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/text-patching.h>
15 #include <asm/alternative.h>
16 #include <asm/sections.h>
17 #include <asm/pgtable.h>
20 #include <asm/cacheflush.h>
21 #include <asm/tlbflush.h>
23 #include <asm/fixmap.h>
25 int __read_mostly alternatives_patched;
27 EXPORT_SYMBOL_GPL(alternatives_patched);
29 #define MAX_PATCH_LEN (255-1)
31 static int __initdata_or_module debug_alternative;
33 static int __init debug_alt(char *str)
35 debug_alternative = 1;
38 __setup("debug-alternative", debug_alt);
40 static int noreplace_smp;
42 static int __init setup_noreplace_smp(char *str)
47 __setup("noreplace-smp", setup_noreplace_smp);
49 #ifdef CONFIG_PARAVIRT
50 static int __initdata_or_module noreplace_paravirt = 0;
52 static int __init setup_noreplace_paravirt(char *str)
54 noreplace_paravirt = 1;
57 __setup("noreplace-paravirt", setup_noreplace_paravirt);
60 #define DPRINTK(fmt, args...) \
62 if (debug_alternative) \
63 printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args); \
66 #define DUMP_BYTES(buf, len, fmt, args...) \
68 if (unlikely(debug_alternative)) { \
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]); \
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.
87 * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
88 * nice symmetry of sizes of the previous nops.
90 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
91 static const unsigned char intelnops[] =
103 static const unsigned char * const intel_nops[ASM_NOP_MAX+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,
119 static const unsigned char k8nops[] =
131 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
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,
146 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
147 static const unsigned char k7nops[] =
159 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
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,
175 static const unsigned char p6nops[] =
187 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
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,
202 /* Initialize these to a safe default */
204 const unsigned char * const *ideal_nops = p6_nops;
206 const unsigned char * const *ideal_nops = intel_nops;
209 void __init arch_init_ideal_nops(void)
211 switch (boot_cpu_data.x86_vendor) {
212 case X86_VENDOR_INTEL:
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.
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;
229 ideal_nops = k8_nops;
231 ideal_nops = intel_nops;
237 if (boot_cpu_data.x86 > 0xf) {
238 ideal_nops = p6_nops;
246 ideal_nops = k8_nops;
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;
253 ideal_nops = intel_nops;
258 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
259 static void __init_or_module add_nops(void *insns, unsigned int len)
262 unsigned int noplen = len;
263 if (noplen > ASM_NOP_MAX)
264 noplen = ASM_NOP_MAX;
265 memcpy(insns, ideal_nops[noplen], noplen);
271 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
272 extern s32 __smp_locks[], __smp_locks_end[];
273 void *text_poke_early(void *addr, const void *opcode, size_t len);
276 * Are we looking at a near JMP with a 1 or 4-byte displacement.
278 static inline bool is_jmp(const u8 opcode)
280 return opcode == 0xeb || opcode == 0xe9;
283 static void __init_or_module
284 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf)
286 u8 *next_rip, *tgt_rip;
290 if (a->replacementlen != 5)
293 o_dspl = *(s32 *)(insnbuf + 1);
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;
301 DPRINTK("target RIP: %p, new_displ: 0x%x", tgt_rip, n_dspl);
303 if (tgt_rip - orig_insn >= 0) {
304 if (n_dspl - 2 <= 127)
308 /* negative offset */
310 if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
320 insnbuf[1] = (s8)n_dspl;
321 add_nops(insnbuf + 2, 3);
330 *(s32 *)&insnbuf[1] = n_dspl;
336 DPRINTK("final displ: 0x%08x, JMP 0x%lx",
337 n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
340 static void __init_or_module optimize_nops(struct alt_instr *a, u8 *instr)
344 if (instr[0] != 0x90)
347 local_irq_save(flags);
348 add_nops(instr + (a->instrlen - a->padlen), a->padlen);
350 local_irq_restore(flags);
352 DUMP_BYTES(instr, a->instrlen, "%p: [%d:%d) optimized NOPs: ",
353 instr, a->instrlen - a->padlen, a->padlen);
357 * Replace instructions with better alternatives for this CPU type. This runs
358 * before SMP is initialized to avoid SMP problems with self modifying code.
359 * This implies that asymmetric systems where APs have less capabilities than
360 * the boot processor are not handled. Tough. Make sure you disable such
363 void __init_or_module apply_alternatives(struct alt_instr *start,
364 struct alt_instr *end)
367 u8 *instr, *replacement;
368 u8 insnbuf[MAX_PATCH_LEN];
370 DPRINTK("alt table %p -> %p", start, end);
372 * The scan order should be from start to end. A later scanned
373 * alternative code can overwrite previously scanned alternative code.
