arch/x86/kernel/alternative.c

   1 #include <linux/module.h>
   2 #include <linux/sched.h>
   3 #include <linux/mutex.h>
   4 #include <linux/list.h>
   5 #include <linux/stringify.h>
   6 #include <linux/kprobes.h>
   7 #include <linux/mm.h>
   8 #include <linux/vmalloc.h>
   9 #include <linux/memory.h>
  10 #include <asm/alternative.h>
  11 #include <asm/sections.h>
  12 #include <asm/pgtable.h>
  13 #include <asm/mce.h>
  14 #include <asm/nmi.h>
  15 #include <asm/vsyscall.h>
  16 #include <asm/cacheflush.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/io.h>
  19 #include <asm/fixmap.h>
  20
  21 #define MAX_PATCH_LEN (255-1)
  22
  23 #ifdef CONFIG_HOTPLUG_CPU
  24 static int smp_alt_once;
  25
  26 static int __init bootonly(char *str)
  27 {
  28         smp_alt_once = 1;
  29         return 1;
  30 }
  31 __setup("smp-alt-boot", bootonly);
  32 #else
  33 #define smp_alt_once 1
  34 #endif
  35
  36 static int __initdata_or_module debug_alternative;
  37
  38 static int __init debug_alt(char *str)
  39 {
  40         debug_alternative = 1;
  41         return 1;
  42 }
  43 __setup("debug-alternative", debug_alt);
  44
  45 static int noreplace_smp;
  46
  47 static int __init setup_noreplace_smp(char *str)
  48 {
  49         noreplace_smp = 1;
  50         return 1;
  51 }
  52 __setup("noreplace-smp", setup_noreplace_smp);
  53
  54 #ifdef CONFIG_PARAVIRT
  55 static int __initdata_or_module noreplace_paravirt = 0;
  56
  57 static int __init setup_noreplace_paravirt(char *str)
  58 {
  59         noreplace_paravirt = 1;
  60         return 1;
  61 }
  62 __setup("noreplace-paravirt", setup_noreplace_paravirt);
  63 #endif
  64
  65 #define DPRINTK(fmt, args...) if (debug_alternative) \
  66         printk(KERN_DEBUG fmt, args)
  67
  68 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
  69 /* Use inline assembly to define this because the nops are defined
  70    as inline assembly strings in the include files and we cannot
  71    get them easily into strings. */
  72 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nintelnops: "
  73         GENERIC_NOP1 GENERIC_NOP2 GENERIC_NOP3 GENERIC_NOP4 GENERIC_NOP5 GENERIC_NOP6
  74         GENERIC_NOP7 GENERIC_NOP8
  75     "\t.previous");
  76 extern const unsigned char intelnops[];
  77 static const unsigned char *const __initconst_or_module
  78 intel_nops[ASM_NOP_MAX+1] = {
  79         NULL,
  80         intelnops,
  81         intelnops + 1,
  82         intelnops + 1 + 2,
  83         intelnops + 1 + 2 + 3,
  84         intelnops + 1 + 2 + 3 + 4,
  85         intelnops + 1 + 2 + 3 + 4 + 5,
  86         intelnops + 1 + 2 + 3 + 4 + 5 + 6,
  87         intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
  88 };
  89 #endif
  90
  91 #ifdef K8_NOP1
  92 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nk8nops: "
  93         K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
  94         K8_NOP7 K8_NOP8
  95     "\t.previous");
  96 extern const unsigned char k8nops[];
  97 static const unsigned char *const __initconst_or_module
  98 k8_nops[ASM_NOP_MAX+1] = {
  99         NULL,
 100         k8nops,
 101         k8nops + 1,
 102         k8nops + 1 + 2,
 103         k8nops + 1 + 2 + 3,
 104         k8nops + 1 + 2 + 3 + 4,
 105         k8nops + 1 + 2 + 3 + 4 + 5,
 106         k8nops + 1 + 2 + 3 + 4 + 5 + 6,
 107         k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 108 };
 109 #endif
 110
 111 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
 112 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\nk7nops: "
 113         K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
 114         K7_NOP7 K7_NOP8
 115     "\t.previous");
 116 extern const unsigned char k7nops[];
 117 static const unsigned char *const __initconst_or_module
 118 k7_nops[ASM_NOP_MAX+1] = {
 119         NULL,
 120         k7nops,
 121         k7nops + 1,
 122         k7nops + 1 + 2,
 123         k7nops + 1 + 2 + 3,
 124         k7nops + 1 + 2 + 3 + 4,
 125         k7nops + 1 + 2 + 3 + 4 + 5,
 126         k7nops + 1 + 2 + 3 + 4 + 5 + 6,
 127         k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 128 };
 129 #endif
 130
 131 #ifdef P6_NOP1
 132 asm("\t" __stringify(__INITRODATA_OR_MODULE) "\np6nops: "
 133         P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
 134         P6_NOP7 P6_NOP8
 135     "\t.previous");
 136 extern const unsigned char p6nops[];
 137 static const unsigned char *const __initconst_or_module
 138 p6_nops[ASM_NOP_MAX+1] = {
 139         NULL,
 140         p6nops,
 141         p6nops + 1,
 142         p6nops + 1 + 2,
 143         p6nops + 1 + 2 + 3,
 144         p6nops + 1 + 2 + 3 + 4,
 145         p6nops + 1 + 2 + 3 + 4 + 5,
 146         p6nops + 1 + 2 + 3 + 4 + 5 + 6,
 147         p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 148 };
 149 #endif
 150
 151 #ifdef CONFIG_X86_64
 152
 153 extern char __vsyscall_0;
 154 static const unsigned char *const *__init_or_module find_nop_table(void)
 155 {
 156         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 157             boot_cpu_has(X86_FEATURE_NOPL))
 158                 return p6_nops;
 159         else
 160                 return k8_nops;
 161 }
 162
 163 #else /* CONFIG_X86_64 */
 164
 165 static const unsigned char *const *__init_or_module find_nop_table(void)
 166 {
 167         if (boot_cpu_has(X86_FEATURE_K8))
 168                 return k8_nops;
 169         else if (boot_cpu_has(X86_FEATURE_K7))
 170                 return k7_nops;
 171         else if (boot_cpu_has(X86_FEATURE_NOPL))
 172                 return p6_nops;
 173         else
 174                 return intel_nops;
 175 }
 176
 177 #endif /* CONFIG_X86_64 */
 178
 179 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
 180 static void __init_or_module add_nops(void *insns, unsigned int len)
 181 {
 182         const unsigned char *const *noptable = find_nop_table();
 183
 184         while (len > 0) {
 185                 unsigned int noplen = len;
 186                 if (noplen > ASM_NOP_MAX)
 187                         noplen = ASM_NOP_MAX;
 188                 memcpy(insns, noptable[noplen], noplen);
 189                 insns += noplen;
 190                 len -= noplen;
 191         }
 192 }
 193
 194 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 195 extern u8 *__smp_locks[], *__smp_locks_end[];
 196 static void *text_poke_early(void *addr, const void *opcode, size_t len);
 197
 198 /* Replace instructions with better alternatives for this CPU type.
 199    This runs before SMP is initialized to avoid SMP problems with
 200    self modifying code. This implies that assymetric systems where
 201    APs have less capabilities than the boot processor are not handled.
 202    Tough. Make sure you disable such features by hand. */
 203
 204 void __init_or_module apply_alternatives(struct alt_instr *start,
 205                                          struct alt_instr *end)
 206 {
 207         struct alt_instr *a;
 208         u8 insnbuf[MAX_PATCH_LEN];
 209
 210         DPRINTK("%s: alt table %p -> %p\n", __func__, start, end);
 211         for (a = start; a < end; a++) {
 212                 u8 *instr = a->instr;
 213                 BUG_ON(a->replacementlen > a->instrlen);
 214                 BUG_ON(a->instrlen > sizeof(insnbuf));
 215                 if (!boot_cpu_has(a->cpuid))
 216                         continue;
 217 #ifdef CONFIG_X86_64
 218                 /* vsyscall code is not mapped yet. resolve it manually. */
 219                 if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
 220                         instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
 221                         DPRINTK("%s: vsyscall fixup: %p => %p\n",
 222                                 __func__, a->instr, instr);
 223                 }
 224 #endif
 225                 memcpy(insnbuf, a->replacement, a->replacementlen);
 226                 if (*insnbuf == 0xe8 && a->replacementlen == 5)
 227                     *(s32 *)(insnbuf + 1) += a->replacement - a->instr;
 228                 add_nops(insnbuf + a->replacementlen,
 229                          a->instrlen - a->replacementlen);
 230                 text_poke_early(instr, insnbuf, a->instrlen);
 231         }
 232 }
 233
 234 #ifdef CONFIG_SMP
 235
 236 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 237 {
 238         u8 **ptr;
 239
 240         mutex_lock(&text_mutex);
 241         for (ptr = start; ptr < end; ptr++) {
 242                 if (*ptr < text)
 243                         continue;
 244                 if (*ptr > text_end)
 245                         continue;
 246                 /* turn DS segment override prefix into lock prefix */
 247                 text_poke(*ptr, ((unsigned char []){0xf0}), 1);
 248         };
 249         mutex_unlock(&text_mutex);
 250 }
 251
 252 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 253 {
 254         u8 **ptr;
 255
 256         if (noreplace_smp)
 257                 return;
 258
 259         mutex_lock(&text_mutex);
 260         for (ptr = start; ptr < end; ptr++) {
 261                 if (*ptr < text)
 262                         continue;
 263                 if (*ptr > text_end)
 264                         continue;
 265                 /* turn lock prefix into DS segment override prefix */
 266                 text_poke(*ptr, ((unsigned char []){0x3E}), 1);
 267         };
 268         mutex_unlock(&text_mutex);
 269 }
 270
 271 struct smp_alt_module {
 272         /* what is this ??? */
 273         struct module   *mod;
 274         char            *name;
 275
 276         /* ptrs to lock prefixes */
 277         u8              **locks;
 278         u8              **locks_end;
 279
 280         /* .text segment, needed to avoid patching init code ;) */
 281         u8              *text;
 282         u8              *text_end;
 283
 284         struct list_head next;
 285 };
 286 static LIST_HEAD(smp_alt_modules);
 287 static DEFINE_MUTEX(smp_alt);
 288 static int smp_mode = 1;        /* protected by smp_alt */
 289
 290 void __init_or_module alternatives_smp_module_add(struct module *mod,
 291                                                   char *name,
 292                                                   void *locks, void *locks_end,
 293                                                   void *text,  void *text_end)
 294 {
 295         struct smp_alt_module *smp;
 296
 297         if (noreplace_smp)
 298                 return;
 299
 300         if (smp_alt_once) {
 301                 if (boot_cpu_has(X86_FEATURE_UP))
 302                         alternatives_smp_unlock(locks, locks_end,
 303                                                 text, text_end);
 304                 return;
 305         }
 306
 307         smp = kzalloc(sizeof(*smp), GFP_KERNEL);
 308         if (NULL == smp)
 309                 return; /* we'll run the (safe but slow) SMP code then ... */
 310
 311         smp->mod        = mod;
 312         smp->name       = name;
 313         smp->locks      = locks;
 314         smp->locks_end  = locks_end;
 315         smp->text       = text;
 316         smp->text_end   = text_end;
 317         DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
 318                 __func__, smp->locks, smp->locks_end,
 319                 smp->text, smp->text_end, smp->name);
 320
 321         mutex_lock(&smp_alt);
 322         list_add_tail(&smp->next, &smp_alt_modules);
 323         if (boot_cpu_has(X86_FEATURE_UP))
 324                 alternatives_smp_unlock(smp->locks, smp->locks_end,
 325                                         smp->text, smp->text_end);
 326         mutex_unlock(&smp_alt);
 327 }
 328
 329 void __init_or_module alternatives_smp_module_del(struct module *mod)
 330 {
 331         struct smp_alt_module *item;
 332
 333         if (smp_alt_once || noreplace_smp)
 334                 return;
 335
 336         mutex_lock(&smp_alt);
 337         list_for_each_entry(item, &smp_alt_modules, next) {
 338                 if (mod != item->mod)
 339                         continue;
 340                 list_del(&item->next);
 341                 mutex_unlock(&smp_alt);
 342                 DPRINTK("%s: %s\n", __func__, item->name);
 343                 kfree(item);
 344                 return;
 345         }
 346         mutex_unlock(&smp_alt);
 347 }
 348
 349 void alternatives_smp_switch(int smp)
 350 {
 351         struct smp_alt_module *mod;
 352
 353 #ifdef CONFIG_LOCKDEP
 354         /*
 355          * Older binutils section handling bug prevented
 356          * alternatives-replacement from working reliably.
