4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/module.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
62 #include <asm/pgalloc.h>
63 #include <asm/uaccess.h>
65 #include <asm/tlbflush.h>
66 #include <asm/pgtable.h>
70 #ifndef CONFIG_NEED_MULTIPLE_NODES
71 /* use the per-pgdat data instead for discontigmem - mbligh */
72 unsigned long max_mapnr;
75 EXPORT_SYMBOL(max_mapnr);
76 EXPORT_SYMBOL(mem_map);
79 unsigned long num_physpages;
81 * A number of key systems in x86 including ioremap() rely on the assumption
82 * that high_memory defines the upper bound on direct map memory, then end
83 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
84 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
89 EXPORT_SYMBOL(num_physpages);
90 EXPORT_SYMBOL(high_memory);
93 * Randomize the address space (stacks, mmaps, brk, etc.).
95 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
96 * as ancient (libc5 based) binaries can segfault. )
98 int randomize_va_space __read_mostly =
99 #ifdef CONFIG_COMPAT_BRK
105 static int __init disable_randmaps(char *s)
107 randomize_va_space = 0;
110 __setup("norandmaps", disable_randmaps);
112 unsigned long zero_pfn __read_mostly;
113 unsigned long highest_memmap_pfn __read_mostly;
116 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
118 static int __init init_zero_pfn(void)
120 zero_pfn = page_to_pfn(ZERO_PAGE(0));
123 core_initcall(init_zero_pfn);
126 #if defined(SPLIT_RSS_COUNTING)
128 static void __sync_task_rss_stat(struct task_struct *task, struct mm_struct *mm)
132 for (i = 0; i < NR_MM_COUNTERS; i++) {
133 if (task->rss_stat.count[i]) {
134 add_mm_counter(mm, i, task->rss_stat.count[i]);
135 task->rss_stat.count[i] = 0;
138 task->rss_stat.events = 0;
141 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
143 struct task_struct *task = current;
145 if (likely(task->mm == mm))
146 task->rss_stat.count[member] += val;
148 add_mm_counter(mm, member, val);
150 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
151 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
153 /* sync counter once per 64 page faults */
154 #define TASK_RSS_EVENTS_THRESH (64)
155 static void check_sync_rss_stat(struct task_struct *task)
157 if (unlikely(task != current))
159 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
160 __sync_task_rss_stat(task, task->mm);
163 unsigned long get_mm_counter(struct mm_struct *mm, int member)
168 * Don't use task->mm here...for avoiding to use task_get_mm()..
169 * The caller must guarantee task->mm is not invalid.
171 val = atomic_long_read(&mm->rss_stat.count[member]);
173 * counter is updated in asynchronous manner and may go to minus.
174 * But it's never be expected number for users.
178 return (unsigned long)val;
181 void sync_mm_rss(struct task_struct *task, struct mm_struct *mm)
183 __sync_task_rss_stat(task, mm);
185 #else /* SPLIT_RSS_COUNTING */
187 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
188 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
190 static void check_sync_rss_stat(struct task_struct *task)
194 #endif /* SPLIT_RSS_COUNTING */
196 #ifdef HAVE_GENERIC_MMU_GATHER
198 static int tlb_next_batch(struct mmu_gather *tlb)
200 struct mmu_gather_batch *batch;
204 tlb->active = batch->next;
208 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
214 batch->max = MAX_GATHER_BATCH;
216 tlb->active->next = batch;
223 * Called to initialize an (on-stack) mmu_gather structure for page-table
224 * tear-down from @mm. The @fullmm argument is used when @mm is without
225 * users and we're going to destroy the full address space (exit/execve).
227 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
231 tlb->fullmm = fullmm;
233 tlb->fast_mode = (num_possible_cpus() == 1);
234 tlb->local.next = NULL;
236 tlb->local.max = ARRAY_SIZE(tlb->__pages);
237 tlb->active = &tlb->local;
239 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 void tlb_flush_mmu(struct mmu_gather *tlb)
246 struct mmu_gather_batch *batch;
248 if (!tlb->need_flush)
252 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
253 tlb_table_flush(tlb);
256 if (tlb_fast_mode(tlb))
259 for (batch = &tlb->local; batch; batch = batch->next) {
260 free_pages_and_swap_cache(batch->pages, batch->nr);
263 tlb->active = &tlb->local;
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
270 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
272 struct mmu_gather_batch *batch, *next;
276 /* keep the page table cache within bounds */
279 for (batch = tlb->local.next; batch; batch = next) {
281 free_pages((unsigned long)batch, 0);
283 tlb->local.next = NULL;
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
292 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
294 struct mmu_gather_batch *batch;
298 if (tlb_fast_mode(tlb)) {
299 free_page_and_swap_cache(page);
300 return 1; /* avoid calling tlb_flush_mmu() */
304 batch->pages[batch->nr++] = page;
305 if (batch->nr == batch->max) {
306 if (!tlb_next_batch(tlb))
310 VM_BUG_ON(batch->nr > batch->max);
312 return batch->max - batch->nr;
315 #endif /* HAVE_GENERIC_MMU_GATHER */
317 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
320 * See the comment near struct mmu_table_batch.
323 static void tlb_remove_table_smp_sync(void *arg)
325 /* Simply deliver the interrupt */
328 static void tlb_remove_table_one(void *table)
331 * This isn't an RCU grace period and hence the page-tables cannot be
332 * assumed to be actually RCU-freed.
334 * It is however sufficient for software page-table walkers that rely on
335 * IRQ disabling. See the comment near struct mmu_table_batch.
337 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
338 __tlb_remove_table(table);
341 static void tlb_remove_table_rcu(struct rcu_head *head)
343 struct mmu_table_batch *batch;
346 batch = container_of(head, struct mmu_table_batch, rcu);
348 for (i = 0; i < batch->nr; i++)
349 __tlb_remove_table(batch->tables[i]);
351 free_page((unsigned long)batch);
354 void tlb_table_flush(struct mmu_gather *tlb)
356 struct mmu_table_batch **batch = &tlb->batch;
359 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
364 void tlb_remove_table(struct mmu_gather *tlb, void *table)
366 struct mmu_table_batch **batch = &tlb->batch;
371 * When there's less then two users of this mm there cannot be a
372 * concurrent page-table walk.
374 if (atomic_read(&tlb->mm->mm_users) < 2) {
375 __tlb_remove_table(table);
379 if (*batch == NULL) {
380 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
381 if (*batch == NULL) {
382 tlb_remove_table_one(table);
387 (*batch)->tables[(*batch)->nr++] = table;
388 if ((*batch)->nr == MAX_TABLE_BATCH)
389 tlb_table_flush(tlb);
392 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
395 * If a p?d_bad entry is found while walking page tables, report
396 * the error, before resetting entry to p?d_none. Usually (but
397 * very seldom) called out from the p?d_none_or_clear_bad macros.
400 void pgd_clear_bad(pgd_t *pgd)
406 void pud_clear_bad(pud_t *pud)
412 void pmd_clear_bad(pmd_t *pmd)
419 * Note: this doesn't free the actual pages themselves. That
420 * has been handled earlier when unmapping all the memory regions.
422 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
425 pgtable_t token = pmd_pgtable(*pmd);
427 pte_free_tlb(tlb, token, addr);
431 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
432 unsigned long addr, unsigned long end,
433 unsigned long floor, unsigned long ceiling)
440 pmd = pmd_offset(pud, addr);
442 next = pmd_addr_end(addr, end);
443 if (pmd_none_or_clear_bad(pmd))
445 free_pte_range(tlb, pmd, addr);
446 } while (pmd++, addr = next, addr != end);
456 if (end - 1 > ceiling - 1)
459 pmd = pmd_offset(pud, start);
461 pmd_free_tlb(tlb, pmd, start);
464 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
465 unsigned long addr, unsigned long end,
466 unsigned long floor, unsigned long ceiling)
473 pud = pud_offset(pgd, addr);
475 next = pud_addr_end(addr, end);
476 if (pud_none_or_clear_bad(pud))
478 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
479 } while (pud++, addr = next, addr != end);
485 ceiling &= PGDIR_MASK;
489 if (end - 1 > ceiling - 1)
492 pud = pud_offset(pgd, start);
494 pud_free_tlb(tlb, pud, start);
498 * This function frees user-level page tables of a process.
500 * Must be called with pagetable lock held.
502 void free_pgd_range(struct mmu_gather *tlb,
503 unsigned long addr, unsigned long end,
504 unsigned long floor, unsigned long ceiling)
510 * The next few lines have given us lots of grief...
512 * Why are we testing PMD* at this top level? Because often
513 * there will be no work to do at all, and we'd prefer not to
514 * go all the way down to the bottom just to discover that.
516 * Why all these "- 1"s? Because 0 represents both the bottom
517 * of the address space and the top of it (using -1 for the
518 * top wouldn't help much: the masks would do the wrong thing).
519 * The rule is that addr 0 and floor 0 refer to the bottom of
520 * the address space, but end 0 and ceiling 0 refer to the top
521 * Comparisons need to use "end - 1" and "ceiling - 1" (though
522 * that end 0 case should be mythical).
524 * Wherever addr is brought up or ceiling brought down, we must
525 * be careful to reject "the opposite 0" before it confuses the
526 * subsequent tests. But what about where end is brought down
527 * by PMD_SIZE below? no, end can't go down to 0 there.
529 * Whereas we round start (addr) and ceiling down, by different
530 * masks at different levels, in order to test whether a table
531 * now has no other vmas using it, so can be freed, we don't
532 * bother to round floor or end up - the tests don't need that.
546 if (end - 1 > ceiling - 1)
551 pgd = pgd_offset(tlb->mm, addr);
553 next = pgd_addr_end(addr, end);
554 if (pgd_none_or_clear_bad(pgd))
556 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
557 } while (pgd++, addr = next, addr != end);
560 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
561 unsigned long floor, unsigned long ceiling)
564 struct vm_area_struct *next = vma->vm_next;
565 unsigned long addr = vma->vm_start;
568 * Hide vma from rmap and truncate_pagecache before freeing
571 unlink_anon_vmas(vma);
572 unlink_file_vma(vma);
574 if (is_vm_hugetlb_page(vma)) {
575 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
576 floor, next? next->vm_start: ceiling);
579 * Optimization: gather nearby vmas into one call down
581 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
582 && !is_vm_hugetlb_page(next)) {
585 unlink_anon_vmas(vma);
586 unlink_file_vma(vma);
588 free_pgd_range(tlb, addr, vma->vm_end,
589 floor, next? next->vm_start: ceiling);
595 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
596 pmd_t *pmd, unsigned long address)
598 pgtable_t new = pte_alloc_one(mm, address);
599 int wait_split_huge_page;
604 * Ensure all pte setup (eg. pte page lock and page clearing) are
605 * visible before the pte is made visible to other CPUs by being
606 * put into page tables.
608 * The other side of the story is the pointer chasing in the page
609 * table walking code (when walking the page table without locking;
610 * ie. most of the time). Fortunately, these data accesses consist
611 * of a chain of data-dependent loads, meaning most CPUs (alpha
612 * being the notable exception) will already guarantee loads are
613 * seen in-order. See the alpha page table accessors for the
614 * smp_read_barrier_depends() barriers in page table walking code.
