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
  16 #include <asm/pgtable.h>
  17 #include <asm/pgalloc.h>
  18 #include <asm/tlbflush.h>
  19
  20 /*
  21  * Called with mm->page_table_lock held to protect against other
  22  * threads/the swapper from ripping pte's out from under us.
  23  */
  24 static int filemap_sync_pte(pte_t *ptep, struct vm_area_struct *vma,
  25         unsigned long address, unsigned int flags)
  26 {
  27         pte_t pte = *ptep;
  28         unsigned long pfn = pte_pfn(pte);
  29         struct page *page;
  30
  31         if (pte_present(pte) && pfn_valid(pfn)) {
  32                 page = pfn_to_page(pfn);
  33                 if (!PageReserved(page) &&
  34                     (ptep_clear_flush_dirty(vma, address, ptep) ||
  35                      page_test_and_clear_dirty(page)))
  36                         set_page_dirty(page);
  37         }
  38         return 0;
  39 }
  40
  41 static int filemap_sync_pte_range(pmd_t * pmd,
  42         unsigned long address, unsigned long end,
  43         struct vm_area_struct *vma, unsigned int flags)
  44 {
  45         pte_t *pte;
  46         int error;
  47
  48         if (pmd_none(*pmd))
  49                 return 0;
  50         if (pmd_bad(*pmd)) {
  51                 pmd_ERROR(*pmd);
  52                 pmd_clear(pmd);
  53                 return 0;
  54         }
  55         pte = pte_offset_map(pmd, address);
  56         if ((address & PMD_MASK) != (end & PMD_MASK))
  57                 end = (address & PMD_MASK) + PMD_SIZE;
  58         error = 0;
  59         do {
  60                 error |= filemap_sync_pte(pte, vma, address, flags);
  61                 address += PAGE_SIZE;
  62                 pte++;
  63         } while (address && (address < end));
  64
  65         pte_unmap(pte - 1);
  66
  67         return error;
  68 }
  69
  70 static inline int filemap_sync_pmd_range(pgd_t * pgd,
  71         unsigned long address, unsigned long end,
  72         struct vm_area_struct *vma, unsigned int flags)
  73 {
  74         pmd_t * pmd;
  75         int error;
  76
  77         if (pgd_none(*pgd))
  78                 return 0;
  79         if (pgd_bad(*pgd)) {
  80                 pgd_ERROR(*pgd);
  81                 pgd_clear(pgd);
  82                 return 0;
  83         }
  84         pmd = pmd_offset(pgd, address);
  85         if ((address & PGDIR_MASK) != (end & PGDIR_MASK))
  86                 end = (address & PGDIR_MASK) + PGDIR_SIZE;
  87         error = 0;
  88         do {
  89                 error |= filemap_sync_pte_range(pmd, address, end, vma, flags);
  90                 address = (address + PMD_SIZE) & PMD_MASK;
  91                 pmd++;
  92         } while (address && (address < end));
  93         return error;
  94 }
  95
  96 static int filemap_sync(struct vm_area_struct * vma, unsigned long address,
  97         size_t size, unsigned int flags)
  98 {
  99         pgd_t * dir;
 100         unsigned long end = address + size;
 101         int error = 0;
 102
 103         /* Aquire the lock early; it may be possible to avoid dropping
 104          * and reaquiring it repeatedly.
 105          */
 106         spin_lock(&vma->vm_mm->page_table_lock);
 107
 108         dir = pgd_offset(vma->vm_mm, address);
 109         flush_cache_range(vma, address, end);
 110
 111         /* For hugepages we can't go walking the page table normally,
 112          * but that's ok, hugetlbfs is memory based, so we don't need
 113          * to do anything more on an msync() */
 114         if (is_vm_hugetlb_page(vma))
 115                 goto out;
 116
 117         if (address >= end)
 118                 BUG();
 119         do {
 120                 error |= filemap_sync_pmd_range(dir, address, end, vma, flags);
 121                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 122                 dir++;
 123         } while (address && (address < end));
 124         /*
 125          * Why flush ? filemap_sync_pte already flushed the tlbs with the
 126          * dirty bits.
 127          */
 128         flush_tlb_range(vma, end - size, end);
 129  out:
 130         spin_unlock(&vma->vm_mm->page_table_lock);
 131
 132         return error;
 133 }
 134
 135 /*
 136  * MS_SYNC syncs the entire file - including mappings.
 137  *
 138  * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 139  * marks the relevant pages dirty.  The application may now run fsync() to
 140  * write out the dirty pages and wait on the writeout and check the result.
 141  * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 142  * async writeout immediately.
 143  * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 144  * applications.
 145  */
 146 static int msync_interval(struct vm_area_struct * vma,
 147         unsigned long start, unsigned long end, int flags)
 148 {
 149         int ret = 0;
 150         struct file * file = vma->vm_file;
 151
 152         if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 153                 return -EBUSY;
 154
 155         if (file && (vma->vm_flags & VM_SHARED)) {
 156                 ret = filemap_sync(vma, start, end-start, flags);
 157
 158                 if (!ret && (flags & MS_SYNC)) {
 159                         struct address_space *mapping = file->f_mapping;
 160                         int err;
 161
 162                         down(&mapping->host->i_sem);
 163                         ret = filemap_fdatawrite(mapping);
 164                         if (file->f_op && file->f_op->fsync) {
 165                                 err = file->f_op->fsync(file,file->f_dentry,1);
 166                                 if (err && !ret)
 167                                         ret = err;
 168                         }
 169                         err = filemap_fdatawait(mapping);
 170                         if (!ret)
 171                                 ret = err;
 172                         up(&mapping->host->i_sem);
 173                 }
 174         }
 175         return ret;
 176 }
 177
 178 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 179 {
 180         unsigned long end;
 181         struct vm_area_struct * vma;
 182         int unmapped_error, error = -EINVAL;
 183
 184         down_read(&current->mm->mmap_sem);
 185         if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
 186                 goto out;
 187         if (start & ~PAGE_MASK)
 188                 goto out;
 189         if ((flags & MS_ASYNC) && (flags & MS_SYNC))
 190                 goto out;
 191         error = -ENOMEM;
 192         len = (len + ~PAGE_MASK) & PAGE_MASK;
 193         end = start + len;
 194         if (end < start)
 195                 goto out;
 196         error = 0;
 197         if (end == start)
 198                 goto out;
 199         /*
 200          * If the interval [start,end) covers some unmapped address ranges,
 201          * just ignore them, but return -ENOMEM at the end.
 202          */
 203         vma = find_vma(current->mm, start);
 204         unmapped_error = 0;
 205         for (;;) {
 206                 /* Still start < end. */
 207                 error = -ENOMEM;
 208                 if (!vma)
 209                         goto out;
 210                 /* Here start < vma->vm_end. */
 211                 if (start < vma->vm_start) {
 212                         unmapped_error = -ENOMEM;
 213                         start = vma->vm_start;
 214                 }
 215                 /* Here vma->vm_start <= start < vma->vm_end. */
 216                 if (end <= vma->vm_end) {
 217                         if (start < end) {
 218                                 error = msync_interval(vma, start, end, flags);
 219                                 if (error)
 220                                         goto out;
 221                         }
 222                         error = unmapped_error;
 223                         goto out;
 224                 }
 225                 /* Here vma->vm_start <= start < vma->vm_end < end. */
 226                 error = msync_interval(vma, start, vma->vm_end, flags);
 227                 if (error)
 228                         goto out;
 229                 start = vma->vm_end;
 230                 vma = vma->vm_next;
 231         }
 232 out:
 233         up_read(&current->mm->mmap_sem);
 234         return error;
 235 }