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 int __read_mostly alternatives_patched;
26 EXPORT_SYMBOL_GPL(alternatives_patched);
28 #define MAX_PATCH_LEN (255-1)
30 static int __initdata_or_module debug_alternative;
32 static int __init debug_alt(char *str)
34 debug_alternative = 1;
37 __setup("debug-alternative", debug_alt);
39 static int noreplace_smp;
41 static int __init setup_noreplace_smp(char *str)
46 __setup("noreplace-smp", setup_noreplace_smp);
48 #ifdef CONFIG_PARAVIRT
49 static int __initdata_or_module noreplace_paravirt = 0;
51 static int __init setup_noreplace_paravirt(char *str)
53 noreplace_paravirt = 1;
56 __setup("noreplace-paravirt", setup_noreplace_paravirt);
59 #define DPRINTK(fmt, args...) \
61 if (debug_alternative) \
62 printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args); \
65 #define DUMP_BYTES(buf, len, fmt, args...) \
67 if (unlikely(debug_alternative)) { \
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]); \
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.
86 * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
87 * nice symmetry of sizes of the previous nops.
89 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
90 static const unsigned char intelnops[] =
102 static const unsigned char * const intel_nops[ASM_NOP_MAX+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,
118 static const unsigned char k8nops[] =
130 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
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,
145 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
146 static const unsigned char k7nops[] =
158 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
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,
174 static const unsigned char p6nops[] =
186 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
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,
201 /* Initialize these to a safe default */
203 const unsigned char * const *ideal_nops = p6_nops;
205 const unsigned char * const *ideal_nops = intel_nops;
208 void __init arch_init_ideal_nops(void)
210 switch (boot_cpu_data.x86_vendor) {
211 case X86_VENDOR_INTEL:
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.
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;
228 ideal_nops = k8_nops;
230 ideal_nops = intel_nops;
236 if (boot_cpu_data.x86 > 0xf) {
237 ideal_nops = p6_nops;
245 ideal_nops = k8_nops;
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;
252 ideal_nops = intel_nops;
257 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
258 static void __init_or_module add_nops(void *insns, unsigned int len)
261 unsigned int noplen = len;
262 if (noplen > ASM_NOP_MAX)
263 noplen = ASM_NOP_MAX;
264 memcpy(insns, ideal_nops[noplen], noplen);
270 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
271 extern s32 __smp_locks[], __smp_locks_end[];
272 void *text_poke_early(void *addr, const void *opcode, size_t len);
275 * Are we looking at a near JMP with a 1 or 4-byte displacement.
277 static inline bool is_jmp(const u8 opcode)
279 return opcode == 0xeb || opcode == 0xe9;
282 static void __init_or_module
283 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf)
285 u8 *next_rip, *tgt_rip;
289 if (a->replacementlen != 5)
292 o_dspl = *(s32 *)(insnbuf + 1);
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;
300 DPRINTK("target RIP: %p, new_displ: 0x%x", tgt_rip, n_dspl);
302 if (tgt_rip - orig_insn >= 0) {
303 if (n_dspl - 2 <= 127)
307 /* negative offset */
309 if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
319 insnbuf[1] = (s8)n_dspl;
320 add_nops(insnbuf + 2, 3);
329 *(s32 *)&insnbuf[1] = n_dspl;
335 DPRINTK("final displ: 0x%08x, JMP 0x%lx",
336 n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
339 static void __init_or_module optimize_nops(struct alt_instr *a, u8 *instr)
341 if (instr[0] != 0x90)
344 add_nops(instr + (a->instrlen - a->padlen), a->padlen);
346 DUMP_BYTES(instr, a->instrlen, "%p: [%d:%d) optimized NOPs: ",
347 instr, a->instrlen - a->padlen, a->padlen);
351 * Replace instructions with better alternatives for this CPU type. This runs
352 * before SMP is initialized to avoid SMP problems with self modifying code.
353 * This implies that asymmetric systems where APs have less capabilities than
354 * the boot processor are not handled. Tough. Make sure you disable such
357 void __init_or_module apply_alternatives(struct alt_instr *start,
358 struct alt_instr *end)
361 u8 *instr, *replacement;
362 u8 insnbuf[MAX_PATCH_LEN];
364 DPRINTK("alt table %p -> %p", start, end);
366 * The scan order should be from start to end. A later scanned
367 * alternative code can overwrite previously scanned alternative code.
368 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
371 * So be careful if you want to change the scan order to any other
374 for (a = start; a < end; a++) {
377 instr = (u8 *)&a->instr_offset + a->instr_offset;
378 replacement = (u8 *)&a->repl_offset + a->repl_offset;
379 BUG_ON(a->instrlen > sizeof(insnbuf));
380 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
381 if (!boot_cpu_has(a->cpuid)) {
383 optimize_nops(a, instr);
388 DPRINTK("feat: %d*32+%d, old: (%p, len: %d), repl: (%p, len: %d), pad: %d",
392 replacement, a->replacementlen, a->padlen);
394 DUMP_BYTES(instr, a->instrlen, "%p: old_insn: ", instr);
395 DUMP_BYTES(replacement, a->replacementlen, "%p: rpl_insn: ", replacement);
397 memcpy(insnbuf, replacement, a->replacementlen);
398 insnbuf_sz = a->replacementlen;
400 /* 0xe8 is a relative jump; fix the offset. */
401 if (*insnbuf == 0xe8 && a->replacementlen == 5) {
402 *(s32 *)(insnbuf + 1) += replacement - instr;
403 DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
404 *(s32 *)(insnbuf + 1),
405 (unsigned long)instr + *(s32 *)(insnbuf + 1) + 5);
408 if (a->replacementlen && is_jmp(replacement[0]))
409 recompute_jump(a, instr, replacement, insnbuf);
411 if (a->instrlen > a->replacementlen) {
412 add_nops(insnbuf + a->replacementlen,
413 a->instrlen - a->replacementlen);
414 insnbuf_sz += a->instrlen - a->replacementlen;
416 DUMP_BYTES(insnbuf, insnbuf_sz, "%p: final_insn: ", instr);
418 text_poke_early(instr, insnbuf, insnbuf_sz);
423 static void alternatives_smp_lock(const s32 *start, const s32 *end,
424 u8 *text, u8 *text_end)
428 mutex_lock(&text_mutex);
429 for (poff = start; poff < end; poff++) {
430 u8 *ptr = (u8 *)poff + *poff;
432 if (!*poff || ptr < text || ptr >= text_end)
434 /* turn DS segment override prefix into lock prefix */
436 text_poke(ptr, ((unsigned char []){0xf0}), 1);
438 mutex_unlock(&text_mutex);
441 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
442 u8 *text, u8 *text_end)
446 mutex_lock(&text_mutex);
447 for (poff = start; poff < end; poff++) {
448 u8 *ptr = (u8 *)poff + *poff;
450 if (!*poff || ptr < text || ptr >= text_end)
452 /* turn lock prefix into DS segment override prefix */
454 text_poke(ptr, ((unsigned char []){0x3E}), 1);
456 mutex_unlock(&text_mutex);
459 struct smp_alt_module {
460 /* what is this ??? */
464 /* ptrs to lock prefixes */
466 const s32 *locks_end;
468 /* .text segment, needed to avoid patching init code ;) */
472 struct list_head next;
474 static LIST_HEAD(smp_alt_modules);
475 static DEFINE_MUTEX(smp_alt);
476 static bool uniproc_patched = false; /* protected by smp_alt */
478 void __init_or_module alternatives_smp_module_add(struct module *mod,
480 void *locks, void *locks_end,
481 void *text, void *text_end)
483 struct smp_alt_module *smp;
485 mutex_lock(&smp_alt);
486 if (!uniproc_patched)
489 if (num_possible_cpus() == 1)
490 /* Don't bother remembering, we'll never have to undo it. */
493 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
495 /* we'll run the (safe but slow) SMP code then ... */
501 smp->locks_end = locks_end;
503 smp->text_end = text_end;
504 DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
505 smp->locks, smp->locks_end,
506 smp->text, smp->text_end, smp->name);
508 list_add_tail(&smp->next, &smp_alt_modules);
510 alternatives_smp_unlock(locks, locks_end, text, text_end);
512 mutex_unlock(&smp_alt);
515 void __init_or_module alternatives_smp_module_del(struct module *mod)
517 struct smp_alt_module *item;
519 mutex_lock(&smp_alt);
520 list_for_each_entry(item, &smp_alt_modules, next) {
521 if (mod != item->mod)
523 list_del(&item->next);
527 mutex_unlock(&smp_alt);
530 void alternatives_enable_smp(void)
532 struct smp_alt_module *mod;
534 /* Why bother if there are no other CPUs? */
535 BUG_ON(num_possible_cpus() == 1);
537 mutex_lock(&smp_alt);
539 if (uniproc_patched) {
540 pr_info("switching to SMP code\n");
541 BUG_ON(num_online_cpus() != 1);
542 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
543 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
544 list_for_each_entry(mod, &smp_alt_modules, next)
545 alternatives_smp_lock(mod->locks, mod->locks_end,
546 mod->text, mod->text_end);
547 uniproc_patched = false;
549 mutex_unlock(&smp_alt);
552 /* Return 1 if the address range is reserved for smp-alternatives */
553 int alternatives_text_reserved(void *start, void *end)
555 struct smp_alt_module *mod;
557 u8 *text_start = start;
560 list_for_each_entry(mod, &smp_alt_modules, next) {
561 if (mod->text > text_end || mod->text_end < text_start)
563 for (poff = mod->locks; poff < mod->locks_end; poff++) {
564 const u8 *ptr = (const u8 *)poff + *poff;
566 if (text_start <= ptr && text_end > ptr)
573 #endif /* CONFIG_SMP */
575 #ifdef CONFIG_PARAVIRT
576 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
577 struct paravirt_patch_site *end)
579 struct paravirt_patch_site *p;
580 char insnbuf[MAX_PATCH_LEN];
582 if (noreplace_paravirt)
585 for (p = start; p < end; p++) {
588 BUG_ON(p->len > MAX_PATCH_LEN);
589 /* prep the buffer with the original instructions */
590 memcpy(insnbuf, p->instr, p->len);
591 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
592 (unsigned long)p->instr, p->len);
594 BUG_ON(used > p->len);
596 /* Pad the rest with nops */
597 add_nops(insnbuf + used, p->len - used);
598 text_poke_early(p->instr, insnbuf, p->len);
601 extern struct paravirt_patch_site __start_parainstructions[],
602 __stop_parainstructions[];
603 #endif /* CONFIG_PARAVIRT */
605 void __init alternative_instructions(void)
607 /* The patching is not fully atomic, so try to avoid local interruptions
608 that might execute the to be patched code.
609 Other CPUs are not running. */
613 * Don't stop machine check exceptions while patching.
614 * MCEs only happen when something got corrupted and in this
615 * case we must do something about the corruption.
616 * Ignoring it is worse than a unlikely patching race.
617 * Also machine checks tend to be broadcast and if one CPU
618 * goes into machine check the others follow quickly, so we don't
619 * expect a machine check to cause undue problems during to code
623 apply_alternatives(__alt_instructions, __alt_instructions_end);
626 /* Patch to UP if other cpus not imminent. */
627 if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
628 uniproc_patched = true;
629 alternatives_smp_module_add(NULL, "core kernel",
630 __smp_locks, __smp_locks_end,
634 if (!uniproc_patched || num_possible_cpus() == 1)
635 free_init_pages("SMP alternatives",
636 (unsigned long)__smp_locks,
637 (unsigned long)__smp_locks_end);
640 apply_paravirt(__parainstructions, __parainstructions_end);
643 alternatives_patched = 1;
647 * text_poke_early - Update instructions on a live kernel at boot time
648 * @addr: address to modify
649 * @opcode: source of the copy
650 * @len: length to copy
652 * When you use this code to patch more than one byte of an instruction
653 * you need to make sure that other CPUs cannot execute this code in parallel.
654 * Also no thread must be currently preempted in the middle of these
655 * instructions. And on the local CPU you need to be protected again NMI or MCE
656 * handlers seeing an inconsistent instruction while you patch.
658 void *__init_or_module text_poke_early(void *addr, const void *opcode,
662 local_irq_save(flags);
663 memcpy(addr, opcode, len);
665 local_irq_restore(flags);
666 /* Could also do a CLFLUSH here to speed up CPU recovery; but
667 that causes hangs on some VIA CPUs. */
672 * text_poke - Update instructions on a live kernel
673 * @addr: address to modify
674 * @opcode: source of the copy
675 * @len: length to copy
677 * Only atomic text poke/set should be allowed when not doing early patching.
678 * It means the size must be writable atomically and the address must be aligned
679 * in a way that permits an atomic write. It also makes sure we fit on a single
682 * Note: Must be called under text_mutex.
684 void *text_poke(void *addr, const void *opcode, size_t len)
688 struct page *pages[2];
691 if (!core_kernel_text((unsigned long)addr)) {
692 pages[0] = vmalloc_to_page(addr);
693 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
695 pages[0] = virt_to_page(addr);
696 WARN_ON(!PageReserved(pages[0]));
697 pages[1] = virt_to_page(addr + PAGE_SIZE);
700 local_irq_save(flags);
701 set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
703 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
704 vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
705 memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
706 clear_fixmap(FIX_TEXT_POKE0);
708 clear_fixmap(FIX_TEXT_POKE1);
711 /* Could also do a CLFLUSH here to speed up CPU recovery; but
712 that causes hangs on some VIA CPUs. */
713 for (i = 0; i < len; i++)
714 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
715 local_irq_restore(flags);
719 static void do_sync_core(void *info)
724 static bool bp_patching_in_progress;
725 static void *bp_int3_handler, *bp_int3_addr;
727 int poke_int3_handler(struct pt_regs *regs)
729 /* bp_patching_in_progress */
732 if (likely(!bp_patching_in_progress))
735 if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr)
738 /* set up the specified breakpoint handler */
739 regs->ip = (unsigned long) bp_int3_handler;
746 * text_poke_bp() -- update instructions on live kernel on SMP
747 * @addr: address to patch
748 * @opcode: opcode of new instruction
749 * @len: length to copy
750 * @handler: address to jump to when the temporary breakpoint is hit
752 * Modify multi-byte instruction by using int3 breakpoint on SMP.
753 * We completely avoid stop_machine() here, and achieve the
754 * synchronization using int3 breakpoint.
756 * The way it is done:
757 * - add a int3 trap to the address that will be patched
759 * - update all but the first byte of the patched range
761 * - replace the first byte (int3) by the first byte of
765 * Note: must be called under text_mutex.
767 void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler)
769 unsigned char int3 = 0xcc;
771 bp_int3_handler = handler;
772 bp_int3_addr = (u8 *)addr + sizeof(int3);
773 bp_patching_in_progress = true;
775 * Corresponding read barrier in int3 notifier for
776 * making sure the in_progress flags is correctly ordered wrt.
781 text_poke(addr, &int3, sizeof(int3));
783 on_each_cpu(do_sync_core, NULL, 1);
785 if (len - sizeof(int3) > 0) {
786 /* patch all but the first byte */
787 text_poke((char *)addr + sizeof(int3),
788 (const char *) opcode + sizeof(int3),
791 * According to Intel, this core syncing is very likely
792 * not necessary and we'd be safe even without it. But
793 * better safe than sorry (plus there's not only Intel).
795 on_each_cpu(do_sync_core, NULL, 1);
798 /* patch the first byte */
799 text_poke(addr, opcode, sizeof(int3));
801 on_each_cpu(do_sync_core, NULL, 1);
803 bp_patching_in_progress = false;