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);
85 int randomize_va_space __read_mostly = 1;
87 static int __init disable_randmaps(char *s)
89 randomize_va_space = 0;
92 __setup("norandmaps", disable_randmaps);
96 * If a p?d_bad entry is found while walking page tables, report
97 * the error, before resetting entry to p?d_none. Usually (but
98 * very seldom) called out from the p?d_none_or_clear_bad macros.
101 void pgd_clear_bad(pgd_t *pgd)
107 void pud_clear_bad(pud_t *pud)
113 void pmd_clear_bad(pmd_t *pmd)
120 * Note: this doesn't free the actual pages themselves. That
121 * has been handled earlier when unmapping all the memory regions.
123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
125 struct page *page = pmd_page(*pmd);
127 pte_lock_deinit(page);
128 pte_free_tlb(tlb, page);
129 dec_page_state(nr_page_table_pages);
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134 unsigned long addr, unsigned long end,
135 unsigned long floor, unsigned long ceiling)
142 pmd = pmd_offset(pud, addr);
144 next = pmd_addr_end(addr, end);
145 if (pmd_none_or_clear_bad(pmd))
147 free_pte_range(tlb, pmd);
148 } while (pmd++, addr = next, addr != end);
158 if (end - 1 > ceiling - 1)
161 pmd = pmd_offset(pud, start);
163 pmd_free_tlb(tlb, pmd);
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167 unsigned long addr, unsigned long end,
168 unsigned long floor, unsigned long ceiling)
175 pud = pud_offset(pgd, addr);
177 next = pud_addr_end(addr, end);
178 if (pud_none_or_clear_bad(pud))
180 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181 } while (pud++, addr = next, addr != end);
187 ceiling &= PGDIR_MASK;
191 if (end - 1 > ceiling - 1)
194 pud = pud_offset(pgd, start);
196 pud_free_tlb(tlb, pud);
200 * This function frees user-level page tables of a process.
202 * Must be called with pagetable lock held.
204 void free_pgd_range(struct mmu_gather **tlb,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
213 * The next few lines have given us lots of grief...
215 * Why are we testing PMD* at this top level? Because often
216 * there will be no work to do at all, and we'd prefer not to
217 * go all the way down to the bottom just to discover that.
219 * Why all these "- 1"s? Because 0 represents both the bottom
220 * of the address space and the top of it (using -1 for the
221 * top wouldn't help much: the masks would do the wrong thing).
222 * The rule is that addr 0 and floor 0 refer to the bottom of
223 * the address space, but end 0 and ceiling 0 refer to the top
224 * Comparisons need to use "end - 1" and "ceiling - 1" (though
225 * that end 0 case should be mythical).
227 * Wherever addr is brought up or ceiling brought down, we must
228 * be careful to reject "the opposite 0" before it confuses the
229 * subsequent tests. But what about where end is brought down
230 * by PMD_SIZE below? no, end can't go down to 0 there.
232 * Whereas we round start (addr) and ceiling down, by different
233 * masks at different levels, in order to test whether a table
234 * now has no other vmas using it, so can be freed, we don't
235 * bother to round floor or end up - the tests don't need that.
249 if (end - 1 > ceiling - 1)
255 pgd = pgd_offset((*tlb)->mm, addr);
257 next = pgd_addr_end(addr, end);
258 if (pgd_none_or_clear_bad(pgd))
260 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261 } while (pgd++, addr = next, addr != end);
264 flush_tlb_pgtables((*tlb)->mm, start, end);
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268 unsigned long floor, unsigned long ceiling)
271 struct vm_area_struct *next = vma->vm_next;
272 unsigned long addr = vma->vm_start;
275 * Hide vma from rmap and vmtruncate before freeing pgtables
277 anon_vma_unlink(vma);
278 unlink_file_vma(vma);
280 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
281 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282 floor, next? next->vm_start: ceiling);
285 * Optimization: gather nearby vmas into one call down
287 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288 && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
292 anon_vma_unlink(vma);
293 unlink_file_vma(vma);
295 free_pgd_range(tlb, addr, vma->vm_end,
296 floor, next? next->vm_start: ceiling);
302 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
304 struct page *new = pte_alloc_one(mm, address);
309 spin_lock(&mm->page_table_lock);
310 if (pmd_present(*pmd)) { /* Another has populated it */
311 pte_lock_deinit(new);
315 inc_page_state(nr_page_table_pages);
316 pmd_populate(mm, pmd, new);
318 spin_unlock(&mm->page_table_lock);
322 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
324 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
328 spin_lock(&init_mm.page_table_lock);
329 if (pmd_present(*pmd)) /* Another has populated it */
330 pte_free_kernel(new);
332 pmd_populate_kernel(&init_mm, pmd, new);
333 spin_unlock(&init_mm.page_table_lock);
337 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
340 add_mm_counter(mm, file_rss, file_rss);
342 add_mm_counter(mm, anon_rss, anon_rss);
346 * This function is called to print an error when a bad pte
347 * is found. For example, we might have a PFN-mapped pte in
348 * a region that doesn't allow it.
350 * The calling function must still handle the error.
352 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
354 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
355 "vm_flags = %lx, vaddr = %lx\n",
356 (long long)pte_val(pte),
357 (vma->vm_mm == current->mm ? current->comm : "???"),
358 vma->vm_flags, vaddr);
362 static inline int is_cow_mapping(unsigned int flags)
364 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
368 * This function gets the "struct page" associated with a pte.
370 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
371 * will have each page table entry just pointing to a raw page frame
372 * number, and as far as the VM layer is concerned, those do not have
373 * pages associated with them - even if the PFN might point to memory
374 * that otherwise is perfectly fine and has a "struct page".
376 * The way we recognize those mappings is through the rules set up
377 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
378 * and the vm_pgoff will point to the first PFN mapped: thus every
379 * page that is a raw mapping will always honor the rule
381 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
383 * and if that isn't true, the page has been COW'ed (in which case it
384 * _does_ have a "struct page" associated with it even if it is in a
387 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
389 unsigned long pfn = pte_pfn(pte);
391 if (vma->vm_flags & VM_PFNMAP) {
392 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
393 if (pfn == vma->vm_pgoff + off)
395 if (!is_cow_mapping(vma->vm_flags))
400 * Add some anal sanity checks for now. Eventually,
401 * we should just do "return pfn_to_page(pfn)", but
402 * in the meantime we check that we get a valid pfn,
403 * and that the resulting page looks ok.
405 * Remove this test eventually!
407 if (unlikely(!pfn_valid(pfn))) {
408 print_bad_pte(vma, pte, addr);
413 * NOTE! We still have PageReserved() pages in the page
416 * The PAGE_ZERO() pages and various VDSO mappings can
417 * cause them to exist.
419 return pfn_to_page(pfn);
423 * copy one vm_area from one task to the other. Assumes the page tables
424 * already present in the new task to be cleared in the whole range
425 * covered by this vma.
429 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
430 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
431 unsigned long addr, int *rss)
433 unsigned long vm_flags = vma->vm_flags;
434 pte_t pte = *src_pte;
437 /* pte contains position in swap or file, so copy. */
438 if (unlikely(!pte_present(pte))) {
439 if (!pte_file(pte)) {
440 swap_duplicate(pte_to_swp_entry(pte));
441 /* make sure dst_mm is on swapoff's mmlist. */
442 if (unlikely(list_empty(&dst_mm->mmlist))) {
443 spin_lock(&mmlist_lock);
444 if (list_empty(&dst_mm->mmlist))
445 list_add(&dst_mm->mmlist,
447 spin_unlock(&mmlist_lock);
454 * If it's a COW mapping, write protect it both
455 * in the parent and the child
457 if (is_cow_mapping(vm_flags)) {
458 ptep_set_wrprotect(src_mm, addr, src_pte);
463 * If it's a shared mapping, mark it clean in
466 if (vm_flags & VM_SHARED)
467 pte = pte_mkclean(pte);
468 pte = pte_mkold(pte);
470 page = vm_normal_page(vma, addr, pte);
474 rss[!!PageAnon(page)]++;
478 set_pte_at(dst_mm, addr, dst_pte, pte);
481 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
482 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
483 unsigned long addr, unsigned long end)
485 pte_t *src_pte, *dst_pte;
486 spinlock_t *src_ptl, *dst_ptl;
492 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
495 src_pte = pte_offset_map_nested(src_pmd, addr);
496 src_ptl = pte_lockptr(src_mm, src_pmd);
501 * We are holding two locks at this point - either of them
502 * could generate latencies in another task on another CPU.
504 if (progress >= 32) {
506 if (need_resched() ||
507 need_lockbreak(src_ptl) ||
508 need_lockbreak(dst_ptl))
511 if (pte_none(*src_pte)) {
515 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
517 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
519 spin_unlock(src_ptl);
520 pte_unmap_nested(src_pte - 1);
521 add_mm_rss(dst_mm, rss[0], rss[1]);
522 pte_unmap_unlock(dst_pte - 1, dst_ptl);
529 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
530 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
531 unsigned long addr, unsigned long end)
533 pmd_t *src_pmd, *dst_pmd;
536 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
539 src_pmd = pmd_offset(src_pud, addr);
541 next = pmd_addr_end(addr, end);
542 if (pmd_none_or_clear_bad(src_pmd))
544 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
547 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
551 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
552 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
553 unsigned long addr, unsigned long end)
555 pud_t *src_pud, *dst_pud;
558 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
561 src_pud = pud_offset(src_pgd, addr);
563 next = pud_addr_end(addr, end);
564 if (pud_none_or_clear_bad(src_pud))
566 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
569 } while (dst_pud++, src_pud++, addr = next, addr != end);
573 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
574 struct vm_area_struct *vma)
576 pgd_t *src_pgd, *dst_pgd;
578 unsigned long addr = vma->vm_start;
579 unsigned long end = vma->vm_end;
582 * Don't copy ptes where a page fault will fill them correctly.
583 * Fork becomes much lighter when there are big shared or private
584 * readonly mappings. The tradeoff is that copy_page_range is more
585 * efficient than faulting.
587 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
592 if (is_vm_hugetlb_page(vma))
593 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
595 dst_pgd = pgd_offset(dst_mm, addr);
596 src_pgd = pgd_offset(src_mm, addr);
598 next = pgd_addr_end(addr, end);
599 if (pgd_none_or_clear_bad(src_pgd))
601 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
604 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
608 static unsigned long zap_pte_range(struct mmu_gather *tlb,
609 struct vm_area_struct *vma, pmd_t *pmd,
610 unsigned long addr, unsigned long end,
611 long *zap_work, struct zap_details *details)
613 struct mm_struct *mm = tlb->mm;
619 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
622 if (pte_none(ptent)) {
627 (*zap_work) -= PAGE_SIZE;
629 if (pte_present(ptent)) {
632 page = vm_normal_page(vma, addr, ptent);
633 if (unlikely(details) && page) {
635 * unmap_shared_mapping_pages() wants to
636 * invalidate cache without truncating:
637 * unmap shared but keep private pages.
639 if (details->check_mapping &&
640 details->check_mapping != page->mapping)
643 * Each page->index must be checked when
644 * invalidating or truncating nonlinear.
646 if (details->nonlinear_vma &&
647 (page->index < details->first_index ||
648 page->index > details->last_index))
651 ptent = ptep_get_and_clear_full(mm, addr, pte,
653 tlb_remove_tlb_entry(tlb, pte, addr);
656 if (unlikely(details) && details->nonlinear_vma
657 && linear_page_index(details->nonlinear_vma,
658 addr) != page->index)
659 set_pte_at(mm, addr, pte,
660 pgoff_to_pte(page->index));
664 if (pte_dirty(ptent))
665 set_page_dirty(page);
666 if (pte_young(ptent))
667 mark_page_accessed(page);
670 page_remove_rmap(page);
671 tlb_remove_page(tlb, page);
675 * If details->check_mapping, we leave swap entries;
676 * if details->nonlinear_vma, we leave file entries.
678 if (unlikely(details))
680 if (!pte_file(ptent))
681 free_swap_and_cache(pte_to_swp_entry(ptent));
682 pte_clear_full(mm, addr, pte, tlb->fullmm);
683 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
685 add_mm_rss(mm, file_rss, anon_rss);
686 pte_unmap_unlock(pte - 1, ptl);
691 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
692 struct vm_area_struct *vma, pud_t *pud,
693 unsigned long addr, unsigned long end,
694 long *zap_work, struct zap_details *details)
699 pmd = pmd_offset(pud, addr);
701 next = pmd_addr_end(addr, end);
702 if (pmd_none_or_clear_bad(pmd)) {
706 next = zap_pte_range(tlb, vma, pmd, addr, next,
708 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
713 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
714 struct vm_area_struct *vma, pgd_t *pgd,
715 unsigned long addr, unsigned long end,
716 long *zap_work, struct zap_details *details)
721 pud = pud_offset(pgd, addr);
723 next = pud_addr_end(addr, end);
724 if (pud_none_or_clear_bad(pud)) {
728 next = zap_pmd_range(tlb, vma, pud, addr, next,
730 } while (pud++, addr = next, (addr != end && *zap_work > 0));
735 static unsigned long unmap_page_range(struct mmu_gather *tlb,
736 struct vm_area_struct *vma,
737 unsigned long addr, unsigned long end,
738 long *zap_work, struct zap_details *details)
743 if (details && !details->check_mapping && !details->nonlinear_vma)
747 tlb_start_vma(tlb, vma);
748 pgd = pgd_offset(vma->vm_mm, addr);
750 next = pgd_addr_end(addr, end);
751 if (pgd_none_or_clear_bad(pgd)) {
755 next = zap_pud_range(tlb, vma, pgd, addr, next,
757 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
758 tlb_end_vma(tlb, vma);
763 #ifdef CONFIG_PREEMPT
764 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
766 /* No preempt: go for improved straight-line efficiency */
767 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
771 * unmap_vmas - unmap a range of memory covered by a list of vma's
772 * @tlbp: address of the caller's struct mmu_gather
773 * @vma: the starting vma
774 * @start_addr: virtual address at which to start unmapping
775 * @end_addr: virtual address at which to end unmapping
776 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
777 * @details: details of nonlinear truncation or shared cache invalidation
779 * Returns the end address of the unmapping (restart addr if interrupted).
781 * Unmap all pages in the vma list.
783 * We aim to not hold locks for too long (for scheduling latency reasons).
784 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
785 * return the ending mmu_gather to the caller.
787 * Only addresses between `start' and `end' will be unmapped.
789 * The VMA list must be sorted in ascending virtual address order.
791 * unmap_vmas() assumes that the caller will flush the whole unmapped address
792 * range after unmap_vmas() returns. So the only responsibility here is to
793 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
794 * drops the lock and schedules.
796 unsigned long unmap_vmas(struct mmu_gather **tlbp,
797 struct vm_area_struct *vma, unsigned long start_addr,
798 unsigned long end_addr, unsigned long *nr_accounted,
799 struct zap_details *details)
801 long zap_work = ZAP_BLOCK_SIZE;
802 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
803 int tlb_start_valid = 0;
804 unsigned long start = start_addr;
805 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
806 int fullmm = (*tlbp)->fullmm;
808 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
811 start = max(vma->vm_start, start_addr);
812 if (start >= vma->vm_end)
814 end = min(vma->vm_end, end_addr);
815 if (end <= vma->vm_start)
818 if (vma->vm_flags & VM_ACCOUNT)
819 *nr_accounted += (end - start) >> PAGE_SHIFT;
821 while (start != end) {
822 if (!tlb_start_valid) {
827 if (unlikely(is_vm_hugetlb_page(vma))) {
828 unmap_hugepage_range(vma, start, end);
829 zap_work -= (end - start) /
830 (HPAGE_SIZE / PAGE_SIZE);
833 start = unmap_page_range(*tlbp, vma,
834 start, end, &zap_work, details);
837 BUG_ON(start != end);
841 tlb_finish_mmu(*tlbp, tlb_start, start);
843 if (need_resched() ||
844 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
852 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
854 zap_work = ZAP_BLOCK_SIZE;
858 return start; /* which is now the end (or restart) address */
862 * zap_page_range - remove user pages in a given range
863 * @vma: vm_area_struct holding the applicable pages
864 * @address: starting address of pages to zap
865 * @size: number of bytes to zap
866 * @details: details of nonlinear truncation or shared cache invalidation
868 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
869 unsigned long size, struct zap_details *details)
871 struct mm_struct *mm = vma->vm_mm;
872 struct mmu_gather *tlb;
873 unsigned long end = address + size;
874 unsigned long nr_accounted = 0;
877 tlb = tlb_gather_mmu(mm, 0);
878 update_hiwater_rss(mm);
879 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
881 tlb_finish_mmu(tlb, address, end);
886 * Do a quick page-table lookup for a single page.
888 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
897 struct mm_struct *mm = vma->vm_mm;
899 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
901 BUG_ON(flags & FOLL_GET);
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 pmd = pmd_offset(pud, address);
915 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
918 if (pmd_huge(*pmd)) {
919 BUG_ON(flags & FOLL_GET);
920 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
924 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
929 if (!pte_present(pte))
931 if ((flags & FOLL_WRITE) && !pte_write(pte))
933 page = vm_normal_page(vma, address, pte);
937 if (flags & FOLL_GET)
939 if (flags & FOLL_TOUCH) {
940 if ((flags & FOLL_WRITE) &&
941 !pte_dirty(pte) && !PageDirty(page))
942 set_page_dirty(page);
943 mark_page_accessed(page);
946 pte_unmap_unlock(ptep, ptl);
952 * When core dumping an enormous anonymous area that nobody
953 * has touched so far, we don't want to allocate page tables.
955 if (flags & FOLL_ANON) {
956 page = ZERO_PAGE(address);
957 if (flags & FOLL_GET)
959 BUG_ON(flags & FOLL_WRITE);
964 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
965 unsigned long start, int len, int write, int force,
966 struct page **pages, struct vm_area_struct **vmas)
969 unsigned int vm_flags;
972 * Require read or write permissions.
973 * If 'force' is set, we only require the "MAY" flags.
975 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
976 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
980 struct vm_area_struct *vma;
981 unsigned int foll_flags;
983 vma = find_extend_vma(mm, start);
984 if (!vma && in_gate_area(tsk, start)) {
985 unsigned long pg = start & PAGE_MASK;
986 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
991 if (write) /* user gate pages are read-only */
992 return i ? : -EFAULT;
994 pgd = pgd_offset_k(pg);
996 pgd = pgd_offset_gate(mm, pg);
997 BUG_ON(pgd_none(*pgd));
998 pud = pud_offset(pgd, pg);
999 BUG_ON(pud_none(*pud));
1000 pmd = pmd_offset(pud, pg);
1002 return i ? : -EFAULT;
1003 pte = pte_offset_map(pmd, pg);
1004 if (pte_none(*pte)) {
1006 return i ? : -EFAULT;
1009 struct page *page = vm_normal_page(gate_vma, start, *pte);
1023 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1024 || !(vm_flags & vma->vm_flags))
1025 return i ? : -EFAULT;
1027 if (is_vm_hugetlb_page(vma)) {
1028 i = follow_hugetlb_page(mm, vma, pages, vmas,
1033 foll_flags = FOLL_TOUCH;
1035 foll_flags |= FOLL_GET;
1036 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1037 (!vma->vm_ops || !vma->vm_ops->nopage))
1038 foll_flags |= FOLL_ANON;
1044 foll_flags |= FOLL_WRITE;
1047 while (!(page = follow_page(vma, start, foll_flags))) {
1049 ret = __handle_mm_fault(mm, vma, start,
1050 foll_flags & FOLL_WRITE);
1052 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1053 * broken COW when necessary, even if maybe_mkwrite
1054 * decided not to set pte_write. We can thus safely do
1055 * subsequent page lookups as if they were reads.
1057 if (ret & VM_FAULT_WRITE)
1058 foll_flags &= ~FOLL_WRITE;
1060 switch (ret & ~VM_FAULT_WRITE) {
1061 case VM_FAULT_MINOR:
1064 case VM_FAULT_MAJOR:
1067 case VM_FAULT_SIGBUS:
1068 return i ? i : -EFAULT;
1070 return i ? i : -ENOMEM;
1077 flush_dcache_page(page);
1084 } while (len && start < vma->vm_end);
1088 EXPORT_SYMBOL(get_user_pages);
1090 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1091 unsigned long addr, unsigned long end, pgprot_t prot)
1096 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1100 struct page *page = ZERO_PAGE(addr);
1101 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1102 page_cache_get(page);
1103 page_add_file_rmap(page);
1104 inc_mm_counter(mm, file_rss);
1105 BUG_ON(!pte_none(*pte));
1106 set_pte_at(mm, addr, pte, zero_pte);
1107 } while (pte++, addr += PAGE_SIZE, addr != end);
1108 pte_unmap_unlock(pte - 1, ptl);
1112 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1113 unsigned long addr, unsigned long end, pgprot_t prot)
1118 pmd = pmd_alloc(mm, pud, addr);
1122 next = pmd_addr_end(addr, end);
1123 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1125 } while (pmd++, addr = next, addr != end);
1129 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1130 unsigned long addr, unsigned long end, pgprot_t prot)
1135 pud = pud_alloc(mm, pgd, addr);
1139 next = pud_addr_end(addr, end);
1140 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1142 } while (pud++, addr = next, addr != end);
1146 int zeromap_page_range(struct vm_area_struct *vma,
1147 unsigned long addr, unsigned long size, pgprot_t prot)
1151 unsigned long end = addr + size;
1152 struct mm_struct *mm = vma->vm_mm;
1155 BUG_ON(addr >= end);
1156 pgd = pgd_offset(mm, addr);
1157 flush_cache_range(vma, addr, end);
1159 next = pgd_addr_end(addr, end);
1160 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1163 } while (pgd++, addr = next, addr != end);
1167 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1169 pgd_t * pgd = pgd_offset(mm, addr);
1170 pud_t * pud = pud_alloc(mm, pgd, addr);
1172 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1174 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1180 * This is the old fallback for page remapping.
1182 * For historical reasons, it only allows reserved pages. Only
1183 * old drivers should use this, and they needed to mark their
1184 * pages reserved for the old functions anyway.
1186 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1196 flush_dcache_page(page);
1197 pte = get_locked_pte(mm, addr, &ptl);
1201 if (!pte_none(*pte))
1204 /* Ok, finally just insert the thing.. */
1206 inc_mm_counter(mm, file_rss);
1207 page_add_file_rmap(page);
1208 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1212 pte_unmap_unlock(pte, ptl);
1218 * This allows drivers to insert individual pages they've allocated
1221 * The page has to be a nice clean _individual_ kernel allocation.
1222 * If you allocate a compound page, you need to have marked it as
1223 * such (__GFP_COMP), or manually just split the page up yourself
1224 * (which is mainly an issue of doing "set_page_count(page, 1)" for
1225 * each sub-page, and then freeing them one by one when you free
1226 * them rather than freeing it as a compound page).
1228 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1229 * took an arbitrary page protection parameter. This doesn't allow
1230 * that. Your vma protection will have to be set up correctly, which
1231 * means that if you want a shared writable mapping, you'd better
1232 * ask for a shared writable mapping!
1234 * The page does not need to be reserved.
1236 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1238 if (addr < vma->vm_start || addr >= vma->vm_end)
1240 if (!page_count(page))
1242 vma->vm_flags |= VM_INSERTPAGE;
1243 return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1245 EXPORT_SYMBOL(vm_insert_page);
1248 * maps a range of physical memory into the requested pages. the old
1249 * mappings are removed. any references to nonexistent pages results
1250 * in null mappings (currently treated as "copy-on-access")
1252 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1253 unsigned long addr, unsigned long end,
1254 unsigned long pfn, pgprot_t prot)
1259 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1263 BUG_ON(!pte_none(*pte));
1264 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1266 } while (pte++, addr += PAGE_SIZE, addr != end);
1267 pte_unmap_unlock(pte - 1, ptl);
1271 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1272 unsigned long addr, unsigned long end,
1273 unsigned long pfn, pgprot_t prot)
1278 pfn -= addr >> PAGE_SHIFT;
1279 pmd = pmd_alloc(mm, pud, addr);
1283 next = pmd_addr_end(addr, end);
1284 if (remap_pte_range(mm, pmd, addr, next,
1285 pfn + (addr >> PAGE_SHIFT), prot))
1287 } while (pmd++, addr = next, addr != end);
1291 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1292 unsigned long addr, unsigned long end,
1293 unsigned long pfn, pgprot_t prot)
1298 pfn -= addr >> PAGE_SHIFT;
1299 pud = pud_alloc(mm, pgd, addr);
1303 next = pud_addr_end(addr, end);
1304 if (remap_pmd_range(mm, pud, addr, next,
1305 pfn + (addr >> PAGE_SHIFT), prot))
1307 } while (pud++, addr = next, addr != end);
1311 /* Note: this is only safe if the mm semaphore is held when called. */
1312 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1313 unsigned long pfn, unsigned long size, pgprot_t prot)
1317 unsigned long end = addr + PAGE_ALIGN(size);
1318 struct mm_struct *mm = vma->vm_mm;
1322 * Physically remapped pages are special. Tell the
1323 * rest of the world about it:
1324 * VM_IO tells people not to look at these pages
1325 * (accesses can have side effects).
1326 * VM_RESERVED is specified all over the place, because
1327 * in 2.4 it kept swapout's vma scan off this vma; but
1328 * in 2.6 the LRU scan won't even find its pages, so this
1329 * flag means no more than count its pages in reserved_vm,
1330 * and omit it from core dump, even when VM_IO turned off.
1331 * VM_PFNMAP tells the core MM that the base pages are just
1332 * raw PFN mappings, and do not have a "struct page" associated
1335 * There's a horrible special case to handle copy-on-write
1336 * behaviour that some programs depend on. We mark the "original"
1337 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1339 if (is_cow_mapping(vma->vm_flags)) {
1340 if (addr != vma->vm_start || end != vma->vm_end)
1342 vma->vm_pgoff = pfn;
1345 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1347 BUG_ON(addr >= end);
1348 pfn -= addr >> PAGE_SHIFT;
1349 pgd = pgd_offset(mm, addr);
1350 flush_cache_range(vma, addr, end);
1352 next = pgd_addr_end(addr, end);
1353 err = remap_pud_range(mm, pgd, addr, next,
1354 pfn + (addr >> PAGE_SHIFT), prot);
1357 } while (pgd++, addr = next, addr != end);
1360 EXPORT_SYMBOL(remap_pfn_range);
1363 * handle_pte_fault chooses page fault handler according to an entry
1364 * which was read non-atomically. Before making any commitment, on
1365 * those architectures or configurations (e.g. i386 with PAE) which
1366 * might give a mix of unmatched parts, do_swap_page and do_file_page
1367 * must check under lock before unmapping the pte and proceeding
1368 * (but do_wp_page is only called after already making such a check;
1369 * and do_anonymous_page and do_no_page can safely check later on).
1371 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1372 pte_t *page_table, pte_t orig_pte)
1375 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1376 if (sizeof(pte_t) > sizeof(unsigned long)) {
1377 spinlock_t *ptl = pte_lockptr(mm, pmd);
1379 same = pte_same(*page_table, orig_pte);
1383 pte_unmap(page_table);
1388 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1389 * servicing faults for write access. In the normal case, do always want
1390 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1391 * that do not have writing enabled, when used by access_process_vm.
1393 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1395 if (likely(vma->vm_flags & VM_WRITE))
1396 pte = pte_mkwrite(pte);
1400 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1403 * If the source page was a PFN mapping, we don't have
1404 * a "struct page" for it. We do a best-effort copy by
1405 * just copying from the original user address. If that
1406 * fails, we just zero-fill it. Live with it.
1408 if (unlikely(!src)) {
1409 void *kaddr = kmap_atomic(dst, KM_USER0);
1410 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1413 * This really shouldn't fail, because the page is there
1414 * in the page tables. But it might just be unreadable,
1415 * in which case we just give up and fill the result with
1418 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1419 memset(kaddr, 0, PAGE_SIZE);
1420 kunmap_atomic(kaddr, KM_USER0);
1421 flush_dcache_page(dst);
1425 copy_user_highpage(dst, src, va);
1429 * This routine handles present pages, when users try to write
1430 * to a shared page. It is done by copying the page to a new address
1431 * and decrementing the shared-page counter for the old page.
1433 * Note that this routine assumes that the protection checks have been
1434 * done by the caller (the low-level page fault routine in most cases).
1435 * Thus we can safely just mark it writable once we've done any necessary
1438 * We also mark the page dirty at this point even though the page will
1439 * change only once the write actually happens. This avoids a few races,
1440 * and potentially makes it more efficient.
1442 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1443 * but allow concurrent faults), with pte both mapped and locked.
1444 * We return with mmap_sem still held, but pte unmapped and unlocked.
1446 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1447 unsigned long address, pte_t *page_table, pmd_t *pmd,
1448 spinlock_t *ptl, pte_t orig_pte)
1450 struct page *old_page, *new_page;
1452 int ret = VM_FAULT_MINOR;
1454 old_page = vm_normal_page(vma, address, orig_pte);
1458 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1459 int reuse = can_share_swap_page(old_page);
1460 unlock_page(old_page);
1462 flush_cache_page(vma, address, pte_pfn(orig_pte));
1463 entry = pte_mkyoung(orig_pte);
1464 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1465 ptep_set_access_flags(vma, address, page_table, entry, 1);
1466 update_mmu_cache(vma, address, entry);
1467 lazy_mmu_prot_update(entry);
1468 ret |= VM_FAULT_WRITE;
1474 * Ok, we need to copy. Oh, well..
1476 page_cache_get(old_page);
1478 pte_unmap_unlock(page_table, ptl);
1480 if (unlikely(anon_vma_prepare(vma)))
1482 if (old_page == ZERO_PAGE(address)) {
1483 new_page = alloc_zeroed_user_highpage(vma, address);
1487 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1490 cow_user_page(new_page, old_page, address);
1494 * Re-check the pte - we dropped the lock
1496 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1497 if (likely(pte_same(*page_table, orig_pte))) {
1499 page_remove_rmap(old_page);
1500 if (!PageAnon(old_page)) {
1501 dec_mm_counter(mm, file_rss);
1502 inc_mm_counter(mm, anon_rss);
1505 inc_mm_counter(mm, anon_rss);
1506 flush_cache_page(vma, address, pte_pfn(orig_pte));
1507 entry = mk_pte(new_page, vma->vm_page_prot);
1508 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1509 ptep_establish(vma, address, page_table, entry);
1510 update_mmu_cache(vma, address, entry);
1511 lazy_mmu_prot_update(entry);
1512 lru_cache_add_active(new_page);
1513 page_add_new_anon_rmap(new_page, vma, address);
1515 /* Free the old page.. */
1516 new_page = old_page;
1517 ret |= VM_FAULT_WRITE;
1520 page_cache_release(new_page);
1522 page_cache_release(old_page);
1524 pte_unmap_unlock(page_table, ptl);
1528 page_cache_release(old_page);
1529 return VM_FAULT_OOM;
1533 * Helper functions for unmap_mapping_range().
1535 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1537 * We have to restart searching the prio_tree whenever we drop the lock,
1538 * since the iterator is only valid while the lock is held, and anyway
1539 * a later vma might be split and reinserted earlier while lock dropped.
1541 * The list of nonlinear vmas could be handled more efficiently, using
1542 * a placeholder, but handle it in the same way until a need is shown.
1543 * It is important to search the prio_tree before nonlinear list: a vma
1544 * may become nonlinear and be shifted from prio_tree to nonlinear list
1545 * while the lock is dropped; but never shifted from list to prio_tree.
1547 * In order to make forward progress despite restarting the search,
1548 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1549 * quickly skip it next time around. Since the prio_tree search only
1550 * shows us those vmas affected by unmapping the range in question, we
1551 * can't efficiently keep all vmas in step with mapping->truncate_count:
1552 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1553 * mapping->truncate_count and vma->vm_truncate_count are protected by
1556 * In order to make forward progress despite repeatedly restarting some
1557 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1558 * and restart from that address when we reach that vma again. It might
1559 * have been split or merged, shrunk or extended, but never shifted: so
1560 * restart_addr remains valid so long as it remains in the vma's range.
1561 * unmap_mapping_range forces truncate_count to leap over page-aligned
1562 * values so we can save vma's restart_addr in its truncate_count field.
1564 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1566 static void reset_vma_truncate_counts(struct address_space *mapping)
1568 struct vm_area_struct *vma;
1569 struct prio_tree_iter iter;
1571 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1572 vma->vm_truncate_count = 0;
1573 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1574 vma->vm_truncate_count = 0;
1577 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1578 unsigned long start_addr, unsigned long end_addr,
1579 struct zap_details *details)
1581 unsigned long restart_addr;
1585 restart_addr = vma->vm_truncate_count;
1586 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1587 start_addr = restart_addr;
1588 if (start_addr >= end_addr) {
1589 /* Top of vma has been split off since last time */
1590 vma->vm_truncate_count = details->truncate_count;
1595 restart_addr = zap_page_range(vma, start_addr,
1596 end_addr - start_addr, details);
1597 need_break = need_resched() ||
1598 need_lockbreak(details->i_mmap_lock);
1600 if (restart_addr >= end_addr) {
1601 /* We have now completed this vma: mark it so */
1602 vma->vm_truncate_count = details->truncate_count;
1606 /* Note restart_addr in vma's truncate_count field */
1607 vma->vm_truncate_count = restart_addr;
1612 spin_unlock(details->i_mmap_lock);
1614 spin_lock(details->i_mmap_lock);
1618 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1619 struct zap_details *details)
1621 struct vm_area_struct *vma;
1622 struct prio_tree_iter iter;
1623 pgoff_t vba, vea, zba, zea;
1626 vma_prio_tree_foreach(vma, &iter, root,
1627 details->first_index, details->last_index) {
1628 /* Skip quickly over those we have already dealt with */
1629 if (vma->vm_truncate_count == details->truncate_count)
1632 vba = vma->vm_pgoff;
1633 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1634 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1635 zba = details->first_index;
1638 zea = details->last_index;
1642 if (unmap_mapping_range_vma(vma,
1643 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1644 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1650 static inline void unmap_mapping_range_list(struct list_head *head,
1651 struct zap_details *details)
1653 struct vm_area_struct *vma;
1656 * In nonlinear VMAs there is no correspondence between virtual address
1657 * offset and file offset. So we must perform an exhaustive search
1658 * across *all* the pages in each nonlinear VMA, not just the pages
1659 * whose virtual address lies outside the file truncation point.
1662 list_for_each_entry(vma, head, shared.vm_set.list) {
1663 /* Skip quickly over those we have already dealt with */
1664 if (vma->vm_truncate_count == details->truncate_count)
1666 details->nonlinear_vma = vma;
1667 if (unmap_mapping_range_vma(vma, vma->vm_start,
1668 vma->vm_end, details) < 0)
1674 * unmap_mapping_range - unmap the portion of all mmaps
1675 * in the specified address_space corresponding to the specified
1676 * page range in the underlying file.
1677 * @mapping: the address space containing mmaps to be unmapped.
1678 * @holebegin: byte in first page to unmap, relative to the start of
1679 * the underlying file. This will be rounded down to a PAGE_SIZE
1680 * boundary. Note that this is different from vmtruncate(), which
1681 * must keep the partial page. In contrast, we must get rid of
1683 * @holelen: size of prospective hole in bytes. This will be rounded
1684 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1686 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1687 * but 0 when invalidating pagecache, don't throw away private data.
1689 void unmap_mapping_range(struct address_space *mapping,
1690 loff_t const holebegin, loff_t const holelen, int even_cows)
1692 struct zap_details details;
1693 pgoff_t hba = holebegin >> PAGE_SHIFT;
1694 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1696 /* Check for overflow. */
1697 if (sizeof(holelen) > sizeof(hlen)) {
1699 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1700 if (holeend & ~(long long)ULONG_MAX)
1701 hlen = ULONG_MAX - hba + 1;
1704 details.check_mapping = even_cows? NULL: mapping;
1705 details.nonlinear_vma = NULL;
1706 details.first_index = hba;
1707 details.last_index = hba + hlen - 1;
1708 if (details.last_index < details.first_index)
1709 details.last_index = ULONG_MAX;
1710 details.i_mmap_lock = &mapping->i_mmap_lock;
1712 spin_lock(&mapping->i_mmap_lock);
1714 /* serialize i_size write against truncate_count write */
1716 /* Protect against page faults, and endless unmapping loops */
1717 mapping->truncate_count++;
1719 * For archs where spin_lock has inclusive semantics like ia64
1720 * this smp_mb() will prevent to read pagetable contents
1721 * before the truncate_count increment is visible to
1725 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1726 if (mapping->truncate_count == 0)
1727 reset_vma_truncate_counts(mapping);
1728 mapping->truncate_count++;
1730 details.truncate_count = mapping->truncate_count;
1732 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1733 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1734 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1735 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1736 spin_unlock(&mapping->i_mmap_lock);
1738 EXPORT_SYMBOL(unmap_mapping_range);
1741 * Handle all mappings that got truncated by a "truncate()"
1744 * NOTE! We have to be ready to update the memory sharing
1745 * between the file and the memory map for a potential last
1746 * incomplete page. Ugly, but necessary.
1748 int vmtruncate(struct inode * inode, loff_t offset)
1750 struct address_space *mapping = inode->i_mapping;
1751 unsigned long limit;
1753 if (inode->i_size < offset)
1756 * truncation of in-use swapfiles is disallowed - it would cause
1757 * subsequent swapout to scribble on the now-freed blocks.
1759 if (IS_SWAPFILE(inode))
1761 i_size_write(inode, offset);
1762 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1763 truncate_inode_pages(mapping, offset);
1767 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1768 if (limit != RLIM_INFINITY && offset > limit)
1770 if (offset > inode->i_sb->s_maxbytes)
1772 i_size_write(inode, offset);
1775 if (inode->i_op && inode->i_op->truncate)
1776 inode->i_op->truncate(inode);
1779 send_sig(SIGXFSZ, current, 0);
1785 EXPORT_SYMBOL(vmtruncate);
1787 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1789 struct address_space *mapping = inode->i_mapping;
1792 * If the underlying filesystem is not going to provide
1793 * a way to truncate a range of blocks (punch a hole) -
1794 * we should return failure right now.
1796 if (!inode->i_op || !inode->i_op->truncate_range)
1799 mutex_lock(&inode->i_mutex);
1800 down_write(&inode->i_alloc_sem);
1801 unmap_mapping_range(mapping, offset, (end - offset), 1);
1802 truncate_inode_pages_range(mapping, offset, end);
1803 inode->i_op->truncate_range(inode, offset, end);
1804 up_write(&inode->i_alloc_sem);
1805 mutex_unlock(&inode->i_mutex);
1809 EXPORT_SYMBOL(vmtruncate_range);
1812 * Primitive swap readahead code. We simply read an aligned block of
1813 * (1 << page_cluster) entries in the swap area. This method is chosen
1814 * because it doesn't cost us any seek time. We also make sure to queue
1815 * the 'original' request together with the readahead ones...
1817 * This has been extended to use the NUMA policies from the mm triggering
1820 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1822 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1825 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1828 struct page *new_page;
1829 unsigned long offset;
1832 * Get the number of handles we should do readahead io to.
1834 num = valid_swaphandles(entry, &offset);
1835 for (i = 0; i < num; offset++, i++) {
1836 /* Ok, do the async read-ahead now */
1837 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1838 offset), vma, addr);
1841 page_cache_release(new_page);
1844 * Find the next applicable VMA for the NUMA policy.
1850 if (addr >= vma->vm_end) {
1852 next_vma = vma ? vma->vm_next : NULL;
1854 if (vma && addr < vma->vm_start)
1857 if (next_vma && addr >= next_vma->vm_start) {
1859 next_vma = vma->vm_next;
1864 lru_add_drain(); /* Push any new pages onto the LRU now */
1868 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1869 * but allow concurrent faults), and pte mapped but not yet locked.
1870 * We return with mmap_sem still held, but pte unmapped and unlocked.
1872 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1873 unsigned long address, pte_t *page_table, pmd_t *pmd,
1874 int write_access, pte_t orig_pte)
1880 int ret = VM_FAULT_MINOR;
1882 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1885 entry = pte_to_swp_entry(orig_pte);
1887 page = lookup_swap_cache(entry);
1889 swapin_readahead(entry, address, vma);
1890 page = read_swap_cache_async(entry, vma, address);
1893 * Back out if somebody else faulted in this pte
1894 * while we released the pte lock.
1896 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1897 if (likely(pte_same(*page_table, orig_pte)))
1902 /* Had to read the page from swap area: Major fault */
1903 ret = VM_FAULT_MAJOR;
1904 inc_page_state(pgmajfault);
1908 if (!vx_rsspages_avail(mm, 1)) {
1912 mark_page_accessed(page);
1914 if (!PageSwapCache(page)) {
1915 /* Page migration has occured */
1917 page_cache_release(page);
1922 * Back out if somebody else already faulted in this pte.
1924 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1925 if (unlikely(!pte_same(*page_table, orig_pte)))
1928 if (unlikely(!PageUptodate(page))) {
1929 ret = VM_FAULT_SIGBUS;
1933 /* The page isn't present yet, go ahead with the fault. */
1935 inc_mm_counter(mm, anon_rss);
1936 pte = mk_pte(page, vma->vm_page_prot);
1937 if (write_access && can_share_swap_page(page)) {
1938 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1942 flush_icache_page(vma, page);
1943 set_pte_at(mm, address, page_table, pte);
1944 page_add_anon_rmap(page, vma, address);
1948 remove_exclusive_swap_page(page);
1952 if (do_wp_page(mm, vma, address,
1953 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1958 /* No need to invalidate - it was non-present before */
1959 update_mmu_cache(vma, address, pte);
1960 lazy_mmu_prot_update(pte);
1962 pte_unmap_unlock(page_table, ptl);
1966 pte_unmap_unlock(page_table, ptl);
1968 page_cache_release(page);
1973 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1974 * but allow concurrent faults), and pte mapped but not yet locked.
1975 * We return with mmap_sem still held, but pte unmapped and unlocked.
1977 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1978 unsigned long address, pte_t *page_table, pmd_t *pmd,
1986 /* Allocate our own private page. */
1987 pte_unmap(page_table);
1989 if (!vx_rsspages_avail(mm, 1))
1991 if (unlikely(anon_vma_prepare(vma)))
1993 page = alloc_zeroed_user_highpage(vma, address);
1997 entry = mk_pte(page, vma->vm_page_prot);
1998 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2000 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2001 if (!pte_none(*page_table))
2003 inc_mm_counter(mm, anon_rss);
2004 lru_cache_add_active(page);
2005 page_add_new_anon_rmap(page, vma, address);
2007 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2008 page = ZERO_PAGE(address);
2009 page_cache_get(page);
2010 entry = mk_pte(page, vma->vm_page_prot);
2012 ptl = pte_lockptr(mm, pmd);
2014 if (!pte_none(*page_table))
2016 inc_mm_counter(mm, file_rss);
2017 page_add_file_rmap(page);
2020 set_pte_at(mm, address, page_table, entry);
2022 /* No need to invalidate - it was non-present before */
2023 update_mmu_cache(vma, address, entry);
2024 lazy_mmu_prot_update(entry);
2026 pte_unmap_unlock(page_table, ptl);
2027 return VM_FAULT_MINOR;
2029 page_cache_release(page);
2032 return VM_FAULT_OOM;
2036 * do_no_page() tries to create a new page mapping. It aggressively
2037 * tries to share with existing pages, but makes a separate copy if
2038 * the "write_access" parameter is true in order to avoid the next
2041 * As this is called only for pages that do not currently exist, we
2042 * do not need to flush old virtual caches or the TLB.
2044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2045 * but allow concurrent faults), and pte mapped but not yet locked.
2046 * We return with mmap_sem still held, but pte unmapped and unlocked.
2048 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2049 unsigned long address, pte_t *page_table, pmd_t *pmd,
2053 struct page *new_page;
2054 struct address_space *mapping = NULL;
2056 unsigned int sequence = 0;
2057 int ret = VM_FAULT_MINOR;
2060 pte_unmap(page_table);
2061 BUG_ON(vma->vm_flags & VM_PFNMAP);
2064 mapping = vma->vm_file->f_mapping;
2065 sequence = mapping->truncate_count;
2066 smp_rmb(); /* serializes i_size against truncate_count */
2069 /* FIXME: is that check useful here? */
2070 if (!vx_rsspages_avail(mm, 1))
2071 return VM_FAULT_OOM;
2072 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2074 * No smp_rmb is needed here as long as there's a full
2075 * spin_lock/unlock sequence inside the ->nopage callback
2076 * (for the pagecache lookup) that acts as an implicit
2077 * smp_mb() and prevents the i_size read to happen
2078 * after the next truncate_count read.
2081 /* no page was available -- either SIGBUS or OOM */
2082 if (new_page == NOPAGE_SIGBUS)
2083 return VM_FAULT_SIGBUS;
2084 if (new_page == NOPAGE_OOM)
2085 return VM_FAULT_OOM;
2088 * Should we do an early C-O-W break?
2090 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2093 if (unlikely(anon_vma_prepare(vma)))
2095 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2098 copy_user_highpage(page, new_page, address);
2099 page_cache_release(new_page);
2104 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2106 * For a file-backed vma, someone could have truncated or otherwise
2107 * invalidated this page. If unmap_mapping_range got called,
2108 * retry getting the page.
2110 if (mapping && unlikely(sequence != mapping->truncate_count)) {
2111 pte_unmap_unlock(page_table, ptl);
2112 page_cache_release(new_page);
2114 sequence = mapping->truncate_count;
2120 * This silly early PAGE_DIRTY setting removes a race
2121 * due to the bad i386 page protection. But it's valid
2122 * for other architectures too.
2124 * Note that if write_access is true, we either now have
2125 * an exclusive copy of the page, or this is a shared mapping,
2126 * so we can make it writable and dirty to avoid having to
2127 * handle that later.
2129 /* Only go through if we didn't race with anybody else... */
2130 if (pte_none(*page_table)) {
2131 flush_icache_page(vma, new_page);
2132 entry = mk_pte(new_page, vma->vm_page_prot);
2134 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2135 set_pte_at(mm, address, page_table, entry);
2137 inc_mm_counter(mm, anon_rss);
2138 lru_cache_add_active(new_page);
2139 page_add_new_anon_rmap(new_page, vma, address);
2141 inc_mm_counter(mm, file_rss);
2142 page_add_file_rmap(new_page);
2145 /* One of our sibling threads was faster, back out. */
2146 page_cache_release(new_page);
2150 /* no need to invalidate: a not-present page shouldn't be cached */
2151 update_mmu_cache(vma, address, entry);
2152 lazy_mmu_prot_update(entry);
2154 pte_unmap_unlock(page_table, ptl);
2157 page_cache_release(new_page);
2158 return VM_FAULT_OOM;
2162 * Fault of a previously existing named mapping. Repopulate the pte
2163 * from the encoded file_pte if possible. This enables swappable
2166 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2167 * but allow concurrent faults), and pte mapped but not yet locked.
2168 * We return with mmap_sem still held, but pte unmapped and unlocked.
2170 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2171 unsigned long address, pte_t *page_table, pmd_t *pmd,
2172 int write_access, pte_t orig_pte)
2177 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2178 return VM_FAULT_MINOR;
2180 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2182 * Page table corrupted: show pte and kill process.
2184 print_bad_pte(vma, orig_pte, address);
2185 return VM_FAULT_OOM;
2187 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2189 pgoff = pte_to_pgoff(orig_pte);
2190 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2191 vma->vm_page_prot, pgoff, 0);
2193 return VM_FAULT_OOM;
2195 return VM_FAULT_SIGBUS;
2196 return VM_FAULT_MAJOR;
2200 * These routines also need to handle stuff like marking pages dirty
2201 * and/or accessed for architectures that don't do it in hardware (most
2202 * RISC architectures). The early dirtying is also good on the i386.
2204 * There is also a hook called "update_mmu_cache()" that architectures
2205 * with external mmu caches can use to update those (ie the Sparc or
2206 * PowerPC hashed page tables that act as extended TLBs).
2208 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2209 * but allow concurrent faults), and pte mapped but not yet locked.
2210 * We return with mmap_sem still held, but pte unmapped and unlocked.
2212 static inline int handle_pte_fault(struct mm_struct *mm,
2213 struct vm_area_struct *vma, unsigned long address,
2214 pte_t *pte, pmd_t *pmd, int write_access)
2220 old_entry = entry = *pte;
2221 if (!pte_present(entry)) {
2222 if (pte_none(entry)) {
2223 if (!vma->vm_ops || !vma->vm_ops->nopage)
2224 return do_anonymous_page(mm, vma, address,
2225 pte, pmd, write_access);
2226 return do_no_page(mm, vma, address,
2227 pte, pmd, write_access);
2229 if (pte_file(entry))
2230 return do_file_page(mm, vma, address,
2231 pte, pmd, write_access, entry);
2232 return do_swap_page(mm, vma, address,
2233 pte, pmd, write_access, entry);
2236 ptl = pte_lockptr(mm, pmd);
2238 if (unlikely(!pte_same(*pte, entry)))
2241 if (!pte_write(entry))
2242 return do_wp_page(mm, vma, address,
2243 pte, pmd, ptl, entry);
2244 entry = pte_mkdirty(entry);
2246 entry = pte_mkyoung(entry);
2247 if (!pte_same(old_entry, entry)) {
2248 ptep_set_access_flags(vma, address, pte, entry, write_access);
2249 update_mmu_cache(vma, address, entry);
2250 lazy_mmu_prot_update(entry);
2253 * This is needed only for protection faults but the arch code
2254 * is not yet telling us if this is a protection fault or not.
2255 * This still avoids useless tlb flushes for .text page faults
2259 flush_tlb_page(vma, address);
2262 pte_unmap_unlock(pte, ptl);
2263 return VM_FAULT_MINOR;
2267 * By the time we get here, we already hold the mm semaphore
2269 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2270 unsigned long address, int write_access)
2277 __set_current_state(TASK_RUNNING);
2279 inc_page_state(pgfault);
2281 if (unlikely(is_vm_hugetlb_page(vma)))
2282 return hugetlb_fault(mm, vma, address, write_access);
2284 pgd = pgd_offset(mm, address);
2285 pud = pud_alloc(mm, pgd, address);
2287 return VM_FAULT_OOM;
2288 pmd = pmd_alloc(mm, pud, address);
2290 return VM_FAULT_OOM;
2291 pte = pte_alloc_map(mm, pmd, address);
2293 return VM_FAULT_OOM;
2295 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2298 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2300 #ifndef __PAGETABLE_PUD_FOLDED
2302 * Allocate page upper directory.
2303 * We've already handled the fast-path in-line.
2305 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2307 pud_t *new = pud_alloc_one(mm, address);
2311 spin_lock(&mm->page_table_lock);
2312 if (pgd_present(*pgd)) /* Another has populated it */
2315 pgd_populate(mm, pgd, new);
2316 spin_unlock(&mm->page_table_lock);
2320 /* Workaround for gcc 2.96 */
2321 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2325 #endif /* __PAGETABLE_PUD_FOLDED */
2327 #ifndef __PAGETABLE_PMD_FOLDED
2329 * Allocate page middle directory.
2330 * We've already handled the fast-path in-line.
2332 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2334 pmd_t *new = pmd_alloc_one(mm, address);
2338 spin_lock(&mm->page_table_lock);
2339 #ifndef __ARCH_HAS_4LEVEL_HACK
2340 if (pud_present(*pud)) /* Another has populated it */
2343 pud_populate(mm, pud, new);
2345 if (pgd_present(*pud)) /* Another has populated it */
2348 pgd_populate(mm, pud, new);
2349 #endif /* __ARCH_HAS_4LEVEL_HACK */
2350 spin_unlock(&mm->page_table_lock);
2354 /* Workaround for gcc 2.96 */
2355 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2359 #endif /* __PAGETABLE_PMD_FOLDED */
2361 int make_pages_present(unsigned long addr, unsigned long end)
2363 int ret, len, write;
2364 struct vm_area_struct * vma;
2366 vma = find_vma(current->mm, addr);
2369 write = (vma->vm_flags & VM_WRITE) != 0;
2372 if (end > vma->vm_end)
2374 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2375 ret = get_user_pages(current, current->mm, addr,
2376 len, write, 0, NULL, NULL);
2379 return ret == len ? 0 : -1;
2383 * Map a vmalloc()-space virtual address to the physical page.
2385 struct page * vmalloc_to_page(void * vmalloc_addr)
2387 unsigned long addr = (unsigned long) vmalloc_addr;
2388 struct page *page = NULL;
2389 pgd_t *pgd = pgd_offset_k(addr);
2394 if (!pgd_none(*pgd)) {
2395 pud = pud_offset(pgd, addr);
2396 if (!pud_none(*pud)) {
2397 pmd = pmd_offset(pud, addr);
2398 if (!pmd_none(*pmd)) {
2399 ptep = pte_offset_map(pmd, addr);
2401 if (pte_present(pte))
2402 page = pte_page(pte);
2410 EXPORT_SYMBOL(vmalloc_to_page);
2413 * Map a vmalloc()-space virtual address to the physical page frame number.
2415 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2417 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2420 EXPORT_SYMBOL(vmalloc_to_pfn);
2422 #if !defined(__HAVE_ARCH_GATE_AREA)
2424 #if defined(AT_SYSINFO_EHDR)
2425 static struct vm_area_struct gate_vma;
2427 static int __init gate_vma_init(void)
2429 gate_vma.vm_mm = NULL;
2430 gate_vma.vm_start = FIXADDR_USER_START;
2431 gate_vma.vm_end = FIXADDR_USER_END;
2432 gate_vma.vm_page_prot = PAGE_READONLY;
2433 gate_vma.vm_flags = 0;
2436 __initcall(gate_vma_init);
2439 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2441 #ifdef AT_SYSINFO_EHDR
2448 int in_gate_area_no_task(unsigned long addr)
2450 #ifdef AT_SYSINFO_EHDR
2451 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2457 #endif /* __HAVE_ARCH_GATE_AREA */