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 #include <linux/sched.h>
  14
  15 /*
  16  * Any behaviour which results in changes to the vma->vm_flags needs to
  17  * take mmap_sem for writing. Others, which simply traverse vmas, need
  18  * to only take it for reading.
  19  */
  20 static int madvise_need_mmap_write(int behavior)
  21 {
  22         switch (behavior) {
  23         case MADV_REMOVE:
  24         case MADV_WILLNEED:
  25         case MADV_DONTNEED:
  26                 return 0;
  27         default:
  28                 /* be safe, default to 1. list exceptions explicitly */
  29                 return 1;
  30         }
  31 }
  32
  33 /*
  34  * We can potentially split a vm area into separate
  35  * areas, each area with its own behavior.
  36  */
  37 static long madvise_behavior(struct vm_area_struct * vma,
  38                      struct vm_area_struct **prev,
  39                      unsigned long start, unsigned long end, int behavior)
  40 {
  41         struct mm_struct * mm = vma->vm_mm;
  42         int error = 0;
  43         pgoff_t pgoff;
  44         int new_flags = vma->vm_flags;
  45
  46         switch (behavior) {
  47         case MADV_NORMAL:
  48                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  49                 break;
  50         case MADV_SEQUENTIAL:
  51                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  52                 break;
  53         case MADV_RANDOM:
  54                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  55                 break;
  56         case MADV_DONTFORK:
  57                 new_flags |= VM_DONTCOPY;
  58                 break;
  59         case MADV_DOFORK:
  60                 new_flags &= ~VM_DONTCOPY;
  61                 break;
  62         }
  63
  64         if (new_flags == vma->vm_flags) {
  65                 *prev = vma;
  66                 goto out;
  67         }
  68
  69         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
  70         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
  71                                 vma->vm_file, pgoff, vma_policy(vma));
  72         if (*prev) {
  73                 vma = *prev;
  74                 goto success;
  75         }
  76
  77         *prev = vma;
  78
  79         if (start != vma->vm_start) {
  80                 error = split_vma(mm, vma, start, 1);
  81                 if (error)
  82                         goto out;
  83         }
  84
  85         if (end != vma->vm_end) {
  86                 error = split_vma(mm, vma, end, 0);
  87                 if (error)
  88                         goto out;
  89         }
  90
  91 success:
  92         /*
  93          * vm_flags is protected by the mmap_sem held in write mode.
  94          */
  95         vma->vm_flags = new_flags;
  96
  97 out:
  98         if (error == -ENOMEM)
  99                 error = -EAGAIN;
 100         return error;
 101 }
 102
 103 /*
 104  * Schedule all required I/O operations.  Do not wait for completion.
 105  */
 106 static long madvise_willneed(struct vm_area_struct * vma,
 107                              struct vm_area_struct ** prev,
 108                              unsigned long start, unsigned long end)
 109 {
 110         struct file *file = vma->vm_file;
 111
 112         if (!file)
 113                 return -EBADF;
 114
 115         if (file->f_mapping->a_ops->get_xip_mem) {
 116                 /* no bad return value, but ignore advice */
 117                 return 0;
 118         }
 119
 120         *prev = vma;
 121         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 122         if (end > vma->vm_end)
 123                 end = vma->vm_end;
 124         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 125
 126         force_page_cache_readahead(file->f_mapping,
 127                         file, start, max_sane_readahead(end - start));
 128         return 0;
 129 }
 130
 131 /*
 132  * Application no longer needs these pages.  If the pages are dirty,
 133  * it's OK to just throw them away.  The app will be more careful about
 134  * data it wants to keep.  Be sure to free swap resources too.  The
 135  * zap_page_range call sets things up for shrink_active_list to actually free
 136  * these pages later if no one else has touched them in the meantime,
 137  * although we could add these pages to a global reuse list for
 138  * shrink_active_list to pick up before reclaiming other pages.
 139  *
 140  * NB: This interface discards data rather than pushes it out to swap,
 141  * as some implementations do.  This has performance implications for
 142  * applications like large transactional databases which want to discard
 143  * pages in anonymous maps after committing to backing store the data
 144  * that was kept in them.  There is no reason to write this data out to
 145  * the swap area if the application is discarding it.
 146  *
 147  * An interface that causes the system to free clean pages and flush
 148  * dirty pages is already available as msync(MS_INVALIDATE).
 149  */
 150 static long madvise_dontneed(struct vm_area_struct * vma,
 151                              struct vm_area_struct ** prev,
 152                              unsigned long start, unsigned long end)
 153 {
 154         *prev = vma;
 155         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 156                 return -EINVAL;
 157
 158         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 159                 struct zap_details details = {
 160                         .nonlinear_vma = vma,
 161                         .last_index = ULONG_MAX,
 162                 };
 163                 zap_page_range(vma, start, end - start, &details);
 164         } else
 165                 zap_page_range(vma, start, end - start, NULL);
 166         return 0;
 167 }
 168
 169 /*
 170  * Application wants to free up the pages and associated backing store.
 171  * This is effectively punching a hole into the middle of a file.
 172  *
 173  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 174  * Other filesystems return -ENOSYS.
 175  */
 176 static long madvise_remove(struct vm_area_struct *vma,
 177                                 struct vm_area_struct **prev,
 178                                 unsigned long start, unsigned long end)
 179 {
 180         struct address_space *mapping;
 181         loff_t offset, endoff;
 182         int error;
 183
 184         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 185
 186         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 187                 return -EINVAL;
 188
 189         if (!vma->vm_file || !vma->vm_file->f_mapping
 190                 || !vma->vm_file->f_mapping->host) {
 191                         return -EINVAL;
 192         }
 193
 194         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 195                 return -EACCES;
 196
 197         mapping = vma->vm_file->f_mapping;
 198
 199         offset = (loff_t)(start - vma->vm_start)
 200                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 201         endoff = (loff_t)(end - vma->vm_start - 1)
 202                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 203
 204         /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
 205         up_read(&current->mm->mmap_sem);
 206         error = vmtruncate_range(mapping->host, offset, endoff);
 207         down_read(&current->mm->mmap_sem);
 208         return error;
 209 }
 210
 211 static long
 212 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 213                 unsigned long start, unsigned long end, int behavior)
 214 {
 215         long error;
 216
 217         switch (behavior) {
 218         case MADV_DOFORK:
 219                 if (vma->vm_flags & VM_IO) {
 220                         error = -EINVAL;
 221                         break;
 222                 }
 223         case MADV_DONTFORK:
 224         case MADV_NORMAL:
 225         case MADV_SEQUENTIAL:
 226         case MADV_RANDOM:
 227                 error = madvise_behavior(vma, prev, start, end, behavior);
 228                 break;
 229         case MADV_REMOVE:
 230                 error = madvise_remove(vma, prev, start, end);
 231                 break;
 232
 233         case MADV_WILLNEED:
 234                 error = madvise_willneed(vma, prev, start, end);
 235                 break;
 236
 237         case MADV_DONTNEED:
 238                 error = madvise_dontneed(vma, prev, start, end);
 239                 break;
 240
 241         default:
 242                 error = -EINVAL;
 243                 break;
 244         }
 245         return error;
 246 }
 247
 248 /*
 249  * The madvise(2) system call.
 250  *
 251  * Applications can use madvise() to advise the kernel how it should
 252  * handle paging I/O in this VM area.  The idea is to help the kernel
 253  * use appropriate read-ahead and caching techniques.  The information
 254  * provided is advisory only, and can be safely disregarded by the
 255  * kernel without affecting the correct operation of the application.
 256  *
 257  * behavior values:
 258  *  MADV_NORMAL - the default behavior is to read clusters.  This
 259  *              results in some read-ahead and read-behind.
 260  *  MADV_RANDOM - the system should read the minimum amount of data
 261  *              on any access, since it is unlikely that the appli-
 262  *              cation will need more than what it asks for.
 263  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 264  *              once, so they can be aggressively read ahead, and
 265  *              can be freed soon after they are accessed.
 266  *  MADV_WILLNEED - the application is notifying the system to read
 267  *              some pages ahead.
 268  *  MADV_DONTNEED - the application is finished with the given range,
 269  *              so the kernel can free resources associated with it.
 270  *  MADV_REMOVE - the application wants to free up the given range of
 271  *              pages and associated backing store.
 272  *
 273  * return values:
 274  *  zero    - success
 275  *  -EINVAL - start + len < 0, start is not page-aligned,
 276  *              "behavior" is not a valid value, or application
 277  *              is attempting to release locked or shared pages.
 278  *  -ENOMEM - addresses in the specified range are not currently
 279  *              mapped, or are outside the AS of the process.
 280  *  -EIO    - an I/O error occurred while paging in data.
 281  *  -EBADF  - map exists, but area maps something that isn't a file.
 282  *  -EAGAIN - a kernel resource was temporarily unavailable.
 283  */
 284 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 285 {
 286         unsigned long end, tmp;
 287         struct vm_area_struct * vma, *prev;
 288         int unmapped_error = 0;
 289         int error = -EINVAL;
 290         int write;
 291         size_t len;
 292
 293         write = madvise_need_mmap_write(behavior);
 294         if (write)
 295                 down_write(&current->mm->mmap_sem);
 296         else
 297                 down_read(&current->mm->mmap_sem);
 298
 299         if (start & ~PAGE_MASK)
 300                 goto out;
 301         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 302
 303         /* Check to see whether len was rounded up from small -ve to zero */
 304         if (len_in && !len)
 305                 goto out;
 306
 307         end = start + len;
 308         if (end < start)
 309                 goto out;
 310
 311         error = 0;
 312         if (end == start)
 313                 goto out;
 314
 315         /*
 316          * If the interval [start,end) covers some unmapped address
 317          * ranges, just ignore them, but return -ENOMEM at the end.
 318          * - different from the way of handling in mlock etc.
 319          */
 320         vma = find_vma_prev(current->mm, start, &prev);
 321         if (vma && start > vma->vm_start)
 322                 prev = vma;
 323
 324         for (;;) {
 325                 /* Still start < end. */
 326                 error = -ENOMEM;
 327                 if (!vma)
 328                         goto out;
 329
 330                 /* Here start < (end|vma->vm_end). */
 331                 if (start < vma->vm_start) {
 332                         unmapped_error = -ENOMEM;
 333                         start = vma->vm_start;
 334                         if (start >= end)
 335                                 goto out;
 336                 }
 337
 338                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 339                 tmp = vma->vm_end;
 340                 if (end < tmp)
 341                         tmp = end;
 342
 343                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 344                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 345                 if (error)
 346                         goto out;
 347                 start = tmp;
 348                 if (prev && start < prev->vm_end)
 349                         start = prev->vm_end;
 350                 error = unmapped_error;
 351                 if (start >= end)
 352                         goto out;
 353                 if (prev)
 354                         vma = prev->vm_next;
 355                 else    /* madvise_remove dropped mmap_sem */
 356                         vma = find_vma(current->mm, start);
 357         }
 358 out:
 359         if (write)
 360                 up_write(&current->mm->mmap_sem);
 361         else
 362                 up_read(&current->mm->mmap_sem);
 363
 364         return error;
 365 }