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/kprobes.h>
   6 #include <linux/mm.h>
   7 #include <linux/vmalloc.h>
   8 #include <asm/alternative.h>
   9 #include <asm/sections.h>
  10 #include <asm/pgtable.h>
  11 #include <asm/mce.h>
  12 #include <asm/nmi.h>
  13 #include <asm/vsyscall.h>
  14 #include <asm/cacheflush.h>
  15 #include <asm/io.h>
  16
  17 #define MAX_PATCH_LEN (255-1)
  18
  19 #ifdef CONFIG_HOTPLUG_CPU
  20 static int smp_alt_once;
  21
  22 static int __init bootonly(char *str)
  23 {
  24         smp_alt_once = 1;
  25         return 1;
  26 }
  27 __setup("smp-alt-boot", bootonly);
  28 #else
  29 #define smp_alt_once 1
  30 #endif
  31
  32 static int debug_alternative;
  33
  34 static int __init debug_alt(char *str)
  35 {
  36         debug_alternative = 1;
  37         return 1;
  38 }
  39 __setup("debug-alternative", debug_alt);
  40
  41 static int noreplace_smp;
  42
  43 static int __init setup_noreplace_smp(char *str)
  44 {
  45         noreplace_smp = 1;
  46         return 1;
  47 }
  48 __setup("noreplace-smp", setup_noreplace_smp);
  49
  50 #ifdef CONFIG_PARAVIRT
  51 static int noreplace_paravirt = 0;
  52
  53 static int __init setup_noreplace_paravirt(char *str)
  54 {
  55         noreplace_paravirt = 1;
  56         return 1;
  57 }
  58 __setup("noreplace-paravirt", setup_noreplace_paravirt);
  59 #endif
  60
  61 #define DPRINTK(fmt, args...) if (debug_alternative) \
  62         printk(KERN_DEBUG fmt, args)
  63
  64 #ifdef GENERIC_NOP1
  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
  71     "\t.previous");
  72 extern const unsigned char intelnops[];
  73 static const unsigned char *const intel_nops[ASM_NOP_MAX+1] = {
  74         NULL,
  75         intelnops,
  76         intelnops + 1,
  77         intelnops + 1 + 2,
  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,
  83 };
  84 #endif
  85
  86 #ifdef K8_NOP1
  87 asm("\t.section .rodata, \"a\"\nk8nops: "
  88         K8_NOP1 K8_NOP2 K8_NOP3 K8_NOP4 K8_NOP5 K8_NOP6
  89         K8_NOP7 K8_NOP8
  90     "\t.previous");
  91 extern const unsigned char k8nops[];
  92 static const unsigned char *const k8_nops[ASM_NOP_MAX+1] = {
  93         NULL,
  94         k8nops,
  95         k8nops + 1,
  96         k8nops + 1 + 2,
  97         k8nops + 1 + 2 + 3,
  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,
 102 };
 103 #endif
 104
 105 #ifdef K7_NOP1
 106 asm("\t.section .rodata, \"a\"\nk7nops: "
 107         K7_NOP1 K7_NOP2 K7_NOP3 K7_NOP4 K7_NOP5 K7_NOP6
 108         K7_NOP7 K7_NOP8
 109     "\t.previous");
 110 extern const unsigned char k7nops[];
 111 static const unsigned char *const k7_nops[ASM_NOP_MAX+1] = {
 112         NULL,
 113         k7nops,
 114         k7nops + 1,
 115         k7nops + 1 + 2,
 116         k7nops + 1 + 2 + 3,
 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,
 121 };
 122 #endif
 123
 124 #ifdef P6_NOP1
 125 asm("\t.section .rodata, \"a\"\np6nops: "
 126         P6_NOP1 P6_NOP2 P6_NOP3 P6_NOP4 P6_NOP5 P6_NOP6
 127         P6_NOP7 P6_NOP8
 128     "\t.previous");
 129 extern const unsigned char p6nops[];
 130 static const unsigned char *const p6_nops[ASM_NOP_MAX+1] = {
 131         NULL,
 132         p6nops,
 133         p6nops + 1,
 134         p6nops + 1 + 2,
 135         p6nops + 1 + 2 + 3,
 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,
 140 };
 141 #endif
 142
 143 #ifdef CONFIG_X86_64
 144
 145 extern char __vsyscall_0;
 146 const unsigned char *const *find_nop_table(void)
 147 {
 148         if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 149             boot_cpu_has(X86_FEATURE_NOPL))
 150                 return p6_nops;
 151         else
 152                 return k8_nops;
 153 }
 154
 155 #else /* CONFIG_X86_64 */
 156
 157 const unsigned char *const *find_nop_table(void)
 158 {
 159         if (boot_cpu_has(X86_FEATURE_K8))
 160                 return k8_nops;
 161         else if (boot_cpu_has(X86_FEATURE_K7))
 162                 return k7_nops;
 163         else if (boot_cpu_has(X86_FEATURE_NOPL))
 164                 return p6_nops;
 165         else
 166                 return intel_nops;
 167 }
 168
 169 #endif /* CONFIG_X86_64 */
 170
 171 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
 172 void add_nops(void *insns, unsigned int len)
 173 {
 174         const unsigned char *const *noptable = find_nop_table();
 175
 176         while (len > 0) {
 177                 unsigned int noplen = len;
 178                 if (noplen > ASM_NOP_MAX)
 179                         noplen = ASM_NOP_MAX;
 180                 memcpy(insns, noptable[noplen], noplen);
 181                 insns += noplen;
 182                 len -= noplen;
 183         }
 184 }
 185 EXPORT_SYMBOL_GPL(add_nops);
 186
 187 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 188 extern u8 *__smp_locks[], *__smp_locks_end[];
 189
 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. */
 195
 196 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
 197 {
 198         struct alt_instr *a;
 199         char insnbuf[MAX_PATCH_LEN];
 200
 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))
 207                         continue;
 208 #ifdef CONFIG_X86_64
 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);
 214                 }
 215 #endif
 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);
 220         }
 221 }
 222
 223 #ifdef CONFIG_SMP
 224
 225 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 226 {
 227         u8 **ptr;
 228
 229         for (ptr = start; ptr < end; ptr++) {
 230                 if (*ptr < text)
 231                         continue;
 232                 if (*ptr > text_end)
 233                         continue;
 234                 /* turn DS segment override prefix into lock prefix */
 235                 text_poke(*ptr, ((unsigned char []){0xf0}), 1);
 236         };
 237 }
 238
 239 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 240 {
 241         u8 **ptr;
 242
 243         if (noreplace_smp)
 244                 return;
 245
 246         for (ptr = start; ptr < end; ptr++) {
 247                 if (*ptr < text)
 248                         continue;
 249                 if (*ptr > text_end)
 250                         continue;
 251                 /* turn lock prefix into DS segment override prefix */
 252                 text_poke(*ptr, ((unsigned char []){0x3E}), 1);
 253         };
 254 }
 255
 256 struct smp_alt_module {
 257         /* what is this ??? */
 258         struct module   *mod;
 259         char            *name;
 260
 261         /* ptrs to lock prefixes */
 262         u8              **locks;
 263         u8              **locks_end;
 264
 265         /* .text segment, needed to avoid patching init code ;) */
 266         u8              *text;
 267         u8              *text_end;
 268
 269         struct list_head next;
 270 };
 271 static LIST_HEAD(smp_alt_modules);
 272 static DEFINE_MUTEX(smp_alt);
 273 static int smp_mode = 1;        /* protected by smp_alt */
 274
 275 void alternatives_smp_module_add(struct module *mod, char *name,
 276                                  void *locks, void *locks_end,
 277                                  void *text,  void *text_end)
 278 {
 279         struct smp_alt_module *smp;
 280
 281         if (noreplace_smp)
 282                 return;
 283
 284         if (smp_alt_once) {
 285                 if (boot_cpu_has(X86_FEATURE_UP))
 286                         alternatives_smp_unlock(locks, locks_end,
 287                                                 text, text_end);
 288                 return;
 289         }
 290
 291         smp = kzalloc(sizeof(*smp), GFP_KERNEL);
 292         if (NULL == smp)
 293                 return; /* we'll run the (safe but slow) SMP code then ... */
 294
 295         smp->mod        = mod;
 296         smp->name       = name;
 297         smp->locks      = locks;
 298         smp->locks_end  = locks_end;
 299         smp->text       = text;
 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);
 304
 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);
 311 }
 312
 313 void alternatives_smp_module_del(struct module *mod)
 314 {
 315         struct smp_alt_module *item;
 316
 317         if (smp_alt_once || noreplace_smp)
 318                 return;
 319
 320         mutex_lock(&smp_alt);
 321         list_for_each_entry(item, &smp_alt_modules, next) {
 322                 if (mod != item->mod)
 323                         continue;
 324                 list_del(&item->next);
 325                 mutex_unlock(&smp_alt);
 326                 DPRINTK("%s: %s\n", __func__, item->name);
 327                 kfree(item);
 328                 return;
 329         }
 330         mutex_unlock(&smp_alt);
 331 }
 332
 333 void alternatives_smp_switch(int smp)
 334 {
 335         struct smp_alt_module *mod;
 336
 337 #ifdef CONFIG_LOCKDEP
 338         /*
 339          * Older binutils section handling bug prevented
 340          * alternatives-replacement from working reliably.
 341          *
 342          * If this still occurs then you should see a hang
 343          * or crash shortly after this line:
 344          */
 345         printk("lockdep: fixing up alternatives.\n");
 346 #endif
 347
 348         if (noreplace_smp || smp_alt_once)
 349                 return;
 350         BUG_ON(!smp && (num_online_cpus() > 1));
 351
 352         mutex_lock(&smp_alt);
 353
 354         /*
 355          * Avoid unnecessary switches because it forces JIT based VMs to
 356          * throw away all cached translations, which can be quite costly.
 357          */
 358         if (smp == smp_mode) {
 359                 /* nothing */
 360         } else if (smp) {
 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);
 367         } else {
 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);
 374         }
 375         smp_mode = smp;
 376         mutex_unlock(&smp_alt);
 377 }
 378
 379 #endif
 380
 381 #ifdef CONFIG_PARAVIRT
 382 void apply_paravirt(struct paravirt_patch_site *start,
 383                     struct paravirt_patch_site *end)
 384 {
 385         struct paravirt_patch_site *p;
 386         char insnbuf[MAX_PATCH_LEN];
 387
 388         if (noreplace_paravirt)
 389                 return;
 390
 391         for (p = start; p < end; p++) {
 392                 unsigned int used;
 393
 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);
 399
 400                 BUG_ON(used > p->len);
 401
 402                 /* Pad the rest with nops */
 403                 add_nops(insnbuf + used, p->len - used);
 404                 text_poke_early(p->instr, insnbuf, p->len);
 405         }
 406 }
 407 extern struct paravirt_patch_site __start_parainstructions[],
 408         __stop_parainstructions[];
 409 #endif  /* CONFIG_PARAVIRT */
 410
 411 void __init alternative_instructions(void)
 412 {
 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. */
 416         stop_nmi();
 417 #ifdef CONFIG_X86_MCE
 418         stop_mce();
 419 #endif
 420
 421         apply_alternatives(__alt_instructions, __alt_instructions_end);
 422
 423         /* switch to patch-once-at-boottime-only mode and free the
 424          * tables in case we know the number of CPUs will never ever
 425          * change */
 426 #ifdef CONFIG_HOTPLUG_CPU
 427         if (num_possible_cpus() < 2)
 428                 smp_alt_once = 1;
 429 #endif
 430
 431 #ifdef CONFIG_SMP
 432         if (smp_alt_once) {
 433                 if (1 == num_possible_cpus()) {
 434                         printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 435                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 436                         set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 437
 438                         alternatives_smp_unlock(__smp_locks, __smp_locks_end,
 439                                                 _text, _etext);
 440                 }
 441         } else {
 442                 alternatives_smp_module_add(NULL, "core kernel",
 443                                             __smp_locks, __smp_locks_end,
 444                                             _text, _etext);
 445
 446                 /* Only switch to UP mode if we don't immediately boot others */
 447                 if (num_present_cpus() == 1 || setup_max_cpus <= 1)
 448                         alternatives_smp_switch(0);
 449         }
 450 #endif
 451         apply_paravirt(__parainstructions, __parainstructions_end);
 452
 453         if (smp_alt_once)
 454                 free_init_pages("SMP alternatives",
 455                                 (unsigned long)__smp_locks,
 456                                 (unsigned long)__smp_locks_end);
 457
 458         restart_nmi();
 459 #ifdef CONFIG_X86_MCE
 460         restart_mce();
 461 #endif
 462 }
 463
 464 /**
 465  * text_poke_early - Update instructions on a live kernel at boot time
 466  * @addr: address to modify
 467  * @opcode: source of the copy
 468  * @len: length to copy
 469  *
 470  * When you use this code to patch more than one byte of an instruction
 471  * you need to make sure that other CPUs cannot execute this code in parallel.
 472  * Also no thread must be currently preempted in the middle of these
 473  * instructions. And on the local CPU you need to be protected again NMI or MCE
 474  * handlers seeing an inconsistent instruction while you patch.
 475  */
 476 void *text_poke_early(void *addr, const void *opcode, size_t len)
 477 {
 478         unsigned long flags;
 479         local_irq_save(flags);
 480         memcpy(addr, opcode, len);
 481         local_irq_restore(flags);
 482         sync_core();
 483         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 484            that causes hangs on some VIA CPUs. */
 485         return addr;
 486 }
 487
 488 /**
 489  * text_poke - Update instructions on a live kernel
 490  * @addr: address to modify
 491  * @opcode: source of the copy
 492  * @len: length to copy
 493  *
 494  * Only atomic text poke/set should be allowed when not doing early patching.
 495  * It means the size must be writable atomically and the address must be aligned
 496  * in a way that permits an atomic write. It also makes sure we fit on a single
 497  * page.
 498  */
 499 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
 500 {
 501         unsigned long flags;
 502         char *vaddr;
 503         int nr_pages = 2;
 504         struct page *pages[2];
 505         int i;
 506
 507         if (!core_kernel_text((unsigned long)addr)) {
 508                 pages[0] = vmalloc_to_page(addr);
 509                 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
 510         } else {
 511                 pages[0] = virt_to_page(addr);
 512                 WARN_ON(!PageReserved(pages[0]));
 513                 pages[1] = virt_to_page(addr + PAGE_SIZE);
 514         }
 515         BUG_ON(!pages[0]);
 516         if (!pages[1])
 517                 nr_pages = 1;
 518         vaddr = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
 519         BUG_ON(!vaddr);
 520         local_irq_save(flags);
 521         memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
 522         local_irq_restore(flags);
 523         vunmap(vaddr);
 524         sync_core();
 525         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 526            that causes hangs on some VIA CPUs. */
 527         for (i = 0; i < len; i++)
 528                 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
 529         return addr;
 530 }