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         char 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                 add_nops(insnbuf + a->replacementlen,
 227                          a->instrlen - a->replacementlen);
 228                 text_poke_early(instr, insnbuf, a->instrlen);
 229         }
 230 }
 231
 232 #ifdef CONFIG_SMP
 233
 234 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 235 {
 236         u8 **ptr;
 237
 238         mutex_lock(&text_mutex);
 239         for (ptr = start; ptr < end; ptr++) {
 240                 if (*ptr < text)
 241                         continue;
 242                 if (*ptr > text_end)
 243                         continue;
 244                 /* turn DS segment override prefix into lock prefix */
 245                 text_poke(*ptr, ((unsigned char []){0xf0}), 1);
 246         };
 247         mutex_unlock(&text_mutex);
 248 }
 249
 250 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 251 {
 252         u8 **ptr;
 253
 254         if (noreplace_smp)
 255                 return;
 256
 257         mutex_lock(&text_mutex);
 258         for (ptr = start; ptr < end; ptr++) {
 259                 if (*ptr < text)
 260                         continue;
 261                 if (*ptr > text_end)
 262                         continue;
 263                 /* turn lock prefix into DS segment override prefix */
 264                 text_poke(*ptr, ((unsigned char []){0x3E}), 1);
 265         };
 266         mutex_unlock(&text_mutex);
 267 }
 268
 269 struct smp_alt_module {
 270         /* what is this ??? */
 271         struct module   *mod;
 272         char            *name;
 273
 274         /* ptrs to lock prefixes */
 275         u8              **locks;
 276         u8              **locks_end;
 277
 278         /* .text segment, needed to avoid patching init code ;) */
 279         u8              *text;
 280         u8              *text_end;
 281
 282         struct list_head next;
 283 };
 284 static LIST_HEAD(smp_alt_modules);
 285 static DEFINE_MUTEX(smp_alt);
 286 static int smp_mode = 1;        /* protected by smp_alt */
 287
 288 void __init_or_module alternatives_smp_module_add(struct module *mod,
 289                                                   char *name,
 290                                                   void *locks, void *locks_end,
 291                                                   void *text,  void *text_end)
 292 {
 293         struct smp_alt_module *smp;
 294
 295         if (noreplace_smp)
 296                 return;
 297
 298         if (smp_alt_once) {
 299                 if (boot_cpu_has(X86_FEATURE_UP))
 300                         alternatives_smp_unlock(locks, locks_end,
 301                                                 text, text_end);
 302                 return;
 303         }
 304
 305         smp = kzalloc(sizeof(*smp), GFP_KERNEL);
 306         if (NULL == smp)
 307                 return; /* we'll run the (safe but slow) SMP code then ... */
 308
 309         smp->mod        = mod;
 310         smp->name       = name;
 311         smp->locks      = locks;
 312         smp->locks_end  = locks_end;
 313         smp->text       = text;
 314         smp->text_end   = text_end;
 315         DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
 316                 __func__, smp->locks, smp->locks_end,
 317                 smp->text, smp->text_end, smp->name);
 318
 319         mutex_lock(&smp_alt);
 320         list_add_tail(&smp->next, &smp_alt_modules);
 321         if (boot_cpu_has(X86_FEATURE_UP))
 322                 alternatives_smp_unlock(smp->locks, smp->locks_end,
 323                                         smp->text, smp->text_end);
 324         mutex_unlock(&smp_alt);
 325 }
 326
 327 void __init_or_module alternatives_smp_module_del(struct module *mod)
 328 {
 329         struct smp_alt_module *item;
 330
 331         if (smp_alt_once || noreplace_smp)
 332                 return;
 333
 334         mutex_lock(&smp_alt);
 335         list_for_each_entry(item, &smp_alt_modules, next) {
 336                 if (mod != item->mod)
 337                         continue;
 338                 list_del(&item->next);
 339                 mutex_unlock(&smp_alt);
 340                 DPRINTK("%s: %s\n", __func__, item->name);
 341                 kfree(item);
 342                 return;
 343         }
 344         mutex_unlock(&smp_alt);
 345 }
 346
 347 void alternatives_smp_switch(int smp)
 348 {
 349         struct smp_alt_module *mod;
 350
 351 #ifdef CONFIG_LOCKDEP
 352         /*
 353          * Older binutils section handling bug prevented
 354          * alternatives-replacement from working reliably.
 355          *
 356          * If this still occurs then you should see a hang
 357          * or crash shortly after this line:
 358          */
 359         printk("lockdep: fixing up alternatives.\n");
 360 #endif
 361
 362         if (noreplace_smp || smp_alt_once)
 363                 return;
 364         BUG_ON(!smp && (num_online_cpus() > 1));
 365
 366         mutex_lock(&smp_alt);
 367
 368         /*
 369          * Avoid unnecessary switches because it forces JIT based VMs to
 370          * throw away all cached translations, which can be quite costly.
 371          */
 372         if (smp == smp_mode) {
 373                 /* nothing */
 374         } else if (smp) {
 375                 printk(KERN_INFO "SMP alternatives: switching to SMP code\n");
 376                 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 377                 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 378                 list_for_each_entry(mod, &smp_alt_modules, next)
 379                         alternatives_smp_lock(mod->locks, mod->locks_end,
 380                                               mod->text, mod->text_end);
 381         } else {
 382                 printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 383                 set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 384                 set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 385                 list_for_each_entry(mod, &smp_alt_modules, next)
 386                         alternatives_smp_unlock(mod->locks, mod->locks_end,
 387                                                 mod->text, mod->text_end);
 388         }
 389         smp_mode = smp;
 390         mutex_unlock(&smp_alt);
 391 }
 392
 393 #endif
 394
 395 #ifdef CONFIG_PARAVIRT
 396 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
 397                                      struct paravirt_patch_site *end)
 398 {
 399         struct paravirt_patch_site *p;
 400         char insnbuf[MAX_PATCH_LEN];
 401
 402         if (noreplace_paravirt)
 403                 return;
 404
 405         for (p = start; p < end; p++) {
 406                 unsigned int used;
 407
 408                 BUG_ON(p->len > MAX_PATCH_LEN);
 409                 /* prep the buffer with the original instructions */
 410                 memcpy(insnbuf, p->instr, p->len);
 411                 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
 412                                          (unsigned long)p->instr, p->len);
 413
 414                 BUG_ON(used > p->len);
 415
 416                 /* Pad the rest with nops */
 417                 add_nops(insnbuf + used, p->len - used);
 418                 text_poke_early(p->instr, insnbuf, p->len);
 419         }
 420 }
 421 extern struct paravirt_patch_site __start_parainstructions[],
 422         __stop_parainstructions[];
 423 #endif  /* CONFIG_PARAVIRT */
 424
 425 void __init alternative_instructions(void)
 426 {
 427         /* The patching is not fully atomic, so try to avoid local interruptions
 428            that might execute the to be patched code.
 429            Other CPUs are not running. */
 430         stop_nmi();
 431
 432         /*
 433          * Don't stop machine check exceptions while patching.
 434          * MCEs only happen when something got corrupted and in this
 435          * case we must do something about the corruption.
 436          * Ignoring it is worse than a unlikely patching race.
 437          * Also machine checks tend to be broadcast and if one CPU
 438          * goes into machine check the others follow quickly, so we don't
 439          * expect a machine check to cause undue problems during to code
 440          * patching.
 441          */
 442
 443         apply_alternatives(__alt_instructions, __alt_instructions_end);
 444
 445         /* switch to patch-once-at-boottime-only mode and free the
 446          * tables in case we know the number of CPUs will never ever
 447          * change */
 448 #ifdef CONFIG_HOTPLUG_CPU
 449         if (num_possible_cpus() < 2)
 450                 smp_alt_once = 1;
 451 #endif
 452
 453 #ifdef CONFIG_SMP
 454         if (smp_alt_once) {
 455                 if (1 == num_possible_cpus()) {
 456                         printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 457                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 458                         set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 459
 460                         alternatives_smp_unlock(__smp_locks, __smp_locks_end,
 461                                                 _text, _etext);
 462                 }
 463         } else {
 464                 alternatives_smp_module_add(NULL, "core kernel",
 465                                             __smp_locks, __smp_locks_end,
 466                                             _text, _etext);
 467
 468                 /* Only switch to UP mode if we don't immediately boot others */
 469                 if (num_present_cpus() == 1 || setup_max_cpus <= 1)
 470                         alternatives_smp_switch(0);
 471         }
 472 #endif
 473         apply_paravirt(__parainstructions, __parainstructions_end);
 474
 475         if (smp_alt_once)
 476                 free_init_pages("SMP alternatives",
 477                                 (unsigned long)__smp_locks,
 478                                 (unsigned long)__smp_locks_end);
 479
 480         restart_nmi();
 481 }
 482
 483 /**
 484  * text_poke_early - Update instructions on a live kernel at boot time
 485  * @addr: address to modify
 486  * @opcode: source of the copy
 487  * @len: length to copy
 488  *
 489  * When you use this code to patch more than one byte of an instruction
 490  * you need to make sure that other CPUs cannot execute this code in parallel.
 491  * Also no thread must be currently preempted in the middle of these
 492  * instructions. And on the local CPU you need to be protected again NMI or MCE
 493  * handlers seeing an inconsistent instruction while you patch.
 494  */
 495 static void *__init_or_module text_poke_early(void *addr, const void *opcode,
 496                                               size_t len)
 497 {
 498         unsigned long flags;
 499         local_irq_save(flags);
 500         memcpy(addr, opcode, len);
 501         sync_core();
 502         local_irq_restore(flags);
 503         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 504            that causes hangs on some VIA CPUs. */
 505         return addr;
 506 }
 507
 508 /**
 509  * text_poke - Update instructions on a live kernel
 510  * @addr: address to modify
 511  * @opcode: source of the copy
 512  * @len: length to copy
 513  *
 514  * Only atomic text poke/set should be allowed when not doing early patching.
 515  * It means the size must be writable atomically and the address must be aligned
 516  * in a way that permits an atomic write. It also makes sure we fit on a single
 517  * page.
 518  *
 519  * Note: Must be called under text_mutex.
 520  */
 521 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
 522 {
 523         unsigned long flags;
 524         char *vaddr;
 525         struct page *pages[2];
 526         int i;
 527
 528         if (!core_kernel_text((unsigned long)addr)) {
 529                 pages[0] = vmalloc_to_page(addr);
 530                 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
 531         } else {
 532                 pages[0] = virt_to_page(addr);
 533                 WARN_ON(!PageReserved(pages[0]));
 534                 pages[1] = virt_to_page(addr + PAGE_SIZE);
 535         }
 536         BUG_ON(!pages[0]);
 537         local_irq_save(flags);
 538         set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
 539         if (pages[1])
 540                 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
 541         vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
 542         memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
 543         clear_fixmap(FIX_TEXT_POKE0);
 544         if (pages[1])
 545                 clear_fixmap(FIX_TEXT_POKE1);
 546         local_flush_tlb();
 547         sync_core();
 548         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 549            that causes hangs on some VIA CPUs. */
 550         for (i = 0; i < len; i++)
 551                 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
 552         local_irq_restore(flags);
 553         return addr;
 554 }