 357          *
 358          * If this still occurs then you should see a hang
 359          * or crash shortly after this line:
 360          */
 361         printk("lockdep: fixing up alternatives.\n");
 362 #endif
 363
 364         if (noreplace_smp || smp_alt_once)
 365                 return;
 366         BUG_ON(!smp && (num_online_cpus() > 1));
 367
 368         mutex_lock(&smp_alt);
 369
 370         /*
 371          * Avoid unnecessary switches because it forces JIT based VMs to
 372          * throw away all cached translations, which can be quite costly.
 373          */
 374         if (smp == smp_mode) {
 375                 /* nothing */
 376         } else if (smp) {
 377                 printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
 378                 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 379                 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 380                 list_for_each_entry(mod, &smp_alt_modules, next)
 381                         alternatives_smp_lock(mod->locks, mod->locks_end,
 382                                               mod->text, mod->text_end);
 383         } else {
 384                 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 385                 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 386                 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 387                 list_for_each_entry(mod, &smp_alt_modules, next)
 388                         alternatives_smp_unlock(mod->locks, mod->locks_end,
 389                                                 mod->text, mod->text_end);
 390         }
 391         smp_mode = smp;
 392         mutex_unlock(&smp_alt);
 393 }
 394
 395 /* Return 1 if the address range is reserved for smp-alternatives */
 396 int alternatives_text_reserved(void *start, void *end)
 397 {
 398         struct smp_alt_module *mod;
 399         u8 **ptr;
 400         u8 *text_start = start;
 401         u8 *text_end = end;
 402
 403         list_for_each_entry(mod, &smp_alt_modules, next) {
 404                 if (mod->text > text_end || mod->text_end < text_start)
 405                         continue;
 406                 for (ptr = mod->locks; ptr < mod->locks_end; ptr++)
 407                         if (text_start <= *ptr && text_end >= *ptr)
 408                                 return 1;
 409         }
 410
 411         return 0;
 412 }
 413 #endif
 414
 415 #ifdef CONFIG_PARAVIRT
 416 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
 417                                      struct paravirt_patch_site *end)
 418 {
 419         struct paravirt_patch_site *p;
 420         char insnbuf[MAX_PATCH_LEN];
 421
 422         if (noreplace_paravirt)
 423                 return;
 424
 425         for (p = start; p < end; p++) {
 426                 unsigned int used;
 427
 428                 BUG_ON(p->len > MAX_PATCH_LEN);
 429                 /* prep the buffer with the original instructions */
 430                 memcpy(insnbuf, p->instr, p->len);
 431                 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
 432                                          (unsigned long)p->instr, p->len);
 433
 434                 BUG_ON(used > p->len);
 435
 436                 /* Pad the rest with nops */
 437                 add_nops(insnbuf + used, p->len - used);
 438                 text_poke_early(p->instr, insnbuf, p->len);
 439         }
 440 }
 441 extern struct paravirt_patch_site __start_parainstructions[],
 442         __stop_parainstructions[];
 443 #endif  /* CONFIG_PARAVIRT */
 444
 445 void __init alternative_instructions(void)
 446 {
 447         /* The patching is not fully atomic, so try to avoid local interruptions
 448            that might execute the to be patched code.
 449            Other CPUs are not running. */
 450         stop_nmi();
 451
 452         /*
 453          * Don't stop machine check exceptions while patching.
 454          * MCEs only happen when something got corrupted and in this
 455          * case we must do something about the corruption.
 456          * Ignoring it is worse than a unlikely patching race.
 457          * Also machine checks tend to be broadcast and if one CPU
 458          * goes into machine check the others follow quickly, so we don't
 459          * expect a machine check to cause undue problems during to code
 460          * patching.
 461          */
 462
 463         apply_alternatives(__alt_instructions, __alt_instructions_end);
 464
 465         /* switch to patch-once-at-boottime-only mode and free the
 466          * tables in case we know the number of CPUs will never ever
 467          * change */
 468 #ifdef CONFIG_HOTPLUG_CPU
 469         if (num_possible_cpus() < 2)
 470                 smp_alt_once = 1;
 471 #endif
 472
 473 #ifdef CONFIG_SMP
 474         if (smp_alt_once) {
 475                 if (1 == num_possible_cpus()) {
 476                         printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 477                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 478                         set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 479
 480                         alternatives_smp_unlock(__smp_locks, __smp_locks_end,
 481                                                 _text, _etext);
 482                 }
 483         } else {
 484                 alternatives_smp_module_add(NULL, "core kernel",
 485                                             __smp_locks, __smp_locks_end,
 486                                             _text, _etext);
 487
 488                 /* Only switch to UP mode if we don't immediately boot others */
 489                 if (num_present_cpus() == 1 || setup_max_cpus <= 1)
 490                         alternatives_smp_switch(0);
 491         }
 492 #endif
 493         apply_paravirt(__parainstructions, __parainstructions_end);
 494
 495         if (smp_alt_once)
 496                 free_init_pages("SMP alternatives",
 497                                 (unsigned long)__smp_locks,
 498                                 (unsigned long)__smp_locks_end);
 499
 500         restart_nmi();
 501 }
 502
 503 /**
 504  * text_poke_early - Update instructions on a live kernel at boot time
 505  * @addr: address to modify
 506  * @opcode: source of the copy
 507  * @len: length to copy
 508  *
 509  * When you use this code to patch more than one byte of an instruction
 510  * you need to make sure that other CPUs cannot execute this code in parallel.
 511  * Also no thread must be currently preempted in the middle of these
 512  * instructions. And on the local CPU you need to be protected again NMI or MCE
 513  * handlers seeing an inconsistent instruction while you patch.
 514  */
 515 static void *__init_or_module text_poke_early(void *addr, const void *opcode,
 516                                               size_t len)
 517 {
 518         unsigned long flags;
 519         local_irq_save(flags);
 520         memcpy(addr, opcode, len);
 521         sync_core();
 522         local_irq_restore(flags);
 523         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 524            that causes hangs on some VIA CPUs. */
 525         return addr;
 526 }
 527
 528 /**
 529  * text_poke - Update instructions on a live kernel
 530  * @addr: address to modify
 531  * @opcode: source of the copy
 532  * @len: length to copy
 533  *
 534  * Only atomic text poke/set should be allowed when not doing early patching.
 535  * It means the size must be writable atomically and the address must be aligned
 536  * in a way that permits an atomic write. It also makes sure we fit on a single
 537  * page.
 538  *
 539  * Note: Must be called under text_mutex.
 540  */
 541 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
 542 {
 543         unsigned long flags;
 544         char *vaddr;
 545         struct page *pages[2];
 546         int i;
 547
 548         if (!core_kernel_text((unsigned long)addr)) {
 549                 pages[0] = vmalloc_to_page(addr);
 550                 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
 551         } else {
 552                 pages[0] = virt_to_page(addr);
 553                 WARN_ON(!PageReserved(pages[0]));
 554                 pages[1] = virt_to_page(addr + PAGE_SIZE);
 555         }
 556         BUG_ON(!pages[0]);
 557         local_irq_save(flags);
 558         set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
 559         if (pages[1])
 560                 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
 561         vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
 562         memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
 563         clear_fixmap(FIX_TEXT_POKE0);
 564         if (pages[1])
 565                 clear_fixmap(FIX_TEXT_POKE1);
 566         local_flush_tlb();
 567         sync_core();
 568         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 569            that causes hangs on some VIA CPUs. */
 570         for (i = 0; i < len; i++)
 571                 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
 572         local_irq_restore(flags);
 573         return addr;
 574 }