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>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
79 unsigned long num_physpages;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
105 static int __init disable_randmaps(char *s)
107 randomize_va_space = 0;
110 __setup("norandmaps", disable_randmaps);
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init init_zero_pfn(void)
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
123 core_initcall(init_zero_pfn);
126 #if defined(SPLIT_RSS_COUNTING)
128 void sync_mm_rss(struct mm_struct *mm)
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (current->rss_stat.count[i]) {
134 add_mm_counter(mm, i, current->rss_stat.count[i]);
135 current->rss_stat.count[i] = 0;
138 current->rss_stat.events = 0;
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 struct task_struct *task = current;
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
148 add_mm_counter(mm, member, val);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct *task)
157 if (unlikely(task != current))
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 sync_mm_rss(task->mm);
162 #else /* SPLIT_RSS_COUNTING */
164 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
165 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
167 static void check_sync_rss_stat(struct task_struct *task)
171 #endif /* SPLIT_RSS_COUNTING */
173 #ifdef HAVE_GENERIC_MMU_GATHER
175 static int tlb_next_batch(struct mmu_gather *tlb)
177 struct mmu_gather_batch *batch;
181 tlb->active = batch->next;
185 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
191 batch->max = MAX_GATHER_BATCH;
193 tlb->active->next = batch;
200 * Called to initialize an (on-stack) mmu_gather structure for page-table
201 * tear-down from @mm. The @fullmm argument is used when @mm is without
202 * users and we're going to destroy the full address space (exit/execve).
204 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
208 tlb->fullmm = fullmm;
210 tlb->fast_mode = (num_possible_cpus() == 1);
211 tlb->local.next = NULL;
213 tlb->local.max = ARRAY_SIZE(tlb->__pages);
214 tlb->active = &tlb->local;
216 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
221 void tlb_flush_mmu(struct mmu_gather *tlb)
223 struct mmu_gather_batch *batch;
225 if (!tlb->need_flush)
229 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb_table_flush(tlb);
233 if (tlb_fast_mode(tlb))
236 for (batch = &tlb->local; batch; batch = batch->next) {
237 free_pages_and_swap_cache(batch->pages, batch->nr);
240 tlb->active = &tlb->local;
244 * Called at the end of the shootdown operation to free up any resources
245 * that were required.
247 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
249 struct mmu_gather_batch *batch, *next;
253 /* keep the page table cache within bounds */
256 for (batch = tlb->local.next; batch; batch = next) {
258 free_pages((unsigned long)batch, 0);
260 tlb->local.next = NULL;
264 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
265 * handling the additional races in SMP caused by other CPUs caching valid
266 * mappings in their TLBs. Returns the number of free page slots left.
267 * When out of page slots we must call tlb_flush_mmu().
269 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
271 struct mmu_gather_batch *batch;
273 VM_BUG_ON(!tlb->need_flush);
275 if (tlb_fast_mode(tlb)) {
276 free_page_and_swap_cache(page);
277 return 1; /* avoid calling tlb_flush_mmu() */
281 batch->pages[batch->nr++] = page;
282 if (batch->nr == batch->max) {
283 if (!tlb_next_batch(tlb))
287 VM_BUG_ON(batch->nr > batch->max);
289 return batch->max - batch->nr;
292 #endif /* HAVE_GENERIC_MMU_GATHER */
294 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
297 * See the comment near struct mmu_table_batch.
300 static void tlb_remove_table_smp_sync(void *arg)
302 /* Simply deliver the interrupt */
305 static void tlb_remove_table_one(void *table)
308 * This isn't an RCU grace period and hence the page-tables cannot be
309 * assumed to be actually RCU-freed.
311 * It is however sufficient for software page-table walkers that rely on
312 * IRQ disabling. See the comment near struct mmu_table_batch.
314 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
315 __tlb_remove_table(table);
318 static void tlb_remove_table_rcu(struct rcu_head *head)
320 struct mmu_table_batch *batch;
323 batch = container_of(head, struct mmu_table_batch, rcu);
325 for (i = 0; i < batch->nr; i++)
326 __tlb_remove_table(batch->tables[i]);
328 free_page((unsigned long)batch);
331 void tlb_table_flush(struct mmu_gather *tlb)
333 struct mmu_table_batch **batch = &tlb->batch;
336 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
341 void tlb_remove_table(struct mmu_gather *tlb, void *table)
343 struct mmu_table_batch **batch = &tlb->batch;
348 * When there's less then two users of this mm there cannot be a
349 * concurrent page-table walk.
351 if (atomic_read(&tlb->mm->mm_users) < 2) {
352 __tlb_remove_table(table);
356 if (*batch == NULL) {
357 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
358 if (*batch == NULL) {
359 tlb_remove_table_one(table);
364 (*batch)->tables[(*batch)->nr++] = table;
365 if ((*batch)->nr == MAX_TABLE_BATCH)
366 tlb_table_flush(tlb);
369 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
372 * If a p?d_bad entry is found while walking page tables, report
373 * the error, before resetting entry to p?d_none. Usually (but
374 * very seldom) called out from the p?d_none_or_clear_bad macros.
377 void pgd_clear_bad(pgd_t *pgd)
383 void pud_clear_bad(pud_t *pud)
389 void pmd_clear_bad(pmd_t *pmd)
396 * Note: this doesn't free the actual pages themselves. That
397 * has been handled earlier when unmapping all the memory regions.
399 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
402 pgtable_t token = pmd_pgtable(*pmd);
404 pte_free_tlb(tlb, token, addr);
408 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
409 unsigned long addr, unsigned long end,
410 unsigned long floor, unsigned long ceiling)
417 pmd = pmd_offset(pud, addr);
419 next = pmd_addr_end(addr, end);
420 if (pmd_none_or_clear_bad(pmd))
422 free_pte_range(tlb, pmd, addr);
423 } while (pmd++, addr = next, addr != end);
433 if (end - 1 > ceiling - 1)
436 pmd = pmd_offset(pud, start);
438 pmd_free_tlb(tlb, pmd, start);
441 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
450 pud = pud_offset(pgd, addr);
452 next = pud_addr_end(addr, end);
453 if (pud_none_or_clear_bad(pud))
455 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
456 } while (pud++, addr = next, addr != end);
462 ceiling &= PGDIR_MASK;
466 if (end - 1 > ceiling - 1)
469 pud = pud_offset(pgd, start);
471 pud_free_tlb(tlb, pud, start);
475 * This function frees user-level page tables of a process.
477 * Must be called with pagetable lock held.
479 void free_pgd_range(struct mmu_gather *tlb,
480 unsigned long addr, unsigned long end,
481 unsigned long floor, unsigned long ceiling)
487 * The next few lines have given us lots of grief...
489 * Why are we testing PMD* at this top level? Because often
490 * there will be no work to do at all, and we'd prefer not to
491 * go all the way down to the bottom just to discover that.
493 * Why all these "- 1"s? Because 0 represents both the bottom
494 * of the address space and the top of it (using -1 for the
495 * top wouldn't help much: the masks would do the wrong thing).
496 * The rule is that addr 0 and floor 0 refer to the bottom of
497 * the address space, but end 0 and ceiling 0 refer to the top
498 * Comparisons need to use "end - 1" and "ceiling - 1" (though
499 * that end 0 case should be mythical).
501 * Wherever addr is brought up or ceiling brought down, we must
502 * be careful to reject "the opposite 0" before it confuses the
503 * subsequent tests. But what about where end is brought down
504 * by PMD_SIZE below? no, end can't go down to 0 there.
506 * Whereas we round start (addr) and ceiling down, by different
507 * masks at different levels, in order to test whether a table
508 * now has no other vmas using it, so can be freed, we don't
509 * bother to round floor or end up - the tests don't need that.
523 if (end - 1 > ceiling - 1)
528 pgd = pgd_offset(tlb->mm, addr);
530 next = pgd_addr_end(addr, end);
531 if (pgd_none_or_clear_bad(pgd))
533 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
534 } while (pgd++, addr = next, addr != end);
537 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
538 unsigned long floor, unsigned long ceiling)
541 struct vm_area_struct *next = vma->vm_next;
542 unsigned long addr = vma->vm_start;
545 * Hide vma from rmap and truncate_pagecache before freeing
548 unlink_anon_vmas(vma);
549 unlink_file_vma(vma);
551 if (is_vm_hugetlb_page(vma)) {
552 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
553 floor, next? next->vm_start: ceiling);
556 * Optimization: gather nearby vmas into one call down
558 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
559 && !is_vm_hugetlb_page(next)) {
562 unlink_anon_vmas(vma);
563 unlink_file_vma(vma);
565 free_pgd_range(tlb, addr, vma->vm_end,
566 floor, next? next->vm_start: ceiling);
572 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
573 pmd_t *pmd, unsigned long address)
575 pgtable_t new = pte_alloc_one(mm, address);
576 int wait_split_huge_page;
581 * Ensure all pte setup (eg. pte page lock and page clearing) are
582 * visible before the pte is made visible to other CPUs by being
583 * put into page tables.
585 * The other side of the story is the pointer chasing in the page
586 * table walking code (when walking the page table without locking;
587 * ie. most of the time). Fortunately, these data accesses consist
588 * of a chain of data-dependent loads, meaning most CPUs (alpha
589 * being the notable exception) will already guarantee loads are
590 * seen in-order. See the alpha page table accessors for the
591 * smp_read_barrier_depends() barriers in page table walking code.
593 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
595 spin_lock(&mm->page_table_lock);
596 wait_split_huge_page = 0;
597 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
599 pmd_populate(mm, pmd, new);
601 } else if (unlikely(pmd_trans_splitting(*pmd)))
602 wait_split_huge_page = 1;
603 spin_unlock(&mm->page_table_lock);
606 if (wait_split_huge_page)
607 wait_split_huge_page(vma->anon_vma, pmd);
611 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
613 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
617 smp_wmb(); /* See comment in __pte_alloc */
619 spin_lock(&init_mm.page_table_lock);
620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
621 pmd_populate_kernel(&init_mm, pmd, new);
624 VM_BUG_ON(pmd_trans_splitting(*pmd));
625 spin_unlock(&init_mm.page_table_lock);
627 pte_free_kernel(&init_mm, new);
631 static inline void init_rss_vec(int *rss)
633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
636 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
640 if (current->mm == mm)
642 for (i = 0; i < NR_MM_COUNTERS; i++)
644 add_mm_counter(mm, i, rss[i]);
648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
652 * The calling function must still handle the error.
654 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655 pte_t pte, struct page *page)
657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658 pud_t *pud = pud_offset(pgd, addr);
659 pmd_t *pmd = pmd_offset(pud, addr);
660 struct address_space *mapping;
662 static unsigned long resume;
663 static unsigned long nr_shown;
664 static unsigned long nr_unshown;
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
670 if (nr_shown == 60) {
671 if (time_before(jiffies, resume)) {
677 "BUG: Bad page map: %lu messages suppressed\n",
684 resume = jiffies + 60 * HZ;
686 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
687 index = linear_page_index(vma, addr);
690 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
692 (long long)pte_val(pte), (long long)pmd_val(*pmd));
696 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
702 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
703 (unsigned long)vma->vm_ops->fault);
704 if (vma->vm_file && vma->vm_file->f_op)
705 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
706 (unsigned long)vma->vm_file->f_op->mmap);
708 add_taint(TAINT_BAD_PAGE);
711 static inline int is_cow_mapping(vm_flags_t flags)
713 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
717 static inline int is_zero_pfn(unsigned long pfn)
719 return pfn == zero_pfn;
724 static inline unsigned long my_zero_pfn(unsigned long addr)
731 * vm_normal_page -- This function gets the "struct page" associated with a pte.
733 * "Special" mappings do not wish to be associated with a "struct page" (either
734 * it doesn't exist, or it exists but they don't want to touch it). In this
735 * case, NULL is returned here. "Normal" mappings do have a struct page.
737 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
738 * pte bit, in which case this function is trivial. Secondly, an architecture
739 * may not have a spare pte bit, which requires a more complicated scheme,
742 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
743 * special mapping (even if there are underlying and valid "struct pages").
744 * COWed pages of a VM_PFNMAP are always normal.
746 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
747 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
748 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
749 * mapping will always honor the rule
751 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
753 * And for normal mappings this is false.
755 * This restricts such mappings to be a linear translation from virtual address
756 * to pfn. To get around this restriction, we allow arbitrary mappings so long
757 * as the vma is not a COW mapping; in that case, we know that all ptes are
758 * special (because none can have been COWed).
761 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
763 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
764 * page" backing, however the difference is that _all_ pages with a struct
765 * page (that is, those where pfn_valid is true) are refcounted and considered
766 * normal pages by the VM. The disadvantage is that pages are refcounted
767 * (which can be slower and simply not an option for some PFNMAP users). The
768 * advantage is that we don't have to follow the strict linearity rule of
769 * PFNMAP mappings in order to support COWable mappings.
772 #ifdef __HAVE_ARCH_PTE_SPECIAL
773 # define HAVE_PTE_SPECIAL 1
775 # define HAVE_PTE_SPECIAL 0
777 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
780 unsigned long pfn = pte_pfn(pte);
782 #if defined(CONFIG_XEN) && defined(CONFIG_X86)
783 /* XEN: Covers user-space grant mappings (even of local pages). */
784 if (unlikely(vma->vm_flags & VM_FOREIGN))
788 if (HAVE_PTE_SPECIAL) {
789 if (likely(!pte_special(pte)))
791 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
793 if (!is_zero_pfn(pfn))
794 print_bad_pte(vma, addr, pte, NULL);
798 /* !HAVE_PTE_SPECIAL case follows: */
800 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
801 if (vma->vm_flags & VM_MIXEDMAP) {
807 off = (addr - vma->vm_start) >> PAGE_SHIFT;
808 if (pfn == vma->vm_pgoff + off)
810 if (!is_cow_mapping(vma->vm_flags))
815 if (is_zero_pfn(pfn))
818 if (unlikely(pfn > highest_memmap_pfn)) {
820 if (!(vma->vm_flags & VM_RESERVED))
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;
1050 * Don't copy ptes where a page fault will fill them correctly.
1051 * Fork becomes much lighter when there are big shared or private
1052 * readonly mappings. The tradeoff is that copy_page_range is more
1053 * efficient than faulting.
1055 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1060 if (is_vm_hugetlb_page(vma))
1061 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1063 if (unlikely(is_pfn_mapping(vma))) {
1065 * We do not free on error cases below as remove_vma
1066 * gets called on error from higher level routine
1068 ret = track_pfn_vma_copy(vma);
1074 * We need to invalidate the secondary MMU mappings only when
1075 * there could be a permission downgrade on the ptes of the
1076 * parent mm. And a permission downgrade will only happen if
1077 * is_cow_mapping() returns true.
1079 if (is_cow_mapping(vma->vm_flags))
1080 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1083 dst_pgd = pgd_offset(dst_mm, addr);
1084 src_pgd = pgd_offset(src_mm, addr);
1086 next = pgd_addr_end(addr, end);
1087 if (pgd_none_or_clear_bad(src_pgd))
1089 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1090 vma, addr, next))) {
1094 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1096 if (is_cow_mapping(vma->vm_flags))
1097 mmu_notifier_invalidate_range_end(src_mm,
1098 vma->vm_start, end);
1102 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1103 struct vm_area_struct *vma, pmd_t *pmd,
1104 unsigned long addr, unsigned long end,
1105 struct zap_details *details)
1107 struct mm_struct *mm = tlb->mm;
1108 int force_flush = 0;
1109 int rss[NR_MM_COUNTERS];
1116 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1118 arch_enter_lazy_mmu_mode();
1121 if (pte_none(ptent)) {
1125 if (pte_present(ptent)) {
1128 page = vm_normal_page(vma, addr, ptent);
1129 if (unlikely(details) && page) {
1131 * unmap_shared_mapping_pages() wants to
1132 * invalidate cache without truncating:
1133 * unmap shared but keep private pages.
1135 if (details->check_mapping &&
1136 details->check_mapping != page->mapping)
1139 * Each page->index must be checked when
1140 * invalidating or truncating nonlinear.
1142 if (details->nonlinear_vma &&
1143 (page->index < details->first_index ||
1144 page->index > details->last_index))
1148 if (unlikely(vma->vm_ops && vma->vm_ops->zap_pte))
1149 ptent = vma->vm_ops->zap_pte(vma, addr, pte,
1153 ptent = ptep_get_and_clear_full(mm, addr, pte,
1155 tlb_remove_tlb_entry(tlb, pte, addr);
1156 if (unlikely(!page))
1158 if (unlikely(details) && details->nonlinear_vma
1159 && linear_page_index(details->nonlinear_vma,
1160 addr) != page->index)
1161 set_pte_at(mm, addr, pte,
1162 pgoff_to_pte(page->index));
1164 rss[MM_ANONPAGES]--;
1166 if (pte_dirty(ptent))
1167 set_page_dirty(page);
1168 if (pte_young(ptent) &&
1169 likely(!VM_SequentialReadHint(vma)))
1170 mark_page_accessed(page);
1171 rss[MM_FILEPAGES]--;
1173 page_remove_rmap(page);
1174 if (unlikely(page_mapcount(page) < 0))
1175 print_bad_pte(vma, addr, ptent, page);
1176 force_flush = !__tlb_remove_page(tlb, page);
1182 * If details->check_mapping, we leave swap entries;
1183 * if details->nonlinear_vma, we leave file entries.
1185 if (unlikely(details))
1187 if (pte_file(ptent)) {
1188 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1189 print_bad_pte(vma, addr, ptent, NULL);
1191 swp_entry_t entry = pte_to_swp_entry(ptent);
1193 if (!non_swap_entry(entry))
1195 else if (is_migration_entry(entry)) {
1198 page = migration_entry_to_page(entry);
1201 rss[MM_ANONPAGES]--;
1203 rss[MM_FILEPAGES]--;
1205 if (unlikely(!free_swap_and_cache(entry)))
1206 print_bad_pte(vma, addr, ptent, NULL);
1208 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1209 } while (pte++, addr += PAGE_SIZE, addr != end);
1211 add_mm_rss_vec(mm, rss);
1212 arch_leave_lazy_mmu_mode();
1213 pte_unmap_unlock(start_pte, ptl);
1216 * mmu_gather ran out of room to batch pages, we break out of
1217 * the PTE lock to avoid doing the potential expensive TLB invalidate
1218 * and page-free while holding it.
1230 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1231 struct vm_area_struct *vma, pud_t *pud,
1232 unsigned long addr, unsigned long end,
1233 struct zap_details *details)
1238 pmd = pmd_offset(pud, addr);
1240 next = pmd_addr_end(addr, end);
1241 if (pmd_trans_huge(*pmd)) {
1242 if (next - addr != HPAGE_PMD_SIZE) {
1243 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1244 split_huge_page_pmd(vma->vm_mm, pmd);
1245 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1250 * Here there can be other concurrent MADV_DONTNEED or
1251 * trans huge page faults running, and if the pmd is
1252 * none or trans huge it can change under us. This is
1253 * because MADV_DONTNEED holds the mmap_sem in read
1256 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1258 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261 } while (pmd++, addr = next, addr != end);
1266 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1267 struct vm_area_struct *vma, pgd_t *pgd,
1268 unsigned long addr, unsigned long end,
1269 struct zap_details *details)
1274 pud = pud_offset(pgd, addr);
1276 next = pud_addr_end(addr, end);
1277 if (pud_none_or_clear_bad(pud))
1279 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1280 } while (pud++, addr = next, addr != end);
1285 static void unmap_page_range(struct mmu_gather *tlb,
1286 struct vm_area_struct *vma,
1287 unsigned long addr, unsigned long end,
1288 struct zap_details *details)
1293 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 BUG_ON(addr >= end);
1297 mem_cgroup_uncharge_start();
1298 tlb_start_vma(tlb, vma);
1299 pgd = pgd_offset(vma->vm_mm, addr);
1301 next = pgd_addr_end(addr, end);
1302 if (pgd_none_or_clear_bad(pgd))
1304 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1305 } while (pgd++, addr = next, addr != end);
1306 tlb_end_vma(tlb, vma);
1307 mem_cgroup_uncharge_end();
1311 static void unmap_single_vma(struct mmu_gather *tlb,
1312 struct vm_area_struct *vma, unsigned long start_addr,
1313 unsigned long end_addr, unsigned long *nr_accounted,
1314 struct zap_details *details)
1316 unsigned long start = max(vma->vm_start, start_addr);
1319 if (start >= vma->vm_end)
1321 end = min(vma->vm_end, end_addr);
1322 if (end <= vma->vm_start)
1325 if (vma->vm_flags & VM_ACCOUNT)
1326 *nr_accounted += (end - start) >> PAGE_SHIFT;
1328 if (unlikely(is_pfn_mapping(vma)))
1329 untrack_pfn_vma(vma, 0, 0);
1332 if (unlikely(is_vm_hugetlb_page(vma))) {
1334 * It is undesirable to test vma->vm_file as it
1335 * should be non-null for valid hugetlb area.
1336 * However, vm_file will be NULL in the error
1337 * cleanup path of do_mmap_pgoff. When
1338 * hugetlbfs ->mmap method fails,
1339 * do_mmap_pgoff() nullifies vma->vm_file
1340 * before calling this function to clean up.
1341 * Since no pte has actually been setup, it is
1342 * safe to do nothing in this case.
1345 unmap_hugepage_range(vma, start, end, NULL);
1347 unmap_page_range(tlb, vma, start, end, details);
1352 * unmap_vmas - unmap a range of memory covered by a list of vma's
1353 * @tlb: address of the caller's struct mmu_gather
1354 * @vma: the starting vma
1355 * @start_addr: virtual address at which to start unmapping
1356 * @end_addr: virtual address at which to end unmapping
1357 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1358 * @details: details of nonlinear truncation or shared cache invalidation
1360 * Unmap all pages in the vma list.
1362 * Only addresses between `start' and `end' will be unmapped.
1364 * The VMA list must be sorted in ascending virtual address order.
1366 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1367 * range after unmap_vmas() returns. So the only responsibility here is to
1368 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1369 * drops the lock and schedules.
1371 void unmap_vmas(struct mmu_gather *tlb,
1372 struct vm_area_struct *vma, unsigned long start_addr,
1373 unsigned long end_addr, unsigned long *nr_accounted,
1374 struct zap_details *details)
1376 struct mm_struct *mm = vma->vm_mm;
1378 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1379 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1380 unmap_single_vma(tlb, vma, start_addr, end_addr, nr_accounted,
1382 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1386 * zap_page_range - remove user pages in a given range
1387 * @vma: vm_area_struct holding the applicable pages
1388 * @address: starting address of pages to zap
1389 * @size: number of bytes to zap
1390 * @details: details of nonlinear truncation or shared cache invalidation
1392 * Caller must protect the VMA list
1394 void zap_page_range(struct vm_area_struct *vma, unsigned long address,
1395 unsigned long size, struct zap_details *details)
1397 struct mm_struct *mm = vma->vm_mm;
1398 struct mmu_gather tlb;
1399 unsigned long end = address + size;
1400 unsigned long nr_accounted = 0;
1403 tlb_gather_mmu(&tlb, mm, 0);
1404 update_hiwater_rss(mm);
1405 unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1406 tlb_finish_mmu(&tlb, address, end);
1408 EXPORT_SYMBOL(zap_page_range);
1411 * zap_page_range_single - remove user pages in a given range
1412 * @vma: vm_area_struct holding the applicable pages
1413 * @address: starting address of pages to zap
1414 * @size: number of bytes to zap
1415 * @details: details of nonlinear truncation or shared cache invalidation
1417 * The range must fit into one VMA.
1419 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1420 unsigned long size, struct zap_details *details)
1422 struct mm_struct *mm = vma->vm_mm;
1423 struct mmu_gather tlb;
1424 unsigned long end = address + size;
1425 unsigned long nr_accounted = 0;
1428 tlb_gather_mmu(&tlb, mm, 0);
1429 update_hiwater_rss(mm);
1430 mmu_notifier_invalidate_range_start(mm, address, end);
1431 unmap_single_vma(&tlb, vma, address, end, &nr_accounted, details);
1432 mmu_notifier_invalidate_range_end(mm, address, end);
1433 tlb_finish_mmu(&tlb, address, end);
1437 * zap_vma_ptes - remove ptes mapping the vma
1438 * @vma: vm_area_struct holding ptes to be zapped
1439 * @address: starting address of pages to zap
1440 * @size: number of bytes to zap
1442 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 * The entire address range must be fully contained within the vma.
1446 * Returns 0 if successful.
1448 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1451 if (address < vma->vm_start || address + size > vma->vm_end ||
1452 !(vma->vm_flags & VM_PFNMAP))
1454 zap_page_range_single(vma, address, size, NULL);
1457 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1460 * follow_page - look up a page descriptor from a user-virtual address
1461 * @vma: vm_area_struct mapping @address
1462 * @address: virtual address to look up
1463 * @flags: flags modifying lookup behaviour
1465 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1467 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1468 * an error pointer if there is a mapping to something not represented
1469 * by a page descriptor (see also vm_normal_page()).
1471 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1480 struct mm_struct *mm = vma->vm_mm;
1482 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1483 if (!IS_ERR(page)) {
1484 BUG_ON(flags & FOLL_GET);
1489 pgd = pgd_offset(mm, address);
1490 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1493 pud = pud_offset(pgd, address);
1496 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1497 BUG_ON(flags & FOLL_GET);
1498 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1501 if (unlikely(pud_bad(*pud)))
1504 pmd = pmd_offset(pud, address);
1507 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1508 BUG_ON(flags & FOLL_GET);
1509 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1512 if (pmd_trans_huge(*pmd)) {
1513 if (flags & FOLL_SPLIT) {
1514 split_huge_page_pmd(mm, pmd);
1515 goto split_fallthrough;
1517 spin_lock(&mm->page_table_lock);
1518 if (likely(pmd_trans_huge(*pmd))) {
1519 if (unlikely(pmd_trans_splitting(*pmd))) {
1520 spin_unlock(&mm->page_table_lock);
1521 wait_split_huge_page(vma->anon_vma, pmd);
1523 page = follow_trans_huge_pmd(mm, address,
1525 spin_unlock(&mm->page_table_lock);
1529 spin_unlock(&mm->page_table_lock);
1533 if (unlikely(pmd_bad(*pmd)))
1536 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1539 if (!pte_present(pte))
1541 if ((flags & FOLL_WRITE) && !pte_write(pte))
1544 page = vm_normal_page(vma, address, pte);
1545 if (unlikely(!page)) {
1546 if ((flags & FOLL_DUMP) ||
1547 !is_zero_pfn(pte_pfn(pte)))
1549 page = pte_page(pte);
1552 if (flags & FOLL_GET)
1553 get_page_foll(page);
1554 if (flags & FOLL_TOUCH) {
1555 if ((flags & FOLL_WRITE) &&
1556 !pte_dirty(pte) && !PageDirty(page))
1557 set_page_dirty(page);
1559 * pte_mkyoung() would be more correct here, but atomic care
1560 * is needed to avoid losing the dirty bit: it is easier to use
1561 * mark_page_accessed().
1563 mark_page_accessed(page);
1565 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1567 * The preliminary mapping check is mainly to avoid the
1568 * pointless overhead of lock_page on the ZERO_PAGE
1569 * which might bounce very badly if there is contention.
1571 * If the page is already locked, we don't need to
1572 * handle it now - vmscan will handle it later if and
1573 * when it attempts to reclaim the page.
1575 if (page->mapping && trylock_page(page)) {
1576 lru_add_drain(); /* push cached pages to LRU */
1578 * Because we lock page here and migration is
1579 * blocked by the pte's page reference, we need
1580 * only check for file-cache page truncation.
1583 mlock_vma_page(page);
1588 pte_unmap_unlock(ptep, ptl);
1593 pte_unmap_unlock(ptep, ptl);
1594 return ERR_PTR(-EFAULT);
1597 pte_unmap_unlock(ptep, ptl);
1603 * When core dumping an enormous anonymous area that nobody
1604 * has touched so far, we don't want to allocate unnecessary pages or
1605 * page tables. Return error instead of NULL to skip handle_mm_fault,
1606 * then get_dump_page() will return NULL to leave a hole in the dump.
1607 * But we can only make this optimization where a hole would surely
1608 * be zero-filled if handle_mm_fault() actually did handle it.
1610 if ((flags & FOLL_DUMP) &&
1611 (!vma->vm_ops || !vma->vm_ops->fault))
1612 return ERR_PTR(-EFAULT);
1616 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1618 return stack_guard_page_start(vma, addr) ||
1619 stack_guard_page_end(vma, addr+PAGE_SIZE);
1623 * __get_user_pages() - pin user pages in memory
1624 * @tsk: task_struct of target task
1625 * @mm: mm_struct of target mm
1626 * @start: starting user address
1627 * @nr_pages: number of pages from start to pin
1628 * @gup_flags: flags modifying pin behaviour
1629 * @pages: array that receives pointers to the pages pinned.
1630 * Should be at least nr_pages long. Or NULL, if caller
1631 * only intends to ensure the pages are faulted in.
1632 * @vmas: array of pointers to vmas corresponding to each page.
1633 * Or NULL if the caller does not require them.
1634 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1636 * Returns number of pages pinned. This may be fewer than the number
1637 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1638 * were pinned, returns -errno. Each page returned must be released
1639 * with a put_page() call when it is finished with. vmas will only
1640 * remain valid while mmap_sem is held.
1642 * Must be called with mmap_sem held for read or write.
1644 * __get_user_pages walks a process's page tables and takes a reference to
1645 * each struct page that each user address corresponds to at a given
1646 * instant. That is, it takes the page that would be accessed if a user
1647 * thread accesses the given user virtual address at that instant.
1649 * This does not guarantee that the page exists in the user mappings when
1650 * __get_user_pages returns, and there may even be a completely different
1651 * page there in some cases (eg. if mmapped pagecache has been invalidated
1652 * and subsequently re faulted). However it does guarantee that the page
1653 * won't be freed completely. And mostly callers simply care that the page
1654 * contains data that was valid *at some point in time*. Typically, an IO
1655 * or similar operation cannot guarantee anything stronger anyway because
1656 * locks can't be held over the syscall boundary.
1658 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1659 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1660 * appropriate) must be called after the page is finished with, and
1661 * before put_page is called.
1663 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1664 * or mmap_sem contention, and if waiting is needed to pin all pages,
1665 * *@nonblocking will be set to 0.
1667 * In most cases, get_user_pages or get_user_pages_fast should be used
1668 * instead of __get_user_pages. __get_user_pages should be used only if
1669 * you need some special @gup_flags.
1671 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1672 unsigned long start, int nr_pages, unsigned int gup_flags,
1673 struct page **pages, struct vm_area_struct **vmas,
1677 unsigned long vm_flags;
1682 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1685 * Require read or write permissions.
1686 * If FOLL_FORCE is set, we only require the "MAY" flags.
1688 vm_flags = (gup_flags & FOLL_WRITE) ?
1689 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1690 vm_flags &= (gup_flags & FOLL_FORCE) ?
1691 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1695 struct vm_area_struct *vma;
1697 vma = find_extend_vma(mm, start);
1698 if (!vma && in_gate_area(mm, start)) {
1699 unsigned long pg = start & PAGE_MASK;
1705 /* user gate pages are read-only */
1706 if (gup_flags & FOLL_WRITE)
1707 return i ? : -EFAULT;
1709 pgd = pgd_offset_k(pg);
1711 pgd = pgd_offset_gate(mm, pg);
1712 BUG_ON(pgd_none(*pgd));
1713 pud = pud_offset(pgd, pg);
1714 BUG_ON(pud_none(*pud));
1715 pmd = pmd_offset(pud, pg);
1717 return i ? : -EFAULT;
1718 VM_BUG_ON(pmd_trans_huge(*pmd));
1719 pte = pte_offset_map(pmd, pg);
1720 if (pte_none(*pte)) {
1722 return i ? : -EFAULT;
1724 vma = get_gate_vma(mm);
1728 page = vm_normal_page(vma, start, *pte);
1730 if (!(gup_flags & FOLL_DUMP) &&
1731 is_zero_pfn(pte_pfn(*pte)))
1732 page = pte_page(*pte);
1735 return i ? : -EFAULT;
1746 if (vma && (vma->vm_flags & VM_FOREIGN)) {
1747 struct vm_foreign_map *foreign_map =
1748 vma->vm_private_data;
1749 struct page **map = foreign_map->map;
1750 int offset = (start - vma->vm_start) >> PAGE_SHIFT;
1751 if (map[offset] != NULL) {
1753 struct page *page = map[offset];
1768 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1769 !(vm_flags & vma->vm_flags))
1770 return i ? : -EFAULT;
1772 if (is_vm_hugetlb_page(vma)) {
1773 i = follow_hugetlb_page(mm, vma, pages, vmas,
1774 &start, &nr_pages, i, gup_flags);
1780 unsigned int foll_flags = gup_flags;
1783 * If we have a pending SIGKILL, don't keep faulting
1784 * pages and potentially allocating memory.
1786 if (unlikely(fatal_signal_pending(current)))
1787 return i ? i : -ERESTARTSYS;
1790 while (!(page = follow_page(vma, start, foll_flags))) {
1792 unsigned int fault_flags = 0;
1794 /* For mlock, just skip the stack guard page. */
1795 if (foll_flags & FOLL_MLOCK) {
1796 if (stack_guard_page(vma, start))
1799 if (foll_flags & FOLL_WRITE)
1800 fault_flags |= FAULT_FLAG_WRITE;
1802 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1803 if (foll_flags & FOLL_NOWAIT)
1804 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1806 ret = handle_mm_fault(mm, vma, start,
1809 if (ret & VM_FAULT_ERROR) {
1810 if (ret & VM_FAULT_OOM)
1811 return i ? i : -ENOMEM;
1812 if (ret & (VM_FAULT_HWPOISON |
1813 VM_FAULT_HWPOISON_LARGE)) {
1816 else if (gup_flags & FOLL_HWPOISON)
1821 if (ret & VM_FAULT_SIGBUS)
1822 return i ? i : -EFAULT;
1827 if (ret & VM_FAULT_MAJOR)
1833 if (ret & VM_FAULT_RETRY) {
1840 * The VM_FAULT_WRITE bit tells us that
1841 * do_wp_page has broken COW when necessary,
1842 * even if maybe_mkwrite decided not to set
1843 * pte_write. We can thus safely do subsequent
1844 * page lookups as if they were reads. But only
1845 * do so when looping for pte_write is futile:
1846 * in some cases userspace may also be wanting
1847 * to write to the gotten user page, which a
1848 * read fault here might prevent (a readonly
1849 * page might get reCOWed by userspace write).
1851 if ((ret & VM_FAULT_WRITE) &&
1852 !(vma->vm_flags & VM_WRITE))
1853 foll_flags &= ~FOLL_WRITE;
1858 return i ? i : PTR_ERR(page);
1862 flush_anon_page(vma, page, start);
1863 flush_dcache_page(page);
1871 } while (nr_pages && start < vma->vm_end);
1875 EXPORT_SYMBOL(__get_user_pages);
1878 * fixup_user_fault() - manually resolve a user page fault
1879 * @tsk: the task_struct to use for page fault accounting, or
1880 * NULL if faults are not to be recorded.
1881 * @mm: mm_struct of target mm
1882 * @address: user address
1883 * @fault_flags:flags to pass down to handle_mm_fault()
1885 * This is meant to be called in the specific scenario where for locking reasons
1886 * we try to access user memory in atomic context (within a pagefault_disable()
1887 * section), this returns -EFAULT, and we want to resolve the user fault before
1890 * Typically this is meant to be used by the futex code.
1892 * The main difference with get_user_pages() is that this function will
1893 * unconditionally call handle_mm_fault() which will in turn perform all the
1894 * necessary SW fixup of the dirty and young bits in the PTE, while
1895 * handle_mm_fault() only guarantees to update these in the struct page.
1897 * This is important for some architectures where those bits also gate the
1898 * access permission to the page because they are maintained in software. On
1899 * such architectures, gup() will not be enough to make a subsequent access
1902 * This should be called with the mm_sem held for read.
1904 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1905 unsigned long address, unsigned int fault_flags)
1907 struct vm_area_struct *vma;
1910 vma = find_extend_vma(mm, address);
1911 if (!vma || address < vma->vm_start)
1914 ret = handle_mm_fault(mm, vma, address, fault_flags);
1915 if (ret & VM_FAULT_ERROR) {
1916 if (ret & VM_FAULT_OOM)
1918 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1920 if (ret & VM_FAULT_SIGBUS)
1925 if (ret & VM_FAULT_MAJOR)
1934 * get_user_pages() - pin user pages in memory
1935 * @tsk: the task_struct to use for page fault accounting, or
1936 * NULL if faults are not to be recorded.
1937 * @mm: mm_struct of target mm
1938 * @start: starting user address
1939 * @nr_pages: number of pages from start to pin
1940 * @write: whether pages will be written to by the caller
1941 * @force: whether to force write access even if user mapping is
1942 * readonly. This will result in the page being COWed even
1943 * in MAP_SHARED mappings. You do not want this.
1944 * @pages: array that receives pointers to the pages pinned.
1945 * Should be at least nr_pages long. Or NULL, if caller
1946 * only intends to ensure the pages are faulted in.
1947 * @vmas: array of pointers to vmas corresponding to each page.
1948 * Or NULL if the caller does not require them.
1950 * Returns number of pages pinned. This may be fewer than the number
1951 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1952 * were pinned, returns -errno. Each page returned must be released
1953 * with a put_page() call when it is finished with. vmas will only
1954 * remain valid while mmap_sem is held.
1956 * Must be called with mmap_sem held for read or write.
1958 * get_user_pages walks a process's page tables and takes a reference to
1959 * each struct page that each user address corresponds to at a given
1960 * instant. That is, it takes the page that would be accessed if a user
1961 * thread accesses the given user virtual address at that instant.
1963 * This does not guarantee that the page exists in the user mappings when
1964 * get_user_pages returns, and there may even be a completely different
1965 * page there in some cases (eg. if mmapped pagecache has been invalidated
1966 * and subsequently re faulted). However it does guarantee that the page
1967 * won't be freed completely. And mostly callers simply care that the page
1968 * contains data that was valid *at some point in time*. Typically, an IO
1969 * or similar operation cannot guarantee anything stronger anyway because
1970 * locks can't be held over the syscall boundary.
1972 * If write=0, the page must not be written to. If the page is written to,
1973 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1974 * after the page is finished with, and before put_page is called.
1976 * get_user_pages is typically used for fewer-copy IO operations, to get a
1977 * handle on the memory by some means other than accesses via the user virtual
1978 * addresses. The pages may be submitted for DMA to devices or accessed via
1979 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1980 * use the correct cache flushing APIs.
1982 * See also get_user_pages_fast, for performance critical applications.
1984 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1985 unsigned long start, int nr_pages, int write, int force,
1986 struct page **pages, struct vm_area_struct **vmas)
1988 int flags = FOLL_TOUCH;
1993 flags |= FOLL_WRITE;
1995 flags |= FOLL_FORCE;
1997 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2000 EXPORT_SYMBOL(get_user_pages);
2003 * get_dump_page() - pin user page in memory while writing it to core dump
2004 * @addr: user address
2006 * Returns struct page pointer of user page pinned for dump,
2007 * to be freed afterwards by page_cache_release() or put_page().
2009 * Returns NULL on any kind of failure - a hole must then be inserted into
2010 * the corefile, to preserve alignment with its headers; and also returns
2011 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2012 * allowing a hole to be left in the corefile to save diskspace.
2014 * Called without mmap_sem, but after all other threads have been killed.
2016 #ifdef CONFIG_ELF_CORE
2017 struct page *get_dump_page(unsigned long addr)
2019 struct vm_area_struct *vma;
2022 if (__get_user_pages(current, current->mm, addr, 1,
2023 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2026 flush_cache_page(vma, addr, page_to_pfn(page));
2029 #endif /* CONFIG_ELF_CORE */
2031 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2034 pgd_t * pgd = pgd_offset(mm, addr);
2035 pud_t * pud = pud_alloc(mm, pgd, addr);
2037 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2039 VM_BUG_ON(pmd_trans_huge(*pmd));
2040 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2047 * This is the old fallback for page remapping.
2049 * For historical reasons, it only allows reserved pages. Only
2050 * old drivers should use this, and they needed to mark their
2051 * pages reserved for the old functions anyway.
2053 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2054 struct page *page, pgprot_t prot)
2056 struct mm_struct *mm = vma->vm_mm;
2065 flush_dcache_page(page);
2066 pte = get_locked_pte(mm, addr, &ptl);
2070 if (!pte_none(*pte))
2073 /* Ok, finally just insert the thing.. */
2075 inc_mm_counter_fast(mm, MM_FILEPAGES);
2076 page_add_file_rmap(page);
2077 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2080 pte_unmap_unlock(pte, ptl);
2083 pte_unmap_unlock(pte, ptl);
2089 * vm_insert_page - insert single page into user vma
2090 * @vma: user vma to map to
2091 * @addr: target user address of this page
2092 * @page: source kernel page
2094 * This allows drivers to insert individual pages they've allocated
2097 * The page has to be a nice clean _individual_ kernel allocation.
2098 * If you allocate a compound page, you need to have marked it as
2099 * such (__GFP_COMP), or manually just split the page up yourself
2100 * (see split_page()).
2102 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2103 * took an arbitrary page protection parameter. This doesn't allow
2104 * that. Your vma protection will have to be set up correctly, which
2105 * means that if you want a shared writable mapping, you'd better
2106 * ask for a shared writable mapping!
2108 * The page does not need to be reserved.
2110 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2113 if (addr < vma->vm_start || addr >= vma->vm_end)
2115 if (!page_count(page))
2117 vma->vm_flags |= VM_INSERTPAGE;
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_inert_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_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2189 ret = insert_pfn(vma, addr, pfn, pgprot);
2192 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2196 EXPORT_SYMBOL(vm_insert_pfn);
2198 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2201 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2203 if (addr < vma->vm_start || addr >= vma->vm_end)
2207 * If we don't have pte special, then we have to use the pfn_valid()
2208 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2209 * refcount the page if pfn_valid is true (hence insert_page rather
2210 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2211 * without pte special, it would there be refcounted as a normal page.
2213 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2216 page = pfn_to_page(pfn);
2217 return insert_page(vma, addr, page, vma->vm_page_prot);
2219 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2221 EXPORT_SYMBOL(vm_insert_mixed);
2224 * maps a range of physical memory into the requested pages. the old
2225 * mappings are removed. any references to nonexistent pages results
2226 * in null mappings (currently treated as "copy-on-access")
2228 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2229 unsigned long addr, unsigned long end,
2230 unsigned long pfn, pgprot_t prot)
2235 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2238 arch_enter_lazy_mmu_mode();
2240 BUG_ON(!pte_none(*pte));
2241 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2243 } while (pte++, addr += PAGE_SIZE, addr != end);
2244 arch_leave_lazy_mmu_mode();
2245 pte_unmap_unlock(pte - 1, ptl);
2249 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2250 unsigned long addr, unsigned long end,
2251 unsigned long pfn, pgprot_t prot)
2256 pfn -= addr >> PAGE_SHIFT;
2257 pmd = pmd_alloc(mm, pud, addr);
2260 VM_BUG_ON(pmd_trans_huge(*pmd));
2262 next = pmd_addr_end(addr, end);
2263 if (remap_pte_range(mm, pmd, addr, next,
2264 pfn + (addr >> PAGE_SHIFT), prot))
2266 } while (pmd++, addr = next, addr != end);
2270 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2271 unsigned long addr, unsigned long end,
2272 unsigned long pfn, pgprot_t prot)
2277 pfn -= addr >> PAGE_SHIFT;
2278 pud = pud_alloc(mm, pgd, addr);
2282 next = pud_addr_end(addr, end);
2283 if (remap_pmd_range(mm, pud, addr, next,
2284 pfn + (addr >> PAGE_SHIFT), prot))
2286 } while (pud++, addr = next, addr != end);
2291 * remap_pfn_range - remap kernel memory to userspace
2292 * @vma: user vma to map to
2293 * @addr: target user address to start at
2294 * @pfn: physical address of kernel memory
2295 * @size: size of map area
2296 * @prot: page protection flags for this mapping
2298 * Note: this is only safe if the mm semaphore is held when called.
2300 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2301 unsigned long pfn, unsigned long size, pgprot_t prot)
2305 unsigned long end = addr + PAGE_ALIGN(size);
2306 struct mm_struct *mm = vma->vm_mm;
2310 * Physically remapped pages are special. Tell the
2311 * rest of the world about it:
2312 * VM_IO tells people not to look at these pages
2313 * (accesses can have side effects).
2314 * VM_RESERVED is specified all over the place, because
2315 * in 2.4 it kept swapout's vma scan off this vma; but
2316 * in 2.6 the LRU scan won't even find its pages, so this
2317 * flag means no more than count its pages in reserved_vm,
2318 * and omit it from core dump, even when VM_IO turned off.
2319 * VM_PFNMAP tells the core MM that the base pages are just
2320 * raw PFN mappings, and do not have a "struct page" associated
2323 * There's a horrible special case to handle copy-on-write
2324 * behaviour that some programs depend on. We mark the "original"
2325 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2327 if (addr == vma->vm_start && end == vma->vm_end) {
2328 vma->vm_pgoff = pfn;
2329 vma->vm_flags |= VM_PFN_AT_MMAP;
2330 } else if (is_cow_mapping(vma->vm_flags))
2333 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2335 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2338 * To indicate that track_pfn related cleanup is not
2339 * needed from higher level routine calling unmap_vmas
2341 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2342 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2346 BUG_ON(addr >= end);
2347 pfn -= addr >> PAGE_SHIFT;
2348 pgd = pgd_offset(mm, addr);
2349 flush_cache_range(vma, addr, end);
2351 next = pgd_addr_end(addr, end);
2352 err = remap_pud_range(mm, pgd, addr, next,
2353 pfn + (addr >> PAGE_SHIFT), prot);
2356 } while (pgd++, addr = next, addr != end);
2359 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2363 EXPORT_SYMBOL(remap_pfn_range);
2365 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2366 unsigned long addr, unsigned long end,
2367 pte_fn_t fn, void *data)
2372 spinlock_t *uninitialized_var(ptl);
2374 pte = (mm == &init_mm) ?
2375 pte_alloc_kernel(pmd, addr) :
2376 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2380 BUG_ON(pmd_huge(*pmd));
2382 arch_enter_lazy_mmu_mode();
2384 token = pmd_pgtable(*pmd);
2387 err = fn(pte++, token, addr, data);
2390 } while (addr += PAGE_SIZE, addr != end);
2392 arch_leave_lazy_mmu_mode();
2395 pte_unmap_unlock(pte-1, ptl);
2399 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2400 unsigned long addr, unsigned long end,
2401 pte_fn_t fn, void *data)
2407 BUG_ON(pud_huge(*pud));
2409 pmd = pmd_alloc(mm, pud, addr);
2413 next = pmd_addr_end(addr, end);
2414 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2417 } while (pmd++, addr = next, addr != end);
2421 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2422 unsigned long addr, unsigned long end,
2423 pte_fn_t fn, void *data)
2429 pud = pud_alloc(mm, pgd, addr);
2433 next = pud_addr_end(addr, end);
2434 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2437 } while (pud++, addr = next, addr != end);
2442 * Scan a region of virtual memory, filling in page tables as necessary
2443 * and calling a provided function on each leaf page table.
2445 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2446 unsigned long size, pte_fn_t fn, void *data)
2450 unsigned long end = addr + size;
2457 BUG_ON(addr >= end);
2458 pgd = pgd_offset(mm, addr);
2460 next = pgd_addr_end(addr, end);
2461 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2464 } while (pgd++, addr = next, addr != end);
2468 EXPORT_SYMBOL_GPL(apply_to_page_range);
2471 * handle_pte_fault chooses page fault handler according to an entry
2472 * which was read non-atomically. Before making any commitment, on
2473 * those architectures or configurations (e.g. i386 with PAE) which
2474 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2475 * must check under lock before unmapping the pte and proceeding
2476 * (but do_wp_page is only called after already making such a check;
2477 * and do_anonymous_page can safely check later on).
2479 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2480 pte_t *page_table, pte_t orig_pte)
2483 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2484 if (sizeof(pte_t) > sizeof(unsigned long)) {
2485 spinlock_t *ptl = pte_lockptr(mm, pmd);
2487 same = pte_same(*page_table, orig_pte);
2491 pte_unmap(page_table);
2495 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2498 * If the source page was a PFN mapping, we don't have
2499 * a "struct page" for it. We do a best-effort copy by
2500 * just copying from the original user address. If that
2501 * fails, we just zero-fill it. Live with it.
2503 if (unlikely(!src)) {
2504 void *kaddr = kmap_atomic(dst);
2505 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2508 * This really shouldn't fail, because the page is there
2509 * in the page tables. But it might just be unreadable,
2510 * in which case we just give up and fill the result with
2513 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2515 kunmap_atomic(kaddr);
2516 flush_dcache_page(dst);
2518 copy_user_highpage(dst, src, va, vma);
2522 * This routine handles present pages, when users try to write
2523 * to a shared page. It is done by copying the page to a new address
2524 * and decrementing the shared-page counter for the old page.
2526 * Note that this routine assumes that the protection checks have been
2527 * done by the caller (the low-level page fault routine in most cases).
2528 * Thus we can safely just mark it writable once we've done any necessary
2531 * We also mark the page dirty at this point even though the page will
2532 * change only once the write actually happens. This avoids a few races,
2533 * and potentially makes it more efficient.
2535 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2536 * but allow concurrent faults), with pte both mapped and locked.
2537 * We return with mmap_sem still held, but pte unmapped and unlocked.
2539 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2540 unsigned long address, pte_t *page_table, pmd_t *pmd,
2541 spinlock_t *ptl, pte_t orig_pte)
2544 struct page *old_page, *new_page;
2547 int page_mkwrite = 0;
2548 struct page *dirty_page = NULL;
2550 old_page = vm_normal_page(vma, address, orig_pte);
2553 * VM_MIXEDMAP !pfn_valid() case
2555 * We should not cow pages in a shared writeable mapping.
2556 * Just mark the pages writable as we can't do any dirty
2557 * accounting on raw pfn maps.
2559 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2560 (VM_WRITE|VM_SHARED))
2566 * Take out anonymous pages first, anonymous shared vmas are
2567 * not dirty accountable.
2569 if (PageAnon(old_page) && !PageKsm(old_page)) {
2570 if (!trylock_page(old_page)) {
2571 page_cache_get(old_page);
2572 pte_unmap_unlock(page_table, ptl);
2573 lock_page(old_page);
2574 page_table = pte_offset_map_lock(mm, pmd, address,
2576 if (!pte_same(*page_table, orig_pte)) {
2577 unlock_page(old_page);
2580 page_cache_release(old_page);
2582 if (reuse_swap_page(old_page)) {
2584 * The page is all ours. Move it to our anon_vma so
2585 * the rmap code will not search our parent or siblings.
2586 * Protected against the rmap code by the page lock.
2588 page_move_anon_rmap(old_page, vma, address);
2589 unlock_page(old_page);
2592 unlock_page(old_page);
2593 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2594 (VM_WRITE|VM_SHARED))) {
2596 * Only catch write-faults on shared writable pages,
2597 * read-only shared pages can get COWed by
2598 * get_user_pages(.write=1, .force=1).
2600 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2601 struct vm_fault vmf;
2604 vmf.virtual_address = (void __user *)(address &
2606 vmf.pgoff = old_page->index;
2607 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2608 vmf.page = old_page;
2611 * Notify the address space that the page is about to
2612 * become writable so that it can prohibit this or wait
2613 * for the page to get into an appropriate state.
2615 * We do this without the lock held, so that it can
2616 * sleep if it needs to.
2618 page_cache_get(old_page);
2619 pte_unmap_unlock(page_table, ptl);
2621 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2623 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2625 goto unwritable_page;
2627 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2628 lock_page(old_page);
2629 if (!old_page->mapping) {
2630 ret = 0; /* retry the fault */
2631 unlock_page(old_page);
2632 goto unwritable_page;
2635 VM_BUG_ON(!PageLocked(old_page));
2638 * Since we dropped the lock we need to revalidate
2639 * the PTE as someone else may have changed it. If
2640 * they did, we just return, as we can count on the
2641 * MMU to tell us if they didn't also make it writable.
2643 page_table = pte_offset_map_lock(mm, pmd, address,
2645 if (!pte_same(*page_table, orig_pte)) {
2646 unlock_page(old_page);
2652 dirty_page = old_page;
2653 get_page(dirty_page);
2656 flush_cache_page(vma, address, pte_pfn(orig_pte));
2657 entry = pte_mkyoung(orig_pte);
2658 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2659 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2660 update_mmu_cache(vma, address, page_table);
2661 pte_unmap_unlock(page_table, ptl);
2662 ret |= VM_FAULT_WRITE;
2668 * Yes, Virginia, this is actually required to prevent a race
2669 * with clear_page_dirty_for_io() from clearing the page dirty
2670 * bit after it clear all dirty ptes, but before a racing
2671 * do_wp_page installs a dirty pte.
2673 * __do_fault is protected similarly.
2675 if (!page_mkwrite) {
2676 wait_on_page_locked(dirty_page);
2677 set_page_dirty_balance(dirty_page, page_mkwrite);
2679 put_page(dirty_page);
2681 struct address_space *mapping = dirty_page->mapping;
2683 set_page_dirty(dirty_page);
2684 unlock_page(dirty_page);
2685 page_cache_release(dirty_page);
2688 * Some device drivers do not set page.mapping
2689 * but still dirty their pages
2691 balance_dirty_pages_ratelimited(mapping);
2695 /* file_update_time outside page_lock */
2697 file_update_time(vma->vm_file);
2703 * Ok, we need to copy. Oh, well..
2705 page_cache_get(old_page);
2707 pte_unmap_unlock(page_table, ptl);
2709 if (unlikely(anon_vma_prepare(vma)))
2712 if (is_zero_pfn(pte_pfn(orig_pte))) {
2713 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2717 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2720 cow_user_page(new_page, old_page, address, vma);
2722 __SetPageUptodate(new_page);
2724 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2728 * Re-check the pte - we dropped the lock
2730 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2731 if (likely(pte_same(*page_table, orig_pte))) {
2733 if (!PageAnon(old_page)) {
2734 dec_mm_counter_fast(mm, MM_FILEPAGES);
2735 inc_mm_counter_fast(mm, MM_ANONPAGES);
2738 inc_mm_counter_fast(mm, MM_ANONPAGES);
2739 flush_cache_page(vma, address, pte_pfn(orig_pte));
2740 entry = mk_pte(new_page, vma->vm_page_prot);
2741 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2743 * Clear the pte entry and flush it first, before updating the
2744 * pte with the new entry. This will avoid a race condition
2745 * seen in the presence of one thread doing SMC and another
2748 ptep_clear_flush(vma, address, page_table);
2749 page_add_new_anon_rmap(new_page, vma, address);
2751 * We call the notify macro here because, when using secondary
2752 * mmu page tables (such as kvm shadow page tables), we want the
2753 * new page to be mapped directly into the secondary page table.
2755 set_pte_at_notify(mm, address, page_table, entry);
2756 update_mmu_cache(vma, address, page_table);
2759 * Only after switching the pte to the new page may
2760 * we remove the mapcount here. Otherwise another
2761 * process may come and find the rmap count decremented
2762 * before the pte is switched to the new page, and
2763 * "reuse" the old page writing into it while our pte
2764 * here still points into it and can be read by other
2767 * The critical issue is to order this
2768 * page_remove_rmap with the ptp_clear_flush above.
2769 * Those stores are ordered by (if nothing else,)
2770 * the barrier present in the atomic_add_negative
2771 * in page_remove_rmap.
2773 * Then the TLB flush in ptep_clear_flush ensures that
2774 * no process can access the old page before the
2775 * decremented mapcount is visible. And the old page
2776 * cannot be reused until after the decremented
2777 * mapcount is visible. So transitively, TLBs to
2778 * old page will be flushed before it can be reused.
2780 page_remove_rmap(old_page);
2783 /* Free the old page.. */
2784 new_page = old_page;
2785 ret |= VM_FAULT_WRITE;
2787 mem_cgroup_uncharge_page(new_page);
2790 page_cache_release(new_page);
2792 pte_unmap_unlock(page_table, ptl);
2795 * Don't let another task, with possibly unlocked vma,
2796 * keep the mlocked page.
2798 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2799 lock_page(old_page); /* LRU manipulation */
2800 munlock_vma_page(old_page);
2801 unlock_page(old_page);
2803 page_cache_release(old_page);
2807 page_cache_release(new_page);
2811 unlock_page(old_page);
2812 page_cache_release(old_page);
2814 page_cache_release(old_page);
2816 return VM_FAULT_OOM;
2819 page_cache_release(old_page);
2823 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2824 unsigned long start_addr, unsigned long end_addr,
2825 struct zap_details *details)
2827 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2830 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2831 struct zap_details *details)
2833 struct vm_area_struct *vma;
2834 struct prio_tree_iter iter;
2835 pgoff_t vba, vea, zba, zea;
2837 vma_prio_tree_foreach(vma, &iter, root,
2838 details->first_index, details->last_index) {
2840 vba = vma->vm_pgoff;
2841 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2842 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2843 zba = details->first_index;
2846 zea = details->last_index;
2850 unmap_mapping_range_vma(vma,
2851 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2852 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2857 static inline void unmap_mapping_range_list(struct list_head *head,
2858 struct zap_details *details)
2860 struct vm_area_struct *vma;
2863 * In nonlinear VMAs there is no correspondence between virtual address
2864 * offset and file offset. So we must perform an exhaustive search
2865 * across *all* the pages in each nonlinear VMA, not just the pages
2866 * whose virtual address lies outside the file truncation point.
2868 list_for_each_entry(vma, head, shared.vm_set.list) {
2869 details->nonlinear_vma = vma;
2870 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2875 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2876 * @mapping: the address space containing mmaps to be unmapped.
2877 * @holebegin: byte in first page to unmap, relative to the start of
2878 * the underlying file. This will be rounded down to a PAGE_SIZE
2879 * boundary. Note that this is different from truncate_pagecache(), which
2880 * must keep the partial page. In contrast, we must get rid of
2882 * @holelen: size of prospective hole in bytes. This will be rounded
2883 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2885 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2886 * but 0 when invalidating pagecache, don't throw away private data.
2888 void unmap_mapping_range(struct address_space *mapping,
2889 loff_t const holebegin, loff_t const holelen, int even_cows)
2891 struct zap_details details;
2892 pgoff_t hba = holebegin >> PAGE_SHIFT;
2893 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2895 /* Check for overflow. */
2896 if (sizeof(holelen) > sizeof(hlen)) {
2898 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2899 if (holeend & ~(long long)ULONG_MAX)
2900 hlen = ULONG_MAX - hba + 1;
2903 details.check_mapping = even_cows? NULL: mapping;
2904 details.nonlinear_vma = NULL;
2905 details.first_index = hba;
2906 details.last_index = hba + hlen - 1;
2907 if (details.last_index < details.first_index)
2908 details.last_index = ULONG_MAX;
2911 mutex_lock(&mapping->i_mmap_mutex);
2912 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2913 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2914 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2915 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2916 mutex_unlock(&mapping->i_mmap_mutex);
2918 EXPORT_SYMBOL(unmap_mapping_range);
2921 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2922 * but allow concurrent faults), and pte mapped but not yet locked.
2923 * We return with mmap_sem still held, but pte unmapped and unlocked.
2925 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2926 unsigned long address, pte_t *page_table, pmd_t *pmd,
2927 unsigned int flags, pte_t orig_pte)
2930 struct page *page, *swapcache = NULL;
2934 struct mem_cgroup *ptr;
2938 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2941 entry = pte_to_swp_entry(orig_pte);
2942 if (unlikely(non_swap_entry(entry))) {
2943 if (is_migration_entry(entry)) {
2944 migration_entry_wait(mm, pmd, address);
2945 } else if (is_hwpoison_entry(entry)) {
2946 ret = VM_FAULT_HWPOISON;
2948 print_bad_pte(vma, address, orig_pte, NULL);
2949 ret = VM_FAULT_SIGBUS;
2953 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2954 page = lookup_swap_cache(entry);
2956 grab_swap_token(mm); /* Contend for token _before_ read-in */
2957 page = swapin_readahead(entry,
2958 GFP_HIGHUSER_MOVABLE, vma, address);
2961 * Back out if somebody else faulted in this pte
2962 * while we released the pte lock.
2964 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2965 if (likely(pte_same(*page_table, orig_pte)))
2967 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2971 /* Had to read the page from swap area: Major fault */
2972 ret = VM_FAULT_MAJOR;
2973 count_vm_event(PGMAJFAULT);
2974 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2975 } else if (PageHWPoison(page)) {
2977 * hwpoisoned dirty swapcache pages are kept for killing
2978 * owner processes (which may be unknown at hwpoison time)
2980 ret = VM_FAULT_HWPOISON;
2981 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2985 locked = lock_page_or_retry(page, mm, flags);
2986 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2988 ret |= VM_FAULT_RETRY;
2993 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2994 * release the swapcache from under us. The page pin, and pte_same
2995 * test below, are not enough to exclude that. Even if it is still
2996 * swapcache, we need to check that the page's swap has not changed.
2998 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3001 if (ksm_might_need_to_copy(page, vma, address)) {
3003 page = ksm_does_need_to_copy(page, vma, address);
3005 if (unlikely(!page)) {
3013 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3019 * Back out if somebody else already faulted in this pte.
3021 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3022 if (unlikely(!pte_same(*page_table, orig_pte)))
3025 if (unlikely(!PageUptodate(page))) {
3026 ret = VM_FAULT_SIGBUS;
3031 * The page isn't present yet, go ahead with the fault.
3033 * Be careful about the sequence of operations here.
3034 * To get its accounting right, reuse_swap_page() must be called
3035 * while the page is counted on swap but not yet in mapcount i.e.
3036 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3037 * must be called after the swap_free(), or it will never succeed.
3038 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3039 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3040 * in page->private. In this case, a record in swap_cgroup is silently
3041 * discarded at swap_free().
3044 inc_mm_counter_fast(mm, MM_ANONPAGES);
3045 dec_mm_counter_fast(mm, MM_SWAPENTS);
3046 pte = mk_pte(page, vma->vm_page_prot);
3047 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3048 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3049 flags &= ~FAULT_FLAG_WRITE;
3050 ret |= VM_FAULT_WRITE;
3053 flush_icache_page(vma, page);
3054 set_pte_at(mm, address, page_table, pte);
3055 do_page_add_anon_rmap(page, vma, address, exclusive);
3056 /* It's better to call commit-charge after rmap is established */
3057 mem_cgroup_commit_charge_swapin(page, ptr);
3060 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3061 try_to_free_swap(page);
3065 * Hold the lock to avoid the swap entry to be reused
3066 * until we take the PT lock for the pte_same() check
3067 * (to avoid false positives from pte_same). For
3068 * further safety release the lock after the swap_free
3069 * so that the swap count won't change under a
3070 * parallel locked swapcache.
3072 unlock_page(swapcache);
3073 page_cache_release(swapcache);
3076 if (flags & FAULT_FLAG_WRITE) {
3077 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3078 if (ret & VM_FAULT_ERROR)
3079 ret &= VM_FAULT_ERROR;
3083 /* No need to invalidate - it was non-present before */
3084 update_mmu_cache(vma, address, page_table);
3086 pte_unmap_unlock(page_table, ptl);
3090 mem_cgroup_cancel_charge_swapin(ptr);
3091 pte_unmap_unlock(page_table, ptl);
3095 page_cache_release(page);
3097 unlock_page(swapcache);
3098 page_cache_release(swapcache);
3104 * This is like a special single-page "expand_{down|up}wards()",
3105 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3106 * doesn't hit another vma.
3108 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3110 address &= PAGE_MASK;
3111 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3112 struct vm_area_struct *prev = vma->vm_prev;
3115 * Is there a mapping abutting this one below?
3117 * That's only ok if it's the same stack mapping
3118 * that has gotten split..
3120 if (prev && prev->vm_end == address)
3121 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3123 expand_downwards(vma, address - PAGE_SIZE);
3125 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3126 struct vm_area_struct *next = vma->vm_next;
3128 /* As VM_GROWSDOWN but s/below/above/ */
3129 if (next && next->vm_start == address + PAGE_SIZE)
3130 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3132 expand_upwards(vma, address + PAGE_SIZE);
3138 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3139 * but allow concurrent faults), and pte mapped but not yet locked.
3140 * We return with mmap_sem still held, but pte unmapped and unlocked.
3142 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3143 unsigned long address, pte_t *page_table, pmd_t *pmd,
3150 pte_unmap(page_table);
3152 /* Check if we need to add a guard page to the stack */
3153 if (check_stack_guard_page(vma, address) < 0)
3154 return VM_FAULT_SIGBUS;
3156 /* Use the zero-page for reads */
3157 if (!(flags & FAULT_FLAG_WRITE)) {
3158 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3159 vma->vm_page_prot));
3160 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3161 if (!pte_none(*page_table))
3166 /* Allocate our own private page. */
3167 if (unlikely(anon_vma_prepare(vma)))
3169 page = alloc_zeroed_user_highpage_movable(vma, address);
3172 __SetPageUptodate(page);
3174 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3177 entry = mk_pte(page, vma->vm_page_prot);
3178 if (vma->vm_flags & VM_WRITE)
3179 entry = pte_mkwrite(pte_mkdirty(entry));
3181 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3182 if (!pte_none(*page_table))
3185 inc_mm_counter_fast(mm, MM_ANONPAGES);
3186 page_add_new_anon_rmap(page, vma, address);
3188 set_pte_at(mm, address, page_table, entry);
3190 /* No need to invalidate - it was non-present before */
3191 update_mmu_cache(vma, address, page_table);
3193 pte_unmap_unlock(page_table, ptl);
3196 mem_cgroup_uncharge_page(page);
3197 page_cache_release(page);
3200 page_cache_release(page);
3202 return VM_FAULT_OOM;
3206 * __do_fault() tries to create a new page mapping. It aggressively
3207 * tries to share with existing pages, but makes a separate copy if
3208 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3209 * the next page fault.
3211 * As this is called only for pages that do not currently exist, we
3212 * do not need to flush old virtual caches or the TLB.
3214 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3215 * but allow concurrent faults), and pte neither mapped nor locked.
3216 * We return with mmap_sem still held, but pte unmapped and unlocked.
3218 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3219 unsigned long address, pmd_t *pmd,
3220 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3225 struct page *cow_page;
3228 struct page *dirty_page = NULL;
3229 struct vm_fault vmf;
3231 int page_mkwrite = 0;
3234 * If we do COW later, allocate page befor taking lock_page()
3235 * on the file cache page. This will reduce lock holding time.
3237 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3239 if (unlikely(anon_vma_prepare(vma)))
3240 return VM_FAULT_OOM;
3242 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3244 return VM_FAULT_OOM;
3246 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3247 page_cache_release(cow_page);
3248 return VM_FAULT_OOM;
3253 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3258 ret = vma->vm_ops->fault(vma, &vmf);
3259 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3263 if (unlikely(PageHWPoison(vmf.page))) {
3264 if (ret & VM_FAULT_LOCKED)
3265 unlock_page(vmf.page);
3266 ret = VM_FAULT_HWPOISON;
3271 * For consistency in subsequent calls, make the faulted page always
3274 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3275 lock_page(vmf.page);
3277 VM_BUG_ON(!PageLocked(vmf.page));
3280 * Should we do an early C-O-W break?
3283 if (flags & FAULT_FLAG_WRITE) {
3284 if (!(vma->vm_flags & VM_SHARED)) {
3287 copy_user_highpage(page, vmf.page, address, vma);
3288 __SetPageUptodate(page);
3291 * If the page will be shareable, see if the backing
3292 * address space wants to know that the page is about
3293 * to become writable
3295 if (vma->vm_ops->page_mkwrite) {
3299 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3300 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3302 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3304 goto unwritable_page;
3306 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3308 if (!page->mapping) {
3309 ret = 0; /* retry the fault */
3311 goto unwritable_page;
3314 VM_BUG_ON(!PageLocked(page));
3321 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3324 * This silly early PAGE_DIRTY setting removes a race
3325 * due to the bad i386 page protection. But it's valid
3326 * for other architectures too.
3328 * Note that if FAULT_FLAG_WRITE is set, we either now have
3329 * an exclusive copy of the page, or this is a shared mapping,
3330 * so we can make it writable and dirty to avoid having to
3331 * handle that later.
3333 /* Only go through if we didn't race with anybody else... */
3334 if (likely(pte_same(*page_table, orig_pte))) {
3335 flush_icache_page(vma, page);
3336 entry = mk_pte(page, vma->vm_page_prot);
3337 if (flags & FAULT_FLAG_WRITE)
3338 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3340 inc_mm_counter_fast(mm, MM_ANONPAGES);
3341 page_add_new_anon_rmap(page, vma, address);
3343 inc_mm_counter_fast(mm, MM_FILEPAGES);
3344 page_add_file_rmap(page);
3345 if (flags & FAULT_FLAG_WRITE) {
3347 get_page(dirty_page);
3350 set_pte_at(mm, address, page_table, entry);
3352 /* no need to invalidate: a not-present page won't be cached */
3353 update_mmu_cache(vma, address, page_table);
3356 mem_cgroup_uncharge_page(cow_page);
3358 page_cache_release(page);
3360 anon = 1; /* no anon but release faulted_page */
3363 pte_unmap_unlock(page_table, ptl);
3366 struct address_space *mapping = page->mapping;
3368 if (set_page_dirty(dirty_page))
3370 unlock_page(dirty_page);
3371 put_page(dirty_page);
3372 if (page_mkwrite && mapping) {
3374 * Some device drivers do not set page.mapping but still
3377 balance_dirty_pages_ratelimited(mapping);
3380 /* file_update_time outside page_lock */
3382 file_update_time(vma->vm_file);
3384 unlock_page(vmf.page);
3386 page_cache_release(vmf.page);
3392 page_cache_release(page);
3395 /* fs's fault handler get error */
3397 mem_cgroup_uncharge_page(cow_page);
3398 page_cache_release(cow_page);
3403 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3404 unsigned long address, pte_t *page_table, pmd_t *pmd,
3405 unsigned int flags, pte_t orig_pte)
3407 pgoff_t pgoff = (((address & PAGE_MASK)
3408 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3410 pte_unmap(page_table);
3411 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3415 * Fault of a previously existing named mapping. Repopulate the pte
3416 * from the encoded file_pte if possible. This enables swappable
3419 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3420 * but allow concurrent faults), and pte mapped but not yet locked.
3421 * We return with mmap_sem still held, but pte unmapped and unlocked.
3423 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3424 unsigned long address, pte_t *page_table, pmd_t *pmd,
3425 unsigned int flags, pte_t orig_pte)
3429 flags |= FAULT_FLAG_NONLINEAR;
3431 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3434 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3436 * Page table corrupted: show pte and kill process.
3438 print_bad_pte(vma, address, orig_pte, NULL);
3439 return VM_FAULT_SIGBUS;
3442 pgoff = pte_to_pgoff(orig_pte);
3443 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3447 * These routines also need to handle stuff like marking pages dirty
3448 * and/or accessed for architectures that don't do it in hardware (most
3449 * RISC architectures). The early dirtying is also good on the i386.
3451 * There is also a hook called "update_mmu_cache()" that architectures
3452 * with external mmu caches can use to update those (ie the Sparc or
3453 * PowerPC hashed page tables that act as extended TLBs).
3455 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3456 * but allow concurrent faults), and pte mapped but not yet locked.
3457 * We return with mmap_sem still held, but pte unmapped and unlocked.
3459 int handle_pte_fault(struct mm_struct *mm,
3460 struct vm_area_struct *vma, unsigned long address,
3461 pte_t *pte, pmd_t *pmd, unsigned int flags)
3467 if (!pte_present(entry)) {
3468 if (pte_none(entry)) {
3470 if (likely(vma->vm_ops->fault))
3471 return do_linear_fault(mm, vma, address,
3472 pte, pmd, flags, entry);
3474 return do_anonymous_page(mm, vma, address,
3477 if (pte_file(entry))
3478 return do_nonlinear_fault(mm, vma, address,
3479 pte, pmd, flags, entry);
3480 return do_swap_page(mm, vma, address,
3481 pte, pmd, flags, entry);
3484 ptl = pte_lockptr(mm, pmd);
3486 if (unlikely(!pte_same(*pte, entry)))
3488 if (flags & FAULT_FLAG_WRITE) {
3489 if (!pte_write(entry))
3490 return do_wp_page(mm, vma, address,
3491 pte, pmd, ptl, entry);
3492 entry = pte_mkdirty(entry);
3494 entry = pte_mkyoung(entry);
3495 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3496 update_mmu_cache(vma, address, pte);
3499 * This is needed only for protection faults but the arch code
3500 * is not yet telling us if this is a protection fault or not.
3501 * This still avoids useless tlb flushes for .text page faults
3504 if (flags & FAULT_FLAG_WRITE)
3505 flush_tlb_fix_spurious_fault(vma, address);
3508 pte_unmap_unlock(pte, ptl);
3513 * By the time we get here, we already hold the mm semaphore
3515 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3516 unsigned long address, unsigned int flags)
3523 __set_current_state(TASK_RUNNING);
3525 count_vm_event(PGFAULT);
3526 mem_cgroup_count_vm_event(mm, PGFAULT);
3528 /* do counter updates before entering really critical section. */
3529 check_sync_rss_stat(current);
3531 if (unlikely(is_vm_hugetlb_page(vma)))
3532 return hugetlb_fault(mm, vma, address, flags);
3534 pgd = pgd_offset(mm, address);
3535 pud = pud_alloc(mm, pgd, address);
3537 return VM_FAULT_OOM;
3538 pmd = pmd_alloc(mm, pud, address);
3540 return VM_FAULT_OOM;
3541 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3543 return do_huge_pmd_anonymous_page(mm, vma, address,
3546 pmd_t orig_pmd = *pmd;
3548 if (pmd_trans_huge(orig_pmd)) {
3549 if (flags & FAULT_FLAG_WRITE &&
3550 !pmd_write(orig_pmd) &&
3551 !pmd_trans_splitting(orig_pmd))
3552 return do_huge_pmd_wp_page(mm, vma, address,
3559 * Use __pte_alloc instead of pte_alloc_map, because we can't
3560 * run pte_offset_map on the pmd, if an huge pmd could
3561 * materialize from under us from a different thread.
3563 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3564 return VM_FAULT_OOM;
3565 /* if an huge pmd materialized from under us just retry later */
3566 if (unlikely(pmd_trans_huge(*pmd)))
3569 * A regular pmd is established and it can't morph into a huge pmd
3570 * from under us anymore at this point because we hold the mmap_sem
3571 * read mode and khugepaged takes it in write mode. So now it's
3572 * safe to run pte_offset_map().
3574 pte = pte_offset_map(pmd, address);
3576 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3579 #ifndef __PAGETABLE_PUD_FOLDED
3581 * Allocate page upper directory.
3582 * We've already handled the fast-path in-line.
3584 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3586 pud_t *new = pud_alloc_one(mm, address);
3590 smp_wmb(); /* See comment in __pte_alloc */
3592 spin_lock(&mm->page_table_lock);
3593 if (pgd_present(*pgd)) /* Another has populated it */
3596 pgd_populate(mm, pgd, new);
3597 spin_unlock(&mm->page_table_lock);
3600 #endif /* __PAGETABLE_PUD_FOLDED */
3602 #ifndef __PAGETABLE_PMD_FOLDED
3604 * Allocate page middle directory.
3605 * We've already handled the fast-path in-line.
3607 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3609 pmd_t *new = pmd_alloc_one(mm, address);
3613 smp_wmb(); /* See comment in __pte_alloc */
3615 spin_lock(&mm->page_table_lock);
3616 #ifndef __ARCH_HAS_4LEVEL_HACK
3617 if (pud_present(*pud)) /* Another has populated it */
3620 pud_populate(mm, pud, new);
3622 if (pgd_present(*pud)) /* Another has populated it */
3625 pgd_populate(mm, pud, new);
3626 #endif /* __ARCH_HAS_4LEVEL_HACK */
3627 spin_unlock(&mm->page_table_lock);
3630 #endif /* __PAGETABLE_PMD_FOLDED */
3632 int make_pages_present(unsigned long addr, unsigned long end)
3634 int ret, len, write;
3635 struct vm_area_struct * vma;
3637 vma = find_vma(current->mm, addr);
3641 * We want to touch writable mappings with a write fault in order
3642 * to break COW, except for shared mappings because these don't COW
3643 * and we would not want to dirty them for nothing.
3645 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3646 BUG_ON(addr >= end);
3647 BUG_ON(end > vma->vm_end);
3648 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3649 ret = get_user_pages(current, current->mm, addr,
3650 len, write, 0, NULL, NULL);
3653 return ret == len ? 0 : -EFAULT;
3656 #if !defined(__HAVE_ARCH_GATE_AREA)
3658 #if defined(AT_SYSINFO_EHDR)
3659 static struct vm_area_struct gate_vma;
3661 static int __init gate_vma_init(void)
3663 gate_vma.vm_mm = NULL;
3664 gate_vma.vm_start = FIXADDR_USER_START;
3665 gate_vma.vm_end = FIXADDR_USER_END;
3666 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3667 gate_vma.vm_page_prot = __P101;
3671 __initcall(gate_vma_init);
3674 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3676 #ifdef AT_SYSINFO_EHDR
3683 int in_gate_area_no_mm(unsigned long addr)
3685 #ifdef AT_SYSINFO_EHDR
3686 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3692 #endif /* __HAVE_ARCH_GATE_AREA */
3694 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3695 pte_t **ptepp, spinlock_t **ptlp)
3702 pgd = pgd_offset(mm, address);
3703 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3706 pud = pud_offset(pgd, address);
3707 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3710 pmd = pmd_offset(pud, address);
3711 VM_BUG_ON(pmd_trans_huge(*pmd));
3712 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3715 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3719 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3722 if (!pte_present(*ptep))
3727 pte_unmap_unlock(ptep, *ptlp);
3732 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3733 pte_t **ptepp, spinlock_t **ptlp)
3737 /* (void) is needed to make gcc happy */
3738 (void) __cond_lock(*ptlp,
3739 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3744 * follow_pfn - look up PFN at a user virtual address
3745 * @vma: memory mapping
3746 * @address: user virtual address
3747 * @pfn: location to store found PFN
3749 * Only IO mappings and raw PFN mappings are allowed.
3751 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3753 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3760 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3763 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3766 *pfn = pte_pfn(*ptep);
3767 pte_unmap_unlock(ptep, ptl);
3770 EXPORT_SYMBOL(follow_pfn);
3772 #ifdef CONFIG_HAVE_IOREMAP_PROT
3773 int follow_phys(struct vm_area_struct *vma,
3774 unsigned long address, unsigned int flags,
3775 unsigned long *prot, resource_size_t *phys)
3781 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3784 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3788 if ((flags & FOLL_WRITE) && !pte_write(pte))
3791 *prot = pgprot_val(pte_pgprot(pte));
3792 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3796 pte_unmap_unlock(ptep, ptl);
3801 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3802 void *buf, int len, int write)
3804 resource_size_t phys_addr;
3805 unsigned long prot = 0;
3806 void __iomem *maddr;
3807 int offset = addr & (PAGE_SIZE-1);
3809 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3812 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3814 memcpy_toio(maddr + offset, buf, len);
3816 memcpy_fromio(buf, maddr + offset, len);
3824 * Access another process' address space as given in mm. If non-NULL, use the
3825 * given task for page fault accounting.
3827 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3828 unsigned long addr, void *buf, int len, int write)
3830 struct vm_area_struct *vma;
3831 void *old_buf = buf;
3833 down_read(&mm->mmap_sem);
3834 /* ignore errors, just check how much was successfully transferred */
3836 int bytes, ret, offset;
3838 struct page *page = NULL;
3840 ret = get_user_pages(tsk, mm, addr, 1,
3841 write, 1, &page, &vma);
3844 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3845 * we can access using slightly different code.
3847 #ifdef CONFIG_HAVE_IOREMAP_PROT
3848 vma = find_vma(mm, addr);
3849 if (!vma || vma->vm_start > addr)
3851 if (vma->vm_ops && vma->vm_ops->access)
3852 ret = vma->vm_ops->access(vma, addr, buf,
3860 offset = addr & (PAGE_SIZE-1);
3861 if (bytes > PAGE_SIZE-offset)
3862 bytes = PAGE_SIZE-offset;
3866 copy_to_user_page(vma, page, addr,
3867 maddr + offset, buf, bytes);
3868 set_page_dirty_lock(page);
3870 copy_from_user_page(vma, page, addr,
3871 buf, maddr + offset, bytes);
3874 page_cache_release(page);
3880 up_read(&mm->mmap_sem);
3882 return buf - old_buf;
3886 * access_remote_vm - access another process' address space
3887 * @mm: the mm_struct of the target address space
3888 * @addr: start address to access
3889 * @buf: source or destination buffer
3890 * @len: number of bytes to transfer
3891 * @write: whether the access is a write
3893 * The caller must hold a reference on @mm.
3895 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3896 void *buf, int len, int write)
3898 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3902 * Access another process' address space.
3903 * Source/target buffer must be kernel space,
3904 * Do not walk the page table directly, use get_user_pages
3906 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3907 void *buf, int len, int write)
3909 struct mm_struct *mm;
3912 mm = get_task_mm(tsk);
3916 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3923 * Print the name of a VMA.
3925 void print_vma_addr(char *prefix, unsigned long ip)
3927 struct mm_struct *mm = current->mm;
3928 struct vm_area_struct *vma;
3931 * Do not print if we are in atomic
3932 * contexts (in exception stacks, etc.):
3934 if (preempt_count())
3937 down_read(&mm->mmap_sem);
3938 vma = find_vma(mm, ip);
3939 if (vma && vma->vm_file) {
3940 struct file *f = vma->vm_file;
3941 char *buf = (char *)__get_free_page(GFP_KERNEL);
3945 p = d_path(&f->f_path, buf, PAGE_SIZE);
3948 s = strrchr(p, '/');
3951 printk("%s%s[%lx+%lx]", prefix, p,
3953 vma->vm_end - vma->vm_start);
3954 free_page((unsigned long)buf);
3957 up_read(¤t->mm->mmap_sem);
3960 #ifdef CONFIG_PROVE_LOCKING
3961 void might_fault(void)
3964 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3965 * holding the mmap_sem, this is safe because kernel memory doesn't
3966 * get paged out, therefore we'll never actually fault, and the
3967 * below annotations will generate false positives.
3969 if (segment_eq(get_fs(), KERNEL_DS))
3974 * it would be nicer only to annotate paths which are not under
3975 * pagefault_disable, however that requires a larger audit and
3976 * providing helpers like get_user_atomic.
3978 if (!in_atomic() && current->mm)
3979 might_lock_read(¤t->mm->mmap_sem);
3981 EXPORT_SYMBOL(might_fault);
3984 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3985 static void clear_gigantic_page(struct page *page,
3987 unsigned int pages_per_huge_page)
3990 struct page *p = page;
3993 for (i = 0; i < pages_per_huge_page;
3994 i++, p = mem_map_next(p, page, i)) {
3996 clear_user_highpage(p, addr + i * PAGE_SIZE);
3999 void clear_huge_page(struct page *page,
4000 unsigned long addr, unsigned int pages_per_huge_page)
4004 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4005 clear_gigantic_page(page, addr, pages_per_huge_page);
4010 for (i = 0; i < pages_per_huge_page; i++) {
4012 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4016 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4018 struct vm_area_struct *vma,
4019 unsigned int pages_per_huge_page)
4022 struct page *dst_base = dst;
4023 struct page *src_base = src;
4025 for (i = 0; i < pages_per_huge_page; ) {
4027 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4030 dst = mem_map_next(dst, dst_base, i);
4031 src = mem_map_next(src, src_base, i);
4035 void copy_user_huge_page(struct page *dst, struct page *src,
4036 unsigned long addr, struct vm_area_struct *vma,
4037 unsigned int pages_per_huge_page)
4041 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4042 copy_user_gigantic_page(dst, src, addr, vma,
4043 pages_per_huge_page);
4048 for (i = 0; i < pages_per_huge_page; i++) {
4050 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4053 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */