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 #include <linux/syscalls.h>
  18
  19 #include <asm/mmu_context.h>
  20 #include <asm/cacheflush.h>
  21 #include <asm/tlbflush.h>
  22
  23 static int zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  24                         unsigned long addr, pte_t *ptep)
  25 {
  26         pte_t pte = *ptep;
  27         struct page *page = NULL;
  28
  29         if (pte_present(pte)) {
  30                 flush_cache_page(vma, addr, pte_pfn(pte));
  31                 pte = ptep_clear_flush(vma, addr, ptep);
  32                 page = vm_normal_page(vma, addr, pte);
  33                 if (page) {
  34                         if (pte_dirty(pte))
  35                                 set_page_dirty(page);
  36                         page_remove_rmap(page);
  37                         page_cache_release(page);
  38                 }
  39         } else {
  40                 if (!pte_file(pte))
  41                         free_swap_and_cache(pte_to_swp_entry(pte));
  42                 pte_clear(mm, addr, ptep);
  43         }
  44         return !!page;
  45 }
  46
  47 /*
  48  * Install a file page to a given virtual memory address, release any
  49  * previously existing mapping.
  50  */
  51 int install_page(struct mm_struct *mm, struct vm_area_struct *vma,
  52                 unsigned long addr, struct page *page, pgprot_t prot)
  53 {
  54         struct inode *inode;
  55         pgoff_t size;
  56         int err = -ENOMEM;
  57         pte_t *pte;
  58         pmd_t *pmd;
  59         pud_t *pud;
  60         pgd_t *pgd;
  61         pte_t pte_val;
  62         spinlock_t *ptl;
  63
  64         pgd = pgd_offset(mm, addr);
  65         pud = pud_alloc(mm, pgd, addr);
  66         if (!pud)
  67                 goto out;
  68         pmd = pmd_alloc(mm, pud, addr);
  69         if (!pmd)
  70                 goto out;
  71         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
  72         if (!pte)
  73                 goto out;
  74
  75         /*
  76          * This page may have been truncated. Tell the
  77          * caller about it.
  78          */
  79         err = -EINVAL;
  80         inode = vma->vm_file->f_mapping->host;
  81         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
  82         if (!page->mapping || page->index >= size)
  83                 goto unlock;
  84         err = -ENOMEM;
  85         if (page_mapcount(page) > INT_MAX/2)
  86                 goto unlock;
  87
  88         if (pte_none(*pte) || !zap_pte(mm, vma, addr, pte))
  89                 inc_mm_counter(mm, file_rss);
  90
  91         flush_icache_page(vma, page);
  92         set_pte_at(mm, addr, pte, mk_pte(page, prot));
  93         page_add_file_rmap(page);
  94         pte_val = *pte;
  95         update_mmu_cache(vma, addr, pte_val);
  96         err = 0;
  97 unlock:
  98         pte_unmap_unlock(pte, ptl);
  99 out:
 100         return err;
 101 }
 102 EXPORT_SYMBOL(install_page);
 103
 104 /*
 105  * Install a file pte to a given virtual memory address, release any
 106  * previously existing mapping.
 107  */
 108 int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 109                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
 110 {
 111         int err = -ENOMEM;
 112         pte_t *pte;
 113         pmd_t *pmd;
 114         pud_t *pud;
 115         pgd_t *pgd;
 116         pte_t pte_val;
 117         spinlock_t *ptl;
 118
 119         pgd = pgd_offset(mm, addr);
 120         pud = pud_alloc(mm, pgd, addr);
 121         if (!pud)
 122                 goto out;
 123         pmd = pmd_alloc(mm, pud, addr);
 124         if (!pmd)
 125                 goto out;
 126         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
 127         if (!pte)
 128                 goto out;
 129
 130         if (!pte_none(*pte) && zap_pte(mm, vma, addr, pte)) {
 131                 update_hiwater_rss(mm);
 132                 dec_mm_counter(mm, file_rss);
 133         }
 134
 135         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
 136         pte_val = *pte;
 137         update_mmu_cache(vma, addr, pte_val);
 138         pte_unmap_unlock(pte, ptl);
 139         err = 0;
 140 out:
 141         return err;
 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.
 200          */
 201         if (vma && (vma->vm_flags & VM_SHARED) &&
 202                 (!vma->vm_private_data || (vma->vm_flags & VM_NONLINEAR)) &&
 203                 vma->vm_ops && vma->vm_ops->populate &&
 204                         end > start && start >= vma->vm_start &&
 205                                 end <= vma->vm_end) {
 206
 207                 /* Must set VM_NONLINEAR before any pages are populated. */
 208                 if (pgoff != linear_page_index(vma, start) &&
 209                     !(vma->vm_flags & VM_NONLINEAR)) {
 210                         if (!has_write_lock) {
 211                                 up_read(&mm->mmap_sem);
 212                                 down_write(&mm->mmap_sem);
 213                                 has_write_lock = 1;
 214                                 goto retry;
 215                         }
 216                         mapping = vma->vm_file->f_mapping;
 217                         spin_lock(&mapping->i_mmap_lock);
 218                         flush_dcache_mmap_lock(mapping);
 219                         vma->vm_flags |= VM_NONLINEAR;
 220                         vma_prio_tree_remove(vma, &mapping->i_mmap);
 221                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 222                         flush_dcache_mmap_unlock(mapping);
 223                         spin_unlock(&mapping->i_mmap_lock);
 224                 }
 225
 226                 err = vma->vm_ops->populate(vma, start, size,
 227                                             vma->vm_page_prot,
 228                                             pgoff, flags & MAP_NONBLOCK);
 229
 230                 /*
 231                  * We can't clear VM_NONLINEAR because we'd have to do
 232                  * it after ->populate completes, and that would prevent
 233                  * downgrading the lock.  (Locks can't be upgraded).
 234                  */
 235         }
 236         if (likely(!has_write_lock))
 237                 up_read(&mm->mmap_sem);
 238         else
 239                 up_write(&mm->mmap_sem);
 240
 241         return err;
 242 }
 243