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                                 vma_get_anon_name(vma));
 107         if (*prev) {
 108                 vma = *prev;
 109                 goto success;
 110         }
 111
 112         *prev = vma;
 113
 114         if (start != vma->vm_start) {
 115                 error = split_vma(mm, vma, start, 1);
 116                 if (error)
 117                         goto out;
 118         }
 119
 120         if (end != vma->vm_end) {
 121                 error = split_vma(mm, vma, end, 0);
 122                 if (error)
 123                         goto out;
 124         }
 125
 126 success:
 127         /*
 128          * vm_flags is protected by the mmap_sem held in write mode.
 129          */
 130         vma->vm_flags = new_flags;
 131
 132 out:
 133         if (error == -ENOMEM)
 134                 error = -EAGAIN;
 135         return error;
 136 }
 137
 138 #ifdef CONFIG_SWAP
 139 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 140         unsigned long end, struct mm_walk *walk)
 141 {
 142         pte_t *orig_pte;
 143         struct vm_area_struct *vma = walk->private;
 144         unsigned long index;
 145
 146         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 147                 return 0;
 148
 149         for (index = start; index != end; index += PAGE_SIZE) {
 150                 pte_t pte;
 151                 swp_entry_t entry;
 152                 struct page *page;
 153                 spinlock_t *ptl;
 154
 155                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 156                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 157                 pte_unmap_unlock(orig_pte, ptl);
 158
 159                 if (pte_present(pte) || pte_none(pte) || pte_file(pte))
 160                         continue;
 161                 entry = pte_to_swp_entry(pte);
 162                 if (unlikely(non_swap_entry(entry)))
 163                         continue;
 164
 165                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 166                                                                 vma, index);
 167                 if (page)
 168                         page_cache_release(page);
 169         }
 170
 171         return 0;
 172 }
 173
 174 static void force_swapin_readahead(struct vm_area_struct *vma,
 175                 unsigned long start, unsigned long end)
 176 {
 177         struct mm_walk walk = {
 178                 .mm = vma->vm_mm,
 179                 .pmd_entry = swapin_walk_pmd_entry,
 180                 .private = vma,
 181         };
 182
 183         walk_page_range(start, end, &walk);
 184
 185         lru_add_drain();        /* Push any new pages onto the LRU now */
 186 }
 187
 188 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 189                 unsigned long start, unsigned long end,
 190                 struct address_space *mapping)
 191 {
 192         pgoff_t index;
 193         struct page *page;
 194         swp_entry_t swap;
 195
 196         for (; start < end; start += PAGE_SIZE) {
 197                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 198
 199                 page = find_get_page(mapping, index);
 200                 if (!radix_tree_exceptional_entry(page)) {
 201                         if (page)
 202                                 page_cache_release(page);
 203                         continue;
 204                 }
 205                 swap = radix_to_swp_entry(page);
 206                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 207                                                                 NULL, 0);
 208                 if (page)
 209                         page_cache_release(page);
 210         }
 211
 212         lru_add_drain();        /* Push any new pages onto the LRU now */
 213 }
 214 #endif          /* CONFIG_SWAP */
 215
 216 /*
 217  * Schedule all required I/O operations.  Do not wait for completion.
 218  */
 219 static long madvise_willneed(struct vm_area_struct * vma,
 220                              struct vm_area_struct ** prev,
 221                              unsigned long start, unsigned long end)
 222 {
 223         struct file *file = vma->vm_file;
 224
 225 #ifdef CONFIG_SWAP
 226         if (!file || mapping_cap_swap_backed(file->f_mapping)) {
 227                 *prev = vma;
 228                 if (!file)
 229                         force_swapin_readahead(vma, start, end);
 230                 else
 231                         force_shm_swapin_readahead(vma, start, end,
 232                                                 file->f_mapping);
 233                 return 0;
 234         }
 235 #endif
 236
 237         if (!file)
 238                 return -EBADF;
 239
 240         if (file->f_mapping->a_ops->get_xip_mem) {
 241                 /* no bad return value, but ignore advice */
 242                 return 0;
 243         }
 244
 245         *prev = vma;
 246         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 247         if (end > vma->vm_end)
 248                 end = vma->vm_end;
 249         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 250
 251         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 252         return 0;
 253 }
 254
 255 /*
 256  * Application no longer needs these pages.  If the pages are dirty,
 257  * it's OK to just throw them away.  The app will be more careful about
 258  * data it wants to keep.  Be sure to free swap resources too.  The
 259  * zap_page_range call sets things up for shrink_active_list to actually free
 260  * these pages later if no one else has touched them in the meantime,
 261  * although we could add these pages to a global reuse list for
 262  * shrink_active_list to pick up before reclaiming other pages.
