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         pte_t pte_val;
  59         spinlock_t *ptl;
  60
  61         pte = get_locked_pte(mm, addr, &ptl);
  62         if (!pte)
  63                 goto out;
  64
  65         /*
  66          * This page may have been truncated. Tell the
  67          * caller about it.
  68          */
  69         err = -EINVAL;
  70         inode = vma->vm_file->f_mapping->host;
  71         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
  72         if (!page->mapping || page->index >= size)
  73                 goto unlock;
  74         err = -ENOMEM;
  75         if (page_mapcount(page) > INT_MAX/2)
  76                 goto unlock;
  77
  78         if (pte_none(*pte) || !zap_pte(mm, vma, addr, pte))
  79                 inc_mm_counter(mm, file_rss);
  80
  81         flush_icache_page(vma, page);
  82         set_pte_at(mm, addr, pte, mk_pte(page, prot));
  83         page_add_file_rmap(page);
  84         pte_val = *pte;
  85         update_mmu_cache(vma, addr, pte_val);
  86         lazy_mmu_prot_update(pte_val);
  87         err = 0;
  88 unlock:
  89         pte_unmap_unlock(pte, ptl);
  90 out:
  91         return err;
  92 }
  93 EXPORT_SYMBOL(install_page);
  94
  95 /*
  96  * Install a file pte to a given virtual memory address, release any
  97  * previously existing mapping.
  98  */
  99 int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 100                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
 101 {
 102         int err = -ENOMEM;
 103         pte_t *pte;
 104         pte_t pte_val;
 105         spinlock_t *ptl;
 106
 107         pte = get_locked_pte(mm, addr, &ptl);
 108         if (!pte)
 109                 goto out;
 110
 111         if (!pte_none(*pte) && zap_pte(mm, vma, addr, pte)) {
 112                 update_hiwater_rss(mm);
 113                 dec_mm_counter(mm, file_rss);
 114         }
 115
 116         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
 117         pte_val = *pte;
 118         /*
 119          * We don't need to run update_mmu_cache() here because the "file pte"
 120          * being installed by install_file_pte() is not a real pte - it's a
 121          * non-present entry (like a swap entry), noting what file offset should
 122          * be mapped there when there's a fault (in a non-linear vma where
 123          * that's not obvious).
 124          */
 125         pte_unmap_unlock(pte, ptl);
 126         err = 0;
 127 out:
 128         return err;
 129 }
 130
 131 /***
 132  * sys_remap_file_pages - remap arbitrary pages of a shared backing store
 133  *                        file within an existing vma.
 134  * @start: start of the remapped virtual memory range
 135  * @size: size of the remapped virtual memory range
 136  * @prot: new protection bits of the range
 137  * @pgoff: to be mapped page of the backing store file
 138  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 139  *
 140  * this syscall works purely via pagetables, so it's the most efficient
 141  * way to map the same (large) file into a given virtual window. Unlike
 142  * mmap()/mremap() it does not create any new vmas. The new mappings are
 143  * also safe across swapout.
 144  *
 145  * NOTE: the 'prot' parameter right now is ignored, and the vma's default
 146  * protection is used. Arbitrary protections might be implemented in the
 147  * future.
 148  */
 149 asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
 150         unsigned long __prot, unsigned long pgoff, unsigned long flags)
 151 {
 152         struct mm_struct *mm = current->mm;
 153         struct address_space *mapping;
 154         unsigned long end = start + size;
 155         struct vm_area_struct *vma;
 156         int err = -EINVAL;
 157         int has_write_lock = 0;
 158
 159         if (__prot)
 160                 return err;
 161         /*
 162          * Sanitize the syscall parameters:
 163          */
 164         start = start & PAGE_MASK;
 165         size = size & PAGE_MASK;
 166
 167         /* Does the address range wrap, or is the span zero-sized? */
 168         if (start + size <= start)
 169                 return err;
 170
 171         /* Can we represent this offset inside this architecture's pte's? */
 172 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 173         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 174                 return err;
 175 #endif
 176
 177         /* We need down_write() to change vma->vm_flags. */
 178         down_read(&mm->mmap_sem);
 179  retry:
 180         vma = find_vma(mm, start);
 181
 182         /*
 183          * Make sure the vma is shared, that it supports prefaulting,
 184          * and that the remapped range is valid and fully within
 185          * the single existing vma.  vm_private_data is used as a
 186          * swapout cursor in a VM_NONLINEAR vma.
 187          */
 188         if (vma && (vma->vm_flags & VM_SHARED) &&
 189                 (!vma->vm_private_data || (vma->vm_flags & VM_NONLINEAR)) &&
 190                 vma->vm_ops && vma->vm_ops->populate &&
 191                         end > start && start >= vma->vm_start &&
 192                                 end <= vma->vm_end) {
 193
 194                 /* Must set VM_NONLINEAR before any pages are populated. */
 195                 if (pgoff != linear_page_index(vma, start) &&
 196                     !(vma->vm_flags & VM_NONLINEAR)) {
 197                         if (!has_write_lock) {
 198                                 up_read(&mm->mmap_sem);
 199                                 down_write(&mm->mmap_sem);
 200                                 has_write_lock = 1;
 201                                 goto retry;
 202                         }
 203                         mapping = vma->vm_file->f_mapping;
 204                         spin_lock(&mapping->i_mmap_lock);
 205                         flush_dcache_mmap_lock(mapping);
 206                         vma->vm_flags |= VM_NONLINEAR;
 207                         vma_prio_tree_remove(vma, &mapping->i_mmap);
 208                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 209                         flush_dcache_mmap_unlock(mapping);
 210                         spin_unlock(&mapping->i_mmap_lock);
 211                 }
 212
 213                 err = vma->vm_ops->populate(vma, start, size,
 214                                             vma->vm_page_prot,
 215                                             pgoff, flags & MAP_NONBLOCK);
 216
 217                 /*
 218                  * We can't clear VM_NONLINEAR because we'd have to do
 219                  * it after ->populate completes, and that would prevent
 220                  * downgrading the lock.  (Locks can't be upgraded).
 221                  */
 222         }
 223         if (likely(!has_write_lock))
 224                 up_read(&mm->mmap_sem);
 225         else
 226                 up_write(&mm->mmap_sem);
 227
 228         return err;
 229 }
 230