616 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
618 spin_lock(&mm->page_table_lock);
619 wait_split_huge_page = 0;
620 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
622 pmd_populate(mm, pmd, new);
624 } else if (unlikely(pmd_trans_splitting(*pmd)))
625 wait_split_huge_page = 1;
626 spin_unlock(&mm->page_table_lock);
629 if (wait_split_huge_page)
630 wait_split_huge_page(vma->anon_vma, pmd);
634 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
636 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
640 smp_wmb(); /* See comment in __pte_alloc */
642 spin_lock(&init_mm.page_table_lock);
643 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
644 pmd_populate_kernel(&init_mm, pmd, new);
647 VM_BUG_ON(pmd_trans_splitting(*pmd));
648 spin_unlock(&init_mm.page_table_lock);
650 pte_free_kernel(&init_mm, new);
654 static inline void init_rss_vec(int *rss)
656 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
659 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
663 if (current->mm == mm)
664 sync_mm_rss(current, mm);
665 for (i = 0; i < NR_MM_COUNTERS; i++)
667 add_mm_counter(mm, i, rss[i]);
671 * This function is called to print an error when a bad pte
672 * is found. For example, we might have a PFN-mapped pte in
673 * a region that doesn't allow it.
675 * The calling function must still handle the error.
677 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
678 pte_t pte, struct page *page)
680 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
681 pud_t *pud = pud_offset(pgd, addr);
682 pmd_t *pmd = pmd_offset(pud, addr);
683 struct address_space *mapping;
685 static unsigned long resume;
686 static unsigned long nr_shown;
687 static unsigned long nr_unshown;
690 * Allow a burst of 60 reports, then keep quiet for that minute;
691 * or allow a steady drip of one report per second.
693 if (nr_shown == 60) {
694 if (time_before(jiffies, resume)) {
700 "BUG: Bad page map: %lu messages suppressed\n",
707 resume = jiffies + 60 * HZ;
709 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
710 index = linear_page_index(vma, addr);
713 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
715 (long long)pte_val(pte), (long long)pmd_val(*pmd));
719 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
720 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
722 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
725 print_symbol(KERN_ALERT "vma->vm_ops->fault: %s\n",
726 (unsigned long)vma->vm_ops->fault);
727 if (vma->vm_file && vma->vm_file->f_op)
728 print_symbol(KERN_ALERT "vma->vm_file->f_op->mmap: %s\n",
729 (unsigned long)vma->vm_file->f_op->mmap);
731 add_taint(TAINT_BAD_PAGE);
734 static inline int is_cow_mapping(vm_flags_t flags)
736 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
740 static inline int is_zero_pfn(unsigned long pfn)
742 return pfn == zero_pfn;
747 static inline unsigned long my_zero_pfn(unsigned long addr)
754 * vm_normal_page -- This function gets the "struct page" associated with a pte.
756 * "Special" mappings do not wish to be associated with a "struct page" (either
757 * it doesn't exist, or it exists but they don't want to touch it). In this
758 * case, NULL is returned here. "Normal" mappings do have a struct page.
760 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
761 * pte bit, in which case this function is trivial. Secondly, an architecture
762 * may not have a spare pte bit, which requires a more complicated scheme,
765 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
766 * special mapping (even if there are underlying and valid "struct pages").
767 * COWed pages of a VM_PFNMAP are always normal.
769 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
770 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
771 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
772 * mapping will always honor the rule
774 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
776 * And for normal mappings this is false.
778 * This restricts such mappings to be a linear translation from virtual address
779 * to pfn. To get around this restriction, we allow arbitrary mappings so long
780 * as the vma is not a COW mapping; in that case, we know that all ptes are
781 * special (because none can have been COWed).
784 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
786 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
787 * page" backing, however the difference is that _all_ pages with a struct
788 * page (that is, those where pfn_valid is true) are refcounted and considered
789 * normal pages by the VM. The disadvantage is that pages are refcounted
790 * (which can be slower and simply not an option for some PFNMAP users). The
791 * advantage is that we don't have to follow the strict linearity rule of
792 * PFNMAP mappings in order to support COWable mappings.
795 #ifdef __HAVE_ARCH_PTE_SPECIAL
796 # define HAVE_PTE_SPECIAL 1
798 # define HAVE_PTE_SPECIAL 0
800 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
803 unsigned long pfn = pte_pfn(pte);
805 if (HAVE_PTE_SPECIAL) {
806 if (likely(!pte_special(pte)))
808 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
810 if (!is_zero_pfn(pfn))
811 print_bad_pte(vma, addr, pte, NULL);
815 /* !HAVE_PTE_SPECIAL case follows: */
817 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
818 if (vma->vm_flags & VM_MIXEDMAP) {
824 off = (addr - vma->vm_start) >> PAGE_SHIFT;
825 if (pfn == vma->vm_pgoff + off)
827 if (!is_cow_mapping(vma->vm_flags))
832 if (is_zero_pfn(pfn))
835 if (unlikely(pfn > highest_memmap_pfn)) {
836 print_bad_pte(vma, addr, pte, NULL);
841 * NOTE! We still have PageReserved() pages in the page tables.
842 * eg. VDSO mappings can cause them to exist.
845 return pfn_to_page(pfn);
849 * copy one vm_area from one task to the other. Assumes the page tables
850 * already present in the new task to be cleared in the whole range
851 * covered by this vma.
854 static inline unsigned long
855 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
856 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
857 unsigned long addr, int *rss)
859 unsigned long vm_flags = vma->vm_flags;
860 pte_t pte = *src_pte;
863 /* pte contains position in swap or file, so copy. */
864 if (unlikely(!pte_present(pte))) {
865 if (!pte_file(pte)) {
866 swp_entry_t entry = pte_to_swp_entry(pte);
868 if (swap_duplicate(entry) < 0)
871 /* make sure dst_mm is on swapoff's mmlist. */
872 if (unlikely(list_empty(&dst_mm->mmlist))) {
873 spin_lock(&mmlist_lock);
874 if (list_empty(&dst_mm->mmlist))
875 list_add(&dst_mm->mmlist,
877 spin_unlock(&mmlist_lock);
879 if (likely(!non_swap_entry(entry)))
881 else if (is_write_migration_entry(entry) &&
882 is_cow_mapping(vm_flags)) {
884 * COW mappings require pages in both parent
885 * and child to be set to read.
887 make_migration_entry_read(&entry);
888 pte = swp_entry_to_pte(entry);
889 set_pte_at(src_mm, addr, src_pte, pte);
896 * If it's a COW mapping, write protect it both
897 * in the parent and the child
899 if (is_cow_mapping(vm_flags)) {
900 ptep_set_wrprotect(src_mm, addr, src_pte);
901 pte = pte_wrprotect(pte);
905 * If it's a shared mapping, mark it clean in
908 if (vm_flags & VM_SHARED)
909 pte = pte_mkclean(pte);
910 pte = pte_mkold(pte);
912 page = vm_normal_page(vma, addr, pte);
923 set_pte_at(dst_mm, addr, dst_pte, pte);
927 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
928 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
929 unsigned long addr, unsigned long end)
931 pte_t *orig_src_pte, *orig_dst_pte;
932 pte_t *src_pte, *dst_pte;
933 spinlock_t *src_ptl, *dst_ptl;
935 int rss[NR_MM_COUNTERS];
936 swp_entry_t entry = (swp_entry_t){0};
941 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
944 src_pte = pte_offset_map(src_pmd, addr);
945 src_ptl = pte_lockptr(src_mm, src_pmd);
946 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
947 orig_src_pte = src_pte;
948 orig_dst_pte = dst_pte;
949 arch_enter_lazy_mmu_mode();
953 * We are holding two locks at this point - either of them
954 * could generate latencies in another task on another CPU.
956 if (progress >= 32) {
958 if (need_resched() ||
959 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
962 if (pte_none(*src_pte)) {
966 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
971 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
973 arch_leave_lazy_mmu_mode();
974 spin_unlock(src_ptl);
975 pte_unmap(orig_src_pte);
976 add_mm_rss_vec(dst_mm, rss);
977 pte_unmap_unlock(orig_dst_pte, dst_ptl);
981 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
990 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
991 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
992 unsigned long addr, unsigned long end)
994 pmd_t *src_pmd, *dst_pmd;
997 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1000 src_pmd = pmd_offset(src_pud, addr);
1002 next = pmd_addr_end(addr, end);
1003 if (pmd_trans_huge(*src_pmd)) {
1005 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1006 err = copy_huge_pmd(dst_mm, src_mm,
1007 dst_pmd, src_pmd, addr, vma);
1014 if (pmd_none_or_clear_bad(src_pmd))
1016 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1019 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1023 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1024 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1025 unsigned long addr, unsigned long end)
1027 pud_t *src_pud, *dst_pud;
1030 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1033 src_pud = pud_offset(src_pgd, addr);
1035 next = pud_addr_end(addr, end);
1036 if (pud_none_or_clear_bad(src_pud))
1038 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1041 } while (dst_pud++, src_pud++, addr = next, addr != end);
1045 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1046 struct vm_area_struct *vma)
1048 pgd_t *src_pgd, *dst_pgd;
1050 unsigned long addr = vma->vm_start;
1051 unsigned long end = vma->vm_end;
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|VM_PFNMAP|VM_INSERTPAGE))) {
1065 if (is_vm_hugetlb_page(vma))
1066 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1068 if (unlikely(is_pfn_mapping(vma))) {
1070 * We do not free on error cases below as remove_vma
1071 * gets called on error from higher level routine
1073 ret = track_pfn_vma_copy(vma);
1079 * We need to invalidate the secondary MMU mappings only when
1080 * there could be a permission downgrade on the ptes of the
1081 * parent mm. And a permission downgrade will only happen if
1082 * is_cow_mapping() returns true.
1084 if (is_cow_mapping(vma->vm_flags))
1085 mmu_notifier_invalidate_range_start(src_mm, addr, end);
1088 dst_pgd = pgd_offset(dst_mm, addr);
1089 src_pgd = pgd_offset(src_mm, addr);
1091 next = pgd_addr_end(addr, end);
1092 if (pgd_none_or_clear_bad(src_pgd))
1094 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1095 vma, addr, next))) {
1099 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1101 if (is_cow_mapping(vma->vm_flags))
1102 mmu_notifier_invalidate_range_end(src_mm,
1103 vma->vm_start, end);
1107 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1108 struct vm_area_struct *vma, pmd_t *pmd,
1109 unsigned long addr, unsigned long end,
1110 struct zap_details *details)
1112 struct mm_struct *mm = tlb->mm;
1113 int force_flush = 0;
1114 int rss[NR_MM_COUNTERS];
1121 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1123 arch_enter_lazy_mmu_mode();
1126 if (pte_none(ptent)) {
1130 if (pte_present(ptent)) {
1133 page = vm_normal_page(vma, addr, ptent);
1134 if (unlikely(details) && page) {
1136 * unmap_shared_mapping_pages() wants to
1137 * invalidate cache without truncating:
1138 * unmap shared but keep private pages.
1140 if (details->check_mapping &&
1141 details->check_mapping != page->mapping)
1144 * Each page->index must be checked when
1145 * invalidating or truncating nonlinear.
1147 if (details->nonlinear_vma &&
1148 (page->index < details->first_index ||
1149 page->index > details->last_index))
1152 ptent = ptep_get_and_clear_full(mm, addr, pte,
1154 tlb_remove_tlb_entry(tlb, pte, addr);
1155 if (unlikely(!page))
1157 if (unlikely(details) && details->nonlinear_vma
1158 && linear_page_index(details->nonlinear_vma,
1159 addr) != page->index)
1160 set_pte_at(mm, addr, pte,
1161 pgoff_to_pte(page->index));
1163 rss[MM_ANONPAGES]--;
1165 if (pte_dirty(ptent))
1166 set_page_dirty(page);
1167 if (pte_young(ptent) &&
1168 likely(!VM_SequentialReadHint(vma)))
1169 mark_page_accessed(page);
1170 rss[MM_FILEPAGES]--;
1172 page_remove_rmap(page);
1173 if (unlikely(page_mapcount(page) < 0))
1174 print_bad_pte(vma, addr, ptent, page);
1175 force_flush = !__tlb_remove_page(tlb, page);
1181 * If details->check_mapping, we leave swap entries;
1182 * if details->nonlinear_vma, we leave file entries.
