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 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
138 task->rss_stat.events = 0;
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 struct task_struct *task = current;
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
148 add_mm_counter(mm, member, val);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct *task)
157 if (unlikely(task != current))
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
171 val = atomic_long_read(&mm->rss_stat.count[member]);
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
178 return (unsigned long)val;
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
183 __sync_task_rss_stat(task, mm);
185 #else /* SPLIT_RSS_COUNTING */
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
190 static void check_sync_rss_stat(struct task_struct *task)
194 #endif /* SPLIT_RSS_COUNTING */
196 #ifdef HAVE_GENERIC_MMU_GATHER
198 static int tlb_next_batch(struct mmu_gather *tlb)
200 struct mmu_gather_batch *batch;
204 tlb->active = batch->next;
208 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
211 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
218 batch->max = MAX_GATHER_BATCH;
220 tlb->active->next = batch;
227 * Called to initialize an (on-stack) mmu_gather structure for page-table
228 * tear-down from @mm. The @fullmm argument is used when @mm is without
229 * users and we're going to destroy the full address space (exit/execve).
231 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
235 tlb->fullmm = fullmm;
237 tlb->fast_mode = (num_possible_cpus() == 1);
238 tlb->local.next = NULL;
240 tlb->local.max = ARRAY_SIZE(tlb->__pages);
241 tlb->active = &tlb->local;
242 tlb->batch_count = 0;
244 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 void tlb_flush_mmu(struct mmu_gather *tlb)
251 struct mmu_gather_batch *batch;
253 if (!tlb->need_flush)
257 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
258 tlb_table_flush(tlb);
261 if (tlb_fast_mode(tlb))
264 for (batch = &tlb->local; batch; batch = batch->next) {
265 free_pages_and_swap_cache(batch->pages, batch->nr);
268 tlb->active = &tlb->local;
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
275 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
277 struct mmu_gather_batch *batch, *next;
281 /* keep the page table cache within bounds */
284 for (batch = tlb->local.next; batch; batch = next) {
286 free_pages((unsigned long)batch, 0);
288 tlb->local.next = NULL;
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
297 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
299 struct mmu_gather_batch *batch;
303 if (tlb_fast_mode(tlb)) {
304 free_page_and_swap_cache(page);
305 return 1; /* avoid calling tlb_flush_mmu() */
309 batch->pages[batch->nr++] = page;
310 if (batch->nr == batch->max) {
311 if (!tlb_next_batch(tlb))
315 VM_BUG_ON(batch->nr > batch->max);
317 return batch->max - batch->nr;
320 #endif /* HAVE_GENERIC_MMU_GATHER */
322 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
325 * See the comment near struct mmu_table_batch.
328 static void tlb_remove_table_smp_sync(void *arg)
330 /* Simply deliver the interrupt */
333 static void tlb_remove_table_one(void *table)
336 * This isn't an RCU grace period and hence the page-tables cannot be
337 * assumed to be actually RCU-freed.
339 * It is however sufficient for software page-table walkers that rely on
340 * IRQ disabling. See the comment near struct mmu_table_batch.
342 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
343 __tlb_remove_table(table);
346 static void tlb_remove_table_rcu(struct rcu_head *head)
348 struct mmu_table_batch *batch;
351 batch = container_of(head, struct mmu_table_batch, rcu);
353 for (i = 0; i < batch->nr; i++)
354 __tlb_remove_table(batch->tables[i]);
356 free_page((unsigned long)batch);
359 void tlb_table_flush(struct mmu_gather *tlb)
361 struct mmu_table_batch **batch = &tlb->batch;
364 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
369 void tlb_remove_table(struct mmu_gather *tlb, void *table)
371 struct mmu_table_batch **batch = &tlb->batch;
376 * When there's less then two users of this mm there cannot be a
377 * concurrent page-table walk.
379 if (atomic_read(&tlb->mm->mm_users) < 2) {
380 __tlb_remove_table(table);
384 if (*batch == NULL) {
385 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
386 if (*batch == NULL) {
387 tlb_remove_table_one(table);
392 (*batch)->tables[(*batch)->nr++] = table;
393 if ((*batch)->nr == MAX_TABLE_BATCH)
394 tlb_table_flush(tlb);
397 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
400 * If a p?d_bad entry is found while walking page tables, report
401 * the error, before resetting entry to p?d_none. Usually (but
402 * very seldom) called out from the p?d_none_or_clear_bad macros.
405 void pgd_clear_bad(pgd_t *pgd)
411 void pud_clear_bad(pud_t *pud)
417 void pmd_clear_bad(pmd_t *pmd)
424 * Note: this doesn't free the actual pages themselves. That
425 * has been handled earlier when unmapping all the memory regions.
427 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
430 pgtable_t token = pmd_pgtable(*pmd);
432 pte_free_tlb(tlb, token, addr);
436 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
437 unsigned long addr, unsigned long end,
438 unsigned long floor, unsigned long ceiling)
445 pmd = pmd_offset(pud, addr);
447 next = pmd_addr_end(addr, end);
448 if (pmd_none_or_clear_bad(pmd))
450 free_pte_range(tlb, pmd, addr);
451 } while (pmd++, addr = next, addr != end);
461 if (end - 1 > ceiling - 1)
464 pmd = pmd_offset(pud, start);
466 pmd_free_tlb(tlb, pmd, start);
469 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
470 unsigned long addr, unsigned long end,
471 unsigned long floor, unsigned long ceiling)
478 pud = pud_offset(pgd, addr);
480 next = pud_addr_end(addr, end);
481 if (pud_none_or_clear_bad(pud))
483 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
484 } while (pud++, addr = next, addr != end);
490 ceiling &= PGDIR_MASK;
494 if (end - 1 > ceiling - 1)
497 pud = pud_offset(pgd, start);
499 pud_free_tlb(tlb, pud, start);
503 * This function frees user-level page tables of a process.
505 * Must be called with pagetable lock held.
507 void free_pgd_range(struct mmu_gather *tlb,
508 unsigned long addr, unsigned long end,
509 unsigned long floor, unsigned long ceiling)
515 * The next few lines have given us lots of grief...
517 * Why are we testing PMD* at this top level? Because often
518 * there will be no work to do at all, and we'd prefer not to
519 * go all the way down to the bottom just to discover that.
521 * Why all these "- 1"s? Because 0 represents both the bottom
522 * of the address space and the top of it (using -1 for the
523 * top wouldn't help much: the masks would do the wrong thing).
524 * The rule is that addr 0 and floor 0 refer to the bottom of
525 * the address space, but end 0 and ceiling 0 refer to the top
526 * Comparisons need to use "end - 1" and "ceiling - 1" (though
527 * that end 0 case should be mythical).
529 * Wherever addr is brought up or ceiling brought down, we must
530 * be careful to reject "the opposite 0" before it confuses the
531 * subsequent tests. But what about where end is brought down
532 * by PMD_SIZE below? no, end can't go down to 0 there.
534 * Whereas we round start (addr) and ceiling down, by different
535 * masks at different levels, in order to test whether a table
536 * now has no other vmas using it, so can be freed, we don't
537 * bother to round floor or end up - the tests don't need that.
551 if (end - 1 > ceiling - 1)
556 pgd = pgd_offset(tlb->mm, addr);
558 next = pgd_addr_end(addr, end);
559 if (pgd_none_or_clear_bad(pgd))
561 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
562 } while (pgd++, addr = next, addr != end);
565 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
566 unsigned long floor, unsigned long ceiling)
569 struct vm_area_struct *next = vma->vm_next;
570 unsigned long addr = vma->vm_start;
573 * Hide vma from rmap and truncate_pagecache before freeing
576 unlink_anon_vmas(vma);
577 unlink_file_vma(vma);
579 if (is_vm_hugetlb_page(vma)) {
580 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
581 floor, next? next->vm_start: ceiling);
584 * Optimization: gather nearby vmas into one call down
586 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
587 && !is_vm_hugetlb_page(next)) {
590 unlink_anon_vmas(vma);
591 unlink_file_vma(vma);
593 free_pgd_range(tlb, addr, vma->vm_end,
594 floor, next? next->vm_start: ceiling);
600 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
601 pmd_t *pmd, unsigned long address)
603 pgtable_t new = pte_alloc_one(mm, address);
604 int wait_split_huge_page;
609 * Ensure all pte setup (eg. pte page lock and page clearing) are
610 * visible before the pte is made visible to other CPUs by being
611 * put into page tables.
613 * The other side of the story is the pointer chasing in the page
614 * table walking code (when walking the page table without locking;
615 * ie. most of the time). Fortunately, these data accesses consist
616 * of a chain of data-dependent loads, meaning most CPUs (alpha
617 * being the notable exception) will already guarantee loads are
618 * seen in-order. See the alpha page table accessors for the
619 * smp_read_barrier_depends() barriers in page table walking code.
621 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
623 spin_lock(&mm->page_table_lock);
624 wait_split_huge_page = 0;
625 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
627 pmd_populate(mm, pmd, new);
629 } else if (unlikely(pmd_trans_splitting(*pmd)))
630 wait_split_huge_page = 1;
631 spin_unlock(&mm->page_table_lock);
634 if (wait_split_huge_page)
635 wait_split_huge_page(vma->anon_vma, pmd);
639 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
641 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
645 smp_wmb(); /* See comment in __pte_alloc */
647 spin_lock(&init_mm.page_table_lock);
648 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
649 pmd_populate_kernel(&init_mm, pmd, new);
652 VM_BUG_ON(pmd_trans_splitting(*pmd));
653 spin_unlock(&init_mm.page_table_lock);
655 pte_free_kernel(&init_mm, new);
659 static inline void init_rss_vec(int *rss)
661 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
664 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
668 if (current->mm == mm)
669 sync_mm_rss(current, mm);
670 for (i = 0; i < NR_MM_COUNTERS; i++)
672 add_mm_counter(mm, i, rss[i]);
676 * This function is called to print an error when a bad pte
677 * is found. For example, we might have a PFN-mapped pte in
678 * a region that doesn't allow it.
680 * The calling function must still handle the error.
682 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
683 pte_t pte, struct page *page)
685 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
686 pud_t *pud = pud_offset(pgd, addr);
687 pmd_t *pmd = pmd_offset(pud, addr);
688 struct address_space *mapping;
690 static unsigned long resume;
691 static unsigned long nr_shown;
692 static unsigned long nr_unshown;
695 * Allow a burst of 60 reports, then keep quiet for that minute;
696 * or allow a steady drip of one report per second.
698 if (nr_shown == 60) {
699 if (time_before(jiffies, resume)) {
705 "BUG: Bad page map: %lu messages suppressed\n",
712 resume = jiffies + 60 * HZ;
714 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
715 index = linear_page_index(vma, addr);
718 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
720 (long long)pte_val(pte), (long long)pmd_val(*pmd));
724 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
725 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
727 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
730 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
731 (unsigned long)vma->vm_ops->fault);
732 if (vma->vm_file && vma->vm_file->f_op)
733 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
734 (unsigned long)vma->vm_file->f_op->mmap);
736 add_taint(TAINT_BAD_PAGE);
739 static inline int is_cow_mapping(vm_flags_t flags)
741 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
745 static inline int is_zero_pfn(unsigned long pfn)
747 return pfn == zero_pfn;
752 static inline unsigned long my_zero_pfn(unsigned long addr)
759 * vm_normal_page -- This function gets the "struct page" associated with a pte.
761 * "Special" mappings do not wish to be associated with a "struct page" (either
762 * it doesn't exist, or it exists but they don't want to touch it). In this
763 * case, NULL is returned here. "Normal" mappings do have a struct page.
765 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
766 * pte bit, in which case this function is trivial. Secondly, an architecture
767 * may not have a spare pte bit, which requires a more complicated scheme,
770 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
771 * special mapping (even if there are underlying and valid "struct pages").
772 * COWed pages of a VM_PFNMAP are always normal.
774 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
775 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
776 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
777 * mapping will always honor the rule
779 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
781 * And for normal mappings this is false.
783 * This restricts such mappings to be a linear translation from virtual address
784 * to pfn. To get around this restriction, we allow arbitrary mappings so long
785 * as the vma is not a COW mapping; in that case, we know that all ptes are
786 * special (because none can have been COWed).
789 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
791 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
792 * page" backing, however the difference is that _all_ pages with a struct
793 * page (that is, those where pfn_valid is true) are refcounted and considered
794 * normal pages by the VM. The disadvantage is that pages are refcounted
795 * (which can be slower and simply not an option for some PFNMAP users). The
796 * advantage is that we don't have to follow the strict linearity rule of
797 * PFNMAP mappings in order to support COWable mappings.
800 #ifdef __HAVE_ARCH_PTE_SPECIAL
801 # define HAVE_PTE_SPECIAL 1
803 # define HAVE_PTE_SPECIAL 0
805 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
808 unsigned long pfn = pte_pfn(pte);
810 if (HAVE_PTE_SPECIAL) {
811 if (likely(!pte_special(pte)))
813 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
815 if (!is_zero_pfn(pfn))
816 print_bad_pte(vma, addr, pte, NULL);
820 /* !HAVE_PTE_SPECIAL case follows: */
822 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
823 if (vma->vm_flags & VM_MIXEDMAP) {
829 off = (addr - vma->vm_start) >> PAGE_SHIFT;
830 if (pfn == vma->vm_pgoff + off)
832 if (!is_cow_mapping(vma->vm_flags))
837 if (is_zero_pfn(pfn))
840 if (unlikely(pfn > highest_memmap_pfn)) {
841 print_bad_pte(vma, addr, pte, NULL);
846 * NOTE! We still have PageReserved() pages in the page tables.
847 * eg. VDSO mappings can cause them to exist.
850 return pfn_to_page(pfn);
854 * copy one vm_area from one task to the other. Assumes the page tables
855 * already present in the new task to be cleared in the whole range
856 * covered by this vma.
859 static inline unsigned long
860 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
861 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
862 unsigned long addr, int *rss)
864 unsigned long vm_flags = vma->vm_flags;
865 pte_t pte = *src_pte;
868 /* pte contains position in swap or file, so copy. */
869 if (unlikely(!pte_present(pte))) {
870 if (!pte_file(pte)) {
871 swp_entry_t entry = pte_to_swp_entry(pte);
873 if (swap_duplicate(entry) < 0)
876 /* make sure dst_mm is on swapoff's mmlist. */
877 if (unlikely(list_empty(&dst_mm->mmlist))) {
878 spin_lock(&mmlist_lock);
879 if (list_empty(&dst_mm->mmlist))
880 list_add(&dst_mm->mmlist,
882 spin_unlock(&mmlist_lock);
884 if (likely(!non_swap_entry(entry)))
886 else if (is_write_migration_entry(entry) &&
887 is_cow_mapping(vm_flags)) {
889 * COW mappings require pages in both parent
890 * and child to be set to read.
892 make_migration_entry_read(&entry);
893 pte = swp_entry_to_pte(entry);
894 set_pte_at(src_mm, addr, src_pte, pte);
901 * If it's a COW mapping, write protect it both
902 * in the parent and the child
904 if (is_cow_mapping(vm_flags)) {
905 ptep_set_wrprotect(src_mm, addr, src_pte);
906 pte = pte_wrprotect(pte);
910 * If it's a shared mapping, mark it clean in
913 if (vm_flags & VM_SHARED)
914 pte = pte_mkclean(pte);
915 pte = pte_mkold(pte);
917 page = vm_normal_page(vma, addr, pte);
928 set_pte_at(dst_mm, addr, dst_pte, pte);
932 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
933 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
934 unsigned long addr, unsigned long end)
936 pte_t *orig_src_pte, *orig_dst_pte;
937 pte_t *src_pte, *dst_pte;
938 spinlock_t *src_ptl, *dst_ptl;
940 int rss[NR_MM_COUNTERS];
941 swp_entry_t entry = (swp_entry_t){0};
946 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
949 src_pte = pte_offset_map(src_pmd, addr);
950 src_ptl = pte_lockptr(src_mm, src_pmd);
951 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
952 orig_src_pte = src_pte;
953 orig_dst_pte = dst_pte;
954 arch_enter_lazy_mmu_mode();
958 * We are holding two locks at this point - either of them
959 * could generate latencies in another task on another CPU.
961 if (progress >= 32) {
963 if (need_resched() ||
964 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
967 if (pte_none(*src_pte)) {
971 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
976 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
978 arch_leave_lazy_mmu_mode();
979 spin_unlock(src_ptl);
980 pte_unmap(orig_src_pte);
981 add_mm_rss_vec(dst_mm, rss);
982 pte_unmap_unlock(orig_dst_pte, dst_ptl);
986 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
995 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
996 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
997 unsigned long addr, unsigned long end)
999 pmd_t *src_pmd, *dst_pmd;
1002 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1005 src_pmd = pmd_offset(src_pud, addr);
1007 next = pmd_addr_end(addr, end);
1008 if (pmd_trans_huge(*src_pmd)) {
1010 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1011 err = copy_huge_pmd(dst_mm, src_mm,
1012 dst_pmd, src_pmd, addr, vma);
1019 if (pmd_none_or_clear_bad(src_pmd))
1021 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1024 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1028 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1029 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1030 unsigned long addr, unsigned long end)
1032 pud_t *src_pud, *dst_pud;
1035 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1038 src_pud = pud_offset(src_pgd, addr);
1040 next = pud_addr_end(addr, end);
1041 if (pud_none_or_clear_bad(src_pud))
1043 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1046 } while (dst_pud++, src_pud++, addr = next, addr != end);
1050 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1051 struct vm_area_struct *vma)
1053 pgd_t *src_pgd, *dst_pgd;
1055 unsigned long addr = vma->vm_start;
1056 unsigned long end = vma->vm_end;
1060 * Don't copy ptes where a page fault will fill them correctly.
1061 * Fork becomes much lighter when there are big shared or private
1062 * readonly mappings. The tradeoff is that copy_page_range is more
1063 * efficient than faulting.
1065 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
1070 if (is_vm_hugetlb_page(vma))
1071 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1073 if (unlikely(is_pfn_mapping(vma))) {
1075 * We do not free on error cases below as remove_vma
1076 * gets called on error from higher level routine
1078 ret = track_pfn_vma_copy(vma);
1084 * We need to invalidate the secondary MMU mappings only when
1085 * there could be a permission downgrade on the ptes of the
1086 * parent mm. And a permission downgrade will only happen if
1087 * is_cow_mapping() returns true.
1089 if (is_cow_mapping(vma->vm_flags))
1090 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1093 dst_pgd = pgd_offset(dst_mm, addr);
1094 src_pgd = pgd_offset(src_mm, addr);
1096 next = pgd_addr_end(addr, end);
1097 if (pgd_none_or_clear_bad(src_pgd))
1099 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100 vma, addr, next))) {
1104 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1106 if (is_cow_mapping(vma->vm_flags))
1107 mmu_notifier_invalidate_range_end(src_mm,
1108 vma->vm_start, end);
1112 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1113 struct vm_area_struct *vma, pmd_t *pmd,
1114 unsigned long addr, unsigned long end,
1115 struct zap_details *details)
1117 struct mm_struct *mm = tlb->mm;
1118 int force_flush = 0;
1119 int rss[NR_MM_COUNTERS];
1126 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1128 arch_enter_lazy_mmu_mode();
1131 if (pte_none(ptent)) {
1135 if (pte_present(ptent)) {
1138 page = vm_normal_page(vma, addr, ptent);
1139 if (unlikely(details) && page) {
1141 * unmap_shared_mapping_pages() wants to
1142 * invalidate cache without truncating:
1143 * unmap shared but keep private pages.
1145 if (details->check_mapping &&
1146 details->check_mapping != page->mapping)
1149 * Each page->index must be checked when
1150 * invalidating or truncating nonlinear.
1152 if (details->nonlinear_vma &&
1153 (page->index < details->first_index ||
1154 page->index > details->last_index))
1157 ptent = ptep_get_and_clear_full(mm, addr, pte,
1159 tlb_remove_tlb_entry(tlb, pte, addr);
1160 if (unlikely(!page))
1162 if (unlikely(details) && details->nonlinear_vma
1163 && linear_page_index(details->nonlinear_vma,
1164 addr) != page->index)
1165 set_pte_at(mm, addr, pte,
1166 pgoff_to_pte(page->index));
1168 rss[MM_ANONPAGES]--;
1170 if (pte_dirty(ptent))
1171 set_page_dirty(page);
1172 if (pte_young(ptent) &&
1173 likely(!VM_SequentialReadHint(vma)))
1174 mark_page_accessed(page);
1175 rss[MM_FILEPAGES]--;
1177 page_remove_rmap(page);
1178 if (unlikely(page_mapcount(page) < 0))
1179 print_bad_pte(vma, addr, ptent, page);
1180 force_flush = !__tlb_remove_page(tlb, page);
1186 * If details->check_mapping, we leave swap entries;
1187 * if details->nonlinear_vma, we leave file entries.
1189 if (unlikely(details))
1191 if (pte_file(ptent)) {
1192 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1193 print_bad_pte(vma, addr, ptent, NULL);
1195 swp_entry_t entry = pte_to_swp_entry(ptent);
1197 if (!non_swap_entry(entry))
1199 if (unlikely(!free_swap_and_cache(entry)))
1200 print_bad_pte(vma, addr, ptent, NULL);
1202 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1203 } while (pte++, addr += PAGE_SIZE, addr != end);
1205 add_mm_rss_vec(mm, rss);
1206 arch_leave_lazy_mmu_mode();
1207 pte_unmap_unlock(start_pte, ptl);
1210 * mmu_gather ran out of room to batch pages, we break out of
1211 * the PTE lock to avoid doing the potential expensive TLB invalidate
1212 * and page-free while holding it.
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1232 pmd = pmd_offset(pud, addr);
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1238 split_huge_page_pmd(vma->vm_mm, pmd);
1239 } else if (zap_huge_pmd(tlb, vma, pmd))
1244 * Here there can be other concurrent MADV_DONTNEED or
1245 * trans huge page faults running, and if the pmd is
1246 * none or trans huge it can change under us. This is
1247 * because MADV_DONTNEED holds the mmap_sem in read
1250 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1252 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1255 } while (pmd++, addr = next, addr != end);
1260 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1261 struct vm_area_struct *vma, pgd_t *pgd,
1262 unsigned long addr, unsigned long end,
1263 struct zap_details *details)
1268 pud = pud_offset(pgd, addr);
1270 next = pud_addr_end(addr, end);
1271 if (pud_none_or_clear_bad(pud))
1273 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1274 } while (pud++, addr = next, addr != end);
1279 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1280 struct vm_area_struct *vma,
1281 unsigned long addr, unsigned long end,
1282 struct zap_details *details)
1287 if (details && !details->check_mapping && !details->nonlinear_vma)
1290 BUG_ON(addr >= end);
1291 mem_cgroup_uncharge_start();
1292 tlb_start_vma(tlb, vma);
1293 pgd = pgd_offset(vma->vm_mm, addr);
1295 next = pgd_addr_end(addr, end);
1296 if (pgd_none_or_clear_bad(pgd))
1298 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1299 } while (pgd++, addr = next, addr != end);
1300 tlb_end_vma(tlb, vma);
1301 mem_cgroup_uncharge_end();
1307 * unmap_vmas - unmap a range of memory covered by a list of vma's
1308 * @tlb: address of the caller's struct mmu_gather
1309 * @vma: the starting vma
1310 * @start_addr: virtual address at which to start unmapping
1311 * @end_addr: virtual address at which to end unmapping
1312 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1313 * @details: details of nonlinear truncation or shared cache invalidation
1315 * Returns the end address of the unmapping (restart addr if interrupted).
1317 * Unmap all pages in the vma list.
1319 * Only addresses between `start' and `end' will be unmapped.
1321 * The VMA list must be sorted in ascending virtual address order.
1323 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1324 * range after unmap_vmas() returns. So the only responsibility here is to
1325 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1326 * drops the lock and schedules.
1328 unsigned long unmap_vmas(struct mmu_gather *tlb,
1329 struct vm_area_struct *vma, unsigned long start_addr,
1330 unsigned long end_addr, unsigned long *nr_accounted,
1331 struct zap_details *details)
1333 unsigned long start = start_addr;
1334 struct mm_struct *mm = vma->vm_mm;
1336 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1337 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1340 start = max(vma->vm_start, start_addr);
1341 if (start >= vma->vm_end)
1343 end = min(vma->vm_end, end_addr);
1344 if (end <= vma->vm_start)
1347 if (vma->vm_flags & VM_ACCOUNT)
1348 *nr_accounted += (end - start) >> PAGE_SHIFT;
1350 if (unlikely(is_pfn_mapping(vma)))
1351 untrack_pfn_vma(vma, 0, 0);
1353 while (start != end) {
1354 if (unlikely(is_vm_hugetlb_page(vma))) {
1356 * It is undesirable to test vma->vm_file as it
1357 * should be non-null for valid hugetlb area.
1358 * However, vm_file will be NULL in the error
1359 * cleanup path of do_mmap_pgoff. When
1360 * hugetlbfs ->mmap method fails,
1361 * do_mmap_pgoff() nullifies vma->vm_file
1362 * before calling this function to clean up.
1363 * Since no pte has actually been setup, it is
1364 * safe to do nothing in this case.
1367 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1368 __unmap_hugepage_range_final(vma, start, end, NULL);
1369 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1374 start = unmap_page_range(tlb, vma, start, end, details);
1378 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1379 return start; /* which is now the end (or restart) address */
1383 * zap_page_range - remove user pages in a given range
1384 * @vma: vm_area_struct holding the applicable pages
1385 * @address: starting address of pages to zap
1386 * @size: number of bytes to zap
1387 * @details: details of nonlinear truncation or shared cache invalidation
1389 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1390 unsigned long size, struct zap_details *details)
1392 struct mm_struct *mm = vma->vm_mm;
1393 struct mmu_gather tlb;
1394 unsigned long end = address + size;
1395 unsigned long nr_accounted = 0;
1398 tlb_gather_mmu(&tlb, mm, 0);
1399 update_hiwater_rss(mm);
1400 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1401 tlb_finish_mmu(&tlb, address, end);
1404 EXPORT_SYMBOL_GPL(zap_page_range);
1407 * zap_vma_ptes - remove ptes mapping the vma
1408 * @vma: vm_area_struct holding ptes to be zapped
1409 * @address: starting address of pages to zap
1410 * @size: number of bytes to zap
1412 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1414 * The entire address range must be fully contained within the vma.
1416 * Returns 0 if successful.
1418 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1421 if (address < vma->vm_start || address + size > vma->vm_end ||
1422 !(vma->vm_flags & VM_PFNMAP))
1424 zap_page_range(vma, address, size, NULL);
1427 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1430 * follow_page - look up a page descriptor from a user-virtual address
1431 * @vma: vm_area_struct mapping @address
1432 * @address: virtual address to look up
1433 * @flags: flags modifying lookup behaviour
1435 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1437 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1438 * an error pointer if there is a mapping to something not represented
1439 * by a page descriptor (see also vm_normal_page()).
1441 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1450 struct mm_struct *mm = vma->vm_mm;
1452 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1453 if (!IS_ERR(page)) {
1454 BUG_ON(flags & FOLL_GET);
1459 pgd = pgd_offset(mm, address);
1460 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1463 pud = pud_offset(pgd, address);
1466 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1467 BUG_ON(flags & FOLL_GET);
1468 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1471 if (unlikely(pud_bad(*pud)))
1474 pmd = pmd_offset(pud, address);
1477 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1478 BUG_ON(flags & FOLL_GET);
1479 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1482 if (pmd_trans_huge(*pmd)) {
1483 if (flags & FOLL_SPLIT) {
1484 split_huge_page_pmd(mm, pmd);
1485 goto split_fallthrough;
1487 spin_lock(&mm->page_table_lock);
1488 if (likely(pmd_trans_huge(*pmd))) {
1489 if (unlikely(pmd_trans_splitting(*pmd))) {
1490 spin_unlock(&mm->page_table_lock);
1491 wait_split_huge_page(vma->anon_vma, pmd);
1493 page = follow_trans_huge_pmd(mm, address,
1495 spin_unlock(&mm->page_table_lock);
1499 spin_unlock(&mm->page_table_lock);
1503 if (unlikely(pmd_bad(*pmd)))
1506 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1509 if (!pte_present(pte))
1511 if ((flags & FOLL_WRITE) && !pte_write(pte))
1514 page = vm_normal_page(vma, address, pte);
1515 if (unlikely(!page)) {
1516 if ((flags & FOLL_DUMP) ||
1517 !is_zero_pfn(pte_pfn(pte)))
1519 page = pte_page(pte);
1522 if (flags & FOLL_GET)
1523 get_page_foll(page);
1524 if (flags & FOLL_TOUCH) {
1525 if ((flags & FOLL_WRITE) &&
1526 !pte_dirty(pte) && !PageDirty(page))
1527 set_page_dirty(page);
1529 * pte_mkyoung() would be more correct here, but atomic care
1530 * is needed to avoid losing the dirty bit: it is easier to use
1531 * mark_page_accessed().
1533 mark_page_accessed(page);
1535 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1537 * The preliminary mapping check is mainly to avoid the
1538 * pointless overhead of lock_page on the ZERO_PAGE
1539 * which might bounce very badly if there is contention.
1541 * If the page is already locked, we don't need to
1542 * handle it now - vmscan will handle it later if and
1543 * when it attempts to reclaim the page.
1545 if (page->mapping && trylock_page(page)) {
1546 lru_add_drain(); /* push cached pages to LRU */
1548 * Because we lock page here and migration is
1549 * blocked by the pte's page reference, we need
1550 * only check for file-cache page truncation.
1553 mlock_vma_page(page);
1558 pte_unmap_unlock(ptep, ptl);
1563 pte_unmap_unlock(ptep, ptl);
1564 return ERR_PTR(-EFAULT);
1567 pte_unmap_unlock(ptep, ptl);
1573 * When core dumping an enormous anonymous area that nobody
1574 * has touched so far, we don't want to allocate unnecessary pages or
1575 * page tables. Return error instead of NULL to skip handle_mm_fault,
1576 * then get_dump_page() will return NULL to leave a hole in the dump.
1577 * But we can only make this optimization where a hole would surely
1578 * be zero-filled if handle_mm_fault() actually did handle it.
1580 if ((flags & FOLL_DUMP) &&
1581 (!vma->vm_ops || !vma->vm_ops->fault))
1582 return ERR_PTR(-EFAULT);
1586 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1588 return stack_guard_page_start(vma, addr) ||
1589 stack_guard_page_end(vma, addr+PAGE_SIZE);
1593 * __get_user_pages() - pin user pages in memory
1594 * @tsk: task_struct of target task
1595 * @mm: mm_struct of target mm
1596 * @start: starting user address
1597 * @nr_pages: number of pages from start to pin
1598 * @gup_flags: flags modifying pin behaviour
1599 * @pages: array that receives pointers to the pages pinned.
1600 * Should be at least nr_pages long. Or NULL, if caller
1601 * only intends to ensure the pages are faulted in.
1602 * @vmas: array of pointers to vmas corresponding to each page.
1603 * Or NULL if the caller does not require them.
1604 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1606 * Returns number of pages pinned. This may be fewer than the number
1607 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1608 * were pinned, returns -errno. Each page returned must be released
1609 * with a put_page() call when it is finished with. vmas will only
1610 * remain valid while mmap_sem is held.
1612 * Must be called with mmap_sem held for read or write.
1614 * __get_user_pages walks a process's page tables and takes a reference to
1615 * each struct page that each user address corresponds to at a given
1616 * instant. That is, it takes the page that would be accessed if a user
1617 * thread accesses the given user virtual address at that instant.
1619 * This does not guarantee that the page exists in the user mappings when
1620 * __get_user_pages returns, and there may even be a completely different
1621 * page there in some cases (eg. if mmapped pagecache has been invalidated
1622 * and subsequently re faulted). However it does guarantee that the page
1623 * won't be freed completely. And mostly callers simply care that the page
1624 * contains data that was valid *at some point in time*. Typically, an IO
1625 * or similar operation cannot guarantee anything stronger anyway because
1626 * locks can't be held over the syscall boundary.
1628 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1629 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1630 * appropriate) must be called after the page is finished with, and
1631 * before put_page is called.
1633 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1634 * or mmap_sem contention, and if waiting is needed to pin all pages,
1635 * *@nonblocking will be set to 0.
1637 * In most cases, get_user_pages or get_user_pages_fast should be used
1638 * instead of __get_user_pages. __get_user_pages should be used only if
1639 * you need some special @gup_flags.
1641 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1642 unsigned long start, int nr_pages, unsigned int gup_flags,
1643 struct page **pages, struct vm_area_struct **vmas,
1647 unsigned long vm_flags;
1652 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1655 * Require read or write permissions.
1656 * If FOLL_FORCE is set, we only require the "MAY" flags.
1658 vm_flags = (gup_flags & FOLL_WRITE) ?
1659 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1660 vm_flags &= (gup_flags & FOLL_FORCE) ?
1661 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1665 struct vm_area_struct *vma;
1667 vma = find_extend_vma(mm, start);
1668 if (!vma && in_gate_area(mm, start)) {
1669 unsigned long pg = start & PAGE_MASK;
1675 /* user gate pages are read-only */
1676 if (gup_flags & FOLL_WRITE)
1677 return i ? : -EFAULT;
1679 pgd = pgd_offset_k(pg);
1681 pgd = pgd_offset_gate(mm, pg);
1682 BUG_ON(pgd_none(*pgd));
1683 pud = pud_offset(pgd, pg);
1684 BUG_ON(pud_none(*pud));
1685 pmd = pmd_offset(pud, pg);
1687 return i ? : -EFAULT;
1688 VM_BUG_ON(pmd_trans_huge(*pmd));
1689 pte = pte_offset_map(pmd, pg);
1690 if (pte_none(*pte)) {
1692 return i ? : -EFAULT;
1694 vma = get_gate_vma(mm);
1698 page = vm_normal_page(vma, start, *pte);
1700 if (!(gup_flags & FOLL_DUMP) &&
1701 is_zero_pfn(pte_pfn(*pte)))
1702 page = pte_page(*pte);
1705 return i ? : -EFAULT;
1716 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1717 !(vm_flags & vma->vm_flags))
1718 return i ? : -EFAULT;
1720 if (is_vm_hugetlb_page(vma)) {
1721 i = follow_hugetlb_page(mm, vma, pages, vmas,
1722 &start, &nr_pages, i, gup_flags);
1728 unsigned int foll_flags = gup_flags;
1731 * If we have a pending SIGKILL, don't keep faulting
1732 * pages and potentially allocating memory.
1734 if (unlikely(fatal_signal_pending(current)))
1735 return i ? i : -ERESTARTSYS;
1738 while (!(page = follow_page(vma, start, foll_flags))) {
1740 unsigned int fault_flags = 0;
1742 /* For mlock, just skip the stack guard page. */
1743 if (foll_flags & FOLL_MLOCK) {
1744 if (stack_guard_page(vma, start))
1747 if (foll_flags & FOLL_WRITE)
1748 fault_flags |= FAULT_FLAG_WRITE;
1750 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1751 if (foll_flags & FOLL_NOWAIT)
1752 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1754 ret = handle_mm_fault(mm, vma, start,
1757 if (ret & VM_FAULT_ERROR) {
1758 if (ret & VM_FAULT_OOM)
1759 return i ? i : -ENOMEM;
1760 if (ret & (VM_FAULT_HWPOISON |
1761 VM_FAULT_HWPOISON_LARGE)) {
1764 else if (gup_flags & FOLL_HWPOISON)
1769 if (ret & VM_FAULT_SIGBUS)
1770 return i ? i : -EFAULT;
1775 if (ret & VM_FAULT_MAJOR)
1781 if (ret & VM_FAULT_RETRY) {
1788 * The VM_FAULT_WRITE bit tells us that
1789 * do_wp_page has broken COW when necessary,
1790 * even if maybe_mkwrite decided not to set
1791 * pte_write. We can thus safely do subsequent
1792 * page lookups as if they were reads. But only
1793 * do so when looping for pte_write is futile:
1794 * in some cases userspace may also be wanting
1795 * to write to the gotten user page, which a
1796 * read fault here might prevent (a readonly
1797 * page might get reCOWed by userspace write).
1799 if ((ret & VM_FAULT_WRITE) &&
1800 !(vma->vm_flags & VM_WRITE))
1801 foll_flags &= ~FOLL_WRITE;
1806 return i ? i : PTR_ERR(page);
1810 flush_anon_page(vma, page, start);
1811 flush_dcache_page(page);
1819 } while (nr_pages && start < vma->vm_end);
1823 EXPORT_SYMBOL(__get_user_pages);
1826 * fixup_user_fault() - manually resolve a user page fault
1827 * @tsk: the task_struct to use for page fault accounting, or
1828 * NULL if faults are not to be recorded.
1829 * @mm: mm_struct of target mm
1830 * @address: user address
1831 * @fault_flags:flags to pass down to handle_mm_fault()
1833 * This is meant to be called in the specific scenario where for locking reasons
1834 * we try to access user memory in atomic context (within a pagefault_disable()
1835 * section), this returns -EFAULT, and we want to resolve the user fault before
1838 * Typically this is meant to be used by the futex code.
1840 * The main difference with get_user_pages() is that this function will
1841 * unconditionally call handle_mm_fault() which will in turn perform all the
1842 * necessary SW fixup of the dirty and young bits in the PTE, while
1843 * handle_mm_fault() only guarantees to update these in the struct page.
1845 * This is important for some architectures where those bits also gate the
1846 * access permission to the page because they are maintained in software. On
1847 * such architectures, gup() will not be enough to make a subsequent access
1850 * This should be called with the mm_sem held for read.
1852 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1853 unsigned long address, unsigned int fault_flags)
1855 struct vm_area_struct *vma;
1856 vm_flags_t vm_flags;
1859 vma = find_extend_vma(mm, address);
1860 if (!vma || address < vma->vm_start)
1863 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1864 if (!(vm_flags & vma->vm_flags))
1867 ret = handle_mm_fault(mm, vma, address, fault_flags);
1868 if (ret & VM_FAULT_ERROR) {
1869 if (ret & VM_FAULT_OOM)
1871 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1873 if (ret & VM_FAULT_SIGBUS)
1878 if (ret & VM_FAULT_MAJOR)
1887 * get_user_pages() - pin user pages in memory
1888 * @tsk: the task_struct to use for page fault accounting, or
1889 * NULL if faults are not to be recorded.
1890 * @mm: mm_struct of target mm
1891 * @start: starting user address
1892 * @nr_pages: number of pages from start to pin
1893 * @write: whether pages will be written to by the caller
1894 * @force: whether to force write access even if user mapping is
1895 * readonly. This will result in the page being COWed even
1896 * in MAP_SHARED mappings. You do not want this.
1897 * @pages: array that receives pointers to the pages pinned.
1898 * Should be at least nr_pages long. Or NULL, if caller
1899 * only intends to ensure the pages are faulted in.
1900 * @vmas: array of pointers to vmas corresponding to each page.
1901 * Or NULL if the caller does not require them.
1903 * Returns number of pages pinned. This may be fewer than the number
1904 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1905 * were pinned, returns -errno. Each page returned must be released
1906 * with a put_page() call when it is finished with. vmas will only
1907 * remain valid while mmap_sem is held.
1909 * Must be called with mmap_sem held for read or write.
1911 * get_user_pages walks a process's page tables and takes a reference to
1912 * each struct page that each user address corresponds to at a given
1913 * instant. That is, it takes the page that would be accessed if a user
1914 * thread accesses the given user virtual address at that instant.
1916 * This does not guarantee that the page exists in the user mappings when
1917 * get_user_pages returns, and there may even be a completely different
1918 * page there in some cases (eg. if mmapped pagecache has been invalidated
1919 * and subsequently re faulted). However it does guarantee that the page
1920 * won't be freed completely. And mostly callers simply care that the page
1921 * contains data that was valid *at some point in time*. Typically, an IO
1922 * or similar operation cannot guarantee anything stronger anyway because
1923 * locks can't be held over the syscall boundary.
1925 * If write=0, the page must not be written to. If the page is written to,
1926 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1927 * after the page is finished with, and before put_page is called.
1929 * get_user_pages is typically used for fewer-copy IO operations, to get a
1930 * handle on the memory by some means other than accesses via the user virtual
1931 * addresses. The pages may be submitted for DMA to devices or accessed via
1932 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1933 * use the correct cache flushing APIs.
1935 * See also get_user_pages_fast, for performance critical applications.
1937 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1938 unsigned long start, int nr_pages, int write, int force,
1939 struct page **pages, struct vm_area_struct **vmas)
1941 int flags = FOLL_TOUCH;
1946 flags |= FOLL_WRITE;
1948 flags |= FOLL_FORCE;
1950 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1953 EXPORT_SYMBOL(get_user_pages);
1956 * get_dump_page() - pin user page in memory while writing it to core dump
1957 * @addr: user address
1959 * Returns struct page pointer of user page pinned for dump,
1960 * to be freed afterwards by page_cache_release() or put_page().
1962 * Returns NULL on any kind of failure - a hole must then be inserted into
1963 * the corefile, to preserve alignment with its headers; and also returns
1964 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1965 * allowing a hole to be left in the corefile to save diskspace.
1967 * Called without mmap_sem, but after all other threads have been killed.
1969 #ifdef CONFIG_ELF_CORE
1970 struct page *get_dump_page(unsigned long addr)
1972 struct vm_area_struct *vma;
1975 if (__get_user_pages(current, current->mm, addr, 1,
1976 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1979 flush_cache_page(vma, addr, page_to_pfn(page));
1982 #endif /* CONFIG_ELF_CORE */
1984 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1987 pgd_t * pgd = pgd_offset(mm, addr);
1988 pud_t * pud = pud_alloc(mm, pgd, addr);
1990 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1992 VM_BUG_ON(pmd_trans_huge(*pmd));
1993 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2000 * This is the old fallback for page remapping.
2002 * For historical reasons, it only allows reserved pages. Only
2003 * old drivers should use this, and they needed to mark their
2004 * pages reserved for the old functions anyway.
2006 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2007 struct page *page, pgprot_t prot)
2009 struct mm_struct *mm = vma->vm_mm;
2018 flush_dcache_page(page);
2019 pte = get_locked_pte(mm, addr, &ptl);
2023 if (!pte_none(*pte))
2026 /* Ok, finally just insert the thing.. */
2028 inc_mm_counter_fast(mm, MM_FILEPAGES);
2029 page_add_file_rmap(page);
2030 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2033 pte_unmap_unlock(pte, ptl);
2036 pte_unmap_unlock(pte, ptl);
2042 * vm_insert_page - insert single page into user vma
2043 * @vma: user vma to map to
2044 * @addr: target user address of this page
2045 * @page: source kernel page
2047 * This allows drivers to insert individual pages they've allocated
2050 * The page has to be a nice clean _individual_ kernel allocation.
2051 * If you allocate a compound page, you need to have marked it as
2052 * such (__GFP_COMP), or manually just split the page up yourself
2053 * (see split_page()).
2055 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2056 * took an arbitrary page protection parameter. This doesn't allow
2057 * that. Your vma protection will have to be set up correctly, which
2058 * means that if you want a shared writable mapping, you'd better
2059 * ask for a shared writable mapping!
2061 * The page does not need to be reserved.
2063 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2066 if (addr < vma->vm_start || addr >= vma->vm_end)
2068 if (!page_count(page))
2070 vma->vm_flags |= VM_INSERTPAGE;
2071 return insert_page(vma, addr, page, vma->vm_page_prot);
2073 EXPORT_SYMBOL(vm_insert_page);
2075 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2076 unsigned long pfn, pgprot_t prot)
2078 struct mm_struct *mm = vma->vm_mm;
2084 pte = get_locked_pte(mm, addr, &ptl);
2088 if (!pte_none(*pte))
2091 /* Ok, finally just insert the thing.. */
2092 entry = pte_mkspecial(pfn_pte(pfn, prot));
2093 set_pte_at(mm, addr, pte, entry);
2094 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2098 pte_unmap_unlock(pte, ptl);
2104 * vm_insert_pfn - insert single pfn into user vma
2105 * @vma: user vma to map to
2106 * @addr: target user address of this page
2107 * @pfn: source kernel pfn
2109 * Similar to vm_inert_page, this allows drivers to insert individual pages
2110 * they've allocated into a user vma. Same comments apply.
2112 * This function should only be called from a vm_ops->fault handler, and
2113 * in that case the handler should return NULL.
2115 * vma cannot be a COW mapping.
2117 * As this is called only for pages that do not currently exist, we
2118 * do not need to flush old virtual caches or the TLB.
2120 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2124 pgprot_t pgprot = vma->vm_page_prot;
2126 * Technically, architectures with pte_special can avoid all these
2127 * restrictions (same for remap_pfn_range). However we would like
2128 * consistency in testing and feature parity among all, so we should
2129 * try to keep these invariants in place for everybody.
2131 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2132 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2133 (VM_PFNMAP|VM_MIXEDMAP));
2134 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2135 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2137 if (addr < vma->vm_start || addr >= vma->vm_end)
2139 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2142 ret = insert_pfn(vma, addr, pfn, pgprot);
2145 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2149 EXPORT_SYMBOL(vm_insert_pfn);
2151 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2154 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2156 if (addr < vma->vm_start || addr >= vma->vm_end)
2160 * If we don't have pte special, then we have to use the pfn_valid()
2161 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2162 * refcount the page if pfn_valid is true (hence insert_page rather
2163 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2164 * without pte special, it would there be refcounted as a normal page.
2166 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2169 page = pfn_to_page(pfn);
2170 return insert_page(vma, addr, page, vma->vm_page_prot);
2172 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2174 EXPORT_SYMBOL(vm_insert_mixed);
2177 * maps a range of physical memory into the requested pages. the old
2178 * mappings are removed. any references to nonexistent pages results
2179 * in null mappings (currently treated as "copy-on-access")
2181 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2182 unsigned long addr, unsigned long end,
2183 unsigned long pfn, pgprot_t prot)
2188 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2191 arch_enter_lazy_mmu_mode();
2193 BUG_ON(!pte_none(*pte));
2194 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2196 } while (pte++, addr += PAGE_SIZE, addr != end);
2197 arch_leave_lazy_mmu_mode();
2198 pte_unmap_unlock(pte - 1, ptl);
2202 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2203 unsigned long addr, unsigned long end,
2204 unsigned long pfn, pgprot_t prot)
2209 pfn -= addr >> PAGE_SHIFT;
2210 pmd = pmd_alloc(mm, pud, addr);
2213 VM_BUG_ON(pmd_trans_huge(*pmd));
2215 next = pmd_addr_end(addr, end);
2216 if (remap_pte_range(mm, pmd, addr, next,
2217 pfn + (addr >> PAGE_SHIFT), prot))
2219 } while (pmd++, addr = next, addr != end);
2223 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2224 unsigned long addr, unsigned long end,
2225 unsigned long pfn, pgprot_t prot)
2230 pfn -= addr >> PAGE_SHIFT;
2231 pud = pud_alloc(mm, pgd, addr);
2235 next = pud_addr_end(addr, end);
2236 if (remap_pmd_range(mm, pud, addr, next,
2237 pfn + (addr >> PAGE_SHIFT), prot))
2239 } while (pud++, addr = next, addr != end);
2244 * remap_pfn_range - remap kernel memory to userspace
2245 * @vma: user vma to map to
2246 * @addr: target user address to start at
2247 * @pfn: physical address of kernel memory
2248 * @size: size of map area
2249 * @prot: page protection flags for this mapping
2251 * Note: this is only safe if the mm semaphore is held when called.
2253 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2254 unsigned long pfn, unsigned long size, pgprot_t prot)
2258 unsigned long end = addr + PAGE_ALIGN(size);
2259 struct mm_struct *mm = vma->vm_mm;
2263 * Physically remapped pages are special. Tell the
2264 * rest of the world about it:
2265 * VM_IO tells people not to look at these pages
2266 * (accesses can have side effects).
2267 * VM_RESERVED is specified all over the place, because
2268 * in 2.4 it kept swapout's vma scan off this vma; but
2269 * in 2.6 the LRU scan won't even find its pages, so this
2270 * flag means no more than count its pages in reserved_vm,
2271 * and omit it from core dump, even when VM_IO turned off.
2272 * VM_PFNMAP tells the core MM that the base pages are just
2273 * raw PFN mappings, and do not have a "struct page" associated
2276 * There's a horrible special case to handle copy-on-write
2277 * behaviour that some programs depend on. We mark the "original"
2278 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2280 if (addr == vma->vm_start && end == vma->vm_end) {
2281 vma->vm_pgoff = pfn;
2282 vma->vm_flags |= VM_PFN_AT_MMAP;
2283 } else if (is_cow_mapping(vma->vm_flags))
2286 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2288 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2291 * To indicate that track_pfn related cleanup is not
2292 * needed from higher level routine calling unmap_vmas
2294 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2295 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2299 BUG_ON(addr >= end);
2300 pfn -= addr >> PAGE_SHIFT;
2301 pgd = pgd_offset(mm, addr);
2302 flush_cache_range(vma, addr, end);
2304 next = pgd_addr_end(addr, end);
2305 err = remap_pud_range(mm, pgd, addr, next,
2306 pfn + (addr >> PAGE_SHIFT), prot);
2309 } while (pgd++, addr = next, addr != end);
2312 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2316 EXPORT_SYMBOL(remap_pfn_range);
2319 * vm_iomap_memory - remap memory to userspace
2320 * @vma: user vma to map to
2321 * @start: start of area
2322 * @len: size of area
2324 * This is a simplified io_remap_pfn_range() for common driver use. The
2325 * driver just needs to give us the physical memory range to be mapped,
2326 * we'll figure out the rest from the vma information.
2328 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2329 * whatever write-combining details or similar.
2331 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2333 unsigned long vm_len, pfn, pages;
2335 /* Check that the physical memory area passed in looks valid */
2336 if (start + len < start)
2339 * You *really* shouldn't map things that aren't page-aligned,
2340 * but we've historically allowed it because IO memory might
2341 * just have smaller alignment.
2343 len += start & ~PAGE_MASK;
2344 pfn = start >> PAGE_SHIFT;
2345 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2346 if (pfn + pages < pfn)
2349 /* We start the mapping 'vm_pgoff' pages into the area */
2350 if (vma->vm_pgoff > pages)
2352 pfn += vma->vm_pgoff;
2353 pages -= vma->vm_pgoff;
2355 /* Can we fit all of the mapping? */
2356 vm_len = vma->vm_end - vma->vm_start;
2357 if (vm_len >> PAGE_SHIFT > pages)
2360 /* Ok, let it rip */
2361 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2363 EXPORT_SYMBOL(vm_iomap_memory);
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;
2453 BUG_ON(addr >= end);
2454 pgd = pgd_offset(mm, addr);
2456 next = pgd_addr_end(addr, end);
2457 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2460 } while (pgd++, addr = next, addr != end);
2464 EXPORT_SYMBOL_GPL(apply_to_page_range);
2467 * handle_pte_fault chooses page fault handler according to an entry
2468 * which was read non-atomically. Before making any commitment, on
2469 * those architectures or configurations (e.g. i386 with PAE) which
2470 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2471 * must check under lock before unmapping the pte and proceeding
2472 * (but do_wp_page is only called after already making such a check;
2473 * and do_anonymous_page can safely check later on).
2475 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2476 pte_t *page_table, pte_t orig_pte)
2479 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2480 if (sizeof(pte_t) > sizeof(unsigned long)) {
2481 spinlock_t *ptl = pte_lockptr(mm, pmd);
2483 same = pte_same(*page_table, orig_pte);
2487 pte_unmap(page_table);
2491 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2494 * If the source page was a PFN mapping, we don't have
2495 * a "struct page" for it. We do a best-effort copy by
2496 * just copying from the original user address. If that
2497 * fails, we just zero-fill it. Live with it.
2499 if (unlikely(!src)) {
2500 void *kaddr = kmap_atomic(dst, KM_USER0);
2501 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2504 * This really shouldn't fail, because the page is there
2505 * in the page tables. But it might just be unreadable,
2506 * in which case we just give up and fill the result with
2509 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2511 kunmap_atomic(kaddr, KM_USER0);
2512 flush_dcache_page(dst);
2514 copy_user_highpage(dst, src, va, vma);
2518 * This routine handles present pages, when users try to write
2519 * to a shared page. It is done by copying the page to a new address
2520 * and decrementing the shared-page counter for the old page.
2522 * Note that this routine assumes that the protection checks have been
2523 * done by the caller (the low-level page fault routine in most cases).
2524 * Thus we can safely just mark it writable once we've done any necessary
2527 * We also mark the page dirty at this point even though the page will
2528 * change only once the write actually happens. This avoids a few races,
2529 * and potentially makes it more efficient.
2531 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2532 * but allow concurrent faults), with pte both mapped and locked.
2533 * We return with mmap_sem still held, but pte unmapped and unlocked.
2535 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2536 unsigned long address, pte_t *page_table, pmd_t *pmd,
2537 spinlock_t *ptl, pte_t orig_pte)
2540 struct page *old_page, *new_page;
2543 int page_mkwrite = 0;
2544 struct page *dirty_page = NULL;
2546 old_page = vm_normal_page(vma, address, orig_pte);
2549 * VM_MIXEDMAP !pfn_valid() case
2551 * We should not cow pages in a shared writeable mapping.
2552 * Just mark the pages writable as we can't do any dirty
2553 * accounting on raw pfn maps.
2555 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2556 (VM_WRITE|VM_SHARED))
2562 * Take out anonymous pages first, anonymous shared vmas are
2563 * not dirty accountable.
2565 if (PageAnon(old_page) && !PageKsm(old_page)) {
2566 if (!trylock_page(old_page)) {
2567 page_cache_get(old_page);
2568 pte_unmap_unlock(page_table, ptl);
2569 lock_page(old_page);
2570 page_table = pte_offset_map_lock(mm, pmd, address,
2572 if (!pte_same(*page_table, orig_pte)) {
2573 unlock_page(old_page);
2576 page_cache_release(old_page);
2578 if (reuse_swap_page(old_page)) {
2580 * The page is all ours. Move it to our anon_vma so
2581 * the rmap code will not search our parent or siblings.
2582 * Protected against the rmap code by the page lock.
2584 page_move_anon_rmap(old_page, vma, address);
2585 unlock_page(old_page);
2588 unlock_page(old_page);
2589 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2590 (VM_WRITE|VM_SHARED))) {
2592 * Only catch write-faults on shared writable pages,
2593 * read-only shared pages can get COWed by
2594 * get_user_pages(.write=1, .force=1).
2596 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2597 struct vm_fault vmf;
2600 vmf.virtual_address = (void __user *)(address &
2602 vmf.pgoff = old_page->index;
2603 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2604 vmf.page = old_page;
2607 * Notify the address space that the page is about to
2608 * become writable so that it can prohibit this or wait
2609 * for the page to get into an appropriate state.
2611 * We do this without the lock held, so that it can
2612 * sleep if it needs to.
2614 page_cache_get(old_page);
2615 pte_unmap_unlock(page_table, ptl);
2617 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2619 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2621 goto unwritable_page;
2623 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2624 lock_page(old_page);
2625 if (!old_page->mapping) {
2626 ret = 0; /* retry the fault */
2627 unlock_page(old_page);
2628 goto unwritable_page;
2631 VM_BUG_ON(!PageLocked(old_page));
2634 * Since we dropped the lock we need to revalidate
2635 * the PTE as someone else may have changed it. If
2636 * they did, we just return, as we can count on the
2637 * MMU to tell us if they didn't also make it writable.
2639 page_table = pte_offset_map_lock(mm, pmd, address,
2641 if (!pte_same(*page_table, orig_pte)) {
2642 unlock_page(old_page);
2648 dirty_page = old_page;
2649 get_page(dirty_page);
2652 flush_cache_page(vma, address, pte_pfn(orig_pte));
2653 entry = pte_mkyoung(orig_pte);
2654 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2655 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2656 update_mmu_cache(vma, address, page_table);
2657 pte_unmap_unlock(page_table, ptl);
2658 ret |= VM_FAULT_WRITE;
2664 * Yes, Virginia, this is actually required to prevent a race
2665 * with clear_page_dirty_for_io() from clearing the page dirty
2666 * bit after it clear all dirty ptes, but before a racing
2667 * do_wp_page installs a dirty pte.
2669 * __do_fault is protected similarly.
2671 if (!page_mkwrite) {
2672 wait_on_page_locked(dirty_page);
2673 set_page_dirty_balance(dirty_page, page_mkwrite);
2675 put_page(dirty_page);
2677 struct address_space *mapping = dirty_page->mapping;
2679 set_page_dirty(dirty_page);
2680 unlock_page(dirty_page);
2681 page_cache_release(dirty_page);
2684 * Some device drivers do not set page.mapping
2685 * but still dirty their pages
2687 balance_dirty_pages_ratelimited(mapping);
2691 /* file_update_time outside page_lock */
2693 file_update_time(vma->vm_file);
2699 * Ok, we need to copy. Oh, well..
2701 page_cache_get(old_page);
2703 pte_unmap_unlock(page_table, ptl);
2705 if (unlikely(anon_vma_prepare(vma)))
2708 if (is_zero_pfn(pte_pfn(orig_pte))) {
2709 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2713 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2716 cow_user_page(new_page, old_page, address, vma);
2718 __SetPageUptodate(new_page);
2720 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2724 * Re-check the pte - we dropped the lock
2726 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2727 if (likely(pte_same(*page_table, orig_pte))) {
2729 if (!PageAnon(old_page)) {
2730 dec_mm_counter_fast(mm, MM_FILEPAGES);
2731 inc_mm_counter_fast(mm, MM_ANONPAGES);
2734 inc_mm_counter_fast(mm, MM_ANONPAGES);
2735 flush_cache_page(vma, address, pte_pfn(orig_pte));
2736 entry = mk_pte(new_page, vma->vm_page_prot);
2737 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2739 * Clear the pte entry and flush it first, before updating the
2740 * pte with the new entry. This will avoid a race condition
2741 * seen in the presence of one thread doing SMC and another
2744 ptep_clear_flush(vma, address, page_table);
2745 page_add_new_anon_rmap(new_page, vma, address);
2747 * We call the notify macro here because, when using secondary
2748 * mmu page tables (such as kvm shadow page tables), we want the
2749 * new page to be mapped directly into the secondary page table.
2751 set_pte_at_notify(mm, address, page_table, entry);
2752 update_mmu_cache(vma, address, page_table);
2755 * Only after switching the pte to the new page may
2756 * we remove the mapcount here. Otherwise another
2757 * process may come and find the rmap count decremented
2758 * before the pte is switched to the new page, and
2759 * "reuse" the old page writing into it while our pte
2760 * here still points into it and can be read by other
2763 * The critical issue is to order this
2764 * page_remove_rmap with the ptp_clear_flush above.
2765 * Those stores are ordered by (if nothing else,)
2766 * the barrier present in the atomic_add_negative
2767 * in page_remove_rmap.
2769 * Then the TLB flush in ptep_clear_flush ensures that
2770 * no process can access the old page before the
2771 * decremented mapcount is visible. And the old page
2772 * cannot be reused until after the decremented
2773 * mapcount is visible. So transitively, TLBs to
2774 * old page will be flushed before it can be reused.
2776 page_remove_rmap(old_page);
2779 /* Free the old page.. */
2780 new_page = old_page;
2781 ret |= VM_FAULT_WRITE;
2783 mem_cgroup_uncharge_page(new_page);
2786 page_cache_release(new_page);
2788 pte_unmap_unlock(page_table, ptl);
2791 * Don't let another task, with possibly unlocked vma,
2792 * keep the mlocked page.
2794 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2795 lock_page(old_page); /* LRU manipulation */
2796 munlock_vma_page(old_page);
2797 unlock_page(old_page);
2799 page_cache_release(old_page);
2803 page_cache_release(new_page);
2807 unlock_page(old_page);
2808 page_cache_release(old_page);
2810 page_cache_release(old_page);
2812 return VM_FAULT_OOM;
2815 page_cache_release(old_page);
2819 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2820 unsigned long start_addr, unsigned long end_addr,
2821 struct zap_details *details)
2823 zap_page_range(vma, start_addr, end_addr - start_addr, details);
2826 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2827 struct zap_details *details)
2829 struct vm_area_struct *vma;
2830 struct prio_tree_iter iter;
2831 pgoff_t vba, vea, zba, zea;
2833 vma_prio_tree_foreach(vma, &iter, root,
2834 details->first_index, details->last_index) {
2836 vba = vma->vm_pgoff;
2837 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2838 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2839 zba = details->first_index;
2842 zea = details->last_index;
2846 unmap_mapping_range_vma(vma,
2847 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2848 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2853 static inline void unmap_mapping_range_list(struct list_head *head,
2854 struct zap_details *details)
2856 struct vm_area_struct *vma;
2859 * In nonlinear VMAs there is no correspondence between virtual address
2860 * offset and file offset. So we must perform an exhaustive search
2861 * across *all* the pages in each nonlinear VMA, not just the pages
2862 * whose virtual address lies outside the file truncation point.
2864 list_for_each_entry(vma, head, shared.vm_set.list) {
2865 details->nonlinear_vma = vma;
2866 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2871 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2872 * @mapping: the address space containing mmaps to be unmapped.
2873 * @holebegin: byte in first page to unmap, relative to the start of
2874 * the underlying file. This will be rounded down to a PAGE_SIZE
2875 * boundary. Note that this is different from truncate_pagecache(), which
2876 * must keep the partial page. In contrast, we must get rid of
2878 * @holelen: size of prospective hole in bytes. This will be rounded
2879 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2881 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2882 * but 0 when invalidating pagecache, don't throw away private data.
2884 void unmap_mapping_range(struct address_space *mapping,
2885 loff_t const holebegin, loff_t const holelen, int even_cows)
2887 struct zap_details details;
2888 pgoff_t hba = holebegin >> PAGE_SHIFT;
2889 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2891 /* Check for overflow. */
2892 if (sizeof(holelen) > sizeof(hlen)) {
2894 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2895 if (holeend & ~(long long)ULONG_MAX)
2896 hlen = ULONG_MAX - hba + 1;
2899 details.check_mapping = even_cows? NULL: mapping;
2900 details.nonlinear_vma = NULL;
2901 details.first_index = hba;
2902 details.last_index = hba + hlen - 1;
2903 if (details.last_index < details.first_index)
2904 details.last_index = ULONG_MAX;
2907 mutex_lock(&mapping->i_mmap_mutex);
2908 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2909 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2910 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2911 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2912 mutex_unlock(&mapping->i_mmap_mutex);
2914 EXPORT_SYMBOL(unmap_mapping_range);
2917 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2918 * but allow concurrent faults), and pte mapped but not yet locked.
2919 * We return with mmap_sem still held, but pte unmapped and unlocked.
2921 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2922 unsigned long address, pte_t *page_table, pmd_t *pmd,
2923 unsigned int flags, pte_t orig_pte)
2926 struct page *page, *swapcache = NULL;
2930 struct mem_cgroup *ptr;
2934 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2937 entry = pte_to_swp_entry(orig_pte);
2938 if (unlikely(non_swap_entry(entry))) {
2939 if (is_migration_entry(entry)) {
2940 migration_entry_wait(mm, pmd, address);
2941 } else if (is_hwpoison_entry(entry)) {
2942 ret = VM_FAULT_HWPOISON;
2944 print_bad_pte(vma, address, orig_pte, NULL);
2945 ret = VM_FAULT_SIGBUS;
2949 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2950 page = lookup_swap_cache(entry);
2952 grab_swap_token(mm); /* Contend for token _before_ read-in */
2953 page = swapin_readahead(entry,
2954 GFP_HIGHUSER_MOVABLE, vma, address);
2957 * Back out if somebody else faulted in this pte
2958 * while we released the pte lock.
2960 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2961 if (likely(pte_same(*page_table, orig_pte)))
2963 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2967 /* Had to read the page from swap area: Major fault */
2968 ret = VM_FAULT_MAJOR;
2969 count_vm_event(PGMAJFAULT);
2970 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2971 } else if (PageHWPoison(page)) {
2973 * hwpoisoned dirty swapcache pages are kept for killing
2974 * owner processes (which may be unknown at hwpoison time)
2976 ret = VM_FAULT_HWPOISON;
2977 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2981 locked = lock_page_or_retry(page, mm, flags);
2982 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2984 ret |= VM_FAULT_RETRY;
2989 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2990 * release the swapcache from under us. The page pin, and pte_same
2991 * test below, are not enough to exclude that. Even if it is still
2992 * swapcache, we need to check that the page's swap has not changed.
2994 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2997 if (ksm_might_need_to_copy(page, vma, address)) {
2999 page = ksm_does_need_to_copy(page, vma, address);
3001 if (unlikely(!page)) {
3009 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3015 * Back out if somebody else already faulted in this pte.
3017 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3018 if (unlikely(!pte_same(*page_table, orig_pte)))
3021 if (unlikely(!PageUptodate(page))) {
3022 ret = VM_FAULT_SIGBUS;
3027 * The page isn't present yet, go ahead with the fault.
3029 * Be careful about the sequence of operations here.
3030 * To get its accounting right, reuse_swap_page() must be called
3031 * while the page is counted on swap but not yet in mapcount i.e.
3032 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3033 * must be called after the swap_free(), or it will never succeed.
3034 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3035 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3036 * in page->private. In this case, a record in swap_cgroup is silently
3037 * discarded at swap_free().
3040 inc_mm_counter_fast(mm, MM_ANONPAGES);
3041 dec_mm_counter_fast(mm, MM_SWAPENTS);
3042 pte = mk_pte(page, vma->vm_page_prot);
3043 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3044 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3045 flags &= ~FAULT_FLAG_WRITE;
3046 ret |= VM_FAULT_WRITE;
3049 flush_icache_page(vma, page);
3050 set_pte_at(mm, address, page_table, pte);
3051 do_page_add_anon_rmap(page, vma, address, exclusive);
3052 /* It's better to call commit-charge after rmap is established */
3053 mem_cgroup_commit_charge_swapin(page, ptr);
3056 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3057 try_to_free_swap(page);
3061 * Hold the lock to avoid the swap entry to be reused
3062 * until we take the PT lock for the pte_same() check
3063 * (to avoid false positives from pte_same). For
3064 * further safety release the lock after the swap_free
3065 * so that the swap count won't change under a
3066 * parallel locked swapcache.
3068 unlock_page(swapcache);
3069 page_cache_release(swapcache);
3072 if (flags & FAULT_FLAG_WRITE) {
3073 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3074 if (ret & VM_FAULT_ERROR)
3075 ret &= VM_FAULT_ERROR;
3079 /* No need to invalidate - it was non-present before */
3080 update_mmu_cache(vma, address, page_table);
3082 pte_unmap_unlock(page_table, ptl);
3086 mem_cgroup_cancel_charge_swapin(ptr);
3087 pte_unmap_unlock(page_table, ptl);
3091 page_cache_release(page);
3093 unlock_page(swapcache);
3094 page_cache_release(swapcache);
3100 * This is like a special single-page "expand_{down|up}wards()",
3101 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3102 * doesn't hit another vma.
3104 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3106 address &= PAGE_MASK;
3107 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3108 struct vm_area_struct *prev = vma->vm_prev;
3111 * Is there a mapping abutting this one below?
3113 * That's only ok if it's the same stack mapping
3114 * that has gotten split..
3116 if (prev && prev->vm_end == address)
3117 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3119 expand_downwards(vma, address - PAGE_SIZE);
3121 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3122 struct vm_area_struct *next = vma->vm_next;
3124 /* As VM_GROWSDOWN but s/below/above/ */
3125 if (next && next->vm_start == address + PAGE_SIZE)
3126 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3128 expand_upwards(vma, address + PAGE_SIZE);
3134 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3135 * but allow concurrent faults), and pte mapped but not yet locked.
3136 * We return with mmap_sem still held, but pte unmapped and unlocked.
3138 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3139 unsigned long address, pte_t *page_table, pmd_t *pmd,
3146 pte_unmap(page_table);
3148 /* Check if we need to add a guard page to the stack */
3149 if (check_stack_guard_page(vma, address) < 0)
3150 return VM_FAULT_SIGBUS;
3152 /* Use the zero-page for reads */
3153 if (!(flags & FAULT_FLAG_WRITE)) {
3154 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3155 vma->vm_page_prot));
3156 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3157 if (!pte_none(*page_table))
3162 /* Allocate our own private page. */
3163 if (unlikely(anon_vma_prepare(vma)))
3165 page = alloc_zeroed_user_highpage_movable(vma, address);
3168 __SetPageUptodate(page);
3170 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3173 entry = mk_pte(page, vma->vm_page_prot);
3174 if (vma->vm_flags & VM_WRITE)
3175 entry = pte_mkwrite(pte_mkdirty(entry));
3177 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3178 if (!pte_none(*page_table))
3181 inc_mm_counter_fast(mm, MM_ANONPAGES);
3182 page_add_new_anon_rmap(page, vma, address);
3184 set_pte_at(mm, address, page_table, entry);
3186 /* No need to invalidate - it was non-present before */
3187 update_mmu_cache(vma, address, page_table);
3189 pte_unmap_unlock(page_table, ptl);
3192 mem_cgroup_uncharge_page(page);
3193 page_cache_release(page);
3196 page_cache_release(page);
3198 return VM_FAULT_OOM;
3202 * __do_fault() tries to create a new page mapping. It aggressively
3203 * tries to share with existing pages, but makes a separate copy if
3204 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3205 * the next page fault.
3207 * As this is called only for pages that do not currently exist, we
3208 * do not need to flush old virtual caches or the TLB.
3210 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3211 * but allow concurrent faults), and pte neither mapped nor locked.
3212 * We return with mmap_sem still held, but pte unmapped and unlocked.
3214 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3215 unsigned long address, pmd_t *pmd,
3216 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3221 struct page *cow_page;
3224 struct page *dirty_page = NULL;
3225 struct vm_fault vmf;
3227 int page_mkwrite = 0;
3230 * If we do COW later, allocate page befor taking lock_page()
3231 * on the file cache page. This will reduce lock holding time.
3233 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3235 if (unlikely(anon_vma_prepare(vma)))
3236 return VM_FAULT_OOM;
3238 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3240 return VM_FAULT_OOM;
3242 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3243 page_cache_release(cow_page);
3244 return VM_FAULT_OOM;
3249 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3254 ret = vma->vm_ops->fault(vma, &vmf);
3255 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3259 if (unlikely(PageHWPoison(vmf.page))) {
3260 if (ret & VM_FAULT_LOCKED)
3261 unlock_page(vmf.page);
3262 ret = VM_FAULT_HWPOISON;
3267 * For consistency in subsequent calls, make the faulted page always
3270 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3271 lock_page(vmf.page);
3273 VM_BUG_ON(!PageLocked(vmf.page));
3276 * Should we do an early C-O-W break?
3279 if (flags & FAULT_FLAG_WRITE) {
3280 if (!(vma->vm_flags & VM_SHARED)) {
3283 copy_user_highpage(page, vmf.page, address, vma);
3284 __SetPageUptodate(page);
3287 * If the page will be shareable, see if the backing
3288 * address space wants to know that the page is about
3289 * to become writable
3291 if (vma->vm_ops->page_mkwrite) {
3295 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3296 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3298 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3300 goto unwritable_page;
3302 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3304 if (!page->mapping) {
3305 ret = 0; /* retry the fault */
3307 goto unwritable_page;
3310 VM_BUG_ON(!PageLocked(page));
3317 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3320 * This silly early PAGE_DIRTY setting removes a race
3321 * due to the bad i386 page protection. But it's valid
3322 * for other architectures too.
3324 * Note that if FAULT_FLAG_WRITE is set, we either now have
3325 * an exclusive copy of the page, or this is a shared mapping,
3326 * so we can make it writable and dirty to avoid having to
3327 * handle that later.
3329 /* Only go through if we didn't race with anybody else... */
3330 if (likely(pte_same(*page_table, orig_pte))) {
3331 flush_icache_page(vma, page);
3332 entry = mk_pte(page, vma->vm_page_prot);
3333 if (flags & FAULT_FLAG_WRITE)
3334 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3336 inc_mm_counter_fast(mm, MM_ANONPAGES);
3337 page_add_new_anon_rmap(page, vma, address);
3339 inc_mm_counter_fast(mm, MM_FILEPAGES);
3340 page_add_file_rmap(page);
3341 if (flags & FAULT_FLAG_WRITE) {
3343 get_page(dirty_page);
3346 set_pte_at(mm, address, page_table, entry);
3348 /* no need to invalidate: a not-present page won't be cached */
3349 update_mmu_cache(vma, address, page_table);
3352 mem_cgroup_uncharge_page(cow_page);
3354 page_cache_release(page);
3356 anon = 1; /* no anon but release faulted_page */
3359 pte_unmap_unlock(page_table, ptl);
3362 struct address_space *mapping = page->mapping;
3364 if (set_page_dirty(dirty_page))
3366 unlock_page(dirty_page);
3367 put_page(dirty_page);
3368 if (page_mkwrite && mapping) {
3370 * Some device drivers do not set page.mapping but still
3373 balance_dirty_pages_ratelimited(mapping);
3376 /* file_update_time outside page_lock */
3378 file_update_time(vma->vm_file);
3380 unlock_page(vmf.page);
3382 page_cache_release(vmf.page);
3388 page_cache_release(page);
3391 /* fs's fault handler get error */
3393 mem_cgroup_uncharge_page(cow_page);
3394 page_cache_release(cow_page);
3399 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3400 unsigned long address, pte_t *page_table, pmd_t *pmd,
3401 unsigned int flags, pte_t orig_pte)
3403 pgoff_t pgoff = (((address & PAGE_MASK)
3404 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3406 pte_unmap(page_table);
3407 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3411 * Fault of a previously existing named mapping. Repopulate the pte
3412 * from the encoded file_pte if possible. This enables swappable
3415 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3416 * but allow concurrent faults), and pte mapped but not yet locked.
3417 * We return with mmap_sem still held, but pte unmapped and unlocked.
3419 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3420 unsigned long address, pte_t *page_table, pmd_t *pmd,
3421 unsigned int flags, pte_t orig_pte)
3425 flags |= FAULT_FLAG_NONLINEAR;
3427 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3430 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3432 * Page table corrupted: show pte and kill process.
3434 print_bad_pte(vma, address, orig_pte, NULL);
3435 return VM_FAULT_SIGBUS;
3438 pgoff = pte_to_pgoff(orig_pte);
3439 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3443 * These routines also need to handle stuff like marking pages dirty
3444 * and/or accessed for architectures that don't do it in hardware (most
3445 * RISC architectures). The early dirtying is also good on the i386.
3447 * There is also a hook called "update_mmu_cache()" that architectures
3448 * with external mmu caches can use to update those (ie the Sparc or
3449 * PowerPC hashed page tables that act as extended TLBs).
3451 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3452 * but allow concurrent faults), and pte mapped but not yet locked.
3453 * We return with mmap_sem still held, but pte unmapped and unlocked.
3455 int handle_pte_fault(struct mm_struct *mm,
3456 struct vm_area_struct *vma, unsigned long address,
3457 pte_t *pte, pmd_t *pmd, unsigned int flags)
3463 if (!pte_present(entry)) {
3464 if (pte_none(entry)) {
3466 if (likely(vma->vm_ops->fault))
3467 return do_linear_fault(mm, vma, address,
3468 pte, pmd, flags, entry);
3470 return do_anonymous_page(mm, vma, address,
3473 if (pte_file(entry))
3474 return do_nonlinear_fault(mm, vma, address,
3475 pte, pmd, flags, entry);
3476 return do_swap_page(mm, vma, address,
3477 pte, pmd, flags, entry);
3480 ptl = pte_lockptr(mm, pmd);
3482 if (unlikely(!pte_same(*pte, entry)))
3484 if (flags & FAULT_FLAG_WRITE) {
3485 if (!pte_write(entry))
3486 return do_wp_page(mm, vma, address,
3487 pte, pmd, ptl, entry);
3488 entry = pte_mkdirty(entry);
3490 entry = pte_mkyoung(entry);
3491 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3492 update_mmu_cache(vma, address, pte);
3495 * This is needed only for protection faults but the arch code
3496 * is not yet telling us if this is a protection fault or not.
3497 * This still avoids useless tlb flushes for .text page faults
3500 if (flags & FAULT_FLAG_WRITE)
3501 flush_tlb_fix_spurious_fault(vma, address);
3504 pte_unmap_unlock(pte, ptl);
3509 * By the time we get here, we already hold the mm semaphore
3511 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3512 unsigned long address, unsigned int flags)
3519 __set_current_state(TASK_RUNNING);
3521 count_vm_event(PGFAULT);
3522 mem_cgroup_count_vm_event(mm, PGFAULT);
3524 /* do counter updates before entering really critical section. */
3525 check_sync_rss_stat(current);
3527 if (unlikely(is_vm_hugetlb_page(vma)))
3528 return hugetlb_fault(mm, vma, address, flags);
3531 pgd = pgd_offset(mm, address);
3532 pud = pud_alloc(mm, pgd, address);
3534 return VM_FAULT_OOM;
3535 pmd = pmd_alloc(mm, pud, address);
3537 return VM_FAULT_OOM;
3538 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3540 return do_huge_pmd_anonymous_page(mm, vma, address,
3543 pmd_t orig_pmd = *pmd;
3547 if (pmd_trans_huge(orig_pmd)) {
3548 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3550 if (dirty && !pmd_write(orig_pmd) &&
3551 !pmd_trans_splitting(orig_pmd)) {
3552 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3555 * If COW results in an oom, the huge pmd will
3556 * have been split, so retry the fault on the
3557 * pte for a smaller charge.
3559 if (unlikely(ret & VM_FAULT_OOM))
3563 huge_pmd_set_accessed(mm, vma, address, pmd,
3571 * Use __pte_alloc instead of pte_alloc_map, because we can't
3572 * run pte_offset_map on the pmd, if an huge pmd could
3573 * materialize from under us from a different thread.
3575 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3576 return VM_FAULT_OOM;
3577 /* if an huge pmd materialized from under us just retry later */
3578 if (unlikely(pmd_trans_huge(*pmd)))
3581 * A regular pmd is established and it can't morph into a huge pmd
3582 * from under us anymore at this point because we hold the mmap_sem
3583 * read mode and khugepaged takes it in write mode. So now it's
3584 * safe to run pte_offset_map().
3586 pte = pte_offset_map(pmd, address);
3588 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3591 #ifndef __PAGETABLE_PUD_FOLDED
3593 * Allocate page upper directory.
3594 * We've already handled the fast-path in-line.
3596 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3598 pud_t *new = pud_alloc_one(mm, address);
3602 smp_wmb(); /* See comment in __pte_alloc */
3604 spin_lock(&mm->page_table_lock);
3605 if (pgd_present(*pgd)) /* Another has populated it */
3608 pgd_populate(mm, pgd, new);
3609 spin_unlock(&mm->page_table_lock);
3612 #endif /* __PAGETABLE_PUD_FOLDED */
3614 #ifndef __PAGETABLE_PMD_FOLDED
3616 * Allocate page middle directory.
3617 * We've already handled the fast-path in-line.
3619 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3621 pmd_t *new = pmd_alloc_one(mm, address);
3625 smp_wmb(); /* See comment in __pte_alloc */
3627 spin_lock(&mm->page_table_lock);
3628 #ifndef __ARCH_HAS_4LEVEL_HACK
3629 if (pud_present(*pud)) /* Another has populated it */
3632 pud_populate(mm, pud, new);
3634 if (pgd_present(*pud)) /* Another has populated it */
3637 pgd_populate(mm, pud, new);
3638 #endif /* __ARCH_HAS_4LEVEL_HACK */
3639 spin_unlock(&mm->page_table_lock);
3642 #endif /* __PAGETABLE_PMD_FOLDED */
3644 int make_pages_present(unsigned long addr, unsigned long end)
3646 int ret, len, write;
3647 struct vm_area_struct * vma;
3649 vma = find_vma(current->mm, addr);
3653 * We want to touch writable mappings with a write fault in order
3654 * to break COW, except for shared mappings because these don't COW
3655 * and we would not want to dirty them for nothing.
3657 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3658 BUG_ON(addr >= end);
3659 BUG_ON(end > vma->vm_end);
3660 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3661 ret = get_user_pages(current, current->mm, addr,
3662 len, write, 0, NULL, NULL);
3665 return ret == len ? 0 : -EFAULT;
3668 #if !defined(__HAVE_ARCH_GATE_AREA)
3670 #if defined(AT_SYSINFO_EHDR)
3671 static struct vm_area_struct gate_vma;
3673 static int __init gate_vma_init(void)
3675 gate_vma.vm_mm = NULL;
3676 gate_vma.vm_start = FIXADDR_USER_START;
3677 gate_vma.vm_end = FIXADDR_USER_END;
3678 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3679 gate_vma.vm_page_prot = __P101;
3681 * Make sure the vDSO gets into every core dump.
3682 * Dumping its contents makes post-mortem fully interpretable later
3683 * without matching up the same kernel and hardware config to see
3684 * what PC values meant.
3686 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3689 __initcall(gate_vma_init);
3692 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3694 #ifdef AT_SYSINFO_EHDR
3701 int in_gate_area_no_mm(unsigned long addr)
3703 #ifdef AT_SYSINFO_EHDR
3704 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3710 #endif /* __HAVE_ARCH_GATE_AREA */
3712 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3713 pte_t **ptepp, spinlock_t **ptlp)
3720 pgd = pgd_offset(mm, address);
3721 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3724 pud = pud_offset(pgd, address);
3725 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3728 pmd = pmd_offset(pud, address);
3729 VM_BUG_ON(pmd_trans_huge(*pmd));
3730 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3733 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3737 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3740 if (!pte_present(*ptep))
3745 pte_unmap_unlock(ptep, *ptlp);
3750 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3751 pte_t **ptepp, spinlock_t **ptlp)
3755 /* (void) is needed to make gcc happy */
3756 (void) __cond_lock(*ptlp,
3757 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3762 * follow_pfn - look up PFN at a user virtual address
3763 * @vma: memory mapping
3764 * @address: user virtual address
3765 * @pfn: location to store found PFN
3767 * Only IO mappings and raw PFN mappings are allowed.
3769 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3771 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3778 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3781 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3784 *pfn = pte_pfn(*ptep);
3785 pte_unmap_unlock(ptep, ptl);
3788 EXPORT_SYMBOL(follow_pfn);
3790 #ifdef CONFIG_HAVE_IOREMAP_PROT
3791 int follow_phys(struct vm_area_struct *vma,
3792 unsigned long address, unsigned int flags,
3793 unsigned long *prot, resource_size_t *phys)
3799 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3802 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3806 if ((flags & FOLL_WRITE) && !pte_write(pte))
3809 *prot = pgprot_val(pte_pgprot(pte));
3810 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3814 pte_unmap_unlock(ptep, ptl);
3819 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3820 void *buf, int len, int write)
3822 resource_size_t phys_addr;
3823 unsigned long prot = 0;
3824 void __iomem *maddr;
3825 int offset = addr & (PAGE_SIZE-1);
3827 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3830 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3832 memcpy_toio(maddr + offset, buf, len);
3834 memcpy_fromio(buf, maddr + offset, len);
3842 * Access another process' address space as given in mm. If non-NULL, use the
3843 * given task for page fault accounting.
3845 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3846 unsigned long addr, void *buf, int len, int write)
3848 struct vm_area_struct *vma;
3849 void *old_buf = buf;
3851 down_read(&mm->mmap_sem);
3852 /* ignore errors, just check how much was successfully transferred */
3854 int bytes, ret, offset;
3856 struct page *page = NULL;
3858 ret = get_user_pages(tsk, mm, addr, 1,
3859 write, 1, &page, &vma);
3862 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3863 * we can access using slightly different code.
3865 #ifdef CONFIG_HAVE_IOREMAP_PROT
3866 vma = find_vma(mm, addr);
3867 if (!vma || vma->vm_start > addr)
3869 if (vma->vm_ops && vma->vm_ops->access)
3870 ret = vma->vm_ops->access(vma, addr, buf,
3878 offset = addr & (PAGE_SIZE-1);
3879 if (bytes > PAGE_SIZE-offset)
3880 bytes = PAGE_SIZE-offset;
3884 copy_to_user_page(vma, page, addr,
3885 maddr + offset, buf, bytes);
3886 set_page_dirty_lock(page);
3888 copy_from_user_page(vma, page, addr,
3889 buf, maddr + offset, bytes);
3892 page_cache_release(page);
3898 up_read(&mm->mmap_sem);
3900 return buf - old_buf;
3904 * access_remote_vm - access another process' address space
3905 * @mm: the mm_struct of the target address space
3906 * @addr: start address to access
3907 * @buf: source or destination buffer
3908 * @len: number of bytes to transfer
3909 * @write: whether the access is a write
3911 * The caller must hold a reference on @mm.
3913 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3914 void *buf, int len, int write)
3916 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3920 * Access another process' address space.
3921 * Source/target buffer must be kernel space,
3922 * Do not walk the page table directly, use get_user_pages
3924 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3925 void *buf, int len, int write)
3927 struct mm_struct *mm;
3930 mm = get_task_mm(tsk);
3934 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3941 * Print the name of a VMA.
3943 void print_vma_addr(char *prefix, unsigned long ip)
3945 struct mm_struct *mm = current->mm;
3946 struct vm_area_struct *vma;
3949 * Do not print if we are in atomic
3950 * contexts (in exception stacks, etc.):
3952 if (preempt_count())
3955 down_read(&mm->mmap_sem);
3956 vma = find_vma(mm, ip);
3957 if (vma && vma->vm_file) {
3958 struct file *f = vma->vm_file;
3959 char *buf = (char *)__get_free_page(GFP_KERNEL);
3963 p = d_path(&f->f_path, buf, PAGE_SIZE);
3966 s = strrchr(p, '/');
3969 printk("%s%s[%lx+%lx]", prefix, p,
3971 vma->vm_end - vma->vm_start);
3972 free_page((unsigned long)buf);
3975 up_read(¤t->mm->mmap_sem);
3978 #ifdef CONFIG_PROVE_LOCKING
3979 void might_fault(void)
3982 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3983 * holding the mmap_sem, this is safe because kernel memory doesn't
3984 * get paged out, therefore we'll never actually fault, and the
3985 * below annotations will generate false positives.
3987 if (segment_eq(get_fs(), KERNEL_DS))
3992 * it would be nicer only to annotate paths which are not under
3993 * pagefault_disable, however that requires a larger audit and
3994 * providing helpers like get_user_atomic.
3996 if (!in_atomic() && current->mm)
3997 might_lock_read(¤t->mm->mmap_sem);
3999 EXPORT_SYMBOL(might_fault);
4002 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4003 static void clear_gigantic_page(struct page *page,
4005 unsigned int pages_per_huge_page)
4008 struct page *p = page;
4011 for (i = 0; i < pages_per_huge_page;
4012 i++, p = mem_map_next(p, page, i)) {
4014 clear_user_highpage(p, addr + i * PAGE_SIZE);
4017 void clear_huge_page(struct page *page,
4018 unsigned long addr, unsigned int pages_per_huge_page)
4022 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4023 clear_gigantic_page(page, addr, pages_per_huge_page);
4028 for (i = 0; i < pages_per_huge_page; i++) {
4030 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4034 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4036 struct vm_area_struct *vma,
4037 unsigned int pages_per_huge_page)
4040 struct page *dst_base = dst;
4041 struct page *src_base = src;
4043 for (i = 0; i < pages_per_huge_page; ) {
4045 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4048 dst = mem_map_next(dst, dst_base, i);
4049 src = mem_map_next(src, src_base, i);
4053 void copy_user_huge_page(struct page *dst, struct page *src,
4054 unsigned long addr, struct vm_area_struct *vma,
4055 unsigned int pages_per_huge_page)
4059 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4060 copy_user_gigantic_page(dst, src, addr, vma,
4061 pages_per_huge_page);
4066 for (i = 0; i < pages_per_huge_page; i++) {
4068 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4071 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */