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/delayacct.h>
51 #include <linux/init.h>
52 #include <linux/writeback.h>
53 #include <linux/memcontrol.h>
54 #include <linux/mmu_notifier.h>
55 #include <linux/kallsyms.h>
56 #include <linux/swapops.h>
57 #include <linux/elf.h>
59 #include <asm/pgalloc.h>
60 #include <asm/uaccess.h>
62 #include <asm/tlbflush.h>
63 #include <asm/pgtable.h>
67 #ifndef CONFIG_NEED_MULTIPLE_NODES
68 /* use the per-pgdat data instead for discontigmem - mbligh */
69 unsigned long max_mapnr;
72 EXPORT_SYMBOL(max_mapnr);
73 EXPORT_SYMBOL(mem_map);
76 unsigned long num_physpages;
78 * A number of key systems in x86 including ioremap() rely on the assumption
79 * that high_memory defines the upper bound on direct map memory, then end
80 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
81 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
86 EXPORT_SYMBOL(num_physpages);
87 EXPORT_SYMBOL(high_memory);
90 * Randomize the address space (stacks, mmaps, brk, etc.).
92 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
93 * as ancient (libc5 based) binaries can segfault. )
95 int randomize_va_space __read_mostly =
96 #ifdef CONFIG_COMPAT_BRK
102 static int __init disable_randmaps(char *s)
104 randomize_va_space = 0;
107 __setup("norandmaps", disable_randmaps);
111 * If a p?d_bad entry is found while walking page tables, report
112 * the error, before resetting entry to p?d_none. Usually (but
113 * very seldom) called out from the p?d_none_or_clear_bad macros.
116 void pgd_clear_bad(pgd_t *pgd)
122 void pud_clear_bad(pud_t *pud)
128 void pmd_clear_bad(pmd_t *pmd)
135 * Note: this doesn't free the actual pages themselves. That
136 * has been handled earlier when unmapping all the memory regions.
138 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
140 pgtable_t token = pmd_pgtable(*pmd);
142 pte_free_tlb(tlb, token);
146 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
147 unsigned long addr, unsigned long end,
148 unsigned long floor, unsigned long ceiling)
155 pmd = pmd_offset(pud, addr);
157 next = pmd_addr_end(addr, end);
158 if (pmd_none_or_clear_bad(pmd))
160 free_pte_range(tlb, pmd);
161 } while (pmd++, addr = next, addr != end);
171 if (end - 1 > ceiling - 1)
174 pmd = pmd_offset(pud, start);
176 pmd_free_tlb(tlb, pmd);
179 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
180 unsigned long addr, unsigned long end,
181 unsigned long floor, unsigned long ceiling)
188 pud = pud_offset(pgd, addr);
190 next = pud_addr_end(addr, end);
191 if (pud_none_or_clear_bad(pud))
193 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
194 } while (pud++, addr = next, addr != end);
200 ceiling &= PGDIR_MASK;
204 if (end - 1 > ceiling - 1)
207 pud = pud_offset(pgd, start);
209 pud_free_tlb(tlb, pud);
213 * This function frees user-level page tables of a process.
215 * Must be called with pagetable lock held.
217 void free_pgd_range(struct mmu_gather *tlb,
218 unsigned long addr, unsigned long end,
219 unsigned long floor, unsigned long ceiling)
226 * The next few lines have given us lots of grief...
228 * Why are we testing PMD* at this top level? Because often
229 * there will be no work to do at all, and we'd prefer not to
230 * go all the way down to the bottom just to discover that.
232 * Why all these "- 1"s? Because 0 represents both the bottom
233 * of the address space and the top of it (using -1 for the
234 * top wouldn't help much: the masks would do the wrong thing).
235 * The rule is that addr 0 and floor 0 refer to the bottom of
236 * the address space, but end 0 and ceiling 0 refer to the top
237 * Comparisons need to use "end - 1" and "ceiling - 1" (though
238 * that end 0 case should be mythical).
240 * Wherever addr is brought up or ceiling brought down, we must
241 * be careful to reject "the opposite 0" before it confuses the
242 * subsequent tests. But what about where end is brought down
243 * by PMD_SIZE below? no, end can't go down to 0 there.
245 * Whereas we round start (addr) and ceiling down, by different
246 * masks at different levels, in order to test whether a table
247 * now has no other vmas using it, so can be freed, we don't
248 * bother to round floor or end up - the tests don't need that.
262 if (end - 1 > ceiling - 1)
268 pgd = pgd_offset(tlb->mm, addr);
270 next = pgd_addr_end(addr, end);
271 if (pgd_none_or_clear_bad(pgd))
273 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
274 } while (pgd++, addr = next, addr != end);
277 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
278 unsigned long floor, unsigned long ceiling)
281 struct vm_area_struct *next = vma->vm_next;
282 unsigned long addr = vma->vm_start;
285 * Hide vma from rmap and vmtruncate before freeing pgtables
287 anon_vma_unlink(vma);
288 unlink_file_vma(vma);
290 if (is_vm_hugetlb_page(vma)) {
291 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
292 floor, next? next->vm_start: ceiling);
295 * Optimization: gather nearby vmas into one call down
297 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
298 && !is_vm_hugetlb_page(next)) {
301 anon_vma_unlink(vma);
302 unlink_file_vma(vma);
304 free_pgd_range(tlb, addr, vma->vm_end,
305 floor, next? next->vm_start: ceiling);
311 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
313 pgtable_t new = pte_alloc_one(mm, address);
318 * Ensure all pte setup (eg. pte page lock and page clearing) are
319 * visible before the pte is made visible to other CPUs by being
320 * put into page tables.
322 * The other side of the story is the pointer chasing in the page
323 * table walking code (when walking the page table without locking;
324 * ie. most of the time). Fortunately, these data accesses consist
325 * of a chain of data-dependent loads, meaning most CPUs (alpha
326 * being the notable exception) will already guarantee loads are
327 * seen in-order. See the alpha page table accessors for the
328 * smp_read_barrier_depends() barriers in page table walking code.
330 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
332 spin_lock(&mm->page_table_lock);
333 if (!pmd_present(*pmd)) { /* Has another populated it ? */
335 pmd_populate(mm, pmd, new);
338 spin_unlock(&mm->page_table_lock);
344 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
346 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
350 smp_wmb(); /* See comment in __pte_alloc */
352 spin_lock(&init_mm.page_table_lock);
353 if (!pmd_present(*pmd)) { /* Has another populated it ? */
354 pmd_populate_kernel(&init_mm, pmd, new);
357 spin_unlock(&init_mm.page_table_lock);
359 pte_free_kernel(&init_mm, new);
363 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
366 add_mm_counter(mm, file_rss, file_rss);
368 add_mm_counter(mm, anon_rss, anon_rss);
372 * This function is called to print an error when a bad pte
373 * is found. For example, we might have a PFN-mapped pte in
374 * a region that doesn't allow it.
376 * The calling function must still handle the error.
378 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
379 pte_t pte, struct page *page)
381 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
382 pud_t *pud = pud_offset(pgd, addr);
383 pmd_t *pmd = pmd_offset(pud, addr);
384 struct address_space *mapping;
386 static unsigned long resume;
387 static unsigned long nr_shown;
388 static unsigned long nr_unshown;
391 * Allow a burst of 60 reports, then keep quiet for that minute;
392 * or allow a steady drip of one report per second.
394 if (nr_shown == 60) {
395 if (time_before(jiffies, resume)) {
401 "BUG: Bad page map: %lu messages suppressed\n",
408 resume = jiffies + 60 * HZ;
410 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
411 index = linear_page_index(vma, addr);
414 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
416 (long long)pte_val(pte), (long long)pmd_val(*pmd));
419 "page:%p flags:%p count:%d mapcount:%d mapping:%p index:%lx\n",
420 page, (void *)page->flags, page_count(page),
421 page_mapcount(page), page->mapping, page->index);
424 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
425 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
427 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
430 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
431 (unsigned long)vma->vm_ops->fault);
432 if (vma->vm_file && vma->vm_file->f_op)
433 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
434 (unsigned long)vma->vm_file->f_op->mmap);
436 add_taint(TAINT_BAD_PAGE);
439 static inline int is_cow_mapping(unsigned int flags)
441 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
445 * vm_normal_page -- This function gets the "struct page" associated with a pte.
447 * "Special" mappings do not wish to be associated with a "struct page" (either
448 * it doesn't exist, or it exists but they don't want to touch it). In this
449 * case, NULL is returned here. "Normal" mappings do have a struct page.
451 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
452 * pte bit, in which case this function is trivial. Secondly, an architecture
453 * may not have a spare pte bit, which requires a more complicated scheme,
456 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
457 * special mapping (even if there are underlying and valid "struct pages").
458 * COWed pages of a VM_PFNMAP are always normal.
460 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
461 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
462 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
463 * mapping will always honor the rule
465 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
467 * And for normal mappings this is false.
469 * This restricts such mappings to be a linear translation from virtual address
470 * to pfn. To get around this restriction, we allow arbitrary mappings so long
471 * as the vma is not a COW mapping; in that case, we know that all ptes are
472 * special (because none can have been COWed).
475 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
477 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
478 * page" backing, however the difference is that _all_ pages with a struct
479 * page (that is, those where pfn_valid is true) are refcounted and considered
480 * normal pages by the VM. The disadvantage is that pages are refcounted
481 * (which can be slower and simply not an option for some PFNMAP users). The
482 * advantage is that we don't have to follow the strict linearity rule of
483 * PFNMAP mappings in order to support COWable mappings.
486 #ifdef __HAVE_ARCH_PTE_SPECIAL
487 # define HAVE_PTE_SPECIAL 1
489 # define HAVE_PTE_SPECIAL 0
491 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
494 unsigned long pfn = pte_pfn(pte);
496 if (HAVE_PTE_SPECIAL) {
497 if (likely(!pte_special(pte)))
499 if (!(vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)))
500 print_bad_pte(vma, addr, pte, NULL);
504 /* !HAVE_PTE_SPECIAL case follows: */
506 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
507 if (vma->vm_flags & VM_MIXEDMAP) {
513 off = (addr - vma->vm_start) >> PAGE_SHIFT;
514 if (pfn == vma->vm_pgoff + off)
516 if (!is_cow_mapping(vma->vm_flags))
522 if (unlikely(pfn > highest_memmap_pfn)) {
523 print_bad_pte(vma, addr, pte, NULL);
528 * NOTE! We still have PageReserved() pages in the page tables.
529 * eg. VDSO mappings can cause them to exist.
532 return pfn_to_page(pfn);
536 * copy one vm_area from one task to the other. Assumes the page tables
537 * already present in the new task to be cleared in the whole range
538 * covered by this vma.
542 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
543 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
544 unsigned long addr, int *rss)
546 unsigned long vm_flags = vma->vm_flags;
547 pte_t pte = *src_pte;
550 /* pte contains position in swap or file, so copy. */
551 if (unlikely(!pte_present(pte))) {
552 if (!pte_file(pte)) {
553 swp_entry_t entry = pte_to_swp_entry(pte);
555 swap_duplicate(entry);
556 /* make sure dst_mm is on swapoff's mmlist. */
557 if (unlikely(list_empty(&dst_mm->mmlist))) {
558 spin_lock(&mmlist_lock);
559 if (list_empty(&dst_mm->mmlist))
560 list_add(&dst_mm->mmlist,
562 spin_unlock(&mmlist_lock);
564 if (is_write_migration_entry(entry) &&
565 is_cow_mapping(vm_flags)) {
567 * COW mappings require pages in both parent
568 * and child to be set to read.
570 make_migration_entry_read(&entry);
571 pte = swp_entry_to_pte(entry);
572 set_pte_at(src_mm, addr, src_pte, pte);
579 * If it's a COW mapping, write protect it both
580 * in the parent and the child
582 if (is_cow_mapping(vm_flags)) {
583 ptep_set_wrprotect(src_mm, addr, src_pte);
584 pte = pte_wrprotect(pte);
588 * If it's a shared mapping, mark it clean in
591 if (vm_flags & VM_SHARED)
592 pte = pte_mkclean(pte);
593 pte = pte_mkold(pte);
595 page = vm_normal_page(vma, addr, pte);
598 page_dup_rmap(page, vma, addr);
599 rss[!!PageAnon(page)]++;
603 set_pte_at(dst_mm, addr, dst_pte, pte);
606 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
607 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
608 unsigned long addr, unsigned long end)
610 pte_t *src_pte, *dst_pte;
611 spinlock_t *src_ptl, *dst_ptl;
617 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
620 src_pte = pte_offset_map_nested(src_pmd, addr);
621 src_ptl = pte_lockptr(src_mm, src_pmd);
622 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
623 arch_enter_lazy_mmu_mode();
627 * We are holding two locks at this point - either of them
628 * could generate latencies in another task on another CPU.
630 if (progress >= 32) {
632 if (need_resched() ||
633 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
636 if (pte_none(*src_pte)) {
640 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
642 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
644 arch_leave_lazy_mmu_mode();
645 spin_unlock(src_ptl);
646 pte_unmap_nested(src_pte - 1);
647 add_mm_rss(dst_mm, rss[0], rss[1]);
648 pte_unmap_unlock(dst_pte - 1, dst_ptl);
655 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
656 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
657 unsigned long addr, unsigned long end)
659 pmd_t *src_pmd, *dst_pmd;
662 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
665 src_pmd = pmd_offset(src_pud, addr);
667 next = pmd_addr_end(addr, end);
668 if (pmd_none_or_clear_bad(src_pmd))
670 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
673 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
677 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
678 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
679 unsigned long addr, unsigned long end)
681 pud_t *src_pud, *dst_pud;
684 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
687 src_pud = pud_offset(src_pgd, addr);
689 next = pud_addr_end(addr, end);
690 if (pud_none_or_clear_bad(src_pud))
692 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
695 } while (dst_pud++, src_pud++, addr = next, addr != end);
699 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700 struct vm_area_struct *vma)
702 pgd_t *src_pgd, *dst_pgd;
704 unsigned long addr = vma->vm_start;
705 unsigned long end = vma->vm_end;
709 * Don't copy ptes where a page fault will fill them correctly.
710 * Fork becomes much lighter when there are big shared or private
711 * readonly mappings. The tradeoff is that copy_page_range is more
712 * efficient than faulting.
714 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
719 if (is_vm_hugetlb_page(vma))
720 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
722 if (unlikely(is_pfn_mapping(vma))) {
724 * We do not free on error cases below as remove_vma
725 * gets called on error from higher level routine
727 ret = track_pfn_vma_copy(vma);
733 * We need to invalidate the secondary MMU mappings only when
734 * there could be a permission downgrade on the ptes of the
735 * parent mm. And a permission downgrade will only happen if
736 * is_cow_mapping() returns true.
738 if (is_cow_mapping(vma->vm_flags))
739 mmu_notifier_invalidate_range_start(src_mm, addr, end);
742 dst_pgd = pgd_offset(dst_mm, addr);
743 src_pgd = pgd_offset(src_mm, addr);
745 next = pgd_addr_end(addr, end);
746 if (pgd_none_or_clear_bad(src_pgd))
748 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
753 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
755 if (is_cow_mapping(vma->vm_flags))
756 mmu_notifier_invalidate_range_end(src_mm,
761 static unsigned long zap_pte_range(struct mmu_gather *tlb,
762 struct vm_area_struct *vma, pmd_t *pmd,
763 unsigned long addr, unsigned long end,
764 long *zap_work, struct zap_details *details)
766 struct mm_struct *mm = tlb->mm;
772 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
773 arch_enter_lazy_mmu_mode();
776 if (pte_none(ptent)) {
781 (*zap_work) -= PAGE_SIZE;
783 if (pte_present(ptent)) {
786 page = vm_normal_page(vma, addr, ptent);
787 if (unlikely(details) && page) {
789 * unmap_shared_mapping_pages() wants to
790 * invalidate cache without truncating:
791 * unmap shared but keep private pages.
793 if (details->check_mapping &&
794 details->check_mapping != page->mapping)
797 * Each page->index must be checked when
798 * invalidating or truncating nonlinear.
800 if (details->nonlinear_vma &&
801 (page->index < details->first_index ||
802 page->index > details->last_index))
805 ptent = ptep_get_and_clear_full(mm, addr, pte,
807 tlb_remove_tlb_entry(tlb, pte, addr);
810 if (unlikely(details) && details->nonlinear_vma
811 && linear_page_index(details->nonlinear_vma,
812 addr) != page->index)
813 set_pte_at(mm, addr, pte,
814 pgoff_to_pte(page->index));
818 if (pte_dirty(ptent))
819 set_page_dirty(page);
820 if (pte_young(ptent) &&
821 likely(!VM_SequentialReadHint(vma)))
822 mark_page_accessed(page);
825 page_remove_rmap(page);
826 if (unlikely(page_mapcount(page) < 0))
827 print_bad_pte(vma, addr, ptent, page);
828 tlb_remove_page(tlb, page);
832 * If details->check_mapping, we leave swap entries;
833 * if details->nonlinear_vma, we leave file entries.
835 if (unlikely(details))
837 if (pte_file(ptent)) {
838 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
839 print_bad_pte(vma, addr, ptent, NULL);
841 (unlikely(!free_swap_and_cache(pte_to_swp_entry(ptent))))
842 print_bad_pte(vma, addr, ptent, NULL);
843 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
844 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
846 add_mm_rss(mm, file_rss, anon_rss);
847 arch_leave_lazy_mmu_mode();
848 pte_unmap_unlock(pte - 1, ptl);
853 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
854 struct vm_area_struct *vma, pud_t *pud,
855 unsigned long addr, unsigned long end,
856 long *zap_work, struct zap_details *details)
861 pmd = pmd_offset(pud, addr);
863 next = pmd_addr_end(addr, end);
864 if (pmd_none_or_clear_bad(pmd)) {
868 next = zap_pte_range(tlb, vma, pmd, addr, next,
870 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
875 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
876 struct vm_area_struct *vma, pgd_t *pgd,
877 unsigned long addr, unsigned long end,
878 long *zap_work, struct zap_details *details)
883 pud = pud_offset(pgd, addr);
885 next = pud_addr_end(addr, end);
886 if (pud_none_or_clear_bad(pud)) {
890 next = zap_pmd_range(tlb, vma, pud, addr, next,
892 } while (pud++, addr = next, (addr != end && *zap_work > 0));
897 static unsigned long unmap_page_range(struct mmu_gather *tlb,
898 struct vm_area_struct *vma,
899 unsigned long addr, unsigned long end,
900 long *zap_work, struct zap_details *details)
905 if (details && !details->check_mapping && !details->nonlinear_vma)
909 tlb_start_vma(tlb, vma);
910 pgd = pgd_offset(vma->vm_mm, addr);
912 next = pgd_addr_end(addr, end);
913 if (pgd_none_or_clear_bad(pgd)) {
917 next = zap_pud_range(tlb, vma, pgd, addr, next,
919 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
920 tlb_end_vma(tlb, vma);
925 #ifdef CONFIG_PREEMPT
926 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
928 /* No preempt: go for improved straight-line efficiency */
929 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
933 * unmap_vmas - unmap a range of memory covered by a list of vma's
934 * @tlbp: address of the caller's struct mmu_gather
935 * @vma: the starting vma
936 * @start_addr: virtual address at which to start unmapping
937 * @end_addr: virtual address at which to end unmapping
938 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
939 * @details: details of nonlinear truncation or shared cache invalidation
941 * Returns the end address of the unmapping (restart addr if interrupted).
943 * Unmap all pages in the vma list.
945 * We aim to not hold locks for too long (for scheduling latency reasons).
946 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
947 * return the ending mmu_gather to the caller.
949 * Only addresses between `start' and `end' will be unmapped.
951 * The VMA list must be sorted in ascending virtual address order.
953 * unmap_vmas() assumes that the caller will flush the whole unmapped address
954 * range after unmap_vmas() returns. So the only responsibility here is to
955 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
956 * drops the lock and schedules.
958 unsigned long unmap_vmas(struct mmu_gather **tlbp,
959 struct vm_area_struct *vma, unsigned long start_addr,
960 unsigned long end_addr, unsigned long *nr_accounted,
961 struct zap_details *details)
963 long zap_work = ZAP_BLOCK_SIZE;
964 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
965 int tlb_start_valid = 0;
966 unsigned long start = start_addr;
967 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
968 int fullmm = (*tlbp)->fullmm;
969 struct mm_struct *mm = vma->vm_mm;
971 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
972 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
975 start = max(vma->vm_start, start_addr);
976 if (start >= vma->vm_end)
978 end = min(vma->vm_end, end_addr);
979 if (end <= vma->vm_start)
982 if (vma->vm_flags & VM_ACCOUNT)
983 *nr_accounted += (end - start) >> PAGE_SHIFT;
985 if (unlikely(is_pfn_mapping(vma)))
986 untrack_pfn_vma(vma, 0, 0);
988 while (start != end) {
989 if (!tlb_start_valid) {
994 if (unlikely(is_vm_hugetlb_page(vma))) {
996 * It is undesirable to test vma->vm_file as it
997 * should be non-null for valid hugetlb area.
998 * However, vm_file will be NULL in the error
999 * cleanup path of do_mmap_pgoff. When
1000 * hugetlbfs ->mmap method fails,
1001 * do_mmap_pgoff() nullifies vma->vm_file
1002 * before calling this function to clean up.
1003 * Since no pte has actually been setup, it is
1004 * safe to do nothing in this case.
1007 unmap_hugepage_range(vma, start, end, NULL);
1008 zap_work -= (end - start) /
1009 pages_per_huge_page(hstate_vma(vma));
1014 start = unmap_page_range(*tlbp, vma,
1015 start, end, &zap_work, details);
1018 BUG_ON(start != end);
1022 tlb_finish_mmu(*tlbp, tlb_start, start);
1024 if (need_resched() ||
1025 (i_mmap_lock && spin_needbreak(i_mmap_lock))) {
1033 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1034 tlb_start_valid = 0;
1035 zap_work = ZAP_BLOCK_SIZE;
1039 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1040 return start; /* which is now the end (or restart) address */
1044 * zap_page_range - remove user pages in a given range
1045 * @vma: vm_area_struct holding the applicable pages
1046 * @address: starting address of pages to zap
1047 * @size: number of bytes to zap
1048 * @details: details of nonlinear truncation or shared cache invalidation
1050 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1051 unsigned long size, struct zap_details *details)
1053 struct mm_struct *mm = vma->vm_mm;
1054 struct mmu_gather *tlb;
1055 unsigned long end = address + size;
1056 unsigned long nr_accounted = 0;
1059 tlb = tlb_gather_mmu(mm, 0);
1060 update_hiwater_rss(mm);
1061 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1063 tlb_finish_mmu(tlb, address, end);
1068 * zap_vma_ptes - remove ptes mapping the vma
1069 * @vma: vm_area_struct holding ptes to be zapped
1070 * @address: starting address of pages to zap
1071 * @size: number of bytes to zap
1073 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1075 * The entire address range must be fully contained within the vma.
1077 * Returns 0 if successful.
1079 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1082 if (address < vma->vm_start || address + size > vma->vm_end ||
1083 !(vma->vm_flags & VM_PFNMAP))
1085 zap_page_range(vma, address, size, NULL);
1088 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1091 * Do a quick page-table lookup for a single page.
1093 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1102 struct mm_struct *mm = vma->vm_mm;
1104 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1105 if (!IS_ERR(page)) {
1106 BUG_ON(flags & FOLL_GET);
1111 pgd = pgd_offset(mm, address);
1112 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1115 pud = pud_offset(pgd, address);
1118 if (pud_huge(*pud)) {
1119 BUG_ON(flags & FOLL_GET);
1120 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1123 if (unlikely(pud_bad(*pud)))
1126 pmd = pmd_offset(pud, address);
1129 if (pmd_huge(*pmd)) {
1130 BUG_ON(flags & FOLL_GET);
1131 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1134 if (unlikely(pmd_bad(*pmd)))
1137 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1140 if (!pte_present(pte))
1142 if ((flags & FOLL_WRITE) && !pte_write(pte))
1144 page = vm_normal_page(vma, address, pte);
1145 if (unlikely(!page))
1148 if (flags & FOLL_GET)
1150 if (flags & FOLL_TOUCH) {
1151 if ((flags & FOLL_WRITE) &&
1152 !pte_dirty(pte) && !PageDirty(page))
1153 set_page_dirty(page);
1155 * pte_mkyoung() would be more correct here, but atomic care
1156 * is needed to avoid losing the dirty bit: it is easier to use
1157 * mark_page_accessed().
1159 mark_page_accessed(page);
1162 pte_unmap_unlock(ptep, ptl);
1167 pte_unmap_unlock(ptep, ptl);
1168 return ERR_PTR(-EFAULT);
1171 pte_unmap_unlock(ptep, ptl);
1174 /* Fall through to ZERO_PAGE handling */
1177 * When core dumping an enormous anonymous area that nobody
1178 * has touched so far, we don't want to allocate page tables.
1180 if (flags & FOLL_ANON) {
1181 page = ZERO_PAGE(0);
1182 if (flags & FOLL_GET)
1184 BUG_ON(flags & FOLL_WRITE);
1189 /* Can we do the FOLL_ANON optimization? */
1190 static inline int use_zero_page(struct vm_area_struct *vma)
1193 * We don't want to optimize FOLL_ANON for make_pages_present()
1194 * when it tries to page in a VM_LOCKED region. As to VM_SHARED,
1195 * we want to get the page from the page tables to make sure
1196 * that we serialize and update with any other user of that
1199 if (vma->vm_flags & (VM_LOCKED | VM_SHARED))
1202 * And if we have a fault routine, it's not an anonymous region.
1204 return !vma->vm_ops || !vma->vm_ops->fault;
1209 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1210 unsigned long start, int len, int flags,
1211 struct page **pages, struct vm_area_struct **vmas)
1214 unsigned int vm_flags = 0;
1215 int write = !!(flags & GUP_FLAGS_WRITE);
1216 int force = !!(flags & GUP_FLAGS_FORCE);
1217 int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
1218 int ignore_sigkill = !!(flags & GUP_FLAGS_IGNORE_SIGKILL);
1223 * Require read or write permissions.
1224 * If 'force' is set, we only require the "MAY" flags.
1226 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1227 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1231 struct vm_area_struct *vma;
1232 unsigned int foll_flags;
1234 vma = find_extend_vma(mm, start);
1235 if (!vma && in_gate_area(tsk, start)) {
1236 unsigned long pg = start & PAGE_MASK;
1237 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1243 /* user gate pages are read-only */
1244 if (!ignore && write)
1245 return i ? : -EFAULT;
1247 pgd = pgd_offset_k(pg);
1249 pgd = pgd_offset_gate(mm, pg);
1250 BUG_ON(pgd_none(*pgd));
1251 pud = pud_offset(pgd, pg);
1252 BUG_ON(pud_none(*pud));
1253 pmd = pmd_offset(pud, pg);
1255 return i ? : -EFAULT;
1256 pte = pte_offset_map(pmd, pg);
1257 if (pte_none(*pte)) {
1259 return i ? : -EFAULT;
1262 struct page *page = vm_normal_page(gate_vma, start, *pte);
1277 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1278 (!ignore && !(vm_flags & vma->vm_flags)))
1279 return i ? : -EFAULT;
1281 if (is_vm_hugetlb_page(vma)) {
1282 i = follow_hugetlb_page(mm, vma, pages, vmas,
1283 &start, &len, i, write);
1287 foll_flags = FOLL_TOUCH;
1289 foll_flags |= FOLL_GET;
1290 if (!write && use_zero_page(vma))
1291 foll_flags |= FOLL_ANON;
1297 * If we have a pending SIGKILL, don't keep faulting
1298 * pages and potentially allocating memory, unless
1299 * current is handling munlock--e.g., on exit. In
1300 * that case, we are not allocating memory. Rather,
1301 * we're only unlocking already resident/mapped pages.
1303 if (unlikely(!ignore_sigkill &&
1304 fatal_signal_pending(current)))
1305 return i ? i : -ERESTARTSYS;
1308 foll_flags |= FOLL_WRITE;
1311 while (!(page = follow_page(vma, start, foll_flags))) {
1313 ret = handle_mm_fault(mm, vma, start,
1314 foll_flags & FOLL_WRITE);
1315 if (ret & VM_FAULT_ERROR) {
1316 if (ret & VM_FAULT_OOM)
1317 return i ? i : -ENOMEM;
1318 else if (ret & VM_FAULT_SIGBUS)
1319 return i ? i : -EFAULT;
1322 if (ret & VM_FAULT_MAJOR)
1328 * The VM_FAULT_WRITE bit tells us that
1329 * do_wp_page has broken COW when necessary,
1330 * even if maybe_mkwrite decided not to set
1331 * pte_write. We can thus safely do subsequent
1332 * page lookups as if they were reads. But only
1333 * do so when looping for pte_write is futile:
1334 * in some cases userspace may also be wanting
1335 * to write to the gotten user page, which a
1336 * read fault here might prevent (a readonly
1337 * page might get reCOWed by userspace write).
1339 if ((ret & VM_FAULT_WRITE) &&
1340 !(vma->vm_flags & VM_WRITE))
1341 foll_flags &= ~FOLL_WRITE;
1346 return i ? i : PTR_ERR(page);
1350 flush_anon_page(vma, page, start);
1351 flush_dcache_page(page);
1358 } while (len && start < vma->vm_end);
1363 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1364 unsigned long start, int len, int write, int force,
1365 struct page **pages, struct vm_area_struct **vmas)
1370 flags |= GUP_FLAGS_WRITE;
1372 flags |= GUP_FLAGS_FORCE;
1374 return __get_user_pages(tsk, mm,
1379 EXPORT_SYMBOL(get_user_pages);
1381 pte_t *get_locked_pte(struct mm_struct *mm, unsigned long addr,
1384 pgd_t * pgd = pgd_offset(mm, addr);
1385 pud_t * pud = pud_alloc(mm, pgd, addr);
1387 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1389 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1395 * This is the old fallback for page remapping.
1397 * For historical reasons, it only allows reserved pages. Only
1398 * old drivers should use this, and they needed to mark their
1399 * pages reserved for the old functions anyway.
1401 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1402 struct page *page, pgprot_t prot)
1404 struct mm_struct *mm = vma->vm_mm;
1413 flush_dcache_page(page);
1414 pte = get_locked_pte(mm, addr, &ptl);
1418 if (!pte_none(*pte))
1421 /* Ok, finally just insert the thing.. */
1423 inc_mm_counter(mm, file_rss);
1424 page_add_file_rmap(page);
1425 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1428 pte_unmap_unlock(pte, ptl);
1431 pte_unmap_unlock(pte, ptl);
1437 * vm_insert_page - insert single page into user vma
1438 * @vma: user vma to map to
1439 * @addr: target user address of this page
1440 * @page: source kernel page
1442 * This allows drivers to insert individual pages they've allocated
1445 * The page has to be a nice clean _individual_ kernel allocation.
1446 * If you allocate a compound page, you need to have marked it as
1447 * such (__GFP_COMP), or manually just split the page up yourself
1448 * (see split_page()).
1450 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1451 * took an arbitrary page protection parameter. This doesn't allow
1452 * that. Your vma protection will have to be set up correctly, which
1453 * means that if you want a shared writable mapping, you'd better
1454 * ask for a shared writable mapping!
1456 * The page does not need to be reserved.
1458 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1461 if (addr < vma->vm_start || addr >= vma->vm_end)
1463 if (!page_count(page))
1465 vma->vm_flags |= VM_INSERTPAGE;
1466 return insert_page(vma, addr, page, vma->vm_page_prot);
1468 EXPORT_SYMBOL(vm_insert_page);
1470 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1471 unsigned long pfn, pgprot_t prot)
1473 struct mm_struct *mm = vma->vm_mm;
1479 pte = get_locked_pte(mm, addr, &ptl);
1483 if (!pte_none(*pte))
1486 /* Ok, finally just insert the thing.. */
1487 entry = pte_mkspecial(pfn_pte(pfn, prot));
1488 set_pte_at(mm, addr, pte, entry);
1489 update_mmu_cache(vma, addr, entry); /* XXX: why not for insert_page? */
1493 pte_unmap_unlock(pte, ptl);
1499 * vm_insert_pfn - insert single pfn into user vma
1500 * @vma: user vma to map to
1501 * @addr: target user address of this page
1502 * @pfn: source kernel pfn
1504 * Similar to vm_inert_page, this allows drivers to insert individual pages
1505 * they've allocated into a user vma. Same comments apply.
1507 * This function should only be called from a vm_ops->fault handler, and
1508 * in that case the handler should return NULL.
1510 * vma cannot be a COW mapping.
1512 * As this is called only for pages that do not currently exist, we
1513 * do not need to flush old virtual caches or the TLB.
1515 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1519 pgprot_t pgprot = vma->vm_page_prot;
1521 * Technically, architectures with pte_special can avoid all these
1522 * restrictions (same for remap_pfn_range). However we would like
1523 * consistency in testing and feature parity among all, so we should
1524 * try to keep these invariants in place for everybody.
1526 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1527 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1528 (VM_PFNMAP|VM_MIXEDMAP));
1529 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1530 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1532 if (addr < vma->vm_start || addr >= vma->vm_end)
1534 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
1537 ret = insert_pfn(vma, addr, pfn, pgprot);
1540 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
1544 EXPORT_SYMBOL(vm_insert_pfn);
1546 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1549 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1551 if (addr < vma->vm_start || addr >= vma->vm_end)
1555 * If we don't have pte special, then we have to use the pfn_valid()
1556 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1557 * refcount the page if pfn_valid is true (hence insert_page rather
1560 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1563 page = pfn_to_page(pfn);
1564 return insert_page(vma, addr, page, vma->vm_page_prot);
1566 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1568 EXPORT_SYMBOL(vm_insert_mixed);
1571 * maps a range of physical memory into the requested pages. the old
1572 * mappings are removed. any references to nonexistent pages results
1573 * in null mappings (currently treated as "copy-on-access")
1575 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1576 unsigned long addr, unsigned long end,
1577 unsigned long pfn, pgprot_t prot)
1582 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1585 arch_enter_lazy_mmu_mode();
1587 BUG_ON(!pte_none(*pte));
1588 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1590 } while (pte++, addr += PAGE_SIZE, addr != end);
1591 arch_leave_lazy_mmu_mode();
1592 pte_unmap_unlock(pte - 1, ptl);
1596 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1597 unsigned long addr, unsigned long end,
1598 unsigned long pfn, pgprot_t prot)
1603 pfn -= addr >> PAGE_SHIFT;
1604 pmd = pmd_alloc(mm, pud, addr);
1608 next = pmd_addr_end(addr, end);
1609 if (remap_pte_range(mm, pmd, addr, next,
1610 pfn + (addr >> PAGE_SHIFT), prot))
1612 } while (pmd++, addr = next, addr != end);
1616 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1617 unsigned long addr, unsigned long end,
1618 unsigned long pfn, pgprot_t prot)
1623 pfn -= addr >> PAGE_SHIFT;
1624 pud = pud_alloc(mm, pgd, addr);
1628 next = pud_addr_end(addr, end);
1629 if (remap_pmd_range(mm, pud, addr, next,
1630 pfn + (addr >> PAGE_SHIFT), prot))
1632 } while (pud++, addr = next, addr != end);
1637 * remap_pfn_range - remap kernel memory to userspace
1638 * @vma: user vma to map to
1639 * @addr: target user address to start at
1640 * @pfn: physical address of kernel memory
1641 * @size: size of map area
1642 * @prot: page protection flags for this mapping
1644 * Note: this is only safe if the mm semaphore is held when called.
1646 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1647 unsigned long pfn, unsigned long size, pgprot_t prot)
1651 unsigned long end = addr + PAGE_ALIGN(size);
1652 struct mm_struct *mm = vma->vm_mm;
1656 * Physically remapped pages are special. Tell the
1657 * rest of the world about it:
1658 * VM_IO tells people not to look at these pages
1659 * (accesses can have side effects).
1660 * VM_RESERVED is specified all over the place, because
1661 * in 2.4 it kept swapout's vma scan off this vma; but
1662 * in 2.6 the LRU scan won't even find its pages, so this
1663 * flag means no more than count its pages in reserved_vm,
1664 * and omit it from core dump, even when VM_IO turned off.
1665 * VM_PFNMAP tells the core MM that the base pages are just
1666 * raw PFN mappings, and do not have a "struct page" associated
1669 * There's a horrible special case to handle copy-on-write
1670 * behaviour that some programs depend on. We mark the "original"
1671 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1673 if (addr == vma->vm_start && end == vma->vm_end) {
1674 vma->vm_pgoff = pfn;
1675 vma->vm_flags |= VM_PFN_AT_MMAP;
1676 } else if (is_cow_mapping(vma->vm_flags))
1679 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1681 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
1684 * To indicate that track_pfn related cleanup is not
1685 * needed from higher level routine calling unmap_vmas
1687 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
1688 vma->vm_flags &= ~VM_PFN_AT_MMAP;
1692 BUG_ON(addr >= end);
1693 pfn -= addr >> PAGE_SHIFT;
1694 pgd = pgd_offset(mm, addr);
1695 flush_cache_range(vma, addr, end);
1697 next = pgd_addr_end(addr, end);
1698 err = remap_pud_range(mm, pgd, addr, next,
1699 pfn + (addr >> PAGE_SHIFT), prot);
1702 } while (pgd++, addr = next, addr != end);
1705 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
1709 EXPORT_SYMBOL(remap_pfn_range);
1711 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1712 unsigned long addr, unsigned long end,
1713 pte_fn_t fn, void *data)
1718 spinlock_t *uninitialized_var(ptl);
1720 pte = (mm == &init_mm) ?
1721 pte_alloc_kernel(pmd, addr) :
1722 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1726 BUG_ON(pmd_huge(*pmd));
1728 arch_enter_lazy_mmu_mode();
1730 token = pmd_pgtable(*pmd);
1733 err = fn(pte, token, addr, data);
1736 } while (pte++, addr += PAGE_SIZE, addr != end);
1738 arch_leave_lazy_mmu_mode();
1741 pte_unmap_unlock(pte-1, ptl);
1745 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1746 unsigned long addr, unsigned long end,
1747 pte_fn_t fn, void *data)
1753 BUG_ON(pud_huge(*pud));
1755 pmd = pmd_alloc(mm, pud, addr);
1759 next = pmd_addr_end(addr, end);
1760 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1763 } while (pmd++, addr = next, addr != end);
1767 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1768 unsigned long addr, unsigned long end,
1769 pte_fn_t fn, void *data)
1775 pud = pud_alloc(mm, pgd, addr);
1779 next = pud_addr_end(addr, end);
1780 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1783 } while (pud++, addr = next, addr != end);
1788 * Scan a region of virtual memory, filling in page tables as necessary
1789 * and calling a provided function on each leaf page table.
1791 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1792 unsigned long size, pte_fn_t fn, void *data)
1796 unsigned long start = addr, end = addr + size;
1799 BUG_ON(addr >= end);
1800 mmu_notifier_invalidate_range_start(mm, start, end);
1801 pgd = pgd_offset(mm, addr);
1803 next = pgd_addr_end(addr, end);
1804 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1807 } while (pgd++, addr = next, addr != end);
1808 mmu_notifier_invalidate_range_end(mm, start, end);
1811 EXPORT_SYMBOL_GPL(apply_to_page_range);
1814 * handle_pte_fault chooses page fault handler according to an entry
1815 * which was read non-atomically. Before making any commitment, on
1816 * those architectures or configurations (e.g. i386 with PAE) which
1817 * might give a mix of unmatched parts, do_swap_page and do_file_page
1818 * must check under lock before unmapping the pte and proceeding
1819 * (but do_wp_page is only called after already making such a check;
1820 * and do_anonymous_page and do_no_page can safely check later on).
1822 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1823 pte_t *page_table, pte_t orig_pte)
1826 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1827 if (sizeof(pte_t) > sizeof(unsigned long)) {
1828 spinlock_t *ptl = pte_lockptr(mm, pmd);
1830 same = pte_same(*page_table, orig_pte);
1834 pte_unmap(page_table);
1839 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1840 * servicing faults for write access. In the normal case, do always want
1841 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1842 * that do not have writing enabled, when used by access_process_vm.
1844 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1846 if (likely(vma->vm_flags & VM_WRITE))
1847 pte = pte_mkwrite(pte);
1851 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1854 * If the source page was a PFN mapping, we don't have
1855 * a "struct page" for it. We do a best-effort copy by
1856 * just copying from the original user address. If that
1857 * fails, we just zero-fill it. Live with it.
1859 if (unlikely(!src)) {
1860 void *kaddr = kmap_atomic(dst, KM_USER0);
1861 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1864 * This really shouldn't fail, because the page is there
1865 * in the page tables. But it might just be unreadable,
1866 * in which case we just give up and fill the result with
1869 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1870 memset(kaddr, 0, PAGE_SIZE);
1871 kunmap_atomic(kaddr, KM_USER0);
1872 flush_dcache_page(dst);
1874 copy_user_highpage(dst, src, va, vma);
1878 * This routine handles present pages, when users try to write
1879 * to a shared page. It is done by copying the page to a new address
1880 * and decrementing the shared-page counter for the old page.
1882 * Note that this routine assumes that the protection checks have been
1883 * done by the caller (the low-level page fault routine in most cases).
1884 * Thus we can safely just mark it writable once we've done any necessary
1887 * We also mark the page dirty at this point even though the page will
1888 * change only once the write actually happens. This avoids a few races,
1889 * and potentially makes it more efficient.
1891 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1892 * but allow concurrent faults), with pte both mapped and locked.
1893 * We return with mmap_sem still held, but pte unmapped and unlocked.
1895 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1896 unsigned long address, pte_t *page_table, pmd_t *pmd,
1897 spinlock_t *ptl, pte_t orig_pte)
1899 struct page *old_page, *new_page;
1901 int reuse = 0, ret = 0;
1902 int page_mkwrite = 0;
1903 struct page *dirty_page = NULL;
1905 old_page = vm_normal_page(vma, address, orig_pte);
1908 * VM_MIXEDMAP !pfn_valid() case
1910 * We should not cow pages in a shared writeable mapping.
1911 * Just mark the pages writable as we can't do any dirty
1912 * accounting on raw pfn maps.
1914 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1915 (VM_WRITE|VM_SHARED))
1921 * Take out anonymous pages first, anonymous shared vmas are
1922 * not dirty accountable.
1924 if (PageAnon(old_page)) {
1925 if (!trylock_page(old_page)) {
1926 page_cache_get(old_page);
1927 pte_unmap_unlock(page_table, ptl);
1928 lock_page(old_page);
1929 page_table = pte_offset_map_lock(mm, pmd, address,
1931 if (!pte_same(*page_table, orig_pte)) {
1932 unlock_page(old_page);
1933 page_cache_release(old_page);
1936 page_cache_release(old_page);
1938 reuse = reuse_swap_page(old_page);
1939 unlock_page(old_page);
1940 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1941 (VM_WRITE|VM_SHARED))) {
1943 * Only catch write-faults on shared writable pages,
1944 * read-only shared pages can get COWed by
1945 * get_user_pages(.write=1, .force=1).
1947 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1948 struct vm_fault vmf;
1951 vmf.virtual_address = (void __user *)(address &
1953 vmf.pgoff = old_page->index;
1954 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1955 vmf.page = old_page;
1958 * Notify the address space that the page is about to
1959 * become writable so that it can prohibit this or wait
1960 * for the page to get into an appropriate state.
1962 * We do this without the lock held, so that it can
1963 * sleep if it needs to.
1965 page_cache_get(old_page);
1966 pte_unmap_unlock(page_table, ptl);
1968 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
1970 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
1972 goto unwritable_page;
1974 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
1975 lock_page(old_page);
1976 if (!old_page->mapping) {
1977 ret = 0; /* retry the fault */
1978 unlock_page(old_page);
1979 goto unwritable_page;
1982 VM_BUG_ON(!PageLocked(old_page));
1985 * Since we dropped the lock we need to revalidate
1986 * the PTE as someone else may have changed it. If
1987 * they did, we just return, as we can count on the
1988 * MMU to tell us if they didn't also make it writable.
1990 page_table = pte_offset_map_lock(mm, pmd, address,
1992 if (!pte_same(*page_table, orig_pte)) {
1993 unlock_page(old_page);
1994 page_cache_release(old_page);
2000 dirty_page = old_page;
2001 get_page(dirty_page);
2007 flush_cache_page(vma, address, pte_pfn(orig_pte));
2008 entry = pte_mkyoung(orig_pte);
2009 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2010 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2011 update_mmu_cache(vma, address, entry);
2012 ret |= VM_FAULT_WRITE;
2017 * Ok, we need to copy. Oh, well..
2019 page_cache_get(old_page);
2021 pte_unmap_unlock(page_table, ptl);
2023 if (unlikely(anon_vma_prepare(vma)))
2025 VM_BUG_ON(old_page == ZERO_PAGE(0));
2026 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2030 * Don't let another task, with possibly unlocked vma,
2031 * keep the mlocked page.
2033 if ((vma->vm_flags & VM_LOCKED) && old_page) {
2034 lock_page(old_page); /* for LRU manipulation */
2035 clear_page_mlock(old_page);
2036 unlock_page(old_page);
2038 cow_user_page(new_page, old_page, address, vma);
2039 __SetPageUptodate(new_page);
2041 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2045 * Re-check the pte - we dropped the lock
2047 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2048 if (likely(pte_same(*page_table, orig_pte))) {
2050 if (!PageAnon(old_page)) {
2051 dec_mm_counter(mm, file_rss);
2052 inc_mm_counter(mm, anon_rss);
2055 inc_mm_counter(mm, anon_rss);
2056 flush_cache_page(vma, address, pte_pfn(orig_pte));
2057 entry = mk_pte(new_page, vma->vm_page_prot);
2058 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2060 * Clear the pte entry and flush it first, before updating the
2061 * pte with the new entry. This will avoid a race condition
2062 * seen in the presence of one thread doing SMC and another
2065 ptep_clear_flush_notify(vma, address, page_table);
2066 page_add_new_anon_rmap(new_page, vma, address);
2067 set_pte_at(mm, address, page_table, entry);
2068 update_mmu_cache(vma, address, entry);
2071 * Only after switching the pte to the new page may
2072 * we remove the mapcount here. Otherwise another
2073 * process may come and find the rmap count decremented
2074 * before the pte is switched to the new page, and
2075 * "reuse" the old page writing into it while our pte
2076 * here still points into it and can be read by other
2079 * The critical issue is to order this
2080 * page_remove_rmap with the ptp_clear_flush above.
2081 * Those stores are ordered by (if nothing else,)
2082 * the barrier present in the atomic_add_negative
2083 * in page_remove_rmap.
2085 * Then the TLB flush in ptep_clear_flush ensures that
2086 * no process can access the old page before the
2087 * decremented mapcount is visible. And the old page
2088 * cannot be reused until after the decremented
2089 * mapcount is visible. So transitively, TLBs to
2090 * old page will be flushed before it can be reused.
2092 page_remove_rmap(old_page);
2095 /* Free the old page.. */
2096 new_page = old_page;
2097 ret |= VM_FAULT_WRITE;
2099 mem_cgroup_uncharge_page(new_page);
2102 page_cache_release(new_page);
2104 page_cache_release(old_page);
2106 pte_unmap_unlock(page_table, ptl);
2109 * Yes, Virginia, this is actually required to prevent a race
2110 * with clear_page_dirty_for_io() from clearing the page dirty
2111 * bit after it clear all dirty ptes, but before a racing
2112 * do_wp_page installs a dirty pte.
2114 * do_no_page is protected similarly.
2116 if (!page_mkwrite) {
2117 wait_on_page_locked(dirty_page);
2118 set_page_dirty_balance(dirty_page, page_mkwrite);
2120 put_page(dirty_page);
2122 struct address_space *mapping = dirty_page->mapping;
2124 set_page_dirty(dirty_page);
2125 unlock_page(dirty_page);
2126 page_cache_release(dirty_page);
2129 * Some device drivers do not set page.mapping
2130 * but still dirty their pages
2132 balance_dirty_pages_ratelimited(mapping);
2136 /* file_update_time outside page_lock */
2138 file_update_time(vma->vm_file);
2142 page_cache_release(new_page);
2146 unlock_page(old_page);
2147 page_cache_release(old_page);
2149 page_cache_release(old_page);
2151 return VM_FAULT_OOM;
2154 page_cache_release(old_page);
2159 * Helper functions for unmap_mapping_range().
2161 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
2163 * We have to restart searching the prio_tree whenever we drop the lock,
2164 * since the iterator is only valid while the lock is held, and anyway
2165 * a later vma might be split and reinserted earlier while lock dropped.
2167 * The list of nonlinear vmas could be handled more efficiently, using
2168 * a placeholder, but handle it in the same way until a need is shown.
2169 * It is important to search the prio_tree before nonlinear list: a vma
2170 * may become nonlinear and be shifted from prio_tree to nonlinear list
2171 * while the lock is dropped; but never shifted from list to prio_tree.
2173 * In order to make forward progress despite restarting the search,
2174 * vm_truncate_count is used to mark a vma as now dealt with, so we can
2175 * quickly skip it next time around. Since the prio_tree search only
2176 * shows us those vmas affected by unmapping the range in question, we
2177 * can't efficiently keep all vmas in step with mapping->truncate_count:
2178 * so instead reset them all whenever it wraps back to 0 (then go to 1).
2179 * mapping->truncate_count and vma->vm_truncate_count are protected by
2182 * In order to make forward progress despite repeatedly restarting some
2183 * large vma, note the restart_addr from unmap_vmas when it breaks out:
2184 * and restart from that address when we reach that vma again. It might
2185 * have been split or merged, shrunk or extended, but never shifted: so
2186 * restart_addr remains valid so long as it remains in the vma's range.
2187 * unmap_mapping_range forces truncate_count to leap over page-aligned
2188 * values so we can save vma's restart_addr in its truncate_count field.
2190 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
2192 static void reset_vma_truncate_counts(struct address_space *mapping)
2194 struct vm_area_struct *vma;
2195 struct prio_tree_iter iter;
2197 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
2198 vma->vm_truncate_count = 0;
2199 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
2200 vma->vm_truncate_count = 0;
2203 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
2204 unsigned long start_addr, unsigned long end_addr,
2205 struct zap_details *details)
2207 unsigned long restart_addr;
2211 * files that support invalidating or truncating portions of the
2212 * file from under mmaped areas must have their ->fault function
2213 * return a locked page (and set VM_FAULT_LOCKED in the return).
2214 * This provides synchronisation against concurrent unmapping here.
2218 restart_addr = vma->vm_truncate_count;
2219 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
2220 start_addr = restart_addr;
2221 if (start_addr >= end_addr) {
2222 /* Top of vma has been split off since last time */
2223 vma->vm_truncate_count = details->truncate_count;
2228 restart_addr = zap_page_range(vma, start_addr,
2229 end_addr - start_addr, details);
2230 need_break = need_resched() || spin_needbreak(details->i_mmap_lock);
2232 if (restart_addr >= end_addr) {
2233 /* We have now completed this vma: mark it so */
2234 vma->vm_truncate_count = details->truncate_count;
2238 /* Note restart_addr in vma's truncate_count field */
2239 vma->vm_truncate_count = restart_addr;
2244 spin_unlock(details->i_mmap_lock);
2246 spin_lock(details->i_mmap_lock);
2250 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2251 struct zap_details *details)
2253 struct vm_area_struct *vma;
2254 struct prio_tree_iter iter;
2255 pgoff_t vba, vea, zba, zea;
2258 vma_prio_tree_foreach(vma, &iter, root,
2259 details->first_index, details->last_index) {
2260 /* Skip quickly over those we have already dealt with */
2261 if (vma->vm_truncate_count == details->truncate_count)
2264 vba = vma->vm_pgoff;
2265 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2266 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2267 zba = details->first_index;
2270 zea = details->last_index;
2274 if (unmap_mapping_range_vma(vma,
2275 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2276 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2282 static inline void unmap_mapping_range_list(struct list_head *head,
2283 struct zap_details *details)
2285 struct vm_area_struct *vma;
2288 * In nonlinear VMAs there is no correspondence between virtual address
2289 * offset and file offset. So we must perform an exhaustive search
2290 * across *all* the pages in each nonlinear VMA, not just the pages
2291 * whose virtual address lies outside the file truncation point.
2294 list_for_each_entry(vma, head, shared.vm_set.list) {
2295 /* Skip quickly over those we have already dealt with */
2296 if (vma->vm_truncate_count == details->truncate_count)
2298 details->nonlinear_vma = vma;
2299 if (unmap_mapping_range_vma(vma, vma->vm_start,
2300 vma->vm_end, details) < 0)
2306 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2307 * @mapping: the address space containing mmaps to be unmapped.
2308 * @holebegin: byte in first page to unmap, relative to the start of
2309 * the underlying file. This will be rounded down to a PAGE_SIZE
2310 * boundary. Note that this is different from vmtruncate(), which
2311 * must keep the partial page. In contrast, we must get rid of
2313 * @holelen: size of prospective hole in bytes. This will be rounded
2314 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2316 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2317 * but 0 when invalidating pagecache, don't throw away private data.
2319 void unmap_mapping_range(struct address_space *mapping,
2320 loff_t const holebegin, loff_t const holelen, int even_cows)
2322 struct zap_details details;
2323 pgoff_t hba = holebegin >> PAGE_SHIFT;
2324 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2326 /* Check for overflow. */
2327 if (sizeof(holelen) > sizeof(hlen)) {
2329 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2330 if (holeend & ~(long long)ULONG_MAX)
2331 hlen = ULONG_MAX - hba + 1;
2334 details.check_mapping = even_cows? NULL: mapping;
2335 details.nonlinear_vma = NULL;
2336 details.first_index = hba;
2337 details.last_index = hba + hlen - 1;
2338 if (details.last_index < details.first_index)
2339 details.last_index = ULONG_MAX;
2340 details.i_mmap_lock = &mapping->i_mmap_lock;
2342 spin_lock(&mapping->i_mmap_lock);
2344 /* Protect against endless unmapping loops */
2345 mapping->truncate_count++;
2346 if (unlikely(is_restart_addr(mapping->truncate_count))) {
2347 if (mapping->truncate_count == 0)
2348 reset_vma_truncate_counts(mapping);
2349 mapping->truncate_count++;
2351 details.truncate_count = mapping->truncate_count;
2353 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2354 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2355 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2356 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2357 spin_unlock(&mapping->i_mmap_lock);
2359 EXPORT_SYMBOL(unmap_mapping_range);
2362 * vmtruncate - unmap mappings "freed" by truncate() syscall
2363 * @inode: inode of the file used
2364 * @offset: file offset to start truncating
2366 * NOTE! We have to be ready to update the memory sharing
2367 * between the file and the memory map for a potential last
2368 * incomplete page. Ugly, but necessary.
2370 int vmtruncate(struct inode * inode, loff_t offset)
2372 if (inode->i_size < offset) {
2373 unsigned long limit;
2375 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2376 if (limit != RLIM_INFINITY && offset > limit)
2378 if (offset > inode->i_sb->s_maxbytes)
2380 i_size_write(inode, offset);
2382 struct address_space *mapping = inode->i_mapping;
2385 * truncation of in-use swapfiles is disallowed - it would
2386 * cause subsequent swapout to scribble on the now-freed
2389 if (IS_SWAPFILE(inode))
2391 i_size_write(inode, offset);
2394 * unmap_mapping_range is called twice, first simply for
2395 * efficiency so that truncate_inode_pages does fewer
2396 * single-page unmaps. However after this first call, and
2397 * before truncate_inode_pages finishes, it is possible for
2398 * private pages to be COWed, which remain after
2399 * truncate_inode_pages finishes, hence the second
2400 * unmap_mapping_range call must be made for correctness.
2402 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2403 truncate_inode_pages(mapping, offset);
2404 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2407 if (inode->i_op->truncate)
2408 inode->i_op->truncate(inode);
2412 send_sig(SIGXFSZ, current, 0);
2416 EXPORT_SYMBOL(vmtruncate);
2418 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2420 struct address_space *mapping = inode->i_mapping;
2423 * If the underlying filesystem is not going to provide
2424 * a way to truncate a range of blocks (punch a hole) -
2425 * we should return failure right now.
2427 if (!inode->i_op->truncate_range)
2430 mutex_lock(&inode->i_mutex);
2431 down_write(&inode->i_alloc_sem);
2432 unmap_mapping_range(mapping, offset, (end - offset), 1);
2433 truncate_inode_pages_range(mapping, offset, end);
2434 unmap_mapping_range(mapping, offset, (end - offset), 1);
2435 inode->i_op->truncate_range(inode, offset, end);
2436 up_write(&inode->i_alloc_sem);
2437 mutex_unlock(&inode->i_mutex);
2443 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2444 * but allow concurrent faults), and pte mapped but not yet locked.
2445 * We return with mmap_sem still held, but pte unmapped and unlocked.
2447 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2448 unsigned long address, pte_t *page_table, pmd_t *pmd,
2449 int write_access, pte_t orig_pte)
2455 struct mem_cgroup *ptr = NULL;
2458 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2461 entry = pte_to_swp_entry(orig_pte);
2462 if (is_migration_entry(entry)) {
2463 migration_entry_wait(mm, pmd, address);
2466 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2467 page = lookup_swap_cache(entry);
2469 grab_swap_token(); /* Contend for token _before_ read-in */
2470 page = swapin_readahead(entry,
2471 GFP_HIGHUSER_MOVABLE, vma, address);
2474 * Back out if somebody else faulted in this pte
2475 * while we released the pte lock.
2477 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2478 if (likely(pte_same(*page_table, orig_pte)))
2480 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2484 /* Had to read the page from swap area: Major fault */
2485 ret = VM_FAULT_MAJOR;
2486 count_vm_event(PGMAJFAULT);
2490 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2492 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2498 * Back out if somebody else already faulted in this pte.
2500 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2501 if (unlikely(!pte_same(*page_table, orig_pte)))
2504 if (unlikely(!PageUptodate(page))) {
2505 ret = VM_FAULT_SIGBUS;
2510 * The page isn't present yet, go ahead with the fault.
2512 * Be careful about the sequence of operations here.
2513 * To get its accounting right, reuse_swap_page() must be called
2514 * while the page is counted on swap but not yet in mapcount i.e.
2515 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2516 * must be called after the swap_free(), or it will never succeed.
2517 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2518 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2519 * in page->private. In this case, a record in swap_cgroup is silently
2520 * discarded at swap_free().
2523 inc_mm_counter(mm, anon_rss);
2524 pte = mk_pte(page, vma->vm_page_prot);
2525 if (write_access && reuse_swap_page(page)) {
2526 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2529 flush_icache_page(vma, page);
2530 set_pte_at(mm, address, page_table, pte);
2531 page_add_anon_rmap(page, vma, address);
2532 /* It's better to call commit-charge after rmap is established */
2533 mem_cgroup_commit_charge_swapin(page, ptr);
2536 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2537 try_to_free_swap(page);
2541 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2542 if (ret & VM_FAULT_ERROR)
2543 ret &= VM_FAULT_ERROR;
2547 /* No need to invalidate - it was non-present before */
2548 update_mmu_cache(vma, address, pte);
2550 pte_unmap_unlock(page_table, ptl);
2554 mem_cgroup_cancel_charge_swapin(ptr);
2555 pte_unmap_unlock(page_table, ptl);
2558 page_cache_release(page);
2563 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2564 * but allow concurrent faults), and pte mapped but not yet locked.
2565 * We return with mmap_sem still held, but pte unmapped and unlocked.
2567 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2568 unsigned long address, pte_t *page_table, pmd_t *pmd,
2575 /* Allocate our own private page. */
2576 pte_unmap(page_table);
2578 if (unlikely(anon_vma_prepare(vma)))
2580 page = alloc_zeroed_user_highpage_movable(vma, address);
2583 __SetPageUptodate(page);
2585 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
2588 entry = mk_pte(page, vma->vm_page_prot);
2589 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2591 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2592 if (!pte_none(*page_table))
2594 inc_mm_counter(mm, anon_rss);
2595 page_add_new_anon_rmap(page, vma, address);
2596 set_pte_at(mm, address, page_table, entry);
2598 /* No need to invalidate - it was non-present before */
2599 update_mmu_cache(vma, address, entry);
2601 pte_unmap_unlock(page_table, ptl);
2604 mem_cgroup_uncharge_page(page);
2605 page_cache_release(page);
2608 page_cache_release(page);
2610 return VM_FAULT_OOM;
2614 * __do_fault() tries to create a new page mapping. It aggressively
2615 * tries to share with existing pages, but makes a separate copy if
2616 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2617 * the next page fault.
2619 * As this is called only for pages that do not currently exist, we
2620 * do not need to flush old virtual caches or the TLB.
2622 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2623 * but allow concurrent faults), and pte neither mapped nor locked.
2624 * We return with mmap_sem still held, but pte unmapped and unlocked.
2626 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2627 unsigned long address, pmd_t *pmd,
2628 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2636 struct page *dirty_page = NULL;
2637 struct vm_fault vmf;
2639 int page_mkwrite = 0;
2641 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2646 ret = vma->vm_ops->fault(vma, &vmf);
2647 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2651 * For consistency in subsequent calls, make the faulted page always
2654 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2655 lock_page(vmf.page);
2657 VM_BUG_ON(!PageLocked(vmf.page));
2660 * Should we do an early C-O-W break?
2663 if (flags & FAULT_FLAG_WRITE) {
2664 if (!(vma->vm_flags & VM_SHARED)) {
2666 if (unlikely(anon_vma_prepare(vma))) {
2670 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
2676 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
2678 page_cache_release(page);
2683 * Don't let another task, with possibly unlocked vma,
2684 * keep the mlocked page.
2686 if (vma->vm_flags & VM_LOCKED)
2687 clear_page_mlock(vmf.page);
2688 copy_user_highpage(page, vmf.page, address, vma);
2689 __SetPageUptodate(page);
2692 * If the page will be shareable, see if the backing
2693 * address space wants to know that the page is about
2694 * to become writable
2696 if (vma->vm_ops->page_mkwrite) {
2700 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2701 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2703 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2705 goto unwritable_page;
2707 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2709 if (!page->mapping) {
2710 ret = 0; /* retry the fault */
2712 goto unwritable_page;
2715 VM_BUG_ON(!PageLocked(page));
2722 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2725 * This silly early PAGE_DIRTY setting removes a race
2726 * due to the bad i386 page protection. But it's valid
2727 * for other architectures too.
2729 * Note that if write_access is true, we either now have
2730 * an exclusive copy of the page, or this is a shared mapping,
2731 * so we can make it writable and dirty to avoid having to
2732 * handle that later.
2734 /* Only go through if we didn't race with anybody else... */
2735 if (likely(pte_same(*page_table, orig_pte))) {
2736 flush_icache_page(vma, page);
2737 entry = mk_pte(page, vma->vm_page_prot);
2738 if (flags & FAULT_FLAG_WRITE)
2739 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2741 inc_mm_counter(mm, anon_rss);
2742 page_add_new_anon_rmap(page, vma, address);
2744 inc_mm_counter(mm, file_rss);
2745 page_add_file_rmap(page);
2746 if (flags & FAULT_FLAG_WRITE) {
2748 get_page(dirty_page);
2751 set_pte_at(mm, address, page_table, entry);
2753 /* no need to invalidate: a not-present page won't be cached */
2754 update_mmu_cache(vma, address, entry);
2757 mem_cgroup_uncharge_page(page);
2759 page_cache_release(page);
2761 anon = 1; /* no anon but release faulted_page */
2764 pte_unmap_unlock(page_table, ptl);
2768 struct address_space *mapping = page->mapping;
2770 if (set_page_dirty(dirty_page))
2772 unlock_page(dirty_page);
2773 put_page(dirty_page);
2774 if (page_mkwrite && mapping) {
2776 * Some device drivers do not set page.mapping but still
2779 balance_dirty_pages_ratelimited(mapping);
2782 /* file_update_time outside page_lock */
2784 file_update_time(vma->vm_file);
2786 unlock_page(vmf.page);
2788 page_cache_release(vmf.page);
2794 page_cache_release(page);
2798 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2799 unsigned long address, pte_t *page_table, pmd_t *pmd,
2800 int write_access, pte_t orig_pte)
2802 pgoff_t pgoff = (((address & PAGE_MASK)
2803 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
2804 unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2806 pte_unmap(page_table);
2807 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2811 * Fault of a previously existing named mapping. Repopulate the pte
2812 * from the encoded file_pte if possible. This enables swappable
2815 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2816 * but allow concurrent faults), and pte mapped but not yet locked.
2817 * We return with mmap_sem still held, but pte unmapped and unlocked.
2819 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2820 unsigned long address, pte_t *page_table, pmd_t *pmd,
2821 int write_access, pte_t orig_pte)
2823 unsigned int flags = FAULT_FLAG_NONLINEAR |
2824 (write_access ? FAULT_FLAG_WRITE : 0);
2827 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2830 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2832 * Page table corrupted: show pte and kill process.
2834 print_bad_pte(vma, address, orig_pte, NULL);
2835 return VM_FAULT_OOM;
2838 pgoff = pte_to_pgoff(orig_pte);
2839 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
2843 * These routines also need to handle stuff like marking pages dirty
2844 * and/or accessed for architectures that don't do it in hardware (most
2845 * RISC architectures). The early dirtying is also good on the i386.
2847 * There is also a hook called "update_mmu_cache()" that architectures
2848 * with external mmu caches can use to update those (ie the Sparc or
2849 * PowerPC hashed page tables that act as extended TLBs).
2851 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2852 * but allow concurrent faults), and pte mapped but not yet locked.
2853 * We return with mmap_sem still held, but pte unmapped and unlocked.
2855 static inline int handle_pte_fault(struct mm_struct *mm,
2856 struct vm_area_struct *vma, unsigned long address,
2857 pte_t *pte, pmd_t *pmd, int write_access)
2863 if (!pte_present(entry)) {
2864 if (pte_none(entry)) {
2866 if (likely(vma->vm_ops->fault))
2867 return do_linear_fault(mm, vma, address,
2868 pte, pmd, write_access, entry);
2870 return do_anonymous_page(mm, vma, address,
2871 pte, pmd, write_access);
2873 if (pte_file(entry))
2874 return do_nonlinear_fault(mm, vma, address,
2875 pte, pmd, write_access, entry);
2876 return do_swap_page(mm, vma, address,
2877 pte, pmd, write_access, entry);
2880 ptl = pte_lockptr(mm, pmd);
2882 if (unlikely(!pte_same(*pte, entry)))
2885 if (!pte_write(entry))
2886 return do_wp_page(mm, vma, address,
2887 pte, pmd, ptl, entry);
2888 entry = pte_mkdirty(entry);
2890 entry = pte_mkyoung(entry);
2891 if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2892 update_mmu_cache(vma, address, entry);
2895 * This is needed only for protection faults but the arch code
2896 * is not yet telling us if this is a protection fault or not.
2897 * This still avoids useless tlb flushes for .text page faults
2901 flush_tlb_page(vma, address);
2904 pte_unmap_unlock(pte, ptl);
2909 * By the time we get here, we already hold the mm semaphore
2911 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2912 unsigned long address, int write_access)
2919 __set_current_state(TASK_RUNNING);
2921 count_vm_event(PGFAULT);
2923 if (unlikely(is_vm_hugetlb_page(vma)))
2924 return hugetlb_fault(mm, vma, address, write_access);
2926 pgd = pgd_offset(mm, address);
2927 pud = pud_alloc(mm, pgd, address);
2929 return VM_FAULT_OOM;
2930 pmd = pmd_alloc(mm, pud, address);
2932 return VM_FAULT_OOM;
2933 pte = pte_alloc_map(mm, pmd, address);
2935 return VM_FAULT_OOM;
2937 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2940 #ifndef __PAGETABLE_PUD_FOLDED
2942 * Allocate page upper directory.
2943 * We've already handled the fast-path in-line.
2945 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2947 pud_t *new = pud_alloc_one(mm, address);
2951 smp_wmb(); /* See comment in __pte_alloc */
2953 spin_lock(&mm->page_table_lock);
2954 if (pgd_present(*pgd)) /* Another has populated it */
2957 pgd_populate(mm, pgd, new);
2958 spin_unlock(&mm->page_table_lock);
2961 #endif /* __PAGETABLE_PUD_FOLDED */
2963 #ifndef __PAGETABLE_PMD_FOLDED
2965 * Allocate page middle directory.
2966 * We've already handled the fast-path in-line.
2968 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2970 pmd_t *new = pmd_alloc_one(mm, address);
2974 smp_wmb(); /* See comment in __pte_alloc */
2976 spin_lock(&mm->page_table_lock);
2977 #ifndef __ARCH_HAS_4LEVEL_HACK
2978 if (pud_present(*pud)) /* Another has populated it */
2981 pud_populate(mm, pud, new);
2983 if (pgd_present(*pud)) /* Another has populated it */
2986 pgd_populate(mm, pud, new);
2987 #endif /* __ARCH_HAS_4LEVEL_HACK */
2988 spin_unlock(&mm->page_table_lock);
2991 #endif /* __PAGETABLE_PMD_FOLDED */
2993 int make_pages_present(unsigned long addr, unsigned long end)
2995 int ret, len, write;
2996 struct vm_area_struct * vma;
2998 vma = find_vma(current->mm, addr);
3001 write = (vma->vm_flags & VM_WRITE) != 0;
3002 BUG_ON(addr >= end);
3003 BUG_ON(end > vma->vm_end);
3004 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3005 ret = get_user_pages(current, current->mm, addr,
3006 len, write, 0, NULL, NULL);
3009 return ret == len ? 0 : -EFAULT;
3012 #if !defined(__HAVE_ARCH_GATE_AREA)
3014 #if defined(AT_SYSINFO_EHDR)
3015 static struct vm_area_struct gate_vma;
3017 static int __init gate_vma_init(void)
3019 gate_vma.vm_mm = NULL;
3020 gate_vma.vm_start = FIXADDR_USER_START;
3021 gate_vma.vm_end = FIXADDR_USER_END;
3022 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3023 gate_vma.vm_page_prot = __P101;
3025 * Make sure the vDSO gets into every core dump.
3026 * Dumping its contents makes post-mortem fully interpretable later
3027 * without matching up the same kernel and hardware config to see
3028 * what PC values meant.
3030 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3033 __initcall(gate_vma_init);
3036 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
3038 #ifdef AT_SYSINFO_EHDR
3045 int in_gate_area_no_task(unsigned long addr)
3047 #ifdef AT_SYSINFO_EHDR
3048 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3054 #endif /* __HAVE_ARCH_GATE_AREA */
3056 #ifdef CONFIG_HAVE_IOREMAP_PROT
3057 int follow_phys(struct vm_area_struct *vma,
3058 unsigned long address, unsigned int flags,
3059 unsigned long *prot, resource_size_t *phys)
3066 resource_size_t phys_addr = 0;
3067 struct mm_struct *mm = vma->vm_mm;
3070 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3073 pgd = pgd_offset(mm, address);
3074 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3077 pud = pud_offset(pgd, address);
3078 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3081 pmd = pmd_offset(pud, address);
3082 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3085 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3089 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3094 if (!pte_present(pte))
3096 if ((flags & FOLL_WRITE) && !pte_write(pte))
3098 phys_addr = pte_pfn(pte);
3099 phys_addr <<= PAGE_SHIFT; /* Shift here to avoid overflow on PAE */
3101 *prot = pgprot_val(pte_pgprot(pte));
3106 pte_unmap_unlock(ptep, ptl);
3111 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3112 void *buf, int len, int write)
3114 resource_size_t phys_addr;
3115 unsigned long prot = 0;
3116 void __iomem *maddr;
3117 int offset = addr & (PAGE_SIZE-1);
3119 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3122 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3124 memcpy_toio(maddr + offset, buf, len);
3126 memcpy_fromio(buf, maddr + offset, len);
3134 * Access another process' address space.
3135 * Source/target buffer must be kernel space,
3136 * Do not walk the page table directly, use get_user_pages
3138 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
3140 struct mm_struct *mm;
3141 struct vm_area_struct *vma;
3142 void *old_buf = buf;
3144 mm = get_task_mm(tsk);
3148 down_read(&mm->mmap_sem);
3149 /* ignore errors, just check how much was successfully transferred */
3151 int bytes, ret, offset;
3153 struct page *page = NULL;
3155 ret = get_user_pages(tsk, mm, addr, 1,
3156 write, 1, &page, &vma);
3159 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3160 * we can access using slightly different code.
3162 #ifdef CONFIG_HAVE_IOREMAP_PROT
3163 vma = find_vma(mm, addr);
3166 if (vma->vm_ops && vma->vm_ops->access)
3167 ret = vma->vm_ops->access(vma, addr, buf,
3175 offset = addr & (PAGE_SIZE-1);
3176 if (bytes > PAGE_SIZE-offset)
3177 bytes = PAGE_SIZE-offset;
3181 copy_to_user_page(vma, page, addr,
3182 maddr + offset, buf, bytes);
3183 set_page_dirty_lock(page);
3185 copy_from_user_page(vma, page, addr,
3186 buf, maddr + offset, bytes);
3189 page_cache_release(page);
3195 up_read(&mm->mmap_sem);
3198 return buf - old_buf;
3202 * Print the name of a VMA.
3204 void print_vma_addr(char *prefix, unsigned long ip)
3206 struct mm_struct *mm = current->mm;
3207 struct vm_area_struct *vma;
3210 * Do not print if we are in atomic
3211 * contexts (in exception stacks, etc.):
3213 if (preempt_count())
3216 down_read(&mm->mmap_sem);
3217 vma = find_vma(mm, ip);
3218 if (vma && vma->vm_file) {
3219 struct file *f = vma->vm_file;
3220 char *buf = (char *)__get_free_page(GFP_KERNEL);
3224 p = d_path(&f->f_path, buf, PAGE_SIZE);
3227 s = strrchr(p, '/');
3230 printk("%s%s[%lx+%lx]", prefix, p,
3232 vma->vm_end - vma->vm_start);
3233 free_page((unsigned long)buf);
3236 up_read(¤t->mm->mmap_sem);
3239 #ifdef CONFIG_PROVE_LOCKING
3240 void might_fault(void)
3243 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3244 * holding the mmap_sem, this is safe because kernel memory doesn't
3245 * get paged out, therefore we'll never actually fault, and the
3246 * below annotations will generate false positives.
3248 if (segment_eq(get_fs(), KERNEL_DS))
3253 * it would be nicer only to annotate paths which are not under
3254 * pagefault_disable, however that requires a larger audit and
3255 * providing helpers like get_user_atomic.
3257 if (!in_atomic() && current->mm)
3258 might_lock_read(¤t->mm->mmap_sem);
3260 EXPORT_SYMBOL(might_fault);