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 void 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
  28         if (pte_present(pte)) {
  29                 struct page *page;
  30
  31                 flush_cache_page(vma, addr, pte_pfn(pte));
  32                 pte = ptep_clear_flush(vma, addr, ptep);
  33                 page = vm_normal_page(vma, addr, pte);
  34                 if (page) {
  35                         if (pte_dirty(pte))
  36                                 set_page_dirty(page);
  37                         page_remove_rmap(page, vma);
  38                         page_cache_release(page);
  39                         update_hiwater_rss(mm);
  40                         dec_mm_counter(mm, file_rss);
  41                 }
  42         } else {
  43                 if (!pte_file(pte))
  44                         free_swap_and_cache(pte_to_swp_entry(pte));
  45                 pte_clear_not_present_full(mm, addr, ptep, 0);
  46         }
  47 }
  48
  49 /*
  50  * Install a file pte to a given virtual memory address, release any
  51  * previously existing mapping.
  52  */
  53 static int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  54                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
  55 {
  56         int err = -ENOMEM;
  57         pte_t *pte;
  58         spinlock_t *ptl;
  59
  60         pte = get_locked_pte(mm, addr, &ptl);
  61         if (!pte)
  62                 goto out;
  63
  64         if (!pte_none(*pte))
  65                 zap_pte(mm, vma, addr, pte);
  66
  67         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
  68         /*
  69          * We don't need to run update_mmu_cache() here because the "file pte"
  70          * being installed by install_file_pte() is not a real pte - it's a
  71          * non-present entry (like a swap entry), noting what file offset should
  72          * be mapped there when there's a fault (in a non-linear vma where
  73          * that's not obvious).
  74          */
  75         pte_unmap_unlock(pte, ptl);
  76         err = 0;
  77 out:
  78         return err;
  79 }
  80
  81 static int populate_range(struct mm_struct *mm, struct vm_area_struct *vma,
  82                         unsigned long addr, unsigned long size, pgoff_t pgoff)
  83 {
  84         int err;
  85
  86         do {
  87                 err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
  88                 if (err)
  89                         return err;
  90
  91                 size -= PAGE_SIZE;
  92                 addr += PAGE_SIZE;
  93                 pgoff++;
  94         } while (size);
  95
  96         return 0;
  97
  98 }
  99
 100 /***
 101  * sys_remap_file_pages - remap arbitrary pages of a shared backing store
 102  *                        file within an existing vma.
 103  * @start: start of the remapped virtual memory range
 104  * @size: size of the remapped virtual memory range
 105  * @prot: new protection bits of the range
 106  * @pgoff: to be mapped page of the backing store file
 107  * @flags: 0 or MAP_NONBLOCKED - the later will cause no IO.
 108  *
 109  * this syscall works purely via pagetables, so it's the most efficient
 110  * way to map the same (large) file into a given virtual window. Unlike
 111  * mmap()/mremap() it does not create any new vmas. The new mappings are
 112  * also safe across swapout.
 113  *
 114  * NOTE: the 'prot' parameter right now is ignored, and the vma's default
 115  * protection is used. Arbitrary protections might be implemented in the
 116  * future.
 117  */
 118 asmlinkage long sys_remap_file_pages(unsigned long start, unsigned long size,
 119         unsigned long __prot, unsigned long pgoff, unsigned long flags)
 120 {
 121         struct mm_struct *mm = current->mm;
 122         struct address_space *mapping;
 123         unsigned long end = start + size;
 124         struct vm_area_struct *vma;
 125         int err = -EINVAL;
 126         int has_write_lock = 0;
 127
 128         if (__prot)
 129                 return err;
 130         /*
 131          * Sanitize the syscall parameters:
 132          */
 133         start = start & PAGE_MASK;
 134         size = size & PAGE_MASK;
 135
 136         /* Does the address range wrap, or is the span zero-sized? */
 137         if (start + size <= start)
 138                 return err;
 139
 140         /* Can we represent this offset inside this architecture's pte's? */
 141 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 142         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 143                 return err;
 144 #endif
 145
 146         /* We need down_write() to change vma->vm_flags. */
 147         down_read(&mm->mmap_sem);
 148  retry:
 149         vma = find_vma(mm, start);
 150
 151         /*
 152          * Make sure the vma is shared, that it supports prefaulting,
 153          * and that the remapped range is valid and fully within
 154          * the single existing vma.  vm_private_data is used as a
 155          * swapout cursor in a VM_NONLINEAR vma.
 156          */
 157         if (!vma || !(vma->vm_flags & VM_SHARED))
 158                 goto out;
 159
 160         if (vma->vm_private_data && !(vma->vm_flags & VM_NONLINEAR))
 161                 goto out;
 162
 163         if (!(vma->vm_flags & VM_CAN_NONLINEAR))
 164                 goto out;
 165
 166         if (end <= start || start < vma->vm_start || end > vma->vm_end)
 167                 goto out;
 168
 169         /* Must set VM_NONLINEAR before any pages are populated. */
 170         if (!(vma->vm_flags & VM_NONLINEAR)) {
 171                 /* Don't need a nonlinear mapping, exit success */
 172                 if (pgoff == linear_page_index(vma, start)) {
 173                         err = 0;
 174                         goto out;
 175                 }
 176
 177                 if (!has_write_lock) {
 178                         up_read(&mm->mmap_sem);
 179                         down_write(&mm->mmap_sem);
 180                         has_write_lock = 1;
 181                         goto retry;
 182                 }
 183                 mapping = vma->vm_file->f_mapping;
 184                 /*
 185                  * page_mkclean doesn't work on nonlinear vmas, so if
 186                  * dirty pages need to be accounted, emulate with linear
 187                  * vmas.
 188                  */
 189                 if (mapping_cap_account_dirty(mapping)) {
 190                         unsigned long addr;
 191
 192                         flags &= MAP_NONBLOCK;
 193                         addr = mmap_region(vma->vm_file, start, size,
 194                                         flags, vma->vm_flags, pgoff, 1);
 195                         if (IS_ERR_VALUE(addr)) {
 196                                 err = addr;
 197                         } else {
 198                                 BUG_ON(addr != start);
 199                                 err = 0;
 200                         }
 201                         goto out;
 202                 }
 203                 spin_lock(&mapping->i_mmap_lock);
 204                 flush_dcache_mmap_lock(mapping);
 205                 vma->vm_flags |= VM_NONLINEAR;
 206                 vma_prio_tree_remove(vma, &mapping->i_mmap);
 207                 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 208                 flush_dcache_mmap_unlock(mapping);
 209                 spin_unlock(&mapping->i_mmap_lock);
 210         }
 211
 212         err = populate_range(mm, vma, start, size, pgoff);
 213         if (!err && !(flags & MAP_NONBLOCK)) {
 214                 if (unlikely(has_write_lock)) {
 215                         downgrade_write(&mm->mmap_sem);
 216                         has_write_lock = 0;
 217                 }
 218                 make_pages_present(start, start+size);
 219         }
 220
 221         /*
 222          * We can't clear VM_NONLINEAR because we'd have to do
 223          * it after ->populate completes, and that would prevent
 224          * downgrading the lock.  (Locks can't be upgraded).
 225          */
 226
 227 out:
 228         if (likely(!has_write_lock))
 229                 up_read(&mm->mmap_sem);
 230         else
 231                 up_write(&mm->mmap_sem);
 232
 233         return err;
 234 }
 235