mm/msync.c

   1 /*
   2  *      linux/mm/msync.c
   3  *
   4  * Copyright (C) 1994-1999  Linus Torvalds
   5  */
   6
   7 /*
   8  * The msync() system call.
   9  */
  10 #include <linux/slab.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/mm.h>
  13 #include <linux/mman.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/writeback.h>
  16 #include <linux/file.h>
  17 #include <linux/syscalls.h>
  18
  19 #include <asm/pgtable.h>
  20 #include <asm/tlbflush.h>
  21
  22 static unsigned long msync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  23                                 unsigned long addr, unsigned long end)
  24 {
  25         pte_t *pte;
  26         spinlock_t *ptl;
  27         int progress = 0;
  28         unsigned long ret = 0;
  29
  30 again:
  31         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
  32         do {
  33                 struct page *page;
  34
  35                 if (progress >= 64) {
  36                         progress = 0;
  37                         if (need_resched() || need_lockbreak(ptl))
  38                                 break;
  39                 }
  40                 progress++;
  41                 if (!pte_present(*pte))
  42                         continue;
  43                 if (!pte_maybe_dirty(*pte))
  44                         continue;
  45                 page = vm_normal_page(vma, addr, *pte);
  46                 if (!page)
  47                         continue;
  48                 if (ptep_clear_flush_dirty(vma, addr, pte) ||
  49                                 page_test_and_clear_dirty(page))
  50                         ret += set_page_dirty(page);
  51                 progress += 3;
  52         } while (pte++, addr += PAGE_SIZE, addr != end);
  53         pte_unmap_unlock(pte - 1, ptl);
  54         cond_resched();
  55         if (addr != end)
  56                 goto again;
  57         return ret;
  58 }
  59
  60 static inline unsigned long msync_pmd_range(struct vm_area_struct *vma,
  61                         pud_t *pud, unsigned long addr, unsigned long end)
  62 {
  63         pmd_t *pmd;
  64         unsigned long next;
  65         unsigned long ret = 0;
  66
  67         pmd = pmd_offset(pud, addr);
  68         do {
  69                 next = pmd_addr_end(addr, end);
  70                 if (pmd_none_or_clear_bad(pmd))
  71                         continue;
  72                 ret += msync_pte_range(vma, pmd, addr, next);
  73         } while (pmd++, addr = next, addr != end);
  74         return ret;
  75 }
  76
  77 static inline unsigned long msync_pud_range(struct vm_area_struct *vma,
  78                         pgd_t *pgd, unsigned long addr, unsigned long end)
  79 {
  80         pud_t *pud;
  81         unsigned long next;
  82         unsigned long ret = 0;
  83
  84         pud = pud_offset(pgd, addr);
  85         do {
  86                 next = pud_addr_end(addr, end);
  87                 if (pud_none_or_clear_bad(pud))
  88                         continue;
  89                 ret += msync_pmd_range(vma, pud, addr, next);
  90         } while (pud++, addr = next, addr != end);
  91         return ret;
  92 }
  93
  94 static unsigned long msync_page_range(struct vm_area_struct *vma,
  95                                 unsigned long addr, unsigned long end)
  96 {
  97         pgd_t *pgd;
  98         unsigned long next;
  99         unsigned long ret = 0;
 100
 101         /* For hugepages we can't go walking the page table normally,
 102          * but that's ok, hugetlbfs is memory based, so we don't need
 103          * to do anything more on an msync().
 104          */
 105         if (vma->vm_flags & VM_HUGETLB)
 106                 return 0;
 107
 108         BUG_ON(addr >= end);
 109         pgd = pgd_offset(vma->vm_mm, addr);
 110         flush_cache_range(vma, addr, end);
 111         do {
 112                 next = pgd_addr_end(addr, end);
 113                 if (pgd_none_or_clear_bad(pgd))
 114                         continue;
 115                 ret += msync_pud_range(vma, pgd, addr, next);
 116         } while (pgd++, addr = next, addr != end);
 117         return ret;
 118 }
 119
 120 /*
 121  * MS_SYNC syncs the entire file - including mappings.
 122  *
 123  * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 124  * marks the relevant pages dirty.  The application may now run fsync() to
 125  * write out the dirty pages and wait on the writeout and check the result.
 126  * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 127  * async writeout immediately.
 128  * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 129  * applications.
 130  */
 131 static int msync_interval(struct vm_area_struct *vma, unsigned long addr,
 132                         unsigned long end, int flags,
 133                         unsigned long *nr_pages_dirtied)
 134 {
 135         int ret = 0;
 136         struct file *file = vma->vm_file;
 137
 138         if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 139                 return -EBUSY;
 140
 141         if (file && (vma->vm_flags & VM_SHARED)) {
 142                 *nr_pages_dirtied = msync_page_range(vma, addr, end);
 143
 144                 if (flags & MS_SYNC) {
 145                         struct address_space *mapping = file->f_mapping;
 146                         int err;
 147
 148                         ret = filemap_fdatawrite(mapping);
 149                         if (file->f_op && file->f_op->fsync) {
 150                                 /*
 151                                  * We don't take i_mutex here because mmap_sem
 152                                  * is already held.
 153                                  */
 154                                 err = file->f_op->fsync(file,file->f_dentry,1);
 155                                 if (err && !ret)
 156                                         ret = err;
 157                         }
 158                         err = filemap_fdatawait(mapping);
 159                         if (!ret)
 160                                 ret = err;
 161                 }
 162         }
 163         return ret;
 164 }
 165
 166 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 167 {
 168         unsigned long end;
 169         struct vm_area_struct *vma;
 170         int unmapped_error, error = -EINVAL;
 171         int done = 0;
 172
 173         if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
 174                 goto out;
 175         if (start & ~PAGE_MASK)
 176                 goto out;
 177         if ((flags & MS_ASYNC) && (flags & MS_SYNC))
 178                 goto out;
 179         error = -ENOMEM;
 180         len = (len + ~PAGE_MASK) & PAGE_MASK;
 181         end = start + len;
 182         if (end < start)
 183                 goto out;
 184         error = 0;
 185         if (end == start)
 186                 goto out;
 187         /*
 188          * If the interval [start,end) covers some unmapped address ranges,
 189          * just ignore them, but return -ENOMEM at the end.
 190          */
 191         down_read(&current->mm->mmap_sem);
 192         if (flags & MS_SYNC)
 193                 current->flags |= PF_SYNCWRITE;
 194         vma = find_vma(current->mm, start);
 195         unmapped_error = 0;
 196         do {
 197                 unsigned long nr_pages_dirtied = 0;
 198                 struct file *file;
 199
 200                 /* Still start < end. */
 201                 error = -ENOMEM;
 202                 if (!vma)
 203                         goto out_unlock;
 204                 /* Here start < vma->vm_end. */
 205                 if (start < vma->vm_start) {
 206                         unmapped_error = -ENOMEM;
 207                         start = vma->vm_start;
 208                 }
 209                 /* Here vma->vm_start <= start < vma->vm_end. */
 210                 if (end <= vma->vm_end) {
 211                         if (start < end) {
 212                                 error = msync_interval(vma, start, end, flags,
 213                                                         &nr_pages_dirtied);
 214                                 if (error)
 215                                         goto out_unlock;
 216                         }
 217                         error = unmapped_error;
 218                         done = 1;
 219                 } else {
 220                         /* Here vma->vm_start <= start < vma->vm_end < end. */
 221                         error = msync_interval(vma, start, vma->vm_end, flags,
 222                                                 &nr_pages_dirtied);
 223                         if (error)
 224                                 goto out_unlock;
 225                 }
 226                 file = vma->vm_file;
 227                 start = vma->vm_end;
 228                 if ((flags & MS_ASYNC) && file && nr_pages_dirtied) {
 229                         get_file(file);
 230                         up_read(&current->mm->mmap_sem);
 231                         balance_dirty_pages_ratelimited_nr(file->f_mapping,
 232                                                         nr_pages_dirtied);
 233                         fput(file);
 234                         down_read(&current->mm->mmap_sem);
 235                         vma = find_vma(current->mm, start);
 236                 } else {
 237                         vma = vma->vm_next;
 238                 }
 239         } while (!done);
 240 out_unlock:
 241         current->flags &= ~PF_SYNCWRITE;
 242         up_read(&current->mm->mmap_sem);
 243 out:
 244         return error;
 245 }