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/page-isolation.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/falloc.h>
  15 #include <linux/sched.h>
  16 #include <linux/ksm.h>
  17 #include <linux/fs.h>
  18 #include <linux/file.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/swap.h>
  21 #include <linux/swapops.h>
  22
  23 /*
  24  * Any behaviour which results in changes to the vma->vm_flags needs to
  25  * take mmap_sem for writing. Others, which simply traverse vmas, need
  26  * to only take it for reading.
  27  */
  28 static int madvise_need_mmap_write(int behavior)
  29 {
  30         switch (behavior) {
  31         case MADV_REMOVE:
  32         case MADV_WILLNEED:
  33         case MADV_DONTNEED:
  34                 return 0;
  35         default:
  36                 /* be safe, default to 1. list exceptions explicitly */
  37                 return 1;
  38         }
  39 }
  40
  41 /*
  42  * We can potentially split a vm area into separate
  43  * areas, each area with its own behavior.
  44  */
  45 static long madvise_behavior(struct vm_area_struct * vma,
  46                      struct vm_area_struct **prev,
  47                      unsigned long start, unsigned long end, int behavior)
  48 {
  49         struct mm_struct * mm = vma->vm_mm;
  50         int error = 0;
  51         pgoff_t pgoff;
  52         unsigned long new_flags = vma->vm_flags;
  53
  54         switch (behavior) {
  55         case MADV_NORMAL:
  56                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  57                 break;
  58         case MADV_SEQUENTIAL:
  59                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  60                 break;
  61         case MADV_RANDOM:
  62                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  63                 break;
  64         case MADV_DONTFORK:
  65                 new_flags |= VM_DONTCOPY;
  66                 break;
  67         case MADV_DOFORK:
  68                 if (vma->vm_flags & VM_IO) {
  69                         error = -EINVAL;
  70                         goto out;
  71                 }
  72                 new_flags &= ~VM_DONTCOPY;
  73                 break;
  74         case MADV_DONTDUMP:
  75                 new_flags |= VM_DONTDUMP;
  76                 break;
  77         case MADV_DODUMP:
  78                 if (new_flags & VM_SPECIAL) {
  79                         error = -EINVAL;
  80                         goto out;
  81                 }
  82                 new_flags &= ~VM_DONTDUMP;
  83                 break;
  84         case MADV_MERGEABLE:
  85         case MADV_UNMERGEABLE:
  86                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  87                 if (error)
  88                         goto out;
  89                 break;
  90         case MADV_HUGEPAGE:
  91         case MADV_NOHUGEPAGE:
  92                 error = hugepage_madvise(vma, &new_flags, behavior);
  93                 if (error)
  94                         goto out;
  95                 break;
  96         }
  97
  98         if (new_flags == vma->vm_flags) {
  99                 *prev = vma;
 100                 goto out;
 101         }
 102
 103         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 104         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 105                                 vma->vm_file, pgoff, vma_policy(vma));
 106         if (*prev) {
 107                 vma = *prev;
 108                 goto success;
 109         }
 110
 111         *prev = vma;
 112
 113         if (start != vma->vm_start) {
 114                 error = split_vma(mm, vma, start, 1);
 115                 if (error)
 116                         goto out;
 117         }
 118
 119         if (end != vma->vm_end) {
 120                 error = split_vma(mm, vma, end, 0);
 121                 if (error)
 122                         goto out;
 123         }
 124
 125 success:
 126         /*
 127          * vm_flags is protected by the mmap_sem held in write mode.
 128          */
 129         vma->vm_flags = new_flags;
 130
 131 out:
 132         if (error == -ENOMEM)
 133                 error = -EAGAIN;
 134         return error;
 135 }
 136
 137 #ifdef CONFIG_SWAP
 138 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 139         unsigned long end, struct mm_walk *walk)
 140 {
 141         pte_t *orig_pte;
 142         struct vm_area_struct *vma = walk->private;
 143         unsigned long index;
 144
 145         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 146                 return 0;
 147
 148         for (index = start; index != end; index += PAGE_SIZE) {
 149                 pte_t pte;
 150                 swp_entry_t entry;
 151                 struct page *page;
 152                 spinlock_t *ptl;
 153
 154                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 155                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 156                 pte_unmap_unlock(orig_pte, ptl);
 157
 158                 if (pte_present(pte) || pte_none(pte) || pte_file(pte))
 159                         continue;
 160                 entry = pte_to_swp_entry(pte);
 161                 if (unlikely(non_swap_entry(entry)))
 162                         continue;
 163
 164                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 165                                                                 vma, index);
 166                 if (page)
 167                         page_cache_release(page);
 168         }
 169
 170         return 0;
 171 }
 172
 173 static void force_swapin_readahead(struct vm_area_struct *vma,
 174                 unsigned long start, unsigned long end)
 175 {
 176         struct mm_walk walk = {
 177                 .mm = vma->vm_mm,
 178                 .pmd_entry = swapin_walk_pmd_entry,
 179                 .private = vma,
 180         };
 181
 182         walk_page_range(start, end, &walk);
 183
 184         lru_add_drain();        /* Push any new pages onto the LRU now */
 185 }
 186
 187 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 188                 unsigned long start, unsigned long end,
 189                 struct address_space *mapping)
 190 {
 191         pgoff_t index;
 192         struct page *page;
 193         swp_entry_t swap;
 194
 195         for (; start < end; start += PAGE_SIZE) {
 196                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 197
 198                 page = find_get_page(mapping, index);
 199                 if (!radix_tree_exceptional_entry(page)) {
 200                         if (page)
 201                                 page_cache_release(page);
 202                         continue;
 203                 }
 204                 swap = radix_to_swp_entry(page);
 205                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 206                                                                 NULL, 0);
 207                 if (page)
 208                         page_cache_release(page);
 209         }
 210
 211         lru_add_drain();        /* Push any new pages onto the LRU now */
 212 }
 213 #endif          /* CONFIG_SWAP */
 214
 215 /*
 216  * Schedule all required I/O operations.  Do not wait for completion.
 217  */
 218 static long madvise_willneed(struct vm_area_struct * vma,
 219                              struct vm_area_struct ** prev,
 220                              unsigned long start, unsigned long end)
 221 {
 222         struct file *file = vma->vm_file;
 223
 224 #ifdef CONFIG_SWAP
 225         if (!file || mapping_cap_swap_backed(file->f_mapping)) {
 226                 *prev = vma;
 227                 if (!file)
 228                         force_swapin_readahead(vma, start, end);
 229                 else
 230                         force_shm_swapin_readahead(vma, start, end,
 231                                                 file->f_mapping);
 232                 return 0;
 233         }
 234 #endif
 235
 236         if (!file)
 237                 return -EBADF;
 238
 239         if (file->f_mapping->a_ops->get_xip_mem) {
 240                 /* no bad return value, but ignore advice */
 241                 return 0;
 242         }
 243
 244         *prev = vma;
 245         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 246         if (end > vma->vm_end)
 247                 end = vma->vm_end;
 248         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 249
 250         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 251         return 0;
 252 }
 253
 254 /*
 255  * Application no longer needs these pages.  If the pages are dirty,
 256  * it's OK to just throw them away.  The app will be more careful about
 257  * data it wants to keep.  Be sure to free swap resources too.  The
 258  * zap_page_range call sets things up for shrink_active_list to actually free
 259  * these pages later if no one else has touched them in the meantime,
 260  * although we could add these pages to a global reuse list for
 261  * shrink_active_list to pick up before reclaiming other pages.
 262  *
 263  * NB: This interface discards data rather than pushes it out to swap,
 264  * as some implementations do.  This has performance implications for
 265  * applications like large transactional databases which want to discard
 266  * pages in anonymous maps after committing to backing store the data
 267  * that was kept in them.  There is no reason to write this data out to
 268  * the swap area if the application is discarding it.
 269  *
 270  * An interface that causes the system to free clean pages and flush
 271  * dirty pages is already available as msync(MS_INVALIDATE).
 272  */
 273 static long madvise_dontneed(struct vm_area_struct * vma,
 274                              struct vm_area_struct ** prev,
 275                              unsigned long start, unsigned long end)
 276 {
 277         *prev = vma;
 278         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 279                 return -EINVAL;
 280
 281         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 282                 struct zap_details details = {
 283                         .nonlinear_vma = vma,
 284                         .last_index = ULONG_MAX,
 285                 };
 286                 zap_page_range(vma, start, end - start, &details);
 287         } else
 288                 zap_page_range(vma, start, end - start, NULL);
 289         return 0;
 290 }
 291
 292 /*
 293  * Application wants to free up the pages and associated backing store.
 294  * This is effectively punching a hole into the middle of a file.
 295  *
 296  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 297  * Other filesystems return -ENOSYS.
 298  */
 299 static long madvise_remove(struct vm_area_struct *vma,
 300                                 struct vm_area_struct **prev,
 301                                 unsigned long start, unsigned long end)
 302 {
 303         loff_t offset;
 304         int error;
 305         struct file *f;
 306
 307         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 308
 309         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 310                 return -EINVAL;
 311
 312         f = vma->vm_file;
 313
 314         if (!f || !f->f_mapping || !f->f_mapping->host) {
 315                         return -EINVAL;
 316         }
 317
 318         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 319                 return -EACCES;
 320
 321         offset = (loff_t)(start - vma->vm_start)
 322                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 323
 324         /*
 325          * Filesystem's fallocate may need to take i_mutex.  We need to
 326          * explicitly grab a reference because the vma (and hence the
 327          * vma's reference to the file) can go away as soon as we drop
 328          * mmap_sem.
 329          */
 330         get_file(f);
 331         up_read(&current->mm->mmap_sem);
 332         error = do_fallocate(f,
 333                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 334                                 offset, end - start);
 335         fput(f);
 336         down_read(&current->mm->mmap_sem);
 337         return error;
 338 }
 339
 340 #ifdef CONFIG_MEMORY_FAILURE
 341 /*
 342  * Error injection support for memory error handling.
 343  */
 344 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 345 {
 346         int ret = 0;
 347
 348         if (!capable(CAP_SYS_ADMIN))
 349                 return -EPERM;
 350         for (; start < end; start += PAGE_SIZE) {
 351                 struct page *p;
 352                 int ret = get_user_pages_fast(start, 1, 0, &p);
 353                 if (ret != 1)
 354                         return ret;
 355                 if (bhv == MADV_SOFT_OFFLINE) {
 356                         printk(KERN_INFO "Soft offlining page %lx at %lx\n",
 357                                 page_to_pfn(p), start);
 358                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 359                         if (ret)
 360                                 break;
 361                         continue;
 362                 }
 363                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 364                        page_to_pfn(p), start);
 365                 /* Ignore return value for now */
 366                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 367         }
 368         return ret;
 369 }
 370 #endif
 371
 372 static long
 373 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 374                 unsigned long start, unsigned long end, int behavior)
 375 {
 376         switch (behavior) {
 377         case MADV_REMOVE:
 378                 return madvise_remove(vma, prev, start, end);
 379         case MADV_WILLNEED:
 380                 return madvise_willneed(vma, prev, start, end);
 381         case MADV_DONTNEED:
 382                 return madvise_dontneed(vma, prev, start, end);
 383         default:
 384                 return madvise_behavior(vma, prev, start, end, behavior);
 385         }
 386 }
 387
 388 static int
 389 madvise_behavior_valid(int behavior)
 390 {
 391         switch (behavior) {
 392         case MADV_DOFORK:
 393         case MADV_DONTFORK:
 394         case MADV_NORMAL:
 395         case MADV_SEQUENTIAL:
 396         case MADV_RANDOM:
 397         case MADV_REMOVE:
 398         case MADV_WILLNEED:
 399         case MADV_DONTNEED:
 400 #ifdef CONFIG_KSM
 401         case MADV_MERGEABLE:
 402         case MADV_UNMERGEABLE:
 403 #endif
 404 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 405         case MADV_HUGEPAGE:
 406         case MADV_NOHUGEPAGE:
 407 #endif
 408         case MADV_DONTDUMP:
 409         case MADV_DODUMP:
 410                 return 1;
 411
 412         default:
 413                 return 0;
 414         }
 415 }
 416
 417 /*
 418  * The madvise(2) system call.
 419  *
 420  * Applications can use madvise() to advise the kernel how it should
 421  * handle paging I/O in this VM area.  The idea is to help the kernel
 422  * use appropriate read-ahead and caching techniques.  The information
 423  * provided is advisory only, and can be safely disregarded by the
 424  * kernel without affecting the correct operation of the application.
 425  *
 426  * behavior values:
 427  *  MADV_NORMAL - the default behavior is to read clusters.  This
 428  *              results in some read-ahead and read-behind.
 429  *  MADV_RANDOM - the system should read the minimum amount of data
 430  *              on any access, since it is unlikely that the appli-
 431  *              cation will need more than what it asks for.
 432  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 433  *              once, so they can be aggressively read ahead, and
 434  *              can be freed soon after they are accessed.
 435  *  MADV_WILLNEED - the application is notifying the system to read
 436  *              some pages ahead.
 437  *  MADV_DONTNEED - the application is finished with the given range,
 438  *              so the kernel can free resources associated with it.
 439  *  MADV_REMOVE - the application wants to free up the given range of
 440  *              pages and associated backing store.
 441  *  MADV_DONTFORK - omit this area from child's address space when forking:
 442  *              typically, to avoid COWing pages pinned by get_user_pages().
 443  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 444  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 445  *              this area with pages of identical content from other such areas.
 446  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 447  *
 448  * return values:
 449  *  zero    - success
 450  *  -EINVAL - start + len < 0, start is not page-aligned,
 451  *              "behavior" is not a valid value, or application
 452  *              is attempting to release locked or shared pages.
 453  *  -ENOMEM - addresses in the specified range are not currently
 454  *              mapped, or are outside the AS of the process.
 455  *  -EIO    - an I/O error occurred while paging in data.
 456  *  -EBADF  - map exists, but area maps something that isn't a file.
 457  *  -EAGAIN - a kernel resource was temporarily unavailable.
 458  */
 459 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 460 {
 461         unsigned long end, tmp;
 462         struct vm_area_struct * vma, *prev;
 463         int unmapped_error = 0;
 464         int error = -EINVAL;
 465         int write;
 466         size_t len;
 467         struct blk_plug plug;
 468
 469 #ifdef CONFIG_MEMORY_FAILURE
 470         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 471                 return madvise_hwpoison(behavior, start, start+len_in);
 472 #endif
 473         if (!madvise_behavior_valid(behavior))
 474                 return error;
 475
 476         if (start & ~PAGE_MASK)
 477                 return error;
 478         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 479
 480         /* Check to see whether len was rounded up from small -ve to zero */
 481         if (len_in && !len)
 482                 return error;
 483
 484         end = start + len;
 485         if (end < start)
 486                 return error;
 487
 488         error = 0;
 489         if (end == start)
 490                 return error;
 491
 492         write = madvise_need_mmap_write(behavior);
 493         if (write)
 494                 down_write(&current->mm->mmap_sem);
 495         else
 496                 down_read(&current->mm->mmap_sem);
 497
 498         /*
 499          * If the interval [start,end) covers some unmapped address
 500          * ranges, just ignore them, but return -ENOMEM at the end.
 501          * - different from the way of handling in mlock etc.
 502          */
 503         vma = find_vma_prev(current->mm, start, &prev);
 504         if (vma && start > vma->vm_start)
 505                 prev = vma;
 506
 507         blk_start_plug(&plug);
 508         for (;;) {
 509                 /* Still start < end. */
 510                 error = -ENOMEM;
 511                 if (!vma)
 512                         goto out;
 513
 514                 /* Here start < (end|vma->vm_end). */
 515                 if (start < vma->vm_start) {
 516                         unmapped_error = -ENOMEM;
 517                         start = vma->vm_start;
 518                         if (start >= end)
 519                                 goto out;
 520                 }
 521
 522                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 523                 tmp = vma->vm_end;
 524                 if (end < tmp)
 525                         tmp = end;
 526
 527                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 528                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 529                 if (error)
 530                         goto out;
 531                 start = tmp;
 532                 if (prev && start < prev->vm_end)
 533                         start = prev->vm_end;
 534                 error = unmapped_error;
 535                 if (start >= end)
 536                         goto out;
 537                 if (prev)
 538                         vma = prev->vm_next;
 539                 else    /* madvise_remove dropped mmap_sem */
 540                         vma = find_vma(current->mm, start);
 541         }
 542 out:
 543         blk_finish_plug(&plug);
 544         if (write)
 545                 up_write(&current->mm->mmap_sem);
 546         else
 547                 up_read(&current->mm->mmap_sem);
 548
 549         return error;
 550 }