4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
70 unsigned long num_physpages;
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79 unsigned long vmalloc_earlyreserve;
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none. Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
91 void pgd_clear_bad(pgd_t *pgd)
97 void pud_clear_bad(pud_t *pud)
103 void pmd_clear_bad(pmd_t *pmd)
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
115 struct page *page = pmd_page(*pmd);
117 pte_free_tlb(tlb, page);
118 dec_page_state(nr_page_table_pages);
122 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
123 unsigned long addr, unsigned long end,
124 unsigned long floor, unsigned long ceiling)
131 pmd = pmd_offset(pud, addr);
133 next = pmd_addr_end(addr, end);
134 if (pmd_none_or_clear_bad(pmd))
136 free_pte_range(tlb, pmd);
137 } while (pmd++, addr = next, addr != end);
147 if (end - 1 > ceiling - 1)
150 pmd = pmd_offset(pud, start);
152 pmd_free_tlb(tlb, pmd);
155 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
156 unsigned long addr, unsigned long end,
157 unsigned long floor, unsigned long ceiling)
164 pud = pud_offset(pgd, addr);
166 next = pud_addr_end(addr, end);
167 if (pud_none_or_clear_bad(pud))
169 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
170 } while (pud++, addr = next, addr != end);
176 ceiling &= PGDIR_MASK;
180 if (end - 1 > ceiling - 1)
183 pud = pud_offset(pgd, start);
185 pud_free_tlb(tlb, pud);
189 * This function frees user-level page tables of a process.
191 * Must be called with pagetable lock held.
193 void free_pgd_range(struct mmu_gather **tlb,
194 unsigned long addr, unsigned long end,
195 unsigned long floor, unsigned long ceiling)
202 * The next few lines have given us lots of grief...
204 * Why are we testing PMD* at this top level? Because often
205 * there will be no work to do at all, and we'd prefer not to
206 * go all the way down to the bottom just to discover that.
208 * Why all these "- 1"s? Because 0 represents both the bottom
209 * of the address space and the top of it (using -1 for the
210 * top wouldn't help much: the masks would do the wrong thing).
211 * The rule is that addr 0 and floor 0 refer to the bottom of
212 * the address space, but end 0 and ceiling 0 refer to the top
213 * Comparisons need to use "end - 1" and "ceiling - 1" (though
214 * that end 0 case should be mythical).
216 * Wherever addr is brought up or ceiling brought down, we must
217 * be careful to reject "the opposite 0" before it confuses the
218 * subsequent tests. But what about where end is brought down
219 * by PMD_SIZE below? no, end can't go down to 0 there.
221 * Whereas we round start (addr) and ceiling down, by different
222 * masks at different levels, in order to test whether a table
223 * now has no other vmas using it, so can be freed, we don't
224 * bother to round floor or end up - the tests don't need that.
238 if (end - 1 > ceiling - 1)
244 pgd = pgd_offset((*tlb)->mm, addr);
246 next = pgd_addr_end(addr, end);
247 if (pgd_none_or_clear_bad(pgd))
249 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
250 } while (pgd++, addr = next, addr != end);
253 flush_tlb_pgtables((*tlb)->mm, start, end);
256 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
257 unsigned long floor, unsigned long ceiling)
260 struct vm_area_struct *next = vma->vm_next;
261 unsigned long addr = vma->vm_start;
264 * Hide vma from rmap and vmtruncate before freeing pgtables
266 anon_vma_unlink(vma);
267 unlink_file_vma(vma);
269 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
270 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
271 floor, next? next->vm_start: ceiling);
274 * Optimization: gather nearby vmas into one call down
276 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
277 && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
281 anon_vma_unlink(vma);
282 unlink_file_vma(vma);
284 free_pgd_range(tlb, addr, vma->vm_end,
285 floor, next? next->vm_start: ceiling);
291 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
293 struct page *new = pte_alloc_one(mm, address);
297 spin_lock(&mm->page_table_lock);
298 if (pmd_present(*pmd)) /* Another has populated it */
302 inc_page_state(nr_page_table_pages);
303 pmd_populate(mm, pmd, new);
305 spin_unlock(&mm->page_table_lock);
309 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
311 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315 spin_lock(&init_mm.page_table_lock);
316 if (pmd_present(*pmd)) /* Another has populated it */
317 pte_free_kernel(new);
319 pmd_populate_kernel(&init_mm, pmd, new);
320 spin_unlock(&init_mm.page_table_lock);
324 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
327 add_mm_counter(mm, file_rss, file_rss);
329 add_mm_counter(mm, anon_rss, anon_rss);
333 * This function is called to print an error when a pte in a
334 * !VM_RESERVED region is found pointing to an invalid pfn (which
337 * The calling function must still handle the error.
339 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
341 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
342 "vm_flags = %lx, vaddr = %lx\n",
343 (long long)pte_val(pte),
344 (vma->vm_mm == current->mm ? current->comm : "???"),
345 vma->vm_flags, vaddr);
350 * copy one vm_area from one task to the other. Assumes the page tables
351 * already present in the new task to be cleared in the whole range
352 * covered by this vma.
356 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
357 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
358 unsigned long addr, int *rss)
360 unsigned long vm_flags = vma->vm_flags;
361 pte_t pte = *src_pte;
365 /* pte contains position in swap or file, so copy. */
366 if (unlikely(!pte_present(pte))) {
367 if (!pte_file(pte)) {
368 swap_duplicate(pte_to_swp_entry(pte));
369 /* make sure dst_mm is on swapoff's mmlist. */
370 if (unlikely(list_empty(&dst_mm->mmlist))) {
371 spin_lock(&mmlist_lock);
372 list_add(&dst_mm->mmlist, &src_mm->mmlist);
373 spin_unlock(&mmlist_lock);
379 /* If the region is VM_RESERVED, the mapping is not
380 * mapped via rmap - duplicate the pte as is.
382 if (vm_flags & VM_RESERVED)
386 /* If the pte points outside of valid memory but
387 * the region is not VM_RESERVED, we have a problem.