 263  *
 264  * NB: This interface discards data rather than pushes it out to swap,
 265  * as some implementations do.  This has performance implications for
 266  * applications like large transactional databases which want to discard
 267  * pages in anonymous maps after committing to backing store the data
 268  * that was kept in them.  There is no reason to write this data out to
 269  * the swap area if the application is discarding it.
 270  *
 271  * An interface that causes the system to free clean pages and flush
 272  * dirty pages is already available as msync(MS_INVALIDATE).
 273  */
 274 static long madvise_dontneed(struct vm_area_struct * vma,
 275                              struct vm_area_struct ** prev,
 276                              unsigned long start, unsigned long end)
 277 {
 278         *prev = vma;
 279         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 280                 return -EINVAL;
 281
 282         if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
 283                 struct zap_details details = {
 284                         .nonlinear_vma = vma,
 285                         .last_index = ULONG_MAX,
 286                 };
 287                 zap_page_range(vma, start, end - start, &details);
 288         } else
 289                 zap_page_range(vma, start, end - start, NULL);
 290         return 0;
 291 }
 292
 293 /*
 294  * Application wants to free up the pages and associated backing store.
 295  * This is effectively punching a hole into the middle of a file.
 296  *
 297  * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
 298  * Other filesystems return -ENOSYS.
 299  */
 300 static long madvise_remove(struct vm_area_struct *vma,
 301                                 struct vm_area_struct **prev,
 302                                 unsigned long start, unsigned long end)
 303 {
 304         loff_t offset;
 305         int error;
 306         struct file *f;
 307
 308         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 309
 310         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 311                 return -EINVAL;
 312
 313         f = vma->vm_file;
 314
 315         if (!f || !f->f_mapping || !f->f_mapping->host) {
 316                         return -EINVAL;
 317         }
 318
 319         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 320                 return -EACCES;
 321
 322         offset = (loff_t)(start - vma->vm_start)
 323                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 324
 325         /*
 326          * Filesystem's fallocate may need to take i_mutex.  We need to
 327          * explicitly grab a reference because the vma (and hence the
 328          * vma's reference to the file) can go away as soon as we drop
 329          * mmap_sem.
 330          */
 331         get_file(f);
 332         up_read(&current->mm->mmap_sem);
 333         error = do_fallocate(f,
 334                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 335                                 offset, end - start);
 336         fput(f);
 337         down_read(&current->mm->mmap_sem);
 338         return error;
 339 }
 340
 341 #ifdef CONFIG_MEMORY_FAILURE
 342 /*
 343  * Error injection support for memory error handling.
 344  */
 345 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 346 {
 347         int ret = 0;
 348
 349         if (!capable(CAP_SYS_ADMIN))
 350                 return -EPERM;
 351         for (; start < end; start += PAGE_SIZE) {
 352                 struct page *p;
 353                 int ret = get_user_pages_fast(start, 1, 0, &p);
 354                 if (ret != 1)
 355                         return ret;
 356                 if (bhv == MADV_SOFT_OFFLINE) {
 357                         printk(KERN_INFO "Soft offlining page %lx at %lx\n",
 358                                 page_to_pfn(p), start);
 359                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 360                         if (ret)
 361                                 break;
 362                         continue;
 363                 }
 364                 printk(KERN_INFO "Injecting memory failure for page %lx at %lx\n",
 365                        page_to_pfn(p), start);
 366                 /* Ignore return value for now */
 367                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 368         }
 369         return ret;
 370 }
 371 #endif
 372
 373 static long
 374 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 375                 unsigned long start, unsigned long end, int behavior)
 376 {
 377         switch (behavior) {
 378         case MADV_REMOVE:
 379                 return madvise_remove(vma, prev, start, end);
 380         case MADV_WILLNEED:
 381                 return madvise_willneed(vma, prev, start, end);
 382         case MADV_DONTNEED:
 383                 return madvise_dontneed(vma, prev, start, end);
 384         default:
 385                 return madvise_behavior(vma, prev, start, end, behavior);
 386         }
 387 }
 388
 389 static int
 390 madvise_behavior_valid(int behavior)
 391 {
 392         switch (behavior) {
 393         case MADV_DOFORK:
 394         case MADV_DONTFORK:
 395         case MADV_NORMAL:
 396         case MADV_SEQUENTIAL:
 397         case MADV_RANDOM:
 398         case MADV_REMOVE:
 399         case MADV_WILLNEED:
 400         case MADV_DONTNEED:
 401 #ifdef CONFIG_KSM
 402         case MADV_MERGEABLE:
 403         case MADV_UNMERGEABLE:
 404 #endif
 405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 406         case MADV_HUGEPAGE:
 407         case MADV_NOHUGEPAGE:
 408 #endif
 409         case MADV_DONTDUMP:
 410         case MADV_DODUMP:
 411                 return 1;
 412
 413         default:
 414                 return 0;
 415         }
 416 }
 417
 418 /*
 419  * The madvise(2) system call.
 420  *
 421  * Applications can use madvise() to advise the kernel how it should
 422  * handle paging I/O in this VM area.  The idea is to help the kernel
 423  * use appropriate read-ahead and caching techniques.  The information
 424  * provided is advisory only, and can be safely disregarded by the
 425  * kernel without affecting the correct operation of the application.
 426  *
 427  * behavior values:
 428  *  MADV_NORMAL - the default behavior is to read clusters.  This
 429  *              results in some read-ahead and read-behind.
 430  *  MADV_RANDOM - the system should read the minimum amount of data
 431  *              on any access, since it is unlikely that the appli-
 432  *              cation will need more than what it asks for.
 433  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 434  *              once, so they can be aggressively read ahead, and
 435  *              can be freed soon after they are accessed.
 436  *  MADV_WILLNEED - the application is notifying the system to read
 437  *              some pages ahead.
 438  *  MADV_DONTNEED - the application is finished with the given range,
 439  *              so the kernel can free resources associated with it.
 440  *  MADV_REMOVE - the application wants to free up the given range of
 441  *              pages and associated backing store.
 442  *  MADV_DONTFORK - omit this area from child's address space when forking:
 443  *              typically, to avoid COWing pages pinned by get_user_pages().
 444  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 445  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 446  *              this area with pages of identical content from other such areas.
 447  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 448  *
 449  * return values:
 450  *  zero    - success
 451  *  -EINVAL - start + len < 0, start is not page-aligned,
 452  *              "behavior" is not a valid value, or application
 453  *              is attempting to release locked or shared pages.
 454  *  -ENOMEM - addresses in the specified range are not currently
 455  *              mapped, or are outside the AS of the process.
 456  *  -EIO    - an I/O error occurred while paging in data.
 457  *  -EBADF  - map exists, but area maps something that isn't a file.
 458  *  -EAGAIN - a kernel resource was temporarily unavailable.
 459  */
 460 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 461 {
 462         unsigned long end, tmp;
 463         struct vm_area_struct * vma, *prev;
 464         int unmapped_error = 0;
 465         int error = -EINVAL;
 466         int write;
 467         size_t len;
 468         struct blk_plug plug;
 469
 470 #ifdef CONFIG_MEMORY_FAILURE
 471         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 472                 return madvise_hwpoison(behavior, start, start+len_in);
 473 #endif
 474         if (!madvise_behavior_valid(behavior))
 475                 return error;
 476
 477         if (start & ~PAGE_MASK)
 478                 return error;
 479         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 480
 481         /* Check to see whether len was rounded up from small -ve to zero */
 482         if (len_in && !len)
 483                 return error;
 484
 485         end = start + len;
 486         if (end < start)
 487                 return error;
 488
 489         error = 0;
 490         if (end == start)
 491                 return error;
 492
 493         write = madvise_need_mmap_write(behavior);
 494         if (write)
 495                 down_write(&current->mm->mmap_sem);
 496         else
 497                 down_read(&current->mm->mmap_sem);
 498
 499         /*
 500          * If the interval [start,end) covers some unmapped address
 501          * ranges, just ignore them, but return -ENOMEM at the end.
 502          * - different from the way of handling in mlock etc.
 503          */
 504         vma = find_vma_prev(current->mm, start, &prev);
 505         if (vma && start > vma->vm_start)
 506                 prev = vma;
 507
 508         blk_start_plug(&plug);
 509         for (;;) {
 510                 /* Still start < end. */
 511                 error = -ENOMEM;
 512                 if (!vma)
 513                         goto out;
 514
 515                 /* Here start < (end|vma->vm_end). */
 516                 if (start < vma->vm_start) {
 517                         unmapped_error = -ENOMEM;
 518                         start = vma->vm_start;
 519                         if (start >= end)
 520                                 goto out;
 521                 }
 522
 523                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 524                 tmp = vma->vm_end;
 525                 if (end < tmp)
 526                         tmp = end;
 527
 528                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 529                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 530                 if (error)
 531                         goto out;
 532                 start = tmp;
 533                 if (prev && start < prev->vm_end)
 534                         start = prev->vm_end;
 535                 error = unmapped_error;
 536                 if (start >= end)
 537                         goto out;
 538                 if (prev)
 539                         vma = prev->vm_next;
 540                 else    /* madvise_remove dropped mmap_sem */
 541                         vma = find_vma(current->mm, start);
 542         }
 543 out:
 544         blk_finish_plug(&plug);
 545         if (write)
 546                 up_write(&current->mm->mmap_sem);
 547         else
 548                 up_read(&current->mm->mmap_sem);
 549
 550         return error;
 551 }