4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/writeback.h>
54 #include <linux/memcontrol.h>
55 #include <linux/mmu_notifier.h>
56 #include <linux/kallsyms.h>
57 #include <linux/swapops.h>
58 #include <linux/elf.h>
59 #include <linux/gfp.h>
60 #include <linux/migrate.h>
61 #include <linux/string.h>
64 #include <asm/pgalloc.h>
65 #include <asm/uaccess.h>
67 #include <asm/tlbflush.h>
68 #include <asm/pgtable.h>
72 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
73 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
76 #ifndef CONFIG_NEED_MULTIPLE_NODES
77 /* use the per-pgdat data instead for discontigmem - mbligh */
78 unsigned long max_mapnr;
81 EXPORT_SYMBOL(max_mapnr);
82 EXPORT_SYMBOL(mem_map);
86 * A number of key systems in x86 including ioremap() rely on the assumption
87 * that high_memory defines the upper bound on direct map memory, then end
88 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
89 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
94 EXPORT_SYMBOL(high_memory);
97 * Randomize the address space (stacks, mmaps, brk, etc.).
99 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
100 * as ancient (libc5 based) binaries can segfault. )
102 int randomize_va_space __read_mostly =
103 #ifdef CONFIG_COMPAT_BRK
109 static int __init disable_randmaps(char *s)
111 randomize_va_space = 0;
114 __setup("norandmaps", disable_randmaps);
116 unsigned long zero_pfn __read_mostly;
117 unsigned long highest_memmap_pfn __read_mostly;
120 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
122 static int __init init_zero_pfn(void)
124 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 core_initcall(init_zero_pfn);
130 #if defined(SPLIT_RSS_COUNTING)
132 void sync_mm_rss(struct mm_struct *mm)
136 for (i = 0; i < NR_MM_COUNTERS; i++) {
137 if (current->rss_stat.count[i]) {
138 add_mm_counter(mm, i, current->rss_stat.count[i]);
139 current->rss_stat.count[i] = 0;
142 current->rss_stat.events = 0;
145 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
147 struct task_struct *task = current;
149 if (likely(task->mm == mm))
150 task->rss_stat.count[member] += val;
152 add_mm_counter(mm, member, val);
154 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
155 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
157 /* sync counter once per 64 page faults */
158 #define TASK_RSS_EVENTS_THRESH (64)
159 static void check_sync_rss_stat(struct task_struct *task)
161 if (unlikely(task != current))
163 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
164 sync_mm_rss(task->mm);
166 #else /* SPLIT_RSS_COUNTING */
168 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
169 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
171 static void check_sync_rss_stat(struct task_struct *task)
175 #endif /* SPLIT_RSS_COUNTING */
177 #ifdef HAVE_GENERIC_MMU_GATHER
179 static int tlb_next_batch(struct mmu_gather *tlb)
181 struct mmu_gather_batch *batch;
185 tlb->active = batch->next;
189 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
199 batch->max = MAX_GATHER_BATCH;
201 tlb->active->next = batch;
208 * Called to initialize an (on-stack) mmu_gather structure for page-table
209 * tear-down from @mm. The @fullmm argument is used when @mm is without
210 * users and we're going to destroy the full address space (exit/execve).
212 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, bool fullmm)
216 tlb->fullmm = fullmm;
217 tlb->need_flush_all = 0;
221 tlb->local.next = NULL;
223 tlb->local.max = ARRAY_SIZE(tlb->__pages);
224 tlb->active = &tlb->local;
225 tlb->batch_count = 0;
227 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
232 void tlb_flush_mmu(struct mmu_gather *tlb)
234 struct mmu_gather_batch *batch;
236 if (!tlb->need_flush)
240 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
241 tlb_table_flush(tlb);
244 for (batch = &tlb->local; batch; batch = batch->next) {
245 free_pages_and_swap_cache(batch->pages, batch->nr);
248 tlb->active = &tlb->local;
252 * Called at the end of the shootdown operation to free up any resources
253 * that were required.
255 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
257 struct mmu_gather_batch *batch, *next;
263 /* keep the page table cache within bounds */
266 for (batch = tlb->local.next; batch; batch = next) {
268 free_pages((unsigned long)batch, 0);
270 tlb->local.next = NULL;
274 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
275 * handling the additional races in SMP caused by other CPUs caching valid
276 * mappings in their TLBs. Returns the number of free page slots left.
277 * When out of page slots we must call tlb_flush_mmu().
279 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
281 struct mmu_gather_batch *batch;
283 VM_BUG_ON(!tlb->need_flush);
286 batch->pages[batch->nr++] = page;
287 if (batch->nr == batch->max) {
288 if (!tlb_next_batch(tlb))
292 VM_BUG_ON(batch->nr > batch->max);
294 return batch->max - batch->nr;
297 #endif /* HAVE_GENERIC_MMU_GATHER */
299 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
302 * See the comment near struct mmu_table_batch.
305 static void tlb_remove_table_smp_sync(void *arg)
307 /* Simply deliver the interrupt */
310 static void tlb_remove_table_one(void *table)
313 * This isn't an RCU grace period and hence the page-tables cannot be
314 * assumed to be actually RCU-freed.
316 * It is however sufficient for software page-table walkers that rely on
317 * IRQ disabling. See the comment near struct mmu_table_batch.
319 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
320 __tlb_remove_table(table);
323 static void tlb_remove_table_rcu(struct rcu_head *head)
325 struct mmu_table_batch *batch;
328 batch = container_of(head, struct mmu_table_batch, rcu);
330 for (i = 0; i < batch->nr; i++)
331 __tlb_remove_table(batch->tables[i]);
333 free_page((unsigned long)batch);
336 void tlb_table_flush(struct mmu_gather *tlb)
338 struct mmu_table_batch **batch = &tlb->batch;
341 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
346 void tlb_remove_table(struct mmu_gather *tlb, void *table)
348 struct mmu_table_batch **batch = &tlb->batch;
353 * When there's less then two users of this mm there cannot be a
354 * concurrent page-table walk.
356 if (atomic_read(&tlb->mm->mm_users) < 2) {
357 __tlb_remove_table(table);
361 if (*batch == NULL) {
362 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
363 if (*batch == NULL) {
364 tlb_remove_table_one(table);
369 (*batch)->tables[(*batch)->nr++] = table;
370 if ((*batch)->nr == MAX_TABLE_BATCH)
371 tlb_table_flush(tlb);
374 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
377 * If a p?d_bad entry is found while walking page tables, report
378 * the error, before resetting entry to p?d_none. Usually (but
379 * very seldom) called out from the p?d_none_or_clear_bad macros.
382 void pgd_clear_bad(pgd_t *pgd)
388 void pud_clear_bad(pud_t *pud)
394 void pmd_clear_bad(pmd_t *pmd)
401 * Note: this doesn't free the actual pages themselves. That
402 * has been handled earlier when unmapping all the memory regions.
404 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
407 pgtable_t token = pmd_pgtable(*pmd);
409 pte_free_tlb(tlb, token, addr);
413 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414 unsigned long addr, unsigned long end,
415 unsigned long floor, unsigned long ceiling)
422 pmd = pmd_offset(pud, addr);
424 next = pmd_addr_end(addr, end);
425 if (pmd_none_or_clear_bad(pmd))
427 free_pte_range(tlb, pmd, addr);
428 } while (pmd++, addr = next, addr != end);
438 if (end - 1 > ceiling - 1)
441 pmd = pmd_offset(pud, start);
443 pmd_free_tlb(tlb, pmd, start);
446 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
447 unsigned long addr, unsigned long end,
448 unsigned long floor, unsigned long ceiling)
455 pud = pud_offset(pgd, addr);
457 next = pud_addr_end(addr, end);
458 if (pud_none_or_clear_bad(pud))
460 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
461 } while (pud++, addr = next, addr != end);
467 ceiling &= PGDIR_MASK;
471 if (end - 1 > ceiling - 1)
474 pud = pud_offset(pgd, start);
476 pud_free_tlb(tlb, pud, start);
480 * This function frees user-level page tables of a process.
482 * Must be called with pagetable lock held.
484 void free_pgd_range(struct mmu_gather *tlb,
485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
492 * The next few lines have given us lots of grief...
494 * Why are we testing PMD* at this top level? Because often
495 * there will be no work to do at all, and we'd prefer not to
496 * go all the way down to the bottom just to discover that.
498 * Why all these "- 1"s? Because 0 represents both the bottom
499 * of the address space and the top of it (using -1 for the
500 * top wouldn't help much: the masks would do the wrong thing).
501 * The rule is that addr 0 and floor 0 refer to the bottom of
502 * the address space, but end 0 and ceiling 0 refer to the top
503 * Comparisons need to use "end - 1" and "ceiling - 1" (though
504 * that end 0 case should be mythical).
506 * Wherever addr is brought up or ceiling brought down, we must
507 * be careful to reject "the opposite 0" before it confuses the
508 * subsequent tests. But what about where end is brought down
509 * by PMD_SIZE below? no, end can't go down to 0 there.
511 * Whereas we round start (addr) and ceiling down, by different
512 * masks at different levels, in order to test whether a table
513 * now has no other vmas using it, so can be freed, we don't
514 * bother to round floor or end up - the tests don't need that.
528 if (end - 1 > ceiling - 1)
533 pgd = pgd_offset(tlb->mm, addr);
535 next = pgd_addr_end(addr, end);
536 if (pgd_none_or_clear_bad(pgd))
538 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
539 } while (pgd++, addr = next, addr != end);
542 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
543 unsigned long floor, unsigned long ceiling)
546 struct vm_area_struct *next = vma->vm_next;
547 unsigned long addr = vma->vm_start;
550 * Hide vma from rmap and truncate_pagecache before freeing
553 unlink_anon_vmas(vma);
554 unlink_file_vma(vma);
556 if (is_vm_hugetlb_page(vma)) {
557 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
558 floor, next? next->vm_start: ceiling);
561 * Optimization: gather nearby vmas into one call down
563 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
564 && !is_vm_hugetlb_page(next)) {
567 unlink_anon_vmas(vma);
568 unlink_file_vma(vma);
570 free_pgd_range(tlb, addr, vma->vm_end,
571 floor, next? next->vm_start: ceiling);
577 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
578 pmd_t *pmd, unsigned long address)
580 pgtable_t new = pte_alloc_one(mm, address);
581 int wait_split_huge_page;
586 * Ensure all pte setup (eg. pte page lock and page clearing) are
587 * visible before the pte is made visible to other CPUs by being
588 * put into page tables.
590 * The other side of the story is the pointer chasing in the page
591 * table walking code (when walking the page table without locking;
592 * ie. most of the time). Fortunately, these data accesses consist
593 * of a chain of data-dependent loads, meaning most CPUs (alpha
594 * being the notable exception) will already guarantee loads are
595 * seen in-order. See the alpha page table accessors for the
596 * smp_read_barrier_depends() barriers in page table walking code.
598 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
600 spin_lock(&mm->page_table_lock);
601 wait_split_huge_page = 0;
602 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
604 pmd_populate(mm, pmd, new);
606 } else if (unlikely(pmd_trans_splitting(*pmd)))
607 wait_split_huge_page = 1;
608 spin_unlock(&mm->page_table_lock);
611 if (wait_split_huge_page)
612 wait_split_huge_page(vma->anon_vma, pmd);
616 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
618 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
622 smp_wmb(); /* See comment in __pte_alloc */
624 spin_lock(&init_mm.page_table_lock);
625 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
626 pmd_populate_kernel(&init_mm, pmd, new);
629 VM_BUG_ON(pmd_trans_splitting(*pmd));
630 spin_unlock(&init_mm.page_table_lock);
632 pte_free_kernel(&init_mm, new);
636 static inline void init_rss_vec(int *rss)
638 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
641 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
645 if (current->mm == mm)
647 for (i = 0; i < NR_MM_COUNTERS; i++)
649 add_mm_counter(mm, i, rss[i]);
653 * This function is called to print an error when a bad pte
654 * is found. For example, we might have a PFN-mapped pte in
655 * a region that doesn't allow it.
