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)) goto ret;
 244
 245         pmd = pmd_offset(pud, address);
 246         if (bad_address(pmd)) goto bad;
 247         printk("PMD %lx ", pmd_val(*pmd));
 248         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 249
 250         pte = pte_offset_kernel(pmd, address);
 251         if (bad_address(pte)) goto bad;
 252         printk("PTE %lx", pte_val(*pte));
 253 ret:
 254         printk("\n");
 255         return;
 256 bad:
 257         printk("BAD\n");
 258 #endif
 259 }
 260
 261 #ifdef CONFIG_X86_32
 262 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 263 {
 264         unsigned index = pgd_index(address);
 265         pgd_t *pgd_k;
 266         pud_t *pud, *pud_k;
 267         pmd_t *pmd, *pmd_k;
 268
 269         pgd += index;
 270         pgd_k = init_mm.pgd + index;
 271
 272         if (!pgd_present(*pgd_k))
 273                 return NULL;
 274
 275         /*
 276          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 277          * and redundant with the set_pmd() on non-PAE. As would
 278          * set_pud.
 279          */
 280
 281         pud = pud_offset(pgd, address);
 282         pud_k = pud_offset(pgd_k, address);
 283         if (!pud_present(*pud_k))
 284                 return NULL;
 285
 286         pmd = pmd_offset(pud, address);
 287         pmd_k = pmd_offset(pud_k, address);
 288         if (!pmd_present(*pmd_k))
 289                 return NULL;
 290         if (!pmd_present(*pmd)) {
 291                 set_pmd(pmd, *pmd_k);
 292                 arch_flush_lazy_mmu_mode();
 293         } else
 294                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 295         return pmd_k;
 296 }
 297 #endif
 298
 299 #ifdef CONFIG_X86_64
 300 static const char errata93_warning[] =
 301 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 302 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 303 KERN_ERR "******* Please consider a BIOS update.\n"
 304 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 305 #endif
 306
 307 /* Workaround for K8 erratum #93 & buggy BIOS.
 308    BIOS SMM functions are required to use a specific workaround
 309    to avoid corruption of the 64bit RIP register on C stepping K8.
 310    A lot of BIOS that didn't get tested properly miss this.
 311    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 312    Try to work around it here.
 313    Note we only handle faults in kernel here.
 314    Does nothing for X86_32
 315  */
 316 static int is_errata93(struct pt_regs *regs, unsigned long address)
 317 {
 318 #ifdef CONFIG_X86_64
 319         static int warned;
 320         if (address != regs->ip)
 321                 return 0;
 322         if ((address >> 32) != 0)
 323                 return 0;
 324         address |= 0xffffffffUL << 32;
 325         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 326             (address >= MODULES_VADDR && address <= MODULES_END)) {
 327                 if (!warned) {
 328                         printk(errata93_warning);
 329                         warned = 1;
 330                 }
 331                 regs->ip = address;
 332                 return 1;
 333         }
 334 #endif
 335         return 0;
 336 }
 337
 338 /*
 339  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 340  * addresses >4GB.  We catch this in the page fault handler because these
 341  * addresses are not reachable. Just detect this case and return.  Any code
 342  * segment in LDT is compatibility mode.
 343  */
 344 static int is_errata100(struct pt_regs *regs, unsigned long address)
 345 {
 346 #ifdef CONFIG_X86_64
 347         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 348             (address >> 32))
 349                 return 1;
 350 #endif
 351         return 0;
 352 }
 353
 354 void do_invalid_op(struct pt_regs *, unsigned long);
 355
 356 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 357 {
 358 #ifdef CONFIG_X86_F00F_BUG
 359         unsigned long nr;
 360         /*
 361          * Pentium F0 0F C7 C8 bug workaround.
 362          */
 363         if (boot_cpu_data.f00f_bug) {
 364                 nr = (address - idt_descr.address) >> 3;
 365
 366                 if (nr == 6) {
 367                         do_invalid_op(regs, 0);
 368                         return 1;
 369                 }
 370         }
 371 #endif
 372         return 0;
 373 }
 374
 375 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 376                             unsigned long address)
 377 {
 378 #ifdef CONFIG_X86_32
 379         if (!oops_may_print())
 380                 return;
 381
 382 #ifdef CONFIG_X86_PAE
 383         if (error_code & PF_INSTR) {
 384                 int level;
 385                 pte_t *pte = lookup_address(address, &level);
 386
 387                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 388                         printk(KERN_CRIT "kernel tried to execute "
 389                                 "NX-protected page - exploit attempt? "
 390                                 "(uid: %d)\n", current->uid);
 391         }
 392 #endif
 393         printk(KERN_ALERT "BUG: unable to handle kernel ");
 394         if (address < PAGE_SIZE)
 395                 printk(KERN_CONT "NULL pointer dereference");
 396         else
 397                 printk(KERN_CONT "paging request");
 398         printk(KERN_CONT " at %08lx\n", address);
 399
 400         printk(KERN_ALERT "IP:");
 401         printk_address(regs->ip, 1);
 402         dump_pagetable(address);
 403 #else /* CONFIG_X86_64 */
 404         printk(KERN_ALERT "BUG: unable to handle kernel ");
 405         if (address < PAGE_SIZE)
 406                 printk(KERN_CONT "NULL pointer dereference");
 407         else
 408                 printk(KERN_CONT "paging request");
 409         printk(KERN_CONT " at %016lx\n", address);
 410
 411         printk(KERN_ALERT "IP:");
 412         printk_address(regs->ip, 1);
 413         dump_pagetable(address);
 414 #endif
 415 }
 416
 417 #ifdef CONFIG_X86_64
 418 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 419                                  unsigned long error_code)
 420 {
 421         unsigned long flags = oops_begin();
 422         struct task_struct *tsk;
 423
 424         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 425                current->comm, address);
 426         dump_pagetable(address);
 427         tsk = current;
 428         tsk->thread.cr2 = address;
 429         tsk->thread.trap_no = 14;
 430         tsk->thread.error_code = error_code;
 431         if (__die("Bad pagetable", regs, error_code))
 432                 regs = NULL;
 433         oops_end(flags, regs, SIGKILL);
 434 }
 435 #endif
 436
 437 /*
 438  * X86_32
 439  * Handle a fault on the vmalloc or module mapping area
 440  *
 441  * X86_64
 442  * Handle a fault on the vmalloc area
 443  *
 444  * This assumes no large pages in there.
 445  */
 446 static int vmalloc_fault(unsigned long address)
 447 {
 448 #ifdef CONFIG_X86_32
 449         unsigned long pgd_paddr;
 450         pmd_t *pmd_k;
 451         pte_t *pte_k;
 452         /*
 453          * Synchronize this task's top level page-table
 454          * with the 'reference' page table.
 455          *
 456          * Do _not_ use "current" here. We might be inside
 457          * an interrupt in the middle of a task switch..
 458          */
 459         pgd_paddr = read_cr3();
 460         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 461         if (!pmd_k)
 462                 return -1;
 463         pte_k = pte_offset_kernel(pmd_k, address);
 464         if (!pte_present(*pte_k))
 465                 return -1;
 466         return 0;
 467 #else
 468         pgd_t *pgd, *pgd_ref;
 469         pud_t *pud, *pud_ref;
 470         pmd_t *pmd, *pmd_ref;
 471         pte_t *pte, *pte_ref;
 472
 473         /* Copy kernel mappings over when needed. This can also
 474            happen within a race in page table update. In the later
 475            case just flush. */
 476
 477         pgd = pgd_offset(current->mm ?: &init_mm, address);
 478         pgd_ref = pgd_offset_k(address);
 479         if (pgd_none(*pgd_ref))
 480                 return -1;
 481         if (pgd_none(*pgd))
 482                 set_pgd(pgd, *pgd_ref);
 483         else
 484                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 485
 486         /* Below here mismatches are bugs because these lower tables
 487            are shared */
 488
 489         pud = pud_offset(pgd, address);
 490         pud_ref = pud_offset(pgd_ref, address);
 491         if (pud_none(*pud_ref))
 492                 return -1;
 493         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 494                 BUG();
 495         pmd = pmd_offset(pud, address);
 496         pmd_ref = pmd_offset(pud_ref, address);
 497         if (pmd_none(*pmd_ref))
 498                 return -1;
 499         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 500                 BUG();
 501         pte_ref = pte_offset_kernel(pmd_ref, address);
 502         if (!pte_present(*pte_ref))
 503                 return -1;
 504         pte = pte_offset_kernel(pmd, address);
 505         /* Don't use pte_page here, because the mappings can point
 506            outside mem_map, and the NUMA hash lookup cannot handle
 507            that. */
 508         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 509                 BUG();
 510         return 0;
 511 #endif
 512 }
 513
 514 int show_unhandled_signals = 1;
 515
 516 /*
 517  * This routine handles page faults.  It determines the address,
 518  * and the problem, and then passes it off to one of the appropriate
 519  * routines.
 520  */
 521 #ifdef CONFIG_X86_64
 522 asmlinkage
 523 #endif
 524 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 525 {
 526         struct task_struct *tsk;
 527         struct mm_struct *mm;
 528         struct vm_area_struct *vma;
 529         unsigned long address;
 530         int write, si_code;
 531         int fault;
 532 #ifdef CONFIG_X86_64
 533         unsigned long flags;
 534 #endif
 535
 536         /*
 537          * We can fault from pretty much anywhere, with unknown IRQ state.
 538          */
 539         trace_hardirqs_fixup();
 540
 541         tsk = current;
 542         mm = tsk->mm;
 543         prefetchw(&mm->mmap_sem);
 544
 545         /* get the address */
 546         address = read_cr2();
 547
 548         si_code = SEGV_MAPERR;
 549
 550         if (notify_page_fault(regs))
 551                 return;
 552
 553         /*
 554          * We fault-in kernel-space virtual memory on-demand. The
 555          * 'reference' page table is init_mm.pgd.
 556          *
 557          * NOTE! We MUST NOT take any locks for this case. We may
 558          * be in an interrupt or a critical region, and should
 559          * only copy the information from the master page table,
 560          * nothing more.
 561          *
 562          * This verifies that the fault happens in kernel space
 563          * (error_code & 4) == 0, and that the fault was not a
 564          * protection error (error_code & 9) == 0.
 565          */
 566 #ifdef CONFIG_X86_32
 567         if (unlikely(address >= TASK_SIZE)) {
 568                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 569                     vmalloc_fault(address) >= 0)
 570                         return;
 571                 /*
 572                  * Don't take the mm semaphore here. If we fixup a prefetch
 573                  * fault we could otherwise deadlock.
 574                  */
 575                 goto bad_area_nosemaphore;
 576         }
 577
 578         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 579            fault has been handled. */
 580         if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
 581                 local_irq_enable();
 582
 583         /*
 584          * If we're in an interrupt, have no user context or are running in an
 585          * atomic region then we must not take the fault.
 586          */
 587         if (in_atomic() || !mm)
 588                 goto bad_area_nosemaphore;
 589 #else /* CONFIG_X86_64 */
 590         if (unlikely(address >= TASK_SIZE64)) {
 591                 /*
 592                  * Don't check for the module range here: its PML4
 593                  * is always initialized because it's shared with the main
 594                  * kernel text. Only vmalloc may need PML4 syncups.
 595                  */
 596                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 597                       ((address >= VMALLOC_START && address < VMALLOC_END))) {
 598                         if (vmalloc_fault(address) >= 0)
 599                                 return;
 600                 }
 601                 /*
 602                  * Don't take the mm semaphore here. If we fixup a prefetch
 603                  * fault we could otherwise deadlock.
 604                  */
 605                 goto bad_area_nosemaphore;
 606         }
 607         if (likely(regs->flags & X86_EFLAGS_IF))
 608                 local_irq_enable();
 609
 610         if (unlikely(error_code & PF_RSVD))
 611                 pgtable_bad(address, regs, error_code);
 612
 613         /*
 614          * If we're in an interrupt, have no user context or are running in an
 615          * atomic region then we must not take the fault.
 616          */
 617         if (unlikely(in_atomic() || !mm))
 618                 goto bad_area_nosemaphore;
 619
 620         /*
 621          * User-mode registers count as a user access even for any
 622          * potential system fault or CPU buglet.
 623          */
 624         if (user_mode_vm(regs))
 625                 error_code |= PF_USER;
 626 again:
 627 #endif
 628         /* When running in the kernel we expect faults to occur only to
 629          * addresses in user space.  All other faults represent errors in the
 630          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 631          * erroneous fault occurring in a code path which already holds mmap_sem
 632          * we will deadlock attempting to validate the fault against the
 633          * address space.  Luckily the kernel only validly references user
 634          * space from well defined areas of code, which are listed in the
 635          * exceptions table.
 636          *
 637          * As the vast majority of faults will be valid we will only perform
 638          * the source reference check when there is a possibility of a deadlock.
 639          * Attempt to lock the address space, if we cannot we then validate the
 640          * source.  If this is invalid we can skip the address space check,
 641          * thus avoiding the deadlock.
 642          */
 643         if (!down_read_trylock(&mm->mmap_sem)) {
 644                 if ((error_code & PF_USER) == 0 &&
 645                     !search_exception_tables(regs->ip))
 646                         goto bad_area_nosemaphore;
 647                 down_read(&mm->mmap_sem);
 648         }
 649
 650         vma = find_vma(mm, address);
 651         if (!vma)
 652                 goto bad_area;
 653 #ifdef CONFIG_X86_32
 654         if (vma->vm_start <= address)
 655 #else
 656         if (likely(vma->vm_start <= address))
 657 #endif
 658                 goto good_area;
 659         if (!(vma->vm_flags & VM_GROWSDOWN))
 660                 goto bad_area;
 661         if (error_code & PF_USER) {
 662                 /*
 663                  * Accessing the stack below %sp is always a bug.
 664                  * The large cushion allows instructions like enter
 665                  * and pusha to work.  ("enter $65535,$31" pushes
 666                  * 32 pointers and then decrements %sp by 65535.)
 667                  */
 668                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 669                         goto bad_area;
 670         }
 671         if (expand_stack(vma, address))
 672                 goto bad_area;
 673 /*
 674  * Ok, we have a good vm_area for this memory access, so
 675  * we can handle it..
 676  */
 677 good_area:
 678         si_code = SEGV_ACCERR;
 679         write = 0;
 680         switch (error_code & (PF_PROT|PF_WRITE)) {
 681         default:        /* 3: write, present */
 682                 /* fall through */
 683         case PF_WRITE:          /* write, not present */
 684                 if (!(vma->vm_flags & VM_WRITE))
 685                         goto bad_area;
 686                 write++;
 687                 break;
 688         case PF_PROT:           /* read, present */
 689                 goto bad_area;
 690         case 0:                 /* read, not present */
 691                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 692                         goto bad_area;
 693         }
 694
 695 #ifdef CONFIG_X86_32
 696 survive:
 697 #endif
 698         /*
 699          * If for any reason at all we couldn't handle the fault,
 700          * make sure we exit gracefully rather than endlessly redo
 701          * the fault.
 702          */
 703         fault = handle_mm_fault(mm, vma, address, write);
 704         if (unlikely(fault & VM_FAULT_ERROR)) {
 705                 if (fault & VM_FAULT_OOM)
 706                         goto out_of_memory;
 707                 else if (fault & VM_FAULT_SIGBUS)
 708                         goto do_sigbus;
 709                 BUG();
 710         }
 711         if (fault & VM_FAULT_MAJOR)
 712                 tsk->maj_flt++;
 713         else
 714                 tsk->min_flt++;
 715
 716 #ifdef CONFIG_X86_32
 717         /*
 718          * Did it hit the DOS screen memory VA from vm86 mode?
 719          */
 720         if (v8086_mode(regs)) {
 721                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 722                 if (bit < 32)
 723                         tsk->thread.screen_bitmap |= 1 << bit;
 724         }
 725 #endif
 726         up_read(&mm->mmap_sem);
 727         return;
 728
 729 /*
 730  * Something tried to access memory that isn't in our memory map..
 731  * Fix it, but check if it's kernel or user first..
 732  */
 733 bad_area:
 734         up_read(&mm->mmap_sem);
 735
 736 bad_area_nosemaphore:
 737         /* User mode accesses just cause a SIGSEGV */
 738         if (error_code & PF_USER) {
 739                 /*
 740                  * It's possible to have interrupts off here.
 741                  */
 742                 local_irq_enable();
 743
 744                 /*
 745                  * Valid to do another page fault here because this one came
 746                  * from user space.
 747                  */
 748                 if (is_prefetch(regs, address, error_code))
 749                         return;
 750
 751                 if (is_errata100(regs, address))
 752                         return;
 753
 754                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 755                     printk_ratelimit()) {
 756                         printk(
 757 #ifdef CONFIG_X86_32
 758                         "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
 759 #else
 760                         "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
 761 #endif
 762                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 763                         tsk->comm, task_pid_nr(tsk), address, regs->ip,
 764                         regs->sp, error_code);
 765                         print_vma_addr(" in ", regs->ip);
 766                         printk("\n");
 767                 }
 768
 769                 tsk->thread.cr2 = address;
 770                 /* Kernel addresses are always protection faults */
 771                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 772                 tsk->thread.trap_no = 14;
 773                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 774                 return;
 775         }
 776
 777         if (is_f00f_bug(regs, address))
 778                 return;
 779
 780 no_context:
 781         /* Are we prepared to handle this kernel fault?  */
 782         if (fixup_exception(regs))
 783                 return;
 784
 785         /*
 786          * X86_32
 787          * Valid to do another page fault here, because if this fault
 788          * had been triggered by is_prefetch fixup_exception would have
 789          * handled it.
 790          *
 791          * X86_64
 792          * Hall of shame of CPU/BIOS bugs.
 793          */
 794         if (is_prefetch(regs, address, error_code))
 795                 return;
 796
 797         if (is_errata93(regs, address))
 798                 return;
 799
 800 /*
 801  * Oops. The kernel tried to access some bad page. We'll have to
 802  * terminate things with extreme prejudice.
 803  */
 804 #ifdef CONFIG_X86_32
 805         bust_spinlocks(1);
 806
 807         show_fault_oops(regs, error_code, address);
 808
 809         tsk->thread.cr2 = address;
 810         tsk->thread.trap_no = 14;
 811         tsk->thread.error_code = error_code;
 812         die("Oops", regs, error_code);
 813         bust_spinlocks(0);
 814         do_exit(SIGKILL);
 815 #else /* CONFIG_X86_64 */
 816         flags = oops_begin();
 817
 818         show_fault_oops(regs, error_code, address);
 819
 820         tsk->thread.cr2 = address;
 821         tsk->thread.trap_no = 14;
 822         tsk->thread.error_code = error_code;
 823         if (__die("Oops", regs, error_code))
 824                 regs = NULL;
 825         /* Executive summary in case the body of the oops scrolled away */
 826         printk(KERN_EMERG "CR2: %016lx\n", address);
 827         oops_end(flags, regs, SIGKILL);
 828 #endif
 829
 830 /*
 831  * We ran out of memory, or some other thing happened to us that made
 832  * us unable to handle the page fault gracefully.
 833  */
 834 out_of_memory:
 835         up_read(&mm->mmap_sem);
 836 #ifdef CONFIG_X86_32
 837         if (is_global_init(tsk)) {
 838                 yield();
 839                 down_read(&mm->mmap_sem);
 840                 goto survive;
 841         }
 842 #else
 843         if (is_global_init(current)) {
 844                 yield();
 845                 goto again;
 846         }
 847 #endif
 848         printk("VM: killing process %s\n", tsk->comm);
 849         if (error_code & PF_USER)
 850                 do_group_exit(SIGKILL);
 851         goto no_context;
 852
 853 do_sigbus:
 854         up_read(&mm->mmap_sem);
 855
 856         /* Kernel mode? Handle exceptions or die */
 857         if (!(error_code & PF_USER))
 858                 goto no_context;
 859 #ifdef CONFIG_X86_32
 860         /* User space => ok to do another page fault */
 861         if (is_prefetch(regs, address, error_code))
 862                 return;
 863 #endif
 864         tsk->thread.cr2 = address;
 865         tsk->thread.error_code = error_code;
 866         tsk->thread.trap_no = 14;
 867         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 868 }
 869
 870 #ifdef CONFIG_X86_64
 871 DEFINE_SPINLOCK(pgd_lock);
 872 LIST_HEAD(pgd_list);
 873 #endif
 874
 875 void vmalloc_sync_all(void)
 876 {
 877 #ifdef CONFIG_X86_32
 878         /*
 879          * Note that races in the updates of insync and start aren't
 880          * problematic: insync can only get set bits added, and updates to
 881          * start are only improving performance (without affecting correctness
 882          * if undone).
 883          */
 884         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 885         static unsigned long start = TASK_SIZE;
 886         unsigned long address;
 887
 888         if (SHARED_KERNEL_PMD)
 889                 return;
 890
 891         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 892         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 893                 if (!test_bit(pgd_index(address), insync)) {
 894                         unsigned long flags;
 895                         struct page *page;
 896
 897                         spin_lock_irqsave(&pgd_lock, flags);
 898                         for (page = pgd_list; page; page =
 899                                         (struct page *)page->index)
 900                                 if (!vmalloc_sync_one(page_address(page),
 901                                                                 address)) {
 902                                         BUG_ON(page != pgd_list);
 903                                         break;
 904                                 }
 905                         spin_unlock_irqrestore(&pgd_lock, flags);
 906                         if (!page)
 907                                 set_bit(pgd_index(address), insync);
 908                 }
 909                 if (address == start && test_bit(pgd_index(address), insync))
 910                         start = address + PGDIR_SIZE;
 911         }
 912 #else /* CONFIG_X86_64 */
 913         /*
 914          * Note that races in the updates of insync and start aren't
 915          * problematic: insync can only get set bits added, and updates to
 916          * start are only improving performance (without affecting correctness
 917          * if undone).
 918          */
 919         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 920         static unsigned long start = VMALLOC_START & PGDIR_MASK;
 921         unsigned long address;
 922
 923         for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
 924                 if (!test_bit(pgd_index(address), insync)) {
 925                         const pgd_t *pgd_ref = pgd_offset_k(address);
 926                         struct page *page;
 927
 928                         if (pgd_none(*pgd_ref))
 929                                 continue;
 930                         spin_lock(&pgd_lock);
 931                         list_for_each_entry(page, &pgd_list, lru) {
 932                                 pgd_t *pgd;
 933                                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 934                                 if (pgd_none(*pgd))
 935                                         set_pgd(pgd, *pgd_ref);
 936                                 else
 937                                         BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 938                         }
 939                         spin_unlock(&pgd_lock);
 940                         set_bit(pgd_index(address), insync);
 941                 }
 942                 if (address == start)
 943                         start = address + PGDIR_SIZE;
 944         }
 945         /* Check that there is no need to do the same for the modules area. */
 946         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 947         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 948                                 (__START_KERNEL & PGDIR_MASK)));
 949 #endif
 950 }