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 <asm/alternative.h>
9 #include <asm/sections.h>
10 #include <asm/pgtable.h>
13 #include <asm/vsyscall.h>
14 #include <asm/cacheflush.h>
17 #define MAX_PATCH_LEN (255-1)
19 #ifdef CONFIG_HOTPLUG_CPU
20 static int smp_alt_once;
22 static int __init bootonly(char *str)
27 __setup("smp-alt-boot", bootonly);
29 #define smp_alt_once 1
32 static int debug_alternative;
34 static int __init debug_alt(char *str)
36 debug_alternative = 1;
39 __setup("debug-alternative", debug_alt);
41 static int noreplace_smp;
43 static int __init setup_noreplace_smp(char *str)
48 __setup("noreplace-smp", setup_noreplace_smp);
50 #ifdef CONFIG_PARAVIRT
51 static int noreplace_paravirt = 0;
53 static int __init setup_noreplace_paravirt(char *str)
55 noreplace_paravirt = 1;
58 __setup("noreplace-paravirt", setup_noreplace_paravirt);
61 #define DPRINTK(fmt, args...) if (debug_alternative) \
62 printk(KERN_DEBUG fmt, args)
65 /* Use inline assembly to define this because the nops are defined
66 as inline assembly strings in the include files and we cannot
67 get them easily into strings. */
68 asm("\t.section .rodata, \"a\"\nintelnops: "
69 GENERIC_NOP1 GENERIC_NOP2 GENERIC_NOP3 GENERIC_NOP4 GENERIC_NOP5 GENERIC_NOP6
70 GENERIC_NOP7 GENERIC_NOP8
72 extern const unsigned char intelnops[];
73 static const unsigned char *const intel_nops[ASM_NOP_MAX+1] = {
78 intelnops + 1 + 2 + 3,
79 intelnops + 1 + 2 + 3 + 4,
80 intelnops + 1 + 2 + 3 + 4 + 5,
81 intelnops + 1 + 2 + 3 + 4 + 5 + 6,
82 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
87 asm("\t.section .rodata, \"a\"\nk8nops: "
88 K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
91 extern const unsigned char k8nops[];
92 static const unsigned char *const k8_nops[ASM_NOP_MAX+1] = {
98 k8nops + 1 + 2 + 3 + 4,
99 k8nops + 1 + 2 + 3 + 4 + 5,
100 k8nops + 1 + 2 + 3 + 4 + 5 + 6,
101 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
106 asm("\t.section .rodata, \"a\"\nk7nops: "
107 K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
110 extern const unsigned char k7nops[];
111 static const unsigned char *const k7_nops[ASM_NOP_MAX+1] = {
117 k7nops + 1 + 2 + 3 + 4,
118 k7nops + 1 + 2 + 3 + 4 + 5,
119 k7nops + 1 + 2 + 3 + 4 + 5 + 6,
120 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
125 asm("\t.section .rodata, \"a\"\np6nops: "
126 P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
129 extern const unsigned char p6nops[];
130 static const unsigned char *const p6_nops[ASM_NOP_MAX+1] = {
136 p6nops + 1 + 2 + 3 + 4,
137 p6nops + 1 + 2 + 3 + 4 + 5,
138 p6nops + 1 + 2 + 3 + 4 + 5 + 6,
139 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
145 extern char __vsyscall_0;
146 const unsigned char *const *find_nop_table(void)
148 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
149 boot_cpu_has(X86_FEATURE_NOPL))
155 #else /* CONFIG_X86_64 */
157 const unsigned char *const *find_nop_table(void)
159 if (boot_cpu_has(X86_FEATURE_K8))
161 else if (boot_cpu_has(X86_FEATURE_K7))
163 else if (boot_cpu_has(X86_FEATURE_NOPL))
169 #endif /* CONFIG_X86_64 */
171 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
172 void add_nops(void *insns, unsigned int len)
174 const unsigned char *const *noptable = find_nop_table();
177 unsigned int noplen = len;
178 if (noplen > ASM_NOP_MAX)
179 noplen = ASM_NOP_MAX;
180 memcpy(insns, noptable[noplen], noplen);
185 EXPORT_SYMBOL_GPL(add_nops);
187 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
188 extern u8 *__smp_locks[], *__smp_locks_end[];
190 /* Replace instructions with better alternatives for this CPU type.
191 This runs before SMP is initialized to avoid SMP problems with
192 self modifying code. This implies that assymetric systems where
193 APs have less capabilities than the boot processor are not handled.
194 Tough. Make sure you disable such features by hand. */
196 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
199 char insnbuf[MAX_PATCH_LEN];
201 DPRINTK("%s: alt table %p -> %p\n", __func__, start, end);
202 for (a = start; a < end; a++) {
203 u8 *instr = a->instr;
204 BUG_ON(a->replacementlen > a->instrlen);
205 BUG_ON(a->instrlen > sizeof(insnbuf));
206 if (!boot_cpu_has(a->cpuid))
209 /* vsyscall code is not mapped yet. resolve it manually. */
210 if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
211 instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
212 DPRINTK("%s: vsyscall fixup: %p => %p\n",
213 __func__, a->instr, instr);
216 memcpy(insnbuf, a->replacement, a->replacementlen);
217 add_nops(insnbuf + a->replacementlen,
218 a->instrlen - a->replacementlen);
219 text_poke_early(instr, insnbuf, a->instrlen);
225 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
229 for (ptr = start; ptr < end; ptr++) {
234 /* turn DS segment override prefix into lock prefix */
235 text_poke(*ptr, ((unsigned char []){0xf0}), 1);
239 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
246 for (ptr = start; ptr < end; ptr++) {
251 /* turn lock prefix into DS segment override prefix */
252 text_poke(*ptr, ((unsigned char []){0x3E}), 1);
256 struct smp_alt_module {
257 /* what is this ??? */
261 /* ptrs to lock prefixes */
265 /* .text segment, needed to avoid patching init code ;) */
269 struct list_head next;
271 static LIST_HEAD(smp_alt_modules);
272 static DEFINE_MUTEX(smp_alt);
273 static int smp_mode = 1; /* protected by smp_alt */
275 void alternatives_smp_module_add(struct module *mod, char *name,
276 void *locks, void *locks_end,
277 void *text, void *text_end)
279 struct smp_alt_module *smp;
285 if (boot_cpu_has(X86_FEATURE_UP))
286 alternatives_smp_unlock(locks, locks_end,
291 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
293 return; /* we'll run the (safe but slow) SMP code then ... */
298 smp->locks_end = locks_end;
300 smp->text_end = text_end;
301 DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
302 __func__, smp->locks, smp->locks_end,
303 smp->text, smp->text_end, smp->name);
305 mutex_lock(&smp_alt);
306 list_add_tail(&smp->next, &smp_alt_modules);
307 if (boot_cpu_has(X86_FEATURE_UP))
308 alternatives_smp_unlock(smp->locks, smp->locks_end,
309 smp->text, smp->text_end);
310 mutex_unlock(&smp_alt);
313 void alternatives_smp_module_del(struct module *mod)
315 struct smp_alt_module *item;
317 if (smp_alt_once || noreplace_smp)
320 mutex_lock(&smp_alt);
321 list_for_each_entry(item, &smp_alt_modules, next) {
322 if (mod != item->mod)
324 list_del(&item->next);
325 mutex_unlock(&smp_alt);
326 DPRINTK("%s: %s\n", __func__, item->name);
330 mutex_unlock(&smp_alt);
333 void alternatives_smp_switch(int smp)
335 struct smp_alt_module *mod;
337 #ifdef CONFIG_LOCKDEP
339 * Older binutils section handling bug prevented
340 * alternatives-replacement from working reliably.
342 * If this still occurs then you should see a hang
343 * or crash shortly after this line:
345 printk("lockdep: fixing up alternatives.\n");
348 if (noreplace_smp || smp_alt_once)
350 BUG_ON(!smp && (num_online_cpus() > 1));
352 mutex_lock(&smp_alt);
355 * Avoid unnecessary switches because it forces JIT based VMs to
356 * throw away all cached translations, which can be quite costly.
358 if (smp == smp_mode) {
361 printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
362 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
363 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
364 list_for_each_entry(mod, &smp_alt_modules, next)
365 alternatives_smp_lock(mod->locks, mod->locks_end,
366 mod->text, mod->text_end);
368 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
369 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
370 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
371 list_for_each_entry(mod, &smp_alt_modules, next)
372 alternatives_smp_unlock(mod->locks, mod->locks_end,
373 mod->text, mod->text_end);
376 mutex_unlock(&smp_alt);
381 #ifdef CONFIG_PARAVIRT
382 void apply_paravirt(struct paravirt_patch_site *start,
383 struct paravirt_patch_site *end)
385 struct paravirt_patch_site *p;
386 char insnbuf[MAX_PATCH_LEN];
388 if (noreplace_paravirt)
391 for (p = start; p < end; p++) {
394 BUG_ON(p->len > MAX_PATCH_LEN);
395 /* prep the buffer with the original instructions */
396 memcpy(insnbuf, p->instr, p->len);
397 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
398 (unsigned long)p->instr, p->len);
400 BUG_ON(used > p->len);
402 /* Pad the rest with nops */
403 add_nops(insnbuf + used, p->len - used);
404 text_poke_early(p->instr, insnbuf, p->len);
407 extern struct paravirt_patch_site __start_parainstructions[],
408 __stop_parainstructions[];
409 #endif /* CONFIG_PARAVIRT */
411 void __init alternative_instructions(void)
413 /* The patching is not fully atomic, so try to avoid local interruptions
414 that might execute the to be patched code.
415 Other CPUs are not running. */
419 * Don't stop machine check exceptions while patching.
420 * MCEs only happen when something got corrupted and in this
421 * case we must do something about the corruption.
422 * Ignoring it is worse than a unlikely patching race.
423 * Also machine checks tend to be broadcast and if one CPU
424 * goes into machine check the others follow quickly, so we don't
425 * expect a machine check to cause undue problems during to code
429 apply_alternatives(__alt_instructions, __alt_instructions_end);
431 /* switch to patch-once-at-boottime-only mode and free the
432 * tables in case we know the number of CPUs will never ever
434 #ifdef CONFIG_HOTPLUG_CPU
435 if (num_possible_cpus() < 2)
441 if (1 == num_possible_cpus()) {
442 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
443 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
444 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
446 alternatives_smp_unlock(__smp_locks, __smp_locks_end,
450 alternatives_smp_module_add(NULL, "core kernel",
451 __smp_locks, __smp_locks_end,
454 /* Only switch to UP mode if we don't immediately boot others */
455 if (num_present_cpus() == 1 || setup_max_cpus <= 1)
456 alternatives_smp_switch(0);
459 apply_paravirt(__parainstructions, __parainstructions_end);
462 free_init_pages("SMP alternatives",
463 (unsigned long)__smp_locks,
464 (unsigned long)__smp_locks_end);
470 * text_poke_early - Update instructions on a live kernel at boot time
471 * @addr: address to modify
472 * @opcode: source of the copy
473 * @len: length to copy
475 * When you use this code to patch more than one byte of an instruction
476 * you need to make sure that other CPUs cannot execute this code in parallel.
477 * Also no thread must be currently preempted in the middle of these
478 * instructions. And on the local CPU you need to be protected again NMI or MCE
479 * handlers seeing an inconsistent instruction while you patch.
481 void *text_poke_early(void *addr, const void *opcode, size_t len)
484 local_irq_save(flags);
485 memcpy(addr, opcode, len);
486 local_irq_restore(flags);
488 /* Could also do a CLFLUSH here to speed up CPU recovery; but
489 that causes hangs on some VIA CPUs. */
494 * text_poke - Update instructions on a live kernel
495 * @addr: address to modify
496 * @opcode: source of the copy
497 * @len: length to copy
499 * Only atomic text poke/set should be allowed when not doing early patching.
500 * It means the size must be writable atomically and the address must be aligned
501 * in a way that permits an atomic write. It also makes sure we fit on a single
504 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
508 struct page *pages[2];
512 if (!core_kernel_text((unsigned long)addr)) {
513 pages[0] = vmalloc_to_page(addr);
514 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
516 pages[0] = virt_to_page(addr);
517 WARN_ON(!PageReserved(pages[0]));
518 pages[1] = virt_to_page(addr + PAGE_SIZE);
523 vaddr = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
526 memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
530 /* Could also do a CLFLUSH here to speed up CPU recovery; but
531 that causes hangs on some VIA CPUs. */
532 for (i = 0; i < len; i++)
533 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);