1 #define pr_fmt(fmt) "SMP alternatives: " fmt
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>
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>
19 #include <asm/cacheflush.h>
20 #include <asm/tlbflush.h>
22 #include <asm/fixmap.h>
24 #define MAX_PATCH_LEN (255-1)
26 static int __initdata_or_module debug_alternative;
28 static int __init debug_alt(char *str)
30 debug_alternative = 1;
33 __setup("debug-alternative", debug_alt);
35 static int noreplace_smp;
37 static int __init setup_noreplace_smp(char *str)
42 __setup("noreplace-smp", setup_noreplace_smp);
44 #ifdef CONFIG_PARAVIRT
45 static int __initdata_or_module noreplace_paravirt = 0;
47 static int __init setup_noreplace_paravirt(char *str)
49 noreplace_paravirt = 1;
52 __setup("noreplace-paravirt", setup_noreplace_paravirt);
55 #define DPRINTK(fmt, args...) \
57 if (debug_alternative) \
58 printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args); \
61 #define DUMP_BYTES(buf, len, fmt, args...) \
63 if (unlikely(debug_alternative)) { \
69 printk(KERN_DEBUG fmt, ##args); \
70 for (j = 0; j < (len) - 1; j++) \
71 printk(KERN_CONT "%02hhx ", buf[j]); \
72 printk(KERN_CONT "%02hhx\n", buf[j]); \
77 * Each GENERIC_NOPX is of X bytes, and defined as an array of bytes
78 * that correspond to that nop. Getting from one nop to the next, we
79 * add to the array the offset that is equal to the sum of all sizes of
80 * nops preceding the one we are after.
82 * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
83 * nice symmetry of sizes of the previous nops.
85 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
86 static const unsigned char intelnops[] =
98 static const unsigned char * const intel_nops[ASM_NOP_MAX+2] =
104 intelnops + 1 + 2 + 3,
105 intelnops + 1 + 2 + 3 + 4,
106 intelnops + 1 + 2 + 3 + 4 + 5,
107 intelnops + 1 + 2 + 3 + 4 + 5 + 6,
108 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
109 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
114 static const unsigned char k8nops[] =
126 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
133 k8nops + 1 + 2 + 3 + 4,
134 k8nops + 1 + 2 + 3 + 4 + 5,
135 k8nops + 1 + 2 + 3 + 4 + 5 + 6,
136 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
137 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
141 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
142 static const unsigned char k7nops[] =
154 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
161 k7nops + 1 + 2 + 3 + 4,
162 k7nops + 1 + 2 + 3 + 4 + 5,
163 k7nops + 1 + 2 + 3 + 4 + 5 + 6,
164 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
165 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
170 static const unsigned char p6nops[] =
182 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
189 p6nops + 1 + 2 + 3 + 4,
190 p6nops + 1 + 2 + 3 + 4 + 5,
191 p6nops + 1 + 2 + 3 + 4 + 5 + 6,
192 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
193 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
197 /* Initialize these to a safe default */
199 const unsigned char * const *ideal_nops = p6_nops;
201 const unsigned char * const *ideal_nops = intel_nops;
204 void __init arch_init_ideal_nops(void)
206 switch (boot_cpu_data.x86_vendor) {
207 case X86_VENDOR_INTEL:
209 * Due to a decoder implementation quirk, some
210 * specific Intel CPUs actually perform better with
211 * the "k8_nops" than with the SDM-recommended NOPs.
213 if (boot_cpu_data.x86 == 6 &&
214 boot_cpu_data.x86_model >= 0x0f &&
215 boot_cpu_data.x86_model != 0x1c &&
216 boot_cpu_data.x86_model != 0x26 &&
217 boot_cpu_data.x86_model != 0x27 &&
218 boot_cpu_data.x86_model < 0x30) {
219 ideal_nops = k8_nops;
220 } else if (boot_cpu_has(X86_FEATURE_NOPL)) {
221 ideal_nops = p6_nops;
224 ideal_nops = k8_nops;
226 ideal_nops = intel_nops;
232 ideal_nops = k8_nops;
234 if (boot_cpu_has(X86_FEATURE_K8))
235 ideal_nops = k8_nops;
236 else if (boot_cpu_has(X86_FEATURE_K7))
237 ideal_nops = k7_nops;
239 ideal_nops = intel_nops;
244 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
245 static void __init_or_module add_nops(void *insns, unsigned int len)
248 unsigned int noplen = len;
249 if (noplen > ASM_NOP_MAX)
250 noplen = ASM_NOP_MAX;
251 memcpy(insns, ideal_nops[noplen], noplen);
257 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
258 extern s32 __smp_locks[], __smp_locks_end[];
259 void *text_poke_early(void *addr, const void *opcode, size_t len);
262 * Are we looking at a near JMP with a 1 or 4-byte displacement.
264 static inline bool is_jmp(const u8 opcode)
266 return opcode == 0xeb || opcode == 0xe9;
269 static void __init_or_module
270 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf)
272 u8 *next_rip, *tgt_rip;
276 if (a->replacementlen != 5)
279 o_dspl = *(s32 *)(insnbuf + 1);
281 /* next_rip of the replacement JMP */
282 next_rip = repl_insn + a->replacementlen;
283 /* target rip of the replacement JMP */
284 tgt_rip = next_rip + o_dspl;
285 n_dspl = tgt_rip - orig_insn;
287 DPRINTK("target RIP: %p, new_displ: 0x%x", tgt_rip, n_dspl);
289 if (tgt_rip - orig_insn >= 0) {
290 if (n_dspl - 2 <= 127)
294 /* negative offset */
296 if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
306 insnbuf[1] = (s8)n_dspl;
307 add_nops(insnbuf + 2, 3);
316 *(s32 *)&insnbuf[1] = n_dspl;
322 DPRINTK("final displ: 0x%08x, JMP 0x%lx",
323 n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
326 static void __init_or_module optimize_nops(struct alt_instr *a, u8 *instr)
328 if (instr[0] != 0x90)
331 add_nops(instr + (a->instrlen - a->padlen), a->padlen);
333 DUMP_BYTES(instr, a->instrlen, "%p: [%d:%d) optimized NOPs: ",
334 instr, a->instrlen - a->padlen, a->padlen);
338 * Replace instructions with better alternatives for this CPU type. This runs
339 * before SMP is initialized to avoid SMP problems with self modifying code.
340 * This implies that asymmetric systems where APs have less capabilities than
341 * the boot processor are not handled. Tough. Make sure you disable such
344 void __init_or_module apply_alternatives(struct alt_instr *start,
345 struct alt_instr *end)
348 u8 *instr, *replacement;
349 u8 insnbuf[MAX_PATCH_LEN];
351 DPRINTK("alt table %p -> %p", start, end);
353 * The scan order should be from start to end. A later scanned
354 * alternative code can overwrite previously scanned alternative code.
355 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
358 * So be careful if you want to change the scan order to any other
361 for (a = start; a < end; a++) {
364 instr = (u8 *)&a->instr_offset + a->instr_offset;
365 replacement = (u8 *)&a->repl_offset + a->repl_offset;
366 BUG_ON(a->instrlen > sizeof(insnbuf));
367 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
368 if (!boot_cpu_has(a->cpuid)) {
370 optimize_nops(a, instr);
375 DPRINTK("feat: %d*32+%d, old: (%p, len: %d), repl: (%p, len: %d), pad: %d",
379 replacement, a->replacementlen, a->padlen);
381 DUMP_BYTES(instr, a->instrlen, "%p: old_insn: ", instr);
382 DUMP_BYTES(replacement, a->replacementlen, "%p: rpl_insn: ", replacement);
384 memcpy(insnbuf, replacement, a->replacementlen);
385 insnbuf_sz = a->replacementlen;
387 /* 0xe8 is a relative jump; fix the offset. */
388 if (*insnbuf == 0xe8 && a->replacementlen == 5) {
389 *(s32 *)(insnbuf + 1) += replacement - instr;
390 DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
391 *(s32 *)(insnbuf + 1),
392 (unsigned long)instr + *(s32 *)(insnbuf + 1) + 5);
395 if (a->replacementlen && is_jmp(replacement[0]))
396 recompute_jump(a, instr, replacement, insnbuf);
398 if (a->instrlen > a->replacementlen) {
399 add_nops(insnbuf + a->replacementlen,
400 a->instrlen - a->replacementlen);
401 insnbuf_sz += a->instrlen - a->replacementlen;
403 DUMP_BYTES(insnbuf, insnbuf_sz, "%p: final_insn: ", instr);
405 text_poke_early(instr, insnbuf, insnbuf_sz);
410 static void alternatives_smp_lock(const s32 *start, const s32 *end,
411 u8 *text, u8 *text_end)
415 mutex_lock(&text_mutex);
416 for (poff = start; poff < end; poff++) {
417 u8 *ptr = (u8 *)poff + *poff;
419 if (!*poff || ptr < text || ptr >= text_end)
421 /* turn DS segment override prefix into lock prefix */
423 text_poke(ptr, ((unsigned char []){0xf0}), 1);
425 mutex_unlock(&text_mutex);
428 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
429 u8 *text, u8 *text_end)
433 mutex_lock(&text_mutex);
434 for (poff = start; poff < end; poff++) {
435 u8 *ptr = (u8 *)poff + *poff;
437 if (!*poff || ptr < text || ptr >= text_end)
439 /* turn lock prefix into DS segment override prefix */
441 text_poke(ptr, ((unsigned char []){0x3E}), 1);
443 mutex_unlock(&text_mutex);
446 struct smp_alt_module {
447 /* what is this ??? */
451 /* ptrs to lock prefixes */
453 const s32 *locks_end;
455 /* .text segment, needed to avoid patching init code ;) */
459 struct list_head next;
461 static LIST_HEAD(smp_alt_modules);
462 static DEFINE_MUTEX(smp_alt);
463 static bool uniproc_patched = false; /* protected by smp_alt */
465 void __init_or_module alternatives_smp_module_add(struct module *mod,
467 void *locks, void *locks_end,
468 void *text, void *text_end)
470 struct smp_alt_module *smp;
472 mutex_lock(&smp_alt);
473 if (!uniproc_patched)
476 if (num_possible_cpus() == 1)
477 /* Don't bother remembering, we'll never have to undo it. */
480 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
482 /* we'll run the (safe but slow) SMP code then ... */
488 smp->locks_end = locks_end;
490 smp->text_end = text_end;
491 DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
492 smp->locks, smp->locks_end,
493 smp->text, smp->text_end, smp->name);
495 list_add_tail(&smp->next, &smp_alt_modules);
497 alternatives_smp_unlock(locks, locks_end, text, text_end);
499 mutex_unlock(&smp_alt);
502 void __init_or_module alternatives_smp_module_del(struct module *mod)
504 struct smp_alt_module *item;
506 mutex_lock(&smp_alt);
507 list_for_each_entry(item, &smp_alt_modules, next) {
508 if (mod != item->mod)
510 list_del(&item->next);
514 mutex_unlock(&smp_alt);
517 void alternatives_enable_smp(void)
519 struct smp_alt_module *mod;
521 /* Why bother if there are no other CPUs? */
522 BUG_ON(num_possible_cpus() == 1);
524 mutex_lock(&smp_alt);
526 if (uniproc_patched) {
527 pr_info("switching to SMP code\n");
528 BUG_ON(num_online_cpus() != 1);
529 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
530 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
531 list_for_each_entry(mod, &smp_alt_modules, next)
532 alternatives_smp_lock(mod->locks, mod->locks_end,
533 mod->text, mod->text_end);
534 uniproc_patched = false;
536 mutex_unlock(&smp_alt);
539 /* Return 1 if the address range is reserved for smp-alternatives */
540 int alternatives_text_reserved(void *start, void *end)
542 struct smp_alt_module *mod;
544 u8 *text_start = start;
547 list_for_each_entry(mod, &smp_alt_modules, next) {
548 if (mod->text > text_end || mod->text_end < text_start)
550 for (poff = mod->locks; poff < mod->locks_end; poff++) {
551 const u8 *ptr = (const u8 *)poff + *poff;
553 if (text_start <= ptr && text_end > ptr)
560 #endif /* CONFIG_SMP */
562 #ifdef CONFIG_PARAVIRT
563 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
564 struct paravirt_patch_site *end)
566 struct paravirt_patch_site *p;
567 char insnbuf[MAX_PATCH_LEN];
569 if (noreplace_paravirt)
572 for (p = start; p < end; p++) {
575 BUG_ON(p->len > MAX_PATCH_LEN);
576 /* prep the buffer with the original instructions */
577 memcpy(insnbuf, p->instr, p->len);
578 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
579 (unsigned long)p->instr, p->len);
581 BUG_ON(used > p->len);
583 /* Pad the rest with nops */
584 add_nops(insnbuf + used, p->len - used);
585 text_poke_early(p->instr, insnbuf, p->len);
588 extern struct paravirt_patch_site __start_parainstructions[],
589 __stop_parainstructions[];
590 #endif /* CONFIG_PARAVIRT */
592 void __init alternative_instructions(void)
594 /* The patching is not fully atomic, so try to avoid local interruptions
595 that might execute the to be patched code.
596 Other CPUs are not running. */
600 * Don't stop machine check exceptions while patching.
601 * MCEs only happen when something got corrupted and in this
602 * case we must do something about the corruption.
603 * Ignoring it is worse than a unlikely patching race.
604 * Also machine checks tend to be broadcast and if one CPU
605 * goes into machine check the others follow quickly, so we don't
606 * expect a machine check to cause undue problems during to code
610 apply_alternatives(__alt_instructions, __alt_instructions_end);
613 /* Patch to UP if other cpus not imminent. */
614 if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
615 uniproc_patched = true;
616 alternatives_smp_module_add(NULL, "core kernel",
617 __smp_locks, __smp_locks_end,
621 if (!uniproc_patched || num_possible_cpus() == 1)
622 free_init_pages("SMP alternatives",
623 (unsigned long)__smp_locks,
624 (unsigned long)__smp_locks_end);
627 apply_paravirt(__parainstructions, __parainstructions_end);
633 * text_poke_early - Update instructions on a live kernel at boot time
634 * @addr: address to modify
635 * @opcode: source of the copy
636 * @len: length to copy
638 * When you use this code to patch more than one byte of an instruction
639 * you need to make sure that other CPUs cannot execute this code in parallel.
640 * Also no thread must be currently preempted in the middle of these
641 * instructions. And on the local CPU you need to be protected again NMI or MCE
642 * handlers seeing an inconsistent instruction while you patch.
644 void *__init_or_module text_poke_early(void *addr, const void *opcode,
648 local_irq_save(flags);
649 memcpy(addr, opcode, len);
651 local_irq_restore(flags);
652 /* Could also do a CLFLUSH here to speed up CPU recovery; but
653 that causes hangs on some VIA CPUs. */
658 * text_poke - Update instructions on a live kernel
659 * @addr: address to modify
660 * @opcode: source of the copy
661 * @len: length to copy
663 * Only atomic text poke/set should be allowed when not doing early patching.
664 * It means the size must be writable atomically and the address must be aligned
665 * in a way that permits an atomic write. It also makes sure we fit on a single
668 * Note: Must be called under text_mutex.
670 void *text_poke(void *addr, const void *opcode, size_t len)
674 struct page *pages[2];
677 if (!core_kernel_text((unsigned long)addr)) {
678 pages[0] = vmalloc_to_page(addr);
679 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
681 pages[0] = virt_to_page(addr);
682 WARN_ON(!PageReserved(pages[0]));
683 pages[1] = virt_to_page(addr + PAGE_SIZE);
686 local_irq_save(flags);
687 set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
689 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
690 vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
691 memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
692 clear_fixmap(FIX_TEXT_POKE0);
694 clear_fixmap(FIX_TEXT_POKE1);
697 /* Could also do a CLFLUSH here to speed up CPU recovery; but
698 that causes hangs on some VIA CPUs. */
699 for (i = 0; i < len; i++)
700 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
701 local_irq_restore(flags);
705 static void do_sync_core(void *info)
710 static bool bp_patching_in_progress;
711 static void *bp_int3_handler, *bp_int3_addr;
713 int poke_int3_handler(struct pt_regs *regs)
715 /* bp_patching_in_progress */
718 if (likely(!bp_patching_in_progress))
721 if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr)
724 /* set up the specified breakpoint handler */
725 regs->ip = (unsigned long) bp_int3_handler;
732 * text_poke_bp() -- update instructions on live kernel on SMP
733 * @addr: address to patch
734 * @opcode: opcode of new instruction
735 * @len: length to copy
736 * @handler: address to jump to when the temporary breakpoint is hit
738 * Modify multi-byte instruction by using int3 breakpoint on SMP.
739 * We completely avoid stop_machine() here, and achieve the
740 * synchronization using int3 breakpoint.
742 * The way it is done:
743 * - add a int3 trap to the address that will be patched
745 * - update all but the first byte of the patched range
747 * - replace the first byte (int3) by the first byte of
751 * Note: must be called under text_mutex.
753 void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler)
755 unsigned char int3 = 0xcc;
757 bp_int3_handler = handler;
758 bp_int3_addr = (u8 *)addr + sizeof(int3);
759 bp_patching_in_progress = true;
761 * Corresponding read barrier in int3 notifier for
762 * making sure the in_progress flags is correctly ordered wrt.
767 text_poke(addr, &int3, sizeof(int3));
769 on_each_cpu(do_sync_core, NULL, 1);
771 if (len - sizeof(int3) > 0) {
772 /* patch all but the first byte */
773 text_poke((char *)addr + sizeof(int3),
774 (const char *) opcode + sizeof(int3),
777 * According to Intel, this core syncing is very likely
778 * not necessary and we'd be safe even without it. But
779 * better safe than sorry (plus there's not only Intel).
781 on_each_cpu(do_sync_core, NULL, 1);
784 /* patch the first byte */
785 text_poke(addr, opcode, sizeof(int3));
787 on_each_cpu(do_sync_core, NULL, 1);
789 bp_patching_in_progress = false;