657 * The calling function must still handle the error.
659 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
660 pte_t pte, struct page *page)
662 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
663 pud_t *pud = pud_offset(pgd, addr);
664 pmd_t *pmd = pmd_offset(pud, addr);
665 struct address_space *mapping;
667 static unsigned long resume;
668 static unsigned long nr_shown;
669 static unsigned long nr_unshown;
672 * Allow a burst of 60 reports, then keep quiet for that minute;
673 * or allow a steady drip of one report per second.
675 if (nr_shown == 60) {
676 if (time_before(jiffies, resume)) {
682 "BUG: Bad page map: %lu messages suppressed\n",
689 resume = jiffies + 60 * HZ;
691 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
692 index = linear_page_index(vma, addr);
695 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
697 (long long)pte_val(pte), (long long)pmd_val(*pmd));
701 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
702 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
704 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
707 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
709 if (vma->vm_file && vma->vm_file->f_op)
710 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
711 vma->vm_file->f_op->mmap);
713 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
716 static inline bool is_cow_mapping(vm_flags_t flags)
718 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
722 * vm_normal_page -- This function gets the "struct page" associated with a pte.
724 * "Special" mappings do not wish to be associated with a "struct page" (either
725 * it doesn't exist, or it exists but they don't want to touch it). In this
726 * case, NULL is returned here. "Normal" mappings do have a struct page.
728 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
729 * pte bit, in which case this function is trivial. Secondly, an architecture
730 * may not have a spare pte bit, which requires a more complicated scheme,
733 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
734 * special mapping (even if there are underlying and valid "struct pages").
735 * COWed pages of a VM_PFNMAP are always normal.
737 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
738 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
739 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
740 * mapping will always honor the rule
742 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
744 * And for normal mappings this is false.
746 * This restricts such mappings to be a linear translation from virtual address
747 * to pfn. To get around this restriction, we allow arbitrary mappings so long
748 * as the vma is not a COW mapping; in that case, we know that all ptes are
749 * special (because none can have been COWed).
752 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
754 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
755 * page" backing, however the difference is that _all_ pages with a struct
756 * page (that is, those where pfn_valid is true) are refcounted and considered
757 * normal pages by the VM. The disadvantage is that pages are refcounted
758 * (which can be slower and simply not an option for some PFNMAP users). The
759 * advantage is that we don't have to follow the strict linearity rule of
760 * PFNMAP mappings in order to support COWable mappings.
763 #ifdef __HAVE_ARCH_PTE_SPECIAL
764 # define HAVE_PTE_SPECIAL 1
766 # define HAVE_PTE_SPECIAL 0
768 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
771 unsigned long pfn = pte_pfn(pte);
773 if (HAVE_PTE_SPECIAL) {
774 if (likely(!pte_special(pte)))
776 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
778 if (!is_zero_pfn(pfn))
779 print_bad_pte(vma, addr, pte, NULL);
783 /* !HAVE_PTE_SPECIAL case follows: */
785 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
786 if (vma->vm_flags & VM_MIXEDMAP) {
792 off = (addr - vma->vm_start) >> PAGE_SHIFT;
793 if (pfn == vma->vm_pgoff + off)
795 if (!is_cow_mapping(vma->vm_flags))
800 if (is_zero_pfn(pfn))
803 if (unlikely(pfn > highest_memmap_pfn)) {
804 print_bad_pte(vma, addr, pte, NULL);
809 * NOTE! We still have PageReserved() pages in the page tables.
810 * eg. VDSO mappings can cause them to exist.
813 return pfn_to_page(pfn);
817 * copy one vm_area from one task to the other. Assumes the page tables
818 * already present in the new task to be cleared in the whole range
819 * covered by this vma.
822 static inline unsigned long
823 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
824 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
825 unsigned long addr, int *rss)
827 unsigned long vm_flags = vma->vm_flags;
828 pte_t pte = *src_pte;
831 /* pte contains position in swap or file, so copy. */
832 if (unlikely(!pte_present(pte))) {
833 if (!pte_file(pte)) {
834 swp_entry_t entry = pte_to_swp_entry(pte);
836 if (swap_duplicate(entry) < 0)
839 /* make sure dst_mm is on swapoff's mmlist. */
840 if (unlikely(list_empty(&dst_mm->mmlist))) {
841 spin_lock(&mmlist_lock);
842 if (list_empty(&dst_mm->mmlist))
843 list_add(&dst_mm->mmlist,
845 spin_unlock(&mmlist_lock);
847 if (likely(!non_swap_entry(entry)))
849 else if (is_migration_entry(entry)) {
850 page = migration_entry_to_page(entry);
857 if (is_write_migration_entry(entry) &&
858 is_cow_mapping(vm_flags)) {
860 * COW mappings require pages in both
861 * parent and child to be set to read.
863 make_migration_entry_read(&entry);
864 pte = swp_entry_to_pte(entry);
865 set_pte_at(src_mm, addr, src_pte, pte);
873 * If it's a COW mapping, write protect it both
874 * in the parent and the child
876 if (is_cow_mapping(vm_flags)) {
877 ptep_set_wrprotect(src_mm, addr, src_pte);
878 pte = pte_wrprotect(pte);
882 * If it's a shared mapping, mark it clean in
885 if (vm_flags & VM_SHARED)
886 pte = pte_mkclean(pte);
887 pte = pte_mkold(pte);
889 page = vm_normal_page(vma, addr, pte);
900 set_pte_at(dst_mm, addr, dst_pte, pte);
904 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
905 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
906 unsigned long addr, unsigned long end)
908 pte_t *orig_src_pte, *orig_dst_pte;
909 pte_t *src_pte, *dst_pte;
910 spinlock_t *src_ptl, *dst_ptl;
912 int rss[NR_MM_COUNTERS];
913 swp_entry_t entry = (swp_entry_t){0};
918 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
921 src_pte = pte_offset_map(src_pmd, addr);
922 src_ptl = pte_lockptr(src_mm, src_pmd);
923 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
924 orig_src_pte = src_pte;
925 orig_dst_pte = dst_pte;
926 arch_enter_lazy_mmu_mode();
930 * We are holding two locks at this point - either of them
931 * could generate latencies in another task on another CPU.
933 if (progress >= 32) {
935 if (need_resched() ||
936 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
939 if (pte_none(*src_pte)) {
943 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
948 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
950 arch_leave_lazy_mmu_mode();
951 spin_unlock(src_ptl);
952 pte_unmap(orig_src_pte);
953 add_mm_rss_vec(dst_mm, rss);
954 pte_unmap_unlock(orig_dst_pte, dst_ptl);
958 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
967 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
969 unsigned long addr, unsigned long end)
971 pmd_t *src_pmd, *dst_pmd;
974 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
977 src_pmd = pmd_offset(src_pud, addr);
979 next = pmd_addr_end(addr, end);
980 if (pmd_trans_huge(*src_pmd)) {
982 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
983 err = copy_huge_pmd(dst_mm, src_mm,
984 dst_pmd, src_pmd, addr, vma);
991 if (pmd_none_or_clear_bad(src_pmd))
993 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
996 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1000 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1001 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1002 unsigned long addr, unsigned long end)
1004 pud_t *src_pud, *dst_pud;
1007 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1010 src_pud = pud_offset(src_pgd, addr);
1012 next = pud_addr_end(addr, end);
1013 if (pud_none_or_clear_bad(src_pud))
1015 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1018 } while (dst_pud++, src_pud++, addr = next, addr != end);
1022 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 struct vm_area_struct *vma)
1025 pgd_t *src_pgd, *dst_pgd;
1027 unsigned long addr = vma->vm_start;
1028 unsigned long end = vma->vm_end;
1029 unsigned long mmun_start; /* For mmu_notifiers */
1030 unsigned long mmun_end; /* For mmu_notifiers */
1035 * Don't copy ptes where a page fault will fill them correctly.
1036 * Fork becomes much lighter when there are big shared or private
1037 * readonly mappings. The tradeoff is that copy_page_range is more
1038 * efficient than faulting.
1040 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1041 VM_PFNMAP | VM_MIXEDMAP))) {
1046 if (is_vm_hugetlb_page(vma))
1047 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1049 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1051 * We do not free on error cases below as remove_vma
1052 * gets called on error from higher level routine
1054 ret = track_pfn_copy(vma);
1060 * We need to invalidate the secondary MMU mappings only when
1061 * there could be a permission downgrade on the ptes of the
1062 * parent mm. And a permission downgrade will only happen if
1063 * is_cow_mapping() returns true.
1065 is_cow = is_cow_mapping(vma->vm_flags);
1069 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1073 dst_pgd = pgd_offset(dst_mm, addr);
1074 src_pgd = pgd_offset(src_mm, addr);
1076 next = pgd_addr_end(addr, end);
1077 if (pgd_none_or_clear_bad(src_pgd))
1079 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1080 vma, addr, next))) {
1084 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1087 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1091 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1092 struct vm_area_struct *vma, pmd_t *pmd,
1093 unsigned long addr, unsigned long end,
1094 struct zap_details *details)
1096 struct mm_struct *mm = tlb->mm;
1097 int force_flush = 0;
1098 int rss[NR_MM_COUNTERS];
1102 unsigned long range_start = addr;
1106 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1108 arch_enter_lazy_mmu_mode();
1111 if (pte_none(ptent)) {
1115 if (pte_present(ptent)) {
1118 page = vm_normal_page(vma, addr, ptent);
1119 if (unlikely(details) && page) {
1121 * unmap_shared_mapping_pages() wants to
1122 * invalidate cache without truncating:
1123 * unmap shared but keep private pages.
1125 if (details->check_mapping &&
1126 details->check_mapping != page->mapping)
1129 * Each page->index must be checked when
1130 * invalidating or truncating nonlinear.
1132 if (details->nonlinear_vma &&
1133 (page->index < details->first_index ||
1134 page->index > details->last_index))
1137 ptent = ptep_get_and_clear_full(mm, addr, pte,
1139 tlb_remove_tlb_entry(tlb, pte, addr);
1140 if (unlikely(!page))
1142 if (unlikely(details) && details->nonlinear_vma
1143 && linear_page_index(details->nonlinear_vma,
1144 addr) != page->index) {
1145 pte_t ptfile = pgoff_to_pte(page->index);
1146 if (pte_soft_dirty(ptent))
1147 pte_file_mksoft_dirty(ptfile);
1148 set_pte_at(mm, addr, pte, ptfile);
1151 rss[MM_ANONPAGES]--;
1153 if (pte_dirty(ptent))
1154 set_page_dirty(page);
1155 if (pte_young(ptent) &&
1156 likely(!(vma->vm_flags & VM_SEQ_READ)))
1157 mark_page_accessed(page);
1158 rss[MM_FILEPAGES]--;
1160 page_remove_rmap(page);
1161 if (unlikely(page_mapcount(page) < 0))
1162 print_bad_pte(vma, addr, ptent, page);
1163 force_flush = !__tlb_remove_page(tlb, page);
1169 * If details->check_mapping, we leave swap entries;
1170 * if details->nonlinear_vma, we leave file entries.
1172 if (unlikely(details))
1174 if (pte_file(ptent)) {
1175 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1176 print_bad_pte(vma, addr, ptent, NULL);
1178 swp_entry_t entry = pte_to_swp_entry(ptent);
1180 if (!non_swap_entry(entry))
1182 else if (is_migration_entry(entry)) {
1185 page = migration_entry_to_page(entry);
1188 rss[MM_ANONPAGES]--;
1190 rss[MM_FILEPAGES]--;
1192 if (unlikely(!free_swap_and_cache(entry)))
1193 print_bad_pte(vma, addr, ptent, NULL);
1195 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1196 } while (pte++, addr += PAGE_SIZE, addr != end);
1198 add_mm_rss_vec(mm, rss);
1199 arch_leave_lazy_mmu_mode();
1200 pte_unmap_unlock(start_pte, ptl);
1203 * mmu_gather ran out of room to batch pages, we break out of
1204 * the PTE lock to avoid doing the potential expensive TLB invalidate
1205 * and page-free while holding it.
1210 #ifdef HAVE_GENERIC_MMU_GATHER
1211 tlb->start = range_start;
1224 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1232 pmd = pmd_offset(pud, addr);
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237 #ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1263 } while (pmd++, addr = next, addr != end);
1268 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1276 pud = pud_offset(pgd, addr);
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1287 static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1313 static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1318 unsigned long start = max(vma->vm_start, start_addr);
1321 if (start >= vma->vm_end)
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1328 uprobe_munmap(vma, start, end);
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of do_mmap_pgoff. When
1340 * hugetlbfs ->mmap method fails,
1341 * do_mmap_pgoff() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1352 unmap_page_range(tlb, vma, start, end, details);
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1363 * Unmap all pages in the vma list.
1365 * Only addresses between `start' and `end' will be unmapped.
1367 * The VMA list must be sorted in ascending virtual address order.
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1374 void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1378 struct mm_struct *mm = vma->vm_mm;
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1393 * Caller must protect the VMA list
1395 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1403 tlb_gather_mmu(&tlb, mm, 0);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1419 * The range must fit into one VMA.
1421 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1429 tlb_gather_mmu(&tlb, mm, 0);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1445 * The entire address range must be fully contained within the vma.
1447 * Returns 0 if successful.
1449 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1455 zap_page_range_single(vma, address, size, NULL);
1458 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1473 struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1483 struct mm_struct *mm = vma->vm_mm;
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1498 pud = pud_offset(pgd, address);
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 BUG_ON(flags & FOLL_GET);
1503 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1506 if (unlikely(pud_bad(*pud)))
1509 pmd = pmd_offset(pud, address);
1512 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1513 BUG_ON(flags & FOLL_GET);
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1517 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1519 if (pmd_trans_huge(*pmd)) {
1520 if (flags & FOLL_SPLIT) {
1521 split_huge_page_pmd(vma, address, pmd);
1522 goto split_fallthrough;
1524 spin_lock(&mm->page_table_lock);
1525 if (likely(pmd_trans_huge(*pmd))) {
1526 if (unlikely(pmd_trans_splitting(*pmd))) {
1527 spin_unlock(&mm->page_table_lock);
1528 wait_split_huge_page(vma->anon_vma, pmd);
1530 page = follow_trans_huge_pmd(vma, address,
1532 spin_unlock(&mm->page_table_lock);
1533 *page_mask = HPAGE_PMD_NR - 1;
1537 spin_unlock(&mm->page_table_lock);
1541 if (unlikely(pmd_bad(*pmd)))
1544 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1547 if (!pte_present(pte)) {
1550 * KSM's break_ksm() relies upon recognizing a ksm page
1551 * even while it is being migrated, so for that case we
1552 * need migration_entry_wait().
1554 if (likely(!(flags & FOLL_MIGRATION)))
1556 if (pte_none(pte) || pte_file(pte))
1558 entry = pte_to_swp_entry(pte);
1559 if (!is_migration_entry(entry))
1561 pte_unmap_unlock(ptep, ptl);
1562 migration_entry_wait(mm, pmd, address);
1563 goto split_fallthrough;
1565 if ((flags & FOLL_NUMA) && pte_numa(pte))
1567 if ((flags & FOLL_WRITE) && !pte_write(pte))
1570 page = vm_normal_page(vma, address, pte);
1571 if (unlikely(!page)) {
1572 if ((flags & FOLL_DUMP) ||
1573 !is_zero_pfn(pte_pfn(pte)))
1575 page = pte_page(pte);
1578 if (flags & FOLL_GET)
1579 get_page_foll(page);
1580 if (flags & FOLL_TOUCH) {
1581 if ((flags & FOLL_WRITE) &&
1582 !pte_dirty(pte) && !PageDirty(page))
1583 set_page_dirty(page);
1585 * pte_mkyoung() would be more correct here, but atomic care
1586 * is needed to avoid losing the dirty bit: it is easier to use
1587 * mark_page_accessed().
1589 mark_page_accessed(page);
1591 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1593 * The preliminary mapping check is mainly to avoid the
1594 * pointless overhead of lock_page on the ZERO_PAGE
1595 * which might bounce very badly if there is contention.
1597 * If the page is already locked, we don't need to
1598 * handle it now - vmscan will handle it later if and
1599 * when it attempts to reclaim the page.
1601 if (page->mapping && trylock_page(page)) {
1602 lru_add_drain(); /* push cached pages to LRU */
1604 * Because we lock page here, and migration is
1605 * blocked by the pte's page reference, and we
1606 * know the page is still mapped, we don't even
1607 * need to check for file-cache page truncation.
1609 mlock_vma_page(page);
1614 pte_unmap_unlock(ptep, ptl);
1619 pte_unmap_unlock(ptep, ptl);
1620 return ERR_PTR(-EFAULT);
1623 pte_unmap_unlock(ptep, ptl);
1629 * When core dumping an enormous anonymous area that nobody
1630 * has touched so far, we don't want to allocate unnecessary pages or
1631 * page tables. Return error instead of NULL to skip handle_mm_fault,
1632 * then get_dump_page() will return NULL to leave a hole in the dump.
1633 * But we can only make this optimization where a hole would surely
1634 * be zero-filled if handle_mm_fault() actually did handle it.
1636 if ((flags & FOLL_DUMP) &&
1637 (!vma->vm_ops || !vma->vm_ops->fault))
1638 return ERR_PTR(-EFAULT);
1642 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1644 return stack_guard_page_start(vma, addr) ||
1645 stack_guard_page_end(vma, addr+PAGE_SIZE);
1649 * __get_user_pages() - pin user pages in memory
1650 * @tsk: task_struct of target task
1651 * @mm: mm_struct of target mm
1652 * @start: starting user address
1653 * @nr_pages: number of pages from start to pin
1654 * @gup_flags: flags modifying pin behaviour
1655 * @pages: array that receives pointers to the pages pinned.
1656 * Should be at least nr_pages long. Or NULL, if caller
1657 * only intends to ensure the pages are faulted in.
1658 * @vmas: array of pointers to vmas corresponding to each page.
1659 * Or NULL if the caller does not require them.
1660 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1662 * Returns number of pages pinned. This may be fewer than the number
1663 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1664 * were pinned, returns -errno. Each page returned must be released
1665 * with a put_page() call when it is finished with. vmas will only
1666 * remain valid while mmap_sem is held.
1668 * Must be called with mmap_sem held for read or write.
1670 * __get_user_pages walks a process's page tables and takes a reference to
1671 * each struct page that each user address corresponds to at a given
1672 * instant. That is, it takes the page that would be accessed if a user
1673 * thread accesses the given user virtual address at that instant.
1675 * This does not guarantee that the page exists in the user mappings when
1676 * __get_user_pages returns, and there may even be a completely different
1677 * page there in some cases (eg. if mmapped pagecache has been invalidated
1678 * and subsequently re faulted). However it does guarantee that the page
1679 * won't be freed completely. And mostly callers simply care that the page
1680 * contains data that was valid *at some point in time*. Typically, an IO
1681 * or similar operation cannot guarantee anything stronger anyway because
1682 * locks can't be held over the syscall boundary.
1684 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1685 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1686 * appropriate) must be called after the page is finished with, and
1687 * before put_page is called.
1689 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1690 * or mmap_sem contention, and if waiting is needed to pin all pages,
1691 * *@nonblocking will be set to 0.
1693 * In most cases, get_user_pages or get_user_pages_fast should be used
1694 * instead of __get_user_pages. __get_user_pages should be used only if
1695 * you need some special @gup_flags.
1697 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1698 unsigned long start, unsigned long nr_pages,
1699 unsigned int gup_flags, struct page **pages,
1700 struct vm_area_struct **vmas, int *nonblocking)
1703 unsigned long vm_flags;
1704 unsigned int page_mask;
1709 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1712 * Require read or write permissions.
1713 * If FOLL_FORCE is set, we only require the "MAY" flags.
1715 vm_flags = (gup_flags & FOLL_WRITE) ?
1716 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1717 vm_flags &= (gup_flags & FOLL_FORCE) ?
1718 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1721 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1722 * would be called on PROT_NONE ranges. We must never invoke
1723 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1724 * page faults would unprotect the PROT_NONE ranges if
1725 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1726 * bitflag. So to avoid that, don't set FOLL_NUMA if
1727 * FOLL_FORCE is set.
1729 if (!(gup_flags & FOLL_FORCE))
1730 gup_flags |= FOLL_NUMA;
1735 struct vm_area_struct *vma;
1737 vma = find_extend_vma(mm, start);
1738 if (!vma && in_gate_area(mm, start)) {
1739 unsigned long pg = start & PAGE_MASK;
1745 /* user gate pages are read-only */
1746 if (gup_flags & FOLL_WRITE)
1747 return i ? : -EFAULT;
1749 pgd = pgd_offset_k(pg);
1751 pgd = pgd_offset_gate(mm, pg);
1752 BUG_ON(pgd_none(*pgd));
1753 pud = pud_offset(pgd, pg);
1754 BUG_ON(pud_none(*pud));
1755 pmd = pmd_offset(pud, pg);
1757 return i ? : -EFAULT;
1758 VM_BUG_ON(pmd_trans_huge(*pmd));
1759 pte = pte_offset_map(pmd, pg);
1760 if (pte_none(*pte)) {
1762 return i ? : -EFAULT;
1764 vma = get_gate_vma(mm);
1768 page = vm_normal_page(vma, start, *pte);
1770 if (!(gup_flags & FOLL_DUMP) &&
1771 is_zero_pfn(pte_pfn(*pte)))
1772 page = pte_page(*pte);
1775 return i ? : -EFAULT;
1787 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1788 !(vm_flags & vma->vm_flags))
1789 return i ? : -EFAULT;
1791 if (is_vm_hugetlb_page(vma)) {
1792 i = follow_hugetlb_page(mm, vma, pages, vmas,
1793 &start, &nr_pages, i, gup_flags);
1799 unsigned int foll_flags = gup_flags;
1800 unsigned int page_increm;
1803 * If we have a pending SIGKILL, don't keep faulting
1804 * pages and potentially allocating memory.
1806 if (unlikely(fatal_signal_pending(current)))
1807 return i ? i : -ERESTARTSYS;
1810 while (!(page = follow_page_mask(vma, start,
1811 foll_flags, &page_mask))) {
1813 unsigned int fault_flags = 0;
1815 /* For mlock, just skip the stack guard page. */
1816 if (foll_flags & FOLL_MLOCK) {
1817 if (stack_guard_page(vma, start))
1820 if (foll_flags & FOLL_WRITE)
1821 fault_flags |= FAULT_FLAG_WRITE;
1823 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1824 if (foll_flags & FOLL_NOWAIT)
1825 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1827 ret = handle_mm_fault(mm, vma, start,
1830 if (ret & VM_FAULT_ERROR) {
1831 if (ret & VM_FAULT_OOM)
1832 return i ? i : -ENOMEM;
1833 if (ret & (VM_FAULT_HWPOISON |
1834 VM_FAULT_HWPOISON_LARGE)) {
1837 else if (gup_flags & FOLL_HWPOISON)
1842 if (ret & VM_FAULT_SIGBUS)
1843 return i ? i : -EFAULT;
1848 if (ret & VM_FAULT_MAJOR)
1854 if (ret & VM_FAULT_RETRY) {
1861 * The VM_FAULT_WRITE bit tells us that
1862 * do_wp_page has broken COW when necessary,
1863 * even if maybe_mkwrite decided not to set
1864 * pte_write. We can thus safely do subsequent
1865 * page lookups as if they were reads. But only
1866 * do so when looping for pte_write is futile:
1867 * in some cases userspace may also be wanting
1868 * to write to the gotten user page, which a
1869 * read fault here might prevent (a readonly
1870 * page might get reCOWed by userspace write).
1872 if ((ret & VM_FAULT_WRITE) &&
1873 !(vma->vm_flags & VM_WRITE))
1874 foll_flags &= ~FOLL_WRITE;
1879 return i ? i : PTR_ERR(page);
1883 flush_anon_page(vma, page, start);
1884 flush_dcache_page(page);
1892 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1893 if (page_increm > nr_pages)
1894 page_increm = nr_pages;
1896 start += page_increm * PAGE_SIZE;
1897 nr_pages -= page_increm;
1898 } while (nr_pages && start < vma->vm_end);
1902 EXPORT_SYMBOL(__get_user_pages);
1905 * fixup_user_fault() - manually resolve a user page fault
1906 * @tsk: the task_struct to use for page fault accounting, or
1907 * NULL if faults are not to be recorded.
1908 * @mm: mm_struct of target mm
1909 * @address: user address
1910 * @fault_flags:flags to pass down to handle_mm_fault()
1912 * This is meant to be called in the specific scenario where for locking reasons
1913 * we try to access user memory in atomic context (within a pagefault_disable()
1914 * section), this returns -EFAULT, and we want to resolve the user fault before
1917 * Typically this is meant to be used by the futex code.
1919 * The main difference with get_user_pages() is that this function will
1920 * unconditionally call handle_mm_fault() which will in turn perform all the
1921 * necessary SW fixup of the dirty and young bits in the PTE, while
1922 * handle_mm_fault() only guarantees to update these in the struct page.
1924 * This is important for some architectures where those bits also gate the
1925 * access permission to the page because they are maintained in software. On
1926 * such architectures, gup() will not be enough to make a subsequent access
1929 * This should be called with the mm_sem held for read.
1931 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1932 unsigned long address, unsigned int fault_flags)
1934 struct vm_area_struct *vma;
1937 vma = find_extend_vma(mm, address);
1938 if (!vma || address < vma->vm_start)
1941 ret = handle_mm_fault(mm, vma, address, fault_flags);
1942 if (ret & VM_FAULT_ERROR) {
1943 if (ret & VM_FAULT_OOM)
1945 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1947 if (ret & VM_FAULT_SIGBUS)
1952 if (ret & VM_FAULT_MAJOR)
1961 * get_user_pages() - pin user pages in memory
1962 * @tsk: the task_struct to use for page fault accounting, or
1963 * NULL if faults are not to be recorded.
1964 * @mm: mm_struct of target mm
1965 * @start: starting user address
1966 * @nr_pages: number of pages from start to pin
1967 * @write: whether pages will be written to by the caller
1968 * @force: whether to force write access even if user mapping is
1969 * readonly. This will result in the page being COWed even
1970 * in MAP_SHARED mappings. You do not want this.
1971 * @pages: array that receives pointers to the pages pinned.
1972 * Should be at least nr_pages long. Or NULL, if caller
1973 * only intends to ensure the pages are faulted in.
1974 * @vmas: array of pointers to vmas corresponding to each page.
1975 * Or NULL if the caller does not require them.
1977 * Returns number of pages pinned. This may be fewer than the number
1978 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1979 * were pinned, returns -errno. Each page returned must be released
1980 * with a put_page() call when it is finished with. vmas will only
1981 * remain valid while mmap_sem is held.
1983 * Must be called with mmap_sem held for read or write.
1985 * get_user_pages walks a process's page tables and takes a reference to
1986 * each struct page that each user address corresponds to at a given
1987 * instant. That is, it takes the page that would be accessed if a user
1988 * thread accesses the given user virtual address at that instant.
1990 * This does not guarantee that the page exists in the user mappings when
1991 * get_user_pages returns, and there may even be a completely different
1992 * page there in some cases (eg. if mmapped pagecache has been invalidated
1993 * and subsequently re faulted). However it does guarantee that the page
1994 * won't be freed completely. And mostly callers simply care that the page
1995 * contains data that was valid *at some point in time*. Typically, an IO
1996 * or similar operation cannot guarantee anything stronger anyway because
1997 * locks can't be held over the syscall boundary.
1999 * If write=0, the page must not be written to. If the page is written to,
2000 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2001 * after the page is finished with, and before put_page is called.
2003 * get_user_pages is typically used for fewer-copy IO operations, to get a
2004 * handle on the memory by some means other than accesses via the user virtual
2005 * addresses. The pages may be submitted for DMA to devices or accessed via
2006 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2007 * use the correct cache flushing APIs.
2009 * See also get_user_pages_fast, for performance critical applications.
2011 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2012 unsigned long start, unsigned long nr_pages, int write,
2013 int force, struct page **pages, struct vm_area_struct **vmas)
2015 int flags = FOLL_TOUCH;
2020 flags |= FOLL_WRITE;
2022 flags |= FOLL_FORCE;
2024 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2027 EXPORT_SYMBOL(get_user_pages);
2030 * get_dump_page() - pin user page in memory while writing it to core dump
2031 * @addr: user address
2033 * Returns struct page pointer of user page pinned for dump,
2034 * to be freed afterwards by page_cache_release() or put_page().
2036 * Returns NULL on any kind of failure - a hole must then be inserted into
2037 * the corefile, to preserve alignment with its headers; and also returns
2038 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2039 * allowing a hole to be left in the corefile to save diskspace.
2041 * Called without mmap_sem, but after all other threads have been killed.
2043 #ifdef CONFIG_ELF_CORE
2044 struct page *get_dump_page(unsigned long addr)
2046 struct vm_area_struct *vma;
2049 if (__get_user_pages(current, current->mm, addr, 1,
2050 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2053 flush_cache_page(vma, addr, page_to_pfn(page));
2056 #endif /* CONFIG_ELF_CORE */
2058 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2061 pgd_t * pgd = pgd_offset(mm, addr);
2062 pud_t * pud = pud_alloc(mm, pgd, addr);
2064 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2066 VM_BUG_ON(pmd_trans_huge(*pmd));
2067 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2074 * This is the old fallback for page remapping.
2076 * For historical reasons, it only allows reserved pages. Only
2077 * old drivers should use this, and they needed to mark their
2078 * pages reserved for the old functions anyway.
2080 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2081 struct page *page, pgprot_t prot)
2083 struct mm_struct *mm = vma->vm_mm;
2092 flush_dcache_page(page);
2093 pte = get_locked_pte(mm, addr, &ptl);
2097 if (!pte_none(*pte))
2100 /* Ok, finally just insert the thing.. */
2102 inc_mm_counter_fast(mm, MM_FILEPAGES);
2103 page_add_file_rmap(page);
2104 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2107 pte_unmap_unlock(pte, ptl);
2110 pte_unmap_unlock(pte, ptl);
2116 * vm_insert_page - insert single page into user vma
2117 * @vma: user vma to map to
2118 * @addr: target user address of this page
2119 * @page: source kernel page
2121 * This allows drivers to insert individual pages they've allocated
2124 * The page has to be a nice clean _individual_ kernel allocation.
2125 * If you allocate a compound page, you need to have marked it as
2126 * such (__GFP_COMP), or manually just split the page up yourself
2127 * (see split_page()).
2129 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2130 * took an arbitrary page protection parameter. This doesn't allow
2131 * that. Your vma protection will have to be set up correctly, which
2132 * means that if you want a shared writable mapping, you'd better
2133 * ask for a shared writable mapping!
2135 * The page does not need to be reserved.
2137 * Usually this function is called from f_op->mmap() handler
2138 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2139 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2140 * function from other places, for example from page-fault handler.
2142 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2145 if (addr < vma->vm_start || addr >= vma->vm_end)
2147 if (!page_count(page))
2149 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2150 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2151 BUG_ON(vma->vm_flags & VM_PFNMAP);
2152 vma->vm_flags |= VM_MIXEDMAP;
2154 return insert_page(vma, addr, page, vma->vm_page_prot);
2156 EXPORT_SYMBOL(vm_insert_page);
2158 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2159 unsigned long pfn, pgprot_t prot)
2161 struct mm_struct *mm = vma->vm_mm;
2167 pte = get_locked_pte(mm, addr, &ptl);
2171 if (!pte_none(*pte))
2174 /* Ok, finally just insert the thing.. */
2175 entry = pte_mkspecial(pfn_pte(pfn, prot));
2176 set_pte_at(mm, addr, pte, entry);
2177 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2181 pte_unmap_unlock(pte, ptl);
2187 * vm_insert_pfn - insert single pfn into user vma
2188 * @vma: user vma to map to
2189 * @addr: target user address of this page
2190 * @pfn: source kernel pfn
2192 * Similar to vm_insert_page, this allows drivers to insert individual pages
2193 * they've allocated into a user vma. Same comments apply.
2195 * This function should only be called from a vm_ops->fault handler, and
2196 * in that case the handler should return NULL.
2198 * vma cannot be a COW mapping.
2200 * As this is called only for pages that do not currently exist, we
2201 * do not need to flush old virtual caches or the TLB.
2203 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2207 pgprot_t pgprot = vma->vm_page_prot;
2209 * Technically, architectures with pte_special can avoid all these
2210 * restrictions (same for remap_pfn_range). However we would like
2211 * consistency in testing and feature parity among all, so we should
2212 * try to keep these invariants in place for everybody.
2214 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2215 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2216 (VM_PFNMAP|VM_MIXEDMAP));
2217 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2218 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2220 if (addr < vma->vm_start || addr >= vma->vm_end)
2222 if (track_pfn_insert(vma, &pgprot, pfn))
2225 ret = insert_pfn(vma, addr, pfn, pgprot);
2229 EXPORT_SYMBOL(vm_insert_pfn);
2231 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2234 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2236 if (addr < vma->vm_start || addr >= vma->vm_end)
2240 * If we don't have pte special, then we have to use the pfn_valid()
2241 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2242 * refcount the page if pfn_valid is true (hence insert_page rather
2243 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2244 * without pte special, it would there be refcounted as a normal page.
2246 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2249 page = pfn_to_page(pfn);
2250 return insert_page(vma, addr, page, vma->vm_page_prot);
2252 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2254 EXPORT_SYMBOL(vm_insert_mixed);
2257 * maps a range of physical memory into the requested pages. the old
2258 * mappings are removed. any references to nonexistent pages results
2259 * in null mappings (currently treated as "copy-on-access")
2261 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2262 unsigned long addr, unsigned long end,
2263 unsigned long pfn, pgprot_t prot)
2268 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2271 arch_enter_lazy_mmu_mode();
2273 BUG_ON(!pte_none(*pte));
2274 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2276 } while (pte++, addr += PAGE_SIZE, addr != end);
2277 arch_leave_lazy_mmu_mode();
2278 pte_unmap_unlock(pte - 1, ptl);
2282 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2283 unsigned long addr, unsigned long end,
2284 unsigned long pfn, pgprot_t prot)
2289 pfn -= addr >> PAGE_SHIFT;
2290 pmd = pmd_alloc(mm, pud, addr);
2293 VM_BUG_ON(pmd_trans_huge(*pmd));
2295 next = pmd_addr_end(addr, end);
2296 if (remap_pte_range(mm, pmd, addr, next,
2297 pfn + (addr >> PAGE_SHIFT), prot))
2299 } while (pmd++, addr = next, addr != end);
2303 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2310 pfn -= addr >> PAGE_SHIFT;
2311 pud = pud_alloc(mm, pgd, addr);
2315 next = pud_addr_end(addr, end);
2316 if (remap_pmd_range(mm, pud, addr, next,
2317 pfn + (addr >> PAGE_SHIFT), prot))
2319 } while (pud++, addr = next, addr != end);
2324 * remap_pfn_range - remap kernel memory to userspace
2325 * @vma: user vma to map to
2326 * @addr: target user address to start at
2327 * @pfn: physical address of kernel memory
2328 * @size: size of map area
2329 * @prot: page protection flags for this mapping
2331 * Note: this is only safe if the mm semaphore is held when called.
2333 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2334 unsigned long pfn, unsigned long size, pgprot_t prot)
2338 unsigned long end = addr + PAGE_ALIGN(size);
2339 struct mm_struct *mm = vma->vm_mm;
2343 * Physically remapped pages are special. Tell the
2344 * rest of the world about it:
2345 * VM_IO tells people not to look at these pages
2346 * (accesses can have side effects).
2347 * VM_PFNMAP tells the core MM that the base pages are just
2348 * raw PFN mappings, and do not have a "struct page" associated
2351 * Disable vma merging and expanding with mremap().
2353 * Omit vma from core dump, even when VM_IO turned off.
2355 * There's a horrible special case to handle copy-on-write
2356 * behaviour that some programs depend on. We mark the "original"
2357 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2358 * See vm_normal_page() for details.
2360 if (is_cow_mapping(vma->vm_flags)) {
2361 if (addr != vma->vm_start || end != vma->vm_end)
2363 vma->vm_pgoff = pfn;
2366 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2370 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2372 BUG_ON(addr >= end);
2373 pfn -= addr >> PAGE_SHIFT;
2374 pgd = pgd_offset(mm, addr);
2375 flush_cache_range(vma, addr, end);
2377 next = pgd_addr_end(addr, end);
2378 err = remap_pud_range(mm, pgd, addr, next,
2379 pfn + (addr >> PAGE_SHIFT), prot);
2382 } while (pgd++, addr = next, addr != end);
2385 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2389 EXPORT_SYMBOL(remap_pfn_range);
2392 * vm_iomap_memory - remap memory to userspace
2393 * @vma: user vma to map to
2394 * @start: start of area
2395 * @len: size of area
2397 * This is a simplified io_remap_pfn_range() for common driver use. The
2398 * driver just needs to give us the physical memory range to be mapped,
2399 * we'll figure out the rest from the vma information.
2401 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2402 * whatever write-combining details or similar.
2404 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2406 unsigned long vm_len, pfn, pages;
2408 /* Check that the physical memory area passed in looks valid */
2409 if (start + len < start)
2412 * You *really* shouldn't map things that aren't page-aligned,
2413 * but we've historically allowed it because IO memory might
2414 * just have smaller alignment.
2416 len += start & ~PAGE_MASK;
2417 pfn = start >> PAGE_SHIFT;
2418 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2419 if (pfn + pages < pfn)
2422 /* We start the mapping 'vm_pgoff' pages into the area */
2423 if (vma->vm_pgoff > pages)
2425 pfn += vma->vm_pgoff;
2426 pages -= vma->vm_pgoff;
2428 /* Can we fit all of the mapping? */
2429 vm_len = vma->vm_end - vma->vm_start;
2430 if (vm_len >> PAGE_SHIFT > pages)
2433 /* Ok, let it rip */
2434 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2436 EXPORT_SYMBOL(vm_iomap_memory);
2438 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2439 unsigned long addr, unsigned long end,
2440 pte_fn_t fn, void *data)
2445 spinlock_t *uninitialized_var(ptl);
2447 pte = (mm == &init_mm) ?
2448 pte_alloc_kernel(pmd, addr) :
2449 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2453 BUG_ON(pmd_huge(*pmd));
2455 arch_enter_lazy_mmu_mode();
2457 token = pmd_pgtable(*pmd);
2460 err = fn(pte++, token, addr, data);
2463 } while (addr += PAGE_SIZE, addr != end);
2465 arch_leave_lazy_mmu_mode();
2468 pte_unmap_unlock(pte-1, ptl);
2472 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2473 unsigned long addr, unsigned long end,
2474 pte_fn_t fn, void *data)
2480 BUG_ON(pud_huge(*pud));
2482 pmd = pmd_alloc(mm, pud, addr);
2486 next = pmd_addr_end(addr, end);
2487 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2490 } while (pmd++, addr = next, addr != end);
2494 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2495 unsigned long addr, unsigned long end,
2496 pte_fn_t fn, void *data)
2502 pud = pud_alloc(mm, pgd, addr);
2506 next = pud_addr_end(addr, end);
2507 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2510 } while (pud++, addr = next, addr != end);
2515 * Scan a region of virtual memory, filling in page tables as necessary
2516 * and calling a provided function on each leaf page table.
2518 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2519 unsigned long size, pte_fn_t fn, void *data)
2523 unsigned long end = addr + size;
2526 BUG_ON(addr >= end);
2527 pgd = pgd_offset(mm, addr);
2529 next = pgd_addr_end(addr, end);
2530 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2533 } while (pgd++, addr = next, addr != end);
2537 EXPORT_SYMBOL_GPL(apply_to_page_range);
2540 * handle_pte_fault chooses page fault handler according to an entry
2541 * which was read non-atomically. Before making any commitment, on
2542 * those architectures or configurations (e.g. i386 with PAE) which
2543 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2544 * must check under lock before unmapping the pte and proceeding
2545 * (but do_wp_page is only called after already making such a check;
2546 * and do_anonymous_page can safely check later on).
2548 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2549 pte_t *page_table, pte_t orig_pte)
2552 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2553 if (sizeof(pte_t) > sizeof(unsigned long)) {
2554 spinlock_t *ptl = pte_lockptr(mm, pmd);
2556 same = pte_same(*page_table, orig_pte);
2560 pte_unmap(page_table);
2564 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2567 * If the source page was a PFN mapping, we don't have
2568 * a "struct page" for it. We do a best-effort copy by
2569 * just copying from the original user address. If that
2570 * fails, we just zero-fill it. Live with it.
2572 if (unlikely(!src)) {
2573 void *kaddr = kmap_atomic(dst);
2574 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2577 * This really shouldn't fail, because the page is there
2578 * in the page tables. But it might just be unreadable,
2579 * in which case we just give up and fill the result with
2582 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2584 kunmap_atomic(kaddr);
2585 flush_dcache_page(dst);
2587 copy_user_highpage(dst, src, va, vma);
2591 * This routine handles present pages, when users try to write
2592 * to a shared page. It is done by copying the page to a new address
2593 * and decrementing the shared-page counter for the old page.
2595 * Note that this routine assumes that the protection checks have been
2596 * done by the caller (the low-level page fault routine in most cases).
2597 * Thus we can safely just mark it writable once we've done any necessary
2600 * We also mark the page dirty at this point even though the page will
2601 * change only once the write actually happens. This avoids a few races,
2602 * and potentially makes it more efficient.
2604 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2605 * but allow concurrent faults), with pte both mapped and locked.
2606 * We return with mmap_sem still held, but pte unmapped and unlocked.
2608 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2609 unsigned long address, pte_t *page_table, pmd_t *pmd,
2610 spinlock_t *ptl, pte_t orig_pte)
2613 struct page *old_page, *new_page = NULL;
2616 int page_mkwrite = 0;
2617 struct page *dirty_page = NULL;
2618 unsigned long mmun_start = 0; /* For mmu_notifiers */
2619 unsigned long mmun_end = 0; /* For mmu_notifiers */
2621 old_page = vm_normal_page(vma, address, orig_pte);
2624 * VM_MIXEDMAP !pfn_valid() case
2626 * We should not cow pages in a shared writeable mapping.
2627 * Just mark the pages writable as we can't do any dirty
2628 * accounting on raw pfn maps.
2630 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2631 (VM_WRITE|VM_SHARED))
2637 * Take out anonymous pages first, anonymous shared vmas are
2638 * not dirty accountable.
2640 if (PageAnon(old_page) && !PageKsm(old_page)) {
2641 if (!trylock_page(old_page)) {
2642 page_cache_get(old_page);
2643 pte_unmap_unlock(page_table, ptl);
2644 lock_page(old_page);
2645 page_table = pte_offset_map_lock(mm, pmd, address,
2647 if (!pte_same(*page_table, orig_pte)) {
2648 unlock_page(old_page);
2651 page_cache_release(old_page);
2653 if (reuse_swap_page(old_page)) {
2655 * The page is all ours. Move it to our anon_vma so
2656 * the rmap code will not search our parent or siblings.
2657 * Protected against the rmap code by the page lock.
2659 page_move_anon_rmap(old_page, vma, address);
2660 unlock_page(old_page);
2663 unlock_page(old_page);
2664 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2665 (VM_WRITE|VM_SHARED))) {
2667 * Only catch write-faults on shared writable pages,
2668 * read-only shared pages can get COWed by
2669 * get_user_pages(.write=1, .force=1).
2671 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2672 struct vm_fault vmf;
2675 vmf.virtual_address = (void __user *)(address &
2677 vmf.pgoff = old_page->index;
2678 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2679 vmf.page = old_page;
2682 * Notify the address space that the page is about to
2683 * become writable so that it can prohibit this or wait
2684 * for the page to get into an appropriate state.
2686 * We do this without the lock held, so that it can
2687 * sleep if it needs to.
2689 page_cache_get(old_page);
2690 pte_unmap_unlock(page_table, ptl);
2692 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2694 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2696 goto unwritable_page;
2698 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2699 lock_page(old_page);
2700 if (!old_page->mapping) {
2701 ret = 0; /* retry the fault */
2702 unlock_page(old_page);
2703 goto unwritable_page;
2706 VM_BUG_ON(!PageLocked(old_page));
2709 * Since we dropped the lock we need to revalidate
2710 * the PTE as someone else may have changed it. If
2711 * they did, we just return, as we can count on the
2712 * MMU to tell us if they didn't also make it writable.
2714 page_table = pte_offset_map_lock(mm, pmd, address,
2716 if (!pte_same(*page_table, orig_pte)) {
2717 unlock_page(old_page);
2723 dirty_page = old_page;
2724 get_page(dirty_page);
2727 flush_cache_page(vma, address, pte_pfn(orig_pte));
2728 entry = pte_mkyoung(orig_pte);
2729 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2730 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2731 update_mmu_cache(vma, address, page_table);
2732 pte_unmap_unlock(page_table, ptl);
2733 ret |= VM_FAULT_WRITE;
2739 * Yes, Virginia, this is actually required to prevent a race
2740 * with clear_page_dirty_for_io() from clearing the page dirty
2741 * bit after it clear all dirty ptes, but before a racing
2742 * do_wp_page installs a dirty pte.
2744 * __do_fault is protected similarly.
2746 if (!page_mkwrite) {
2747 wait_on_page_locked(dirty_page);
2748 set_page_dirty_balance(dirty_page, page_mkwrite);
2749 /* file_update_time outside page_lock */
2751 file_update_time(vma->vm_file);
2753 put_page(dirty_page);
2755 struct address_space *mapping = dirty_page->mapping;
2757 set_page_dirty(dirty_page);
2758 unlock_page(dirty_page);
2759 page_cache_release(dirty_page);
2762 * Some device drivers do not set page.mapping
2763 * but still dirty their pages
2765 balance_dirty_pages_ratelimited(mapping);
2773 * Ok, we need to copy. Oh, well..
2775 page_cache_get(old_page);
2777 pte_unmap_unlock(page_table, ptl);
2779 if (unlikely(anon_vma_prepare(vma)))
2782 if (is_zero_pfn(pte_pfn(orig_pte))) {
2783 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2787 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2790 cow_user_page(new_page, old_page, address, vma);
2792 __SetPageUptodate(new_page);
2794 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2797 mmun_start = address & PAGE_MASK;
2798 mmun_end = mmun_start + PAGE_SIZE;
2799 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2802 * Re-check the pte - we dropped the lock
2804 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2805 if (likely(pte_same(*page_table, orig_pte))) {
2807 if (!PageAnon(old_page)) {
2808 dec_mm_counter_fast(mm, MM_FILEPAGES);
2809 inc_mm_counter_fast(mm, MM_ANONPAGES);
2812 inc_mm_counter_fast(mm, MM_ANONPAGES);
2813 flush_cache_page(vma, address, pte_pfn(orig_pte));
2814 entry = mk_pte(new_page, vma->vm_page_prot);
2815 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2817 * Clear the pte entry and flush it first, before updating the
2818 * pte with the new entry. This will avoid a race condition
2819 * seen in the presence of one thread doing SMC and another
2822 ptep_clear_flush(vma, address, page_table);
2823 page_add_new_anon_rmap(new_page, vma, address);
2825 * We call the notify macro here because, when using secondary
2826 * mmu page tables (such as kvm shadow page tables), we want the
2827 * new page to be mapped directly into the secondary page table.
2829 set_pte_at_notify(mm, address, page_table, entry);
2830 update_mmu_cache(vma, address, page_table);
2833 * Only after switching the pte to the new page may
2834 * we remove the mapcount here. Otherwise another
2835 * process may come and find the rmap count decremented
2836 * before the pte is switched to the new page, and
2837 * "reuse" the old page writing into it while our pte
2838 * here still points into it and can be read by other
2841 * The critical issue is to order this
2842 * page_remove_rmap with the ptp_clear_flush above.
2843 * Those stores are ordered by (if nothing else,)
2844 * the barrier present in the atomic_add_negative
2845 * in page_remove_rmap.
2847 * Then the TLB flush in ptep_clear_flush ensures that
2848 * no process can access the old page before the
2849 * decremented mapcount is visible. And the old page
2850 * cannot be reused until after the decremented
2851 * mapcount is visible. So transitively, TLBs to
2852 * old page will be flushed before it can be reused.
2854 page_remove_rmap(old_page);
2857 /* Free the old page.. */
2858 new_page = old_page;
2859 ret |= VM_FAULT_WRITE;
2861 mem_cgroup_uncharge_page(new_page);
2864 page_cache_release(new_page);
2866 pte_unmap_unlock(page_table, ptl);
2867 if (mmun_end > mmun_start)
2868 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2871 * Don't let another task, with possibly unlocked vma,
2872 * keep the mlocked page.
2874 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2875 lock_page(old_page); /* LRU manipulation */
2876 munlock_vma_page(old_page);
2877 unlock_page(old_page);
2879 page_cache_release(old_page);
2883 page_cache_release(new_page);
2886 page_cache_release(old_page);
2887 return VM_FAULT_OOM;
2890 page_cache_release(old_page);
2894 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2895 unsigned long start_addr, unsigned long end_addr,
2896 struct zap_details *details)
2898 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2901 static inline void unmap_mapping_range_tree(struct rb_root *root,
2902 struct zap_details *details)
2904 struct vm_area_struct *vma;
2905 pgoff_t vba, vea, zba, zea;
2907 vma_interval_tree_foreach(vma, root,
2908 details->first_index, details->last_index) {
2910 vba = vma->vm_pgoff;
2911 vea = vba + vma_pages(vma) - 1;
2912 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2913 zba = details->first_index;
2916 zea = details->last_index;
2920 unmap_mapping_range_vma(vma,
2921 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2922 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2927 static inline void unmap_mapping_range_list(struct list_head *head,
2928 struct zap_details *details)
2930 struct vm_area_struct *vma;
2933 * In nonlinear VMAs there is no correspondence between virtual address
2934 * offset and file offset. So we must perform an exhaustive search
2935 * across *all* the pages in each nonlinear VMA, not just the pages
2936 * whose virtual address lies outside the file truncation point.
2938 list_for_each_entry(vma, head, shared.nonlinear) {
2939 details->nonlinear_vma = vma;
2940 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2945 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2946 * @mapping: the address space containing mmaps to be unmapped.
2947 * @holebegin: byte in first page to unmap, relative to the start of
2948 * the underlying file. This will be rounded down to a PAGE_SIZE
2949 * boundary. Note that this is different from truncate_pagecache(), which
2950 * must keep the partial page. In contrast, we must get rid of
2952 * @holelen: size of prospective hole in bytes. This will be rounded
2953 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2955 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2956 * but 0 when invalidating pagecache, don't throw away private data.
2958 void unmap_mapping_range(struct address_space *mapping,
2959 loff_t const holebegin, loff_t const holelen, int even_cows)
2961 struct zap_details details;
2962 pgoff_t hba = holebegin >> PAGE_SHIFT;
2963 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2965 /* Check for overflow. */
2966 if (sizeof(holelen) > sizeof(hlen)) {
2968 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2969 if (holeend & ~(long long)ULONG_MAX)
2970 hlen = ULONG_MAX - hba + 1;
2973 details.check_mapping = even_cows? NULL: mapping;
2974 details.nonlinear_vma = NULL;
2975 details.first_index = hba;
2976 details.last_index = hba + hlen - 1;
2977 if (details.last_index < details.first_index)
2978 details.last_index = ULONG_MAX;
2981 mutex_lock(&mapping->i_mmap_mutex);
2982 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2983 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2984 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
2985 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
2986 mutex_unlock(&mapping->i_mmap_mutex);
2988 EXPORT_SYMBOL(unmap_mapping_range);
2991 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2992 * but allow concurrent faults), and pte mapped but not yet locked.
2993 * We return with mmap_sem still held, but pte unmapped and unlocked.
2995 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2996 unsigned long address, pte_t *page_table, pmd_t *pmd,
2997 unsigned int flags, pte_t orig_pte)
3000 struct page *page, *swapcache;
3004 struct mem_cgroup *ptr;
3008 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3011 entry = pte_to_swp_entry(orig_pte);
3012 if (unlikely(non_swap_entry(entry))) {
3013 if (is_migration_entry(entry)) {
3014 migration_entry_wait(mm, pmd, address);
3015 } else if (is_hwpoison_entry(entry)) {
3016 ret = VM_FAULT_HWPOISON;
3018 print_bad_pte(vma, address, orig_pte, NULL);
3019 ret = VM_FAULT_SIGBUS;
3023 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3024 page = lookup_swap_cache(entry);
3026 page = swapin_readahead(entry,
3027 GFP_HIGHUSER_MOVABLE, vma, address);
3030 * Back out if somebody else faulted in this pte
3031 * while we released the pte lock.
3033 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3034 if (likely(pte_same(*page_table, orig_pte)))
3036 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3040 /* Had to read the page from swap area: Major fault */
3041 ret = VM_FAULT_MAJOR;
3042 count_vm_event(PGMAJFAULT);
3043 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3044 } else if (PageHWPoison(page)) {
3046 * hwpoisoned dirty swapcache pages are kept for killing
3047 * owner processes (which may be unknown at hwpoison time)
3049 ret = VM_FAULT_HWPOISON;
3050 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3056 locked = lock_page_or_retry(page, mm, flags);
3058 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3060 ret |= VM_FAULT_RETRY;
3065 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3066 * release the swapcache from under us. The page pin, and pte_same
3067 * test below, are not enough to exclude that. Even if it is still
3068 * swapcache, we need to check that the page's swap has not changed.
3070 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3073 page = ksm_might_need_to_copy(page, vma, address);
3074 if (unlikely(!page)) {
3080 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3086 * Back out if somebody else already faulted in this pte.
3088 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3089 if (unlikely(!pte_same(*page_table, orig_pte)))
3092 if (unlikely(!PageUptodate(page))) {
3093 ret = VM_FAULT_SIGBUS;
3098 * The page isn't present yet, go ahead with the fault.
3100 * Be careful about the sequence of operations here.
3101 * To get its accounting right, reuse_swap_page() must be called
3102 * while the page is counted on swap but not yet in mapcount i.e.
3103 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3104 * must be called after the swap_free(), or it will never succeed.
3105 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3106 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3107 * in page->private. In this case, a record in swap_cgroup is silently
3108 * discarded at swap_free().
3111 inc_mm_counter_fast(mm, MM_ANONPAGES);
3112 dec_mm_counter_fast(mm, MM_SWAPENTS);
3113 pte = mk_pte(page, vma->vm_page_prot);
3114 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3115 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3116 flags &= ~FAULT_FLAG_WRITE;
3117 ret |= VM_FAULT_WRITE;
3120 flush_icache_page(vma, page);
3121 if (pte_swp_soft_dirty(orig_pte))
3122 pte = pte_mksoft_dirty(pte);
3123 set_pte_at(mm, address, page_table, pte);
3124 if (page == swapcache)
3125 do_page_add_anon_rmap(page, vma, address, exclusive);
3126 else /* ksm created a completely new copy */
3127 page_add_new_anon_rmap(page, vma, address);
3128 /* It's better to call commit-charge after rmap is established */
3129 mem_cgroup_commit_charge_swapin(page, ptr);
3132 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3133 try_to_free_swap(page);
3135 if (page != swapcache) {
3137 * Hold the lock to avoid the swap entry to be reused
3138 * until we take the PT lock for the pte_same() check
3139 * (to avoid false positives from pte_same). For
3140 * further safety release the lock after the swap_free
3141 * so that the swap count won't change under a
3142 * parallel locked swapcache.
3144 unlock_page(swapcache);
3145 page_cache_release(swapcache);
3148 if (flags & FAULT_FLAG_WRITE) {
3149 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3150 if (ret & VM_FAULT_ERROR)
3151 ret &= VM_FAULT_ERROR;
3155 /* No need to invalidate - it was non-present before */
3156 update_mmu_cache(vma, address, page_table);
3158 pte_unmap_unlock(page_table, ptl);
3162 mem_cgroup_cancel_charge_swapin(ptr);
3163 pte_unmap_unlock(page_table, ptl);
3167 page_cache_release(page);
3168 if (page != swapcache) {
3169 unlock_page(swapcache);
3170 page_cache_release(swapcache);
3176 * This is like a special single-page "expand_{down|up}wards()",
3177 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3178 * doesn't hit another vma.
3180 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3182 address &= PAGE_MASK;
3183 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3184 struct vm_area_struct *prev = vma->vm_prev;
3187 * Is there a mapping abutting this one below?
3189 * That's only ok if it's the same stack mapping
3190 * that has gotten split..
3192 if (prev && prev->vm_end == address)
3193 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3195 expand_downwards(vma, address - PAGE_SIZE);
3197 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3198 struct vm_area_struct *next = vma->vm_next;
3200 /* As VM_GROWSDOWN but s/below/above/ */
3201 if (next && next->vm_start == address + PAGE_SIZE)
3202 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3204 expand_upwards(vma, address + PAGE_SIZE);
3210 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3211 * but allow concurrent faults), and pte mapped but not yet locked.
3212 * We return with mmap_sem still held, but pte unmapped and unlocked.
3214 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3215 unsigned long address, pte_t *page_table, pmd_t *pmd,
3222 pte_unmap(page_table);
3224 /* Check if we need to add a guard page to the stack */
3225 if (check_stack_guard_page(vma, address) < 0)
3226 return VM_FAULT_SIGBUS;
3228 /* Use the zero-page for reads */
3229 if (!(flags & FAULT_FLAG_WRITE)) {
3230 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3231 vma->vm_page_prot));
3232 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3233 if (!pte_none(*page_table))
3238 /* Allocate our own private page. */
3239 if (unlikely(anon_vma_prepare(vma)))
3241 page = alloc_zeroed_user_highpage_movable(vma, address);
3245 * The memory barrier inside __SetPageUptodate makes sure that
3246 * preceeding stores to the page contents become visible before
3247 * the set_pte_at() write.
3249 __SetPageUptodate(page);
3251 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3254 entry = mk_pte(page, vma->vm_page_prot);
3255 if (vma->vm_flags & VM_WRITE)
3256 entry = pte_mkwrite(pte_mkdirty(entry));
3258 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3259 if (!pte_none(*page_table))
3262 inc_mm_counter_fast(mm, MM_ANONPAGES);
3263 page_add_new_anon_rmap(page, vma, address);
3265 set_pte_at(mm, address, page_table, entry);
3267 /* No need to invalidate - it was non-present before */
3268 update_mmu_cache(vma, address, page_table);
3270 pte_unmap_unlock(page_table, ptl);
3273 mem_cgroup_uncharge_page(page);
3274 page_cache_release(page);
3277 page_cache_release(page);
3279 return VM_FAULT_OOM;
3283 * __do_fault() tries to create a new page mapping. It aggressively
3284 * tries to share with existing pages, but makes a separate copy if
3285 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3286 * the next page fault.
3288 * As this is called only for pages that do not currently exist, we
3289 * do not need to flush old virtual caches or the TLB.
3291 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3292 * but allow concurrent faults), and pte neither mapped nor locked.
3293 * We return with mmap_sem still held, but pte unmapped and unlocked.
3295 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3296 unsigned long address, pmd_t *pmd,
3297 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3302 struct page *cow_page;
3305 struct page *dirty_page = NULL;
3306 struct vm_fault vmf;
3308 int page_mkwrite = 0;
3311 * If we do COW later, allocate page befor taking lock_page()
3312 * on the file cache page. This will reduce lock holding time.
3314 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3316 if (unlikely(anon_vma_prepare(vma)))
3317 return VM_FAULT_OOM;
3319 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3321 return VM_FAULT_OOM;
3323 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3324 page_cache_release(cow_page);
3325 return VM_FAULT_OOM;
3330 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3335 ret = vma->vm_ops->fault(vma, &vmf);
3336 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3340 if (unlikely(PageHWPoison(vmf.page))) {
3341 if (ret & VM_FAULT_LOCKED)
3342 unlock_page(vmf.page);
3343 ret = VM_FAULT_HWPOISON;
3348 * For consistency in subsequent calls, make the faulted page always
3351 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3352 lock_page(vmf.page);
3354 VM_BUG_ON(!PageLocked(vmf.page));
3357 * Should we do an early C-O-W break?
3360 if (flags & FAULT_FLAG_WRITE) {
3361 if (!(vma->vm_flags & VM_SHARED)) {
3364 copy_user_highpage(page, vmf.page, address, vma);
3365 __SetPageUptodate(page);
3368 * If the page will be shareable, see if the backing
3369 * address space wants to know that the page is about
3370 * to become writable
3372 if (vma->vm_ops->page_mkwrite) {
3376 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3377 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3379 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3381 goto unwritable_page;
3383 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3385 if (!page->mapping) {
3386 ret = 0; /* retry the fault */
3388 goto unwritable_page;
3391 VM_BUG_ON(!PageLocked(page));
3398 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3401 * This silly early PAGE_DIRTY setting removes a race
3402 * due to the bad i386 page protection. But it's valid
3403 * for other architectures too.
3405 * Note that if FAULT_FLAG_WRITE is set, we either now have
3406 * an exclusive copy of the page, or this is a shared mapping,
3407 * so we can make it writable and dirty to avoid having to
3408 * handle that later.
3410 /* Only go through if we didn't race with anybody else... */
3411 if (likely(pte_same(*page_table, orig_pte))) {
3412 flush_icache_page(vma, page);
3413 entry = mk_pte(page, vma->vm_page_prot);
3414 if (flags & FAULT_FLAG_WRITE)
3415 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3416 else if (pte_file(orig_pte) && pte_file_soft_dirty(orig_pte))
3417 pte_mksoft_dirty(entry);
3419 inc_mm_counter_fast(mm, MM_ANONPAGES);
3420 page_add_new_anon_rmap(page, vma, address);
3422 inc_mm_counter_fast(mm, MM_FILEPAGES);
3423 page_add_file_rmap(page);
3424 if (flags & FAULT_FLAG_WRITE) {
3426 get_page(dirty_page);
3429 set_pte_at(mm, address, page_table, entry);
3431 /* no need to invalidate: a not-present page won't be cached */
3432 update_mmu_cache(vma, address, page_table);
3435 mem_cgroup_uncharge_page(cow_page);
3437 page_cache_release(page);
3439 anon = 1; /* no anon but release faulted_page */
3442 pte_unmap_unlock(page_table, ptl);
3445 struct address_space *mapping = page->mapping;
3448 if (set_page_dirty(dirty_page))
3450 unlock_page(dirty_page);
3451 put_page(dirty_page);
3452 if ((dirtied || page_mkwrite) && mapping) {
3454 * Some device drivers do not set page.mapping but still
3457 balance_dirty_pages_ratelimited(mapping);
3460 /* file_update_time outside page_lock */
3461 if (vma->vm_file && !page_mkwrite)
3462 file_update_time(vma->vm_file);
3464 unlock_page(vmf.page);
3466 page_cache_release(vmf.page);
3472 page_cache_release(page);
3475 /* fs's fault handler get error */
3477 mem_cgroup_uncharge_page(cow_page);
3478 page_cache_release(cow_page);
3483 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3484 unsigned long address, pte_t *page_table, pmd_t *pmd,
3485 unsigned int flags, pte_t orig_pte)
3487 pgoff_t pgoff = (((address & PAGE_MASK)
3488 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3490 pte_unmap(page_table);
3491 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3495 * Fault of a previously existing named mapping. Repopulate the pte
3496 * from the encoded file_pte if possible. This enables swappable
3499 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3500 * but allow concurrent faults), and pte mapped but not yet locked.
3501 * We return with mmap_sem still held, but pte unmapped and unlocked.
3503 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3504 unsigned long address, pte_t *page_table, pmd_t *pmd,
3505 unsigned int flags, pte_t orig_pte)
3509 flags |= FAULT_FLAG_NONLINEAR;
3511 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3514 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3516 * Page table corrupted: show pte and kill process.
3518 print_bad_pte(vma, address, orig_pte, NULL);
3519 return VM_FAULT_SIGBUS;
3522 pgoff = pte_to_pgoff(orig_pte);
3523 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3526 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3527 unsigned long addr, int current_nid)
3531 count_vm_numa_event(NUMA_HINT_FAULTS);
3532 if (current_nid == numa_node_id())
3533 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3535 return mpol_misplaced(page, vma, addr);
3538 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3539 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3541 struct page *page = NULL;
3543 int current_nid = -1;
3545 bool migrated = false;
3548 * The "pte" at this point cannot be used safely without
3549 * validation through pte_unmap_same(). It's of NUMA type but
3550 * the pfn may be screwed if the read is non atomic.
3552 * ptep_modify_prot_start is not called as this is clearing
3553 * the _PAGE_NUMA bit and it is not really expected that there
3554 * would be concurrent hardware modifications to the PTE.
3556 ptl = pte_lockptr(mm, pmd);
3558 if (unlikely(!pte_same(*ptep, pte))) {
3559 pte_unmap_unlock(ptep, ptl);
3563 pte = pte_mknonnuma(pte);
3564 set_pte_at(mm, addr, ptep, pte);
3565 update_mmu_cache(vma, addr, ptep);
3567 page = vm_normal_page(vma, addr, pte);
3569 pte_unmap_unlock(ptep, ptl);
3573 current_nid = page_to_nid(page);
3574 target_nid = numa_migrate_prep(page, vma, addr, current_nid);
3575 pte_unmap_unlock(ptep, ptl);
3576 if (target_nid == -1) {
3578 * Account for the fault against the current node if it not
3579 * being replaced regardless of where the page is located.
3581 current_nid = numa_node_id();
3586 /* Migrate to the requested node */
3587 migrated = migrate_misplaced_page(page, target_nid);
3589 current_nid = target_nid;
3592 if (current_nid != -1)
3593 task_numa_fault(current_nid, 1, migrated);
3597 /* NUMA hinting page fault entry point for regular pmds */
3598 #ifdef CONFIG_NUMA_BALANCING
3599 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3600 unsigned long addr, pmd_t *pmdp)
3603 pte_t *pte, *orig_pte;
3604 unsigned long _addr = addr & PMD_MASK;
3605 unsigned long offset;
3608 int local_nid = numa_node_id();
3610 spin_lock(&mm->page_table_lock);
3612 if (pmd_numa(pmd)) {
3613 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3616 spin_unlock(&mm->page_table_lock);
3621 /* we're in a page fault so some vma must be in the range */
3623 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3624 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3625 VM_BUG_ON(offset >= PMD_SIZE);
3626 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3627 pte += offset >> PAGE_SHIFT;
3628 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3629 pte_t pteval = *pte;
3631 int curr_nid = local_nid;
3634 if (!pte_present(pteval))
3636 if (!pte_numa(pteval))
3638 if (addr >= vma->vm_end) {
3639 vma = find_vma(mm, addr);
3640 /* there's a pte present so there must be a vma */
3642 BUG_ON(addr < vma->vm_start);
3644 if (pte_numa(pteval)) {
3645 pteval = pte_mknonnuma(pteval);
3646 set_pte_at(mm, addr, pte, pteval);
3648 page = vm_normal_page(vma, addr, pteval);
3649 if (unlikely(!page))
3651 /* only check non-shared pages */
3652 if (unlikely(page_mapcount(page) != 1))
3656 * Note that the NUMA fault is later accounted to either
3657 * the node that is currently running or where the page is
3660 curr_nid = local_nid;
3661 target_nid = numa_migrate_prep(page, vma, addr,
3663 if (target_nid == -1) {
3668 /* Migrate to the requested node */
3669 pte_unmap_unlock(pte, ptl);
3670 migrated = migrate_misplaced_page(page, target_nid);
3672 curr_nid = target_nid;
3673 task_numa_fault(curr_nid, 1, migrated);
3675 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3677 pte_unmap_unlock(orig_pte, ptl);
3682 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3683 unsigned long addr, pmd_t *pmdp)
3688 #endif /* CONFIG_NUMA_BALANCING */
3691 * These routines also need to handle stuff like marking pages dirty
3692 * and/or accessed for architectures that don't do it in hardware (most
3693 * RISC architectures). The early dirtying is also good on the i386.
3695 * There is also a hook called "update_mmu_cache()" that architectures
3696 * with external mmu caches can use to update those (ie the Sparc or
3697 * PowerPC hashed page tables that act as extended TLBs).
3699 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3700 * but allow concurrent faults), and pte mapped but not yet locked.
3701 * We return with mmap_sem still held, but pte unmapped and unlocked.
3703 int handle_pte_fault(struct mm_struct *mm,
3704 struct vm_area_struct *vma, unsigned long address,
3705 pte_t *pte, pmd_t *pmd, unsigned int flags)
3711 if (!pte_present(entry)) {
3712 if (pte_none(entry)) {
3714 if (likely(vma->vm_ops->fault))
3715 return do_linear_fault(mm, vma, address,
3716 pte, pmd, flags, entry);
3718 return do_anonymous_page(mm, vma, address,
3721 if (pte_file(entry))
3722 return do_nonlinear_fault(mm, vma, address,
3723 pte, pmd, flags, entry);
3724 return do_swap_page(mm, vma, address,
3725 pte, pmd, flags, entry);
3728 if (pte_numa(entry))
3729 return do_numa_page(mm, vma, address, entry, pte, pmd);
3731 ptl = pte_lockptr(mm, pmd);
3733 if (unlikely(!pte_same(*pte, entry)))
3735 if (flags & FAULT_FLAG_WRITE) {
3736 if (!pte_write(entry))
3737 return do_wp_page(mm, vma, address,
3738 pte, pmd, ptl, entry);
3739 entry = pte_mkdirty(entry);
3741 entry = pte_mkyoung(entry);
3742 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3743 update_mmu_cache(vma, address, pte);
3746 * This is needed only for protection faults but the arch code
3747 * is not yet telling us if this is a protection fault or not.
3748 * This still avoids useless tlb flushes for .text page faults
3751 if (flags & FAULT_FLAG_WRITE)
3752 flush_tlb_fix_spurious_fault(vma, address);
3755 pte_unmap_unlock(pte, ptl);
3760 * By the time we get here, we already hold the mm semaphore
3762 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3763 unsigned long address, unsigned int flags)
3770 __set_current_state(TASK_RUNNING);
3772 count_vm_event(PGFAULT);
3773 mem_cgroup_count_vm_event(mm, PGFAULT);
3775 /* do counter updates before entering really critical section. */
3776 check_sync_rss_stat(current);
3778 if (unlikely(is_vm_hugetlb_page(vma)))
3779 return hugetlb_fault(mm, vma, address, flags);
3782 pgd = pgd_offset(mm, address);
3783 pud = pud_alloc(mm, pgd, address);
3785 return VM_FAULT_OOM;
3786 pmd = pmd_alloc(mm, pud, address);
3788 return VM_FAULT_OOM;
3789 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3791 return do_huge_pmd_anonymous_page(mm, vma, address,
3794 pmd_t orig_pmd = *pmd;
3798 if (pmd_trans_huge(orig_pmd)) {
3799 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3802 * If the pmd is splitting, return and retry the
3803 * the fault. Alternative: wait until the split
3804 * is done, and goto retry.
3806 if (pmd_trans_splitting(orig_pmd))
3809 if (pmd_numa(orig_pmd))
3810 return do_huge_pmd_numa_page(mm, vma, address,
3813 if (dirty && !pmd_write(orig_pmd)) {
3814 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3817 * If COW results in an oom, the huge pmd will
3818 * have been split, so retry the fault on the
3819 * pte for a smaller charge.
3821 if (unlikely(ret & VM_FAULT_OOM))
3825 huge_pmd_set_accessed(mm, vma, address, pmd,
3834 return do_pmd_numa_page(mm, vma, address, pmd);
3837 * Use __pte_alloc instead of pte_alloc_map, because we can't
3838 * run pte_offset_map on the pmd, if an huge pmd could
3839 * materialize from under us from a different thread.
3841 if (unlikely(pmd_none(*pmd)) &&
3842 unlikely(__pte_alloc(mm, vma, pmd, address)))
3843 return VM_FAULT_OOM;
3844 /* if an huge pmd materialized from under us just retry later */
3845 if (unlikely(pmd_trans_huge(*pmd)))
3848 * A regular pmd is established and it can't morph into a huge pmd
3849 * from under us anymore at this point because we hold the mmap_sem
3850 * read mode and khugepaged takes it in write mode. So now it's
3851 * safe to run pte_offset_map().
3853 pte = pte_offset_map(pmd, address);
3855 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3858 #ifndef __PAGETABLE_PUD_FOLDED
3860 * Allocate page upper directory.
3861 * We've already handled the fast-path in-line.
3863 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3865 pud_t *new = pud_alloc_one(mm, address);
3869 smp_wmb(); /* See comment in __pte_alloc */
3871 spin_lock(&mm->page_table_lock);
3872 if (pgd_present(*pgd)) /* Another has populated it */
3875 pgd_populate(mm, pgd, new);
3876 spin_unlock(&mm->page_table_lock);
3879 #endif /* __PAGETABLE_PUD_FOLDED */
3881 #ifndef __PAGETABLE_PMD_FOLDED
3883 * Allocate page middle directory.
3884 * We've already handled the fast-path in-line.
3886 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3888 pmd_t *new = pmd_alloc_one(mm, address);
3892 smp_wmb(); /* See comment in __pte_alloc */
3894 spin_lock(&mm->page_table_lock);
3895 #ifndef __ARCH_HAS_4LEVEL_HACK
3896 if (pud_present(*pud)) /* Another has populated it */
3899 pud_populate(mm, pud, new);
3901 if (pgd_present(*pud)) /* Another has populated it */
3904 pgd_populate(mm, pud, new);
3905 #endif /* __ARCH_HAS_4LEVEL_HACK */
3906 spin_unlock(&mm->page_table_lock);
3909 #endif /* __PAGETABLE_PMD_FOLDED */
3911 #if !defined(__HAVE_ARCH_GATE_AREA)
3913 #if defined(AT_SYSINFO_EHDR)
3914 static struct vm_area_struct gate_vma;
3916 static int __init gate_vma_init(void)
3918 gate_vma.vm_mm = NULL;
3919 gate_vma.vm_start = FIXADDR_USER_START;
3920 gate_vma.vm_end = FIXADDR_USER_END;
3921 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3922 gate_vma.vm_page_prot = __P101;
3926 __initcall(gate_vma_init);
3929 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3931 #ifdef AT_SYSINFO_EHDR
3938 int in_gate_area_no_mm(unsigned long addr)
3940 #ifdef AT_SYSINFO_EHDR
3941 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
3947 #endif /* __HAVE_ARCH_GATE_AREA */
3949 static int __follow_pte(struct mm_struct *mm, unsigned long address,
3950 pte_t **ptepp, spinlock_t **ptlp)
3957 pgd = pgd_offset(mm, address);
3958 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3961 pud = pud_offset(pgd, address);
3962 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3965 pmd = pmd_offset(pud, address);
3966 VM_BUG_ON(pmd_trans_huge(*pmd));
3967 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3970 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3974 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3977 if (!pte_present(*ptep))
3982 pte_unmap_unlock(ptep, *ptlp);
3987 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3988 pte_t **ptepp, spinlock_t **ptlp)
3992 /* (void) is needed to make gcc happy */
3993 (void) __cond_lock(*ptlp,
3994 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3999 * follow_pfn - look up PFN at a user virtual address
4000 * @vma: memory mapping
4001 * @address: user virtual address
4002 * @pfn: location to store found PFN
4004 * Only IO mappings and raw PFN mappings are allowed.
4006 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4008 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4015 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4018 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4021 *pfn = pte_pfn(*ptep);
4022 pte_unmap_unlock(ptep, ptl);
4025 EXPORT_SYMBOL(follow_pfn);
4027 #ifdef CONFIG_HAVE_IOREMAP_PROT
4028 int follow_phys(struct vm_area_struct *vma,
4029 unsigned long address, unsigned int flags,
4030 unsigned long *prot, resource_size_t *phys)
4036 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4039 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4043 if ((flags & FOLL_WRITE) && !pte_write(pte))
4046 *prot = pgprot_val(pte_pgprot(pte));
4047 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4051 pte_unmap_unlock(ptep, ptl);
4056 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4057 void *buf, int len, int write)
4059 resource_size_t phys_addr;
4060 unsigned long prot = 0;
4061 void __iomem *maddr;
4062 int offset = addr & (PAGE_SIZE-1);
4064 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4067 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4069 memcpy_toio(maddr + offset, buf, len);
4071 memcpy_fromio(buf, maddr + offset, len);
4079 * Access another process' address space as given in mm. If non-NULL, use the
4080 * given task for page fault accounting.
4082 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4083 unsigned long addr, void *buf, int len, int write)
4085 struct vm_area_struct *vma;
4086 void *old_buf = buf;
4088 down_read(&mm->mmap_sem);
4089 /* ignore errors, just check how much was successfully transferred */
4091 int bytes, ret, offset;
4093 struct page *page = NULL;
4095 ret = get_user_pages(tsk, mm, addr, 1,
4096 write, 1, &page, &vma);
4099 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4100 * we can access using slightly different code.
4102 #ifdef CONFIG_HAVE_IOREMAP_PROT
4103 vma = find_vma(mm, addr);
4104 if (!vma || vma->vm_start > addr)
4106 if (vma->vm_ops && vma->vm_ops->access)
4107 ret = vma->vm_ops->access(vma, addr, buf,
4115 offset = addr & (PAGE_SIZE-1);
4116 if (bytes > PAGE_SIZE-offset)
4117 bytes = PAGE_SIZE-offset;
4121 copy_to_user_page(vma, page, addr,
4122 maddr + offset, buf, bytes);
4123 set_page_dirty_lock(page);
4125 copy_from_user_page(vma, page, addr,
4126 buf, maddr + offset, bytes);
4129 page_cache_release(page);
4135 up_read(&mm->mmap_sem);
4137 return buf - old_buf;
4141 * access_remote_vm - access another process' address space
4142 * @mm: the mm_struct of the target address space
4143 * @addr: start address to access
4144 * @buf: source or destination buffer
4145 * @len: number of bytes to transfer
4146 * @write: whether the access is a write
4148 * The caller must hold a reference on @mm.
4150 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4151 void *buf, int len, int write)
4153 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4157 * Access another process' address space.
4158 * Source/target buffer must be kernel space,
4159 * Do not walk the page table directly, use get_user_pages
4161 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4162 void *buf, int len, int write)
4164 struct mm_struct *mm;
4167 mm = get_task_mm(tsk);
4171 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4178 * Print the name of a VMA.
4180 void print_vma_addr(char *prefix, unsigned long ip)
4182 struct mm_struct *mm = current->mm;
4183 struct vm_area_struct *vma;
4186 * Do not print if we are in atomic
4187 * contexts (in exception stacks, etc.):
4189 if (preempt_count())
4192 down_read(&mm->mmap_sem);
4193 vma = find_vma(mm, ip);
4194 if (vma && vma->vm_file) {
4195 struct file *f = vma->vm_file;
4196 char *buf = (char *)__get_free_page(GFP_KERNEL);
4200 p = d_path(&f->f_path, buf, PAGE_SIZE);
4203 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4205 vma->vm_end - vma->vm_start);
4206 free_page((unsigned long)buf);
4209 up_read(&mm->mmap_sem);
4212 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4213 void might_fault(void)
4216 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4217 * holding the mmap_sem, this is safe because kernel memory doesn't
4218 * get paged out, therefore we'll never actually fault, and the
4219 * below annotations will generate false positives.
4221 if (segment_eq(get_fs(), KERNEL_DS))
4225 * it would be nicer only to annotate paths which are not under
4226 * pagefault_disable, however that requires a larger audit and
4227 * providing helpers like get_user_atomic.
4232 __might_sleep(__FILE__, __LINE__, 0);
4235 might_lock_read(¤t->mm->mmap_sem);
4237 EXPORT_SYMBOL(might_fault);
4240 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4241 static void clear_gigantic_page(struct page *page,
4243 unsigned int pages_per_huge_page)
4246 struct page *p = page;
4249 for (i = 0; i < pages_per_huge_page;
4250 i++, p = mem_map_next(p, page, i)) {
4252 clear_user_highpage(p, addr + i * PAGE_SIZE);
4255 void clear_huge_page(struct page *page,
4256 unsigned long addr, unsigned int pages_per_huge_page)
4260 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4261 clear_gigantic_page(page, addr, pages_per_huge_page);
4266 for (i = 0; i < pages_per_huge_page; i++) {
4268 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4272 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4274 struct vm_area_struct *vma,
4275 unsigned int pages_per_huge_page)
4278 struct page *dst_base = dst;
4279 struct page *src_base = src;
4281 for (i = 0; i < pages_per_huge_page; ) {
4283 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4286 dst = mem_map_next(dst, dst_base, i);
4287 src = mem_map_next(src, src_base, i);
4291 void copy_user_huge_page(struct page *dst, struct page *src,
4292 unsigned long addr, struct vm_area_struct *vma,
4293 unsigned int pages_per_huge_page)
4297 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4298 copy_user_gigantic_page(dst, src, addr, vma,
4299 pages_per_huge_page);
4304 for (i = 0; i < pages_per_huge_page; i++) {
4306 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4309 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */