1 #include <linux/module.h>
2 #include <linux/sched.h>
3 #include <linux/spinlock.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>
15 #define MAX_PATCH_LEN (255-1)
17 #ifdef CONFIG_HOTPLUG_CPU
18 static int smp_alt_once;
20 static int __init bootonly(char *str)
25 __setup("smp-alt-boot", bootonly);
27 #define smp_alt_once 1
30 static int 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 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...) if (debug_alternative) \
60 printk(KERN_DEBUG fmt, args)
63 /* Use inline assembly to define this because the nops are defined
64 as inline assembly strings in the include files and we cannot
65 get them easily into strings. */
66 asm("\t.section .rodata, \"a\"\nintelnops: "
67 GENERIC_NOP1 GENERIC_NOP2 GENERIC_NOP3 GENERIC_NOP4 GENERIC_NOP5 GENERIC_NOP6
68 GENERIC_NOP7 GENERIC_NOP8
70 extern const unsigned char intelnops[];
71 static const unsigned char *const intel_nops[ASM_NOP_MAX+1] = {
76 intelnops + 1 + 2 + 3,
77 intelnops + 1 + 2 + 3 + 4,
78 intelnops + 1 + 2 + 3 + 4 + 5,
79 intelnops + 1 + 2 + 3 + 4 + 5 + 6,
80 intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
85 asm("\t.section .rodata, \"a\"\nk8nops: "
86 K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
89 extern const unsigned char k8nops[];
90 static const unsigned char *const k8_nops[ASM_NOP_MAX+1] = {
96 k8nops + 1 + 2 + 3 + 4,
97 k8nops + 1 + 2 + 3 + 4 + 5,
98 k8nops + 1 + 2 + 3 + 4 + 5 + 6,
99 k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
104 asm("\t.section .rodata, \"a\"\nk7nops: "
105 K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
108 extern const unsigned char k7nops[];
109 static const unsigned char *const k7_nops[ASM_NOP_MAX+1] = {
115 k7nops + 1 + 2 + 3 + 4,
116 k7nops + 1 + 2 + 3 + 4 + 5,
117 k7nops + 1 + 2 + 3 + 4 + 5 + 6,
118 k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
123 asm("\t.section .rodata, \"a\"\np6nops: "
124 P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
127 extern const unsigned char p6nops[];
128 static const unsigned char *const p6_nops[ASM_NOP_MAX+1] = {
134 p6nops + 1 + 2 + 3 + 4,
135 p6nops + 1 + 2 + 3 + 4 + 5,
136 p6nops + 1 + 2 + 3 + 4 + 5 + 6,
137 p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
143 extern char __vsyscall_0;
144 static inline const unsigned char*const * find_nop_table(void)
146 return boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
147 boot_cpu_data.x86 < 6 ? k8_nops : p6_nops;
150 #else /* CONFIG_X86_64 */
152 static const struct nop {
154 const unsigned char *const *noptable;
156 { X86_FEATURE_K8, k8_nops },
157 { X86_FEATURE_K7, k7_nops },
158 { X86_FEATURE_P4, p6_nops },
159 { X86_FEATURE_P3, p6_nops },
163 static const unsigned char*const * find_nop_table(void)
165 const unsigned char *const *noptable = intel_nops;
168 for (i = 0; noptypes[i].cpuid >= 0; i++) {
169 if (boot_cpu_has(noptypes[i].cpuid)) {
170 noptable = noptypes[i].noptable;
177 #endif /* CONFIG_X86_64 */
179 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
180 static void add_nops(void *insns, unsigned int len)
182 const unsigned char *const *noptable = find_nop_table();
185 unsigned int noplen = len;
186 if (noplen > ASM_NOP_MAX)
187 noplen = ASM_NOP_MAX;
188 memcpy(insns, noptable[noplen], noplen);
194 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
195 extern u8 *__smp_locks[], *__smp_locks_end[];
197 /* Replace instructions with better alternatives for this CPU type.
198 This runs before SMP is initialized to avoid SMP problems with
199 self modifying code. This implies that assymetric systems where
200 APs have less capabilities than the boot processor are not handled.
201 Tough. Make sure you disable such features by hand. */
203 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
206 char insnbuf[MAX_PATCH_LEN];
208 DPRINTK("%s: alt table %p -> %p\n", __FUNCTION__, start, end);
209 for (a = start; a < end; a++) {
210 u8 *instr = a->instr;
211 BUG_ON(a->replacementlen > a->instrlen);
212 BUG_ON(a->instrlen > sizeof(insnbuf));
213 if (!boot_cpu_has(a->cpuid))
216 /* vsyscall code is not mapped yet. resolve it manually. */
217 if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
218 instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
219 DPRINTK("%s: vsyscall fixup: %p => %p\n",
220 __FUNCTION__, a->instr, instr);
223 memcpy(insnbuf, a->replacement, a->replacementlen);
224 add_nops(insnbuf + a->replacementlen,
225 a->instrlen - a->replacementlen);
226 text_poke(instr, insnbuf, a->instrlen);
232 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
236 for (ptr = start; ptr < end; ptr++) {
241 text_poke(*ptr, ((unsigned char []){0xf0}), 1); /* add lock prefix */
245 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
254 for (ptr = start; ptr < end; ptr++) {
259 text_poke(*ptr, insn, 1);
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_SPINLOCK(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;
293 if (boot_cpu_has(X86_FEATURE_UP))
294 alternatives_smp_unlock(locks, locks_end,
299 smp = kzalloc(sizeof(*smp), GFP_KERNEL);
301 return; /* we'll run the (safe but slow) SMP code then ... */
306 smp->locks_end = locks_end;
308 smp->text_end = text_end;
309 DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
310 __FUNCTION__, smp->locks, smp->locks_end,
311 smp->text, smp->text_end, smp->name);
313 spin_lock_irqsave(&smp_alt, flags);
314 list_add_tail(&smp->next, &smp_alt_modules);
315 if (boot_cpu_has(X86_FEATURE_UP))
316 alternatives_smp_unlock(smp->locks, smp->locks_end,
317 smp->text, smp->text_end);
318 spin_unlock_irqrestore(&smp_alt, flags);
321 void alternatives_smp_module_del(struct module *mod)
323 struct smp_alt_module *item;
326 if (smp_alt_once || noreplace_smp)
329 spin_lock_irqsave(&smp_alt, flags);
330 list_for_each_entry(item, &smp_alt_modules, next) {
331 if (mod != item->mod)
333 list_del(&item->next);
334 spin_unlock_irqrestore(&smp_alt, flags);
335 DPRINTK("%s: %s\n", __FUNCTION__, item->name);
339 spin_unlock_irqrestore(&smp_alt, flags);
342 void alternatives_smp_switch(int smp)
344 struct smp_alt_module *mod;
347 #ifdef CONFIG_LOCKDEP
349 * Older binutils section handling bug prevented
350 * alternatives-replacement from working reliably.
352 * If this still occurs then you should see a hang
353 * or crash shortly after this line:
355 printk("lockdep: fixing up alternatives.\n");
358 if (noreplace_smp || smp_alt_once)
360 BUG_ON(!smp && (num_online_cpus() > 1));
362 spin_lock_irqsave(&smp_alt, flags);
365 * Avoid unnecessary switches because it forces JIT based VMs to
366 * throw away all cached translations, which can be quite costly.
368 if (smp == smp_mode) {
371 printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
372 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
373 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
374 list_for_each_entry(mod, &smp_alt_modules, next)
375 alternatives_smp_lock(mod->locks, mod->locks_end,
376 mod->text, mod->text_end);
378 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
379 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
380 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
381 list_for_each_entry(mod, &smp_alt_modules, next)
382 alternatives_smp_unlock(mod->locks, mod->locks_end,
383 mod->text, mod->text_end);
386 spin_unlock_irqrestore(&smp_alt, flags);
391 #ifdef CONFIG_PARAVIRT
392 void apply_paravirt(struct paravirt_patch_site *start,
393 struct paravirt_patch_site *end)
395 struct paravirt_patch_site *p;
396 char insnbuf[MAX_PATCH_LEN];
398 if (noreplace_paravirt)
401 for (p = start; p < end; p++) {
404 BUG_ON(p->len > MAX_PATCH_LEN);
405 /* prep the buffer with the original instructions */
406 memcpy(insnbuf, p->instr, p->len);
407 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
408 (unsigned long)p->instr, p->len);
410 BUG_ON(used > p->len);
412 /* Pad the rest with nops */
413 add_nops(insnbuf + used, p->len - used);
414 text_poke(p->instr, insnbuf, p->len);
417 extern struct paravirt_patch_site __start_parainstructions[],
418 __stop_parainstructions[];
419 #endif /* CONFIG_PARAVIRT */
421 void __init alternative_instructions(void)
425 /* The patching is not fully atomic, so try to avoid local interruptions
426 that might execute the to be patched code.
427 Other CPUs are not running. */
429 #ifdef CONFIG_X86_MCE
433 local_irq_save(flags);
434 apply_alternatives(__alt_instructions, __alt_instructions_end);
436 /* switch to patch-once-at-boottime-only mode and free the
437 * tables in case we know the number of CPUs will never ever
439 #ifdef CONFIG_HOTPLUG_CPU
440 if (num_possible_cpus() < 2)
446 if (1 == num_possible_cpus()) {
447 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
448 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
449 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
451 alternatives_smp_unlock(__smp_locks, __smp_locks_end,
455 alternatives_smp_module_add(NULL, "core kernel",
456 __smp_locks, __smp_locks_end,
459 /* Only switch to UP mode if we don't immediately boot others */
460 if (num_possible_cpus() == 1 || setup_max_cpus <= 1)
461 alternatives_smp_switch(0);
464 apply_paravirt(__parainstructions, __parainstructions_end);
465 local_irq_restore(flags);
468 free_init_pages("SMP alternatives",
469 (unsigned long)__smp_locks,
470 (unsigned long)__smp_locks_end);
473 #ifdef CONFIG_X86_MCE
480 * When you use this code to patch more than one byte of an instruction
481 * you need to make sure that other CPUs cannot execute this code in parallel.
482 * Also no thread must be currently preempted in the middle of these instructions.
483 * And on the local CPU you need to be protected again NMI or MCE handlers
484 * seeing an inconsistent instruction while you patch.
486 void __kprobes text_poke(void *addr, unsigned char *opcode, int len)
488 memcpy(addr, opcode, len);
490 /* Could also do a CLFLUSH here to speed up CPU recovery; but
491 that causes hangs on some VIA CPUs. */