1 #include <linux/module.h>
2 #include <linux/sched.h>
3 #include <linux/mutex.h>
4 #include <linux/list.h>
5 #include <linux/kprobes.h>
7 #include <linux/vmalloc.h>
8 #include <linux/memory.h>
9 #include <asm/alternative.h>
10 #include <asm/sections.h>
11 #include <asm/pgtable.h>
14 #include <asm/vsyscall.h>
15 #include <asm/cacheflush.h>
16 #include <asm/tlbflush.h>
18 #include <asm/fixmap.h>
20 #define MAX_PATCH_LEN (255-1)
22 #ifdef CONFIG_HOTPLUG_CPU
23 static int smp_alt_once;
25 static int __init bootonly(char *str)
30 __setup("smp-alt-boot", bootonly);
32 #define smp_alt_once 1
35 static int debug_alternative;
37 static int __init debug_alt(char *str)
39 debug_alternative = 1;
42 __setup("debug-alternative", debug_alt);
44 static int noreplace_smp;
46 static int __init setup_noreplace_smp(char *str)
51 __setup("noreplace-smp", setup_noreplace_smp);
53 #ifdef CONFIG_PARAVIRT
54 static int noreplace_paravirt = 0;
56 static int __init setup_noreplace_paravirt(char *str)
58 noreplace_paravirt = 1;
61 __setup("noreplace-paravirt", setup_noreplace_paravirt);
64 #define DPRINTK(fmt, args...) if (debug_alternative) \
65 printk(KERN_DEBUG fmt, args)
68 /* Use inline assembly to define this because the nops are defined
69 as inline assembly strings in the include files and we cannot
70 get them easily into strings. */
71 asm("\t.section .rodata, \"a\"\nintelnops: "
72 GENERIC_NOP1 GENERIC_NOP2 GENERIC_NOP3 GENERIC_NOP4 GENERIC_NOP5 GENERIC_NOP6
73 GENERIC_NOP7 GENERIC_NOP8
75 extern const unsigned char intelnops[];
76 static const unsigned char *const intel_nops[ASM_NOP_MAX+1] = {
81 intelnops + 1 + 2 + 3,
82 intelnops + 1 + 2 + 3 + 4,
83 intelnops + 1 + 2 + 3 + 4 + 5,
84 intelnops + 1 + 2 + 3 + 4 + 5 + 6,
85 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
90 asm("\t.section .rodata, \"a\"\nk8nops: "
91 K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
94 extern const unsigned char k8nops[];
95 static const unsigned char *const k8_nops[ASM_NOP_MAX+1] = {
101 k8nops + 1 + 2 + 3 + 4,
102 k8nops + 1 + 2 + 3 + 4 + 5,
103 k8nops + 1 + 2 + 3 + 4 + 5 + 6,
104 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
109 asm("\t.section .rodata, \"a\"\nk7nops: "
110 K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
113 extern const unsigned char k7nops[];
114 static const unsigned char *const k7_nops[ASM_NOP_MAX+1] = {
120 k7nops + 1 + 2 + 3 + 4,
121 k7nops + 1 + 2 + 3 + 4 + 5,
122 k7nops + 1 + 2 + 3 + 4 + 5 + 6,
123 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
128 asm("\t.section .rodata, \"a\"\np6nops: "
129 P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
132 extern const unsigned char p6nops[];
133 static const unsigned char *const p6_nops[ASM_NOP_MAX+1] = {
139 p6nops + 1 + 2 + 3 + 4,
140 p6nops + 1 + 2 + 3 + 4 + 5,
141 p6nops + 1 + 2 + 3 + 4 + 5 + 6,
142 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
148 extern char __vsyscall_0;
149 const unsigned char *const *find_nop_table(void)
151 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
152 boot_cpu_has(X86_FEATURE_NOPL))
158 #else /* CONFIG_X86_64 */
160 const unsigned char *const *find_nop_table(void)
162 if (boot_cpu_has(X86_FEATURE_K8))
164 else if (boot_cpu_has(X86_FEATURE_K7))
166 else if (boot_cpu_has(X86_FEATURE_NOPL))
172 #endif /* CONFIG_X86_64 */
174 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
175 void add_nops(void *insns, unsigned int len)
177 const unsigned char *const *noptable = find_nop_table();
180 unsigned int noplen = len;
181 if (noplen > ASM_NOP_MAX)
182 noplen = ASM_NOP_MAX;
183 memcpy(insns, noptable[noplen], noplen);
188 EXPORT_SYMBOL_GPL(add_nops);
190 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
191 extern u8 *__smp_locks[], *__smp_locks_end[];
193 /* Replace instructions with better alternatives for this CPU type.
194 This runs before SMP is initialized to avoid SMP problems with
195 self modifying code. This implies that assymetric systems where
196 APs have less capabilities than the boot processor are not handled.
197 Tough. Make sure you disable such features by hand. */
199 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
202 char insnbuf[MAX_PATCH_LEN];
204 DPRINTK("%s: alt table %p -> %p\n", __func__, start, end);
205 for (a = start; a < end; a++) {
206 u8 *instr = a->instr;
207 BUG_ON(a->replacementlen > a->instrlen);
208 BUG_ON(a->instrlen > sizeof(insnbuf));
209 if (!boot_cpu_has(a->cpuid))
212 /* vsyscall code is not mapped yet. resolve it manually. */
213 if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
214 instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
215 DPRINTK("%s: vsyscall fixup: %p => %p\n",
216 __func__, a->instr, instr);
219 memcpy(insnbuf, a->replacement, a->replacementlen);
220 add_nops(insnbuf + a->replacementlen,
221 a->instrlen - a->replacementlen);
222 text_poke_early(instr, insnbuf, a->instrlen);
228 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
232 mutex_lock(&text_mutex);
233 for (ptr = start; ptr < end; ptr++) {
238 /* turn DS segment override prefix into lock prefix */
239 text_poke(*ptr, ((unsigned char []){0xf0}), 1);
241 mutex_unlock(&text_mutex);
244 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
251 mutex_lock(&text_mutex);
252 for (ptr = start; ptr < end; ptr++) {
257 /* turn lock prefix into DS segment override prefix */
258 text_poke(*ptr, ((unsigned char []){0x3E}), 1);
260 mutex_unlock(&text_mutex);
263 struct smp_alt_module {
264 /* what is this ??? */
268 /* ptrs to lock prefixes */
272 /* .text segment, needed to avoid patching init code ;) */
276 struct list_head next;
278 static LIST_HEAD(smp_alt_modules);
279 static DEFINE_MUTEX(smp_alt);
280 static int smp_mode = 1; /* protected by smp_alt */
282 void alternatives_smp_module_add(struct module *mod, char *name,
283 void *locks, void *locks_end,
284 void *text, void *text_end)
286 struct smp_alt_module *smp;
292 if (boot_cpu_has(X86_FEATURE_UP))
293 alternatives_smp_unlock(locks, locks_end,
298 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
300 return; /* we'll run the (safe but slow) SMP code then ... */
305 smp->locks_end = locks_end;
307 smp->text_end = text_end;
308 DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
309 __func__, smp->locks, smp->locks_end,
310 smp->text, smp->text_end, smp->name);
312 mutex_lock(&smp_alt);
313 list_add_tail(&smp->next, &smp_alt_modules);
314 if (boot_cpu_has(X86_FEATURE_UP))
315 alternatives_smp_unlock(smp->locks, smp->locks_end,
316 smp->text, smp->text_end);
317 mutex_unlock(&smp_alt);
320 void alternatives_smp_module_del(struct module *mod)
322 struct smp_alt_module *item;
324 if (smp_alt_once || noreplace_smp)
327 mutex_lock(&smp_alt);
328 list_for_each_entry(item, &smp_alt_modules, next) {
329 if (mod != item->mod)
331 list_del(&item->next);
332 mutex_unlock(&smp_alt);
333 DPRINTK("%s: %s\n", __func__, item->name);
337 mutex_unlock(&smp_alt);
340 void alternatives_smp_switch(int smp)
342 struct smp_alt_module *mod;
344 #ifdef CONFIG_LOCKDEP
346 * Older binutils section handling bug prevented
347 * alternatives-replacement from working reliably.
349 * If this still occurs then you should see a hang
350 * or crash shortly after this line:
352 printk("lockdep: fixing up alternatives.\n");
355 if (noreplace_smp || smp_alt_once)
357 BUG_ON(!smp && (num_online_cpus() > 1));
359 mutex_lock(&smp_alt);
362 * Avoid unnecessary switches because it forces JIT based VMs to
363 * throw away all cached translations, which can be quite costly.
365 if (smp == smp_mode) {
368 printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
369 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
370 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
371 list_for_each_entry(mod, &smp_alt_modules, next)
372 alternatives_smp_lock(mod->locks, mod->locks_end,
373 mod->text, mod->text_end);
375 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
376 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
377 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
378 list_for_each_entry(mod, &smp_alt_modules, next)
379 alternatives_smp_unlock(mod->locks, mod->locks_end,
380 mod->text, mod->text_end);
383 mutex_unlock(&smp_alt);
388 #ifdef CONFIG_PARAVIRT
389 void apply_paravirt(struct paravirt_patch_site *start,
390 struct paravirt_patch_site *end)
392 struct paravirt_patch_site *p;
393 char insnbuf[MAX_PATCH_LEN];
395 if (noreplace_paravirt)
398 for (p = start; p < end; p++) {
401 BUG_ON(p->len > MAX_PATCH_LEN);
402 /* prep the buffer with the original instructions */
403 memcpy(insnbuf, p->instr, p->len);
404 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
405 (unsigned long)p->instr, p->len);
407 BUG_ON(used > p->len);
409 /* Pad the rest with nops */
410 add_nops(insnbuf + used, p->len - used);
411 text_poke_early(p->instr, insnbuf, p->len);
414 extern struct paravirt_patch_site __start_parainstructions[],
415 __stop_parainstructions[];
416 #endif /* CONFIG_PARAVIRT */
418 void __init alternative_instructions(void)
420 /* The patching is not fully atomic, so try to avoid local interruptions
421 that might execute the to be patched code.
422 Other CPUs are not running. */
426 * Don't stop machine check exceptions while patching.
427 * MCEs only happen when something got corrupted and in this
428 * case we must do something about the corruption.
429 * Ignoring it is worse than a unlikely patching race.
430 * Also machine checks tend to be broadcast and if one CPU
431 * goes into machine check the others follow quickly, so we don't
432 * expect a machine check to cause undue problems during to code
436 apply_alternatives(__alt_instructions, __alt_instructions_end);
438 /* switch to patch-once-at-boottime-only mode and free the
439 * tables in case we know the number of CPUs will never ever
441 #ifdef CONFIG_HOTPLUG_CPU
442 if (num_possible_cpus() < 2)
448 if (1 == num_possible_cpus()) {
449 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
450 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
451 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
453 alternatives_smp_unlock(__smp_locks, __smp_locks_end,
457 alternatives_smp_module_add(NULL, "core kernel",
458 __smp_locks, __smp_locks_end,
461 /* Only switch to UP mode if we don't immediately boot others */
462 if (num_present_cpus() == 1 || setup_max_cpus <= 1)
463 alternatives_smp_switch(0);
466 apply_paravirt(__parainstructions, __parainstructions_end);
469 free_init_pages("SMP alternatives",
470 (unsigned long)__smp_locks,
471 (unsigned long)__smp_locks_end);
477 * text_poke_early - Update instructions on a live kernel at boot time
478 * @addr: address to modify
479 * @opcode: source of the copy
480 * @len: length to copy
482 * When you use this code to patch more than one byte of an instruction
483 * you need to make sure that other CPUs cannot execute this code in parallel.
484 * Also no thread must be currently preempted in the middle of these
485 * instructions. And on the local CPU you need to be protected again NMI or MCE
486 * handlers seeing an inconsistent instruction while you patch.
488 void *text_poke_early(void *addr, const void *opcode, size_t len)
491 local_irq_save(flags);
492 memcpy(addr, opcode, len);
493 local_irq_restore(flags);
495 /* Could also do a CLFLUSH here to speed up CPU recovery; but
496 that causes hangs on some VIA CPUs. */
501 * text_poke - Update instructions on a live kernel
502 * @addr: address to modify
503 * @opcode: source of the copy
504 * @len: length to copy
506 * Only atomic text poke/set should be allowed when not doing early patching.
507 * It means the size must be writable atomically and the address must be aligned
508 * in a way that permits an atomic write. It also makes sure we fit on a single
511 * Note: Must be called under text_mutex.
513 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
517 struct page *pages[2];
520 if (!core_kernel_text((unsigned long)addr)) {
521 pages[0] = vmalloc_to_page(addr);
522 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
524 pages[0] = virt_to_page(addr);
525 WARN_ON(!PageReserved(pages[0]));
526 pages[1] = virt_to_page(addr + PAGE_SIZE);
529 local_irq_save(flags);
530 set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
532 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
533 vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
534 memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
535 clear_fixmap(FIX_TEXT_POKE0);
537 clear_fixmap(FIX_TEXT_POKE1);
540 /* Could also do a CLFLUSH here to speed up CPU recovery; but
541 that causes hangs on some VIA CPUs. */
542 for (i = 0; i < len; i++)
543 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
544 local_irq_restore(flags);