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)
40 * 2012 - NUMA placement page faults (Andrea Arcangeli, Peter Zijlstra)
43 #include <linux/kernel_stat.h>
45 #include <linux/hugetlb.h>
46 #include <linux/mman.h>
47 #include <linux/swap.h>
48 #include <linux/highmem.h>
49 #include <linux/pagemap.h>
50 #include <linux/ksm.h>
51 #include <linux/rmap.h>
52 #include <linux/export.h>
53 #include <linux/delayacct.h>
54 #include <linux/init.h>
55 #include <linux/writeback.h>
56 #include <linux/memcontrol.h>
57 #include <linux/mmu_notifier.h>
58 #include <linux/kallsyms.h>
59 #include <linux/swapops.h>
60 #include <linux/elf.h>
61 #include <linux/gfp.h>
62 #include <linux/migrate.h>
63 #include <linux/string.h>
66 #include <asm/pgalloc.h>
67 #include <asm/uaccess.h>
69 #include <asm/tlbflush.h>
70 #include <asm/pgtable.h>
74 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
75 #warning Unfortunate NUMA config, growing page-frame for last_nid.
78 #ifndef CONFIG_NEED_MULTIPLE_NODES
79 /* use the per-pgdat data instead for discontigmem - mbligh */
80 unsigned long max_mapnr;
83 EXPORT_SYMBOL(max_mapnr);
84 EXPORT_SYMBOL(mem_map);
87 unsigned long num_physpages;
89 * A number of key systems in x86 including ioremap() rely on the assumption
90 * that high_memory defines the upper bound on direct map memory, then end
91 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
92 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
97 EXPORT_SYMBOL(num_physpages);
98 EXPORT_SYMBOL(high_memory);
101 * Randomize the address space (stacks, mmaps, brk, etc.).
103 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
104 * as ancient (libc5 based) binaries can segfault. )
106 int randomize_va_space __read_mostly =
107 #ifdef CONFIG_COMPAT_BRK
113 static int __init disable_randmaps(char *s)
115 randomize_va_space = 0;
118 __setup("norandmaps", disable_randmaps);
120 unsigned long zero_pfn __read_mostly;
121 unsigned long highest_memmap_pfn __read_mostly;
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
126 static int __init init_zero_pfn(void)
128 zero_pfn = page_to_pfn(ZERO_PAGE(0));
131 core_initcall(init_zero_pfn);
134 #if defined(SPLIT_RSS_COUNTING)
136 void sync_mm_rss(struct mm_struct *mm)
140 for (i = 0; i < NR_MM_COUNTERS; i++) {
141 if (current->rss_stat.count[i]) {
142 add_mm_counter(mm, i, current->rss_stat.count[i]);
143 current->rss_stat.count[i] = 0;
146 current->rss_stat.events = 0;
149 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
151 struct task_struct *task = current;
153 if (likely(task->mm == mm))
154 task->rss_stat.count[member] += val;
156 add_mm_counter(mm, member, val);
158 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
161 /* sync counter once per 64 page faults */
162 #define TASK_RSS_EVENTS_THRESH (64)
163 static void check_sync_rss_stat(struct task_struct *task)
165 if (unlikely(task != current))
167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
168 sync_mm_rss(task->mm);
170 #else /* SPLIT_RSS_COUNTING */
172 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
175 static void check_sync_rss_stat(struct task_struct *task)
179 #endif /* SPLIT_RSS_COUNTING */
181 #ifdef HAVE_GENERIC_MMU_GATHER
183 static int tlb_next_batch(struct mmu_gather *tlb)
185 struct mmu_gather_batch *batch;
189 tlb->active = batch->next;
193 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
199 batch->max = MAX_GATHER_BATCH;
201 tlb->active->next = batch;
208 * Called to initialize an (on-stack) mmu_gather structure for page-table
209 * tear-down from @mm. The @fullmm argument is used when @mm is without
210 * users and we're going to destroy the full address space (exit/execve).
212 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
216 tlb->fullmm = fullmm;
220 tlb->fast_mode = (num_possible_cpus() == 1);
221 tlb->local.next = NULL;
223 tlb->local.max = ARRAY_SIZE(tlb->__pages);
224 tlb->active = &tlb->local;
226 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 void tlb_flush_mmu(struct mmu_gather *tlb)
233 struct mmu_gather_batch *batch;
235 if (!tlb->need_flush)
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
243 if (tlb_fast_mode(tlb))
246 for (batch = &tlb->local; batch; batch = batch->next) {
247 free_pages_and_swap_cache(batch->pages, batch->nr);
250 tlb->active = &tlb->local;
254 * Called at the end of the shootdown operation to free up any resources
255 * that were required.
257 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
259 struct mmu_gather_batch *batch, *next;
265 /* keep the page table cache within bounds */
268 for (batch = tlb->local.next; batch; batch = next) {
270 free_pages((unsigned long)batch, 0);
272 tlb->local.next = NULL;
276 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
277 * handling the additional races in SMP caused by other CPUs caching valid
278 * mappings in their TLBs. Returns the number of free page slots left.
279 * When out of page slots we must call tlb_flush_mmu().
281 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
283 struct mmu_gather_batch *batch;
285 VM_BUG_ON(!tlb->need_flush);
287 if (tlb_fast_mode(tlb)) {
288 free_page_and_swap_cache(page);
289 return 1; /* avoid calling tlb_flush_mmu() */
293 batch->pages[batch->nr++] = page;
294 if (batch->nr == batch->max) {
295 if (!tlb_next_batch(tlb))
299 VM_BUG_ON(batch->nr > batch->max);
301 return batch->max - batch->nr;
304 #endif /* HAVE_GENERIC_MMU_GATHER */
306 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
309 * See the comment near struct mmu_table_batch.
312 static void tlb_remove_table_smp_sync(void *arg)
314 /* Simply deliver the interrupt */
317 static void tlb_remove_table_one(void *table)
320 * This isn't an RCU grace period and hence the page-tables cannot be
321 * assumed to be actually RCU-freed.
323 * It is however sufficient for software page-table walkers that rely on
324 * IRQ disabling. See the comment near struct mmu_table_batch.
326 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
327 __tlb_remove_table(table);
330 static void tlb_remove_table_rcu(struct rcu_head *head)
332 struct mmu_table_batch *batch;
335 batch = container_of(head, struct mmu_table_batch, rcu);
337 for (i = 0; i < batch->nr; i++)
338 __tlb_remove_table(batch->tables[i]);
340 free_page((unsigned long)batch);
343 void tlb_table_flush(struct mmu_gather *tlb)
345 struct mmu_table_batch **batch = &tlb->batch;
348 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
353 void tlb_remove_table(struct mmu_gather *tlb, void *table)
355 struct mmu_table_batch **batch = &tlb->batch;
360 * When there's less then two users of this mm there cannot be a
361 * concurrent page-table walk.
363 if (atomic_read(&tlb->mm->mm_users) < 2) {
364 __tlb_remove_table(table);
368 if (*batch == NULL) {
369 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
370 if (*batch == NULL) {
371 tlb_remove_table_one(table);
376 (*batch)->tables[(*batch)->nr++] = table;
377 if ((*batch)->nr == MAX_TABLE_BATCH)
378 tlb_table_flush(tlb);
381 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
384 * If a p?d_bad entry is found while walking page tables, report
385 * the error, before resetting entry to p?d_none. Usually (but
386 * very seldom) called out from the p?d_none_or_clear_bad macros.
389 void pgd_clear_bad(pgd_t *pgd)
395 void pud_clear_bad(pud_t *pud)
401 void pmd_clear_bad(pmd_t *pmd)
408 * Note: this doesn't free the actual pages themselves. That
409 * has been handled earlier when unmapping all the memory regions.
411 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
414 pgtable_t token = pmd_pgtable(*pmd);
416 pte_free_tlb(tlb, token, addr);
420 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
421 unsigned long addr, unsigned long end,
422 unsigned long floor, unsigned long ceiling)
429 pmd = pmd_offset(pud, addr);
431 next = pmd_addr_end(addr, end);
432 if (pmd_none_or_clear_bad(pmd))
434 free_pte_range(tlb, pmd, addr);
435 } while (pmd++, addr = next, addr != end);
445 if (end - 1 > ceiling - 1)
448 pmd = pmd_offset(pud, start);
450 pmd_free_tlb(tlb, pmd, start);
453 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
454 unsigned long addr, unsigned long end,
455 unsigned long floor, unsigned long ceiling)
462 pud = pud_offset(pgd, addr);
464 next = pud_addr_end(addr, end);
465 if (pud_none_or_clear_bad(pud))
467 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
468 } while (pud++, addr = next, addr != end);
474 ceiling &= PGDIR_MASK;
478 if (end - 1 > ceiling - 1)
481 pud = pud_offset(pgd, start);
483 pud_free_tlb(tlb, pud, start);
487 * This function frees user-level page tables of a process.
489 * Must be called with pagetable lock held.
491 void free_pgd_range(struct mmu_gather *tlb,
492 unsigned long addr, unsigned long end,
493 unsigned long floor, unsigned long ceiling)
499 * The next few lines have given us lots of grief...
501 * Why are we testing PMD* at this top level? Because often
502 * there will be no work to do at all, and we'd prefer not to
503 * go all the way down to the bottom just to discover that.
505 * Why all these "- 1"s? Because 0 represents both the bottom
506 * of the address space and the top of it (using -1 for the
507 * top wouldn't help much: the masks would do the wrong thing).
508 * The rule is that addr 0 and floor 0 refer to the bottom of
509 * the address space, but end 0 and ceiling 0 refer to the top
510 * Comparisons need to use "end - 1" and "ceiling - 1" (though
511 * that end 0 case should be mythical).
513 * Wherever addr is brought up or ceiling brought down, we must
514 * be careful to reject "the opposite 0" before it confuses the
515 * subsequent tests. But what about where end is brought down
516 * by PMD_SIZE below? no, end can't go down to 0 there.
518 * Whereas we round start (addr) and ceiling down, by different
519 * masks at different levels, in order to test whether a table
520 * now has no other vmas using it, so can be freed, we don't
521 * bother to round floor or end up - the tests don't need that.
535 if (end - 1 > ceiling - 1)
540 pgd = pgd_offset(tlb->mm, addr);
542 next = pgd_addr_end(addr, end);
543 if (pgd_none_or_clear_bad(pgd))
545 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
546 } while (pgd++, addr = next, addr != end);
549 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
550 unsigned long floor, unsigned long ceiling)
553 struct vm_area_struct *next = vma->vm_next;
554 unsigned long addr = vma->vm_start;
557 * Hide vma from rmap and truncate_pagecache before freeing
560 unlink_anon_vmas(vma);
561 unlink_file_vma(vma);
563 if (is_vm_hugetlb_page(vma)) {
564 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
565 floor, next? next->vm_start: ceiling);
568 * Optimization: gather nearby vmas into one call down
570 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
571 && !is_vm_hugetlb_page(next)) {
574 unlink_anon_vmas(vma);
575 unlink_file_vma(vma);
577 free_pgd_range(tlb, addr, vma->vm_end,
578 floor, next? next->vm_start: ceiling);
584 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
585 pmd_t *pmd, unsigned long address)
587 pgtable_t new = pte_alloc_one(mm, address);
588 int wait_split_huge_page;
593 * Ensure all pte setup (eg. pte page lock and page clearing) are
594 * visible before the pte is made visible to other CPUs by being
595 * put into page tables.
597 * The other side of the story is the pointer chasing in the page
598 * table walking code (when walking the page table without locking;
599 * ie. most of the time). Fortunately, these data accesses consist
600 * of a chain of data-dependent loads, meaning most CPUs (alpha
601 * being the notable exception) will already guarantee loads are
602 * seen in-order. See the alpha page table accessors for the
603 * smp_read_barrier_depends() barriers in page table walking code.
605 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
607 spin_lock(&mm->page_table_lock);
608 wait_split_huge_page = 0;
609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate(mm, pmd, new);
613 } else if (unlikely(pmd_trans_splitting(*pmd)))
614 wait_split_huge_page = 1;
615 spin_unlock(&mm->page_table_lock);
618 if (wait_split_huge_page)
619 wait_split_huge_page(vma->anon_vma, pmd);
623 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
625 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
629 smp_wmb(); /* See comment in __pte_alloc */
631 spin_lock(&init_mm.page_table_lock);
632 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
633 pmd_populate_kernel(&init_mm, pmd, new);
636 VM_BUG_ON(pmd_trans_splitting(*pmd));
637 spin_unlock(&init_mm.page_table_lock);
639 pte_free_kernel(&init_mm, new);
643 static inline void init_rss_vec(int *rss)
645 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
648 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
652 if (current->mm == mm)
654 for (i = 0; i < NR_MM_COUNTERS; i++)
656 add_mm_counter(mm, i, rss[i]);
660 * This function is called to print an error when a bad pte
661 * is found. For example, we might have a PFN-mapped pte in
662 * a region that doesn't allow it.
664 * The calling function must still handle the error.
666 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
667 pte_t pte, struct page *page)
669 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
670 pud_t *pud = pud_offset(pgd, addr);
671 pmd_t *pmd = pmd_offset(pud, addr);
672 struct address_space *mapping;
674 static unsigned long resume;
675 static unsigned long nr_shown;
676 static unsigned long nr_unshown;
679 * Allow a burst of 60 reports, then keep quiet for that minute;
680 * or allow a steady drip of one report per second.
682 if (nr_shown == 60) {
683 if (time_before(jiffies, resume)) {
689 "BUG: Bad page map: %lu messages suppressed\n",
696 resume = jiffies + 60 * HZ;
698 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
699 index = linear_page_index(vma, addr);
702 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
704 (long long)pte_val(pte), (long long)pmd_val(*pmd));
708 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
709 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
711 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
714 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
715 (unsigned long)vma->vm_ops->fault);
716 if (vma->vm_file && vma->vm_file->f_op)
717 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
718 (unsigned long)vma->vm_file->f_op->mmap);
720 add_taint(TAINT_BAD_PAGE);
723 static inline bool is_cow_mapping(vm_flags_t flags)
725 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
729 static inline int is_zero_pfn(unsigned long pfn)
731 return pfn == zero_pfn;
736 static inline unsigned long my_zero_pfn(unsigned long addr)
743 * vm_normal_page -- This function gets the "struct page" associated with a pte.
745 * "Special" mappings do not wish to be associated with a "struct page" (either
746 * it doesn't exist, or it exists but they don't want to touch it). In this
747 * case, NULL is returned here. "Normal" mappings do have a struct page.
749 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
750 * pte bit, in which case this function is trivial. Secondly, an architecture
751 * may not have a spare pte bit, which requires a more complicated scheme,
754 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
755 * special mapping (even if there are underlying and valid "struct pages").
756 * COWed pages of a VM_PFNMAP are always normal.
758 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
759 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
760 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
761 * mapping will always honor the rule
763 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
765 * And for normal mappings this is false.
767 * This restricts such mappings to be a linear translation from virtual address
768 * to pfn. To get around this restriction, we allow arbitrary mappings so long
769 * as the vma is not a COW mapping; in that case, we know that all ptes are
770 * special (because none can have been COWed).
773 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
775 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
776 * page" backing, however the difference is that _all_ pages with a struct
777 * page (that is, those where pfn_valid is true) are refcounted and considered
778 * normal pages by the VM. The disadvantage is that pages are refcounted
779 * (which can be slower and simply not an option for some PFNMAP users). The
780 * advantage is that we don't have to follow the strict linearity rule of
781 * PFNMAP mappings in order to support COWable mappings.
784 #ifdef __HAVE_ARCH_PTE_SPECIAL
785 # define HAVE_PTE_SPECIAL 1
787 # define HAVE_PTE_SPECIAL 0
789 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
792 unsigned long pfn = pte_pfn(pte);
794 if (HAVE_PTE_SPECIAL) {
795 if (likely(!pte_special(pte)))
797 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
799 if (!is_zero_pfn(pfn))
800 print_bad_pte(vma, addr, pte, NULL);
804 /* !HAVE_PTE_SPECIAL case follows: */
806 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
807 if (vma->vm_flags & VM_MIXEDMAP) {
813 off = (addr - vma->vm_start) >> PAGE_SHIFT;
814 if (pfn == vma->vm_pgoff + off)
816 if (!is_cow_mapping(vma->vm_flags))
821 if (is_zero_pfn(pfn))
824 if (unlikely(pfn > highest_memmap_pfn)) {
825 print_bad_pte(vma, addr, pte, NULL);
830 * NOTE! We still have PageReserved() pages in the page tables.
831 * eg. VDSO mappings can cause them to exist.
834 return pfn_to_page(pfn);
838 * copy one vm_area from one task to the other. Assumes the page tables
839 * already present in the new task to be cleared in the whole range
840 * covered by this vma.
843 static inline unsigned long
844 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
845 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
846 unsigned long addr, int *rss)
848 unsigned long vm_flags = vma->vm_flags;
849 pte_t pte = *src_pte;
852 /* pte contains position in swap or file, so copy. */
853 if (unlikely(!pte_present(pte))) {
854 if (!pte_file(pte)) {
855 swp_entry_t entry = pte_to_swp_entry(pte);
857 if (swap_duplicate(entry) < 0)
860 /* make sure dst_mm is on swapoff's mmlist. */
861 if (unlikely(list_empty(&dst_mm->mmlist))) {
862 spin_lock(&mmlist_lock);
863 if (list_empty(&dst_mm->mmlist))
864 list_add(&dst_mm->mmlist,
866 spin_unlock(&mmlist_lock);
868 if (likely(!non_swap_entry(entry)))
870 else if (is_migration_entry(entry)) {
871 page = migration_entry_to_page(entry);
878 if (is_write_migration_entry(entry) &&
879 is_cow_mapping(vm_flags)) {
881 * COW mappings require pages in both
882 * parent and child to be set to read.
884 make_migration_entry_read(&entry);
885 pte = swp_entry_to_pte(entry);
886 set_pte_at(src_mm, addr, src_pte, pte);
894 * If it's a COW mapping, write protect it both
895 * in the parent and the child
897 if (is_cow_mapping(vm_flags)) {
898 ptep_set_wrprotect(src_mm, addr, src_pte);
899 pte = pte_wrprotect(pte);
903 * If it's a shared mapping, mark it clean in
906 if (vm_flags & VM_SHARED)
907 pte = pte_mkclean(pte);
908 pte = pte_mkold(pte);
910 page = vm_normal_page(vma, addr, pte);
921 set_pte_at(dst_mm, addr, dst_pte, pte);
925 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
926 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
927 unsigned long addr, unsigned long end)
929 pte_t *orig_src_pte, *orig_dst_pte;
930 pte_t *src_pte, *dst_pte;
931 spinlock_t *src_ptl, *dst_ptl;
933 int rss[NR_MM_COUNTERS];
934 swp_entry_t entry = (swp_entry_t){0};
939 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
942 src_pte = pte_offset_map(src_pmd, addr);
943 src_ptl = pte_lockptr(src_mm, src_pmd);
944 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
945 orig_src_pte = src_pte;
946 orig_dst_pte = dst_pte;
947 arch_enter_lazy_mmu_mode();
951 * We are holding two locks at this point - either of them
952 * could generate latencies in another task on another CPU.
954 if (progress >= 32) {
956 if (need_resched() ||
957 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
960 if (pte_none(*src_pte)) {
964 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
969 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
971 arch_leave_lazy_mmu_mode();
972 spin_unlock(src_ptl);
973 pte_unmap(orig_src_pte);
974 add_mm_rss_vec(dst_mm, rss);
975 pte_unmap_unlock(orig_dst_pte, dst_ptl);
979 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
988 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
989 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
990 unsigned long addr, unsigned long end)
992 pmd_t *src_pmd, *dst_pmd;
995 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
998 src_pmd = pmd_offset(src_pud, addr);
1000 next = pmd_addr_end(addr, end);
1001 if (pmd_trans_huge(*src_pmd)) {
1003 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1004 err = copy_huge_pmd(dst_mm, src_mm,
1005 dst_pmd, src_pmd, addr, vma);
1012 if (pmd_none_or_clear_bad(src_pmd))
1014 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1017 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1021 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1022 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1023 unsigned long addr, unsigned long end)
1025 pud_t *src_pud, *dst_pud;
1028 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1031 src_pud = pud_offset(src_pgd, addr);
1033 next = pud_addr_end(addr, end);
1034 if (pud_none_or_clear_bad(src_pud))
1036 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1039 } while (dst_pud++, src_pud++, addr = next, addr != end);
1043 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1044 struct vm_area_struct *vma)
1046 pgd_t *src_pgd, *dst_pgd;
1048 unsigned long addr = vma->vm_start;
1049 unsigned long end = vma->vm_end;
1050 unsigned long mmun_start; /* For mmu_notifiers */
1051 unsigned long mmun_end; /* For mmu_notifiers */
1056 * Don't copy ptes where a page fault will fill them correctly.
1057 * Fork becomes much lighter when there are big shared or private
1058 * readonly mappings. The tradeoff is that copy_page_range is more
1059 * efficient than faulting.
1061 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1062 VM_PFNMAP | VM_MIXEDMAP))) {
1067 if (is_vm_hugetlb_page(vma))
1068 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1070 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1072 * We do not free on error cases below as remove_vma
1073 * gets called on error from higher level routine
1075 ret = track_pfn_copy(vma);
1081 * We need to invalidate the secondary MMU mappings only when
1082 * there could be a permission downgrade on the ptes of the
1083 * parent mm. And a permission downgrade will only happen if
1084 * is_cow_mapping() returns true.
1086 is_cow = is_cow_mapping(vma->vm_flags);
1090 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1094 dst_pgd = pgd_offset(dst_mm, addr);
1095 src_pgd = pgd_offset(src_mm, addr);
1097 next = pgd_addr_end(addr, end);
1098 if (pgd_none_or_clear_bad(src_pgd))
1100 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1101 vma, addr, next))) {
1105 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1108 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_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 else if (is_migration_entry(entry)) {
1202 page = migration_entry_to_page(entry);
1205 rss[MM_ANONPAGES]--;
1207 rss[MM_FILEPAGES]--;
1209 if (unlikely(!free_swap_and_cache(entry)))
1210 print_bad_pte(vma, addr, ptent, NULL);
1212 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1213 } while (pte++, addr += PAGE_SIZE, addr != end);
1215 add_mm_rss_vec(mm, rss);
1216 arch_leave_lazy_mmu_mode();
1217 pte_unmap_unlock(start_pte, ptl);
1220 * mmu_gather ran out of room to batch pages, we break out of
1221 * the PTE lock to avoid doing the potential expensive TLB invalidate
1222 * and page-free while holding it.
1227 #ifdef HAVE_GENERIC_MMU_GATHER
1239 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1240 struct vm_area_struct *vma, pud_t *pud,
1241 unsigned long addr, unsigned long end,
1242 struct zap_details *details)
1247 pmd = pmd_offset(pud, addr);
1249 next = pmd_addr_end(addr, end);
1250 if (pmd_trans_huge(*pmd)) {
1251 if (next - addr != HPAGE_PMD_SIZE) {
1252 #ifdef CONFIG_DEBUG_VM
1253 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1254 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1255 __func__, addr, end,
1261 split_huge_page_pmd(vma->vm_mm, pmd);
1262 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1267 * Here there can be other concurrent MADV_DONTNEED or
1268 * trans huge page faults running, and if the pmd is
1269 * none or trans huge it can change under us. This is
1270 * because MADV_DONTNEED holds the mmap_sem in read
1273 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1275 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1278 } while (pmd++, addr = next, addr != end);
1283 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1284 struct vm_area_struct *vma, pgd_t *pgd,
1285 unsigned long addr, unsigned long end,
1286 struct zap_details *details)
1291 pud = pud_offset(pgd, addr);
1293 next = pud_addr_end(addr, end);
1294 if (pud_none_or_clear_bad(pud))
1296 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1297 } while (pud++, addr = next, addr != end);
1302 static void unmap_page_range(struct mmu_gather *tlb,
1303 struct vm_area_struct *vma,
1304 unsigned long addr, unsigned long end,
1305 struct zap_details *details)
1310 if (details && !details->check_mapping && !details->nonlinear_vma)
1313 BUG_ON(addr >= end);
1314 mem_cgroup_uncharge_start();
1315 tlb_start_vma(tlb, vma);
1316 pgd = pgd_offset(vma->vm_mm, addr);
1318 next = pgd_addr_end(addr, end);
1319 if (pgd_none_or_clear_bad(pgd))
1321 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1322 } while (pgd++, addr = next, addr != end);
1323 tlb_end_vma(tlb, vma);
1324 mem_cgroup_uncharge_end();
1328 static void unmap_single_vma(struct mmu_gather *tlb,
1329 struct vm_area_struct *vma, unsigned long start_addr,
1330 unsigned long end_addr,
1331 struct zap_details *details)
1333 unsigned long start = max(vma->vm_start, start_addr);
1336 if (start >= vma->vm_end)
1338 end = min(vma->vm_end, end_addr);
1339 if (end <= vma->vm_start)
1343 uprobe_munmap(vma, start, end);
1345 if (unlikely(vma->vm_flags & VM_PFNMAP))
1346 untrack_pfn(vma, 0, 0);
1349 if (unlikely(is_vm_hugetlb_page(vma))) {
1351 * It is undesirable to test vma->vm_file as it
1352 * should be non-null for valid hugetlb area.
1353 * However, vm_file will be NULL in the error
1354 * cleanup path of do_mmap_pgoff. When
1355 * hugetlbfs ->mmap method fails,
1356 * do_mmap_pgoff() nullifies vma->vm_file
1357 * before calling this function to clean up.
1358 * Since no pte has actually been setup, it is
1359 * safe to do nothing in this case.
1362 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1363 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1364 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1367 unmap_page_range(tlb, vma, start, end, details);
1372 * unmap_vmas - unmap a range of memory covered by a list of vma's
1373 * @tlb: address of the caller's struct mmu_gather
1374 * @vma: the starting vma
1375 * @start_addr: virtual address at which to start unmapping
1376 * @end_addr: virtual address at which to end unmapping
1378 * Unmap all pages in the vma list.
1380 * Only addresses between `start' and `end' will be unmapped.
1382 * The VMA list must be sorted in ascending virtual address order.
1384 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1385 * range after unmap_vmas() returns. So the only responsibility here is to
1386 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1387 * drops the lock and schedules.
1389 void unmap_vmas(struct mmu_gather *tlb,
1390 struct vm_area_struct *vma, unsigned long start_addr,
1391 unsigned long end_addr)
1393 struct mm_struct *mm = vma->vm_mm;
1395 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1396 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1397 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1398 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1402 * zap_page_range - remove user pages in a given range
1403 * @vma: vm_area_struct holding the applicable pages
1404 * @start: starting address of pages to zap
1405 * @size: number of bytes to zap
1406 * @details: details of nonlinear truncation or shared cache invalidation
1408 * Caller must protect the VMA list
1410 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1411 unsigned long size, struct zap_details *details)
1413 struct mm_struct *mm = vma->vm_mm;
1414 struct mmu_gather tlb;
1415 unsigned long end = start + size;
1418 tlb_gather_mmu(&tlb, mm, 0);
1419 update_hiwater_rss(mm);
1420 mmu_notifier_invalidate_range_start(mm, start, end);
1421 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1422 unmap_single_vma(&tlb, vma, start, end, details);
1423 mmu_notifier_invalidate_range_end(mm, start, end);
1424 tlb_finish_mmu(&tlb, start, end);
1428 * zap_page_range_single - remove user pages in a given range
1429 * @vma: vm_area_struct holding the applicable pages
1430 * @address: starting address of pages to zap
1431 * @size: number of bytes to zap
1432 * @details: details of nonlinear truncation or shared cache invalidation
1434 * The range must fit into one VMA.
1436 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1437 unsigned long size, struct zap_details *details)
1439 struct mm_struct *mm = vma->vm_mm;
1440 struct mmu_gather tlb;
1441 unsigned long end = address + size;
1444 tlb_gather_mmu(&tlb, mm, 0);
1445 update_hiwater_rss(mm);
1446 mmu_notifier_invalidate_range_start(mm, address, end);
1447 unmap_single_vma(&tlb, vma, address, end, details);
1448 mmu_notifier_invalidate_range_end(mm, address, end);
1449 tlb_finish_mmu(&tlb, address, end);
1453 * zap_vma_ptes - remove ptes mapping the vma
1454 * @vma: vm_area_struct holding ptes to be zapped
1455 * @address: starting address of pages to zap
1456 * @size: number of bytes to zap
1458 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1460 * The entire address range must be fully contained within the vma.
1462 * Returns 0 if successful.
1464 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1467 if (address < vma->vm_start || address + size > vma->vm_end ||
1468 !(vma->vm_flags & VM_PFNMAP))
1470 zap_page_range_single(vma, address, size, NULL);
1473 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1475 static bool pte_numa(struct vm_area_struct *vma, pte_t pte)
1478 * For NUMA page faults, we use PROT_NONE ptes in VMAs with
1479 * "normal" vma->vm_page_prot protections. Genuine PROT_NONE
1480 * VMAs should never get here, because the fault handling code
1481 * will notice that the VMA has no read or write permissions.
1483 * This means we cannot get 'special' PROT_NONE faults from genuine
1484 * PROT_NONE maps, nor from PROT_WRITE file maps that do dirty
1487 * Neither case is really interesting for our current use though so we
1490 if (pte_same(pte, pte_modify(pte, vma->vm_page_prot)))
1493 return pte_same(pte, pte_modify(pte, vma_prot_none(vma)));
1497 * follow_page - look up a page descriptor from a user-virtual address
1498 * @vma: vm_area_struct mapping @address
1499 * @address: virtual address to look up
1500 * @flags: flags modifying lookup behaviour
1502 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1504 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1505 * an error pointer if there is a mapping to something not represented
1506 * by a page descriptor (see also vm_normal_page()).
1508 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1517 struct mm_struct *mm = vma->vm_mm;
1519 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1520 if (!IS_ERR(page)) {
1521 BUG_ON(flags & FOLL_GET);
1526 pgd = pgd_offset(mm, address);
1527 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1530 pud = pud_offset(pgd, address);
1533 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1534 BUG_ON(flags & FOLL_GET);
1535 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1538 if (unlikely(pud_bad(*pud)))
1541 pmd = pmd_offset(pud, address);
1544 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1545 BUG_ON(flags & FOLL_GET);
1546 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1549 if ((flags & FOLL_NUMA) && pmd_numa(vma, *pmd))
1551 if (pmd_trans_huge(*pmd)) {
1552 if (flags & FOLL_SPLIT) {
1553 split_huge_page_pmd(mm, pmd);
1554 goto split_fallthrough;
1556 spin_lock(&mm->page_table_lock);
1557 if (likely(pmd_trans_huge(*pmd))) {
1558 if (unlikely(pmd_trans_splitting(*pmd))) {
1559 spin_unlock(&mm->page_table_lock);
1560 wait_split_huge_page(vma->anon_vma, pmd);
1562 page = follow_trans_huge_pmd(vma, address,
1564 spin_unlock(&mm->page_table_lock);
1568 spin_unlock(&mm->page_table_lock);
1572 if (unlikely(pmd_bad(*pmd)))
1575 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1578 if (!pte_present(pte))
1580 if ((flags & FOLL_NUMA) && pte_numa(vma, pte))
1582 if ((flags & FOLL_WRITE) && !pte_write(pte))
1585 page = vm_normal_page(vma, address, pte);
1586 if (unlikely(!page)) {
1587 if ((flags & FOLL_DUMP) ||
1588 !is_zero_pfn(pte_pfn(pte)))
1590 page = pte_page(pte);
1593 if (flags & FOLL_GET)
1594 get_page_foll(page);
1595 if (flags & FOLL_TOUCH) {
1596 if ((flags & FOLL_WRITE) &&
1597 !pte_dirty(pte) && !PageDirty(page))
1598 set_page_dirty(page);
1600 * pte_mkyoung() would be more correct here, but atomic care
1601 * is needed to avoid losing the dirty bit: it is easier to use
1602 * mark_page_accessed().
1604 mark_page_accessed(page);
1606 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1608 * The preliminary mapping check is mainly to avoid the
1609 * pointless overhead of lock_page on the ZERO_PAGE
1610 * which might bounce very badly if there is contention.
1612 * If the page is already locked, we don't need to
1613 * handle it now - vmscan will handle it later if and
1614 * when it attempts to reclaim the page.
1616 if (page->mapping && trylock_page(page)) {
1617 lru_add_drain(); /* push cached pages to LRU */
1619 * Because we lock page here, and migration is
1620 * blocked by the pte's page reference, and we
1621 * know the page is still mapped, we don't even
1622 * need to check for file-cache page truncation.
1624 mlock_vma_page(page);
1629 pte_unmap_unlock(ptep, ptl);
1634 pte_unmap_unlock(ptep, ptl);
1635 return ERR_PTR(-EFAULT);
1638 pte_unmap_unlock(ptep, ptl);
1644 * When core dumping an enormous anonymous area that nobody
1645 * has touched so far, we don't want to allocate unnecessary pages or
1646 * page tables. Return error instead of NULL to skip handle_mm_fault,
1647 * then get_dump_page() will return NULL to leave a hole in the dump.
1648 * But we can only make this optimization where a hole would surely
1649 * be zero-filled if handle_mm_fault() actually did handle it.
1651 if ((flags & FOLL_DUMP) &&
1652 (!vma->vm_ops || !vma->vm_ops->fault))
1653 return ERR_PTR(-EFAULT);
1657 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1659 return stack_guard_page_start(vma, addr) ||
1660 stack_guard_page_end(vma, addr+PAGE_SIZE);
1664 * __get_user_pages() - pin user pages in memory
1665 * @tsk: task_struct of target task
1666 * @mm: mm_struct of target mm
1667 * @start: starting user address
1668 * @nr_pages: number of pages from start to pin
1669 * @gup_flags: flags modifying pin behaviour
1670 * @pages: array that receives pointers to the pages pinned.
1671 * Should be at least nr_pages long. Or NULL, if caller
1672 * only intends to ensure the pages are faulted in.
1673 * @vmas: array of pointers to vmas corresponding to each page.
1674 * Or NULL if the caller does not require them.
1675 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1677 * Returns number of pages pinned. This may be fewer than the number
1678 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1679 * were pinned, returns -errno. Each page returned must be released
1680 * with a put_page() call when it is finished with. vmas will only
1681 * remain valid while mmap_sem is held.
1683 * Must be called with mmap_sem held for read or write.
1685 * __get_user_pages walks a process's page tables and takes a reference to
1686 * each struct page that each user address corresponds to at a given
1687 * instant. That is, it takes the page that would be accessed if a user
1688 * thread accesses the given user virtual address at that instant.
1690 * This does not guarantee that the page exists in the user mappings when
1691 * __get_user_pages returns, and there may even be a completely different
1692 * page there in some cases (eg. if mmapped pagecache has been invalidated
1693 * and subsequently re faulted). However it does guarantee that the page
1694 * won't be freed completely. And mostly callers simply care that the page
1695 * contains data that was valid *at some point in time*. Typically, an IO
1696 * or similar operation cannot guarantee anything stronger anyway because
1697 * locks can't be held over the syscall boundary.
1699 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1700 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1701 * appropriate) must be called after the page is finished with, and
1702 * before put_page is called.
1704 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1705 * or mmap_sem contention, and if waiting is needed to pin all pages,
1706 * *@nonblocking will be set to 0.
1708 * In most cases, get_user_pages or get_user_pages_fast should be used
1709 * instead of __get_user_pages. __get_user_pages should be used only if
1710 * you need some special @gup_flags.
1712 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1713 unsigned long start, int nr_pages, unsigned int gup_flags,
1714 struct page **pages, struct vm_area_struct **vmas,
1718 unsigned long vm_flags;
1723 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1726 * Require read or write permissions.
1727 * If FOLL_FORCE is set, we only require the "MAY" flags.
1729 vm_flags = (gup_flags & FOLL_WRITE) ?
1730 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1731 vm_flags &= (gup_flags & FOLL_FORCE) ?
1732 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1735 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1736 * would be called on PROT_NONE ranges. We must never invoke
1737 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1738 * page faults would unprotect the PROT_NONE ranges if
1739 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1740 * bitflag. So to avoid that, don't set FOLL_NUMA if
1741 * FOLL_FORCE is set.
1743 if (!(gup_flags & FOLL_FORCE))
1744 gup_flags |= FOLL_NUMA;
1749 struct vm_area_struct *vma;
1751 vma = find_extend_vma(mm, start);
1752 if (!vma && in_gate_area(mm, start)) {
1753 unsigned long pg = start & PAGE_MASK;
1759 /* user gate pages are read-only */
1760 if (gup_flags & FOLL_WRITE)
1761 return i ? : -EFAULT;
1763 pgd = pgd_offset_k(pg);
1765 pgd = pgd_offset_gate(mm, pg);
1766 BUG_ON(pgd_none(*pgd));
1767 pud = pud_offset(pgd, pg);
1768 BUG_ON(pud_none(*pud));
1769 pmd = pmd_offset(pud, pg);
1771 return i ? : -EFAULT;
1772 VM_BUG_ON(pmd_trans_huge(*pmd));
1773 pte = pte_offset_map(pmd, pg);
1774 if (pte_none(*pte)) {
1776 return i ? : -EFAULT;
1778 vma = get_gate_vma(mm);
1782 page = vm_normal_page(vma, start, *pte);
1784 if (!(gup_flags & FOLL_DUMP) &&
1785 is_zero_pfn(pte_pfn(*pte)))
1786 page = pte_page(*pte);
1789 return i ? : -EFAULT;
1800 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1801 !(vm_flags & vma->vm_flags))
1802 return i ? : -EFAULT;
1804 if (is_vm_hugetlb_page(vma)) {
1805 i = follow_hugetlb_page(mm, vma, pages, vmas,
1806 &start, &nr_pages, i, gup_flags);
1812 unsigned int foll_flags = gup_flags;
1815 * If we have a pending SIGKILL, don't keep faulting
1816 * pages and potentially allocating memory.
1818 if (unlikely(fatal_signal_pending(current)))
1819 return i ? i : -ERESTARTSYS;
1822 while (!(page = follow_page(vma, start, foll_flags))) {
1824 unsigned int fault_flags = 0;
1826 /* For mlock, just skip the stack guard page. */
1827 if (foll_flags & FOLL_MLOCK) {
1828 if (stack_guard_page(vma, start))
1831 if (foll_flags & FOLL_WRITE)
1832 fault_flags |= FAULT_FLAG_WRITE;
1834 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1835 if (foll_flags & FOLL_NOWAIT)
1836 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1838 ret = handle_mm_fault(mm, vma, start,
1841 if (ret & VM_FAULT_ERROR) {
1842 if (ret & VM_FAULT_OOM)
1843 return i ? i : -ENOMEM;
1844 if (ret & (VM_FAULT_HWPOISON |
1845 VM_FAULT_HWPOISON_LARGE)) {
1848 else if (gup_flags & FOLL_HWPOISON)
1853 if (ret & VM_FAULT_SIGBUS)
1854 return i ? i : -EFAULT;
1859 if (ret & VM_FAULT_MAJOR)
1865 if (ret & VM_FAULT_RETRY) {
1872 * The VM_FAULT_WRITE bit tells us that
1873 * do_wp_page has broken COW when necessary,
1874 * even if maybe_mkwrite decided not to set
1875 * pte_write. We can thus safely do subsequent
1876 * page lookups as if they were reads. But only
1877 * do so when looping for pte_write is futile:
1878 * in some cases userspace may also be wanting
1879 * to write to the gotten user page, which a
1880 * read fault here might prevent (a readonly
1881 * page might get reCOWed by userspace write).
1883 if ((ret & VM_FAULT_WRITE) &&
1884 !(vma->vm_flags & VM_WRITE))
1885 foll_flags &= ~FOLL_WRITE;
1890 return i ? i : PTR_ERR(page);
1894 flush_anon_page(vma, page, start);
1895 flush_dcache_page(page);
1903 } while (nr_pages && start < vma->vm_end);
1907 EXPORT_SYMBOL(__get_user_pages);
1910 * fixup_user_fault() - manually resolve a user page fault
1911 * @tsk: the task_struct to use for page fault accounting, or
1912 * NULL if faults are not to be recorded.
1913 * @mm: mm_struct of target mm
1914 * @address: user address
1915 * @fault_flags:flags to pass down to handle_mm_fault()
1917 * This is meant to be called in the specific scenario where for locking reasons
1918 * we try to access user memory in atomic context (within a pagefault_disable()
1919 * section), this returns -EFAULT, and we want to resolve the user fault before
1922 * Typically this is meant to be used by the futex code.
1924 * The main difference with get_user_pages() is that this function will
1925 * unconditionally call handle_mm_fault() which will in turn perform all the
1926 * necessary SW fixup of the dirty and young bits in the PTE, while
1927 * handle_mm_fault() only guarantees to update these in the struct page.
1929 * This is important for some architectures where those bits also gate the
1930 * access permission to the page because they are maintained in software. On
1931 * such architectures, gup() will not be enough to make a subsequent access
1934 * This should be called with the mm_sem held for read.
1936 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1937 unsigned long address, unsigned int fault_flags)
1939 struct vm_area_struct *vma;
1942 vma = find_extend_vma(mm, address);
1943 if (!vma || address < vma->vm_start)
1946 ret = handle_mm_fault(mm, vma, address, fault_flags);
1947 if (ret & VM_FAULT_ERROR) {
1948 if (ret & VM_FAULT_OOM)
1950 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1952 if (ret & VM_FAULT_SIGBUS)
1957 if (ret & VM_FAULT_MAJOR)
1966 * get_user_pages() - pin user pages in memory
1967 * @tsk: the task_struct to use for page fault accounting, or
1968 * NULL if faults are not to be recorded.
1969 * @mm: mm_struct of target mm
1970 * @start: starting user address
1971 * @nr_pages: number of pages from start to pin
1972 * @write: whether pages will be written to by the caller
1973 * @force: whether to force write access even if user mapping is
1974 * readonly. This will result in the page being COWed even
1975 * in MAP_SHARED mappings. You do not want this.
1976 * @pages: array that receives pointers to the pages pinned.
1977 * Should be at least nr_pages long. Or NULL, if caller
1978 * only intends to ensure the pages are faulted in.
1979 * @vmas: array of pointers to vmas corresponding to each page.
1980 * Or NULL if the caller does not require them.
1982 * Returns number of pages pinned. This may be fewer than the number
1983 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1984 * were pinned, returns -errno. Each page returned must be released
1985 * with a put_page() call when it is finished with. vmas will only
1986 * remain valid while mmap_sem is held.
1988 * Must be called with mmap_sem held for read or write.
1990 * get_user_pages walks a process's page tables and takes a reference to
1991 * each struct page that each user address corresponds to at a given
1992 * instant. That is, it takes the page that would be accessed if a user
1993 * thread accesses the given user virtual address at that instant.
1995 * This does not guarantee that the page exists in the user mappings when
1996 * get_user_pages returns, and there may even be a completely different
1997 * page there in some cases (eg. if mmapped pagecache has been invalidated
1998 * and subsequently re faulted). However it does guarantee that the page
1999 * won't be freed completely. And mostly callers simply care that the page
2000 * contains data that was valid *at some point in time*. Typically, an IO
2001 * or similar operation cannot guarantee anything stronger anyway because
2002 * locks can't be held over the syscall boundary.
2004 * If write=0, the page must not be written to. If the page is written to,
2005 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2006 * after the page is finished with, and before put_page is called.
2008 * get_user_pages is typically used for fewer-copy IO operations, to get a
2009 * handle on the memory by some means other than accesses via the user virtual
2010 * addresses. The pages may be submitted for DMA to devices or accessed via
2011 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2012 * use the correct cache flushing APIs.
2014 * See also get_user_pages_fast, for performance critical applications.
2016 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2017 unsigned long start, int nr_pages, int write, int force,
2018 struct page **pages, struct vm_area_struct **vmas)
2020 int flags = FOLL_TOUCH;
2025 flags |= FOLL_WRITE;
2027 flags |= FOLL_FORCE;
2029 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2032 EXPORT_SYMBOL(get_user_pages);
2035 * get_dump_page() - pin user page in memory while writing it to core dump
2036 * @addr: user address
2038 * Returns struct page pointer of user page pinned for dump,
2039 * to be freed afterwards by page_cache_release() or put_page().
2041 * Returns NULL on any kind of failure - a hole must then be inserted into
2042 * the corefile, to preserve alignment with its headers; and also returns
2043 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2044 * allowing a hole to be left in the corefile to save diskspace.
2046 * Called without mmap_sem, but after all other threads have been killed.
2048 #ifdef CONFIG_ELF_CORE
2049 struct page *get_dump_page(unsigned long addr)
2051 struct vm_area_struct *vma;
2054 if (__get_user_pages(current, current->mm, addr, 1,
2055 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2058 flush_cache_page(vma, addr, page_to_pfn(page));
2061 #endif /* CONFIG_ELF_CORE */
2063 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2066 pgd_t * pgd = pgd_offset(mm, addr);
2067 pud_t * pud = pud_alloc(mm, pgd, addr);
2069 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2071 VM_BUG_ON(pmd_trans_huge(*pmd));
2072 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2079 * This is the old fallback for page remapping.
2081 * For historical reasons, it only allows reserved pages. Only
2082 * old drivers should use this, and they needed to mark their
2083 * pages reserved for the old functions anyway.
2085 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2086 struct page *page, pgprot_t prot)
2088 struct mm_struct *mm = vma->vm_mm;
2097 flush_dcache_page(page);
2098 pte = get_locked_pte(mm, addr, &ptl);
2102 if (!pte_none(*pte))
2105 /* Ok, finally just insert the thing.. */
2107 inc_mm_counter_fast(mm, MM_FILEPAGES);
2108 page_add_file_rmap(page);
2109 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2112 pte_unmap_unlock(pte, ptl);
2115 pte_unmap_unlock(pte, ptl);
2121 * vm_insert_page - insert single page into user vma
2122 * @vma: user vma to map to
2123 * @addr: target user address of this page
2124 * @page: source kernel page
2126 * This allows drivers to insert individual pages they've allocated
2129 * The page has to be a nice clean _individual_ kernel allocation.
2130 * If you allocate a compound page, you need to have marked it as
2131 * such (__GFP_COMP), or manually just split the page up yourself
2132 * (see split_page()).
2134 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2135 * took an arbitrary page protection parameter. This doesn't allow
2136 * that. Your vma protection will have to be set up correctly, which
2137 * means that if you want a shared writable mapping, you'd better
2138 * ask for a shared writable mapping!
2140 * The page does not need to be reserved.
2142 * Usually this function is called from f_op->mmap() handler
2143 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2144 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2145 * function from other places, for example from page-fault handler.
2147 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2150 if (addr < vma->vm_start || addr >= vma->vm_end)
2152 if (!page_count(page))
2154 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2155 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2156 BUG_ON(vma->vm_flags & VM_PFNMAP);
2157 vma->vm_flags |= VM_MIXEDMAP;
2159 return insert_page(vma, addr, page, vma->vm_page_prot);
2161 EXPORT_SYMBOL(vm_insert_page);
2163 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2164 unsigned long pfn, pgprot_t prot)
2166 struct mm_struct *mm = vma->vm_mm;
2172 pte = get_locked_pte(mm, addr, &ptl);
2176 if (!pte_none(*pte))
2179 /* Ok, finally just insert the thing.. */
2180 entry = pte_mkspecial(pfn_pte(pfn, prot));
2181 set_pte_at(mm, addr, pte, entry);
2182 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2186 pte_unmap_unlock(pte, ptl);
2192 * vm_insert_pfn - insert single pfn into user vma
2193 * @vma: user vma to map to
2194 * @addr: target user address of this page
2195 * @pfn: source kernel pfn
2197 * Similar to vm_insert_page, this allows drivers to insert individual pages
2198 * they've allocated into a user vma. Same comments apply.
2200 * This function should only be called from a vm_ops->fault handler, and
2201 * in that case the handler should return NULL.
2203 * vma cannot be a COW mapping.
2205 * As this is called only for pages that do not currently exist, we
2206 * do not need to flush old virtual caches or the TLB.
2208 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2212 pgprot_t pgprot = vma->vm_page_prot;
2214 * Technically, architectures with pte_special can avoid all these
2215 * restrictions (same for remap_pfn_range). However we would like
2216 * consistency in testing and feature parity among all, so we should
2217 * try to keep these invariants in place for everybody.
2219 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2220 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2221 (VM_PFNMAP|VM_MIXEDMAP));
2222 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2223 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2225 if (addr < vma->vm_start || addr >= vma->vm_end)
2227 if (track_pfn_insert(vma, &pgprot, pfn))
2230 ret = insert_pfn(vma, addr, pfn, pgprot);
2234 EXPORT_SYMBOL(vm_insert_pfn);
2236 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2239 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2241 if (addr < vma->vm_start || addr >= vma->vm_end)
2245 * If we don't have pte special, then we have to use the pfn_valid()
2246 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2247 * refcount the page if pfn_valid is true (hence insert_page rather
2248 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2249 * without pte special, it would there be refcounted as a normal page.
2251 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2254 page = pfn_to_page(pfn);
2255 return insert_page(vma, addr, page, vma->vm_page_prot);
2257 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2259 EXPORT_SYMBOL(vm_insert_mixed);
2262 * maps a range of physical memory into the requested pages. the old
2263 * mappings are removed. any references to nonexistent pages results
2264 * in null mappings (currently treated as "copy-on-access")
2266 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2267 unsigned long addr, unsigned long end,
2268 unsigned long pfn, pgprot_t prot)
2273 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2276 arch_enter_lazy_mmu_mode();
2278 BUG_ON(!pte_none(*pte));
2279 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2281 } while (pte++, addr += PAGE_SIZE, addr != end);
2282 arch_leave_lazy_mmu_mode();
2283 pte_unmap_unlock(pte - 1, ptl);
2287 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2288 unsigned long addr, unsigned long end,
2289 unsigned long pfn, pgprot_t prot)
2294 pfn -= addr >> PAGE_SHIFT;
2295 pmd = pmd_alloc(mm, pud, addr);
2298 VM_BUG_ON(pmd_trans_huge(*pmd));
2300 next = pmd_addr_end(addr, end);
2301 if (remap_pte_range(mm, pmd, addr, next,
2302 pfn + (addr >> PAGE_SHIFT), prot))
2304 } while (pmd++, addr = next, addr != end);
2308 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2309 unsigned long addr, unsigned long end,
2310 unsigned long pfn, pgprot_t prot)
2315 pfn -= addr >> PAGE_SHIFT;
2316 pud = pud_alloc(mm, pgd, addr);
2320 next = pud_addr_end(addr, end);
2321 if (remap_pmd_range(mm, pud, addr, next,
2322 pfn + (addr >> PAGE_SHIFT), prot))
2324 } while (pud++, addr = next, addr != end);
2329 * remap_pfn_range - remap kernel memory to userspace
2330 * @vma: user vma to map to
2331 * @addr: target user address to start at
2332 * @pfn: physical address of kernel memory
2333 * @size: size of map area
2334 * @prot: page protection flags for this mapping
2336 * Note: this is only safe if the mm semaphore is held when called.
2338 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2339 unsigned long pfn, unsigned long size, pgprot_t prot)
2343 unsigned long end = addr + PAGE_ALIGN(size);
2344 struct mm_struct *mm = vma->vm_mm;
2348 * Physically remapped pages are special. Tell the
2349 * rest of the world about it:
2350 * VM_IO tells people not to look at these pages
2351 * (accesses can have side effects).
2352 * VM_PFNMAP tells the core MM that the base pages are just
2353 * raw PFN mappings, and do not have a "struct page" associated
2356 * Disable vma merging and expanding with mremap().
2358 * Omit vma from core dump, even when VM_IO turned off.
2360 * There's a horrible special case to handle copy-on-write
2361 * behaviour that some programs depend on. We mark the "original"
2362 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2363 * See vm_normal_page() for details.
2365 if (is_cow_mapping(vma->vm_flags)) {
2366 if (addr != vma->vm_start || end != vma->vm_end)
2368 vma->vm_pgoff = pfn;
2371 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2375 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2377 BUG_ON(addr >= end);
2378 pfn -= addr >> PAGE_SHIFT;
2379 pgd = pgd_offset(mm, addr);
2380 flush_cache_range(vma, addr, end);
2382 next = pgd_addr_end(addr, end);
2383 err = remap_pud_range(mm, pgd, addr, next,
2384 pfn + (addr >> PAGE_SHIFT), prot);
2387 } while (pgd++, addr = next, addr != end);
2390 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2394 EXPORT_SYMBOL(remap_pfn_range);
2396 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2397 unsigned long addr, unsigned long end,
2398 pte_fn_t fn, void *data)
2403 spinlock_t *uninitialized_var(ptl);
2405 pte = (mm == &init_mm) ?
2406 pte_alloc_kernel(pmd, addr) :
2407 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2411 BUG_ON(pmd_huge(*pmd));
2413 arch_enter_lazy_mmu_mode();
2415 token = pmd_pgtable(*pmd);
2418 err = fn(pte++, token, addr, data);
2421 } while (addr += PAGE_SIZE, addr != end);
2423 arch_leave_lazy_mmu_mode();
2426 pte_unmap_unlock(pte-1, ptl);
2430 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2431 unsigned long addr, unsigned long end,
2432 pte_fn_t fn, void *data)
2438 BUG_ON(pud_huge(*pud));
2440 pmd = pmd_alloc(mm, pud, addr);
2444 next = pmd_addr_end(addr, end);
2445 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2448 } while (pmd++, addr = next, addr != end);
2452 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2453 unsigned long addr, unsigned long end,
2454 pte_fn_t fn, void *data)
2460 pud = pud_alloc(mm, pgd, addr);
2464 next = pud_addr_end(addr, end);
2465 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2468 } while (pud++, addr = next, addr != end);
2473 * Scan a region of virtual memory, filling in page tables as necessary
2474 * and calling a provided function on each leaf page table.
2476 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2477 unsigned long size, pte_fn_t fn, void *data)
2481 unsigned long end = addr + size;
2484 BUG_ON(addr >= end);
2485 pgd = pgd_offset(mm, addr);
2487 next = pgd_addr_end(addr, end);
2488 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2491 } while (pgd++, addr = next, addr != end);
2495 EXPORT_SYMBOL_GPL(apply_to_page_range);
2498 * handle_pte_fault chooses page fault handler according to an entry
2499 * which was read non-atomically. Before making any commitment, on
2500 * those architectures or configurations (e.g. i386 with PAE) which
2501 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2502 * must check under lock before unmapping the pte and proceeding
2503 * (but do_wp_page is only called after already making such a check;
2504 * and do_anonymous_page can safely check later on).
2506 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2507 pte_t *page_table, pte_t orig_pte)
2510 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2511 if (sizeof(pte_t) > sizeof(unsigned long)) {
2512 spinlock_t *ptl = pte_lockptr(mm, pmd);
2514 same = pte_same(*page_table, orig_pte);
2518 pte_unmap(page_table);
2522 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2525 * If the source page was a PFN mapping, we don't have
2526 * a "struct page" for it. We do a best-effort copy by
2527 * just copying from the original user address. If that
2528 * fails, we just zero-fill it. Live with it.
2530 if (unlikely(!src)) {
2531 void *kaddr = kmap_atomic(dst);
2532 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2535 * This really shouldn't fail, because the page is there
2536 * in the page tables. But it might just be unreadable,
2537 * in which case we just give up and fill the result with
2540 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2542 kunmap_atomic(kaddr);
2543 flush_dcache_page(dst);
2545 copy_user_highpage(dst, src, va, vma);
2549 * This routine handles present pages, when users try to write
2550 * to a shared page. It is done by copying the page to a new address
2551 * and decrementing the shared-page counter for the old page.
2553 * Note that this routine assumes that the protection checks have been
2554 * done by the caller (the low-level page fault routine in most cases).
2555 * Thus we can safely just mark it writable once we've done any necessary
2558 * We also mark the page dirty at this point even though the page will
2559 * change only once the write actually happens. This avoids a few races,
2560 * and potentially makes it more efficient.
2562 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2563 * but allow concurrent faults), with pte both mapped and locked.
2564 * We return with mmap_sem still held, but pte unmapped and unlocked.
2566 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2567 unsigned long address, pte_t *page_table, pmd_t *pmd,
2568 spinlock_t *ptl, pte_t orig_pte)
2571 struct page *old_page, *new_page = NULL;
2574 int page_mkwrite = 0;
2575 struct page *dirty_page = NULL;
2576 unsigned long mmun_start; /* For mmu_notifiers */
2577 unsigned long mmun_end; /* For mmu_notifiers */
2578 bool mmun_called = false; /* For mmu_notifiers */
2580 old_page = vm_normal_page(vma, address, orig_pte);
2583 * VM_MIXEDMAP !pfn_valid() case
2585 * We should not cow pages in a shared writeable mapping.
2586 * Just mark the pages writable as we can't do any dirty
2587 * accounting on raw pfn maps.
2589 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2590 (VM_WRITE|VM_SHARED))
2596 * Take out anonymous pages first, anonymous shared vmas are
2597 * not dirty accountable.
2599 if (PageAnon(old_page) && !PageKsm(old_page)) {
2600 if (!trylock_page(old_page)) {
2601 page_cache_get(old_page);
2602 pte_unmap_unlock(page_table, ptl);
2603 lock_page(old_page);
2604 page_table = pte_offset_map_lock(mm, pmd, address,
2606 if (!pte_same(*page_table, orig_pte)) {
2607 unlock_page(old_page);
2610 page_cache_release(old_page);
2612 if (reuse_swap_page(old_page)) {
2614 * The page is all ours. Move it to our anon_vma so
2615 * the rmap code will not search our parent or siblings.
2616 * Protected against the rmap code by the page lock.
2618 page_move_anon_rmap(old_page, vma, address);
2619 unlock_page(old_page);
2622 unlock_page(old_page);
2623 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2624 (VM_WRITE|VM_SHARED))) {
2626 * Only catch write-faults on shared writable pages,
2627 * read-only shared pages can get COWed by
2628 * get_user_pages(.write=1, .force=1).
2630 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2631 struct vm_fault vmf;
2634 vmf.virtual_address = (void __user *)(address &
2636 vmf.pgoff = old_page->index;
2637 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2638 vmf.page = old_page;
2641 * Notify the address space that the page is about to
2642 * become writable so that it can prohibit this or wait
2643 * for the page to get into an appropriate state.
2645 * We do this without the lock held, so that it can
2646 * sleep if it needs to.
2648 page_cache_get(old_page);
2649 pte_unmap_unlock(page_table, ptl);
2651 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2653 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2655 goto unwritable_page;
2657 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2658 lock_page(old_page);
2659 if (!old_page->mapping) {
2660 ret = 0; /* retry the fault */
2661 unlock_page(old_page);
2662 goto unwritable_page;
2665 VM_BUG_ON(!PageLocked(old_page));
2668 * Since we dropped the lock we need to revalidate
2669 * the PTE as someone else may have changed it. If
2670 * they did, we just return, as we can count on the
2671 * MMU to tell us if they didn't also make it writable.
2673 page_table = pte_offset_map_lock(mm, pmd, address,
2675 if (!pte_same(*page_table, orig_pte)) {
2676 unlock_page(old_page);
2682 dirty_page = old_page;
2683 get_page(dirty_page);
2686 flush_cache_page(vma, address, pte_pfn(orig_pte));
2687 entry = pte_mkyoung(orig_pte);
2688 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2689 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2690 update_mmu_cache(vma, address, page_table);
2691 pte_unmap_unlock(page_table, ptl);
2692 ret |= VM_FAULT_WRITE;
2698 * Yes, Virginia, this is actually required to prevent a race
2699 * with clear_page_dirty_for_io() from clearing the page dirty
2700 * bit after it clear all dirty ptes, but before a racing
2701 * do_wp_page installs a dirty pte.
2703 * __do_fault is protected similarly.
2705 if (!page_mkwrite) {
2706 wait_on_page_locked(dirty_page);
2707 set_page_dirty_balance(dirty_page, page_mkwrite);
2708 /* file_update_time outside page_lock */
2710 file_update_time(vma->vm_file);
2712 put_page(dirty_page);
2714 struct address_space *mapping = dirty_page->mapping;
2716 set_page_dirty(dirty_page);
2717 unlock_page(dirty_page);
2718 page_cache_release(dirty_page);
2721 * Some device drivers do not set page.mapping
2722 * but still dirty their pages
2724 balance_dirty_pages_ratelimited(mapping);
2732 * Ok, we need to copy. Oh, well..
2734 page_cache_get(old_page);
2736 pte_unmap_unlock(page_table, ptl);
2738 if (unlikely(anon_vma_prepare(vma)))
2741 if (is_zero_pfn(pte_pfn(orig_pte))) {
2742 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2746 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2749 cow_user_page(new_page, old_page, address, vma);
2751 __SetPageUptodate(new_page);
2753 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2756 mmun_start = address & PAGE_MASK;
2757 mmun_end = (address & PAGE_MASK) + PAGE_SIZE;
2759 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2762 * Re-check the pte - we dropped the lock
2764 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2765 if (likely(pte_same(*page_table, orig_pte))) {
2767 if (!PageAnon(old_page)) {
2768 dec_mm_counter_fast(mm, MM_FILEPAGES);
2769 inc_mm_counter_fast(mm, MM_ANONPAGES);
2772 inc_mm_counter_fast(mm, MM_ANONPAGES);
2773 flush_cache_page(vma, address, pte_pfn(orig_pte));
2774 entry = mk_pte(new_page, vma->vm_page_prot);
2775 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2777 * Clear the pte entry and flush it first, before updating the
2778 * pte with the new entry. This will avoid a race condition
2779 * seen in the presence of one thread doing SMC and another
2782 ptep_clear_flush(vma, address, page_table);
2783 page_add_new_anon_rmap(new_page, vma, address);
2785 * We call the notify macro here because, when using secondary
2786 * mmu page tables (such as kvm shadow page tables), we want the
2787 * new page to be mapped directly into the secondary page table.
2789 set_pte_at_notify(mm, address, page_table, entry);
2790 update_mmu_cache(vma, address, page_table);
2793 * Only after switching the pte to the new page may
2794 * we remove the mapcount here. Otherwise another
2795 * process may come and find the rmap count decremented
2796 * before the pte is switched to the new page, and
2797 * "reuse" the old page writing into it while our pte
2798 * here still points into it and can be read by other
2801 * The critical issue is to order this
2802 * page_remove_rmap with the ptp_clear_flush above.
2803 * Those stores are ordered by (if nothing else,)
2804 * the barrier present in the atomic_add_negative
2805 * in page_remove_rmap.
2807 * Then the TLB flush in ptep_clear_flush ensures that
2808 * no process can access the old page before the
2809 * decremented mapcount is visible. And the old page
2810 * cannot be reused until after the decremented
2811 * mapcount is visible. So transitively, TLBs to
2812 * old page will be flushed before it can be reused.
2814 page_remove_rmap(old_page);
2817 /* Free the old page.. */
2818 new_page = old_page;
2819 ret |= VM_FAULT_WRITE;
2821 mem_cgroup_uncharge_page(new_page);
2824 page_cache_release(new_page);
2826 pte_unmap_unlock(page_table, ptl);
2828 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2831 * Don't let another task, with possibly unlocked vma,
2832 * keep the mlocked page.
2834 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2835 lock_page(old_page); /* LRU manipulation */
2836 munlock_vma_page(old_page);
2837 unlock_page(old_page);
2839 page_cache_release(old_page);
2843 page_cache_release(new_page);
2847 unlock_page(old_page);
2848 page_cache_release(old_page);
2850 page_cache_release(old_page);
2852 return VM_FAULT_OOM;
2855 page_cache_release(old_page);
2859 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2860 unsigned long start_addr, unsigned long end_addr,
2861 struct zap_details *details)
2863 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2866 static inline void unmap_mapping_range_tree(struct rb_root *root,
2867 struct zap_details *details)
2869 struct vm_area_struct *vma;
2870 pgoff_t vba, vea, zba, zea;
2872 vma_interval_tree_foreach(vma, root,
2873 details->first_index, details->last_index) {
2875 vba = vma->vm_pgoff;
2876 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2877 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2878 zba = details->first_index;
2881 zea = details->last_index;
2885 unmap_mapping_range_vma(vma,
2886 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2887 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2892 static inline void unmap_mapping_range_list(struct list_head *head,
2893 struct zap_details *details)
2895 struct vm_area_struct *vma;
2898 * In nonlinear VMAs there is no correspondence between virtual address
2899 * offset and file offset. So we must perform an exhaustive search
2900 * across *all* the pages in each nonlinear VMA, not just the pages
2901 * whose virtual address lies outside the file truncation point.
2903 list_for_each_entry(vma, head, shared.nonlinear) {
2904 details->nonlinear_vma = vma;
2905 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2910 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2911 * @mapping: the address space containing mmaps to be unmapped.
2912 * @holebegin: byte in first page to unmap, relative to the start of
2913 * the underlying file. This will be rounded down to a PAGE_SIZE
2914 * boundary. Note that this is different from truncate_pagecache(), which
2915 * must keep the partial page. In contrast, we must get rid of
2917 * @holelen: size of prospective hole in bytes. This will be rounded
2918 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2920 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2921 * but 0 when invalidating pagecache, don't throw away private data.
2923 void unmap_mapping_range(struct address_space *mapping,
2924 loff_t const holebegin, loff_t const holelen, int even_cows)
2926 struct zap_details details;
2927 pgoff_t hba = holebegin >> PAGE_SHIFT;
2928 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2930 /* Check for overflow. */
2931 if (sizeof(holelen) > sizeof(hlen)) {
2933 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2934 if (holeend & ~(long long)ULONG_MAX)
2935 hlen = ULONG_MAX - hba + 1;
2938 details.check_mapping = even_cows? NULL: mapping;
2939 details.nonlinear_vma = NULL;
2940 details.first_index = hba;
2941 details.last_index = hba + hlen - 1;
2942 if (details.last_index < details.first_index)
2943 details.last_index = ULONG_MAX;
2946 mutex_lock(&mapping->i_mmap_mutex);
2947 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2948 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2949 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2950 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2951 mutex_unlock(&mapping->i_mmap_mutex);
2953 EXPORT_SYMBOL(unmap_mapping_range);
2956 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2957 * but allow concurrent faults), and pte mapped but not yet locked.
2958 * We return with mmap_sem still held, but pte unmapped and unlocked.
2960 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2961 unsigned long address, pte_t *page_table, pmd_t *pmd,
2962 unsigned int flags, pte_t orig_pte)
2965 struct page *page, *swapcache = NULL;
2969 struct mem_cgroup *ptr;
2973 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2976 entry = pte_to_swp_entry(orig_pte);
2977 if (unlikely(non_swap_entry(entry))) {
2978 if (is_migration_entry(entry)) {
2979 migration_entry_wait(mm, pmd, address);
2980 } else if (is_hwpoison_entry(entry)) {
2981 ret = VM_FAULT_HWPOISON;
2983 print_bad_pte(vma, address, orig_pte, NULL);
2984 ret = VM_FAULT_SIGBUS;
2988 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2989 page = lookup_swap_cache(entry);
2991 page = swapin_readahead(entry,
2992 GFP_HIGHUSER_MOVABLE, vma, address);
2995 * Back out if somebody else faulted in this pte
2996 * while we released the pte lock.
2998 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2999 if (likely(pte_same(*page_table, orig_pte)))
3001 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3005 /* Had to read the page from swap area: Major fault */
3006 ret = VM_FAULT_MAJOR;
3007 count_vm_event(PGMAJFAULT);
3008 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3009 } else if (PageHWPoison(page)) {
3011 * hwpoisoned dirty swapcache pages are kept for killing
3012 * owner processes (which may be unknown at hwpoison time)
3014 ret = VM_FAULT_HWPOISON;
3015 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3017 } else if (!(flags & FAULT_FLAG_TRIED))
3018 swap_cache_hit(vma);
3020 locked = lock_page_or_retry(page, mm, flags);
3022 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3024 ret |= VM_FAULT_RETRY;
3029 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3030 * release the swapcache from under us. The page pin, and pte_same
3031 * test below, are not enough to exclude that. Even if it is still
3032 * swapcache, we need to check that the page's swap has not changed.
3034 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3037 if (ksm_might_need_to_copy(page, vma, address)) {
3039 page = ksm_does_need_to_copy(page, vma, address);
3041 if (unlikely(!page)) {
3049 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3055 * Back out if somebody else already faulted in this pte.
3057 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3058 if (unlikely(!pte_same(*page_table, orig_pte)))
3061 if (unlikely(!PageUptodate(page))) {
3062 ret = VM_FAULT_SIGBUS;
3067 * The page isn't present yet, go ahead with the fault.
3069 * Be careful about the sequence of operations here.
3070 * To get its accounting right, reuse_swap_page() must be called
3071 * while the page is counted on swap but not yet in mapcount i.e.
3072 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3073 * must be called after the swap_free(), or it will never succeed.
3074 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3075 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3076 * in page->private. In this case, a record in swap_cgroup is silently
3077 * discarded at swap_free().
3080 inc_mm_counter_fast(mm, MM_ANONPAGES);
3081 dec_mm_counter_fast(mm, MM_SWAPENTS);
3082 pte = mk_pte(page, vma->vm_page_prot);
3083 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3084 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3085 flags &= ~FAULT_FLAG_WRITE;
3086 ret |= VM_FAULT_WRITE;
3089 flush_icache_page(vma, page);
3090 set_pte_at(mm, address, page_table, pte);
3091 do_page_add_anon_rmap(page, vma, address, exclusive);
3092 /* It's better to call commit-charge after rmap is established */
3093 mem_cgroup_commit_charge_swapin(page, ptr);
3096 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3097 try_to_free_swap(page);
3101 * Hold the lock to avoid the swap entry to be reused
3102 * until we take the PT lock for the pte_same() check
3103 * (to avoid false positives from pte_same). For
3104 * further safety release the lock after the swap_free
3105 * so that the swap count won't change under a
3106 * parallel locked swapcache.
3108 unlock_page(swapcache);
3109 page_cache_release(swapcache);
3112 if (flags & FAULT_FLAG_WRITE) {
3113 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3114 if (ret & VM_FAULT_ERROR)
3115 ret &= VM_FAULT_ERROR;
3119 /* No need to invalidate - it was non-present before */
3120 update_mmu_cache(vma, address, page_table);
3122 pte_unmap_unlock(page_table, ptl);
3126 mem_cgroup_cancel_charge_swapin(ptr);
3127 pte_unmap_unlock(page_table, ptl);
3131 page_cache_release(page);
3133 unlock_page(swapcache);
3134 page_cache_release(swapcache);
3140 * This is like a special single-page "expand_{down|up}wards()",
3141 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3142 * doesn't hit another vma.
3144 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3146 address &= PAGE_MASK;
3147 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3148 struct vm_area_struct *prev = vma->vm_prev;
3151 * Is there a mapping abutting this one below?
3153 * That's only ok if it's the same stack mapping
3154 * that has gotten split..
3156 if (prev && prev->vm_end == address)
3157 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3159 expand_downwards(vma, address - PAGE_SIZE);
3161 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3162 struct vm_area_struct *next = vma->vm_next;
3164 /* As VM_GROWSDOWN but s/below/above/ */
3165 if (next && next->vm_start == address + PAGE_SIZE)
3166 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3168 expand_upwards(vma, address + PAGE_SIZE);
3174 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3175 * but allow concurrent faults), and pte mapped but not yet locked.
3176 * We return with mmap_sem still held, but pte unmapped and unlocked.
3178 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3179 unsigned long address, pte_t *page_table, pmd_t *pmd,
3186 pte_unmap(page_table);
3188 /* Check if we need to add a guard page to the stack */
3189 if (check_stack_guard_page(vma, address) < 0)
3190 return VM_FAULT_SIGBUS;
3192 /* Use the zero-page for reads */
3193 if (!(flags & FAULT_FLAG_WRITE)) {
3194 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3195 vma->vm_page_prot));
3196 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3197 if (!pte_none(*page_table))
3202 /* Allocate our own private page. */
3203 if (unlikely(anon_vma_prepare(vma)))
3205 page = alloc_zeroed_user_highpage_movable(vma, address);
3208 __SetPageUptodate(page);
3210 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3213 entry = mk_pte(page, vma->vm_page_prot);
3214 if (vma->vm_flags & VM_WRITE)
3215 entry = pte_mkwrite(pte_mkdirty(entry));
3217 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3218 if (!pte_none(*page_table))
3221 inc_mm_counter_fast(mm, MM_ANONPAGES);
3222 page_add_new_anon_rmap(page, vma, address);
3224 set_pte_at(mm, address, page_table, entry);
3226 /* No need to invalidate - it was non-present before */
3227 update_mmu_cache(vma, address, page_table);
3229 pte_unmap_unlock(page_table, ptl);
3232 mem_cgroup_uncharge_page(page);
3233 page_cache_release(page);
3236 page_cache_release(page);
3238 return VM_FAULT_OOM;
3242 * __do_fault() tries to create a new page mapping. It aggressively
3243 * tries to share with existing pages, but makes a separate copy if
3244 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3245 * the next page fault.
3247 * As this is called only for pages that do not currently exist, we
3248 * do not need to flush old virtual caches or the TLB.
3250 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3251 * but allow concurrent faults), and pte neither mapped nor locked.
3252 * We return with mmap_sem still held, but pte unmapped and unlocked.
3254 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3255 unsigned long address, pmd_t *pmd,
3256 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3261 struct page *cow_page;
3264 struct page *dirty_page = NULL;
3265 struct vm_fault vmf;
3267 int page_mkwrite = 0;
3270 * If we do COW later, allocate page befor taking lock_page()
3271 * on the file cache page. This will reduce lock holding time.
3273 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3275 if (unlikely(anon_vma_prepare(vma)))
3276 return VM_FAULT_OOM;
3278 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3280 return VM_FAULT_OOM;
3282 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3283 page_cache_release(cow_page);
3284 return VM_FAULT_OOM;
3289 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3294 ret = vma->vm_ops->fault(vma, &vmf);
3295 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3299 if (unlikely(PageHWPoison(vmf.page))) {
3300 if (ret & VM_FAULT_LOCKED)
3301 unlock_page(vmf.page);
3302 ret = VM_FAULT_HWPOISON;
3307 * For consistency in subsequent calls, make the faulted page always
3310 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3311 lock_page(vmf.page);
3313 VM_BUG_ON(!PageLocked(vmf.page));
3316 * Should we do an early C-O-W break?
3319 if (flags & FAULT_FLAG_WRITE) {
3320 if (!(vma->vm_flags & VM_SHARED)) {
3323 copy_user_highpage(page, vmf.page, address, vma);
3324 __SetPageUptodate(page);
3327 * If the page will be shareable, see if the backing
3328 * address space wants to know that the page is about
3329 * to become writable
3331 if (vma->vm_ops->page_mkwrite) {
3335 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3336 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3338 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3340 goto unwritable_page;
3342 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3344 if (!page->mapping) {
3345 ret = 0; /* retry the fault */
3347 goto unwritable_page;
3350 VM_BUG_ON(!PageLocked(page));
3357 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3360 * This silly early PAGE_DIRTY setting removes a race
3361 * due to the bad i386 page protection. But it's valid
3362 * for other architectures too.
3364 * Note that if FAULT_FLAG_WRITE is set, we either now have
3365 * an exclusive copy of the page, or this is a shared mapping,
3366 * so we can make it writable and dirty to avoid having to
3367 * handle that later.
3369 /* Only go through if we didn't race with anybody else... */
3370 if (likely(pte_same(*page_table, orig_pte))) {
3371 flush_icache_page(vma, page);
3372 entry = mk_pte(page, vma->vm_page_prot);
3373 if (flags & FAULT_FLAG_WRITE)
3374 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3376 inc_mm_counter_fast(mm, MM_ANONPAGES);
3377 page_add_new_anon_rmap(page, vma, address);
3379 inc_mm_counter_fast(mm, MM_FILEPAGES);
3380 page_add_file_rmap(page);
3381 if (flags & FAULT_FLAG_WRITE) {
3383 get_page(dirty_page);
3386 set_pte_at(mm, address, page_table, entry);
3388 /* no need to invalidate: a not-present page won't be cached */
3389 update_mmu_cache(vma, address, page_table);
3392 mem_cgroup_uncharge_page(cow_page);
3394 page_cache_release(page);
3396 anon = 1; /* no anon but release faulted_page */
3399 pte_unmap_unlock(page_table, ptl);
3402 struct address_space *mapping = page->mapping;
3405 if (set_page_dirty(dirty_page))
3407 unlock_page(dirty_page);
3408 put_page(dirty_page);
3409 if ((dirtied || page_mkwrite) && mapping) {
3411 * Some device drivers do not set page.mapping but still
3414 balance_dirty_pages_ratelimited(mapping);
3417 /* file_update_time outside page_lock */
3418 if (vma->vm_file && !page_mkwrite)
3419 file_update_time(vma->vm_file);
3421 unlock_page(vmf.page);
3423 page_cache_release(vmf.page);
3429 page_cache_release(page);
3432 /* fs's fault handler get error */
3434 mem_cgroup_uncharge_page(cow_page);
3435 page_cache_release(cow_page);
3440 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3441 unsigned long address, pte_t *page_table, pmd_t *pmd,
3442 unsigned int flags, pte_t orig_pte)
3444 pgoff_t pgoff = (((address & PAGE_MASK)
3445 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3447 pte_unmap(page_table);
3448 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3452 * Fault of a previously existing named mapping. Repopulate the pte
3453 * from the encoded file_pte if possible. This enables swappable
3456 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3457 * but allow concurrent faults), and pte mapped but not yet locked.
3458 * We return with mmap_sem still held, but pte unmapped and unlocked.
3460 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3461 unsigned long address, pte_t *page_table, pmd_t *pmd,
3462 unsigned int flags, pte_t orig_pte)
3466 flags |= FAULT_FLAG_NONLINEAR;
3468 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3471 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3473 * Page table corrupted: show pte and kill process.
3475 print_bad_pte(vma, address, orig_pte, NULL);
3476 return VM_FAULT_SIGBUS;
3479 pgoff = pte_to_pgoff(orig_pte);
3480 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3483 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3484 unsigned long address, pte_t *ptep, pmd_t *pmd,
3485 unsigned int flags, pte_t entry)
3487 struct page *page = NULL;
3488 int node, page_nid = -1;
3491 ptl = pte_lockptr(mm, pmd);
3493 if (unlikely(!pte_same(*ptep, entry)))
3496 page = vm_normal_page(vma, address, entry);
3499 page_nid = page_to_nid(page);
3500 node = mpol_misplaced(page, vma, address);
3505 out_pte_upgrade_unlock:
3506 flush_cache_page(vma, address, pte_pfn(entry));
3508 ptep_modify_prot_start(mm, address, ptep);
3509 entry = pte_modify(entry, vma->vm_page_prot);
3510 ptep_modify_prot_commit(mm, address, ptep, entry);
3512 /* No TLB flush needed because we upgraded the PTE */
3514 update_mmu_cache(vma, address, ptep);
3517 pte_unmap_unlock(ptep, ptl);
3520 task_numa_fault(page_nid, 1);
3527 pte_unmap_unlock(ptep, ptl);
3529 if (!migrate_misplaced_page(page, node)) {
3534 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3535 if (!pte_same(*ptep, entry)) {
3541 goto out_pte_upgrade_unlock;
3545 * These routines also need to handle stuff like marking pages dirty
3546 * and/or accessed for architectures that don't do it in hardware (most
3547 * RISC architectures). The early dirtying is also good on the i386.
3549 * There is also a hook called "update_mmu_cache()" that architectures
3550 * with external mmu caches can use to update those (ie the Sparc or
3551 * PowerPC hashed page tables that act as extended TLBs).
3553 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3554 * but allow concurrent faults), and pte mapped but not yet locked.
3555 * We return with mmap_sem still held, but pte unmapped and unlocked.
3557 int handle_pte_fault(struct mm_struct *mm,
3558 struct vm_area_struct *vma, unsigned long address,
3559 pte_t *pte, pmd_t *pmd, unsigned int flags)
3564 entry = ACCESS_ONCE(*pte);
3565 if (!pte_present(entry)) {
3566 if (pte_none(entry)) {
3568 if (likely(vma->vm_ops->fault))
3569 return do_linear_fault(mm, vma, address,
3570 pte, pmd, flags, entry);
3572 return do_anonymous_page(mm, vma, address,
3575 if (pte_file(entry))
3576 return do_nonlinear_fault(mm, vma, address,
3577 pte, pmd, flags, entry);
3578 return do_swap_page(mm, vma, address,
3579 pte, pmd, flags, entry);
3582 if (pte_numa(vma, entry))
3583 return do_numa_page(mm, vma, address, pte, pmd, flags, entry);
3585 ptl = pte_lockptr(mm, pmd);
3587 if (unlikely(!pte_same(*pte, entry)))
3589 if (flags & FAULT_FLAG_WRITE) {
3590 if (!pte_write(entry))
3591 return do_wp_page(mm, vma, address,
3592 pte, pmd, ptl, entry);
3593 entry = pte_mkdirty(entry);
3595 entry = pte_mkyoung(entry);
3596 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3597 update_mmu_cache(vma, address, pte);
3600 * This is needed only for protection faults but the arch code
3601 * is not yet telling us if this is a protection fault or not.
3602 * This still avoids useless tlb flushes for .text page faults
3605 if (flags & FAULT_FLAG_WRITE)
3606 flush_tlb_fix_spurious_fault(vma, address);
3609 pte_unmap_unlock(pte, ptl);
3614 * By the time we get here, we already hold the mm semaphore
3616 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3617 unsigned long address, unsigned int flags)
3624 __set_current_state(TASK_RUNNING);
3626 count_vm_event(PGFAULT);
3627 mem_cgroup_count_vm_event(mm, PGFAULT);
3629 /* do counter updates before entering really critical section. */
3630 check_sync_rss_stat(current);
3632 if (unlikely(is_vm_hugetlb_page(vma)))
3633 return hugetlb_fault(mm, vma, address, flags);
3636 pgd = pgd_offset(mm, address);
3637 pud = pud_alloc(mm, pgd, address);
3639 return VM_FAULT_OOM;
3640 pmd = pmd_alloc(mm, pud, address);
3642 return VM_FAULT_OOM;
3643 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3645 return do_huge_pmd_anonymous_page(mm, vma, address,
3648 pmd_t orig_pmd = *pmd;
3652 if (pmd_trans_huge(orig_pmd) && !pmd_trans_splitting(orig_pmd)) {
3653 if (pmd_numa(vma, orig_pmd)) {
3654 do_huge_pmd_numa_page(mm, vma, address, pmd,
3658 if ((flags & FAULT_FLAG_WRITE) && !pmd_write(orig_pmd)) {
3659 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3662 * If COW results in an oom, the huge pmd will
3663 * have been split, so retry the fault on the
3664 * pte for a smaller charge.
3666 if (unlikely(ret & VM_FAULT_OOM))
3676 * Use __pte_alloc instead of pte_alloc_map, because we can't
3677 * run pte_offset_map on the pmd, if an huge pmd could
3678 * materialize from under us from a different thread.
3680 if (unlikely(pmd_none(*pmd)) &&
3681 unlikely(__pte_alloc(mm, vma, pmd, address)))
3682 return VM_FAULT_OOM;
3683 /* if an huge pmd materialized from under us just retry later */
3684 if (unlikely(pmd_trans_huge(*pmd)))
3687 * A regular pmd is established and it can't morph into a huge pmd
3688 * from under us anymore at this point because we hold the mmap_sem
3689 * read mode and khugepaged takes it in write mode. So now it's
3690 * safe to run pte_offset_map().
3692 pte = pte_offset_map(pmd, address);
3694 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3697 #ifndef __PAGETABLE_PUD_FOLDED
3699 * Allocate page upper directory.
3700 * We've already handled the fast-path in-line.
3702 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3704 pud_t *new = pud_alloc_one(mm, address);
3708 smp_wmb(); /* See comment in __pte_alloc */
3710 spin_lock(&mm->page_table_lock);
3711 if (pgd_present(*pgd)) /* Another has populated it */
3714 pgd_populate(mm, pgd, new);
3715 spin_unlock(&mm->page_table_lock);
3718 #endif /* __PAGETABLE_PUD_FOLDED */
3720 #ifndef __PAGETABLE_PMD_FOLDED
3722 * Allocate page middle directory.
3723 * We've already handled the fast-path in-line.
3725 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3727 pmd_t *new = pmd_alloc_one(mm, address);
3731 smp_wmb(); /* See comment in __pte_alloc */
3733 spin_lock(&mm->page_table_lock);
3734 #ifndef __ARCH_HAS_4LEVEL_HACK
3735 if (pud_present(*pud)) /* Another has populated it */
3738 pud_populate(mm, pud, new);
3740 if (pgd_present(*pud)) /* Another has populated it */
3743 pgd_populate(mm, pud, new);
3744 #endif /* __ARCH_HAS_4LEVEL_HACK */
3745 spin_unlock(&mm->page_table_lock);
3748 #endif /* __PAGETABLE_PMD_FOLDED */
3750 int make_pages_present(unsigned long addr, unsigned long end)
3752 int ret, len, write;
3753 struct vm_area_struct * vma;
3755 vma = find_vma(current->mm, addr);
3759 * We want to touch writable mappings with a write fault in order
3760 * to break COW, except for shared mappings because these don't COW
3761 * and we would not want to dirty them for nothing.
3763 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3764 BUG_ON(addr >= end);
3765 BUG_ON(end > vma->vm_end);
3766 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3767 ret = get_user_pages(current, current->mm, addr,
3768 len, write, 0, NULL, NULL);
3771 return ret == len ? 0 : -EFAULT;
3774 #if !defined(__HAVE_ARCH_GATE_AREA)
3776 #if defined(AT_SYSINFO_EHDR)
3777 static struct vm_area_struct gate_vma;
3779 static int __init gate_vma_init(void)
3781 gate_vma.vm_mm = NULL;
3782 gate_vma.vm_start = FIXADDR_USER_START;
3783 gate_vma.vm_end = FIXADDR_USER_END;
3784 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3785 gate_vma.vm_page_prot = __P101;
3789 __initcall(gate_vma_init);
3792 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3794 #ifdef AT_SYSINFO_EHDR
3801 int in_gate_area_no_mm(unsigned long addr)
3803 #ifdef AT_SYSINFO_EHDR
3804 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3810 #endif /* __HAVE_ARCH_GATE_AREA */
3812 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3813 pte_t **ptepp, spinlock_t **ptlp)
3820 pgd = pgd_offset(mm, address);
3821 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3824 pud = pud_offset(pgd, address);
3825 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3828 pmd = pmd_offset(pud, address);
3829 VM_BUG_ON(pmd_trans_huge(*pmd));
3830 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3833 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3837 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3840 if (!pte_present(*ptep))
3845 pte_unmap_unlock(ptep, *ptlp);
3850 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3851 pte_t **ptepp, spinlock_t **ptlp)
3855 /* (void) is needed to make gcc happy */
3856 (void) __cond_lock(*ptlp,
3857 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3862 * follow_pfn - look up PFN at a user virtual address
3863 * @vma: memory mapping
3864 * @address: user virtual address
3865 * @pfn: location to store found PFN
3867 * Only IO mappings and raw PFN mappings are allowed.
3869 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3871 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3878 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3881 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3884 *pfn = pte_pfn(*ptep);
3885 pte_unmap_unlock(ptep, ptl);
3888 EXPORT_SYMBOL(follow_pfn);
3890 #ifdef CONFIG_HAVE_IOREMAP_PROT
3891 int follow_phys(struct vm_area_struct *vma,
3892 unsigned long address, unsigned int flags,
3893 unsigned long *prot, resource_size_t *phys)
3899 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3902 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3906 if ((flags & FOLL_WRITE) && !pte_write(pte))
3909 *prot = pgprot_val(pte_pgprot(pte));
3910 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3914 pte_unmap_unlock(ptep, ptl);
3919 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3920 void *buf, int len, int write)
3922 resource_size_t phys_addr;
3923 unsigned long prot = 0;
3924 void __iomem *maddr;
3925 int offset = addr & (PAGE_SIZE-1);
3927 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3930 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3932 memcpy_toio(maddr + offset, buf, len);
3934 memcpy_fromio(buf, maddr + offset, len);
3942 * Access another process' address space as given in mm. If non-NULL, use the
3943 * given task for page fault accounting.
3945 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3946 unsigned long addr, void *buf, int len, int write)
3948 struct vm_area_struct *vma;
3949 void *old_buf = buf;
3951 down_read(&mm->mmap_sem);
3952 /* ignore errors, just check how much was successfully transferred */
3954 int bytes, ret, offset;
3956 struct page *page = NULL;
3958 ret = get_user_pages(tsk, mm, addr, 1,
3959 write, 1, &page, &vma);
3962 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3963 * we can access using slightly different code.
3965 #ifdef CONFIG_HAVE_IOREMAP_PROT
3966 vma = find_vma(mm, addr);
3967 if (!vma || vma->vm_start > addr)
3969 if (vma->vm_ops && vma->vm_ops->access)
3970 ret = vma->vm_ops->access(vma, addr, buf,
3978 offset = addr & (PAGE_SIZE-1);
3979 if (bytes > PAGE_SIZE-offset)
3980 bytes = PAGE_SIZE-offset;
3984 copy_to_user_page(vma, page, addr,
3985 maddr + offset, buf, bytes);
3986 set_page_dirty_lock(page);
3988 copy_from_user_page(vma, page, addr,
3989 buf, maddr + offset, bytes);
3992 page_cache_release(page);
3998 up_read(&mm->mmap_sem);
4000 return buf - old_buf;
4004 * access_remote_vm - access another process' address space
4005 * @mm: the mm_struct of the target address space
4006 * @addr: start address to access
4007 * @buf: source or destination buffer
4008 * @len: number of bytes to transfer
4009 * @write: whether the access is a write
4011 * The caller must hold a reference on @mm.
4013 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4014 void *buf, int len, int write)
4016 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4020 * Access another process' address space.
4021 * Source/target buffer must be kernel space,
4022 * Do not walk the page table directly, use get_user_pages
4024 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4025 void *buf, int len, int write)
4027 struct mm_struct *mm;
4030 mm = get_task_mm(tsk);
4034 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4041 * Print the name of a VMA.
4043 void print_vma_addr(char *prefix, unsigned long ip)
4045 struct mm_struct *mm = current->mm;
4046 struct vm_area_struct *vma;
4049 * Do not print if we are in atomic
4050 * contexts (in exception stacks, etc.):
4052 if (preempt_count())
4055 down_read(&mm->mmap_sem);
4056 vma = find_vma(mm, ip);
4057 if (vma && vma->vm_file) {
4058 struct file *f = vma->vm_file;
4059 char *buf = (char *)__get_free_page(GFP_KERNEL);
4063 p = d_path(&f->f_path, buf, PAGE_SIZE);
4066 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4068 vma->vm_end - vma->vm_start);
4069 free_page((unsigned long)buf);
4072 up_read(&mm->mmap_sem);
4075 #ifdef CONFIG_PROVE_LOCKING
4076 void might_fault(void)
4079 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4080 * holding the mmap_sem, this is safe because kernel memory doesn't
4081 * get paged out, therefore we'll never actually fault, and the
4082 * below annotations will generate false positives.
4084 if (segment_eq(get_fs(), KERNEL_DS))
4089 * it would be nicer only to annotate paths which are not under
4090 * pagefault_disable, however that requires a larger audit and
4091 * providing helpers like get_user_atomic.
4093 if (!in_atomic() && current->mm)
4094 might_lock_read(¤t->mm->mmap_sem);
4096 EXPORT_SYMBOL(might_fault);
4099 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4100 static void clear_gigantic_page(struct page *page,
4102 unsigned int pages_per_huge_page)
4105 struct page *p = page;
4108 for (i = 0; i < pages_per_huge_page;
4109 i++, p = mem_map_next(p, page, i)) {
4111 clear_user_highpage(p, addr + i * PAGE_SIZE);
4114 void clear_huge_page(struct page *page,
4115 unsigned long addr, unsigned int pages_per_huge_page)
4119 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4120 clear_gigantic_page(page, addr, pages_per_huge_page);
4125 for (i = 0; i < pages_per_huge_page; i++) {
4127 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4131 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4133 struct vm_area_struct *vma,
4134 unsigned int pages_per_huge_page)
4137 struct page *dst_base = dst;
4138 struct page *src_base = src;
4140 for (i = 0; i < pages_per_huge_page; ) {
4142 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4145 dst = mem_map_next(dst, dst_base, i);
4146 src = mem_map_next(src, src_base, i);
4150 void copy_user_huge_page(struct page *dst, struct page *src,
4151 unsigned long addr, struct vm_area_struct *vma,
4152 unsigned int pages_per_huge_page)
4156 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4157 copy_user_gigantic_page(dst, src, addr, vma,
4158 pages_per_huge_page);
4163 for (i = 0; i < pages_per_huge_page; i++) {
4165 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4168 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */