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 #include <linux/backing-dev.h>
   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 #include <linux/mmu_notifier.h>
  19
  20 #include <asm/mmu_context.h>
  21 #include <asm/cacheflush.h>
  22 #include <asm/tlbflush.h>
  23
  24 static void zap_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  25                         unsigned long addr, pte_t *ptep)
  26 {
  27         pte_t pte = *ptep;
  28
  29         if (pte_present(pte)) {
  30                 struct page *page;
  31
  32                 flush_cache_page(vma, addr, pte_pfn(pte));
  33                 pte = ptep_clear_flush(vma, addr, ptep);
  34                 page = vm_normal_page(vma, addr, pte);
  35                 if (page) {
  36                         if (pte_dirty(pte))
  37                                 set_page_dirty(page);
  38                         page_remove_rmap(page, vma);
  39                         page_cache_release(page);
  40                         update_hiwater_rss(mm);
  41                         dec_mm_counter(mm, file_rss);
  42                 }
  43         } else {
  44                 if (!pte_file(pte))
  45                         free_swap_and_cache(pte_to_swp_entry(pte));
  46                 pte_clear_not_present_full(mm, addr, ptep, 0);
  47         }
  48 }
  49
  50 /*
  51  * Install a file pte to a given virtual memory address, release any
  52  * previously existing mapping.
  53  */
  54 static int install_file_pte(struct mm_struct *mm, struct vm_area_struct *vma,
  55                 unsigned long addr, unsigned long pgoff, pgprot_t prot)
  56 {
  57         int err = -ENOMEM;
  58         pte_t *pte;
  59         spinlock_t *ptl;
  60
  61         pte = get_locked_pte(mm, addr, &ptl);
  62         if (!pte)
  63                 goto out;
  64
  65         if (!pte_none(*pte))
  66                 zap_pte(mm, vma, addr, pte);
  67
  68         set_pte_at(mm, addr, pte, pgoff_to_pte(pgoff));
  69         /*
  70          * We don't need to run update_mmu_cache() here because the "file pte"
  71          * being installed by install_file_pte() is not a real pte - it's a
  72          * non-present entry (like a swap entry), noting what file offset should
  73          * be mapped there when there's a fault (in a non-linear vma where
  74          * that's not obvious).
  75          */
  76         pte_unmap_unlock(pte, ptl);
  77         err = 0;
  78 out:
  79         return err;
  80 }
  81
  82 static int populate_range(struct mm_struct *mm, struct vm_area_struct *vma,
  83                         unsigned long addr, unsigned long size, pgoff_t pgoff)
  84 {
  85         int err;
  86
  87         do {
  88                 err = install_file_pte(mm, vma, addr, pgoff, vma->vm_page_prot);
  89                 if (err)
  90                         return err;
  91
  92                 size -= PAGE_SIZE;
  93                 addr += PAGE_SIZE;
  94                 pgoff++;
  95         } while (size);
  96
  97         return 0;
  98
  99 }
 100
 101 /**
 102  * sys_remap_file_pages - remap arbitrary pages of an existing VM_SHARED 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 (see NOTE)
 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  * sys_remap_file_pages remaps arbitrary pages of an existing VM_SHARED vma
 110  * (shared backing store file).
 111  *
 112  * This syscall works purely via pagetables, so it's the most efficient
 113  * way to map the same (large) file into a given virtual window. Unlike
 114  * mmap()/mremap() it does not create any new vmas. The new mappings are
 115  * also safe across swapout.
 116  *
 117  * NOTE: the @prot parameter right now is ignored (but must be zero),
 118  * and the vma's default protection is used. Arbitrary protections
 119  * might be implemented in the future.
 120  */
 121 SYSCALL_DEFINE5(remap_file_pages, unsigned long, start, unsigned long, size,
 122                 unsigned long, prot, unsigned long, pgoff, unsigned long, flags)
 123 {
 124         struct mm_struct *mm = current->mm;
 125         struct address_space *mapping;
 126         unsigned long end = start + size;
 127         struct vm_area_struct *vma;
 128         int err = -EINVAL;
 129         int has_write_lock = 0;
 130
 131         if (prot)
 132                 return err;
 133         /*
 134          * Sanitize the syscall parameters:
 135          */
 136         start = start & PAGE_MASK;
 137         size = size & PAGE_MASK;
 138
 139         /* Does the address range wrap, or is the span zero-sized? */
 140         if (start + size <= start)
 141                 return err;
 142
 143         /* Can we represent this offset inside this architecture's pte's? */
 144 #if PTE_FILE_MAX_BITS < BITS_PER_LONG
 145         if (pgoff + (size >> PAGE_SHIFT) >= (1UL << PTE_FILE_MAX_BITS))
 146                 return err;
 147 #endif
 148
 149         /* We need down_write() to change vma->vm_flags. */
 150         down_read(&mm->mmap_sem);
 151  retry:
 152         vma = find_vma(mm, start);
 153
 154         /*
 155          * Make sure the vma is shared, that it supports prefaulting,
 156          * and that the remapped range is valid and fully within
 157          * the single existing vma.  vm_private_data is used as a
 158          * swapout cursor in a VM_NONLINEAR vma.
 159          */
 160         if (!vma || !(vma->vm_flags & VM_SHARED))
 161                 goto out;
 162
 163         if (vma->vm_private_data && !(vma->vm_flags & VM_NONLINEAR))
 164                 goto out;
 165
 166         if (!(vma->vm_flags & VM_CAN_NONLINEAR))
 167                 goto out;
 168
 169         if (end <= start || start < vma->vm_start || end > vma->vm_end)
 170                 goto out;
 171
 172         /* Must set VM_NONLINEAR before any pages are populated. */
 173         if (!(vma->vm_flags & VM_NONLINEAR)) {
 174                 /* Don't need a nonlinear mapping, exit success */
 175                 if (pgoff == linear_page_index(vma, start)) {
 176                         err = 0;
 177                         goto out;
 178                 }
 179
 180                 if (!has_write_lock) {
 181                         up_read(&mm->mmap_sem);
 182                         down_write(&mm->mmap_sem);
 183                         has_write_lock = 1;
 184                         goto retry;
 185                 }
 186                 mapping = vma->vm_file->f_mapping;
 187                 /*
 188                  * page_mkclean doesn't work on nonlinear vmas, so if
 189                  * dirty pages need to be accounted, emulate with linear
 190                  * vmas.
 191                  */
 192                 if (mapping_cap_account_dirty(mapping)) {
 193                         unsigned long addr;
 194                         struct file *file = vma->vm_file;
 195
 196                         flags &= MAP_NONBLOCK;
 197                         get_file(file);
 198                         addr = mmap_region(file, start, size,
 199                                         flags, vma->vm_flags, pgoff, 1);
 200                         fput(file);
 201                         if (IS_ERR_VALUE(addr)) {
 202                                 err = addr;
 203                         } else {
 204                                 BUG_ON(addr != start);
 205                                 err = 0;
 206                         }
 207                         goto out;
 208                 }
 209                 spin_lock(&mapping->i_mmap_lock);
 210                 flush_dcache_mmap_lock(mapping);
 211                 vma->vm_flags |= VM_NONLINEAR;
 212                 vma_prio_tree_remove(vma, &mapping->i_mmap);
 213                 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
 214                 flush_dcache_mmap_unlock(mapping);
 215                 spin_unlock(&mapping->i_mmap_lock);
 216         }
 217
 218         mmu_notifier_invalidate_range_start(mm, start, start + size);
 219         err = populate_range(mm, vma, start, size, pgoff);
 220         mmu_notifier_invalidate_range_end(mm, start, start + size);
 221         if (!err && !(flags & MAP_NONBLOCK)) {
 222                 if (unlikely(has_write_lock)) {
 223                         downgrade_write(&mm->mmap_sem);
 224                         has_write_lock = 0;
 225                 }
 226                 make_pages_present(start, start+size);
 227         }
 228
 229         /*
 230          * We can't clear VM_NONLINEAR because we'd have to do
 231          * it after ->populate completes, and that would prevent
 232          * downgrading the lock.  (Locks can't be upgraded).
 233          */
 234
 235 out:
 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