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