mm/madvise.c

   1 /*
   2  *      linux/mm/madvise.c
   3  *
   4  * Copyright (C) 1999  Linus Torvalds
   5  * Copyright (C) 2002  Christoph Hellwig
   6  */
   7
   8 #include <linux/mman.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/syscalls.h>
  11 #include <linux/mempolicy.h>
  12 #include <linux/hugetlb.h>
  13
  14 /*
  15  * We can potentially split a vm area into separate
  16  * areas, each area with its own behavior.
  17  */
  18 static long madvise_behavior(struct vm_area_struct * vma,
  19                      struct vm_area_struct **prev,
  20                      unsigned long start, unsigned long end, int behavior)
  21 {
  22         struct mm_struct * mm = vma->vm_mm;
  23         int error = 0;
  24         pgoff_t pgoff;
  25         int new_flags = vma->vm_flags & ~VM_READHINTMASK;
  26
  27         switch (behavior) {
  28         case MADV_SEQUENTIAL:
  29                 new_flags |= VM_SEQ_READ;
  30                 break;
  31         case MADV_RANDOM:
  32                 new_flags |= VM_RAND_READ;
  33                 break;
  34         default:
  35                 break;
  36         }
  37
  38         if (new_flags == vma->vm_flags) {
  39                 *prev = vma;
  40                 goto success;
  41         }
  42
  43         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  44         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  45                                 vma->vm_file, pgoff, vma_policy(vma));
  46         if (*prev) {
  47                 vma = *prev;
  48                 goto success;
  49         }
  50
  51         *prev = vma;
  52
  53         if (start != vma->vm_start) {
  54                 error = split_vma(mm, vma, start, 1);
  55                 if (error)
  56                         goto out;
  57         }
  58
  59         if (end != vma->vm_end) {
  60                 error = split_vma(mm, vma, end, 0);
  61                 if (error)
  62                         goto out;
  63         }
  64
  65         /*
  66          * vm_flags is protected by the mmap_sem held in write mode.
  67          */
  68         vma->vm_flags = new_flags;
  69
  70 out:
  71         if (error == -ENOMEM)
  72                 error = -EAGAIN;
  73 success:
  74         return error;
  75 }
  76
  77 /*
  78  * Schedule all required I/O operations.  Do not wait for completion.
  79  */
  80 static long madvise_willneed(struct vm_area_struct * vma,
  81                              struct vm_area_struct ** prev,
  82                              unsigned long start, unsigned long end)
  83 {
  84         struct file *file = vma->vm_file;
  85
  86         if (!file)
  87                 return -EBADF;
  88
  89         if (file->f_mapping->a_ops->get_xip_page) {
  90                 /* no bad return value, but ignore advice */
  91                 return 0;
  92         }
  93
  94         *prev = vma;
  95         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
  96         if (end > vma->vm_end)
  97                 end = vma->vm_end;
  98         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
  99
 100         force_page_cache_readahead(file->f_mapping,
 101                         file, start, max_sane_readahead(end - start));
 102         return 0;
 103 }
 104
 105 /*
 106  * Application no longer needs these pages.  If the pages are dirty,
 107  * it's OK to just throw them away.  The app will be more careful about
 108  * data it wants to keep.  Be sure to free swap resources too.  The
 109  * zap_page_range call sets things up for refill_inactive to actually free
 110  * these pages later if no one else has touched them in the meantime,
 111  * although we could add these pages to a global reuse list for
 112  * refill_inactive to pick up before reclaiming other pages.
 113  *
 114  * NB: This interface discards data rather than pushes it out to swap,
 115  * as some implementations do.  This has performance implications for
 116  * applications like large transactional databases which want to discard
 117  * pages in anonymous maps after committing to backing store the data
 118  * that was kept in them.  There is no reason to write this data out to
 119  * the swap area if the application is discarding it.
 120  *
 121  * An interface that causes the system to free clean pages and flush
 122  * dirty pages is already available as msync(MS_INVALIDATE).
 123  */
 124 static long madvise_dontneed(struct vm_area_struct * vma,
 125                              struct vm_area_struct ** prev,
 126                              unsigned long start, unsigned long end)
 127 {
 128         *prev = vma;
 129         if ((vma->vm_flags & VM_LOCKED) || is_vm_hugetlb_page(vma))
 130                 return -EINVAL;
 131
 132         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 133                 struct zap_details details = {
 134                         .nonlinear_vma = vma,
 135                         .last_index = ULONG_MAX,
 136                 };
 137                 zap_page_range(vma, start, end - start, &details);
 138         } else
 139                 zap_page_range(vma, start, end - start, NULL);
 140         return 0;
 141 }
 142
 143 static long madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 144                         unsigned long start, unsigned long end, int behavior)
 145 {
 146         long error = -EBADF;
 147
 148         switch (behavior) {
 149         case MADV_NORMAL:
 150         case MADV_SEQUENTIAL:
 151         case MADV_RANDOM:
 152                 error = madvise_behavior(vma, prev, start, end, behavior);
 153                 break;
 154
 155         case MADV_WILLNEED:
 156                 error = madvise_willneed(vma, prev, start, end);
 157                 break;
 158
 159         case MADV_DONTNEED:
 160                 error = madvise_dontneed(vma, prev, start, end);
 161                 break;
 162
 163         default:
 164                 error = -EINVAL;
 165                 break;
 166         }
 167
 168         return error;
 169 }
 170
 171 /*
 172  * The madvise(2) system call.
 173  *
 174  * Applications can use madvise() to advise the kernel how it should
 175  * handle paging I/O in this VM area.  The idea is to help the kernel
 176  * use appropriate read-ahead and caching techniques.  The information
 177  * provided is advisory only, and can be safely disregarded by the
 178  * kernel without affecting the correct operation of the application.
 179  *
 180  * behavior values:
 181  *  MADV_NORMAL - the default behavior is to read clusters.  This
 182  *              results in some read-ahead and read-behind.
 183  *  MADV_RANDOM - the system should read the minimum amount of data
 184  *              on any access, since it is unlikely that the appli-
 185  *              cation will need more than what it asks for.
 186  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 187  *              once, so they can be aggressively read ahead, and
 188  *              can be freed soon after they are accessed.
 189  *  MADV_WILLNEED - the application is notifying the system to read
 190  *              some pages ahead.
 191  *  MADV_DONTNEED - the application is finished with the given range,
 192  *              so the kernel can free resources associated with it.
 193  *
 194  * return values:
 195  *  zero    - success
 196  *  -EINVAL - start + len < 0, start is not page-aligned,
 197  *              "behavior" is not a valid value, or application
 198  *              is attempting to release locked or shared pages.
 199  *  -ENOMEM - addresses in the specified range are not currently
 200  *              mapped, or are outside the AS of the process.
 201  *  -EIO    - an I/O error occurred while paging in data.
 202  *  -EBADF  - map exists, but area maps something that isn't a file.
 203  *  -EAGAIN - a kernel resource was temporarily unavailable.
 204  */
 205 asmlinkage long sys_madvise(unsigned long start, size_t len_in, int behavior)
 206 {
 207         unsigned long end, tmp;
 208         struct vm_area_struct * vma, *prev;
 209         int unmapped_error = 0;
 210         int error = -EINVAL;
 211         size_t len;
 212
 213         down_write(&current->mm->mmap_sem);
 214
 215         if (start & ~PAGE_MASK)
 216                 goto out;
 217         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 218
 219         /* Check to see whether len was rounded up from small -ve to zero */
 220         if (len_in && !len)
 221                 goto out;
 222
 223         end = start + len;
 224         if (end < start)
 225                 goto out;
 226
 227         error = 0;
 228         if (end == start)
 229                 goto out;
 230
 231         /*
 232          * If the interval [start,end) covers some unmapped address
 233          * ranges, just ignore them, but return -ENOMEM at the end.
 234          * - different from the way of handling in mlock etc.
 235          */
 236         vma = find_vma_prev(current->mm, start, &prev);
 237         if (!vma && prev)
 238                 vma = prev->vm_next;
 239         for (;;) {
 240                 /* Still start < end. */
 241                 error = -ENOMEM;
 242                 if (!vma)
 243                         goto out;
 244
 245                 /* Here start < (end|vma->vm_end). */
 246                 if (start < vma->vm_start) {
 247                         unmapped_error = -ENOMEM;
 248                         start = vma->vm_start;
 249                         if (start >= end)
 250                                 goto out;
 251                 }
 252
 253                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 254                 tmp = vma->vm_end;
 255                 if (end < tmp)
 256                         tmp = end;
 257
 258                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 259                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 260                 if (error)
 261                         goto out;
 262                 start = tmp;
 263                 if (start < prev->vm_end)
 264                         start = prev->vm_end;
 265                 error = unmapped_error;
 266                 if (start >= end)
 267                         goto out;
 268                 vma = prev->vm_next;
 269         }
 270 out:
 271         up_write(&current->mm->mmap_sem);
 272         return error;
 273 }