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