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>
65 #include <asm/pgalloc.h>
66 #include <asm/uaccess.h>
68 #include <asm/tlbflush.h>
69 #include <asm/pgtable.h>
73 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
74 #warning Unfortunate NUMA config, growing page-frame for last_nid.
77 #ifndef CONFIG_NEED_MULTIPLE_NODES
78 /* use the per-pgdat data instead for discontigmem - mbligh */
79 unsigned long max_mapnr;
82 EXPORT_SYMBOL(max_mapnr);
83 EXPORT_SYMBOL(mem_map);
86 unsigned long num_physpages;
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 EXPORT_SYMBOL(num_physpages);
97 EXPORT_SYMBOL(high_memory);
100 * Randomize the address space (stacks, mmaps, brk, etc.).
102 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103 * as ancient (libc5 based) binaries can segfault. )
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
112 static int __init disable_randmaps(char *s)
114 randomize_va_space = 0;
117 __setup("norandmaps", disable_randmaps);
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
123 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
125 static int __init init_zero_pfn(void)
127 zero_pfn = page_to_pfn(ZERO_PAGE(0));
130 core_initcall(init_zero_pfn);
133 #if defined(SPLIT_RSS_COUNTING)
135 void sync_mm_rss(struct mm_struct *mm)
139 for (i = 0; i < NR_MM_COUNTERS; i++) {
140 if (current->rss_stat.count[i]) {
141 add_mm_counter(mm, i, current->rss_stat.count[i]);
142 current->rss_stat.count[i] = 0;
145 current->rss_stat.events = 0;
148 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
150 struct task_struct *task = current;
152 if (likely(task->mm == mm))
153 task->rss_stat.count[member] += val;
155 add_mm_counter(mm, member, val);
157 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
158 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
160 /* sync counter once per 64 page faults */
161 #define TASK_RSS_EVENTS_THRESH (64)
162 static void check_sync_rss_stat(struct task_struct *task)
164 if (unlikely(task != current))
166 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
167 sync_mm_rss(task->mm);
169 #else /* SPLIT_RSS_COUNTING */
171 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
172 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
174 static void check_sync_rss_stat(struct task_struct *task)
178 #endif /* SPLIT_RSS_COUNTING */
180 #ifdef HAVE_GENERIC_MMU_GATHER
182 static int tlb_next_batch(struct mmu_gather *tlb)
184 struct mmu_gather_batch *batch;
188 tlb->active = batch->next;
192 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
198 batch->max = MAX_GATHER_BATCH;
200 tlb->active->next = batch;
207 * Called to initialize an (on-stack) mmu_gather structure for page-table
208 * tear-down from @mm. The @fullmm argument is used when @mm is without
209 * users and we're going to destroy the full address space (exit/execve).
211 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
215 tlb->fullmm = fullmm;
219 tlb->fast_mode = (num_possible_cpus() == 1);
220 tlb->local.next = NULL;
222 tlb->local.max = ARRAY_SIZE(tlb->__pages);
223 tlb->active = &tlb->local;
225 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 void tlb_flush_mmu(struct mmu_gather *tlb)
232 struct mmu_gather_batch *batch;
234 if (!tlb->need_flush)
238 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
239 tlb_table_flush(tlb);
242 if (tlb_fast_mode(tlb))
245 for (batch = &tlb->local; batch; batch = batch->next) {
246 free_pages_and_swap_cache(batch->pages, batch->nr);
249 tlb->active = &tlb->local;
253 * Called at the end of the shootdown operation to free up any resources
254 * that were required.
256 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
258 struct mmu_gather_batch *batch, *next;
264 /* keep the page table cache within bounds */
267 for (batch = tlb->local.next; batch; batch = next) {
269 free_pages((unsigned long)batch, 0);
271 tlb->local.next = NULL;
275 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
276 * handling the additional races in SMP caused by other CPUs caching valid
277 * mappings in their TLBs. Returns the number of free page slots left.
278 * When out of page slots we must call tlb_flush_mmu().
280 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
282 struct mmu_gather_batch *batch;
284 VM_BUG_ON(!tlb->need_flush);
286 if (tlb_fast_mode(tlb)) {
287 free_page_and_swap_cache(page);
288 return 1; /* avoid calling tlb_flush_mmu() */
292 batch->pages[batch->nr++] = page;
293 if (batch->nr == batch->max) {
294 if (!tlb_next_batch(tlb))
298 VM_BUG_ON(batch->nr > batch->max);
300 return batch->max - batch->nr;
303 #endif /* HAVE_GENERIC_MMU_GATHER */
305 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
308 * See the comment near struct mmu_table_batch.
311 static void tlb_remove_table_smp_sync(void *arg)
313 /* Simply deliver the interrupt */
316 static void tlb_remove_table_one(void *table)
319 * This isn't an RCU grace period and hence the page-tables cannot be
320 * assumed to be actually RCU-freed.
322 * It is however sufficient for software page-table walkers that rely on
323 * IRQ disabling. See the comment near struct mmu_table_batch.
325 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
326 __tlb_remove_table(table);
329 static void tlb_remove_table_rcu(struct rcu_head *head)
331 struct mmu_table_batch *batch;
334 batch = container_of(head, struct mmu_table_batch, rcu);
336 for (i = 0; i < batch->nr; i++)
337 __tlb_remove_table(batch->tables[i]);
339 free_page((unsigned long)batch);
342 void tlb_table_flush(struct mmu_gather *tlb)
344 struct mmu_table_batch **batch = &tlb->batch;
347 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 void tlb_remove_table(struct mmu_gather *tlb, void *table)
354 struct mmu_table_batch **batch = &tlb->batch;
359 * When there's less then two users of this mm there cannot be a
360 * concurrent page-table walk.
362 if (atomic_read(&tlb->mm->mm_users) < 2) {
363 __tlb_remove_table(table);
367 if (*batch == NULL) {
368 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
369 if (*batch == NULL) {
370 tlb_remove_table_one(table);
375 (*batch)->tables[(*batch)->nr++] = table;
376 if ((*batch)->nr == MAX_TABLE_BATCH)
377 tlb_table_flush(tlb);
380 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
383 * If a p?d_bad entry is found while walking page tables, report
384 * the error, before resetting entry to p?d_none. Usually (but
385 * very seldom) called out from the p?d_none_or_clear_bad macros.
388 void pgd_clear_bad(pgd_t *pgd)
394 void pud_clear_bad(pud_t *pud)
400 void pmd_clear_bad(pmd_t *pmd)
407 * Note: this doesn't free the actual pages themselves. That
408 * has been handled earlier when unmapping all the memory regions.
410 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
413 pgtable_t token = pmd_pgtable(*pmd);
415 pte_free_tlb(tlb, token, addr);
419 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
420 unsigned long addr, unsigned long end,
421 unsigned long floor, unsigned long ceiling)
428 pmd = pmd_offset(pud, addr);
430 next = pmd_addr_end(addr, end);
431 if (pmd_none_or_clear_bad(pmd))
433 free_pte_range(tlb, pmd, addr);
434 } while (pmd++, addr = next, addr != end);
444 if (end - 1 > ceiling - 1)
447 pmd = pmd_offset(pud, start);
449 pmd_free_tlb(tlb, pmd, start);
452 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
453 unsigned long addr, unsigned long end,
454 unsigned long floor, unsigned long ceiling)
461 pud = pud_offset(pgd, addr);
463 next = pud_addr_end(addr, end);
464 if (pud_none_or_clear_bad(pud))
466 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
467 } while (pud++, addr = next, addr != end);
473 ceiling &= PGDIR_MASK;
477 if (end - 1 > ceiling - 1)
480 pud = pud_offset(pgd, start);
482 pud_free_tlb(tlb, pud, start);
486 * This function frees user-level page tables of a process.
488 * Must be called with pagetable lock held.
490 void free_pgd_range(struct mmu_gather *tlb,
491 unsigned long addr, unsigned long end,
492 unsigned long floor, unsigned long ceiling)
498 * The next few lines have given us lots of grief...
500 * Why are we testing PMD* at this top level? Because often
501 * there will be no work to do at all, and we'd prefer not to
502 * go all the way down to the bottom just to discover that.
504 * Why all these "- 1"s? Because 0 represents both the bottom
505 * of the address space and the top of it (using -1 for the
506 * top wouldn't help much: the masks would do the wrong thing).
507 * The rule is that addr 0 and floor 0 refer to the bottom of
508 * the address space, but end 0 and ceiling 0 refer to the top
509 * Comparisons need to use "end - 1" and "ceiling - 1" (though
510 * that end 0 case should be mythical).
512 * Wherever addr is brought up or ceiling brought down, we must
513 * be careful to reject "the opposite 0" before it confuses the
514 * subsequent tests. But what about where end is brought down
515 * by PMD_SIZE below? no, end can't go down to 0 there.
517 * Whereas we round start (addr) and ceiling down, by different
518 * masks at different levels, in order to test whether a table
519 * now has no other vmas using it, so can be freed, we don't
520 * bother to round floor or end up - the tests don't need that.
534 if (end - 1 > ceiling - 1)
539 pgd = pgd_offset(tlb->mm, addr);
541 next = pgd_addr_end(addr, end);
542 if (pgd_none_or_clear_bad(pgd))
544 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
545 } while (pgd++, addr = next, addr != end);
548 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
549 unsigned long floor, unsigned long ceiling)
552 struct vm_area_struct *next = vma->vm_next;
553 unsigned long addr = vma->vm_start;
556 * Hide vma from rmap and truncate_pagecache before freeing
559 unlink_anon_vmas(vma);
560 unlink_file_vma(vma);
562 if (is_vm_hugetlb_page(vma)) {
563 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
564 floor, next? next->vm_start: ceiling);
567 * Optimization: gather nearby vmas into one call down
569 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
570 && !is_vm_hugetlb_page(next)) {
573 unlink_anon_vmas(vma);
574 unlink_file_vma(vma);
576 free_pgd_range(tlb, addr, vma->vm_end,
577 floor, next? next->vm_start: ceiling);
583 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
584 pmd_t *pmd, unsigned long address)
586 pgtable_t new = pte_alloc_one(mm, address);
587 int wait_split_huge_page;
592 * Ensure all pte setup (eg. pte page lock and page clearing) are
593 * visible before the pte is made visible to other CPUs by being
594 * put into page tables.
596 * The other side of the story is the pointer chasing in the page
597 * table walking code (when walking the page table without locking;
598 * ie. most of the time). Fortunately, these data accesses consist
599 * of a chain of data-dependent loads, meaning most CPUs (alpha
600 * being the notable exception) will already guarantee loads are
601 * seen in-order. See the alpha page table accessors for the
602 * smp_read_barrier_depends() barriers in page table walking code.
604 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
606 spin_lock(&mm->page_table_lock);
607 wait_split_huge_page = 0;
608 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
610 pmd_populate(mm, pmd, new);
612 } else if (unlikely(pmd_trans_splitting(*pmd)))
613 wait_split_huge_page = 1;
614 spin_unlock(&mm->page_table_lock);
617 if (wait_split_huge_page)
618 wait_split_huge_page(vma->anon_vma, pmd);
622 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
624 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
628 smp_wmb(); /* See comment in __pte_alloc */
630 spin_lock(&init_mm.page_table_lock);
631 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
632 pmd_populate_kernel(&init_mm, pmd, new);
635 VM_BUG_ON(pmd_trans_splitting(*pmd));
636 spin_unlock(&init_mm.page_table_lock);
638 pte_free_kernel(&init_mm, new);
642 static inline void init_rss_vec(int *rss)
644 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
647 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
651 if (current->mm == mm)
653 for (i = 0; i < NR_MM_COUNTERS; i++)
655 add_mm_counter(mm, i, rss[i]);
659 * This function is called to print an error when a bad pte
660 * is found. For example, we might have a PFN-mapped pte in
661 * a region that doesn't allow it.
663 * The calling function must still handle the error.
665 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
666 pte_t pte, struct page *page)
668 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
669 pud_t *pud = pud_offset(pgd, addr);
670 pmd_t *pmd = pmd_offset(pud, addr);
671 struct address_space *mapping;
673 static unsigned long resume;
674 static unsigned long nr_shown;
675 static unsigned long nr_unshown;
678 * Allow a burst of 60 reports, then keep quiet for that minute;
679 * or allow a steady drip of one report per second.
681 if (nr_shown == 60) {
682 if (time_before(jiffies, resume)) {
688 "BUG: Bad page map: %lu messages suppressed\n",
695 resume = jiffies + 60 * HZ;
697 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
698 index = linear_page_index(vma, addr);
701 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
703 (long long)pte_val(pte), (long long)pmd_val(*pmd));
707 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
708 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
710 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
713 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
714 (unsigned long)vma->vm_ops->fault);
715 if (vma->vm_file && vma->vm_file->f_op)
716 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
717 (unsigned long)vma->vm_file->f_op->mmap);
719 add_taint(TAINT_BAD_PAGE);
722 static inline bool is_cow_mapping(vm_flags_t flags)
724 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
728 static inline int is_zero_pfn(unsigned long pfn)
730 return pfn == zero_pfn;
735 static inline unsigned long my_zero_pfn(unsigned long addr)
742 * vm_normal_page -- This function gets the "struct page" associated with a pte.
744 * "Special" mappings do not wish to be associated with a "struct page" (either
745 * it doesn't exist, or it exists but they don't want to touch it). In this
746 * case, NULL is returned here. "Normal" mappings do have a struct page.
748 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
749 * pte bit, in which case this function is trivial. Secondly, an architecture
750 * may not have a spare pte bit, which requires a more complicated scheme,
753 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
754 * special mapping (even if there are underlying and valid "struct pages").
755 * COWed pages of a VM_PFNMAP are always normal.
757 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
758 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
759 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
760 * mapping will always honor the rule
762 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
764 * And for normal mappings this is false.
766 * This restricts such mappings to be a linear translation from virtual address
767 * to pfn. To get around this restriction, we allow arbitrary mappings so long
768 * as the vma is not a COW mapping; in that case, we know that all ptes are
769 * special (because none can have been COWed).
772 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
774 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
775 * page" backing, however the difference is that _all_ pages with a struct
776 * page (that is, those where pfn_valid is true) are refcounted and considered
777 * normal pages by the VM. The disadvantage is that pages are refcounted
778 * (which can be slower and simply not an option for some PFNMAP users). The
779 * advantage is that we don't have to follow the strict linearity rule of
780 * PFNMAP mappings in order to support COWable mappings.
783 #ifdef __HAVE_ARCH_PTE_SPECIAL
784 # define HAVE_PTE_SPECIAL 1
786 # define HAVE_PTE_SPECIAL 0
788 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
791 unsigned long pfn = pte_pfn(pte);
793 if (HAVE_PTE_SPECIAL) {
794 if (likely(!pte_special(pte)))
796 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
798 if (!is_zero_pfn(pfn))
799 print_bad_pte(vma, addr, pte, NULL);
803 /* !HAVE_PTE_SPECIAL case follows: */
805 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
806 if (vma->vm_flags & VM_MIXEDMAP) {
812 off = (addr - vma->vm_start) >> PAGE_SHIFT;
813 if (pfn == vma->vm_pgoff + off)
815 if (!is_cow_mapping(vma->vm_flags))
820 if (is_zero_pfn(pfn))
823 if (unlikely(pfn > highest_memmap_pfn)) {
824 print_bad_pte(vma, addr, pte, NULL);
829 * NOTE! We still have PageReserved() pages in the page tables.
830 * eg. VDSO mappings can cause them to exist.
833 return pfn_to_page(pfn);
837 * copy one vm_area from one task to the other. Assumes the page tables
838 * already present in the new task to be cleared in the whole range
839 * covered by this vma.
842 static inline unsigned long
843 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
844 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
845 unsigned long addr, int *rss)
847 unsigned long vm_flags = vma->vm_flags;
848 pte_t pte = *src_pte;
851 /* pte contains position in swap or file, so copy. */
852 if (unlikely(!pte_present(pte))) {
853 if (!pte_file(pte)) {
854 swp_entry_t entry = pte_to_swp_entry(pte);
856 if (swap_duplicate(entry) < 0)
859 /* make sure dst_mm is on swapoff's mmlist. */
860 if (unlikely(list_empty(&dst_mm->mmlist))) {
861 spin_lock(&mmlist_lock);
862 if (list_empty(&dst_mm->mmlist))
863 list_add(&dst_mm->mmlist,
865 spin_unlock(&mmlist_lock);
867 if (likely(!non_swap_entry(entry)))
869 else if (is_migration_entry(entry)) {
870 page = migration_entry_to_page(entry);
877 if (is_write_migration_entry(entry) &&
878 is_cow_mapping(vm_flags)) {
880 * COW mappings require pages in both
881 * parent and child to be set to read.
883 make_migration_entry_read(&entry);
884 pte = swp_entry_to_pte(entry);
885 set_pte_at(src_mm, addr, src_pte, pte);
893 * If it's a COW mapping, write protect it both
894 * in the parent and the child
896 if (is_cow_mapping(vm_flags)) {
897 ptep_set_wrprotect(src_mm, addr, src_pte);
898 pte = pte_wrprotect(pte);
902 * If it's a shared mapping, mark it clean in
905 if (vm_flags & VM_SHARED)
906 pte = pte_mkclean(pte);
907 pte = pte_mkold(pte);
909 page = vm_normal_page(vma, addr, pte);
920 set_pte_at(dst_mm, addr, dst_pte, pte);
924 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
925 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
926 unsigned long addr, unsigned long end)
928 pte_t *orig_src_pte, *orig_dst_pte;
929 pte_t *src_pte, *dst_pte;
930 spinlock_t *src_ptl, *dst_ptl;
932 int rss[NR_MM_COUNTERS];
933 swp_entry_t entry = (swp_entry_t){0};
938 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
941 src_pte = pte_offset_map(src_pmd, addr);
942 src_ptl = pte_lockptr(src_mm, src_pmd);
943 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
944 orig_src_pte = src_pte;
945 orig_dst_pte = dst_pte;
946 arch_enter_lazy_mmu_mode();
950 * We are holding two locks at this point - either of them
951 * could generate latencies in another task on another CPU.
953 if (progress >= 32) {
955 if (need_resched() ||
956 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
959 if (pte_none(*src_pte)) {
963 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
968 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
970 arch_leave_lazy_mmu_mode();
971 spin_unlock(src_ptl);
972 pte_unmap(orig_src_pte);
973 add_mm_rss_vec(dst_mm, rss);
974 pte_unmap_unlock(orig_dst_pte, dst_ptl);
978 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
987 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
991 pmd_t *src_pmd, *dst_pmd;
994 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
997 src_pmd = pmd_offset(src_pud, addr);
999 next = pmd_addr_end(addr, end);
1000 if (pmd_trans_huge(*src_pmd)) {
1002 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1003 err = copy_huge_pmd(dst_mm, src_mm,
1004 dst_pmd, src_pmd, addr, vma);
1011 if (pmd_none_or_clear_bad(src_pmd))
1013 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1020 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1021 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1022 unsigned long addr, unsigned long end)
1024 pud_t *src_pud, *dst_pud;
1027 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030 src_pud = pud_offset(src_pgd, addr);
1032 next = pud_addr_end(addr, end);
1033 if (pud_none_or_clear_bad(src_pud))
1035 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038 } while (dst_pud++, src_pud++, addr = next, addr != end);
1042 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1043 struct vm_area_struct *vma)
1045 pgd_t *src_pgd, *dst_pgd;
1047 unsigned long addr = vma->vm_start;
1048 unsigned long end = vma->vm_end;
1049 unsigned long mmun_start; /* For mmu_notifiers */
1050 unsigned long mmun_end; /* For mmu_notifiers */
1055 * Don't copy ptes where a page fault will fill them correctly.
1056 * Fork becomes much lighter when there are big shared or private
1057 * readonly mappings. The tradeoff is that copy_page_range is more
1058 * efficient than faulting.
1060 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1061 VM_PFNMAP | VM_MIXEDMAP))) {
1066 if (is_vm_hugetlb_page(vma))
1067 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1069 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1071 * We do not free on error cases below as remove_vma
1072 * gets called on error from higher level routine
1074 ret = track_pfn_copy(vma);
1080 * We need to invalidate the secondary MMU mappings only when
1081 * there could be a permission downgrade on the ptes of the
1082 * parent mm. And a permission downgrade will only happen if
1083 * is_cow_mapping() returns true.
1085 is_cow = is_cow_mapping(vma->vm_flags);
1089 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1093 dst_pgd = pgd_offset(dst_mm, addr);
1094 src_pgd = pgd_offset(src_mm, addr);
1096 next = pgd_addr_end(addr, end);
1097 if (pgd_none_or_clear_bad(src_pgd))
1099 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100 vma, addr, next))) {
1104 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1107 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1111 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1112 struct vm_area_struct *vma, pmd_t *pmd,
1113 unsigned long addr, unsigned long end,
1114 struct zap_details *details)
1116 struct mm_struct *mm = tlb->mm;
1117 int force_flush = 0;
1118 int rss[NR_MM_COUNTERS];
1125 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1127 arch_enter_lazy_mmu_mode();
1130 if (pte_none(ptent)) {
1134 if (pte_present(ptent)) {
1137 page = vm_normal_page(vma, addr, ptent);
1138 if (unlikely(details) && page) {
1140 * unmap_shared_mapping_pages() wants to
1141 * invalidate cache without truncating:
1142 * unmap shared but keep private pages.
1144 if (details->check_mapping &&
1145 details->check_mapping != page->mapping)
1148 * Each page->index must be checked when
1149 * invalidating or truncating nonlinear.
1151 if (details->nonlinear_vma &&
1152 (page->index < details->first_index ||
1153 page->index > details->last_index))
1156 ptent = ptep_get_and_clear_full(mm, addr, pte,
1158 tlb_remove_tlb_entry(tlb, pte, addr);
1159 if (unlikely(!page))
1161 if (unlikely(details) && details->nonlinear_vma
1162 && linear_page_index(details->nonlinear_vma,
1163 addr) != page->index)
1164 set_pte_at(mm, addr, pte,
1165 pgoff_to_pte(page->index));
1167 rss[MM_ANONPAGES]--;
1169 if (pte_dirty(ptent))
1170 set_page_dirty(page);
1171 if (pte_young(ptent) &&
1172 likely(!VM_SequentialReadHint(vma)))
1173 mark_page_accessed(page);
1174 rss[MM_FILEPAGES]--;
1176 page_remove_rmap(page);
1177 if (unlikely(page_mapcount(page) < 0))
1178 print_bad_pte(vma, addr, ptent, page);
1179 force_flush = !__tlb_remove_page(tlb, page);
1185 * If details->check_mapping, we leave swap entries;
1186 * if details->nonlinear_vma, we leave file entries.
1188 if (unlikely(details))
1190 if (pte_file(ptent)) {
1191 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1192 print_bad_pte(vma, addr, ptent, NULL);
1194 swp_entry_t entry = pte_to_swp_entry(ptent);
1196 if (!non_swap_entry(entry))
1198 else if (is_migration_entry(entry)) {
1201 page = migration_entry_to_page(entry);
1204 rss[MM_ANONPAGES]--;
1206 rss[MM_FILEPAGES]--;
1208 if (unlikely(!free_swap_and_cache(entry)))
1209 print_bad_pte(vma, addr, ptent, NULL);
1211 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1212 } while (pte++, addr += PAGE_SIZE, addr != end);
1214 add_mm_rss_vec(mm, rss);
1215 arch_leave_lazy_mmu_mode();
1216 pte_unmap_unlock(start_pte, ptl);
1219 * mmu_gather ran out of room to batch pages, we break out of
1220 * the PTE lock to avoid doing the potential expensive TLB invalidate
1221 * and page-free while holding it.
1226 #ifdef HAVE_GENERIC_MMU_GATHER
1238 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1239 struct vm_area_struct *vma, pud_t *pud,
1240 unsigned long addr, unsigned long end,
1241 struct zap_details *details)
1246 pmd = pmd_offset(pud, addr);
1248 next = pmd_addr_end(addr, end);
1249 if (pmd_trans_huge(*pmd)) {
1250 if (next - addr != HPAGE_PMD_SIZE) {
1251 #ifdef CONFIG_DEBUG_VM
1252 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1253 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1254 __func__, addr, end,
1260 split_huge_page_pmd(vma->vm_mm, pmd);
1261 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1266 * Here there can be other concurrent MADV_DONTNEED or
1267 * trans huge page faults running, and if the pmd is
1268 * none or trans huge it can change under us. This is
1269 * because MADV_DONTNEED holds the mmap_sem in read
1272 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1274 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1277 } while (pmd++, addr = next, addr != end);
1282 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1283 struct vm_area_struct *vma, pgd_t *pgd,
1284 unsigned long addr, unsigned long end,
1285 struct zap_details *details)
1290 pud = pud_offset(pgd, addr);
1292 next = pud_addr_end(addr, end);
1293 if (pud_none_or_clear_bad(pud))
1295 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1296 } while (pud++, addr = next, addr != end);
1301 static void unmap_page_range(struct mmu_gather *tlb,
1302 struct vm_area_struct *vma,
1303 unsigned long addr, unsigned long end,
1304 struct zap_details *details)
1309 if (details && !details->check_mapping && !details->nonlinear_vma)
1312 BUG_ON(addr >= end);
1313 mem_cgroup_uncharge_start();
1314 tlb_start_vma(tlb, vma);
1315 pgd = pgd_offset(vma->vm_mm, addr);
1317 next = pgd_addr_end(addr, end);
1318 if (pgd_none_or_clear_bad(pgd))
1320 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1321 } while (pgd++, addr = next, addr != end);
1322 tlb_end_vma(tlb, vma);
1323 mem_cgroup_uncharge_end();
1327 static void unmap_single_vma(struct mmu_gather *tlb,
1328 struct vm_area_struct *vma, unsigned long start_addr,
1329 unsigned long end_addr,
1330 struct zap_details *details)
1332 unsigned long start = max(vma->vm_start, start_addr);
1335 if (start >= vma->vm_end)
1337 end = min(vma->vm_end, end_addr);
1338 if (end <= vma->vm_start)
1342 uprobe_munmap(vma, start, end);
1344 if (unlikely(vma->vm_flags & VM_PFNMAP))
1345 untrack_pfn(vma, 0, 0);
1348 if (unlikely(is_vm_hugetlb_page(vma))) {
1350 * It is undesirable to test vma->vm_file as it
1351 * should be non-null for valid hugetlb area.
1352 * However, vm_file will be NULL in the error
1353 * cleanup path of do_mmap_pgoff. When
1354 * hugetlbfs ->mmap method fails,
1355 * do_mmap_pgoff() nullifies vma->vm_file
1356 * before calling this function to clean up.
1357 * Since no pte has actually been setup, it is
1358 * safe to do nothing in this case.
1361 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1362 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1363 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1366 unmap_page_range(tlb, vma, start, end, details);
1371 * unmap_vmas - unmap a range of memory covered by a list of vma's
1372 * @tlb: address of the caller's struct mmu_gather
1373 * @vma: the starting vma
1374 * @start_addr: virtual address at which to start unmapping
1375 * @end_addr: virtual address at which to end unmapping
1377 * Unmap all pages in the vma list.
1379 * Only addresses between `start' and `end' will be unmapped.
1381 * The VMA list must be sorted in ascending virtual address order.
1383 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1384 * range after unmap_vmas() returns. So the only responsibility here is to
1385 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1386 * drops the lock and schedules.
1388 void unmap_vmas(struct mmu_gather *tlb,
1389 struct vm_area_struct *vma, unsigned long start_addr,
1390 unsigned long end_addr)
1392 struct mm_struct *mm = vma->vm_mm;
1394 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1395 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1396 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1397 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1401 * zap_page_range - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @start: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of nonlinear truncation or shared cache invalidation
1407 * Caller must protect the VMA list
1409 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1410 unsigned long size, struct zap_details *details)
1412 struct mm_struct *mm = vma->vm_mm;
1413 struct mmu_gather tlb;
1414 unsigned long end = start + size;
1417 tlb_gather_mmu(&tlb, mm, 0);
1418 update_hiwater_rss(mm);
1419 mmu_notifier_invalidate_range_start(mm, start, end);
1420 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1421 unmap_single_vma(&tlb, vma, start, end, details);
1422 mmu_notifier_invalidate_range_end(mm, start, end);
1423 tlb_finish_mmu(&tlb, start, end);
1427 * zap_page_range_single - remove user pages in a given range
1428 * @vma: vm_area_struct holding the applicable pages
1429 * @address: starting address of pages to zap
1430 * @size: number of bytes to zap
1431 * @details: details of nonlinear truncation or shared cache invalidation
1433 * The range must fit into one VMA.
1435 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1436 unsigned long size, struct zap_details *details)
1438 struct mm_struct *mm = vma->vm_mm;
1439 struct mmu_gather tlb;
1440 unsigned long end = address + size;
1443 tlb_gather_mmu(&tlb, mm, 0);
1444 update_hiwater_rss(mm);
1445 mmu_notifier_invalidate_range_start(mm, address, end);
1446 unmap_single_vma(&tlb, vma, address, end, details);
1447 mmu_notifier_invalidate_range_end(mm, address, end);
1448 tlb_finish_mmu(&tlb, address, end);
1452 * zap_vma_ptes - remove ptes mapping the vma
1453 * @vma: vm_area_struct holding ptes to be zapped
1454 * @address: starting address of pages to zap
1455 * @size: number of bytes to zap
1457 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1459 * The entire address range must be fully contained within the vma.
1461 * Returns 0 if successful.
1463 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1466 if (address < vma->vm_start || address + size > vma->vm_end ||
1467 !(vma->vm_flags & VM_PFNMAP))
1469 zap_page_range_single(vma, address, size, NULL);
1472 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1474 static bool pte_numa(struct vm_area_struct *vma, pte_t pte)
1477 * For NUMA page faults, we use PROT_NONE ptes in VMAs with
1478 * "normal" vma->vm_page_prot protections. Genuine PROT_NONE
1479 * VMAs should never get here, because the fault handling code
1480 * will notice that the VMA has no read or write permissions.
1482 * This means we cannot get 'special' PROT_NONE faults from genuine
1483 * PROT_NONE maps, nor from PROT_WRITE file maps that do dirty
1486 * Neither case is really interesting for our current use though so we
1489 if (pte_same(pte, pte_modify(pte, vma->vm_page_prot)))
1492 return pte_same(pte, pte_modify(pte, vma_prot_none(vma)));
1496 * follow_page - look up a page descriptor from a user-virtual address
1497 * @vma: vm_area_struct mapping @address
1498 * @address: virtual address to look up
1499 * @flags: flags modifying lookup behaviour
1501 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1503 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1504 * an error pointer if there is a mapping to something not represented
1505 * by a page descriptor (see also vm_normal_page()).
1507 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1516 struct mm_struct *mm = vma->vm_mm;
1518 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1519 if (!IS_ERR(page)) {
1520 BUG_ON(flags & FOLL_GET);
1525 pgd = pgd_offset(mm, address);
1526 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1529 pud = pud_offset(pgd, address);
1532 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1533 BUG_ON(flags & FOLL_GET);
1534 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1537 if (unlikely(pud_bad(*pud)))
1540 pmd = pmd_offset(pud, address);
1543 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1544 BUG_ON(flags & FOLL_GET);
1545 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1548 if ((flags & FOLL_NUMA) && pmd_numa(vma, *pmd))
1550 if (pmd_trans_huge(*pmd)) {
1551 if (flags & FOLL_SPLIT) {
1552 split_huge_page_pmd(mm, pmd);
1553 goto split_fallthrough;
1555 spin_lock(&mm->page_table_lock);
1556 if (likely(pmd_trans_huge(*pmd))) {
1557 if (unlikely(pmd_trans_splitting(*pmd))) {
1558 spin_unlock(&mm->page_table_lock);
1559 wait_split_huge_page(vma->anon_vma, pmd);
1561 page = follow_trans_huge_pmd(vma, address,
1563 spin_unlock(&mm->page_table_lock);
1567 spin_unlock(&mm->page_table_lock);
1571 if (unlikely(pmd_bad(*pmd)))
1574 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1577 if (!pte_present(pte))
1579 if ((flags & FOLL_NUMA) && pte_numa(vma, pte))
1581 if ((flags & FOLL_WRITE) && !pte_write(pte))
1584 page = vm_normal_page(vma, address, pte);
1585 if (unlikely(!page)) {
1586 if ((flags & FOLL_DUMP) ||
1587 !is_zero_pfn(pte_pfn(pte)))
1589 page = pte_page(pte);
1592 if (flags & FOLL_GET)
1593 get_page_foll(page);
1594 if (flags & FOLL_TOUCH) {
1595 if ((flags & FOLL_WRITE) &&
1596 !pte_dirty(pte) && !PageDirty(page))
1597 set_page_dirty(page);
1599 * pte_mkyoung() would be more correct here, but atomic care
1600 * is needed to avoid losing the dirty bit: it is easier to use
1601 * mark_page_accessed().
1603 mark_page_accessed(page);
1605 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1607 * The preliminary mapping check is mainly to avoid the
1608 * pointless overhead of lock_page on the ZERO_PAGE
1609 * which might bounce very badly if there is contention.
1611 * If the page is already locked, we don't need to
1612 * handle it now - vmscan will handle it later if and
1613 * when it attempts to reclaim the page.
1615 if (page->mapping && trylock_page(page)) {
1616 lru_add_drain(); /* push cached pages to LRU */
1618 * Because we lock page here, and migration is
1619 * blocked by the pte's page reference, and we
1620 * know the page is still mapped, we don't even
1621 * need to check for file-cache page truncation.
1623 mlock_vma_page(page);
1628 pte_unmap_unlock(ptep, ptl);
1633 pte_unmap_unlock(ptep, ptl);
1634 return ERR_PTR(-EFAULT);
1637 pte_unmap_unlock(ptep, ptl);
1643 * When core dumping an enormous anonymous area that nobody
1644 * has touched so far, we don't want to allocate unnecessary pages or
1645 * page tables. Return error instead of NULL to skip handle_mm_fault,
1646 * then get_dump_page() will return NULL to leave a hole in the dump.
1647 * But we can only make this optimization where a hole would surely
1648 * be zero-filled if handle_mm_fault() actually did handle it.
1650 if ((flags & FOLL_DUMP) &&
1651 (!vma->vm_ops || !vma->vm_ops->fault))
1652 return ERR_PTR(-EFAULT);
1656 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1658 return stack_guard_page_start(vma, addr) ||
1659 stack_guard_page_end(vma, addr+PAGE_SIZE);
1663 * __get_user_pages() - pin user pages in memory
1664 * @tsk: task_struct of target task
1665 * @mm: mm_struct of target mm
1666 * @start: starting user address
1667 * @nr_pages: number of pages from start to pin
1668 * @gup_flags: flags modifying pin behaviour
1669 * @pages: array that receives pointers to the pages pinned.
1670 * Should be at least nr_pages long. Or NULL, if caller
1671 * only intends to ensure the pages are faulted in.
1672 * @vmas: array of pointers to vmas corresponding to each page.
1673 * Or NULL if the caller does not require them.
1674 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1676 * Returns number of pages pinned. This may be fewer than the number
1677 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1678 * were pinned, returns -errno. Each page returned must be released
1679 * with a put_page() call when it is finished with. vmas will only
1680 * remain valid while mmap_sem is held.
1682 * Must be called with mmap_sem held for read or write.
