SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 msync_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
  26                                 unsigned long addr, unsigned long end)
  27 {
  28         pte_t *pte;
  29         int progress = 0;
  30
  31 again:
  32         pte = pte_offset_map(pmd, addr);
  33         do {
  34                 unsigned long pfn;
  35                 struct page *page;
  36
  37                 if (progress >= 64) {
  38                         progress = 0;
  39                         if (need_resched() ||
  40                             need_lockbreak(&vma->vm_mm->page_table_lock))
  41                                 break;
  42                 }
  43                 progress++;
  44                 if (!pte_present(*pte))
  45                         continue;
  46                 if (!pte_maybe_dirty(*pte))
  47                         continue;
  48                 pfn = pte_pfn(*pte);
  49                 if (!pfn_valid(pfn))
  50                         continue;
  51                 page = pfn_to_page(pfn);
  52                 if (PageReserved(page))
  53                         continue;
  54
  55                 if (ptep_clear_flush_dirty(vma, addr, pte) ||
  56                     page_test_and_clear_dirty(page))
  57                         set_page_dirty(page);
  58                 progress += 3;
  59         } while (pte++, addr += PAGE_SIZE, addr != end);
  60         pte_unmap(pte - 1);
  61         cond_resched_lock(&vma->vm_mm->page_table_lock);
  62         if (addr != end)
  63                 goto again;
  64 }
  65
  66 static inline void msync_pmd_range(struct vm_area_struct *vma, pud_t *pud,
  67                                 unsigned long addr, unsigned long end)
  68 {
  69         pmd_t *pmd;
  70         unsigned long next;
  71
  72         pmd = pmd_offset(pud, addr);
  73         do {
  74                 next = pmd_addr_end(addr, end);
  75                 if (pmd_none_or_clear_bad(pmd))
  76                         continue;
  77                 msync_pte_range(vma, pmd, addr, next);
  78         } while (pmd++, addr = next, addr != end);
  79 }
  80
  81 static inline void msync_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
  82                                 unsigned long addr, unsigned long end)
  83 {
  84         pud_t *pud;
  85         unsigned long next;
  86
  87         pud = pud_offset(pgd, addr);
  88         do {
  89                 next = pud_addr_end(addr, end);
  90                 if (pud_none_or_clear_bad(pud))
  91                         continue;
  92                 msync_pmd_range(vma, pud, addr, next);
  93         } while (pud++, addr = next, addr != end);
  94 }
  95
  96 static void msync_page_range(struct vm_area_struct *vma,
  97                                 unsigned long addr, unsigned long end)
  98 {
  99         struct mm_struct *mm = vma->vm_mm;
 100         pgd_t *pgd;
 101         unsigned long next;
 102
 103         /* For hugepages we can't go walking the page table normally,
 104          * but that's ok, hugetlbfs is memory based, so we don't need
 105          * to do anything more on an msync() */
 106         if (is_vm_hugetlb_page(vma))
 107                 return;
 108
 109         BUG_ON(addr >= end);
 110         pgd = pgd_offset(mm, addr);
 111         flush_cache_range(vma, addr, end);
 112         spin_lock(&mm->page_table_lock);
 113         do {
 114                 next = pgd_addr_end(addr, end);
 115                 if (pgd_none_or_clear_bad(pgd))
 116                         continue;
 117                 msync_pud_range(vma, pgd, addr, next);
 118         } while (pgd++, addr = next, addr != end);
 119         spin_unlock(&mm->page_table_lock);
 120 }
 121
 122 /*
 123  * MS_SYNC syncs the entire file - including mappings.
 124  *
 125  * MS_ASYNC does not start I/O (it used to, up to 2.5.67).  Instead, it just
 126  * marks the relevant pages dirty.  The application may now run fsync() to
 127  * write out the dirty pages and wait on the writeout and check the result.
 128  * Or the application may run fadvise(FADV_DONTNEED) against the fd to start
 129  * async writeout immediately.
 130  * So my _not_ starting I/O in MS_ASYNC we provide complete flexibility to
 131  * applications.
 132  */
 133 static int msync_interval(struct vm_area_struct *vma,
 134                         unsigned long addr, unsigned long end, int flags)
 135 {
 136         int ret = 0;
 137         struct file *file = vma->vm_file;
 138
 139         if ((flags & MS_INVALIDATE) && (vma->vm_flags & VM_LOCKED))
 140                 return -EBUSY;
 141
 142         if (file && (vma->vm_flags & VM_SHARED)) {
 143                 msync_page_range(vma, addr, end);
 144
 145                 if (flags & MS_SYNC) {
 146                         struct address_space *mapping = file->f_mapping;
 147                         int err;
 148
 149                         ret = filemap_fdatawrite(mapping);
 150                         if (file->f_op && file->f_op->fsync) {
 151                                 /*
 152                                  * We don't take i_sem here because mmap_sem
 153                                  * is already held.
 154                                  */
 155                                 err = file->f_op->fsync(file,file->f_dentry,1);
 156                                 if (err && !ret)
 157                                         ret = err;
 158                         }
 159                         err = filemap_fdatawait(mapping);
 160                         if (!ret)
 161                                 ret = err;
 162                 }
 163         }
 164         return ret;
 165 }
 166
 167 asmlinkage long sys_msync(unsigned long start, size_t len, int flags)
 168 {
 169         unsigned long end;
 170         struct vm_area_struct *vma;
 171         int unmapped_error, error = -EINVAL;
 172
 173         if (flags & MS_SYNC)
 174                 current->flags |= PF_SYNCWRITE;
 175
 176         down_read(&current->mm->mmap_sem);
 177         if (flags & ~(MS_ASYNC | MS_INVALIDATE | MS_SYNC))
 178                 goto out;
 179         if (start & ~PAGE_MASK)
 180                 goto out;
 181         if ((flags & MS_ASYNC) && (flags & MS_SYNC))
 182                 goto out;
 183         error = -ENOMEM;
 184         len = (len + ~PAGE_MASK) & PAGE_MASK;
 185         end = start + len;
 186         if (end < start)
 187                 goto out;
 188         error = 0;
 189         if (end == start)
 190                 goto out;
 191         /*
 192          * If the interval [start,end) covers some unmapped address ranges,
 193          * just ignore them, but return -ENOMEM at the end.
 194          */
 195         vma = find_vma(current->mm, start);
 196         unmapped_error = 0;
 197         for (;;) {
 198                 /* Still start < end. */
 199                 error = -ENOMEM;
 200                 if (!vma)
 201                         goto out;
 202                 /* Here start < vma->vm_end. */
 203                 if (start < vma->vm_start) {
 204                         unmapped_error = -ENOMEM;
 205                         start = vma->vm_start;
 206                 }
 207                 /* Here vma->vm_start <= start < vma->vm_end. */
 208                 if (end <= vma->vm_end) {
 209                         if (start < end) {
 210                                 error = msync_interval(vma, start, end, flags);
 211                                 if (error)
 212                                         goto out;
 213                         }
 214                         error = unmapped_error;
 215                         goto out;
 216                 }
 217                 /* Here vma->vm_start <= start < vma->vm_end < end. */
 218                 error = msync_interval(vma, start, vma->vm_end, flags);
 219                 if (error)
 220                         goto out;
 221                 start = vma->vm_end;
 222                 vma = vma->vm_next;
 223         }
 224 out:
 225         up_read(&current->mm->mmap_sem);
 226         current->flags &= ~PF_SYNCWRITE;
 227         return error;
 228 }