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/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
79 unsigned long num_physpages;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
105 static int __init disable_randmaps(char *s)
107 randomize_va_space = 0;
110 __setup("norandmaps", disable_randmaps);
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init init_zero_pfn(void)
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
123 core_initcall(init_zero_pfn);
126 #if defined(SPLIT_RSS_COUNTING)
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
138 task->rss_stat.events = 0;
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 struct task_struct *task = current;
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
148 add_mm_counter(mm, member, val);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct *task)
157 if (unlikely(task != current))
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
171 val = atomic_long_read(&mm->rss_stat.count[member]);
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
178 return (unsigned long)val;
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
183 __sync_task_rss_stat(task, mm);
185 #else /* SPLIT_RSS_COUNTING */
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
190 static void check_sync_rss_stat(struct task_struct *task)
194 #endif /* SPLIT_RSS_COUNTING */
196 #ifdef HAVE_GENERIC_MMU_GATHER
198 static int tlb_next_batch(struct mmu_gather *tlb)
200 struct mmu_gather_batch *batch;
204 tlb->active = batch->next;
208 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
211 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
218 batch->max = MAX_GATHER_BATCH;
220 tlb->active->next = batch;
227 * Called to initialize an (on-stack) mmu_gather structure for page-table
228 * tear-down from @mm. The @fullmm argument is used when @mm is without
229 * users and we're going to destroy the full address space (exit/execve).
231 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
235 tlb->fullmm = fullmm;
237 tlb->fast_mode = (num_possible_cpus() == 1);
238 tlb->local.next = NULL;
240 tlb->local.max = ARRAY_SIZE(tlb->__pages);
241 tlb->active = &tlb->local;
242 tlb->batch_count = 0;
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 void tlb_flush_mmu(struct mmu_gather *tlb)
251 struct mmu_gather_batch *batch;
253 if (!tlb->need_flush)
257 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
258 tlb_table_flush(tlb);
261 if (tlb_fast_mode(tlb))
264 for (batch = &tlb->local; batch; batch = batch->next) {
265 free_pages_and_swap_cache(batch->pages, batch->nr);
268 tlb->active = &tlb->local;
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
277 struct mmu_gather_batch *batch, *next;
281 /* keep the page table cache within bounds */
284 for (batch = tlb->local.next; batch; batch = next) {
286 free_pages((unsigned long)batch, 0);
288 tlb->local.next = NULL;
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
297 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
299 struct mmu_gather_batch *batch;
303 if (tlb_fast_mode(tlb)) {
304 free_page_and_swap_cache(page);
305 return 1; /* avoid calling tlb_flush_mmu() */
309 batch->pages[batch->nr++] = page;
310 if (batch->nr == batch->max) {
311 if (!tlb_next_batch(tlb))
315 VM_BUG_ON(batch->nr > batch->max);
317 return batch->max - batch->nr;
320 #endif /* HAVE_GENERIC_MMU_GATHER */
322 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
325 * See the comment near struct mmu_table_batch.
328 static void tlb_remove_table_smp_sync(void *arg)
330 /* Simply deliver the interrupt */
333 static void tlb_remove_table_one(void *table)
336 * This isn't an RCU grace period and hence the page-tables cannot be
337 * assumed to be actually RCU-freed.
339 * It is however sufficient for software page-table walkers that rely on
340 * IRQ disabling. See the comment near struct mmu_table_batch.
342 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
343 __tlb_remove_table(table);
346 static void tlb_remove_table_rcu(struct rcu_head *head)
348 struct mmu_table_batch *batch;
351 batch = container_of(head, struct mmu_table_batch, rcu);
353 for (i = 0; i < batch->nr; i++)
354 __tlb_remove_table(batch->tables[i]);
356 free_page((unsigned long)batch);
359 void tlb_table_flush(struct mmu_gather *tlb)
361 struct mmu_table_batch **batch = &tlb->batch;
364 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
369 void tlb_remove_table(struct mmu_gather *tlb, void *table)
371 struct mmu_table_batch **batch = &tlb->batch;
376 * When there's less then two users of this mm there cannot be a
377 * concurrent page-table walk.
379 if (atomic_read(&tlb->mm->mm_users) < 2) {
380 __tlb_remove_table(table);
384 if (*batch == NULL) {
385 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
386 if (*batch == NULL) {
387 tlb_remove_table_one(table);
392 (*batch)->tables[(*batch)->nr++] = table;
393 if ((*batch)->nr == MAX_TABLE_BATCH)
394 tlb_table_flush(tlb);
397 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
400 * If a p?d_bad entry is found while walking page tables, report
401 * the error, before resetting entry to p?d_none. Usually (but
402 * very seldom) called out from the p?d_none_or_clear_bad macros.
405 void pgd_clear_bad(pgd_t *pgd)
411 void pud_clear_bad(pud_t *pud)
417 void pmd_clear_bad(pmd_t *pmd)
424 * Note: this doesn't free the actual pages themselves. That
425 * has been handled earlier when unmapping all the memory regions.
427 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
430 pgtable_t token = pmd_pgtable(*pmd);
432 pte_free_tlb(tlb, token, addr);
436 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
437 unsigned long addr, unsigned long end,
438 unsigned long floor, unsigned long ceiling)
445 pmd = pmd_offset(pud, addr);
447 next = pmd_addr_end(addr, end);
448 if (pmd_none_or_clear_bad(pmd))
450 free_pte_range(tlb, pmd, addr);
451 } while (pmd++, addr = next, addr != end);
461 if (end - 1 > ceiling - 1)
464 pmd = pmd_offset(pud, start);
466 pmd_free_tlb(tlb, pmd, start);
469 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
470 unsigned long addr, unsigned long end,
471 unsigned long floor, unsigned long ceiling)
478 pud = pud_offset(pgd, addr);
480 next = pud_addr_end(addr, end);
481 if (pud_none_or_clear_bad(pud))
483 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
484 } while (pud++, addr = next, addr != end);
490 ceiling &= PGDIR_MASK;
494 if (end - 1 > ceiling - 1)
497 pud = pud_offset(pgd, start);
499 pud_free_tlb(tlb, pud, start);
503 * This function frees user-level page tables of a process.
505 * Must be called with pagetable lock held.
507 void free_pgd_range(struct mmu_gather *tlb,
508 unsigned long addr, unsigned long end,
509 unsigned long floor, unsigned long ceiling)
515 * The next few lines have given us lots of grief...
517 * Why are we testing PMD* at this top level? Because often
518 * there will be no work to do at all, and we'd prefer not to
519 * go all the way down to the bottom just to discover that.
521 * Why all these "- 1"s? Because 0 represents both the bottom
522 * of the address space and the top of it (using -1 for the
523 * top wouldn't help much: the masks would do the wrong thing).
524 * The rule is that addr 0 and floor 0 refer to the bottom of
525 * the address space, but end 0 and ceiling 0 refer to the top
526 * Comparisons need to use "end - 1" and "ceiling - 1" (though
527 * that end 0 case should be mythical).
529 * Wherever addr is brought up or ceiling brought down, we must
530 * be careful to reject "the opposite 0" before it confuses the
531 * subsequent tests. But what about where end is brought down
532 * by PMD_SIZE below? no, end can't go down to 0 there.
534 * Whereas we round start (addr) and ceiling down, by different
535 * masks at different levels, in order to test whether a table
536 * now has no other vmas using it, so can be freed, we don't
537 * bother to round floor or end up - the tests don't need that.
551 if (end - 1 > ceiling - 1)
556 pgd = pgd_offset(tlb->mm, addr);
558 next = pgd_addr_end(addr, end);
559 if (pgd_none_or_clear_bad(pgd))
561 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
562 } while (pgd++, addr = next, addr != end);
565 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
566 unsigned long floor, unsigned long ceiling)
569 struct vm_area_struct *next = vma->vm_next;
570 unsigned long addr = vma->vm_start;
573 * Hide vma from rmap and truncate_pagecache before freeing
576 unlink_anon_vmas(vma);
577 unlink_file_vma(vma);
579 if (is_vm_hugetlb_page(vma)) {
580 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
581 floor, next? next->vm_start: ceiling);
584 * Optimization: gather nearby vmas into one call down
586 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
587 && !is_vm_hugetlb_page(next)) {
590 unlink_anon_vmas(vma);
591 unlink_file_vma(vma);
593 free_pgd_range(tlb, addr, vma->vm_end,
594 floor, next? next->vm_start: ceiling);
600 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
601 pmd_t *pmd, unsigned long address)
603 pgtable_t new = pte_alloc_one(mm, address);
604 int wait_split_huge_page;
609 * Ensure all pte setup (eg. pte page lock and page clearing) are
610 * visible before the pte is made visible to other CPUs by being
611 * put into page tables.
613 * The other side of the story is the pointer chasing in the page
614 * table walking code (when walking the page table without locking;
615 * ie. most of the time). Fortunately, these data accesses consist
616 * of a chain of data-dependent loads, meaning most CPUs (alpha
617 * being the notable exception) will already guarantee loads are
618 * seen in-order. See the alpha page table accessors for the
619 * smp_read_barrier_depends() barriers in page table walking code.
621 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
623 spin_lock(&mm->page_table_lock);
624 wait_split_huge_page = 0;
625 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
627 pmd_populate(mm, pmd, new);
629 } else if (unlikely(pmd_trans_splitting(*pmd)))
630 wait_split_huge_page = 1;
631 spin_unlock(&mm->page_table_lock);
634 if (wait_split_huge_page)
635 wait_split_huge_page(vma->anon_vma, pmd);
639 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
641 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
645 smp_wmb(); /* See comment in __pte_alloc */
647 spin_lock(&init_mm.page_table_lock);
648 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
649 pmd_populate_kernel(&init_mm, pmd, new);
652 VM_BUG_ON(pmd_trans_splitting(*pmd));
653 spin_unlock(&init_mm.page_table_lock);
655 pte_free_kernel(&init_mm, new);
659 static inline void init_rss_vec(int *rss)
661 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
664 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
668 if (current->mm == mm)
669 sync_mm_rss(current, mm);
670 for (i = 0; i < NR_MM_COUNTERS; i++)
672 add_mm_counter(mm, i, rss[i]);
676 * This function is called to print an error when a bad pte
677 * is found. For example, we might have a PFN-mapped pte in
678 * a region that doesn't allow it.
680 * The calling function must still handle the error.
682 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
683 pte_t pte, struct page *page)
685 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
686 pud_t *pud = pud_offset(pgd, addr);
687 pmd_t *pmd = pmd_offset(pud, addr);
688 struct address_space *mapping;
690 static unsigned long resume;
691 static unsigned long nr_shown;
692 static unsigned long nr_unshown;
695 * Allow a burst of 60 reports, then keep quiet for that minute;
696 * or allow a steady drip of one report per second.
698 if (nr_shown == 60) {
699 if (time_before(jiffies, resume)) {
705 "BUG: Bad page map: %lu messages suppressed\n",
712 resume = jiffies + 60 * HZ;
714 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
715 index = linear_page_index(vma, addr);
718 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
720 (long long)pte_val(pte), (long long)pmd_val(*pmd));
724 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
725 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
727 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
730 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
731 (unsigned long)vma->vm_ops->fault);
732 if (vma->vm_file && vma->vm_file->f_op)
733 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
734 (unsigned long)vma->vm_file->f_op->mmap);
736 add_taint(TAINT_BAD_PAGE);
739 static inline int is_cow_mapping(vm_flags_t flags)
741 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
745 static inline int is_zero_pfn(unsigned long pfn)
747 return pfn == zero_pfn;
752 static inline unsigned long my_zero_pfn(unsigned long addr)
759 * vm_normal_page -- This function gets the "struct page" associated with a pte.
761 * "Special" mappings do not wish to be associated with a "struct page" (either
762 * it doesn't exist, or it exists but they don't want to touch it). In this
763 * case, NULL is returned here. "Normal" mappings do have a struct page.
765 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
766 * pte bit, in which case this function is trivial. Secondly, an architecture
767 * may not have a spare pte bit, which requires a more complicated scheme,
770 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
771 * special mapping (even if there are underlying and valid "struct pages").
772 * COWed pages of a VM_PFNMAP are always normal.
774 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
775 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
776 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
777 * mapping will always honor the rule
779 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
781 * And for normal mappings this is false.
783 * This restricts such mappings to be a linear translation from virtual address
784 * to pfn. To get around this restriction, we allow arbitrary mappings so long
785 * as the vma is not a COW mapping; in that case, we know that all ptes are
786 * special (because none can have been COWed).
789 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
791 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
792 * page" backing, however the difference is that _all_ pages with a struct
793 * page (that is, those where pfn_valid is true) are refcounted and considered
794 * normal pages by the VM. The disadvantage is that pages are refcounted
795 * (which can be slower and simply not an option for some PFNMAP users). The
796 * advantage is that we don't have to follow the strict linearity rule of
797 * PFNMAP mappings in order to support COWable mappings.
800 #ifdef __HAVE_ARCH_PTE_SPECIAL
801 # define HAVE_PTE_SPECIAL 1
803 # define HAVE_PTE_SPECIAL 0
805 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
808 unsigned long pfn = pte_pfn(pte);
810 if (HAVE_PTE_SPECIAL) {
811 if (likely(!pte_special(pte)))
813 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
815 if (!is_zero_pfn(pfn))
816 print_bad_pte(vma, addr, pte, NULL);
820 /* !HAVE_PTE_SPECIAL case follows: */
822 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
823 if (vma->vm_flags & VM_MIXEDMAP) {
829 off = (addr - vma->vm_start) >> PAGE_SHIFT;
830 if (pfn == vma->vm_pgoff + off)
832 if (!is_cow_mapping(vma->vm_flags))
837 if (is_zero_pfn(pfn))
840 if (unlikely(pfn > highest_memmap_pfn)) {
841 print_bad_pte(vma, addr, pte, NULL);
846 * NOTE! We still have PageReserved() pages in the page tables.
847 * eg. VDSO mappings can cause them to exist.
850 return pfn_to_page(pfn);
854 * copy one vm_area from one task to the other. Assumes the page tables
855 * already present in the new task to be cleared in the whole range
856 * covered by this vma.
859 static inline unsigned long
860 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
861 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
862 unsigned long addr, int *rss)
864 unsigned long vm_flags = vma->vm_flags;
865 pte_t pte = *src_pte;
868 /* pte contains position in swap or file, so copy. */
869 if (unlikely(!pte_present(pte))) {
870 if (!pte_file(pte)) {
871 swp_entry_t entry = pte_to_swp_entry(pte);
873 if (swap_duplicate(entry) < 0)
876 /* make sure dst_mm is on swapoff's mmlist. */
877 if (unlikely(list_empty(&dst_mm->mmlist))) {
878 spin_lock(&mmlist_lock);
879 if (list_empty(&dst_mm->mmlist))
880 list_add(&dst_mm->mmlist,
882 spin_unlock(&mmlist_lock);
884 if (likely(!non_swap_entry(entry)))
886 else if (is_write_migration_entry(entry) &&
887 is_cow_mapping(vm_flags)) {
889 * COW mappings require pages in both parent
890 * and child to be set to read.
892 make_migration_entry_read(&entry);
893 pte = swp_entry_to_pte(entry);
894 set_pte_at(src_mm, addr, src_pte, pte);
901 * If it's a COW mapping, write protect it both
902 * in the parent and the child
904 if (is_cow_mapping(vm_flags)) {
905 ptep_set_wrprotect(src_mm, addr, src_pte);
906 pte = pte_wrprotect(pte);
910 * If it's a shared mapping, mark it clean in
913 if (vm_flags & VM_SHARED)
914 pte = pte_mkclean(pte);
915 pte = pte_mkold(pte);
917 page = vm_normal_page(vma, addr, pte);
928 set_pte_at(dst_mm, addr, dst_pte, pte);
932 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
933 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
934 unsigned long addr, unsigned long end)
936 pte_t *orig_src_pte, *orig_dst_pte;
937 pte_t *src_pte, *dst_pte;
938 spinlock_t *src_ptl, *dst_ptl;
940 int rss[NR_MM_COUNTERS];
941 swp_entry_t entry = (swp_entry_t){0};
946 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
949 src_pte = pte_offset_map(src_pmd, addr);
950 src_ptl = pte_lockptr(src_mm, src_pmd);
951 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
952 orig_src_pte = src_pte;
953 orig_dst_pte = dst_pte;
954 arch_enter_lazy_mmu_mode();
958 * We are holding two locks at this point - either of them
959 * could generate latencies in another task on another CPU.
961 if (progress >= 32) {
963 if (need_resched() ||
964 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
967 if (pte_none(*src_pte)) {
971 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
976 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
978 arch_leave_lazy_mmu_mode();
979 spin_unlock(src_ptl);
980 pte_unmap(orig_src_pte);
981 add_mm_rss_vec(dst_mm, rss);
982 pte_unmap_unlock(orig_dst_pte, dst_ptl);
986 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
995 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
996 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
997 unsigned long addr, unsigned long end)
999 pmd_t *src_pmd, *dst_pmd;
1002 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1005 src_pmd = pmd_offset(src_pud, addr);
1007 next = pmd_addr_end(addr, end);
1008 if (pmd_trans_huge(*src_pmd)) {
1010 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1011 err = copy_huge_pmd(dst_mm, src_mm,
1012 dst_pmd, src_pmd, addr, vma);
1019 if (pmd_none_or_clear_bad(src_pmd))
1021 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1024 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1028 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1029 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1030 unsigned long addr, unsigned long end)
1032 pud_t *src_pud, *dst_pud;
1035 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1038 src_pud = pud_offset(src_pgd, addr);
1040 next = pud_addr_end(addr, end);
1041 if (pud_none_or_clear_bad(src_pud))
1043 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1046 } while (dst_pud++, src_pud++, addr = next, addr != end);
1050 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1051 struct vm_area_struct *vma)
1053 pgd_t *src_pgd, *dst_pgd;
1055 unsigned long addr = vma->vm_start;
1056 unsigned long end = vma->vm_end;
1060 * Don't copy ptes where a page fault will fill them correctly.
1061 * Fork becomes much lighter when there are big shared or private
1062 * readonly mappings. The tradeoff is that copy_page_range is more
1063 * efficient than faulting.
1065 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1070 if (is_vm_hugetlb_page(vma))
1071 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1073 if (unlikely(is_pfn_mapping(vma))) {
1075 * We do not free on error cases below as remove_vma
1076 * gets called on error from higher level routine
1078 ret = track_pfn_vma_copy(vma);
1084 * We need to invalidate the secondary MMU mappings only when
1085 * there could be a permission downgrade on the ptes of the
1086 * parent mm. And a permission downgrade will only happen if
1087 * is_cow_mapping() returns true.
1089 if (is_cow_mapping(vma->vm_flags))
1090 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1093 dst_pgd = pgd_offset(dst_mm, addr);
1094 src_pgd = pgd_offset(src_mm, addr);
1096 next = pgd_addr_end(addr, end);
1097 if (pgd_none_or_clear_bad(src_pgd))
1099 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100 vma, addr, next))) {
1104 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1106 if (is_cow_mapping(vma->vm_flags))
1107 mmu_notifier_invalidate_range_end(src_mm,
1108 vma->vm_start, end);
1112 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1113 struct vm_area_struct *vma, pmd_t *pmd,
1114 unsigned long addr, unsigned long end,
1115 struct zap_details *details)
1117 struct mm_struct *mm = tlb->mm;
1118 int force_flush = 0;
1119 int rss[NR_MM_COUNTERS];
1126 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1128 arch_enter_lazy_mmu_mode();
1131 if (pte_none(ptent)) {
1135 if (pte_present(ptent)) {
1138 page = vm_normal_page(vma, addr, ptent);
1139 if (unlikely(details) && page) {
1141 * unmap_shared_mapping_pages() wants to
1142 * invalidate cache without truncating:
1143 * unmap shared but keep private pages.
1145 if (details->check_mapping &&
1146 details->check_mapping != page->mapping)
1149 * Each page->index must be checked when
1150 * invalidating or truncating nonlinear.
1152 if (details->nonlinear_vma &&
1153 (page->index < details->first_index ||
1154 page->index > details->last_index))
1157 ptent = ptep_get_and_clear_full(mm, addr, pte,
1159 tlb_remove_tlb_entry(tlb, pte, addr);
1160 if (unlikely(!page))
1162 if (unlikely(details) && details->nonlinear_vma
1163 && linear_page_index(details->nonlinear_vma,
1164 addr) != page->index)
1165 set_pte_at(mm, addr, pte,
1166 pgoff_to_pte(page->index));
1168 rss[MM_ANONPAGES]--;
1170 if (pte_dirty(ptent))
1171 set_page_dirty(page);
1172 if (pte_young(ptent) &&
1173 likely(!VM_SequentialReadHint(vma)))
1174 mark_page_accessed(page);
1175 rss[MM_FILEPAGES]--;
1177 page_remove_rmap(page);
1178 if (unlikely(page_mapcount(page) < 0))
1179 print_bad_pte(vma, addr, ptent, page);
1180 force_flush = !__tlb_remove_page(tlb, page);
1186 * If details->check_mapping, we leave swap entries;
1187 * if details->nonlinear_vma, we leave file entries.
1189 if (unlikely(details))
1191 if (pte_file(ptent)) {
1192 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1193 print_bad_pte(vma, addr, ptent, NULL);
1195 swp_entry_t entry = pte_to_swp_entry(ptent);
1197 if (!non_swap_entry(entry))
1199 if (unlikely(!free_swap_and_cache(entry)))
1200 print_bad_pte(vma, addr, ptent, NULL);
1202 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1203 } while (pte++, addr += PAGE_SIZE, addr != end);
1205 add_mm_rss_vec(mm, rss);
1206 arch_leave_lazy_mmu_mode();
1207 pte_unmap_unlock(start_pte, ptl);
1210 * mmu_gather ran out of room to batch pages, we break out of
1211 * the PTE lock to avoid doing the potential expensive TLB invalidate
1212 * and page-free while holding it.
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1232 pmd = pmd_offset(pud, addr);
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1238 split_huge_page_pmd(vma->vm_mm, pmd);
1239 } else if (zap_huge_pmd(tlb, vma, pmd))
1244 * Here there can be other concurrent MADV_DONTNEED or
1245 * trans huge page faults running, and if the pmd is
1246 * none or trans huge it can change under us. This is
1247 * because MADV_DONTNEED holds the mmap_sem in read
1250 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1252 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1255 } while (pmd++, addr = next, addr != end);
1260 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1261 struct vm_area_struct *vma, pgd_t *pgd,
1262 unsigned long addr, unsigned long end,
1263 struct zap_details *details)
1268 pud = pud_offset(pgd, addr);
1270 next = pud_addr_end(addr, end);
1271 if (pud_none_or_clear_bad(pud))
1273 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1274 } while (pud++, addr = next, addr != end);
1279 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1280 struct vm_area_struct *vma,
1281 unsigned long addr, unsigned long end,
1282 struct zap_details *details)
1287 if (details && !details->check_mapping && !details->nonlinear_vma)
1290 BUG_ON(addr >= end);
1291 mem_cgroup_uncharge_start();
1292 tlb_start_vma(tlb, vma);
1293 pgd = pgd_offset(vma->vm_mm, addr);
1295 next = pgd_addr_end(addr, end);
1296 if (pgd_none_or_clear_bad(pgd))
1298 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1299 } while (pgd++, addr = next, addr != end);
1300 tlb_end_vma(tlb, vma);
1301 mem_cgroup_uncharge_end();
1306 #ifdef CONFIG_PREEMPT
1307 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
1309 /* No preempt: go for improved straight-line efficiency */
1310 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
1314 * unmap_vmas - unmap a range of memory covered by a list of vma's
1315 * @tlb: address of the caller's struct mmu_gather
1316 * @vma: the starting vma
1317 * @start_addr: virtual address at which to start unmapping
1318 * @end_addr: virtual address at which to end unmapping
1319 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1320 * @details: details of nonlinear truncation or shared cache invalidation
1322 * Returns the end address of the unmapping (restart addr if interrupted).
1324 * Unmap all pages in the vma list.
1326 * We aim to not hold locks for too long (for scheduling latency reasons).
1327 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1328 * return the ending mmu_gather to the caller.
1330 * Only addresses between `start' and `end' will be unmapped.
1332 * The VMA list must be sorted in ascending virtual address order.
1334 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1335 * range after unmap_vmas() returns. So the only responsibility here is to
1336 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1337 * drops the lock and schedules.
1339 unsigned long unmap_vmas(struct mmu_gather *tlb,
1340 struct vm_area_struct *vma, unsigned long start_addr,
1341 unsigned long end_addr, unsigned long *nr_accounted,
1342 struct zap_details *details)
1344 unsigned long start = start_addr;
1345 struct mm_struct *mm = vma->vm_mm;
1347 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1348 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1351 start = max(vma->vm_start, start_addr);
1352 if (start >= vma->vm_end)
1354 end = min(vma->vm_end, end_addr);
1355 if (end <= vma->vm_start)
1358 if (vma->vm_flags & VM_ACCOUNT)
1359 *nr_accounted += (end - start) >> PAGE_SHIFT;
1361 if (unlikely(is_pfn_mapping(vma)))
1362 untrack_pfn_vma(vma, 0, 0);
1364 while (start != end) {
1365 if (unlikely(is_vm_hugetlb_page(vma))) {
1367 * It is undesirable to test vma->vm_file as it
1368 * should be non-null for valid hugetlb area.
1369 * However, vm_file will be NULL in the error
1370 * cleanup path of do_mmap_pgoff. When
1371 * hugetlbfs ->mmap method fails,
1372 * do_mmap_pgoff() nullifies vma->vm_file
1373 * before calling this function to clean up.
1374 * Since no pte has actually been setup, it is
1375 * safe to do nothing in this case.
1378 unmap_hugepage_range(vma, start, end, NULL);
1382 start = unmap_page_range(tlb, vma, start, end, details);
1386 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1387 return start; /* which is now the end (or restart) address */
1391 * zap_page_range - remove user pages in a given range
1392 * @vma: vm_area_struct holding the applicable pages
1393 * @address: starting address of pages to zap
1394 * @size: number of bytes to zap
1395 * @details: details of nonlinear truncation or shared cache invalidation
1397 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1398 unsigned long size, struct zap_details *details)
1400 struct mm_struct *mm = vma->vm_mm;
1401 struct mmu_gather tlb;
1402 unsigned long end = address + size;
1403 unsigned long nr_accounted = 0;
1406 tlb_gather_mmu(&tlb, mm, 0);
1407 update_hiwater_rss(mm);
1408 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1409 tlb_finish_mmu(&tlb, address, end);
1414 * zap_vma_ptes - remove ptes mapping the vma
1415 * @vma: vm_area_struct holding ptes to be zapped
1416 * @address: starting address of pages to zap
1417 * @size: number of bytes to zap
1419 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1421 * The entire address range must be fully contained within the vma.
1423 * Returns 0 if successful.
1425 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1428 if (address < vma->vm_start || address + size > vma->vm_end ||
1429 !(vma->vm_flags & VM_PFNMAP))
1431 zap_page_range(vma, address, size, NULL);
1434 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1437 * follow_page - look up a page descriptor from a user-virtual address
1438 * @vma: vm_area_struct mapping @address
1439 * @address: virtual address to look up
1440 * @flags: flags modifying lookup behaviour
1442 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1444 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1445 * an error pointer if there is a mapping to something not represented
1446 * by a page descriptor (see also vm_normal_page()).
1448 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1457 struct mm_struct *mm = vma->vm_mm;
1459 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1460 if (!IS_ERR(page)) {
1461 BUG_ON(flags & FOLL_GET);
1466 pgd = pgd_offset(mm, address);
1467 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1470 pud = pud_offset(pgd, address);
1473 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1474 BUG_ON(flags & FOLL_GET);
1475 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1478 if (unlikely(pud_bad(*pud)))
1481 pmd = pmd_offset(pud, address);
1484 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1485 BUG_ON(flags & FOLL_GET);
1486 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1489 if (pmd_trans_huge(*pmd)) {
1490 if (flags & FOLL_SPLIT) {
1491 split_huge_page_pmd(mm, pmd);
1492 goto split_fallthrough;
1494 spin_lock(&mm->page_table_lock);
1495 if (likely(pmd_trans_huge(*pmd))) {
1496 if (unlikely(pmd_trans_splitting(*pmd))) {
1497 spin_unlock(&mm->page_table_lock);
1498 wait_split_huge_page(vma->anon_vma, pmd);
1500 page = follow_trans_huge_pmd(mm, address,
1502 spin_unlock(&mm->page_table_lock);
1506 spin_unlock(&mm->page_table_lock);
1510 if (unlikely(pmd_bad(*pmd)))
1513 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1516 if (!pte_present(pte))
1518 if ((flags & FOLL_WRITE) && !pte_write(pte))
1521 page = vm_normal_page(vma, address, pte);
1522 if (unlikely(!page)) {
1523 if ((flags & FOLL_DUMP) ||
1524 !is_zero_pfn(pte_pfn(pte)))
1526 page = pte_page(pte);
1529 if (flags & FOLL_GET)
1530 get_page_foll(page);
1531 if (flags & FOLL_TOUCH) {
1532 if ((flags & FOLL_WRITE) &&
1533 !pte_dirty(pte) && !PageDirty(page))
1534 set_page_dirty(page);
1536 * pte_mkyoung() would be more correct here, but atomic care
1537 * is needed to avoid losing the dirty bit: it is easier to use
1538 * mark_page_accessed().
1540 mark_page_accessed(page);
1542 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1544 * The preliminary mapping check is mainly to avoid the
1545 * pointless overhead of lock_page on the ZERO_PAGE
1546 * which might bounce very badly if there is contention.
1548 * If the page is already locked, we don't need to
1549 * handle it now - vmscan will handle it later if and
1550 * when it attempts to reclaim the page.
1552 if (page->mapping && trylock_page(page)) {
1553 lru_add_drain(); /* push cached pages to LRU */
1555 * Because we lock page here and migration is
1556 * blocked by the pte's page reference, we need
1557 * only check for file-cache page truncation.
1560 mlock_vma_page(page);
1565 pte_unmap_unlock(ptep, ptl);
1570 pte_unmap_unlock(ptep, ptl);
1571 return ERR_PTR(-EFAULT);
1574 pte_unmap_unlock(ptep, ptl);
1580 * When core dumping an enormous anonymous area that nobody
1581 * has touched so far, we don't want to allocate unnecessary pages or
1582 * page tables. Return error instead of NULL to skip handle_mm_fault,
1583 * then get_dump_page() will return NULL to leave a hole in the dump.
1584 * But we can only make this optimization where a hole would surely
1585 * be zero-filled if handle_mm_fault() actually did handle it.
1587 if ((flags & FOLL_DUMP) &&
1588 (!vma->vm_ops || !vma->vm_ops->fault))
1589 return ERR_PTR(-EFAULT);
1593 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1595 return stack_guard_page_start(vma, addr) ||
1596 stack_guard_page_end(vma, addr+PAGE_SIZE);
1600 * __get_user_pages() - pin user pages in memory
1601 * @tsk: task_struct of target task
1602 * @mm: mm_struct of target mm
1603 * @start: starting user address
1604 * @nr_pages: number of pages from start to pin
1605 * @gup_flags: flags modifying pin behaviour
1606 * @pages: array that receives pointers to the pages pinned.
1607 * Should be at least nr_pages long. Or NULL, if caller
1608 * only intends to ensure the pages are faulted in.
1609 * @vmas: array of pointers to vmas corresponding to each page.
1610 * Or NULL if the caller does not require them.
1611 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1613 * Returns number of pages pinned. This may be fewer than the number
1614 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1615 * were pinned, returns -errno. Each page returned must be released
1616 * with a put_page() call when it is finished with. vmas will only
1617 * remain valid while mmap_sem is held.
1619 * Must be called with mmap_sem held for read or write.
1621 * __get_user_pages walks a process's page tables and takes a reference to
1622 * each struct page that each user address corresponds to at a given
1623 * instant. That is, it takes the page that would be accessed if a user
1624 * thread accesses the given user virtual address at that instant.
1626 * This does not guarantee that the page exists in the user mappings when
1627 * __get_user_pages returns, and there may even be a completely different
1628 * page there in some cases (eg. if mmapped pagecache has been invalidated
1629 * and subsequently re faulted). However it does guarantee that the page
1630 * won't be freed completely. And mostly callers simply care that the page
1631 * contains data that was valid *at some point in time*. Typically, an IO
1632 * or similar operation cannot guarantee anything stronger anyway because
1633 * locks can't be held over the syscall boundary.
1635 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1636 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1637 * appropriate) must be called after the page is finished with, and
1638 * before put_page is called.
1640 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1641 * or mmap_sem contention, and if waiting is needed to pin all pages,
1642 * *@nonblocking will be set to 0.
1644 * In most cases, get_user_pages or get_user_pages_fast should be used
1645 * instead of __get_user_pages. __get_user_pages should be used only if
1646 * you need some special @gup_flags.
1648 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1649 unsigned long start, int nr_pages, unsigned int gup_flags,
1650 struct page **pages, struct vm_area_struct **vmas,
1654 unsigned long vm_flags;
1659 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1662 * Require read or write permissions.
1663 * If FOLL_FORCE is set, we only require the "MAY" flags.
1665 vm_flags = (gup_flags & FOLL_WRITE) ?
1666 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1667 vm_flags &= (gup_flags & FOLL_FORCE) ?
1668 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1672 struct vm_area_struct *vma;
1674 vma = find_extend_vma(mm, start);
1675 if (!vma && in_gate_area(mm, start)) {
1676 unsigned long pg = start & PAGE_MASK;
1682 /* user gate pages are read-only */
1683 if (gup_flags & FOLL_WRITE)
1684 return i ? : -EFAULT;
1686 pgd = pgd_offset_k(pg);
1688 pgd = pgd_offset_gate(mm, pg);
1689 BUG_ON(pgd_none(*pgd));
1690 pud = pud_offset(pgd, pg);
1691 BUG_ON(pud_none(*pud));
1692 pmd = pmd_offset(pud, pg);
1694 return i ? : -EFAULT;
1695 VM_BUG_ON(pmd_trans_huge(*pmd));
1696 pte = pte_offset_map(pmd, pg);
1697 if (pte_none(*pte)) {
1699 return i ? : -EFAULT;
1701 vma = get_gate_vma(mm);
1705 page = vm_normal_page(vma, start, *pte);
1707 if (!(gup_flags & FOLL_DUMP) &&
1708 is_zero_pfn(pte_pfn(*pte)))
1709 page = pte_page(*pte);
1712 return i ? : -EFAULT;
1723 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1724 !(vm_flags & vma->vm_flags))
1725 return i ? : -EFAULT;
1727 if (is_vm_hugetlb_page(vma)) {
1728 i = follow_hugetlb_page(mm, vma, pages, vmas,
1729 &start, &nr_pages, i, gup_flags);
1735 unsigned int foll_flags = gup_flags;
1738 * If we have a pending SIGKILL, don't keep faulting
1739 * pages and potentially allocating memory.
1741 if (unlikely(fatal_signal_pending(current)))
1742 return i ? i : -ERESTARTSYS;
1745 while (!(page = follow_page(vma, start, foll_flags))) {
1747 unsigned int fault_flags = 0;
1749 /* For mlock, just skip the stack guard page. */
1750 if (foll_flags & FOLL_MLOCK) {
1751 if (stack_guard_page(vma, start))
1754 if (foll_flags & FOLL_WRITE)
1755 fault_flags |= FAULT_FLAG_WRITE;
1757 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1758 if (foll_flags & FOLL_NOWAIT)
1759 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1761 ret = handle_mm_fault(mm, vma, start,
1764 if (ret & VM_FAULT_ERROR) {
1765 if (ret & VM_FAULT_OOM)
1766 return i ? i : -ENOMEM;
1767 if (ret & (VM_FAULT_HWPOISON |
1768 VM_FAULT_HWPOISON_LARGE)) {
1771 else if (gup_flags & FOLL_HWPOISON)
1776 if (ret & VM_FAULT_SIGBUS)
1777 return i ? i : -EFAULT;
1782 if (ret & VM_FAULT_MAJOR)
1788 if (ret & VM_FAULT_RETRY) {
1795 * The VM_FAULT_WRITE bit tells us that
1796 * do_wp_page has broken COW when necessary,
1797 * even if maybe_mkwrite decided not to set
1798 * pte_write. We can thus safely do subsequent
1799 * page lookups as if they were reads. But only
1800 * do so when looping for pte_write is futile:
1801 * in some cases userspace may also be wanting
1802 * to write to the gotten user page, which a
1803 * read fault here might prevent (a readonly
1804 * page might get reCOWed by userspace write).
1806 if ((ret & VM_FAULT_WRITE) &&
1807 !(vma->vm_flags & VM_WRITE))
1808 foll_flags &= ~FOLL_WRITE;
1813 return i ? i : PTR_ERR(page);
1817 flush_anon_page(vma, page, start);
1818 flush_dcache_page(page);
1826 } while (nr_pages && start < vma->vm_end);
1830 EXPORT_SYMBOL(__get_user_pages);
1833 * fixup_user_fault() - manually resolve a user page fault
1834 * @tsk: the task_struct to use for page fault accounting, or
1835 * NULL if faults are not to be recorded.
1836 * @mm: mm_struct of target mm
1837 * @address: user address
1838 * @fault_flags:flags to pass down to handle_mm_fault()
1840 * This is meant to be called in the specific scenario where for locking reasons
1841 * we try to access user memory in atomic context (within a pagefault_disable()
1842 * section), this returns -EFAULT, and we want to resolve the user fault before
1845 * Typically this is meant to be used by the futex code.
1847 * The main difference with get_user_pages() is that this function will
1848 * unconditionally call handle_mm_fault() which will in turn perform all the
1849 * necessary SW fixup of the dirty and young bits in the PTE, while
1850 * handle_mm_fault() only guarantees to update these in the struct page.
1852 * This is important for some architectures where those bits also gate the
1853 * access permission to the page because they are maintained in software. On
1854 * such architectures, gup() will not be enough to make a subsequent access
1857 * This should be called with the mm_sem held for read.
1859 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1860 unsigned long address, unsigned int fault_flags)
1862 struct vm_area_struct *vma;
1865 vma = find_extend_vma(mm, address);
1866 if (!vma || address < vma->vm_start)
1869 ret = handle_mm_fault(mm, vma, address, fault_flags);
1870 if (ret & VM_FAULT_ERROR) {
1871 if (ret & VM_FAULT_OOM)
1873 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1875 if (ret & VM_FAULT_SIGBUS)
1880 if (ret & VM_FAULT_MAJOR)
1889 * get_user_pages() - pin user pages in memory
1890 * @tsk: the task_struct to use for page fault accounting, or
1891 * NULL if faults are not to be recorded.
1892 * @mm: mm_struct of target mm
1893 * @start: starting user address
1894 * @nr_pages: number of pages from start to pin
1895 * @write: whether pages will be written to by the caller
1896 * @force: whether to force write access even if user mapping is
1897 * readonly. This will result in the page being COWed even
1898 * in MAP_SHARED mappings. You do not want this.
1899 * @pages: array that receives pointers to the pages pinned.
1900 * Should be at least nr_pages long. Or NULL, if caller
1901 * only intends to ensure the pages are faulted in.
1902 * @vmas: array of pointers to vmas corresponding to each page.
1903 * Or NULL if the caller does not require them.
1905 * Returns number of pages pinned. This may be fewer than the number
1906 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1907 * were pinned, returns -errno. Each page returned must be released
1908 * with a put_page() call when it is finished with. vmas will only
1909 * remain valid while mmap_sem is held.
1911 * Must be called with mmap_sem held for read or write.
1913 * get_user_pages walks a process's page tables and takes a reference to
1914 * each struct page that each user address corresponds to at a given
1915 * instant. That is, it takes the page that would be accessed if a user
1916 * thread accesses the given user virtual address at that instant.
1918 * This does not guarantee that the page exists in the user mappings when
1919 * get_user_pages returns, and there may even be a completely different
1920 * page there in some cases (eg. if mmapped pagecache has been invalidated
1921 * and subsequently re faulted). However it does guarantee that the page
1922 * won't be freed completely. And mostly callers simply care that the page
1923 * contains data that was valid *at some point in time*. Typically, an IO
1924 * or similar operation cannot guarantee anything stronger anyway because
1925 * locks can't be held over the syscall boundary.
1927 * If write=0, the page must not be written to. If the page is written to,
1928 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1929 * after the page is finished with, and before put_page is called.
1931 * get_user_pages is typically used for fewer-copy IO operations, to get a
1932 * handle on the memory by some means other than accesses via the user virtual
1933 * addresses. The pages may be submitted for DMA to devices or accessed via
1934 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1935 * use the correct cache flushing APIs.
1937 * See also get_user_pages_fast, for performance critical applications.
1939 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1940 unsigned long start, int nr_pages, int write, int force,
1941 struct page **pages, struct vm_area_struct **vmas)
1943 int flags = FOLL_TOUCH;
1948 flags |= FOLL_WRITE;
1950 flags |= FOLL_FORCE;
1952 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1955 EXPORT_SYMBOL(get_user_pages);
1958 * get_dump_page() - pin user page in memory while writing it to core dump
1959 * @addr: user address
1961 * Returns struct page pointer of user page pinned for dump,
1962 * to be freed afterwards by page_cache_release() or put_page().
1964 * Returns NULL on any kind of failure - a hole must then be inserted into
1965 * the corefile, to preserve alignment with its headers; and also returns
1966 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1967 * allowing a hole to be left in the corefile to save diskspace.
1969 * Called without mmap_sem, but after all other threads have been killed.
1971 #ifdef CONFIG_ELF_CORE
1972 struct page *get_dump_page(unsigned long addr)
1974 struct vm_area_struct *vma;
1977 if (__get_user_pages(current, current->mm, addr, 1,
1978 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1981 flush_cache_page(vma, addr, page_to_pfn(page));
1984 #endif /* CONFIG_ELF_CORE */
1986 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1989 pgd_t * pgd = pgd_offset(mm, addr);
1990 pud_t * pud = pud_alloc(mm, pgd, addr);
1992 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1994 VM_BUG_ON(pmd_trans_huge(*pmd));
1995 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2002 * This is the old fallback for page remapping.
2004 * For historical reasons, it only allows reserved pages. Only
2005 * old drivers should use this, and they needed to mark their
2006 * pages reserved for the old functions anyway.
2008 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2009 struct page *page, pgprot_t prot)
2011 struct mm_struct *mm = vma->vm_mm;
2020 flush_dcache_page(page);
2021 pte = get_locked_pte(mm, addr, &ptl);
2025 if (!pte_none(*pte))
2028 /* Ok, finally just insert the thing.. */
2030 inc_mm_counter_fast(mm, MM_FILEPAGES);
2031 page_add_file_rmap(page);
2032 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2035 pte_unmap_unlock(pte, ptl);
2038 pte_unmap_unlock(pte, ptl);
2044 * vm_insert_page - insert single page into user vma
2045 * @vma: user vma to map to
2046 * @addr: target user address of this page
2047 * @page: source kernel page
2049 * This allows drivers to insert individual pages they've allocated
2052 * The page has to be a nice clean _individual_ kernel allocation.
2053 * If you allocate a compound page, you need to have marked it as
2054 * such (__GFP_COMP), or manually just split the page up yourself
2055 * (see split_page()).
2057 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2058 * took an arbitrary page protection parameter. This doesn't allow
2059 * that. Your vma protection will have to be set up correctly, which
2060 * means that if you want a shared writable mapping, you'd better
2061 * ask for a shared writable mapping!
2063 * The page does not need to be reserved.
2065 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2068 if (addr < vma->vm_start || addr >= vma->vm_end)
2070 if (!page_count(page))
2072 vma->vm_flags |= VM_INSERTPAGE;
2073 return insert_page(vma, addr, page, vma->vm_page_prot);
2075 EXPORT_SYMBOL(vm_insert_page);
2077 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2078 unsigned long pfn, pgprot_t prot)
2080 struct mm_struct *mm = vma->vm_mm;
2086 pte = get_locked_pte(mm, addr, &ptl);
2090 if (!pte_none(*pte))
2093 /* Ok, finally just insert the thing.. */
2094 entry = pte_mkspecial(pfn_pte(pfn, prot));
2095 set_pte_at(mm, addr, pte, entry);
2096 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2100 pte_unmap_unlock(pte, ptl);
2106 * vm_insert_pfn - insert single pfn into user vma
2107 * @vma: user vma to map to
2108 * @addr: target user address of this page
2109 * @pfn: source kernel pfn
2111 * Similar to vm_inert_page, this allows drivers to insert individual pages
2112 * they've allocated into a user vma. Same comments apply.
2114 * This function should only be called from a vm_ops->fault handler, and
2115 * in that case the handler should return NULL.
2117 * vma cannot be a COW mapping.
2119 * As this is called only for pages that do not currently exist, we
2120 * do not need to flush old virtual caches or the TLB.
2122 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2126 pgprot_t pgprot = vma->vm_page_prot;
2128 * Technically, architectures with pte_special can avoid all these
2129 * restrictions (same for remap_pfn_range). However we would like
2130 * consistency in testing and feature parity among all, so we should
2131 * try to keep these invariants in place for everybody.
2133 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2134 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2135 (VM_PFNMAP|VM_MIXEDMAP));
2136 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2137 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2139 if (addr < vma->vm_start || addr >= vma->vm_end)
2141 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2144 ret = insert_pfn(vma, addr, pfn, pgprot);
2147 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2151 EXPORT_SYMBOL(vm_insert_pfn);
2153 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2156 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2158 if (addr < vma->vm_start || addr >= vma->vm_end)
2162 * If we don't have pte special, then we have to use the pfn_valid()
2163 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2164 * refcount the page if pfn_valid is true (hence insert_page rather
2165 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2166 * without pte special, it would there be refcounted as a normal page.
2168 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2171 page = pfn_to_page(pfn);
2172 return insert_page(vma, addr, page, vma->vm_page_prot);
2174 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2176 EXPORT_SYMBOL(vm_insert_mixed);
2179 * maps a range of physical memory into the requested pages. the old
2180 * mappings are removed. any references to nonexistent pages results
2181 * in null mappings (currently treated as "copy-on-access")
2183 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2184 unsigned long addr, unsigned long end,
2185 unsigned long pfn, pgprot_t prot)
2190 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2193 arch_enter_lazy_mmu_mode();
2195 BUG_ON(!pte_none(*pte));
2196 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2198 } while (pte++, addr += PAGE_SIZE, addr != end);
2199 arch_leave_lazy_mmu_mode();
2200 pte_unmap_unlock(pte - 1, ptl);
2204 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2205 unsigned long addr, unsigned long end,
2206 unsigned long pfn, pgprot_t prot)
2211 pfn -= addr >> PAGE_SHIFT;
2212 pmd = pmd_alloc(mm, pud, addr);
2215 VM_BUG_ON(pmd_trans_huge(*pmd));
2217 next = pmd_addr_end(addr, end);
2218 if (remap_pte_range(mm, pmd, addr, next,
2219 pfn + (addr >> PAGE_SHIFT), prot))
2221 } while (pmd++, addr = next, addr != end);
2225 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2226 unsigned long addr, unsigned long end,
2227 unsigned long pfn, pgprot_t prot)
2232 pfn -= addr >> PAGE_SHIFT;
2233 pud = pud_alloc(mm, pgd, addr);
2237 next = pud_addr_end(addr, end);
2238 if (remap_pmd_range(mm, pud, addr, next,
2239 pfn + (addr >> PAGE_SHIFT), prot))
2241 } while (pud++, addr = next, addr != end);
2246 * remap_pfn_range - remap kernel memory to userspace
2247 * @vma: user vma to map to
2248 * @addr: target user address to start at
2249 * @pfn: physical address of kernel memory
2250 * @size: size of map area
2251 * @prot: page protection flags for this mapping
2253 * Note: this is only safe if the mm semaphore is held when called.
2255 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2256 unsigned long pfn, unsigned long size, pgprot_t prot)
2260 unsigned long end = addr + PAGE_ALIGN(size);
2261 struct mm_struct *mm = vma->vm_mm;
2265 * Physically remapped pages are special. Tell the
2266 * rest of the world about it:
2267 * VM_IO tells people not to look at these pages
2268 * (accesses can have side effects).
2269 * VM_RESERVED is specified all over the place, because
2270 * in 2.4 it kept swapout's vma scan off this vma; but
2271 * in 2.6 the LRU scan won't even find its pages, so this
2272 * flag means no more than count its pages in reserved_vm,
2273 * and omit it from core dump, even when VM_IO turned off.
2274 * VM_PFNMAP tells the core MM that the base pages are just
2275 * raw PFN mappings, and do not have a "struct page" associated
2278 * There's a horrible special case to handle copy-on-write
2279 * behaviour that some programs depend on. We mark the "original"
2280 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2282 if (addr == vma->vm_start && end == vma->vm_end) {
2283 vma->vm_pgoff = pfn;
2284 vma->vm_flags |= VM_PFN_AT_MMAP;
2285 } else if (is_cow_mapping(vma->vm_flags))
2288 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2290 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2293 * To indicate that track_pfn related cleanup is not
2294 * needed from higher level routine calling unmap_vmas
2296 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2297 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2301 BUG_ON(addr >= end);
2302 pfn -= addr >> PAGE_SHIFT;
2303 pgd = pgd_offset(mm, addr);
2304 flush_cache_range(vma, addr, end);
2306 next = pgd_addr_end(addr, end);
2307 err = remap_pud_range(mm, pgd, addr, next,
2308 pfn + (addr >> PAGE_SHIFT), prot);
2311 } while (pgd++, addr = next, addr != end);
2314 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2318 EXPORT_SYMBOL(remap_pfn_range);
2320 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2321 unsigned long addr, unsigned long end,
2322 pte_fn_t fn, void *data)
2327 spinlock_t *uninitialized_var(ptl);
2329 pte = (mm == &init_mm) ?
2330 pte_alloc_kernel(pmd, addr) :
2331 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2335 BUG_ON(pmd_huge(*pmd));
2337 arch_enter_lazy_mmu_mode();
2339 token = pmd_pgtable(*pmd);
2342 err = fn(pte++, token, addr, data);
2345 } while (addr += PAGE_SIZE, addr != end);
2347 arch_leave_lazy_mmu_mode();
2350 pte_unmap_unlock(pte-1, ptl);
2354 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2355 unsigned long addr, unsigned long end,
2356 pte_fn_t fn, void *data)
2362 BUG_ON(pud_huge(*pud));
2364 pmd = pmd_alloc(mm, pud, addr);
2368 next = pmd_addr_end(addr, end);
2369 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2372 } while (pmd++, addr = next, addr != end);
2376 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2377 unsigned long addr, unsigned long end,
2378 pte_fn_t fn, void *data)
2384 pud = pud_alloc(mm, pgd, addr);
2388 next = pud_addr_end(addr, end);
2389 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2392 } while (pud++, addr = next, addr != end);
2397 * Scan a region of virtual memory, filling in page tables as necessary
2398 * and calling a provided function on each leaf page table.
2400 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2401 unsigned long size, pte_fn_t fn, void *data)
2405 unsigned long end = addr + size;
2408 BUG_ON(addr >= end);
2409 pgd = pgd_offset(mm, addr);
2411 next = pgd_addr_end(addr, end);
2412 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2415 } while (pgd++, addr = next, addr != end);
2419 EXPORT_SYMBOL_GPL(apply_to_page_range);
2422 * handle_pte_fault chooses page fault handler according to an entry
2423 * which was read non-atomically. Before making any commitment, on
2424 * those architectures or configurations (e.g. i386 with PAE) which
2425 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2426 * must check under lock before unmapping the pte and proceeding
2427 * (but do_wp_page is only called after already making such a check;
2428 * and do_anonymous_page can safely check later on).
2430 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2431 pte_t *page_table, pte_t orig_pte)
2434 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2435 if (sizeof(pte_t) > sizeof(unsigned long)) {
2436 spinlock_t *ptl = pte_lockptr(mm, pmd);
2438 same = pte_same(*page_table, orig_pte);
2442 pte_unmap(page_table);
2446 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2449 * If the source page was a PFN mapping, we don't have
2450 * a "struct page" for it. We do a best-effort copy by
2451 * just copying from the original user address. If that
2452 * fails, we just zero-fill it. Live with it.
2454 if (unlikely(!src)) {
2455 void *kaddr = kmap_atomic(dst, KM_USER0);
2456 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2459 * This really shouldn't fail, because the page is there
2460 * in the page tables. But it might just be unreadable,
2461 * in which case we just give up and fill the result with
2464 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2466 kunmap_atomic(kaddr, KM_USER0);
2467 flush_dcache_page(dst);
2469 copy_user_highpage(dst, src, va, vma);
2473 * This routine handles present pages, when users try to write
2474 * to a shared page. It is done by copying the page to a new address
2475 * and decrementing the shared-page counter for the old page.
2477 * Note that this routine assumes that the protection checks have been
2478 * done by the caller (the low-level page fault routine in most cases).
2479 * Thus we can safely just mark it writable once we've done any necessary
2482 * We also mark the page dirty at this point even though the page will
2483 * change only once the write actually happens. This avoids a few races,
2484 * and potentially makes it more efficient.
2486 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2487 * but allow concurrent faults), with pte both mapped and locked.
2488 * We return with mmap_sem still held, but pte unmapped and unlocked.
2490 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2491 unsigned long address, pte_t *page_table, pmd_t *pmd,
2492 spinlock_t *ptl, pte_t orig_pte)
2495 struct page *old_page, *new_page;
2498 int page_mkwrite = 0;
2499 struct page *dirty_page = NULL;
2501 old_page = vm_normal_page(vma, address, orig_pte);
2504 * VM_MIXEDMAP !pfn_valid() case
2506 * We should not cow pages in a shared writeable mapping.
2507 * Just mark the pages writable as we can't do any dirty
2508 * accounting on raw pfn maps.
2510 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2511 (VM_WRITE|VM_SHARED))
2517 * Take out anonymous pages first, anonymous shared vmas are
2518 * not dirty accountable.
2520 if (PageAnon(old_page) && !PageKsm(old_page)) {
2521 if (!trylock_page(old_page)) {
2522 page_cache_get(old_page);
2523 pte_unmap_unlock(page_table, ptl);
2524 lock_page(old_page);
2525 page_table = pte_offset_map_lock(mm, pmd, address,
2527 if (!pte_same(*page_table, orig_pte)) {
2528 unlock_page(old_page);
2531 page_cache_release(old_page);
2533 if (reuse_swap_page(old_page)) {
2535 * The page is all ours. Move it to our anon_vma so
2536 * the rmap code will not search our parent or siblings.
2537 * Protected against the rmap code by the page lock.
2539 page_move_anon_rmap(old_page, vma, address);
2540 unlock_page(old_page);
2543 unlock_page(old_page);
2544 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2545 (VM_WRITE|VM_SHARED))) {
2547 * Only catch write-faults on shared writable pages,
2548 * read-only shared pages can get COWed by
2549 * get_user_pages(.write=1, .force=1).
2551 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2552 struct vm_fault vmf;
2555 vmf.virtual_address = (void __user *)(address &
2557 vmf.pgoff = old_page->index;
2558 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2559 vmf.page = old_page;
2562 * Notify the address space that the page is about to
2563 * become writable so that it can prohibit this or wait
2564 * for the page to get into an appropriate state.
2566 * We do this without the lock held, so that it can
2567 * sleep if it needs to.
2569 page_cache_get(old_page);
2570 pte_unmap_unlock(page_table, ptl);
2572 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2574 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2576 goto unwritable_page;
2578 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2579 lock_page(old_page);
2580 if (!old_page->mapping) {
2581 ret = 0; /* retry the fault */
2582 unlock_page(old_page);
2583 goto unwritable_page;
2586 VM_BUG_ON(!PageLocked(old_page));
2589 * Since we dropped the lock we need to revalidate
2590 * the PTE as someone else may have changed it. If
2591 * they did, we just return, as we can count on the
2592 * MMU to tell us if they didn't also make it writable.
2594 page_table = pte_offset_map_lock(mm, pmd, address,
2596 if (!pte_same(*page_table, orig_pte)) {
2597 unlock_page(old_page);
2603 dirty_page = old_page;
2604 get_page(dirty_page);
2607 flush_cache_page(vma, address, pte_pfn(orig_pte));
2608 entry = pte_mkyoung(orig_pte);
2609 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2610 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2611 update_mmu_cache(vma, address, page_table);
2612 pte_unmap_unlock(page_table, ptl);
2613 ret |= VM_FAULT_WRITE;
2619 * Yes, Virginia, this is actually required to prevent a race
2620 * with clear_page_dirty_for_io() from clearing the page dirty
2621 * bit after it clear all dirty ptes, but before a racing
2622 * do_wp_page installs a dirty pte.
2624 * __do_fault is protected similarly.
2626 if (!page_mkwrite) {
2627 wait_on_page_locked(dirty_page);
2628 set_page_dirty_balance(dirty_page, page_mkwrite);
2630 put_page(dirty_page);
2632 struct address_space *mapping = dirty_page->mapping;
2634 set_page_dirty(dirty_page);
2635 unlock_page(dirty_page);
2636 page_cache_release(dirty_page);
2639 * Some device drivers do not set page.mapping
2640 * but still dirty their pages
2642 balance_dirty_pages_ratelimited(mapping);
2646 /* file_update_time outside page_lock */
2648 file_update_time(vma->vm_file);
2654 * Ok, we need to copy. Oh, well..
2656 page_cache_get(old_page);
2658 pte_unmap_unlock(page_table, ptl);
2660 if (unlikely(anon_vma_prepare(vma)))
2663 if (is_zero_pfn(pte_pfn(orig_pte))) {
2664 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2668 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2671 cow_user_page(new_page, old_page, address, vma);
2673 __SetPageUptodate(new_page);
2675 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2679 * Re-check the pte - we dropped the lock
2681 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2682 if (likely(pte_same(*page_table, orig_pte))) {
2684 if (!PageAnon(old_page)) {
2685 dec_mm_counter_fast(mm, MM_FILEPAGES);
2686 inc_mm_counter_fast(mm, MM_ANONPAGES);
2689 inc_mm_counter_fast(mm, MM_ANONPAGES);
2690 flush_cache_page(vma, address, pte_pfn(orig_pte));
2691 entry = mk_pte(new_page, vma->vm_page_prot);
2692 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2694 * Clear the pte entry and flush it first, before updating the
2695 * pte with the new entry. This will avoid a race condition
2696 * seen in the presence of one thread doing SMC and another
2699 ptep_clear_flush(vma, address, page_table);
2700 page_add_new_anon_rmap(new_page, vma, address);
2702 * We call the notify macro here because, when using secondary
2703 * mmu page tables (such as kvm shadow page tables), we want the
2704 * new page to be mapped directly into the secondary page table.
2706 set_pte_at_notify(mm, address, page_table, entry);
2707 update_mmu_cache(vma, address, page_table);
2710 * Only after switching the pte to the new page may
2711 * we remove the mapcount here. Otherwise another
2712 * process may come and find the rmap count decremented
2713 * before the pte is switched to the new page, and
2714 * "reuse" the old page writing into it while our pte
2715 * here still points into it and can be read by other
2718 * The critical issue is to order this
2719 * page_remove_rmap with the ptp_clear_flush above.
2720 * Those stores are ordered by (if nothing else,)
2721 * the barrier present in the atomic_add_negative
2722 * in page_remove_rmap.
2724 * Then the TLB flush in ptep_clear_flush ensures that
2725 * no process can access the old page before the
2726 * decremented mapcount is visible. And the old page
2727 * cannot be reused until after the decremented
2728 * mapcount is visible. So transitively, TLBs to
2729 * old page will be flushed before it can be reused.
2731 page_remove_rmap(old_page);
2734 /* Free the old page.. */
2735 new_page = old_page;
2736 ret |= VM_FAULT_WRITE;
2738 mem_cgroup_uncharge_page(new_page);
2741 page_cache_release(new_page);
2743 pte_unmap_unlock(page_table, ptl);
2746 * Don't let another task, with possibly unlocked vma,
2747 * keep the mlocked page.
2749 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2750 lock_page(old_page); /* LRU manipulation */
2751 munlock_vma_page(old_page);
2752 unlock_page(old_page);
2754 page_cache_release(old_page);
2758 page_cache_release(new_page);
2762 unlock_page(old_page);
2763 page_cache_release(old_page);
2765 page_cache_release(old_page);
2767 return VM_FAULT_OOM;
2770 page_cache_release(old_page);
2774 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2775 unsigned long start_addr, unsigned long end_addr,
2776 struct zap_details *details)
2778 zap_page_range(vma, start_addr, end_addr - start_addr, details);
2781 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2782 struct zap_details *details)
2784 struct vm_area_struct *vma;
2785 struct prio_tree_iter iter;
2786 pgoff_t vba, vea, zba, zea;
2788 vma_prio_tree_foreach(vma, &iter, root,
2789 details->first_index, details->last_index) {
2791 vba = vma->vm_pgoff;
2792 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2793 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2794 zba = details->first_index;
2797 zea = details->last_index;
2801 unmap_mapping_range_vma(vma,
2802 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2803 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2808 static inline void unmap_mapping_range_list(struct list_head *head,
2809 struct zap_details *details)
2811 struct vm_area_struct *vma;
2814 * In nonlinear VMAs there is no correspondence between virtual address
2815 * offset and file offset. So we must perform an exhaustive search
2816 * across *all* the pages in each nonlinear VMA, not just the pages
2817 * whose virtual address lies outside the file truncation point.
2819 list_for_each_entry(vma, head, shared.vm_set.list) {
2820 details->nonlinear_vma = vma;
2821 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2826 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2827 * @mapping: the address space containing mmaps to be unmapped.
2828 * @holebegin: byte in first page to unmap, relative to the start of
2829 * the underlying file. This will be rounded down to a PAGE_SIZE
2830 * boundary. Note that this is different from truncate_pagecache(), which
2831 * must keep the partial page. In contrast, we must get rid of
2833 * @holelen: size of prospective hole in bytes. This will be rounded
2834 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2836 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2837 * but 0 when invalidating pagecache, don't throw away private data.
2839 void unmap_mapping_range(struct address_space *mapping,
2840 loff_t const holebegin, loff_t const holelen, int even_cows)
2842 struct zap_details details;
2843 pgoff_t hba = holebegin >> PAGE_SHIFT;
2844 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2846 /* Check for overflow. */
2847 if (sizeof(holelen) > sizeof(hlen)) {
2849 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2850 if (holeend & ~(long long)ULONG_MAX)
2851 hlen = ULONG_MAX - hba + 1;
2854 details.check_mapping = even_cows? NULL: mapping;
2855 details.nonlinear_vma = NULL;
2856 details.first_index = hba;
2857 details.last_index = hba + hlen - 1;
2858 if (details.last_index < details.first_index)
2859 details.last_index = ULONG_MAX;
2862 mutex_lock(&mapping->i_mmap_mutex);
2863 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2864 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2865 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2866 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2867 mutex_unlock(&mapping->i_mmap_mutex);
2869 EXPORT_SYMBOL(unmap_mapping_range);
2872 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2873 * but allow concurrent faults), and pte mapped but not yet locked.
2874 * We return with mmap_sem still held, but pte unmapped and unlocked.
2876 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2877 unsigned long address, pte_t *page_table, pmd_t *pmd,
2878 unsigned int flags, pte_t orig_pte)
2881 struct page *page, *swapcache = NULL;
2885 struct mem_cgroup *ptr;
2889 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2892 entry = pte_to_swp_entry(orig_pte);
2893 if (unlikely(non_swap_entry(entry))) {
2894 if (is_migration_entry(entry)) {
2895 migration_entry_wait(mm, pmd, address);
2896 } else if (is_hwpoison_entry(entry)) {
2897 ret = VM_FAULT_HWPOISON;
2899 print_bad_pte(vma, address, orig_pte, NULL);
2900 ret = VM_FAULT_SIGBUS;
2904 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2905 page = lookup_swap_cache(entry);
2907 grab_swap_token(mm); /* Contend for token _before_ read-in */
2908 page = swapin_readahead(entry,
2909 GFP_HIGHUSER_MOVABLE, vma, address);
2912 * Back out if somebody else faulted in this pte
2913 * while we released the pte lock.
2915 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2916 if (likely(pte_same(*page_table, orig_pte)))
2918 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2922 /* Had to read the page from swap area: Major fault */
2923 ret = VM_FAULT_MAJOR;
2924 count_vm_event(PGMAJFAULT);
2925 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2926 } else if (PageHWPoison(page)) {
2928 * hwpoisoned dirty swapcache pages are kept for killing
2929 * owner processes (which may be unknown at hwpoison time)
2931 ret = VM_FAULT_HWPOISON;
2932 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2936 locked = lock_page_or_retry(page, mm, flags);
2937 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2939 ret |= VM_FAULT_RETRY;
2944 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2945 * release the swapcache from under us. The page pin, and pte_same
2946 * test below, are not enough to exclude that. Even if it is still
2947 * swapcache, we need to check that the page's swap has not changed.
2949 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2952 if (ksm_might_need_to_copy(page, vma, address)) {
2954 page = ksm_does_need_to_copy(page, vma, address);
2956 if (unlikely(!page)) {
2964 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2970 * Back out if somebody else already faulted in this pte.
2972 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2973 if (unlikely(!pte_same(*page_table, orig_pte)))
2976 if (unlikely(!PageUptodate(page))) {
2977 ret = VM_FAULT_SIGBUS;
2982 * The page isn't present yet, go ahead with the fault.
2984 * Be careful about the sequence of operations here.
2985 * To get its accounting right, reuse_swap_page() must be called
2986 * while the page is counted on swap but not yet in mapcount i.e.
2987 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2988 * must be called after the swap_free(), or it will never succeed.
2989 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2990 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2991 * in page->private. In this case, a record in swap_cgroup is silently
2992 * discarded at swap_free().
2995 inc_mm_counter_fast(mm, MM_ANONPAGES);
2996 dec_mm_counter_fast(mm, MM_SWAPENTS);
2997 pte = mk_pte(page, vma->vm_page_prot);
2998 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2999 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3000 flags &= ~FAULT_FLAG_WRITE;
3001 ret |= VM_FAULT_WRITE;
3004 flush_icache_page(vma, page);
3005 set_pte_at(mm, address, page_table, pte);
3006 do_page_add_anon_rmap(page, vma, address, exclusive);
3007 /* It's better to call commit-charge after rmap is established */
3008 mem_cgroup_commit_charge_swapin(page, ptr);
3011 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3012 try_to_free_swap(page);
3016 * Hold the lock to avoid the swap entry to be reused
3017 * until we take the PT lock for the pte_same() check
3018 * (to avoid false positives from pte_same). For
3019 * further safety release the lock after the swap_free
3020 * so that the swap count won't change under a
3021 * parallel locked swapcache.
3023 unlock_page(swapcache);
3024 page_cache_release(swapcache);
3027 if (flags & FAULT_FLAG_WRITE) {
3028 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3029 if (ret & VM_FAULT_ERROR)
3030 ret &= VM_FAULT_ERROR;
3034 /* No need to invalidate - it was non-present before */
3035 update_mmu_cache(vma, address, page_table);
3037 pte_unmap_unlock(page_table, ptl);
3041 mem_cgroup_cancel_charge_swapin(ptr);
3042 pte_unmap_unlock(page_table, ptl);
3046 page_cache_release(page);
3048 unlock_page(swapcache);
3049 page_cache_release(swapcache);
3055 * This is like a special single-page "expand_{down|up}wards()",
3056 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3057 * doesn't hit another vma.
3059 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3061 address &= PAGE_MASK;
3062 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3063 struct vm_area_struct *prev = vma->vm_prev;
3066 * Is there a mapping abutting this one below?
3068 * That's only ok if it's the same stack mapping
3069 * that has gotten split..
3071 if (prev && prev->vm_end == address)
3072 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3074 expand_downwards(vma, address - PAGE_SIZE);
3076 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3077 struct vm_area_struct *next = vma->vm_next;
3079 /* As VM_GROWSDOWN but s/below/above/ */
3080 if (next && next->vm_start == address + PAGE_SIZE)
3081 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3083 expand_upwards(vma, address + PAGE_SIZE);
3089 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3090 * but allow concurrent faults), and pte mapped but not yet locked.
3091 * We return with mmap_sem still held, but pte unmapped and unlocked.
3093 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3094 unsigned long address, pte_t *page_table, pmd_t *pmd,
3101 pte_unmap(page_table);
3103 /* Check if we need to add a guard page to the stack */
3104 if (check_stack_guard_page(vma, address) < 0)
3105 return VM_FAULT_SIGBUS;
3107 /* Use the zero-page for reads */
3108 if (!(flags & FAULT_FLAG_WRITE)) {
3109 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3110 vma->vm_page_prot));
3111 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3112 if (!pte_none(*page_table))
3117 /* Allocate our own private page. */
3118 if (unlikely(anon_vma_prepare(vma)))
3120 page = alloc_zeroed_user_highpage_movable(vma, address);
3123 __SetPageUptodate(page);
3125 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3128 entry = mk_pte(page, vma->vm_page_prot);
3129 if (vma->vm_flags & VM_WRITE)
3130 entry = pte_mkwrite(pte_mkdirty(entry));
3132 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3133 if (!pte_none(*page_table))
3136 inc_mm_counter_fast(mm, MM_ANONPAGES);
3137 page_add_new_anon_rmap(page, vma, address);
3139 set_pte_at(mm, address, page_table, entry);
3141 /* No need to invalidate - it was non-present before */
3142 update_mmu_cache(vma, address, page_table);
3144 pte_unmap_unlock(page_table, ptl);
3147 mem_cgroup_uncharge_page(page);
3148 page_cache_release(page);
3151 page_cache_release(page);
3153 return VM_FAULT_OOM;
3157 * __do_fault() tries to create a new page mapping. It aggressively
3158 * tries to share with existing pages, but makes a separate copy if
3159 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3160 * the next page fault.
3162 * As this is called only for pages that do not currently exist, we
3163 * do not need to flush old virtual caches or the TLB.
3165 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3166 * but allow concurrent faults), and pte neither mapped nor locked.
3167 * We return with mmap_sem still held, but pte unmapped and unlocked.
3169 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3170 unsigned long address, pmd_t *pmd,
3171 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3179 struct page *dirty_page = NULL;
3180 struct vm_fault vmf;
3182 int page_mkwrite = 0;
3184 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3189 ret = vma->vm_ops->fault(vma, &vmf);
3190 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3194 if (unlikely(PageHWPoison(vmf.page))) {
3195 if (ret & VM_FAULT_LOCKED)
3196 unlock_page(vmf.page);
3197 return VM_FAULT_HWPOISON;
3201 * For consistency in subsequent calls, make the faulted page always
3204 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3205 lock_page(vmf.page);
3207 VM_BUG_ON(!PageLocked(vmf.page));
3210 * Should we do an early C-O-W break?
3213 if (flags & FAULT_FLAG_WRITE) {
3214 if (!(vma->vm_flags & VM_SHARED)) {
3216 if (unlikely(anon_vma_prepare(vma))) {
3220 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE,
3226 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL)) {
3228 page_cache_release(page);
3232 copy_user_highpage(page, vmf.page, address, vma);
3233 __SetPageUptodate(page);
3236 * If the page will be shareable, see if the backing
3237 * address space wants to know that the page is about
3238 * to become writable
3240 if (vma->vm_ops->page_mkwrite) {
3244 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3245 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3247 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3249 goto unwritable_page;
3251 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3253 if (!page->mapping) {
3254 ret = 0; /* retry the fault */
3256 goto unwritable_page;
3259 VM_BUG_ON(!PageLocked(page));
3266 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3269 * This silly early PAGE_DIRTY setting removes a race
3270 * due to the bad i386 page protection. But it's valid
3271 * for other architectures too.
3273 * Note that if FAULT_FLAG_WRITE is set, we either now have
3274 * an exclusive copy of the page, or this is a shared mapping,
3275 * so we can make it writable and dirty to avoid having to
3276 * handle that later.
3278 /* Only go through if we didn't race with anybody else... */
3279 if (likely(pte_same(*page_table, orig_pte))) {
3280 flush_icache_page(vma, page);
3281 entry = mk_pte(page, vma->vm_page_prot);
3282 if (flags & FAULT_FLAG_WRITE)
3283 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3285 inc_mm_counter_fast(mm, MM_ANONPAGES);
3286 page_add_new_anon_rmap(page, vma, address);
3288 inc_mm_counter_fast(mm, MM_FILEPAGES);
3289 page_add_file_rmap(page);
3290 if (flags & FAULT_FLAG_WRITE) {
3292 get_page(dirty_page);
3295 set_pte_at(mm, address, page_table, entry);
3297 /* no need to invalidate: a not-present page won't be cached */
3298 update_mmu_cache(vma, address, page_table);
3301 mem_cgroup_uncharge_page(page);
3303 page_cache_release(page);
3305 anon = 1; /* no anon but release faulted_page */
3308 pte_unmap_unlock(page_table, ptl);
3312 struct address_space *mapping = page->mapping;
3314 if (set_page_dirty(dirty_page))
3316 unlock_page(dirty_page);
3317 put_page(dirty_page);
3318 if (page_mkwrite && mapping) {
3320 * Some device drivers do not set page.mapping but still
3323 balance_dirty_pages_ratelimited(mapping);
3326 /* file_update_time outside page_lock */
3328 file_update_time(vma->vm_file);
3330 unlock_page(vmf.page);
3332 page_cache_release(vmf.page);
3338 page_cache_release(page);
3342 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3343 unsigned long address, pte_t *page_table, pmd_t *pmd,
3344 unsigned int flags, pte_t orig_pte)
3346 pgoff_t pgoff = (((address & PAGE_MASK)
3347 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3349 pte_unmap(page_table);
3350 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3354 * Fault of a previously existing named mapping. Repopulate the pte
3355 * from the encoded file_pte if possible. This enables swappable
3358 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3359 * but allow concurrent faults), and pte mapped but not yet locked.
3360 * We return with mmap_sem still held, but pte unmapped and unlocked.
3362 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3363 unsigned long address, pte_t *page_table, pmd_t *pmd,
3364 unsigned int flags, pte_t orig_pte)
3368 flags |= FAULT_FLAG_NONLINEAR;
3370 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3373 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3375 * Page table corrupted: show pte and kill process.
3377 print_bad_pte(vma, address, orig_pte, NULL);
3378 return VM_FAULT_SIGBUS;
3381 pgoff = pte_to_pgoff(orig_pte);
3382 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3386 * These routines also need to handle stuff like marking pages dirty
3387 * and/or accessed for architectures that don't do it in hardware (most
3388 * RISC architectures). The early dirtying is also good on the i386.
3390 * There is also a hook called "update_mmu_cache()" that architectures
3391 * with external mmu caches can use to update those (ie the Sparc or
3392 * PowerPC hashed page tables that act as extended TLBs).
3394 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3395 * but allow concurrent faults), and pte mapped but not yet locked.
3396 * We return with mmap_sem still held, but pte unmapped and unlocked.
3398 int handle_pte_fault(struct mm_struct *mm,
3399 struct vm_area_struct *vma, unsigned long address,
3400 pte_t *pte, pmd_t *pmd, unsigned int flags)
3406 if (!pte_present(entry)) {
3407 if (pte_none(entry)) {
3409 if (likely(vma->vm_ops->fault))
3410 return do_linear_fault(mm, vma, address,
3411 pte, pmd, flags, entry);
3413 return do_anonymous_page(mm, vma, address,
3416 if (pte_file(entry))
3417 return do_nonlinear_fault(mm, vma, address,
3418 pte, pmd, flags, entry);
3419 return do_swap_page(mm, vma, address,
3420 pte, pmd, flags, entry);
3423 ptl = pte_lockptr(mm, pmd);
3425 if (unlikely(!pte_same(*pte, entry)))
3427 if (flags & FAULT_FLAG_WRITE) {
3428 if (!pte_write(entry))
3429 return do_wp_page(mm, vma, address,
3430 pte, pmd, ptl, entry);
3431 entry = pte_mkdirty(entry);
3433 entry = pte_mkyoung(entry);
3434 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3435 update_mmu_cache(vma, address, pte);
3438 * This is needed only for protection faults but the arch code
3439 * is not yet telling us if this is a protection fault or not.
3440 * This still avoids useless tlb flushes for .text page faults
3443 if (flags & FAULT_FLAG_WRITE)
3444 flush_tlb_fix_spurious_fault(vma, address);
3447 pte_unmap_unlock(pte, ptl);
3452 * By the time we get here, we already hold the mm semaphore
3454 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3455 unsigned long address, unsigned int flags)
3462 __set_current_state(TASK_RUNNING);
3464 count_vm_event(PGFAULT);
3465 mem_cgroup_count_vm_event(mm, PGFAULT);
3467 /* do counter updates before entering really critical section. */
3468 check_sync_rss_stat(current);
3470 if (unlikely(is_vm_hugetlb_page(vma)))
3471 return hugetlb_fault(mm, vma, address, flags);
3474 pgd = pgd_offset(mm, address);
3475 pud = pud_alloc(mm, pgd, address);
3477 return VM_FAULT_OOM;
3478 pmd = pmd_alloc(mm, pud, address);
3480 return VM_FAULT_OOM;
3481 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3483 return do_huge_pmd_anonymous_page(mm, vma, address,
3486 pmd_t orig_pmd = *pmd;
3490 if (pmd_trans_huge(orig_pmd)) {
3491 if (flags & FAULT_FLAG_WRITE &&
3492 !pmd_write(orig_pmd) &&
3493 !pmd_trans_splitting(orig_pmd)) {
3494 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3497 * If COW results in an oom, the huge pmd will
3498 * have been split, so retry the fault on the
3499 * pte for a smaller charge.
3501 if (unlikely(ret & VM_FAULT_OOM))
3510 * Use __pte_alloc instead of pte_alloc_map, because we can't
3511 * run pte_offset_map on the pmd, if an huge pmd could
3512 * materialize from under us from a different thread.
3514 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3515 return VM_FAULT_OOM;
3516 /* if an huge pmd materialized from under us just retry later */
3517 if (unlikely(pmd_trans_huge(*pmd)))
3520 * A regular pmd is established and it can't morph into a huge pmd
3521 * from under us anymore at this point because we hold the mmap_sem
3522 * read mode and khugepaged takes it in write mode. So now it's
3523 * safe to run pte_offset_map().
3525 pte = pte_offset_map(pmd, address);
3527 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3530 #ifndef __PAGETABLE_PUD_FOLDED
3532 * Allocate page upper directory.
3533 * We've already handled the fast-path in-line.
3535 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3537 pud_t *new = pud_alloc_one(mm, address);
3541 smp_wmb(); /* See comment in __pte_alloc */
3543 spin_lock(&mm->page_table_lock);
3544 if (pgd_present(*pgd)) /* Another has populated it */
3547 pgd_populate(mm, pgd, new);
3548 spin_unlock(&mm->page_table_lock);
3551 #endif /* __PAGETABLE_PUD_FOLDED */
3553 #ifndef __PAGETABLE_PMD_FOLDED
3555 * Allocate page middle directory.
3556 * We've already handled the fast-path in-line.
3558 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3560 pmd_t *new = pmd_alloc_one(mm, address);
3564 smp_wmb(); /* See comment in __pte_alloc */
3566 spin_lock(&mm->page_table_lock);
3567 #ifndef __ARCH_HAS_4LEVEL_HACK
3568 if (pud_present(*pud)) /* Another has populated it */
3571 pud_populate(mm, pud, new);
3573 if (pgd_present(*pud)) /* Another has populated it */
3576 pgd_populate(mm, pud, new);
3577 #endif /* __ARCH_HAS_4LEVEL_HACK */
3578 spin_unlock(&mm->page_table_lock);
3581 #endif /* __PAGETABLE_PMD_FOLDED */
3583 int make_pages_present(unsigned long addr, unsigned long end)
3585 int ret, len, write;
3586 struct vm_area_struct * vma;
3588 vma = find_vma(current->mm, addr);
3592 * We want to touch writable mappings with a write fault in order
3593 * to break COW, except for shared mappings because these don't COW
3594 * and we would not want to dirty them for nothing.
3596 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3597 BUG_ON(addr >= end);
3598 BUG_ON(end > vma->vm_end);
3599 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3600 ret = get_user_pages(current, current->mm, addr,
3601 len, write, 0, NULL, NULL);
3604 return ret == len ? 0 : -EFAULT;
3607 #if !defined(__HAVE_ARCH_GATE_AREA)
3609 #if defined(AT_SYSINFO_EHDR)
3610 static struct vm_area_struct gate_vma;
3612 static int __init gate_vma_init(void)
3614 gate_vma.vm_mm = NULL;
3615 gate_vma.vm_start = FIXADDR_USER_START;
3616 gate_vma.vm_end = FIXADDR_USER_END;
3617 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3618 gate_vma.vm_page_prot = __P101;
3620 * Make sure the vDSO gets into every core dump.
3621 * Dumping its contents makes post-mortem fully interpretable later
3622 * without matching up the same kernel and hardware config to see
3623 * what PC values meant.
3625 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3628 __initcall(gate_vma_init);
3631 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3633 #ifdef AT_SYSINFO_EHDR
3640 int in_gate_area_no_mm(unsigned long addr)
3642 #ifdef AT_SYSINFO_EHDR
3643 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3649 #endif /* __HAVE_ARCH_GATE_AREA */
3651 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3652 pte_t **ptepp, spinlock_t **ptlp)
3659 pgd = pgd_offset(mm, address);
3660 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3663 pud = pud_offset(pgd, address);
3664 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3667 pmd = pmd_offset(pud, address);
3668 VM_BUG_ON(pmd_trans_huge(*pmd));
3669 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3672 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3676 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3679 if (!pte_present(*ptep))
3684 pte_unmap_unlock(ptep, *ptlp);
3689 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3690 pte_t **ptepp, spinlock_t **ptlp)
3694 /* (void) is needed to make gcc happy */
3695 (void) __cond_lock(*ptlp,
3696 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3701 * follow_pfn - look up PFN at a user virtual address
3702 * @vma: memory mapping
3703 * @address: user virtual address
3704 * @pfn: location to store found PFN
3706 * Only IO mappings and raw PFN mappings are allowed.
3708 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3710 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3717 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3720 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3723 *pfn = pte_pfn(*ptep);
3724 pte_unmap_unlock(ptep, ptl);
3727 EXPORT_SYMBOL(follow_pfn);
3729 #ifdef CONFIG_HAVE_IOREMAP_PROT
3730 int follow_phys(struct vm_area_struct *vma,
3731 unsigned long address, unsigned int flags,
3732 unsigned long *prot, resource_size_t *phys)
3738 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3741 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3745 if ((flags & FOLL_WRITE) && !pte_write(pte))
3748 *prot = pgprot_val(pte_pgprot(pte));
3749 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3753 pte_unmap_unlock(ptep, ptl);
3758 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3759 void *buf, int len, int write)
3761 resource_size_t phys_addr;
3762 unsigned long prot = 0;
3763 void __iomem *maddr;
3764 int offset = addr & (PAGE_SIZE-1);
3766 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3769 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3771 memcpy_toio(maddr + offset, buf, len);
3773 memcpy_fromio(buf, maddr + offset, len);
3781 * Access another process' address space as given in mm. If non-NULL, use the
3782 * given task for page fault accounting.
3784 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3785 unsigned long addr, void *buf, int len, int write)
3787 struct vm_area_struct *vma;
3788 void *old_buf = buf;
3790 down_read(&mm->mmap_sem);
3791 /* ignore errors, just check how much was successfully transferred */
3793 int bytes, ret, offset;
3795 struct page *page = NULL;
3797 ret = get_user_pages(tsk, mm, addr, 1,
3798 write, 1, &page, &vma);
3801 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3802 * we can access using slightly different code.
3804 #ifdef CONFIG_HAVE_IOREMAP_PROT
3805 vma = find_vma(mm, addr);
3806 if (!vma || vma->vm_start > addr)
3808 if (vma->vm_ops && vma->vm_ops->access)
3809 ret = vma->vm_ops->access(vma, addr, buf,
3817 offset = addr & (PAGE_SIZE-1);
3818 if (bytes > PAGE_SIZE-offset)
3819 bytes = PAGE_SIZE-offset;
3823 copy_to_user_page(vma, page, addr,
3824 maddr + offset, buf, bytes);
3825 set_page_dirty_lock(page);
3827 copy_from_user_page(vma, page, addr,
3828 buf, maddr + offset, bytes);
3831 page_cache_release(page);
3837 up_read(&mm->mmap_sem);
3839 return buf - old_buf;
3843 * access_remote_vm - access another process' address space
3844 * @mm: the mm_struct of the target address space
3845 * @addr: start address to access
3846 * @buf: source or destination buffer
3847 * @len: number of bytes to transfer
3848 * @write: whether the access is a write
3850 * The caller must hold a reference on @mm.
3852 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3853 void *buf, int len, int write)
3855 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3859 * Access another process' address space.
3860 * Source/target buffer must be kernel space,
3861 * Do not walk the page table directly, use get_user_pages
3863 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3864 void *buf, int len, int write)
3866 struct mm_struct *mm;
3869 mm = get_task_mm(tsk);
3873 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3880 * Print the name of a VMA.
3882 void print_vma_addr(char *prefix, unsigned long ip)
3884 struct mm_struct *mm = current->mm;
3885 struct vm_area_struct *vma;
3888 * Do not print if we are in atomic
3889 * contexts (in exception stacks, etc.):
3891 if (preempt_count())
3894 down_read(&mm->mmap_sem);
3895 vma = find_vma(mm, ip);
3896 if (vma && vma->vm_file) {
3897 struct file *f = vma->vm_file;
3898 char *buf = (char *)__get_free_page(GFP_KERNEL);
3902 p = d_path(&f->f_path, buf, PAGE_SIZE);
3905 s = strrchr(p, '/');
3908 printk("%s%s[%lx+%lx]", prefix, p,
3910 vma->vm_end - vma->vm_start);
3911 free_page((unsigned long)buf);
3914 up_read(¤t->mm->mmap_sem);
3917 #ifdef CONFIG_PROVE_LOCKING
3918 void might_fault(void)
3921 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3922 * holding the mmap_sem, this is safe because kernel memory doesn't
3923 * get paged out, therefore we'll never actually fault, and the
3924 * below annotations will generate false positives.
3926 if (segment_eq(get_fs(), KERNEL_DS))
3931 * it would be nicer only to annotate paths which are not under
3932 * pagefault_disable, however that requires a larger audit and
3933 * providing helpers like get_user_atomic.
3935 if (!in_atomic() && current->mm)
3936 might_lock_read(¤t->mm->mmap_sem);
3938 EXPORT_SYMBOL(might_fault);
3941 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3942 static void clear_gigantic_page(struct page *page,
3944 unsigned int pages_per_huge_page)
3947 struct page *p = page;
3950 for (i = 0; i < pages_per_huge_page;
3951 i++, p = mem_map_next(p, page, i)) {
3953 clear_user_highpage(p, addr + i * PAGE_SIZE);
3956 void clear_huge_page(struct page *page,
3957 unsigned long addr, unsigned int pages_per_huge_page)
3961 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3962 clear_gigantic_page(page, addr, pages_per_huge_page);
3967 for (i = 0; i < pages_per_huge_page; i++) {
3969 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3973 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3975 struct vm_area_struct *vma,
3976 unsigned int pages_per_huge_page)
3979 struct page *dst_base = dst;
3980 struct page *src_base = src;
3982 for (i = 0; i < pages_per_huge_page; ) {
3984 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3987 dst = mem_map_next(dst, dst_base, i);
3988 src = mem_map_next(src, src_base, i);
3992 void copy_user_huge_page(struct page *dst, struct page *src,
3993 unsigned long addr, struct vm_area_struct *vma,
3994 unsigned int pages_per_huge_page)
3998 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3999 copy_user_gigantic_page(dst, src, addr, vma,
4000 pages_per_huge_page);
4005 for (i = 0; i < pages_per_huge_page; i++) {
4007 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4010 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */