226 		memcpy(insnbuf, a->replacement, a->replacementlen);
227 		if (*insnbuf == 0xe8 && a->replacementlen == 5)
228 		    *(s32 *)(insnbuf + 1) += a->replacement - a->instr;
229 		add_nops(insnbuf + a->replacementlen,
230 			 a->instrlen - a->replacementlen);
231 		text_poke_early(instr, insnbuf, a->instrlen);
232 	}
233 }
234 
235 #ifdef CONFIG_SMP
236 
237 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
238 {
239 	u8 **ptr;
240 
241 	mutex_lock(&text_mutex);
242 	for (ptr = start; ptr < end; ptr++) {
243 		if (*ptr < text)
244 			continue;
245 		if (*ptr > text_end)
246 			continue;
247 		/* turn DS segment override prefix into lock prefix */
248 		text_poke(*ptr, ((unsigned char []){0xf0}), 1);
249 	};
250 	mutex_unlock(&text_mutex);
251 }
252 
253 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
254 {
255 	u8 **ptr;
256 
257 	if (noreplace_smp)
258 		return;
259 
260 	mutex_lock(&text_mutex);
261 	for (ptr = start; ptr < end; ptr++) {
262 		if (*ptr < text)
263 			continue;
264 		if (*ptr > text_end)
265 			continue;
266 		/* turn lock prefix into DS segment override prefix */
267 		text_poke(*ptr, ((unsigned char []){0x3E}), 1);
268 	};
269 	mutex_unlock(&text_mutex);
270 }
271 
272 struct smp_alt_module {
273 	/* what is this ??? */
274 	struct module	*mod;
275 	char		*name;
276 
277 	/* ptrs to lock prefixes */
278 	u8		**locks;
279 	u8		**locks_end;
280 
281 	/* .text segment, needed to avoid patching init code ;) */
282 	u8		*text;
283 	u8		*text_end;
284 
285 	struct list_head next;
286 };
287 static LIST_HEAD(smp_alt_modules);
288 static DEFINE_MUTEX(smp_alt);
289 static int smp_mode = 1;	/* protected by smp_alt */
290 
291 void __init_or_module alternatives_smp_module_add(struct module *mod,
292 						  char *name,
293 						  void *locks, void *locks_end,
294 						  void *text,  void *text_end)
295 {
296 	struct smp_alt_module *smp;
297 
298 	if (noreplace_smp)
299 		return;
300 
301 	if (smp_alt_once) {
302 		if (boot_cpu_has(X86_FEATURE_UP))
303 			alternatives_smp_unlock(locks, locks_end,
304 						text, text_end);
305 		return;
306 	}
307 
308 	smp = kzalloc(sizeof(*smp), GFP_KERNEL);
309 	if (NULL == smp)
310 		return; /* we'll run the (safe but slow) SMP code then ... */
311 
312 	smp->mod	= mod;
313 	smp->name	= name;
314 	smp->locks	= locks;
315 	smp->locks_end	= locks_end;
316 	smp->text	= text;
317 	smp->text_end	= text_end;
318 	DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
319 		__func__, smp->locks, smp->locks_end,
320 		smp->text, smp->text_end, smp->name);
321 
322 	mutex_lock(&smp_alt);
323 	list_add_tail(&smp->next, &smp_alt_modules);
324 	if (boot_cpu_has(X86_FEATURE_UP))
325 		alternatives_smp_unlock(smp->locks, smp->locks_end,
326 					smp->text, smp->text_end);
327 	mutex_unlock(&smp_alt);
328 }
329 
330 void __init_or_module alternatives_smp_module_del(struct module *mod)
331 {
332 	struct smp_alt_module *item;
333 
334 	if (smp_alt_once || noreplace_smp)
335 		return;
336 
337 	mutex_lock(&smp_alt);
338 	list_for_each_entry(item, &smp_alt_modules, next) {
339 		if (mod != item->mod)
340 			continue;
341 		list_del(&item->next);
342 		mutex_unlock(&smp_alt);
343 		DPRINTK("%s: %s\n", __func__, item->name);
344 		kfree(item);
345 		return;
346 	}
347 	mutex_unlock(&smp_alt);
348 }
349 
350 void alternatives_smp_switch(int smp)
351 {
352 	struct smp_alt_module *mod;
353 
354 #ifdef CONFIG_LOCKDEP
355 	/*
356 	 * Older binutils section handling bug prevented
357 	 * alternatives-replacement from working reliably.
358 	 *
359 	 * If this still occurs then you should see a hang
360 	 * or crash shortly after this line:
361 	 */
362 	printk("lockdep: fixing up alternatives.\n");
363 #endif
364 
365 	if (noreplace_smp || smp_alt_once)
366 		return;
367 	BUG_ON(!smp && (num_online_cpus() > 1));
368 
369 	mutex_lock(&smp_alt);
370 
371 	/*
372 	 * Avoid unnecessary switches because it forces JIT based VMs to
373 	 * throw away all cached translations, which can be quite costly.
374 	 */
375 	if (smp == smp_mode) {
376 		/* nothing */
377 	} else if (smp) {
378 		printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
379 		clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
380 		clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
381 		list_for_each_entry(mod, &smp_alt_modules, next)
382 			alternatives_smp_lock(mod->locks, mod->locks_end,
383 					      mod->text, mod->text_end);
384 	} else {
385 		printk(KERN_INFO "SMP alternatives: switching to UP code\n");
386 		set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
387 		set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
388 		list_for_each_entry(mod, &smp_alt_modules, next)
389 			alternatives_smp_unlock(mod->locks, mod->locks_end,
390 						mod->text, mod->text_end);
391 	}
392 	smp_mode = smp;
393 	mutex_unlock(&smp_alt);
394 }
395 
396 /* Return 1 if the address range is reserved for smp-alternatives */
397 int alternatives_text_reserved(void *start, void *end)
398 {
399 	struct smp_alt_module *mod;
400 	u8 **ptr;
401 	u8 *text_start = start;
402 	u8 *text_end = end;
403 
404 	list_for_each_entry(mod, &smp_alt_modules, next) {
405 		if (mod->text > text_end || mod->text_end < text_start)
406 			continue;
407 		for (ptr = mod->locks; ptr < mod->locks_end; ptr++)
408 			if (text_start <= *ptr && text_end >= *ptr)
409 				return 1;
410 	}
411 
412 	return 0;
413 }
414 #endif
415 
416 #ifdef CONFIG_PARAVIRT
417 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
418 				     struct paravirt_patch_site *end)
419 {
420 	struct paravirt_patch_site *p;
421 	char insnbuf[MAX_PATCH_LEN];
422 
423 	if (noreplace_paravirt)
424 		return;
425 
426 	for (p = start; p < end; p++) {
427 		unsigned int used;
428 
429 		BUG_ON(p->len > MAX_PATCH_LEN);
430 		/* prep the buffer with the original instructions */
431 		memcpy(insnbuf, p->instr, p->len);
432 		used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
433 					 (unsigned long)p->instr, p->len);
434 
435 		BUG_ON(used > p->len);
436 
437 		/* Pad the rest with nops */
438 		add_nops(insnbuf + used, p->len - used);
439 		text_poke_early(p->instr, insnbuf, p->len);
440 	}
441 }
442 extern struct paravirt_patch_site __start_parainstructions[],
443 	__stop_parainstructions[];
444 #endif	/* CONFIG_PARAVIRT */
445 
446 void __init alternative_instructions(void)
447 {
448 	/* The patching is not fully atomic, so try to avoid local interruptions
449 	   that might execute the to be patched code.
450 	   Other CPUs are not running. */
451 	stop_nmi();
452 
453 	/*
454 	 * Don't stop machine check exceptions while patching.
455 	 * MCEs only happen when something got corrupted and in this
456 	 * case we must do something about the corruption.
457 	 * Ignoring it is worse than a unlikely patching race.
458 	 * Also machine checks tend to be broadcast and if one CPU
459 	 * goes into machine check the others follow quickly, so we don't
460 	 * expect a machine check to cause undue problems during to code
461 	 * patching.
462 	 */
463 
464 	apply_alternatives(__alt_instructions, __alt_instructions_end);
465 
466 	/* switch to patch-once-at-boottime-only mode and free the
467 	 * tables in case we know the number of CPUs will never ever
468 	 * change */
469 #ifdef CONFIG_HOTPLUG_CPU
470 	if (num_possible_cpus() < 2)
471 		smp_alt_once = 1;
472 #endif
473 
474 #ifdef CONFIG_SMP
475 	if (smp_alt_once) {
476 		if (1 == num_possible_cpus()) {
477 			printk(KERN_INFO "SMP alternatives: switching to UP code\n");
478 			set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
479 			set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
480 
481 			alternatives_smp_unlock(__smp_locks, __smp_locks_end,
482 						_text, _etext);
483 		}
484 	} else {
485 		alternatives_smp_module_add(NULL, "core kernel",
486 					    __smp_locks, __smp_locks_end,
487 					    _text, _etext);
488 
489 		/* Only switch to UP mode if we don't immediately boot others */
490 		if (num_present_cpus() == 1 || setup_max_cpus <= 1)
491 			alternatives_smp_switch(0);
492 	}
493 #endif
494  	apply_paravirt(__parainstructions, __parainstructions_end);
495 
496 	if (smp_alt_once)
497 		free_init_pages("SMP alternatives",
498 				(unsigned long)__smp_locks,
499 				(unsigned long)__smp_locks_end);
500 
501 	restart_nmi();
502 }
503 
504 /**
505  * text_poke_early - Update instructions on a live kernel at boot time
506  * @addr: address to modify
507  * @opcode: source of the copy
508  * @len: length to copy
509  *
510  * When you use this code to patch more than one byte of an instruction
511  * you need to make sure that other CPUs cannot execute this code in parallel.
512  * Also no thread must be currently preempted in the middle of these
513  * instructions. And on the local CPU you need to be protected again NMI or MCE
514  * handlers seeing an inconsistent instruction while you patch.
515  */
516 static void *__init_or_module text_poke_early(void *addr, const void *opcode,
517 					      size_t len)
518 {
519 	unsigned long flags;
520 	local_irq_save(flags);
521 	memcpy(addr, opcode, len);
522 	sync_core();
523 	local_irq_restore(flags);
524 	/* Could also do a CLFLUSH here to speed up CPU recovery; but
525 	   that causes hangs on some VIA CPUs. */
526 	return addr;
527 }
528 
529 /**
530  * text_poke - Update instructions on a live kernel
531  * @addr: address to modify
532  * @opcode: source of the copy
533  * @len: length to copy
534  *
535  * Only atomic text poke/set should be allowed when not doing early patching.
536  * It means the size must be writable atomically and the address must be aligned
537  * in a way that permits an atomic write. It also makes sure we fit on a single
538  * page.
539  *
540  * Note: Must be called under text_mutex.
541  */
542 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
543 {
544 	unsigned long flags;
545 	char *vaddr;
546 	struct page *pages[2];
547 	int i;
548 
549 	if (!core_kernel_text((unsigned long)addr)) {
550 		pages[0] = vmalloc_to_page(addr);
551 		pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
552 	} else {
553 		pages[0] = virt_to_page(addr);
554 		WARN_ON(!PageReserved(pages[0]));
555 		pages[1] = virt_to_page(addr + PAGE_SIZE);
556 	}
557 	BUG_ON(!pages[0]);
558 	local_irq_save(flags);
559 	set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
560 	if (pages[1])
561 		set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
562 	vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
563 	memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
564 	clear_fixmap(FIX_TEXT_POKE0);
565 	if (pages[1])
566 		clear_fixmap(FIX_TEXT_POKE1);
567 	local_flush_tlb();
568 	sync_core();
569 	/* Could also do a CLFLUSH here to speed up CPU recovery; but
570 	   that causes hangs on some VIA CPUs. */
571 	for (i = 0; i < len; i++)
572 		BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
573 	local_irq_restore(flags);
574 	return addr;
575 }
576 
577 /*
578  * Cross-modifying kernel text with stop_machine().
579  * This code originally comes from immediate value.
580  */
581 static atomic_t stop_machine_first;
582 static int wrote_text;
583 
584 struct text_poke_params {
585 	void *addr;
586 	const void *opcode;
587 	size_t len;
588 };
589 
590 static int __kprobes stop_machine_text_poke(void *data)
591 {
592 	struct text_poke_params *tpp = data;
593 
594 	if (atomic_dec_and_test(&stop_machine_first)) {
595 		text_poke(tpp->addr, tpp->opcode, tpp->len);
596 		smp_wmb();	/* Make sure other cpus see that this has run */
597 		wrote_text = 1;
598 	} else {
599 		while (!wrote_text)
600 			cpu_relax();
601 		smp_mb();	/* Load wrote_text before following execution */
602 	}
603 
604 	flush_icache_range((unsigned long)tpp->addr,
605 			   (unsigned long)tpp->addr + tpp->len);
606 	return 0;
607 }
608 
609 /**
610  * text_poke_smp - Update instructions on a live kernel on SMP
611  * @addr: address to modify
612  * @opcode: source of the copy
613  * @len: length to copy
614  *
615  * Modify multi-byte instruction by using stop_machine() on SMP. This allows
616  * user to poke/set multi-byte text on SMP. Only non-NMI/MCE code modifying
617  * should be allowed, since stop_machine() does _not_ protect code against
618  * NMI and MCE.
619  *
620  * Note: Must be called under get_online_cpus() and text_mutex.
621  */
622 void *__kprobes text_poke_smp(void *addr, const void *opcode, size_t len)
623 {
624 	struct text_poke_params tpp;
625 
626 	tpp.addr = addr;
627 	tpp.opcode = opcode;
628 	tpp.len = len;
629 	atomic_set(&stop_machine_first, 1);
630 	wrote_text = 0;
631 	stop_machine(stop_machine_text_poke, (void *)&tpp, NULL);
632 	return addr;
633 }
634 
635