arch/x86/mm/kmemcheck/kmemcheck.c

   1 /**
   2  * kmemcheck - a heavyweight memory checker for the linux kernel
   3  * Copyright (C) 2007, 2008  Vegard Nossum <vegardno@ifi.uio.no>
   4  * (With a lot of help from Ingo Molnar and Pekka Enberg.)
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License (version 2) as
   8  * published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/init.h>
  12 #include <linux/interrupt.h>
  13 #include <linux/kallsyms.h>
  14 #include <linux/kernel.h>
  15 #include <linux/kmemcheck.h>
  16 #include <linux/mm.h>
  17 #include <linux/module.h>
  18 #include <linux/page-flags.h>
  19 #include <linux/percpu.h>
  20 #include <linux/ptrace.h>
  21 #include <linux/string.h>
  22 #include <linux/types.h>
  23
  24 #include <asm/cacheflush.h>
  25 #include <asm/kmemcheck.h>
  26 #include <asm/pgtable.h>
  27 #include <asm/tlbflush.h>
  28
  29 #include "error.h"
  30 #include "opcode.h"
  31 #include "pte.h"
  32 #include "selftest.h"
  33 #include "shadow.h"
  34
  35
  36 #ifdef CONFIG_KMEMCHECK_DISABLED_BY_DEFAULT
  37 #  define KMEMCHECK_ENABLED 0
  38 #endif
  39
  40 #ifdef CONFIG_KMEMCHECK_ENABLED_BY_DEFAULT
  41 #  define KMEMCHECK_ENABLED 1
  42 #endif
  43
  44 #ifdef CONFIG_KMEMCHECK_ONESHOT_BY_DEFAULT
  45 #  define KMEMCHECK_ENABLED 2
  46 #endif
  47
  48 int kmemcheck_enabled = KMEMCHECK_ENABLED;
  49
  50 int __init kmemcheck_init(void)
  51 {
  52 #ifdef CONFIG_SMP
  53         /*
  54          * Limit SMP to use a single CPU. We rely on the fact that this code
  55          * runs before SMP is set up.
  56          */
  57         if (setup_max_cpus > 1) {
  58                 printk(KERN_INFO
  59                         "kmemcheck: Limiting number of CPUs to 1.\n");
  60                 setup_max_cpus = 1;
  61         }
  62 #endif
  63
  64         if (!kmemcheck_selftest()) {
  65                 printk(KERN_INFO "kmemcheck: self-tests failed; disabling\n");
  66                 kmemcheck_enabled = 0;
  67                 return -EINVAL;
  68         }
  69
  70         printk(KERN_INFO "kmemcheck: Initialized\n");
  71         return 0;
  72 }
  73
  74 early_initcall(kmemcheck_init);
  75
  76 /*
  77  * We need to parse the kmemcheck= option before any memory is allocated.
  78  */
  79 static int __init param_kmemcheck(char *str)
  80 {
  81         if (!str)
  82                 return -EINVAL;
  83
  84         sscanf(str, "%d", &kmemcheck_enabled);
  85         return 0;
  86 }
  87
  88 early_param("kmemcheck", param_kmemcheck);
  89
  90 int kmemcheck_show_addr(unsigned long address)
  91 {
  92         pte_t *pte;
  93
  94         pte = kmemcheck_pte_lookup(address);
  95         if (!pte)
  96                 return 0;
  97
  98         set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
  99         __flush_tlb_one(address);
 100         return 1;
 101 }
 102
 103 int kmemcheck_hide_addr(unsigned long address)
 104 {
 105         pte_t *pte;
 106
 107         pte = kmemcheck_pte_lookup(address);
 108         if (!pte)
 109                 return 0;
 110
 111         set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
 112         __flush_tlb_one(address);
 113         return 1;
 114 }
 115
 116 struct kmemcheck_context {
 117         bool busy;
 118         int balance;
 119
 120         /*
 121          * There can be at most two memory operands to an instruction, but
 122          * each address can cross a page boundary -- so we may need up to
 123          * four addresses that must be hidden/revealed for each fault.
 124          */
 125         unsigned long addr[4];
 126         unsigned long n_addrs;
 127         unsigned long flags;
 128
 129         /* Data size of the instruction that caused a fault. */
 130         unsigned int size;
 131 };
 132
 133 static DEFINE_PER_CPU(struct kmemcheck_context, kmemcheck_context);
 134
 135 bool kmemcheck_active(struct pt_regs *regs)
 136 {
 137         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 138
 139         return data->balance > 0;
 140 }
 141
 142 /* Save an address that needs to be shown/hidden */
 143 static void kmemcheck_save_addr(unsigned long addr)
 144 {
 145         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 146
 147         BUG_ON(data->n_addrs >= ARRAY_SIZE(data->addr));
 148         data->addr[data->n_addrs++] = addr;
 149 }
 150
 151 static unsigned int kmemcheck_show_all(void)
 152 {
 153         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 154         unsigned int i;
 155         unsigned int n;
 156
 157         n = 0;
 158         for (i = 0; i < data->n_addrs; ++i)
 159                 n += kmemcheck_show_addr(data->addr[i]);
 160
 161         return n;
 162 }
 163
 164 static unsigned int kmemcheck_hide_all(void)
 165 {
 166         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 167         unsigned int i;
 168         unsigned int n;
 169
 170         n = 0;
 171         for (i = 0; i < data->n_addrs; ++i)
 172                 n += kmemcheck_hide_addr(data->addr[i]);
 173
 174         return n;
 175 }
 176
 177 /*
 178  * Called from the #PF handler.
 179  */
 180 void kmemcheck_show(struct pt_regs *regs)
 181 {
 182         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 183
 184         BUG_ON(!irqs_disabled());
 185
 186         if (unlikely(data->balance != 0)) {
 187                 kmemcheck_show_all();
 188                 kmemcheck_error_save_bug(regs);
 189                 data->balance = 0;
 190                 return;
 191         }
 192
 193         /*
 194          * None of the addresses actually belonged to kmemcheck. Note that
 195          * this is not an error.
 196          */
 197         if (kmemcheck_show_all() == 0)
 198                 return;
 199
 200         ++data->balance;
 201
 202         /*
 203          * The IF needs to be cleared as well, so that the faulting
 204          * instruction can run "uninterrupted". Otherwise, we might take
 205          * an interrupt and start executing that before we've had a chance
 206          * to hide the page again.
 207          *
 208          * NOTE: In the rare case of multiple faults, we must not override
 209          * the original flags:
 210          */
 211         if (!(regs->flags & X86_EFLAGS_TF))
 212                 data->flags = regs->flags;
 213
 214         regs->flags |= X86_EFLAGS_TF;
 215         regs->flags &= ~X86_EFLAGS_IF;
 216 }
 217
 218 /*
 219  * Called from the #DB handler.
 220  */
 221 void kmemcheck_hide(struct pt_regs *regs)
 222 {
 223         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 224         int n;
 225
 226         BUG_ON(!irqs_disabled());
 227
 228         if (data->balance == 0)
 229                 return;
 230
 231         if (unlikely(data->balance != 1)) {
 232                 kmemcheck_show_all();
 233                 kmemcheck_error_save_bug(regs);
 234                 data->n_addrs = 0;
 235                 data->balance = 0;
 236
 237                 if (!(data->flags & X86_EFLAGS_TF))
 238                         regs->flags &= ~X86_EFLAGS_TF;
 239                 if (data->flags & X86_EFLAGS_IF)
 240                         regs->flags |= X86_EFLAGS_IF;
 241                 return;
 242         }
 243
 244         if (kmemcheck_enabled)
 245                 n = kmemcheck_hide_all();
 246         else
 247                 n = kmemcheck_show_all();
 248
 249         if (n == 0)
 250                 return;
 251
 252         --data->balance;
 253
 254         data->n_addrs = 0;
 255
 256         if (!(data->flags & X86_EFLAGS_TF))
 257                 regs->flags &= ~X86_EFLAGS_TF;
 258         if (data->flags & X86_EFLAGS_IF)
 259                 regs->flags |= X86_EFLAGS_IF;
 260 }
 261
 262 void kmemcheck_show_pages(struct page *p, unsigned int n)
 263 {
 264         unsigned int i;
 265
 266         for (i = 0; i < n; ++i) {
 267                 unsigned long address;
 268                 pte_t *pte;
 269                 unsigned int level;
 270
 271                 address = (unsigned long) page_address(&p[i]);
 272                 pte = lookup_address(address, &level);
 273                 BUG_ON(!pte);
 274                 BUG_ON(level != PG_LEVEL_4K);
 275
 276                 set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
 277                 set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_HIDDEN));
 278                 __flush_tlb_one(address);
 279         }
 280 }
 281
 282 bool kmemcheck_page_is_tracked(struct page *p)
 283 {
 284         /* This will also check the "hidden" flag of the PTE. */
 285         return kmemcheck_pte_lookup((unsigned long) page_address(p));
 286 }
 287
 288 void kmemcheck_hide_pages(struct page *p, unsigned int n)
 289 {
 290         unsigned int i;
 291
 292         for (i = 0; i < n; ++i) {
 293                 unsigned long address;
 294                 pte_t *pte;
 295                 unsigned int level;
 296
 297                 address = (unsigned long) page_address(&p[i]);
 298                 pte = lookup_address(address, &level);
 299                 BUG_ON(!pte);
 300                 BUG_ON(level != PG_LEVEL_4K);
 301
 302                 set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
 303                 set_pte(pte, __pte(pte_val(*pte) | _PAGE_HIDDEN));
 304                 __flush_tlb_one(address);
 305         }
 306 }
 307
 308 /* Access may NOT cross page boundary */
 309 static void kmemcheck_read_strict(struct pt_regs *regs,
 310         unsigned long addr, unsigned int size)
 311 {
 312         void *shadow;
 313         enum kmemcheck_shadow status;
 314
 315         shadow = kmemcheck_shadow_lookup(addr);
 316         if (!shadow)
 317                 return;
 318
 319         kmemcheck_save_addr(addr);
 320         status = kmemcheck_shadow_test(shadow, size);
 321         if (status == KMEMCHECK_SHADOW_INITIALIZED)
 322                 return;
 323
 324         if (kmemcheck_enabled)
 325                 kmemcheck_error_save(status, addr, size, regs);
 326
 327         if (kmemcheck_enabled == 2)
 328                 kmemcheck_enabled = 0;
 329
 330         /* Don't warn about it again. */
 331         kmemcheck_shadow_set(shadow, size);
 332 }
 333
 334 /* Access may cross page boundary */
 335 static void kmemcheck_read(struct pt_regs *regs,
 336         unsigned long addr, unsigned int size)
 337 {
 338         unsigned long page = addr & PAGE_MASK;
 339         unsigned long next_addr = addr + size - 1;
 340         unsigned long next_page = next_addr & PAGE_MASK;
 341
 342         if (likely(page == next_page)) {
 343                 kmemcheck_read_strict(regs, addr, size);
 344                 return;
 345         }
 346
 347         /*
 348          * What we do is basically to split the access across the
 349          * two pages and handle each part separately. Yes, this means
 350          * that we may now see reads that are 3 + 5 bytes, for
 351          * example (and if both are uninitialized, there will be two
 352          * reports), but it makes the code a lot simpler.
 353          */
 354         kmemcheck_read_strict(regs, addr, next_page - addr);
 355         kmemcheck_read_strict(regs, next_page, next_addr - next_page);
 356 }
 357
 358 static void kmemcheck_write_strict(struct pt_regs *regs,
 359         unsigned long addr, unsigned int size)
 360 {
 361         void *shadow;
 362
 363         shadow = kmemcheck_shadow_lookup(addr);
 364         if (!shadow)
 365                 return;
 366
 367         kmemcheck_save_addr(addr);
 368         kmemcheck_shadow_set(shadow, size);
 369 }
 370
 371 static void kmemcheck_write(struct pt_regs *regs,
 372         unsigned long addr, unsigned int size)
 373 {
 374         unsigned long page = addr & PAGE_MASK;
 375         unsigned long next_addr = addr + size - 1;
 376         unsigned long next_page = next_addr & PAGE_MASK;
 377
 378         if (likely(page == next_page)) {
 379                 kmemcheck_write_strict(regs, addr, size);
 380                 return;
 381         }
 382
 383         /* See comment in kmemcheck_read(). */
 384         kmemcheck_write_strict(regs, addr, next_page - addr);
 385         kmemcheck_write_strict(regs, next_page, next_addr - next_page);
 386 }
 387
 388 /*
 389  * Copying is hard. We have two addresses, each of which may be split across
 390  * a page (and each page will have different shadow addresses).
 391  */
 392 static void kmemcheck_copy(struct pt_regs *regs,
 393         unsigned long src_addr, unsigned long dst_addr, unsigned int size)
 394 {
 395         uint8_t shadow[8];
 396         enum kmemcheck_shadow status;
 397
 398         unsigned long page;
 399         unsigned long next_addr;
 400         unsigned long next_page;
 401
 402         uint8_t *x;
 403         unsigned int i;
 404         unsigned int n;
 405
 406         BUG_ON(size > sizeof(shadow));
 407
 408         page = src_addr & PAGE_MASK;
 409         next_addr = src_addr + size - 1;
 410         next_page = next_addr & PAGE_MASK;
 411
 412         if (likely(page == next_page)) {
 413                 /* Same page */
 414                 x = kmemcheck_shadow_lookup(src_addr);
 415                 if (x) {
 416                         kmemcheck_save_addr(src_addr);
 417                         for (i = 0; i < size; ++i)
 418                                 shadow[i] = x[i];
 419                 } else {
 420                         for (i = 0; i < size; ++i)
 421                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 422                 }
 423         } else {
 424                 n = next_page - src_addr;
 425                 BUG_ON(n > sizeof(shadow));
 426
 427                 /* First page */
 428                 x = kmemcheck_shadow_lookup(src_addr);
 429                 if (x) {
 430                         kmemcheck_save_addr(src_addr);
 431                         for (i = 0; i < n; ++i)
 432                                 shadow[i] = x[i];
 433                 } else {
 434                         /* Not tracked */
 435                         for (i = 0; i < n; ++i)
 436                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 437                 }
 438
 439                 /* Second page */
 440                 x = kmemcheck_shadow_lookup(next_page);
 441                 if (x) {
 442                         kmemcheck_save_addr(next_page);
 443                         for (i = n; i < size; ++i)
 444                                 shadow[i] = x[i - n];
 445                 } else {
 446                         /* Not tracked */
 447                         for (i = n; i < size; ++i)
 448                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 449                 }
 450         }
 451
 452         page = dst_addr & PAGE_MASK;
 453         next_addr = dst_addr + size - 1;
 454         next_page = next_addr & PAGE_MASK;
 455
 456         if (likely(page == next_page)) {
 457                 /* Same page */
 458                 x = kmemcheck_shadow_lookup(dst_addr);
 459                 if (x) {
 460                         kmemcheck_save_addr(dst_addr);
 461                         for (i = 0; i < size; ++i) {
 462                                 x[i] = shadow[i];
 463                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 464                         }
 465                 }
 466         } else {
 467                 n = next_page - dst_addr;
 468                 BUG_ON(n > sizeof(shadow));
 469
 470                 /* First page */
 471                 x = kmemcheck_shadow_lookup(dst_addr);
 472                 if (x) {
 473                         kmemcheck_save_addr(dst_addr);
 474                         for (i = 0; i < n; ++i) {
 475                                 x[i] = shadow[i];
 476                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 477                         }
 478                 }
 479
 480                 /* Second page */
 481                 x = kmemcheck_shadow_lookup(next_page);
 482                 if (x) {
 483                         kmemcheck_save_addr(next_page);
 484                         for (i = n; i < size; ++i) {
 485                                 x[i - n] = shadow[i];
 486                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 487                         }
 488                 }
 489         }
 490
 491         status = kmemcheck_shadow_test(shadow, size);
 492         if (status == KMEMCHECK_SHADOW_INITIALIZED)
 493                 return;
 494
 495         if (kmemcheck_enabled)
 496                 kmemcheck_error_save(status, src_addr, size, regs);
 497
 498         if (kmemcheck_enabled == 2)
 499                 kmemcheck_enabled = 0;
 500 }
 501
 502 enum kmemcheck_method {
 503         KMEMCHECK_READ,
 504         KMEMCHECK_WRITE,
 505 };
 506
 507 static void kmemcheck_access(struct pt_regs *regs,
 508         unsigned long fallback_address, enum kmemcheck_method fallback_method)
 509 {
 510         const uint8_t *insn;
 511         const uint8_t *insn_primary;
 512         unsigned int size;
 513
 514         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 515
 516         /* Recursive fault -- ouch. */
 517         if (data->busy) {
 518                 kmemcheck_show_addr(fallback_address);
 519                 kmemcheck_error_save_bug(regs);
 520                 return;
 521         }
 522
 523         data->busy = true;
 524
 525         insn = (const uint8_t *) regs->ip;
 526         insn_primary = kmemcheck_opcode_get_primary(insn);
 527
 528         kmemcheck_opcode_decode(insn, &size);
 529
 530         switch (insn_primary[0]) {
 531 #ifdef CONFIG_KMEMCHECK_BITOPS_OK
 532                 /* AND, OR, XOR */
 533                 /*
 534                  * Unfortunately, these instructions have to be excluded from
 535                  * our regular checking since they access only some (and not
 536                  * all) bits. This clears out "bogus" bitfield-access warnings.
 537                  */
 538         case 0x80:
 539         case 0x81:
 540         case 0x82:
 541         case 0x83:
 542                 switch ((insn_primary[1] >> 3) & 7) {
 543                         /* OR */
 544                 case 1:
 545                         /* AND */
 546                 case 4:
 547                         /* XOR */
 548                 case 6:
 549                         kmemcheck_write(regs, fallback_address, size);
 550                         goto out;
 551
 552                         /* ADD */
 553                 case 0:
 554                         /* ADC */
 555                 case 2:
 556                         /* SBB */
 557                 case 3:
 558                         /* SUB */
 559                 case 5:
 560                         /* CMP */
 561                 case 7:
 562                         break;
 563                 }
 564                 break;
 565 #endif
 566
 567                 /* MOVS, MOVSB, MOVSW, MOVSD */
 568         case 0xa4:
 569         case 0xa5:
 570                 /*
 571                  * These instructions are special because they take two
 572                  * addresses, but we only get one page fault.
 573                  */
 574                 kmemcheck_copy(regs, regs->si, regs->di, size);
 575                 goto out;
 576
 577                 /* CMPS, CMPSB, CMPSW, CMPSD */
 578         case 0xa6:
 579         case 0xa7:
 580                 kmemcheck_read(regs, regs->si, size);
 581                 kmemcheck_read(regs, regs->di, size);
 582                 goto out;
 583         }
 584
 585         /*
 586          * If the opcode isn't special in any way, we use the data from the
 587          * page fault handler to determine the address and type of memory
 588          * access.
 589          */
 590         switch (fallback_method) {
 591         case KMEMCHECK_READ:
 592                 kmemcheck_read(regs, fallback_address, size);
 593                 goto out;
 594         case KMEMCHECK_WRITE:
 595                 kmemcheck_write(regs, fallback_address, size);
 596                 goto out;
 597         }
 598
 599 out:
 600         data->busy = false;
 601 }
 602
 603 bool kmemcheck_fault(struct pt_regs *regs, unsigned long address,
 604         unsigned long error_code)
 605 {
 606         pte_t *pte;
 607
 608         /*
 609          * XXX: Is it safe to assume that memory accesses from virtual 86
 610          * mode or non-kernel code segments will _never_ access kernel
 611          * memory (e.g. tracked pages)? For now, we need this to avoid
 612          * invoking kmemcheck for PnP BIOS calls.
 613          */
 614         if (regs->flags & X86_VM_MASK)
 615                 return false;
 616         if (regs->cs != __KERNEL_CS)
 617                 return false;
 618
 619         pte = kmemcheck_pte_lookup(address);
 620         if (!pte)
 621                 return false;
 622
 623         if (error_code & 2)
 624                 kmemcheck_access(regs, address, KMEMCHECK_WRITE);
 625         else
 626                 kmemcheck_access(regs, address, KMEMCHECK_READ);
 627
 628         kmemcheck_show(regs);
 629         return true;
 630 }
 631
 632 bool kmemcheck_trap(struct pt_regs *regs)
 633 {
 634         if (!kmemcheck_active(regs))
 635                 return false;
 636
 637         /* We're done. */
 638         kmemcheck_hide(regs);
 639         return true;
 640 }