arch/x86/mm/fault_32.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  */
   4
   5 #include <linux/signal.h>
   6 #include <linux/sched.h>
   7 #include <linux/kernel.h>
   8 #include <linux/errno.h>
   9 #include <linux/string.h>
  10 #include <linux/types.h>
  11 #include <linux/ptrace.h>
  12 #include <linux/mman.h>
  13 #include <linux/mm.h>
  14 #include <linux/smp.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/init.h>
  17 #include <linux/tty.h>
  18 #include <linux/vt_kern.h>              /* For unblank_screen() */
  19 #include <linux/highmem.h>
  20 #include <linux/bootmem.h>              /* for max_low_pfn */
  21 #include <linux/vmalloc.h>
  22 #include <linux/module.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/kdebug.h>
  26
  27 #include <asm/system.h>
  28 #include <asm/desc.h>
  29 #include <asm/segment.h>
  30
  31 /*
  32  * Page fault error code bits
  33  *      bit 0 == 0 means no page found, 1 means protection fault
  34  *      bit 1 == 0 means read, 1 means write
  35  *      bit 2 == 0 means kernel, 1 means user-mode
  36  *      bit 3 == 1 means use of reserved bit detected
  37  *      bit 4 == 1 means fault was an instruction fetch
  38  */
  39 #define PF_PROT         (1<<0)
  40 #define PF_WRITE        (1<<1)
  41 #define PF_USER         (1<<2)
  42 #define PF_RSVD         (1<<3)
  43 #define PF_INSTR        (1<<4)
  44
  45 static inline int notify_page_fault(struct pt_regs *regs)
  46 {
  47 #ifdef CONFIG_KPROBES
  48         int ret = 0;
  49
  50         /* kprobe_running() needs smp_processor_id() */
  51         if (!user_mode_vm(regs)) {
  52                 preempt_disable();
  53                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  54                         ret = 1;
  55                 preempt_enable();
  56         }
  57
  58         return ret;
  59 #else
  60         return 0;
  61 #endif
  62 }
  63
  64 /*
  65  * X86_32
  66  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  67  * Check that here and ignore it.
  68  *
  69  * X86_64
  70  * Sometimes the CPU reports invalid exceptions on prefetch.
  71  * Check that here and ignore it.
  72  *
  73  * Opcode checker based on code by Richard Brunner
  74  */
  75 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
  76                        unsigned long error_code)
  77 {
  78         unsigned char *instr;
  79         int scan_more = 1;
  80         int prefetch = 0;
  81         unsigned char *max_instr;
  82
  83 #ifdef CONFIG_X86_32
  84         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
  85                      boot_cpu_data.x86 >= 6)) {
  86                 /* Catch an obscure case of prefetch inside an NX page. */
  87                 if (nx_enabled && (error_code & PF_INSTR))
  88                         return 0;
  89         } else {
  90                 return 0;
  91         }
  92 #else
  93         /* If it was a exec fault ignore */
  94         if (error_code & PF_INSTR)
  95                 return 0;
  96 #endif
  97
  98         instr = (unsigned char *)convert_ip_to_linear(current, regs);
  99         max_instr = instr + 15;
 100
 101         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 102                 return 0;
 103
 104         while (scan_more && instr < max_instr) {
 105                 unsigned char opcode;
 106                 unsigned char instr_hi;
 107                 unsigned char instr_lo;
 108
 109                 if (probe_kernel_address(instr, opcode))
 110                         break;
 111
 112                 instr_hi = opcode & 0xf0;
 113                 instr_lo = opcode & 0x0f;
 114                 instr++;
 115
 116                 switch (instr_hi) {
 117                 case 0x20:
 118                 case 0x30:
 119                         /*
 120                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 121                          * In X86_64 long mode, the CPU will signal invalid
 122                          * opcode if some of these prefixes are present so
 123                          * X86_64 will never get here anyway
 124                          */
 125                         scan_more = ((instr_lo & 7) == 0x6);
 126                         break;
 127 #ifdef CONFIG_X86_64
 128                 case 0x40:
 129                         /*
 130                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 131                          * Need to figure out under what instruction mode the
 132                          * instruction was issued. Could check the LDT for lm,
 133                          * but for now it's good enough to assume that long
 134                          * mode only uses well known segments or kernel.
 135                          */
 136                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 137                         break;
 138 #endif
 139                 case 0x60:
 140                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 141                         scan_more = (instr_lo & 0xC) == 0x4;
 142                         break;
 143                 case 0xF0:
 144                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 145                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 146                         break;
 147                 case 0x00:
 148                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 149                         scan_more = 0;
 150
 151                         if (probe_kernel_address(instr, opcode))
 152                                 break;
 153                         prefetch = (instr_lo == 0xF) &&
 154                                 (opcode == 0x0D || opcode == 0x18);
 155                         break;
 156                 default:
 157                         scan_more = 0;
 158                         break;
 159                 }
 160         }
 161         return prefetch;
 162 }
 163
 164 static void force_sig_info_fault(int si_signo, int si_code,
 165         unsigned long address, struct task_struct *tsk)
 166 {
 167         siginfo_t info;
 168
 169         info.si_signo = si_signo;
 170         info.si_errno = 0;
 171         info.si_code = si_code;
 172         info.si_addr = (void __user *)address;
 173         force_sig_info(si_signo, &info, tsk);
 174 }
 175
 176 void do_invalid_op(struct pt_regs *, unsigned long);
 177
 178 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 179 {
 180         unsigned index = pgd_index(address);
 181         pgd_t *pgd_k;
 182         pud_t *pud, *pud_k;
 183         pmd_t *pmd, *pmd_k;
 184
 185         pgd += index;
 186         pgd_k = init_mm.pgd + index;
 187
 188         if (!pgd_present(*pgd_k))
 189                 return NULL;
 190
 191         /*
 192          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 193          * and redundant with the set_pmd() on non-PAE. As would
 194          * set_pud.
 195          */
 196
 197         pud = pud_offset(pgd, address);
 198         pud_k = pud_offset(pgd_k, address);
 199         if (!pud_present(*pud_k))
 200                 return NULL;
 201
 202         pmd = pmd_offset(pud, address);
 203         pmd_k = pmd_offset(pud_k, address);
 204         if (!pmd_present(*pmd_k))
 205                 return NULL;
 206         if (!pmd_present(*pmd)) {
 207                 set_pmd(pmd, *pmd_k);
 208                 arch_flush_lazy_mmu_mode();
 209         } else
 210                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 211         return pmd_k;
 212 }
 213
 214 #ifdef CONFIG_X86_64
 215 static const char errata93_warning[] =
 216 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 217 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 218 KERN_ERR "******* Please consider a BIOS update.\n"
 219 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 220 #endif
 221
 222 /* Workaround for K8 erratum #93 & buggy BIOS.
 223    BIOS SMM functions are required to use a specific workaround
 224    to avoid corruption of the 64bit RIP register on C stepping K8.
 225    A lot of BIOS that didn't get tested properly miss this.
 226    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 227    Try to work around it here.
 228    Note we only handle faults in kernel here.
 229    Does nothing for X86_32
 230  */
 231 static int is_errata93(struct pt_regs *regs, unsigned long address)
 232 {
 233 #ifdef CONFIG_X86_64
 234         static int warned;
 235         if (address != regs->ip)
 236                 return 0;
 237         if ((address >> 32) != 0)
 238                 return 0;
 239         address |= 0xffffffffUL << 32;
 240         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 241             (address >= MODULES_VADDR && address <= MODULES_END)) {
 242                 if (!warned) {
 243                         printk(errata93_warning);
 244                         warned = 1;
 245                 }
 246                 regs->ip = address;
 247                 return 1;
 248         }
 249 #endif
 250         return 0;
 251 }
 252
 253
 254 /*
 255  * Handle a fault on the vmalloc or module mapping area
 256  *
 257  * This assumes no large pages in there.
 258  */
 259 static inline int vmalloc_fault(unsigned long address)
 260 {
 261 #ifdef CONFIG_X86_32
 262         unsigned long pgd_paddr;
 263         pmd_t *pmd_k;
 264         pte_t *pte_k;
 265         /*
 266          * Synchronize this task's top level page-table
 267          * with the 'reference' page table.
 268          *
 269          * Do _not_ use "current" here. We might be inside
 270          * an interrupt in the middle of a task switch..
 271          */
 272         pgd_paddr = read_cr3();
 273         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 274         if (!pmd_k)
 275                 return -1;
 276         pte_k = pte_offset_kernel(pmd_k, address);
 277         if (!pte_present(*pte_k))
 278                 return -1;
 279         return 0;
 280 #else
 281         pgd_t *pgd, *pgd_ref;
 282         pud_t *pud, *pud_ref;
 283         pmd_t *pmd, *pmd_ref;
 284         pte_t *pte, *pte_ref;
 285
 286         /* Copy kernel mappings over when needed. This can also
 287            happen within a race in page table update. In the later
 288            case just flush. */
 289
 290         pgd = pgd_offset(current->mm ?: &init_mm, address);
 291         pgd_ref = pgd_offset_k(address);
 292         if (pgd_none(*pgd_ref))
 293                 return -1;
 294         if (pgd_none(*pgd))
 295                 set_pgd(pgd, *pgd_ref);
 296         else
 297                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 298
 299         /* Below here mismatches are bugs because these lower tables
 300            are shared */
 301
 302         pud = pud_offset(pgd, address);
 303         pud_ref = pud_offset(pgd_ref, address);
 304         if (pud_none(*pud_ref))
 305                 return -1;
 306         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 307                 BUG();
 308         pmd = pmd_offset(pud, address);
 309         pmd_ref = pmd_offset(pud_ref, address);
 310         if (pmd_none(*pmd_ref))
 311                 return -1;
 312         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 313                 BUG();
 314         pte_ref = pte_offset_kernel(pmd_ref, address);
 315         if (!pte_present(*pte_ref))
 316                 return -1;
 317         pte = pte_offset_kernel(pmd, address);
 318         /* Don't use pte_page here, because the mappings can point
 319            outside mem_map, and the NUMA hash lookup cannot handle
 320            that. */
 321         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 322                 BUG();
 323         return 0;
 324 #endif
 325 }
 326
 327 int show_unhandled_signals = 1;
 328
 329 /*
 330  * This routine handles page faults.  It determines the address,
 331  * and the problem, and then passes it off to one of the appropriate
 332  * routines.
 333  */
 334 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 335 {
 336         struct task_struct *tsk;
 337         struct mm_struct *mm;
 338         struct vm_area_struct *vma;
 339         unsigned long address;
 340         int write, si_code;
 341         int fault;
 342
 343         /*
 344          * We can fault from pretty much anywhere, with unknown IRQ state.
 345          */
 346         trace_hardirqs_fixup();
 347
 348         tsk = current;
 349         mm = tsk->mm;
 350         prefetchw(&mm->mmap_sem);
 351
 352         /* get the address */
 353         address = read_cr2();
 354
 355         si_code = SEGV_MAPERR;
 356
 357         if (notify_page_fault(regs))
 358                 return;
 359
 360         /*
 361          * We fault-in kernel-space virtual memory on-demand. The
 362          * 'reference' page table is init_mm.pgd.
 363          *
 364          * NOTE! We MUST NOT take any locks for this case. We may
 365          * be in an interrupt or a critical region, and should
 366          * only copy the information from the master page table,
 367          * nothing more.
 368          *
 369          * This verifies that the fault happens in kernel space
 370          * (error_code & 4) == 0, and that the fault was not a
 371          * protection error (error_code & 9) == 0.
 372          */
 373         if (unlikely(address >= TASK_SIZE)) {
 374                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 375                     vmalloc_fault(address) >= 0)
 376                         return;
 377                 /*
 378                  * Don't take the mm semaphore here. If we fixup a prefetch
 379                  * fault we could otherwise deadlock.
 380                  */
 381                 goto bad_area_nosemaphore;
 382         }
 383
 384         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 385            fault has been handled. */
 386         if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
 387                 local_irq_enable();
 388
 389         /*
 390          * If we're in an interrupt, have no user context or are running in an
 391          * atomic region then we must not take the fault.
 392          */
 393         if (in_atomic() || !mm)
 394                 goto bad_area_nosemaphore;
 395
 396         /* When running in the kernel we expect faults to occur only to
 397          * addresses in user space.  All other faults represent errors in the
 398          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 399          * erroneous fault occurring in a code path which already holds mmap_sem
 400          * we will deadlock attempting to validate the fault against the
 401          * address space.  Luckily the kernel only validly references user
 402          * space from well defined areas of code, which are listed in the
 403          * exceptions table.
 404          *
 405          * As the vast majority of faults will be valid we will only perform
 406          * the source reference check when there is a possibility of a deadlock.
 407          * Attempt to lock the address space, if we cannot we then validate the
 408          * source.  If this is invalid we can skip the address space check,
 409          * thus avoiding the deadlock.
 410          */
 411         if (!down_read_trylock(&mm->mmap_sem)) {
 412                 if ((error_code & PF_USER) == 0 &&
 413                     !search_exception_tables(regs->ip))
 414                         goto bad_area_nosemaphore;
 415                 down_read(&mm->mmap_sem);
 416         }
 417
 418         vma = find_vma(mm, address);
 419         if (!vma)
 420                 goto bad_area;
 421         if (vma->vm_start <= address)
 422                 goto good_area;
 423         if (!(vma->vm_flags & VM_GROWSDOWN))
 424                 goto bad_area;
 425         if (error_code & PF_USER) {
 426                 /*
 427                  * Accessing the stack below %sp is always a bug.
 428                  * The large cushion allows instructions like enter
 429                  * and pusha to work.  ("enter $65535,$31" pushes
 430                  * 32 pointers and then decrements %sp by 65535.)
 431                  */
 432                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 433                         goto bad_area;
 434         }
 435         if (expand_stack(vma, address))
 436                 goto bad_area;
 437 /*
 438  * Ok, we have a good vm_area for this memory access, so
 439  * we can handle it..
 440  */
 441 good_area:
 442         si_code = SEGV_ACCERR;
 443         write = 0;
 444         switch (error_code & (PF_PROT|PF_WRITE)) {
 445         default:        /* 3: write, present */
 446                 /* fall through */
 447         case PF_WRITE:          /* write, not present */
 448                 if (!(vma->vm_flags & VM_WRITE))
 449                         goto bad_area;
 450                 write++;
 451                 break;
 452         case PF_PROT:           /* read, present */
 453                 goto bad_area;
 454         case 0:                 /* read, not present */
 455                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 456                         goto bad_area;
 457         }
 458
 459  survive:
 460         /*
 461          * If for any reason at all we couldn't handle the fault,
 462          * make sure we exit gracefully rather than endlessly redo
 463          * the fault.
 464          */
 465         fault = handle_mm_fault(mm, vma, address, write);
 466         if (unlikely(fault & VM_FAULT_ERROR)) {
 467                 if (fault & VM_FAULT_OOM)
 468                         goto out_of_memory;
 469                 else if (fault & VM_FAULT_SIGBUS)
 470                         goto do_sigbus;
 471                 BUG();
 472         }
 473         if (fault & VM_FAULT_MAJOR)
 474                 tsk->maj_flt++;
 475         else
 476                 tsk->min_flt++;
 477
 478         /*
 479          * Did it hit the DOS screen memory VA from vm86 mode?
 480          */
 481         if (regs->flags & VM_MASK) {
 482                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 483                 if (bit < 32)
 484                         tsk->thread.screen_bitmap |= 1 << bit;
 485         }
 486         up_read(&mm->mmap_sem);
 487         return;
 488
 489 /*
 490  * Something tried to access memory that isn't in our memory map..
 491  * Fix it, but check if it's kernel or user first..
 492  */
 493 bad_area:
 494         up_read(&mm->mmap_sem);
 495
 496 bad_area_nosemaphore:
 497         /* User mode accesses just cause a SIGSEGV */
 498         if (error_code & PF_USER) {
 499                 /*
 500                  * It's possible to have interrupts off here.
 501                  */
 502                 local_irq_enable();
 503
 504                 /*
 505                  * Valid to do another page fault here because this one came
 506                  * from user space.
 507                  */
 508                 if (is_prefetch(regs, address, error_code))
 509                         return;
 510
 511                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 512                     printk_ratelimit()) {
 513                         printk(
 514 #ifdef CONFIG_X86_32
 515                         "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
 516 #else
 517                         "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx\n",
 518 #endif
 519                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 520                         tsk->comm, task_pid_nr(tsk), address, regs->ip,
 521                         regs->sp, error_code);
 522                 }
 523                 tsk->thread.cr2 = address;
 524                 /* Kernel addresses are always protection faults */
 525                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 526                 tsk->thread.trap_no = 14;
 527                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 528                 return;
 529         }
 530
 531 #ifdef CONFIG_X86_F00F_BUG
 532         /*
 533          * Pentium F0 0F C7 C8 bug workaround.
 534          */
 535         if (boot_cpu_data.f00f_bug) {
 536                 unsigned long nr;
 537
 538                 nr = (address - idt_descr.address) >> 3;
 539
 540                 if (nr == 6) {
 541                         do_invalid_op(regs, 0);
 542                         return;
 543                 }
 544         }
 545 #endif
 546
 547 no_context:
 548         /* Are we prepared to handle this kernel fault?  */
 549         if (fixup_exception(regs))
 550                 return;
 551
 552         /*
 553          * Valid to do another page fault here, because if this fault
 554          * had been triggered by is_prefetch fixup_exception would have
 555          * handled it.
 556          */
 557         if (is_prefetch(regs, address, error_code))
 558                 return;
 559
 560         if (is_errata93(regs, address))
 561                 return;
 562
 563 /*
 564  * Oops. The kernel tried to access some bad page. We'll have to
 565  * terminate things with extreme prejudice.
 566  */
 567
 568         bust_spinlocks(1);
 569
 570         if (oops_may_print()) {
 571                 __typeof__(pte_val(__pte(0))) page;
 572
 573 #ifdef CONFIG_X86_PAE
 574                 if (error_code & PF_INSTR) {
 575                         pte_t *pte = lookup_address(address);
 576
 577                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 578                                 printk(KERN_CRIT "kernel tried to execute "
 579                                         "NX-protected page - exploit attempt? "
 580                                         "(uid: %d)\n", current->uid);
 581                 }
 582 #endif
 583                 if (address < PAGE_SIZE)
 584                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 585                                         "pointer dereference");
 586                 else
 587                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 588                                         " request");
 589                 printk(" at virtual address %08lx\n", address);
 590                 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
 591
 592                 page = read_cr3();
 593                 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 594 #ifdef CONFIG_X86_PAE
 595                 printk("*pdpt = %016Lx ", page);
 596                 if ((page >> PAGE_SHIFT) < max_low_pfn
 597                     && page & _PAGE_PRESENT) {
 598                         page &= PAGE_MASK;
 599                         page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 600                                                                  & (PTRS_PER_PMD - 1)];
 601                         printk(KERN_CONT "*pde = %016Lx ", page);
 602                         page &= ~_PAGE_NX;
 603                 }
 604 #else
 605                 printk("*pde = %08lx ", page);
 606 #endif
 607
 608                 /*
 609                  * We must not directly access the pte in the highpte
 610                  * case if the page table is located in highmem.
 611                  * And let's rather not kmap-atomic the pte, just in case
 612                  * it's allocated already.
 613                  */
 614                 if ((page >> PAGE_SHIFT) < max_low_pfn
 615                     && (page & _PAGE_PRESENT)
 616                     && !(page & _PAGE_PSE)) {
 617                         page &= PAGE_MASK;
 618                         page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 619                                                                  & (PTRS_PER_PTE - 1)];
 620                         printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 621                 }
 622
 623                 printk("\n");
 624         }
 625
 626         tsk->thread.cr2 = address;
 627         tsk->thread.trap_no = 14;
 628         tsk->thread.error_code = error_code;
 629         die("Oops", regs, error_code);
 630         bust_spinlocks(0);
 631         do_exit(SIGKILL);
 632
 633 /*
 634  * We ran out of memory, or some other thing happened to us that made
 635  * us unable to handle the page fault gracefully.
 636  */
 637 out_of_memory:
 638         up_read(&mm->mmap_sem);
 639         if (is_global_init(tsk)) {
 640                 yield();
 641                 down_read(&mm->mmap_sem);
 642                 goto survive;
 643         }
 644         printk("VM: killing process %s\n", tsk->comm);
 645         if (error_code & PF_USER)
 646                 do_group_exit(SIGKILL);
 647         goto no_context;
 648
 649 do_sigbus:
 650         up_read(&mm->mmap_sem);
 651
 652         /* Kernel mode? Handle exceptions or die */
 653         if (!(error_code & PF_USER))
 654                 goto no_context;
 655
 656         /* User space => ok to do another page fault */
 657         if (is_prefetch(regs, address, error_code))
 658                 return;
 659
 660         tsk->thread.cr2 = address;
 661         tsk->thread.error_code = error_code;
 662         tsk->thread.trap_no = 14;
 663         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 664 }
 665
 666 void vmalloc_sync_all(void)
 667 {
 668         /*
 669          * Note that races in the updates of insync and start aren't
 670          * problematic: insync can only get set bits added, and updates to
 671          * start are only improving performance (without affecting correctness
 672          * if undone).
 673          */
 674         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 675         static unsigned long start = TASK_SIZE;
 676         unsigned long address;
 677
 678         if (SHARED_KERNEL_PMD)
 679                 return;
 680
 681         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 682         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 683                 if (!test_bit(pgd_index(address), insync)) {
 684                         unsigned long flags;
 685                         struct page *page;
 686
 687                         spin_lock_irqsave(&pgd_lock, flags);
 688                         for (page = pgd_list; page; page =
 689                                         (struct page *)page->index)
 690                                 if (!vmalloc_sync_one(page_address(page),
 691                                                                 address)) {
 692                                         BUG_ON(page != pgd_list);
 693                                         break;
 694                                 }
 695                         spin_unlock_irqrestore(&pgd_lock, flags);
 696                         if (!page)
 697                                 set_bit(pgd_index(address), insync);
 698                 }
 699                 if (address == start && test_bit(pgd_index(address), insync))
 700                         start = address + PGDIR_SIZE;
 701         }
 702 }