374 * Some kernel functions (e.g. memcpy, memset, etc) use this order to
377 * So be careful if you want to change the scan order to any other
380 for (a = start; a < end; a++) {
383 instr = (u8 *)&a->instr_offset + a->instr_offset;
384 replacement = (u8 *)&a->repl_offset + a->repl_offset;
385 BUG_ON(a->instrlen > sizeof(insnbuf));
386 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
387 if (!boot_cpu_has(a->cpuid)) {
389 optimize_nops(a, instr);
394 DPRINTK("feat: %d*32+%d, old: (%p, len: %d), repl: (%p, len: %d), pad: %d",
398 replacement, a->replacementlen, a->padlen);
400 DUMP_BYTES(instr, a->instrlen, "%p: old_insn: ", instr);
401 DUMP_BYTES(replacement, a->replacementlen, "%p: rpl_insn: ", replacement);
403 memcpy(insnbuf, replacement, a->replacementlen);
404 insnbuf_sz = a->replacementlen;
406 /* 0xe8 is a relative jump; fix the offset. */
407 if (*insnbuf == 0xe8 && a->replacementlen == 5) {
408 *(s32 *)(insnbuf + 1) += replacement - instr;
409 DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
410 *(s32 *)(insnbuf + 1),
411 (unsigned long)instr + *(s32 *)(insnbuf + 1) + 5);
414 if (a->replacementlen && is_jmp(replacement[0]))
415 recompute_jump(a, instr, replacement, insnbuf);
417 if (a->instrlen > a->replacementlen) {
418 add_nops(insnbuf + a->replacementlen,
419 a->instrlen - a->replacementlen);
420 insnbuf_sz += a->instrlen - a->replacementlen;
422 DUMP_BYTES(insnbuf, insnbuf_sz, "%p: final_insn: ", instr);
424 text_poke_early(instr, insnbuf, insnbuf_sz);
429 static void alternatives_smp_lock(const s32 *start, const s32 *end,
430 u8 *text, u8 *text_end)
434 mutex_lock(&text_mutex);
435 for (poff = start; poff < end; poff++) {
436 u8 *ptr = (u8 *)poff + *poff;
438 if (!*poff || ptr < text || ptr >= text_end)
440 /* turn DS segment override prefix into lock prefix */
442 text_poke(ptr, ((unsigned char []){0xf0}), 1);
444 mutex_unlock(&text_mutex);
447 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
448 u8 *text, u8 *text_end)
452 mutex_lock(&text_mutex);
453 for (poff = start; poff < end; poff++) {
454 u8 *ptr = (u8 *)poff + *poff;
456 if (!*poff || ptr < text || ptr >= text_end)
458 /* turn lock prefix into DS segment override prefix */
460 text_poke(ptr, ((unsigned char []){0x3E}), 1);
462 mutex_unlock(&text_mutex);
465 struct smp_alt_module {
466 /* what is this ??? */
470 /* ptrs to lock prefixes */
472 const s32 *locks_end;
474 /* .text segment, needed to avoid patching init code ;) */
478 struct list_head next;
480 static LIST_HEAD(smp_alt_modules);
481 static DEFINE_MUTEX(smp_alt);
482 static bool uniproc_patched = false; /* protected by smp_alt */
484 void __init_or_module alternatives_smp_module_add(struct module *mod,
486 void *locks, void *locks_end,
487 void *text, void *text_end)
489 struct smp_alt_module *smp;
491 mutex_lock(&smp_alt);
492 if (!uniproc_patched)
495 if (num_possible_cpus() == 1)
496 /* Don't bother remembering, we'll never have to undo it. */
499 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
501 /* we'll run the (safe but slow) SMP code then ... */
507 smp->locks_end = locks_end;
509 smp->text_end = text_end;
510 DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
511 smp->locks, smp->locks_end,
512 smp->text, smp->text_end, smp->name);
514 list_add_tail(&smp->next, &smp_alt_modules);
516 alternatives_smp_unlock(locks, locks_end, text, text_end);
518 mutex_unlock(&smp_alt);
521 void __init_or_module alternatives_smp_module_del(struct module *mod)
523 struct smp_alt_module *item;
525 mutex_lock(&smp_alt);
526 list_for_each_entry(item, &smp_alt_modules, next) {
527 if (mod != item->mod)
529 list_del(&item->next);
533 mutex_unlock(&smp_alt);
536 void alternatives_enable_smp(void)
538 struct smp_alt_module *mod;
540 /* Why bother if there are no other CPUs? */
541 BUG_ON(num_possible_cpus() == 1);
543 mutex_lock(&smp_alt);
545 if (uniproc_patched) {
546 pr_info("switching to SMP code\n");
547 BUG_ON(num_online_cpus() != 1);
548 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
549 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
550 list_for_each_entry(mod, &smp_alt_modules, next)
551 alternatives_smp_lock(mod->locks, mod->locks_end,
552 mod->text, mod->text_end);
553 uniproc_patched = false;
555 mutex_unlock(&smp_alt);
558 /* Return 1 if the address range is reserved for smp-alternatives */
559 int alternatives_text_reserved(void *start, void *end)
561 struct smp_alt_module *mod;
563 u8 *text_start = start;
566 list_for_each_entry(mod, &smp_alt_modules, next) {
567 if (mod->text > text_end || mod->text_end < text_start)
569 for (poff = mod->locks; poff < mod->locks_end; poff++) {
570 const u8 *ptr = (const u8 *)poff + *poff;
572 if (text_start <= ptr && text_end > ptr)
579 #endif /* CONFIG_SMP */
581 #ifdef CONFIG_PARAVIRT
582 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
583 struct paravirt_patch_site *end)
585 struct paravirt_patch_site *p;
586 char insnbuf[MAX_PATCH_LEN];
588 if (noreplace_paravirt)
591 for (p = start; p < end; p++) {
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);
600 BUG_ON(used > p->len);
602 /* Pad the rest with nops */
603 add_nops(insnbuf + used, p->len - used);
604 text_poke_early(p->instr, insnbuf, p->len);
607 extern struct paravirt_patch_site __start_parainstructions[],
608 __stop_parainstructions[];
609 #endif /* CONFIG_PARAVIRT */
611 void __init alternative_instructions(void)
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. */
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
629 apply_alternatives(__alt_instructions, __alt_instructions_end);
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,
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);
646 apply_paravirt(__parainstructions, __parainstructions_end);
649 alternatives_patched = 1;
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
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.
664 void *__init_or_module text_poke_early(void *addr, const void *opcode,
668 local_irq_save(flags);
669 memcpy(addr, opcode, len);
671 local_irq_restore(flags);
672 /* Could also do a CLFLUSH here to speed up CPU recovery; but
673 that causes hangs on some VIA CPUs. */
678 * text_poke - Update instructions on a live kernel
679 * @addr: address to modify
680 * @opcode: source of the copy
681 * @len: length to copy
683 * Only atomic text poke/set should be allowed when not doing early patching.
684 * It means the size must be writable atomically and the address must be aligned
685 * in a way that permits an atomic write. It also makes sure we fit on a single
688 * Note: Must be called under text_mutex.
690 void *text_poke(void *addr, const void *opcode, size_t len)
694 struct page *pages[2];
697 if (!core_kernel_text((unsigned long)addr)) {
698 pages[0] = vmalloc_to_page(addr);
699 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
701 pages[0] = virt_to_page(addr);
702 WARN_ON(!PageReserved(pages[0]));
703 pages[1] = virt_to_page(addr + PAGE_SIZE);
706 local_irq_save(flags);
707 set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
709 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
710 vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
711 memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
712 clear_fixmap(FIX_TEXT_POKE0);
714 clear_fixmap(FIX_TEXT_POKE1);
717 /* Could also do a CLFLUSH here to speed up CPU recovery; but
718 that causes hangs on some VIA CPUs. */
719 for (i = 0; i < len; i++)
720 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
721 local_irq_restore(flags);
725 static void do_sync_core(void *info)
730 static bool bp_patching_in_progress;
731 static void *bp_int3_handler, *bp_int3_addr;
733 int poke_int3_handler(struct pt_regs *regs)
735 /* bp_patching_in_progress */
738 if (likely(!bp_patching_in_progress))
741 if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr)
744 /* set up the specified breakpoint handler */
745 regs->ip = (unsigned long) bp_int3_handler;
752 * text_poke_bp() -- update instructions on live kernel on SMP
753 * @addr: address to patch
754 * @opcode: opcode of new instruction
755 * @len: length to copy
756 * @handler: address to jump to when the temporary breakpoint is hit
758 * Modify multi-byte instruction by using int3 breakpoint on SMP.
759 * We completely avoid stop_machine() here, and achieve the
760 * synchronization using int3 breakpoint.
762 * The way it is done:
763 * - add a int3 trap to the address that will be patched
765 * - update all but the first byte of the patched range
767 * - replace the first byte (int3) by the first byte of
771 * Note: must be called under text_mutex.
773 void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler)
775 unsigned char int3 = 0xcc;
777 bp_int3_handler = handler;
778 bp_int3_addr = (u8 *)addr + sizeof(int3);
779 bp_patching_in_progress = true;
781 * Corresponding read barrier in int3 notifier for
782 * making sure the in_progress flags is correctly ordered wrt.
787 text_poke(addr, &int3, sizeof(int3));
789 on_each_cpu(do_sync_core, NULL, 1);
791 if (len - sizeof(int3) > 0) {
792 /* patch all but the first byte */
793 text_poke((char *)addr + sizeof(int3),
794 (const char *) opcode + sizeof(int3),
797 * According to Intel, this core syncing is very likely
798 * not necessary and we'd be safe even without it. But
799 * better safe than sorry (plus there's not only Intel).
801 on_each_cpu(do_sync_core, NULL, 1);
804 /* patch the first byte */
805 text_poke(addr, opcode, sizeof(int3));
807 on_each_cpu(do_sync_core, NULL, 1);
809 bp_patching_in_progress = false;