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         /*
 116          * Page cache readahead assumes page cache pages are order-0 which
 117          * is not the case for hugetlbfs. Do not give a bad return value
 118          * but ignore the advice.
 119          */
 120         if (vma->vm_flags & VM_HUGETLB)
 121                 return 0;
 122
 123         if (file->f_mapping->a_ops->get_xip_mem) {
 124                 /* no bad return value, but ignore advice */
 125                 return 0;
 126         }
 127
 128         *prev = vma;
 129         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 130         if (end > vma->vm_end)
 131                 end = vma->vm_end;
 132         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 133
 134         force_page_cache_readahead(file->f_mapping,
 135                         file, start, max_sane_readahead(end - start));
 136         return 0;
 137 }
 138
 139 /*
 140  * Application no longer needs these pages.  If the pages are dirty,
 141  * it's OK to just throw them away.  The app will be more careful about
 142  * data it wants to keep.  Be sure to free swap resources too.  The
 143  * zap_page_range call sets things up for shrink_active_list to actually free
 144  * these pages later if no one else has touched them in the meantime,
 145  * although we could add these pages to a global reuse list for
 146  * shrink_active_list to pick up before reclaiming other pages.
 147  *
 148  * NB: This interface discards data rather than pushes it out to swap,
 149  * as some implementations do.  This has performance implications for
 150  * applications like large transactional databases which want to discard
 151  * pages in anonymous maps after committing to backing store the data
 152  * that was kept in them.  There is no reason to write this data out to
 153  * the swap area if the application is discarding it.
 154  *
 155  * An interface that causes the system to free clean pages and flush
 156  * dirty pages is already available as msync(MS_INVALIDATE).
 157  */
 158 static long madvise_dontneed(struct vm_area_struct * vma,
 159                              struct vm_area_struct ** prev,
 160                              unsigned long start, unsigned long end)
 161 {
 162         *prev = vma;
 163         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 164                 return -EINVAL;
 165
 166         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 167                 struct zap_details details = {
 168                         .nonlinear_vma = vma,
 169                         .last_index = ULONG_MAX,
 170                 };
 171                 zap_page_range(vma, start, end - start, &details);
 172         } else
 173                 zap_page_range(vma, start, end - start, NULL);
 174         return 0;
 175 }
 176
 177 /*
 178  * Application wants to free up the pages and associated backing store.
 179  * This is effectively punching a hole into the middle of a file.
 180  *
 181  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 182  * Other filesystems return -ENOSYS.
 183  */
 184 static long madvise_remove(struct vm_area_struct *vma,
 185                                 struct vm_area_struct **prev,
 186                                 unsigned long start, unsigned long end)
 187 {
 188         struct address_space *mapping;
 189         loff_t offset, endoff;
 190         int error;
 191
 192         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 193
 194         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 195                 return -EINVAL;
 196
 197         if (!vma->vm_file || !vma->vm_file->f_mapping
 198                 || !vma->vm_file->f_mapping->host) {
 199                         return -EINVAL;
 200         }
 201
 202         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 203                 return -EACCES;
 204
 205         mapping = vma->vm_file->f_mapping;
 206
 207         offset = (loff_t)(start - vma->vm_start)
 208                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 209         endoff = (loff_t)(end - vma->vm_start - 1)
 210                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 211
 212         /* vmtruncate_range needs to take i_mutex and i_alloc_sem */
 213         up_read(&current->mm->mmap_sem);
 214         error = vmtruncate_range(mapping->host, offset, endoff);
 215         down_read(&current->mm->mmap_sem);
 216         return error;
 217 }
 218
 219 static long
 220 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 221                 unsigned long start, unsigned long end, int behavior)
 222 {
 223         long error;
 224
 225         switch (behavior) {
 226         case MADV_DOFORK:
 227                 if (vma->vm_flags & VM_IO) {
 228                         error = -EINVAL;
 229                         break;
 230                 }
 231         case MADV_DONTFORK:
 232         case MADV_NORMAL:
 233         case MADV_SEQUENTIAL:
 234         case MADV_RANDOM:
 235                 error = madvise_behavior(vma, prev, start, end, behavior);
 236                 break;
 237         case MADV_REMOVE:
 238                 error = madvise_remove(vma, prev, start, end);
 239                 break;
 240
 241         case MADV_WILLNEED:
 242                 error = madvise_willneed(vma, prev, start, end);
 243                 break;
 244
 245         case MADV_DONTNEED:
 246                 error = madvise_dontneed(vma, prev, start, end);
 247                 break;
 248
 249         default:
 250                 error = -EINVAL;
 251                 break;
 252         }
 253         return error;
 254 }
 255
 256 /*
 257  * The madvise(2) system call.
 258  *
 259  * Applications can use madvise() to advise the kernel how it should
 260  * handle paging I/O in this VM area.  The idea is to help the kernel
 261  * use appropriate read-ahead and caching techniques.  The information
 262  * provided is advisory only, and can be safely disregarded by the
 263  * kernel without affecting the correct operation of the application.
 264  *
 265  * behavior values:
 266  *  MADV_NORMAL - the default behavior is to read clusters.  This
 267  *              results in some read-ahead and read-behind.
 268  *  MADV_RANDOM - the system should read the minimum amount of data
 269  *              on any access, since it is unlikely that the appli-
 270  *              cation will need more than what it asks for.
 271  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 272  *              once, so they can be aggressively read ahead, and
 273  *              can be freed soon after they are accessed.
 274  *  MADV_WILLNEED - the application is notifying the system to read
 275  *              some pages ahead.
 276  *  MADV_DONTNEED - the application is finished with the given range,
 277  *              so the kernel can free resources associated with it.
 278  *  MADV_REMOVE - the application wants to free up the given range of
 279  *              pages and associated backing store.
 280  *
 281  * return values:
 282  *  zero    - success
 283  *  -EINVAL - start + len < 0, start is not page-aligned,
 284  *              "behavior" is not a valid value, or application
 285  *              is attempting to release locked or shared pages.
 286  *  -ENOMEM - addresses in the specified range are not currently
 287  *              mapped, or are outside the AS of the process.
 288  *  -EIO    - an I/O error occurred while paging in data.
 289  *  -EBADF  - map exists, but area maps something that isn't a file.
 290  *  -EAGAIN - a kernel resource was temporarily unavailable.
 291  */
 292 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 293 {
 294         unsigned long end, tmp;
 295         struct vm_area_struct * vma, *prev;
 296         int unmapped_error = 0;
 297         int error = -EINVAL;
 298         int write;
 299         size_t len;
 300
 301         write = madvise_need_mmap_write(behavior);
 302         if (write)
 303                 down_write(&current->mm->mmap_sem);
 304         else
 305                 down_read(&current->mm->mmap_sem);
 306
 307         if (start & ~PAGE_MASK)
 308                 goto out;
 309         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 310
 311         /* Check to see whether len was rounded up from small -ve to zero */
 312         if (len_in && !len)
 313                 goto out;
 314
 315         end = start + len;
 316         if (end < start)
 317                 goto out;
 318
 319         error = 0;
 320         if (end == start)
 321                 goto out;
 322
 323         /*
 324          * If the interval [start,end) covers some unmapped address
 325          * ranges, just ignore them, but return -ENOMEM at the end.
 326          * - different from the way of handling in mlock etc.
 327          */
 328         vma = find_vma_prev(current->mm, start, &prev);
 329         if (vma && start > vma->vm_start)
 330                 prev = vma;
 331
 332         for (;;) {
 333                 /* Still start < end. */
 334                 error = -ENOMEM;
 335                 if (!vma)
 336                         goto out;
 337
 338                 /* Here start < (end|vma->vm_end). */
 339                 if (start < vma->vm_start) {
 340                         unmapped_error = -ENOMEM;
 341                         start = vma->vm_start;
 342                         if (start >= end)
 343                                 goto out;
 344                 }
 345
 346                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 347                 tmp = vma->vm_end;
 348                 if (end < tmp)
 349                         tmp = end;
 350
 351                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 352                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 353                 if (error)
 354                         goto out;
 355                 start = tmp;
 356                 if (prev && start < prev->vm_end)
 357                         start = prev->vm_end;
 358                 error = unmapped_error;
 359                 if (start >= end)
 360                         goto out;
 361                 if (prev)
 362                         vma = prev->vm_next;
 363                 else    /* madvise_remove dropped mmap_sem */
 364                         vma = find_vma(current->mm, start);
 365         }
 366 out:
 367         if (write)
 368                 up_write(&current->mm->mmap_sem);
 369         else
 370                 up_read(&current->mm->mmap_sem);
 371
 372         return error;
 373 }