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 (unlikely(data->balance != 1)) {
 229                 kmemcheck_show_all();
 230                 kmemcheck_error_save_bug(regs);
 231                 data->n_addrs = 0;
 232                 data->balance = 0;
 233
 234                 if (!(data->flags & X86_EFLAGS_TF))
 235                         regs->flags &= ~X86_EFLAGS_TF;
 236                 if (data->flags & X86_EFLAGS_IF)
 237                         regs->flags |= X86_EFLAGS_IF;
 238                 return;
 239         }
 240
 241         if (kmemcheck_enabled)
 242                 n = kmemcheck_hide_all();
 243         else
 244                 n = kmemcheck_show_all();
 245
 246         if (n == 0)
 247                 return;
 248
 249         --data->balance;
 250
 251         data->n_addrs = 0;
 252
 253         if (!(data->flags & X86_EFLAGS_TF))
 254                 regs->flags &= ~X86_EFLAGS_TF;
 255         if (data->flags & X86_EFLAGS_IF)
 256                 regs->flags |= X86_EFLAGS_IF;
 257 }
 258
 259 void kmemcheck_show_pages(struct page *p, unsigned int n)
 260 {
 261         unsigned int i;
 262
 263         for (i = 0; i < n; ++i) {
 264                 unsigned long address;
 265                 pte_t *pte;
 266                 unsigned int level;
 267
 268                 address = (unsigned long) page_address(&p[i]);
 269                 pte = lookup_address(address, &level);
 270                 BUG_ON(!pte);
 271                 BUG_ON(level != PG_LEVEL_4K);
 272
 273                 set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
 274                 set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_HIDDEN));
 275                 __flush_tlb_one(address);
 276         }
 277 }
 278
 279 bool kmemcheck_page_is_tracked(struct page *p)
 280 {
 281         /* This will also check the "hidden" flag of the PTE. */
 282         return kmemcheck_pte_lookup((unsigned long) page_address(p));
 283 }
 284
 285 void kmemcheck_hide_pages(struct page *p, unsigned int n)
 286 {
 287         unsigned int i;
 288
 289         for (i = 0; i < n; ++i) {
 290                 unsigned long address;
 291                 pte_t *pte;
 292                 unsigned int level;
 293
 294                 address = (unsigned long) page_address(&p[i]);
 295                 pte = lookup_address(address, &level);
 296                 BUG_ON(!pte);
 297                 BUG_ON(level != PG_LEVEL_4K);
 298
 299                 set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
 300                 set_pte(pte, __pte(pte_val(*pte) | _PAGE_HIDDEN));
 301                 __flush_tlb_one(address);
 302         }
 303 }
 304
 305 /* Access may NOT cross page boundary */
 306 static void kmemcheck_read_strict(struct pt_regs *regs,
 307         unsigned long addr, unsigned int size)
 308 {
 309         void *shadow;
 310         enum kmemcheck_shadow status;
 311
 312         shadow = kmemcheck_shadow_lookup(addr);
 313         if (!shadow)
 314                 return;
 315
 316         kmemcheck_save_addr(addr);
 317         status = kmemcheck_shadow_test(shadow, size);
 318         if (status == KMEMCHECK_SHADOW_INITIALIZED)
 319                 return;
 320
 321         if (kmemcheck_enabled)
 322                 kmemcheck_error_save(status, addr, size, regs);
 323
 324         if (kmemcheck_enabled == 2)
 325                 kmemcheck_enabled = 0;
 326
 327         /* Don't warn about it again. */
 328         kmemcheck_shadow_set(shadow, size);
 329 }
 330
 331 bool kmemcheck_is_obj_initialized(unsigned long addr, size_t size)
 332 {
 333         enum kmemcheck_shadow status;
 334         void *shadow;
 335
 336         shadow = kmemcheck_shadow_lookup(addr);
 337         if (!shadow)
 338                 return true;
 339
 340         status = kmemcheck_shadow_test_all(shadow, size);
 341
 342         return status == KMEMCHECK_SHADOW_INITIALIZED;
 343 }
 344
 345 /* Access may cross page boundary */
 346 static void kmemcheck_read(struct pt_regs *regs,
 347         unsigned long addr, unsigned int size)
 348 {
 349         unsigned long page = addr & PAGE_MASK;
 350         unsigned long next_addr = addr + size - 1;
 351         unsigned long next_page = next_addr & PAGE_MASK;
 352
 353         if (likely(page == next_page)) {
 354                 kmemcheck_read_strict(regs, addr, size);
 355                 return;
 356         }
 357
 358         /*
 359          * What we do is basically to split the access across the
 360          * two pages and handle each part separately. Yes, this means
 361          * that we may now see reads that are 3 + 5 bytes, for
 362          * example (and if both are uninitialized, there will be two
 363          * reports), but it makes the code a lot simpler.
 364          */
 365         kmemcheck_read_strict(regs, addr, next_page - addr);
 366         kmemcheck_read_strict(regs, next_page, next_addr - next_page);
 367 }
 368
 369 static void kmemcheck_write_strict(struct pt_regs *regs,
 370         unsigned long addr, unsigned int size)
 371 {
 372         void *shadow;
 373
 374         shadow = kmemcheck_shadow_lookup(addr);
 375         if (!shadow)
 376                 return;
 377
 378         kmemcheck_save_addr(addr);
 379         kmemcheck_shadow_set(shadow, size);
 380 }
 381
 382 static void kmemcheck_write(struct pt_regs *regs,
 383         unsigned long addr, unsigned int size)
 384 {
 385         unsigned long page = addr & PAGE_MASK;
 386         unsigned long next_addr = addr + size - 1;
 387         unsigned long next_page = next_addr & PAGE_MASK;
 388
 389         if (likely(page == next_page)) {
 390                 kmemcheck_write_strict(regs, addr, size);
 391                 return;
 392         }
 393
 394         /* See comment in kmemcheck_read(). */
 395         kmemcheck_write_strict(regs, addr, next_page - addr);
 396         kmemcheck_write_strict(regs, next_page, next_addr - next_page);
 397 }
 398
 399 /*
 400  * Copying is hard. We have two addresses, each of which may be split across
 401  * a page (and each page will have different shadow addresses).
 402  */
 403 static void kmemcheck_copy(struct pt_regs *regs,
 404         unsigned long src_addr, unsigned long dst_addr, unsigned int size)
 405 {
 406         uint8_t shadow[8];
 407         enum kmemcheck_shadow status;
 408
 409         unsigned long page;
 410         unsigned long next_addr;
 411         unsigned long next_page;
 412
 413         uint8_t *x;
 414         unsigned int i;
 415         unsigned int n;
 416
 417         BUG_ON(size > sizeof(shadow));
 418
 419         page = src_addr & PAGE_MASK;
 420         next_addr = src_addr + size - 1;
 421         next_page = next_addr & PAGE_MASK;
 422
 423         if (likely(page == next_page)) {
 424                 /* Same page */
 425                 x = kmemcheck_shadow_lookup(src_addr);
 426                 if (x) {
 427                         kmemcheck_save_addr(src_addr);
 428                         for (i = 0; i < size; ++i)
 429                                 shadow[i] = x[i];
 430                 } else {
 431                         for (i = 0; i < size; ++i)
 432                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 433                 }
 434         } else {
 435                 n = next_page - src_addr;
 436                 BUG_ON(n > sizeof(shadow));
 437
 438                 /* First page */
 439                 x = kmemcheck_shadow_lookup(src_addr);
 440                 if (x) {
 441                         kmemcheck_save_addr(src_addr);
 442                         for (i = 0; i < n; ++i)
 443                                 shadow[i] = x[i];
 444                 } else {
 445                         /* Not tracked */
 446                         for (i = 0; i < n; ++i)
 447                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 448                 }
 449
 450                 /* Second page */
 451                 x = kmemcheck_shadow_lookup(next_page);
 452                 if (x) {
 453                         kmemcheck_save_addr(next_page);
 454                         for (i = n; i < size; ++i)
 455                                 shadow[i] = x[i - n];
 456                 } else {
 457                         /* Not tracked */
 458                         for (i = n; i < size; ++i)
 459                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 460                 }
 461         }
 462
 463         page = dst_addr & PAGE_MASK;
 464         next_addr = dst_addr + size - 1;
 465         next_page = next_addr & PAGE_MASK;
 466
 467         if (likely(page == next_page)) {
 468                 /* Same page */
 469                 x = kmemcheck_shadow_lookup(dst_addr);
 470                 if (x) {
 471                         kmemcheck_save_addr(dst_addr);
 472                         for (i = 0; i < size; ++i) {
 473                                 x[i] = shadow[i];
 474                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 475                         }
 476                 }
 477         } else {
 478                 n = next_page - dst_addr;
 479                 BUG_ON(n > sizeof(shadow));
 480
 481                 /* First page */
 482                 x = kmemcheck_shadow_lookup(dst_addr);
 483                 if (x) {
 484                         kmemcheck_save_addr(dst_addr);
 485                         for (i = 0; i < n; ++i) {
 486                                 x[i] = shadow[i];
 487                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 488                         }
 489                 }
 490
 491                 /* Second page */
 492                 x = kmemcheck_shadow_lookup(next_page);
 493                 if (x) {
 494                         kmemcheck_save_addr(next_page);
 495                         for (i = n; i < size; ++i) {
 496                                 x[i - n] = shadow[i];
 497                                 shadow[i] = KMEMCHECK_SHADOW_INITIALIZED;
 498                         }
 499                 }
 500         }
 501
 502         status = kmemcheck_shadow_test(shadow, size);
 503         if (status == KMEMCHECK_SHADOW_INITIALIZED)
 504                 return;
 505
 506         if (kmemcheck_enabled)
 507                 kmemcheck_error_save(status, src_addr, size, regs);
 508
 509         if (kmemcheck_enabled == 2)
 510                 kmemcheck_enabled = 0;
 511 }
 512
 513 enum kmemcheck_method {
 514         KMEMCHECK_READ,
 515         KMEMCHECK_WRITE,
 516 };
 517
 518 static void kmemcheck_access(struct pt_regs *regs,
 519         unsigned long fallback_address, enum kmemcheck_method fallback_method)
 520 {
 521         const uint8_t *insn;
 522         const uint8_t *insn_primary;
 523         unsigned int size;
 524
 525         struct kmemcheck_context *data = &__get_cpu_var(kmemcheck_context);
 526
 527         /* Recursive fault -- ouch. */
 528         if (data->busy) {
 529                 kmemcheck_show_addr(fallback_address);
 530                 kmemcheck_error_save_bug(regs);
 531                 return;
 532         }
 533
 534         data->busy = true;
 535
 536         insn = (const uint8_t *) regs->ip;
 537         insn_primary = kmemcheck_opcode_get_primary(insn);
 538
 539         kmemcheck_opcode_decode(insn, &size);
 540
 541         switch (insn_primary[0]) {
 542 #ifdef CONFIG_KMEMCHECK_BITOPS_OK
 543                 /* AND, OR, XOR */
 544                 /*
 545                  * Unfortunately, these instructions have to be excluded from
 546                  * our regular checking since they access only some (and not
 547                  * all) bits. This clears out "bogus" bitfield-access warnings.
 548                  */
 549         case 0x80:
 550         case 0x81:
 551         case 0x82:
 552         case 0x83:
 553                 switch ((insn_primary[1] >> 3) & 7) {
 554                         /* OR */
 555                 case 1:
 556                         /* AND */
 557                 case 4:
 558                         /* XOR */
 559                 case 6:
 560                         kmemcheck_write(regs, fallback_address, size);
 561                         goto out;
 562
 563                         /* ADD */
 564                 case 0:
 565                         /* ADC */
 566                 case 2:
 567                         /* SBB */
 568                 case 3:
 569                         /* SUB */
 570                 case 5:
 571                         /* CMP */
 572                 case 7:
 573                         break;
 574                 }
 575                 break;
 576 #endif
 577
 578                 /* MOVS, MOVSB, MOVSW, MOVSD */
 579         case 0xa4:
 580         case 0xa5:
 581                 /*
 582                  * These instructions are special because they take two
 583                  * addresses, but we only get one page fault.
 584                  */
 585                 kmemcheck_copy(regs, regs->si, regs->di, size);
 586                 goto out;
 587
 588                 /* CMPS, CMPSB, CMPSW, CMPSD */
 589         case 0xa6:
 590         case 0xa7:
 591                 kmemcheck_read(regs, regs->si, size);
 592                 kmemcheck_read(regs, regs->di, size);
 593                 goto out;
 594         }
 595
 596         /*
 597          * If the opcode isn't special in any way, we use the data from the
 598          * page fault handler to determine the address and type of memory
 599          * access.
 600          */
 601         switch (fallback_method) {
 602         case KMEMCHECK_READ:
 603                 kmemcheck_read(regs, fallback_address, size);
 604                 goto out;
 605         case KMEMCHECK_WRITE:
 606                 kmemcheck_write(regs, fallback_address, size);
 607                 goto out;
 608         }
 609
 610 out:
 611         data->busy = false;
 612 }
 613
 614 bool kmemcheck_fault(struct pt_regs *regs, unsigned long address,
 615         unsigned long error_code)
 616 {
 617         pte_t *pte;
 618
 619         /*
 620          * XXX: Is it safe to assume that memory accesses from virtual 86
 621          * mode or non-kernel code segments will _never_ access kernel
 622          * memory (e.g. tracked pages)? For now, we need this to avoid
 623          * invoking kmemcheck for PnP BIOS calls.
 624          */
 625         if (regs->flags & X86_VM_MASK)
 626                 return false;
 627         if (regs->cs != __KERNEL_CS)
 628                 return false;
 629
 630         pte = kmemcheck_pte_lookup(address);
 631         if (!pte)
 632                 return false;
 633
 634         WARN_ON_ONCE(in_nmi());
 635
 636         if (error_code & 2)
 637                 kmemcheck_access(regs, address, KMEMCHECK_WRITE);
 638         else
 639                 kmemcheck_access(regs, address, KMEMCHECK_READ);
 640
 641         kmemcheck_show(regs);
 642         return true;
 643 }
 644
 645 bool kmemcheck_trap(struct pt_regs *regs)
 646 {
 647         if (!kmemcheck_active(regs))
 648                 return false;
 649
 650         /* We're done. */
 651         kmemcheck_hide(regs);
 652         return true;
 653 }