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/delay.h>
53 #include <linux/init.h>
54 #include <linux/writeback.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/kallsyms.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60 #include <linux/gfp.h>
61 #include <linux/migrate.h>
62 #include <linux/string.h>
65 #include <asm/pgalloc.h>
66 #include <asm/uaccess.h>
68 #include <asm/tlbflush.h>
69 #include <asm/pgtable.h>
73 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
74 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_nid.
77 #ifndef CONFIG_NEED_MULTIPLE_NODES
78 /* use the per-pgdat data instead for discontigmem - mbligh */
79 unsigned long max_mapnr;
82 EXPORT_SYMBOL(max_mapnr);
83 EXPORT_SYMBOL(mem_map);
86 unsigned long num_physpages;
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 EXPORT_SYMBOL(num_physpages);
97 EXPORT_SYMBOL(high_memory);
100 * Randomize the address space (stacks, mmaps, brk, etc.).
102 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
103 * as ancient (libc5 based) binaries can segfault. )
105 int randomize_va_space __read_mostly =
106 #ifdef CONFIG_COMPAT_BRK
112 static int __init disable_randmaps(char *s)
114 randomize_va_space = 0;
117 __setup("norandmaps", disable_randmaps);
119 unsigned long zero_pfn __read_mostly;
120 unsigned long highest_memmap_pfn __read_mostly;
122 EXPORT_SYMBOL(zero_pfn);
125 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
127 static int __init init_zero_pfn(void)
129 zero_pfn = page_to_pfn(ZERO_PAGE(0));
132 core_initcall(init_zero_pfn);
135 #if defined(SPLIT_RSS_COUNTING)
137 void sync_mm_rss(struct mm_struct *mm)
141 for (i = 0; i < NR_MM_COUNTERS; i++) {
142 if (current->rss_stat.count[i]) {
143 add_mm_counter(mm, i, current->rss_stat.count[i]);
144 current->rss_stat.count[i] = 0;
147 current->rss_stat.events = 0;
150 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
152 struct task_struct *task = current;
154 if (likely(task->mm == mm))
155 task->rss_stat.count[member] += val;
157 add_mm_counter(mm, member, val);
159 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
160 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
162 /* sync counter once per 64 page faults */
163 #define TASK_RSS_EVENTS_THRESH (64)
164 static void check_sync_rss_stat(struct task_struct *task)
166 if (unlikely(task != current))
168 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
169 sync_mm_rss(task->mm);
171 #else /* SPLIT_RSS_COUNTING */
173 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
174 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
176 static void check_sync_rss_stat(struct task_struct *task)
180 #endif /* SPLIT_RSS_COUNTING */
182 #ifdef HAVE_GENERIC_MMU_GATHER
184 static int tlb_next_batch(struct mmu_gather *tlb)
186 struct mmu_gather_batch *batch;
190 tlb->active = batch->next;
194 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
197 #ifndef CONFIG_SPRD_PAGERECORDER
198 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
200 batch = (void *)__get_free_pages_nopagedebug(GFP_NOWAIT | __GFP_NOWARN, 0);
208 batch->max = MAX_GATHER_BATCH;
210 tlb->active->next = batch;
217 * Called to initialize an (on-stack) mmu_gather structure for page-table
218 * tear-down from @mm. The @fullmm argument is used when @mm is without
219 * users and we're going to destroy the full address space (exit/execve).
221 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
225 /* Is it from 0 to ~0? */
226 tlb->fullmm = !(start | (end+1));
227 tlb->need_flush_all = 0;
231 tlb->local.next = NULL;
233 tlb->local.max = ARRAY_SIZE(tlb->__pages);
234 tlb->active = &tlb->local;
235 tlb->batch_count = 0;
237 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
242 void tlb_flush_mmu(struct mmu_gather *tlb)
244 struct mmu_gather_batch *batch;
246 if (!tlb->need_flush)
250 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
251 tlb_table_flush(tlb);
254 for (batch = &tlb->local; batch; batch = batch->next) {
255 free_pages_and_swap_cache(batch->pages, batch->nr);
258 tlb->active = &tlb->local;
262 * Called at the end of the shootdown operation to free up any resources
263 * that were required.
265 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
267 struct mmu_gather_batch *batch, *next;
271 /* keep the page table cache within bounds */
274 for (batch = tlb->local.next; batch; batch = next) {
276 free_pages((unsigned long)batch, 0);
278 tlb->local.next = NULL;
282 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
283 * handling the additional races in SMP caused by other CPUs caching valid
284 * mappings in their TLBs. Returns the number of free page slots left.
285 * When out of page slots we must call tlb_flush_mmu().
287 int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
289 struct mmu_gather_batch *batch;
291 VM_BUG_ON(!tlb->need_flush);
294 batch->pages[batch->nr++] = page;
295 if (batch->nr == batch->max) {
296 if (!tlb_next_batch(tlb))
300 VM_BUG_ON(batch->nr > batch->max);
302 return batch->max - batch->nr;
305 #endif /* HAVE_GENERIC_MMU_GATHER */
307 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
310 * See the comment near struct mmu_table_batch.
313 static void tlb_remove_table_smp_sync(void *arg)
315 /* Simply deliver the interrupt */
318 static void tlb_remove_table_one(void *table)
321 * This isn't an RCU grace period and hence the page-tables cannot be
322 * assumed to be actually RCU-freed.
324 * It is however sufficient for software page-table walkers that rely on
325 * IRQ disabling. See the comment near struct mmu_table_batch.
327 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
328 __tlb_remove_table(table);
331 static void tlb_remove_table_rcu(struct rcu_head *head)
333 struct mmu_table_batch *batch;
336 batch = container_of(head, struct mmu_table_batch, rcu);
338 for (i = 0; i < batch->nr; i++)
339 __tlb_remove_table(batch->tables[i]);
341 free_page((unsigned long)batch);
344 void tlb_table_flush(struct mmu_gather *tlb)
346 struct mmu_table_batch **batch = &tlb->batch;
349 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
354 void tlb_remove_table(struct mmu_gather *tlb, void *table)
356 struct mmu_table_batch **batch = &tlb->batch;
361 * When there's less then two users of this mm there cannot be a
362 * concurrent page-table walk.
364 if (atomic_read(&tlb->mm->mm_users) < 2) {
365 __tlb_remove_table(table);
369 if (*batch == NULL) {
370 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
371 if (*batch == NULL) {
372 tlb_remove_table_one(table);
377 (*batch)->tables[(*batch)->nr++] = table;
378 if ((*batch)->nr == MAX_TABLE_BATCH)
379 tlb_table_flush(tlb);
382 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385 * If a p?d_bad entry is found while walking page tables, report
386 * the error, before resetting entry to p?d_none. Usually (but
387 * very seldom) called out from the p?d_none_or_clear_bad macros.
390 void pgd_clear_bad(pgd_t *pgd)
396 void pud_clear_bad(pud_t *pud)
402 void pmd_clear_bad(pmd_t *pmd)
409 * Note: this doesn't free the actual pages themselves. That
410 * has been handled earlier when unmapping all the memory regions.
412 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
415 pgtable_t token = pmd_pgtable(*pmd);
417 pte_free_tlb(tlb, token, addr);
421 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
422 unsigned long addr, unsigned long end,
423 unsigned long floor, unsigned long ceiling)
430 pmd = pmd_offset(pud, addr);
432 next = pmd_addr_end(addr, end);
433 if (pmd_none_or_clear_bad(pmd))
435 free_pte_range(tlb, pmd, addr);
436 } while (pmd++, addr = next, addr != end);
446 if (end - 1 > ceiling - 1)
449 pmd = pmd_offset(pud, start);
451 pmd_free_tlb(tlb, pmd, start);
454 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
455 unsigned long addr, unsigned long end,
456 unsigned long floor, unsigned long ceiling)
463 pud = pud_offset(pgd, addr);
465 next = pud_addr_end(addr, end);
466 if (pud_none_or_clear_bad(pud))
468 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
469 } while (pud++, addr = next, addr != end);
475 ceiling &= PGDIR_MASK;
479 if (end - 1 > ceiling - 1)
482 pud = pud_offset(pgd, start);
484 pud_free_tlb(tlb, pud, start);
488 * This function frees user-level page tables of a process.
490 * Must be called with pagetable lock held.
492 void free_pgd_range(struct mmu_gather *tlb,
493 unsigned long addr, unsigned long end,
494 unsigned long floor, unsigned long ceiling)
500 * The next few lines have given us lots of grief...
502 * Why are we testing PMD* at this top level? Because often
503 * there will be no work to do at all, and we'd prefer not to
504 * go all the way down to the bottom just to discover that.
506 * Why all these "- 1"s? Because 0 represents both the bottom
507 * of the address space and the top of it (using -1 for the
508 * top wouldn't help much: the masks would do the wrong thing).
509 * The rule is that addr 0 and floor 0 refer to the bottom of
510 * the address space, but end 0 and ceiling 0 refer to the top
511 * Comparisons need to use "end - 1" and "ceiling - 1" (though
512 * that end 0 case should be mythical).
514 * Wherever addr is brought up or ceiling brought down, we must
515 * be careful to reject "the opposite 0" before it confuses the
516 * subsequent tests. But what about where end is brought down
517 * by PMD_SIZE below? no, end can't go down to 0 there.
519 * Whereas we round start (addr) and ceiling down, by different
520 * masks at different levels, in order to test whether a table
521 * now has no other vmas using it, so can be freed, we don't
522 * bother to round floor or end up - the tests don't need that.
536 if (end - 1 > ceiling - 1)
541 pgd = pgd_offset(tlb->mm, addr);
543 next = pgd_addr_end(addr, end);
544 if (pgd_none_or_clear_bad(pgd))
546 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
547 } while (pgd++, addr = next, addr != end);
550 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
551 unsigned long floor, unsigned long ceiling)
554 struct vm_area_struct *next = vma->vm_next;
555 unsigned long addr = vma->vm_start;
558 * Hide vma from rmap and truncate_pagecache before freeing
561 unlink_anon_vmas(vma);
562 unlink_file_vma(vma);
564 if (is_vm_hugetlb_page(vma)) {
565 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
566 floor, next? next->vm_start: ceiling);
569 * Optimization: gather nearby vmas into one call down
571 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
572 && !is_vm_hugetlb_page(next)) {
575 unlink_anon_vmas(vma);
576 unlink_file_vma(vma);
578 free_pgd_range(tlb, addr, vma->vm_end,
579 floor, next? next->vm_start: ceiling);
585 int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
586 pmd_t *pmd, unsigned long address)
588 pgtable_t new = pte_alloc_one(mm, address);
589 int wait_split_huge_page;
594 * Ensure all pte setup (eg. pte page lock and page clearing) are
595 * visible before the pte is made visible to other CPUs by being
596 * put into page tables.
598 * The other side of the story is the pointer chasing in the page
599 * table walking code (when walking the page table without locking;
600 * ie. most of the time). Fortunately, these data accesses consist
601 * of a chain of data-dependent loads, meaning most CPUs (alpha
602 * being the notable exception) will already guarantee loads are
603 * seen in-order. See the alpha page table accessors for the
604 * smp_read_barrier_depends() barriers in page table walking code.
606 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
608 spin_lock(&mm->page_table_lock);
609 wait_split_huge_page = 0;
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
612 pmd_populate(mm, pmd, new);
614 } else if (unlikely(pmd_trans_splitting(*pmd)))
615 wait_split_huge_page = 1;
616 spin_unlock(&mm->page_table_lock);
619 if (wait_split_huge_page)
620 wait_split_huge_page(vma->anon_vma, pmd);
624 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
626 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
630 smp_wmb(); /* See comment in __pte_alloc */
632 spin_lock(&init_mm.page_table_lock);
633 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
634 pmd_populate_kernel(&init_mm, pmd, new);
637 VM_BUG_ON(pmd_trans_splitting(*pmd));
638 spin_unlock(&init_mm.page_table_lock);
640 pte_free_kernel(&init_mm, new);
644 static inline void init_rss_vec(int *rss)
646 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
649 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
653 if (current->mm == mm)
655 for (i = 0; i < NR_MM_COUNTERS; i++)
657 add_mm_counter(mm, i, rss[i]);
661 * This function is called to print an error when a bad pte
662 * is found. For example, we might have a PFN-mapped pte in
663 * a region that doesn't allow it.
665 * The calling function must still handle the error.
667 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
668 pte_t pte, struct page *page)
670 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
671 pud_t *pud = pud_offset(pgd, addr);
672 pmd_t *pmd = pmd_offset(pud, addr);
673 struct address_space *mapping;
675 static unsigned long resume;
676 static unsigned long nr_shown;
677 static unsigned long nr_unshown;
680 * Allow a burst of 60 reports, then keep quiet for that minute;
681 * or allow a steady drip of one report per second.
683 if (nr_shown == 60) {
684 if (time_before(jiffies, resume)) {
690 "BUG: Bad page map: %lu messages suppressed\n",
697 resume = jiffies + 60 * HZ;
699 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
700 index = linear_page_index(vma, addr);
703 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
705 (long long)pte_val(pte), (long long)pmd_val(*pmd));
709 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
710 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
712 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
715 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
717 if (vma->vm_file && vma->vm_file->f_op)
718 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
719 vma->vm_file->f_op->mmap);
721 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
724 static inline bool is_cow_mapping(vm_flags_t flags)
726 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
730 * vm_normal_page -- This function gets the "struct page" associated with a pte.
732 * "Special" mappings do not wish to be associated with a "struct page" (either
733 * it doesn't exist, or it exists but they don't want to touch it). In this
734 * case, NULL is returned here. "Normal" mappings do have a struct page.
736 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
737 * pte bit, in which case this function is trivial. Secondly, an architecture
738 * may not have a spare pte bit, which requires a more complicated scheme,
741 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
742 * special mapping (even if there are underlying and valid "struct pages").
743 * COWed pages of a VM_PFNMAP are always normal.
745 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
746 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
747 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
748 * mapping will always honor the rule
750 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
752 * And for normal mappings this is false.
754 * This restricts such mappings to be a linear translation from virtual address
755 * to pfn. To get around this restriction, we allow arbitrary mappings so long
756 * as the vma is not a COW mapping; in that case, we know that all ptes are
757 * special (because none can have been COWed).
760 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
762 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
763 * page" backing, however the difference is that _all_ pages with a struct
764 * page (that is, those where pfn_valid is true) are refcounted and considered
765 * normal pages by the VM. The disadvantage is that pages are refcounted
766 * (which can be slower and simply not an option for some PFNMAP users). The
767 * advantage is that we don't have to follow the strict linearity rule of
768 * PFNMAP mappings in order to support COWable mappings.
771 #ifdef __HAVE_ARCH_PTE_SPECIAL
772 # define HAVE_PTE_SPECIAL 1
774 # define HAVE_PTE_SPECIAL 0
776 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
779 unsigned long pfn = pte_pfn(pte);
781 if (HAVE_PTE_SPECIAL) {
782 if (likely(!pte_special(pte)))
784 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
786 if (!is_zero_pfn(pfn))
787 print_bad_pte(vma, addr, pte, NULL);
791 /* !HAVE_PTE_SPECIAL case follows: */
793 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
794 if (vma->vm_flags & VM_MIXEDMAP) {
800 off = (addr - vma->vm_start) >> PAGE_SHIFT;
801 if (pfn == vma->vm_pgoff + off)
803 if (!is_cow_mapping(vma->vm_flags))
808 if (is_zero_pfn(pfn))
811 if (unlikely(pfn > highest_memmap_pfn)) {
812 print_bad_pte(vma, addr, pte, NULL);
817 * NOTE! We still have PageReserved() pages in the page tables.
818 * eg. VDSO mappings can cause them to exist.
821 return pfn_to_page(pfn);
825 * copy one vm_area from one task to the other. Assumes the page tables
826 * already present in the new task to be cleared in the whole range
827 * covered by this vma.
830 static inline unsigned long
831 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
832 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
833 unsigned long addr, int *rss)
835 unsigned long vm_flags = vma->vm_flags;
836 pte_t pte = *src_pte;
839 /* pte contains position in swap or file, so copy. */
840 if (unlikely(!pte_present(pte))) {
841 if (!pte_file(pte)) {
842 swp_entry_t entry = pte_to_swp_entry(pte);
844 if (likely(!non_swap_entry(entry))) {
845 if (swap_duplicate(entry) < 0)
848 /* make sure dst_mm is on swapoff's mmlist. */
849 if (unlikely(list_empty(&dst_mm->mmlist))) {
850 spin_lock(&mmlist_lock);
851 if (list_empty(&dst_mm->mmlist))
852 list_add(&dst_mm->mmlist,
854 spin_unlock(&mmlist_lock);
857 } else if (is_migration_entry(entry)) {
858 page = migration_entry_to_page(entry);
865 if (is_write_migration_entry(entry) &&
866 is_cow_mapping(vm_flags)) {
868 * COW mappings require pages in both
869 * parent and child to be set to read.
871 make_migration_entry_read(&entry);
872 pte = swp_entry_to_pte(entry);
873 set_pte_at(src_mm, addr, src_pte, pte);
881 * If it's a COW mapping, write protect it both
882 * in the parent and the child
884 if (is_cow_mapping(vm_flags)) {
885 ptep_set_wrprotect(src_mm, addr, src_pte);
886 pte = pte_wrprotect(pte);
890 * If it's a shared mapping, mark it clean in
893 if (vm_flags & VM_SHARED)
894 pte = pte_mkclean(pte);
895 pte = pte_mkold(pte);
897 page = vm_normal_page(vma, addr, pte);
908 set_pte_at(dst_mm, addr, dst_pte, pte);
912 int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
913 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
914 unsigned long addr, unsigned long end)
916 pte_t *orig_src_pte, *orig_dst_pte;
917 pte_t *src_pte, *dst_pte;
918 spinlock_t *src_ptl, *dst_ptl;
920 int rss[NR_MM_COUNTERS];
921 swp_entry_t entry = (swp_entry_t){0};
926 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
929 src_pte = pte_offset_map(src_pmd, addr);
930 src_ptl = pte_lockptr(src_mm, src_pmd);
931 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
932 orig_src_pte = src_pte;
933 orig_dst_pte = dst_pte;
934 arch_enter_lazy_mmu_mode();
938 * We are holding two locks at this point - either of them
939 * could generate latencies in another task on another CPU.
941 if (progress >= 32) {
943 if (need_resched() ||
944 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
947 if (pte_none(*src_pte)) {
951 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
956 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
958 arch_leave_lazy_mmu_mode();
959 spin_unlock(src_ptl);
960 pte_unmap(orig_src_pte);
961 add_mm_rss_vec(dst_mm, rss);
962 pte_unmap_unlock(orig_dst_pte, dst_ptl);
966 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
975 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
976 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
977 unsigned long addr, unsigned long end)
979 pmd_t *src_pmd, *dst_pmd;
982 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
985 src_pmd = pmd_offset(src_pud, addr);
987 next = pmd_addr_end(addr, end);
988 if (pmd_trans_huge(*src_pmd)) {
990 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
991 err = copy_huge_pmd(dst_mm, src_mm,
992 dst_pmd, src_pmd, addr, vma);
999 if (pmd_none_or_clear_bad(src_pmd))
1001 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1004 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1008 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1009 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1010 unsigned long addr, unsigned long end)
1012 pud_t *src_pud, *dst_pud;
1015 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1018 src_pud = pud_offset(src_pgd, addr);
1020 next = pud_addr_end(addr, end);
1021 if (pud_none_or_clear_bad(src_pud))
1023 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1026 } while (dst_pud++, src_pud++, addr = next, addr != end);
1030 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1031 struct vm_area_struct *vma)
1033 pgd_t *src_pgd, *dst_pgd;
1035 unsigned long addr = vma->vm_start;
1036 unsigned long end = vma->vm_end;
1037 unsigned long mmun_start; /* For mmu_notifiers */
1038 unsigned long mmun_end; /* For mmu_notifiers */
1043 * Don't copy ptes where a page fault will fill them correctly.
1044 * Fork becomes much lighter when there are big shared or private
1045 * readonly mappings. The tradeoff is that copy_page_range is more
1046 * efficient than faulting.
1048 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1049 VM_PFNMAP | VM_MIXEDMAP))) {
1054 if (is_vm_hugetlb_page(vma))
1055 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1057 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1059 * We do not free on error cases below as remove_vma
1060 * gets called on error from higher level routine
1062 ret = track_pfn_copy(vma);
1068 * We need to invalidate the secondary MMU mappings only when
1069 * there could be a permission downgrade on the ptes of the
1070 * parent mm. And a permission downgrade will only happen if
1071 * is_cow_mapping() returns true.
1073 is_cow = is_cow_mapping(vma->vm_flags);
1077 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1081 dst_pgd = pgd_offset(dst_mm, addr);
1082 src_pgd = pgd_offset(src_mm, addr);
1084 next = pgd_addr_end(addr, end);
1085 if (pgd_none_or_clear_bad(src_pgd))
1087 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1088 vma, addr, next))) {
1092 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1095 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1099 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1100 struct vm_area_struct *vma, pmd_t *pmd,
1101 unsigned long addr, unsigned long end,
1102 struct zap_details *details)
1104 struct mm_struct *mm = tlb->mm;
1105 int force_flush = 0;
1106 int rss[NR_MM_COUNTERS];
1113 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1115 arch_enter_lazy_mmu_mode();
1118 if (pte_none(ptent)) {
1122 if (pte_present(ptent)) {
1125 page = vm_normal_page(vma, addr, ptent);
1126 if (unlikely(details) && page) {
1128 * unmap_shared_mapping_pages() wants to
1129 * invalidate cache without truncating:
1130 * unmap shared but keep private pages.
1132 if (details->check_mapping &&
1133 details->check_mapping != page->mapping)
1136 * Each page->index must be checked when
1137 * invalidating or truncating nonlinear.
1139 if (details->nonlinear_vma &&
1140 (page->index < details->first_index ||
1141 page->index > details->last_index))
1144 ptent = ptep_get_and_clear_full(mm, addr, pte,
1146 tlb_remove_tlb_entry(tlb, pte, addr);
1147 if (unlikely(!page))
1149 if (unlikely(details) && details->nonlinear_vma
1150 && linear_page_index(details->nonlinear_vma,
1151 addr) != page->index)
1152 set_pte_at(mm, addr, pte,
1153 pgoff_to_pte(page->index));
1155 rss[MM_ANONPAGES]--;
1157 if (pte_dirty(ptent))
1158 set_page_dirty(page);
1159 if (pte_young(ptent) &&
1160 likely(!VM_SequentialReadHint(vma)))
1161 mark_page_accessed(page);
1162 rss[MM_FILEPAGES]--;
1164 page_remove_rmap(page);
1165 if (unlikely(page_mapcount(page) < 0))
1166 print_bad_pte(vma, addr, ptent, page);
1167 force_flush = !__tlb_remove_page(tlb, page);
1173 * If details->check_mapping, we leave swap entries;
1174 * if details->nonlinear_vma, we leave file entries.
1176 if (unlikely(details))
1178 if (pte_file(ptent)) {
1179 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1180 print_bad_pte(vma, addr, ptent, NULL);
1182 swp_entry_t entry = pte_to_swp_entry(ptent);
1184 if (!non_swap_entry(entry))
1186 else if (is_migration_entry(entry)) {
1189 page = migration_entry_to_page(entry);
1192 rss[MM_ANONPAGES]--;
1194 rss[MM_FILEPAGES]--;
1196 if (unlikely(!free_swap_and_cache(entry)))
1197 print_bad_pte(vma, addr, ptent, NULL);
1199 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1200 } while (pte++, addr += PAGE_SIZE, addr != end);
1202 add_mm_rss_vec(mm, rss);
1203 arch_leave_lazy_mmu_mode();
1204 pte_unmap_unlock(start_pte, ptl);
1207 * mmu_gather ran out of room to batch pages, we break out of
1208 * the PTE lock to avoid doing the potential expensive TLB invalidate
1209 * and page-free while holding it.
1212 unsigned long old_end;
1217 * Flush the TLB just for the previous segment,
1218 * then update the range to be the remaining
1236 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1237 struct vm_area_struct *vma, pud_t *pud,
1238 unsigned long addr, unsigned long end,
1239 struct zap_details *details)
1244 pmd = pmd_offset(pud, addr);
1246 next = pmd_addr_end(addr, end);
1247 if (pmd_trans_huge(*pmd)) {
1248 if (next - addr != HPAGE_PMD_SIZE) {
1249 #ifdef CONFIG_DEBUG_VM
1250 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1251 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1252 __func__, addr, end,
1258 split_huge_page_pmd(vma, addr, pmd);
1259 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1264 * Here there can be other concurrent MADV_DONTNEED or
1265 * trans huge page faults running, and if the pmd is
1266 * none or trans huge it can change under us. This is
1267 * because MADV_DONTNEED holds the mmap_sem in read
1270 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1272 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1275 } while (pmd++, addr = next, addr != end);
1280 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1281 struct vm_area_struct *vma, pgd_t *pgd,
1282 unsigned long addr, unsigned long end,
1283 struct zap_details *details)
1288 pud = pud_offset(pgd, addr);
1290 next = pud_addr_end(addr, end);
1291 if (pud_none_or_clear_bad(pud))
1293 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1294 } while (pud++, addr = next, addr != end);
1299 static void unmap_page_range(struct mmu_gather *tlb,
1300 struct vm_area_struct *vma,
1301 unsigned long addr, unsigned long end,
1302 struct zap_details *details)
1307 if (details && !details->check_mapping && !details->nonlinear_vma)
1310 BUG_ON(addr >= end);
1311 mem_cgroup_uncharge_start();
1312 tlb_start_vma(tlb, vma);
1313 pgd = pgd_offset(vma->vm_mm, addr);
1315 next = pgd_addr_end(addr, end);
1316 if (pgd_none_or_clear_bad(pgd))
1318 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1319 } while (pgd++, addr = next, addr != end);
1320 tlb_end_vma(tlb, vma);
1321 mem_cgroup_uncharge_end();
1325 static void unmap_single_vma(struct mmu_gather *tlb,
1326 struct vm_area_struct *vma, unsigned long start_addr,
1327 unsigned long end_addr,
1328 struct zap_details *details)
1330 unsigned long start = max(vma->vm_start, start_addr);
1333 if (start >= vma->vm_end)
1335 end = min(vma->vm_end, end_addr);
1336 if (end <= vma->vm_start)
1340 uprobe_munmap(vma, start, end);
1342 if (unlikely(vma->vm_flags & VM_PFNMAP))
1343 untrack_pfn(vma, 0, 0);
1346 if (unlikely(is_vm_hugetlb_page(vma))) {
1348 * It is undesirable to test vma->vm_file as it
1349 * should be non-null for valid hugetlb area.
1350 * However, vm_file will be NULL in the error
1351 * cleanup path of do_mmap_pgoff. When
1352 * hugetlbfs ->mmap method fails,
1353 * do_mmap_pgoff() nullifies vma->vm_file
1354 * before calling this function to clean up.
1355 * Since no pte has actually been setup, it is
1356 * safe to do nothing in this case.
1359 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1360 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1361 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1364 unmap_page_range(tlb, vma, start, end, details);
1369 * unmap_vmas - unmap a range of memory covered by a list of vma's
1370 * @tlb: address of the caller's struct mmu_gather
1371 * @vma: the starting vma
1372 * @start_addr: virtual address at which to start unmapping
1373 * @end_addr: virtual address at which to end unmapping
1375 * Unmap all pages in the vma list.
1377 * Only addresses between `start' and `end' will be unmapped.
1379 * The VMA list must be sorted in ascending virtual address order.
1381 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1382 * range after unmap_vmas() returns. So the only responsibility here is to
1383 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1384 * drops the lock and schedules.
1386 void unmap_vmas(struct mmu_gather *tlb,
1387 struct vm_area_struct *vma, unsigned long start_addr,
1388 unsigned long end_addr)
1390 struct mm_struct *mm = vma->vm_mm;
1392 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1393 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1394 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1395 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1399 * zap_page_range - remove user pages in a given range
1400 * @vma: vm_area_struct holding the applicable pages
1401 * @start: starting address of pages to zap
1402 * @size: number of bytes to zap
1403 * @details: details of nonlinear truncation or shared cache invalidation
1405 * Caller must protect the VMA list
1407 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1408 unsigned long size, struct zap_details *details)
1410 struct mm_struct *mm = vma->vm_mm;
1411 struct mmu_gather tlb;
1412 unsigned long end = start + size;
1415 tlb_gather_mmu(&tlb, mm, start, end);
1416 update_hiwater_rss(mm);
1417 mmu_notifier_invalidate_range_start(mm, start, end);
1418 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1419 unmap_single_vma(&tlb, vma, start, end, details);
1420 mmu_notifier_invalidate_range_end(mm, start, end);
1421 tlb_finish_mmu(&tlb, start, end);
1425 * zap_page_range_single - remove user pages in a given range
1426 * @vma: vm_area_struct holding the applicable pages
1427 * @address: starting address of pages to zap
1428 * @size: number of bytes to zap
1429 * @details: details of nonlinear truncation or shared cache invalidation
1431 * The range must fit into one VMA.
1433 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1434 unsigned long size, struct zap_details *details)
1436 struct mm_struct *mm = vma->vm_mm;
1437 struct mmu_gather tlb;
1438 unsigned long end = address + size;
1441 tlb_gather_mmu(&tlb, mm, address, end);
1442 update_hiwater_rss(mm);
1443 mmu_notifier_invalidate_range_start(mm, address, end);
1444 unmap_single_vma(&tlb, vma, address, end, details);
1445 mmu_notifier_invalidate_range_end(mm, address, end);
1446 tlb_finish_mmu(&tlb, address, end);
1450 * zap_vma_ptes - remove ptes mapping the vma
1451 * @vma: vm_area_struct holding ptes to be zapped
1452 * @address: starting address of pages to zap
1453 * @size: number of bytes to zap
1455 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1457 * The entire address range must be fully contained within the vma.
1459 * Returns 0 if successful.
1461 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1464 if (address < vma->vm_start || address + size > vma->vm_end ||
1465 !(vma->vm_flags & VM_PFNMAP))
1467 zap_page_range_single(vma, address, size, NULL);
1470 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1473 * FOLL_FORCE can write to even unwritable pte's, but only
1474 * after we've gone through a COW cycle and they are dirty.
1476 static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
1478 return pte_write(pte) ||
1479 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
1483 * follow_page_mask - look up a page descriptor from a user-virtual address
1484 * @vma: vm_area_struct mapping @address
1485 * @address: virtual address to look up
1486 * @flags: flags modifying lookup behaviour
1487 * @page_mask: on output, *page_mask is set according to the size of the page
1489 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1491 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1492 * an error pointer if there is a mapping to something not represented
1493 * by a page descriptor (see also vm_normal_page()).
1495 struct page *follow_page_mask(struct vm_area_struct *vma,
1496 unsigned long address, unsigned int flags,
1497 unsigned int *page_mask)
1505 struct mm_struct *mm = vma->vm_mm;
1509 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1510 if (!IS_ERR(page)) {
1511 BUG_ON(flags & FOLL_GET);
1516 pgd = pgd_offset(mm, address);
1517 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1520 pud = pud_offset(pgd, address);
1523 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1524 BUG_ON(flags & FOLL_GET);
1525 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1528 if (unlikely(pud_bad(*pud)))
1531 pmd = pmd_offset(pud, address);
1534 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1535 BUG_ON(flags & FOLL_GET);
1536 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1539 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1541 if (pmd_trans_huge(*pmd)) {
1542 if (flags & FOLL_SPLIT) {
1543 split_huge_page_pmd(vma, address, pmd);
1544 goto split_fallthrough;
1546 spin_lock(&mm->page_table_lock);
1547 if (likely(pmd_trans_huge(*pmd))) {
1548 if (unlikely(pmd_trans_splitting(*pmd))) {
1549 spin_unlock(&mm->page_table_lock);
1550 wait_split_huge_page(vma->anon_vma, pmd);
1552 page = follow_trans_huge_pmd(vma, address,
1554 spin_unlock(&mm->page_table_lock);
1555 *page_mask = HPAGE_PMD_NR - 1;
1559 spin_unlock(&mm->page_table_lock);
1563 if (unlikely(pmd_bad(*pmd)))
1566 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1569 if (!pte_present(pte)) {
1572 * KSM's break_ksm() relies upon recognizing a ksm page
1573 * even while it is being migrated, so for that case we
1574 * need migration_entry_wait().
1576 if (likely(!(flags & FOLL_MIGRATION)))
1578 if (pte_none(pte) || pte_file(pte))
1580 entry = pte_to_swp_entry(pte);
1581 if (!is_migration_entry(entry))
1583 pte_unmap_unlock(ptep, ptl);
1584 migration_entry_wait(mm, pmd, address);
1585 goto split_fallthrough;
1587 if ((flags & FOLL_NUMA) && pte_numa(pte))
1589 if ((flags & FOLL_WRITE) && !can_follow_write_pte(pte, flags))
1592 page = vm_normal_page(vma, address, pte);
1593 if (unlikely(!page)) {
1594 if ((flags & FOLL_DUMP) ||
1595 !is_zero_pfn(pte_pfn(pte)))
1597 page = pte_page(pte);
1600 if (flags & FOLL_GET)
1601 get_page_foll(page);
1602 if (flags & FOLL_TOUCH) {
1603 if ((flags & FOLL_WRITE) &&
1604 !pte_dirty(pte) && !PageDirty(page))
1605 set_page_dirty(page);
1607 * pte_mkyoung() would be more correct here, but atomic care
1608 * is needed to avoid losing the dirty bit: it is easier to use
1609 * mark_page_accessed().
1611 mark_page_accessed(page);
1613 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1615 * The preliminary mapping check is mainly to avoid the
1616 * pointless overhead of lock_page on the ZERO_PAGE
1617 * which might bounce very badly if there is contention.
1619 * If the page is already locked, we don't need to
1620 * handle it now - vmscan will handle it later if and
1621 * when it attempts to reclaim the page.
1623 if (page->mapping && trylock_page(page)) {
1624 lru_add_drain(); /* push cached pages to LRU */
1626 * Because we lock page here, and migration is
1627 * blocked by the pte's page reference, and we
1628 * know the page is still mapped, we don't even
1629 * need to check for file-cache page truncation.
1631 mlock_vma_page(page);
1636 pte_unmap_unlock(ptep, ptl);
1641 pte_unmap_unlock(ptep, ptl);
1642 return ERR_PTR(-EFAULT);
1645 pte_unmap_unlock(ptep, ptl);
1651 * When core dumping an enormous anonymous area that nobody
1652 * has touched so far, we don't want to allocate unnecessary pages or
1653 * page tables. Return error instead of NULL to skip handle_mm_fault,
1654 * then get_dump_page() will return NULL to leave a hole in the dump.
1655 * But we can only make this optimization where a hole would surely
1656 * be zero-filled if handle_mm_fault() actually did handle it.
1658 if ((flags & FOLL_DUMP) &&
1659 (!vma->vm_ops || !vma->vm_ops->fault))
1660 return ERR_PTR(-EFAULT);
1664 static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1666 return stack_guard_page_start(vma, addr) ||
1667 stack_guard_page_end(vma, addr+PAGE_SIZE);
1671 * __get_user_pages() - pin user pages in memory
1672 * @tsk: task_struct of target task
1673 * @mm: mm_struct of target mm
1674 * @start: starting user address
1675 * @nr_pages: number of pages from start to pin
1676 * @gup_flags: flags modifying pin behaviour
1677 * @pages: array that receives pointers to the pages pinned.
1678 * Should be at least nr_pages long. Or NULL, if caller
1679 * only intends to ensure the pages are faulted in.
1680 * @vmas: array of pointers to vmas corresponding to each page.
1681 * Or NULL if the caller does not require them.
1682 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1684 * Returns number of pages pinned. This may be fewer than the number
1685 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1686 * were pinned, returns -errno. Each page returned must be released
1687 * with a put_page() call when it is finished with. vmas will only
1688 * remain valid while mmap_sem is held.
1690 * Must be called with mmap_sem held for read or write.
1692 * __get_user_pages walks a process's page tables and takes a reference to
1693 * each struct page that each user address corresponds to at a given
1694 * instant. That is, it takes the page that would be accessed if a user
1695 * thread accesses the given user virtual address at that instant.
1697 * This does not guarantee that the page exists in the user mappings when
1698 * __get_user_pages returns, and there may even be a completely different
1699 * page there in some cases (eg. if mmapped pagecache has been invalidated
1700 * and subsequently re faulted). However it does guarantee that the page
1701 * won't be freed completely. And mostly callers simply care that the page
1702 * contains data that was valid *at some point in time*. Typically, an IO
1703 * or similar operation cannot guarantee anything stronger anyway because
1704 * locks can't be held over the syscall boundary.
1706 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1707 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1708 * appropriate) must be called after the page is finished with, and
1709 * before put_page is called.
1711 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1712 * or mmap_sem contention, and if waiting is needed to pin all pages,
1713 * *@nonblocking will be set to 0.
1715 * In most cases, get_user_pages or get_user_pages_fast should be used
1716 * instead of __get_user_pages. __get_user_pages should be used only if
1717 * you need some special @gup_flags.
1719 long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1720 unsigned long start, unsigned long nr_pages,
1721 unsigned int gup_flags, struct page **pages,
1722 struct vm_area_struct **vmas, int *nonblocking)
1725 unsigned long vm_flags;
1726 unsigned int page_mask;
1731 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1734 * Require read or write permissions.
1735 * If FOLL_FORCE is set, we only require the "MAY" flags.
1737 vm_flags = (gup_flags & FOLL_WRITE) ?
1738 (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
1739 vm_flags &= (gup_flags & FOLL_FORCE) ?
1740 (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1743 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1744 * would be called on PROT_NONE ranges. We must never invoke
1745 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1746 * page faults would unprotect the PROT_NONE ranges if
1747 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1748 * bitflag. So to avoid that, don't set FOLL_NUMA if
1749 * FOLL_FORCE is set.
1751 if (!(gup_flags & FOLL_FORCE))
1752 gup_flags |= FOLL_NUMA;
1757 struct vm_area_struct *vma;
1759 vma = find_extend_vma(mm, start);
1760 if (!vma && in_gate_area(mm, start)) {
1761 unsigned long pg = start & PAGE_MASK;
1767 /* user gate pages are read-only */
1768 if (gup_flags & FOLL_WRITE)
1769 return i ? : -EFAULT;
1771 pgd = pgd_offset_k(pg);
1773 pgd = pgd_offset_gate(mm, pg);
1774 BUG_ON(pgd_none(*pgd));
1775 pud = pud_offset(pgd, pg);
1776 BUG_ON(pud_none(*pud));
1777 pmd = pmd_offset(pud, pg);
1779 return i ? : -EFAULT;
1780 VM_BUG_ON(pmd_trans_huge(*pmd));
1781 pte = pte_offset_map(pmd, pg);
1782 if (pte_none(*pte)) {
1784 return i ? : -EFAULT;
1786 vma = get_gate_vma(mm);
1790 page = vm_normal_page(vma, start, *pte);
1792 if (!(gup_flags & FOLL_DUMP) &&
1793 is_zero_pfn(pte_pfn(*pte)))
1794 page = pte_page(*pte);
1797 return i ? : -EFAULT;
1809 (vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
1810 !(vm_flags & vma->vm_flags))
1811 return i ? : -EFAULT;
1813 if (is_vm_hugetlb_page(vma)) {
1814 i = follow_hugetlb_page(mm, vma, pages, vmas,
1815 &start, &nr_pages, i, gup_flags);
1821 unsigned int foll_flags = gup_flags;
1822 unsigned int page_increm;
1825 * If we have a pending SIGKILL, don't keep faulting
1826 * pages and potentially allocating memory.
1828 if (unlikely(fatal_signal_pending(current)))
1829 return i ? i : -ERESTARTSYS;
1832 while (!(page = follow_page_mask(vma, start,
1833 foll_flags, &page_mask))) {
1835 unsigned int fault_flags = 0;
1837 /* For mlock, just skip the stack guard page. */
1838 if (foll_flags & FOLL_MLOCK) {
1839 if (stack_guard_page(vma, start))
1842 if (foll_flags & FOLL_WRITE)
1843 fault_flags |= FAULT_FLAG_WRITE;
1845 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1846 if (foll_flags & FOLL_NOWAIT)
1847 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1849 ret = handle_mm_fault(mm, vma, start,
1852 if (ret & VM_FAULT_ERROR) {
1853 if (ret & VM_FAULT_OOM)
1854 return i ? i : -ENOMEM;
1855 if (ret & (VM_FAULT_HWPOISON |
1856 VM_FAULT_HWPOISON_LARGE)) {
1859 else if (gup_flags & FOLL_HWPOISON)
1864 if (ret & VM_FAULT_SIGBUS)
1865 return i ? i : -EFAULT;
1870 if (ret & VM_FAULT_MAJOR)
1876 if (ret & VM_FAULT_RETRY) {
1883 * The VM_FAULT_WRITE bit tells us that
1884 * do_wp_page has broken COW when necessary,
1885 * even if maybe_mkwrite decided not to set
1886 * pte_write. We can thus safely do subsequent
1887 * page lookups as if they were reads. But only
1888 * do so when looping for pte_write is futile:
1889 * in some cases userspace may also be wanting
1890 * to write to the gotten user page, which a
1891 * read fault here might prevent (a readonly
1892 * page might get reCOWed by userspace write).
1894 if ((ret & VM_FAULT_WRITE) &&
1895 !(vma->vm_flags & VM_WRITE))
1896 foll_flags |= FOLL_COW;
1901 return i ? i : PTR_ERR(page);
1905 flush_anon_page(vma, page, start);
1906 flush_dcache_page(page);
1914 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1915 if (page_increm > nr_pages)
1916 page_increm = nr_pages;
1918 start += page_increm * PAGE_SIZE;
1919 nr_pages -= page_increm;
1920 } while (nr_pages && start < vma->vm_end);
1924 EXPORT_SYMBOL(__get_user_pages);
1927 * fixup_user_fault() - manually resolve a user page fault
1928 * @tsk: the task_struct to use for page fault accounting, or
1929 * NULL if faults are not to be recorded.
1930 * @mm: mm_struct of target mm
1931 * @address: user address
1932 * @fault_flags:flags to pass down to handle_mm_fault()
1934 * This is meant to be called in the specific scenario where for locking reasons
1935 * we try to access user memory in atomic context (within a pagefault_disable()
1936 * section), this returns -EFAULT, and we want to resolve the user fault before
1939 * Typically this is meant to be used by the futex code.
1941 * The main difference with get_user_pages() is that this function will
1942 * unconditionally call handle_mm_fault() which will in turn perform all the
1943 * necessary SW fixup of the dirty and young bits in the PTE, while
1944 * handle_mm_fault() only guarantees to update these in the struct page.
1946 * This is important for some architectures where those bits also gate the
1947 * access permission to the page because they are maintained in software. On
1948 * such architectures, gup() will not be enough to make a subsequent access
1951 * This should be called with the mm_sem held for read.
1953 int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1954 unsigned long address, unsigned int fault_flags)
1956 struct vm_area_struct *vma;
1957 vm_flags_t vm_flags;
1960 vma = find_extend_vma(mm, address);
1961 if (!vma || address < vma->vm_start)
1964 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1965 if (!(vm_flags & vma->vm_flags))
1968 ret = handle_mm_fault(mm, vma, address, fault_flags);
1969 if (ret & VM_FAULT_ERROR) {
1970 if (ret & VM_FAULT_OOM)
1972 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1974 if (ret & VM_FAULT_SIGBUS)
1979 if (ret & VM_FAULT_MAJOR)
1988 * get_user_pages() - pin user pages in memory
1989 * @tsk: the task_struct to use for page fault accounting, or
1990 * NULL if faults are not to be recorded.
1991 * @mm: mm_struct of target mm
1992 * @start: starting user address
1993 * @nr_pages: number of pages from start to pin
1994 * @write: whether pages will be written to by the caller
1995 * @force: whether to force write access even if user mapping is
1996 * readonly. This will result in the page being COWed even
1997 * in MAP_SHARED mappings. You do not want this.
1998 * @pages: array that receives pointers to the pages pinned.
1999 * Should be at least nr_pages long. Or NULL, if caller
2000 * only intends to ensure the pages are faulted in.
2001 * @vmas: array of pointers to vmas corresponding to each page.
2002 * Or NULL if the caller does not require them.
2004 * Returns number of pages pinned. This may be fewer than the number
2005 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2006 * were pinned, returns -errno. Each page returned must be released
2007 * with a put_page() call when it is finished with. vmas will only
2008 * remain valid while mmap_sem is held.
2010 * Must be called with mmap_sem held for read or write.
2012 * get_user_pages walks a process's page tables and takes a reference to
2013 * each struct page that each user address corresponds to at a given
2014 * instant. That is, it takes the page that would be accessed if a user
2015 * thread accesses the given user virtual address at that instant.
2017 * This does not guarantee that the page exists in the user mappings when
2018 * get_user_pages returns, and there may even be a completely different
2019 * page there in some cases (eg. if mmapped pagecache has been invalidated
2020 * and subsequently re faulted). However it does guarantee that the page
2021 * won't be freed completely. And mostly callers simply care that the page
2022 * contains data that was valid *at some point in time*. Typically, an IO
2023 * or similar operation cannot guarantee anything stronger anyway because
2024 * locks can't be held over the syscall boundary.
2026 * If write=0, the page must not be written to. If the page is written to,
2027 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2028 * after the page is finished with, and before put_page is called.
2030 * get_user_pages is typically used for fewer-copy IO operations, to get a
2031 * handle on the memory by some means other than accesses via the user virtual
2032 * addresses. The pages may be submitted for DMA to devices or accessed via
2033 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2034 * use the correct cache flushing APIs.
2036 * See also get_user_pages_fast, for performance critical applications.
2038 long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2039 unsigned long start, unsigned long nr_pages, int write,
2040 int force, struct page **pages, struct vm_area_struct **vmas)
2042 int flags = FOLL_TOUCH;
2047 flags |= FOLL_WRITE;
2049 flags |= FOLL_FORCE;
2051 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2054 EXPORT_SYMBOL(get_user_pages);
2057 * get_dump_page() - pin user page in memory while writing it to core dump
2058 * @addr: user address
2060 * Returns struct page pointer of user page pinned for dump,
2061 * to be freed afterwards by page_cache_release() or put_page().
2063 * Returns NULL on any kind of failure - a hole must then be inserted into
2064 * the corefile, to preserve alignment with its headers; and also returns
2065 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2066 * allowing a hole to be left in the corefile to save diskspace.
2068 * Called without mmap_sem, but after all other threads have been killed.
2070 #ifdef CONFIG_ELF_CORE
2071 struct page *get_dump_page(unsigned long addr)
2073 struct vm_area_struct *vma;
2076 if (__get_user_pages(current, current->mm, addr, 1,
2077 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2080 flush_cache_page(vma, addr, page_to_pfn(page));
2083 #endif /* CONFIG_ELF_CORE */
2085 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2088 pgd_t * pgd = pgd_offset(mm, addr);
2089 pud_t * pud = pud_alloc(mm, pgd, addr);
2091 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2093 VM_BUG_ON(pmd_trans_huge(*pmd));
2094 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2101 * This is the old fallback for page remapping.
2103 * For historical reasons, it only allows reserved pages. Only
2104 * old drivers should use this, and they needed to mark their
2105 * pages reserved for the old functions anyway.
2107 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2108 struct page *page, pgprot_t prot)
2110 struct mm_struct *mm = vma->vm_mm;
2119 flush_dcache_page(page);
2120 pte = get_locked_pte(mm, addr, &ptl);
2124 if (!pte_none(*pte))
2127 /* Ok, finally just insert the thing.. */
2129 inc_mm_counter_fast(mm, MM_FILEPAGES);
2130 page_add_file_rmap(page);
2131 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2134 pte_unmap_unlock(pte, ptl);
2137 pte_unmap_unlock(pte, ptl);
2143 * vm_insert_page - insert single page into user vma
2144 * @vma: user vma to map to
2145 * @addr: target user address of this page
2146 * @page: source kernel page
2148 * This allows drivers to insert individual pages they've allocated
2151 * The page has to be a nice clean _individual_ kernel allocation.
2152 * If you allocate a compound page, you need to have marked it as
2153 * such (__GFP_COMP), or manually just split the page up yourself
2154 * (see split_page()).
2156 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2157 * took an arbitrary page protection parameter. This doesn't allow
2158 * that. Your vma protection will have to be set up correctly, which
2159 * means that if you want a shared writable mapping, you'd better
2160 * ask for a shared writable mapping!
2162 * The page does not need to be reserved.
2164 * Usually this function is called from f_op->mmap() handler
2165 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2166 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2167 * function from other places, for example from page-fault handler.
2169 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2172 if (addr < vma->vm_start || addr >= vma->vm_end)
2174 if (!page_count(page))
2176 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2177 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2178 BUG_ON(vma->vm_flags & VM_PFNMAP);
2179 vma->vm_flags |= VM_MIXEDMAP;
2181 return insert_page(vma, addr, page, vma->vm_page_prot);
2183 EXPORT_SYMBOL(vm_insert_page);
2185 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2186 unsigned long pfn, pgprot_t prot)
2188 struct mm_struct *mm = vma->vm_mm;
2194 pte = get_locked_pte(mm, addr, &ptl);
2198 if (!pte_none(*pte))
2201 /* Ok, finally just insert the thing.. */
2202 entry = pte_mkspecial(pfn_pte(pfn, prot));
2203 set_pte_at(mm, addr, pte, entry);
2204 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2208 pte_unmap_unlock(pte, ptl);
2214 * vm_insert_pfn - insert single pfn into user vma
2215 * @vma: user vma to map to
2216 * @addr: target user address of this page
2217 * @pfn: source kernel pfn
2219 * Similar to vm_insert_page, this allows drivers to insert individual pages
2220 * they've allocated into a user vma. Same comments apply.
2222 * This function should only be called from a vm_ops->fault handler, and
2223 * in that case the handler should return NULL.
2225 * vma cannot be a COW mapping.
2227 * As this is called only for pages that do not currently exist, we
2228 * do not need to flush old virtual caches or the TLB.
2230 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2234 pgprot_t pgprot = vma->vm_page_prot;
2236 * Technically, architectures with pte_special can avoid all these
2237 * restrictions (same for remap_pfn_range). However we would like
2238 * consistency in testing and feature parity among all, so we should
2239 * try to keep these invariants in place for everybody.
2241 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2242 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2243 (VM_PFNMAP|VM_MIXEDMAP));
2244 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2245 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2247 if (addr < vma->vm_start || addr >= vma->vm_end)
2249 if (track_pfn_insert(vma, &pgprot, pfn))
2252 ret = insert_pfn(vma, addr, pfn, pgprot);
2256 EXPORT_SYMBOL(vm_insert_pfn);
2258 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2261 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2263 if (addr < vma->vm_start || addr >= vma->vm_end)
2267 * If we don't have pte special, then we have to use the pfn_valid()
2268 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2269 * refcount the page if pfn_valid is true (hence insert_page rather
2270 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2271 * without pte special, it would there be refcounted as a normal page.
2273 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2276 page = pfn_to_page(pfn);
2277 return insert_page(vma, addr, page, vma->vm_page_prot);
2279 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2281 EXPORT_SYMBOL(vm_insert_mixed);
2284 * maps a range of physical memory into the requested pages. the old
2285 * mappings are removed. any references to nonexistent pages results
2286 * in null mappings (currently treated as "copy-on-access")
2288 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2289 unsigned long addr, unsigned long end,
2290 unsigned long pfn, pgprot_t prot)
2295 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2298 arch_enter_lazy_mmu_mode();
2300 BUG_ON(!pte_none(*pte));
2301 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2303 } while (pte++, addr += PAGE_SIZE, addr != end);
2304 arch_leave_lazy_mmu_mode();
2305 pte_unmap_unlock(pte - 1, ptl);
2309 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2310 unsigned long addr, unsigned long end,
2311 unsigned long pfn, pgprot_t prot)
2316 pfn -= addr >> PAGE_SHIFT;
2317 pmd = pmd_alloc(mm, pud, addr);
2320 VM_BUG_ON(pmd_trans_huge(*pmd));
2322 next = pmd_addr_end(addr, end);
2323 if (remap_pte_range(mm, pmd, addr, next,
2324 pfn + (addr >> PAGE_SHIFT), prot))
2326 } while (pmd++, addr = next, addr != end);
2330 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2331 unsigned long addr, unsigned long end,
2332 unsigned long pfn, pgprot_t prot)
2337 pfn -= addr >> PAGE_SHIFT;
2338 pud = pud_alloc(mm, pgd, addr);
2342 next = pud_addr_end(addr, end);
2343 if (remap_pmd_range(mm, pud, addr, next,
2344 pfn + (addr >> PAGE_SHIFT), prot))
2346 } while (pud++, addr = next, addr != end);
2351 * remap_pfn_range - remap kernel memory to userspace
2352 * @vma: user vma to map to
2353 * @addr: target user address to start at
2354 * @pfn: physical address of kernel memory
2355 * @size: size of map area
2356 * @prot: page protection flags for this mapping
2358 * Note: this is only safe if the mm semaphore is held when called.
2360 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2361 unsigned long pfn, unsigned long size, pgprot_t prot)
2365 unsigned long end = addr + PAGE_ALIGN(size);
2366 struct mm_struct *mm = vma->vm_mm;
2370 * Physically remapped pages are special. Tell the
2371 * rest of the world about it:
2372 * VM_IO tells people not to look at these pages
2373 * (accesses can have side effects).
2374 * VM_PFNMAP tells the core MM that the base pages are just
2375 * raw PFN mappings, and do not have a "struct page" associated
2378 * Disable vma merging and expanding with mremap().
2380 * Omit vma from core dump, even when VM_IO turned off.
2382 * There's a horrible special case to handle copy-on-write
2383 * behaviour that some programs depend on. We mark the "original"
2384 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2385 * See vm_normal_page() for details.
2387 if (is_cow_mapping(vma->vm_flags)) {
2388 if (addr != vma->vm_start || end != vma->vm_end)
2390 vma->vm_pgoff = pfn;
2393 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2397 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2399 BUG_ON(addr >= end);
2400 pfn -= addr >> PAGE_SHIFT;
2401 pgd = pgd_offset(mm, addr);
2402 flush_cache_range(vma, addr, end);
2404 next = pgd_addr_end(addr, end);
2405 err = remap_pud_range(mm, pgd, addr, next,
2406 pfn + (addr >> PAGE_SHIFT), prot);
2409 } while (pgd++, addr = next, addr != end);
2412 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2416 EXPORT_SYMBOL(remap_pfn_range);
2419 * vm_iomap_memory - remap memory to userspace
2420 * @vma: user vma to map to
2421 * @start: start of area
2422 * @len: size of area
2424 * This is a simplified io_remap_pfn_range() for common driver use. The
2425 * driver just needs to give us the physical memory range to be mapped,
2426 * we'll figure out the rest from the vma information.
2428 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2429 * whatever write-combining details or similar.
2431 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2433 unsigned long vm_len, pfn, pages;
2435 /* Check that the physical memory area passed in looks valid */
2436 if (start + len < start)
2439 * You *really* shouldn't map things that aren't page-aligned,
2440 * but we've historically allowed it because IO memory might
2441 * just have smaller alignment.
2443 len += start & ~PAGE_MASK;
2444 pfn = start >> PAGE_SHIFT;
2445 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2446 if (pfn + pages < pfn)
2449 /* We start the mapping 'vm_pgoff' pages into the area */
2450 if (vma->vm_pgoff > pages)
2452 pfn += vma->vm_pgoff;
2453 pages -= vma->vm_pgoff;
2455 /* Can we fit all of the mapping? */
2456 vm_len = vma->vm_end - vma->vm_start;
2457 if (vm_len >> PAGE_SHIFT > pages)
2460 /* Ok, let it rip */
2461 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2463 EXPORT_SYMBOL(vm_iomap_memory);
2465 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2466 unsigned long addr, unsigned long end,
2467 pte_fn_t fn, void *data)
2472 spinlock_t *uninitialized_var(ptl);
2474 pte = (mm == &init_mm) ?
2475 pte_alloc_kernel(pmd, addr) :
2476 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2480 BUG_ON(pmd_huge(*pmd));
2482 arch_enter_lazy_mmu_mode();
2484 token = pmd_pgtable(*pmd);
2487 err = fn(pte++, token, addr, data);
2490 } while (addr += PAGE_SIZE, addr != end);
2492 arch_leave_lazy_mmu_mode();
2495 pte_unmap_unlock(pte-1, ptl);
2499 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2500 unsigned long addr, unsigned long end,
2501 pte_fn_t fn, void *data)
2507 BUG_ON(pud_huge(*pud));
2509 pmd = pmd_alloc(mm, pud, addr);
2513 next = pmd_addr_end(addr, end);
2514 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2517 } while (pmd++, addr = next, addr != end);
2521 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2522 unsigned long addr, unsigned long end,
2523 pte_fn_t fn, void *data)
2529 pud = pud_alloc(mm, pgd, addr);
2533 next = pud_addr_end(addr, end);
2534 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2537 } while (pud++, addr = next, addr != end);
2542 * Scan a region of virtual memory, filling in page tables as necessary
2543 * and calling a provided function on each leaf page table.
2545 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2546 unsigned long size, pte_fn_t fn, void *data)
2550 unsigned long end = addr + size;
2553 BUG_ON(addr >= end);
2554 pgd = pgd_offset(mm, addr);
2556 next = pgd_addr_end(addr, end);
2557 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2560 } while (pgd++, addr = next, addr != end);
2564 EXPORT_SYMBOL_GPL(apply_to_page_range);
2567 * handle_pte_fault chooses page fault handler according to an entry
2568 * which was read non-atomically. Before making any commitment, on
2569 * those architectures or configurations (e.g. i386 with PAE) which
2570 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2571 * must check under lock before unmapping the pte and proceeding
2572 * (but do_wp_page is only called after already making such a check;
2573 * and do_anonymous_page can safely check later on).
2575 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2576 pte_t *page_table, pte_t orig_pte)
2579 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2580 if (sizeof(pte_t) > sizeof(unsigned long)) {
2581 spinlock_t *ptl = pte_lockptr(mm, pmd);
2583 same = pte_same(*page_table, orig_pte);
2587 pte_unmap(page_table);
2591 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2594 * If the source page was a PFN mapping, we don't have
2595 * a "struct page" for it. We do a best-effort copy by
2596 * just copying from the original user address. If that
2597 * fails, we just zero-fill it. Live with it.
2599 if (unlikely(!src)) {
2600 void *kaddr = kmap_atomic(dst);
2601 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2604 * This really shouldn't fail, because the page is there
2605 * in the page tables. But it might just be unreadable,
2606 * in which case we just give up and fill the result with
2609 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2611 kunmap_atomic(kaddr);
2612 flush_dcache_page(dst);
2614 copy_user_highpage(dst, src, va, vma);
2618 * This routine handles present pages, when users try to write
2619 * to a shared page. It is done by copying the page to a new address
2620 * and decrementing the shared-page counter for the old page.
2622 * Note that this routine assumes that the protection checks have been
2623 * done by the caller (the low-level page fault routine in most cases).
2624 * Thus we can safely just mark it writable once we've done any necessary
2627 * We also mark the page dirty at this point even though the page will
2628 * change only once the write actually happens. This avoids a few races,
2629 * and potentially makes it more efficient.
2631 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2632 * but allow concurrent faults), with pte both mapped and locked.
2633 * We return with mmap_sem still held, but pte unmapped and unlocked.
2635 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2636 unsigned long address, pte_t *page_table, pmd_t *pmd,
2637 spinlock_t *ptl, pte_t orig_pte)
2640 struct page *old_page, *new_page = NULL;
2643 int page_mkwrite = 0;
2644 struct page *dirty_page = NULL;
2645 unsigned long mmun_start = 0; /* For mmu_notifiers */
2646 unsigned long mmun_end = 0; /* For mmu_notifiers */
2648 old_page = vm_normal_page(vma, address, orig_pte);
2651 * VM_MIXEDMAP !pfn_valid() case
2653 * We should not cow pages in a shared writeable mapping.
2654 * Just mark the pages writable as we can't do any dirty
2655 * accounting on raw pfn maps.
2657 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2658 (VM_WRITE|VM_SHARED))
2664 * Take out anonymous pages first, anonymous shared vmas are
2665 * not dirty accountable.
2667 if (PageAnon(old_page) && !PageKsm(old_page)) {
2668 if (!trylock_page(old_page)) {
2669 page_cache_get(old_page);
2670 pte_unmap_unlock(page_table, ptl);
2671 lock_page(old_page);
2672 page_table = pte_offset_map_lock(mm, pmd, address,
2674 if (!pte_same(*page_table, orig_pte)) {
2675 unlock_page(old_page);
2678 page_cache_release(old_page);
2680 if (reuse_swap_page(old_page)) {
2682 * The page is all ours. Move it to our anon_vma so
2683 * the rmap code will not search our parent or siblings.
2684 * Protected against the rmap code by the page lock.
2686 page_move_anon_rmap(old_page, vma, address);
2687 unlock_page(old_page);
2690 unlock_page(old_page);
2691 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2692 (VM_WRITE|VM_SHARED))) {
2694 * Only catch write-faults on shared writable pages,
2695 * read-only shared pages can get COWed by
2696 * get_user_pages(.write=1, .force=1).
2698 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2699 struct vm_fault vmf;
2702 vmf.virtual_address = (void __user *)(address &
2704 vmf.pgoff = old_page->index;
2705 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2706 vmf.page = old_page;
2709 * Notify the address space that the page is about to
2710 * become writable so that it can prohibit this or wait
2711 * for the page to get into an appropriate state.
2713 * We do this without the lock held, so that it can
2714 * sleep if it needs to.
2716 page_cache_get(old_page);
2717 pte_unmap_unlock(page_table, ptl);
2719 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
2721 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2723 goto unwritable_page;
2725 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
2726 lock_page(old_page);
2727 if (!old_page->mapping) {
2728 ret = 0; /* retry the fault */
2729 unlock_page(old_page);
2730 goto unwritable_page;
2733 VM_BUG_ON(!PageLocked(old_page));
2736 * Since we dropped the lock we need to revalidate
2737 * the PTE as someone else may have changed it. If
2738 * they did, we just return, as we can count on the
2739 * MMU to tell us if they didn't also make it writable.
2741 page_table = pte_offset_map_lock(mm, pmd, address,
2743 if (!pte_same(*page_table, orig_pte)) {
2744 unlock_page(old_page);
2750 dirty_page = old_page;
2751 get_page(dirty_page);
2754 flush_cache_page(vma, address, pte_pfn(orig_pte));
2755 entry = pte_mkyoung(orig_pte);
2756 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2757 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2758 update_mmu_cache(vma, address, page_table);
2759 pte_unmap_unlock(page_table, ptl);
2760 ret |= VM_FAULT_WRITE;
2766 * Yes, Virginia, this is actually required to prevent a race
2767 * with clear_page_dirty_for_io() from clearing the page dirty
2768 * bit after it clear all dirty ptes, but before a racing
2769 * do_wp_page installs a dirty pte.
2771 * __do_fault is protected similarly.
2773 if (!page_mkwrite) {
2774 wait_on_page_locked(dirty_page);
2775 set_page_dirty_balance(dirty_page, page_mkwrite);
2776 /* file_update_time outside page_lock */
2778 file_update_time(vma->vm_file);
2780 put_page(dirty_page);
2782 struct address_space *mapping = dirty_page->mapping;
2784 set_page_dirty(dirty_page);
2785 unlock_page(dirty_page);
2786 page_cache_release(dirty_page);
2789 * Some device drivers do not set page.mapping
2790 * but still dirty their pages
2792 balance_dirty_pages_ratelimited(mapping);
2800 * Ok, we need to copy. Oh, well..
2802 page_cache_get(old_page);
2804 pte_unmap_unlock(page_table, ptl);
2806 if (unlikely(anon_vma_prepare(vma)))
2809 if (is_zero_pfn(pte_pfn(orig_pte))) {
2810 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2814 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2817 cow_user_page(new_page, old_page, address, vma);
2819 __SetPageUptodate(new_page);
2821 if (mem_cgroup_newpage_charge(new_page, mm, GFP_KERNEL))
2824 mmun_start = address & PAGE_MASK;
2825 mmun_end = mmun_start + PAGE_SIZE;
2826 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2829 * Re-check the pte - we dropped the lock
2831 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2832 if (likely(pte_same(*page_table, orig_pte))) {
2834 if (!PageAnon(old_page)) {
2835 dec_mm_counter_fast(mm, MM_FILEPAGES);
2836 inc_mm_counter_fast(mm, MM_ANONPAGES);
2839 inc_mm_counter_fast(mm, MM_ANONPAGES);
2840 flush_cache_page(vma, address, pte_pfn(orig_pte));
2841 entry = mk_pte(new_page, vma->vm_page_prot);
2842 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2844 * Clear the pte entry and flush it first, before updating the
2845 * pte with the new entry. This will avoid a race condition
2846 * seen in the presence of one thread doing SMC and another
2849 ptep_clear_flush(vma, address, page_table);
2850 page_add_new_anon_rmap(new_page, vma, address);
2852 * We call the notify macro here because, when using secondary
2853 * mmu page tables (such as kvm shadow page tables), we want the
2854 * new page to be mapped directly into the secondary page table.
2856 set_pte_at_notify(mm, address, page_table, entry);
2857 update_mmu_cache(vma, address, page_table);
2860 * Only after switching the pte to the new page may
2861 * we remove the mapcount here. Otherwise another
2862 * process may come and find the rmap count decremented
2863 * before the pte is switched to the new page, and
2864 * "reuse" the old page writing into it while our pte
2865 * here still points into it and can be read by other
2868 * The critical issue is to order this
2869 * page_remove_rmap with the ptp_clear_flush above.
2870 * Those stores are ordered by (if nothing else,)
2871 * the barrier present in the atomic_add_negative
2872 * in page_remove_rmap.
2874 * Then the TLB flush in ptep_clear_flush ensures that
2875 * no process can access the old page before the
2876 * decremented mapcount is visible. And the old page
2877 * cannot be reused until after the decremented
2878 * mapcount is visible. So transitively, TLBs to
2879 * old page will be flushed before it can be reused.
2881 page_remove_rmap(old_page);
2884 /* Free the old page.. */
2885 new_page = old_page;
2886 ret |= VM_FAULT_WRITE;
2888 mem_cgroup_uncharge_page(new_page);
2891 page_cache_release(new_page);
2893 pte_unmap_unlock(page_table, ptl);
2894 if (mmun_end > mmun_start)
2895 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2898 * Don't let another task, with possibly unlocked vma,
2899 * keep the mlocked page.
2901 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2902 lock_page(old_page); /* LRU manipulation */
2903 munlock_vma_page(old_page);
2904 unlock_page(old_page);
2906 page_cache_release(old_page);
2910 page_cache_release(new_page);
2913 page_cache_release(old_page);
2914 return VM_FAULT_OOM;
2917 page_cache_release(old_page);
2921 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2922 unsigned long start_addr, unsigned long end_addr,
2923 struct zap_details *details)
2925 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2928 static inline void unmap_mapping_range_tree(struct rb_root *root,
2929 struct zap_details *details)
2931 struct vm_area_struct *vma;
2932 pgoff_t vba, vea, zba, zea;
2934 vma_interval_tree_foreach(vma, root,
2935 details->first_index, details->last_index) {
2937 vba = vma->vm_pgoff;
2938 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
2939 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2940 zba = details->first_index;
2943 zea = details->last_index;
2947 unmap_mapping_range_vma(vma,
2948 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2949 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2954 static inline void unmap_mapping_range_list(struct list_head *head,
2955 struct zap_details *details)
2957 struct vm_area_struct *vma;
2960 * In nonlinear VMAs there is no correspondence between virtual address
2961 * offset and file offset. So we must perform an exhaustive search
2962 * across *all* the pages in each nonlinear VMA, not just the pages
2963 * whose virtual address lies outside the file truncation point.
2965 list_for_each_entry(vma, head, shared.nonlinear) {
2966 details->nonlinear_vma = vma;
2967 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2972 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2973 * @mapping: the address space containing mmaps to be unmapped.
2974 * @holebegin: byte in first page to unmap, relative to the start of
2975 * the underlying file. This will be rounded down to a PAGE_SIZE
2976 * boundary. Note that this is different from truncate_pagecache(), which
2977 * must keep the partial page. In contrast, we must get rid of
2979 * @holelen: size of prospective hole in bytes. This will be rounded
2980 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2982 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2983 * but 0 when invalidating pagecache, don't throw away private data.
2985 void unmap_mapping_range(struct address_space *mapping,
2986 loff_t const holebegin, loff_t const holelen, int even_cows)
2988 struct zap_details details;
2989 pgoff_t hba = holebegin >> PAGE_SHIFT;
2990 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2992 /* Check for overflow. */
2993 if (sizeof(holelen) > sizeof(hlen)) {
2995 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2996 if (holeend & ~(long long)ULONG_MAX)
2997 hlen = ULONG_MAX - hba + 1;
3000 details.check_mapping = even_cows? NULL: mapping;
3001 details.nonlinear_vma = NULL;
3002 details.first_index = hba;
3003 details.last_index = hba + hlen - 1;
3004 if (details.last_index < details.first_index)
3005 details.last_index = ULONG_MAX;
3008 mutex_lock(&mapping->i_mmap_mutex);
3009 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3010 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3011 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3012 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3013 mutex_unlock(&mapping->i_mmap_mutex);
3015 EXPORT_SYMBOL(unmap_mapping_range);
3018 #include <linux/module.h>
3019 // always try to free swap in zram swap case
3020 static int keep_to_free_swap = 1;
3021 module_param_named(keep_to_free_swap,keep_to_free_swap, int, S_IRUGO | S_IWUSR);
3026 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3027 * but allow concurrent faults), and pte mapped but not yet locked.
3028 * We return with mmap_sem still held, but pte unmapped and unlocked.
3030 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3031 unsigned long address, pte_t *page_table, pmd_t *pmd,
3032 unsigned int flags, pte_t orig_pte)
3035 struct page *page, *swapcache;
3039 struct mem_cgroup *ptr;
3043 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3046 entry = pte_to_swp_entry(orig_pte);
3047 if (unlikely(non_swap_entry(entry))) {
3048 if (is_migration_entry(entry)) {
3051 * FIXME: mszyprow: cruel, brute-force method for
3052 * letting cma/migration to finish it's job without
3053 * stealing the lock migration_entry_wait() and creating
3054 * a live-lock on the faulted page
3055 * (page->_count == 2 migration failure issue)
3059 migration_entry_wait(mm, pmd, address);
3060 } else if (is_hwpoison_entry(entry)) {
3061 ret = VM_FAULT_HWPOISON;
3063 print_bad_pte(vma, address, orig_pte, NULL);
3064 ret = VM_FAULT_SIGBUS;
3068 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3069 page = lookup_swap_cache(entry);
3071 page = swapin_readahead(entry,
3072 GFP_HIGHUSER_MOVABLE, vma, address);
3075 * Back out if somebody else faulted in this pte
3076 * while we released the pte lock.
3078 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3079 if (likely(pte_same(*page_table, orig_pte)))
3081 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3085 /* Had to read the page from swap area: Major fault */
3086 ret = VM_FAULT_MAJOR;
3087 count_vm_event(PGMAJFAULT);
3088 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3089 } else if (PageHWPoison(page)) {
3091 * hwpoisoned dirty swapcache pages are kept for killing
3092 * owner processes (which may be unknown at hwpoison time)
3094 ret = VM_FAULT_HWPOISON;
3095 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3101 locked = lock_page_or_retry(page, mm, flags);
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3105 ret |= VM_FAULT_RETRY;
3110 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3111 * release the swapcache from under us. The page pin, and pte_same
3112 * test below, are not enough to exclude that. Even if it is still
3113 * swapcache, we need to check that the page's swap has not changed.
3115 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3118 page = ksm_might_need_to_copy(page, vma, address);
3119 if (unlikely(!page)) {
3125 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3131 * Back out if somebody else already faulted in this pte.
3133 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3134 if (unlikely(!pte_same(*page_table, orig_pte)))
3137 if (unlikely(!PageUptodate(page))) {
3138 ret = VM_FAULT_SIGBUS;
3143 * The page isn't present yet, go ahead with the fault.
3145 * Be careful about the sequence of operations here.
3146 * To get its accounting right, reuse_swap_page() must be called
3147 * while the page is counted on swap but not yet in mapcount i.e.
3148 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3149 * must be called after the swap_free(), or it will never succeed.
3150 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3151 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3152 * in page->private. In this case, a record in swap_cgroup is silently
3153 * discarded at swap_free().
3156 inc_mm_counter_fast(mm, MM_ANONPAGES);
3157 dec_mm_counter_fast(mm, MM_SWAPENTS);
3158 pte = mk_pte(page, vma->vm_page_prot);
3159 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3160 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3161 flags &= ~FAULT_FLAG_WRITE;
3162 ret |= VM_FAULT_WRITE;
3165 flush_icache_page(vma, page);
3166 set_pte_at(mm, address, page_table, pte);
3167 if (page == swapcache)
3168 do_page_add_anon_rmap(page, vma, address, exclusive);
3169 else /* ksm created a completely new copy */
3170 page_add_new_anon_rmap(page, vma, address);
3171 /* It's better to call commit-charge after rmap is established */
3172 mem_cgroup_commit_charge_swapin(page, ptr);
3177 if (keep_to_free_swap || vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3179 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3181 try_to_free_swap(page);
3183 if (page != swapcache) {
3185 * Hold the lock to avoid the swap entry to be reused
3186 * until we take the PT lock for the pte_same() check
3187 * (to avoid false positives from pte_same). For
3188 * further safety release the lock after the swap_free
3189 * so that the swap count won't change under a
3190 * parallel locked swapcache.
3192 unlock_page(swapcache);
3193 page_cache_release(swapcache);
3196 if (flags & FAULT_FLAG_WRITE) {
3197 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3198 if (ret & VM_FAULT_ERROR)
3199 ret &= VM_FAULT_ERROR;
3203 /* No need to invalidate - it was non-present before */
3204 update_mmu_cache(vma, address, page_table);
3206 pte_unmap_unlock(page_table, ptl);
3210 mem_cgroup_cancel_charge_swapin(ptr);
3211 pte_unmap_unlock(page_table, ptl);
3215 page_cache_release(page);
3216 if (page != swapcache) {
3217 unlock_page(swapcache);
3218 page_cache_release(swapcache);
3224 * This is like a special single-page "expand_{down|up}wards()",
3225 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3226 * doesn't hit another vma.
3228 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3230 address &= PAGE_MASK;
3231 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3232 struct vm_area_struct *prev = vma->vm_prev;
3235 * Is there a mapping abutting this one below?
3237 * That's only ok if it's the same stack mapping
3238 * that has gotten split..
3240 if (prev && prev->vm_end == address)
3241 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3243 return expand_downwards(vma, address - PAGE_SIZE);
3245 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3246 struct vm_area_struct *next = vma->vm_next;
3248 /* As VM_GROWSDOWN but s/below/above/ */
3249 if (next && next->vm_start == address + PAGE_SIZE)
3250 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3252 return expand_upwards(vma, address + PAGE_SIZE);
3258 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3259 * but allow concurrent faults), and pte mapped but not yet locked.
3260 * We return with mmap_sem still held, but pte unmapped and unlocked.
3262 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3263 unsigned long address, pte_t *page_table, pmd_t *pmd,
3270 pte_unmap(page_table);
3272 /* Check if we need to add a guard page to the stack */
3273 if (check_stack_guard_page(vma, address) < 0)
3274 return VM_FAULT_SIGBUS;
3276 /* Use the zero-page for reads */
3277 if (!(flags & FAULT_FLAG_WRITE)) {
3278 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3279 vma->vm_page_prot));
3280 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3281 if (!pte_none(*page_table))
3286 /* Allocate our own private page. */
3287 if (unlikely(anon_vma_prepare(vma)))
3289 page = alloc_zeroed_user_highpage_movable(vma, address);
3293 * The memory barrier inside __SetPageUptodate makes sure that
3294 * preceeding stores to the page contents become visible before
3295 * the set_pte_at() write.
3297 __SetPageUptodate(page);
3299 if (mem_cgroup_newpage_charge(page, mm, GFP_KERNEL))
3302 entry = mk_pte(page, vma->vm_page_prot);
3303 if (vma->vm_flags & VM_WRITE)
3304 entry = pte_mkwrite(pte_mkdirty(entry));
3306 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3307 if (!pte_none(*page_table))
3310 inc_mm_counter_fast(mm, MM_ANONPAGES);
3311 page_add_new_anon_rmap(page, vma, address);
3313 set_pte_at(mm, address, page_table, entry);
3315 /* No need to invalidate - it was non-present before */
3316 update_mmu_cache(vma, address, page_table);
3318 pte_unmap_unlock(page_table, ptl);
3321 mem_cgroup_uncharge_page(page);
3322 page_cache_release(page);
3325 page_cache_release(page);
3327 return VM_FAULT_OOM;
3331 * __do_fault() tries to create a new page mapping. It aggressively
3332 * tries to share with existing pages, but makes a separate copy if
3333 * the FAULT_FLAG_WRITE is set in the flags parameter in order to avoid
3334 * the next page fault.
3336 * As this is called only for pages that do not currently exist, we
3337 * do not need to flush old virtual caches or the TLB.
3339 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3340 * but allow concurrent faults), and pte neither mapped nor locked.
3341 * We return with mmap_sem still held, but pte unmapped and unlocked.
3343 static int __do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3344 unsigned long address, pmd_t *pmd,
3345 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3350 struct page *cow_page;
3353 struct page *dirty_page = NULL;
3354 struct vm_fault vmf;
3356 int page_mkwrite = 0;
3359 * If we do COW later, allocate page befor taking lock_page()
3360 * on the file cache page. This will reduce lock holding time.
3362 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3364 if (unlikely(anon_vma_prepare(vma)))
3365 return VM_FAULT_OOM;
3367 cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3369 return VM_FAULT_OOM;
3371 if (mem_cgroup_newpage_charge(cow_page, mm, GFP_KERNEL)) {
3372 page_cache_release(cow_page);
3373 return VM_FAULT_OOM;
3378 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3383 ret = vma->vm_ops->fault(vma, &vmf);
3384 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3388 if (unlikely(PageHWPoison(vmf.page))) {
3389 if (ret & VM_FAULT_LOCKED)
3390 unlock_page(vmf.page);
3391 ret = VM_FAULT_HWPOISON;
3396 * For consistency in subsequent calls, make the faulted page always
3399 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3400 lock_page(vmf.page);
3402 VM_BUG_ON(!PageLocked(vmf.page));
3405 * Should we do an early C-O-W break?
3408 if (flags & FAULT_FLAG_WRITE) {
3409 if (!(vma->vm_flags & VM_SHARED)) {
3412 copy_user_highpage(page, vmf.page, address, vma);
3413 __SetPageUptodate(page);
3416 * If the page will be shareable, see if the backing
3417 * address space wants to know that the page is about
3418 * to become writable
3420 if (vma->vm_ops->page_mkwrite) {
3424 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
3425 tmp = vma->vm_ops->page_mkwrite(vma, &vmf);
3427 (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3429 goto unwritable_page;
3431 if (unlikely(!(tmp & VM_FAULT_LOCKED))) {
3433 if (!page->mapping) {
3434 ret = 0; /* retry the fault */
3436 goto unwritable_page;
3439 VM_BUG_ON(!PageLocked(page));
3446 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3449 * This silly early PAGE_DIRTY setting removes a race
3450 * due to the bad i386 page protection. But it's valid
3451 * for other architectures too.
3453 * Note that if FAULT_FLAG_WRITE is set, we either now have
3454 * an exclusive copy of the page, or this is a shared mapping,
3455 * so we can make it writable and dirty to avoid having to
3456 * handle that later.
3458 /* Only go through if we didn't race with anybody else... */
3459 if (likely(pte_same(*page_table, orig_pte))) {
3460 flush_icache_page(vma, page);
3461 entry = mk_pte(page, vma->vm_page_prot);
3462 if (flags & FAULT_FLAG_WRITE)
3463 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3465 inc_mm_counter_fast(mm, MM_ANONPAGES);
3466 page_add_new_anon_rmap(page, vma, address);
3468 inc_mm_counter_fast(mm, MM_FILEPAGES);
3469 page_add_file_rmap(page);
3470 if (flags & FAULT_FLAG_WRITE) {
3472 get_page(dirty_page);
3475 set_pte_at(mm, address, page_table, entry);
3477 /* no need to invalidate: a not-present page won't be cached */
3478 update_mmu_cache(vma, address, page_table);
3481 mem_cgroup_uncharge_page(cow_page);
3483 page_cache_release(page);
3485 anon = 1; /* no anon but release faulted_page */
3488 pte_unmap_unlock(page_table, ptl);
3491 struct address_space *mapping = page->mapping;
3494 if (set_page_dirty(dirty_page))
3496 unlock_page(dirty_page);
3497 put_page(dirty_page);
3498 if ((dirtied || page_mkwrite) && mapping) {
3500 * Some device drivers do not set page.mapping but still
3503 balance_dirty_pages_ratelimited(mapping);
3506 /* file_update_time outside page_lock */
3507 if (vma->vm_file && !page_mkwrite)
3508 file_update_time(vma->vm_file);
3510 unlock_page(vmf.page);
3512 page_cache_release(vmf.page);
3518 page_cache_release(page);
3521 /* fs's fault handler get error */
3523 mem_cgroup_uncharge_page(cow_page);
3524 page_cache_release(cow_page);
3529 static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3530 unsigned long address, pte_t *page_table, pmd_t *pmd,
3531 unsigned int flags, pte_t orig_pte)
3533 pgoff_t pgoff = (((address & PAGE_MASK)
3534 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3536 pte_unmap(page_table);
3537 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3541 * Fault of a previously existing named mapping. Repopulate the pte
3542 * from the encoded file_pte if possible. This enables swappable
3545 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3546 * but allow concurrent faults), and pte mapped but not yet locked.
3547 * We return with mmap_sem still held, but pte unmapped and unlocked.
3549 static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3550 unsigned long address, pte_t *page_table, pmd_t *pmd,
3551 unsigned int flags, pte_t orig_pte)
3555 flags |= FAULT_FLAG_NONLINEAR;
3557 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3560 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3562 * Page table corrupted: show pte and kill process.
3564 print_bad_pte(vma, address, orig_pte, NULL);
3565 return VM_FAULT_SIGBUS;
3568 pgoff = pte_to_pgoff(orig_pte);
3569 return __do_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3572 int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3573 unsigned long addr, int page_nid)
3577 count_vm_numa_event(NUMA_HINT_FAULTS);
3578 if (page_nid == numa_node_id())
3579 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3581 return mpol_misplaced(page, vma, addr);
3584 int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3585 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3587 struct page *page = NULL;
3591 bool migrated = false;
3594 * The "pte" at this point cannot be used safely without
3595 * validation through pte_unmap_same(). It's of NUMA type but
3596 * the pfn may be screwed if the read is non atomic.
3598 * ptep_modify_prot_start is not called as this is clearing
3599 * the _PAGE_NUMA bit and it is not really expected that there
3600 * would be concurrent hardware modifications to the PTE.
3602 ptl = pte_lockptr(mm, pmd);
3604 if (unlikely(!pte_same(*ptep, pte))) {
3605 pte_unmap_unlock(ptep, ptl);
3609 pte = pte_mknonnuma(pte);
3610 set_pte_at(mm, addr, ptep, pte);
3611 update_mmu_cache(vma, addr, ptep);
3613 page = vm_normal_page(vma, addr, pte);
3615 pte_unmap_unlock(ptep, ptl);
3619 page_nid = page_to_nid(page);
3620 target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3621 pte_unmap_unlock(ptep, ptl);
3622 if (target_nid == -1) {
3627 /* Migrate to the requested node */
3628 migrated = migrate_misplaced_page(page, target_nid);
3630 page_nid = target_nid;
3634 task_numa_fault(page_nid, 1, migrated);
3638 /* NUMA hinting page fault entry point for regular pmds */
3639 #ifdef CONFIG_NUMA_BALANCING
3640 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3641 unsigned long addr, pmd_t *pmdp)
3644 pte_t *pte, *orig_pte;
3645 unsigned long _addr = addr & PMD_MASK;
3646 unsigned long offset;
3650 spin_lock(&mm->page_table_lock);
3652 if (pmd_numa(pmd)) {
3653 set_pmd_at(mm, _addr, pmdp, pmd_mknonnuma(pmd));
3656 spin_unlock(&mm->page_table_lock);
3661 /* we're in a page fault so some vma must be in the range */
3663 BUG_ON(vma->vm_start >= _addr + PMD_SIZE);
3664 offset = max(_addr, vma->vm_start) & ~PMD_MASK;
3665 VM_BUG_ON(offset >= PMD_SIZE);
3666 orig_pte = pte = pte_offset_map_lock(mm, pmdp, _addr, &ptl);
3667 pte += offset >> PAGE_SHIFT;
3668 for (addr = _addr + offset; addr < _addr + PMD_SIZE; pte++, addr += PAGE_SIZE) {
3669 pte_t pteval = *pte;
3673 bool migrated = false;
3675 if (!pte_present(pteval))
3677 if (!pte_numa(pteval))
3679 if (addr >= vma->vm_end) {
3680 vma = find_vma(mm, addr);
3681 /* there's a pte present so there must be a vma */
3683 BUG_ON(addr < vma->vm_start);
3685 if (pte_numa(pteval)) {
3686 pteval = pte_mknonnuma(pteval);
3687 set_pte_at(mm, addr, pte, pteval);
3689 page = vm_normal_page(vma, addr, pteval);
3690 if (unlikely(!page))
3692 /* only check non-shared pages */
3693 if (unlikely(page_mapcount(page) != 1))
3696 page_nid = page_to_nid(page);
3697 target_nid = numa_migrate_prep(page, vma, addr, page_nid);
3698 pte_unmap_unlock(pte, ptl);
3699 if (target_nid != -1) {
3700 migrated = migrate_misplaced_page(page, target_nid);
3702 page_nid = target_nid;
3708 task_numa_fault(page_nid, 1, migrated);
3710 pte = pte_offset_map_lock(mm, pmdp, addr, &ptl);
3712 pte_unmap_unlock(orig_pte, ptl);
3717 static int do_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3718 unsigned long addr, pmd_t *pmdp)
3723 #endif /* CONFIG_NUMA_BALANCING */
3726 * These routines also need to handle stuff like marking pages dirty
3727 * and/or accessed for architectures that don't do it in hardware (most
3728 * RISC architectures). The early dirtying is also good on the i386.
3730 * There is also a hook called "update_mmu_cache()" that architectures
3731 * with external mmu caches can use to update those (ie the Sparc or
3732 * PowerPC hashed page tables that act as extended TLBs).
3734 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3735 * but allow concurrent faults), and pte mapped but not yet locked.
3736 * We return with mmap_sem still held, but pte unmapped and unlocked.
3738 int handle_pte_fault(struct mm_struct *mm,
3739 struct vm_area_struct *vma, unsigned long address,
3740 pte_t *pte, pmd_t *pmd, unsigned int flags)
3746 if (!pte_present(entry)) {
3747 if (pte_none(entry)) {
3749 if (likely(vma->vm_ops->fault))
3750 return do_linear_fault(mm, vma, address,
3751 pte, pmd, flags, entry);
3753 return do_anonymous_page(mm, vma, address,
3756 if (pte_file(entry))
3757 return do_nonlinear_fault(mm, vma, address,
3758 pte, pmd, flags, entry);
3759 return do_swap_page(mm, vma, address,
3760 pte, pmd, flags, entry);
3763 if (pte_numa(entry))
3764 return do_numa_page(mm, vma, address, entry, pte, pmd);
3766 ptl = pte_lockptr(mm, pmd);
3768 if (unlikely(!pte_same(*pte, entry)))
3770 if (flags & FAULT_FLAG_WRITE) {
3771 if (!pte_write(entry))
3772 return do_wp_page(mm, vma, address,
3773 pte, pmd, ptl, entry);
3774 entry = pte_mkdirty(entry);
3776 entry = pte_mkyoung(entry);
3777 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3778 update_mmu_cache(vma, address, pte);
3781 * This is needed only for protection faults but the arch code
3782 * is not yet telling us if this is a protection fault or not.
3783 * This still avoids useless tlb flushes for .text page faults
3786 if (flags & FAULT_FLAG_WRITE)
3787 flush_tlb_fix_spurious_fault(vma, address);
3790 pte_unmap_unlock(pte, ptl);
3795 * By the time we get here, we already hold the mm semaphore
3797 static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3798 unsigned long address, unsigned int flags)
3805 if (unlikely(is_vm_hugetlb_page(vma)))
3806 return hugetlb_fault(mm, vma, address, flags);
3809 pgd = pgd_offset(mm, address);
3810 pud = pud_alloc(mm, pgd, address);
3812 return VM_FAULT_OOM;
3813 pmd = pmd_alloc(mm, pud, address);
3815 return VM_FAULT_OOM;
3816 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3818 return do_huge_pmd_anonymous_page(mm, vma, address,
3821 pmd_t orig_pmd = *pmd;
3825 if (pmd_trans_huge(orig_pmd)) {
3826 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3829 * If the pmd is splitting, return and retry the
3830 * the fault. Alternative: wait until the split
3831 * is done, and goto retry.
3833 if (pmd_trans_splitting(orig_pmd))
3836 if (pmd_numa(orig_pmd))
3837 return do_huge_pmd_numa_page(mm, vma, address,
3840 if (dirty && !pmd_write(orig_pmd)) {
3841 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3844 * If COW results in an oom, the huge pmd will
3845 * have been split, so retry the fault on the
3846 * pte for a smaller charge.
3848 if (unlikely(ret & VM_FAULT_OOM))
3852 huge_pmd_set_accessed(mm, vma, address, pmd,
3861 return do_pmd_numa_page(mm, vma, address, pmd);
3864 * Use __pte_alloc instead of pte_alloc_map, because we can't
3865 * run pte_offset_map on the pmd, if an huge pmd could
3866 * materialize from under us from a different thread.
3868 if (unlikely(pmd_none(*pmd)) &&
3869 unlikely(__pte_alloc(mm, vma, pmd, address)))
3870 return VM_FAULT_OOM;
3871 /* if an huge pmd materialized from under us just retry later */
3872 if (unlikely(pmd_trans_huge(*pmd)))
3875 * A regular pmd is established and it can't morph into a huge pmd
3876 * from under us anymore at this point because we hold the mmap_sem
3877 * read mode and khugepaged takes it in write mode. So now it's
3878 * safe to run pte_offset_map().
3880 pte = pte_offset_map(pmd, address);
3882 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3885 int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3886 unsigned long address, unsigned int flags)
3890 __set_current_state(TASK_RUNNING);
3892 count_vm_event(PGFAULT);
3893 mem_cgroup_count_vm_event(mm, PGFAULT);
3895 /* do counter updates before entering really critical section. */
3896 check_sync_rss_stat(current);
3899 * Enable the memcg OOM handling for faults triggered in user
3900 * space. Kernel faults are handled more gracefully.
3902 if (flags & FAULT_FLAG_USER)
3903 mem_cgroup_oom_enable();
3905 ret = __handle_mm_fault(mm, vma, address, flags);
3907 if (flags & FAULT_FLAG_USER) {
3908 mem_cgroup_oom_disable();
3910 * The task may have entered a memcg OOM situation but
3911 * if the allocation error was handled gracefully (no
3912 * VM_FAULT_OOM), there is no need to kill anything.
3913 * Just clean up the OOM state peacefully.
3915 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3916 mem_cgroup_oom_synchronize(false);
3922 #ifndef __PAGETABLE_PUD_FOLDED
3924 * Allocate page upper directory.
3925 * We've already handled the fast-path in-line.
3927 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3929 pud_t *new = pud_alloc_one(mm, address);
3933 smp_wmb(); /* See comment in __pte_alloc */
3935 spin_lock(&mm->page_table_lock);
3936 if (pgd_present(*pgd)) /* Another has populated it */
3939 pgd_populate(mm, pgd, new);
3940 spin_unlock(&mm->page_table_lock);
3943 #endif /* __PAGETABLE_PUD_FOLDED */
3945 #ifndef __PAGETABLE_PMD_FOLDED
3947 * Allocate page middle directory.
3948 * We've already handled the fast-path in-line.
3950 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3952 pmd_t *new = pmd_alloc_one(mm, address);
3956 smp_wmb(); /* See comment in __pte_alloc */
3958 spin_lock(&mm->page_table_lock);
3959 #ifndef __ARCH_HAS_4LEVEL_HACK
3960 if (pud_present(*pud)) /* Another has populated it */
3963 pud_populate(mm, pud, new);
3965 if (pgd_present(*pud)) /* Another has populated it */
3968 pgd_populate(mm, pud, new);
3969 #endif /* __ARCH_HAS_4LEVEL_HACK */
3970 spin_unlock(&mm->page_table_lock);
3973 #endif /* __PAGETABLE_PMD_FOLDED */
3975 #if !defined(__HAVE_ARCH_GATE_AREA)
3977 #if defined(AT_SYSINFO_EHDR)
3978 static struct vm_area_struct gate_vma;
3980 static int __init gate_vma_init(void)
3982 gate_vma.vm_mm = NULL;
3983 gate_vma.vm_start = FIXADDR_USER_START;
3984 gate_vma.vm_end = FIXADDR_USER_END;
3985 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
3986 gate_vma.vm_page_prot = __P101;
3990 __initcall(gate_vma_init);
3993 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
3995 #ifdef AT_SYSINFO_EHDR
4002 int in_gate_area_no_mm(unsigned long addr)
4004 #ifdef AT_SYSINFO_EHDR
4005 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4011 #endif /* __HAVE_ARCH_GATE_AREA */
4013 static int __follow_pte(struct mm_struct *mm, unsigned long address,
4014 pte_t **ptepp, spinlock_t **ptlp)
4021 pgd = pgd_offset(mm, address);
4022 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4025 pud = pud_offset(pgd, address);
4026 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4029 pmd = pmd_offset(pud, address);
4030 VM_BUG_ON(pmd_trans_huge(*pmd));
4031 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4034 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4038 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4041 if (!pte_present(*ptep))
4046 pte_unmap_unlock(ptep, *ptlp);
4051 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4052 pte_t **ptepp, spinlock_t **ptlp)
4056 /* (void) is needed to make gcc happy */
4057 (void) __cond_lock(*ptlp,
4058 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4063 * follow_pfn - look up PFN at a user virtual address
4064 * @vma: memory mapping
4065 * @address: user virtual address
4066 * @pfn: location to store found PFN
4068 * Only IO mappings and raw PFN mappings are allowed.
4070 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4072 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4079 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4082 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4085 *pfn = pte_pfn(*ptep);
4086 pte_unmap_unlock(ptep, ptl);
4089 EXPORT_SYMBOL(follow_pfn);
4091 #ifdef CONFIG_HAVE_IOREMAP_PROT
4092 int follow_phys(struct vm_area_struct *vma,
4093 unsigned long address, unsigned int flags,
4094 unsigned long *prot, resource_size_t *phys)
4100 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4103 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4107 if ((flags & FOLL_WRITE) && !pte_write(pte))
4110 *prot = pgprot_val(pte_pgprot(pte));
4111 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4115 pte_unmap_unlock(ptep, ptl);
4120 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4121 void *buf, int len, int write)
4123 resource_size_t phys_addr;
4124 unsigned long prot = 0;
4125 void __iomem *maddr;
4126 int offset = addr & (PAGE_SIZE-1);
4128 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4131 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4133 memcpy_toio(maddr + offset, buf, len);
4135 memcpy_fromio(buf, maddr + offset, len);
4140 EXPORT_SYMBOL_GPL(generic_access_phys);
4144 * Access another process' address space as given in mm. If non-NULL, use the
4145 * given task for page fault accounting.
4147 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4148 unsigned long addr, void *buf, int len, int write)
4150 struct vm_area_struct *vma;
4151 void *old_buf = buf;
4153 down_read(&mm->mmap_sem);
4154 /* ignore errors, just check how much was successfully transferred */
4156 int bytes, ret, offset;
4158 struct page *page = NULL;
4160 ret = get_user_pages(tsk, mm, addr, 1,
4161 write, 1, &page, &vma);
4164 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4165 * we can access using slightly different code.
4167 #ifdef CONFIG_HAVE_IOREMAP_PROT
4168 vma = find_vma(mm, addr);
4169 if (!vma || vma->vm_start > addr)
4171 if (vma->vm_ops && vma->vm_ops->access)
4172 ret = vma->vm_ops->access(vma, addr, buf,
4180 offset = addr & (PAGE_SIZE-1);
4181 if (bytes > PAGE_SIZE-offset)
4182 bytes = PAGE_SIZE-offset;
4186 copy_to_user_page(vma, page, addr,
4187 maddr + offset, buf, bytes);
4188 set_page_dirty_lock(page);
4190 copy_from_user_page(vma, page, addr,
4191 buf, maddr + offset, bytes);
4194 page_cache_release(page);
4200 up_read(&mm->mmap_sem);
4202 return buf - old_buf;
4206 * access_remote_vm - access another process' address space
4207 * @mm: the mm_struct of the target address space
4208 * @addr: start address to access
4209 * @buf: source or destination buffer
4210 * @len: number of bytes to transfer
4211 * @write: whether the access is a write
4213 * The caller must hold a reference on @mm.
4215 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4216 void *buf, int len, int write)
4218 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4222 * Access another process' address space.
4223 * Source/target buffer must be kernel space,
4224 * Do not walk the page table directly, use get_user_pages
4226 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4227 void *buf, int len, int write)
4229 struct mm_struct *mm;
4232 mm = get_task_mm(tsk);
4236 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4243 * Print the name of a VMA.
4245 void print_vma_addr(char *prefix, unsigned long ip)
4247 struct mm_struct *mm = current->mm;
4248 struct vm_area_struct *vma;
4251 * Do not print if we are in atomic
4252 * contexts (in exception stacks, etc.):
4254 if (preempt_count())
4257 down_read(&mm->mmap_sem);
4258 vma = find_vma(mm, ip);
4259 if (vma && vma->vm_file) {
4260 struct file *f = vma->vm_file;
4261 char *buf = (char *)__get_free_page(GFP_KERNEL);
4265 p = d_path(&f->f_path, buf, PAGE_SIZE);
4268 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4270 vma->vm_end - vma->vm_start);
4271 free_page((unsigned long)buf);
4274 up_read(&mm->mmap_sem);
4277 #ifdef CONFIG_PROVE_LOCKING
4278 void might_fault(void)
4281 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4282 * holding the mmap_sem, this is safe because kernel memory doesn't
4283 * get paged out, therefore we'll never actually fault, and the
4284 * below annotations will generate false positives.
4286 if (segment_eq(get_fs(), KERNEL_DS))
4291 * it would be nicer only to annotate paths which are not under
4292 * pagefault_disable, however that requires a larger audit and
4293 * providing helpers like get_user_atomic.
4295 if (!in_atomic() && current->mm)
4296 might_lock_read(¤t->mm->mmap_sem);
4298 EXPORT_SYMBOL(might_fault);
4301 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4302 static void clear_gigantic_page(struct page *page,
4304 unsigned int pages_per_huge_page)
4307 struct page *p = page;
4310 for (i = 0; i < pages_per_huge_page;
4311 i++, p = mem_map_next(p, page, i)) {
4313 clear_user_highpage(p, addr + i * PAGE_SIZE);
4316 void clear_huge_page(struct page *page,
4317 unsigned long addr, unsigned int pages_per_huge_page)
4321 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4322 clear_gigantic_page(page, addr, pages_per_huge_page);
4327 for (i = 0; i < pages_per_huge_page; i++) {
4329 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4333 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4335 struct vm_area_struct *vma,
4336 unsigned int pages_per_huge_page)
4339 struct page *dst_base = dst;
4340 struct page *src_base = src;
4342 for (i = 0; i < pages_per_huge_page; ) {
4344 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4347 dst = mem_map_next(dst, dst_base, i);
4348 src = mem_map_next(src, src_base, i);
4352 void copy_user_huge_page(struct page *dst, struct page *src,
4353 unsigned long addr, struct vm_area_struct *vma,
4354 unsigned int pages_per_huge_page)
4358 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4359 copy_user_gigantic_page(dst, src, addr, vma,
4360 pages_per_huge_page);
4365 for (i = 0; i < pages_per_huge_page; i++) {
4367 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4370 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */