mm/fremap.c

   1 /*
   2  *   linux/mm/fremap.c
   3  *
   4  * Explicit pagetable population and nonlinear (random) mappings support.
   5  *
   6  * started by Ingo Molnar, Copyright (C) 2002, 2003
   7  */
   8
   9 #include <linux/mm.h>
  10 #include <linux/swap.h>
  11 #include <linux/file.h>
  12 #include <linux/mman.h>
  13 #include <linux/pagemap.h>
  14 #include <linux/swapops.h>
  15 #include <linux/rmap.h>
  16 #include <linux/module.h>
  17
  18 #include <asm/mmu_context.h>
  19 #include <asm/cacheflush.h>
  20 #include <asm/tlbflush.h>
  21
  22 static inline void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  23                         unsigned long addr, pte_t *ptep)
  24 {
  25         pte_t pte = *ptep;
  26
  27         if (pte_none(pte))
  28                 return;
  29         if (pte_present(pte)) {
  30                 unsigned long pfn = pte_pfn(pte);
  31
  32                 flush_cache_page(vma, addr);
  33                 pte = ptep_clear_flush(vma, addr, ptep);
  34                 if (pfn_valid(pfn)) {
  35                         struct page *page = pfn_to_page(pfn);
  36                         if (!PageReserved(page)) {
  37                                 if (pte_dirty(pte))
  38                                         set_page_dirty(page);
  39                                 page_remove_rmap(page);
  40                                 page_cache_release(page);
  41                                 mm->rss--;
  42                         }
  43                 }
  44         } else {
  45                 if (!pte_file(pte))
  46                         free_swap_and_cache(pte_to_swp_entry(pte));
  47                 pte_clear(ptep);
  48         }
  49 }
  50
  51 /*
  52  * Install a file page to a given virtual memory address, release any
  53  * previously existing mapping.
  54  */
  55 int install_page(struct mm_struct *mm, struct vm_area_struct *vma,
  56                 unsigned long addr, struct page *page, pgprot_t prot)
  57 {
  58         struct inode *inode;
  59         pgoff_t size;
  60         int err = -ENOMEM;
  61         pte_t *pte;
  62         pgd_t *pgd;
  63         pmd_t *pmd;
  64         pte_t pte_val;
  65
  66         pgd = pgd_offset(mm, addr);
  67         spin_lock(&mm->page_table_lock);
  68
  69         pmd = pmd_alloc(mm, pgd, addr);
  70         if (!pmd)
  71                 goto err_unlock;
  72
  73         pte = pte_alloc_map(mm, pmd, addr);
  74         if (!pte)
  75                 goto err_unlock;
  76
  77         /*
  78          * This page may have been truncated. Tell the
  79          * caller about it.
  80          */
  81         err = -EINVAL;
  82         inode = vma->vm_file->f_mapping->host;
  83         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
  84         if (!page->mapping || page->index >= size)
  85                 goto err_unlock;
  86
  87         zap_pte(mm, vma, addr, pte);
  88
  89         mm->rss++;
  90         flush_icache_page(vma, page);
  91         set_pte(pte, mk_pte(page, prot));
  92         page_add_file_rmap(page);
  93         pte_val = *pte;
  94         pte_unmap(pte);
  95         update_mmu_cache(vma, addr, pte_val);
  96
  97         err = 0;
  98 err_unlock:
  99         spin_unlock(&mm->page_table_lock);
 100         return err;
 101 }
 102 EXPORT_SYMBOL(install_page);
 103
 104
 105 /*
 106  * Install a file pte to a given virtual memory address, release any
 107  * previously existing mapping.
 108  */
 109 int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 110                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
 111 {
 112         int err = -ENOMEM;
 113         pte_t *pte;
 114         pgd_t *pgd;
 115         pmd_t *pmd;
 116         pte_t pte_val;
 117
 118         pgd = pgd_offset(mm, addr);
 119         spin_lock(&mm->page_table_lock);
 120
 121         pmd = pmd_alloc(mm, pgd, addr);
 122         if (!pmd)
 123                 goto err_unlock;
 124
 125         pte = pte_alloc_map(mm, pmd, addr);
 126         if (!pte)
 127                 goto err_unlock;
 128
 129         zap_pte(mm, vma, addr, pte);
 130
 131         set_pte(pte, pgoff_to_pte(pgoff));
 132         pte_val = *pte;
 133         pte_unmap(pte);
 134         update_mmu_cache(vma, addr, pte_val);
 135         spin_unlock(&mm->page_table_lock);
 136         return 0;
 137
 138 err_unlock:
 139         spin_unlock(&mm->page_table_lock);
 140         return err;
 141 }
 142
 143
 144 /***
 145  * sys_remap_file_pages - remap arbitrary pages of a shared backing store
 146  *                        file within an existing vma.
 147  * @start: start of the remapped virtual memory range
 148  * @size: size of the remapped virtual memory range
 149  * @prot: new protection bits of the range
 150  * @pgoff: to be mapped page of the backing store file
 151  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 152  *
 153  * this syscall works purely via pagetables, so it's the most efficient
 154  * way to map the same (large) file into a given virtual window. Unlike
 155  * mmap()/mremap() it does not create any new vmas. The new mappings are
 156  * also safe across swapout.
 157  *
 158  * NOTE: the 'prot' parameter right now is ignored, and the vma's default
 159  * protection is used. Arbitrary protections might be implemented in the
 160  * future.
 161  */
 162 asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
 163         unsigned long __prot, unsigned long pgoff, unsigned long flags)
 164 {
 165         struct mm_struct *mm = current->mm;
 166         struct address_space *mapping;
 167         unsigned long end = start + size;
 168         struct vm_area_struct *vma;
 169         int err = -EINVAL;
 170         int has_write_lock = 0;
 171
 172         if (__prot)
 173                 return err;
 174         /*
 175          * Sanitize the syscall parameters:
 176          */
 177         start = start & PAGE_MASK;
 178         size = size & PAGE_MASK;
 179
 180         /* Does the address range wrap, or is the span zero-sized? */
 181         if (start + size <= start)
 182                 return err;
 183
 184         /* Can we represent this offset inside this architecture's pte's? */
 185 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 186         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 187                 return err;
 188 #endif
 189
 190         /* We need down_write() to change vma->vm_flags. */
 191         down_read(&mm->mmap_sem);
 192  retry:
 193         vma = find_vma(mm, start);
 194
 195         /*
 196          * Make sure the vma is shared, that it supports prefaulting,
 197          * and that the remapped range is valid and fully within
 198          * the single existing vma.  vm_private_data is used as a
 199          * swapout cursor in a VM_NONLINEAR vma (unless VM_RESERVED
 200          * or VM_LOCKED, but VM_LOCKED could be revoked later on).
 201          */
 202         if (vma && (vma->vm_flags & VM_SHARED) &&
 203                 (!vma->vm_private_data ||
 204                         (vma->vm_flags & (VM_NONLINEAR|VM_RESERVED))) &&
 205                 vma->vm_ops && vma->vm_ops->populate &&
 206                         end > start && start >= vma->vm_start &&
 207                                 end <= vma->vm_end) {
 208
 209                 /* Must set VM_NONLINEAR before any pages are populated. */
 210                 if (pgoff != linear_page_index(vma, start) &&
 211                     !(vma->vm_flags & VM_NONLINEAR)) {
 212                         if (!has_write_lock) {
 213                                 up_read(&mm->mmap_sem);
 214                                 down_write(&mm->mmap_sem);
 215                                 has_write_lock = 1;
 216                                 goto retry;
 217                         }
 218                         mapping = vma->vm_file->f_mapping;
 219                         spin_lock(&mapping->i_mmap_lock);
 220                         flush_dcache_mmap_lock(mapping);
 221                         vma->vm_flags |= VM_NONLINEAR;
 222                         vma_prio_tree_remove(vma, &mapping->i_mmap);
 223                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 224                         flush_dcache_mmap_unlock(mapping);
 225                         spin_unlock(&mapping->i_mmap_lock);
 226                 }
 227
 228                 err = vma->vm_ops->populate(vma, start, size,
 229                                             vma->vm_page_prot,
 230                                             pgoff, flags & MAP_NONBLOCK);
 231
 232                 /*
 233                  * We can't clear VM_NONLINEAR because we'd have to do
 234                  * it after ->populate completes, and that would prevent
 235                  * downgrading the lock.  (Locks can't be upgraded).
 236                  */
 237         }
 238         if (likely(!has_write_lock))
 239                 up_read(&mm->mmap_sem);
 240         else
 241                 up_write(&mm->mmap_sem);
 242
 243         return err;
 244 }
 245