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