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