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/export.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>
60 #include <linux/migrate.h>
63 #include <asm/pgalloc.h>
64 #include <asm/uaccess.h>
66 #include <asm/tlbflush.h>
67 #include <asm/pgtable.h>
71 #ifdef LAST_CPU_NOT_IN_PAGE_FLAGS
72 #warning Unfortunate NUMA config, growing page-frame for last_cpu.
75 #ifndef CONFIG_NEED_MULTIPLE_NODES
76 /* use the per-pgdat data instead for discontigmem - mbligh */
77 unsigned long max_mapnr;
80 EXPORT_SYMBOL(max_mapnr);
81 EXPORT_SYMBOL(mem_map);
84 unsigned long num_physpages;
86 * A number of key systems in x86 including ioremap() rely on the assumption
87 * that high_memory defines the upper bound on direct map memory, then end
88 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
94 EXPORT_SYMBOL(num_physpages);
95 EXPORT_SYMBOL(high_memory);
98 * Randomize the address space (stacks, mmaps, brk, etc.).
100 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
101 * as ancient (libc5 based) binaries can segfault. )
103 int randomize_va_space __read_mostly =
104 #ifdef CONFIG_COMPAT_BRK
110 static int __init disable_randmaps(char *s)
112 randomize_va_space = 0;
115 __setup("norandmaps", disable_randmaps);
117 unsigned long zero_pfn __read_mostly;
118 unsigned long highest_memmap_pfn __read_mostly;
121 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 static int __init init_zero_pfn(void)
125 zero_pfn = page_to_pfn(ZERO_PAGE(0));
128 core_initcall(init_zero_pfn);
131 #if defined(SPLIT_RSS_COUNTING)
133 void sync_mm_rss(struct mm_struct *mm)
137 for (i = 0; i < NR_MM_COUNTERS; i++) {
138 if (current->rss_stat.count[i]) {
139 add_mm_counter(mm, i, current->rss_stat.count[i]);
140 current->rss_stat.count[i] = 0;
143 current->rss_stat.events = 0;
146 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148 struct task_struct *task = current;
150 if (likely(task->mm == mm))
151 task->rss_stat.count[member] += val;
153 add_mm_counter(mm, member, val);
155 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
156 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158 /* sync counter once per 64 page faults */
159 #define TASK_RSS_EVENTS_THRESH (64)
160 static void check_sync_rss_stat(struct task_struct *task)
162 if (unlikely(task != current))
164 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
165 sync_mm_rss(task->mm);
167 #else /* SPLIT_RSS_COUNTING */
169 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
170 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172 static void check_sync_rss_stat(struct task_struct *task)
176 #endif /* SPLIT_RSS_COUNTING */
178 #ifdef HAVE_GENERIC_MMU_GATHER
180 static int tlb_next_batch(struct mmu_gather *tlb)
182 struct mmu_gather_batch *batch;
186 tlb->active = batch->next;
190 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
196 batch->max = MAX_GATHER_BATCH;
198 tlb->active->next = batch;
205 * Called to initialize an (on-stack) mmu_gather structure for page-table
206 * tear-down from @mm. The @fullmm argument is used when @mm is without
207 * users and we're going to destroy the full address space (exit/execve).
209 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
213 tlb->fullmm = fullmm;
217 tlb->fast_mode = (num_possible_cpus() == 1);
218 tlb->local.next = NULL;
220 tlb->local.max = ARRAY_SIZE(tlb->__pages);
221 tlb->active = &tlb->local;
223 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
228 void tlb_flush_mmu(struct mmu_gather *tlb)
230 struct mmu_gather_batch *batch;
232 if (!tlb->need_flush)
236 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237 tlb_table_flush(tlb);
240 if (tlb_fast_mode(tlb))
243 for (batch = &tlb->local; batch; batch = batch->next) {
244 free_pages_and_swap_cache(batch->pages, batch->nr);
247 tlb->active = &tlb->local;
251 * Called at the end of the shootdown operation to free up any resources
252 * that were required.
254 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
256 struct mmu_gather_batch *batch, *next;
262 /* keep the page table cache within bounds */
265 for (batch = tlb->local.next; batch; batch = next) {
267 free_pages((unsigned long)batch, 0);
269 tlb->local.next = NULL;
273 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
274 * handling the additional races in SMP caused by other CPUs caching valid
275 * mappings in their TLBs. Returns the number of free page slots left.
276 * When out of page slots we must call tlb_flush_mmu().
278 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
280 struct mmu_gather_batch *batch;
282 VM_BUG_ON(!tlb->need_flush);
284 if (tlb_fast_mode(tlb)) {
285 free_page_and_swap_cache(page);
286 return 1; /* avoid calling tlb_flush_mmu() */
290 batch->pages[batch->nr++] = page;
291 if (batch->nr == batch->max) {
292 if (!tlb_next_batch(tlb))
296 VM_BUG_ON(batch->nr > batch->max);
298 return batch->max - batch->nr;
301 #endif /* HAVE_GENERIC_MMU_GATHER */
303 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
306 * See the comment near struct mmu_table_batch.
309 static void tlb_remove_table_smp_sync(void *arg)
311 /* Simply deliver the interrupt */
314 static void tlb_remove_table_one(void *table)
317 * This isn't an RCU grace period and hence the page-tables cannot be
318 * assumed to be actually RCU-freed.
320 * It is however sufficient for software page-table walkers that rely on
321 * IRQ disabling. See the comment near struct mmu_table_batch.
323 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
324 __tlb_remove_table(table);
327 static void tlb_remove_table_rcu(struct rcu_head *head)
329 struct mmu_table_batch *batch;
332 batch = container_of(head, struct mmu_table_batch, rcu);
334 for (i = 0; i < batch->nr; i++)
335 __tlb_remove_table(batch->tables[i]);
337 free_page((unsigned long)batch);
340 void tlb_table_flush(struct mmu_gather *tlb)
342 struct mmu_table_batch **batch = &tlb->batch;
345 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
350 void tlb_remove_table(struct mmu_gather *tlb, void *table)
352 struct mmu_table_batch **batch = &tlb->batch;
357 * When there's less then two users of this mm there cannot be a
358 * concurrent page-table walk.
360 if (atomic_read(&tlb->mm->mm_users) < 2) {
361 __tlb_remove_table(table);
365 if (*batch == NULL) {
366 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
367 if (*batch == NULL) {
368 tlb_remove_table_one(table);
373 (*batch)->tables[(*batch)->nr++] = table;
374 if ((*batch)->nr == MAX_TABLE_BATCH)
375 tlb_table_flush(tlb);
378 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
381 * If a p?d_bad entry is found while walking page tables, report
382 * the error, before resetting entry to p?d_none. Usually (but
383 * very seldom) called out from the p?d_none_or_clear_bad macros.
386 void pgd_clear_bad(pgd_t *pgd)
392 void pud_clear_bad(pud_t *pud)
398 void pmd_clear_bad(pmd_t *pmd)
405 * Note: this doesn't free the actual pages themselves. That
406 * has been handled earlier when unmapping all the memory regions.
408 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
411 pgtable_t token = pmd_pgtable(*pmd);
413 pte_free_tlb(tlb, token, addr);
417 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
418 unsigned long addr, unsigned long end,
419 unsigned long floor, unsigned long ceiling)
426 pmd = pmd_offset(pud, addr);
428 next = pmd_addr_end(addr, end);
429 if (pmd_none_or_clear_bad(pmd))
431 free_pte_range(tlb, pmd, addr);
432 } while (pmd++, addr = next, addr != end);
442 if (end - 1 > ceiling - 1)
445 pmd = pmd_offset(pud, start);
447 pmd_free_tlb(tlb, pmd, start);
450 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
451 unsigned long addr, unsigned long end,
452 unsigned long floor, unsigned long ceiling)
459 pud = pud_offset(pgd, addr);
461 next = pud_addr_end(addr, end);
462 if (pud_none_or_clear_bad(pud))
464 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
465 } while (pud++, addr = next, addr != end);
471 ceiling &= PGDIR_MASK;
475 if (end - 1 > ceiling - 1)
478 pud = pud_offset(pgd, start);
480 pud_free_tlb(tlb, pud, start);
484 * This function frees user-level page tables of a process.
486 * Must be called with pagetable lock held.
488 void free_pgd_range(struct mmu_gather *tlb,
489 unsigned long addr, unsigned long end,
490 unsigned long floor, unsigned long ceiling)
496 * The next few lines have given us lots of grief...
498 * Why are we testing PMD* at this top level? Because often
499 * there will be no work to do at all, and we'd prefer not to
500 * go all the way down to the bottom just to discover that.
502 * Why all these "- 1"s? Because 0 represents both the bottom
503 * of the address space and the top of it (using -1 for the
504 * top wouldn't help much: the masks would do the wrong thing).
505 * The rule is that addr 0 and floor 0 refer to the bottom of
506 * the address space, but end 0 and ceiling 0 refer to the top
507 * Comparisons need to use "end - 1" and "ceiling - 1" (though
508 * that end 0 case should be mythical).
510 * Wherever addr is brought up or ceiling brought down, we must
511 * be careful to reject "the opposite 0" before it confuses the
512 * subsequent tests. But what about where end is brought down
513 * by PMD_SIZE below? no, end can't go down to 0 there.
515 * Whereas we round start (addr) and ceiling down, by different
516 * masks at different levels, in order to test whether a table
517 * now has no other vmas using it, so can be freed, we don't
518 * bother to round floor or end up - the tests don't need that.
532 if (end - 1 > ceiling - 1)
537 pgd = pgd_offset(tlb->mm, addr);
539 next = pgd_addr_end(addr, end);
540 if (pgd_none_or_clear_bad(pgd))
542 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
543 } while (pgd++, addr = next, addr != end);
546 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
547 unsigned long floor, unsigned long ceiling)
550 struct vm_area_struct *next = vma->vm_next;
551 unsigned long addr = vma->vm_start;
554 * Hide vma from rmap and truncate_pagecache before freeing
557 unlink_anon_vmas(vma);
558 unlink_file_vma(vma);
560 if (is_vm_hugetlb_page(vma)) {
561 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
562 floor, next? next->vm_start: ceiling);
565 * Optimization: gather nearby vmas into one call down
567 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
568 && !is_vm_hugetlb_page(next)) {
571 unlink_anon_vmas(vma);
572 unlink_file_vma(vma);
574 free_pgd_range(tlb, addr, vma->vm_end,
575 floor, next? next->vm_start: ceiling);
581 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
582 pmd_t *pmd, unsigned long address)
584 pgtable_t new = pte_alloc_one(mm, address);
585 int wait_split_huge_page;
590 * Ensure all pte setup (eg. pte page lock and page clearing) are
591 * visible before the pte is made visible to other CPUs by being
592 * put into page tables.
594 * The other side of the story is the pointer chasing in the page
595 * table walking code (when walking the page table without locking;
596 * ie. most of the time). Fortunately, these data accesses consist
597 * of a chain of data-dependent loads, meaning most CPUs (alpha
598 * being the notable exception) will already guarantee loads are
599 * seen in-order. See the alpha page table accessors for the
600 * smp_read_barrier_depends() barriers in page table walking code.
602 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
604 spin_lock(&mm->page_table_lock);
605 wait_split_huge_page = 0;
606 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
608 pmd_populate(mm, pmd, new);
610 } else if (unlikely(pmd_trans_splitting(*pmd)))
611 wait_split_huge_page = 1;
612 spin_unlock(&mm->page_table_lock);
615 if (wait_split_huge_page)
616 wait_split_huge_page(vma->anon_vma, pmd);
620 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
622 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
626 smp_wmb(); /* See comment in __pte_alloc */
628 spin_lock(&init_mm.page_table_lock);
629 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
630 pmd_populate_kernel(&init_mm, pmd, new);
633 VM_BUG_ON(pmd_trans_splitting(*pmd));
634 spin_unlock(&init_mm.page_table_lock);
636 pte_free_kernel(&init_mm, new);
640 static inline void init_rss_vec(int *rss)
642 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
645 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
649 if (current->mm == mm)
651 for (i = 0; i < NR_MM_COUNTERS; i++)
653 add_mm_counter(mm, i, rss[i]);
657 * This function is called to print an error when a bad pte
658 * is found. For example, we might have a PFN-mapped pte in
659 * a region that doesn't allow it.
661 * The calling function must still handle the error.
663 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
664 pte_t pte, struct page *page)
666 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
667 pud_t *pud = pud_offset(pgd, addr);
668 pmd_t *pmd = pmd_offset(pud, addr);
669 struct address_space *mapping;
671 static unsigned long resume;
672 static unsigned long nr_shown;
673 static unsigned long nr_unshown;
676 * Allow a burst of 60 reports, then keep quiet for that minute;
677 * or allow a steady drip of one report per second.
679 if (nr_shown == 60) {
680 if (time_before(jiffies, resume)) {
686 "BUG: Bad page map: %lu messages suppressed\n",
693 resume = jiffies + 60 * HZ;
695 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
696 index = linear_page_index(vma, addr);
699 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
701 (long long)pte_val(pte), (long long)pmd_val(*pmd));
705 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
706 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
708 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
711 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
712 (unsigned long)vma->vm_ops->fault);
713 if (vma->vm_file && vma->vm_file->f_op)
714 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
715 (unsigned long)vma->vm_file->f_op->mmap);
717 add_taint(TAINT_BAD_PAGE);
720 static inline bool is_cow_mapping(vm_flags_t flags)
722 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
726 static inline int is_zero_pfn(unsigned long pfn)
728 return pfn == zero_pfn;
733 static inline unsigned long my_zero_pfn(unsigned long addr)
740 * vm_normal_page -- This function gets the "struct page" associated with a pte.
742 * "Special" mappings do not wish to be associated with a "struct page" (either
743 * it doesn't exist, or it exists but they don't want to touch it). In this
744 * case, NULL is returned here. "Normal" mappings do have a struct page.
746 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
747 * pte bit, in which case this function is trivial. Secondly, an architecture
748 * may not have a spare pte bit, which requires a more complicated scheme,
751 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
752 * special mapping (even if there are underlying and valid "struct pages").
753 * COWed pages of a VM_PFNMAP are always normal.
755 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
756 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
757 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
758 * mapping will always honor the rule
760 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
762 * And for normal mappings this is false.
764 * This restricts such mappings to be a linear translation from virtual address
765 * to pfn. To get around this restriction, we allow arbitrary mappings so long
766 * as the vma is not a COW mapping; in that case, we know that all ptes are
767 * special (because none can have been COWed).
770 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
772 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
773 * page" backing, however the difference is that _all_ pages with a struct
774 * page (that is, those where pfn_valid is true) are refcounted and considered
775 * normal pages by the VM. The disadvantage is that pages are refcounted
776 * (which can be slower and simply not an option for some PFNMAP users). The
777 * advantage is that we don't have to follow the strict linearity rule of
778 * PFNMAP mappings in order to support COWable mappings.
781 #ifdef __HAVE_ARCH_PTE_SPECIAL
782 # define HAVE_PTE_SPECIAL 1
784 # define HAVE_PTE_SPECIAL 0
786 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
789 unsigned long pfn = pte_pfn(pte);
791 if (HAVE_PTE_SPECIAL) {
792 if (likely(!pte_special(pte)))
794 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
796 if (!is_zero_pfn(pfn))
797 print_bad_pte(vma, addr, pte, NULL);
801 /* !HAVE_PTE_SPECIAL case follows: */
803 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
804 if (vma->vm_flags & VM_MIXEDMAP) {
810 off = (addr - vma->vm_start) >> PAGE_SHIFT;
811 if (pfn == vma->vm_pgoff + off)
813 if (!is_cow_mapping(vma->vm_flags))
818 if (is_zero_pfn(pfn))
821 if (unlikely(pfn > highest_memmap_pfn)) {
822 print_bad_pte(vma, addr, pte, NULL);
827 * NOTE! We still have PageReserved() pages in the page tables.
828 * eg. VDSO mappings can cause them to exist.
831 return pfn_to_page(pfn);
835 * copy one vm_area from one task to the other. Assumes the page tables
836 * already present in the new task to be cleared in the whole range
837 * covered by this vma.
840 static inline unsigned long
841 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
842 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
843 unsigned long addr, int *rss)
845 unsigned long vm_flags = vma->vm_flags;
846 pte_t pte = *src_pte;
849 /* pte contains position in swap or file, so copy. */
850 if (unlikely(!pte_present(pte))) {
851 if (!pte_file(pte)) {
852 swp_entry_t entry = pte_to_swp_entry(pte);
854 if (swap_duplicate(entry) < 0)
857 /* make sure dst_mm is on swapoff's mmlist. */
858 if (unlikely(list_empty(&dst_mm->mmlist))) {
859 spin_lock(&mmlist_lock);
860 if (list_empty(&dst_mm->mmlist))
861 list_add(&dst_mm->mmlist,
863 spin_unlock(&mmlist_lock);
865 if (likely(!non_swap_entry(entry)))
867 else if (is_migration_entry(entry)) {
868 page = migration_entry_to_page(entry);
875 if (is_write_migration_entry(entry) &&
876 is_cow_mapping(vm_flags)) {
878 * COW mappings require pages in both
879 * parent and child to be set to read.
881 make_migration_entry_read(&entry);
882 pte = swp_entry_to_pte(entry);
883 set_pte_at(src_mm, addr, src_pte, pte);
891 * If it's a COW mapping, write protect it both
892 * in the parent and the child
894 if (is_cow_mapping(vm_flags)) {
895 ptep_set_wrprotect(src_mm, addr, src_pte);
896 pte = pte_wrprotect(pte);
900 * If it's a shared mapping, mark it clean in
903 if (vm_flags & VM_SHARED)
904 pte = pte_mkclean(pte);
905 pte = pte_mkold(pte);
907 page = vm_normal_page(vma, addr, pte);
918 set_pte_at(dst_mm, addr, dst_pte, pte);
922 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
923 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
924 unsigned long addr, unsigned long end)
926 pte_t *orig_src_pte, *orig_dst_pte;
927 pte_t *src_pte, *dst_pte;
928 spinlock_t *src_ptl, *dst_ptl;
930 int rss[NR_MM_COUNTERS];
931 swp_entry_t entry = (swp_entry_t){0};
936 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
939 src_pte = pte_offset_map(src_pmd, addr);
940 src_ptl = pte_lockptr(src_mm, src_pmd);
941 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
942 orig_src_pte = src_pte;
943 orig_dst_pte = dst_pte;
944 arch_enter_lazy_mmu_mode();
948 * We are holding two locks at this point - either of them
949 * could generate latencies in another task on another CPU.
951 if (progress >= 32) {
953 if (need_resched() ||
954 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
957 if (pte_none(*src_pte)) {
961 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
966 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
968 arch_leave_lazy_mmu_mode();
969 spin_unlock(src_ptl);
970 pte_unmap(orig_src_pte);
971 add_mm_rss_vec(dst_mm, rss);
972 pte_unmap_unlock(orig_dst_pte, dst_ptl);
976 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
985 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
986 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
987 unsigned long addr, unsigned long end)
989 pmd_t *src_pmd, *dst_pmd;
992 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
995 src_pmd = pmd_offset(src_pud, addr);
997 next = pmd_addr_end(addr, end);
998 if (pmd_trans_huge(*src_pmd)) {
1000 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1001 err = copy_huge_pmd(dst_mm, src_mm,
1002 dst_pmd, src_pmd, addr, vma);
1009 if (pmd_none_or_clear_bad(src_pmd))
1011 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1014 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1018 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1019 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1020 unsigned long addr, unsigned long end)
1022 pud_t *src_pud, *dst_pud;
1025 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1028 src_pud = pud_offset(src_pgd, addr);
1030 next = pud_addr_end(addr, end);
1031 if (pud_none_or_clear_bad(src_pud))
1033 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1036 } while (dst_pud++, src_pud++, addr = next, addr != end);
1040 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1041 struct vm_area_struct *vma)
1043 pgd_t *src_pgd, *dst_pgd;
1045 unsigned long addr = vma->vm_start;
1046 unsigned long end = vma->vm_end;
1047 unsigned long mmun_start; /* For mmu_notifiers */
1048 unsigned long mmun_end; /* For mmu_notifiers */
1053 * Don't copy ptes where a page fault will fill them correctly.
1054 * Fork becomes much lighter when there are big shared or private
1055 * readonly mappings. The tradeoff is that copy_page_range is more
1056 * efficient than faulting.
1058 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1059 VM_PFNMAP | VM_MIXEDMAP))) {
1064 if (is_vm_hugetlb_page(vma))
1065 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1067 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1069 * We do not free on error cases below as remove_vma
1070 * gets called on error from higher level routine
1072 ret = track_pfn_copy(vma);
1078 * We need to invalidate the secondary MMU mappings only when
1079 * there could be a permission downgrade on the ptes of the
1080 * parent mm. And a permission downgrade will only happen if
1081 * is_cow_mapping() returns true.
1083 is_cow = is_cow_mapping(vma->vm_flags);
1087 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1091 dst_pgd = pgd_offset(dst_mm, addr);
1092 src_pgd = pgd_offset(src_mm, addr);
1094 next = pgd_addr_end(addr, end);
1095 if (pgd_none_or_clear_bad(src_pgd))
1097 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1098 vma, addr, next))) {
1102 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1105 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1109 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1110 struct vm_area_struct *vma, pmd_t *pmd,
1111 unsigned long addr, unsigned long end,
1112 struct zap_details *details)
1114 struct mm_struct *mm = tlb->mm;
1115 int force_flush = 0;
1116 int rss[NR_MM_COUNTERS];
1123 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1125 arch_enter_lazy_mmu_mode();
1128 if (pte_none(ptent)) {
1132 if (pte_present(ptent)) {
1135 page = vm_normal_page(vma, addr, ptent);
1136 if (unlikely(details) && page) {
1138 * unmap_shared_mapping_pages() wants to
1139 * invalidate cache without truncating:
1140 * unmap shared but keep private pages.
1142 if (details->check_mapping &&
1143 details->check_mapping != page->mapping)
1146 * Each page->index must be checked when
1147 * invalidating or truncating nonlinear.
1149 if (details->nonlinear_vma &&
1150 (page->index < details->first_index ||
1151 page->index > details->last_index))
1154 ptent = ptep_get_and_clear_full(mm, addr, pte,
1156 tlb_remove_tlb_entry(tlb, pte, addr);
1157 if (unlikely(!page))
1159 if (unlikely(details) && details->nonlinear_vma
1160 && linear_page_index(details->nonlinear_vma,
1161 addr) != page->index)
1162 set_pte_at(mm, addr, pte,
1163 pgoff_to_pte(page->index));
1165 rss[MM_ANONPAGES]--;
1167 if (pte_dirty(ptent))
1168 set_page_dirty(page);
1169 if (pte_young(ptent) &&
1170 likely(!VM_SequentialReadHint(vma)))
1171 mark_page_accessed(page);
1172 rss[MM_FILEPAGES]--;
1174 page_remove_rmap(page);
1175 if (unlikely(page_mapcount(page) < 0))
1176 print_bad_pte(vma, addr, ptent, page);
1177 force_flush = !__tlb_remove_page(tlb, page);
1183 * If details->check_mapping, we leave swap entries;
1184 * if details->nonlinear_vma, we leave file entries.
1186 if (unlikely(details))
1188 if (pte_file(ptent)) {
1189 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1190 print_bad_pte(vma, addr, ptent, NULL);
1192 swp_entry_t entry = pte_to_swp_entry(ptent);
1194 if (!non_swap_entry(entry))
1196 else if (is_migration_entry(entry)) {
1199 page = migration_entry_to_page(entry);
1202 rss[MM_ANONPAGES]--;
1204 rss[MM_FILEPAGES]--;
1206 if (unlikely(!free_swap_and_cache(entry)))
1207 print_bad_pte(vma, addr, ptent, NULL);
1209 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1210 } while (pte++, addr += PAGE_SIZE, addr != end);
1212 add_mm_rss_vec(mm, rss);
1213 arch_leave_lazy_mmu_mode();
1214 pte_unmap_unlock(start_pte, ptl);
1217 * mmu_gather ran out of room to batch pages, we break out of
1218 * the PTE lock to avoid doing the potential expensive TLB invalidate
1219 * and page-free while holding it.
1224 #ifdef HAVE_GENERIC_MMU_GATHER
1236 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1237 struct vm_area_struct *vma, pud_t *pud,
1238 unsigned long addr, unsigned long end,
1239 struct zap_details *details)
1244 pmd = pmd_offset(pud, addr);
1246 next = pmd_addr_end(addr, end);
1247 if (pmd_trans_huge(*pmd)) {
1248 if (next - addr != HPAGE_PMD_SIZE) {
1249 #ifdef CONFIG_DEBUG_VM
1250 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1251 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1252 __func__, addr, end,
1258 split_huge_page_pmd(vma->vm_mm, pmd);
1259 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1264 * Here there can be other concurrent MADV_DONTNEED or
1265 * trans huge page faults running, and if the pmd is
1266 * none or trans huge it can change under us. This is
1267 * because MADV_DONTNEED holds the mmap_sem in read
1270 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1272 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1275 } while (pmd++, addr = next, addr != end);
1280 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1281 struct vm_area_struct *vma, pgd_t *pgd,
1282 unsigned long addr, unsigned long end,
1283 struct zap_details *details)
1288 pud = pud_offset(pgd, addr);
1290 next = pud_addr_end(addr, end);
1291 if (pud_none_or_clear_bad(pud))
1293 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1294 } while (pud++, addr = next, addr != end);
1299 static void unmap_page_range(struct mmu_gather *tlb,
1300 struct vm_area_struct *vma,
1301 unsigned long addr, unsigned long end,
1302 struct zap_details *details)
1307 if (details && !details->check_mapping && !details->nonlinear_vma)
1310 BUG_ON(addr >= end);
1311 mem_cgroup_uncharge_start();
1312 tlb_start_vma(tlb, vma);
1313 pgd = pgd_offset(vma->vm_mm, addr);
1315 next = pgd_addr_end(addr, end);
1316 if (pgd_none_or_clear_bad(pgd))
1318 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1319 } while (pgd++, addr = next, addr != end);
1320 tlb_end_vma(tlb, vma);
1321 mem_cgroup_uncharge_end();
1325 static void unmap_single_vma(struct mmu_gather *tlb,
1326 struct vm_area_struct *vma, unsigned long start_addr,
1327 unsigned long end_addr,
1328 struct zap_details *details)
1330 unsigned long start = max(vma->vm_start, start_addr);
1333 if (start >= vma->vm_end)
1335 end = min(vma->vm_end, end_addr);
1336 if (end <= vma->vm_start)
1340 uprobe_munmap(vma, start, end);
1342 if (unlikely(vma->vm_flags & VM_PFNMAP))
1343 untrack_pfn(vma, 0, 0);
1346 if (unlikely(is_vm_hugetlb_page(vma))) {
1348 * It is undesirable to test vma->vm_file as it
1349 * should be non-null for valid hugetlb area.
1350 * However, vm_file will be NULL in the error
1351 * cleanup path of do_mmap_pgoff. When
1352 * hugetlbfs ->mmap method fails,
1353 * do_mmap_pgoff() nullifies vma->vm_file
1354 * before calling this function to clean up.
1355 * Since no pte has actually been setup, it is
1356 * safe to do nothing in this case.
1359 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1360 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1361 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1364 unmap_page_range(tlb, vma, start, end, details);
1369 * unmap_vmas - unmap a range of memory covered by a list of vma's
1370 * @tlb: address of the caller's struct mmu_gather
1371 * @vma: the starting vma
1372 * @start_addr: virtual address at which to start unmapping
1373 * @end_addr: virtual address at which to end unmapping
1375 * Unmap all pages in the vma list.
1377 * Only addresses between `start' and `end' will be unmapped.
1379 * The VMA list must be sorted in ascending virtual address order.
1381 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1382 * range after unmap_vmas() returns. So the only responsibility here is to
1383 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1384 * drops the lock and schedules.
1386 void unmap_vmas(struct mmu_gather *tlb,
1387 struct vm_area_struct *vma, unsigned long start_addr,
1388 unsigned long end_addr)
1390 struct mm_struct *mm = vma->vm_mm;
1392 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1393 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1394 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1395 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1399 * zap_page_range - remove user pages in a given range
1400 * @vma: vm_area_struct holding the applicable pages
1401 * @start: starting address of pages to zap
1402 * @size: number of bytes to zap
1403 * @details: details of nonlinear truncation or shared cache invalidation
1405 * Caller must protect the VMA list
1407 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1408 unsigned long size, struct zap_details *details)
1410 struct mm_struct *mm = vma->vm_mm;
1411 struct mmu_gather tlb;
1412 unsigned long end = start + size;
1415 tlb_gather_mmu(&tlb, mm, 0);
1416 update_hiwater_rss(mm);
1417 mmu_notifier_invalidate_range_start(mm, start, end);
1418 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1419 unmap_single_vma(&tlb, vma, start, end, details);
1420 mmu_notifier_invalidate_range_end(mm, start, end);
1421 tlb_finish_mmu(&tlb, start, end);
1425 * zap_page_range_single - remove user pages in a given range
1426 * @vma: vm_area_struct holding the applicable pages
1427 * @address: starting address of pages to zap
1428 * @size: number of bytes to zap
1429 * @details: details of nonlinear truncation or shared cache invalidation
1431 * The range must fit into one VMA.
1433 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1434 unsigned long size, struct zap_details *details)
1436 struct mm_struct *mm = vma->vm_mm;
1437 struct mmu_gather tlb;
1438 unsigned long end = address + size;
1441 tlb_gather_mmu(&tlb, mm, 0);
1442 update_hiwater_rss(mm);
1443 mmu_notifier_invalidate_range_start(mm, address, end);
1444 unmap_single_vma(&tlb, vma, address, end, details);
1445 mmu_notifier_invalidate_range_end(mm, address, end);
1446 tlb_finish_mmu(&tlb, address, end);
1450 * zap_vma_ptes - remove ptes mapping the vma
1451 * @vma: vm_area_struct holding ptes to be zapped
1452 * @address: starting address of pages to zap
1453 * @size: number of bytes to zap
1455 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1457 * The entire address range must be fully contained within the vma.
1459 * Returns 0 if successful.
1461 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1464 if (address < vma->vm_start || address + size > vma->vm_end ||
1465 !(vma->vm_flags & VM_PFNMAP))
1467 zap_page_range_single(vma, address, size, NULL);
1470 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1473 * follow_page - look up a page descriptor from a user-virtual address
1474 * @vma: vm_area_struct mapping @address
1475 * @address: virtual address to look up
1476 * @flags: flags modifying lookup behaviour
1478 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1480 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1481 * an error pointer if there is a mapping to something not represented
1482 * by a page descriptor (see also vm_normal_page()).
1484 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1493 struct mm_struct *mm = vma->vm_mm;
1495 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1496 if (!IS_ERR(page)) {
1497 BUG_ON(flags & FOLL_GET);
1502 pgd = pgd_offset(mm, address);
1503 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1506 pud = pud_offset(pgd, address);
1509 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1510 BUG_ON(flags & FOLL_GET);
1511 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1514 if (unlikely(pud_bad(*pud)))
1517 pmd = pmd_offset(pud, address);
1520 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1521 BUG_ON(flags & FOLL_GET);
1522 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1525 if (pmd_trans_huge(*pmd)) {
1526 if (flags & FOLL_SPLIT) {
1527 split_huge_page_pmd(mm, pmd);
1528 goto split_fallthrough;
1530 spin_lock(&mm->page_table_lock);
1531 if (likely(pmd_trans_huge(*pmd))) {
1532 if (unlikely(pmd_trans_splitting(*pmd))) {
1533 spin_unlock(&mm->page_table_lock);
1534 wait_split_huge_page(vma->anon_vma, pmd);
1536 page = follow_trans_huge_pmd(vma, address,
1538 spin_unlock(&mm->page_table_lock);
1542 spin_unlock(&mm->page_table_lock);
1546 if (unlikely(pmd_bad(*pmd)))
1549 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1552 if (!pte_present(pte))
1554 if ((flags & FOLL_WRITE) && !pte_write(pte))
1557 page = vm_normal_page(vma, address, pte);
1558 if (unlikely(!page)) {
1559 if ((flags & FOLL_DUMP) ||
1560 !is_zero_pfn(pte_pfn(pte)))
1562 page = pte_page(pte);
1565 if (flags & FOLL_GET)
1566 get_page_foll(page);
1567 if (flags & FOLL_TOUCH) {
1568 if ((flags & FOLL_WRITE) &&
1569 !pte_dirty(pte) && !PageDirty(page))
1570 set_page_dirty(page);
1572 * pte_mkyoung() would be more correct here, but atomic care
1573 * is needed to avoid losing the dirty bit: it is easier to use
1574 * mark_page_accessed().
1576 mark_page_accessed(page);
1578 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1580 * The preliminary mapping check is mainly to avoid the
1581 * pointless overhead of lock_page on the ZERO_PAGE
1582 * which might bounce very badly if there is contention.
1584 * If the page is already locked, we don't need to
1585 * handle it now - vmscan will handle it later if and
1586 * when it attempts to reclaim the page.
1588 if (page->mapping && trylock_page(page)) {
1589 lru_add_drain(); /* push cached pages to LRU */
1591 * Because we lock page here, and migration is
1592 * blocked by the pte's page reference, and we
1593 * know the page is still mapped, we don't even
1594 * need to check for file-cache page truncation.
1596 mlock_vma_page(page);
1601 pte_unmap_unlock(ptep, ptl);
1606 pte_unmap_unlock(ptep, ptl);
1607 return ERR_PTR(-EFAULT);
1610 pte_unmap_unlock(ptep, ptl);
1616 * When core dumping an enormous anonymous area that nobody
1617 * has touched so far, we don't want to allocate unnecessary pages or
1618 * page tables. Return error instead of NULL to skip handle_mm_fault,
1619 * then get_dump_page() will return NULL to leave a hole in the dump.
1620 * But we can only make this optimization where a hole would surely
1621 * be zero-filled if handle_mm_fault() actually did handle it.
1623 if ((flags & FOLL_DUMP) &&
1624 (!vma->vm_ops || !vma->vm_ops->fault))
1625 return ERR_PTR(-EFAULT);
1629 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1631 return stack_guard_page_start(vma, addr) ||
1632 stack_guard_page_end(vma, addr+PAGE_SIZE);
1636 * __get_user_pages() - pin user pages in memory
1637 * @tsk: task_struct of target task
1638 * @mm: mm_struct of target mm
1639 * @start: starting user address
1640 * @nr_pages: number of pages from start to pin
1641 * @gup_flags: flags modifying pin behaviour
1642 * @pages: array that receives pointers to the pages pinned.
1643 * Should be at least nr_pages long. Or NULL, if caller
1644 * only intends to ensure the pages are faulted in.
1645 * @vmas: array of pointers to vmas corresponding to each page.
1646 * Or NULL if the caller does not require them.
1647 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1649 * Returns number of pages pinned. This may be fewer than the number
1650 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1651 * were pinned, returns -errno. Each page returned must be released
1652 * with a put_page() call when it is finished with. vmas will only
1653 * remain valid while mmap_sem is held.
1655 * Must be called with mmap_sem held for read or write.
1657 * __get_user_pages walks a process's page tables and takes a reference to
1658 * each struct page that each user address corresponds to at a given
1659 * instant. That is, it takes the page that would be accessed if a user
1660 * thread accesses the given user virtual address at that instant.
1662 * This does not guarantee that the page exists in the user mappings when
1663 * __get_user_pages returns, and there may even be a completely different
1664 * page there in some cases (eg. if mmapped pagecache has been invalidated
1665 * and subsequently re faulted). However it does guarantee that the page
1666 * won't be freed completely. And mostly callers simply care that the page
1667 * contains data that was valid *at some point in time*. Typically, an IO
1668 * or similar operation cannot guarantee anything stronger anyway because
1669 * locks can't be held over the syscall boundary.
1671 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1672 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1673 * appropriate) must be called after the page is finished with, and
1674 * before put_page is called.
1676 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1677 * or mmap_sem contention, and if waiting is needed to pin all pages,
1678 * *@nonblocking will be set to 0.
1680 * In most cases, get_user_pages or get_user_pages_fast should be used
1681 * instead of __get_user_pages. __get_user_pages should be used only if
1682 * you need some special @gup_flags.
1684 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1685 unsigned long start, int nr_pages, unsigned int gup_flags,
1686 struct page **pages, struct vm_area_struct **vmas,
1690 unsigned long vm_flags;
1695 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1698 * Require read or write permissions.
1699 * If FOLL_FORCE is set, we only require the "MAY" flags.
1701 vm_flags = (gup_flags & FOLL_WRITE) ?
1702 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1703 vm_flags &= (gup_flags & FOLL_FORCE) ?
1704 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1708 struct vm_area_struct *vma;
1710 vma = find_extend_vma(mm, start);
1711 if (!vma && in_gate_area(mm, start)) {
1712 unsigned long pg = start & PAGE_MASK;
1718 /* user gate pages are read-only */
1719 if (gup_flags & FOLL_WRITE)
1720 return i ? : -EFAULT;
1722 pgd = pgd_offset_k(pg);
1724 pgd = pgd_offset_gate(mm, pg);
1725 BUG_ON(pgd_none(*pgd));
1726 pud = pud_offset(pgd, pg);
1727 BUG_ON(pud_none(*pud));
1728 pmd = pmd_offset(pud, pg);
1730 return i ? : -EFAULT;
1731 VM_BUG_ON(pmd_trans_huge(*pmd));
1732 pte = pte_offset_map(pmd, pg);
1733 if (pte_none(*pte)) {
1735 return i ? : -EFAULT;
1737 vma = get_gate_vma(mm);
1741 page = vm_normal_page(vma, start, *pte);
1743 if (!(gup_flags & FOLL_DUMP) &&
1744 is_zero_pfn(pte_pfn(*pte)))
1745 page = pte_page(*pte);
1748 return i ? : -EFAULT;
1759 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1760 !(vm_flags & vma->vm_flags))
1761 return i ? : -EFAULT;
1763 if (is_vm_hugetlb_page(vma)) {
1764 i = follow_hugetlb_page(mm, vma, pages, vmas,
1765 &start, &nr_pages, i, gup_flags);
1771 unsigned int foll_flags = gup_flags;
1774 * If we have a pending SIGKILL, don't keep faulting
1775 * pages and potentially allocating memory.
1777 if (unlikely(fatal_signal_pending(current)))
1778 return i ? i : -ERESTARTSYS;
1781 while (!(page = follow_page(vma, start, foll_flags))) {
1783 unsigned int fault_flags = 0;
1785 /* For mlock, just skip the stack guard page. */
1786 if (foll_flags & FOLL_MLOCK) {
1787 if (stack_guard_page(vma, start))
1790 if (foll_flags & FOLL_WRITE)
1791 fault_flags |= FAULT_FLAG_WRITE;
1793 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1794 if (foll_flags & FOLL_NOWAIT)
1795 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1797 ret = handle_mm_fault(mm, vma, start,
1800 if (ret & VM_FAULT_ERROR) {
1801 if (ret & VM_FAULT_OOM)
1802 return i ? i : -ENOMEM;
1803 if (ret & (VM_FAULT_HWPOISON |
1804 VM_FAULT_HWPOISON_LARGE)) {
1807 else if (gup_flags & FOLL_HWPOISON)
1812 if (ret & VM_FAULT_SIGBUS)
1813 return i ? i : -EFAULT;
1818 if (ret & VM_FAULT_MAJOR)
1824 if (ret & VM_FAULT_RETRY) {
1831 * The VM_FAULT_WRITE bit tells us that
1832 * do_wp_page has broken COW when necessary,
1833 * even if maybe_mkwrite decided not to set
1834 * pte_write. We can thus safely do subsequent
1835 * page lookups as if they were reads. But only
1836 * do so when looping for pte_write is futile:
1837 * in some cases userspace may also be wanting
1838 * to write to the gotten user page, which a
1839 * read fault here might prevent (a readonly
1840 * page might get reCOWed by userspace write).
1842 if ((ret & VM_FAULT_WRITE) &&
1843 !(vma->vm_flags & VM_WRITE))
1844 foll_flags &= ~FOLL_WRITE;
1849 return i ? i : PTR_ERR(page);
1853 flush_anon_page(vma, page, start);
1854 flush_dcache_page(page);
1862 } while (nr_pages && start < vma->vm_end);
1866 EXPORT_SYMBOL(__get_user_pages);
1869 * fixup_user_fault() - manually resolve a user page fault
1870 * @tsk: the task_struct to use for page fault accounting, or
1871 * NULL if faults are not to be recorded.
1872 * @mm: mm_struct of target mm
1873 * @address: user address
1874 * @fault_flags:flags to pass down to handle_mm_fault()
1876 * This is meant to be called in the specific scenario where for locking reasons
1877 * we try to access user memory in atomic context (within a pagefault_disable()
1878 * section), this returns -EFAULT, and we want to resolve the user fault before
1881 * Typically this is meant to be used by the futex code.
1883 * The main difference with get_user_pages() is that this function will
1884 * unconditionally call handle_mm_fault() which will in turn perform all the
1885 * necessary SW fixup of the dirty and young bits in the PTE, while
1886 * handle_mm_fault() only guarantees to update these in the struct page.
1888 * This is important for some architectures where those bits also gate the
1889 * access permission to the page because they are maintained in software. On
1890 * such architectures, gup() will not be enough to make a subsequent access
1893 * This should be called with the mm_sem held for read.
1895 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1896 unsigned long address, unsigned int fault_flags)
1898 struct vm_area_struct *vma;
1901 vma = find_extend_vma(mm, address);
1902 if (!vma || address < vma->vm_start)
1905 ret = handle_mm_fault(mm, vma, address, fault_flags);
1906 if (ret & VM_FAULT_ERROR) {
1907 if (ret & VM_FAULT_OOM)
1909 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1911 if (ret & VM_FAULT_SIGBUS)
1916 if (ret & VM_FAULT_MAJOR)
1925 * get_user_pages() - pin user pages in memory
1926 * @tsk: the task_struct to use for page fault accounting, or
1927 * NULL if faults are not to be recorded.
1928 * @mm: mm_struct of target mm
1929 * @start: starting user address
1930 * @nr_pages: number of pages from start to pin
1931 * @write: whether pages will be written to by the caller
1932 * @force: whether to force write access even if user mapping is
1933 * readonly. This will result in the page being COWed even
1934 * in MAP_SHARED mappings. You do not want this.
1935 * @pages: array that receives pointers to the pages pinned.
1936 * Should be at least nr_pages long. Or NULL, if caller
1937 * only intends to ensure the pages are faulted in.
1938 * @vmas: array of pointers to vmas corresponding to each page.
1939 * Or NULL if the caller does not require them.
1941 * Returns number of pages pinned. This may be fewer than the number
1942 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1943 * were pinned, returns -errno. Each page returned must be released
1944 * with a put_page() call when it is finished with. vmas will only
1945 * remain valid while mmap_sem is held.
1947 * Must be called with mmap_sem held for read or write.
1949 * get_user_pages walks a process's page tables and takes a reference to
1950 * each struct page that each user address corresponds to at a given
1951 * instant. That is, it takes the page that would be accessed if a user
1952 * thread accesses the given user virtual address at that instant.
1954 * This does not guarantee that the page exists in the user mappings when
1955 * get_user_pages returns, and there may even be a completely different
1956 * page there in some cases (eg. if mmapped pagecache has been invalidated
1957 * and subsequently re faulted). However it does guarantee that the page
1958 * won't be freed completely. And mostly callers simply care that the page
1959 * contains data that was valid *at some point in time*. Typically, an IO
1960 * or similar operation cannot guarantee anything stronger anyway because
1961 * locks can't be held over the syscall boundary.
1963 * If write=0, the page must not be written to. If the page is written to,
1964 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1965 * after the page is finished with, and before put_page is called.
1967 * get_user_pages is typically used for fewer-copy IO operations, to get a
1968 * handle on the memory by some means other than accesses via the user virtual
1969 * addresses. The pages may be submitted for DMA to devices or accessed via
1970 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1971 * use the correct cache flushing APIs.
1973 * See also get_user_pages_fast, for performance critical applications.
1975 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1976 unsigned long start, int nr_pages, int write, int force,
1977 struct page **pages, struct vm_area_struct **vmas)
1979 int flags = FOLL_TOUCH;
1984 flags |= FOLL_WRITE;
1986 flags |= FOLL_FORCE;
1988 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1991 EXPORT_SYMBOL(get_user_pages);
1994 * get_dump_page() - pin user page in memory while writing it to core dump
1995 * @addr: user address
1997 * Returns struct page pointer of user page pinned for dump,
1998 * to be freed afterwards by page_cache_release() or put_page().
2000 * Returns NULL on any kind of failure - a hole must then be inserted into
2001 * the corefile, to preserve alignment with its headers; and also returns
2002 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2003 * allowing a hole to be left in the corefile to save diskspace.
2005 * Called without mmap_sem, but after all other threads have been killed.
2007 #ifdef CONFIG_ELF_CORE
2008 struct page *get_dump_page(unsigned long addr)
2010 struct vm_area_struct *vma;
2013 if (__get_user_pages(current, current->mm, addr, 1,
2014 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2017 flush_cache_page(vma, addr, page_to_pfn(page));
2020 #endif /* CONFIG_ELF_CORE */
2022 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2025 pgd_t * pgd = pgd_offset(mm, addr);
2026 pud_t * pud = pud_alloc(mm, pgd, addr);
2028 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2030 VM_BUG_ON(pmd_trans_huge(*pmd));
2031 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2038 * This is the old fallback for page remapping.
2040 * For historical reasons, it only allows reserved pages. Only
2041 * old drivers should use this, and they needed to mark their
2042 * pages reserved for the old functions anyway.
2044 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2045 struct page *page, pgprot_t prot)
2047 struct mm_struct *mm = vma->vm_mm;
2056 flush_dcache_page(page);
2057 pte = get_locked_pte(mm, addr, &ptl);
2061 if (!pte_none(*pte))
2064 /* Ok, finally just insert the thing.. */
2066 inc_mm_counter_fast(mm, MM_FILEPAGES);
2067 page_add_file_rmap(page);
2068 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2071 pte_unmap_unlock(pte, ptl);
2074 pte_unmap_unlock(pte, ptl);
2080 * vm_insert_page - insert single page into user vma
2081 * @vma: user vma to map to
2082 * @addr: target user address of this page
2083 * @page: source kernel page
2085 * This allows drivers to insert individual pages they've allocated
2088 * The page has to be a nice clean _individual_ kernel allocation.
2089 * If you allocate a compound page, you need to have marked it as
2090 * such (__GFP_COMP), or manually just split the page up yourself
2091 * (see split_page()).
2093 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2094 * took an arbitrary page protection parameter. This doesn't allow
2095 * that. Your vma protection will have to be set up correctly, which
2096 * means that if you want a shared writable mapping, you'd better
2097 * ask for a shared writable mapping!
2099 * The page does not need to be reserved.
2101 * Usually this function is called from f_op->mmap() handler
2102 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2103 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2104 * function from other places, for example from page-fault handler.
2106 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2109 if (addr < vma->vm_start || addr >= vma->vm_end)
2111 if (!page_count(page))
2113 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2114 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2115 BUG_ON(vma->vm_flags & VM_PFNMAP);
2116 vma->vm_flags |= VM_MIXEDMAP;
2118 return insert_page(vma, addr, page, vma->vm_page_prot);
2120 EXPORT_SYMBOL(vm_insert_page);
2122 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2123 unsigned long pfn, pgprot_t prot)
2125 struct mm_struct *mm = vma->vm_mm;
2131 pte = get_locked_pte(mm, addr, &ptl);
2135 if (!pte_none(*pte))
2138 /* Ok, finally just insert the thing.. */
2139 entry = pte_mkspecial(pfn_pte(pfn, prot));
2140 set_pte_at(mm, addr, pte, entry);
2141 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2145 pte_unmap_unlock(pte, ptl);
2151 * vm_insert_pfn - insert single pfn into user vma
2152 * @vma: user vma to map to
2153 * @addr: target user address of this page
2154 * @pfn: source kernel pfn
2156 * Similar to vm_insert_page, this allows drivers to insert individual pages
2157 * they've allocated into a user vma. Same comments apply.
2159 * This function should only be called from a vm_ops->fault handler, and
2160 * in that case the handler should return NULL.
2162 * vma cannot be a COW mapping.
2164 * As this is called only for pages that do not currently exist, we
2165 * do not need to flush old virtual caches or the TLB.
2167 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2171 pgprot_t pgprot = vma->vm_page_prot;
2173 * Technically, architectures with pte_special can avoid all these
2174 * restrictions (same for remap_pfn_range). However we would like
2175 * consistency in testing and feature parity among all, so we should
2176 * try to keep these invariants in place for everybody.
2178 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2179 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2180 (VM_PFNMAP|VM_MIXEDMAP));
2181 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2182 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2184 if (addr < vma->vm_start || addr >= vma->vm_end)
2186 if (track_pfn_insert(vma, &pgprot, pfn))
2189 ret = insert_pfn(vma, addr, pfn, pgprot);
2193 EXPORT_SYMBOL(vm_insert_pfn);
2195 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2198 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2200 if (addr < vma->vm_start || addr >= vma->vm_end)
2204 * If we don't have pte special, then we have to use the pfn_valid()
2205 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2206 * refcount the page if pfn_valid is true (hence insert_page rather
2207 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2208 * without pte special, it would there be refcounted as a normal page.
2210 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2213 page = pfn_to_page(pfn);
2214 return insert_page(vma, addr, page, vma->vm_page_prot);
2216 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2218 EXPORT_SYMBOL(vm_insert_mixed);
2221 * maps a range of physical memory into the requested pages. the old
2222 * mappings are removed. any references to nonexistent pages results
2223 * in null mappings (currently treated as "copy-on-access")
2225 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2226 unsigned long addr, unsigned long end,
2227 unsigned long pfn, pgprot_t prot)
2232 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2235 arch_enter_lazy_mmu_mode();
2237 BUG_ON(!pte_none(*pte));
2238 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2240 } while (pte++, addr += PAGE_SIZE, addr != end);
2241 arch_leave_lazy_mmu_mode();
2242 pte_unmap_unlock(pte - 1, ptl);
2246 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2247 unsigned long addr, unsigned long end,
2248 unsigned long pfn, pgprot_t prot)
2253 pfn -= addr >> PAGE_SHIFT;
2254 pmd = pmd_alloc(mm, pud, addr);
2257 VM_BUG_ON(pmd_trans_huge(*pmd));
2259 next = pmd_addr_end(addr, end);
2260 if (remap_pte_range(mm, pmd, addr, next,
2261 pfn + (addr >> PAGE_SHIFT), prot))
2263 } while (pmd++, addr = next, addr != end);
2267 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2268 unsigned long addr, unsigned long end,
2269 unsigned long pfn, pgprot_t prot)
2274 pfn -= addr >> PAGE_SHIFT;
2275 pud = pud_alloc(mm, pgd, addr);
2279 next = pud_addr_end(addr, end);
2280 if (remap_pmd_range(mm, pud, addr, next,
2281 pfn + (addr >> PAGE_SHIFT), prot))
2283 } while (pud++, addr = next, addr != end);
2288 * remap_pfn_range - remap kernel memory to userspace
2289 * @vma: user vma to map to
2290 * @addr: target user address to start at
2291 * @pfn: physical address of kernel memory
2292 * @size: size of map area
2293 * @prot: page protection flags for this mapping
2295 * Note: this is only safe if the mm semaphore is held when called.
2297 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2298 unsigned long pfn, unsigned long size, pgprot_t prot)
2302 unsigned long end = addr + PAGE_ALIGN(size);
2303 struct mm_struct *mm = vma->vm_mm;
2307 * Physically remapped pages are special. Tell the
2308 * rest of the world about it:
2309 * VM_IO tells people not to look at these pages
2310 * (accesses can have side effects).
2311 * VM_PFNMAP tells the core MM that the base pages are just
2312 * raw PFN mappings, and do not have a "struct page" associated
2315 * Disable vma merging and expanding with mremap().
2317 * Omit vma from core dump, even when VM_IO turned off.
2319 * There's a horrible special case to handle copy-on-write
2320 * behaviour that some programs depend on. We mark the "original"
2321 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2322 * See vm_normal_page() for details.
2324 if (is_cow_mapping(vma->vm_flags)) {
2325 if (addr != vma->vm_start || end != vma->vm_end)
2327 vma->vm_pgoff = pfn;
2330 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2334 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2336 BUG_ON(addr >= end);
2337 pfn -= addr >> PAGE_SHIFT;
2338 pgd = pgd_offset(mm, addr);
2339 flush_cache_range(vma, addr, end);
2341 next = pgd_addr_end(addr, end);
2342 err = remap_pud_range(mm, pgd, addr, next,
2343 pfn + (addr >> PAGE_SHIFT), prot);
2346 } while (pgd++, addr = next, addr != end);
2349 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2353 EXPORT_SYMBOL(remap_pfn_range);
2355 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2356 unsigned long addr, unsigned long end,
2357 pte_fn_t fn, void *data)
2362 spinlock_t *uninitialized_var(ptl);
2364 pte = (mm == &init_mm) ?
2365 pte_alloc_kernel(pmd, addr) :
2366 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2370 BUG_ON(pmd_huge(*pmd));
2372 arch_enter_lazy_mmu_mode();
2374 token = pmd_pgtable(*pmd);
2377 err = fn(pte++, token, addr, data);
2380 } while (addr += PAGE_SIZE, addr != end);
2382 arch_leave_lazy_mmu_mode();
2385 pte_unmap_unlock(pte-1, ptl);
2389 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2390 unsigned long addr, unsigned long end,
2391 pte_fn_t fn, void *data)
2397 BUG_ON(pud_huge(*pud));
2399 pmd = pmd_alloc(mm, pud, addr);
2403 next = pmd_addr_end(addr, end);
2404 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2407 } while (pmd++, addr = next, addr != end);
2411 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2412 unsigned long addr, unsigned long end,
2413 pte_fn_t fn, void *data)
2419 pud = pud_alloc(mm, pgd, addr);
2423 next = pud_addr_end(addr, end);
2424 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2427 } while (pud++, addr = next, addr != end);
2432 * Scan a region of virtual memory, filling in page tables as necessary
2433 * and calling a provided function on each leaf page table.
2435 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2436 unsigned long size, pte_fn_t fn, void *data)
2440 unsigned long end = addr + size;
2443 BUG_ON(addr >= end);
2444 pgd = pgd_offset(mm, addr);
2446 next = pgd_addr_end(addr, end);
2447 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2450 } while (pgd++, addr = next, addr != end);
2454 EXPORT_SYMBOL_GPL(apply_to_page_range);
2457 * handle_pte_fault chooses page fault handler according to an entry
2458 * which was read non-atomically. Before making any commitment, on
2459 * those architectures or configurations (e.g. i386 with PAE) which
2460 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2461 * must check under lock before unmapping the pte and proceeding
2462 * (but do_wp_page is only called after already making such a check;
2463 * and do_anonymous_page can safely check later on).
2465 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2466 pte_t *page_table, pte_t orig_pte)
2469 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2470 if (sizeof(pte_t) > sizeof(unsigned long)) {
2471 spinlock_t *ptl = pte_lockptr(mm, pmd);
2473 same = pte_same(*page_table, orig_pte);
2477 pte_unmap(page_table);
2481 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2484 * If the source page was a PFN mapping, we don't have
2485 * a "struct page" for it. We do a best-effort copy by
2486 * just copying from the original user address. If that
2487 * fails, we just zero-fill it. Live with it.
2489 if (unlikely(!src)) {
2490 void *kaddr = kmap_atomic(dst);
2491 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2494 * This really shouldn't fail, because the page is there
2495 * in the page tables. But it might just be unreadable,
2496 * in which case we just give up and fill the result with
2499 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2501 kunmap_atomic(kaddr);
2502 flush_dcache_page(dst);
2504 copy_user_highpage(dst, src, va, vma);
2508 * This routine handles present pages, when users try to write
2509 * to a shared page. It is done by copying the page to a new address
2510 * and decrementing the shared-page counter for the old page.
2512 * Note that this routine assumes that the protection checks have been
2513 * done by the caller (the low-level page fault routine in most cases).
2514 * Thus we can safely just mark it writable once we've done any necessary
2517 * We also mark the page dirty at this point even though the page will
2518 * change only once the write actually happens. This avoids a few races,
2519 * and potentially makes it more efficient.
2521 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2522 * but allow concurrent faults), with pte both mapped and locked.
2523 * We return with mmap_sem still held, but pte unmapped and unlocked.
2525 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2526 unsigned long address, pte_t *page_table, pmd_t *pmd,
2527 spinlock_t *ptl, pte_t orig_pte)
2530 struct page *old_page, *new_page = NULL;
2533 int page_mkwrite = 0;
2534 struct page *dirty_page = NULL;
2535 unsigned long mmun_start; /* For mmu_notifiers */
2536 unsigned long mmun_end; /* For mmu_notifiers */
2537 bool mmun_called = false; /* For mmu_notifiers */
2539 old_page = vm_normal_page(vma, address, orig_pte);
2542 * VM_MIXEDMAP !pfn_valid() case
2544 * We should not cow pages in a shared writeable mapping.
2545 * Just mark the pages writable as we can't do any dirty
2546 * accounting on raw pfn maps.
2548 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2549 (VM_WRITE|VM_SHARED))
2555 * Take out anonymous pages first, anonymous shared vmas are
2556 * not dirty accountable.
2558 if (PageAnon(old_page) && !PageKsm(old_page)) {
2559 if (!trylock_page(old_page)) {
2560 page_cache_get(old_page);
2561 pte_unmap_unlock(page_table, ptl);
2562 lock_page(old_page);
2563 page_table = pte_offset_map_lock(mm, pmd, address,
2565 if (!pte_same(*page_table, orig_pte)) {
2566 unlock_page(old_page);
2569 page_cache_release(old_page);
2571 if (reuse_swap_page(old_page)) {
2573 * The page is all ours. Move it to our anon_vma so
2574 * the rmap code will not search our parent or siblings.
2575 * Protected against the rmap code by the page lock.
2577 page_move_anon_rmap(old_page, vma, address);
2578 unlock_page(old_page);
2581 unlock_page(old_page);
2582 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2583 (VM_WRITE|VM_SHARED))) {
2585 * Only catch write-faults on shared writable pages,
2586 * read-only shared pages can get COWed by
2587 * get_user_pages(.write=1, .force=1).
2589 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2590 struct vm_fault vmf;
2593 vmf.virtual_address = (void __user *)(address &
2595 vmf.pgoff = old_page->index;
2596 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2597 vmf.page = old_page;
2600 * Notify the address space that the page is about to
2601 * become writable so that it can prohibit this or wait
2602 * for the page to get into an appropriate state.
2604 * We do this without the lock held, so that it can
2605 * sleep if it needs to.
2607 page_cache_get(old_page);
2608 pte_unmap_unlock(page_table, ptl);
2610 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2612 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2614 goto unwritable_page;
2616 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2617 lock_page(old_page);
2618 if (!old_page->mapping) {
2619 ret = 0; /* retry the fault */
2620 unlock_page(old_page);
2621 goto unwritable_page;
2624 VM_BUG_ON(!PageLocked(old_page));
2627 * Since we dropped the lock we need to revalidate
2628 * the PTE as someone else may have changed it. If
2629 * they did, we just return, as we can count on the
2630 * MMU to tell us if they didn't also make it writable.
2632 page_table = pte_offset_map_lock(mm, pmd, address,
2634 if (!pte_same(*page_table, orig_pte)) {
2635 unlock_page(old_page);
2641 dirty_page = old_page;
2642 get_page(dirty_page);
2645 flush_cache_page(vma, address, pte_pfn(orig_pte));
2646 entry = pte_mkyoung(orig_pte);
2647 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2648 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2649 update_mmu_cache(vma, address, page_table);
2650 pte_unmap_unlock(page_table, ptl);
2651 ret |= VM_FAULT_WRITE;
2657 * Yes, Virginia, this is actually required to prevent a race
2658 * with clear_page_dirty_for_io() from clearing the page dirty
2659 * bit after it clear all dirty ptes, but before a racing
2660 * do_wp_page installs a dirty pte.
2662 * __do_fault is protected similarly.
2664 if (!page_mkwrite) {
2665 wait_on_page_locked(dirty_page);
2666 set_page_dirty_balance(dirty_page, page_mkwrite);
2667 /* file_update_time outside page_lock */
2669 file_update_time(vma->vm_file);
2671 put_page(dirty_page);
2673 struct address_space *mapping = dirty_page->mapping;
2675 set_page_dirty(dirty_page);
2676 unlock_page(dirty_page);
2677 page_cache_release(dirty_page);
2680 * Some device drivers do not set page.mapping
2681 * but still dirty their pages
2683 balance_dirty_pages_ratelimited(mapping);
2691 * Ok, we need to copy. Oh, well..
2693 page_cache_get(old_page);
2695 pte_unmap_unlock(page_table, ptl);
2697 if (unlikely(anon_vma_prepare(vma)))
2700 if (is_zero_pfn(pte_pfn(orig_pte))) {
2701 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2705 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2708 cow_user_page(new_page, old_page, address, vma);
2710 __SetPageUptodate(new_page);
2712 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2715 mmun_start = address & PAGE_MASK;
2716 mmun_end = (address & PAGE_MASK) + PAGE_SIZE;
2718 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2721 * Re-check the pte - we dropped the lock
2723 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2724 if (likely(pte_same(*page_table, orig_pte))) {
2726 if (!PageAnon(old_page)) {
2727 dec_mm_counter_fast(mm, MM_FILEPAGES);
2728 inc_mm_counter_fast(mm, MM_ANONPAGES);
2731 inc_mm_counter_fast(mm, MM_ANONPAGES);
2732 flush_cache_page(vma, address, pte_pfn(orig_pte));
2733 entry = mk_pte(new_page, vma->vm_page_prot);
2734 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2736 * Clear the pte entry and flush it first, before updating the
2737 * pte with the new entry. This will avoid a race condition
2738 * seen in the presence of one thread doing SMC and another
2741 ptep_clear_flush(vma, address, page_table);
2742 page_add_new_anon_rmap(new_page, vma, address);
2744 * We call the notify macro here because, when using secondary
2745 * mmu page tables (such as kvm shadow page tables), we want the
2746 * new page to be mapped directly into the secondary page table.
2748 set_pte_at_notify(mm, address, page_table, entry);
2749 update_mmu_cache(vma, address, page_table);
2752 * Only after switching the pte to the new page may
2753 * we remove the mapcount here. Otherwise another
2754 * process may come and find the rmap count decremented
2755 * before the pte is switched to the new page, and
2756 * "reuse" the old page writing into it while our pte
2757 * here still points into it and can be read by other
2760 * The critical issue is to order this
2761 * page_remove_rmap with the ptp_clear_flush above.
2762 * Those stores are ordered by (if nothing else,)
2763 * the barrier present in the atomic_add_negative
2764 * in page_remove_rmap.
2766 * Then the TLB flush in ptep_clear_flush ensures that
2767 * no process can access the old page before the
2768 * decremented mapcount is visible. And the old page
2769 * cannot be reused until after the decremented
2770 * mapcount is visible. So transitively, TLBs to
2771 * old page will be flushed before it can be reused.
2773 page_remove_rmap(old_page);
2776 /* Free the old page.. */
2777 new_page = old_page;
2778 ret |= VM_FAULT_WRITE;
2780 mem_cgroup_uncharge_page(new_page);
2783 page_cache_release(new_page);
2785 pte_unmap_unlock(page_table, ptl);
2787 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2790 * Don't let another task, with possibly unlocked vma,
2791 * keep the mlocked page.
2793 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2794 lock_page(old_page); /* LRU manipulation */
2795 munlock_vma_page(old_page);
2796 unlock_page(old_page);
2798 page_cache_release(old_page);
2802 page_cache_release(new_page);
2806 unlock_page(old_page);
2807 page_cache_release(old_page);
2809 page_cache_release(old_page);
2811 return VM_FAULT_OOM;
2814 page_cache_release(old_page);
2818 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2819 unsigned long start_addr, unsigned long end_addr,
2820 struct zap_details *details)
2822 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2825 static inline void unmap_mapping_range_tree(struct rb_root *root,
2826 struct zap_details *details)
2828 struct vm_area_struct *vma;
2829 pgoff_t vba, vea, zba, zea;
2831 vma_interval_tree_foreach(vma, root,
2832 details->first_index, details->last_index) {
2834 vba = vma->vm_pgoff;
2835 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2836 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2837 zba = details->first_index;
2840 zea = details->last_index;
2844 unmap_mapping_range_vma(vma,
2845 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2846 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2851 static inline void unmap_mapping_range_list(struct list_head *head,
2852 struct zap_details *details)
2854 struct vm_area_struct *vma;
2857 * In nonlinear VMAs there is no correspondence between virtual address
2858 * offset and file offset. So we must perform an exhaustive search
2859 * across *all* the pages in each nonlinear VMA, not just the pages
2860 * whose virtual address lies outside the file truncation point.
2862 list_for_each_entry(vma, head, shared.nonlinear) {
2863 details->nonlinear_vma = vma;
2864 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2869 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2870 * @mapping: the address space containing mmaps to be unmapped.
2871 * @holebegin: byte in first page to unmap, relative to the start of
2872 * the underlying file. This will be rounded down to a PAGE_SIZE
2873 * boundary. Note that this is different from truncate_pagecache(), which
2874 * must keep the partial page. In contrast, we must get rid of
2876 * @holelen: size of prospective hole in bytes. This will be rounded
2877 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2879 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2880 * but 0 when invalidating pagecache, don't throw away private data.
2882 void unmap_mapping_range(struct address_space *mapping,
2883 loff_t const holebegin, loff_t const holelen, int even_cows)
2885 struct zap_details details;
2886 pgoff_t hba = holebegin >> PAGE_SHIFT;
2887 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2889 /* Check for overflow. */
2890 if (sizeof(holelen) > sizeof(hlen)) {
2892 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2893 if (holeend & ~(long long)ULONG_MAX)
2894 hlen = ULONG_MAX - hba + 1;
2897 details.check_mapping = even_cows? NULL: mapping;
2898 details.nonlinear_vma = NULL;
2899 details.first_index = hba;
2900 details.last_index = hba + hlen - 1;
2901 if (details.last_index < details.first_index)
2902 details.last_index = ULONG_MAX;
2905 mutex_lock(&mapping->i_mmap_mutex);
2906 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2907 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2908 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2909 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2910 mutex_unlock(&mapping->i_mmap_mutex);
2912 EXPORT_SYMBOL(unmap_mapping_range);
2915 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2916 * but allow concurrent faults), and pte mapped but not yet locked.
2917 * We return with mmap_sem still held, but pte unmapped and unlocked.
2919 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2920 unsigned long address, pte_t *page_table, pmd_t *pmd,
2921 unsigned int flags, pte_t orig_pte)
2924 struct page *page, *swapcache = NULL;
2928 struct mem_cgroup *ptr;
2932 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2935 entry = pte_to_swp_entry(orig_pte);
2936 if (unlikely(non_swap_entry(entry))) {
2937 if (is_migration_entry(entry)) {
2938 migration_entry_wait(mm, pmd, address);
2939 } else if (is_hwpoison_entry(entry)) {
2940 ret = VM_FAULT_HWPOISON;
2942 print_bad_pte(vma, address, orig_pte, NULL);
2943 ret = VM_FAULT_SIGBUS;
2947 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2948 page = lookup_swap_cache(entry);
2950 page = swapin_readahead(entry,
2951 GFP_HIGHUSER_MOVABLE, vma, address);
2954 * Back out if somebody else faulted in this pte
2955 * while we released the pte lock.
2957 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2958 if (likely(pte_same(*page_table, orig_pte)))
2960 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2964 /* Had to read the page from swap area: Major fault */
2965 ret = VM_FAULT_MAJOR;
2966 count_vm_event(PGMAJFAULT);
2967 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2968 } else if (PageHWPoison(page)) {
2970 * hwpoisoned dirty swapcache pages are kept for killing
2971 * owner processes (which may be unknown at hwpoison time)
2973 ret = VM_FAULT_HWPOISON;
2974 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2978 locked = lock_page_or_retry(page, mm, flags);
2980 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2982 ret |= VM_FAULT_RETRY;
2987 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2988 * release the swapcache from under us. The page pin, and pte_same
2989 * test below, are not enough to exclude that. Even if it is still
2990 * swapcache, we need to check that the page's swap has not changed.
2992 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2995 if (ksm_might_need_to_copy(page, vma, address)) {
2997 page = ksm_does_need_to_copy(page, vma, address);
2999 if (unlikely(!page)) {
3007 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3013 * Back out if somebody else already faulted in this pte.
3015 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3016 if (unlikely(!pte_same(*page_table, orig_pte)))
3019 if (unlikely(!PageUptodate(page))) {
3020 ret = VM_FAULT_SIGBUS;
3025 * The page isn't present yet, go ahead with the fault.
3027 * Be careful about the sequence of operations here.
3028 * To get its accounting right, reuse_swap_page() must be called
3029 * while the page is counted on swap but not yet in mapcount i.e.
3030 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3031 * must be called after the swap_free(), or it will never succeed.
3032 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3033 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3034 * in page->private. In this case, a record in swap_cgroup is silently
3035 * discarded at swap_free().
3038 inc_mm_counter_fast(mm, MM_ANONPAGES);
3039 dec_mm_counter_fast(mm, MM_SWAPENTS);
3040 pte = mk_pte(page, vma->vm_page_prot);
3041 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3042 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3043 flags &= ~FAULT_FLAG_WRITE;
3044 ret |= VM_FAULT_WRITE;
3047 flush_icache_page(vma, page);
3048 set_pte_at(mm, address, page_table, pte);
3049 do_page_add_anon_rmap(page, vma, address, exclusive);
3050 /* It's better to call commit-charge after rmap is established */
3051 mem_cgroup_commit_charge_swapin(page, ptr);
3054 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3055 try_to_free_swap(page);
3059 * Hold the lock to avoid the swap entry to be reused
3060 * until we take the PT lock for the pte_same() check
3061 * (to avoid false positives from pte_same). For
3062 * further safety release the lock after the swap_free
3063 * so that the swap count won't change under a
3064 * parallel locked swapcache.
3066 unlock_page(swapcache);
3067 page_cache_release(swapcache);
3070 if (flags & FAULT_FLAG_WRITE) {
3071 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3072 if (ret & VM_FAULT_ERROR)
3073 ret &= VM_FAULT_ERROR;
3077 /* No need to invalidate - it was non-present before */
3078 update_mmu_cache(vma, address, page_table);
3080 pte_unmap_unlock(page_table, ptl);
3084 mem_cgroup_cancel_charge_swapin(ptr);
3085 pte_unmap_unlock(page_table, ptl);
3089 page_cache_release(page);
3091 unlock_page(swapcache);
3092 page_cache_release(swapcache);
3098 * This is like a special single-page "expand_{down|up}wards()",
3099 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3100 * doesn't hit another vma.
3102 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3104 address &= PAGE_MASK;
3105 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3106 struct vm_area_struct *prev = vma->vm_prev;
3109 * Is there a mapping abutting this one below?
3111 * That's only ok if it's the same stack mapping
3112 * that has gotten split..
3114 if (prev && prev->vm_end == address)
3115 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3117 expand_downwards(vma, address - PAGE_SIZE);
3119 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3120 struct vm_area_struct *next = vma->vm_next;
3122 /* As VM_GROWSDOWN but s/below/above/ */
3123 if (next && next->vm_start == address + PAGE_SIZE)
3124 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3126 expand_upwards(vma, address + PAGE_SIZE);
3132 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3133 * but allow concurrent faults), and pte mapped but not yet locked.
3134 * We return with mmap_sem still held, but pte unmapped and unlocked.
3136 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3137 unsigned long address, pte_t *page_table, pmd_t *pmd,
3144 pte_unmap(page_table);
3146 /* Check if we need to add a guard page to the stack */
3147 if (check_stack_guard_page(vma, address) < 0)
3148 return VM_FAULT_SIGBUS;
3150 /* Use the zero-page for reads */
3151 if (!(flags & FAULT_FLAG_WRITE)) {
3152 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3153 vma->vm_page_prot));
3154 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3155 if (!pte_none(*page_table))
3160 /* Allocate our own private page. */
3161 if (unlikely(anon_vma_prepare(vma)))
3163 page = alloc_zeroed_user_highpage_movable(vma, address);
3166 __SetPageUptodate(page);
3168 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3171 entry = mk_pte(page, vma->vm_page_prot);
3172 if (vma->vm_flags & VM_WRITE)
3173 entry = pte_mkwrite(pte_mkdirty(entry));
3175 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3176 if (!pte_none(*page_table))
3179 inc_mm_counter_fast(mm, MM_ANONPAGES);
3180 page_add_new_anon_rmap(page, vma, address);
3182 set_pte_at(mm, address, page_table, entry);
3184 /* No need to invalidate - it was non-present before */
3185 update_mmu_cache(vma, address, page_table);
3187 pte_unmap_unlock(page_table, ptl);
3190 mem_cgroup_uncharge_page(page);
3191 page_cache_release(page);
3194 page_cache_release(page);
3196 return VM_FAULT_OOM;
3200 * __do_fault() tries to create a new page mapping. It aggressively
3201 * tries to share with existing pages, but makes a separate copy if
3202 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3203 * the next page fault.
3205 * As this is called only for pages that do not currently exist, we
3206 * do not need to flush old virtual caches or the TLB.
3208 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3209 * but allow concurrent faults), and pte neither mapped nor locked.
3210 * We return with mmap_sem still held, but pte unmapped and unlocked.
3212 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3213 unsigned long address, pmd_t *pmd,
3214 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3219 struct page *cow_page;
3222 struct page *dirty_page = NULL;
3223 struct vm_fault vmf;
3225 int page_mkwrite = 0;
3228 * If we do COW later, allocate page befor taking lock_page()
3229 * on the file cache page. This will reduce lock holding time.
3231 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3233 if (unlikely(anon_vma_prepare(vma)))
3234 return VM_FAULT_OOM;
3236 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3238 return VM_FAULT_OOM;
3240 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3241 page_cache_release(cow_page);
3242 return VM_FAULT_OOM;
3247 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3252 ret = vma->vm_ops->fault(vma, &vmf);
3253 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3257 if (unlikely(PageHWPoison(vmf.page))) {
3258 if (ret & VM_FAULT_LOCKED)
3259 unlock_page(vmf.page);
3260 ret = VM_FAULT_HWPOISON;
3265 * For consistency in subsequent calls, make the faulted page always
3268 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3269 lock_page(vmf.page);
3271 VM_BUG_ON(!PageLocked(vmf.page));
3274 * Should we do an early C-O-W break?
3277 if (flags & FAULT_FLAG_WRITE) {
3278 if (!(vma->vm_flags & VM_SHARED)) {
3281 copy_user_highpage(page, vmf.page, address, vma);
3282 __SetPageUptodate(page);
3285 * If the page will be shareable, see if the backing
3286 * address space wants to know that the page is about
3287 * to become writable
3289 if (vma->vm_ops->page_mkwrite) {
3293 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3294 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3296 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3298 goto unwritable_page;
3300 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3302 if (!page->mapping) {
3303 ret = 0; /* retry the fault */
3305 goto unwritable_page;
3308 VM_BUG_ON(!PageLocked(page));
3315 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3318 * This silly early PAGE_DIRTY setting removes a race
3319 * due to the bad i386 page protection. But it's valid
3320 * for other architectures too.
3322 * Note that if FAULT_FLAG_WRITE is set, we either now have
3323 * an exclusive copy of the page, or this is a shared mapping,
3324 * so we can make it writable and dirty to avoid having to
3325 * handle that later.
3327 /* Only go through if we didn't race with anybody else... */
3328 if (likely(pte_same(*page_table, orig_pte))) {
3329 flush_icache_page(vma, page);
3330 entry = mk_pte(page, vma->vm_page_prot);
3331 if (flags & FAULT_FLAG_WRITE)
3332 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3334 inc_mm_counter_fast(mm, MM_ANONPAGES);
3335 page_add_new_anon_rmap(page, vma, address);
3337 inc_mm_counter_fast(mm, MM_FILEPAGES);
3338 page_add_file_rmap(page);
3339 if (flags & FAULT_FLAG_WRITE) {
3341 get_page(dirty_page);
3344 set_pte_at(mm, address, page_table, entry);
3346 /* no need to invalidate: a not-present page won't be cached */
3347 update_mmu_cache(vma, address, page_table);
3350 mem_cgroup_uncharge_page(cow_page);
3352 page_cache_release(page);
3354 anon = 1; /* no anon but release faulted_page */
3357 pte_unmap_unlock(page_table, ptl);
3360 struct address_space *mapping = page->mapping;
3363 if (set_page_dirty(dirty_page))
3365 unlock_page(dirty_page);
3366 put_page(dirty_page);
3367 if ((dirtied || page_mkwrite) && mapping) {
3369 * Some device drivers do not set page.mapping but still
3372 balance_dirty_pages_ratelimited(mapping);
3375 /* file_update_time outside page_lock */
3376 if (vma->vm_file && !page_mkwrite)
3377 file_update_time(vma->vm_file);
3379 unlock_page(vmf.page);
3381 page_cache_release(vmf.page);
3387 page_cache_release(page);
3390 /* fs's fault handler get error */
3392 mem_cgroup_uncharge_page(cow_page);
3393 page_cache_release(cow_page);
3398 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3399 unsigned long address, pte_t *page_table, pmd_t *pmd,
3400 unsigned int flags, pte_t orig_pte)
3402 pgoff_t pgoff = (((address & PAGE_MASK)
3403 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3405 pte_unmap(page_table);
3406 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3410 * Fault of a previously existing named mapping. Repopulate the pte
3411 * from the encoded file_pte if possible. This enables swappable
3414 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3415 * but allow concurrent faults), and pte mapped but not yet locked.
3416 * We return with mmap_sem still held, but pte unmapped and unlocked.
3418 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3419 unsigned long address, pte_t *page_table, pmd_t *pmd,
3420 unsigned int flags, pte_t orig_pte)
3424 flags |= FAULT_FLAG_NONLINEAR;
3426 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3429 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3431 * Page table corrupted: show pte and kill process.
3433 print_bad_pte(vma, address, orig_pte, NULL);
3434 return VM_FAULT_SIGBUS;
3437 pgoff = pte_to_pgoff(orig_pte);
3438 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3441 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3442 unsigned long address, pte_t *ptep, pmd_t *pmd,
3443 unsigned int flags, pte_t entry)
3445 struct page *page = NULL;
3446 int node, page_nid = -1;
3450 ptl = pte_lockptr(mm, pmd);
3452 if (unlikely(!pte_same(*ptep, entry)))
3455 page = vm_normal_page(vma, address, entry);
3458 page_nid = page_to_nid(page);
3459 last_cpu = page_last_cpu(page);
3460 node = mpol_misplaced(page, vma, address);
3461 if (node != -1 && node != page_nid)
3465 out_pte_upgrade_unlock:
3466 flush_cache_page(vma, address, pte_pfn(entry));
3468 ptep_modify_prot_start(mm, address, ptep);
3469 entry = pte_modify(entry, vma->vm_page_prot);
3470 ptep_modify_prot_commit(mm, address, ptep, entry);
3472 /* No TLB flush needed because we upgraded the PTE */
3474 update_mmu_cache(vma, address, ptep);
3477 pte_unmap_unlock(ptep, ptl);
3480 task_numa_fault(page_nid, last_cpu, 1);
3487 pte_unmap_unlock(ptep, ptl);
3489 if (!migrate_misplaced_page(page, node)) {
3494 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3495 if (!pte_same(*ptep, entry)) {
3501 goto out_pte_upgrade_unlock;
3505 * These routines also need to handle stuff like marking pages dirty
3506 * and/or accessed for architectures that don't do it in hardware (most
3507 * RISC architectures). The early dirtying is also good on the i386.
3509 * There is also a hook called "update_mmu_cache()" that architectures
3510 * with external mmu caches can use to update those (ie the Sparc or
3511 * PowerPC hashed page tables that act as extended TLBs).
3513 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3514 * but allow concurrent faults), and pte mapped but not yet locked.
3515 * We return with mmap_sem still held, but pte unmapped and unlocked.
3517 int handle_pte_fault(struct mm_struct *mm,
3518 struct vm_area_struct *vma, unsigned long address,
3519 pte_t *pte, pmd_t *pmd, unsigned int flags)
3525 if (!pte_present(entry)) {
3526 if (pte_none(entry)) {
3528 if (likely(vma->vm_ops->fault))
3529 return do_linear_fault(mm, vma, address,
3530 pte, pmd, flags, entry);
3532 return do_anonymous_page(mm, vma, address,
3535 if (pte_file(entry))
3536 return do_nonlinear_fault(mm, vma, address,
3537 pte, pmd, flags, entry);
3538 return do_swap_page(mm, vma, address,
3539 pte, pmd, flags, entry);
3542 if (pte_numa(vma, entry))
3543 return do_numa_page(mm, vma, address, pte, pmd, flags, entry);
3545 ptl = pte_lockptr(mm, pmd);
3547 if (unlikely(!pte_same(*pte, entry)))
3549 if (flags & FAULT_FLAG_WRITE) {
3550 if (!pte_write(entry))
3551 return do_wp_page(mm, vma, address,
3552 pte, pmd, ptl, entry);
3553 entry = pte_mkdirty(entry);
3555 entry = pte_mkyoung(entry);
3556 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3557 update_mmu_cache(vma, address, pte);
3560 * This is needed only for protection faults but the arch code
3561 * is not yet telling us if this is a protection fault or not.
3562 * This still avoids useless tlb flushes for .text page faults
3565 if (flags & FAULT_FLAG_WRITE)
3566 flush_tlb_fix_spurious_fault(vma, address);
3569 pte_unmap_unlock(pte, ptl);
3574 * By the time we get here, we already hold the mm semaphore
3576 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3577 unsigned long address, unsigned int flags)
3584 __set_current_state(TASK_RUNNING);
3586 count_vm_event(PGFAULT);
3587 mem_cgroup_count_vm_event(mm, PGFAULT);
3589 /* do counter updates before entering really critical section. */
3590 check_sync_rss_stat(current);
3592 if (unlikely(is_vm_hugetlb_page(vma)))
3593 return hugetlb_fault(mm, vma, address, flags);
3596 pgd = pgd_offset(mm, address);
3597 pud = pud_alloc(mm, pgd, address);
3599 return VM_FAULT_OOM;
3600 pmd = pmd_alloc(mm, pud, address);
3602 return VM_FAULT_OOM;
3603 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3605 return do_huge_pmd_anonymous_page(mm, vma, address,
3608 pmd_t orig_pmd = *pmd;
3612 if (pmd_trans_huge(orig_pmd) && !pmd_trans_splitting(orig_pmd)) {
3613 if (pmd_numa(vma, orig_pmd)) {
3614 do_huge_pmd_numa_page(mm, vma, address, pmd,
3618 if ((flags & FAULT_FLAG_WRITE) && !pmd_write(orig_pmd)) {
3619 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3622 * If COW results in an oom, the huge pmd will
3623 * have been split, so retry the fault on the
3624 * pte for a smaller charge.
3626 if (unlikely(ret & VM_FAULT_OOM))
3636 * Use __pte_alloc instead of pte_alloc_map, because we can't
3637 * run pte_offset_map on the pmd, if an huge pmd could
3638 * materialize from under us from a different thread.
3640 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3641 return VM_FAULT_OOM;
3642 /* if an huge pmd materialized from under us just retry later */
3643 if (unlikely(pmd_trans_huge(*pmd)))
3646 * A regular pmd is established and it can't morph into a huge pmd
3647 * from under us anymore at this point because we hold the mmap_sem
3648 * read mode and khugepaged takes it in write mode. So now it's
3649 * safe to run pte_offset_map().
3651 pte = pte_offset_map(pmd, address);
3653 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3656 #ifndef __PAGETABLE_PUD_FOLDED
3658 * Allocate page upper directory.
3659 * We've already handled the fast-path in-line.
3661 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3663 pud_t *new = pud_alloc_one(mm, address);
3667 smp_wmb(); /* See comment in __pte_alloc */
3669 spin_lock(&mm->page_table_lock);
3670 if (pgd_present(*pgd)) /* Another has populated it */
3673 pgd_populate(mm, pgd, new);
3674 spin_unlock(&mm->page_table_lock);
3677 #endif /* __PAGETABLE_PUD_FOLDED */
3679 #ifndef __PAGETABLE_PMD_FOLDED
3681 * Allocate page middle directory.
3682 * We've already handled the fast-path in-line.
3684 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3686 pmd_t *new = pmd_alloc_one(mm, address);
3690 smp_wmb(); /* See comment in __pte_alloc */
3692 spin_lock(&mm->page_table_lock);
3693 #ifndef __ARCH_HAS_4LEVEL_HACK
3694 if (pud_present(*pud)) /* Another has populated it */
3697 pud_populate(mm, pud, new);
3699 if (pgd_present(*pud)) /* Another has populated it */
3702 pgd_populate(mm, pud, new);
3703 #endif /* __ARCH_HAS_4LEVEL_HACK */
3704 spin_unlock(&mm->page_table_lock);
3707 #endif /* __PAGETABLE_PMD_FOLDED */
3709 int make_pages_present(unsigned long addr, unsigned long end)
3711 int ret, len, write;
3712 struct vm_area_struct * vma;
3714 vma = find_vma(current->mm, addr);
3718 * We want to touch writable mappings with a write fault in order
3719 * to break COW, except for shared mappings because these don't COW
3720 * and we would not want to dirty them for nothing.
3722 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3723 BUG_ON(addr >= end);
3724 BUG_ON(end > vma->vm_end);
3725 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3726 ret = get_user_pages(current, current->mm, addr,
3727 len, write, 0, NULL, NULL);
3730 return ret == len ? 0 : -EFAULT;
3733 #if !defined(__HAVE_ARCH_GATE_AREA)
3735 #if defined(AT_SYSINFO_EHDR)
3736 static struct vm_area_struct gate_vma;
3738 static int __init gate_vma_init(void)
3740 gate_vma.vm_mm = NULL;
3741 gate_vma.vm_start = FIXADDR_USER_START;
3742 gate_vma.vm_end = FIXADDR_USER_END;
3743 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3744 gate_vma.vm_page_prot = __P101;
3748 __initcall(gate_vma_init);
3751 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3753 #ifdef AT_SYSINFO_EHDR
3760 int in_gate_area_no_mm(unsigned long addr)
3762 #ifdef AT_SYSINFO_EHDR
3763 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3769 #endif /* __HAVE_ARCH_GATE_AREA */
3771 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3772 pte_t **ptepp, spinlock_t **ptlp)
3779 pgd = pgd_offset(mm, address);
3780 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3783 pud = pud_offset(pgd, address);
3784 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3787 pmd = pmd_offset(pud, address);
3788 VM_BUG_ON(pmd_trans_huge(*pmd));
3789 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3792 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3796 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3799 if (!pte_present(*ptep))
3804 pte_unmap_unlock(ptep, *ptlp);
3809 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3810 pte_t **ptepp, spinlock_t **ptlp)
3814 /* (void) is needed to make gcc happy */
3815 (void) __cond_lock(*ptlp,
3816 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3821 * follow_pfn - look up PFN at a user virtual address
3822 * @vma: memory mapping
3823 * @address: user virtual address
3824 * @pfn: location to store found PFN
3826 * Only IO mappings and raw PFN mappings are allowed.
3828 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3830 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3837 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3840 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3843 *pfn = pte_pfn(*ptep);
3844 pte_unmap_unlock(ptep, ptl);
3847 EXPORT_SYMBOL(follow_pfn);
3849 #ifdef CONFIG_HAVE_IOREMAP_PROT
3850 int follow_phys(struct vm_area_struct *vma,
3851 unsigned long address, unsigned int flags,
3852 unsigned long *prot, resource_size_t *phys)
3858 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3861 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3865 if ((flags & FOLL_WRITE) && !pte_write(pte))
3868 *prot = pgprot_val(pte_pgprot(pte));
3869 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3873 pte_unmap_unlock(ptep, ptl);
3878 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3879 void *buf, int len, int write)
3881 resource_size_t phys_addr;
3882 unsigned long prot = 0;
3883 void __iomem *maddr;
3884 int offset = addr & (PAGE_SIZE-1);
3886 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3889 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3891 memcpy_toio(maddr + offset, buf, len);
3893 memcpy_fromio(buf, maddr + offset, len);
3901 * Access another process' address space as given in mm. If non-NULL, use the
3902 * given task for page fault accounting.
3904 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3905 unsigned long addr, void *buf, int len, int write)
3907 struct vm_area_struct *vma;
3908 void *old_buf = buf;
3910 down_read(&mm->mmap_sem);
3911 /* ignore errors, just check how much was successfully transferred */
3913 int bytes, ret, offset;
3915 struct page *page = NULL;
3917 ret = get_user_pages(tsk, mm, addr, 1,
3918 write, 1, &page, &vma);
3921 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3922 * we can access using slightly different code.
3924 #ifdef CONFIG_HAVE_IOREMAP_PROT
3925 vma = find_vma(mm, addr);
3926 if (!vma || vma->vm_start > addr)
3928 if (vma->vm_ops && vma->vm_ops->access)
3929 ret = vma->vm_ops->access(vma, addr, buf,
3937 offset = addr & (PAGE_SIZE-1);
3938 if (bytes > PAGE_SIZE-offset)
3939 bytes = PAGE_SIZE-offset;
3943 copy_to_user_page(vma, page, addr,
3944 maddr + offset, buf, bytes);
3945 set_page_dirty_lock(page);
3947 copy_from_user_page(vma, page, addr,
3948 buf, maddr + offset, bytes);
3951 page_cache_release(page);
3957 up_read(&mm->mmap_sem);
3959 return buf - old_buf;
3963 * access_remote_vm - access another process' address space
3964 * @mm: the mm_struct of the target address space
3965 * @addr: start address to access
3966 * @buf: source or destination buffer
3967 * @len: number of bytes to transfer
3968 * @write: whether the access is a write
3970 * The caller must hold a reference on @mm.
3972 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3973 void *buf, int len, int write)
3975 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3979 * Access another process' address space.
3980 * Source/target buffer must be kernel space,
3981 * Do not walk the page table directly, use get_user_pages
3983 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3984 void *buf, int len, int write)
3986 struct mm_struct *mm;
3989 mm = get_task_mm(tsk);
3993 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4000 * Print the name of a VMA.
4002 void print_vma_addr(char *prefix, unsigned long ip)
4004 struct mm_struct *mm = current->mm;
4005 struct vm_area_struct *vma;
4008 * Do not print if we are in atomic
4009 * contexts (in exception stacks, etc.):
4011 if (preempt_count())
4014 down_read(&mm->mmap_sem);
4015 vma = find_vma(mm, ip);
4016 if (vma && vma->vm_file) {
4017 struct file *f = vma->vm_file;
4018 char *buf = (char *)__get_free_page(GFP_KERNEL);
4022 p = d_path(&f->f_path, buf, PAGE_SIZE);
4025 s = strrchr(p, '/');
4028 printk("%s%s[%lx+%lx]", prefix, p,
4030 vma->vm_end - vma->vm_start);
4031 free_page((unsigned long)buf);
4034 up_read(&mm->mmap_sem);
4037 #ifdef CONFIG_PROVE_LOCKING
4038 void might_fault(void)
4041 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4042 * holding the mmap_sem, this is safe because kernel memory doesn't
4043 * get paged out, therefore we'll never actually fault, and the
4044 * below annotations will generate false positives.
4046 if (segment_eq(get_fs(), KERNEL_DS))
4051 * it would be nicer only to annotate paths which are not under
4052 * pagefault_disable, however that requires a larger audit and
4053 * providing helpers like get_user_atomic.
4055 if (!in_atomic() && current->mm)
4056 might_lock_read(¤t->mm->mmap_sem);
4058 EXPORT_SYMBOL(might_fault);
4061 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4062 static void clear_gigantic_page(struct page *page,
4064 unsigned int pages_per_huge_page)
4067 struct page *p = page;
4070 for (i = 0; i < pages_per_huge_page;
4071 i++, p = mem_map_next(p, page, i)) {
4073 clear_user_highpage(p, addr + i * PAGE_SIZE);
4076 void clear_huge_page(struct page *page,
4077 unsigned long addr, unsigned int pages_per_huge_page)
4081 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4082 clear_gigantic_page(page, addr, pages_per_huge_page);
4087 for (i = 0; i < pages_per_huge_page; i++) {
4089 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4093 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4095 struct vm_area_struct *vma,
4096 unsigned int pages_per_huge_page)
4099 struct page *dst_base = dst;
4100 struct page *src_base = src;
4102 for (i = 0; i < pages_per_huge_page; ) {
4104 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4107 dst = mem_map_next(dst, dst_base, i);
4108 src = mem_map_next(src, src_base, i);
4112 void copy_user_huge_page(struct page *dst, struct page *src,
4113 unsigned long addr, struct vm_area_struct *vma,
4114 unsigned int pages_per_huge_page)
4118 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4119 copy_user_gigantic_page(dst, src, addr, vma,
4120 pages_per_huge_page);
4125 for (i = 0; i < pages_per_huge_page; i++) {
4127 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4130 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */