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 out;
  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 success:
  66         /*
  67          * vm_flags is protected by the mmap_sem held in write mode.
  68          */
  69         vma->vm_flags = new_flags;
  70
  71 out:
  72         if (error == -ENOMEM)
  73                 error = -EAGAIN;
  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|VM_HUGETLB|VM_PFNMAP))
 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 /*
 144  * Application wants to free up the pages and associated backing store.
 145  * This is effectively punching a hole into the middle of a file.
 146  *
 147  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 148  * Other filesystems return -ENOSYS.
 149  */
 150 static long madvise_remove(struct vm_area_struct *vma,
 151                                 unsigned long start, unsigned long end)
 152 {
 153         struct address_space *mapping;
 154         loff_t offset, endoff;
 155
 156         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 157                 return -EINVAL;
 158
 159         if (!vma->vm_file || !vma->vm_file->f_mapping
 160                 || !vma->vm_file->f_mapping->host) {
 161                         return -EINVAL;
 162         }
 163
 164         mapping = vma->vm_file->f_mapping;
 165
 166         offset = (loff_t)(start - vma->vm_start)
 167                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 168         endoff = (loff_t)(end - vma->vm_start - 1)
 169                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 170         return  vmtruncate_range(mapping->host, offset, endoff);
 171 }
 172
 173 static long
 174 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 175                 unsigned long start, unsigned long end, int behavior)
 176 {
 177         long error;
 178
 179         switch (behavior) {
 180         case MADV_NORMAL:
 181         case MADV_SEQUENTIAL:
 182         case MADV_RANDOM:
 183                 error = madvise_behavior(vma, prev, start, end, behavior);
 184                 break;
 185         case MADV_REMOVE:
 186                 error = madvise_remove(vma, start, end);
 187                 break;
 188
 189         case MADV_WILLNEED:
 190                 error = madvise_willneed(vma, prev, start, end);
 191                 break;
 192
 193         case MADV_DONTNEED:
 194                 error = madvise_dontneed(vma, prev, start, end);
 195                 break;
 196
 197         default:
 198                 error = -EINVAL;
 199                 break;
 200         }
 201         return error;
 202 }
 203
 204 /*
 205  * The madvise(2) system call.
 206  *
 207  * Applications can use madvise() to advise the kernel how it should
 208  * handle paging I/O in this VM area.  The idea is to help the kernel
 209  * use appropriate read-ahead and caching techniques.  The information
 210  * provided is advisory only, and can be safely disregarded by the
 211  * kernel without affecting the correct operation of the application.
 212  *
 213  * behavior values:
 214  *  MADV_NORMAL - the default behavior is to read clusters.  This
 215  *              results in some read-ahead and read-behind.
 216  *  MADV_RANDOM - the system should read the minimum amount of data
 217  *              on any access, since it is unlikely that the appli-
 218  *              cation will need more than what it asks for.
 219  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 220  *              once, so they can be aggressively read ahead, and
 221  *              can be freed soon after they are accessed.
 222  *  MADV_WILLNEED - the application is notifying the system to read
 223  *              some pages ahead.
 224  *  MADV_DONTNEED - the application is finished with the given range,
 225  *              so the kernel can free resources associated with it.
 226  *  MADV_REMOVE - the application wants to free up the given range of
 227  *              pages and associated backing store.
 228  *
 229  * return values:
 230  *  zero    - success
 231  *  -EINVAL - start + len < 0, start is not page-aligned,
 232  *              "behavior" is not a valid value, or application
 233  *              is attempting to release locked or shared pages.
 234  *  -ENOMEM - addresses in the specified range are not currently
 235  *              mapped, or are outside the AS of the process.
 236  *  -EIO    - an I/O error occurred while paging in data.
 237  *  -EBADF  - map exists, but area maps something that isn't a file.
 238  *  -EAGAIN - a kernel resource was temporarily unavailable.
 239  */
 240 asmlinkage long sys_madvise(unsigned long start, size_t len_in, int behavior)
 241 {
 242         unsigned long end, tmp;
 243         struct vm_area_struct * vma, *prev;
 244         int unmapped_error = 0;
 245         int error = -EINVAL;
 246         size_t len;
 247
 248         down_write(&current->mm->mmap_sem);
 249
 250         if (start & ~PAGE_MASK)
 251                 goto out;
 252         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 253
 254         /* Check to see whether len was rounded up from small -ve to zero */
 255         if (len_in && !len)
 256                 goto out;
 257
 258         end = start + len;
 259         if (end < start)
 260                 goto out;
 261
 262         error = 0;
 263         if (end == start)
 264                 goto out;
 265
 266         /*
 267          * If the interval [start,end) covers some unmapped address
 268          * ranges, just ignore them, but return -ENOMEM at the end.
 269          * - different from the way of handling in mlock etc.
 270          */
 271         vma = find_vma_prev(current->mm, start, &prev);
 272         if (vma && start > vma->vm_start)
 273                 prev = vma;
 274
 275         for (;;) {
 276                 /* Still start < end. */
 277                 error = -ENOMEM;
 278                 if (!vma)
 279                         goto out;
 280
 281                 /* Here start < (end|vma->vm_end). */
 282                 if (start < vma->vm_start) {
 283                         unmapped_error = -ENOMEM;
 284                         start = vma->vm_start;
 285                         if (start >= end)
 286                                 goto out;
 287                 }
 288
 289                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 290                 tmp = vma->vm_end;
 291                 if (end < tmp)
 292                         tmp = end;
 293
 294                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 295                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 296                 if (error)
 297                         goto out;
 298                 start = tmp;
 299                 if (start < prev->vm_end)
 300                         start = prev->vm_end;
 301                 error = unmapped_error;
 302                 if (start >= end)
 303                         goto out;
 304                 vma = prev->vm_next;
 305         }
 306 out:
 307         up_write(&current->mm->mmap_sem);
 308         return error;
 309 }