389 if (unlikely(!pfn_valid(pfn))) {
390 print_bad_pte(vma, pte, addr);
391 goto out_set_pte; /* try to do something sane */
394 page = pfn_to_page(pfn);
397 * If it's a COW mapping, write protect it both
398 * in the parent and the child
400 if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
401 ptep_set_wrprotect(src_mm, addr, src_pte);
406 * If it's a shared mapping, mark it clean in
409 if (vm_flags & VM_SHARED)
410 pte = pte_mkclean(pte);
411 pte = pte_mkold(pte);
414 rss[!!PageAnon(page)]++;
417 set_pte_at(dst_mm, addr, dst_pte, pte);
420 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
421 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
422 unsigned long addr, unsigned long end)
424 pte_t *src_pte, *dst_pte;
425 spinlock_t *src_ptl, *dst_ptl;
431 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
434 src_pte = pte_offset_map_nested(src_pmd, addr);
435 src_ptl = &src_mm->page_table_lock;
440 * We are holding two locks at this point - either of them
441 * could generate latencies in another task on another CPU.
443 if (progress >= 32) {
445 if (need_resched() ||
446 need_lockbreak(src_ptl) ||
447 need_lockbreak(dst_ptl))
450 if (pte_none(*src_pte)) {
454 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
456 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
458 spin_unlock(src_ptl);
459 pte_unmap_nested(src_pte - 1);
460 add_mm_rss(dst_mm, rss[0], rss[1]);
461 pte_unmap_unlock(dst_pte - 1, dst_ptl);
468 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
469 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
470 unsigned long addr, unsigned long end)
472 pmd_t *src_pmd, *dst_pmd;
475 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
478 src_pmd = pmd_offset(src_pud, addr);
480 next = pmd_addr_end(addr, end);
481 if (pmd_none_or_clear_bad(src_pmd))
483 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
486 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
490 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
492 unsigned long addr, unsigned long end)
494 pud_t *src_pud, *dst_pud;
497 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
500 src_pud = pud_offset(src_pgd, addr);
502 next = pud_addr_end(addr, end);
503 if (pud_none_or_clear_bad(src_pud))
505 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
508 } while (dst_pud++, src_pud++, addr = next, addr != end);
512 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
513 struct vm_area_struct *vma)
515 pgd_t *src_pgd, *dst_pgd;
517 unsigned long addr = vma->vm_start;
518 unsigned long end = vma->vm_end;
521 * Don't copy ptes where a page fault will fill them correctly.
522 * Fork becomes much lighter when there are big shared or private
523 * readonly mappings. The tradeoff is that copy_page_range is more
524 * efficient than faulting.
526 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_RESERVED))) {
531 if (is_vm_hugetlb_page(vma))
532 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
534 dst_pgd = pgd_offset(dst_mm, addr);
535 src_pgd = pgd_offset(src_mm, addr);
537 next = pgd_addr_end(addr, end);
538 if (pgd_none_or_clear_bad(src_pgd))
540 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
543 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
547 static void zap_pte_range(struct mmu_gather *tlb,
548 struct vm_area_struct *vma, pmd_t *pmd,
549 unsigned long addr, unsigned long end,
550 struct zap_details *details)
552 struct mm_struct *mm = tlb->mm;
557 pte = pte_offset_map(pmd, addr);
562 if (pte_present(ptent)) {
563 struct page *page = NULL;
564 if (!(vma->vm_flags & VM_RESERVED)) {
565 unsigned long pfn = pte_pfn(ptent);
566 if (unlikely(!pfn_valid(pfn)))
567 print_bad_pte(vma, ptent, addr);
569 page = pfn_to_page(pfn);
571 if (unlikely(details) && page) {
573 * unmap_shared_mapping_pages() wants to
574 * invalidate cache without truncating:
575 * unmap shared but keep private pages.
577 if (details->check_mapping &&
578 details->check_mapping != page->mapping)
581 * Each page->index must be checked when
582 * invalidating or truncating nonlinear.
584 if (details->nonlinear_vma &&
585 (page->index < details->first_index ||
586 page->index > details->last_index))
589 ptent = ptep_get_and_clear_full(mm, addr, pte,
591 tlb_remove_tlb_entry(tlb, pte, addr);
594 if (unlikely(details) && details->nonlinear_vma
595 && linear_page_index(details->nonlinear_vma,
596 addr) != page->index)
597 set_pte_at(mm, addr, pte,
598 pgoff_to_pte(page->index));
602 if (pte_dirty(ptent))
603 set_page_dirty(page);
604 if (pte_young(ptent))
605 mark_page_accessed(page);
608 page_remove_rmap(page);
609 tlb_remove_page(tlb, page);
613 * If details->check_mapping, we leave swap entries;
614 * if details->nonlinear_vma, we leave file entries.
616 if (unlikely(details))
618 if (!pte_file(ptent))
619 free_swap_and_cache(pte_to_swp_entry(ptent));
620 pte_clear_full(mm, addr, pte, tlb->fullmm);
621 } while (pte++, addr += PAGE_SIZE, addr != end);
623 add_mm_rss(mm, file_rss, anon_rss);
627 static inline void zap_pmd_range(struct mmu_gather *tlb,
628 struct vm_area_struct *vma, pud_t *pud,
629 unsigned long addr, unsigned long end,
630 struct zap_details *details)
635 pmd = pmd_offset(pud, addr);
637 next = pmd_addr_end(addr, end);
638 if (pmd_none_or_clear_bad(pmd))
640 zap_pte_range(tlb, vma, pmd, addr, next, details);
641 } while (pmd++, addr = next, addr != end);
644 static inline void zap_pud_range(struct mmu_gather *tlb,
645 struct vm_area_struct *vma, pgd_t *pgd,
646 unsigned long addr, unsigned long end,
647 struct zap_details *details)
652 pud = pud_offset(pgd, addr);
654 next = pud_addr_end(addr, end);
655 if (pud_none_or_clear_bad(pud))
657 zap_pmd_range(tlb, vma, pud, addr, next, details);
658 } while (pud++, addr = next, addr != end);
661 static void unmap_page_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
662 unsigned long addr, unsigned long end,
663 struct zap_details *details)
668 if (details && !details->check_mapping && !details->nonlinear_vma)
672 tlb_start_vma(tlb, vma);
673 pgd = pgd_offset(vma->vm_mm, addr);
675 next = pgd_addr_end(addr, end);
676 if (pgd_none_or_clear_bad(pgd))
678 zap_pud_range(tlb, vma, pgd, addr, next, details);
679 } while (pgd++, addr = next, addr != end);
680 tlb_end_vma(tlb, vma);
683 #ifdef CONFIG_PREEMPT
684 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
686 /* No preempt: go for improved straight-line efficiency */
687 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
691 * unmap_vmas - unmap a range of memory covered by a list of vma's
692 * @tlbp: address of the caller's struct mmu_gather
693 * @mm: the controlling mm_struct
694 * @vma: the starting vma
695 * @start_addr: virtual address at which to start unmapping
696 * @end_addr: virtual address at which to end unmapping
697 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
698 * @details: details of nonlinear truncation or shared cache invalidation
700 * Returns the end address of the unmapping (restart addr if interrupted).
702 * Unmap all pages in the vma list. Called under page_table_lock.
704 * We aim to not hold page_table_lock for too long (for scheduling latency
705 * reasons). So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
706 * return the ending mmu_gather to the caller.
708 * Only addresses between `start' and `end' will be unmapped.
710 * The VMA list must be sorted in ascending virtual address order.
712 * unmap_vmas() assumes that the caller will flush the whole unmapped address
713 * range after unmap_vmas() returns. So the only responsibility here is to
714 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
715 * drops the lock and schedules.
717 unsigned long unmap_vmas(struct mmu_gather **tlbp, struct mm_struct *mm,
718 struct vm_area_struct *vma, unsigned long start_addr,
719 unsigned long end_addr, unsigned long *nr_accounted,
720 struct zap_details *details)
722 unsigned long zap_bytes = ZAP_BLOCK_SIZE;
723 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
724 int tlb_start_valid = 0;
725 unsigned long start = start_addr;
726 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
727 int fullmm = (*tlbp)->fullmm;
729 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
732 start = max(vma->vm_start, start_addr);
733 if (start >= vma->vm_end)
735 end = min(vma->vm_end, end_addr);
736 if (end <= vma->vm_start)
739 if (vma->vm_flags & VM_ACCOUNT)
740 *nr_accounted += (end - start) >> PAGE_SHIFT;
742 while (start != end) {
745 if (!tlb_start_valid) {
750 if (is_vm_hugetlb_page(vma)) {
752 unmap_hugepage_range(vma, start, end);
754 block = min(zap_bytes, end - start);
755 unmap_page_range(*tlbp, vma, start,
756 start + block, details);
761 if ((long)zap_bytes > 0)
764 tlb_finish_mmu(*tlbp, tlb_start, start);
766 if (need_resched() ||
767 need_lockbreak(&mm->page_table_lock) ||
768 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
770 /* must reset count of rss freed */
771 *tlbp = tlb_gather_mmu(mm, fullmm);
774 spin_unlock(&mm->page_table_lock);
776 spin_lock(&mm->page_table_lock);
779 *tlbp = tlb_gather_mmu(mm, fullmm);
781 zap_bytes = ZAP_BLOCK_SIZE;
785 return start; /* which is now the end (or restart) address */
789 * zap_page_range - remove user pages in a given range
790 * @vma: vm_area_struct holding the applicable pages
791 * @address: starting address of pages to zap
792 * @size: number of bytes to zap
793 * @details: details of nonlinear truncation or shared cache invalidation
795 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
796 unsigned long size, struct zap_details *details)
798 struct mm_struct *mm = vma->vm_mm;
799 struct mmu_gather *tlb;
800 unsigned long end = address + size;
801 unsigned long nr_accounted = 0;
803 if (is_vm_hugetlb_page(vma)) {
804 zap_hugepage_range(vma, address, size);
809 tlb = tlb_gather_mmu(mm, 0);
810 update_hiwater_rss(mm);
811 spin_lock(&mm->page_table_lock);
812 end = unmap_vmas(&tlb, mm, vma, address, end, &nr_accounted, details);
813 spin_unlock(&mm->page_table_lock);
814 tlb_finish_mmu(tlb, address, end);
819 * Do a quick page-table lookup for a single page.
820 * mm->page_table_lock must be held.
822 static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
823 int read, int write, int accessed)
832 page = follow_huge_addr(mm, address, write);
836 pgd = pgd_offset(mm, address);
837 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
840 pud = pud_offset(pgd, address);
841 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
844 pmd = pmd_offset(pud, address);
845 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
848 return follow_huge_pmd(mm, address, pmd, write);
850 ptep = pte_offset_map(pmd, address);
856 if (pte_present(pte)) {
857 if (write && !pte_write(pte))
859 if (read && !pte_read(pte))
862 if (pfn_valid(pfn)) {
863 page = pfn_to_page(pfn);
865 if (write && !pte_dirty(pte) &&!PageDirty(page))
866 set_page_dirty(page);
867 mark_page_accessed(page);
878 follow_page(struct mm_struct *mm, unsigned long address, int write)
880 return __follow_page(mm, address, 0, write, 1);
884 * check_user_page_readable() can be called frm niterrupt context by oprofile,
885 * so we need to avoid taking any non-irq-safe locks
887 int check_user_page_readable(struct mm_struct *mm, unsigned long address)
889 return __follow_page(mm, address, 1, 0, 0) != NULL;
891 EXPORT_SYMBOL(check_user_page_readable);
894 untouched_anonymous_page(struct mm_struct* mm, struct vm_area_struct *vma,
895 unsigned long address)
901 /* Check if the vma is for an anonymous mapping. */
902 if (vma->vm_ops && vma->vm_ops->nopage)
905 /* Check if page directory entry exists. */
906 pgd = pgd_offset(mm, address);
907 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
910 pud = pud_offset(pgd, address);
911 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
914 /* Check if page middle directory entry exists. */
915 pmd = pmd_offset(pud, address);
916 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
919 /* There is a pte slot for 'address' in 'mm'. */
923 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
924 unsigned long start, int len, int write, int force,
925 struct page **pages, struct vm_area_struct **vmas)
931 * Require read or write permissions.
932 * If 'force' is set, we only require the "MAY" flags.
934 flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
935 flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
939 struct vm_area_struct * vma;
941 vma = find_extend_vma(mm, start);
942 if (!vma && in_gate_area(tsk, start)) {
943 unsigned long pg = start & PAGE_MASK;
944 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
949 if (write) /* user gate pages are read-only */
950 return i ? : -EFAULT;
952 pgd = pgd_offset_k(pg);
954 pgd = pgd_offset_gate(mm, pg);
955 BUG_ON(pgd_none(*pgd));
956 pud = pud_offset(pgd, pg);
957 BUG_ON(pud_none(*pud));
958 pmd = pmd_offset(pud, pg);
960 return i ? : -EFAULT;
961 pte = pte_offset_map(pmd, pg);
962 if (pte_none(*pte)) {
964 return i ? : -EFAULT;
967 pages[i] = pte_page(*pte);
979 if (!vma || (vma->vm_flags & (VM_IO | VM_RESERVED))
980 || !(flags & vma->vm_flags))
981 return i ? : -EFAULT;
983 if (is_vm_hugetlb_page(vma)) {
984 i = follow_hugetlb_page(mm, vma, pages, vmas,
988 spin_lock(&mm->page_table_lock);
990 int write_access = write;
993 cond_resched_lock(&mm->page_table_lock);
994 while (!(page = follow_page(mm, start, write_access))) {
998 * Shortcut for anonymous pages. We don't want
999 * to force the creation of pages tables for
1000 * insanely big anonymously mapped areas that
1001 * nobody touched so far. This is important
1002 * for doing a core dump for these mappings.
1004 if (!write && untouched_anonymous_page(mm,vma,start)) {
1005 page = ZERO_PAGE(start);
1008 spin_unlock(&mm->page_table_lock);
1009 ret = __handle_mm_fault(mm, vma, start, write_access);
1012 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1013 * broken COW when necessary, even if maybe_mkwrite
1014 * decided not to set pte_write. We can thus safely do
1015 * subsequent page lookups as if they were reads.
1017 if (ret & VM_FAULT_WRITE)
1020 switch (ret & ~VM_FAULT_WRITE) {
1021 case VM_FAULT_MINOR:
1024 case VM_FAULT_MAJOR:
1027 case VM_FAULT_SIGBUS:
1028 return i ? i : -EFAULT;
1030 return i ? i : -ENOMEM;
1034 spin_lock(&mm->page_table_lock);
1038 flush_dcache_page(page);
1039 page_cache_get(page);
1046 } while (len && start < vma->vm_end);
1047 spin_unlock(&mm->page_table_lock);
1051 EXPORT_SYMBOL(get_user_pages);
1053 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1054 unsigned long addr, unsigned long end, pgprot_t prot)
1059 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1063 struct page *page = ZERO_PAGE(addr);
1064 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1065 page_cache_get(page);
1066 page_add_file_rmap(page);
1067 inc_mm_counter(mm, file_rss);
1068 BUG_ON(!pte_none(*pte));
1069 set_pte_at(mm, addr, pte, zero_pte);
1070 } while (pte++, addr += PAGE_SIZE, addr != end);
1071 pte_unmap_unlock(pte - 1, ptl);
1075 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1076 unsigned long addr, unsigned long end, pgprot_t prot)
1081 pmd = pmd_alloc(mm, pud, addr);
1085 next = pmd_addr_end(addr, end);
1086 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1088 } while (pmd++, addr = next, addr != end);
1092 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1093 unsigned long addr, unsigned long end, pgprot_t prot)
1098 pud = pud_alloc(mm, pgd, addr);
1102 next = pud_addr_end(addr, end);
1103 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1105 } while (pud++, addr = next, addr != end);
1109 int zeromap_page_range(struct vm_area_struct *vma,
1110 unsigned long addr, unsigned long size, pgprot_t prot)
1114 unsigned long end = addr + size;
1115 struct mm_struct *mm = vma->vm_mm;
1118 BUG_ON(addr >= end);
1119 pgd = pgd_offset(mm, addr);
1120 flush_cache_range(vma, addr, end);
1122 next = pgd_addr_end(addr, end);
1123 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1126 } while (pgd++, addr = next, addr != end);
1131 * maps a range of physical memory into the requested pages. the old
1132 * mappings are removed. any references to nonexistent pages results
1133 * in null mappings (currently treated as "copy-on-access")
1135 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1136 unsigned long addr, unsigned long end,
1137 unsigned long pfn, pgprot_t prot)
1142 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1146 BUG_ON(!pte_none(*pte));
1147 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1149 } while (pte++, addr += PAGE_SIZE, addr != end);
1150 pte_unmap_unlock(pte - 1, ptl);
1154 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1155 unsigned long addr, unsigned long end,
1156 unsigned long pfn, pgprot_t prot)
1161 pfn -= addr >> PAGE_SHIFT;
1162 pmd = pmd_alloc(mm, pud, addr);
1166 next = pmd_addr_end(addr, end);
1167 if (remap_pte_range(mm, pmd, addr, next,
1168 pfn + (addr >> PAGE_SHIFT), prot))
1170 } while (pmd++, addr = next, addr != end);
1174 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1175 unsigned long addr, unsigned long end,
1176 unsigned long pfn, pgprot_t prot)
1181 pfn -= addr >> PAGE_SHIFT;
1182 pud = pud_alloc(mm, pgd, addr);
1186 next = pud_addr_end(addr, end);
1187 if (remap_pmd_range(mm, pud, addr, next,
1188 pfn + (addr >> PAGE_SHIFT), prot))
1190 } while (pud++, addr = next, addr != end);
1194 /* Note: this is only safe if the mm semaphore is held when called. */
1195 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1196 unsigned long pfn, unsigned long size, pgprot_t prot)
1200 unsigned long end = addr + PAGE_ALIGN(size);
1201 struct mm_struct *mm = vma->vm_mm;
1205 * Physically remapped pages are special. Tell the
1206 * rest of the world about it:
1207 * VM_IO tells people not to look at these pages
1208 * (accesses can have side effects).
1209 * VM_RESERVED tells the core MM not to "manage" these pages
1210 * (e.g. refcount, mapcount, try to swap them out).
1212 vma->vm_flags |= VM_IO | VM_RESERVED;
1214 BUG_ON(addr >= end);
1215 pfn -= addr >> PAGE_SHIFT;
1216 pgd = pgd_offset(mm, addr);
1217 flush_cache_range(vma, addr, end);
1219 next = pgd_addr_end(addr, end);
1220 err = remap_pud_range(mm, pgd, addr, next,
1221 pfn + (addr >> PAGE_SHIFT), prot);
1224 } while (pgd++, addr = next, addr != end);
1227 EXPORT_SYMBOL(remap_pfn_range);
1230 * handle_pte_fault chooses page fault handler according to an entry
1231 * which was read non-atomically. Before making any commitment, on
1232 * those architectures or configurations (e.g. i386 with PAE) which
1233 * might give a mix of unmatched parts, do_swap_page and do_file_page
1234 * must check under lock before unmapping the pte and proceeding
1235 * (but do_wp_page is only called after already making such a check;
1236 * and do_anonymous_page and do_no_page can safely check later on).
1238 static inline int pte_unmap_same(struct mm_struct *mm,
1239 pte_t *page_table, pte_t orig_pte)
1242 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1243 if (sizeof(pte_t) > sizeof(unsigned long)) {
1244 spin_lock(&mm->page_table_lock);
1245 same = pte_same(*page_table, orig_pte);
1246 spin_unlock(&mm->page_table_lock);
1249 pte_unmap(page_table);
1254 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1255 * servicing faults for write access. In the normal case, do always want
1256 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1257 * that do not have writing enabled, when used by access_process_vm.
1259 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1261 if (likely(vma->vm_flags & VM_WRITE))
1262 pte = pte_mkwrite(pte);
1267 * This routine handles present pages, when users try to write
1268 * to a shared page. It is done by copying the page to a new address
1269 * and decrementing the shared-page counter for the old page.
1271 * Note that this routine assumes that the protection checks have been
1272 * done by the caller (the low-level page fault routine in most cases).
1273 * Thus we can safely just mark it writable once we've done any necessary
1276 * We also mark the page dirty at this point even though the page will
1277 * change only once the write actually happens. This avoids a few races,
1278 * and potentially makes it more efficient.
1280 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1281 * but allow concurrent faults), with pte both mapped and locked.
1282 * We return with mmap_sem still held, but pte unmapped and unlocked.
1284 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1285 unsigned long address, pte_t *page_table, pmd_t *pmd,
1286 spinlock_t *ptl, pte_t orig_pte)
1288 struct page *old_page, *new_page;
1289 unsigned long pfn = pte_pfn(orig_pte);
1291 int ret = VM_FAULT_MINOR;
1293 BUG_ON(vma->vm_flags & VM_RESERVED);
1295 if (unlikely(!pfn_valid(pfn))) {
1297 * Page table corrupted: show pte and kill process.
1299 print_bad_pte(vma, orig_pte, address);
1303 old_page = pfn_to_page(pfn);
1305 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1306 int reuse = can_share_swap_page(old_page);
1307 unlock_page(old_page);
1309 flush_cache_page(vma, address, pfn);
1310 entry = pte_mkyoung(orig_pte);
1311 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1312 ptep_set_access_flags(vma, address, page_table, entry, 1);
1313 update_mmu_cache(vma, address, entry);
1314 lazy_mmu_prot_update(entry);
1315 ret |= VM_FAULT_WRITE;
1321 * Ok, we need to copy. Oh, well..
1323 page_cache_get(old_page);
1324 pte_unmap_unlock(page_table, ptl);
1326 if (unlikely(anon_vma_prepare(vma)))
1328 if (old_page == ZERO_PAGE(address)) {
1329 new_page = alloc_zeroed_user_highpage(vma, address);
1333 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1336 copy_user_highpage(new_page, old_page, address);
1340 * Re-check the pte - we dropped the lock
1342 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1343 if (likely(pte_same(*page_table, orig_pte))) {
1344 page_remove_rmap(old_page);
1345 if (!PageAnon(old_page)) {
1346 inc_mm_counter(mm, anon_rss);
1347 dec_mm_counter(mm, file_rss);
1349 flush_cache_page(vma, address, pfn);
1350 entry = mk_pte(new_page, vma->vm_page_prot);
1351 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1352 ptep_establish(vma, address, page_table, entry);
1353 update_mmu_cache(vma, address, entry);
1354 lazy_mmu_prot_update(entry);
1355 lru_cache_add_active(new_page);
1356 page_add_anon_rmap(new_page, vma, address);
1358 /* Free the old page.. */
1359 new_page = old_page;
1360 ret |= VM_FAULT_WRITE;
1362 page_cache_release(new_page);
1363 page_cache_release(old_page);
1365 pte_unmap_unlock(page_table, ptl);
1368 page_cache_release(old_page);
1369 return VM_FAULT_OOM;
1373 * Helper functions for unmap_mapping_range().
1375 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1377 * We have to restart searching the prio_tree whenever we drop the lock,
1378 * since the iterator is only valid while the lock is held, and anyway
1379 * a later vma might be split and reinserted earlier while lock dropped.
1381 * The list of nonlinear vmas could be handled more efficiently, using
1382 * a placeholder, but handle it in the same way until a need is shown.
1383 * It is important to search the prio_tree before nonlinear list: a vma
1384 * may become nonlinear and be shifted from prio_tree to nonlinear list
1385 * while the lock is dropped; but never shifted from list to prio_tree.
1387 * In order to make forward progress despite restarting the search,
1388 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1389 * quickly skip it next time around. Since the prio_tree search only
1390 * shows us those vmas affected by unmapping the range in question, we
1391 * can't efficiently keep all vmas in step with mapping->truncate_count:
1392 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1393 * mapping->truncate_count and vma->vm_truncate_count are protected by
1396 * In order to make forward progress despite repeatedly restarting some
1397 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1398 * and restart from that address when we reach that vma again. It might
1399 * have been split or merged, shrunk or extended, but never shifted: so
1400 * restart_addr remains valid so long as it remains in the vma's range.
1401 * unmap_mapping_range forces truncate_count to leap over page-aligned
1402 * values so we can save vma's restart_addr in its truncate_count field.
1404 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1406 static void reset_vma_truncate_counts(struct address_space *mapping)
1408 struct vm_area_struct *vma;
1409 struct prio_tree_iter iter;
1411 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1412 vma->vm_truncate_count = 0;
1413 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1414 vma->vm_truncate_count = 0;
1417 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1418 unsigned long start_addr, unsigned long end_addr,
1419 struct zap_details *details)
1421 unsigned long restart_addr;
1425 restart_addr = vma->vm_truncate_count;
1426 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1427 start_addr = restart_addr;
1428 if (start_addr >= end_addr) {
1429 /* Top of vma has been split off since last time */
1430 vma->vm_truncate_count = details->truncate_count;
1435 restart_addr = zap_page_range(vma, start_addr,
1436 end_addr - start_addr, details);
1439 * We cannot rely on the break test in unmap_vmas:
1440 * on the one hand, we don't want to restart our loop
1441 * just because that broke out for the page_table_lock;
1442 * on the other hand, it does no test when vma is small.
1444 need_break = need_resched() ||
1445 need_lockbreak(details->i_mmap_lock);
1447 if (restart_addr >= end_addr) {
1448 /* We have now completed this vma: mark it so */
1449 vma->vm_truncate_count = details->truncate_count;
1453 /* Note restart_addr in vma's truncate_count field */
1454 vma->vm_truncate_count = restart_addr;
1459 spin_unlock(details->i_mmap_lock);
1461 spin_lock(details->i_mmap_lock);
1465 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1466 struct zap_details *details)
1468 struct vm_area_struct *vma;
1469 struct prio_tree_iter iter;
1470 pgoff_t vba, vea, zba, zea;
1473 vma_prio_tree_foreach(vma, &iter, root,
1474 details->first_index, details->last_index) {
1475 /* Skip quickly over those we have already dealt with */
1476 if (vma->vm_truncate_count == details->truncate_count)
1479 vba = vma->vm_pgoff;
1480 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1481 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1482 zba = details->first_index;
1485 zea = details->last_index;
1489 if (unmap_mapping_range_vma(vma,
1490 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1491 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1497 static inline void unmap_mapping_range_list(struct list_head *head,
1498 struct zap_details *details)
1500 struct vm_area_struct *vma;
1503 * In nonlinear VMAs there is no correspondence between virtual address
1504 * offset and file offset. So we must perform an exhaustive search
1505 * across *all* the pages in each nonlinear VMA, not just the pages
1506 * whose virtual address lies outside the file truncation point.
1509 list_for_each_entry(vma, head, shared.vm_set.list) {
1510 /* Skip quickly over those we have already dealt with */
1511 if (vma->vm_truncate_count == details->truncate_count)
1513 details->nonlinear_vma = vma;
1514 if (unmap_mapping_range_vma(vma, vma->vm_start,
1515 vma->vm_end, details) < 0)
1521 * unmap_mapping_range - unmap the portion of all mmaps
1522 * in the specified address_space corresponding to the specified
1523 * page range in the underlying file.
1524 * @mapping: the address space containing mmaps to be unmapped.
1525 * @holebegin: byte in first page to unmap, relative to the start of
1526 * the underlying file. This will be rounded down to a PAGE_SIZE
1527 * boundary. Note that this is different from vmtruncate(), which
1528 * must keep the partial page. In contrast, we must get rid of
1530 * @holelen: size of prospective hole in bytes. This will be rounded
1531 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1533 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1534 * but 0 when invalidating pagecache, don't throw away private data.
1536 void unmap_mapping_range(struct address_space *mapping,
1537 loff_t const holebegin, loff_t const holelen, int even_cows)
1539 struct zap_details details;
1540 pgoff_t hba = holebegin >> PAGE_SHIFT;
1541 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1543 /* Check for overflow. */
1544 if (sizeof(holelen) > sizeof(hlen)) {
1546 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1547 if (holeend & ~(long long)ULONG_MAX)
1548 hlen = ULONG_MAX - hba + 1;
1551 details.check_mapping = even_cows? NULL: mapping;
1552 details.nonlinear_vma = NULL;
1553 details.first_index = hba;
1554 details.last_index = hba + hlen - 1;
1555 if (details.last_index < details.first_index)
1556 details.last_index = ULONG_MAX;
1557 details.i_mmap_lock = &mapping->i_mmap_lock;
1559 spin_lock(&mapping->i_mmap_lock);
1561 /* serialize i_size write against truncate_count write */
1563 /* Protect against page faults, and endless unmapping loops */
1564 mapping->truncate_count++;
1566 * For archs where spin_lock has inclusive semantics like ia64
1567 * this smp_mb() will prevent to read pagetable contents
1568 * before the truncate_count increment is visible to
1572 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1573 if (mapping->truncate_count == 0)
1574 reset_vma_truncate_counts(mapping);
1575 mapping->truncate_count++;
1577 details.truncate_count = mapping->truncate_count;
1579 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1580 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1581 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1582 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1583 spin_unlock(&mapping->i_mmap_lock);
1585 EXPORT_SYMBOL(unmap_mapping_range);
1588 * Handle all mappings that got truncated by a "truncate()"
1591 * NOTE! We have to be ready to update the memory sharing
1592 * between the file and the memory map for a potential last
1593 * incomplete page. Ugly, but necessary.
1595 int vmtruncate(struct inode * inode, loff_t offset)
1597 struct address_space *mapping = inode->i_mapping;
1598 unsigned long limit;
1600 if (inode->i_size < offset)
1603 * truncation of in-use swapfiles is disallowed - it would cause
1604 * subsequent swapout to scribble on the now-freed blocks.
1606 if (IS_SWAPFILE(inode))
1608 i_size_write(inode, offset);
1609 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1610 truncate_inode_pages(mapping, offset);
1614 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1615 if (limit != RLIM_INFINITY && offset > limit)
1617 if (offset > inode->i_sb->s_maxbytes)
1619 i_size_write(inode, offset);
1622 if (inode->i_op && inode->i_op->truncate)
1623 inode->i_op->truncate(inode);
1626 send_sig(SIGXFSZ, current, 0);
1633 EXPORT_SYMBOL(vmtruncate);
1636 * Primitive swap readahead code. We simply read an aligned block of
1637 * (1 << page_cluster) entries in the swap area. This method is chosen
1638 * because it doesn't cost us any seek time. We also make sure to queue
1639 * the 'original' request together with the readahead ones...
1641 * This has been extended to use the NUMA policies from the mm triggering
1644 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1646 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1649 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1652 struct page *new_page;
1653 unsigned long offset;
1656 * Get the number of handles we should do readahead io to.
1658 num = valid_swaphandles(entry, &offset);
1659 for (i = 0; i < num; offset++, i++) {
1660 /* Ok, do the async read-ahead now */
1661 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1662 offset), vma, addr);
1665 page_cache_release(new_page);
1668 * Find the next applicable VMA for the NUMA policy.
1674 if (addr >= vma->vm_end) {
1676 next_vma = vma ? vma->vm_next : NULL;
1678 if (vma && addr < vma->vm_start)
1681 if (next_vma && addr >= next_vma->vm_start) {
1683 next_vma = vma->vm_next;
1688 lru_add_drain(); /* Push any new pages onto the LRU now */
1692 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1693 * but allow concurrent faults), and pte mapped but not yet locked.
1694 * We return with mmap_sem still held, but pte unmapped and unlocked.
1696 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1697 unsigned long address, pte_t *page_table, pmd_t *pmd,
1698 int write_access, pte_t orig_pte)
1704 int ret = VM_FAULT_MINOR;
1706 if (!pte_unmap_same(mm, page_table, orig_pte))
1709 entry = pte_to_swp_entry(orig_pte);
1710 page = lookup_swap_cache(entry);
1712 swapin_readahead(entry, address, vma);
1713 page = read_swap_cache_async(entry, vma, address);
1716 * Back out if somebody else faulted in this pte
1717 * while we released the pte lock.
1719 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1720 if (likely(pte_same(*page_table, orig_pte)))
1725 /* Had to read the page from swap area: Major fault */
1726 ret = VM_FAULT_MAJOR;
1727 inc_page_state(pgmajfault);
1731 mark_page_accessed(page);
1735 * Back out if somebody else already faulted in this pte.
1737 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1738 if (unlikely(!pte_same(*page_table, orig_pte)))
1741 if (unlikely(!PageUptodate(page))) {
1742 ret = VM_FAULT_SIGBUS;
1746 /* The page isn't present yet, go ahead with the fault. */
1748 inc_mm_counter(mm, anon_rss);
1749 pte = mk_pte(page, vma->vm_page_prot);
1750 if (write_access && can_share_swap_page(page)) {
1751 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1755 flush_icache_page(vma, page);
1756 set_pte_at(mm, address, page_table, pte);
1757 page_add_anon_rmap(page, vma, address);
1761 remove_exclusive_swap_page(page);
1765 if (do_wp_page(mm, vma, address,
1766 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1771 /* No need to invalidate - it was non-present before */
1772 update_mmu_cache(vma, address, pte);
1773 lazy_mmu_prot_update(pte);
1775 pte_unmap_unlock(page_table, ptl);
1779 pte_unmap_unlock(page_table, ptl);
1781 page_cache_release(page);
1786 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1787 * but allow concurrent faults), and pte mapped but not yet locked.
1788 * We return with mmap_sem still held, but pte unmapped and unlocked.
1790 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1791 unsigned long address, pte_t *page_table, pmd_t *pmd,
1799 /* Allocate our own private page. */
1800 pte_unmap(page_table);
1802 if (unlikely(anon_vma_prepare(vma)))
1804 page = alloc_zeroed_user_highpage(vma, address);
1808 entry = mk_pte(page, vma->vm_page_prot);
1809 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1811 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1812 if (!pte_none(*page_table))
1814 inc_mm_counter(mm, anon_rss);
1815 lru_cache_add_active(page);
1816 SetPageReferenced(page);
1817 page_add_anon_rmap(page, vma, address);
1819 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1820 page = ZERO_PAGE(address);
1821 page_cache_get(page);
1822 entry = mk_pte(page, vma->vm_page_prot);
1824 ptl = &mm->page_table_lock;
1826 if (!pte_none(*page_table))
1828 inc_mm_counter(mm, file_rss);
1829 page_add_file_rmap(page);
1832 set_pte_at(mm, address, page_table, entry);
1834 /* No need to invalidate - it was non-present before */
1835 update_mmu_cache(vma, address, entry);
1836 lazy_mmu_prot_update(entry);
1838 pte_unmap_unlock(page_table, ptl);
1839 return VM_FAULT_MINOR;
1841 page_cache_release(page);
1844 return VM_FAULT_OOM;
1848 * do_no_page() tries to create a new page mapping. It aggressively
1849 * tries to share with existing pages, but makes a separate copy if
1850 * the "write_access" parameter is true in order to avoid the next
1853 * As this is called only for pages that do not currently exist, we
1854 * do not need to flush old virtual caches or the TLB.
1856 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1857 * but allow concurrent faults), and pte mapped but not yet locked.
1858 * We return with mmap_sem still held, but pte unmapped and unlocked.
1860 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1861 unsigned long address, pte_t *page_table, pmd_t *pmd,
1865 struct page *new_page;
1866 struct address_space *mapping = NULL;
1868 unsigned int sequence = 0;
1869 int ret = VM_FAULT_MINOR;
1872 pte_unmap(page_table);
1875 mapping = vma->vm_file->f_mapping;
1876 sequence = mapping->truncate_count;
1877 smp_rmb(); /* serializes i_size against truncate_count */
1880 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1882 * No smp_rmb is needed here as long as there's a full
1883 * spin_lock/unlock sequence inside the ->nopage callback
1884 * (for the pagecache lookup) that acts as an implicit
1885 * smp_mb() and prevents the i_size read to happen
1886 * after the next truncate_count read.
1889 /* no page was available -- either SIGBUS or OOM */
1890 if (new_page == NOPAGE_SIGBUS)
1891 return VM_FAULT_SIGBUS;
1892 if (new_page == NOPAGE_OOM)
1893 return VM_FAULT_OOM;
1896 * Should we do an early C-O-W break?
1898 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1901 if (unlikely(anon_vma_prepare(vma)))
1903 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1906 copy_user_highpage(page, new_page, address);
1907 page_cache_release(new_page);
1912 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1914 * For a file-backed vma, someone could have truncated or otherwise
1915 * invalidated this page. If unmap_mapping_range got called,
1916 * retry getting the page.
1918 if (mapping && unlikely(sequence != mapping->truncate_count)) {
1919 pte_unmap_unlock(page_table, ptl);
1920 page_cache_release(new_page);
1922 sequence = mapping->truncate_count;
1928 * This silly early PAGE_DIRTY setting removes a race
1929 * due to the bad i386 page protection. But it's valid
1930 * for other architectures too.
1932 * Note that if write_access is true, we either now have
1933 * an exclusive copy of the page, or this is a shared mapping,
1934 * so we can make it writable and dirty to avoid having to
1935 * handle that later.
1937 /* Only go through if we didn't race with anybody else... */
1938 if (pte_none(*page_table)) {
1939 flush_icache_page(vma, new_page);
1940 entry = mk_pte(new_page, vma->vm_page_prot);
1942 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1943 set_pte_at(mm, address, page_table, entry);
1945 inc_mm_counter(mm, anon_rss);
1946 lru_cache_add_active(new_page);
1947 page_add_anon_rmap(new_page, vma, address);
1948 } else if (!(vma->vm_flags & VM_RESERVED)) {
1949 inc_mm_counter(mm, file_rss);
1950 page_add_file_rmap(new_page);
1953 /* One of our sibling threads was faster, back out. */
1954 page_cache_release(new_page);
1958 /* no need to invalidate: a not-present page shouldn't be cached */
1959 update_mmu_cache(vma, address, entry);
1960 lazy_mmu_prot_update(entry);
1962 pte_unmap_unlock(page_table, ptl);
1965 page_cache_release(new_page);
1966 return VM_FAULT_OOM;
1970 * Fault of a previously existing named mapping. Repopulate the pte
1971 * from the encoded file_pte if possible. This enables swappable
1974 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1975 * but allow concurrent faults), and pte mapped but not yet locked.
1976 * We return with mmap_sem still held, but pte unmapped and unlocked.
1978 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1979 unsigned long address, pte_t *page_table, pmd_t *pmd,
1980 int write_access, pte_t orig_pte)
1985 if (!pte_unmap_same(mm, page_table, orig_pte))
1986 return VM_FAULT_MINOR;
1988 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1990 * Page table corrupted: show pte and kill process.
1992 print_bad_pte(vma, orig_pte, address);
1993 return VM_FAULT_OOM;
1995 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1997 pgoff = pte_to_pgoff(orig_pte);
1998 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1999 vma->vm_page_prot, pgoff, 0);
2001 return VM_FAULT_OOM;
2003 return VM_FAULT_SIGBUS;
2004 return VM_FAULT_MAJOR;
2008 * These routines also need to handle stuff like marking pages dirty
2009 * and/or accessed for architectures that don't do it in hardware (most
2010 * RISC architectures). The early dirtying is also good on the i386.
2012 * There is also a hook called "update_mmu_cache()" that architectures
2013 * with external mmu caches can use to update those (ie the Sparc or
2014 * PowerPC hashed page tables that act as extended TLBs).
2016 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2017 * but allow concurrent faults), and pte mapped but not yet locked.
2018 * We return with mmap_sem still held, but pte unmapped and unlocked.
2020 static inline int handle_pte_fault(struct mm_struct *mm,
2021 struct vm_area_struct *vma, unsigned long address,
2022 pte_t *pte, pmd_t *pmd, int write_access)
2028 if (!pte_present(entry)) {
2029 if (pte_none(entry)) {
2030 if (!vma->vm_ops || !vma->vm_ops->nopage)
2031 return do_anonymous_page(mm, vma, address,
2032 pte, pmd, write_access);
2033 return do_no_page(mm, vma, address,
2034 pte, pmd, write_access);
2036 if (pte_file(entry))
2037 return do_file_page(mm, vma, address,
2038 pte, pmd, write_access, entry);
2039 return do_swap_page(mm, vma, address,
2040 pte, pmd, write_access, entry);
2043 ptl = &mm->page_table_lock;
2045 if (unlikely(!pte_same(*pte, entry)))
2048 if (!pte_write(entry))
2049 return do_wp_page(mm, vma, address,
2050 pte, pmd, ptl, entry);
2051 entry = pte_mkdirty(entry);
2053 entry = pte_mkyoung(entry);
2054 ptep_set_access_flags(vma, address, pte, entry, write_access);
2055 update_mmu_cache(vma, address, entry);
2056 lazy_mmu_prot_update(entry);
2058 pte_unmap_unlock(pte, ptl);
2059 return VM_FAULT_MINOR;
2063 * By the time we get here, we already hold the mm semaphore
2065 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2066 unsigned long address, int write_access)
2073 __set_current_state(TASK_RUNNING);
2075 inc_page_state(pgfault);
2077 if (unlikely(is_vm_hugetlb_page(vma)))
2078 return hugetlb_fault(mm, vma, address, write_access);
2080 pgd = pgd_offset(mm, address);
2081 pud = pud_alloc(mm, pgd, address);
2083 return VM_FAULT_OOM;
2084 pmd = pmd_alloc(mm, pud, address);
2086 return VM_FAULT_OOM;
2087 pte = pte_alloc_map(mm, pmd, address);
2089 return VM_FAULT_OOM;
2091 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2094 #ifndef __PAGETABLE_PUD_FOLDED
2096 * Allocate page upper directory.
2097 * We've already handled the fast-path in-line.
2099 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2101 pud_t *new = pud_alloc_one(mm, address);
2105 spin_lock(&mm->page_table_lock);
2106 if (pgd_present(*pgd)) /* Another has populated it */
2109 pgd_populate(mm, pgd, new);
2110 spin_unlock(&mm->page_table_lock);
2113 #endif /* __PAGETABLE_PUD_FOLDED */
2115 #ifndef __PAGETABLE_PMD_FOLDED
2117 * Allocate page middle directory.
2118 * We've already handled the fast-path in-line.
2120 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2122 pmd_t *new = pmd_alloc_one(mm, address);
2126 spin_lock(&mm->page_table_lock);
2127 #ifndef __ARCH_HAS_4LEVEL_HACK
2128 if (pud_present(*pud)) /* Another has populated it */
2131 pud_populate(mm, pud, new);
2133 if (pgd_present(*pud)) /* Another has populated it */
2136 pgd_populate(mm, pud, new);
2137 #endif /* __ARCH_HAS_4LEVEL_HACK */
2138 spin_unlock(&mm->page_table_lock);
2141 #endif /* __PAGETABLE_PMD_FOLDED */
2143 int make_pages_present(unsigned long addr, unsigned long end)
2145 int ret, len, write;
2146 struct vm_area_struct * vma;
2148 vma = find_vma(current->mm, addr);
2151 write = (vma->vm_flags & VM_WRITE) != 0;
2154 if (end > vma->vm_end)
2156 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2157 ret = get_user_pages(current, current->mm, addr,
2158 len, write, 0, NULL, NULL);
2161 return ret == len ? 0 : -1;
2165 * Map a vmalloc()-space virtual address to the physical page.
2167 struct page * vmalloc_to_page(void * vmalloc_addr)
2169 unsigned long addr = (unsigned long) vmalloc_addr;
2170 struct page *page = NULL;
2171 pgd_t *pgd = pgd_offset_k(addr);
2176 if (!pgd_none(*pgd)) {
2177 pud = pud_offset(pgd, addr);
2178 if (!pud_none(*pud)) {
2179 pmd = pmd_offset(pud, addr);
2180 if (!pmd_none(*pmd)) {
2181 ptep = pte_offset_map(pmd, addr);
2183 if (pte_present(pte))
2184 page = pte_page(pte);
2192 EXPORT_SYMBOL(vmalloc_to_page);
2195 * Map a vmalloc()-space virtual address to the physical page frame number.
2197 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2199 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2202 EXPORT_SYMBOL(vmalloc_to_pfn);
2204 #if !defined(__HAVE_ARCH_GATE_AREA)
2206 #if defined(AT_SYSINFO_EHDR)
2207 static struct vm_area_struct gate_vma;
2209 static int __init gate_vma_init(void)
2211 gate_vma.vm_mm = NULL;
2212 gate_vma.vm_start = FIXADDR_USER_START;
2213 gate_vma.vm_end = FIXADDR_USER_END;
2214 gate_vma.vm_page_prot = PAGE_READONLY;
2215 gate_vma.vm_flags = VM_RESERVED;
2218 __initcall(gate_vma_init);
2221 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2223 #ifdef AT_SYSINFO_EHDR
2230 int in_gate_area_no_task(unsigned long addr)
2232 #ifdef AT_SYSINFO_EHDR
2233 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2239 #endif /* __HAVE_ARCH_GATE_AREA */