1684 * __get_user_pages walks a process's page tables and takes a reference to
1685 * each struct page that each user address corresponds to at a given
1686 * instant. That is, it takes the page that would be accessed if a user
1687 * thread accesses the given user virtual address at that instant.
1689 * This does not guarantee that the page exists in the user mappings when
1690 * __get_user_pages returns, and there may even be a completely different
1691 * page there in some cases (eg. if mmapped pagecache has been invalidated
1692 * and subsequently re faulted). However it does guarantee that the page
1693 * won't be freed completely. And mostly callers simply care that the page
1694 * contains data that was valid *at some point in time*. Typically, an IO
1695 * or similar operation cannot guarantee anything stronger anyway because
1696 * locks can't be held over the syscall boundary.
1698 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1699 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1700 * appropriate) must be called after the page is finished with, and
1701 * before put_page is called.
1703 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1704 * or mmap_sem contention, and if waiting is needed to pin all pages,
1705 * *@nonblocking will be set to 0.
1707 * In most cases, get_user_pages or get_user_pages_fast should be used
1708 * instead of __get_user_pages. __get_user_pages should be used only if
1709 * you need some special @gup_flags.
1711 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1712 unsigned long start, int nr_pages, unsigned int gup_flags,
1713 struct page **pages, struct vm_area_struct **vmas,
1717 unsigned long vm_flags;
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1725 * Require read or write permissions.
1726 * If FOLL_FORCE is set, we only require the "MAY" flags.
1728 vm_flags = (gup_flags & FOLL_WRITE) ?
1729 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1730 vm_flags &= (gup_flags & FOLL_FORCE) ?
1731 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1734 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1735 * would be called on PROT_NONE ranges. We must never invoke
1736 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1737 * page faults would unprotect the PROT_NONE ranges if
1738 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1739 * bitflag. So to avoid that, don't set FOLL_NUMA if
1740 * FOLL_FORCE is set.
1742 if (!(gup_flags & FOLL_FORCE))
1743 gup_flags |= FOLL_NUMA;
1748 struct vm_area_struct *vma;
1750 vma = find_extend_vma(mm, start);
1751 if (!vma && in_gate_area(mm, start)) {
1752 unsigned long pg = start & PAGE_MASK;
1758 /* user gate pages are read-only */
1759 if (gup_flags & FOLL_WRITE)
1760 return i ? : -EFAULT;
1762 pgd = pgd_offset_k(pg);
1764 pgd = pgd_offset_gate(mm, pg);
1765 BUG_ON(pgd_none(*pgd));
1766 pud = pud_offset(pgd, pg);
1767 BUG_ON(pud_none(*pud));
1768 pmd = pmd_offset(pud, pg);
1770 return i ? : -EFAULT;
1771 VM_BUG_ON(pmd_trans_huge(*pmd));
1772 pte = pte_offset_map(pmd, pg);
1773 if (pte_none(*pte)) {
1775 return i ? : -EFAULT;
1777 vma = get_gate_vma(mm);
1781 page = vm_normal_page(vma, start, *pte);
1783 if (!(gup_flags & FOLL_DUMP) &&
1784 is_zero_pfn(pte_pfn(*pte)))
1785 page = pte_page(*pte);
1788 return i ? : -EFAULT;
1799 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1800 !(vm_flags & vma->vm_flags))
1801 return i ? : -EFAULT;
1803 if (is_vm_hugetlb_page(vma)) {
1804 i = follow_hugetlb_page(mm, vma, pages, vmas,
1805 &start, &nr_pages, i, gup_flags);
1811 unsigned int foll_flags = gup_flags;
1814 * If we have a pending SIGKILL, don't keep faulting
1815 * pages and potentially allocating memory.
1817 if (unlikely(fatal_signal_pending(current)))
1818 return i ? i : -ERESTARTSYS;
1821 while (!(page = follow_page(vma, start, foll_flags))) {
1823 unsigned int fault_flags = 0;
1825 /* For mlock, just skip the stack guard page. */
1826 if (foll_flags & FOLL_MLOCK) {
1827 if (stack_guard_page(vma, start))
1830 if (foll_flags & FOLL_WRITE)
1831 fault_flags |= FAULT_FLAG_WRITE;
1833 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1834 if (foll_flags & FOLL_NOWAIT)
1835 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1837 ret = handle_mm_fault(mm, vma, start,
1840 if (ret & VM_FAULT_ERROR) {
1841 if (ret & VM_FAULT_OOM)
1842 return i ? i : -ENOMEM;
1843 if (ret & (VM_FAULT_HWPOISON |
1844 VM_FAULT_HWPOISON_LARGE)) {
1847 else if (gup_flags & FOLL_HWPOISON)
1852 if (ret & VM_FAULT_SIGBUS)
1853 return i ? i : -EFAULT;
1858 if (ret & VM_FAULT_MAJOR)
1864 if (ret & VM_FAULT_RETRY) {
1871 * The VM_FAULT_WRITE bit tells us that
1872 * do_wp_page has broken COW when necessary,
1873 * even if maybe_mkwrite decided not to set
1874 * pte_write. We can thus safely do subsequent
1875 * page lookups as if they were reads. But only
1876 * do so when looping for pte_write is futile:
1877 * in some cases userspace may also be wanting
1878 * to write to the gotten user page, which a
1879 * read fault here might prevent (a readonly
1880 * page might get reCOWed by userspace write).
1882 if ((ret & VM_FAULT_WRITE) &&
1883 !(vma->vm_flags & VM_WRITE))
1884 foll_flags &= ~FOLL_WRITE;
1889 return i ? i : PTR_ERR(page);
1893 flush_anon_page(vma, page, start);
1894 flush_dcache_page(page);
1902 } while (nr_pages && start < vma->vm_end);
1906 EXPORT_SYMBOL(__get_user_pages);
1909 * fixup_user_fault() - manually resolve a user page fault
1910 * @tsk: the task_struct to use for page fault accounting, or
1911 * NULL if faults are not to be recorded.
1912 * @mm: mm_struct of target mm
1913 * @address: user address
1914 * @fault_flags:flags to pass down to handle_mm_fault()
1916 * This is meant to be called in the specific scenario where for locking reasons
1917 * we try to access user memory in atomic context (within a pagefault_disable()
1918 * section), this returns -EFAULT, and we want to resolve the user fault before
1921 * Typically this is meant to be used by the futex code.
1923 * The main difference with get_user_pages() is that this function will
1924 * unconditionally call handle_mm_fault() which will in turn perform all the
1925 * necessary SW fixup of the dirty and young bits in the PTE, while
1926 * handle_mm_fault() only guarantees to update these in the struct page.
1928 * This is important for some architectures where those bits also gate the
1929 * access permission to the page because they are maintained in software. On
1930 * such architectures, gup() will not be enough to make a subsequent access
1933 * This should be called with the mm_sem held for read.
1935 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1936 unsigned long address, unsigned int fault_flags)
1938 struct vm_area_struct *vma;
1941 vma = find_extend_vma(mm, address);
1942 if (!vma || address < vma->vm_start)
1945 ret = handle_mm_fault(mm, vma, address, fault_flags);
1946 if (ret & VM_FAULT_ERROR) {
1947 if (ret & VM_FAULT_OOM)
1949 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1951 if (ret & VM_FAULT_SIGBUS)
1956 if (ret & VM_FAULT_MAJOR)
1965 * get_user_pages() - pin user pages in memory
1966 * @tsk: the task_struct to use for page fault accounting, or
1967 * NULL if faults are not to be recorded.
1968 * @mm: mm_struct of target mm
1969 * @start: starting user address
1970 * @nr_pages: number of pages from start to pin
1971 * @write: whether pages will be written to by the caller
1972 * @force: whether to force write access even if user mapping is
1973 * readonly. This will result in the page being COWed even
1974 * in MAP_SHARED mappings. You do not want this.
1975 * @pages: array that receives pointers to the pages pinned.
1976 * Should be at least nr_pages long. Or NULL, if caller
1977 * only intends to ensure the pages are faulted in.
1978 * @vmas: array of pointers to vmas corresponding to each page.
1979 * Or NULL if the caller does not require them.
1981 * Returns number of pages pinned. This may be fewer than the number
1982 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1983 * were pinned, returns -errno. Each page returned must be released
1984 * with a put_page() call when it is finished with. vmas will only
1985 * remain valid while mmap_sem is held.
1987 * Must be called with mmap_sem held for read or write.
1989 * get_user_pages walks a process's page tables and takes a reference to
1990 * each struct page that each user address corresponds to at a given
1991 * instant. That is, it takes the page that would be accessed if a user
1992 * thread accesses the given user virtual address at that instant.
1994 * This does not guarantee that the page exists in the user mappings when
1995 * get_user_pages returns, and there may even be a completely different
1996 * page there in some cases (eg. if mmapped pagecache has been invalidated
1997 * and subsequently re faulted). However it does guarantee that the page
1998 * won't be freed completely. And mostly callers simply care that the page
1999 * contains data that was valid *at some point in time*. Typically, an IO
2000 * or similar operation cannot guarantee anything stronger anyway because
2001 * locks can't be held over the syscall boundary.
2003 * If write=0, the page must not be written to. If the page is written to,
2004 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2005 * after the page is finished with, and before put_page is called.
2007 * get_user_pages is typically used for fewer-copy IO operations, to get a
2008 * handle on the memory by some means other than accesses via the user virtual
2009 * addresses. The pages may be submitted for DMA to devices or accessed via
2010 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2011 * use the correct cache flushing APIs.
2013 * See also get_user_pages_fast, for performance critical applications.
2015 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2016 unsigned long start, int nr_pages, int write, int force,
2017 struct page **pages, struct vm_area_struct **vmas)
2019 int flags = FOLL_TOUCH;
2024 flags |= FOLL_WRITE;
2026 flags |= FOLL_FORCE;
2028 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2031 EXPORT_SYMBOL(get_user_pages);
2034 * get_dump_page() - pin user page in memory while writing it to core dump
2035 * @addr: user address
2037 * Returns struct page pointer of user page pinned for dump,
2038 * to be freed afterwards by page_cache_release() or put_page().
2040 * Returns NULL on any kind of failure - a hole must then be inserted into
2041 * the corefile, to preserve alignment with its headers; and also returns
2042 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2043 * allowing a hole to be left in the corefile to save diskspace.
2045 * Called without mmap_sem, but after all other threads have been killed.
2047 #ifdef CONFIG_ELF_CORE
2048 struct page *get_dump_page(unsigned long addr)
2050 struct vm_area_struct *vma;
2053 if (__get_user_pages(current, current->mm, addr, 1,
2054 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2057 flush_cache_page(vma, addr, page_to_pfn(page));
2060 #endif /* CONFIG_ELF_CORE */
2062 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2065 pgd_t * pgd = pgd_offset(mm, addr);
2066 pud_t * pud = pud_alloc(mm, pgd, addr);
2068 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2070 VM_BUG_ON(pmd_trans_huge(*pmd));
2071 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2078 * This is the old fallback for page remapping.
2080 * For historical reasons, it only allows reserved pages. Only
2081 * old drivers should use this, and they needed to mark their
2082 * pages reserved for the old functions anyway.
2084 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2085 struct page *page, pgprot_t prot)
2087 struct mm_struct *mm = vma->vm_mm;
2096 flush_dcache_page(page);
2097 pte = get_locked_pte(mm, addr, &ptl);
2101 if (!pte_none(*pte))
2104 /* Ok, finally just insert the thing.. */
2106 inc_mm_counter_fast(mm, MM_FILEPAGES);
2107 page_add_file_rmap(page);
2108 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2111 pte_unmap_unlock(pte, ptl);
2114 pte_unmap_unlock(pte, ptl);
2120 * vm_insert_page - insert single page into user vma
2121 * @vma: user vma to map to
2122 * @addr: target user address of this page
2123 * @page: source kernel page
2125 * This allows drivers to insert individual pages they've allocated
2128 * The page has to be a nice clean _individual_ kernel allocation.
2129 * If you allocate a compound page, you need to have marked it as
2130 * such (__GFP_COMP), or manually just split the page up yourself
2131 * (see split_page()).
2133 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2134 * took an arbitrary page protection parameter. This doesn't allow
2135 * that. Your vma protection will have to be set up correctly, which
2136 * means that if you want a shared writable mapping, you'd better
2137 * ask for a shared writable mapping!
2139 * The page does not need to be reserved.
2141 * Usually this function is called from f_op->mmap() handler
2142 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2143 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2144 * function from other places, for example from page-fault handler.
2146 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2149 if (addr < vma->vm_start || addr >= vma->vm_end)
2151 if (!page_count(page))
2153 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2154 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2155 BUG_ON(vma->vm_flags & VM_PFNMAP);
2156 vma->vm_flags |= VM_MIXEDMAP;
2158 return insert_page(vma, addr, page, vma->vm_page_prot);
2160 EXPORT_SYMBOL(vm_insert_page);
2162 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2163 unsigned long pfn, pgprot_t prot)
2165 struct mm_struct *mm = vma->vm_mm;
2171 pte = get_locked_pte(mm, addr, &ptl);
2175 if (!pte_none(*pte))
2178 /* Ok, finally just insert the thing.. */
2179 entry = pte_mkspecial(pfn_pte(pfn, prot));
2180 set_pte_at(mm, addr, pte, entry);
2181 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2185 pte_unmap_unlock(pte, ptl);
2191 * vm_insert_pfn - insert single pfn into user vma
2192 * @vma: user vma to map to
2193 * @addr: target user address of this page
2194 * @pfn: source kernel pfn
2196 * Similar to vm_insert_page, this allows drivers to insert individual pages
2197 * they've allocated into a user vma. Same comments apply.
2199 * This function should only be called from a vm_ops->fault handler, and
2200 * in that case the handler should return NULL.
2202 * vma cannot be a COW mapping.
2204 * As this is called only for pages that do not currently exist, we
2205 * do not need to flush old virtual caches or the TLB.
2207 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2211 pgprot_t pgprot = vma->vm_page_prot;
2213 * Technically, architectures with pte_special can avoid all these
2214 * restrictions (same for remap_pfn_range). However we would like
2215 * consistency in testing and feature parity among all, so we should
2216 * try to keep these invariants in place for everybody.
2218 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2219 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2220 (VM_PFNMAP|VM_MIXEDMAP));
2221 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2222 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2224 if (addr < vma->vm_start || addr >= vma->vm_end)
2226 if (track_pfn_insert(vma, &pgprot, pfn))
2229 ret = insert_pfn(vma, addr, pfn, pgprot);
2233 EXPORT_SYMBOL(vm_insert_pfn);
2235 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2238 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2240 if (addr < vma->vm_start || addr >= vma->vm_end)
2244 * If we don't have pte special, then we have to use the pfn_valid()
2245 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2246 * refcount the page if pfn_valid is true (hence insert_page rather
2247 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2248 * without pte special, it would there be refcounted as a normal page.
2250 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2253 page = pfn_to_page(pfn);
2254 return insert_page(vma, addr, page, vma->vm_page_prot);
2256 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2258 EXPORT_SYMBOL(vm_insert_mixed);
2261 * maps a range of physical memory into the requested pages. the old
2262 * mappings are removed. any references to nonexistent pages results
2263 * in null mappings (currently treated as "copy-on-access")
2265 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2266 unsigned long addr, unsigned long end,
2267 unsigned long pfn, pgprot_t prot)
2272 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2275 arch_enter_lazy_mmu_mode();
2277 BUG_ON(!pte_none(*pte));
2278 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2280 } while (pte++, addr += PAGE_SIZE, addr != end);
2281 arch_leave_lazy_mmu_mode();
2282 pte_unmap_unlock(pte - 1, ptl);
2286 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2287 unsigned long addr, unsigned long end,
2288 unsigned long pfn, pgprot_t prot)
2293 pfn -= addr >> PAGE_SHIFT;
2294 pmd = pmd_alloc(mm, pud, addr);
2297 VM_BUG_ON(pmd_trans_huge(*pmd));
2299 next = pmd_addr_end(addr, end);
2300 if (remap_pte_range(mm, pmd, addr, next,
2301 pfn + (addr >> PAGE_SHIFT), prot))
2303 } while (pmd++, addr = next, addr != end);
2307 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2308 unsigned long addr, unsigned long end,
2309 unsigned long pfn, pgprot_t prot)
2314 pfn -= addr >> PAGE_SHIFT;
2315 pud = pud_alloc(mm, pgd, addr);
2319 next = pud_addr_end(addr, end);
2320 if (remap_pmd_range(mm, pud, addr, next,
2321 pfn + (addr >> PAGE_SHIFT), prot))
2323 } while (pud++, addr = next, addr != end);
2328 * remap_pfn_range - remap kernel memory to userspace
2329 * @vma: user vma to map to
2330 * @addr: target user address to start at
2331 * @pfn: physical address of kernel memory
2332 * @size: size of map area
2333 * @prot: page protection flags for this mapping
2335 * Note: this is only safe if the mm semaphore is held when called.
2337 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2338 unsigned long pfn, unsigned long size, pgprot_t prot)
2342 unsigned long end = addr + PAGE_ALIGN(size);
2343 struct mm_struct *mm = vma->vm_mm;
2347 * Physically remapped pages are special. Tell the
2348 * rest of the world about it:
2349 * VM_IO tells people not to look at these pages
2350 * (accesses can have side effects).
2351 * VM_PFNMAP tells the core MM that the base pages are just
2352 * raw PFN mappings, and do not have a "struct page" associated
2355 * Disable vma merging and expanding with mremap().
2357 * Omit vma from core dump, even when VM_IO turned off.
2359 * There's a horrible special case to handle copy-on-write
2360 * behaviour that some programs depend on. We mark the "original"
2361 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2362 * See vm_normal_page() for details.
2364 if (is_cow_mapping(vma->vm_flags)) {
2365 if (addr != vma->vm_start || end != vma->vm_end)
2367 vma->vm_pgoff = pfn;
2370 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2374 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2376 BUG_ON(addr >= end);
2377 pfn -= addr >> PAGE_SHIFT;
2378 pgd = pgd_offset(mm, addr);
2379 flush_cache_range(vma, addr, end);
2381 next = pgd_addr_end(addr, end);
2382 err = remap_pud_range(mm, pgd, addr, next,
2383 pfn + (addr >> PAGE_SHIFT), prot);
2386 } while (pgd++, addr = next, addr != end);
2389 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2393 EXPORT_SYMBOL(remap_pfn_range);
2395 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2396 unsigned long addr, unsigned long end,
2397 pte_fn_t fn, void *data)
2402 spinlock_t *uninitialized_var(ptl);
2404 pte = (mm == &init_mm) ?
2405 pte_alloc_kernel(pmd, addr) :
2406 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2410 BUG_ON(pmd_huge(*pmd));
2412 arch_enter_lazy_mmu_mode();
2414 token = pmd_pgtable(*pmd);
2417 err = fn(pte++, token, addr, data);
2420 } while (addr += PAGE_SIZE, addr != end);
2422 arch_leave_lazy_mmu_mode();
2425 pte_unmap_unlock(pte-1, ptl);
2429 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2430 unsigned long addr, unsigned long end,
2431 pte_fn_t fn, void *data)
2437 BUG_ON(pud_huge(*pud));
2439 pmd = pmd_alloc(mm, pud, addr);
2443 next = pmd_addr_end(addr, end);
2444 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2447 } while (pmd++, addr = next, addr != end);
2451 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2452 unsigned long addr, unsigned long end,
2453 pte_fn_t fn, void *data)
2459 pud = pud_alloc(mm, pgd, addr);
2463 next = pud_addr_end(addr, end);
2464 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2467 } while (pud++, addr = next, addr != end);
2472 * Scan a region of virtual memory, filling in page tables as necessary
2473 * and calling a provided function on each leaf page table.
2475 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2476 unsigned long size, pte_fn_t fn, void *data)
2480 unsigned long end = addr + size;
2483 BUG_ON(addr >= end);
2484 pgd = pgd_offset(mm, addr);
2486 next = pgd_addr_end(addr, end);
2487 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2490 } while (pgd++, addr = next, addr != end);
2494 EXPORT_SYMBOL_GPL(apply_to_page_range);
2497 * handle_pte_fault chooses page fault handler according to an entry
2498 * which was read non-atomically. Before making any commitment, on
2499 * those architectures or configurations (e.g. i386 with PAE) which
2500 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2501 * must check under lock before unmapping the pte and proceeding
2502 * (but do_wp_page is only called after already making such a check;
2503 * and do_anonymous_page can safely check later on).
2505 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2506 pte_t *page_table, pte_t orig_pte)
2509 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2510 if (sizeof(pte_t) > sizeof(unsigned long)) {
2511 spinlock_t *ptl = pte_lockptr(mm, pmd);
2513 same = pte_same(*page_table, orig_pte);
2517 pte_unmap(page_table);
2521 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2524 * If the source page was a PFN mapping, we don't have
2525 * a "struct page" for it. We do a best-effort copy by
2526 * just copying from the original user address. If that
2527 * fails, we just zero-fill it. Live with it.
2529 if (unlikely(!src)) {
2530 void *kaddr = kmap_atomic(dst);
2531 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2534 * This really shouldn't fail, because the page is there
2535 * in the page tables. But it might just be unreadable,
2536 * in which case we just give up and fill the result with
2539 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2541 kunmap_atomic(kaddr);
2542 flush_dcache_page(dst);
2544 copy_user_highpage(dst, src, va, vma);
2548 * This routine handles present pages, when users try to write
2549 * to a shared page. It is done by copying the page to a new address
2550 * and decrementing the shared-page counter for the old page.
2552 * Note that this routine assumes that the protection checks have been
2553 * done by the caller (the low-level page fault routine in most cases).
2554 * Thus we can safely just mark it writable once we've done any necessary
2557 * We also mark the page dirty at this point even though the page will
2558 * change only once the write actually happens. This avoids a few races,
2559 * and potentially makes it more efficient.
2561 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2562 * but allow concurrent faults), with pte both mapped and locked.
2563 * We return with mmap_sem still held, but pte unmapped and unlocked.
2565 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2566 unsigned long address, pte_t *page_table, pmd_t *pmd,
2567 spinlock_t *ptl, pte_t orig_pte)
2570 struct page *old_page, *new_page = NULL;
2573 int page_mkwrite = 0;
2574 struct page *dirty_page = NULL;
2575 unsigned long mmun_start = 0; /* For mmu_notifiers */
2576 unsigned long mmun_end = 0; /* For mmu_notifiers */
2578 old_page = vm_normal_page(vma, address, orig_pte);
2581 * VM_MIXEDMAP !pfn_valid() case
2583 * We should not cow pages in a shared writeable mapping.
2584 * Just mark the pages writable as we can't do any dirty
2585 * accounting on raw pfn maps.
2587 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2588 (VM_WRITE|VM_SHARED))
2594 * Take out anonymous pages first, anonymous shared vmas are
2595 * not dirty accountable.
2597 if (PageAnon(old_page) && !PageKsm(old_page)) {
2598 if (!trylock_page(old_page)) {
2599 page_cache_get(old_page);
2600 pte_unmap_unlock(page_table, ptl);
2601 lock_page(old_page);
2602 page_table = pte_offset_map_lock(mm, pmd, address,
2604 if (!pte_same(*page_table, orig_pte)) {
2605 unlock_page(old_page);
2608 page_cache_release(old_page);
2610 if (reuse_swap_page(old_page)) {
2612 * The page is all ours. Move it to our anon_vma so
2613 * the rmap code will not search our parent or siblings.
2614 * Protected against the rmap code by the page lock.
2616 page_move_anon_rmap(old_page, vma, address);
2617 unlock_page(old_page);
2620 unlock_page(old_page);
2621 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2622 (VM_WRITE|VM_SHARED))) {
2624 * Only catch write-faults on shared writable pages,
2625 * read-only shared pages can get COWed by
2626 * get_user_pages(.write=1, .force=1).
2628 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2629 struct vm_fault vmf;
2632 vmf.virtual_address = (void __user *)(address &
2634 vmf.pgoff = old_page->index;
2635 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2636 vmf.page = old_page;
2639 * Notify the address space that the page is about to
2640 * become writable so that it can prohibit this or wait
2641 * for the page to get into an appropriate state.
2643 * We do this without the lock held, so that it can
2644 * sleep if it needs to.
2646 page_cache_get(old_page);
2647 pte_unmap_unlock(page_table, ptl);
2649 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2651 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2653 goto unwritable_page;
2655 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2656 lock_page(old_page);
2657 if (!old_page->mapping) {
2658 ret = 0; /* retry the fault */
2659 unlock_page(old_page);
2660 goto unwritable_page;
2663 VM_BUG_ON(!PageLocked(old_page));
2666 * Since we dropped the lock we need to revalidate
2667 * the PTE as someone else may have changed it. If
2668 * they did, we just return, as we can count on the
2669 * MMU to tell us if they didn't also make it writable.
2671 page_table = pte_offset_map_lock(mm, pmd, address,
2673 if (!pte_same(*page_table, orig_pte)) {
2674 unlock_page(old_page);
2680 dirty_page = old_page;
2681 get_page(dirty_page);
2684 flush_cache_page(vma, address, pte_pfn(orig_pte));
2685 entry = pte_mkyoung(orig_pte);
2686 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2687 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2688 update_mmu_cache(vma, address, page_table);
2689 pte_unmap_unlock(page_table, ptl);
2690 ret |= VM_FAULT_WRITE;
2696 * Yes, Virginia, this is actually required to prevent a race
2697 * with clear_page_dirty_for_io() from clearing the page dirty
2698 * bit after it clear all dirty ptes, but before a racing
2699 * do_wp_page installs a dirty pte.
2701 * __do_fault is protected similarly.
2703 if (!page_mkwrite) {
2704 wait_on_page_locked(dirty_page);
2705 set_page_dirty_balance(dirty_page, page_mkwrite);
2706 /* file_update_time outside page_lock */
2708 file_update_time(vma->vm_file);
2710 put_page(dirty_page);
2712 struct address_space *mapping = dirty_page->mapping;
2714 set_page_dirty(dirty_page);
2715 unlock_page(dirty_page);
2716 page_cache_release(dirty_page);
2719 * Some device drivers do not set page.mapping
2720 * but still dirty their pages
2722 balance_dirty_pages_ratelimited(mapping);
2730 * Ok, we need to copy. Oh, well..
2732 page_cache_get(old_page);
2734 pte_unmap_unlock(page_table, ptl);
2736 if (unlikely(anon_vma_prepare(vma)))
2739 if (is_zero_pfn(pte_pfn(orig_pte))) {
2740 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2744 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2747 cow_user_page(new_page, old_page, address, vma);
2749 __SetPageUptodate(new_page);
2751 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2754 mmun_start = address & PAGE_MASK;
2755 mmun_end = mmun_start + PAGE_SIZE;
2756 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2759 * Re-check the pte - we dropped the lock
2761 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2762 if (likely(pte_same(*page_table, orig_pte))) {
2764 if (!PageAnon(old_page)) {
2765 dec_mm_counter_fast(mm, MM_FILEPAGES);
2766 inc_mm_counter_fast(mm, MM_ANONPAGES);
2769 inc_mm_counter_fast(mm, MM_ANONPAGES);
2770 flush_cache_page(vma, address, pte_pfn(orig_pte));
2771 entry = mk_pte(new_page, vma->vm_page_prot);
2772 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2774 * Clear the pte entry and flush it first, before updating the
2775 * pte with the new entry. This will avoid a race condition
2776 * seen in the presence of one thread doing SMC and another
2779 ptep_clear_flush(vma, address, page_table);
2780 page_add_new_anon_rmap(new_page, vma, address);
2782 * We call the notify macro here because, when using secondary
2783 * mmu page tables (such as kvm shadow page tables), we want the
2784 * new page to be mapped directly into the secondary page table.
2786 set_pte_at_notify(mm, address, page_table, entry);
2787 update_mmu_cache(vma, address, page_table);
2790 * Only after switching the pte to the new page may
2791 * we remove the mapcount here. Otherwise another
2792 * process may come and find the rmap count decremented
2793 * before the pte is switched to the new page, and
2794 * "reuse" the old page writing into it while our pte
2795 * here still points into it and can be read by other
2798 * The critical issue is to order this
2799 * page_remove_rmap with the ptp_clear_flush above.
2800 * Those stores are ordered by (if nothing else,)
2801 * the barrier present in the atomic_add_negative
2802 * in page_remove_rmap.
2804 * Then the TLB flush in ptep_clear_flush ensures that
2805 * no process can access the old page before the
2806 * decremented mapcount is visible. And the old page
2807 * cannot be reused until after the decremented
2808 * mapcount is visible. So transitively, TLBs to
2809 * old page will be flushed before it can be reused.
2811 page_remove_rmap(old_page);
2814 /* Free the old page.. */
2815 new_page = old_page;
2816 ret |= VM_FAULT_WRITE;
2818 mem_cgroup_uncharge_page(new_page);
2821 page_cache_release(new_page);
2823 pte_unmap_unlock(page_table, ptl);
2824 if (mmun_end > mmun_start)
2825 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2828 * Don't let another task, with possibly unlocked vma,
2829 * keep the mlocked page.
2831 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2832 lock_page(old_page); /* LRU manipulation */
2833 munlock_vma_page(old_page);
2834 unlock_page(old_page);
2836 page_cache_release(old_page);
2840 page_cache_release(new_page);
2844 unlock_page(old_page);
2845 page_cache_release(old_page);
2847 page_cache_release(old_page);
2849 return VM_FAULT_OOM;
2852 page_cache_release(old_page);
2856 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2857 unsigned long start_addr, unsigned long end_addr,
2858 struct zap_details *details)
2860 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2863 static inline void unmap_mapping_range_tree(struct rb_root *root,
2864 struct zap_details *details)
2866 struct vm_area_struct *vma;
2867 pgoff_t vba, vea, zba, zea;
2869 vma_interval_tree_foreach(vma, root,
2870 details->first_index, details->last_index) {
2872 vba = vma->vm_pgoff;
2873 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2874 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2875 zba = details->first_index;
2878 zea = details->last_index;
2882 unmap_mapping_range_vma(vma,
2883 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2884 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2889 static inline void unmap_mapping_range_list(struct list_head *head,
2890 struct zap_details *details)
2892 struct vm_area_struct *vma;
2895 * In nonlinear VMAs there is no correspondence between virtual address
2896 * offset and file offset. So we must perform an exhaustive search
2897 * across *all* the pages in each nonlinear VMA, not just the pages
2898 * whose virtual address lies outside the file truncation point.
2900 list_for_each_entry(vma, head, shared.nonlinear) {
2901 details->nonlinear_vma = vma;
2902 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2907 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2908 * @mapping: the address space containing mmaps to be unmapped.
2909 * @holebegin: byte in first page to unmap, relative to the start of
2910 * the underlying file. This will be rounded down to a PAGE_SIZE
2911 * boundary. Note that this is different from truncate_pagecache(), which
2912 * must keep the partial page. In contrast, we must get rid of
2914 * @holelen: size of prospective hole in bytes. This will be rounded
2915 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2917 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2918 * but 0 when invalidating pagecache, don't throw away private data.
2920 void unmap_mapping_range(struct address_space *mapping,
2921 loff_t const holebegin, loff_t const holelen, int even_cows)
2923 struct zap_details details;
2924 pgoff_t hba = holebegin >> PAGE_SHIFT;
2925 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2927 /* Check for overflow. */
2928 if (sizeof(holelen) > sizeof(hlen)) {
2930 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2931 if (holeend & ~(long long)ULONG_MAX)
2932 hlen = ULONG_MAX - hba + 1;
2935 details.check_mapping = even_cows? NULL: mapping;
2936 details.nonlinear_vma = NULL;
2937 details.first_index = hba;
2938 details.last_index = hba + hlen - 1;
2939 if (details.last_index < details.first_index)
2940 details.last_index = ULONG_MAX;
2943 mutex_lock(&mapping->i_mmap_mutex);
2944 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2945 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2946 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2947 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2948 mutex_unlock(&mapping->i_mmap_mutex);
2950 EXPORT_SYMBOL(unmap_mapping_range);
2953 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2954 * but allow concurrent faults), and pte mapped but not yet locked.
2955 * We return with mmap_sem still held, but pte unmapped and unlocked.
2957 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2958 unsigned long address, pte_t *page_table, pmd_t *pmd,
2959 unsigned int flags, pte_t orig_pte)
2962 struct page *page, *swapcache = NULL;
2966 struct mem_cgroup *ptr;
2970 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2973 entry = pte_to_swp_entry(orig_pte);
2974 if (unlikely(non_swap_entry(entry))) {
2975 if (is_migration_entry(entry)) {
2976 migration_entry_wait(mm, pmd, address);
2977 } else if (is_hwpoison_entry(entry)) {
2978 ret = VM_FAULT_HWPOISON;
2980 print_bad_pte(vma, address, orig_pte, NULL);
2981 ret = VM_FAULT_SIGBUS;
2985 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2986 page = lookup_swap_cache(entry);
2988 page = swapin_readahead(entry,
2989 GFP_HIGHUSER_MOVABLE, vma, address);
2992 * Back out if somebody else faulted in this pte
2993 * while we released the pte lock.
2995 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2996 if (likely(pte_same(*page_table, orig_pte)))
2998 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3002 /* Had to read the page from swap area: Major fault */
3003 ret = VM_FAULT_MAJOR;
3004 count_vm_event(PGMAJFAULT);
3005 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3006 } else if (PageHWPoison(page)) {
3008 * hwpoisoned dirty swapcache pages are kept for killing
3009 * owner processes (which may be unknown at hwpoison time)
3011 ret = VM_FAULT_HWPOISON;
3012 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3014 } else if (!(flags & FAULT_FLAG_TRIED))
3015 swap_cache_hit(vma);
3017 locked = lock_page_or_retry(page, mm, flags);
3019 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3021 ret |= VM_FAULT_RETRY;
3026 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3027 * release the swapcache from under us. The page pin, and pte_same
3028 * test below, are not enough to exclude that. Even if it is still
3029 * swapcache, we need to check that the page's swap has not changed.
3031 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3034 if (ksm_might_need_to_copy(page, vma, address)) {
3036 page = ksm_does_need_to_copy(page, vma, address);
3038 if (unlikely(!page)) {
3046 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3052 * Back out if somebody else already faulted in this pte.
3054 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3055 if (unlikely(!pte_same(*page_table, orig_pte)))
3058 if (unlikely(!PageUptodate(page))) {
3059 ret = VM_FAULT_SIGBUS;
3064 * The page isn't present yet, go ahead with the fault.
3066 * Be careful about the sequence of operations here.
3067 * To get its accounting right, reuse_swap_page() must be called
3068 * while the page is counted on swap but not yet in mapcount i.e.
3069 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3070 * must be called after the swap_free(), or it will never succeed.
3071 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3072 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3073 * in page->private. In this case, a record in swap_cgroup is silently
3074 * discarded at swap_free().
3077 inc_mm_counter_fast(mm, MM_ANONPAGES);
3078 dec_mm_counter_fast(mm, MM_SWAPENTS);
3079 pte = mk_pte(page, vma->vm_page_prot);
3080 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3081 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3082 flags &= ~FAULT_FLAG_WRITE;
3083 ret |= VM_FAULT_WRITE;
3086 flush_icache_page(vma, page);
3087 set_pte_at(mm, address, page_table, pte);
3088 do_page_add_anon_rmap(page, vma, address, exclusive);
3089 /* It's better to call commit-charge after rmap is established */
3090 mem_cgroup_commit_charge_swapin(page, ptr);
3093 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3094 try_to_free_swap(page);
3098 * Hold the lock to avoid the swap entry to be reused
3099 * until we take the PT lock for the pte_same() check
3100 * (to avoid false positives from pte_same). For
3101 * further safety release the lock after the swap_free
3102 * so that the swap count won't change under a
3103 * parallel locked swapcache.
3105 unlock_page(swapcache);
3106 page_cache_release(swapcache);
3109 if (flags & FAULT_FLAG_WRITE) {
3110 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3111 if (ret & VM_FAULT_ERROR)
3112 ret &= VM_FAULT_ERROR;
3116 /* No need to invalidate - it was non-present before */
3117 update_mmu_cache(vma, address, page_table);
3119 pte_unmap_unlock(page_table, ptl);
3123 mem_cgroup_cancel_charge_swapin(ptr);
3124 pte_unmap_unlock(page_table, ptl);
3128 page_cache_release(page);
3130 unlock_page(swapcache);
3131 page_cache_release(swapcache);
3137 * This is like a special single-page "expand_{down|up}wards()",
3138 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3139 * doesn't hit another vma.
3141 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3143 address &= PAGE_MASK;
3144 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3145 struct vm_area_struct *prev = vma->vm_prev;
3148 * Is there a mapping abutting this one below?
3150 * That's only ok if it's the same stack mapping
3151 * that has gotten split..
3153 if (prev && prev->vm_end == address)
3154 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3156 expand_downwards(vma, address - PAGE_SIZE);
3158 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3159 struct vm_area_struct *next = vma->vm_next;
3161 /* As VM_GROWSDOWN but s/below/above/ */
3162 if (next && next->vm_start == address + PAGE_SIZE)
3163 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3165 expand_upwards(vma, address + PAGE_SIZE);
3171 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3172 * but allow concurrent faults), and pte mapped but not yet locked.
3173 * We return with mmap_sem still held, but pte unmapped and unlocked.
3175 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3176 unsigned long address, pte_t *page_table, pmd_t *pmd,
3183 pte_unmap(page_table);
3185 /* Check if we need to add a guard page to the stack */
3186 if (check_stack_guard_page(vma, address) < 0)
3187 return VM_FAULT_SIGBUS;
3189 /* Use the zero-page for reads */
3190 if (!(flags & FAULT_FLAG_WRITE)) {
3191 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3192 vma->vm_page_prot));
3193 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3194 if (!pte_none(*page_table))
3199 /* Allocate our own private page. */
3200 if (unlikely(anon_vma_prepare(vma)))
3202 page = alloc_zeroed_user_highpage_movable(vma, address);
3205 __SetPageUptodate(page);
3207 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3210 entry = mk_pte(page, vma->vm_page_prot);
3211 if (vma->vm_flags & VM_WRITE)
3212 entry = pte_mkwrite(pte_mkdirty(entry));
3214 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3215 if (!pte_none(*page_table))
3218 inc_mm_counter_fast(mm, MM_ANONPAGES);
3219 page_add_new_anon_rmap(page, vma, address);
3221 set_pte_at(mm, address, page_table, entry);
3223 /* No need to invalidate - it was non-present before */
3224 update_mmu_cache(vma, address, page_table);
3226 pte_unmap_unlock(page_table, ptl);
3229 mem_cgroup_uncharge_page(page);
3230 page_cache_release(page);
3233 page_cache_release(page);
3235 return VM_FAULT_OOM;
3239 * __do_fault() tries to create a new page mapping. It aggressively
3240 * tries to share with existing pages, but makes a separate copy if
3241 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3242 * the next page fault.
3244 * As this is called only for pages that do not currently exist, we
3245 * do not need to flush old virtual caches or the TLB.
3247 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3248 * but allow concurrent faults), and pte neither mapped nor locked.
3249 * We return with mmap_sem still held, but pte unmapped and unlocked.
3251 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3252 unsigned long address, pmd_t *pmd,
3253 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3258 struct page *cow_page;
3261 struct page *dirty_page = NULL;
3262 struct vm_fault vmf;
3264 int page_mkwrite = 0;
3267 * If we do COW later, allocate page befor taking lock_page()
3268 * on the file cache page. This will reduce lock holding time.
3270 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3272 if (unlikely(anon_vma_prepare(vma)))
3273 return VM_FAULT_OOM;
3275 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3277 return VM_FAULT_OOM;
3279 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3280 page_cache_release(cow_page);
3281 return VM_FAULT_OOM;
3286 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3291 ret = vma->vm_ops->fault(vma, &vmf);
3292 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3296 if (unlikely(PageHWPoison(vmf.page))) {
3297 if (ret & VM_FAULT_LOCKED)
3298 unlock_page(vmf.page);
3299 ret = VM_FAULT_HWPOISON;
3304 * For consistency in subsequent calls, make the faulted page always
3307 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3308 lock_page(vmf.page);
3310 VM_BUG_ON(!PageLocked(vmf.page));
3313 * Should we do an early C-O-W break?
3316 if (flags & FAULT_FLAG_WRITE) {
3317 if (!(vma->vm_flags & VM_SHARED)) {
3320 copy_user_highpage(page, vmf.page, address, vma);
3321 __SetPageUptodate(page);
3324 * If the page will be shareable, see if the backing
3325 * address space wants to know that the page is about
3326 * to become writable
3328 if (vma->vm_ops->page_mkwrite) {
3332 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3333 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3335 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3337 goto unwritable_page;
3339 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3341 if (!page->mapping) {
3342 ret = 0; /* retry the fault */
3344 goto unwritable_page;
3347 VM_BUG_ON(!PageLocked(page));
3354 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3357 * This silly early PAGE_DIRTY setting removes a race
3358 * due to the bad i386 page protection. But it's valid
3359 * for other architectures too.
3361 * Note that if FAULT_FLAG_WRITE is set, we either now have
3362 * an exclusive copy of the page, or this is a shared mapping,
3363 * so we can make it writable and dirty to avoid having to
3364 * handle that later.
3366 /* Only go through if we didn't race with anybody else... */
3367 if (likely(pte_same(*page_table, orig_pte))) {
3368 flush_icache_page(vma, page);
3369 entry = mk_pte(page, vma->vm_page_prot);
3370 if (flags & FAULT_FLAG_WRITE)
3371 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3373 inc_mm_counter_fast(mm, MM_ANONPAGES);
3374 page_add_new_anon_rmap(page, vma, address);
3376 inc_mm_counter_fast(mm, MM_FILEPAGES);
3377 page_add_file_rmap(page);
3378 if (flags & FAULT_FLAG_WRITE) {
3380 get_page(dirty_page);
3383 set_pte_at(mm, address, page_table, entry);
3385 /* no need to invalidate: a not-present page won't be cached */
3386 update_mmu_cache(vma, address, page_table);
3389 mem_cgroup_uncharge_page(cow_page);
3391 page_cache_release(page);
3393 anon = 1; /* no anon but release faulted_page */
3396 pte_unmap_unlock(page_table, ptl);
3399 struct address_space *mapping = page->mapping;
3402 if (set_page_dirty(dirty_page))
3404 unlock_page(dirty_page);
3405 put_page(dirty_page);
3406 if ((dirtied || page_mkwrite) && mapping) {
3408 * Some device drivers do not set page.mapping but still
3411 balance_dirty_pages_ratelimited(mapping);
3414 /* file_update_time outside page_lock */
3415 if (vma->vm_file && !page_mkwrite)
3416 file_update_time(vma->vm_file);
3418 unlock_page(vmf.page);
3420 page_cache_release(vmf.page);
3426 page_cache_release(page);
3429 /* fs's fault handler get error */
3431 mem_cgroup_uncharge_page(cow_page);
3432 page_cache_release(cow_page);
3437 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3438 unsigned long address, pte_t *page_table, pmd_t *pmd,
3439 unsigned int flags, pte_t orig_pte)
3441 pgoff_t pgoff = (((address & PAGE_MASK)
3442 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3444 pte_unmap(page_table);
3445 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3449 * Fault of a previously existing named mapping. Repopulate the pte
3450 * from the encoded file_pte if possible. This enables swappable
3453 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3454 * but allow concurrent faults), and pte mapped but not yet locked.
3455 * We return with mmap_sem still held, but pte unmapped and unlocked.
3457 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3458 unsigned long address, pte_t *page_table, pmd_t *pmd,
3459 unsigned int flags, pte_t orig_pte)
3463 flags |= FAULT_FLAG_NONLINEAR;
3465 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3468 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3470 * Page table corrupted: show pte and kill process.
3472 print_bad_pte(vma, address, orig_pte, NULL);
3473 return VM_FAULT_SIGBUS;
3476 pgoff = pte_to_pgoff(orig_pte);
3477 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3480 static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3481 unsigned long address, pte_t *ptep, pmd_t *pmd,
3482 unsigned int flags, pte_t entry)
3484 struct page *page = NULL;
3485 int node, page_nid = -1;
3488 ptl = pte_lockptr(mm, pmd);
3490 if (unlikely(!pte_same(*ptep, entry)))
3493 page = vm_normal_page(vma, address, entry);
3496 page_nid = page_to_nid(page);
3497 node = mpol_misplaced(page, vma, address);
3502 out_pte_upgrade_unlock:
3503 flush_cache_page(vma, address, pte_pfn(entry));
3505 ptep_modify_prot_start(mm, address, ptep);
3506 entry = pte_modify(entry, vma->vm_page_prot);
3507 ptep_modify_prot_commit(mm, address, ptep, entry);
3509 /* No TLB flush needed because we upgraded the PTE */
3511 update_mmu_cache(vma, address, ptep);
3514 pte_unmap_unlock(ptep, ptl);
3517 task_numa_fault(page_nid, 1);
3524 pte_unmap_unlock(ptep, ptl);
3526 if (!migrate_misplaced_page(page, node)) {
3531 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
3532 if (!pte_same(*ptep, entry)) {
3538 goto out_pte_upgrade_unlock;
3542 * These routines also need to handle stuff like marking pages dirty
3543 * and/or accessed for architectures that don't do it in hardware (most
3544 * RISC architectures). The early dirtying is also good on the i386.
3546 * There is also a hook called "update_mmu_cache()" that architectures
3547 * with external mmu caches can use to update those (ie the Sparc or
3548 * PowerPC hashed page tables that act as extended TLBs).
3550 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3551 * but allow concurrent faults), and pte mapped but not yet locked.
3552 * We return with mmap_sem still held, but pte unmapped and unlocked.
3554 int handle_pte_fault(struct mm_struct *mm,
3555 struct vm_area_struct *vma, unsigned long address,
3556 pte_t *pte, pmd_t *pmd, unsigned int flags)
3561 entry = ACCESS_ONCE(*pte);
3562 if (!pte_present(entry)) {
3563 if (pte_none(entry)) {
3565 if (likely(vma->vm_ops->fault))
3566 return do_linear_fault(mm, vma, address,
3567 pte, pmd, flags, entry);
3569 return do_anonymous_page(mm, vma, address,
3572 if (pte_file(entry))
3573 return do_nonlinear_fault(mm, vma, address,
3574 pte, pmd, flags, entry);
3575 return do_swap_page(mm, vma, address,
3576 pte, pmd, flags, entry);
3579 if (pte_numa(vma, entry))
3580 return do_numa_page(mm, vma, address, pte, pmd, flags, entry);
3582 ptl = pte_lockptr(mm, pmd);
3584 if (unlikely(!pte_same(*pte, entry)))
3586 if (flags & FAULT_FLAG_WRITE) {
3587 if (!pte_write(entry))
3588 return do_wp_page(mm, vma, address,
3589 pte, pmd, ptl, entry);
3590 entry = pte_mkdirty(entry);
3592 entry = pte_mkyoung(entry);
3593 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3594 update_mmu_cache(vma, address, pte);
3597 * This is needed only for protection faults but the arch code
3598 * is not yet telling us if this is a protection fault or not.
3599 * This still avoids useless tlb flushes for .text page faults
3602 if (flags & FAULT_FLAG_WRITE)
3603 flush_tlb_fix_spurious_fault(vma, address);
3606 pte_unmap_unlock(pte, ptl);
3611 * By the time we get here, we already hold the mm semaphore
3613 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3614 unsigned long address, unsigned int flags)
3621 __set_current_state(TASK_RUNNING);
3623 count_vm_event(PGFAULT);
3624 mem_cgroup_count_vm_event(mm, PGFAULT);
3626 /* do counter updates before entering really critical section. */
3627 check_sync_rss_stat(current);
3629 if (unlikely(is_vm_hugetlb_page(vma)))
3630 return hugetlb_fault(mm, vma, address, flags);
3633 pgd = pgd_offset(mm, address);
3634 pud = pud_alloc(mm, pgd, address);
3636 return VM_FAULT_OOM;
3637 pmd = pmd_alloc(mm, pud, address);
3639 return VM_FAULT_OOM;
3640 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3642 return do_huge_pmd_anonymous_page(mm, vma, address,
3645 pmd_t orig_pmd = *pmd;
3649 if (pmd_trans_huge(orig_pmd) && !pmd_trans_splitting(orig_pmd)) {
3650 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3652 if (pmd_numa(vma, orig_pmd)) {
3653 do_huge_pmd_numa_page(mm, vma, address, pmd,
3657 if (dirty && !pmd_write(orig_pmd)) {
3658 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3661 * If COW results in an oom, the huge pmd will
3662 * have been split, so retry the fault on the
3663 * pte for a smaller charge.
3665 if (unlikely(ret & VM_FAULT_OOM))
3668 huge_pmd_set_accessed(mm, vma, address, pmd,
3678 * Use __pte_alloc instead of pte_alloc_map, because we can't
3679 * run pte_offset_map on the pmd, if an huge pmd could
3680 * materialize from under us from a different thread.
3682 if (unlikely(pmd_none(*pmd)) &&
3683 unlikely(__pte_alloc(mm, vma, pmd, address)))
3684 return VM_FAULT_OOM;
3685 /* if an huge pmd materialized from under us just retry later */
3686 if (unlikely(pmd_trans_huge(*pmd)))
3689 * A regular pmd is established and it can't morph into a huge pmd
3690 * from under us anymore at this point because we hold the mmap_sem
3691 * read mode and khugepaged takes it in write mode. So now it's
3692 * safe to run pte_offset_map().
3694 pte = pte_offset_map(pmd, address);
3696 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3699 #ifndef __PAGETABLE_PUD_FOLDED
3701 * Allocate page upper directory.
3702 * We've already handled the fast-path in-line.
3704 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3706 pud_t *new = pud_alloc_one(mm, address);
3710 smp_wmb(); /* See comment in __pte_alloc */
3712 spin_lock(&mm->page_table_lock);
3713 if (pgd_present(*pgd)) /* Another has populated it */
3716 pgd_populate(mm, pgd, new);
3717 spin_unlock(&mm->page_table_lock);
3720 #endif /* __PAGETABLE_PUD_FOLDED */
3722 #ifndef __PAGETABLE_PMD_FOLDED
3724 * Allocate page middle directory.
3725 * We've already handled the fast-path in-line.
3727 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3729 pmd_t *new = pmd_alloc_one(mm, address);
3733 smp_wmb(); /* See comment in __pte_alloc */
3735 spin_lock(&mm->page_table_lock);
3736 #ifndef __ARCH_HAS_4LEVEL_HACK
3737 if (pud_present(*pud)) /* Another has populated it */
3740 pud_populate(mm, pud, new);
3742 if (pgd_present(*pud)) /* Another has populated it */
3745 pgd_populate(mm, pud, new);
3746 #endif /* __ARCH_HAS_4LEVEL_HACK */
3747 spin_unlock(&mm->page_table_lock);
3750 #endif /* __PAGETABLE_PMD_FOLDED */
3752 int make_pages_present(unsigned long addr, unsigned long end)
3754 int ret, len, write;
3755 struct vm_area_struct * vma;
3757 vma = find_vma(current->mm, addr);
3761 * We want to touch writable mappings with a write fault in order
3762 * to break COW, except for shared mappings because these don't COW
3763 * and we would not want to dirty them for nothing.
3765 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3766 BUG_ON(addr >= end);
3767 BUG_ON(end > vma->vm_end);
3768 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3769 ret = get_user_pages(current, current->mm, addr,
3770 len, write, 0, NULL, NULL);
3773 return ret == len ? 0 : -EFAULT;
3776 #if !defined(__HAVE_ARCH_GATE_AREA)
3778 #if defined(AT_SYSINFO_EHDR)
3779 static struct vm_area_struct gate_vma;
3781 static int __init gate_vma_init(void)
3783 gate_vma.vm_mm = NULL;
3784 gate_vma.vm_start = FIXADDR_USER_START;
3785 gate_vma.vm_end = FIXADDR_USER_END;
3786 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3787 gate_vma.vm_page_prot = __P101;
3791 __initcall(gate_vma_init);
3794 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3796 #ifdef AT_SYSINFO_EHDR
3803 int in_gate_area_no_mm(unsigned long addr)
3805 #ifdef AT_SYSINFO_EHDR
3806 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3812 #endif /* __HAVE_ARCH_GATE_AREA */
3814 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3815 pte_t **ptepp, spinlock_t **ptlp)
3822 pgd = pgd_offset(mm, address);
3823 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3826 pud = pud_offset(pgd, address);
3827 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3830 pmd = pmd_offset(pud, address);
3831 VM_BUG_ON(pmd_trans_huge(*pmd));
3832 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3835 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3839 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3842 if (!pte_present(*ptep))
3847 pte_unmap_unlock(ptep, *ptlp);
3852 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3853 pte_t **ptepp, spinlock_t **ptlp)
3857 /* (void) is needed to make gcc happy */
3858 (void) __cond_lock(*ptlp,
3859 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3864 * follow_pfn - look up PFN at a user virtual address
3865 * @vma: memory mapping
3866 * @address: user virtual address
3867 * @pfn: location to store found PFN
3869 * Only IO mappings and raw PFN mappings are allowed.
3871 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3873 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3880 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3883 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3886 *pfn = pte_pfn(*ptep);
3887 pte_unmap_unlock(ptep, ptl);
3890 EXPORT_SYMBOL(follow_pfn);
3892 #ifdef CONFIG_HAVE_IOREMAP_PROT
3893 int follow_phys(struct vm_area_struct *vma,
3894 unsigned long address, unsigned int flags,
3895 unsigned long *prot, resource_size_t *phys)
3901 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3904 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3908 if ((flags & FOLL_WRITE) && !pte_write(pte))
3911 *prot = pgprot_val(pte_pgprot(pte));
3912 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3916 pte_unmap_unlock(ptep, ptl);
3921 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3922 void *buf, int len, int write)
3924 resource_size_t phys_addr;
3925 unsigned long prot = 0;
3926 void __iomem *maddr;
3927 int offset = addr & (PAGE_SIZE-1);
3929 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3932 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3934 memcpy_toio(maddr + offset, buf, len);
3936 memcpy_fromio(buf, maddr + offset, len);
3944 * Access another process' address space as given in mm. If non-NULL, use the
3945 * given task for page fault accounting.
3947 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3948 unsigned long addr, void *buf, int len, int write)
3950 struct vm_area_struct *vma;
3951 void *old_buf = buf;
3953 down_read(&mm->mmap_sem);
3954 /* ignore errors, just check how much was successfully transferred */
3956 int bytes, ret, offset;
3958 struct page *page = NULL;
3960 ret = get_user_pages(tsk, mm, addr, 1,
3961 write, 1, &page, &vma);
3964 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3965 * we can access using slightly different code.
3967 #ifdef CONFIG_HAVE_IOREMAP_PROT
3968 vma = find_vma(mm, addr);
3969 if (!vma || vma->vm_start > addr)
3971 if (vma->vm_ops && vma->vm_ops->access)
3972 ret = vma->vm_ops->access(vma, addr, buf,
3980 offset = addr & (PAGE_SIZE-1);
3981 if (bytes > PAGE_SIZE-offset)
3982 bytes = PAGE_SIZE-offset;
3986 copy_to_user_page(vma, page, addr,
3987 maddr + offset, buf, bytes);
3988 set_page_dirty_lock(page);
3990 copy_from_user_page(vma, page, addr,
3991 buf, maddr + offset, bytes);
3994 page_cache_release(page);
4000 up_read(&mm->mmap_sem);
4002 return buf - old_buf;
4006 * access_remote_vm - access another process' address space
4007 * @mm: the mm_struct of the target address space
4008 * @addr: start address to access
4009 * @buf: source or destination buffer
4010 * @len: number of bytes to transfer
4011 * @write: whether the access is a write
4013 * The caller must hold a reference on @mm.
4015 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4016 void *buf, int len, int write)
4018 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4022 * Access another process' address space.
4023 * Source/target buffer must be kernel space,
4024 * Do not walk the page table directly, use get_user_pages
4026 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4027 void *buf, int len, int write)
4029 struct mm_struct *mm;
4032 mm = get_task_mm(tsk);
4036 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4043 * Print the name of a VMA.
4045 void print_vma_addr(char *prefix, unsigned long ip)
4047 struct mm_struct *mm = current->mm;
4048 struct vm_area_struct *vma;
4051 * Do not print if we are in atomic
4052 * contexts (in exception stacks, etc.):
4054 if (preempt_count())
4057 down_read(&mm->mmap_sem);
4058 vma = find_vma(mm, ip);
4059 if (vma && vma->vm_file) {
4060 struct file *f = vma->vm_file;
4061 char *buf = (char *)__get_free_page(GFP_KERNEL);
4065 p = d_path(&f->f_path, buf, PAGE_SIZE);
4068 s = strrchr(p, '/');
4071 printk("%s%s[%lx+%lx]", prefix, p,
4073 vma->vm_end - vma->vm_start);
4074 free_page((unsigned long)buf);
4077 up_read(&mm->mmap_sem);
4080 #ifdef CONFIG_PROVE_LOCKING
4081 void might_fault(void)
4084 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4085 * holding the mmap_sem, this is safe because kernel memory doesn't
4086 * get paged out, therefore we'll never actually fault, and the
4087 * below annotations will generate false positives.
4089 if (segment_eq(get_fs(), KERNEL_DS))
4094 * it would be nicer only to annotate paths which are not under
4095 * pagefault_disable, however that requires a larger audit and
4096 * providing helpers like get_user_atomic.
4098 if (!in_atomic() && current->mm)
4099 might_lock_read(¤t->mm->mmap_sem);
4101 EXPORT_SYMBOL(might_fault);
4104 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4105 static void clear_gigantic_page(struct page *page,
4107 unsigned int pages_per_huge_page)
4110 struct page *p = page;
4113 for (i = 0; i < pages_per_huge_page;
4114 i++, p = mem_map_next(p, page, i)) {
4116 clear_user_highpage(p, addr + i * PAGE_SIZE);
4119 void clear_huge_page(struct page *page,
4120 unsigned long addr, unsigned int pages_per_huge_page)
4124 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4125 clear_gigantic_page(page, addr, pages_per_huge_page);
4130 for (i = 0; i < pages_per_huge_page; i++) {
4132 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4136 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4138 struct vm_area_struct *vma,
4139 unsigned int pages_per_huge_page)
4142 struct page *dst_base = dst;
4143 struct page *src_base = src;
4145 for (i = 0; i < pages_per_huge_page; ) {
4147 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4150 dst = mem_map_next(dst, dst_base, i);
4151 src = mem_map_next(src, src_base, i);
4155 void copy_user_huge_page(struct page *dst, struct page *src,
4156 unsigned long addr, struct vm_area_struct *vma,
4157 unsigned int pages_per_huge_page)
4161 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4162 copy_user_gigantic_page(dst, src, addr, vma,
4163 pages_per_huge_page);
4168 for (i = 0; i < pages_per_huge_page; i++) {
4170 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4173 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */