arch/x86/mm/fault.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
   4  *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
   5  */
   6 #include <linux/magic.h>                /* STACK_END_MAGIC              */
   7 #include <linux/sched.h>                /* test_thread_flag(), ...      */
   8 #include <linux/kdebug.h>               /* oops_begin/end, ...          */
   9 #include <linux/module.h>               /* search_exception_table       */
  10 #include <linux/bootmem.h>              /* max_low_pfn                  */
  11 #include <linux/kprobes.h>              /* __kprobes, ...               */
  12 #include <linux/mmiotrace.h>            /* kmmio_handler, ...           */
  13 #include <linux/perf_counter.h>         /* perf_swcounter_event         */
  14
  15 #include <asm/traps.h>                  /* dotraplinkage, ...           */
  16 #include <asm/pgalloc.h>                /* pgd_*(), ...                 */
  17 #include <asm/kmemcheck.h>              /* kmemcheck_*(), ...           */
  18
  19 /*
  20  * Page fault error code bits:
  21  *
  22  *   bit 0 ==    0: no page found       1: protection fault
  23  *   bit 1 ==    0: read access         1: write access
  24  *   bit 2 ==    0: kernel-mode access  1: user-mode access
  25  *   bit 3 ==                           1: use of reserved bit detected
  26  *   bit 4 ==                           1: fault was an instruction fetch
  27  */
  28 enum x86_pf_error_code {
  29
  30         PF_PROT         =               1 << 0,
  31         PF_WRITE        =               1 << 1,
  32         PF_USER         =               1 << 2,
  33         PF_RSVD         =               1 << 3,
  34         PF_INSTR        =               1 << 4,
  35 };
  36
  37 /*
  38  * Returns 0 if mmiotrace is disabled, or if the fault is not
  39  * handled by mmiotrace:
  40  */
  41 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  42 {
  43         if (unlikely(is_kmmio_active()))
  44                 if (kmmio_handler(regs, addr) == 1)
  45                         return -1;
  46         return 0;
  47 }
  48
  49 static inline int notify_page_fault(struct pt_regs *regs)
  50 {
  51         int ret = 0;
  52
  53         /* kprobe_running() needs smp_processor_id() */
  54         if (kprobes_built_in() && !user_mode_vm(regs)) {
  55                 preempt_disable();
  56                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  57                         ret = 1;
  58                 preempt_enable();
  59         }
  60
  61         return ret;
  62 }
  63
  64 /*
  65  * Prefetch quirks:
  66  *
  67  * 32-bit mode:
  68  *
  69  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  70  *   Check that here and ignore it.
  71  *
  72  * 64-bit mode:
  73  *
  74  *   Sometimes the CPU reports invalid exceptions on prefetch.
  75  *   Check that here and ignore it.
  76  *
  77  * Opcode checker based on code by Richard Brunner.
  78  */
  79 static inline int
  80 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  81                       unsigned char opcode, int *prefetch)
  82 {
  83         unsigned char instr_hi = opcode & 0xf0;
  84         unsigned char instr_lo = opcode & 0x0f;
  85
  86         switch (instr_hi) {
  87         case 0x20:
  88         case 0x30:
  89                 /*
  90                  * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  91                  * In X86_64 long mode, the CPU will signal invalid
  92                  * opcode if some of these prefixes are present so
  93                  * X86_64 will never get here anyway
  94                  */
  95                 return ((instr_lo & 7) == 0x6);
  96 #ifdef CONFIG_X86_64
  97         case 0x40:
  98                 /*
  99                  * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 100                  * Need to figure out under what instruction mode the
 101                  * instruction was issued. Could check the LDT for lm,
 102                  * but for now it's good enough to assume that long
 103                  * mode only uses well known segments or kernel.
 104                  */
 105                 return (!user_mode(regs)) || (regs->cs == __USER_CS);
 106 #endif
 107         case 0x60:
 108                 /* 0x64 thru 0x67 are valid prefixes in all modes. */
 109                 return (instr_lo & 0xC) == 0x4;
 110         case 0xF0:
 111                 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 112                 return !instr_lo || (instr_lo>>1) == 1;
 113         case 0x00:
 114                 /* Prefetch instruction is 0x0F0D or 0x0F18 */
 115                 if (probe_kernel_address(instr, opcode))
 116                         return 0;
 117
 118                 *prefetch = (instr_lo == 0xF) &&
 119                         (opcode == 0x0D || opcode == 0x18);
 120                 return 0;
 121         default:
 122                 return 0;
 123         }
 124 }
 125
 126 static int
 127 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 128 {
 129         unsigned char *max_instr;
 130         unsigned char *instr;
 131         int prefetch = 0;
 132
 133         /*
 134          * If it was a exec (instruction fetch) fault on NX page, then
 135          * do not ignore the fault:
 136          */
 137         if (error_code & PF_INSTR)
 138                 return 0;
 139
 140         instr = (void *)convert_ip_to_linear(current, regs);
 141         max_instr = instr + 15;
 142
 143         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 144                 return 0;
 145
 146         while (instr < max_instr) {
 147                 unsigned char opcode;
 148
 149                 if (probe_kernel_address(instr, opcode))
 150                         break;
 151
 152                 instr++;
 153
 154                 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 155                         break;
 156         }
 157         return prefetch;
 158 }
 159
 160 static void
 161 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 162                      struct task_struct *tsk)
 163 {
 164         siginfo_t info;
 165
 166         info.si_signo   = si_signo;
 167         info.si_errno   = 0;
 168         info.si_code    = si_code;
 169         info.si_addr    = (void __user *)address;
 170
 171         force_sig_info(si_signo, &info, tsk);
 172 }
 173
 174 DEFINE_SPINLOCK(pgd_lock);
 175 LIST_HEAD(pgd_list);
 176
 177 #ifdef CONFIG_X86_32
 178 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 179 {
 180         unsigned index = pgd_index(address);
 181         pgd_t *pgd_k;
 182         pud_t *pud, *pud_k;
 183         pmd_t *pmd, *pmd_k;
 184
 185         pgd += index;
 186         pgd_k = init_mm.pgd + index;
 187
 188         if (!pgd_present(*pgd_k))
 189                 return NULL;
 190
 191         /*
 192          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 193          * and redundant with the set_pmd() on non-PAE. As would
 194          * set_pud.
 195          */
 196         pud = pud_offset(pgd, address);
 197         pud_k = pud_offset(pgd_k, address);
 198         if (!pud_present(*pud_k))
 199                 return NULL;
 200
 201         pmd = pmd_offset(pud, address);
 202         pmd_k = pmd_offset(pud_k, address);
 203         if (!pmd_present(*pmd_k))
 204                 return NULL;
 205
 206         if (!pmd_present(*pmd))
 207                 set_pmd(pmd, *pmd_k);
 208         else
 209                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 210
 211         return pmd_k;
 212 }
 213
 214 void vmalloc_sync_all(void)
 215 {
 216         unsigned long address;
 217
 218         if (SHARED_KERNEL_PMD)
 219                 return;
 220
 221         for (address = VMALLOC_START & PMD_MASK;
 222              address >= TASK_SIZE && address < FIXADDR_TOP;
 223              address += PMD_SIZE) {
 224
 225                 unsigned long flags;
 226                 struct page *page;
 227
 228                 spin_lock_irqsave(&pgd_lock, flags);
 229                 list_for_each_entry(page, &pgd_list, lru) {
 230                         if (!vmalloc_sync_one(page_address(page), address))
 231                                 break;
 232                 }
 233                 spin_unlock_irqrestore(&pgd_lock, flags);
 234         }
 235 }
 236
 237 /*
 238  * 32-bit:
 239  *
 240  *   Handle a fault on the vmalloc or module mapping area
 241  */
 242 static noinline int vmalloc_fault(unsigned long address)
 243 {
 244         unsigned long pgd_paddr;
 245         pmd_t *pmd_k;
 246         pte_t *pte_k;
 247
 248         /* Make sure we are in vmalloc area: */
 249         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 250                 return -1;
 251
 252         /*
 253          * Synchronize this task's top level page-table
 254          * with the 'reference' page table.
 255          *
 256          * Do _not_ use "current" here. We might be inside
 257          * an interrupt in the middle of a task switch..
 258          */
 259         pgd_paddr = read_cr3();
 260         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 261         if (!pmd_k)
 262                 return -1;
 263
 264         pte_k = pte_offset_kernel(pmd_k, address);
 265         if (!pte_present(*pte_k))
 266                 return -1;
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * Did it hit the DOS screen memory VA from vm86 mode?
 273  */
 274 static inline void
 275 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 276                  struct task_struct *tsk)
 277 {
 278         unsigned long bit;
 279
 280         if (!v8086_mode(regs))
 281                 return;
 282
 283         bit = (address - 0xA0000) >> PAGE_SHIFT;
 284         if (bit < 32)
 285                 tsk->thread.screen_bitmap |= 1 << bit;
 286 }
 287
 288 static bool low_pfn(unsigned long pfn)
 289 {
 290         return pfn < max_low_pfn;
 291 }
 292
 293 static void dump_pagetable(unsigned long address)
 294 {
 295         pgd_t *base = __va(read_cr3());
 296         pgd_t *pgd = &base[pgd_index(address)];
 297         pmd_t *pmd;
 298         pte_t *pte;
 299
 300 #ifdef CONFIG_X86_PAE
 301         printk("*pdpt = %016Lx ", pgd_val(*pgd));
 302         if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 303                 goto out;
 304 #endif
 305         pmd = pmd_offset(pud_offset(pgd, address), address);
 306         printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 307
 308         /*
 309          * We must not directly access the pte in the highpte
 310          * case if the page table is located in highmem.
 311          * And let's rather not kmap-atomic the pte, just in case
 312          * it's allocated already:
 313          */
 314         if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 315                 goto out;
 316
 317         pte = pte_offset_kernel(pmd, address);
 318         printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 319 out:
 320         printk("\n");
 321 }
 322
 323 #else /* CONFIG_X86_64: */
 324
 325 void vmalloc_sync_all(void)
 326 {
 327         unsigned long address;
 328
 329         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 330              address += PGDIR_SIZE) {
 331
 332                 const pgd_t *pgd_ref = pgd_offset_k(address);
 333                 unsigned long flags;
 334                 struct page *page;
 335
 336                 if (pgd_none(*pgd_ref))
 337                         continue;
 338
 339                 spin_lock_irqsave(&pgd_lock, flags);
 340                 list_for_each_entry(page, &pgd_list, lru) {
 341                         pgd_t *pgd;
 342                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 343                         if (pgd_none(*pgd))
 344                                 set_pgd(pgd, *pgd_ref);
 345                         else
 346                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 347                 }
 348                 spin_unlock_irqrestore(&pgd_lock, flags);
 349         }
 350 }
 351
 352 /*
 353  * 64-bit:
 354  *
 355  *   Handle a fault on the vmalloc area
 356  *
 357  * This assumes no large pages in there.
 358  */
 359 static noinline int vmalloc_fault(unsigned long address)
 360 {
 361         pgd_t *pgd, *pgd_ref;
 362         pud_t *pud, *pud_ref;
 363         pmd_t *pmd, *pmd_ref;
 364         pte_t *pte, *pte_ref;
 365
 366         /* Make sure we are in vmalloc area: */
 367         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 368                 return -1;
 369
 370         /*
 371          * Copy kernel mappings over when needed. This can also
 372          * happen within a race in page table update. In the later
 373          * case just flush:
 374          */
 375         pgd = pgd_offset(current->active_mm, address);
 376         pgd_ref = pgd_offset_k(address);
 377         if (pgd_none(*pgd_ref))
 378                 return -1;
 379
 380         if (pgd_none(*pgd))
 381                 set_pgd(pgd, *pgd_ref);
 382         else
 383                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 384
 385         /*
 386          * Below here mismatches are bugs because these lower tables
 387          * are shared:
 388          */
 389
 390         pud = pud_offset(pgd, address);
 391         pud_ref = pud_offset(pgd_ref, address);
 392         if (pud_none(*pud_ref))
 393                 return -1;
 394
 395         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 396                 BUG();
 397
 398         pmd = pmd_offset(pud, address);
 399         pmd_ref = pmd_offset(pud_ref, address);
 400         if (pmd_none(*pmd_ref))
 401                 return -1;
 402
 403         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 404                 BUG();
 405
 406         pte_ref = pte_offset_kernel(pmd_ref, address);
 407         if (!pte_present(*pte_ref))
 408                 return -1;
 409
 410         pte = pte_offset_kernel(pmd, address);
 411
 412         /*
 413          * Don't use pte_page here, because the mappings can point
 414          * outside mem_map, and the NUMA hash lookup cannot handle
 415          * that:
 416          */
 417         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 418                 BUG();
 419
 420         return 0;
 421 }
 422
 423 static const char errata93_warning[] =
 424 KERN_ERR
 425 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 426 "******* Working around it, but it may cause SEGVs or burn power.\n"
 427 "******* Please consider a BIOS update.\n"
 428 "******* Disabling USB legacy in the BIOS may also help.\n";
 429
 430 /*
 431  * No vm86 mode in 64-bit mode:
 432  */
 433 static inline void
 434 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 435                  struct task_struct *tsk)
 436 {
 437 }
 438
 439 static int bad_address(void *p)
 440 {
 441         unsigned long dummy;
 442
 443         return probe_kernel_address((unsigned long *)p, dummy);
 444 }
 445
 446 static void dump_pagetable(unsigned long address)
 447 {
 448         pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 449         pgd_t *pgd = base + pgd_index(address);
 450         pud_t *pud;
 451         pmd_t *pmd;
 452         pte_t *pte;
 453
 454         if (bad_address(pgd))
 455                 goto bad;
 456
 457         printk("PGD %lx ", pgd_val(*pgd));
 458
 459         if (!pgd_present(*pgd))
 460                 goto out;
 461
 462         pud = pud_offset(pgd, address);
 463         if (bad_address(pud))
 464                 goto bad;
 465
 466         printk("PUD %lx ", pud_val(*pud));
 467         if (!pud_present(*pud) || pud_large(*pud))
 468                 goto out;
 469
 470         pmd = pmd_offset(pud, address);
 471         if (bad_address(pmd))
 472                 goto bad;
 473
 474         printk("PMD %lx ", pmd_val(*pmd));
 475         if (!pmd_present(*pmd) || pmd_large(*pmd))
 476                 goto out;
 477
 478         pte = pte_offset_kernel(pmd, address);
 479         if (bad_address(pte))
 480                 goto bad;
 481
 482         printk("PTE %lx", pte_val(*pte));
 483 out:
 484         printk("\n");
 485         return;
 486 bad:
 487         printk("BAD\n");
 488 }
 489
 490 #endif /* CONFIG_X86_64 */
 491
 492 /*
 493  * Workaround for K8 erratum #93 & buggy BIOS.
 494  *
 495  * BIOS SMM functions are required to use a specific workaround
 496  * to avoid corruption of the 64bit RIP register on C stepping K8.
 497  *
 498  * A lot of BIOS that didn't get tested properly miss this.
 499  *
 500  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 501  * Try to work around it here.
 502  *
 503  * Note we only handle faults in kernel here.
 504  * Does nothing on 32-bit.
 505  */
 506 static int is_errata93(struct pt_regs *regs, unsigned long address)
 507 {
 508 #ifdef CONFIG_X86_64
 509         if (address != regs->ip)
 510                 return 0;
 511
 512         if ((address >> 32) != 0)
 513                 return 0;
 514
 515         address |= 0xffffffffUL << 32;
 516         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 517             (address >= MODULES_VADDR && address <= MODULES_END)) {
 518                 printk_once(errata93_warning);
 519                 regs->ip = address;
 520                 return 1;
 521         }
 522 #endif
 523         return 0;
 524 }
 525
 526 /*
 527  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 528  * to illegal addresses >4GB.
 529  *
 530  * We catch this in the page fault handler because these addresses
 531  * are not reachable. Just detect this case and return.  Any code
 532  * segment in LDT is compatibility mode.
 533  */
 534 static int is_errata100(struct pt_regs *regs, unsigned long address)
 535 {
 536 #ifdef CONFIG_X86_64
 537         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 538                 return 1;
 539 #endif
 540         return 0;
 541 }
 542
 543 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 544 {
 545 #ifdef CONFIG_X86_F00F_BUG
 546         unsigned long nr;
 547
 548         /*
 549          * Pentium F0 0F C7 C8 bug workaround:
 550          */
 551         if (boot_cpu_data.f00f_bug) {
 552                 nr = (address - idt_descr.address) >> 3;
 553
 554                 if (nr == 6) {
 555                         do_invalid_op(regs, 0);
 556                         return 1;
 557                 }
 558         }
 559 #endif
 560         return 0;
 561 }
 562
 563 static const char nx_warning[] = KERN_CRIT
 564 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 565
 566 static void
 567 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 568                 unsigned long address)
 569 {
 570         if (!oops_may_print())
 571                 return;
 572
 573         if (error_code & PF_INSTR) {
 574                 unsigned int level;
 575
 576                 pte_t *pte = lookup_address(address, &level);
 577
 578                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 579                         printk(nx_warning, current_uid());
 580         }
 581
 582         printk(KERN_ALERT "BUG: unable to handle kernel ");
 583         if (address < PAGE_SIZE)
 584                 printk(KERN_CONT "NULL pointer dereference");
 585         else
 586                 printk(KERN_CONT "paging request");
 587
 588         printk(KERN_CONT " at %p\n", (void *) address);
 589         printk(KERN_ALERT "IP:");
 590         printk_address(regs->ip, 1);
 591
 592         dump_pagetable(address);
 593 }
 594
 595 static noinline void
 596 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 597             unsigned long address)
 598 {
 599         struct task_struct *tsk;
 600         unsigned long flags;
 601         int sig;
 602
 603         flags = oops_begin();
 604         tsk = current;
 605         sig = SIGKILL;
 606
 607         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 608                tsk->comm, address);
 609         dump_pagetable(address);
 610
 611         tsk->thread.cr2         = address;
 612         tsk->thread.trap_no     = 14;
 613         tsk->thread.error_code  = error_code;
 614
 615         if (__die("Bad pagetable", regs, error_code))
 616                 sig = 0;
 617
 618         oops_end(flags, regs, sig);
 619 }
 620
 621 static noinline void
 622 no_context(struct pt_regs *regs, unsigned long error_code,
 623            unsigned long address)
 624 {
 625         struct task_struct *tsk = current;
 626         unsigned long *stackend;
 627         unsigned long flags;
 628         int sig;
 629
 630         /* Are we prepared to handle this kernel fault? */
 631         if (fixup_exception(regs))
 632                 return;
 633
 634         /*
 635          * 32-bit:
 636          *
 637          *   Valid to do another page fault here, because if this fault
 638          *   had been triggered by is_prefetch fixup_exception would have
 639          *   handled it.
 640          *
 641          * 64-bit:
 642          *
 643          *   Hall of shame of CPU/BIOS bugs.
 644          */
 645         if (is_prefetch(regs, error_code, address))
 646                 return;
 647
 648         if (is_errata93(regs, address))
 649                 return;
 650
 651         /*
 652          * Oops. The kernel tried to access some bad page. We'll have to
 653          * terminate things with extreme prejudice:
 654          */
 655         flags = oops_begin();
 656
 657         show_fault_oops(regs, error_code, address);
 658
 659         stackend = end_of_stack(tsk);
 660         if (*stackend != STACK_END_MAGIC)
 661                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 662
 663         tsk->thread.cr2         = address;
 664         tsk->thread.trap_no     = 14;
 665         tsk->thread.error_code  = error_code;
 666
 667         sig = SIGKILL;
 668         if (__die("Oops", regs, error_code))
 669                 sig = 0;
 670
 671         /* Executive summary in case the body of the oops scrolled away */
 672         printk(KERN_EMERG "CR2: %016lx\n", address);
 673
 674         oops_end(flags, regs, sig);
 675 }
 676
 677 /*
 678  * Print out info about fatal segfaults, if the show_unhandled_signals
 679  * sysctl is set:
 680  */
 681 static inline void
 682 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 683                 unsigned long address, struct task_struct *tsk)
 684 {
 685         if (!unhandled_signal(tsk, SIGSEGV))
 686                 return;
 687
 688         if (!printk_ratelimit())
 689                 return;
 690
 691         printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 692                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 693                 tsk->comm, task_pid_nr(tsk), address,
 694                 (void *)regs->ip, (void *)regs->sp, error_code);
 695
 696         print_vma_addr(KERN_CONT " in ", regs->ip);
 697
 698         printk(KERN_CONT "\n");
 699 }
 700
 701 static void
 702 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 703                        unsigned long address, int si_code)
 704 {
 705         struct task_struct *tsk = current;
 706
 707         /* User mode accesses just cause a SIGSEGV */
 708         if (error_code & PF_USER) {
 709                 /*
 710                  * It's possible to have interrupts off here:
 711                  */
 712                 local_irq_enable();
 713
 714                 /*
 715                  * Valid to do another page fault here because this one came
 716                  * from user space:
 717                  */
 718                 if (is_prefetch(regs, error_code, address))
 719                         return;
 720
 721                 if (is_errata100(regs, address))
 722                         return;
 723
 724                 if (unlikely(show_unhandled_signals))
 725                         show_signal_msg(regs, error_code, address, tsk);
 726
 727                 /* Kernel addresses are always protection faults: */
 728                 tsk->thread.cr2         = address;
 729                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 730                 tsk->thread.trap_no     = 14;
 731
 732                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 733
 734                 return;
 735         }
 736
 737         if (is_f00f_bug(regs, address))
 738                 return;
 739
 740         no_context(regs, error_code, address);
 741 }
 742
 743 static noinline void
 744 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 745                      unsigned long address)
 746 {
 747         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 748 }
 749
 750 static void
 751 __bad_area(struct pt_regs *regs, unsigned long error_code,
 752            unsigned long address, int si_code)
 753 {
 754         struct mm_struct *mm = current->mm;
 755
 756         /*
 757          * Something tried to access memory that isn't in our memory map..
 758          * Fix it, but check if it's kernel or user first..
 759          */
 760         up_read(&mm->mmap_sem);
 761
 762         __bad_area_nosemaphore(regs, error_code, address, si_code);
 763 }
 764
 765 static noinline void
 766 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 767 {
 768         __bad_area(regs, error_code, address, SEGV_MAPERR);
 769 }
 770
 771 static noinline void
 772 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 773                       unsigned long address)
 774 {
 775         __bad_area(regs, error_code, address, SEGV_ACCERR);
 776 }
 777
 778 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 779 static void
 780 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 781               unsigned long address)
 782 {
 783         /*
 784          * We ran out of memory, call the OOM killer, and return the userspace
 785          * (which will retry the fault, or kill us if we got oom-killed):
 786          */
 787         up_read(&current->mm->mmap_sem);
 788
 789         pagefault_out_of_memory();
 790 }
 791
 792 static void
 793 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 794 {
 795         struct task_struct *tsk = current;
 796         struct mm_struct *mm = tsk->mm;
 797
 798         up_read(&mm->mmap_sem);
 799
 800         /* Kernel mode? Handle exceptions or die: */
 801         if (!(error_code & PF_USER))
 802                 no_context(regs, error_code, address);
 803
 804         /* User-space => ok to do another page fault: */
 805         if (is_prefetch(regs, error_code, address))
 806                 return;
 807
 808         tsk->thread.cr2         = address;
 809         tsk->thread.error_code  = error_code;
 810         tsk->thread.trap_no     = 14;
 811
 812         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 813 }
 814
 815 static noinline void
 816 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 817                unsigned long address, unsigned int fault)
 818 {
 819         if (fault & VM_FAULT_OOM) {
 820                 out_of_memory(regs, error_code, address);
 821         } else {
 822                 if (fault & VM_FAULT_SIGBUS)
 823                         do_sigbus(regs, error_code, address);
 824                 else
 825                         BUG();
 826         }
 827 }
 828
 829 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 830 {
 831         if ((error_code & PF_WRITE) && !pte_write(*pte))
 832                 return 0;
 833
 834         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 835                 return 0;
 836
 837         return 1;
 838 }
 839
 840 /*
 841  * Handle a spurious fault caused by a stale TLB entry.
 842  *
 843  * This allows us to lazily refresh the TLB when increasing the
 844  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 845  * eagerly is very expensive since that implies doing a full
 846  * cross-processor TLB flush, even if no stale TLB entries exist
 847  * on other processors.
 848  *
 849  * There are no security implications to leaving a stale TLB when
 850  * increasing the permissions on a page.
 851  */
 852 static noinline int
 853 spurious_fault(unsigned long error_code, unsigned long address)
 854 {
 855         pgd_t *pgd;
 856         pud_t *pud;
 857         pmd_t *pmd;
 858         pte_t *pte;
 859         int ret;
 860
 861         /* Reserved-bit violation or user access to kernel space? */
 862         if (error_code & (PF_USER | PF_RSVD))
 863                 return 0;
 864
 865         pgd = init_mm.pgd + pgd_index(address);
 866         if (!pgd_present(*pgd))
 867                 return 0;
 868
 869         pud = pud_offset(pgd, address);
 870         if (!pud_present(*pud))
 871                 return 0;
 872
 873         if (pud_large(*pud))
 874                 return spurious_fault_check(error_code, (pte_t *) pud);
 875
 876         pmd = pmd_offset(pud, address);
 877         if (!pmd_present(*pmd))
 878                 return 0;
 879
 880         if (pmd_large(*pmd))
 881                 return spurious_fault_check(error_code, (pte_t *) pmd);
 882
 883         pte = pte_offset_kernel(pmd, address);
 884         if (!pte_present(*pte))
 885                 return 0;
 886
 887         ret = spurious_fault_check(error_code, pte);
 888         if (!ret)
 889                 return 0;
 890
 891         /*
 892          * Make sure we have permissions in PMD.
 893          * If not, then there's a bug in the page tables:
 894          */
 895         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 896         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 897
 898         return ret;
 899 }
 900
 901 int show_unhandled_signals = 1;
 902
 903 static inline int
 904 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
 905 {
 906         if (write) {
 907                 /* write, present and write, not present: */
 908                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 909                         return 1;
 910                 return 0;
 911         }
 912
 913         /* read, present: */
 914         if (unlikely(error_code & PF_PROT))
 915                 return 1;
 916
 917         /* read, not present: */
 918         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 919                 return 1;
 920
 921         return 0;
 922 }
 923
 924 static int fault_in_kernel_space(unsigned long address)
 925 {
 926         return address >= TASK_SIZE_MAX;
 927 }
 928
 929 /*
 930  * This routine handles page faults.  It determines the address,
 931  * and the problem, and then passes it off to one of the appropriate
 932  * routines.
 933  */
 934 dotraplinkage void __kprobes
 935 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 936 {
 937         struct vm_area_struct *vma;
 938         struct task_struct *tsk;
 939         unsigned long address;
 940         struct mm_struct *mm;
 941         int write;
 942         int fault;
 943
 944         tsk = current;
 945         mm = tsk->mm;
 946
 947         /* Get the faulting address: */
 948         address = read_cr2();
 949
 950         /*
 951          * Detect and handle instructions that would cause a page fault for
 952          * both a tracked kernel page and a userspace page.
 953          */
 954         if (kmemcheck_active(regs))
 955                 kmemcheck_hide(regs);
 956         prefetchw(&mm->mmap_sem);
 957
 958         if (unlikely(kmmio_fault(regs, address)))
 959                 return;
 960
 961         /*
 962          * We fault-in kernel-space virtual memory on-demand. The
 963          * 'reference' page table is init_mm.pgd.
 964          *
 965          * NOTE! We MUST NOT take any locks for this case. We may
 966          * be in an interrupt or a critical region, and should
 967          * only copy the information from the master page table,
 968          * nothing more.
 969          *
 970          * This verifies that the fault happens in kernel space
 971          * (error_code & 4) == 0, and that the fault was not a
 972          * protection error (error_code & 9) == 0.
 973          */
 974         if (unlikely(fault_in_kernel_space(address))) {
 975                 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
 976                         if (vmalloc_fault(address) >= 0)
 977                                 return;
 978
 979                         if (kmemcheck_fault(regs, address, error_code))
 980                                 return;
 981                 }
 982
 983                 /* Can handle a stale RO->RW TLB: */
 984                 if (spurious_fault(error_code, address))
 985                         return;
 986
 987                 /* kprobes don't want to hook the spurious faults: */
 988                 if (notify_page_fault(regs))
 989                         return;
 990                 /*
 991                  * Don't take the mm semaphore here. If we fixup a prefetch
 992                  * fault we could otherwise deadlock:
 993                  */
 994                 bad_area_nosemaphore(regs, error_code, address);
 995
 996                 return;
 997         }
 998
 999         /* kprobes don't want to hook the spurious faults: */
1000         if (unlikely(notify_page_fault(regs)))
1001                 return;
1002         /*
1003          * It's safe to allow irq's after cr2 has been saved and the
1004          * vmalloc fault has been handled.
1005          *
1006          * User-mode registers count as a user access even for any
1007          * potential system fault or CPU buglet:
1008          */
1009         if (user_mode_vm(regs)) {
1010                 local_irq_enable();
1011                 error_code |= PF_USER;
1012         } else {
1013                 if (regs->flags & X86_EFLAGS_IF)
1014                         local_irq_enable();
1015         }
1016
1017         if (unlikely(error_code & PF_RSVD))
1018                 pgtable_bad(regs, error_code, address);
1019
1020         perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1021
1022         /*
1023          * If we're in an interrupt, have no user context or are running
1024          * in an atomic region then we must not take the fault:
1025          */
1026         if (unlikely(in_atomic() || !mm)) {
1027                 bad_area_nosemaphore(regs, error_code, address);
1028                 return;
1029         }
1030
1031         /*
1032          * When running in the kernel we expect faults to occur only to
1033          * addresses in user space.  All other faults represent errors in
1034          * the kernel and should generate an OOPS.  Unfortunately, in the
1035          * case of an erroneous fault occurring in a code path which already
1036          * holds mmap_sem we will deadlock attempting to validate the fault
1037          * against the address space.  Luckily the kernel only validly
1038          * references user space from well defined areas of code, which are
1039          * listed in the exceptions table.
1040          *
1041          * As the vast majority of faults will be valid we will only perform
1042          * the source reference check when there is a possibility of a
1043          * deadlock. Attempt to lock the address space, if we cannot we then
1044          * validate the source. If this is invalid we can skip the address
1045          * space check, thus avoiding the deadlock:
1046          */
1047         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1048                 if ((error_code & PF_USER) == 0 &&
1049                     !search_exception_tables(regs->ip)) {
1050                         bad_area_nosemaphore(regs, error_code, address);
1051                         return;
1052                 }
1053                 down_read(&mm->mmap_sem);
1054         } else {
1055                 /*
1056                  * The above down_read_trylock() might have succeeded in
1057                  * which case we'll have missed the might_sleep() from
1058                  * down_read():
1059                  */
1060                 might_sleep();
1061         }
1062
1063         vma = find_vma(mm, address);
1064         if (unlikely(!vma)) {
1065                 bad_area(regs, error_code, address);
1066                 return;
1067         }
1068         if (likely(vma->vm_start <= address))
1069                 goto good_area;
1070         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1071                 bad_area(regs, error_code, address);
1072                 return;
1073         }
1074         if (error_code & PF_USER) {
1075                 /*
1076                  * Accessing the stack below %sp is always a bug.
1077                  * The large cushion allows instructions like enter
1078                  * and pusha to work. ("enter $65535, $31" pushes
1079                  * 32 pointers and then decrements %sp by 65535.)
1080                  */
1081                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1082                         bad_area(regs, error_code, address);
1083                         return;
1084                 }
1085         }
1086         if (unlikely(expand_stack(vma, address))) {
1087                 bad_area(regs, error_code, address);
1088                 return;
1089         }
1090
1091         /*
1092          * Ok, we have a good vm_area for this memory access, so
1093          * we can handle it..
1094          */
1095 good_area:
1096         write = error_code & PF_WRITE;
1097
1098         if (unlikely(access_error(error_code, write, vma))) {
1099                 bad_area_access_error(regs, error_code, address);
1100                 return;
1101         }
1102
1103         /*
1104          * If for any reason at all we couldn't handle the fault,
1105          * make sure we exit gracefully rather than endlessly redo
1106          * the fault:
1107          */
1108         fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
1109
1110         if (unlikely(fault & VM_FAULT_ERROR)) {
1111                 mm_fault_error(regs, error_code, address, fault);
1112                 return;
1113         }
1114
1115         if (fault & VM_FAULT_MAJOR) {
1116                 tsk->maj_flt++;
1117                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1118                                      regs, address);
1119         } else {
1120                 tsk->min_flt++;
1121                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1122                                      regs, address);
1123         }
1124
1125         check_v8086_mode(regs, address, tsk);
1126
1127         up_read(&mm->mmap_sem);
1128 }