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_vm_hugetlb_page(vma)) {
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_vm_hugetlb_page(next)) {
291 anon_vma_unlink(vma);
292 unlink_file_vma(vma);
294 free_pgd_range(tlb, addr, vma->vm_end,
295 floor, next? next->vm_start: ceiling);
301 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
303 struct page *new = pte_alloc_one(mm, address);
308 spin_lock(&mm->page_table_lock);
309 if (pmd_present(*pmd)) { /* Another has populated it */
310 pte_lock_deinit(new);
314 inc_page_state(nr_page_table_pages);
315 pmd_populate(mm, pmd, new);
317 spin_unlock(&mm->page_table_lock);
321 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
323 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
327 spin_lock(&init_mm.page_table_lock);
328 if (pmd_present(*pmd)) /* Another has populated it */
329 pte_free_kernel(new);
331 pmd_populate_kernel(&init_mm, pmd, new);
332 spin_unlock(&init_mm.page_table_lock);
336 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
339 add_mm_counter(mm, file_rss, file_rss);
341 add_mm_counter(mm, anon_rss, anon_rss);
345 * This function is called to print an error when a bad pte
346 * is found. For example, we might have a PFN-mapped pte in
347 * a region that doesn't allow it.
349 * The calling function must still handle the error.
351 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
353 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
354 "vm_flags = %lx, vaddr = %lx\n",
355 (long long)pte_val(pte),
356 (vma->vm_mm == current->mm ? current->comm : "???"),
357 vma->vm_flags, vaddr);
361 static inline int is_cow_mapping(unsigned int flags)
363 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
367 * This function gets the "struct page" associated with a pte.
369 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
370 * will have each page table entry just pointing to a raw page frame
371 * number, and as far as the VM layer is concerned, those do not have
372 * pages associated with them - even if the PFN might point to memory
373 * that otherwise is perfectly fine and has a "struct page".
375 * The way we recognize those mappings is through the rules set up
376 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
377 * and the vm_pgoff will point to the first PFN mapped: thus every
378 * page that is a raw mapping will always honor the rule
380 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
382 * and if that isn't true, the page has been COW'ed (in which case it
383 * _does_ have a "struct page" associated with it even if it is in a
386 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
388 unsigned long pfn = pte_pfn(pte);
390 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
391 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
392 if (pfn == vma->vm_pgoff + off)
394 if (!is_cow_mapping(vma->vm_flags))
399 * Add some anal sanity checks for now. Eventually,
400 * we should just do "return pfn_to_page(pfn)", but
401 * in the meantime we check that we get a valid pfn,
402 * and that the resulting page looks ok.
404 if (unlikely(!pfn_valid(pfn))) {
405 if (!(vma->vm_flags & VM_RESERVED))
406 print_bad_pte(vma, pte, addr);
411 * NOTE! We still have PageReserved() pages in the page
414 * The PAGE_ZERO() pages and various VDSO mappings can
415 * cause them to exist.
417 return pfn_to_page(pfn);
421 * copy one vm_area from one task to the other. Assumes the page tables
422 * already present in the new task to be cleared in the whole range
423 * covered by this vma.
427 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
428 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
429 unsigned long addr, int *rss)
431 unsigned long vm_flags = vma->vm_flags;
432 pte_t pte = *src_pte;
435 /* pte contains position in swap or file, so copy. */
436 if (unlikely(!pte_present(pte))) {
437 if (!pte_file(pte)) {
438 swap_duplicate(pte_to_swp_entry(pte));
439 /* make sure dst_mm is on swapoff's mmlist. */
440 if (unlikely(list_empty(&dst_mm->mmlist))) {
441 spin_lock(&mmlist_lock);
442 if (list_empty(&dst_mm->mmlist))
443 list_add(&dst_mm->mmlist,
445 spin_unlock(&mmlist_lock);
452 * If it's a COW mapping, write protect it both
453 * in the parent and the child
455 if (is_cow_mapping(vm_flags)) {
456 ptep_set_wrprotect(src_mm, addr, src_pte);
461 * If it's a shared mapping, mark it clean in
464 if (vm_flags & VM_SHARED)
465 pte = pte_mkclean(pte);
466 pte = pte_mkold(pte);
468 page = vm_normal_page(vma, addr, pte);
472 rss[!!PageAnon(page)]++;
476 set_pte_at(dst_mm, addr, dst_pte, pte);
479 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
480 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
481 unsigned long addr, unsigned long end)
483 pte_t *src_pte, *dst_pte;
484 spinlock_t *src_ptl, *dst_ptl;
490 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
493 src_pte = pte_offset_map_nested(src_pmd, addr);
494 src_ptl = pte_lockptr(src_mm, src_pmd);
499 * We are holding two locks at this point - either of them
500 * could generate latencies in another task on another CPU.
502 if (progress >= 32) {
504 if (need_resched() ||
505 need_lockbreak(src_ptl) ||
506 need_lockbreak(dst_ptl))
509 if (pte_none(*src_pte)) {
513 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
515 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
517 spin_unlock(src_ptl);
518 pte_unmap_nested(src_pte - 1);
519 add_mm_rss(dst_mm, rss[0], rss[1]);
520 pte_unmap_unlock(dst_pte - 1, dst_ptl);
527 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
528 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
529 unsigned long addr, unsigned long end)
531 pmd_t *src_pmd, *dst_pmd;
534 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
537 src_pmd = pmd_offset(src_pud, addr);
539 next = pmd_addr_end(addr, end);
540 if (pmd_none_or_clear_bad(src_pmd))
542 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
545 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
549 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
550 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
551 unsigned long addr, unsigned long end)
553 pud_t *src_pud, *dst_pud;
556 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
559 src_pud = pud_offset(src_pgd, addr);
561 next = pud_addr_end(addr, end);
562 if (pud_none_or_clear_bad(src_pud))
564 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
567 } while (dst_pud++, src_pud++, addr = next, addr != end);
571 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
572 struct vm_area_struct *vma)
574 pgd_t *src_pgd, *dst_pgd;
576 unsigned long addr = vma->vm_start;
577 unsigned long end = vma->vm_end;
580 * Don't copy ptes where a page fault will fill them correctly.
581 * Fork becomes much lighter when there are big shared or private
582 * readonly mappings. The tradeoff is that copy_page_range is more
583 * efficient than faulting.
585 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
590 if (is_vm_hugetlb_page(vma))
591 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
593 dst_pgd = pgd_offset(dst_mm, addr);
594 src_pgd = pgd_offset(src_mm, addr);
596 next = pgd_addr_end(addr, end);
597 if (pgd_none_or_clear_bad(src_pgd))
599 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
602 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
606 static unsigned long zap_pte_range(struct mmu_gather *tlb,
607 struct vm_area_struct *vma, pmd_t *pmd,
608 unsigned long addr, unsigned long end,
609 long *zap_work, struct zap_details *details)
611 struct mm_struct *mm = tlb->mm;
617 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
620 if (pte_none(ptent)) {
625 (*zap_work) -= PAGE_SIZE;
627 if (pte_present(ptent)) {
630 page = vm_normal_page(vma, addr, ptent);
631 if (unlikely(details) && page) {
633 * unmap_shared_mapping_pages() wants to
634 * invalidate cache without truncating:
635 * unmap shared but keep private pages.
637 if (details->check_mapping &&
638 details->check_mapping != page->mapping)
641 * Each page->index must be checked when
642 * invalidating or truncating nonlinear.
644 if (details->nonlinear_vma &&
645 (page->index < details->first_index ||
646 page->index > details->last_index))
649 ptent = ptep_get_and_clear_full(mm, addr, pte,
651 tlb_remove_tlb_entry(tlb, pte, addr);
654 if (unlikely(details) && details->nonlinear_vma
655 && linear_page_index(details->nonlinear_vma,
656 addr) != page->index)
657 set_pte_at(mm, addr, pte,
658 pgoff_to_pte(page->index));
662 if (pte_dirty(ptent))
663 set_page_dirty(page);
664 if (pte_young(ptent))
665 mark_page_accessed(page);
668 page_remove_rmap(page);
669 tlb_remove_page(tlb, page);
673 * If details->check_mapping, we leave swap entries;
674 * if details->nonlinear_vma, we leave file entries.
676 if (unlikely(details))
678 if (!pte_file(ptent))
679 free_swap_and_cache(pte_to_swp_entry(ptent));
680 pte_clear_full(mm, addr, pte, tlb->fullmm);
681 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
683 add_mm_rss(mm, file_rss, anon_rss);
684 pte_unmap_unlock(pte - 1, ptl);
689 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
690 struct vm_area_struct *vma, pud_t *pud,
691 unsigned long addr, unsigned long end,
692 long *zap_work, struct zap_details *details)
697 pmd = pmd_offset(pud, addr);
699 next = pmd_addr_end(addr, end);
700 if (pmd_none_or_clear_bad(pmd)) {
704 next = zap_pte_range(tlb, vma, pmd, addr, next,
706 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
711 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
712 struct vm_area_struct *vma, pgd_t *pgd,
713 unsigned long addr, unsigned long end,
714 long *zap_work, struct zap_details *details)
719 pud = pud_offset(pgd, addr);
721 next = pud_addr_end(addr, end);
722 if (pud_none_or_clear_bad(pud)) {
726 next = zap_pmd_range(tlb, vma, pud, addr, next,
728 } while (pud++, addr = next, (addr != end && *zap_work > 0));
733 static unsigned long unmap_page_range(struct mmu_gather *tlb,
734 struct vm_area_struct *vma,
735 unsigned long addr, unsigned long end,
736 long *zap_work, struct zap_details *details)
741 if (details && !details->check_mapping && !details->nonlinear_vma)
745 tlb_start_vma(tlb, vma);
746 pgd = pgd_offset(vma->vm_mm, addr);
748 next = pgd_addr_end(addr, end);
749 if (pgd_none_or_clear_bad(pgd)) {
753 next = zap_pud_range(tlb, vma, pgd, addr, next,
755 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
756 tlb_end_vma(tlb, vma);
761 #ifdef CONFIG_PREEMPT
762 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
764 /* No preempt: go for improved straight-line efficiency */
765 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
769 * unmap_vmas - unmap a range of memory covered by a list of vma's
770 * @tlbp: address of the caller's struct mmu_gather
771 * @vma: the starting vma
772 * @start_addr: virtual address at which to start unmapping
773 * @end_addr: virtual address at which to end unmapping
774 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
775 * @details: details of nonlinear truncation or shared cache invalidation
777 * Returns the end address of the unmapping (restart addr if interrupted).
779 * Unmap all pages in the vma list.
781 * We aim to not hold locks for too long (for scheduling latency reasons).
782 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
783 * return the ending mmu_gather to the caller.
785 * Only addresses between `start' and `end' will be unmapped.
787 * The VMA list must be sorted in ascending virtual address order.
789 * unmap_vmas() assumes that the caller will flush the whole unmapped address
790 * range after unmap_vmas() returns. So the only responsibility here is to
791 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
792 * drops the lock and schedules.
794 unsigned long unmap_vmas(struct mmu_gather **tlbp,
795 struct vm_area_struct *vma, unsigned long start_addr,
796 unsigned long end_addr, unsigned long *nr_accounted,
797 struct zap_details *details)
799 long zap_work = ZAP_BLOCK_SIZE;
800 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
801 int tlb_start_valid = 0;
802 unsigned long start = start_addr;
803 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
804 int fullmm = (*tlbp)->fullmm;
806 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
809 start = max(vma->vm_start, start_addr);
810 if (start >= vma->vm_end)
812 end = min(vma->vm_end, end_addr);
813 if (end <= vma->vm_start)
816 if (vma->vm_flags & VM_ACCOUNT)
817 *nr_accounted += (end - start) >> PAGE_SHIFT;
819 while (start != end) {
820 if (!tlb_start_valid) {
825 if (unlikely(is_vm_hugetlb_page(vma))) {
826 unmap_hugepage_range(vma, start, end);
827 zap_work -= (end - start) /
828 (HPAGE_SIZE / PAGE_SIZE);
831 start = unmap_page_range(*tlbp, vma,
832 start, end, &zap_work, details);
835 BUG_ON(start != end);
839 tlb_finish_mmu(*tlbp, tlb_start, start);
841 if (need_resched() ||
842 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
850 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
852 zap_work = ZAP_BLOCK_SIZE;
856 return start; /* which is now the end (or restart) address */
860 * zap_page_range - remove user pages in a given range
861 * @vma: vm_area_struct holding the applicable pages
862 * @address: starting address of pages to zap
863 * @size: number of bytes to zap
864 * @details: details of nonlinear truncation or shared cache invalidation
866 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
867 unsigned long size, struct zap_details *details)
869 struct mm_struct *mm = vma->vm_mm;
870 struct mmu_gather *tlb;
871 unsigned long end = address + size;
872 unsigned long nr_accounted = 0;
875 tlb = tlb_gather_mmu(mm, 0);
876 update_hiwater_rss(mm);
877 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
879 tlb_finish_mmu(tlb, address, end);
884 * Do a quick page-table lookup for a single page.
886 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
895 struct mm_struct *mm = vma->vm_mm;
897 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
899 BUG_ON(flags & FOLL_GET);
904 pgd = pgd_offset(mm, address);
905 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
908 pud = pud_offset(pgd, address);
909 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
912 pmd = pmd_offset(pud, address);
913 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
916 if (pmd_huge(*pmd)) {
917 BUG_ON(flags & FOLL_GET);
918 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
922 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
927 if (!pte_present(pte))
929 if ((flags & FOLL_WRITE) && !pte_write(pte))
931 page = vm_normal_page(vma, address, pte);
935 if (flags & FOLL_GET)
937 if (flags & FOLL_TOUCH) {
938 if ((flags & FOLL_WRITE) &&
939 !pte_dirty(pte) && !PageDirty(page))
940 set_page_dirty(page);
941 mark_page_accessed(page);
944 pte_unmap_unlock(ptep, ptl);
950 * When core dumping an enormous anonymous area that nobody
951 * has touched so far, we don't want to allocate page tables.
953 if (flags & FOLL_ANON) {
954 page = ZERO_PAGE(address);
955 if (flags & FOLL_GET)
957 BUG_ON(flags & FOLL_WRITE);
962 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
963 unsigned long start, int len, int write, int force,
964 struct page **pages, struct vm_area_struct **vmas)
967 unsigned int vm_flags;
970 * Require read or write permissions.
971 * If 'force' is set, we only require the "MAY" flags.
973 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
974 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
978 struct vm_area_struct *vma;
979 unsigned int foll_flags;
981 vma = find_extend_vma(mm, start);
982 if (!vma && in_gate_area(tsk, start)) {
983 unsigned long pg = start & PAGE_MASK;
984 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
989 if (write) /* user gate pages are read-only */
990 return i ? : -EFAULT;
992 pgd = pgd_offset_k(pg);
994 pgd = pgd_offset_gate(mm, pg);
995 BUG_ON(pgd_none(*pgd));
996 pud = pud_offset(pgd, pg);
997 BUG_ON(pud_none(*pud));
998 pmd = pmd_offset(pud, pg);
1000 return i ? : -EFAULT;
1001 pte = pte_offset_map(pmd, pg);
1002 if (pte_none(*pte)) {
1004 return i ? : -EFAULT;
1007 struct page *page = vm_normal_page(gate_vma, start, *pte);
1022 if (vma && (vma->vm_flags & VM_FOREIGN)) {
1023 struct page **map = vma->vm_private_data;
1024 int offset = (start - vma->vm_start) >> PAGE_SHIFT;
1025 if (map[offset] != NULL) {
1027 struct page *page = map[offset];
1041 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1042 || !(vm_flags & vma->vm_flags))
1043 return i ? : -EFAULT;
1045 if (is_vm_hugetlb_page(vma)) {
1046 i = follow_hugetlb_page(mm, vma, pages, vmas,
1051 foll_flags = FOLL_TOUCH;
1053 foll_flags |= FOLL_GET;
1054 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1055 (!vma->vm_ops || !vma->vm_ops->nopage))
1056 foll_flags |= FOLL_ANON;
1062 foll_flags |= FOLL_WRITE;
1065 while (!(page = follow_page(vma, start, foll_flags))) {
1067 ret = __handle_mm_fault(mm, vma, start,
1068 foll_flags & FOLL_WRITE);
1070 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1071 * broken COW when necessary, even if maybe_mkwrite
1072 * decided not to set pte_write. We can thus safely do
1073 * subsequent page lookups as if they were reads.
1075 if (ret & VM_FAULT_WRITE)
1076 foll_flags &= ~FOLL_WRITE;
1078 switch (ret & ~VM_FAULT_WRITE) {
1079 case VM_FAULT_MINOR:
1082 case VM_FAULT_MAJOR:
1085 case VM_FAULT_SIGBUS:
1086 return i ? i : -EFAULT;
1088 return i ? i : -ENOMEM;
1096 flush_anon_page(page, start);
1097 flush_dcache_page(page);
1104 } while (len && start < vma->vm_end);
1108 EXPORT_SYMBOL(get_user_pages);
1110 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1111 unsigned long addr, unsigned long end, pgprot_t prot)
1116 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1120 struct page *page = ZERO_PAGE(addr);
1121 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1122 page_cache_get(page);
1123 page_add_file_rmap(page);
1124 inc_mm_counter(mm, file_rss);
1125 BUG_ON(!pte_none(*pte));
1126 set_pte_at(mm, addr, pte, zero_pte);
1127 } while (pte++, addr += PAGE_SIZE, addr != end);
1128 pte_unmap_unlock(pte - 1, ptl);
1132 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1133 unsigned long addr, unsigned long end, pgprot_t prot)
1138 pmd = pmd_alloc(mm, pud, addr);
1142 next = pmd_addr_end(addr, end);
1143 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1145 } while (pmd++, addr = next, addr != end);
1149 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1150 unsigned long addr, unsigned long end, pgprot_t prot)
1155 pud = pud_alloc(mm, pgd, addr);
1159 next = pud_addr_end(addr, end);
1160 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1162 } while (pud++, addr = next, addr != end);
1166 int zeromap_page_range(struct vm_area_struct *vma,
1167 unsigned long addr, unsigned long size, pgprot_t prot)
1171 unsigned long end = addr + size;
1172 struct mm_struct *mm = vma->vm_mm;
1175 BUG_ON(addr >= end);
1176 pgd = pgd_offset(mm, addr);
1177 flush_cache_range(vma, addr, end);
1179 next = pgd_addr_end(addr, end);
1180 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1183 } while (pgd++, addr = next, addr != end);
1187 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1189 pgd_t * pgd = pgd_offset(mm, addr);
1190 pud_t * pud = pud_alloc(mm, pgd, addr);
1192 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1194 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1200 * This is the old fallback for page remapping.
1202 * For historical reasons, it only allows reserved pages. Only
1203 * old drivers should use this, and they needed to mark their
1204 * pages reserved for the old functions anyway.
1206 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1216 flush_dcache_page(page);
1217 pte = get_locked_pte(mm, addr, &ptl);
1221 if (!pte_none(*pte))
1224 /* Ok, finally just insert the thing.. */
1226 inc_mm_counter(mm, file_rss);
1227 page_add_file_rmap(page);
1228 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1232 pte_unmap_unlock(pte, ptl);
1238 * This allows drivers to insert individual pages they've allocated
1241 * The page has to be a nice clean _individual_ kernel allocation.
1242 * If you allocate a compound page, you need to have marked it as
1243 * such (__GFP_COMP), or manually just split the page up yourself
1244 * (see split_page()).
1246 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1247 * took an arbitrary page protection parameter. This doesn't allow
1248 * that. Your vma protection will have to be set up correctly, which
1249 * means that if you want a shared writable mapping, you'd better
1250 * ask for a shared writable mapping!
1252 * The page does not need to be reserved.
1254 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1256 if (addr < vma->vm_start || addr >= vma->vm_end)
1258 if (!page_count(page))
1260 vma->vm_flags |= VM_INSERTPAGE;
1261 return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1263 EXPORT_SYMBOL(vm_insert_page);
1266 * maps a range of physical memory into the requested pages. the old
1267 * mappings are removed. any references to nonexistent pages results
1268 * in null mappings (currently treated as "copy-on-access")
1270 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1271 unsigned long addr, unsigned long end,
1272 unsigned long pfn, pgprot_t prot)
1277 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1281 BUG_ON(!pte_none(*pte));
1282 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1284 } while (pte++, addr += PAGE_SIZE, addr != end);
1285 pte_unmap_unlock(pte - 1, ptl);
1289 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1290 unsigned long addr, unsigned long end,
1291 unsigned long pfn, pgprot_t prot)
1296 pfn -= addr >> PAGE_SHIFT;
1297 pmd = pmd_alloc(mm, pud, addr);
1301 next = pmd_addr_end(addr, end);
1302 if (remap_pte_range(mm, pmd, addr, next,
1303 pfn + (addr >> PAGE_SHIFT), prot))
1305 } while (pmd++, addr = next, addr != end);
1309 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1310 unsigned long addr, unsigned long end,
1311 unsigned long pfn, pgprot_t prot)
1316 pfn -= addr >> PAGE_SHIFT;
1317 pud = pud_alloc(mm, pgd, addr);
1321 next = pud_addr_end(addr, end);
1322 if (remap_pmd_range(mm, pud, addr, next,
1323 pfn + (addr >> PAGE_SHIFT), prot))
1325 } while (pud++, addr = next, addr != end);
1329 /* Note: this is only safe if the mm semaphore is held when called. */
1330 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1331 unsigned long pfn, unsigned long size, pgprot_t prot)
1335 unsigned long end = addr + PAGE_ALIGN(size);
1336 struct mm_struct *mm = vma->vm_mm;
1340 * Physically remapped pages are special. Tell the
1341 * rest of the world about it:
1342 * VM_IO tells people not to look at these pages
1343 * (accesses can have side effects).
1344 * VM_RESERVED is specified all over the place, because
1345 * in 2.4 it kept swapout's vma scan off this vma; but
1346 * in 2.6 the LRU scan won't even find its pages, so this
1347 * flag means no more than count its pages in reserved_vm,
1348 * and omit it from core dump, even when VM_IO turned off.
1349 * VM_PFNMAP tells the core MM that the base pages are just
1350 * raw PFN mappings, and do not have a "struct page" associated
1353 * There's a horrible special case to handle copy-on-write
1354 * behaviour that some programs depend on. We mark the "original"
1355 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1357 if (is_cow_mapping(vma->vm_flags)) {
1358 if (addr != vma->vm_start || end != vma->vm_end)
1360 vma->vm_pgoff = pfn;
1363 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1365 BUG_ON(addr >= end);
1366 pfn -= addr >> PAGE_SHIFT;
1367 pgd = pgd_offset(mm, addr);
1368 flush_cache_range(vma, addr, end);
1370 next = pgd_addr_end(addr, end);
1371 err = remap_pud_range(mm, pgd, addr, next,
1372 pfn + (addr >> PAGE_SHIFT), prot);
1375 } while (pgd++, addr = next, addr != end);
1378 EXPORT_SYMBOL(remap_pfn_range);
1381 static inline int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1382 unsigned long addr, unsigned long end,
1383 pte_fn_t fn, void *data)
1387 struct page *pmd_page;
1390 pte = (mm == &init_mm) ?
1391 pte_alloc_kernel(pmd, addr) :
1392 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1396 BUG_ON(pmd_huge(*pmd));
1398 pmd_page = pmd_page(*pmd);
1401 err = fn(pte, pmd_page, addr, data);
1404 } while (pte++, addr += PAGE_SIZE, addr != end);
1407 pte_unmap_unlock(pte-1, ptl);
1411 static inline int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1412 unsigned long addr, unsigned long end,
1413 pte_fn_t fn, void *data)
1419 pmd = pmd_alloc(mm, pud, addr);
1423 next = pmd_addr_end(addr, end);
1424 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1427 } while (pmd++, addr = next, addr != end);
1431 static inline int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1432 unsigned long addr, unsigned long end,
1433 pte_fn_t fn, void *data)
1439 pud = pud_alloc(mm, pgd, addr);
1443 next = pud_addr_end(addr, end);
1444 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1447 } while (pud++, addr = next, addr != end);
1452 * Scan a region of virtual memory, filling in page tables as necessary
1453 * and calling a provided function on each leaf page table.
1455 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1456 unsigned long size, pte_fn_t fn, void *data)
1460 unsigned long end = addr + size;
1463 BUG_ON(addr >= end);
1464 pgd = pgd_offset(mm, addr);
1466 next = pgd_addr_end(addr, end);
1467 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1470 } while (pgd++, addr = next, addr != end);
1473 EXPORT_SYMBOL_GPL(apply_to_page_range);
1477 * handle_pte_fault chooses page fault handler according to an entry
1478 * which was read non-atomically. Before making any commitment, on
1479 * those architectures or configurations (e.g. i386 with PAE) which
1480 * might give a mix of unmatched parts, do_swap_page and do_file_page
1481 * must check under lock before unmapping the pte and proceeding
1482 * (but do_wp_page is only called after already making such a check;
1483 * and do_anonymous_page and do_no_page can safely check later on).
1485 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1486 pte_t *page_table, pte_t orig_pte)
1489 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1490 if (sizeof(pte_t) > sizeof(unsigned long)) {
1491 spinlock_t *ptl = pte_lockptr(mm, pmd);
1493 same = pte_same(*page_table, orig_pte);
1497 pte_unmap(page_table);
1502 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1503 * servicing faults for write access. In the normal case, do always want
1504 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1505 * that do not have writing enabled, when used by access_process_vm.
1507 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1509 if (likely(vma->vm_flags & VM_WRITE))
1510 pte = pte_mkwrite(pte);
1514 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1517 * If the source page was a PFN mapping, we don't have
1518 * a "struct page" for it. We do a best-effort copy by
1519 * just copying from the original user address. If that
1520 * fails, we just zero-fill it. Live with it.
1522 if (unlikely(!src)) {
1523 void *kaddr = kmap_atomic(dst, KM_USER0);
1524 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1527 * This really shouldn't fail, because the page is there
1528 * in the page tables. But it might just be unreadable,
1529 * in which case we just give up and fill the result with
1532 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1533 memset(kaddr, 0, PAGE_SIZE);
1534 kunmap_atomic(kaddr, KM_USER0);
1538 copy_user_highpage(dst, src, va);
1542 * This routine handles present pages, when users try to write
1543 * to a shared page. It is done by copying the page to a new address
1544 * and decrementing the shared-page counter for the old page.
1546 * Note that this routine assumes that the protection checks have been
1547 * done by the caller (the low-level page fault routine in most cases).
1548 * Thus we can safely just mark it writable once we've done any necessary
1551 * We also mark the page dirty at this point even though the page will
1552 * change only once the write actually happens. This avoids a few races,
1553 * and potentially makes it more efficient.
1555 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1556 * but allow concurrent faults), with pte both mapped and locked.
1557 * We return with mmap_sem still held, but pte unmapped and unlocked.
1559 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1560 unsigned long address, pte_t *page_table, pmd_t *pmd,
1561 spinlock_t *ptl, pte_t orig_pte)
1563 struct page *old_page, *new_page;
1565 int ret = VM_FAULT_MINOR;
1567 old_page = vm_normal_page(vma, address, orig_pte);
1571 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1572 int reuse = can_share_swap_page(old_page);
1573 unlock_page(old_page);
1575 flush_cache_page(vma, address, pte_pfn(orig_pte));
1576 entry = pte_mkyoung(orig_pte);
1577 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1578 ptep_set_access_flags(vma, address, page_table, entry, 1);
1579 update_mmu_cache(vma, address, entry);
1580 lazy_mmu_prot_update(entry);
1581 ret |= VM_FAULT_WRITE;
1587 * Ok, we need to copy. Oh, well..
1589 page_cache_get(old_page);
1591 pte_unmap_unlock(page_table, ptl);
1593 if (unlikely(anon_vma_prepare(vma)))
1595 if (old_page == ZERO_PAGE(address)) {
1596 new_page = alloc_zeroed_user_highpage(vma, address);
1600 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1603 cow_user_page(new_page, old_page, address);
1607 * Re-check the pte - we dropped the lock
1609 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1610 if (likely(pte_same(*page_table, orig_pte))) {
1612 page_remove_rmap(old_page);
1613 if (!PageAnon(old_page)) {
1614 dec_mm_counter(mm, file_rss);
1615 inc_mm_counter(mm, anon_rss);
1618 inc_mm_counter(mm, anon_rss);
1619 flush_cache_page(vma, address, pte_pfn(orig_pte));
1620 entry = mk_pte(new_page, vma->vm_page_prot);
1621 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1622 ptep_establish(vma, address, page_table, entry);
1623 update_mmu_cache(vma, address, entry);
1624 lazy_mmu_prot_update(entry);
1625 lru_cache_add_active(new_page);
1626 page_add_new_anon_rmap(new_page, vma, address);
1628 /* Free the old page.. */
1629 new_page = old_page;
1630 ret |= VM_FAULT_WRITE;
1633 page_cache_release(new_page);
1635 page_cache_release(old_page);
1637 pte_unmap_unlock(page_table, ptl);
1641 page_cache_release(old_page);
1642 return VM_FAULT_OOM;
1646 * Helper functions for unmap_mapping_range().
1648 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1650 * We have to restart searching the prio_tree whenever we drop the lock,
1651 * since the iterator is only valid while the lock is held, and anyway
1652 * a later vma might be split and reinserted earlier while lock dropped.
1654 * The list of nonlinear vmas could be handled more efficiently, using
1655 * a placeholder, but handle it in the same way until a need is shown.
1656 * It is important to search the prio_tree before nonlinear list: a vma
1657 * may become nonlinear and be shifted from prio_tree to nonlinear list
1658 * while the lock is dropped; but never shifted from list to prio_tree.
1660 * In order to make forward progress despite restarting the search,
1661 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1662 * quickly skip it next time around. Since the prio_tree search only
1663 * shows us those vmas affected by unmapping the range in question, we
1664 * can't efficiently keep all vmas in step with mapping->truncate_count:
1665 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1666 * mapping->truncate_count and vma->vm_truncate_count are protected by
1669 * In order to make forward progress despite repeatedly restarting some
1670 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1671 * and restart from that address when we reach that vma again. It might
1672 * have been split or merged, shrunk or extended, but never shifted: so
1673 * restart_addr remains valid so long as it remains in the vma's range.
1674 * unmap_mapping_range forces truncate_count to leap over page-aligned
1675 * values so we can save vma's restart_addr in its truncate_count field.
1677 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1679 static void reset_vma_truncate_counts(struct address_space *mapping)
1681 struct vm_area_struct *vma;
1682 struct prio_tree_iter iter;
1684 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1685 vma->vm_truncate_count = 0;
1686 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1687 vma->vm_truncate_count = 0;
1690 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1691 unsigned long start_addr, unsigned long end_addr,
1692 struct zap_details *details)
1694 unsigned long restart_addr;
1698 restart_addr = vma->vm_truncate_count;
1699 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1700 start_addr = restart_addr;
1701 if (start_addr >= end_addr) {
1702 /* Top of vma has been split off since last time */
1703 vma->vm_truncate_count = details->truncate_count;
1708 restart_addr = zap_page_range(vma, start_addr,
1709 end_addr - start_addr, details);
1710 need_break = need_resched() ||
1711 need_lockbreak(details->i_mmap_lock);
1713 if (restart_addr >= end_addr) {
1714 /* We have now completed this vma: mark it so */
1715 vma->vm_truncate_count = details->truncate_count;
1719 /* Note restart_addr in vma's truncate_count field */
1720 vma->vm_truncate_count = restart_addr;
1725 spin_unlock(details->i_mmap_lock);
1727 spin_lock(details->i_mmap_lock);
1731 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1732 struct zap_details *details)
1734 struct vm_area_struct *vma;
1735 struct prio_tree_iter iter;
1736 pgoff_t vba, vea, zba, zea;
1739 vma_prio_tree_foreach(vma, &iter, root,
1740 details->first_index, details->last_index) {
1741 /* Skip quickly over those we have already dealt with */
1742 if (vma->vm_truncate_count == details->truncate_count)
1745 vba = vma->vm_pgoff;
1746 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1747 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1748 zba = details->first_index;
1751 zea = details->last_index;
1755 if (unmap_mapping_range_vma(vma,
1756 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1757 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1763 static inline void unmap_mapping_range_list(struct list_head *head,
1764 struct zap_details *details)
1766 struct vm_area_struct *vma;
1769 * In nonlinear VMAs there is no correspondence between virtual address
1770 * offset and file offset. So we must perform an exhaustive search
1771 * across *all* the pages in each nonlinear VMA, not just the pages
1772 * whose virtual address lies outside the file truncation point.
1775 list_for_each_entry(vma, head, shared.vm_set.list) {
1776 /* Skip quickly over those we have already dealt with */
1777 if (vma->vm_truncate_count == details->truncate_count)
1779 details->nonlinear_vma = vma;
1780 if (unmap_mapping_range_vma(vma, vma->vm_start,
1781 vma->vm_end, details) < 0)
1787 * unmap_mapping_range - unmap the portion of all mmaps
1788 * in the specified address_space corresponding to the specified
1789 * page range in the underlying file.
1790 * @mapping: the address space containing mmaps to be unmapped.
1791 * @holebegin: byte in first page to unmap, relative to the start of
1792 * the underlying file. This will be rounded down to a PAGE_SIZE
1793 * boundary. Note that this is different from vmtruncate(), which
1794 * must keep the partial page. In contrast, we must get rid of
1796 * @holelen: size of prospective hole in bytes. This will be rounded
1797 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1799 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1800 * but 0 when invalidating pagecache, don't throw away private data.
1802 void unmap_mapping_range(struct address_space *mapping,
1803 loff_t const holebegin, loff_t const holelen, int even_cows)
1805 struct zap_details details;
1806 pgoff_t hba = holebegin >> PAGE_SHIFT;
1807 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1809 /* Check for overflow. */
1810 if (sizeof(holelen) > sizeof(hlen)) {
1812 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1813 if (holeend & ~(long long)ULONG_MAX)
1814 hlen = ULONG_MAX - hba + 1;
1817 details.check_mapping = even_cows? NULL: mapping;
1818 details.nonlinear_vma = NULL;
1819 details.first_index = hba;
1820 details.last_index = hba + hlen - 1;
1821 if (details.last_index < details.first_index)
1822 details.last_index = ULONG_MAX;
1823 details.i_mmap_lock = &mapping->i_mmap_lock;
1825 spin_lock(&mapping->i_mmap_lock);
1827 /* serialize i_size write against truncate_count write */
1829 /* Protect against page faults, and endless unmapping loops */
1830 mapping->truncate_count++;
1832 * For archs where spin_lock has inclusive semantics like ia64
1833 * this smp_mb() will prevent to read pagetable contents
1834 * before the truncate_count increment is visible to
1838 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1839 if (mapping->truncate_count == 0)
1840 reset_vma_truncate_counts(mapping);
1841 mapping->truncate_count++;
1843 details.truncate_count = mapping->truncate_count;
1845 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1846 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1847 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1848 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1849 spin_unlock(&mapping->i_mmap_lock);
1851 EXPORT_SYMBOL(unmap_mapping_range);
1854 * Handle all mappings that got truncated by a "truncate()"
1857 * NOTE! We have to be ready to update the memory sharing
1858 * between the file and the memory map for a potential last
1859 * incomplete page. Ugly, but necessary.
1861 int vmtruncate(struct inode * inode, loff_t offset)
1863 struct address_space *mapping = inode->i_mapping;
1864 unsigned long limit;
1866 if (inode->i_size < offset)
1869 * truncation of in-use swapfiles is disallowed - it would cause
1870 * subsequent swapout to scribble on the now-freed blocks.
1872 if (IS_SWAPFILE(inode))
1874 i_size_write(inode, offset);
1875 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1876 truncate_inode_pages(mapping, offset);
1880 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1881 if (limit != RLIM_INFINITY && offset > limit)
1883 if (offset > inode->i_sb->s_maxbytes)
1885 i_size_write(inode, offset);
1888 if (inode->i_op && inode->i_op->truncate)
1889 inode->i_op->truncate(inode);
1892 send_sig(SIGXFSZ, current, 0);
1898 EXPORT_SYMBOL(vmtruncate);
1900 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1902 struct address_space *mapping = inode->i_mapping;
1905 * If the underlying filesystem is not going to provide
1906 * a way to truncate a range of blocks (punch a hole) -
1907 * we should return failure right now.
1909 if (!inode->i_op || !inode->i_op->truncate_range)
1912 mutex_lock(&inode->i_mutex);
1913 down_write(&inode->i_alloc_sem);
1914 unmap_mapping_range(mapping, offset, (end - offset), 1);
1915 truncate_inode_pages_range(mapping, offset, end);
1916 inode->i_op->truncate_range(inode, offset, end);
1917 up_write(&inode->i_alloc_sem);
1918 mutex_unlock(&inode->i_mutex);
1922 EXPORT_SYMBOL(vmtruncate_range);
1925 * Primitive swap readahead code. We simply read an aligned block of
1926 * (1 << page_cluster) entries in the swap area. This method is chosen
1927 * because it doesn't cost us any seek time. We also make sure to queue
1928 * the 'original' request together with the readahead ones...
1930 * This has been extended to use the NUMA policies from the mm triggering
1933 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1935 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1938 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1941 struct page *new_page;
1942 unsigned long offset;
1945 * Get the number of handles we should do readahead io to.
1947 num = valid_swaphandles(entry, &offset);
1948 for (i = 0; i < num; offset++, i++) {
1949 /* Ok, do the async read-ahead now */
1950 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1951 offset), vma, addr);
1954 page_cache_release(new_page);
1957 * Find the next applicable VMA for the NUMA policy.
1963 if (addr >= vma->vm_end) {
1965 next_vma = vma ? vma->vm_next : NULL;
1967 if (vma && addr < vma->vm_start)
1970 if (next_vma && addr >= next_vma->vm_start) {
1972 next_vma = vma->vm_next;
1977 lru_add_drain(); /* Push any new pages onto the LRU now */
1981 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1982 * but allow concurrent faults), and pte mapped but not yet locked.
1983 * We return with mmap_sem still held, but pte unmapped and unlocked.
1985 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1986 unsigned long address, pte_t *page_table, pmd_t *pmd,
1987 int write_access, pte_t orig_pte)
1993 int ret = VM_FAULT_MINOR;
1995 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1998 entry = pte_to_swp_entry(orig_pte);
2000 page = lookup_swap_cache(entry);
2002 swapin_readahead(entry, address, vma);
2003 page = read_swap_cache_async(entry, vma, address);
2006 * Back out if somebody else faulted in this pte
2007 * while we released the pte lock.
2009 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2010 if (likely(pte_same(*page_table, orig_pte)))
2015 /* Had to read the page from swap area: Major fault */
2016 ret = VM_FAULT_MAJOR;
2017 inc_page_state(pgmajfault);
2021 if (!vx_rsspages_avail(mm, 1)) {
2025 mark_page_accessed(page);
2027 if (!PageSwapCache(page)) {
2028 /* Page migration has occured */
2030 page_cache_release(page);
2035 * Back out if somebody else already faulted in this pte.
2037 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2038 if (unlikely(!pte_same(*page_table, orig_pte)))
2041 if (unlikely(!PageUptodate(page))) {
2042 ret = VM_FAULT_SIGBUS;
2046 /* The page isn't present yet, go ahead with the fault. */
2048 inc_mm_counter(mm, anon_rss);
2049 pte = mk_pte(page, vma->vm_page_prot);
2050 if (write_access && can_share_swap_page(page)) {
2051 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2055 flush_icache_page(vma, page);
2056 set_pte_at(mm, address, page_table, pte);
2057 page_add_anon_rmap(page, vma, address);
2061 remove_exclusive_swap_page(page);
2065 if (do_wp_page(mm, vma, address,
2066 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2071 /* No need to invalidate - it was non-present before */
2072 update_mmu_cache(vma, address, pte);
2073 lazy_mmu_prot_update(pte);
2075 pte_unmap_unlock(page_table, ptl);
2079 pte_unmap_unlock(page_table, ptl);
2081 page_cache_release(page);
2086 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2087 * but allow concurrent faults), and pte mapped but not yet locked.
2088 * We return with mmap_sem still held, but pte unmapped and unlocked.
2090 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2091 unsigned long address, pte_t *page_table, pmd_t *pmd,
2099 /* Allocate our own private page. */
2100 pte_unmap(page_table);
2102 if (!vx_rsspages_avail(mm, 1))
2104 if (unlikely(anon_vma_prepare(vma)))
2106 page = alloc_zeroed_user_highpage(vma, address);
2110 entry = mk_pte(page, vma->vm_page_prot);
2111 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2113 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2114 if (!pte_none(*page_table))
2116 inc_mm_counter(mm, anon_rss);
2117 lru_cache_add_active(page);
2118 page_add_new_anon_rmap(page, vma, address);
2120 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2121 page = ZERO_PAGE(address);
2122 page_cache_get(page);
2123 entry = mk_pte(page, vma->vm_page_prot);
2125 ptl = pte_lockptr(mm, pmd);
2127 if (!pte_none(*page_table))
2129 inc_mm_counter(mm, file_rss);
2130 page_add_file_rmap(page);
2133 set_pte_at(mm, address, page_table, entry);
2135 /* No need to invalidate - it was non-present before */
2136 update_mmu_cache(vma, address, entry);
2137 lazy_mmu_prot_update(entry);
2139 pte_unmap_unlock(page_table, ptl);
2140 return VM_FAULT_MINOR;
2142 page_cache_release(page);
2145 return VM_FAULT_OOM;
2149 * do_no_page() tries to create a new page mapping. It aggressively
2150 * tries to share with existing pages, but makes a separate copy if
2151 * the "write_access" parameter is true in order to avoid the next
2154 * As this is called only for pages that do not currently exist, we
2155 * do not need to flush old virtual caches or the TLB.
2157 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2158 * but allow concurrent faults), and pte mapped but not yet locked.
2159 * We return with mmap_sem still held, but pte unmapped and unlocked.
2161 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2162 unsigned long address, pte_t *page_table, pmd_t *pmd,
2166 struct page *new_page;
2167 struct address_space *mapping = NULL;
2169 unsigned int sequence = 0;
2170 int ret = VM_FAULT_MINOR;
2173 pte_unmap(page_table);
2174 BUG_ON(vma->vm_flags & VM_PFNMAP);
2177 mapping = vma->vm_file->f_mapping;
2178 sequence = mapping->truncate_count;
2179 smp_rmb(); /* serializes i_size against truncate_count */
2182 /* FIXME: is that check useful here? */
2183 if (!vx_rsspages_avail(mm, 1))
2184 return VM_FAULT_OOM;
2185 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2187 * No smp_rmb is needed here as long as there's a full
2188 * spin_lock/unlock sequence inside the ->nopage callback
2189 * (for the pagecache lookup) that acts as an implicit
2190 * smp_mb() and prevents the i_size read to happen
2191 * after the next truncate_count read.
2194 /* no page was available -- either SIGBUS or OOM */
2195 if (new_page == NOPAGE_SIGBUS)
2196 return VM_FAULT_SIGBUS;
2197 if (new_page == NOPAGE_OOM)
2198 return VM_FAULT_OOM;
2201 * Should we do an early C-O-W break?
2203 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2206 if (unlikely(anon_vma_prepare(vma)))
2208 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2211 copy_user_highpage(page, new_page, address);
2212 page_cache_release(new_page);
2217 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2219 * For a file-backed vma, someone could have truncated or otherwise
2220 * invalidated this page. If unmap_mapping_range got called,
2221 * retry getting the page.
2223 if (mapping && unlikely(sequence != mapping->truncate_count)) {
2224 pte_unmap_unlock(page_table, ptl);
2225 page_cache_release(new_page);
2227 sequence = mapping->truncate_count;
2233 * This silly early PAGE_DIRTY setting removes a race
2234 * due to the bad i386 page protection. But it's valid
2235 * for other architectures too.
2237 * Note that if write_access is true, we either now have
2238 * an exclusive copy of the page, or this is a shared mapping,
2239 * so we can make it writable and dirty to avoid having to
2240 * handle that later.
2242 /* Only go through if we didn't race with anybody else... */
2243 if (pte_none(*page_table)) {
2244 flush_icache_page(vma, new_page);
2245 entry = mk_pte(new_page, vma->vm_page_prot);
2247 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2248 set_pte_at(mm, address, page_table, entry);
2250 inc_mm_counter(mm, anon_rss);
2251 lru_cache_add_active(new_page);
2252 page_add_new_anon_rmap(new_page, vma, address);
2254 inc_mm_counter(mm, file_rss);
2255 page_add_file_rmap(new_page);
2258 /* One of our sibling threads was faster, back out. */
2259 page_cache_release(new_page);
2263 /* no need to invalidate: a not-present page shouldn't be cached */
2264 update_mmu_cache(vma, address, entry);
2265 lazy_mmu_prot_update(entry);
2267 pte_unmap_unlock(page_table, ptl);
2270 page_cache_release(new_page);
2271 return VM_FAULT_OOM;
2275 * Fault of a previously existing named mapping. Repopulate the pte
2276 * from the encoded file_pte if possible. This enables swappable
2279 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2280 * but allow concurrent faults), and pte mapped but not yet locked.
2281 * We return with mmap_sem still held, but pte unmapped and unlocked.
2283 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2284 unsigned long address, pte_t *page_table, pmd_t *pmd,
2285 int write_access, pte_t orig_pte)
2290 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2291 return VM_FAULT_MINOR;
2293 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2295 * Page table corrupted: show pte and kill process.
2297 print_bad_pte(vma, orig_pte, address);
2298 return VM_FAULT_OOM;
2300 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2302 pgoff = pte_to_pgoff(orig_pte);
2303 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2304 vma->vm_page_prot, pgoff, 0);
2306 return VM_FAULT_OOM;
2308 return VM_FAULT_SIGBUS;
2309 return VM_FAULT_MAJOR;
2313 * These routines also need to handle stuff like marking pages dirty
2314 * and/or accessed for architectures that don't do it in hardware (most
2315 * RISC architectures). The early dirtying is also good on the i386.
2317 * There is also a hook called "update_mmu_cache()" that architectures
2318 * with external mmu caches can use to update those (ie the Sparc or
2319 * PowerPC hashed page tables that act as extended TLBs).
2321 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2322 * but allow concurrent faults), and pte mapped but not yet locked.
2323 * We return with mmap_sem still held, but pte unmapped and unlocked.
2325 static inline int handle_pte_fault(struct mm_struct *mm,
2326 struct vm_area_struct *vma, unsigned long address,
2327 pte_t *pte, pmd_t *pmd, int write_access)
2333 old_entry = entry = *pte;
2334 if (!pte_present(entry)) {
2335 if (pte_none(entry)) {
2336 if (!vma->vm_ops || !vma->vm_ops->nopage)
2337 return do_anonymous_page(mm, vma, address,
2338 pte, pmd, write_access);
2339 return do_no_page(mm, vma, address,
2340 pte, pmd, write_access);
2342 if (pte_file(entry))
2343 return do_file_page(mm, vma, address,
2344 pte, pmd, write_access, entry);
2345 return do_swap_page(mm, vma, address,
2346 pte, pmd, write_access, entry);
2349 ptl = pte_lockptr(mm, pmd);
2351 if (unlikely(!pte_same(*pte, entry)))
2354 if (!pte_write(entry))
2355 return do_wp_page(mm, vma, address,
2356 pte, pmd, ptl, entry);
2357 entry = pte_mkdirty(entry);
2359 entry = pte_mkyoung(entry);
2360 if (!pte_same(old_entry, entry)) {
2361 ptep_set_access_flags(vma, address, pte, entry, write_access);
2362 update_mmu_cache(vma, address, entry);
2363 lazy_mmu_prot_update(entry);
2366 * This is needed only for protection faults but the arch code
2367 * is not yet telling us if this is a protection fault or not.
2368 * This still avoids useless tlb flushes for .text page faults
2372 flush_tlb_page(vma, address);
2375 pte_unmap_unlock(pte, ptl);
2376 return VM_FAULT_MINOR;
2380 * By the time we get here, we already hold the mm semaphore
2382 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2383 unsigned long address, int write_access)
2390 __set_current_state(TASK_RUNNING);
2392 inc_page_state(pgfault);
2394 if (unlikely(is_vm_hugetlb_page(vma)))
2395 return hugetlb_fault(mm, vma, address, write_access);
2397 pgd = pgd_offset(mm, address);
2398 pud = pud_alloc(mm, pgd, address);
2400 return VM_FAULT_OOM;
2401 pmd = pmd_alloc(mm, pud, address);
2403 return VM_FAULT_OOM;
2404 pte = pte_alloc_map(mm, pmd, address);
2406 return VM_FAULT_OOM;
2408 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2411 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2413 #ifndef __PAGETABLE_PUD_FOLDED
2415 * Allocate page upper directory.
2416 * We've already handled the fast-path in-line.
2418 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2420 pud_t *new = pud_alloc_one(mm, address);
2424 spin_lock(&mm->page_table_lock);
2425 if (pgd_present(*pgd)) /* Another has populated it */
2428 pgd_populate(mm, pgd, new);
2429 spin_unlock(&mm->page_table_lock);
2433 /* Workaround for gcc 2.96 */
2434 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2438 #endif /* __PAGETABLE_PUD_FOLDED */
2440 #ifndef __PAGETABLE_PMD_FOLDED
2442 * Allocate page middle directory.
2443 * We've already handled the fast-path in-line.
2445 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2447 pmd_t *new = pmd_alloc_one(mm, address);
2451 spin_lock(&mm->page_table_lock);
2452 #ifndef __ARCH_HAS_4LEVEL_HACK
2453 if (pud_present(*pud)) /* Another has populated it */
2456 pud_populate(mm, pud, new);
2458 if (pgd_present(*pud)) /* Another has populated it */
2461 pgd_populate(mm, pud, new);
2462 #endif /* __ARCH_HAS_4LEVEL_HACK */
2463 spin_unlock(&mm->page_table_lock);
2467 /* Workaround for gcc 2.96 */
2468 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2472 #endif /* __PAGETABLE_PMD_FOLDED */
2474 int make_pages_present(unsigned long addr, unsigned long end)
2476 int ret, len, write;
2477 struct vm_area_struct * vma;
2479 vma = find_vma(current->mm, addr);
2482 write = (vma->vm_flags & VM_WRITE) != 0;
2483 BUG_ON(addr >= end);
2484 BUG_ON(end > vma->vm_end);
2485 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2486 ret = get_user_pages(current, current->mm, addr,
2487 len, write, 0, NULL, NULL);
2490 return ret == len ? 0 : -1;
2494 * Map a vmalloc()-space virtual address to the physical page.
2496 struct page * vmalloc_to_page(void * vmalloc_addr)
2498 unsigned long addr = (unsigned long) vmalloc_addr;
2499 struct page *page = NULL;
2500 pgd_t *pgd = pgd_offset_k(addr);
2505 if (!pgd_none(*pgd)) {
2506 pud = pud_offset(pgd, addr);
2507 if (!pud_none(*pud)) {
2508 pmd = pmd_offset(pud, addr);
2509 if (!pmd_none(*pmd)) {
2510 ptep = pte_offset_map(pmd, addr);
2512 if (pte_present(pte))
2513 page = pte_page(pte);
2521 EXPORT_SYMBOL(vmalloc_to_page);
2524 * Map a vmalloc()-space virtual address to the physical page frame number.
2526 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2528 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2531 EXPORT_SYMBOL(vmalloc_to_pfn);
2533 #if !defined(__HAVE_ARCH_GATE_AREA)
2535 #if defined(AT_SYSINFO_EHDR)
2536 static struct vm_area_struct gate_vma;
2538 static int __init gate_vma_init(void)
2540 gate_vma.vm_mm = NULL;
2541 gate_vma.vm_start = FIXADDR_USER_START;
2542 gate_vma.vm_end = FIXADDR_USER_END;
2543 gate_vma.vm_page_prot = PAGE_READONLY;
2544 gate_vma.vm_flags = 0;
2547 __initcall(gate_vma_init);
2550 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2552 #ifdef AT_SYSINFO_EHDR
2559 int in_gate_area_no_task(unsigned long addr)
2561 #ifdef AT_SYSINFO_EHDR
2562 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2568 #endif /* __HAVE_ARCH_GATE_AREA */