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  * Any behaviour which results in changes to the vma->vm_flags needs to
  16  * take mmap_sem for writing. Others, which simply traverse vmas, need
  17  * to only take it for reading.
  18  */
  19 static int madvise_need_mmap_write(int behavior)
  20 {
  21         switch (behavior) {
  22         case MADV_REMOVE:
  23         case MADV_WILLNEED:
  24         case MADV_DONTNEED:
  25                 return 0;
  26         default:
  27                 /* be safe, default to 1. list exceptions explicitly */
  28                 return 1;
  29         }
  30 }
  31
  32 /*
  33  * We can potentially split a vm area into separate
  34  * areas, each area with its own behavior.
  35  */
  36 static long madvise_behavior(struct vm_area_struct * vma,
  37                      struct vm_area_struct **prev,
  38                      unsigned long start, unsigned long end, int behavior)
  39 {
  40         struct mm_struct * mm = vma->vm_mm;
  41         int error = 0;
  42         pgoff_t pgoff;
  43         int new_flags = vma->vm_flags;
  44
  45         switch (behavior) {
  46         case MADV_NORMAL:
  47                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  48                 break;
  49         case MADV_SEQUENTIAL:
  50                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  51                 break;
  52         case MADV_RANDOM:
  53                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  54                 break;
  55         case MADV_DONTFORK:
  56                 new_flags |= VM_DONTCOPY;
  57                 break;
  58         case MADV_DOFORK:
  59                 new_flags &= ~VM_DONTCOPY;
  60                 break;
  61         }
  62
  63         if (new_flags == vma->vm_flags) {
  64                 *prev = vma;
  65                 goto out;
  66         }
  67
  68         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  69         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  70                                 vma->vm_file, pgoff, vma_policy(vma));
  71         if (*prev) {
  72                 vma = *prev;
  73                 goto success;
  74         }
  75
  76         *prev = vma;
  77
  78         if (start != vma->vm_start) {
  79                 error = split_vma(mm, vma, start, 1);
  80                 if (error)
  81                         goto out;
  82         }
  83
  84         if (end != vma->vm_end) {
  85                 error = split_vma(mm, vma, end, 0);
  86                 if (error)
  87                         goto out;
  88         }
  89
  90 success:
  91         /*
  92          * vm_flags is protected by the mmap_sem held in write mode.
  93          */
  94         vma->vm_flags = new_flags;
  95
  96 out:
  97         if (error == -ENOMEM)
  98                 error = -EAGAIN;
  99         return error;
 100 }
 101
 102 /*
 103  * Schedule all required I/O operations.  Do not wait for completion.
 104  */
 105 static long madvise_willneed(struct vm_area_struct * vma,
 106                              struct vm_area_struct ** prev,
 107                              unsigned long start, unsigned long end)
 108 {
 109         struct file *file = vma->vm_file;
 110
 111         if (!file)
 112                 return -EBADF;
 113
 114         if (file->f_mapping->a_ops->get_xip_page) {
 115                 /* no bad return value, but ignore advice */
 116                 return 0;
 117         }
 118
 119         *prev = vma;
 120         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 121         if (end > vma->vm_end)
 122                 end = vma->vm_end;
 123         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 124
 125         force_page_cache_readahead(file->f_mapping,
 126                         file, start, max_sane_readahead(end - start));
 127         return 0;
 128 }
 129
 130 /*
 131  * Application no longer needs these pages.  If the pages are dirty,
 132  * it's OK to just throw them away.  The app will be more careful about
 133  * data it wants to keep.  Be sure to free swap resources too.  The
 134  * zap_page_range call sets things up for refill_inactive to actually free
 135  * these pages later if no one else has touched them in the meantime,
 136  * although we could add these pages to a global reuse list for
 137  * refill_inactive to pick up before reclaiming other pages.
 138  *
 139  * NB: This interface discards data rather than pushes it out to swap,
 140  * as some implementations do.  This has performance implications for
 141  * applications like large transactional databases which want to discard
 142  * pages in anonymous maps after committing to backing store the data
 143  * that was kept in them.  There is no reason to write this data out to
 144  * the swap area if the application is discarding it.
 145  *
 146  * An interface that causes the system to free clean pages and flush
 147  * dirty pages is already available as msync(MS_INVALIDATE).
 148  */
 149 static long madvise_dontneed(struct vm_area_struct * vma,
 150                              struct vm_area_struct ** prev,
 151                              unsigned long start, unsigned long end)
 152 {
 153         *prev = vma;
 154         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 155                 return -EINVAL;
 156
 157         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 158                 struct zap_details details = {
 159                         .nonlinear_vma = vma,
 160                         .last_index = ULONG_MAX,
 161                 };
 162                 zap_page_range(vma, start, end - start, &details);
 163         } else
 164                 zap_page_range(vma, start, end - start, NULL);
 165         return 0;
 166 }
 167
 168 /*
 169  * Application wants to free up the pages and associated backing store.
 170  * This is effectively punching a hole into the middle of a file.
 171  *
 172  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 173  * Other filesystems return -ENOSYS.
 174  */
 175 static long madvise_remove(struct vm_area_struct *vma,
 176                                 struct vm_area_struct **prev,
 177                                 unsigned long start, unsigned long end)
 178 {
 179         struct address_space *mapping;
 180         loff_t offset, endoff;
 181         int error;
 182
 183         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 184
 185         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 186                 return -EINVAL;
 187
 188         if (!vma->vm_file || !vma->vm_file->f_mapping
 189                 || !vma->vm_file->f_mapping->host) {
 190                         return -EINVAL;
 191         }
 192
 193         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 194                 return -EACCES;
 195
 196         mapping = vma->vm_file->f_mapping;
 197
 198         offset = (loff_t)(start - vma->vm_start)
 199                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 200         endoff = (loff_t)(end - vma->vm_start - 1)
 201                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 202
 203         /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
 204         up_read(&current->mm->mmap_sem);
 205         error = vmtruncate_range(mapping->host, offset, endoff);
 206         down_read(&current->mm->mmap_sem);
 207         return error;
 208 }
 209
 210 static long
 211 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 212                 unsigned long start, unsigned long end, int behavior)
 213 {
 214         long error;
 215
 216         switch (behavior) {
 217         case MADV_DOFORK:
 218                 if (vma->vm_flags & VM_IO) {
 219                         error = -EINVAL;
 220                         break;
 221                 }
 222         case MADV_DONTFORK:
 223         case MADV_NORMAL:
 224         case MADV_SEQUENTIAL:
 225         case MADV_RANDOM:
 226                 error = madvise_behavior(vma, prev, start, end, behavior);
 227                 break;
 228         case MADV_REMOVE:
 229                 error = madvise_remove(vma, prev, start, end);
 230                 break;
 231
 232         case MADV_WILLNEED:
 233                 error = madvise_willneed(vma, prev, start, end);
 234                 break;
 235
 236         case MADV_DONTNEED:
 237                 error = madvise_dontneed(vma, prev, start, end);
 238                 break;
 239
 240         default:
 241                 error = -EINVAL;
 242                 break;
 243         }
 244         return error;
 245 }
 246
 247 /*
 248  * The madvise(2) system call.
 249  *
 250  * Applications can use madvise() to advise the kernel how it should
 251  * handle paging I/O in this VM area.  The idea is to help the kernel
 252  * use appropriate read-ahead and caching techniques.  The information
 253  * provided is advisory only, and can be safely disregarded by the
 254  * kernel without affecting the correct operation of the application.
 255  *
 256  * behavior values:
 257  *  MADV_NORMAL - the default behavior is to read clusters.  This
 258  *              results in some read-ahead and read-behind.
 259  *  MADV_RANDOM - the system should read the minimum amount of data
 260  *              on any access, since it is unlikely that the appli-
 261  *              cation will need more than what it asks for.
 262  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 263  *              once, so they can be aggressively read ahead, and
 264  *              can be freed soon after they are accessed.
 265  *  MADV_WILLNEED - the application is notifying the system to read
 266  *              some pages ahead.
 267  *  MADV_DONTNEED - the application is finished with the given range,
 268  *              so the kernel can free resources associated with it.
 269  *  MADV_REMOVE - the application wants to free up the given range of
 270  *              pages and associated backing store.
 271  *
 272  * return values:
 273  *  zero    - success
 274  *  -EINVAL - start + len < 0, start is not page-aligned,
 275  *              "behavior" is not a valid value, or application
 276  *              is attempting to release locked or shared pages.
 277  *  -ENOMEM - addresses in the specified range are not currently
 278  *              mapped, or are outside the AS of the process.
 279  *  -EIO    - an I/O error occurred while paging in data.
 280  *  -EBADF  - map exists, but area maps something that isn't a file.
 281  *  -EAGAIN - a kernel resource was temporarily unavailable.
 282  */
 283 asmlinkage long sys_madvise(unsigned long start, size_t len_in, int behavior)
 284 {
 285         unsigned long end, tmp;
 286         struct vm_area_struct * vma, *prev;
 287         int unmapped_error = 0;
 288         int error = -EINVAL;
 289         size_t len;
 290
 291         if (madvise_need_mmap_write(behavior))
 292                 down_write(&current->mm->mmap_sem);
 293         else
 294                 down_read(&current->mm->mmap_sem);
 295
 296         if (start & ~PAGE_MASK)
 297                 goto out;
 298         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 299
 300         /* Check to see whether len was rounded up from small -ve to zero */
 301         if (len_in && !len)
 302                 goto out;
 303
 304         end = start + len;
 305         if (end < start)
 306                 goto out;
 307
 308         error = 0;
 309         if (end == start)
 310                 goto out;
 311
 312         /*
 313          * If the interval [start,end) covers some unmapped address
 314          * ranges, just ignore them, but return -ENOMEM at the end.
 315          * - different from the way of handling in mlock etc.
 316          */
 317         vma = find_vma_prev(current->mm, start, &prev);
 318         if (vma && start > vma->vm_start)
 319                 prev = vma;
 320
 321         for (;;) {
 322                 /* Still start < end. */
 323                 error = -ENOMEM;
 324                 if (!vma)
 325                         goto out;
 326
 327                 /* Here start < (end|vma->vm_end). */
 328                 if (start < vma->vm_start) {
 329                         unmapped_error = -ENOMEM;
 330                         start = vma->vm_start;
 331                         if (start >= end)
 332                                 goto out;
 333                 }
 334
 335                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 336                 tmp = vma->vm_end;
 337                 if (end < tmp)
 338                         tmp = end;
 339
 340                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 341                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 342                 if (error)
 343                         goto out;
 344                 start = tmp;
 345                 if (prev && start < prev->vm_end)
 346                         start = prev->vm_end;
 347                 error = unmapped_error;
 348                 if (start >= end)
 349                         goto out;
 350                 if (prev)
 351                         vma = prev->vm_next;
 352                 else    /* madvise_remove dropped mmap_sem */
 353                         vma = find_vma(current->mm, start);
 354         }
 355 out:
 356         if (madvise_need_mmap_write(behavior))
 357                 up_write(&current->mm->mmap_sem);
 358         else
 359                 up_read(&current->mm->mmap_sem);
 360
 361         return error;
 362 }