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/syscalls.h>
  16
  17 #include <asm/pgtable.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(pud_t * pud,
  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 (pud_none(*pud))
  78                 return 0;
  79         if (pud_bad(*pud)) {
  80                 pud_ERROR(*pud);
  81                 pud_clear(pud);
  82                 return 0;
  83         }
  84         pmd = pmd_offset(pud, address);
  85         if ((address & PUD_MASK) != (end & PUD_MASK))
  86                 end = (address & PUD_MASK) + PUD_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 inline int filemap_sync_pud_range(pgd_t *pgd,
  97         unsigned long address, unsigned long end,
  98         struct vm_area_struct *vma, unsigned int flags)
  99 {
 100         pud_t *pud;
 101         int error;
 102
 103         if (pgd_none(*pgd))
 104                 return 0;
 105         if (pgd_bad(*pgd)) {
 106                 pgd_ERROR(*pgd);
 107                 pgd_clear(pgd);
 108                 return 0;
 109         }
 110         pud = pud_offset(pgd, address);
 111         if ((address & PGDIR_MASK) != (end & PGDIR_MASK))
 112                 end = (address & PGDIR_MASK) + PGDIR_SIZE;
 113         error = 0;
 114         do {
 115                 error |= filemap_sync_pmd_range(pud, address, end, vma, flags);
 116                 address = (address + PUD_SIZE) & PUD_MASK;
 117                 pud++;
 118         } while (address && (address < end));
 119         return error;
 120 }
 121
 122 static int __filemap_sync(struct vm_area_struct *vma, unsigned long address,
 123                         size_t size, unsigned int flags)
 124 {
 125         pgd_t *pgd;
 126         unsigned long end = address + size;
 127         unsigned long next;
 128         int i;
 129         int error = 0;
 130
 131         /* Aquire the lock early; it may be possible to avoid dropping
 132          * and reaquiring it repeatedly.
 133          */
 134         spin_lock(&vma->vm_mm->page_table_lock);
 135
 136         pgd = pgd_offset(vma->vm_mm, address);
 137         flush_cache_range(vma, address, end);
 138
 139         /* For hugepages we can't go walking the page table normally,
 140          * but that's ok, hugetlbfs is memory based, so we don't need
 141          * to do anything more on an msync() */
 142         if (is_vm_hugetlb_page(vma))
 143                 goto out;
 144
 145         if (address >= end)
 146                 BUG();
 147         for (i = pgd_index(address); i <= pgd_index(end-1); i++) {
 148                 next = (address + PGDIR_SIZE) & PGDIR_MASK;
 149                 if (next <= address || next > end)
 150                         next = end;
 151                 error |= filemap_sync_pud_range(pgd, address, next, vma, flags);
 152                 address = next;
 153                 pgd++;
 154         }
 155         /*
 156          * Why flush ? filemap_sync_pte already flushed the tlbs with the
 157          * dirty bits.
 158          */
 159         flush_tlb_range(vma, end - size, end);
 160  out:
 161         spin_unlock(&vma->vm_mm->page_table_lock);
 162
 163         return error;
 164 }
 165
 166 #ifdef CONFIG_PREEMPT
 167 static int filemap_sync(struct vm_area_struct *vma, unsigned long address,
 168                         size_t size, unsigned int flags)
 169 {
 170         const size_t chunk = 64 * 1024; /* bytes */
 171         int error = 0;
 172
 173         while (size) {
 174                 size_t sz = min(size, chunk);
 175
 176                 error |= __filemap_sync(vma, address, sz, flags);
 177                 cond_resched();
 178                 address += sz;
 179                 size -= sz;
 180         }
 181         return error;
 182 }
 183 #else
 184 static int filemap_sync(struct vm_area_struct *vma, unsigned long address,
 185                         size_t size, unsigned int flags)
 186 {
 187         return __filemap_sync(vma, address, size, flags);
 188 }
 189 #endif
 190
 191 /*
 192  * MS_SYNC syncs the entire file - including mappings.
 193  *
 194  * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 195  * marks the relevant pages dirty.  The application may now run fsync() to
 196  * write out the dirty pages and wait on the writeout and check the result.
 197  * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 198  * async writeout immediately.
 199  * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 200  * applications.
 201  */
 202 static int msync_interval(struct vm_area_struct * vma,
 203         unsigned long start, unsigned long end, int flags)
 204 {
 205         int ret = 0;
 206         struct file * file = vma->vm_file;
 207
 208         if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 209                 return -EBUSY;
 210
 211         if (file && (vma->vm_flags & VM_SHARED)) {
 212                 ret = filemap_sync(vma, start, end-start, flags);
 213
 214                 if (!ret && (flags & MS_SYNC)) {
 215                         struct address_space *mapping = file->f_mapping;
 216                         int err;
 217
 218                         ret = filemap_fdatawrite(mapping);
 219                         if (file->f_op && file->f_op->fsync) {
 220                                 /*
 221                                  * We don't take i_sem here because mmap_sem
 222                                  * is already held.
 223                                  */
 224                                 err = file->f_op->fsync(file,file->f_dentry,1);
 225                                 if (err && !ret)
 226                                         ret = err;
 227                         }
 228                         err = filemap_fdatawait(mapping);
 229                         if (!ret)
 230                                 ret = err;
 231                 }
 232         }
 233         return ret;
 234 }
 235
 236 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 237 {
 238         unsigned long end;
 239         struct vm_area_struct * vma;
 240         int unmapped_error, error = -EINVAL;
 241
 242         if (flags & MS_SYNC)
 243                 current->flags |= PF_SYNCWRITE;
 244
 245         down_read(&current->mm->mmap_sem);
 246         if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
 247                 goto out;
 248         if (start & ~PAGE_MASK)
 249                 goto out;
 250         if ((flags & MS_ASYNC) && (flags & MS_SYNC))
 251                 goto out;
 252         error = -ENOMEM;
 253         len = (len + ~PAGE_MASK) & PAGE_MASK;
 254         end = start + len;
 255         if (end < start)
 256                 goto out;
 257         error = 0;
 258         if (end == start)
 259                 goto out;
 260         /*
 261          * If the interval [start,end) covers some unmapped address ranges,
 262          * just ignore them, but return -ENOMEM at the end.
 263          */
 264         vma = find_vma(current->mm, start);
 265         unmapped_error = 0;
 266         for (;;) {
 267                 /* Still start < end. */
 268                 error = -ENOMEM;
 269                 if (!vma)
 270                         goto out;
 271                 /* Here start < vma->vm_end. */
 272                 if (start < vma->vm_start) {
 273                         unmapped_error = -ENOMEM;
 274                         start = vma->vm_start;
 275                 }
 276                 /* Here vma->vm_start <= start < vma->vm_end. */
 277                 if (end <= vma->vm_end) {
 278                         if (start < end) {
 279                                 error = msync_interval(vma, start, end, flags);
 280                                 if (error)
 281                                         goto out;
 282                         }
 283                         error = unmapped_error;
 284                         goto out;
 285                 }
 286                 /* Here vma->vm_start <= start < vma->vm_end < end. */
 287                 error = msync_interval(vma, start, vma->vm_end, flags);
 288                 if (error)
 289                         goto out;
 290                 start = vma->vm_end;
 291                 vma = vma->vm_next;
 292         }
 293 out:
 294         up_read(&current->mm->mmap_sem);
 295         current->flags &= ~PF_SYNCWRITE;
 296         return error;
 297 }