arch/x86/kernel/alternative.c

   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>
   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 static inline const unsigned char*const * find_nop_table(void)
 147 {
 148         return boot_cpu_data.x86_vendor != X86_VENDOR_INTEL ||
 149                boot_cpu_data.x86 < 6 ? k8_nops : p6_nops;
 150 }
 151
 152 #else /* CONFIG_X86_64 */
 153
 154 static const struct nop {
 155         int cpuid;
 156         const unsigned char *const *noptable;
 157 } noptypes[] = {
 158         { X86_FEATURE_K8, k8_nops },
 159         { X86_FEATURE_K7, k7_nops },
 160         { X86_FEATURE_P4, p6_nops },
 161         { X86_FEATURE_P3, p6_nops },
 162         { -1, NULL }
 163 };
 164
 165 static const unsigned char*const * find_nop_table(void)
 166 {
 167         const unsigned char *const *noptable = intel_nops;
 168         int i;
 169
 170         for (i = 0; noptypes[i].cpuid >= 0; i++) {
 171                 if (boot_cpu_has(noptypes[i].cpuid)) {
 172                         noptable = noptypes[i].noptable;
 173                         break;
 174                 }
 175         }
 176         return noptable;
 177 }
 178
 179 #endif /* CONFIG_X86_64 */
 180
 181 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
 182 void add_nops(void *insns, unsigned int len)
 183 {
 184         const unsigned char *const *noptable = find_nop_table();
 185
 186         while (len > 0) {
 187                 unsigned int noplen = len;
 188                 if (noplen > ASM_NOP_MAX)
 189                         noplen = ASM_NOP_MAX;
 190                 memcpy(insns, noptable[noplen], noplen);
 191                 insns += noplen;
 192                 len -= noplen;
 193         }
 194 }
 195 EXPORT_SYMBOL_GPL(add_nops);
 196
 197 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 198 extern u8 *__smp_locks[], *__smp_locks_end[];
 199
 200 /* Replace instructions with better alternatives for this CPU type.
 201    This runs before SMP is initialized to avoid SMP problems with
 202    self modifying code. This implies that assymetric systems where
 203    APs have less capabilities than the boot processor are not handled.
 204    Tough. Make sure you disable such features by hand. */
 205
 206 void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
 207 {
 208         struct alt_instr *a;
 209         char insnbuf[MAX_PATCH_LEN];
 210
 211         DPRINTK("%s: alt table %p -> %p\n", __func__, start, end);
 212         for (a = start; a < end; a++) {
 213                 u8 *instr = a->instr;
 214                 BUG_ON(a->replacementlen > a->instrlen);
 215                 BUG_ON(a->instrlen > sizeof(insnbuf));
 216                 if (!boot_cpu_has(a->cpuid))
 217                         continue;
 218 #ifdef CONFIG_X86_64
 219                 /* vsyscall code is not mapped yet. resolve it manually. */
 220                 if (instr >= (u8 *)VSYSCALL_START && instr < (u8*)VSYSCALL_END) {
 221                         instr = __va(instr - (u8*)VSYSCALL_START + (u8*)__pa_symbol(&__vsyscall_0));
 222                         DPRINTK("%s: vsyscall fixup: %p => %p\n",
 223                                 __func__, a->instr, instr);
 224                 }
 225 #endif
 226                 memcpy(insnbuf, a->replacement, a->replacementlen);
 227                 add_nops(insnbuf + a->replacementlen,
 228                          a->instrlen - a->replacementlen);
 229                 text_poke_early(instr, insnbuf, a->instrlen);
 230         }
 231 }
 232
 233 #ifdef CONFIG_SMP
 234
 235 static void alternatives_smp_lock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 236 {
 237         u8 **ptr;
 238
 239         for (ptr = start; ptr < end; ptr++) {
 240                 if (*ptr < text)
 241                         continue;
 242                 if (*ptr > text_end)
 243                         continue;
 244                 text_poke(*ptr, ((unsigned char []){0xf0}), 1); /* add lock prefix */
 245         };
 246 }
 247
 248 static void alternatives_smp_unlock(u8 **start, u8 **end, u8 *text, u8 *text_end)
 249 {
 250         u8 **ptr;
 251         char insn[1];
 252
 253         if (noreplace_smp)
 254                 return;
 255
 256         add_nops(insn, 1);
 257         for (ptr = start; ptr < end; ptr++) {
 258                 if (*ptr < text)
 259                         continue;
 260                 if (*ptr > text_end)
 261                         continue;
 262                 text_poke(*ptr, insn, 1);
 263         };
 264 }
 265
 266 struct smp_alt_module {
 267         /* what is this ??? */
 268         struct module   *mod;
 269         char            *name;
 270
 271         /* ptrs to lock prefixes */
 272         u8              **locks;
 273         u8              **locks_end;
 274
 275         /* .text segment, needed to avoid patching init code ;) */
 276         u8              *text;
 277         u8              *text_end;
 278
 279         struct list_head next;
 280 };
 281 static LIST_HEAD(smp_alt_modules);
 282 static DEFINE_SPINLOCK(smp_alt);
 283 static int smp_mode = 1;        /* protected by smp_alt */
 284
 285 void alternatives_smp_module_add(struct module *mod, char *name,
 286                                  void *locks, void *locks_end,
 287                                  void *text,  void *text_end)
 288 {
 289         struct smp_alt_module *smp;
 290
 291         if (noreplace_smp)
 292                 return;
 293
 294         if (smp_alt_once) {
 295                 if (boot_cpu_has(X86_FEATURE_UP))
 296                         alternatives_smp_unlock(locks, locks_end,
 297                                                 text, text_end);
 298                 return;
 299         }
 300
 301         smp = kzalloc(sizeof(*smp), GFP_KERNEL);
 302         if (NULL == smp)
 303                 return; /* we'll run the (safe but slow) SMP code then ... */
 304
 305         smp->mod        = mod;
 306         smp->name       = name;
 307         smp->locks      = locks;
 308         smp->locks_end  = locks_end;
 309         smp->text       = text;
 310         smp->text_end   = text_end;
 311         DPRINTK("%s: locks %p -> %p, text %p -> %p, name %s\n",
 312                 __func__, smp->locks, smp->locks_end,
 313                 smp->text, smp->text_end, smp->name);
 314
 315         spin_lock(&smp_alt);
 316         list_add_tail(&smp->next, &smp_alt_modules);
 317         if (boot_cpu_has(X86_FEATURE_UP))
 318                 alternatives_smp_unlock(smp->locks, smp->locks_end,
 319                                         smp->text, smp->text_end);
 320         spin_unlock(&smp_alt);
 321 }
 322
 323 void alternatives_smp_module_del(struct module *mod)
 324 {
 325         struct smp_alt_module *item;
 326
 327         if (smp_alt_once || noreplace_smp)
 328                 return;
 329
 330         spin_lock(&smp_alt);
 331         list_for_each_entry(item, &smp_alt_modules, next) {
 332                 if (mod != item->mod)
 333                         continue;
 334                 list_del(&item->next);
 335                 spin_unlock(&smp_alt);
 336                 DPRINTK("%s: %s\n", __func__, item->name);
 337                 kfree(item);
 338                 return;
 339         }
 340         spin_unlock(&smp_alt);
 341 }
 342
 343 void alternatives_smp_switch(int smp)
 344 {
 345         struct smp_alt_module *mod;
 346
 347 #ifdef CONFIG_LOCKDEP
 348         /*
 349          * Older binutils section handling bug prevented
 350          * alternatives-replacement from working reliably.
 351          *
 352          * If this still occurs then you should see a hang
 353          * or crash shortly after this line:
 354          */
 355         printk("lockdep: fixing up alternatives.\n");
 356 #endif
 357
 358         if (noreplace_smp || smp_alt_once)
 359                 return;
 360         BUG_ON(!smp && (num_online_cpus() > 1));
 361
 362         spin_lock(&smp_alt);
 363
 364         /*
 365          * Avoid unnecessary switches because it forces JIT based VMs to
 366          * throw away all cached translations, which can be quite costly.
 367          */
 368         if (smp == smp_mode) {
 369                 /* nothing */
 370         } else if (smp) {
 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);
 377         } else {
 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);
 384         }
 385         smp_mode = smp;
 386         spin_unlock(&smp_alt);
 387 }
 388
 389 #endif
 390
 391 #ifdef CONFIG_PARAVIRT
 392 void apply_paravirt(struct paravirt_patch_site *start,
 393                     struct paravirt_patch_site *end)
 394 {
 395         struct paravirt_patch_site *p;
 396         char insnbuf[MAX_PATCH_LEN];
 397
 398         if (noreplace_paravirt)
 399                 return;
 400
 401         for (p = start; p < end; p++) {
 402                 unsigned int used;
 403
 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);
 409
 410                 BUG_ON(used > p->len);
 411
 412                 /* Pad the rest with nops */
 413                 add_nops(insnbuf + used, p->len - used);
 414                 text_poke_early(p->instr, insnbuf, p->len);
 415         }
 416 }
 417 extern struct paravirt_patch_site __start_parainstructions[],
 418         __stop_parainstructions[];
 419 #endif  /* CONFIG_PARAVIRT */
 420
 421 void __init alternative_instructions(void)
 422 {
 423         /* The patching is not fully atomic, so try to avoid local interruptions
 424            that might execute the to be patched code.
 425            Other CPUs are not running. */
 426         stop_nmi();
 427 #ifdef CONFIG_X86_MCE
 428         stop_mce();
 429 #endif
 430
 431         apply_alternatives(__alt_instructions, __alt_instructions_end);
 432
 433         /* switch to patch-once-at-boottime-only mode and free the
 434          * tables in case we know the number of CPUs will never ever
 435          * change */
 436 #ifdef CONFIG_HOTPLUG_CPU
 437         if (num_possible_cpus() < 2)
 438                 smp_alt_once = 1;
 439 #endif
 440
 441 #ifdef CONFIG_SMP
 442         if (smp_alt_once) {
 443                 if (1 == num_possible_cpus()) {
 444                         printk(KERN_INFO "SMP alternatives: switching to UP code\n");
 445                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 446                         set_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 447
 448                         alternatives_smp_unlock(__smp_locks, __smp_locks_end,
 449                                                 _text, _etext);
 450                 }
 451         } else {
 452                 alternatives_smp_module_add(NULL, "core kernel",
 453                                             __smp_locks, __smp_locks_end,
 454                                             _text, _etext);
 455
 456                 /* Only switch to UP mode if we don't immediately boot others */
 457                 if (num_possible_cpus() == 1 || setup_max_cpus <= 1)
 458                         alternatives_smp_switch(0);
 459         }
 460 #endif
 461         apply_paravirt(__parainstructions, __parainstructions_end);
 462
 463         if (smp_alt_once)
 464                 free_init_pages("SMP alternatives",
 465                                 (unsigned long)__smp_locks,
 466                                 (unsigned long)__smp_locks_end);
 467
 468         restart_nmi();
 469 #ifdef CONFIG_X86_MCE
 470         restart_mce();
 471 #endif
 472 }
 473
 474 /**
 475  * text_poke_early - Update instructions on a live kernel at boot time
 476  * @addr: address to modify
 477  * @opcode: source of the copy
 478  * @len: length to copy
 479  *
 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
 483  * instructions. And on the local CPU you need to be protected again NMI or MCE
 484  * handlers seeing an inconsistent instruction while you patch.
 485  */
 486 void *text_poke_early(void *addr, const void *opcode, size_t len)
 487 {
 488         unsigned long flags;
 489         local_irq_save(flags);
 490         memcpy(addr, opcode, len);
 491         local_irq_restore(flags);
 492         sync_core();
 493         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 494            that causes hangs on some VIA CPUs. */
 495         return addr;
 496 }
 497
 498 /**
 499  * text_poke - Update instructions on a live kernel
 500  * @addr: address to modify
 501  * @opcode: source of the copy
 502  * @len: length to copy
 503  *
 504  * Only atomic text poke/set should be allowed when not doing early patching.
 505  * It means the size must be writable atomically and the address must be aligned
 506  * in a way that permits an atomic write. It also makes sure we fit on a single
 507  * page.
 508  */
 509 void *__kprobes text_poke(void *addr, const void *opcode, size_t len)
 510 {
 511         unsigned long flags;
 512         char *vaddr;
 513         int nr_pages = 2;
 514         struct page *pages[2];
 515         int i;
 516
 517         if (!core_kernel_text((unsigned long)addr)) {
 518                 pages[0] = vmalloc_to_page(addr);
 519                 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
 520         } else {
 521                 pages[0] = virt_to_page(addr);
 522                 WARN_ON(!PageReserved(pages[0]));
 523                 pages[1] = virt_to_page(addr + PAGE_SIZE);
 524         }
 525         BUG_ON(!pages[0]);
 526         if (!pages[1])
 527                 nr_pages = 1;
 528         vaddr = vmap(pages, nr_pages, VM_MAP, PAGE_KERNEL);
 529         BUG_ON(!vaddr);
 530         local_irq_save(flags);
 531         memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
 532         local_irq_restore(flags);
 533         vunmap(vaddr);
 534         sync_core();
 535         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 536            that causes hangs on some VIA CPUs. */
 537         for (i = 0; i < len; i++)
 538                 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
 539         return addr;
 540 }