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>
54 #include <asm/pgalloc.h>
55 #include <asm/uaccess.h>
57 #include <asm/tlbflush.h>
58 #include <asm/pgtable.h>
60 #include <linux/swapops.h>
61 #include <linux/elf.h>
63 #ifndef CONFIG_NEED_MULTIPLE_NODES
64 /* use the per-pgdat data instead for discontigmem - mbligh */
65 unsigned long max_mapnr;
68 EXPORT_SYMBOL(max_mapnr);
69 EXPORT_SYMBOL(mem_map);
72 unsigned long num_physpages;
74 * A number of key systems in x86 including ioremap() rely on the assumption
75 * that high_memory defines the upper bound on direct map memory, then end
76 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
77 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
82 EXPORT_SYMBOL(num_physpages);
83 EXPORT_SYMBOL(high_memory);
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_zone_page_state(page, NR_PAGETABLE);
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_zone_page_state(new, NR_PAGETABLE);
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 print_bad_pte(vma, pte, addr);
410 * NOTE! We still have PageReserved() pages in the page
413 * The PAGE_ZERO() pages and various VDSO mappings can
414 * cause them to exist.
416 return pfn_to_page(pfn);
420 * copy one vm_area from one task to the other. Assumes the page tables
421 * already present in the new task to be cleared in the whole range
422 * covered by this vma.
426 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
427 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
428 unsigned long addr, int *rss)
430 unsigned long vm_flags = vma->vm_flags;
431 pte_t pte = *src_pte;
434 /* pte contains position in swap or file, so copy. */
435 if (unlikely(!pte_present(pte))) {
436 if (!pte_file(pte)) {
437 swp_entry_t entry = pte_to_swp_entry(pte);
439 swap_duplicate(entry);
440 /* make sure dst_mm is on swapoff's mmlist. */
441 if (unlikely(list_empty(&dst_mm->mmlist))) {
442 spin_lock(&mmlist_lock);
443 if (list_empty(&dst_mm->mmlist))
444 list_add(&dst_mm->mmlist,
446 spin_unlock(&mmlist_lock);
448 if (is_write_migration_entry(entry) &&
449 is_cow_mapping(vm_flags)) {
451 * COW mappings require pages in both parent
452 * and child to be set to read.
454 make_migration_entry_read(&entry);
455 pte = swp_entry_to_pte(entry);
456 set_pte_at(src_mm, addr, src_pte, pte);
463 * If it's a COW mapping, write protect it both
464 * in the parent and the child
466 if (is_cow_mapping(vm_flags)) {
467 ptep_set_wrprotect(src_mm, addr, src_pte);
468 pte = pte_wrprotect(pte);
472 * If it's a shared mapping, mark it clean in
475 if (vm_flags & VM_SHARED)
476 pte = pte_mkclean(pte);
477 pte = pte_mkold(pte);
479 page = vm_normal_page(vma, addr, pte);
482 page_dup_rmap(page, vma, addr);
483 rss[!!PageAnon(page)]++;
487 set_pte_at(dst_mm, addr, dst_pte, pte);
490 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
491 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
492 unsigned long addr, unsigned long end)
494 pte_t *src_pte, *dst_pte;
495 spinlock_t *src_ptl, *dst_ptl;
501 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
504 src_pte = pte_offset_map_nested(src_pmd, addr);
505 src_ptl = pte_lockptr(src_mm, src_pmd);
506 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
507 arch_enter_lazy_mmu_mode();
511 * We are holding two locks at this point - either of them
512 * could generate latencies in another task on another CPU.
514 if (progress >= 32) {
516 if (need_resched() ||
517 need_lockbreak(src_ptl) ||
518 need_lockbreak(dst_ptl))
521 if (pte_none(*src_pte)) {
525 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
527 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
529 arch_leave_lazy_mmu_mode();
530 spin_unlock(src_ptl);
531 pte_unmap_nested(src_pte - 1);
532 add_mm_rss(dst_mm, rss[0], rss[1]);
533 pte_unmap_unlock(dst_pte - 1, dst_ptl);
540 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
541 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
542 unsigned long addr, unsigned long end)
544 pmd_t *src_pmd, *dst_pmd;
547 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
550 src_pmd = pmd_offset(src_pud, addr);
552 next = pmd_addr_end(addr, end);
553 if (pmd_none_or_clear_bad(src_pmd))
555 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
558 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
562 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
563 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
564 unsigned long addr, unsigned long end)
566 pud_t *src_pud, *dst_pud;
569 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
572 src_pud = pud_offset(src_pgd, addr);
574 next = pud_addr_end(addr, end);
575 if (pud_none_or_clear_bad(src_pud))
577 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
580 } while (dst_pud++, src_pud++, addr = next, addr != end);
584 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
585 struct vm_area_struct *vma)
587 pgd_t *src_pgd, *dst_pgd;
589 unsigned long addr = vma->vm_start;
590 unsigned long end = vma->vm_end;
593 * Don't copy ptes where a page fault will fill them correctly.
594 * Fork becomes much lighter when there are big shared or private
595 * readonly mappings. The tradeoff is that copy_page_range is more
596 * efficient than faulting.
598 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
603 if (is_vm_hugetlb_page(vma))
604 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
606 dst_pgd = pgd_offset(dst_mm, addr);
607 src_pgd = pgd_offset(src_mm, addr);
609 next = pgd_addr_end(addr, end);
610 if (pgd_none_or_clear_bad(src_pgd))
612 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
615 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
619 static unsigned long zap_pte_range(struct mmu_gather *tlb,
620 struct vm_area_struct *vma, pmd_t *pmd,
621 unsigned long addr, unsigned long end,
622 long *zap_work, struct zap_details *details)
624 struct mm_struct *mm = tlb->mm;
630 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
631 arch_enter_lazy_mmu_mode();
634 if (pte_none(ptent)) {
639 (*zap_work) -= PAGE_SIZE;
641 if (pte_present(ptent)) {
644 page = vm_normal_page(vma, addr, ptent);
645 if (unlikely(details) && page) {
647 * unmap_shared_mapping_pages() wants to
648 * invalidate cache without truncating:
649 * unmap shared but keep private pages.
651 if (details->check_mapping &&
652 details->check_mapping != page->mapping)
655 * Each page->index must be checked when
656 * invalidating or truncating nonlinear.
658 if (details->nonlinear_vma &&
659 (page->index < details->first_index ||
660 page->index > details->last_index))
663 ptent = ptep_get_and_clear_full(mm, addr, pte,
665 tlb_remove_tlb_entry(tlb, pte, addr);
668 if (unlikely(details) && details->nonlinear_vma
669 && linear_page_index(details->nonlinear_vma,
670 addr) != page->index)
671 set_pte_at(mm, addr, pte,
672 pgoff_to_pte(page->index));
676 if (pte_dirty(ptent))
677 set_page_dirty(page);
678 if (pte_young(ptent))
679 SetPageReferenced(page);
682 page_remove_rmap(page, vma);
683 tlb_remove_page(tlb, page);
687 * If details->check_mapping, we leave swap entries;
688 * if details->nonlinear_vma, we leave file entries.
690 if (unlikely(details))
692 if (!pte_file(ptent))
693 free_swap_and_cache(pte_to_swp_entry(ptent));
694 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
695 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
697 add_mm_rss(mm, file_rss, anon_rss);
698 arch_leave_lazy_mmu_mode();
699 pte_unmap_unlock(pte - 1, ptl);
704 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
705 struct vm_area_struct *vma, pud_t *pud,
706 unsigned long addr, unsigned long end,
707 long *zap_work, struct zap_details *details)
712 pmd = pmd_offset(pud, addr);
714 next = pmd_addr_end(addr, end);
715 if (pmd_none_or_clear_bad(pmd)) {
719 next = zap_pte_range(tlb, vma, pmd, addr, next,
721 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
726 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
727 struct vm_area_struct *vma, pgd_t *pgd,
728 unsigned long addr, unsigned long end,
729 long *zap_work, struct zap_details *details)
734 pud = pud_offset(pgd, addr);
736 next = pud_addr_end(addr, end);
737 if (pud_none_or_clear_bad(pud)) {
741 next = zap_pmd_range(tlb, vma, pud, addr, next,
743 } while (pud++, addr = next, (addr != end && *zap_work > 0));
748 static unsigned long unmap_page_range(struct mmu_gather *tlb,
749 struct vm_area_struct *vma,
750 unsigned long addr, unsigned long end,
751 long *zap_work, struct zap_details *details)
756 if (details && !details->check_mapping && !details->nonlinear_vma)
760 tlb_start_vma(tlb, vma);
761 pgd = pgd_offset(vma->vm_mm, addr);
763 next = pgd_addr_end(addr, end);
764 if (pgd_none_or_clear_bad(pgd)) {
768 next = zap_pud_range(tlb, vma, pgd, addr, next,
770 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
771 tlb_end_vma(tlb, vma);
776 #ifdef CONFIG_PREEMPT
777 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
779 /* No preempt: go for improved straight-line efficiency */
780 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
784 * unmap_vmas - unmap a range of memory covered by a list of vma's
785 * @tlbp: address of the caller's struct mmu_gather
786 * @vma: the starting vma
787 * @start_addr: virtual address at which to start unmapping
788 * @end_addr: virtual address at which to end unmapping
789 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
790 * @details: details of nonlinear truncation or shared cache invalidation
792 * Returns the end address of the unmapping (restart addr if interrupted).
794 * Unmap all pages in the vma list.
796 * We aim to not hold locks for too long (for scheduling latency reasons).
797 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
798 * return the ending mmu_gather to the caller.
800 * Only addresses between `start' and `end' will be unmapped.
802 * The VMA list must be sorted in ascending virtual address order.
804 * unmap_vmas() assumes that the caller will flush the whole unmapped address
805 * range after unmap_vmas() returns. So the only responsibility here is to
806 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
807 * drops the lock and schedules.
809 unsigned long unmap_vmas(struct mmu_gather **tlbp,
810 struct vm_area_struct *vma, unsigned long start_addr,
811 unsigned long end_addr, unsigned long *nr_accounted,
812 struct zap_details *details)
814 long zap_work = ZAP_BLOCK_SIZE;
815 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
816 int tlb_start_valid = 0;
817 unsigned long start = start_addr;
818 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
819 int fullmm = (*tlbp)->fullmm;
821 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
824 start = max(vma->vm_start, start_addr);
825 if (start >= vma->vm_end)
827 end = min(vma->vm_end, end_addr);
828 if (end <= vma->vm_start)
831 if (vma->vm_flags & VM_ACCOUNT)
832 *nr_accounted += (end - start) >> PAGE_SHIFT;
834 while (start != end) {
835 if (!tlb_start_valid) {
840 if (unlikely(is_vm_hugetlb_page(vma))) {
841 unmap_hugepage_range(vma, start, end);
842 zap_work -= (end - start) /
843 (HPAGE_SIZE / PAGE_SIZE);
846 start = unmap_page_range(*tlbp, vma,
847 start, end, &zap_work, details);
850 BUG_ON(start != end);
854 tlb_finish_mmu(*tlbp, tlb_start, start);
856 if (need_resched() ||
857 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
865 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
867 zap_work = ZAP_BLOCK_SIZE;
871 return start; /* which is now the end (or restart) address */
875 * zap_page_range - remove user pages in a given range
876 * @vma: vm_area_struct holding the applicable pages
877 * @address: starting address of pages to zap
878 * @size: number of bytes to zap
879 * @details: details of nonlinear truncation or shared cache invalidation
881 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
882 unsigned long size, struct zap_details *details)
884 struct mm_struct *mm = vma->vm_mm;
885 struct mmu_gather *tlb;
886 unsigned long end = address + size;
887 unsigned long nr_accounted = 0;
890 tlb = tlb_gather_mmu(mm, 0);
891 update_hiwater_rss(mm);
892 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
894 tlb_finish_mmu(tlb, address, end);
899 * Do a quick page-table lookup for a single page.
901 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
910 struct mm_struct *mm = vma->vm_mm;
912 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
914 BUG_ON(flags & FOLL_GET);
919 pgd = pgd_offset(mm, address);
920 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
923 pud = pud_offset(pgd, address);
924 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
927 pmd = pmd_offset(pud, address);
928 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
931 if (pmd_huge(*pmd)) {
932 BUG_ON(flags & FOLL_GET);
933 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
937 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
942 if (!pte_present(pte))
944 if ((flags & FOLL_WRITE) && !pte_write(pte))
946 page = vm_normal_page(vma, address, pte);
950 if (flags & FOLL_GET)
952 if (flags & FOLL_TOUCH) {
953 if ((flags & FOLL_WRITE) &&
954 !pte_dirty(pte) && !PageDirty(page))
955 set_page_dirty(page);
956 mark_page_accessed(page);
959 pte_unmap_unlock(ptep, ptl);
965 * When core dumping an enormous anonymous area that nobody
966 * has touched so far, we don't want to allocate page tables.
968 if (flags & FOLL_ANON) {
969 page = ZERO_PAGE(address);
970 if (flags & FOLL_GET)
972 BUG_ON(flags & FOLL_WRITE);
977 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
978 unsigned long start, int len, int write, int force,
979 struct page **pages, struct vm_area_struct **vmas)
982 unsigned int vm_flags;
985 * Require read or write permissions.
986 * If 'force' is set, we only require the "MAY" flags.
988 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
989 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
993 struct vm_area_struct *vma;
994 unsigned int foll_flags;
996 vma = find_extend_vma(mm, start);
997 if (!vma && in_gate_area(tsk, start)) {
998 unsigned long pg = start & PAGE_MASK;
999 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
1004 if (write) /* user gate pages are read-only */
1005 return i ? : -EFAULT;
1007 pgd = pgd_offset_k(pg);
1009 pgd = pgd_offset_gate(mm, pg);
1010 BUG_ON(pgd_none(*pgd));
1011 pud = pud_offset(pgd, pg);
1012 BUG_ON(pud_none(*pud));
1013 pmd = pmd_offset(pud, pg);
1015 return i ? : -EFAULT;
1016 pte = pte_offset_map(pmd, pg);
1017 if (pte_none(*pte)) {
1019 return i ? : -EFAULT;
1022 struct page *page = vm_normal_page(gate_vma, start, *pte);
1036 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1037 || !(vm_flags & vma->vm_flags))
1038 return i ? : -EFAULT;
1040 if (is_vm_hugetlb_page(vma)) {
1041 i = follow_hugetlb_page(mm, vma, pages, vmas,
1046 foll_flags = FOLL_TOUCH;
1048 foll_flags |= FOLL_GET;
1049 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1050 (!vma->vm_ops || (!vma->vm_ops->nopage &&
1051 !vma->vm_ops->fault)))
1052 foll_flags |= FOLL_ANON;
1058 * If tsk is ooming, cut off its access to large memory
1059 * allocations. It has a pending SIGKILL, but it can't
1060 * be processed until returning to user space.
1062 if (unlikely(test_tsk_thread_flag(tsk, TIF_MEMDIE)))
1066 foll_flags |= FOLL_WRITE;
1069 while (!(page = follow_page(vma, start, foll_flags))) {
1071 ret = __handle_mm_fault(mm, vma, start,
1072 foll_flags & FOLL_WRITE);
1074 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1075 * broken COW when necessary, even if maybe_mkwrite
1076 * decided not to set pte_write. We can thus safely do
1077 * subsequent page lookups as if they were reads.
1079 if (ret & VM_FAULT_WRITE)
1080 foll_flags &= ~FOLL_WRITE;
1082 switch (ret & ~VM_FAULT_WRITE) {
1083 case VM_FAULT_MINOR:
1086 case VM_FAULT_MAJOR:
1089 case VM_FAULT_SIGBUS:
1090 return i ? i : -EFAULT;
1092 return i ? i : -ENOMEM;
1101 flush_anon_page(vma, page, start);
1102 flush_dcache_page(page);
1109 } while (len && start < vma->vm_end);
1113 EXPORT_SYMBOL(get_user_pages);
1115 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1116 unsigned long addr, unsigned long end, pgprot_t prot)
1122 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1125 arch_enter_lazy_mmu_mode();
1127 struct page *page = ZERO_PAGE(addr);
1128 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1130 if (unlikely(!pte_none(*pte))) {
1135 page_cache_get(page);
1136 page_add_file_rmap(page);
1137 inc_mm_counter(mm, file_rss);
1138 set_pte_at(mm, addr, pte, zero_pte);
1139 } while (pte++, addr += PAGE_SIZE, addr != end);
1140 arch_leave_lazy_mmu_mode();
1141 pte_unmap_unlock(pte - 1, ptl);
1145 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1146 unsigned long addr, unsigned long end, pgprot_t prot)
1152 pmd = pmd_alloc(mm, pud, addr);
1156 next = pmd_addr_end(addr, end);
1157 err = zeromap_pte_range(mm, pmd, addr, next, prot);
1160 } while (pmd++, addr = next, addr != end);
1164 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1165 unsigned long addr, unsigned long end, pgprot_t prot)
1171 pud = pud_alloc(mm, pgd, addr);
1175 next = pud_addr_end(addr, end);
1176 err = zeromap_pmd_range(mm, pud, addr, next, prot);
1179 } while (pud++, addr = next, addr != end);
1183 int zeromap_page_range(struct vm_area_struct *vma,
1184 unsigned long addr, unsigned long size, pgprot_t prot)
1188 unsigned long end = addr + size;
1189 struct mm_struct *mm = vma->vm_mm;
1192 BUG_ON(addr >= end);
1193 pgd = pgd_offset(mm, addr);
1194 flush_cache_range(vma, addr, end);
1196 next = pgd_addr_end(addr, end);
1197 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1200 } while (pgd++, addr = next, addr != end);
1204 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1206 pgd_t * pgd = pgd_offset(mm, addr);
1207 pud_t * pud = pud_alloc(mm, pgd, addr);
1209 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1211 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1217 * This is the old fallback for page remapping.
1219 * For historical reasons, it only allows reserved pages. Only
1220 * old drivers should use this, and they needed to mark their
1221 * pages reserved for the old functions anyway.
1223 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1233 flush_dcache_page(page);
1234 pte = get_locked_pte(mm, addr, &ptl);
1238 if (!pte_none(*pte))
1241 /* Ok, finally just insert the thing.. */
1243 inc_mm_counter(mm, file_rss);
1244 page_add_file_rmap(page);
1245 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1249 pte_unmap_unlock(pte, ptl);
1255 * vm_insert_page - insert single page into user vma
1256 * @vma: user vma to map to
1257 * @addr: target user address of this page
1258 * @page: source kernel page
1260 * This allows drivers to insert individual pages they've allocated
1263 * The page has to be a nice clean _individual_ kernel allocation.
1264 * If you allocate a compound page, you need to have marked it as
1265 * such (__GFP_COMP), or manually just split the page up yourself
1266 * (see split_page()).
1268 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1269 * took an arbitrary page protection parameter. This doesn't allow
1270 * that. Your vma protection will have to be set up correctly, which
1271 * means that if you want a shared writable mapping, you'd better
1272 * ask for a shared writable mapping!
1274 * The page does not need to be reserved.
1276 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1278 if (addr < vma->vm_start || addr >= vma->vm_end)
1280 if (!page_count(page))
1282 vma->vm_flags |= VM_INSERTPAGE;
1283 return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1285 EXPORT_SYMBOL(vm_insert_page);
1288 * vm_insert_pfn - insert single pfn into user vma
1289 * @vma: user vma to map to
1290 * @addr: target user address of this page
1291 * @pfn: source kernel pfn
1293 * Similar to vm_inert_page, this allows drivers to insert individual pages
1294 * they've allocated into a user vma. Same comments apply.
1296 * This function should only be called from a vm_ops->fault handler, and
1297 * in that case the handler should return NULL.
1299 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1302 struct mm_struct *mm = vma->vm_mm;
1307 BUG_ON(!(vma->vm_flags & VM_PFNMAP));
1308 BUG_ON(is_cow_mapping(vma->vm_flags));
1311 pte = get_locked_pte(mm, addr, &ptl);
1315 if (!pte_none(*pte))
1318 /* Ok, finally just insert the thing.. */
1319 entry = pfn_pte(pfn, vma->vm_page_prot);
1320 set_pte_at(mm, addr, pte, entry);
1321 update_mmu_cache(vma, addr, entry);
1325 pte_unmap_unlock(pte, ptl);
1330 EXPORT_SYMBOL(vm_insert_pfn);
1333 * maps a range of physical memory into the requested pages. the old
1334 * mappings are removed. any references to nonexistent pages results
1335 * in null mappings (currently treated as "copy-on-access")
1337 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1338 unsigned long addr, unsigned long end,
1339 unsigned long pfn, pgprot_t prot)
1344 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1347 arch_enter_lazy_mmu_mode();
1349 BUG_ON(!pte_none(*pte));
1350 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1352 } while (pte++, addr += PAGE_SIZE, addr != end);
1353 arch_leave_lazy_mmu_mode();
1354 pte_unmap_unlock(pte - 1, ptl);
1358 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1359 unsigned long addr, unsigned long end,
1360 unsigned long pfn, pgprot_t prot)
1365 pfn -= addr >> PAGE_SHIFT;
1366 pmd = pmd_alloc(mm, pud, addr);
1370 next = pmd_addr_end(addr, end);
1371 if (remap_pte_range(mm, pmd, addr, next,
1372 pfn + (addr >> PAGE_SHIFT), prot))
1374 } while (pmd++, addr = next, addr != end);
1378 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1379 unsigned long addr, unsigned long end,
1380 unsigned long pfn, pgprot_t prot)
1385 pfn -= addr >> PAGE_SHIFT;
1386 pud = pud_alloc(mm, pgd, addr);
1390 next = pud_addr_end(addr, end);
1391 if (remap_pmd_range(mm, pud, addr, next,
1392 pfn + (addr >> PAGE_SHIFT), prot))
1394 } while (pud++, addr = next, addr != end);
1399 * remap_pfn_range - remap kernel memory to userspace
1400 * @vma: user vma to map to
1401 * @addr: target user address to start at
1402 * @pfn: physical address of kernel memory
1403 * @size: size of map area
1404 * @prot: page protection flags for this mapping
1406 * Note: this is only safe if the mm semaphore is held when called.
1408 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1409 unsigned long pfn, unsigned long size, pgprot_t prot)
1413 unsigned long end = addr + PAGE_ALIGN(size);
1414 struct mm_struct *mm = vma->vm_mm;
1418 * Physically remapped pages are special. Tell the
1419 * rest of the world about it:
1420 * VM_IO tells people not to look at these pages
1421 * (accesses can have side effects).
1422 * VM_RESERVED is specified all over the place, because
1423 * in 2.4 it kept swapout's vma scan off this vma; but
1424 * in 2.6 the LRU scan won't even find its pages, so this
1425 * flag means no more than count its pages in reserved_vm,
1426 * and omit it from core dump, even when VM_IO turned off.
1427 * VM_PFNMAP tells the core MM that the base pages are just
1428 * raw PFN mappings, and do not have a "struct page" associated
1431 * There's a horrible special case to handle copy-on-write
1432 * behaviour that some programs depend on. We mark the "original"
1433 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1435 if (is_cow_mapping(vma->vm_flags)) {
1436 if (addr != vma->vm_start || end != vma->vm_end)
1438 vma->vm_pgoff = pfn;
1441 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1443 BUG_ON(addr >= end);
1444 pfn -= addr >> PAGE_SHIFT;
1445 pgd = pgd_offset(mm, addr);
1446 flush_cache_range(vma, addr, end);
1448 next = pgd_addr_end(addr, end);
1449 err = remap_pud_range(mm, pgd, addr, next,
1450 pfn + (addr >> PAGE_SHIFT), prot);
1453 } while (pgd++, addr = next, addr != end);
1456 EXPORT_SYMBOL(remap_pfn_range);
1458 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1459 unsigned long addr, unsigned long end,
1460 pte_fn_t fn, void *data)
1464 struct page *pmd_page;
1465 spinlock_t *uninitialized_var(ptl);
1467 pte = (mm == &init_mm) ?
1468 pte_alloc_kernel(pmd, addr) :
1469 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1473 BUG_ON(pmd_huge(*pmd));
1475 pmd_page = pmd_page(*pmd);
1478 err = fn(pte, pmd_page, addr, data);
1481 } while (pte++, addr += PAGE_SIZE, addr != end);
1484 pte_unmap_unlock(pte-1, ptl);
1488 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1489 unsigned long addr, unsigned long end,
1490 pte_fn_t fn, void *data)
1496 pmd = pmd_alloc(mm, pud, addr);
1500 next = pmd_addr_end(addr, end);
1501 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1504 } while (pmd++, addr = next, addr != end);
1508 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1509 unsigned long addr, unsigned long end,
1510 pte_fn_t fn, void *data)
1516 pud = pud_alloc(mm, pgd, addr);
1520 next = pud_addr_end(addr, end);
1521 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1524 } while (pud++, addr = next, addr != end);
1529 * Scan a region of virtual memory, filling in page tables as necessary
1530 * and calling a provided function on each leaf page table.
1532 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1533 unsigned long size, pte_fn_t fn, void *data)
1537 unsigned long end = addr + size;
1540 BUG_ON(addr >= end);
1541 pgd = pgd_offset(mm, addr);
1543 next = pgd_addr_end(addr, end);
1544 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1547 } while (pgd++, addr = next, addr != end);
1550 EXPORT_SYMBOL_GPL(apply_to_page_range);
1553 * handle_pte_fault chooses page fault handler according to an entry
1554 * which was read non-atomically. Before making any commitment, on
1555 * those architectures or configurations (e.g. i386 with PAE) which
1556 * might give a mix of unmatched parts, do_swap_page and do_file_page
1557 * must check under lock before unmapping the pte and proceeding
1558 * (but do_wp_page is only called after already making such a check;
1559 * and do_anonymous_page and do_no_page can safely check later on).
1561 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1562 pte_t *page_table, pte_t orig_pte)
1565 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1566 if (sizeof(pte_t) > sizeof(unsigned long)) {
1567 spinlock_t *ptl = pte_lockptr(mm, pmd);
1569 same = pte_same(*page_table, orig_pte);
1573 pte_unmap(page_table);
1578 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1579 * servicing faults for write access. In the normal case, do always want
1580 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1581 * that do not have writing enabled, when used by access_process_vm.
1583 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1585 if (likely(vma->vm_flags & VM_WRITE))
1586 pte = pte_mkwrite(pte);
1590 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1593 * If the source page was a PFN mapping, we don't have
1594 * a "struct page" for it. We do a best-effort copy by
1595 * just copying from the original user address. If that
1596 * fails, we just zero-fill it. Live with it.
1598 if (unlikely(!src)) {
1599 void *kaddr = kmap_atomic(dst, KM_USER0);
1600 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1603 * This really shouldn't fail, because the page is there
1604 * in the page tables. But it might just be unreadable,
1605 * in which case we just give up and fill the result with
1608 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1609 memset(kaddr, 0, PAGE_SIZE);
1610 kunmap_atomic(kaddr, KM_USER0);
1611 flush_dcache_page(dst);
1615 copy_user_highpage(dst, src, va, vma);
1619 * This routine handles present pages, when users try to write
1620 * to a shared page. It is done by copying the page to a new address
1621 * and decrementing the shared-page counter for the old page.
1623 * Note that this routine assumes that the protection checks have been
1624 * done by the caller (the low-level page fault routine in most cases).
1625 * Thus we can safely just mark it writable once we've done any necessary
1628 * We also mark the page dirty at this point even though the page will
1629 * change only once the write actually happens. This avoids a few races,
1630 * and potentially makes it more efficient.
1632 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1633 * but allow concurrent faults), with pte both mapped and locked.
1634 * We return with mmap_sem still held, but pte unmapped and unlocked.
1636 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1637 unsigned long address, pte_t *page_table, pmd_t *pmd,
1638 spinlock_t *ptl, pte_t orig_pte)
1640 struct page *old_page, *new_page;
1642 int reuse = 0, ret = VM_FAULT_MINOR;
1643 struct page *dirty_page = NULL;
1645 old_page = vm_normal_page(vma, address, orig_pte);
1650 * Take out anonymous pages first, anonymous shared vmas are
1651 * not dirty accountable.
1653 if (PageAnon(old_page)) {
1654 if (!TestSetPageLocked(old_page)) {
1655 reuse = can_share_swap_page(old_page);
1656 unlock_page(old_page);
1658 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
1659 (VM_WRITE|VM_SHARED))) {
1661 * Only catch write-faults on shared writable pages,
1662 * read-only shared pages can get COWed by
1663 * get_user_pages(.write=1, .force=1).
1665 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
1667 * Notify the address space that the page is about to
1668 * become writable so that it can prohibit this or wait
1669 * for the page to get into an appropriate state.
1671 * We do this without the lock held, so that it can
1672 * sleep if it needs to.
1674 page_cache_get(old_page);
1675 pte_unmap_unlock(page_table, ptl);
1677 if (vma->vm_ops->page_mkwrite(vma, old_page) < 0)
1678 goto unwritable_page;
1681 * Since we dropped the lock we need to revalidate
1682 * the PTE as someone else may have changed it. If
1683 * they did, we just return, as we can count on the
1684 * MMU to tell us if they didn't also make it writable.
1686 page_table = pte_offset_map_lock(mm, pmd, address,
1688 page_cache_release(old_page);
1689 if (!pte_same(*page_table, orig_pte))
1692 dirty_page = old_page;
1693 get_page(dirty_page);
1698 flush_cache_page(vma, address, pte_pfn(orig_pte));
1699 entry = pte_mkyoung(orig_pte);
1700 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1701 if (ptep_set_access_flags(vma, address, page_table, entry,1)) {
1702 update_mmu_cache(vma, address, entry);
1703 lazy_mmu_prot_update(entry);
1705 ret |= VM_FAULT_WRITE;
1710 * Ok, we need to copy. Oh, well..
1712 page_cache_get(old_page);
1714 pte_unmap_unlock(page_table, ptl);
1716 if (unlikely(anon_vma_prepare(vma)))
1718 if (old_page == ZERO_PAGE(address)) {
1719 new_page = alloc_zeroed_user_highpage_movable(vma, address);
1723 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1726 cow_user_page(new_page, old_page, address, vma);
1730 * Re-check the pte - we dropped the lock
1732 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1733 if (likely(pte_same(*page_table, orig_pte))) {
1735 page_remove_rmap(old_page, vma);
1736 if (!PageAnon(old_page)) {
1737 dec_mm_counter(mm, file_rss);
1738 inc_mm_counter(mm, anon_rss);
1741 inc_mm_counter(mm, anon_rss);
1742 flush_cache_page(vma, address, pte_pfn(orig_pte));
1743 entry = mk_pte(new_page, vma->vm_page_prot);
1744 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1745 lazy_mmu_prot_update(entry);
1747 * Clear the pte entry and flush it first, before updating the
1748 * pte with the new entry. This will avoid a race condition
1749 * seen in the presence of one thread doing SMC and another
1752 ptep_clear_flush(vma, address, page_table);
1753 set_pte_at(mm, address, page_table, entry);
1754 update_mmu_cache(vma, address, entry);
1755 lru_cache_add_active(new_page);
1756 page_add_new_anon_rmap(new_page, vma, address);
1758 /* Free the old page.. */
1759 new_page = old_page;
1760 ret |= VM_FAULT_WRITE;
1763 page_cache_release(new_page);
1765 page_cache_release(old_page);
1767 pte_unmap_unlock(page_table, ptl);
1769 set_page_dirty_balance(dirty_page);
1770 put_page(dirty_page);
1775 page_cache_release(old_page);
1776 return VM_FAULT_OOM;
1779 page_cache_release(old_page);
1780 return VM_FAULT_SIGBUS;
1784 * Helper functions for unmap_mapping_range().
1786 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1788 * We have to restart searching the prio_tree whenever we drop the lock,
1789 * since the iterator is only valid while the lock is held, and anyway
1790 * a later vma might be split and reinserted earlier while lock dropped.
1792 * The list of nonlinear vmas could be handled more efficiently, using
1793 * a placeholder, but handle it in the same way until a need is shown.
1794 * It is important to search the prio_tree before nonlinear list: a vma
1795 * may become nonlinear and be shifted from prio_tree to nonlinear list
1796 * while the lock is dropped; but never shifted from list to prio_tree.
1798 * In order to make forward progress despite restarting the search,
1799 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1800 * quickly skip it next time around. Since the prio_tree search only
1801 * shows us those vmas affected by unmapping the range in question, we
1802 * can't efficiently keep all vmas in step with mapping->truncate_count:
1803 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1804 * mapping->truncate_count and vma->vm_truncate_count are protected by
1807 * In order to make forward progress despite repeatedly restarting some
1808 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1809 * and restart from that address when we reach that vma again. It might
1810 * have been split or merged, shrunk or extended, but never shifted: so
1811 * restart_addr remains valid so long as it remains in the vma's range.
1812 * unmap_mapping_range forces truncate_count to leap over page-aligned
1813 * values so we can save vma's restart_addr in its truncate_count field.
1815 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1817 static void reset_vma_truncate_counts(struct address_space *mapping)
1819 struct vm_area_struct *vma;
1820 struct prio_tree_iter iter;
1822 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1823 vma->vm_truncate_count = 0;
1824 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1825 vma->vm_truncate_count = 0;
1828 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1829 unsigned long start_addr, unsigned long end_addr,
1830 struct zap_details *details)
1832 unsigned long restart_addr;
1836 * files that support invalidating or truncating portions of the
1837 * file from under mmaped areas must have their ->fault function
1838 * return a locked page (and FAULT_RET_LOCKED code). This provides
1839 * synchronisation against concurrent unmapping here.
1843 restart_addr = vma->vm_truncate_count;
1844 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1845 start_addr = restart_addr;
1846 if (start_addr >= end_addr) {
1847 /* Top of vma has been split off since last time */
1848 vma->vm_truncate_count = details->truncate_count;
1853 restart_addr = zap_page_range(vma, start_addr,
1854 end_addr - start_addr, details);
1855 need_break = need_resched() ||
1856 need_lockbreak(details->i_mmap_lock);
1858 if (restart_addr >= end_addr) {
1859 /* We have now completed this vma: mark it so */
1860 vma->vm_truncate_count = details->truncate_count;
1864 /* Note restart_addr in vma's truncate_count field */
1865 vma->vm_truncate_count = restart_addr;
1870 spin_unlock(details->i_mmap_lock);
1872 spin_lock(details->i_mmap_lock);
1876 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1877 struct zap_details *details)
1879 struct vm_area_struct *vma;
1880 struct prio_tree_iter iter;
1881 pgoff_t vba, vea, zba, zea;
1884 vma_prio_tree_foreach(vma, &iter, root,
1885 details->first_index, details->last_index) {
1886 /* Skip quickly over those we have already dealt with */
1887 if (vma->vm_truncate_count == details->truncate_count)
1890 vba = vma->vm_pgoff;
1891 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1892 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1893 zba = details->first_index;
1896 zea = details->last_index;
1900 if (unmap_mapping_range_vma(vma,
1901 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1902 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1908 static inline void unmap_mapping_range_list(struct list_head *head,
1909 struct zap_details *details)
1911 struct vm_area_struct *vma;
1914 * In nonlinear VMAs there is no correspondence between virtual address
1915 * offset and file offset. So we must perform an exhaustive search
1916 * across *all* the pages in each nonlinear VMA, not just the pages
1917 * whose virtual address lies outside the file truncation point.
1920 list_for_each_entry(vma, head, shared.vm_set.list) {
1921 /* Skip quickly over those we have already dealt with */
1922 if (vma->vm_truncate_count == details->truncate_count)
1924 details->nonlinear_vma = vma;
1925 if (unmap_mapping_range_vma(vma, vma->vm_start,
1926 vma->vm_end, details) < 0)
1932 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
1933 * @mapping: the address space containing mmaps to be unmapped.
1934 * @holebegin: byte in first page to unmap, relative to the start of
1935 * the underlying file. This will be rounded down to a PAGE_SIZE
1936 * boundary. Note that this is different from vmtruncate(), which
1937 * must keep the partial page. In contrast, we must get rid of
1939 * @holelen: size of prospective hole in bytes. This will be rounded
1940 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1942 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1943 * but 0 when invalidating pagecache, don't throw away private data.
1945 void unmap_mapping_range(struct address_space *mapping,
1946 loff_t const holebegin, loff_t const holelen, int even_cows)
1948 struct zap_details details;
1949 pgoff_t hba = holebegin >> PAGE_SHIFT;
1950 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1952 /* Check for overflow. */
1953 if (sizeof(holelen) > sizeof(hlen)) {
1955 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1956 if (holeend & ~(long long)ULONG_MAX)
1957 hlen = ULONG_MAX - hba + 1;
1960 details.check_mapping = even_cows? NULL: mapping;
1961 details.nonlinear_vma = NULL;
1962 details.first_index = hba;
1963 details.last_index = hba + hlen - 1;
1964 if (details.last_index < details.first_index)
1965 details.last_index = ULONG_MAX;
1966 details.i_mmap_lock = &mapping->i_mmap_lock;
1968 spin_lock(&mapping->i_mmap_lock);
1970 /* Protect against endless unmapping loops */
1971 mapping->truncate_count++;
1972 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1973 if (mapping->truncate_count == 0)
1974 reset_vma_truncate_counts(mapping);
1975 mapping->truncate_count++;
1977 details.truncate_count = mapping->truncate_count;
1979 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1980 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1981 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1982 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1983 spin_unlock(&mapping->i_mmap_lock);
1985 EXPORT_SYMBOL(unmap_mapping_range);
1988 * vmtruncate - unmap mappings "freed" by truncate() syscall
1989 * @inode: inode of the file used
1990 * @offset: file offset to start truncating
1992 * NOTE! We have to be ready to update the memory sharing
1993 * between the file and the memory map for a potential last
1994 * incomplete page. Ugly, but necessary.
1996 int vmtruncate(struct inode * inode, loff_t offset)
1998 struct address_space *mapping = inode->i_mapping;
1999 unsigned long limit;
2001 if (inode->i_size < offset)
2004 * truncation of in-use swapfiles is disallowed - it would cause
2005 * subsequent swapout to scribble on the now-freed blocks.
2007 if (IS_SWAPFILE(inode))
2009 i_size_write(inode, offset);
2012 * unmap_mapping_range is called twice, first simply for efficiency
2013 * so that truncate_inode_pages does fewer single-page unmaps. However
2014 * after this first call, and before truncate_inode_pages finishes,
2015 * it is possible for private pages to be COWed, which remain after
2016 * truncate_inode_pages finishes, hence the second unmap_mapping_range
2017 * call must be made for correctness.
2019 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2020 truncate_inode_pages(mapping, offset);
2021 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
2025 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2026 if (limit != RLIM_INFINITY && offset > limit)
2028 if (offset > inode->i_sb->s_maxbytes)
2030 i_size_write(inode, offset);
2033 if (inode->i_op && inode->i_op->truncate)
2034 inode->i_op->truncate(inode);
2037 send_sig(SIGXFSZ, current, 0);
2043 EXPORT_SYMBOL(vmtruncate);
2045 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
2047 struct address_space *mapping = inode->i_mapping;
2050 * If the underlying filesystem is not going to provide
2051 * a way to truncate a range of blocks (punch a hole) -
2052 * we should return failure right now.
2054 if (!inode->i_op || !inode->i_op->truncate_range)
2057 mutex_lock(&inode->i_mutex);
2058 down_write(&inode->i_alloc_sem);
2059 unmap_mapping_range(mapping, offset, (end - offset), 1);
2060 truncate_inode_pages_range(mapping, offset, end);
2061 unmap_mapping_range(mapping, offset, (end - offset), 1);
2062 inode->i_op->truncate_range(inode, offset, end);
2063 up_write(&inode->i_alloc_sem);
2064 mutex_unlock(&inode->i_mutex);
2070 * swapin_readahead - swap in pages in hope we need them soon
2071 * @entry: swap entry of this memory
2072 * @addr: address to start
2073 * @vma: user vma this addresses belong to
2075 * Primitive swap readahead code. We simply read an aligned block of
2076 * (1 << page_cluster) entries in the swap area. This method is chosen
2077 * because it doesn't cost us any seek time. We also make sure to queue
2078 * the 'original' request together with the readahead ones...
2080 * This has been extended to use the NUMA policies from the mm triggering
2083 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
2085 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
2088 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
2091 struct page *new_page;
2092 unsigned long offset;
2095 * Get the number of handles we should do readahead io to.
2097 num = valid_swaphandles(entry, &offset);
2098 for (i = 0; i < num; offset++, i++) {
2099 /* Ok, do the async read-ahead now */
2100 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
2101 offset), vma, addr);
2104 page_cache_release(new_page);
2107 * Find the next applicable VMA for the NUMA policy.
2113 if (addr >= vma->vm_end) {
2115 next_vma = vma ? vma->vm_next : NULL;
2117 if (vma && addr < vma->vm_start)
2120 if (next_vma && addr >= next_vma->vm_start) {
2122 next_vma = vma->vm_next;
2127 lru_add_drain(); /* Push any new pages onto the LRU now */
2131 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2132 * but allow concurrent faults), and pte mapped but not yet locked.
2133 * We return with mmap_sem still held, but pte unmapped and unlocked.
2135 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2136 unsigned long address, pte_t *page_table, pmd_t *pmd,
2137 int write_access, pte_t orig_pte)
2143 int ret = VM_FAULT_MINOR;
2145 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2148 entry = pte_to_swp_entry(orig_pte);
2149 if (is_migration_entry(entry)) {
2150 migration_entry_wait(mm, pmd, address);
2153 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2154 page = lookup_swap_cache(entry);
2156 grab_swap_token(); /* Contend for token _before_ read-in */
2157 swapin_readahead(entry, address, vma);
2158 page = read_swap_cache_async(entry, vma, address);
2161 * Back out if somebody else faulted in this pte
2162 * while we released the pte lock.
2164 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2165 if (likely(pte_same(*page_table, orig_pte)))
2167 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2171 /* Had to read the page from swap area: Major fault */
2172 ret = VM_FAULT_MAJOR;
2173 count_vm_event(PGMAJFAULT);
2176 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2177 mark_page_accessed(page);
2181 * Back out if somebody else already faulted in this pte.
2183 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2184 if (unlikely(!pte_same(*page_table, orig_pte)))
2187 if (unlikely(!PageUptodate(page))) {
2188 ret = VM_FAULT_SIGBUS;
2192 /* The page isn't present yet, go ahead with the fault. */
2194 inc_mm_counter(mm, anon_rss);
2195 pte = mk_pte(page, vma->vm_page_prot);
2196 if (write_access && can_share_swap_page(page)) {
2197 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2201 flush_icache_page(vma, page);
2202 set_pte_at(mm, address, page_table, pte);
2203 page_add_anon_rmap(page, vma, address);
2207 remove_exclusive_swap_page(page);
2211 if (do_wp_page(mm, vma, address,
2212 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2217 /* No need to invalidate - it was non-present before */
2218 update_mmu_cache(vma, address, pte);
2220 pte_unmap_unlock(page_table, ptl);
2224 pte_unmap_unlock(page_table, ptl);
2226 page_cache_release(page);
2231 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2232 * but allow concurrent faults), and pte mapped but not yet locked.
2233 * We return with mmap_sem still held, but pte unmapped and unlocked.
2235 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2236 unsigned long address, pte_t *page_table, pmd_t *pmd,
2244 /* Allocate our own private page. */
2245 pte_unmap(page_table);
2247 if (unlikely(anon_vma_prepare(vma)))
2249 page = alloc_zeroed_user_highpage_movable(vma, address);
2253 entry = mk_pte(page, vma->vm_page_prot);
2254 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2256 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2257 if (!pte_none(*page_table))
2259 inc_mm_counter(mm, anon_rss);
2260 lru_cache_add_active(page);
2261 page_add_new_anon_rmap(page, vma, address);
2263 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2264 page = ZERO_PAGE(address);
2265 page_cache_get(page);
2266 entry = mk_pte(page, vma->vm_page_prot);
2268 ptl = pte_lockptr(mm, pmd);
2270 if (!pte_none(*page_table))
2272 inc_mm_counter(mm, file_rss);
2273 page_add_file_rmap(page);
2276 set_pte_at(mm, address, page_table, entry);
2278 /* No need to invalidate - it was non-present before */
2279 update_mmu_cache(vma, address, entry);
2280 lazy_mmu_prot_update(entry);
2282 pte_unmap_unlock(page_table, ptl);
2283 return VM_FAULT_MINOR;
2285 page_cache_release(page);
2288 return VM_FAULT_OOM;
2292 * __do_fault() tries to create a new page mapping. It aggressively
2293 * tries to share with existing pages, but makes a separate copy if
2294 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
2295 * the next page fault.
2297 * As this is called only for pages that do not currently exist, we
2298 * do not need to flush old virtual caches or the TLB.
2300 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2301 * but allow concurrent faults), and pte mapped but not yet locked.
2302 * We return with mmap_sem still held, but pte unmapped and unlocked.
2304 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2305 unsigned long address, pte_t *page_table, pmd_t *pmd,
2306 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2312 struct page *dirty_page = NULL;
2313 struct vm_fault vmf;
2316 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2321 pte_unmap(page_table);
2322 BUG_ON(vma->vm_flags & VM_PFNMAP);
2324 if (likely(vma->vm_ops->fault)) {
2325 ret = vma->vm_ops->fault(vma, &vmf);
2326 if (unlikely(ret & (VM_FAULT_ERROR | FAULT_RET_NOPAGE)))
2327 return (ret & VM_FAULT_MASK);
2329 /* Legacy ->nopage path */
2330 ret = VM_FAULT_MINOR;
2331 vmf.page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2332 /* no page was available -- either SIGBUS or OOM */
2333 if (unlikely(vmf.page == NOPAGE_SIGBUS))
2334 return VM_FAULT_SIGBUS;
2335 else if (unlikely(vmf.page == NOPAGE_OOM))
2336 return VM_FAULT_OOM;
2340 * For consistency in subsequent calls, make the faulted page always
2343 if (unlikely(!(ret & FAULT_RET_LOCKED)))
2344 lock_page(vmf.page);
2346 VM_BUG_ON(!PageLocked(vmf.page));
2349 * Should we do an early C-O-W break?
2352 if (flags & FAULT_FLAG_WRITE) {
2353 if (!(vma->vm_flags & VM_SHARED)) {
2355 if (unlikely(anon_vma_prepare(vma))) {
2359 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2364 copy_user_highpage(page, vmf.page, address, vma);
2367 * If the page will be shareable, see if the backing
2368 * address space wants to know that the page is about
2369 * to become writable
2371 if (vma->vm_ops->page_mkwrite) {
2373 if (vma->vm_ops->page_mkwrite(vma, page) < 0) {
2374 ret = VM_FAULT_SIGBUS;
2375 anon = 1; /* no anon but release vmf.page */
2380 * XXX: this is not quite right (racy vs
2381 * invalidate) to unlock and relock the page
2382 * like this, however a better fix requires
2383 * reworking page_mkwrite locking API, which
2384 * is better done later.
2386 if (!page->mapping) {
2387 ret = VM_FAULT_MINOR;
2388 anon = 1; /* no anon but release vmf.page */
2396 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2399 * This silly early PAGE_DIRTY setting removes a race
2400 * due to the bad i386 page protection. But it's valid
2401 * for other architectures too.
2403 * Note that if write_access is true, we either now have
2404 * an exclusive copy of the page, or this is a shared mapping,
2405 * so we can make it writable and dirty to avoid having to
2406 * handle that later.
2408 /* Only go through if we didn't race with anybody else... */
2409 if (likely(pte_same(*page_table, orig_pte))) {
2410 flush_icache_page(vma, page);
2411 entry = mk_pte(page, vma->vm_page_prot);
2412 if (flags & FAULT_FLAG_WRITE)
2413 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2414 set_pte_at(mm, address, page_table, entry);
2416 inc_mm_counter(mm, anon_rss);
2417 lru_cache_add_active(page);
2418 page_add_new_anon_rmap(page, vma, address);
2420 inc_mm_counter(mm, file_rss);
2421 page_add_file_rmap(page);
2422 if (flags & FAULT_FLAG_WRITE) {
2424 get_page(dirty_page);
2428 /* no need to invalidate: a not-present page won't be cached */
2429 update_mmu_cache(vma, address, entry);
2430 lazy_mmu_prot_update(entry);
2433 page_cache_release(page);
2435 anon = 1; /* no anon but release faulted_page */
2438 pte_unmap_unlock(page_table, ptl);
2441 unlock_page(vmf.page);
2444 page_cache_release(vmf.page);
2445 else if (dirty_page) {
2446 set_page_dirty_balance(dirty_page);
2447 put_page(dirty_page);
2450 return (ret & VM_FAULT_MASK);
2453 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2454 unsigned long address, pte_t *page_table, pmd_t *pmd,
2455 int write_access, pte_t orig_pte)
2457 pgoff_t pgoff = (((address & PAGE_MASK)
2458 - vma->vm_start) >> PAGE_CACHE_SHIFT) + vma->vm_pgoff;
2459 unsigned int flags = (write_access ? FAULT_FLAG_WRITE : 0);
2461 return __do_fault(mm, vma, address, page_table, pmd, pgoff,
2467 * do_no_pfn() tries to create a new page mapping for a page without
2468 * a struct_page backing it
2470 * As this is called only for pages that do not currently exist, we
2471 * do not need to flush old virtual caches or the TLB.
2473 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2474 * but allow concurrent faults), and pte mapped but not yet locked.
2475 * We return with mmap_sem still held, but pte unmapped and unlocked.
2477 * It is expected that the ->nopfn handler always returns the same pfn
2478 * for a given virtual mapping.
2480 * Mark this `noinline' to prevent it from bloating the main pagefault code.
2482 static noinline int do_no_pfn(struct mm_struct *mm, struct vm_area_struct *vma,
2483 unsigned long address, pte_t *page_table, pmd_t *pmd,
2489 int ret = VM_FAULT_MINOR;
2491 pte_unmap(page_table);
2492 BUG_ON(!(vma->vm_flags & VM_PFNMAP));
2493 BUG_ON(is_cow_mapping(vma->vm_flags));
2495 pfn = vma->vm_ops->nopfn(vma, address & PAGE_MASK);
2496 if (unlikely(pfn == NOPFN_OOM))
2497 return VM_FAULT_OOM;
2498 else if (unlikely(pfn == NOPFN_SIGBUS))
2499 return VM_FAULT_SIGBUS;
2500 else if (unlikely(pfn == NOPFN_REFAULT))
2501 return VM_FAULT_MINOR;
2503 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2505 /* Only go through if we didn't race with anybody else... */
2506 if (pte_none(*page_table)) {
2507 entry = pfn_pte(pfn, vma->vm_page_prot);
2509 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2510 set_pte_at(mm, address, page_table, entry);
2512 pte_unmap_unlock(page_table, ptl);
2517 * Fault of a previously existing named mapping. Repopulate the pte
2518 * from the encoded file_pte if possible. This enables swappable
2521 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2522 * but allow concurrent faults), and pte mapped but not yet locked.
2523 * We return with mmap_sem still held, but pte unmapped and unlocked.
2525 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2526 unsigned long address, pte_t *page_table, pmd_t *pmd,
2527 int write_access, pte_t orig_pte)
2529 unsigned int flags = FAULT_FLAG_NONLINEAR |
2530 (write_access ? FAULT_FLAG_WRITE : 0);
2533 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2534 return VM_FAULT_MINOR;
2536 if (unlikely(!(vma->vm_flags & VM_NONLINEAR) ||
2537 !(vma->vm_flags & VM_CAN_NONLINEAR))) {
2539 * Page table corrupted: show pte and kill process.
2541 print_bad_pte(vma, orig_pte, address);
2542 return VM_FAULT_OOM;
2545 pgoff = pte_to_pgoff(orig_pte);
2547 return __do_fault(mm, vma, address, page_table, pmd, pgoff,
2552 * These routines also need to handle stuff like marking pages dirty
2553 * and/or accessed for architectures that don't do it in hardware (most
2554 * RISC architectures). The early dirtying is also good on the i386.
2556 * There is also a hook called "update_mmu_cache()" that architectures
2557 * with external mmu caches can use to update those (ie the Sparc or
2558 * PowerPC hashed page tables that act as extended TLBs).
2560 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2561 * but allow concurrent faults), and pte mapped but not yet locked.
2562 * We return with mmap_sem still held, but pte unmapped and unlocked.
2564 static inline int handle_pte_fault(struct mm_struct *mm,
2565 struct vm_area_struct *vma, unsigned long address,
2566 pte_t *pte, pmd_t *pmd, int write_access)
2572 if (!pte_present(entry)) {
2573 if (pte_none(entry)) {
2575 if (vma->vm_ops->fault || vma->vm_ops->nopage)
2576 return do_linear_fault(mm, vma, address,
2577 pte, pmd, write_access, entry);
2578 if (unlikely(vma->vm_ops->nopfn))
2579 return do_no_pfn(mm, vma, address, pte,
2582 return do_anonymous_page(mm, vma, address,
2583 pte, pmd, write_access);
2585 if (pte_file(entry))
2586 return do_nonlinear_fault(mm, vma, address,
2587 pte, pmd, write_access, entry);
2588 return do_swap_page(mm, vma, address,
2589 pte, pmd, write_access, entry);
2592 ptl = pte_lockptr(mm, pmd);
2594 if (unlikely(!pte_same(*pte, entry)))
2597 if (!pte_write(entry))
2598 return do_wp_page(mm, vma, address,
2599 pte, pmd, ptl, entry);
2600 entry = pte_mkdirty(entry);
2602 entry = pte_mkyoung(entry);
2603 if (ptep_set_access_flags(vma, address, pte, entry, write_access)) {
2604 update_mmu_cache(vma, address, entry);
2605 lazy_mmu_prot_update(entry);
2608 * This is needed only for protection faults but the arch code
2609 * is not yet telling us if this is a protection fault or not.
2610 * This still avoids useless tlb flushes for .text page faults
2614 flush_tlb_page(vma, address);
2617 pte_unmap_unlock(pte, ptl);
2618 return VM_FAULT_MINOR;
2622 * By the time we get here, we already hold the mm semaphore
2624 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2625 unsigned long address, int write_access)
2632 __set_current_state(TASK_RUNNING);
2634 count_vm_event(PGFAULT);
2636 if (unlikely(is_vm_hugetlb_page(vma)))
2637 return hugetlb_fault(mm, vma, address, write_access);
2639 pgd = pgd_offset(mm, address);
2640 pud = pud_alloc(mm, pgd, address);
2642 return VM_FAULT_OOM;
2643 pmd = pmd_alloc(mm, pud, address);
2645 return VM_FAULT_OOM;
2646 pte = pte_alloc_map(mm, pmd, address);
2648 return VM_FAULT_OOM;
2650 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2653 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2655 #ifndef __PAGETABLE_PUD_FOLDED
2657 * Allocate page upper directory.
2658 * We've already handled the fast-path in-line.
2660 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2662 pud_t *new = pud_alloc_one(mm, address);
2666 spin_lock(&mm->page_table_lock);
2667 if (pgd_present(*pgd)) /* Another has populated it */
2670 pgd_populate(mm, pgd, new);
2671 spin_unlock(&mm->page_table_lock);
2674 #endif /* __PAGETABLE_PUD_FOLDED */
2676 #ifndef __PAGETABLE_PMD_FOLDED
2678 * Allocate page middle directory.
2679 * We've already handled the fast-path in-line.
2681 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2683 pmd_t *new = pmd_alloc_one(mm, address);
2687 spin_lock(&mm->page_table_lock);
2688 #ifndef __ARCH_HAS_4LEVEL_HACK
2689 if (pud_present(*pud)) /* Another has populated it */
2692 pud_populate(mm, pud, new);
2694 if (pgd_present(*pud)) /* Another has populated it */
2697 pgd_populate(mm, pud, new);
2698 #endif /* __ARCH_HAS_4LEVEL_HACK */
2699 spin_unlock(&mm->page_table_lock);
2702 #endif /* __PAGETABLE_PMD_FOLDED */
2704 int make_pages_present(unsigned long addr, unsigned long end)
2706 int ret, len, write;
2707 struct vm_area_struct * vma;
2709 vma = find_vma(current->mm, addr);
2712 write = (vma->vm_flags & VM_WRITE) != 0;
2713 BUG_ON(addr >= end);
2714 BUG_ON(end > vma->vm_end);
2715 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
2716 ret = get_user_pages(current, current->mm, addr,
2717 len, write, 0, NULL, NULL);
2720 return ret == len ? 0 : -1;
2724 * Map a vmalloc()-space virtual address to the physical page.
2726 struct page * vmalloc_to_page(void * vmalloc_addr)
2728 unsigned long addr = (unsigned long) vmalloc_addr;
2729 struct page *page = NULL;
2730 pgd_t *pgd = pgd_offset_k(addr);
2735 if (!pgd_none(*pgd)) {
2736 pud = pud_offset(pgd, addr);
2737 if (!pud_none(*pud)) {
2738 pmd = pmd_offset(pud, addr);
2739 if (!pmd_none(*pmd)) {
2740 ptep = pte_offset_map(pmd, addr);
2742 if (pte_present(pte))
2743 page = pte_page(pte);
2751 EXPORT_SYMBOL(vmalloc_to_page);
2754 * Map a vmalloc()-space virtual address to the physical page frame number.
2756 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2758 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2761 EXPORT_SYMBOL(vmalloc_to_pfn);
2763 #if !defined(__HAVE_ARCH_GATE_AREA)
2765 #if defined(AT_SYSINFO_EHDR)
2766 static struct vm_area_struct gate_vma;
2768 static int __init gate_vma_init(void)
2770 gate_vma.vm_mm = NULL;
2771 gate_vma.vm_start = FIXADDR_USER_START;
2772 gate_vma.vm_end = FIXADDR_USER_END;
2773 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
2774 gate_vma.vm_page_prot = __P101;
2776 * Make sure the vDSO gets into every core dump.
2777 * Dumping its contents makes post-mortem fully interpretable later
2778 * without matching up the same kernel and hardware config to see
2779 * what PC values meant.
2781 gate_vma.vm_flags |= VM_ALWAYSDUMP;
2784 __initcall(gate_vma_init);
2787 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2789 #ifdef AT_SYSINFO_EHDR
2796 int in_gate_area_no_task(unsigned long addr)
2798 #ifdef AT_SYSINFO_EHDR
2799 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2805 #endif /* __HAVE_ARCH_GATE_AREA */
2808 * Access another process' address space.
2809 * Source/target buffer must be kernel space,
2810 * Do not walk the page table directly, use get_user_pages
2812 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2814 struct mm_struct *mm;
2815 struct vm_area_struct *vma;
2817 void *old_buf = buf;
2819 mm = get_task_mm(tsk);
2823 down_read(&mm->mmap_sem);
2824 /* ignore errors, just check how much was sucessfully transfered */
2826 int bytes, ret, offset;
2829 ret = get_user_pages(tsk, mm, addr, 1,
2830 write, 1, &page, &vma);
2835 offset = addr & (PAGE_SIZE-1);
2836 if (bytes > PAGE_SIZE-offset)
2837 bytes = PAGE_SIZE-offset;
2841 copy_to_user_page(vma, page, addr,
2842 maddr + offset, buf, bytes);
2843 set_page_dirty_lock(page);
2845 copy_from_user_page(vma, page, addr,
2846 buf, maddr + offset, bytes);
2849 page_cache_release(page);
2854 up_read(&mm->mmap_sem);
2857 return buf - old_buf;