2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
6 #include <linux/signal.h>
7 #include <linux/sched.h>
8 #include <linux/kernel.h>
9 #include <linux/errno.h>
10 #include <linux/string.h>
11 #include <linux/types.h>
12 #include <linux/ptrace.h>
13 #include <linux/mmiotrace.h>
14 #include <linux/mman.h>
16 #include <linux/smp.h>
17 #include <linux/interrupt.h>
18 #include <linux/init.h>
19 #include <linux/tty.h>
20 #include <linux/vt_kern.h> /* For unblank_screen() */
21 #include <linux/compiler.h>
22 #include <linux/highmem.h>
23 #include <linux/bootmem.h> /* for max_low_pfn */
24 #include <linux/vmalloc.h>
25 #include <linux/module.h>
26 #include <linux/kprobes.h>
27 #include <linux/uaccess.h>
28 #include <linux/kdebug.h>
30 #include <asm/system.h>
32 #include <asm/segment.h>
33 #include <asm/pgalloc.h>
35 #include <asm/tlbflush.h>
36 #include <asm/proto.h>
37 #include <asm-generic/sections.h>
40 * Page fault error code bits
41 * bit 0 == 0 means no page found, 1 means protection fault
42 * bit 1 == 0 means read, 1 means write
43 * bit 2 == 0 means kernel, 1 means user-mode
44 * bit 3 == 1 means use of reserved bit detected
45 * bit 4 == 1 means fault was an instruction fetch
47 #define PF_PROT (1<<0)
48 #define PF_WRITE (1<<1)
49 #define PF_USER (1<<2)
50 #define PF_RSVD (1<<3)
51 #define PF_INSTR (1<<4)
53 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
55 #ifdef CONFIG_MMIOTRACE_HOOKS
56 if (unlikely(is_kmmio_active()))
57 if (kmmio_handler(regs, addr) == 1)
63 static inline int notify_page_fault(struct pt_regs *regs)
68 /* kprobe_running() needs smp_processor_id() */
70 if (!user_mode_vm(regs)) {
72 if (!user_mode(regs)) {
75 if (kprobe_running() && kprobe_fault_handler(regs, 14))
88 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
89 * Check that here and ignore it.
92 * Sometimes the CPU reports invalid exceptions on prefetch.
93 * Check that here and ignore it.
95 * Opcode checker based on code by Richard Brunner
97 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
98 unsigned long error_code)
100 unsigned char *instr;
103 unsigned char *max_instr;
106 * If it was a exec (instruction fetch) fault on NX page, then
107 * do not ignore the fault:
109 if (error_code & PF_INSTR)
112 instr = (unsigned char *)convert_ip_to_linear(current, regs);
113 max_instr = instr + 15;
115 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
118 while (scan_more && instr < max_instr) {
119 unsigned char opcode;
120 unsigned char instr_hi;
121 unsigned char instr_lo;
123 if (probe_kernel_address(instr, opcode))
126 instr_hi = opcode & 0xf0;
127 instr_lo = opcode & 0x0f;
134 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
135 * In X86_64 long mode, the CPU will signal invalid
136 * opcode if some of these prefixes are present so
137 * X86_64 will never get here anyway
139 scan_more = ((instr_lo & 7) == 0x6);
144 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
145 * Need to figure out under what instruction mode the
146 * instruction was issued. Could check the LDT for lm,
147 * but for now it's good enough to assume that long
148 * mode only uses well known segments or kernel.
150 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
154 /* 0x64 thru 0x67 are valid prefixes in all modes. */
155 scan_more = (instr_lo & 0xC) == 0x4;
158 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
159 scan_more = !instr_lo || (instr_lo>>1) == 1;
162 /* Prefetch instruction is 0x0F0D or 0x0F18 */
165 if (probe_kernel_address(instr, opcode))
167 prefetch = (instr_lo == 0xF) &&
168 (opcode == 0x0D || opcode == 0x18);
178 static void force_sig_info_fault(int si_signo, int si_code,
179 unsigned long address, struct task_struct *tsk)
183 info.si_signo = si_signo;
185 info.si_code = si_code;
186 info.si_addr = (void __user *)address;
187 force_sig_info(si_signo, &info, tsk);
191 static int bad_address(void *p)
194 return probe_kernel_address((unsigned long *)p, dummy);
198 static void dump_pagetable(unsigned long address)
201 __typeof__(pte_val(__pte(0))) page;
204 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
205 #ifdef CONFIG_X86_PAE
206 printk("*pdpt = %016Lx ", page);
207 if ((page >> PAGE_SHIFT) < max_low_pfn
208 && page & _PAGE_PRESENT) {
210 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
211 & (PTRS_PER_PMD - 1)];
212 printk(KERN_CONT "*pde = %016Lx ", page);
216 printk("*pde = %08lx ", page);
220 * We must not directly access the pte in the highpte
221 * case if the page table is located in highmem.
222 * And let's rather not kmap-atomic the pte, just in case
223 * it's allocated already.
225 if ((page >> PAGE_SHIFT) < max_low_pfn
226 && (page & _PAGE_PRESENT)
227 && !(page & _PAGE_PSE)) {
229 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
230 & (PTRS_PER_PTE - 1)];
231 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
235 #else /* CONFIG_X86_64 */
241 pgd = (pgd_t *)read_cr3();
243 pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
244 pgd += pgd_index(address);
245 if (bad_address(pgd)) goto bad;
246 printk("PGD %lx ", pgd_val(*pgd));
247 if (!pgd_present(*pgd)) goto ret;
249 pud = pud_offset(pgd, address);
250 if (bad_address(pud)) goto bad;
251 printk("PUD %lx ", pud_val(*pud));
252 if (!pud_present(*pud) || pud_large(*pud))
255 pmd = pmd_offset(pud, address);
256 if (bad_address(pmd)) goto bad;
257 printk("PMD %lx ", pmd_val(*pmd));
258 if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
260 pte = pte_offset_kernel(pmd, address);
261 if (bad_address(pte)) goto bad;
262 printk("PTE %lx", pte_val(*pte));
272 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
274 unsigned index = pgd_index(address);
280 pgd_k = init_mm.pgd + index;
282 if (!pgd_present(*pgd_k))
286 * set_pgd(pgd, *pgd_k); here would be useless on PAE
287 * and redundant with the set_pmd() on non-PAE. As would
291 pud = pud_offset(pgd, address);
292 pud_k = pud_offset(pgd_k, address);
293 if (!pud_present(*pud_k))
296 pmd = pmd_offset(pud, address);
297 pmd_k = pmd_offset(pud_k, address);
298 if (!pmd_present(*pmd_k))
300 if (!pmd_present(*pmd)) {
301 set_pmd(pmd, *pmd_k);
302 arch_flush_lazy_mmu_mode();
304 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
310 static const char errata93_warning[] =
311 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
312 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
313 KERN_ERR "******* Please consider a BIOS update.\n"
314 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
317 /* Workaround for K8 erratum #93 & buggy BIOS.
318 BIOS SMM functions are required to use a specific workaround
319 to avoid corruption of the 64bit RIP register on C stepping K8.
320 A lot of BIOS that didn't get tested properly miss this.
321 The OS sees this as a page fault with the upper 32bits of RIP cleared.
322 Try to work around it here.
323 Note we only handle faults in kernel here.
324 Does nothing for X86_32
326 static int is_errata93(struct pt_regs *regs, unsigned long address)
330 if (address != regs->ip)
332 if ((address >> 32) != 0)
334 address |= 0xffffffffUL << 32;
335 if ((address >= (u64)_stext && address <= (u64)_etext) ||
336 (address >= MODULES_VADDR && address <= MODULES_END)) {
338 printk(errata93_warning);
349 * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
350 * addresses >4GB. We catch this in the page fault handler because these
351 * addresses are not reachable. Just detect this case and return. Any code
352 * segment in LDT is compatibility mode.
354 static int is_errata100(struct pt_regs *regs, unsigned long address)
357 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
364 void do_invalid_op(struct pt_regs *, unsigned long);
366 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
368 #ifdef CONFIG_X86_F00F_BUG
371 * Pentium F0 0F C7 C8 bug workaround.
373 if (boot_cpu_data.f00f_bug) {
374 nr = (address - idt_descr.address) >> 3;
377 do_invalid_op(regs, 0);
385 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
386 unsigned long address)
389 if (!oops_may_print())
393 #ifdef CONFIG_X86_PAE
394 if (error_code & PF_INSTR) {
396 pte_t *pte = lookup_address(address, &level);
398 if (pte && pte_present(*pte) && !pte_exec(*pte))
399 printk(KERN_CRIT "kernel tried to execute "
400 "NX-protected page - exploit attempt? "
401 "(uid: %d)\n", current->uid);
405 printk(KERN_ALERT "BUG: unable to handle kernel ");
406 if (address < PAGE_SIZE)
407 printk(KERN_CONT "NULL pointer dereference");
409 printk(KERN_CONT "paging request");
411 printk(KERN_CONT " at %08lx\n", address);
413 printk(KERN_CONT " at %016lx\n", address);
415 printk(KERN_ALERT "IP:");
416 printk_address(regs->ip, 1);
417 dump_pagetable(address);
421 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
422 unsigned long error_code)
424 unsigned long flags = oops_begin();
425 struct task_struct *tsk;
427 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
428 current->comm, address);
429 dump_pagetable(address);
431 tsk->thread.cr2 = address;
432 tsk->thread.trap_no = 14;
433 tsk->thread.error_code = error_code;
434 if (__die("Bad pagetable", regs, error_code))
436 oops_end(flags, regs, SIGKILL);
440 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
442 if ((error_code & PF_WRITE) && !pte_write(*pte))
444 if ((error_code & PF_INSTR) && !pte_exec(*pte))
451 * Handle a spurious fault caused by a stale TLB entry. This allows
452 * us to lazily refresh the TLB when increasing the permissions of a
453 * kernel page (RO -> RW or NX -> X). Doing it eagerly is very
454 * expensive since that implies doing a full cross-processor TLB
455 * flush, even if no stale TLB entries exist on other processors.
456 * There are no security implications to leaving a stale TLB when
457 * increasing the permissions on a page.
459 static int spurious_fault(unsigned long address,
460 unsigned long error_code)
467 /* Reserved-bit violation or user access to kernel space? */
468 if (error_code & (PF_USER | PF_RSVD))
471 pgd = init_mm.pgd + pgd_index(address);
472 if (!pgd_present(*pgd))
475 pud = pud_offset(pgd, address);
476 if (!pud_present(*pud))
480 return spurious_fault_check(error_code, (pte_t *) pud);
482 pmd = pmd_offset(pud, address);
483 if (!pmd_present(*pmd))
487 return spurious_fault_check(error_code, (pte_t *) pmd);
489 pte = pte_offset_kernel(pmd, address);
490 if (!pte_present(*pte))
493 return spurious_fault_check(error_code, pte);
498 * Handle a fault on the vmalloc or module mapping area
501 * Handle a fault on the vmalloc area
503 * This assumes no large pages in there.
505 static int vmalloc_fault(unsigned long address)
508 unsigned long pgd_paddr;
512 * Synchronize this task's top level page-table
513 * with the 'reference' page table.
515 * Do _not_ use "current" here. We might be inside
516 * an interrupt in the middle of a task switch..
518 pgd_paddr = read_cr3();
519 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
522 pte_k = pte_offset_kernel(pmd_k, address);
523 if (!pte_present(*pte_k))
527 pgd_t *pgd, *pgd_ref;
528 pud_t *pud, *pud_ref;
529 pmd_t *pmd, *pmd_ref;
530 pte_t *pte, *pte_ref;
532 /* Make sure we are in vmalloc area */
533 if (!(address >= VMALLOC_START && address < VMALLOC_END))
536 /* Copy kernel mappings over when needed. This can also
537 happen within a race in page table update. In the later
540 pgd = pgd_offset(current->mm ?: &init_mm, address);
541 pgd_ref = pgd_offset_k(address);
542 if (pgd_none(*pgd_ref))
545 set_pgd(pgd, *pgd_ref);
547 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
549 /* Below here mismatches are bugs because these lower tables
552 pud = pud_offset(pgd, address);
553 pud_ref = pud_offset(pgd_ref, address);
554 if (pud_none(*pud_ref))
556 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
558 pmd = pmd_offset(pud, address);
559 pmd_ref = pmd_offset(pud_ref, address);
560 if (pmd_none(*pmd_ref))
562 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
564 pte_ref = pte_offset_kernel(pmd_ref, address);
565 if (!pte_present(*pte_ref))
567 pte = pte_offset_kernel(pmd, address);
568 /* Don't use pte_page here, because the mappings can point
569 outside mem_map, and the NUMA hash lookup cannot handle
571 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
577 int show_unhandled_signals = 1;
580 * This routine handles page faults. It determines the address,
581 * and the problem, and then passes it off to one of the appropriate
587 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
589 struct task_struct *tsk;
590 struct mm_struct *mm;
591 struct vm_area_struct *vma;
592 unsigned long address;
600 * We can fault from pretty much anywhere, with unknown IRQ state.
602 trace_hardirqs_fixup();
606 prefetchw(&mm->mmap_sem);
608 /* get the address */
609 address = read_cr2();
611 si_code = SEGV_MAPERR;
613 if (notify_page_fault(regs))
615 if (unlikely(kmmio_fault(regs, address)))
619 * We fault-in kernel-space virtual memory on-demand. The
620 * 'reference' page table is init_mm.pgd.
622 * NOTE! We MUST NOT take any locks for this case. We may
623 * be in an interrupt or a critical region, and should
624 * only copy the information from the master page table,
627 * This verifies that the fault happens in kernel space
628 * (error_code & 4) == 0, and that the fault was not a
629 * protection error (error_code & 9) == 0.
632 if (unlikely(address >= TASK_SIZE)) {
634 if (unlikely(address >= TASK_SIZE64)) {
636 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
637 vmalloc_fault(address) >= 0)
640 /* Can handle a stale RO->RW TLB */
641 if (spurious_fault(address, error_code))
645 * Don't take the mm semaphore here. If we fixup a prefetch
646 * fault we could otherwise deadlock.
648 goto bad_area_nosemaphore;
653 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
654 fault has been handled. */
655 if (regs->flags & (X86_EFLAGS_IF | X86_VM_MASK))
659 * If we're in an interrupt, have no user context or are running in an
660 * atomic region then we must not take the fault.
662 if (in_atomic() || !mm)
663 goto bad_area_nosemaphore;
664 #else /* CONFIG_X86_64 */
665 if (likely(regs->flags & X86_EFLAGS_IF))
668 if (unlikely(error_code & PF_RSVD))
669 pgtable_bad(address, regs, error_code);
672 * If we're in an interrupt, have no user context or are running in an
673 * atomic region then we must not take the fault.
675 if (unlikely(in_atomic() || !mm))
676 goto bad_area_nosemaphore;
679 * User-mode registers count as a user access even for any
680 * potential system fault or CPU buglet.
682 if (user_mode_vm(regs))
683 error_code |= PF_USER;
686 /* When running in the kernel we expect faults to occur only to
687 * addresses in user space. All other faults represent errors in the
688 * kernel and should generate an OOPS. Unfortunately, in the case of an
689 * erroneous fault occurring in a code path which already holds mmap_sem
690 * we will deadlock attempting to validate the fault against the
691 * address space. Luckily the kernel only validly references user
692 * space from well defined areas of code, which are listed in the
695 * As the vast majority of faults will be valid we will only perform
696 * the source reference check when there is a possibility of a deadlock.
697 * Attempt to lock the address space, if we cannot we then validate the
698 * source. If this is invalid we can skip the address space check,
699 * thus avoiding the deadlock.
701 if (!down_read_trylock(&mm->mmap_sem)) {
702 if ((error_code & PF_USER) == 0 &&
703 !search_exception_tables(regs->ip))
704 goto bad_area_nosemaphore;
705 down_read(&mm->mmap_sem);
708 vma = find_vma(mm, address);
711 if (vma->vm_start <= address)
713 if (!(vma->vm_flags & VM_GROWSDOWN))
715 if (error_code & PF_USER) {
717 * Accessing the stack below %sp is always a bug.
718 * The large cushion allows instructions like enter
719 * and pusha to work. ("enter $65535,$31" pushes
720 * 32 pointers and then decrements %sp by 65535.)
722 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
725 if (expand_stack(vma, address))
728 * Ok, we have a good vm_area for this memory access, so
732 si_code = SEGV_ACCERR;
734 switch (error_code & (PF_PROT|PF_WRITE)) {
735 default: /* 3: write, present */
737 case PF_WRITE: /* write, not present */
738 if (!(vma->vm_flags & VM_WRITE))
742 case PF_PROT: /* read, present */
744 case 0: /* read, not present */
745 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
753 * If for any reason at all we couldn't handle the fault,
754 * make sure we exit gracefully rather than endlessly redo
757 fault = handle_mm_fault(mm, vma, address, write);
758 if (unlikely(fault & VM_FAULT_ERROR)) {
759 if (fault & VM_FAULT_OOM)
761 else if (fault & VM_FAULT_SIGBUS)
765 if (fault & VM_FAULT_MAJOR)
772 * Did it hit the DOS screen memory VA from vm86 mode?
774 if (v8086_mode(regs)) {
775 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
777 tsk->thread.screen_bitmap |= 1 << bit;
780 up_read(&mm->mmap_sem);
784 * Something tried to access memory that isn't in our memory map..
785 * Fix it, but check if it's kernel or user first..
788 up_read(&mm->mmap_sem);
790 bad_area_nosemaphore:
791 /* User mode accesses just cause a SIGSEGV */
792 if (error_code & PF_USER) {
794 * It's possible to have interrupts off here.
799 * Valid to do another page fault here because this one came
802 if (is_prefetch(regs, address, error_code))
805 if (is_errata100(regs, address))
808 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
809 printk_ratelimit()) {
812 "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
814 "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
816 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
817 tsk->comm, task_pid_nr(tsk), address, regs->ip,
818 regs->sp, error_code);
819 print_vma_addr(" in ", regs->ip);
823 tsk->thread.cr2 = address;
824 /* Kernel addresses are always protection faults */
825 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
826 tsk->thread.trap_no = 14;
827 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
831 if (is_f00f_bug(regs, address))
835 /* Are we prepared to handle this kernel fault? */
836 if (fixup_exception(regs))
841 * Valid to do another page fault here, because if this fault
842 * had been triggered by is_prefetch fixup_exception would have
846 * Hall of shame of CPU/BIOS bugs.
848 if (is_prefetch(regs, address, error_code))
851 if (is_errata93(regs, address))
855 * Oops. The kernel tried to access some bad page. We'll have to
856 * terminate things with extreme prejudice.
861 flags = oops_begin();
864 show_fault_oops(regs, error_code, address);
866 tsk->thread.cr2 = address;
867 tsk->thread.trap_no = 14;
868 tsk->thread.error_code = error_code;
871 die("Oops", regs, error_code);
875 if (__die("Oops", regs, error_code))
877 /* Executive summary in case the body of the oops scrolled away */
878 printk(KERN_EMERG "CR2: %016lx\n", address);
879 oops_end(flags, regs, SIGKILL);
883 * We ran out of memory, or some other thing happened to us that made
884 * us unable to handle the page fault gracefully.
887 up_read(&mm->mmap_sem);
888 if (is_global_init(tsk)) {
891 down_read(&mm->mmap_sem);
898 printk("VM: killing process %s\n", tsk->comm);
899 if (error_code & PF_USER)
900 do_group_exit(SIGKILL);
904 up_read(&mm->mmap_sem);
906 /* Kernel mode? Handle exceptions or die */
907 if (!(error_code & PF_USER))
910 /* User space => ok to do another page fault */
911 if (is_prefetch(regs, address, error_code))
914 tsk->thread.cr2 = address;
915 tsk->thread.error_code = error_code;
916 tsk->thread.trap_no = 14;
917 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
920 DEFINE_SPINLOCK(pgd_lock);
923 void vmalloc_sync_all(void)
927 * Note that races in the updates of insync and start aren't
928 * problematic: insync can only get set bits added, and updates to
929 * start are only improving performance (without affecting correctness
932 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
933 static unsigned long start = TASK_SIZE;
934 unsigned long address;
936 if (SHARED_KERNEL_PMD)
939 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
940 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
941 if (!test_bit(pgd_index(address), insync)) {
945 spin_lock_irqsave(&pgd_lock, flags);
946 list_for_each_entry(page, &pgd_list, lru) {
947 if (!vmalloc_sync_one(page_address(page),
951 spin_unlock_irqrestore(&pgd_lock, flags);
953 set_bit(pgd_index(address), insync);
955 if (address == start && test_bit(pgd_index(address), insync))
956 start = address + PGDIR_SIZE;
958 #else /* CONFIG_X86_64 */
960 * Note that races in the updates of insync and start aren't
961 * problematic: insync can only get set bits added, and updates to
962 * start are only improving performance (without affecting correctness
965 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
966 static unsigned long start = VMALLOC_START & PGDIR_MASK;
967 unsigned long address;
969 for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
970 if (!test_bit(pgd_index(address), insync)) {
971 const pgd_t *pgd_ref = pgd_offset_k(address);
975 if (pgd_none(*pgd_ref))
977 spin_lock_irqsave(&pgd_lock, flags);
978 list_for_each_entry(page, &pgd_list, lru) {
980 pgd = (pgd_t *)page_address(page) + pgd_index(address);
982 set_pgd(pgd, *pgd_ref);
984 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
986 spin_unlock_irqrestore(&pgd_lock, flags);
987 set_bit(pgd_index(address), insync);
989 if (address == start)
990 start = address + PGDIR_SIZE;