arch/x86/mm/fault.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  *  Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
   4  */
   5
   6 #include <linux/signal.h>
   7 #include <linux/sched.h>
   8 #include <linux/kernel.h>
   9 #include <linux/errno.h>
  10 #include <linux/string.h>
  11 #include <linux/types.h>
  12 #include <linux/ptrace.h>
  13 #include <linux/mmiotrace.h>
  14 #include <linux/mman.h>
  15 #include <linux/mm.h>
  16 #include <linux/smp.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/init.h>
  19 #include <linux/tty.h>
  20 #include <linux/vt_kern.h>              /* For unblank_screen() */
  21 #include <linux/compiler.h>
  22 #include <linux/highmem.h>
  23 #include <linux/bootmem.h>              /* for max_low_pfn */
  24 #include <linux/vmalloc.h>
  25 #include <linux/module.h>
  26 #include <linux/kprobes.h>
  27 #include <linux/uaccess.h>
  28 #include <linux/kdebug.h>
  29
  30 #include <asm/system.h>
  31 #include <asm/desc.h>
  32 #include <asm/segment.h>
  33 #include <asm/pgalloc.h>
  34 #include <asm/smp.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/proto.h>
  37 #include <asm-generic/sections.h>
  38 #include <asm/traps.h>
  39
  40 /*
  41  * Page fault error code bits
  42  *      bit 0 == 0 means no page found, 1 means protection fault
  43  *      bit 1 == 0 means read, 1 means write
  44  *      bit 2 == 0 means kernel, 1 means user-mode
  45  *      bit 3 == 1 means use of reserved bit detected
  46  *      bit 4 == 1 means fault was an instruction fetch
  47  */
  48 #define PF_PROT         (1<<0)
  49 #define PF_WRITE        (1<<1)
  50 #define PF_USER         (1<<2)
  51 #define PF_RSVD         (1<<3)
  52 #define PF_INSTR        (1<<4)
  53
  54 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  55 {
  56 #ifdef CONFIG_MMIOTRACE
  57         if (unlikely(is_kmmio_active()))
  58                 if (kmmio_handler(regs, addr) == 1)
  59                         return -1;
  60 #endif
  61         return 0;
  62 }
  63
  64 static inline int notify_page_fault(struct pt_regs *regs)
  65 {
  66 #ifdef CONFIG_KPROBES
  67         int ret = 0;
  68
  69         /* kprobe_running() needs smp_processor_id() */
  70         if (!user_mode_vm(regs)) {
  71                 preempt_disable();
  72                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  73                         ret = 1;
  74                 preempt_enable();
  75         }
  76
  77         return ret;
  78 #else
  79         return 0;
  80 #endif
  81 }
  82
  83 /*
  84  * X86_32
  85  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  86  * Check that here and ignore it.
  87  *
  88  * X86_64
  89  * Sometimes the CPU reports invalid exceptions on prefetch.
  90  * Check that here and ignore it.
  91  *
  92  * Opcode checker based on code by Richard Brunner
  93  */
  94 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
  95                        unsigned long error_code)
  96 {
  97         unsigned char *instr;
  98         int scan_more = 1;
  99         int prefetch = 0;
 100         unsigned char *max_instr;
 101
 102         /*
 103          * If it was a exec (instruction fetch) fault on NX page, then
 104          * do not ignore the fault:
 105          */
 106         if (error_code & PF_INSTR)
 107                 return 0;
 108
 109         instr = (unsigned char *)convert_ip_to_linear(current, regs);
 110         max_instr = instr + 15;
 111
 112         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 113                 return 0;
 114
 115         while (scan_more && instr < max_instr) {
 116                 unsigned char opcode;
 117                 unsigned char instr_hi;
 118                 unsigned char instr_lo;
 119
 120                 if (probe_kernel_address(instr, opcode))
 121                         break;
 122
 123                 instr_hi = opcode & 0xf0;
 124                 instr_lo = opcode & 0x0f;
 125                 instr++;
 126
 127                 switch (instr_hi) {
 128                 case 0x20:
 129                 case 0x30:
 130                         /*
 131                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 132                          * In X86_64 long mode, the CPU will signal invalid
 133                          * opcode if some of these prefixes are present so
 134                          * X86_64 will never get here anyway
 135                          */
 136                         scan_more = ((instr_lo & 7) == 0x6);
 137                         break;
 138 #ifdef CONFIG_X86_64
 139                 case 0x40:
 140                         /*
 141                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 142                          * Need to figure out under what instruction mode the
 143                          * instruction was issued. Could check the LDT for lm,
 144                          * but for now it's good enough to assume that long
 145                          * mode only uses well known segments or kernel.
 146                          */
 147                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 148                         break;
 149 #endif
 150                 case 0x60:
 151                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 152                         scan_more = (instr_lo & 0xC) == 0x4;
 153                         break;
 154                 case 0xF0:
 155                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 156                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 157                         break;
 158                 case 0x00:
 159                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 160                         scan_more = 0;
 161
 162                         if (probe_kernel_address(instr, opcode))
 163                                 break;
 164                         prefetch = (instr_lo == 0xF) &&
 165                                 (opcode == 0x0D || opcode == 0x18);
 166                         break;
 167                 default:
 168                         scan_more = 0;
 169                         break;
 170                 }
 171         }
 172         return prefetch;
 173 }
 174
 175 static void force_sig_info_fault(int si_signo, int si_code,
 176         unsigned long address, struct task_struct *tsk)
 177 {
 178         siginfo_t info;
 179
 180         info.si_signo = si_signo;
 181         info.si_errno = 0;
 182         info.si_code = si_code;
 183         info.si_addr = (void __user *)address;
 184         force_sig_info(si_signo, &info, tsk);
 185 }
 186
 187 #ifdef CONFIG_X86_64
 188 static int bad_address(void *p)
 189 {
 190         unsigned long dummy;
 191         return probe_kernel_address((unsigned long *)p, dummy);
 192 }
 193 #endif
 194
 195 static void dump_pagetable(unsigned long address)
 196 {
 197 #ifdef CONFIG_X86_32
 198         __typeof__(pte_val(__pte(0))) page;
 199
 200         page = read_cr3();
 201         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 202 #ifdef CONFIG_X86_PAE
 203         printk("*pdpt = %016Lx ", page);
 204         if ((page >> PAGE_SHIFT) < max_low_pfn
 205             && page & _PAGE_PRESENT) {
 206                 page &= PAGE_MASK;
 207                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 208                                                          & (PTRS_PER_PMD - 1)];
 209                 printk(KERN_CONT "*pde = %016Lx ", page);
 210                 page &= ~_PAGE_NX;
 211         }
 212 #else
 213         printk("*pde = %08lx ", page);
 214 #endif
 215
 216         /*
 217          * We must not directly access the pte in the highpte
 218          * case if the page table is located in highmem.
 219          * And let's rather not kmap-atomic the pte, just in case
 220          * it's allocated already.
 221          */
 222         if ((page >> PAGE_SHIFT) < max_low_pfn
 223             && (page & _PAGE_PRESENT)
 224             && !(page & _PAGE_PSE)) {
 225                 page &= PAGE_MASK;
 226                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 227                                                          & (PTRS_PER_PTE - 1)];
 228                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 229         }
 230
 231         printk("\n");
 232 #else /* CONFIG_X86_64 */
 233         pgd_t *pgd;
 234         pud_t *pud;
 235         pmd_t *pmd;
 236         pte_t *pte;
 237
 238         pgd = (pgd_t *)read_cr3();
 239
 240         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 241         pgd += pgd_index(address);
 242         if (bad_address(pgd)) goto bad;
 243         printk("PGD %lx ", pgd_val(*pgd));
 244         if (!pgd_present(*pgd)) goto ret;
 245
 246         pud = pud_offset(pgd, address);
 247         if (bad_address(pud)) goto bad;
 248         printk("PUD %lx ", pud_val(*pud));
 249         if (!pud_present(*pud) || pud_large(*pud))
 250                 goto ret;
 251
 252         pmd = pmd_offset(pud, address);
 253         if (bad_address(pmd)) goto bad;
 254         printk("PMD %lx ", pmd_val(*pmd));
 255         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 256
 257         pte = pte_offset_kernel(pmd, address);
 258         if (bad_address(pte)) goto bad;
 259         printk("PTE %lx", pte_val(*pte));
 260 ret:
 261         printk("\n");
 262         return;
 263 bad:
 264         printk("BAD\n");
 265 #endif
 266 }
 267
 268 #ifdef CONFIG_X86_32
 269 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 270 {
 271         unsigned index = pgd_index(address);
 272         pgd_t *pgd_k;
 273         pud_t *pud, *pud_k;
 274         pmd_t *pmd, *pmd_k;
 275
 276         pgd += index;
 277         pgd_k = init_mm.pgd + index;
 278
 279         if (!pgd_present(*pgd_k))
 280                 return NULL;
 281
 282         /*
 283          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 284          * and redundant with the set_pmd() on non-PAE. As would
 285          * set_pud.
 286          */
 287
 288         pud = pud_offset(pgd, address);
 289         pud_k = pud_offset(pgd_k, address);
 290         if (!pud_present(*pud_k))
 291                 return NULL;
 292
 293         pmd = pmd_offset(pud, address);
 294         pmd_k = pmd_offset(pud_k, address);
 295         if (!pmd_present(*pmd_k))
 296                 return NULL;
 297         if (!pmd_present(*pmd)) {
 298                 set_pmd(pmd, *pmd_k);
 299                 arch_flush_lazy_mmu_mode();
 300         } else
 301                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 302         return pmd_k;
 303 }
 304 #endif
 305
 306 #ifdef CONFIG_X86_64
 307 static const char errata93_warning[] =
 308 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 309 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 310 KERN_ERR "******* Please consider a BIOS update.\n"
 311 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 312 #endif
 313
 314 /* Workaround for K8 erratum #93 & buggy BIOS.
 315    BIOS SMM functions are required to use a specific workaround
 316    to avoid corruption of the 64bit RIP register on C stepping K8.
 317    A lot of BIOS that didn't get tested properly miss this.
 318    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 319    Try to work around it here.
 320    Note we only handle faults in kernel here.
 321    Does nothing for X86_32
 322  */
 323 static int is_errata93(struct pt_regs *regs, unsigned long address)
 324 {
 325 #ifdef CONFIG_X86_64
 326         static int warned;
 327         if (address != regs->ip)
 328                 return 0;
 329         if ((address >> 32) != 0)
 330                 return 0;
 331         address |= 0xffffffffUL << 32;
 332         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 333             (address >= MODULES_VADDR && address <= MODULES_END)) {
 334                 if (!warned) {
 335                         printk(errata93_warning);
 336                         warned = 1;
 337                 }
 338                 regs->ip = address;
 339                 return 1;
 340         }
 341 #endif
 342         return 0;
 343 }
 344
 345 /*
 346  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 347  * addresses >4GB.  We catch this in the page fault handler because these
 348  * addresses are not reachable. Just detect this case and return.  Any code
 349  * segment in LDT is compatibility mode.
 350  */
 351 static int is_errata100(struct pt_regs *regs, unsigned long address)
 352 {
 353 #ifdef CONFIG_X86_64
 354         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 355             (address >> 32))
 356                 return 1;
 357 #endif
 358         return 0;
 359 }
 360
 361 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 362 {
 363 #ifdef CONFIG_X86_F00F_BUG
 364         unsigned long nr;
 365         /*
 366          * Pentium F0 0F C7 C8 bug workaround.
 367          */
 368         if (boot_cpu_data.f00f_bug) {
 369                 nr = (address - idt_descr.address) >> 3;
 370
 371                 if (nr == 6) {
 372                         do_invalid_op(regs, 0);
 373                         return 1;
 374                 }
 375         }
 376 #endif
 377         return 0;
 378 }
 379
 380 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 381                             unsigned long address)
 382 {
 383 #ifdef CONFIG_X86_32
 384         if (!oops_may_print())
 385                 return;
 386 #endif
 387
 388 #ifdef CONFIG_X86_PAE
 389         if (error_code & PF_INSTR) {
 390                 unsigned int level;
 391                 pte_t *pte = lookup_address(address, &level);
 392
 393                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 394                         printk(KERN_CRIT "kernel tried to execute "
 395                                 "NX-protected page - exploit attempt? "
 396                                 "(uid: %d)\n", current_uid());
 397         }
 398 #endif
 399
 400         printk(KERN_ALERT "BUG: unable to handle kernel ");
 401         if (address < PAGE_SIZE)
 402                 printk(KERN_CONT "NULL pointer dereference");
 403         else
 404                 printk(KERN_CONT "paging request");
 405         printk(KERN_CONT " at %p\n", (void *) address);
 406         printk(KERN_ALERT "IP:");
 407         printk_address(regs->ip, 1);
 408         dump_pagetable(address);
 409 }
 410
 411 #ifdef CONFIG_X86_64
 412 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 413                                  unsigned long error_code)
 414 {
 415         unsigned long flags = oops_begin();
 416         int sig = SIGKILL;
 417         struct task_struct *tsk;
 418
 419         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 420                current->comm, address);
 421         dump_pagetable(address);
 422         tsk = current;
 423         tsk->thread.cr2 = address;
 424         tsk->thread.trap_no = 14;
 425         tsk->thread.error_code = error_code;
 426         if (__die("Bad pagetable", regs, error_code))
 427                 sig = 0;
 428         oops_end(flags, regs, sig);
 429 }
 430 #endif
 431
 432 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 433 {
 434         if ((error_code & PF_WRITE) && !pte_write(*pte))
 435                 return 0;
 436         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 437                 return 0;
 438
 439         return 1;
 440 }
 441
 442 /*
 443  * Handle a spurious fault caused by a stale TLB entry.  This allows
 444  * us to lazily refresh the TLB when increasing the permissions of a
 445  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
 446  * expensive since that implies doing a full cross-processor TLB
 447  * flush, even if no stale TLB entries exist on other processors.
 448  * There are no security implications to leaving a stale TLB when
 449  * increasing the permissions on a page.
 450  */
 451 static int spurious_fault(unsigned long address,
 452                           unsigned long error_code)
 453 {
 454         pgd_t *pgd;
 455         pud_t *pud;
 456         pmd_t *pmd;
 457         pte_t *pte;
 458         int ret;
 459
 460         /* Reserved-bit violation or user access to kernel space? */
 461         if (error_code & (PF_USER | PF_RSVD))
 462                 return 0;
 463
 464         pgd = init_mm.pgd + pgd_index(address);
 465         if (!pgd_present(*pgd))
 466                 return 0;
 467
 468         pud = pud_offset(pgd, address);
 469         if (!pud_present(*pud))
 470                 return 0;
 471
 472         if (pud_large(*pud))
 473                 return spurious_fault_check(error_code, (pte_t *) pud);
 474
 475         pmd = pmd_offset(pud, address);
 476         if (!pmd_present(*pmd))
 477                 return 0;
 478
 479         if (pmd_large(*pmd))
 480                 return spurious_fault_check(error_code, (pte_t *) pmd);
 481
 482         pte = pte_offset_kernel(pmd, address);
 483         if (!pte_present(*pte))
 484                 return 0;
 485
 486         ret = spurious_fault_check(error_code, pte);
 487         if (!ret)
 488                 return 0;
 489
 490         /*
 491          * Make sure we have permissions in PMD
 492          * If not, then there's a bug in the page tables.
 493          */
 494         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 495         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 496         return ret;
 497 }
 498
 499 /*
 500  * X86_32
 501  * Handle a fault on the vmalloc or module mapping area
 502  *
 503  * X86_64
 504  * Handle a fault on the vmalloc area
 505  *
 506  * This assumes no large pages in there.
 507  */
 508 static int vmalloc_fault(unsigned long address)
 509 {
 510 #ifdef CONFIG_X86_32
 511         unsigned long pgd_paddr;
 512         pmd_t *pmd_k;
 513         pte_t *pte_k;
 514
 515         /* Make sure we are in vmalloc area */
 516         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 517                 return -1;
 518
 519         /*
 520          * Synchronize this task's top level page-table
 521          * with the 'reference' page table.
 522          *
 523          * Do _not_ use "current" here. We might be inside
 524          * an interrupt in the middle of a task switch..
 525          */
 526         pgd_paddr = read_cr3();
 527         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 528         if (!pmd_k)
 529                 return -1;
 530         pte_k = pte_offset_kernel(pmd_k, address);
 531         if (!pte_present(*pte_k))
 532                 return -1;
 533         return 0;
 534 #else
 535         pgd_t *pgd, *pgd_ref;
 536         pud_t *pud, *pud_ref;
 537         pmd_t *pmd, *pmd_ref;
 538         pte_t *pte, *pte_ref;
 539
 540         /* Make sure we are in vmalloc area */
 541         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 542                 return -1;
 543
 544         /* Copy kernel mappings over when needed. This can also
 545            happen within a race in page table update. In the later
 546            case just flush. */
 547
 548         pgd = pgd_offset(current->active_mm, address);
 549         pgd_ref = pgd_offset_k(address);
 550         if (pgd_none(*pgd_ref))
 551                 return -1;
 552         if (pgd_none(*pgd))
 553                 set_pgd(pgd, *pgd_ref);
 554         else
 555                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 556
 557         /* Below here mismatches are bugs because these lower tables
 558            are shared */
 559
 560         pud = pud_offset(pgd, address);
 561         pud_ref = pud_offset(pgd_ref, address);
 562         if (pud_none(*pud_ref))
 563                 return -1;
 564         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 565                 BUG();
 566         pmd = pmd_offset(pud, address);
 567         pmd_ref = pmd_offset(pud_ref, address);
 568         if (pmd_none(*pmd_ref))
 569                 return -1;
 570         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 571                 BUG();
 572         pte_ref = pte_offset_kernel(pmd_ref, address);
 573         if (!pte_present(*pte_ref))
 574                 return -1;
 575         pte = pte_offset_kernel(pmd, address);
 576         /* Don't use pte_page here, because the mappings can point
 577            outside mem_map, and the NUMA hash lookup cannot handle
 578            that. */
 579         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 580                 BUG();
 581         return 0;
 582 #endif
 583 }
 584
 585 int show_unhandled_signals = 1;
 586
 587 /*
 588  * This routine handles page faults.  It determines the address,
 589  * and the problem, and then passes it off to one of the appropriate
 590  * routines.
 591  */
 592 #ifdef CONFIG_X86_64
 593 asmlinkage
 594 #endif
 595 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 596 {
 597         struct task_struct *tsk;
 598         struct mm_struct *mm;
 599         struct vm_area_struct *vma;
 600         unsigned long address;
 601         int write, si_code;
 602         int fault;
 603 #ifdef CONFIG_X86_64
 604         unsigned long flags;
 605         int sig;
 606 #endif
 607
 608         tsk = current;
 609         mm = tsk->mm;
 610         prefetchw(&mm->mmap_sem);
 611
 612         /* get the address */
 613         address = read_cr2();
 614
 615         si_code = SEGV_MAPERR;
 616
 617         if (unlikely(kmmio_fault(regs, address)))
 618                 return;
 619
 620         /*
 621          * We fault-in kernel-space virtual memory on-demand. The
 622          * 'reference' page table is init_mm.pgd.
 623          *
 624          * NOTE! We MUST NOT take any locks for this case. We may
 625          * be in an interrupt or a critical region, and should
 626          * only copy the information from the master page table,
 627          * nothing more.
 628          *
 629          * This verifies that the fault happens in kernel space
 630          * (error_code & 4) == 0, and that the fault was not a
 631          * protection error (error_code & 9) == 0.
 632          */
 633 #ifdef CONFIG_X86_32
 634         if (unlikely(address >= TASK_SIZE)) {
 635 #else
 636         if (unlikely(address >= TASK_SIZE64)) {
 637 #endif
 638                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 639                     vmalloc_fault(address) >= 0)
 640                         return;
 641
 642                 /* Can handle a stale RO->RW TLB */
 643                 if (spurious_fault(address, error_code))
 644                         return;
 645
 646                 /* kprobes don't want to hook the spurious faults. */
 647                 if (notify_page_fault(regs))
 648                         return;
 649                 /*
 650                  * Don't take the mm semaphore here. If we fixup a prefetch
 651                  * fault we could otherwise deadlock.
 652                  */
 653                 goto bad_area_nosemaphore;
 654         }
 655
 656         /* kprobes don't want to hook the spurious faults. */
 657         if (notify_page_fault(regs))
 658                 return;
 659
 660         /*
 661          * It's safe to allow irq's after cr2 has been saved and the
 662          * vmalloc fault has been handled.
 663          *
 664          * User-mode registers count as a user access even for any
 665          * potential system fault or CPU buglet.
 666          */
 667         if (user_mode_vm(regs)) {
 668                 local_irq_enable();
 669                 error_code |= PF_USER;
 670         } else if (regs->flags & X86_EFLAGS_IF)
 671                 local_irq_enable();
 672
 673 #ifdef CONFIG_X86_64
 674         if (unlikely(error_code & PF_RSVD))
 675                 pgtable_bad(address, regs, error_code);
 676 #endif
 677
 678         /*
 679          * If we're in an interrupt, have no user context or are running in an
 680          * atomic region then we must not take the fault.
 681          */
 682         if (unlikely(in_atomic() || !mm))
 683                 goto bad_area_nosemaphore;
 684
 685         /*
 686          * When running in the kernel we expect faults to occur only to
 687          * addresses in user space.  All other faults represent errors in the
 688          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 689          * erroneous fault occurring in a code path which already holds mmap_sem
 690          * we will deadlock attempting to validate the fault against the
 691          * address space.  Luckily the kernel only validly references user
 692          * space from well defined areas of code, which are listed in the
 693          * exceptions table.
 694          *
 695          * As the vast majority of faults will be valid we will only perform
 696          * the source reference check when there is a possibility of a deadlock.
 697          * Attempt to lock the address space, if we cannot we then validate the
 698          * source.  If this is invalid we can skip the address space check,
 699          * thus avoiding the deadlock.
 700          */
 701         if (!down_read_trylock(&mm->mmap_sem)) {
 702                 if ((error_code & PF_USER) == 0 &&
 703                     !search_exception_tables(regs->ip))
 704                         goto bad_area_nosemaphore;
 705                 down_read(&mm->mmap_sem);
 706         }
 707
 708         vma = find_vma(mm, address);
 709         if (!vma)
 710                 goto bad_area;
 711         if (vma->vm_start <= address)
 712                 goto good_area;
 713         if (!(vma->vm_flags & VM_GROWSDOWN))
 714                 goto bad_area;
 715         if (error_code & PF_USER) {
 716                 /*
 717                  * Accessing the stack below %sp is always a bug.
 718                  * The large cushion allows instructions like enter
 719                  * and pusha to work.  ("enter $65535,$31" pushes
 720                  * 32 pointers and then decrements %sp by 65535.)
 721                  */
 722                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 723                         goto bad_area;
 724         }
 725         if (expand_stack(vma, address))
 726                 goto bad_area;
 727 /*
 728  * Ok, we have a good vm_area for this memory access, so
 729  * we can handle it..
 730  */
 731 good_area:
 732         si_code = SEGV_ACCERR;
 733         write = 0;
 734         switch (error_code & (PF_PROT|PF_WRITE)) {
 735         default:        /* 3: write, present */
 736                 /* fall through */
 737         case PF_WRITE:          /* write, not present */
 738                 if (!(vma->vm_flags & VM_WRITE))
 739                         goto bad_area;
 740                 write++;
 741                 break;
 742         case PF_PROT:           /* read, present */
 743                 goto bad_area;
 744         case 0:                 /* read, not present */
 745                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 746                         goto bad_area;
 747         }
 748
 749         /*
 750          * If for any reason at all we couldn't handle the fault,
 751          * make sure we exit gracefully rather than endlessly redo
 752          * the fault.
 753          */
 754         fault = handle_mm_fault(mm, vma, address, write);
 755         if (unlikely(fault & VM_FAULT_ERROR)) {
 756                 if (fault & VM_FAULT_OOM)
 757                         goto out_of_memory;
 758                 else if (fault & VM_FAULT_SIGBUS)
 759                         goto do_sigbus;
 760                 BUG();
 761         }
 762         if (fault & VM_FAULT_MAJOR)
 763                 tsk->maj_flt++;
 764         else
 765                 tsk->min_flt++;
 766
 767 #ifdef CONFIG_X86_32
 768         /*
 769          * Did it hit the DOS screen memory VA from vm86 mode?
 770          */
 771         if (v8086_mode(regs)) {
 772                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 773                 if (bit < 32)
 774                         tsk->thread.screen_bitmap |= 1 << bit;
 775         }
 776 #endif
 777         up_read(&mm->mmap_sem);
 778         return;
 779
 780 /*
 781  * Something tried to access memory that isn't in our memory map..
 782  * Fix it, but check if it's kernel or user first..
 783  */
 784 bad_area:
 785         up_read(&mm->mmap_sem);
 786
 787 bad_area_nosemaphore:
 788         /* User mode accesses just cause a SIGSEGV */
 789         if (error_code & PF_USER) {
 790                 /*
 791                  * It's possible to have interrupts off here.
 792                  */
 793                 local_irq_enable();
 794
 795                 /*
 796                  * Valid to do another page fault here because this one came
 797                  * from user space.
 798                  */
 799                 if (is_prefetch(regs, address, error_code))
 800                         return;
 801
 802                 if (is_errata100(regs, address))
 803                         return;
 804
 805                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 806                     printk_ratelimit()) {
 807                         printk(
 808                         "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 809                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 810                         tsk->comm, task_pid_nr(tsk), address,
 811                         (void *) regs->ip, (void *) regs->sp, error_code);
 812                         print_vma_addr(" in ", regs->ip);
 813                         printk("\n");
 814                 }
 815
 816                 tsk->thread.cr2 = address;
 817                 /* Kernel addresses are always protection faults */
 818                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 819                 tsk->thread.trap_no = 14;
 820                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 821                 return;
 822         }
 823
 824         if (is_f00f_bug(regs, address))
 825                 return;
 826
 827 no_context:
 828         /* Are we prepared to handle this kernel fault?  */
 829         if (fixup_exception(regs))
 830                 return;
 831
 832         /*
 833          * X86_32
 834          * Valid to do another page fault here, because if this fault
 835          * had been triggered by is_prefetch fixup_exception would have
 836          * handled it.
 837          *
 838          * X86_64
 839          * Hall of shame of CPU/BIOS bugs.
 840          */
 841         if (is_prefetch(regs, address, error_code))
 842                 return;
 843
 844         if (is_errata93(regs, address))
 845                 return;
 846
 847 /*
 848  * Oops. The kernel tried to access some bad page. We'll have to
 849  * terminate things with extreme prejudice.
 850  */
 851 #ifdef CONFIG_X86_32
 852         bust_spinlocks(1);
 853 #else
 854         flags = oops_begin();
 855 #endif
 856
 857         show_fault_oops(regs, error_code, address);
 858
 859         tsk->thread.cr2 = address;
 860         tsk->thread.trap_no = 14;
 861         tsk->thread.error_code = error_code;
 862
 863 #ifdef CONFIG_X86_32
 864         die("Oops", regs, error_code);
 865         bust_spinlocks(0);
 866         do_exit(SIGKILL);
 867 #else
 868         sig = SIGKILL;
 869         if (__die("Oops", regs, error_code))
 870                 sig = 0;
 871         /* Executive summary in case the body of the oops scrolled away */
 872         printk(KERN_EMERG "CR2: %016lx\n", address);
 873         oops_end(flags, regs, sig);
 874 #endif
 875
 876 out_of_memory:
 877         /*
 878          * We ran out of memory, call the OOM killer, and return the userspace
 879          * (which will retry the fault, or kill us if we got oom-killed).
 880          */
 881         up_read(&mm->mmap_sem);
 882         pagefault_out_of_memory();
 883         return;
 884
 885 do_sigbus:
 886         up_read(&mm->mmap_sem);
 887
 888         /* Kernel mode? Handle exceptions or die */
 889         if (!(error_code & PF_USER))
 890                 goto no_context;
 891 #ifdef CONFIG_X86_32
 892         /* User space => ok to do another page fault */
 893         if (is_prefetch(regs, address, error_code))
 894                 return;
 895 #endif
 896         tsk->thread.cr2 = address;
 897         tsk->thread.error_code = error_code;
 898         tsk->thread.trap_no = 14;
 899         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 900 }
 901
 902 DEFINE_SPINLOCK(pgd_lock);
 903 LIST_HEAD(pgd_list);
 904
 905 void vmalloc_sync_all(void)
 906 {
 907         unsigned long address;
 908
 909 #ifdef CONFIG_X86_32
 910         if (SHARED_KERNEL_PMD)
 911                 return;
 912
 913         for (address = VMALLOC_START & PMD_MASK;
 914              address >= TASK_SIZE && address < FIXADDR_TOP;
 915              address += PMD_SIZE) {
 916                 unsigned long flags;
 917                 struct page *page;
 918
 919                 spin_lock_irqsave(&pgd_lock, flags);
 920                 list_for_each_entry(page, &pgd_list, lru) {
 921                         if (!vmalloc_sync_one(page_address(page),
 922                                               address))
 923                                 break;
 924                 }
 925                 spin_unlock_irqrestore(&pgd_lock, flags);
 926         }
 927 #else /* CONFIG_X86_64 */
 928         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 929              address += PGDIR_SIZE) {
 930                 const pgd_t *pgd_ref = pgd_offset_k(address);
 931                 unsigned long flags;
 932                 struct page *page;
 933
 934                 if (pgd_none(*pgd_ref))
 935                         continue;
 936                 spin_lock_irqsave(&pgd_lock, flags);
 937                 list_for_each_entry(page, &pgd_list, lru) {
 938                         pgd_t *pgd;
 939                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 940                         if (pgd_none(*pgd))
 941                                 set_pgd(pgd, *pgd_ref);
 942                         else
 943                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 944                 }
 945                 spin_unlock_irqrestore(&pgd_lock, flags);
 946         }
 947 #endif
 948 }