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, file, start, 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 shrink_active_list 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  * shrink_active_list 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 #ifdef CONFIG_MEMORY_FAILURE
 211 /*
 212  * Error injection support for memory error handling.
 213  */
 214 static int madvise_hwpoison(unsigned long start, unsigned long end)
 215 {
 216         int ret = 0;
 217
 218         if (!capable(CAP_SYS_ADMIN))
 219                 return -EPERM;
 220         for (; start < end; start += PAGE_SIZE) {
 221                 struct page *p;
 222                 int ret = get_user_pages(current, current->mm, start, 1,
 223                                                 0, 0, &p, NULL);
 224                 if (ret != 1)
 225                         return ret;
 226                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 227                        page_to_pfn(p), start);
 228                 /* Ignore return value for now */
 229                 __memory_failure(page_to_pfn(p), 0, 1);
 230                 put_page(p);
 231         }
 232         return ret;
 233 }
 234 #endif
 235
 236 static long
 237 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 238                 unsigned long start, unsigned long end, int behavior)
 239 {
 240         long error;
 241
 242         switch (behavior) {
 243         case MADV_DOFORK:
 244                 if (vma->vm_flags & VM_IO) {
 245                         error = -EINVAL;
 246                         break;
 247                 }
 248         case MADV_DONTFORK:
 249         case MADV_NORMAL:
 250         case MADV_SEQUENTIAL:
 251         case MADV_RANDOM:
 252                 error = madvise_behavior(vma, prev, start, end, behavior);
 253                 break;
 254         case MADV_REMOVE:
 255                 error = madvise_remove(vma, prev, start, end);
 256                 break;
 257
 258         case MADV_WILLNEED:
 259                 error = madvise_willneed(vma, prev, start, end);
 260                 break;
 261
 262         case MADV_DONTNEED:
 263                 error = madvise_dontneed(vma, prev, start, end);
 264                 break;
 265
 266         default:
 267                 BUG();
 268                 break;
 269         }
 270         return error;
 271 }
 272
 273 static int
 274 madvise_behavior_valid(int behavior)
 275 {
 276         switch (behavior) {
 277         case MADV_DOFORK:
 278         case MADV_DONTFORK:
 279         case MADV_NORMAL:
 280         case MADV_SEQUENTIAL:
 281         case MADV_RANDOM:
 282         case MADV_REMOVE:
 283         case MADV_WILLNEED:
 284         case MADV_DONTNEED:
 285                 return 1;
 286
 287         default:
 288                 return 0;
 289         }
 290 }
 291 /*
 292  * The madvise(2) system call.
 293  *
 294  * Applications can use madvise() to advise the kernel how it should
 295  * handle paging I/O in this VM area.  The idea is to help the kernel
 296  * use appropriate read-ahead and caching techniques.  The information
 297  * provided is advisory only, and can be safely disregarded by the
 298  * kernel without affecting the correct operation of the application.
 299  *
 300  * behavior values:
 301  *  MADV_NORMAL - the default behavior is to read clusters.  This
 302  *              results in some read-ahead and read-behind.
 303  *  MADV_RANDOM - the system should read the minimum amount of data
 304  *              on any access, since it is unlikely that the appli-
 305  *              cation will need more than what it asks for.
 306  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 307  *              once, so they can be aggressively read ahead, and
 308  *              can be freed soon after they are accessed.
 309  *  MADV_WILLNEED - the application is notifying the system to read
 310  *              some pages ahead.
 311  *  MADV_DONTNEED - the application is finished with the given range,
 312  *              so the kernel can free resources associated with it.
 313  *  MADV_REMOVE - the application wants to free up the given range of
 314  *              pages and associated backing store.
 315  *
 316  * return values:
 317  *  zero    - success
 318  *  -EINVAL - start + len < 0, start is not page-aligned,
 319  *              "behavior" is not a valid value, or application
 320  *              is attempting to release locked or shared pages.
 321  *  -ENOMEM - addresses in the specified range are not currently
 322  *              mapped, or are outside the AS of the process.
 323  *  -EIO    - an I/O error occurred while paging in data.
 324  *  -EBADF  - map exists, but area maps something that isn't a file.
 325  *  -EAGAIN - a kernel resource was temporarily unavailable.
 326  */
 327 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 328 {
 329         unsigned long end, tmp;
 330         struct vm_area_struct * vma, *prev;
 331         int unmapped_error = 0;
 332         int error = -EINVAL;
 333         int write;
 334         size_t len;
 335
 336 #ifdef CONFIG_MEMORY_FAILURE
 337         if (behavior == MADV_HWPOISON)
 338                 return madvise_hwpoison(start, start+len_in);
 339 #endif
 340         if (!madvise_behavior_valid(behavior))
 341                 return error;
 342
 343         write = madvise_need_mmap_write(behavior);
 344         if (write)
 345                 down_write(&current->mm->mmap_sem);
 346         else
 347                 down_read(&current->mm->mmap_sem);
 348
 349         if (start & ~PAGE_MASK)
 350                 goto out;
 351         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 352
 353         /* Check to see whether len was rounded up from small -ve to zero */
 354         if (len_in && !len)
 355                 goto out;
 356
 357         end = start + len;
 358         if (end < start)
 359                 goto out;
 360
 361         error = 0;
 362         if (end == start)
 363                 goto out;
 364
 365         /*
 366          * If the interval [start,end) covers some unmapped address
 367          * ranges, just ignore them, but return -ENOMEM at the end.
 368          * - different from the way of handling in mlock etc.
 369          */
 370         vma = find_vma_prev(current->mm, start, &prev);
 371         if (vma && start > vma->vm_start)
 372                 prev = vma;
 373
 374         for (;;) {
 375                 /* Still start < end. */
 376                 error = -ENOMEM;
 377                 if (!vma)
 378                         goto out;
 379
 380                 /* Here start < (end|vma->vm_end). */
 381                 if (start < vma->vm_start) {
 382                         unmapped_error = -ENOMEM;
 383                         start = vma->vm_start;
 384                         if (start >= end)
 385                                 goto out;
 386                 }
 387
 388                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 389                 tmp = vma->vm_end;
 390                 if (end < tmp)
 391                         tmp = end;
 392
 393                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 394                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 395                 if (error)
 396                         goto out;
 397                 start = tmp;
 398                 if (prev && start < prev->vm_end)
 399                         start = prev->vm_end;
 400                 error = unmapped_error;
 401                 if (start >= end)
 402                         goto out;
 403                 if (prev)
 404                         vma = prev->vm_next;
 405                 else    /* madvise_remove dropped mmap_sem */
 406                         vma = find_vma(current->mm, start);
 407         }
 408 out:
 409         if (write)
 410                 up_write(&current->mm->mmap_sem);
 411         else
 412                 up_read(&current->mm->mmap_sem);
 413
 414         return error;
 415 }