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
  25 static void sync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  26                                 unsigned long addr, unsigned long end)
  27 {
  28         pte_t *pte;
  29
  30         pte = pte_offset_map(pmd, addr);
  31         do {
  32                 unsigned long pfn;
  33                 struct page *page;
  34
  35                 if (!pte_present(*pte))
  36                         continue;
  37                 pfn = pte_pfn(*pte);
  38                 if (!pfn_valid(pfn))
  39                         continue;
  40                 page = pfn_to_page(pfn);
  41                 if (PageReserved(page))
  42                         continue;
  43
  44                 if (ptep_clear_flush_dirty(vma, addr, pte) ||
  45                     page_test_and_clear_dirty(page))
  46                         set_page_dirty(page);
  47         } while (pte++, addr += PAGE_SIZE, addr != end);
  48         pte_unmap(pte - 1);
  49 }
  50
  51 static inline void sync_pmd_range(struct vm_area_struct *vma, pud_t *pud,
  52                                 unsigned long addr, unsigned long end)
  53 {
  54         pmd_t *pmd;
  55         unsigned long next;
  56
  57         pmd = pmd_offset(pud, addr);
  58         do {
  59                 next = pmd_addr_end(addr, end);
  60                 if (pmd_none_or_clear_bad(pmd))
  61                         continue;
  62                 sync_pte_range(vma, pmd, addr, next);
  63         } while (pmd++, addr = next, addr != end);
  64 }
  65
  66 static inline void sync_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
  67                                 unsigned long addr, unsigned long end)
  68 {
  69         pud_t *pud;
  70         unsigned long next;
  71
  72         pud = pud_offset(pgd, addr);
  73         do {
  74                 next = pud_addr_end(addr, end);
  75                 if (pud_none_or_clear_bad(pud))
  76                         continue;
  77                 sync_pmd_range(vma, pud, addr, next);
  78         } while (pud++, addr = next, addr != end);
  79 }
  80
  81 static void sync_page_range(struct vm_area_struct *vma,
  82                                 unsigned long addr, unsigned long end)
  83 {
  84         struct mm_struct *mm = vma->vm_mm;
  85         pgd_t *pgd;
  86         unsigned long next;
  87
  88         /* For hugepages we can't go walking the page table normally,
  89          * but that's ok, hugetlbfs is memory based, so we don't need
  90          * to do anything more on an msync() */
  91         if (is_vm_hugetlb_page(vma))
  92                 return;
  93
  94         BUG_ON(addr >= end);
  95         pgd = pgd_offset(mm, addr);
  96         flush_cache_range(vma, addr, end);
  97         spin_lock(&mm->page_table_lock);
  98         do {
  99                 next = pgd_addr_end(addr, end);
 100                 if (pgd_none_or_clear_bad(pgd))
 101                         continue;
 102                 sync_pud_range(vma, pgd, addr, next);
 103         } while (pgd++, addr = next, addr != end);
 104         spin_unlock(&mm->page_table_lock);
 105 }
 106
 107 #ifdef CONFIG_PREEMPT
 108 static inline void filemap_sync(struct vm_area_struct *vma,
 109                                 unsigned long addr, unsigned long end)
 110 {
 111         const size_t chunk = 64 * 1024; /* bytes */
 112         unsigned long next;
 113
 114         do {
 115                 next = addr + chunk;
 116                 if (next > end || next < addr)
 117                         next = end;
 118                 sync_page_range(vma, addr, next);
 119                 cond_resched();
 120         } while (addr = next, addr != end);
 121 }
 122 #else
 123 static inline void filemap_sync(struct vm_area_struct *vma,
 124                                 unsigned long addr, unsigned long end)
 125 {
 126         sync_page_range(vma, addr, end);
 127 }
 128 #endif
 129
 130 /*
 131  * MS_SYNC syncs the entire file - including mappings.
 132  *
 133  * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 134  * marks the relevant pages dirty.  The application may now run fsync() to
 135  * write out the dirty pages and wait on the writeout and check the result.
 136  * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 137  * async writeout immediately.
 138  * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 139  * applications.
 140  */
 141 static int msync_interval(struct vm_area_struct *vma,
 142                         unsigned long addr, unsigned long end, int flags)
 143 {
 144         int ret = 0;
 145         struct file *file = vma->vm_file;
 146
 147         if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 148                 return -EBUSY;
 149
 150         if (file && (vma->vm_flags & VM_SHARED)) {
 151                 filemap_sync(vma, addr, end);
 152
 153                 if (flags & MS_SYNC) {
 154                         struct address_space *mapping = file->f_mapping;
 155                         int err;
 156
 157                         ret = filemap_fdatawrite(mapping);
 158                         if (file->f_op && file->f_op->fsync) {
 159                                 /*
 160                                  * We don't take i_sem here because mmap_sem
 161                                  * is already held.
 162                                  */
 163                                 err = file->f_op->fsync(file,file->f_dentry,1);
 164                                 if (err && !ret)
 165                                         ret = err;
 166                         }
 167                         err = filemap_fdatawait(mapping);
 168                         if (!ret)
 169                                 ret = err;
 170                 }
 171         }
 172         return ret;
 173 }
 174
 175 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 176 {
 177         unsigned long end;
 178         struct vm_area_struct *vma;
 179         int unmapped_error, error = -EINVAL;
 180
 181         if (flags & MS_SYNC)
 182                 current->flags |= PF_SYNCWRITE;
 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         current->flags &= ~PF_SYNCWRITE;
 235         return error;
 236 }