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/mman.h>
  14 #include <linux/mm.h>
  15 #include <linux/smp.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/init.h>
  18 #include <linux/tty.h>
  19 #include <linux/vt_kern.h>              /* For unblank_screen() */
  20 #include <linux/compiler.h>
  21 #include <linux/highmem.h>
  22 #include <linux/bootmem.h>              /* for max_low_pfn */
  23 #include <linux/vmalloc.h>
  24 #include <linux/module.h>
  25 #include <linux/kprobes.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/kdebug.h>
  28
  29 #include <asm/system.h>
  30 #include <asm/desc.h>
  31 #include <asm/segment.h>
  32 #include <asm/pgalloc.h>
  33 #include <asm/smp.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/proto.h>
  36 #include <asm-generic/sections.h>
  37
  38 /*
  39  * Page fault error code bits
  40  *      bit 0 == 0 means no page found, 1 means protection fault
  41  *      bit 1 == 0 means read, 1 means write
  42  *      bit 2 == 0 means kernel, 1 means user-mode
  43  *      bit 3 == 1 means use of reserved bit detected
  44  *      bit 4 == 1 means fault was an instruction fetch
  45  */
  46 #define PF_PROT         (1<<0)
  47 #define PF_WRITE        (1<<1)
  48 #define PF_USER         (1<<2)
  49 #define PF_RSVD         (1<<3)
  50 #define PF_INSTR        (1<<4)
  51
  52 static inline int notify_page_fault(struct pt_regs *regs)
  53 {
  54 #ifdef CONFIG_KPROBES
  55         int ret = 0;
  56
  57         /* kprobe_running() needs smp_processor_id() */
  58 #ifdef CONFIG_X86_32
  59         if (!user_mode_vm(regs)) {
  60 #else
  61         if (!user_mode(regs)) {
  62 #endif
  63                 preempt_disable();
  64                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  65                         ret = 1;
  66                 preempt_enable();
  67         }
  68
  69         return ret;
  70 #else
  71         return 0;
  72 #endif
  73 }
  74
  75 /*
  76  * X86_32
  77  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  78  * Check that here and ignore it.
  79  *
  80  * X86_64
  81  * Sometimes the CPU reports invalid exceptions on prefetch.
  82  * Check that here and ignore it.
  83  *
  84  * Opcode checker based on code by Richard Brunner
  85  */
  86 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
  87                        unsigned long error_code)
  88 {
  89         unsigned char *instr;
  90         int scan_more = 1;
  91         int prefetch = 0;
  92         unsigned char *max_instr;
  93
  94 #ifdef CONFIG_X86_32
  95         if (!(__supported_pte_mask & _PAGE_NX))
  96                 return 0;
  97 #endif
  98
  99         /* If it was a exec fault on NX page, ignore */
 100         if (error_code & PF_INSTR)
 101                 return 0;
 102
 103         instr = (unsigned char *)convert_ip_to_linear(current, regs);
 104         max_instr = instr + 15;
 105
 106         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 107                 return 0;
 108
 109         while (scan_more && instr < max_instr) {
 110                 unsigned char opcode;
 111                 unsigned char instr_hi;
 112                 unsigned char instr_lo;
 113
 114                 if (probe_kernel_address(instr, opcode))
 115                         break;
 116
 117                 instr_hi = opcode & 0xf0;
 118                 instr_lo = opcode & 0x0f;
 119                 instr++;
 120
 121                 switch (instr_hi) {
 122                 case 0x20:
 123                 case 0x30:
 124                         /*
 125                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 126                          * In X86_64 long mode, the CPU will signal invalid
 127                          * opcode if some of these prefixes are present so
 128                          * X86_64 will never get here anyway
 129                          */
 130                         scan_more = ((instr_lo & 7) == 0x6);
 131                         break;
 132 #ifdef CONFIG_X86_64
 133                 case 0x40:
 134                         /*
 135                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 136                          * Need to figure out under what instruction mode the
 137                          * instruction was issued. Could check the LDT for lm,
 138                          * but for now it's good enough to assume that long
 139                          * mode only uses well known segments or kernel.
 140                          */
 141                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 142                         break;
 143 #endif
 144                 case 0x60:
 145                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 146                         scan_more = (instr_lo & 0xC) == 0x4;
 147                         break;
 148                 case 0xF0:
 149                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 150                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 151                         break;
 152                 case 0x00:
 153                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 154                         scan_more = 0;
 155
 156                         if (probe_kernel_address(instr, opcode))
 157                                 break;
 158                         prefetch = (instr_lo == 0xF) &&
 159                                 (opcode == 0x0D || opcode == 0x18);
 160                         break;
 161                 default:
 162                         scan_more = 0;
 163                         break;
 164                 }
 165         }
 166         return prefetch;
 167 }
 168
 169 static void force_sig_info_fault(int si_signo, int si_code,
 170         unsigned long address, struct task_struct *tsk)
 171 {
 172         siginfo_t info;
 173
 174         info.si_signo = si_signo;
 175         info.si_errno = 0;
 176         info.si_code = si_code;
 177         info.si_addr = (void __user *)address;
 178         force_sig_info(si_signo, &info, tsk);
 179 }
 180
 181 #ifdef CONFIG_X86_64
 182 static int bad_address(void *p)
 183 {
 184         unsigned long dummy;
 185         return probe_kernel_address((unsigned long *)p, dummy);
 186 }
 187 #endif
 188
 189 void dump_pagetable(unsigned long address)
 190 {
 191 #ifdef CONFIG_X86_32
 192         __typeof__(pte_val(__pte(0))) page;
 193
 194         page = read_cr3();
 195         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 196 #ifdef CONFIG_X86_PAE
 197         printk("*pdpt = %016Lx ", page);
 198         if ((page >> PAGE_SHIFT) < max_low_pfn
 199             && page & _PAGE_PRESENT) {
 200                 page &= PAGE_MASK;
 201                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 202                                                          & (PTRS_PER_PMD - 1)];
 203                 printk(KERN_CONT "*pde = %016Lx ", page);
 204                 page &= ~_PAGE_NX;
 205         }
 206 #else
 207         printk("*pde = %08lx ", page);
 208 #endif
 209
 210         /*
 211          * We must not directly access the pte in the highpte
 212          * case if the page table is located in highmem.
 213          * And let's rather not kmap-atomic the pte, just in case
 214          * it's allocated already.
 215          */
 216         if ((page >> PAGE_SHIFT) < max_low_pfn
 217             && (page & _PAGE_PRESENT)
 218             && !(page & _PAGE_PSE)) {
 219                 page &= PAGE_MASK;
 220                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 221                                                          & (PTRS_PER_PTE - 1)];
 222                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 223         }
 224
 225         printk("\n");
 226 #else /* CONFIG_X86_64 */
 227         pgd_t *pgd;
 228         pud_t *pud;
 229         pmd_t *pmd;
 230         pte_t *pte;
 231
 232         pgd = (pgd_t *)read_cr3();
 233
 234         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 235         pgd += pgd_index(address);
 236         if (bad_address(pgd)) goto bad;
 237         printk("PGD %lx ", pgd_val(*pgd));
 238         if (!pgd_present(*pgd)) goto ret;
 239
 240         pud = pud_offset(pgd, address);
 241         if (bad_address(pud)) goto bad;
 242         printk("PUD %lx ", pud_val(*pud));
 243         if (!pud_present(*pud) || pud_large(*pud))
 244                 goto ret;
 245
 246         pmd = pmd_offset(pud, address);
 247         if (bad_address(pmd)) goto bad;
 248         printk("PMD %lx ", pmd_val(*pmd));
 249         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 250
 251         pte = pte_offset_kernel(pmd, address);
 252         if (bad_address(pte)) goto bad;
 253         printk("PTE %lx", pte_val(*pte));
 254 ret:
 255         printk("\n");
 256         return;
 257 bad:
 258         printk("BAD\n");
 259 #endif
 260 }
 261
 262 #ifdef CONFIG_X86_32
 263 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 264 {
 265         unsigned index = pgd_index(address);
 266         pgd_t *pgd_k;
 267         pud_t *pud, *pud_k;
 268         pmd_t *pmd, *pmd_k;
 269
 270         pgd += index;
 271         pgd_k = init_mm.pgd + index;
 272
 273         if (!pgd_present(*pgd_k))
 274                 return NULL;
 275
 276         /*
 277          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 278          * and redundant with the set_pmd() on non-PAE. As would
 279          * set_pud.
 280          */
 281
 282         pud = pud_offset(pgd, address);
 283         pud_k = pud_offset(pgd_k, address);
 284         if (!pud_present(*pud_k))
 285                 return NULL;
 286
 287         pmd = pmd_offset(pud, address);
 288         pmd_k = pmd_offset(pud_k, address);
 289         if (!pmd_present(*pmd_k))
 290                 return NULL;
 291         if (!pmd_present(*pmd)) {
 292                 set_pmd(pmd, *pmd_k);
 293                 arch_flush_lazy_mmu_mode();
 294         } else
 295                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 296         return pmd_k;
 297 }
 298 #endif
 299
 300 #ifdef CONFIG_X86_64
 301 static const char errata93_warning[] =
 302 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 303 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 304 KERN_ERR "******* Please consider a BIOS update.\n"
 305 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 306 #endif
 307
 308 /* Workaround for K8 erratum #93 & buggy BIOS.
 309    BIOS SMM functions are required to use a specific workaround
 310    to avoid corruption of the 64bit RIP register on C stepping K8.
 311    A lot of BIOS that didn't get tested properly miss this.
 312    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 313    Try to work around it here.
 314    Note we only handle faults in kernel here.
 315    Does nothing for X86_32
 316  */
 317 static int is_errata93(struct pt_regs *regs, unsigned long address)
 318 {
 319 #ifdef CONFIG_X86_64
 320         static int warned;
 321         if (address != regs->ip)
 322                 return 0;
 323         if ((address >> 32) != 0)
 324                 return 0;
 325         address |= 0xffffffffUL << 32;
 326         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 327             (address >= MODULES_VADDR && address <= MODULES_END)) {
 328                 if (!warned) {
 329                         printk(errata93_warning);
 330                         warned = 1;
 331                 }
 332                 regs->ip = address;
 333                 return 1;
 334         }
 335 #endif
 336         return 0;
 337 }
 338
 339 /*
 340  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 341  * addresses >4GB.  We catch this in the page fault handler because these
 342  * addresses are not reachable. Just detect this case and return.  Any code
 343  * segment in LDT is compatibility mode.
 344  */
 345 static int is_errata100(struct pt_regs *regs, unsigned long address)
 346 {
 347 #ifdef CONFIG_X86_64
 348         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 349             (address >> 32))
 350                 return 1;
 351 #endif
 352         return 0;
 353 }
 354
 355 void do_invalid_op(struct pt_regs *, unsigned long);
 356
 357 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 358 {
 359 #ifdef CONFIG_X86_F00F_BUG
 360         unsigned long nr;
 361         /*
 362          * Pentium F0 0F C7 C8 bug workaround.
 363          */
 364         if (boot_cpu_data.f00f_bug) {
 365                 nr = (address - idt_descr.address) >> 3;
 366
 367                 if (nr == 6) {
 368                         do_invalid_op(regs, 0);
 369                         return 1;
 370                 }
 371         }
 372 #endif
 373         return 0;
 374 }
 375
 376 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 377                             unsigned long address)
 378 {
 379 #ifdef CONFIG_X86_32
 380         if (!oops_may_print())
 381                 return;
 382 #endif
 383
 384 #ifdef CONFIG_X86_PAE
 385         if (error_code & PF_INSTR) {
 386                 unsigned int level;
 387                 pte_t *pte = lookup_address(address, &level);
 388
 389                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 390                         printk(KERN_CRIT "kernel tried to execute "
 391                                 "NX-protected page - exploit attempt? "
 392                                 "(uid: %d)\n", current->uid);
 393         }
 394 #endif
 395
 396         printk(KERN_ALERT "BUG: unable to handle kernel ");
 397         if (address < PAGE_SIZE)
 398                 printk(KERN_CONT "NULL pointer dereference");
 399         else
 400                 printk(KERN_CONT "paging request");
 401 #ifdef CONFIG_X86_32
 402         printk(KERN_CONT " at %08lx\n", address);
 403 #else
 404         printk(KERN_CONT " at %016lx\n", address);
 405 #endif
 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         struct task_struct *tsk;
 417
 418         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 419                current->comm, address);
 420         dump_pagetable(address);
 421         tsk = current;
 422         tsk->thread.cr2 = address;
 423         tsk->thread.trap_no = 14;
 424         tsk->thread.error_code = error_code;
 425         if (__die("Bad pagetable", regs, error_code))
 426                 regs = NULL;
 427         oops_end(flags, regs, SIGKILL);
 428 }
 429 #endif
 430
 431 /*
 432  * Handle a spurious fault caused by a stale TLB entry.  This allows
 433  * us to lazily refresh the TLB when increasing the permissions of a
 434  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
 435  * expensive since that implies doing a full cross-processor TLB
 436  * flush, even if no stale TLB entries exist on other processors.
 437  * There are no security implications to leaving a stale TLB when
 438  * increasing the permissions on a page.
 439  */
 440 static int spurious_fault(unsigned long address,
 441                           unsigned long error_code)
 442 {
 443         pgd_t *pgd;
 444         pud_t *pud;
 445         pmd_t *pmd;
 446         pte_t *pte;
 447
 448         /* Reserved-bit violation or user access to kernel space? */
 449         if (error_code & (PF_USER | PF_RSVD))
 450                 return 0;
 451
 452         pgd = init_mm.pgd + pgd_index(address);
 453         if (!pgd_present(*pgd))
 454                 return 0;
 455
 456         pud = pud_offset(pgd, address);
 457         if (!pud_present(*pud))
 458                 return 0;
 459
 460         pmd = pmd_offset(pud, address);
 461         if (!pmd_present(*pmd))
 462                 return 0;
 463
 464         pte = pte_offset_kernel(pmd, address);
 465         if (!pte_present(*pte))
 466                 return 0;
 467
 468         if ((error_code & PF_WRITE) && !pte_write(*pte))
 469                 return 0;
 470         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 471                 return 0;
 472
 473         return 1;
 474 }
 475
 476 /*
 477  * X86_32
 478  * Handle a fault on the vmalloc or module mapping area
 479  *
 480  * X86_64
 481  * Handle a fault on the vmalloc area
 482  *
 483  * This assumes no large pages in there.
 484  */
 485 static int vmalloc_fault(unsigned long address)
 486 {
 487 #ifdef CONFIG_X86_32
 488         unsigned long pgd_paddr;
 489         pmd_t *pmd_k;
 490         pte_t *pte_k;
 491         /*
 492          * Synchronize this task's top level page-table
 493          * with the 'reference' page table.
 494          *
 495          * Do _not_ use "current" here. We might be inside
 496          * an interrupt in the middle of a task switch..
 497          */
 498         pgd_paddr = read_cr3();
 499         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 500         if (!pmd_k)
 501                 return -1;
 502         pte_k = pte_offset_kernel(pmd_k, address);
 503         if (!pte_present(*pte_k))
 504                 return -1;
 505         return 0;
 506 #else
 507         pgd_t *pgd, *pgd_ref;
 508         pud_t *pud, *pud_ref;
 509         pmd_t *pmd, *pmd_ref;
 510         pte_t *pte, *pte_ref;
 511
 512         /* Make sure we are in vmalloc area */
 513         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 514                 return -1;
 515
 516         /* Copy kernel mappings over when needed. This can also
 517            happen within a race in page table update. In the later
 518            case just flush. */
 519
 520         pgd = pgd_offset(current->mm ?: &init_mm, address);
 521         pgd_ref = pgd_offset_k(address);
 522         if (pgd_none(*pgd_ref))
 523                 return -1;
 524         if (pgd_none(*pgd))
 525                 set_pgd(pgd, *pgd_ref);
 526         else
 527                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 528
 529         /* Below here mismatches are bugs because these lower tables
 530            are shared */
 531
 532         pud = pud_offset(pgd, address);
 533         pud_ref = pud_offset(pgd_ref, address);
 534         if (pud_none(*pud_ref))
 535                 return -1;
 536         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 537                 BUG();
 538         pmd = pmd_offset(pud, address);
 539         pmd_ref = pmd_offset(pud_ref, address);
 540         if (pmd_none(*pmd_ref))
 541                 return -1;
 542         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 543                 BUG();
 544         pte_ref = pte_offset_kernel(pmd_ref, address);
 545         if (!pte_present(*pte_ref))
 546                 return -1;
 547         pte = pte_offset_kernel(pmd, address);
 548         /* Don't use pte_page here, because the mappings can point
 549            outside mem_map, and the NUMA hash lookup cannot handle
 550            that. */
 551         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 552                 BUG();
 553         return 0;
 554 #endif
 555 }
 556
 557 int show_unhandled_signals = 1;
 558
 559 /*
 560  * This routine handles page faults.  It determines the address,
 561  * and the problem, and then passes it off to one of the appropriate
 562  * routines.
 563  */
 564 #ifdef CONFIG_X86_64
 565 asmlinkage
 566 #endif
 567 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 568 {
 569         struct task_struct *tsk;
 570         struct mm_struct *mm;
 571         struct vm_area_struct *vma;
 572         unsigned long address;
 573         int write, si_code;
 574         int fault;
 575 #ifdef CONFIG_X86_64
 576         unsigned long flags;
 577 #endif
 578
 579         /*
 580          * We can fault from pretty much anywhere, with unknown IRQ state.
 581          */
 582         trace_hardirqs_fixup();
 583
 584         tsk = current;
 585         mm = tsk->mm;
 586         prefetchw(&mm->mmap_sem);
 587
 588         /* get the address */
 589         address = read_cr2();
 590
 591         si_code = SEGV_MAPERR;
 592
 593         if (notify_page_fault(regs))
 594                 return;
 595
 596         /*
 597          * We fault-in kernel-space virtual memory on-demand. The
 598          * 'reference' page table is init_mm.pgd.
 599          *
 600          * NOTE! We MUST NOT take any locks for this case. We may
 601          * be in an interrupt or a critical region, and should
 602          * only copy the information from the master page table,
 603          * nothing more.
 604          *
 605          * This verifies that the fault happens in kernel space
 606          * (error_code & 4) == 0, and that the fault was not a
 607          * protection error (error_code & 9) == 0.
 608          */
 609 #ifdef CONFIG_X86_32
 610         if (unlikely(address >= TASK_SIZE)) {
 611 #else
 612         if (unlikely(address >= TASK_SIZE64)) {
 613 #endif
 614                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 615                     vmalloc_fault(address) >= 0)
 616                         return;
 617
 618                 /* Can handle a stale RO->RW TLB */
 619                 if (spurious_fault(address, error_code))
 620                         return;
 621
 622                 /*
 623                  * Don't take the mm semaphore here. If we fixup a prefetch
 624                  * fault we could otherwise deadlock.
 625                  */
 626                 goto bad_area_nosemaphore;
 627         }
 628
 629
 630 #ifdef CONFIG_X86_32
 631         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 632            fault has been handled. */
 633         if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
 634                 local_irq_enable();
 635
 636         /*
 637          * If we're in an interrupt, have no user context or are running in an
 638          * atomic region then we must not take the fault.
 639          */
 640         if (in_atomic() || !mm)
 641                 goto bad_area_nosemaphore;
 642 #else /* CONFIG_X86_64 */
 643         if (likely(regs->flags & X86_EFLAGS_IF))
 644                 local_irq_enable();
 645
 646         if (unlikely(error_code & PF_RSVD))
 647                 pgtable_bad(address, regs, error_code);
 648
 649         /*
 650          * If we're in an interrupt, have no user context or are running in an
 651          * atomic region then we must not take the fault.
 652          */
 653         if (unlikely(in_atomic() || !mm))
 654                 goto bad_area_nosemaphore;
 655
 656         /*
 657          * User-mode registers count as a user access even for any
 658          * potential system fault or CPU buglet.
 659          */
 660         if (user_mode_vm(regs))
 661                 error_code |= PF_USER;
 662 again:
 663 #endif
 664         /* When running in the kernel we expect faults to occur only to
 665          * addresses in user space.  All other faults represent errors in the
 666          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 667          * erroneous fault occurring in a code path which already holds mmap_sem
 668          * we will deadlock attempting to validate the fault against the
 669          * address space.  Luckily the kernel only validly references user
 670          * space from well defined areas of code, which are listed in the
 671          * exceptions table.
 672          *
 673          * As the vast majority of faults will be valid we will only perform
 674          * the source reference check when there is a possibility of a deadlock.
 675          * Attempt to lock the address space, if we cannot we then validate the
 676          * source.  If this is invalid we can skip the address space check,
 677          * thus avoiding the deadlock.
 678          */
 679         if (!down_read_trylock(&mm->mmap_sem)) {
 680                 if ((error_code & PF_USER) == 0 &&
 681                     !search_exception_tables(regs->ip))
 682                         goto bad_area_nosemaphore;
 683                 down_read(&mm->mmap_sem);
 684         }
 685
 686         vma = find_vma(mm, address);
 687         if (!vma)
 688                 goto bad_area;
 689         if (vma->vm_start <= address)
 690                 goto good_area;
 691         if (!(vma->vm_flags & VM_GROWSDOWN))
 692                 goto bad_area;
 693         if (error_code & PF_USER) {
 694                 /*
 695                  * Accessing the stack below %sp is always a bug.
 696                  * The large cushion allows instructions like enter
 697                  * and pusha to work.  ("enter $65535,$31" pushes
 698                  * 32 pointers and then decrements %sp by 65535.)
 699                  */
 700                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 701                         goto bad_area;
 702         }
 703         if (expand_stack(vma, address))
 704                 goto bad_area;
 705 /*
 706  * Ok, we have a good vm_area for this memory access, so
 707  * we can handle it..
 708  */
 709 good_area:
 710         si_code = SEGV_ACCERR;
 711         write = 0;
 712         switch (error_code & (PF_PROT|PF_WRITE)) {
 713         default:        /* 3: write, present */
 714                 /* fall through */
 715         case PF_WRITE:          /* write, not present */
 716                 if (!(vma->vm_flags & VM_WRITE))
 717                         goto bad_area;
 718                 write++;
 719                 break;
 720         case PF_PROT:           /* read, present */
 721                 goto bad_area;
 722         case 0:                 /* read, not present */
 723                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 724                         goto bad_area;
 725         }
 726
 727 #ifdef CONFIG_X86_32
 728 survive:
 729 #endif
 730         /*
 731          * If for any reason at all we couldn't handle the fault,
 732          * make sure we exit gracefully rather than endlessly redo
 733          * the fault.
 734          */
 735         fault = handle_mm_fault(mm, vma, address, write);
 736         if (unlikely(fault & VM_FAULT_ERROR)) {
 737                 if (fault & VM_FAULT_OOM)
 738                         goto out_of_memory;
 739                 else if (fault & VM_FAULT_SIGBUS)
 740                         goto do_sigbus;
 741                 BUG();
 742         }
 743         if (fault & VM_FAULT_MAJOR)
 744                 tsk->maj_flt++;
 745         else
 746                 tsk->min_flt++;
 747
 748 #ifdef CONFIG_X86_32
 749         /*
 750          * Did it hit the DOS screen memory VA from vm86 mode?
 751          */
 752         if (v8086_mode(regs)) {
 753                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 754                 if (bit < 32)
 755                         tsk->thread.screen_bitmap |= 1 << bit;
 756         }
 757 #endif
 758         up_read(&mm->mmap_sem);
 759         return;
 760
 761 /*
 762  * Something tried to access memory that isn't in our memory map..
 763  * Fix it, but check if it's kernel or user first..
 764  */
 765 bad_area:
 766         up_read(&mm->mmap_sem);
 767
 768 bad_area_nosemaphore:
 769         /* User mode accesses just cause a SIGSEGV */
 770         if (error_code & PF_USER) {
 771                 /*
 772                  * It's possible to have interrupts off here.
 773                  */
 774                 local_irq_enable();
 775
 776                 /*
 777                  * Valid to do another page fault here because this one came
 778                  * from user space.
 779                  */
 780                 if (is_prefetch(regs, address, error_code))
 781                         return;
 782
 783                 if (is_errata100(regs, address))
 784                         return;
 785
 786                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 787                     printk_ratelimit()) {
 788                         printk(
 789 #ifdef CONFIG_X86_32
 790                         "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
 791 #else
 792                         "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
 793 #endif
 794                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 795                         tsk->comm, task_pid_nr(tsk), address, regs->ip,
 796                         regs->sp, error_code);
 797                         print_vma_addr(" in ", regs->ip);
 798                         printk("\n");
 799                 }
 800
 801                 tsk->thread.cr2 = address;
 802                 /* Kernel addresses are always protection faults */
 803                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 804                 tsk->thread.trap_no = 14;
 805                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 806                 return;
 807         }
 808
 809         if (is_f00f_bug(regs, address))
 810                 return;
 811
 812 no_context:
 813         /* Are we prepared to handle this kernel fault?  */
 814         if (fixup_exception(regs))
 815                 return;
 816
 817         /*
 818          * X86_32
 819          * Valid to do another page fault here, because if this fault
 820          * had been triggered by is_prefetch fixup_exception would have
 821          * handled it.
 822          *
 823          * X86_64
 824          * Hall of shame of CPU/BIOS bugs.
 825          */
 826         if (is_prefetch(regs, address, error_code))
 827                 return;
 828
 829         if (is_errata93(regs, address))
 830                 return;
 831
 832 /*
 833  * Oops. The kernel tried to access some bad page. We'll have to
 834  * terminate things with extreme prejudice.
 835  */
 836 #ifdef CONFIG_X86_32
 837         bust_spinlocks(1);
 838 #else
 839         flags = oops_begin();
 840 #endif
 841
 842         show_fault_oops(regs, error_code, address);
 843
 844         tsk->thread.cr2 = address;
 845         tsk->thread.trap_no = 14;
 846         tsk->thread.error_code = error_code;
 847
 848 #ifdef CONFIG_X86_32
 849         die("Oops", regs, error_code);
 850         bust_spinlocks(0);
 851         do_exit(SIGKILL);
 852 #else
 853         if (__die("Oops", regs, error_code))
 854                 regs = NULL;
 855         /* Executive summary in case the body of the oops scrolled away */
 856         printk(KERN_EMERG "CR2: %016lx\n", address);
 857         oops_end(flags, regs, SIGKILL);
 858 #endif
 859
 860 /*
 861  * We ran out of memory, or some other thing happened to us that made
 862  * us unable to handle the page fault gracefully.
 863  */
 864 out_of_memory:
 865         up_read(&mm->mmap_sem);
 866         if (is_global_init(tsk)) {
 867                 yield();
 868 #ifdef CONFIG_X86_32
 869                 down_read(&mm->mmap_sem);
 870                 goto survive;
 871 #else
 872                 goto again;
 873 #endif
 874         }
 875
 876         printk("VM: killing process %s\n", tsk->comm);
 877         if (error_code & PF_USER)
 878                 do_group_exit(SIGKILL);
 879         goto no_context;
 880
 881 do_sigbus:
 882         up_read(&mm->mmap_sem);
 883
 884         /* Kernel mode? Handle exceptions or die */
 885         if (!(error_code & PF_USER))
 886                 goto no_context;
 887 #ifdef CONFIG_X86_32
 888         /* User space => ok to do another page fault */
 889         if (is_prefetch(regs, address, error_code))
 890                 return;
 891 #endif
 892         tsk->thread.cr2 = address;
 893         tsk->thread.error_code = error_code;
 894         tsk->thread.trap_no = 14;
 895         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 896 }
 897
 898 DEFINE_SPINLOCK(pgd_lock);
 899 LIST_HEAD(pgd_list);
 900
 901 void vmalloc_sync_all(void)
 902 {
 903 #ifdef CONFIG_X86_32
 904         /*
 905          * Note that races in the updates of insync and start aren't
 906          * problematic: insync can only get set bits added, and updates to
 907          * start are only improving performance (without affecting correctness
 908          * if undone).
 909          */
 910         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 911         static unsigned long start = TASK_SIZE;
 912         unsigned long address;
 913
 914         if (SHARED_KERNEL_PMD)
 915                 return;
 916
 917         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 918         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 919                 if (!test_bit(pgd_index(address), insync)) {
 920                         unsigned long flags;
 921                         struct page *page;
 922
 923                         spin_lock_irqsave(&pgd_lock, flags);
 924                         list_for_each_entry(page, &pgd_list, lru) {
 925                                 if (!vmalloc_sync_one(page_address(page),
 926                                                       address))
 927                                         break;
 928                         }
 929                         spin_unlock_irqrestore(&pgd_lock, flags);
 930                         if (!page)
 931                                 set_bit(pgd_index(address), insync);
 932                 }
 933                 if (address == start && test_bit(pgd_index(address), insync))
 934                         start = address + PGDIR_SIZE;
 935         }
 936 #else /* CONFIG_X86_64 */
 937         /*
 938          * Note that races in the updates of insync and start aren't
 939          * problematic: insync can only get set bits added, and updates to
 940          * start are only improving performance (without affecting correctness
 941          * if undone).
 942          */
 943         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 944         static unsigned long start = VMALLOC_START & PGDIR_MASK;
 945         unsigned long address;
 946
 947         for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
 948                 if (!test_bit(pgd_index(address), insync)) {
 949                         const pgd_t *pgd_ref = pgd_offset_k(address);
 950                         struct page *page;
 951
 952                         if (pgd_none(*pgd_ref))
 953                                 continue;
 954                         spin_lock(&pgd_lock);
 955                         list_for_each_entry(page, &pgd_list, lru) {
 956                                 pgd_t *pgd;
 957                                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 958                                 if (pgd_none(*pgd))
 959                                         set_pgd(pgd, *pgd_ref);
 960                                 else
 961                                         BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 962                         }
 963                         spin_unlock(&pgd_lock);
 964                         set_bit(pgd_index(address), insync);
 965                 }
 966                 if (address == start)
 967                         start = address + PGDIR_SIZE;
 968         }
 969         /* Check that there is no need to do the same for the modules area. */
 970         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 971         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 972                                 (__START_KERNEL & PGDIR_MASK)));
 973 #endif
 974 }