1184 if (unlikely(details))
1186 if (pte_file(ptent)) {
1187 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1188 print_bad_pte(vma, addr, ptent, NULL);
1190 swp_entry_t entry = pte_to_swp_entry(ptent);
1192 if (!non_swap_entry(entry))
1194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
1197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198 } while (pte++, addr += PAGE_SIZE, addr != end);
1200 add_mm_rss_vec(mm, rss);
1201 arch_leave_lazy_mmu_mode();
1202 pte_unmap_unlock(start_pte, ptl);
1205 * mmu_gather ran out of room to batch pages, we break out of
1206 * the PTE lock to avoid doing the potential expensive TLB invalidate
1207 * and page-free while holding it.
1219 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1220 struct vm_area_struct *vma, pud_t *pud,
1221 unsigned long addr, unsigned long end,
1222 struct zap_details *details)
1227 pmd = pmd_offset(pud, addr);
1229 next = pmd_addr_end(addr, end);
1230 if (pmd_trans_huge(*pmd)) {
1231 if (next-addr != HPAGE_PMD_SIZE) {
1232 VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem));
1233 split_huge_page_pmd(vma->vm_mm, pmd);
1234 } else if (zap_huge_pmd(tlb, vma, pmd))
1238 if (pmd_none_or_clear_bad(pmd))
1240 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1242 } while (pmd++, addr = next, addr != end);
1247 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1248 struct vm_area_struct *vma, pgd_t *pgd,
1249 unsigned long addr, unsigned long end,
1250 struct zap_details *details)
1255 pud = pud_offset(pgd, addr);
1257 next = pud_addr_end(addr, end);
1258 if (pud_none_or_clear_bad(pud))
1260 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1261 } while (pud++, addr = next, addr != end);
1266 static unsigned long unmap_page_range(struct mmu_gather *tlb,
1267 struct vm_area_struct *vma,
1268 unsigned long addr, unsigned long end,
1269 struct zap_details *details)
1274 if (details && !details->check_mapping && !details->nonlinear_vma)
1277 BUG_ON(addr >= end);
1278 mem_cgroup_uncharge_start();
1279 tlb_start_vma(tlb, vma);
1280 pgd = pgd_offset(vma->vm_mm, addr);
1282 next = pgd_addr_end(addr, end);
1283 if (pgd_none_or_clear_bad(pgd))
1285 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1286 } while (pgd++, addr = next, addr != end);
1287 tlb_end_vma(tlb, vma);
1288 mem_cgroup_uncharge_end();
1294 * unmap_vmas - unmap a range of memory covered by a list of vma's
1295 * @tlb: address of the caller's struct mmu_gather
1296 * @vma: the starting vma
1297 * @start_addr: virtual address at which to start unmapping
1298 * @end_addr: virtual address at which to end unmapping
1299 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
1300 * @details: details of nonlinear truncation or shared cache invalidation
1302 * Returns the end address of the unmapping (restart addr if interrupted).
1304 * Unmap all pages in the vma list.
1306 * Only addresses between `start' and `end' will be unmapped.
1308 * The VMA list must be sorted in ascending virtual address order.
1310 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1311 * range after unmap_vmas() returns. So the only responsibility here is to
1312 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1313 * drops the lock and schedules.
1315 unsigned long unmap_vmas(struct mmu_gather *tlb,
1316 struct vm_area_struct *vma, unsigned long start_addr,
1317 unsigned long end_addr, unsigned long *nr_accounted,
1318 struct zap_details *details)
1320 unsigned long start = start_addr;
1321 struct mm_struct *mm = vma->vm_mm;
1323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1327 start = max(vma->vm_start, start_addr);
1328 if (start >= vma->vm_end)
1330 end = min(vma->vm_end, end_addr);
1331 if (end <= vma->vm_start)
1334 if (vma->vm_flags & VM_ACCOUNT)
1335 *nr_accounted += (end - start) >> PAGE_SHIFT;
1337 if (unlikely(is_pfn_mapping(vma)))
1338 untrack_pfn_vma(vma, 0, 0);
1340 while (start != end) {
1341 if (unlikely(is_vm_hugetlb_page(vma))) {
1343 * It is undesirable to test vma->vm_file as it
1344 * should be non-null for valid hugetlb area.
1345 * However, vm_file will be NULL in the error
1346 * cleanup path of do_mmap_pgoff. When
1347 * hugetlbfs ->mmap method fails,
1348 * do_mmap_pgoff() nullifies vma->vm_file
1349 * before calling this function to clean up.
1350 * Since no pte has actually been setup, it is
1351 * safe to do nothing in this case.
1354 unmap_hugepage_range(vma, start, end, NULL);
1358 start = unmap_page_range(tlb, vma, start, end, details);
1362 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1363 return start; /* which is now the end (or restart) address */
1367 * zap_page_range - remove user pages in a given range
1368 * @vma: vm_area_struct holding the applicable pages
1369 * @address: starting address of pages to zap
1370 * @size: number of bytes to zap
1371 * @details: details of nonlinear truncation or shared cache invalidation
1373 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1374 unsigned long size, struct zap_details *details)
1376 struct mm_struct *mm = vma->vm_mm;
1377 struct mmu_gather tlb;
1378 unsigned long end = address + size;
1379 unsigned long nr_accounted = 0;
1382 tlb_gather_mmu(&tlb, mm, 0);
1383 update_hiwater_rss(mm);
1384 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
1385 tlb_finish_mmu(&tlb, address, end);
1390 * zap_vma_ptes - remove ptes mapping the vma
1391 * @vma: vm_area_struct holding ptes to be zapped
1392 * @address: starting address of pages to zap
1393 * @size: number of bytes to zap
1395 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1397 * The entire address range must be fully contained within the vma.
1399 * Returns 0 if successful.
1401 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1404 if (address < vma->vm_start || address + size > vma->vm_end ||
1405 !(vma->vm_flags & VM_PFNMAP))
1407 zap_page_range(vma, address, size, NULL);
1410 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1413 * follow_page - look up a page descriptor from a user-virtual address
1414 * @vma: vm_area_struct mapping @address
1415 * @address: virtual address to look up
1416 * @flags: flags modifying lookup behaviour
1418 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1420 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1421 * an error pointer if there is a mapping to something not represented
1422 * by a page descriptor (see also vm_normal_page()).
1424 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1433 struct mm_struct *mm = vma->vm_mm;
1435 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1436 if (!IS_ERR(page)) {
1437 BUG_ON(flags & FOLL_GET);
1442 pgd = pgd_offset(mm, address);
1443 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1446 pud = pud_offset(pgd, address);
1449 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1450 BUG_ON(flags & FOLL_GET);
1451 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1454 if (unlikely(pud_bad(*pud)))
1457 pmd = pmd_offset(pud, address);
1460 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1461 BUG_ON(flags & FOLL_GET);
1462 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1465 if (pmd_trans_huge(*pmd)) {
1466 if (flags & FOLL_SPLIT) {
1467 split_huge_page_pmd(mm, pmd);
1468 goto split_fallthrough;
1470 spin_lock(&mm->page_table_lock);
1471 if (likely(pmd_trans_huge(*pmd))) {
1472 if (unlikely(pmd_trans_splitting(*pmd))) {
1473 spin_unlock(&mm->page_table_lock);
1474 wait_split_huge_page(vma->anon_vma, pmd);
1476 page = follow_trans_huge_pmd(mm, address,
1478 spin_unlock(&mm->page_table_lock);
1482 spin_unlock(&mm->page_table_lock);
1486 if (unlikely(pmd_bad(*pmd)))
1489 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1492 if (!pte_present(pte))
1494 if ((flags & FOLL_WRITE) && !pte_write(pte))
1497 page = vm_normal_page(vma, address, pte);
1498 if (unlikely(!page)) {
1499 if ((flags & FOLL_DUMP) ||
1500 !is_zero_pfn(pte_pfn(pte)))
1502 page = pte_page(pte);
1505 if (flags & FOLL_GET)
1507 if (flags & FOLL_TOUCH) {
1508 if ((flags & FOLL_WRITE) &&
1509 !pte_dirty(pte) && !PageDirty(page))
1510 set_page_dirty(page);
1512 * pte_mkyoung() would be more correct here, but atomic care
1513 * is needed to avoid losing the dirty bit: it is easier to use
1514 * mark_page_accessed().
1516 mark_page_accessed(page);
1518 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1520 * The preliminary mapping check is mainly to avoid the
1521 * pointless overhead of lock_page on the ZERO_PAGE
1522 * which might bounce very badly if there is contention.
1524 * If the page is already locked, we don't need to
1525 * handle it now - vmscan will handle it later if and
1526 * when it attempts to reclaim the page.
1528 if (page->mapping && trylock_page(page)) {
1529 lru_add_drain(); /* push cached pages to LRU */
1531 * Because we lock page here and migration is
1532 * blocked by the pte's page reference, we need
1533 * only check for file-cache page truncation.
1536 mlock_vma_page(page);
1541 pte_unmap_unlock(ptep, ptl);
1546 pte_unmap_unlock(ptep, ptl);
1547 return ERR_PTR(-EFAULT);
1550 pte_unmap_unlock(ptep, ptl);
1556 * When core dumping an enormous anonymous area that nobody
1557 * has touched so far, we don't want to allocate unnecessary pages or
1558 * page tables. Return error instead of NULL to skip handle_mm_fault,
1559 * then get_dump_page() will return NULL to leave a hole in the dump.
1560 * But we can only make this optimization where a hole would surely
1561 * be zero-filled if handle_mm_fault() actually did handle it.
1563 if ((flags & FOLL_DUMP) &&
1564 (!vma->vm_ops || !vma->vm_ops->fault))
1565 return ERR_PTR(-EFAULT);
1569 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1571 return stack_guard_page_start(vma, addr) ||
1572 stack_guard_page_end(vma, addr+PAGE_SIZE);
1576 * __get_user_pages() - pin user pages in memory
1577 * @tsk: task_struct of target task
1578 * @mm: mm_struct of target mm
1579 * @start: starting user address
1580 * @nr_pages: number of pages from start to pin
1581 * @gup_flags: flags modifying pin behaviour
1582 * @pages: array that receives pointers to the pages pinned.
1583 * Should be at least nr_pages long. Or NULL, if caller
1584 * only intends to ensure the pages are faulted in.
1585 * @vmas: array of pointers to vmas corresponding to each page.
1586 * Or NULL if the caller does not require them.
1587 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1589 * Returns number of pages pinned. This may be fewer than the number
1590 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1591 * were pinned, returns -errno. Each page returned must be released
1592 * with a put_page() call when it is finished with. vmas will only
1593 * remain valid while mmap_sem is held.
1595 * Must be called with mmap_sem held for read or write.
1597 * __get_user_pages walks a process's page tables and takes a reference to
1598 * each struct page that each user address corresponds to at a given
1599 * instant. That is, it takes the page that would be accessed if a user
1600 * thread accesses the given user virtual address at that instant.
1602 * This does not guarantee that the page exists in the user mappings when
1603 * __get_user_pages returns, and there may even be a completely different
1604 * page there in some cases (eg. if mmapped pagecache has been invalidated
1605 * and subsequently re faulted). However it does guarantee that the page
1606 * won't be freed completely. And mostly callers simply care that the page
1607 * contains data that was valid *at some point in time*. Typically, an IO
1608 * or similar operation cannot guarantee anything stronger anyway because
1609 * locks can't be held over the syscall boundary.
1611 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1612 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1613 * appropriate) must be called after the page is finished with, and
1614 * before put_page is called.
1616 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1617 * or mmap_sem contention, and if waiting is needed to pin all pages,
1618 * *@nonblocking will be set to 0.
1620 * In most cases, get_user_pages or get_user_pages_fast should be used
1621 * instead of __get_user_pages. __get_user_pages should be used only if
1622 * you need some special @gup_flags.
1624 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1625 unsigned long start, int nr_pages, unsigned int gup_flags,
1626 struct page **pages, struct vm_area_struct **vmas,
1630 unsigned long vm_flags;
1635 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1638 * Require read or write permissions.
1639 * If FOLL_FORCE is set, we only require the "MAY" flags.
1641 vm_flags = (gup_flags & FOLL_WRITE) ?
1642 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1643 vm_flags &= (gup_flags & FOLL_FORCE) ?
1644 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1648 struct vm_area_struct *vma;
1650 vma = find_extend_vma(mm, start);
1651 if (!vma && in_gate_area(mm, start)) {
1652 unsigned long pg = start & PAGE_MASK;
1658 /* user gate pages are read-only */
1659 if (gup_flags & FOLL_WRITE)
1660 return i ? : -EFAULT;
1662 pgd = pgd_offset_k(pg);
1664 pgd = pgd_offset_gate(mm, pg);
1665 BUG_ON(pgd_none(*pgd));
1666 pud = pud_offset(pgd, pg);
1667 BUG_ON(pud_none(*pud));
1668 pmd = pmd_offset(pud, pg);
1670 return i ? : -EFAULT;
1671 VM_BUG_ON(pmd_trans_huge(*pmd));
1672 pte = pte_offset_map(pmd, pg);
1673 if (pte_none(*pte)) {
1675 return i ? : -EFAULT;
1677 vma = get_gate_vma(mm);
1681 page = vm_normal_page(vma, start, *pte);
1683 if (!(gup_flags & FOLL_DUMP) &&
1684 is_zero_pfn(pte_pfn(*pte)))
1685 page = pte_page(*pte);
1688 return i ? : -EFAULT;
1699 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1700 !(vm_flags & vma->vm_flags))
1701 return i ? : -EFAULT;
1703 if (is_vm_hugetlb_page(vma)) {
1704 i = follow_hugetlb_page(mm, vma, pages, vmas,
1705 &start, &nr_pages, i, gup_flags);
1711 unsigned int foll_flags = gup_flags;
1714 * If we have a pending SIGKILL, don't keep faulting
1715 * pages and potentially allocating memory.
1717 if (unlikely(fatal_signal_pending(current)))
1718 return i ? i : -ERESTARTSYS;
1721 while (!(page = follow_page(vma, start, foll_flags))) {
1723 unsigned int fault_flags = 0;
1725 /* For mlock, just skip the stack guard page. */
1726 if (foll_flags & FOLL_MLOCK) {
1727 if (stack_guard_page(vma, start))
1730 if (foll_flags & FOLL_WRITE)
1731 fault_flags |= FAULT_FLAG_WRITE;
1733 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1734 if (foll_flags & FOLL_NOWAIT)
1735 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1737 ret = handle_mm_fault(mm, vma, start,
1740 if (ret & VM_FAULT_ERROR) {
1741 if (ret & VM_FAULT_OOM)
1742 return i ? i : -ENOMEM;
1743 if (ret & (VM_FAULT_HWPOISON |
1744 VM_FAULT_HWPOISON_LARGE)) {
1747 else if (gup_flags & FOLL_HWPOISON)
1752 if (ret & VM_FAULT_SIGBUS)
1753 return i ? i : -EFAULT;
1758 if (ret & VM_FAULT_MAJOR)
1764 if (ret & VM_FAULT_RETRY) {
1771 * The VM_FAULT_WRITE bit tells us that
1772 * do_wp_page has broken COW when necessary,
1773 * even if maybe_mkwrite decided not to set
1774 * pte_write. We can thus safely do subsequent
1775 * page lookups as if they were reads. But only
1776 * do so when looping for pte_write is futile:
1777 * in some cases userspace may also be wanting
1778 * to write to the gotten user page, which a
1779 * read fault here might prevent (a readonly
1780 * page might get reCOWed by userspace write).
1782 if ((ret & VM_FAULT_WRITE) &&
1783 !(vma->vm_flags & VM_WRITE))
1784 foll_flags &= ~FOLL_WRITE;
1789 return i ? i : PTR_ERR(page);
1793 flush_anon_page(vma, page, start);
1794 flush_dcache_page(page);
1802 } while (nr_pages && start < vma->vm_end);
1806 EXPORT_SYMBOL(__get_user_pages);
1809 * fixup_user_fault() - manually resolve a user page fault
1810 * @tsk: the task_struct to use for page fault accounting, or
1811 * NULL if faults are not to be recorded.
1812 * @mm: mm_struct of target mm
1813 * @address: user address
1814 * @fault_flags:flags to pass down to handle_mm_fault()
1816 * This is meant to be called in the specific scenario where for locking reasons
1817 * we try to access user memory in atomic context (within a pagefault_disable()
1818 * section), this returns -EFAULT, and we want to resolve the user fault before
1821 * Typically this is meant to be used by the futex code.
1823 * The main difference with get_user_pages() is that this function will
1824 * unconditionally call handle_mm_fault() which will in turn perform all the
1825 * necessary SW fixup of the dirty and young bits in the PTE, while
1826 * handle_mm_fault() only guarantees to update these in the struct page.
1828 * This is important for some architectures where those bits also gate the
1829 * access permission to the page because they are maintained in software. On
1830 * such architectures, gup() will not be enough to make a subsequent access
1833 * This should be called with the mm_sem held for read.
1835 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1836 unsigned long address, unsigned int fault_flags)
1838 struct vm_area_struct *vma;
1841 vma = find_extend_vma(mm, address);
1842 if (!vma || address < vma->vm_start)
1845 ret = handle_mm_fault(mm, vma, address, fault_flags);
1846 if (ret & VM_FAULT_ERROR) {
1847 if (ret & VM_FAULT_OOM)
1849 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1851 if (ret & VM_FAULT_SIGBUS)
1856 if (ret & VM_FAULT_MAJOR)
1865 * get_user_pages() - pin user pages in memory
1866 * @tsk: the task_struct to use for page fault accounting, or
1867 * NULL if faults are not to be recorded.
1868 * @mm: mm_struct of target mm
1869 * @start: starting user address
1870 * @nr_pages: number of pages from start to pin
1871 * @write: whether pages will be written to by the caller
1872 * @force: whether to force write access even if user mapping is
1873 * readonly. This will result in the page being COWed even
1874 * in MAP_SHARED mappings. You do not want this.
1875 * @pages: array that receives pointers to the pages pinned.
1876 * Should be at least nr_pages long. Or NULL, if caller
1877 * only intends to ensure the pages are faulted in.
1878 * @vmas: array of pointers to vmas corresponding to each page.
1879 * Or NULL if the caller does not require them.
1881 * Returns number of pages pinned. This may be fewer than the number
1882 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1883 * were pinned, returns -errno. Each page returned must be released
1884 * with a put_page() call when it is finished with. vmas will only
1885 * remain valid while mmap_sem is held.
1887 * Must be called with mmap_sem held for read or write.
1889 * get_user_pages walks a process's page tables and takes a reference to
1890 * each struct page that each user address corresponds to at a given
1891 * instant. That is, it takes the page that would be accessed if a user
1892 * thread accesses the given user virtual address at that instant.
1894 * This does not guarantee that the page exists in the user mappings when
1895 * get_user_pages returns, and there may even be a completely different
1896 * page there in some cases (eg. if mmapped pagecache has been invalidated
1897 * and subsequently re faulted). However it does guarantee that the page
1898 * won't be freed completely. And mostly callers simply care that the page
1899 * contains data that was valid *at some point in time*. Typically, an IO
1900 * or similar operation cannot guarantee anything stronger anyway because
1901 * locks can't be held over the syscall boundary.
1903 * If write=0, the page must not be written to. If the page is written to,
1904 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
1905 * after the page is finished with, and before put_page is called.
1907 * get_user_pages is typically used for fewer-copy IO operations, to get a
1908 * handle on the memory by some means other than accesses via the user virtual
1909 * addresses. The pages may be submitted for DMA to devices or accessed via
1910 * their kernel linear mapping (via the kmap APIs). Care should be taken to
1911 * use the correct cache flushing APIs.
1913 * See also get_user_pages_fast, for performance critical applications.
1915 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1916 unsigned long start, int nr_pages, int write, int force,
1917 struct page **pages, struct vm_area_struct **vmas)
1919 int flags = FOLL_TOUCH;
1924 flags |= FOLL_WRITE;
1926 flags |= FOLL_FORCE;
1928 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
1931 EXPORT_SYMBOL(get_user_pages);
1934 * get_dump_page() - pin user page in memory while writing it to core dump
1935 * @addr: user address
1937 * Returns struct page pointer of user page pinned for dump,
1938 * to be freed afterwards by page_cache_release() or put_page().
1940 * Returns NULL on any kind of failure - a hole must then be inserted into
1941 * the corefile, to preserve alignment with its headers; and also returns
1942 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
1943 * allowing a hole to be left in the corefile to save diskspace.
1945 * Called without mmap_sem, but after all other threads have been killed.
1947 #ifdef CONFIG_ELF_CORE
1948 struct page *get_dump_page(unsigned long addr)
1950 struct vm_area_struct *vma;
1953 if (__get_user_pages(current, current->mm, addr, 1,
1954 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
1957 flush_cache_page(vma, addr, page_to_pfn(page));
1960 #endif /* CONFIG_ELF_CORE */
1962 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1965 pgd_t * pgd = pgd_offset(mm, addr);
1966 pud_t * pud = pud_alloc(mm, pgd, addr);
1968 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1970 VM_BUG_ON(pmd_trans_huge(*pmd));
1971 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1978 * This is the old fallback for page remapping.
1980 * For historical reasons, it only allows reserved pages. Only
1981 * old drivers should use this, and they needed to mark their
1982 * pages reserved for the old functions anyway.
1984 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1985 struct page *page, pgprot_t prot)
1987 struct mm_struct *mm = vma->vm_mm;
1996 flush_dcache_page(page);
1997 pte = get_locked_pte(mm, addr, &ptl);
2001 if (!pte_none(*pte))
2004 /* Ok, finally just insert the thing.. */
2006 inc_mm_counter_fast(mm, MM_FILEPAGES);
2007 page_add_file_rmap(page);
2008 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2011 pte_unmap_unlock(pte, ptl);
2014 pte_unmap_unlock(pte, ptl);
2020 * vm_insert_page - insert single page into user vma
2021 * @vma: user vma to map to
2022 * @addr: target user address of this page
2023 * @page: source kernel page
2025 * This allows drivers to insert individual pages they've allocated
2028 * The page has to be a nice clean _individual_ kernel allocation.
2029 * If you allocate a compound page, you need to have marked it as
2030 * such (__GFP_COMP), or manually just split the page up yourself
2031 * (see split_page()).
2033 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2034 * took an arbitrary page protection parameter. This doesn't allow
2035 * that. Your vma protection will have to be set up correctly, which
2036 * means that if you want a shared writable mapping, you'd better
2037 * ask for a shared writable mapping!
2039 * The page does not need to be reserved.
2041 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2044 if (addr < vma->vm_start || addr >= vma->vm_end)
2046 if (!page_count(page))
2048 vma->vm_flags |= VM_INSERTPAGE;
2049 return insert_page(vma, addr, page, vma->vm_page_prot);
2051 EXPORT_SYMBOL(vm_insert_page);
2053 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2054 unsigned long pfn, pgprot_t prot)
2056 struct mm_struct *mm = vma->vm_mm;
2062 pte = get_locked_pte(mm, addr, &ptl);
2066 if (!pte_none(*pte))
2069 /* Ok, finally just insert the thing.. */
2070 entry = pte_mkspecial(pfn_pte(pfn, prot));
2071 set_pte_at(mm, addr, pte, entry);
2072 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2076 pte_unmap_unlock(pte, ptl);
2082 * vm_insert_pfn - insert single pfn into user vma
2083 * @vma: user vma to map to
2084 * @addr: target user address of this page
2085 * @pfn: source kernel pfn
2087 * Similar to vm_inert_page, this allows drivers to insert individual pages
2088 * they've allocated into a user vma. Same comments apply.
2090 * This function should only be called from a vm_ops->fault handler, and
2091 * in that case the handler should return NULL.
2093 * vma cannot be a COW mapping.
2095 * As this is called only for pages that do not currently exist, we
2096 * do not need to flush old virtual caches or the TLB.
2098 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2102 pgprot_t pgprot = vma->vm_page_prot;
2104 * Technically, architectures with pte_special can avoid all these
2105 * restrictions (same for remap_pfn_range). However we would like
2106 * consistency in testing and feature parity among all, so we should
2107 * try to keep these invariants in place for everybody.
2109 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2110 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2111 (VM_PFNMAP|VM_MIXEDMAP));
2112 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2113 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2115 if (addr < vma->vm_start || addr >= vma->vm_end)
2117 if (track_pfn_vma_new(vma, &pgprot, pfn, PAGE_SIZE))
2120 ret = insert_pfn(vma, addr, pfn, pgprot);
2123 untrack_pfn_vma(vma, pfn, PAGE_SIZE);
2127 EXPORT_SYMBOL(vm_insert_pfn);
2129 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2132 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2134 if (addr < vma->vm_start || addr >= vma->vm_end)
2138 * If we don't have pte special, then we have to use the pfn_valid()
2139 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2140 * refcount the page if pfn_valid is true (hence insert_page rather
2141 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2142 * without pte special, it would there be refcounted as a normal page.
2144 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2147 page = pfn_to_page(pfn);
2148 return insert_page(vma, addr, page, vma->vm_page_prot);
2150 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2152 EXPORT_SYMBOL(vm_insert_mixed);
2155 * maps a range of physical memory into the requested pages. the old
2156 * mappings are removed. any references to nonexistent pages results
2157 * in null mappings (currently treated as "copy-on-access")
2159 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2160 unsigned long addr, unsigned long end,
2161 unsigned long pfn, pgprot_t prot)
2166 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2169 arch_enter_lazy_mmu_mode();
2171 BUG_ON(!pte_none(*pte));
2172 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2174 } while (pte++, addr += PAGE_SIZE, addr != end);
2175 arch_leave_lazy_mmu_mode();
2176 pte_unmap_unlock(pte - 1, ptl);
2180 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2181 unsigned long addr, unsigned long end,
2182 unsigned long pfn, pgprot_t prot)
2187 pfn -= addr >> PAGE_SHIFT;
2188 pmd = pmd_alloc(mm, pud, addr);
2191 VM_BUG_ON(pmd_trans_huge(*pmd));
2193 next = pmd_addr_end(addr, end);
2194 if (remap_pte_range(mm, pmd, addr, next,
2195 pfn + (addr >> PAGE_SHIFT), prot))
2197 } while (pmd++, addr = next, addr != end);
2201 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2202 unsigned long addr, unsigned long end,
2203 unsigned long pfn, pgprot_t prot)
2208 pfn -= addr >> PAGE_SHIFT;
2209 pud = pud_alloc(mm, pgd, addr);
2213 next = pud_addr_end(addr, end);
2214 if (remap_pmd_range(mm, pud, addr, next,
2215 pfn + (addr >> PAGE_SHIFT), prot))
2217 } while (pud++, addr = next, addr != end);
2222 * remap_pfn_range - remap kernel memory to userspace
2223 * @vma: user vma to map to
2224 * @addr: target user address to start at
2225 * @pfn: physical address of kernel memory
2226 * @size: size of map area
2227 * @prot: page protection flags for this mapping
2229 * Note: this is only safe if the mm semaphore is held when called.
2231 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2232 unsigned long pfn, unsigned long size, pgprot_t prot)
2236 unsigned long end = addr + PAGE_ALIGN(size);
2237 struct mm_struct *mm = vma->vm_mm;
2241 * Physically remapped pages are special. Tell the
2242 * rest of the world about it:
2243 * VM_IO tells people not to look at these pages
2244 * (accesses can have side effects).
2245 * VM_RESERVED is specified all over the place, because
2246 * in 2.4 it kept swapout's vma scan off this vma; but
2247 * in 2.6 the LRU scan won't even find its pages, so this
2248 * flag means no more than count its pages in reserved_vm,
2249 * and omit it from core dump, even when VM_IO turned off.
2250 * VM_PFNMAP tells the core MM that the base pages are just
2251 * raw PFN mappings, and do not have a "struct page" associated
2254 * There's a horrible special case to handle copy-on-write
2255 * behaviour that some programs depend on. We mark the "original"
2256 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2258 if (addr == vma->vm_start && end == vma->vm_end) {
2259 vma->vm_pgoff = pfn;
2260 vma->vm_flags |= VM_PFN_AT_MMAP;
2261 } else if (is_cow_mapping(vma->vm_flags))
2264 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
2266 err = track_pfn_vma_new(vma, &prot, pfn, PAGE_ALIGN(size));
2269 * To indicate that track_pfn related cleanup is not
2270 * needed from higher level routine calling unmap_vmas
2272 vma->vm_flags &= ~(VM_IO | VM_RESERVED | VM_PFNMAP);
2273 vma->vm_flags &= ~VM_PFN_AT_MMAP;
2277 BUG_ON(addr >= end);
2278 pfn -= addr >> PAGE_SHIFT;
2279 pgd = pgd_offset(mm, addr);
2280 flush_cache_range(vma, addr, end);
2282 next = pgd_addr_end(addr, end);
2283 err = remap_pud_range(mm, pgd, addr, next,
2284 pfn + (addr >> PAGE_SHIFT), prot);
2287 } while (pgd++, addr = next, addr != end);
2290 untrack_pfn_vma(vma, pfn, PAGE_ALIGN(size));
2294 EXPORT_SYMBOL(remap_pfn_range);
2296 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2297 unsigned long addr, unsigned long end,
2298 pte_fn_t fn, void *data)
2303 spinlock_t *uninitialized_var(ptl);
2305 pte = (mm == &init_mm) ?
2306 pte_alloc_kernel(pmd, addr) :
2307 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 BUG_ON(pmd_huge(*pmd));
2313 arch_enter_lazy_mmu_mode();
2315 token = pmd_pgtable(*pmd);
2318 err = fn(pte++, token, addr, data);
2321 } while (addr += PAGE_SIZE, addr != end);
2323 arch_leave_lazy_mmu_mode();
2326 pte_unmap_unlock(pte-1, ptl);
2330 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2331 unsigned long addr, unsigned long end,
2332 pte_fn_t fn, void *data)
2338 BUG_ON(pud_huge(*pud));
2340 pmd = pmd_alloc(mm, pud, addr);
2344 next = pmd_addr_end(addr, end);
2345 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2348 } while (pmd++, addr = next, addr != end);
2352 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2353 unsigned long addr, unsigned long end,
2354 pte_fn_t fn, void *data)
2360 pud = pud_alloc(mm, pgd, addr);
2364 next = pud_addr_end(addr, end);
2365 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2368 } while (pud++, addr = next, addr != end);
2373 * Scan a region of virtual memory, filling in page tables as necessary
2374 * and calling a provided function on each leaf page table.
2376 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2377 unsigned long size, pte_fn_t fn, void *data)
2381 unsigned long end = addr + size;
2384 BUG_ON(addr >= end);
2385 pgd = pgd_offset(mm, addr);
2387 next = pgd_addr_end(addr, end);
2388 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2391 } while (pgd++, addr = next, addr != end);
2395 EXPORT_SYMBOL_GPL(apply_to_page_range);
2398 * handle_pte_fault chooses page fault handler according to an entry
2399 * which was read non-atomically. Before making any commitment, on
2400 * those architectures or configurations (e.g. i386 with PAE) which
2401 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2402 * must check under lock before unmapping the pte and proceeding
2403 * (but do_wp_page is only called after already making such a check;
2404 * and do_anonymous_page can safely check later on).
2406 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2407 pte_t *page_table, pte_t orig_pte)
2410 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2411 if (sizeof(pte_t) > sizeof(unsigned long)) {
2412 spinlock_t *ptl = pte_lockptr(mm, pmd);
2414 same = pte_same(*page_table, orig_pte);
2418 pte_unmap(page_table);
2422 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2425 * If the source page was a PFN mapping, we don't have
2426 * a "struct page" for it. We do a best-effort copy by
2427 * just copying from the original user address. If that
2428 * fails, we just zero-fill it. Live with it.
2430 if (unlikely(!src)) {
2431 void *kaddr = kmap_atomic(dst, KM_USER0);
2432 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2435 * This really shouldn't fail, because the page is there
2436 * in the page tables. But it might just be unreadable,
2437 * in which case we just give up and fill the result with
2440 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2442 kunmap_atomic(kaddr, KM_USER0);
2443 flush_dcache_page(dst);
2445 copy_user_highpage(dst, src, va, vma);
2449 * This routine handles present pages, when users try to write
2450 * to a shared page. It is done by copying the page to a new address
2451 * and decrementing the shared-page counter for the old page.
2453 * Note that this routine assumes that the protection checks have been
2454 * done by the caller (the low-level page fault routine in most cases).
2455 * Thus we can safely just mark it writable once we've done any necessary
2458 * We also mark the page dirty at this point even though the page will
2459 * change only once the write actually happens. This avoids a few races,
2460 * and potentially makes it more efficient.
2462 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2463 * but allow concurrent faults), with pte both mapped and locked.
2464 * We return with mmap_sem still held, but pte unmapped and unlocked.
2466 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2467 unsigned long address, pte_t *page_table, pmd_t *pmd,
2468 spinlock_t *ptl, pte_t orig_pte)
2471 struct page *old_page, *new_page;
2474 int page_mkwrite = 0;
2475 struct page *dirty_page = NULL;
2477 old_page = vm_normal_page(vma, address, orig_pte);
2480 * VM_MIXEDMAP !pfn_valid() case
2482 * We should not cow pages in a shared writeable mapping.
2483 * Just mark the pages writable as we can't do any dirty
2484 * accounting on raw pfn maps.
2486 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2487 (VM_WRITE|VM_SHARED))
2493 * Take out anonymous pages first, anonymous shared vmas are
2494 * not dirty accountable.
2496 if (PageAnon(old_page) && !PageKsm(old_page)) {
2497 if (!trylock_page(old_page)) {
2498 page_cache_get(old_page);
2499 pte_unmap_unlock(page_table, ptl);
2500 lock_page(old_page);
2501 page_table = pte_offset_map_lock(mm, pmd, address,
2503 if (!pte_same(*page_table, orig_pte)) {
2504 unlock_page(old_page);
2507 page_cache_release(old_page);
2509 if (reuse_swap_page(old_page)) {
2511 * The page is all ours. Move it to our anon_vma so
2512 * the rmap code will not search our parent or siblings.
2513 * Protected against the rmap code by the page lock.
2515 page_move_anon_rmap(old_page, vma, address);
2516 unlock_page(old_page);
2519 unlock_page(old_page);
2520 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2521 (VM_WRITE|VM_SHARED))) {
2523 * Only catch write-faults on shared writable pages,
2524 * read-only shared pages can get COWed by
2525 * get_user_pages(.write=1, .force=1).
2527 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2528 struct vm_fault vmf;
2531 vmf.virtual_address = (void __user *)(address &
2533 vmf.pgoff = old_page->index;
2534 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2535 vmf.page = old_page;
2538 * Notify the address space that the page is about to
2539 * become writable so that it can prohibit this or wait
2540 * for the page to get into an appropriate state.
2542 * We do this without the lock held, so that it can
2543 * sleep if it needs to.
2545 page_cache_get(old_page);
2546 pte_unmap_unlock(page_table, ptl);
2548 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2550 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2552 goto unwritable_page;
2554 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2555 lock_page(old_page);
2556 if (!old_page->mapping) {
2557 ret = 0; /* retry the fault */
2558 unlock_page(old_page);
2559 goto unwritable_page;
2562 VM_BUG_ON(!PageLocked(old_page));
2565 * Since we dropped the lock we need to revalidate
2566 * the PTE as someone else may have changed it. If
2567 * they did, we just return, as we can count on the
2568 * MMU to tell us if they didn't also make it writable.
2570 page_table = pte_offset_map_lock(mm, pmd, address,
2572 if (!pte_same(*page_table, orig_pte)) {
2573 unlock_page(old_page);
2579 dirty_page = old_page;
2580 get_page(dirty_page);
2583 flush_cache_page(vma, address, pte_pfn(orig_pte));
2584 entry = pte_mkyoung(orig_pte);
2585 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2586 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2587 update_mmu_cache(vma, address, page_table);
2588 pte_unmap_unlock(page_table, ptl);
2589 ret |= VM_FAULT_WRITE;
2595 * Yes, Virginia, this is actually required to prevent a race
2596 * with clear_page_dirty_for_io() from clearing the page dirty
2597 * bit after it clear all dirty ptes, but before a racing
2598 * do_wp_page installs a dirty pte.
2600 * __do_fault is protected similarly.
2602 if (!page_mkwrite) {
2603 wait_on_page_locked(dirty_page);
2604 set_page_dirty_balance(dirty_page, page_mkwrite);
2606 put_page(dirty_page);
2608 struct address_space *mapping = dirty_page->mapping;
2610 set_page_dirty(dirty_page);
2611 unlock_page(dirty_page);
2612 page_cache_release(dirty_page);
2615 * Some device drivers do not set page.mapping
2616 * but still dirty their pages
2618 balance_dirty_pages_ratelimited(mapping);
2622 /* file_update_time outside page_lock */
2624 file_update_time(vma->vm_file);
2630 * Ok, we need to copy. Oh, well..
2632 page_cache_get(old_page);
2634 pte_unmap_unlock(page_table, ptl);
2636 if (unlikely(anon_vma_prepare(vma)))
2639 if (is_zero_pfn(pte_pfn(orig_pte))) {
2640 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2644 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2647 cow_user_page(new_page, old_page, address, vma);
2649 __SetPageUptodate(new_page);
2651 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2655 * Re-check the pte - we dropped the lock
2657 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2658 if (likely(pte_same(*page_table, orig_pte))) {
2660 if (!PageAnon(old_page)) {
2661 dec_mm_counter_fast(mm, MM_FILEPAGES);
2662 inc_mm_counter_fast(mm, MM_ANONPAGES);
2665 inc_mm_counter_fast(mm, MM_ANONPAGES);
2666 flush_cache_page(vma, address, pte_pfn(orig_pte));
2667 entry = mk_pte(new_page, vma->vm_page_prot);
2668 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2670 * Clear the pte entry and flush it first, before updating the
2671 * pte with the new entry. This will avoid a race condition
2672 * seen in the presence of one thread doing SMC and another
2675 ptep_clear_flush(vma, address, page_table);
2676 page_add_new_anon_rmap(new_page, vma, address);
2678 * We call the notify macro here because, when using secondary
2679 * mmu page tables (such as kvm shadow page tables), we want the
2680 * new page to be mapped directly into the secondary page table.
2682 set_pte_at_notify(mm, address, page_table, entry);
2683 update_mmu_cache(vma, address, page_table);
2686 * Only after switching the pte to the new page may
2687 * we remove the mapcount here. Otherwise another
2688 * process may come and find the rmap count decremented
2689 * before the pte is switched to the new page, and
2690 * "reuse" the old page writing into it while our pte
2691 * here still points into it and can be read by other
2694 * The critical issue is to order this
2695 * page_remove_rmap with the ptp_clear_flush above.
2696 * Those stores are ordered by (if nothing else,)
2697 * the barrier present in the atomic_add_negative
2698 * in page_remove_rmap.
2700 * Then the TLB flush in ptep_clear_flush ensures that
2701 * no process can access the old page before the
2702 * decremented mapcount is visible. And the old page
2703 * cannot be reused until after the decremented
2704 * mapcount is visible. So transitively, TLBs to
2705 * old page will be flushed before it can be reused.
2707 page_remove_rmap(old_page);
2710 /* Free the old page.. */
2711 new_page = old_page;
2712 ret |= VM_FAULT_WRITE;
2714 mem_cgroup_uncharge_page(new_page);
2717 page_cache_release(new_page);
2719 pte_unmap_unlock(page_table, ptl);
2722 * Don't let another task, with possibly unlocked vma,
2723 * keep the mlocked page.
2725 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2726 lock_page(old_page); /* LRU manipulation */
2727 munlock_vma_page(old_page);
2728 unlock_page(old_page);
2730 page_cache_release(old_page);
2734 page_cache_release(new_page);
2738 unlock_page(old_page);
2739 page_cache_release(old_page);
2741 page_cache_release(old_page);
2743 return VM_FAULT_OOM;
2746 page_cache_release(old_page);
2750 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2751 unsigned long start_addr, unsigned long end_addr,
2752 struct zap_details *details)
2754 zap_page_range(vma, start_addr, end_addr - start_addr, details);
2757 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
2758 struct zap_details *details)
2760 struct vm_area_struct *vma;
2761 struct prio_tree_iter iter;
2762 pgoff_t vba, vea, zba, zea;
2764 vma_prio_tree_foreach(vma, &iter, root,
2765 details->first_index, details->last_index) {
2767 vba = vma->vm_pgoff;
2768 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2769 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2770 zba = details->first_index;
2773 zea = details->last_index;
2777 unmap_mapping_range_vma(vma,
2778 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2779 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2784 static inline void unmap_mapping_range_list(struct list_head *head,
2785 struct zap_details *details)
2787 struct vm_area_struct *vma;
2790 * In nonlinear VMAs there is no correspondence between virtual address
2791 * offset and file offset. So we must perform an exhaustive search
2792 * across *all* the pages in each nonlinear VMA, not just the pages
2793 * whose virtual address lies outside the file truncation point.
2795 list_for_each_entry(vma, head, shared.vm_set.list) {
2796 details->nonlinear_vma = vma;
2797 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2802 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2803 * @mapping: the address space containing mmaps to be unmapped.
2804 * @holebegin: byte in first page to unmap, relative to the start of
2805 * the underlying file. This will be rounded down to a PAGE_SIZE
2806 * boundary. Note that this is different from truncate_pagecache(), which
2807 * must keep the partial page. In contrast, we must get rid of
2809 * @holelen: size of prospective hole in bytes. This will be rounded
2810 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2812 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2813 * but 0 when invalidating pagecache, don't throw away private data.
2815 void unmap_mapping_range(struct address_space *mapping,
2816 loff_t const holebegin, loff_t const holelen, int even_cows)
2818 struct zap_details details;
2819 pgoff_t hba = holebegin >> PAGE_SHIFT;
2820 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2822 /* Check for overflow. */
2823 if (sizeof(holelen) > sizeof(hlen)) {
2825 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2826 if (holeend & ~(long long)ULONG_MAX)
2827 hlen = ULONG_MAX - hba + 1;
2830 details.check_mapping = even_cows? NULL: mapping;
2831 details.nonlinear_vma = NULL;
2832 details.first_index = hba;
2833 details.last_index = hba + hlen - 1;
2834 if (details.last_index < details.first_index)
2835 details.last_index = ULONG_MAX;
2838 mutex_lock(&mapping->i_mmap_mutex);
2839 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
2840 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2841 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2842 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2843 mutex_unlock(&mapping->i_mmap_mutex);
2845 EXPORT_SYMBOL(unmap_mapping_range);
2848 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2849 * but allow concurrent faults), and pte mapped but not yet locked.
2850 * We return with mmap_sem still held, but pte unmapped and unlocked.
2852 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2853 unsigned long address, pte_t *page_table, pmd_t *pmd,
2854 unsigned int flags, pte_t orig_pte)
2857 struct page *page, *swapcache = NULL;
2861 struct mem_cgroup *ptr;
2865 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2868 entry = pte_to_swp_entry(orig_pte);
2869 if (unlikely(non_swap_entry(entry))) {
2870 if (is_migration_entry(entry)) {
2871 migration_entry_wait(mm, pmd, address);
2872 } else if (is_hwpoison_entry(entry)) {
2873 ret = VM_FAULT_HWPOISON;
2875 print_bad_pte(vma, address, orig_pte, NULL);
2876 ret = VM_FAULT_SIGBUS;
2880 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2881 page = lookup_swap_cache(entry);
2883 grab_swap_token(mm); /* Contend for token _before_ read-in */
2884 page = swapin_readahead(entry,
2885 GFP_HIGHUSER_MOVABLE, vma, address);
2888 * Back out if somebody else faulted in this pte
2889 * while we released the pte lock.
2891 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2892 if (likely(pte_same(*page_table, orig_pte)))
2894 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2898 /* Had to read the page from swap area: Major fault */
2899 ret = VM_FAULT_MAJOR;
2900 count_vm_event(PGMAJFAULT);
2901 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2902 } else if (PageHWPoison(page)) {
2904 * hwpoisoned dirty swapcache pages are kept for killing
2905 * owner processes (which may be unknown at hwpoison time)
2907 ret = VM_FAULT_HWPOISON;
2908 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2912 locked = lock_page_or_retry(page, mm, flags);
2913 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2915 ret |= VM_FAULT_RETRY;
2920 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2921 * release the swapcache from under us. The page pin, and pte_same
2922 * test below, are not enough to exclude that. Even if it is still
2923 * swapcache, we need to check that the page's swap has not changed.
2925 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2928 if (ksm_might_need_to_copy(page, vma, address)) {
2930 page = ksm_does_need_to_copy(page, vma, address);
2932 if (unlikely(!page)) {
2940 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
2946 * Back out if somebody else already faulted in this pte.
2948 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2949 if (unlikely(!pte_same(*page_table, orig_pte)))
2952 if (unlikely(!PageUptodate(page))) {
2953 ret = VM_FAULT_SIGBUS;
2958 * The page isn't present yet, go ahead with the fault.
2960 * Be careful about the sequence of operations here.
2961 * To get its accounting right, reuse_swap_page() must be called
2962 * while the page is counted on swap but not yet in mapcount i.e.
2963 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2964 * must be called after the swap_free(), or it will never succeed.
2965 * Because delete_from_swap_page() may be called by reuse_swap_page(),
2966 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
2967 * in page->private. In this case, a record in swap_cgroup is silently
2968 * discarded at swap_free().
2971 inc_mm_counter_fast(mm, MM_ANONPAGES);
2972 dec_mm_counter_fast(mm, MM_SWAPENTS);
2973 pte = mk_pte(page, vma->vm_page_prot);
2974 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
2975 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2976 flags &= ~FAULT_FLAG_WRITE;
2977 ret |= VM_FAULT_WRITE;
2980 flush_icache_page(vma, page);
2981 set_pte_at(mm, address, page_table, pte);
2982 do_page_add_anon_rmap(page, vma, address, exclusive);
2983 /* It's better to call commit-charge after rmap is established */
2984 mem_cgroup_commit_charge_swapin(page, ptr);
2987 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2988 try_to_free_swap(page);
2992 * Hold the lock to avoid the swap entry to be reused
2993 * until we take the PT lock for the pte_same() check
2994 * (to avoid false positives from pte_same). For
2995 * further safety release the lock after the swap_free
2996 * so that the swap count won't change under a
2997 * parallel locked swapcache.
2999 unlock_page(swapcache);
3000 page_cache_release(swapcache);
3003 if (flags & FAULT_FLAG_WRITE) {
3004 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3005 if (ret & VM_FAULT_ERROR)
3006 ret &= VM_FAULT_ERROR;
3010 /* No need to invalidate - it was non-present before */
3011 update_mmu_cache(vma, address, page_table);
3013 pte_unmap_unlock(page_table, ptl);
3017 mem_cgroup_cancel_charge_swapin(ptr);
3018 pte_unmap_unlock(page_table, ptl);
3022 page_cache_release(page);
3024 unlock_page(swapcache);
3025 page_cache_release(swapcache);
3031 * This is like a special single-page "expand_{down|up}wards()",
3032 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3033 * doesn't hit another vma.
3035 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3037 address &= PAGE_MASK;
3038 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3039 struct vm_area_struct *prev = vma->vm_prev;
3042 * Is there a mapping abutting this one below?
3044 * That's only ok if it's the same stack mapping
3045 * that has gotten split..
3047 if (prev && prev->vm_end == address)
3048 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3050 expand_downwards(vma, address - PAGE_SIZE);
3052 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3053 struct vm_area_struct *next = vma->vm_next;
3055 /* As VM_GROWSDOWN but s/below/above/ */
3056 if (next && next->vm_start == address + PAGE_SIZE)
3057 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3059 expand_upwards(vma, address + PAGE_SIZE);
3065 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3066 * but allow concurrent faults), and pte mapped but not yet locked.
3067 * We return with mmap_sem still held, but pte unmapped and unlocked.
3069 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3070 unsigned long address, pte_t *page_table, pmd_t *pmd,
3077 pte_unmap(page_table);
3079 /* Check if we need to add a guard page to the stack */
3080 if (check_stack_guard_page(vma, address) < 0)
3081 return VM_FAULT_SIGBUS;
3083 /* Use the zero-page for reads */
3084 if (!(flags & FAULT_FLAG_WRITE)) {
3085 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3086 vma->vm_page_prot));
3087 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3088 if (!pte_none(*page_table))
3093 /* Allocate our own private page. */
3094 if (unlikely(anon_vma_prepare(vma)))
3096 page = alloc_zeroed_user_highpage_movable(vma, address);
3099 __SetPageUptodate(page);
3101 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3104 entry = mk_pte(page, vma->vm_page_prot);
3105 if (vma->vm_flags & VM_WRITE)
3106 entry = pte_mkwrite(pte_mkdirty(entry));
3108 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3109 if (!pte_none(*page_table))
3112 inc_mm_counter_fast(mm, MM_ANONPAGES);
3113 page_add_new_anon_rmap(page, vma, address);
3115 set_pte_at(mm, address, page_table, entry);
3117 /* No need to invalidate - it was non-present before */
3118 update_mmu_cache(vma, address, page_table);
3120 pte_unmap_unlock(page_table, ptl);
3123 mem_cgroup_uncharge_page(page);
3124 page_cache_release(page);
3127 page_cache_release(page);
3129 return VM_FAULT_OOM;
3133 * __do_fault() tries to create a new page mapping. It aggressively
3134 * tries to share with existing pages, but makes a separate copy if
3135 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3136 * the next page fault.
3138 * As this is called only for pages that do not currently exist, we
3139 * do not need to flush old virtual caches or the TLB.
3141 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3142 * but allow concurrent faults), and pte neither mapped nor locked.
3143 * We return with mmap_sem still held, but pte unmapped and unlocked.
3145 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3146 unsigned long address, pmd_t *pmd,
3147 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3152 struct page *cow_page;
3155 struct page *dirty_page = NULL;
3156 struct vm_fault vmf;
3158 int page_mkwrite = 0;
3161 * If we do COW later, allocate page befor taking lock_page()
3162 * on the file cache page. This will reduce lock holding time.
3164 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3166 if (unlikely(anon_vma_prepare(vma)))
3167 return VM_FAULT_OOM;
3169 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3171 return VM_FAULT_OOM;
3173 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3174 page_cache_release(cow_page);
3175 return VM_FAULT_OOM;
3180 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3185 ret = vma->vm_ops->fault(vma, &vmf);
3186 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3190 if (unlikely(PageHWPoison(vmf.page))) {
3191 if (ret & VM_FAULT_LOCKED)
3192 unlock_page(vmf.page);
3193 ret = VM_FAULT_HWPOISON;
3198 * For consistency in subsequent calls, make the faulted page always
3201 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3202 lock_page(vmf.page);
3204 VM_BUG_ON(!PageLocked(vmf.page));
3207 * Should we do an early C-O-W break?
3210 if (flags & FAULT_FLAG_WRITE) {
3211 if (!(vma->vm_flags & VM_SHARED)) {
3214 copy_user_highpage(page, vmf.page, address, vma);
3215 __SetPageUptodate(page);
3218 * If the page will be shareable, see if the backing
3219 * address space wants to know that the page is about
3220 * to become writable
3222 if (vma->vm_ops->page_mkwrite) {
3226 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3227 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3229 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3231 goto unwritable_page;
3233 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3235 if (!page->mapping) {
3236 ret = 0; /* retry the fault */
3238 goto unwritable_page;
3241 VM_BUG_ON(!PageLocked(page));
3248 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3251 * This silly early PAGE_DIRTY setting removes a race
3252 * due to the bad i386 page protection. But it's valid
3253 * for other architectures too.
3255 * Note that if FAULT_FLAG_WRITE is set, we either now have
3256 * an exclusive copy of the page, or this is a shared mapping,
3257 * so we can make it writable and dirty to avoid having to
3258 * handle that later.
3260 /* Only go through if we didn't race with anybody else... */
3261 if (likely(pte_same(*page_table, orig_pte))) {
3262 flush_icache_page(vma, page);
3263 entry = mk_pte(page, vma->vm_page_prot);
3264 if (flags & FAULT_FLAG_WRITE)
3265 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3267 inc_mm_counter_fast(mm, MM_ANONPAGES);
3268 page_add_new_anon_rmap(page, vma, address);
3270 inc_mm_counter_fast(mm, MM_FILEPAGES);
3271 page_add_file_rmap(page);
3272 if (flags & FAULT_FLAG_WRITE) {
3274 get_page(dirty_page);
3277 set_pte_at(mm, address, page_table, entry);
3279 /* no need to invalidate: a not-present page won't be cached */
3280 update_mmu_cache(vma, address, page_table);
3283 mem_cgroup_uncharge_page(cow_page);
3285 page_cache_release(page);
3287 anon = 1; /* no anon but release faulted_page */
3290 pte_unmap_unlock(page_table, ptl);
3293 struct address_space *mapping = page->mapping;
3295 if (set_page_dirty(dirty_page))
3297 unlock_page(dirty_page);
3298 put_page(dirty_page);
3299 if (page_mkwrite && mapping) {
3301 * Some device drivers do not set page.mapping but still
3304 balance_dirty_pages_ratelimited(mapping);
3307 /* file_update_time outside page_lock */
3309 file_update_time(vma->vm_file);
3311 unlock_page(vmf.page);
3313 page_cache_release(vmf.page);
3319 page_cache_release(page);
3322 /* fs's fault handler get error */
3324 mem_cgroup_uncharge_page(cow_page);
3325 page_cache_release(cow_page);
3330 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3331 unsigned long address, pte_t *page_table, pmd_t *pmd,
3332 unsigned int flags, pte_t orig_pte)
3334 pgoff_t pgoff = (((address & PAGE_MASK)
3335 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3337 pte_unmap(page_table);
3338 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3342 * Fault of a previously existing named mapping. Repopulate the pte
3343 * from the encoded file_pte if possible. This enables swappable
3346 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3347 * but allow concurrent faults), and pte mapped but not yet locked.
3348 * We return with mmap_sem still held, but pte unmapped and unlocked.
3350 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3351 unsigned long address, pte_t *page_table, pmd_t *pmd,
3352 unsigned int flags, pte_t orig_pte)
3356 flags |= FAULT_FLAG_NONLINEAR;
3358 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3361 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3363 * Page table corrupted: show pte and kill process.
3365 print_bad_pte(vma, address, orig_pte, NULL);
3366 return VM_FAULT_SIGBUS;
3369 pgoff = pte_to_pgoff(orig_pte);
3370 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3374 * These routines also need to handle stuff like marking pages dirty
3375 * and/or accessed for architectures that don't do it in hardware (most
3376 * RISC architectures). The early dirtying is also good on the i386.
3378 * There is also a hook called "update_mmu_cache()" that architectures
3379 * with external mmu caches can use to update those (ie the Sparc or
3380 * PowerPC hashed page tables that act as extended TLBs).
3382 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3383 * but allow concurrent faults), and pte mapped but not yet locked.
3384 * We return with mmap_sem still held, but pte unmapped and unlocked.
3386 int handle_pte_fault(struct mm_struct *mm,
3387 struct vm_area_struct *vma, unsigned long address,
3388 pte_t *pte, pmd_t *pmd, unsigned int flags)
3394 if (!pte_present(entry)) {
3395 if (pte_none(entry)) {
3397 if (likely(vma->vm_ops->fault))
3398 return do_linear_fault(mm, vma, address,
3399 pte, pmd, flags, entry);
3401 return do_anonymous_page(mm, vma, address,
3404 if (pte_file(entry))
3405 return do_nonlinear_fault(mm, vma, address,
3406 pte, pmd, flags, entry);
3407 return do_swap_page(mm, vma, address,
3408 pte, pmd, flags, entry);
3411 ptl = pte_lockptr(mm, pmd);
3413 if (unlikely(!pte_same(*pte, entry)))
3415 if (flags & FAULT_FLAG_WRITE) {
3416 if (!pte_write(entry))
3417 return do_wp_page(mm, vma, address,
3418 pte, pmd, ptl, entry);
3419 entry = pte_mkdirty(entry);
3421 entry = pte_mkyoung(entry);
3422 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3423 update_mmu_cache(vma, address, pte);
3426 * This is needed only for protection faults but the arch code
3427 * is not yet telling us if this is a protection fault or not.
3428 * This still avoids useless tlb flushes for .text page faults
3431 if (flags & FAULT_FLAG_WRITE)
3432 flush_tlb_fix_spurious_fault(vma, address);
3435 pte_unmap_unlock(pte, ptl);
3440 * By the time we get here, we already hold the mm semaphore
3442 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3443 unsigned long address, unsigned int flags)
3450 __set_current_state(TASK_RUNNING);
3452 count_vm_event(PGFAULT);
3453 mem_cgroup_count_vm_event(mm, PGFAULT);
3455 /* do counter updates before entering really critical section. */
3456 check_sync_rss_stat(current);
3458 if (unlikely(is_vm_hugetlb_page(vma)))
3459 return hugetlb_fault(mm, vma, address, flags);
3461 pgd = pgd_offset(mm, address);
3462 pud = pud_alloc(mm, pgd, address);
3464 return VM_FAULT_OOM;
3465 pmd = pmd_alloc(mm, pud, address);
3467 return VM_FAULT_OOM;
3468 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3470 return do_huge_pmd_anonymous_page(mm, vma, address,
3473 pmd_t orig_pmd = *pmd;
3475 if (pmd_trans_huge(orig_pmd)) {
3476 if (flags & FAULT_FLAG_WRITE &&
3477 !pmd_write(orig_pmd) &&
3478 !pmd_trans_splitting(orig_pmd))
3479 return do_huge_pmd_wp_page(mm, vma, address,
3486 * Use __pte_alloc instead of pte_alloc_map, because we can't
3487 * run pte_offset_map on the pmd, if an huge pmd could
3488 * materialize from under us from a different thread.
3490 if (unlikely(pmd_none(*pmd)) && __pte_alloc(mm, vma, pmd, address))
3491 return VM_FAULT_OOM;
3492 /* if an huge pmd materialized from under us just retry later */
3493 if (unlikely(pmd_trans_huge(*pmd)))
3496 * A regular pmd is established and it can't morph into a huge pmd
3497 * from under us anymore at this point because we hold the mmap_sem
3498 * read mode and khugepaged takes it in write mode. So now it's
3499 * safe to run pte_offset_map().
3501 pte = pte_offset_map(pmd, address);
3503 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3506 #ifndef __PAGETABLE_PUD_FOLDED
3508 * Allocate page upper directory.
3509 * We've already handled the fast-path in-line.
3511 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3513 pud_t *new = pud_alloc_one(mm, address);
3517 smp_wmb(); /* See comment in __pte_alloc */
3519 spin_lock(&mm->page_table_lock);
3520 if (pgd_present(*pgd)) /* Another has populated it */
3523 pgd_populate(mm, pgd, new);
3524 spin_unlock(&mm->page_table_lock);
3527 #endif /* __PAGETABLE_PUD_FOLDED */
3529 #ifndef __PAGETABLE_PMD_FOLDED
3531 * Allocate page middle directory.
3532 * We've already handled the fast-path in-line.
3534 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3536 pmd_t *new = pmd_alloc_one(mm, address);
3540 smp_wmb(); /* See comment in __pte_alloc */
3542 spin_lock(&mm->page_table_lock);
3543 #ifndef __ARCH_HAS_4LEVEL_HACK
3544 if (pud_present(*pud)) /* Another has populated it */
3547 pud_populate(mm, pud, new);
3549 if (pgd_present(*pud)) /* Another has populated it */
3552 pgd_populate(mm, pud, new);
3553 #endif /* __ARCH_HAS_4LEVEL_HACK */
3554 spin_unlock(&mm->page_table_lock);
3557 #endif /* __PAGETABLE_PMD_FOLDED */
3559 int make_pages_present(unsigned long addr, unsigned long end)
3561 int ret, len, write;
3562 struct vm_area_struct * vma;
3564 vma = find_vma(current->mm, addr);
3568 * We want to touch writable mappings with a write fault in order
3569 * to break COW, except for shared mappings because these don't COW
3570 * and we would not want to dirty them for nothing.
3572 write = (vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE;
3573 BUG_ON(addr >= end);
3574 BUG_ON(end > vma->vm_end);
3575 len = DIV_ROUND_UP(end, PAGE_SIZE) - addr/PAGE_SIZE;
3576 ret = get_user_pages(current, current->mm, addr,
3577 len, write, 0, NULL, NULL);
3580 return ret == len ? 0 : -EFAULT;
3583 #if !defined(__HAVE_ARCH_GATE_AREA)
3585 #if defined(AT_SYSINFO_EHDR)
3586 static struct vm_area_struct gate_vma;
3588 static int __init gate_vma_init(void)
3590 gate_vma.vm_mm = NULL;
3591 gate_vma.vm_start = FIXADDR_USER_START;
3592 gate_vma.vm_end = FIXADDR_USER_END;
3593 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3594 gate_vma.vm_page_prot = __P101;
3596 * Make sure the vDSO gets into every core dump.
3597 * Dumping its contents makes post-mortem fully interpretable later
3598 * without matching up the same kernel and hardware config to see
3599 * what PC values meant.
3601 gate_vma.vm_flags |= VM_ALWAYSDUMP;
3604 __initcall(gate_vma_init);
3607 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3609 #ifdef AT_SYSINFO_EHDR
3616 int in_gate_area_no_mm(unsigned long addr)
3618 #ifdef AT_SYSINFO_EHDR
3619 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3625 #endif /* __HAVE_ARCH_GATE_AREA */
3627 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3628 pte_t **ptepp, spinlock_t **ptlp)
3635 pgd = pgd_offset(mm, address);
3636 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3639 pud = pud_offset(pgd, address);
3640 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3643 pmd = pmd_offset(pud, address);
3644 VM_BUG_ON(pmd_trans_huge(*pmd));
3645 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3648 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3652 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3655 if (!pte_present(*ptep))
3660 pte_unmap_unlock(ptep, *ptlp);
3665 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3666 pte_t **ptepp, spinlock_t **ptlp)
3670 /* (void) is needed to make gcc happy */
3671 (void) __cond_lock(*ptlp,
3672 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3677 * follow_pfn - look up PFN at a user virtual address
3678 * @vma: memory mapping
3679 * @address: user virtual address
3680 * @pfn: location to store found PFN
3682 * Only IO mappings and raw PFN mappings are allowed.
3684 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3686 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3693 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3696 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3699 *pfn = pte_pfn(*ptep);
3700 pte_unmap_unlock(ptep, ptl);
3703 EXPORT_SYMBOL(follow_pfn);
3705 #ifdef CONFIG_HAVE_IOREMAP_PROT
3706 int follow_phys(struct vm_area_struct *vma,
3707 unsigned long address, unsigned int flags,
3708 unsigned long *prot, resource_size_t *phys)
3714 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3717 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3721 if ((flags & FOLL_WRITE) && !pte_write(pte))
3724 *prot = pgprot_val(pte_pgprot(pte));
3725 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3729 pte_unmap_unlock(ptep, ptl);
3734 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3735 void *buf, int len, int write)
3737 resource_size_t phys_addr;
3738 unsigned long prot = 0;
3739 void __iomem *maddr;
3740 int offset = addr & (PAGE_SIZE-1);
3742 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3745 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
3747 memcpy_toio(maddr + offset, buf, len);
3749 memcpy_fromio(buf, maddr + offset, len);
3757 * Access another process' address space as given in mm. If non-NULL, use the
3758 * given task for page fault accounting.
3760 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3761 unsigned long addr, void *buf, int len, int write)
3763 struct vm_area_struct *vma;
3764 void *old_buf = buf;
3766 down_read(&mm->mmap_sem);
3767 /* ignore errors, just check how much was successfully transferred */
3769 int bytes, ret, offset;
3771 struct page *page = NULL;
3773 ret = get_user_pages(tsk, mm, addr, 1,
3774 write, 1, &page, &vma);
3777 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3778 * we can access using slightly different code.
3780 #ifdef CONFIG_HAVE_IOREMAP_PROT
3781 vma = find_vma(mm, addr);
3782 if (!vma || vma->vm_start > addr)
3784 if (vma->vm_ops && vma->vm_ops->access)
3785 ret = vma->vm_ops->access(vma, addr, buf,
3793 offset = addr & (PAGE_SIZE-1);
3794 if (bytes > PAGE_SIZE-offset)
3795 bytes = PAGE_SIZE-offset;
3799 copy_to_user_page(vma, page, addr,
3800 maddr + offset, buf, bytes);
3801 set_page_dirty_lock(page);
3803 copy_from_user_page(vma, page, addr,
3804 buf, maddr + offset, bytes);
3807 page_cache_release(page);
3813 up_read(&mm->mmap_sem);
3815 return buf - old_buf;
3819 * access_remote_vm - access another process' address space
3820 * @mm: the mm_struct of the target address space
3821 * @addr: start address to access
3822 * @buf: source or destination buffer
3823 * @len: number of bytes to transfer
3824 * @write: whether the access is a write
3826 * The caller must hold a reference on @mm.
3828 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3829 void *buf, int len, int write)
3831 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3835 * Access another process' address space.
3836 * Source/target buffer must be kernel space,
3837 * Do not walk the page table directly, use get_user_pages
3839 int access_process_vm(struct task_struct *tsk, unsigned long addr,
3840 void *buf, int len, int write)
3842 struct mm_struct *mm;
3845 mm = get_task_mm(tsk);
3849 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3856 * Print the name of a VMA.
3858 void print_vma_addr(char *prefix, unsigned long ip)
3860 struct mm_struct *mm = current->mm;
3861 struct vm_area_struct *vma;
3864 * Do not print if we are in atomic
3865 * contexts (in exception stacks, etc.):
3867 if (preempt_count())
3870 down_read(&mm->mmap_sem);
3871 vma = find_vma(mm, ip);
3872 if (vma && vma->vm_file) {
3873 struct file *f = vma->vm_file;
3874 char *buf = (char *)__get_free_page(GFP_KERNEL);
3878 p = d_path(&f->f_path, buf, PAGE_SIZE);
3881 s = strrchr(p, '/');
3884 printk("%s%s[%lx+%lx]", prefix, p,
3886 vma->vm_end - vma->vm_start);
3887 free_page((unsigned long)buf);
3890 up_read(¤t->mm->mmap_sem);
3893 #ifdef CONFIG_PROVE_LOCKING
3894 void might_fault(void)
3897 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3898 * holding the mmap_sem, this is safe because kernel memory doesn't
3899 * get paged out, therefore we'll never actually fault, and the
3900 * below annotations will generate false positives.
3902 if (segment_eq(get_fs(), KERNEL_DS))
3907 * it would be nicer only to annotate paths which are not under
3908 * pagefault_disable, however that requires a larger audit and
3909 * providing helpers like get_user_atomic.
3911 if (!in_atomic() && current->mm)
3912 might_lock_read(¤t->mm->mmap_sem);
3914 EXPORT_SYMBOL(might_fault);
3917 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3918 static void clear_gigantic_page(struct page *page,
3920 unsigned int pages_per_huge_page)
3923 struct page *p = page;
3926 for (i = 0; i < pages_per_huge_page;
3927 i++, p = mem_map_next(p, page, i)) {
3929 clear_user_highpage(p, addr + i * PAGE_SIZE);
3932 void clear_huge_page(struct page *page,
3933 unsigned long addr, unsigned int pages_per_huge_page)
3937 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3938 clear_gigantic_page(page, addr, pages_per_huge_page);
3943 for (i = 0; i < pages_per_huge_page; i++) {
3945 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3949 static void copy_user_gigantic_page(struct page *dst, struct page *src,
3951 struct vm_area_struct *vma,
3952 unsigned int pages_per_huge_page)
3955 struct page *dst_base = dst;
3956 struct page *src_base = src;
3958 for (i = 0; i < pages_per_huge_page; ) {
3960 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3963 dst = mem_map_next(dst, dst_base, i);
3964 src = mem_map_next(src, src_base, i);
3968 void copy_user_huge_page(struct page *dst, struct page *src,
3969 unsigned long addr, struct vm_area_struct *vma,
3970 unsigned int pages_per_huge_page)
3974 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3975 copy_user_gigantic_page(dst, src, addr, vma,
3976 pages_per_huge_page);
3981 for (i = 0; i < pages_per_huge_page; i++) {
3983 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3986 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */