arch/x86/kernel/alternative.c

   1 #define pr_fmt(fmt) "SMP alternatives: " fmt
   2
   3 #include <linux/module.h>
   4 #include <linux/sched.h>
   5 #include <linux/mutex.h>
   6 #include <linux/list.h>
   7 #include <linux/stringify.h>
   8 #include <linux/mm.h>
   9 #include <linux/vmalloc.h>
  10 #include <linux/memory.h>
  11 #include <linux/stop_machine.h>
  12 #include <linux/slab.h>
  13 #include <linux/kdebug.h>
  14 #include <asm/text-patching.h>
  15 #include <asm/alternative.h>
  16 #include <asm/sections.h>
  17 #include <asm/pgtable.h>
  18 #include <asm/mce.h>
  19 #include <asm/nmi.h>
  20 #include <asm/cacheflush.h>
  21 #include <asm/tlbflush.h>
  22 #include <asm/io.h>
  23 #include <asm/fixmap.h>
  24
  25 int __read_mostly alternatives_patched;
  26
  27 EXPORT_SYMBOL_GPL(alternatives_patched);
  28
  29 #define MAX_PATCH_LEN (255-1)
  30
  31 static int __initdata_or_module debug_alternative;
  32
  33 static int __init debug_alt(char *str)
  34 {
  35         debug_alternative = 1;
  36         return 1;
  37 }
  38 __setup("debug-alternative", debug_alt);
  39
  40 static int noreplace_smp;
  41
  42 static int __init setup_noreplace_smp(char *str)
  43 {
  44         noreplace_smp = 1;
  45         return 1;
  46 }
  47 __setup("noreplace-smp", setup_noreplace_smp);
  48
  49 #define DPRINTK(fmt, args...)                                           \
  50 do {                                                                    \
  51         if (debug_alternative)                                          \
  52                 printk(KERN_DEBUG "%s: " fmt "\n", __func__, ##args);   \
  53 } while (0)
  54
  55 #define DUMP_BYTES(buf, len, fmt, args...)                              \
  56 do {                                                                    \
  57         if (unlikely(debug_alternative)) {                              \
  58                 int j;                                                  \
  59                                                                         \
  60                 if (!(len))                                             \
  61                         break;                                          \
  62                                                                         \
  63                 printk(KERN_DEBUG fmt, ##args);                         \
  64                 for (j = 0; j < (len) - 1; j++)                         \
  65                         printk(KERN_CONT "%02hhx ", buf[j]);            \
  66                 printk(KERN_CONT "%02hhx\n", buf[j]);                   \
  67         }                                                               \
  68 } while (0)
  69
  70 /*
  71  * Each GENERIC_NOPX is of X bytes, and defined as an array of bytes
  72  * that correspond to that nop. Getting from one nop to the next, we
  73  * add to the array the offset that is equal to the sum of all sizes of
  74  * nops preceding the one we are after.
  75  *
  76  * Note: The GENERIC_NOP5_ATOMIC is at the end, as it breaks the
  77  * nice symmetry of sizes of the previous nops.
  78  */
  79 #if defined(GENERIC_NOP1) && !defined(CONFIG_X86_64)
  80 static const unsigned char intelnops[] =
  81 {
  82         GENERIC_NOP1,
  83         GENERIC_NOP2,
  84         GENERIC_NOP3,
  85         GENERIC_NOP4,
  86         GENERIC_NOP5,
  87         GENERIC_NOP6,
  88         GENERIC_NOP7,
  89         GENERIC_NOP8,
  90         GENERIC_NOP5_ATOMIC
  91 };
  92 static const unsigned char * const intel_nops[ASM_NOP_MAX+2] =
  93 {
  94         NULL,
  95         intelnops,
  96         intelnops + 1,
  97         intelnops + 1 + 2,
  98         intelnops + 1 + 2 + 3,
  99         intelnops + 1 + 2 + 3 + 4,
 100         intelnops + 1 + 2 + 3 + 4 + 5,
 101         intelnops + 1 + 2 + 3 + 4 + 5 + 6,
 102         intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 103         intelnops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
 104 };
 105 #endif
 106
 107 #ifdef K8_NOP1
 108 static const unsigned char k8nops[] =
 109 {
 110         K8_NOP1,
 111         K8_NOP2,
 112         K8_NOP3,
 113         K8_NOP4,
 114         K8_NOP5,
 115         K8_NOP6,
 116         K8_NOP7,
 117         K8_NOP8,
 118         K8_NOP5_ATOMIC
 119 };
 120 static const unsigned char * const k8_nops[ASM_NOP_MAX+2] =
 121 {
 122         NULL,
 123         k8nops,
 124         k8nops + 1,
 125         k8nops + 1 + 2,
 126         k8nops + 1 + 2 + 3,
 127         k8nops + 1 + 2 + 3 + 4,
 128         k8nops + 1 + 2 + 3 + 4 + 5,
 129         k8nops + 1 + 2 + 3 + 4 + 5 + 6,
 130         k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 131         k8nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
 132 };
 133 #endif
 134
 135 #if defined(K7_NOP1) && !defined(CONFIG_X86_64)
 136 static const unsigned char k7nops[] =
 137 {
 138         K7_NOP1,
 139         K7_NOP2,
 140         K7_NOP3,
 141         K7_NOP4,
 142         K7_NOP5,
 143         K7_NOP6,
 144         K7_NOP7,
 145         K7_NOP8,
 146         K7_NOP5_ATOMIC
 147 };
 148 static const unsigned char * const k7_nops[ASM_NOP_MAX+2] =
 149 {
 150         NULL,
 151         k7nops,
 152         k7nops + 1,
 153         k7nops + 1 + 2,
 154         k7nops + 1 + 2 + 3,
 155         k7nops + 1 + 2 + 3 + 4,
 156         k7nops + 1 + 2 + 3 + 4 + 5,
 157         k7nops + 1 + 2 + 3 + 4 + 5 + 6,
 158         k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 159         k7nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
 160 };
 161 #endif
 162
 163 #ifdef P6_NOP1
 164 static const unsigned char p6nops[] =
 165 {
 166         P6_NOP1,
 167         P6_NOP2,
 168         P6_NOP3,
 169         P6_NOP4,
 170         P6_NOP5,
 171         P6_NOP6,
 172         P6_NOP7,
 173         P6_NOP8,
 174         P6_NOP5_ATOMIC
 175 };
 176 static const unsigned char * const p6_nops[ASM_NOP_MAX+2] =
 177 {
 178         NULL,
 179         p6nops,
 180         p6nops + 1,
 181         p6nops + 1 + 2,
 182         p6nops + 1 + 2 + 3,
 183         p6nops + 1 + 2 + 3 + 4,
 184         p6nops + 1 + 2 + 3 + 4 + 5,
 185         p6nops + 1 + 2 + 3 + 4 + 5 + 6,
 186         p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7,
 187         p6nops + 1 + 2 + 3 + 4 + 5 + 6 + 7 + 8,
 188 };
 189 #endif
 190
 191 /* Initialize these to a safe default */
 192 #ifdef CONFIG_X86_64
 193 const unsigned char * const *ideal_nops = p6_nops;
 194 #else
 195 const unsigned char * const *ideal_nops = intel_nops;
 196 #endif
 197
 198 void __init arch_init_ideal_nops(void)
 199 {
 200         switch (boot_cpu_data.x86_vendor) {
 201         case X86_VENDOR_INTEL:
 202                 /*
 203                  * Due to a decoder implementation quirk, some
 204                  * specific Intel CPUs actually perform better with
 205                  * the "k8_nops" than with the SDM-recommended NOPs.
 206                  */
 207                 if (boot_cpu_data.x86 == 6 &&
 208                     boot_cpu_data.x86_model >= 0x0f &&
 209                     boot_cpu_data.x86_model != 0x1c &&
 210                     boot_cpu_data.x86_model != 0x26 &&
 211                     boot_cpu_data.x86_model != 0x27 &&
 212                     boot_cpu_data.x86_model < 0x30) {
 213                         ideal_nops = k8_nops;
 214                 } else if (boot_cpu_has(X86_FEATURE_NOPL)) {
 215                            ideal_nops = p6_nops;
 216                 } else {
 217 #ifdef CONFIG_X86_64
 218                         ideal_nops = k8_nops;
 219 #else
 220                         ideal_nops = intel_nops;
 221 #endif
 222                 }
 223                 break;
 224
 225         case X86_VENDOR_AMD:
 226                 if (boot_cpu_data.x86 > 0xf) {
 227                         ideal_nops = p6_nops;
 228                         return;
 229                 }
 230
 231                 /* fall through */
 232
 233         default:
 234 #ifdef CONFIG_X86_64
 235                 ideal_nops = k8_nops;
 236 #else
 237                 if (boot_cpu_has(X86_FEATURE_K8))
 238                         ideal_nops = k8_nops;
 239                 else if (boot_cpu_has(X86_FEATURE_K7))
 240                         ideal_nops = k7_nops;
 241                 else
 242                         ideal_nops = intel_nops;
 243 #endif
 244         }
 245 }
 246
 247 /* Use this to add nops to a buffer, then text_poke the whole buffer. */
 248 static void __init_or_module add_nops(void *insns, unsigned int len)
 249 {
 250         while (len > 0) {
 251                 unsigned int noplen = len;
 252                 if (noplen > ASM_NOP_MAX)
 253                         noplen = ASM_NOP_MAX;
 254                 memcpy(insns, ideal_nops[noplen], noplen);
 255                 insns += noplen;
 256                 len -= noplen;
 257         }
 258 }
 259
 260 extern struct alt_instr __alt_instructions[], __alt_instructions_end[];
 261 extern s32 __smp_locks[], __smp_locks_end[];
 262 void *text_poke_early(void *addr, const void *opcode, size_t len);
 263
 264 /*
 265  * Are we looking at a near JMP with a 1 or 4-byte displacement.
 266  */
 267 static inline bool is_jmp(const u8 opcode)
 268 {
 269         return opcode == 0xeb || opcode == 0xe9;
 270 }
 271
 272 static void __init_or_module
 273 recompute_jump(struct alt_instr *a, u8 *orig_insn, u8 *repl_insn, u8 *insnbuf)
 274 {
 275         u8 *next_rip, *tgt_rip;
 276         s32 n_dspl, o_dspl;
 277         int repl_len;
 278
 279         if (a->replacementlen != 5)
 280                 return;
 281
 282         o_dspl = *(s32 *)(insnbuf + 1);
 283
 284         /* next_rip of the replacement JMP */
 285         next_rip = repl_insn + a->replacementlen;
 286         /* target rip of the replacement JMP */
 287         tgt_rip  = next_rip + o_dspl;
 288         n_dspl = tgt_rip - orig_insn;
 289
 290         DPRINTK("target RIP: %px, new_displ: 0x%x", tgt_rip, n_dspl);
 291
 292         if (tgt_rip - orig_insn >= 0) {
 293                 if (n_dspl - 2 <= 127)
 294                         goto two_byte_jmp;
 295                 else
 296                         goto five_byte_jmp;
 297         /* negative offset */
 298         } else {
 299                 if (((n_dspl - 2) & 0xff) == (n_dspl - 2))
 300                         goto two_byte_jmp;
 301                 else
 302                         goto five_byte_jmp;
 303         }
 304
 305 two_byte_jmp:
 306         n_dspl -= 2;
 307
 308         insnbuf[0] = 0xeb;
 309         insnbuf[1] = (s8)n_dspl;
 310         add_nops(insnbuf + 2, 3);
 311
 312         repl_len = 2;
 313         goto done;
 314
 315 five_byte_jmp:
 316         n_dspl -= 5;
 317
 318         insnbuf[0] = 0xe9;
 319         *(s32 *)&insnbuf[1] = n_dspl;
 320
 321         repl_len = 5;
 322
 323 done:
 324
 325         DPRINTK("final displ: 0x%08x, JMP 0x%lx",
 326                 n_dspl, (unsigned long)orig_insn + n_dspl + repl_len);
 327 }
 328
 329 /*
 330  * "noinline" to cause control flow change and thus invalidate I$ and
 331  * cause refetch after modification.
 332  */
 333 static void __init_or_module noinline optimize_nops(struct alt_instr *a, u8 *instr)
 334 {
 335         unsigned long flags;
 336         int i;
 337
 338         for (i = 0; i < a->padlen; i++) {
 339                 if (instr[i] != 0x90)
 340                         return;
 341         }
 342
 343         local_irq_save(flags);
 344         add_nops(instr + (a->instrlen - a->padlen), a->padlen);
 345         local_irq_restore(flags);
 346
 347         DUMP_BYTES(instr, a->instrlen, "%px: [%d:%d) optimized NOPs: ",
 348                    instr, a->instrlen - a->padlen, a->padlen);
 349 }
 350
 351 /*
 352  * Replace instructions with better alternatives for this CPU type. This runs
 353  * before SMP is initialized to avoid SMP problems with self modifying code.
 354  * This implies that asymmetric systems where APs have less capabilities than
 355  * the boot processor are not handled. Tough. Make sure you disable such
 356  * features by hand.
 357  *
 358  * Marked "noinline" to cause control flow change and thus insn cache
 359  * to refetch changed I$ lines.
 360  */
 361 void __init_or_module noinline apply_alternatives(struct alt_instr *start,
 362                                                   struct alt_instr *end)
 363 {
 364         struct alt_instr *a;
 365         u8 *instr, *replacement;
 366         u8 insnbuf[MAX_PATCH_LEN];
 367
 368         DPRINTK("alt table %px, -> %px", start, end);
 369         /*
 370          * The scan order should be from start to end. A later scanned
 371          * alternative code can overwrite previously scanned alternative code.
 372          * Some kernel functions (e.g. memcpy, memset, etc) use this order to
 373          * patch code.
 374          *
 375          * So be careful if you want to change the scan order to any other
 376          * order.
 377          */
 378         for (a = start; a < end; a++) {
 379                 int insnbuf_sz = 0;
 380
 381                 instr = (u8 *)&a->instr_offset + a->instr_offset;
 382                 replacement = (u8 *)&a->repl_offset + a->repl_offset;
 383                 BUG_ON(a->instrlen > sizeof(insnbuf));
 384                 BUG_ON(a->cpuid >= (NCAPINTS + NBUGINTS) * 32);
 385                 if (!boot_cpu_has(a->cpuid)) {
 386                         if (a->padlen > 1)
 387                                 optimize_nops(a, instr);
 388
 389                         continue;
 390                 }
 391
 392                 DPRINTK("feat: %d*32+%d, old: (%px len: %d), repl: (%px, len: %d), pad: %d",
 393                         a->cpuid >> 5,
 394                         a->cpuid & 0x1f,
 395                         instr, a->instrlen,
 396                         replacement, a->replacementlen, a->padlen);
 397
 398                 DUMP_BYTES(instr, a->instrlen, "%px: old_insn: ", instr);
 399                 DUMP_BYTES(replacement, a->replacementlen, "%px: rpl_insn: ", replacement);
 400
 401                 memcpy(insnbuf, replacement, a->replacementlen);
 402                 insnbuf_sz = a->replacementlen;
 403
 404                 /*
 405                  * 0xe8 is a relative jump; fix the offset.
 406                  *
 407                  * Instruction length is checked before the opcode to avoid
 408                  * accessing uninitialized bytes for zero-length replacements.
 409                  */
 410                 if (a->replacementlen == 5 && *insnbuf == 0xe8) {
 411                         *(s32 *)(insnbuf + 1) += replacement - instr;
 412                         DPRINTK("Fix CALL offset: 0x%x, CALL 0x%lx",
 413                                 *(s32 *)(insnbuf + 1),
 414                                 (unsigned long)instr + *(s32 *)(insnbuf + 1) + 5);
 415                 }
 416
 417                 if (a->replacementlen && is_jmp(replacement[0]))
 418                         recompute_jump(a, instr, replacement, insnbuf);
 419
 420                 if (a->instrlen > a->replacementlen) {
 421                         add_nops(insnbuf + a->replacementlen,
 422                                  a->instrlen - a->replacementlen);
 423                         insnbuf_sz += a->instrlen - a->replacementlen;
 424                 }
 425                 DUMP_BYTES(insnbuf, insnbuf_sz, "%px: final_insn: ", instr);
 426
 427                 text_poke_early(instr, insnbuf, insnbuf_sz);
 428         }
 429 }
 430
 431 #ifdef CONFIG_SMP
 432 static void alternatives_smp_lock(const s32 *start, const s32 *end,
 433                                   u8 *text, u8 *text_end)
 434 {
 435         const s32 *poff;
 436
 437         for (poff = start; poff < end; poff++) {
 438                 u8 *ptr = (u8 *)poff + *poff;
 439
 440                 if (!*poff || ptr < text || ptr >= text_end)
 441                         continue;
 442                 /* turn DS segment override prefix into lock prefix */
 443                 if (*ptr == 0x3e)
 444                         text_poke(ptr, ((unsigned char []){0xf0}), 1);
 445         }
 446 }
 447
 448 static void alternatives_smp_unlock(const s32 *start, const s32 *end,
 449                                     u8 *text, u8 *text_end)
 450 {
 451         const s32 *poff;
 452
 453         for (poff = start; poff < end; poff++) {
 454                 u8 *ptr = (u8 *)poff + *poff;
 455
 456                 if (!*poff || ptr < text || ptr >= text_end)
 457                         continue;
 458                 /* turn lock prefix into DS segment override prefix */
 459                 if (*ptr == 0xf0)
 460                         text_poke(ptr, ((unsigned char []){0x3E}), 1);
 461         }
 462 }
 463
 464 struct smp_alt_module {
 465         /* what is this ??? */
 466         struct module   *mod;
 467         char            *name;
 468
 469         /* ptrs to lock prefixes */
 470         const s32       *locks;
 471         const s32       *locks_end;
 472
 473         /* .text segment, needed to avoid patching init code ;) */
 474         u8              *text;
 475         u8              *text_end;
 476
 477         struct list_head next;
 478 };
 479 static LIST_HEAD(smp_alt_modules);
 480 static bool uniproc_patched = false;    /* protected by text_mutex */
 481
 482 void __init_or_module alternatives_smp_module_add(struct module *mod,
 483                                                   char *name,
 484                                                   void *locks, void *locks_end,
 485                                                   void *text,  void *text_end)
 486 {
 487         struct smp_alt_module *smp;
 488
 489         mutex_lock(&text_mutex);
 490         if (!uniproc_patched)
 491                 goto unlock;
 492
 493         if (num_possible_cpus() == 1)
 494                 /* Don't bother remembering, we'll never have to undo it. */
 495                 goto smp_unlock;
 496
 497         smp = kzalloc(sizeof(*smp), GFP_KERNEL);
 498         if (NULL == smp)
 499                 /* we'll run the (safe but slow) SMP code then ... */
 500                 goto unlock;
 501
 502         smp->mod        = mod;
 503         smp->name       = name;
 504         smp->locks      = locks;
 505         smp->locks_end  = locks_end;
 506         smp->text       = text;
 507         smp->text_end   = text_end;
 508         DPRINTK("locks %p -> %p, text %p -> %p, name %s\n",
 509                 smp->locks, smp->locks_end,
 510                 smp->text, smp->text_end, smp->name);
 511
 512         list_add_tail(&smp->next, &smp_alt_modules);
 513 smp_unlock:
 514         alternatives_smp_unlock(locks, locks_end, text, text_end);
 515 unlock:
 516         mutex_unlock(&text_mutex);
 517 }
 518
 519 void __init_or_module alternatives_smp_module_del(struct module *mod)
 520 {
 521         struct smp_alt_module *item;
 522
 523         mutex_lock(&text_mutex);
 524         list_for_each_entry(item, &smp_alt_modules, next) {
 525                 if (mod != item->mod)
 526                         continue;
 527                 list_del(&item->next);
 528                 kfree(item);
 529                 break;
 530         }
 531         mutex_unlock(&text_mutex);
 532 }
 533
 534 void alternatives_enable_smp(void)
 535 {
 536         struct smp_alt_module *mod;
 537
 538         /* Why bother if there are no other CPUs? */
 539         BUG_ON(num_possible_cpus() == 1);
 540
 541         mutex_lock(&text_mutex);
 542
 543         if (uniproc_patched) {
 544                 pr_info("switching to SMP code\n");
 545                 BUG_ON(num_online_cpus() != 1);
 546                 clear_cpu_cap(&boot_cpu_data, X86_FEATURE_UP);
 547                 clear_cpu_cap(&cpu_data(0), X86_FEATURE_UP);
 548                 list_for_each_entry(mod, &smp_alt_modules, next)
 549                         alternatives_smp_lock(mod->locks, mod->locks_end,
 550                                               mod->text, mod->text_end);
 551                 uniproc_patched = false;
 552         }
 553         mutex_unlock(&text_mutex);
 554 }
 555
 556 /*
 557  * Return 1 if the address range is reserved for SMP-alternatives.
 558  * Must hold text_mutex.
 559  */
 560 int alternatives_text_reserved(void *start, void *end)
 561 {
 562         struct smp_alt_module *mod;
 563         const s32 *poff;
 564         u8 *text_start = start;
 565         u8 *text_end = end;
 566
 567         lockdep_assert_held(&text_mutex);
 568
 569         list_for_each_entry(mod, &smp_alt_modules, next) {
 570                 if (mod->text > text_end || mod->text_end < text_start)
 571                         continue;
 572                 for (poff = mod->locks; poff < mod->locks_end; poff++) {
 573                         const u8 *ptr = (const u8 *)poff + *poff;
 574
 575                         if (text_start <= ptr && text_end > ptr)
 576                                 return 1;
 577                 }
 578         }
 579
 580         return 0;
 581 }
 582 #endif /* CONFIG_SMP */
 583
 584 #ifdef CONFIG_PARAVIRT
 585 void __init_or_module apply_paravirt(struct paravirt_patch_site *start,
 586                                      struct paravirt_patch_site *end)
 587 {
 588         struct paravirt_patch_site *p;
 589         char insnbuf[MAX_PATCH_LEN];
 590
 591         for (p = start; p < end; p++) {
 592                 unsigned int used;
 593
 594                 BUG_ON(p->len > MAX_PATCH_LEN);
 595                 /* prep the buffer with the original instructions */
 596                 memcpy(insnbuf, p->instr, p->len);
 597                 used = pv_init_ops.patch(p->instrtype, p->clobbers, insnbuf,
 598                                          (unsigned long)p->instr, p->len);
 599
 600                 BUG_ON(used > p->len);
 601
 602                 /* Pad the rest with nops */
 603                 add_nops(insnbuf + used, p->len - used);
 604                 text_poke_early(p->instr, insnbuf, p->len);
 605         }
 606 }
 607 extern struct paravirt_patch_site __start_parainstructions[],
 608         __stop_parainstructions[];
 609 #endif  /* CONFIG_PARAVIRT */
 610
 611 void __init alternative_instructions(void)
 612 {
 613         /* The patching is not fully atomic, so try to avoid local interruptions
 614            that might execute the to be patched code.
 615            Other CPUs are not running. */
 616         stop_nmi();
 617
 618         /*
 619          * Don't stop machine check exceptions while patching.
 620          * MCEs only happen when something got corrupted and in this
 621          * case we must do something about the corruption.
 622          * Ignoring it is worse than a unlikely patching race.
 623          * Also machine checks tend to be broadcast and if one CPU
 624          * goes into machine check the others follow quickly, so we don't
 625          * expect a machine check to cause undue problems during to code
 626          * patching.
 627          */
 628
 629         apply_alternatives(__alt_instructions, __alt_instructions_end);
 630
 631 #ifdef CONFIG_SMP
 632         /* Patch to UP if other cpus not imminent. */
 633         if (!noreplace_smp && (num_present_cpus() == 1 || setup_max_cpus <= 1)) {
 634                 uniproc_patched = true;
 635                 alternatives_smp_module_add(NULL, "core kernel",
 636                                             __smp_locks, __smp_locks_end,
 637                                             _text, _etext);
 638         }
 639
 640         if (!uniproc_patched || num_possible_cpus() == 1)
 641                 free_init_pages("SMP alternatives",
 642                                 (unsigned long)__smp_locks,
 643                                 (unsigned long)__smp_locks_end);
 644 #endif
 645
 646         apply_paravirt(__parainstructions, __parainstructions_end);
 647
 648         restart_nmi();
 649         alternatives_patched = 1;
 650 }
 651
 652 /**
 653  * text_poke_early - Update instructions on a live kernel at boot time
 654  * @addr: address to modify
 655  * @opcode: source of the copy
 656  * @len: length to copy
 657  *
 658  * When you use this code to patch more than one byte of an instruction
 659  * you need to make sure that other CPUs cannot execute this code in parallel.
 660  * Also no thread must be currently preempted in the middle of these
 661  * instructions. And on the local CPU you need to be protected again NMI or MCE
 662  * handlers seeing an inconsistent instruction while you patch.
 663  */
 664 void *__init_or_module text_poke_early(void *addr, const void *opcode,
 665                                               size_t len)
 666 {
 667         unsigned long flags;
 668         local_irq_save(flags);
 669         memcpy(addr, opcode, len);
 670         local_irq_restore(flags);
 671         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 672            that causes hangs on some VIA CPUs. */
 673         return addr;
 674 }
 675
 676 /**
 677  * text_poke - Update instructions on a live kernel
 678  * @addr: address to modify
 679  * @opcode: source of the copy
 680  * @len: length to copy
 681  *
 682  * Only atomic text poke/set should be allowed when not doing early patching.
 683  * It means the size must be writable atomically and the address must be aligned
 684  * in a way that permits an atomic write. It also makes sure we fit on a single
 685  * page.
 686  *
 687  * Note: Must be called under text_mutex.
 688  */
 689 void *text_poke(void *addr, const void *opcode, size_t len)
 690 {
 691         unsigned long flags;
 692         char *vaddr;
 693         struct page *pages[2];
 694         int i;
 695
 696         if (!core_kernel_text((unsigned long)addr)) {
 697                 pages[0] = vmalloc_to_page(addr);
 698                 pages[1] = vmalloc_to_page(addr + PAGE_SIZE);
 699         } else {
 700                 pages[0] = virt_to_page(addr);
 701                 WARN_ON(!PageReserved(pages[0]));
 702                 pages[1] = virt_to_page(addr + PAGE_SIZE);
 703         }
 704         BUG_ON(!pages[0]);
 705         local_irq_save(flags);
 706         set_fixmap(FIX_TEXT_POKE0, page_to_phys(pages[0]));
 707         if (pages[1])
 708                 set_fixmap(FIX_TEXT_POKE1, page_to_phys(pages[1]));
 709         vaddr = (char *)fix_to_virt(FIX_TEXT_POKE0);
 710         memcpy(&vaddr[(unsigned long)addr & ~PAGE_MASK], opcode, len);
 711         clear_fixmap(FIX_TEXT_POKE0);
 712         if (pages[1])
 713                 clear_fixmap(FIX_TEXT_POKE1);
 714         local_flush_tlb();
 715         sync_core();
 716         /* Could also do a CLFLUSH here to speed up CPU recovery; but
 717            that causes hangs on some VIA CPUs. */
 718         for (i = 0; i < len; i++)
 719                 BUG_ON(((char *)addr)[i] != ((char *)opcode)[i]);
 720         local_irq_restore(flags);
 721         return addr;
 722 }
 723
 724 static void do_sync_core(void *info)
 725 {
 726         sync_core();
 727 }
 728
 729 static bool bp_patching_in_progress;
 730 static void *bp_int3_handler, *bp_int3_addr;
 731
 732 int poke_int3_handler(struct pt_regs *regs)
 733 {
 734         /*
 735          * Having observed our INT3 instruction, we now must observe
 736          * bp_patching_in_progress.
 737          *
 738          *      in_progress = TRUE              INT3
 739          *      WMB                             RMB
 740          *      write INT3                      if (in_progress)
 741          *
 742          * Idem for bp_int3_handler.
 743          */
 744         smp_rmb();
 745
 746         if (likely(!bp_patching_in_progress))
 747                 return 0;
 748
 749         if (user_mode(regs) || regs->ip != (unsigned long)bp_int3_addr)
 750                 return 0;
 751
 752         /* set up the specified breakpoint handler */
 753         regs->ip = (unsigned long) bp_int3_handler;
 754
 755         return 1;
 756
 757 }
 758
 759 /**
 760  * text_poke_bp() -- update instructions on live kernel on SMP
 761  * @addr:       address to patch
 762  * @opcode:     opcode of new instruction
 763  * @len:        length to copy
 764  * @handler:    address to jump to when the temporary breakpoint is hit
 765  *
 766  * Modify multi-byte instruction by using int3 breakpoint on SMP.
 767  * We completely avoid stop_machine() here, and achieve the
 768  * synchronization using int3 breakpoint.
 769  *
 770  * The way it is done:
 771  *      - add a int3 trap to the address that will be patched
 772  *      - sync cores
 773  *      - update all but the first byte of the patched range
 774  *      - sync cores
 775  *      - replace the first byte (int3) by the first byte of
 776  *        replacing opcode
 777  *      - sync cores
 778  *
 779  * Note: must be called under text_mutex.
 780  */
 781 void *text_poke_bp(void *addr, const void *opcode, size_t len, void *handler)
 782 {
 783         unsigned char int3 = 0xcc;
 784
 785         bp_int3_handler = handler;
 786         bp_int3_addr = (u8 *)addr + sizeof(int3);
 787         bp_patching_in_progress = true;
 788         /*
 789          * Corresponding read barrier in int3 notifier for making sure the
 790          * in_progress and handler are correctly ordered wrt. patching.
 791          */
 792         smp_wmb();
 793
 794         text_poke(addr, &int3, sizeof(int3));
 795
 796         on_each_cpu(do_sync_core, NULL, 1);
 797
 798         if (len - sizeof(int3) > 0) {
 799                 /* patch all but the first byte */
 800                 text_poke((char *)addr + sizeof(int3),
 801                           (const char *) opcode + sizeof(int3),
 802                           len - sizeof(int3));
 803                 /*
 804                  * According to Intel, this core syncing is very likely
 805                  * not necessary and we'd be safe even without it. But
 806                  * better safe than sorry (plus there's not only Intel).
 807                  */
 808                 on_each_cpu(do_sync_core, NULL, 1);
 809         }
 810
 811         /* patch the first byte */
 812         text_poke(addr, opcode, sizeof(int3));
 813
 814         on_each_cpu(do_sync_core, NULL, 1);
 815         /*
 816          * sync_core() implies an smp_mb() and orders this store against
 817          * the writing of the new instruction.
 818          */
 819         bp_patching_in_progress = false;
 820
 821         return addr;
 822 }
 823