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_entry(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 static long madvise_remove(struct vm_area_struct *vma,
 297                                 struct vm_area_struct **prev,
 298                                 unsigned long start, unsigned long end)
 299 {
 300         loff_t offset;
 301         int error;
 302         struct file *f;
 303
 304         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 305
 306         if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
 307                 return -EINVAL;
 308
 309         f = vma->vm_file;
 310
 311         if (!f || !f->f_mapping || !f->f_mapping->host) {
 312                         return -EINVAL;
 313         }
 314
 315         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 316                 return -EACCES;
 317
 318         offset = (loff_t)(start - vma->vm_start)
 319                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 320
 321         /*
 322          * Filesystem's fallocate may need to take i_mutex.  We need to
 323          * explicitly grab a reference because the vma (and hence the
 324          * vma's reference to the file) can go away as soon as we drop
 325          * mmap_sem.
 326          */
 327         get_file(f);
 328         up_read(&current->mm->mmap_sem);
 329         error = do_fallocate(f,
 330                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 331                                 offset, end - start);
 332         fput(f);
 333         down_read(&current->mm->mmap_sem);
 334         return error;
 335 }
 336
 337 #ifdef CONFIG_MEMORY_FAILURE
 338 /*
 339  * Error injection support for memory error handling.
 340  */
 341 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 342 {
 343         struct page *p;
 344         if (!capable(CAP_SYS_ADMIN))
 345                 return -EPERM;
 346         for (; start < end; start += PAGE_SIZE <<
 347                                 compound_order(compound_head(p))) {
 348                 int ret;
 349
 350                 ret = get_user_pages_fast(start, 1, 0, &p);
 351                 if (ret != 1)
 352                         return ret;
 353
 354                 if (PageHWPoison(p)) {
 355                         put_page(p);
 356                         continue;
 357                 }
 358                 if (bhv == MADV_SOFT_OFFLINE) {
 359                         pr_info("Soft offlining page %#lx at %#lx\n",
 360                                 page_to_pfn(p), start);
 361                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 362                         if (ret)
 363                                 return ret;
 364                         continue;
 365                 }
 366                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 367                        page_to_pfn(p), start);
 368                 /* Ignore return value for now */
 369                 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 370         }
 371         return 0;
 372 }
 373 #endif
 374
 375 static long
 376 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 377                 unsigned long start, unsigned long end, int behavior)
 378 {
 379         switch (behavior) {
 380         case MADV_REMOVE:
 381                 return madvise_remove(vma, prev, start, end);
 382         case MADV_WILLNEED:
 383                 return madvise_willneed(vma, prev, start, end);
 384         case MADV_DONTNEED:
 385                 return madvise_dontneed(vma, prev, start, end);
 386         default:
 387                 return madvise_behavior(vma, prev, start, end, behavior);
 388         }
 389 }
 390
 391 static int
 392 madvise_behavior_valid(int behavior)
 393 {
 394         switch (behavior) {
 395         case MADV_DOFORK:
 396         case MADV_DONTFORK:
 397         case MADV_NORMAL:
 398         case MADV_SEQUENTIAL:
 399         case MADV_RANDOM:
 400         case MADV_REMOVE:
 401         case MADV_WILLNEED:
 402         case MADV_DONTNEED:
 403 #ifdef CONFIG_KSM
 404         case MADV_MERGEABLE:
 405         case MADV_UNMERGEABLE:
 406 #endif
 407 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 408         case MADV_HUGEPAGE:
 409         case MADV_NOHUGEPAGE:
 410 #endif
 411         case MADV_DONTDUMP:
 412         case MADV_DODUMP:
 413                 return 1;
 414
 415         default:
 416                 return 0;
 417         }
 418 }
 419
 420 /*
 421  * The madvise(2) system call.
 422  *
 423  * Applications can use madvise() to advise the kernel how it should
 424  * handle paging I/O in this VM area.  The idea is to help the kernel
 425  * use appropriate read-ahead and caching techniques.  The information
 426  * provided is advisory only, and can be safely disregarded by the
 427  * kernel without affecting the correct operation of the application.
 428  *
 429  * behavior values:
 430  *  MADV_NORMAL - the default behavior is to read clusters.  This
 431  *              results in some read-ahead and read-behind.
 432  *  MADV_RANDOM - the system should read the minimum amount of data
 433  *              on any access, since it is unlikely that the appli-
 434  *              cation will need more than what it asks for.
 435  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 436  *              once, so they can be aggressively read ahead, and
 437  *              can be freed soon after they are accessed.
 438  *  MADV_WILLNEED - the application is notifying the system to read
 439  *              some pages ahead.
 440  *  MADV_DONTNEED - the application is finished with the given range,
 441  *              so the kernel can free resources associated with it.
 442  *  MADV_REMOVE - the application wants to free up the given range of
 443  *              pages and associated backing store.
 444  *  MADV_DONTFORK - omit this area from child's address space when forking:
 445  *              typically, to avoid COWing pages pinned by get_user_pages().
 446  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 447  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 448  *              this area with pages of identical content from other such areas.
 449  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 450  *
 451  * return values:
 452  *  zero    - success
 453  *  -EINVAL - start + len < 0, start is not page-aligned,
 454  *              "behavior" is not a valid value, or application
 455  *              is attempting to release locked or shared pages.
 456  *  -ENOMEM - addresses in the specified range are not currently
 457  *              mapped, or are outside the AS of the process.
 458  *  -EIO    - an I/O error occurred while paging in data.
 459  *  -EBADF  - map exists, but area maps something that isn't a file.
 460  *  -EAGAIN - a kernel resource was temporarily unavailable.
 461  */
 462 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 463 {
 464         unsigned long end, tmp;
 465         struct vm_area_struct *vma, *prev;
 466         int unmapped_error = 0;
 467         int error = -EINVAL;
 468         int write;
 469         size_t len;
 470         struct blk_plug plug;
 471
 472 #ifdef CONFIG_MEMORY_FAILURE
 473         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 474                 return madvise_hwpoison(behavior, start, start+len_in);
 475 #endif
 476         if (!madvise_behavior_valid(behavior))
 477                 return error;
 478
 479         if (start & ~PAGE_MASK)
 480                 return error;
 481         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 482
 483         /* Check to see whether len was rounded up from small -ve to zero */
 484         if (len_in && !len)
 485                 return error;
 486
 487         end = start + len;
 488         if (end < start)
 489                 return error;
 490
 491         error = 0;
 492         if (end == start)
 493                 return error;
 494
 495         write = madvise_need_mmap_write(behavior);
 496         if (write)
 497                 down_write(&current->mm->mmap_sem);
 498         else
 499                 down_read(&current->mm->mmap_sem);
 500
 501         /*
 502          * If the interval [start,end) covers some unmapped address
 503          * ranges, just ignore them, but return -ENOMEM at the end.
 504          * - different from the way of handling in mlock etc.
 505          */
 506         vma = find_vma_prev(current->mm, start, &prev);
 507         if (vma && start > vma->vm_start)
 508                 prev = vma;
 509
 510         blk_start_plug(&plug);
 511         for (;;) {
 512                 /* Still start < end. */
 513                 error = -ENOMEM;
 514                 if (!vma)
 515                         goto out;
 516
 517                 /* Here start < (end|vma->vm_end). */
 518                 if (start < vma->vm_start) {
 519                         unmapped_error = -ENOMEM;
 520                         start = vma->vm_start;
 521                         if (start >= end)
 522                                 goto out;
 523                 }
 524
 525                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 526                 tmp = vma->vm_end;
 527                 if (end < tmp)
 528                         tmp = end;
 529
 530                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 531                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 532                 if (error)
 533                         goto out;
 534                 start = tmp;
 535                 if (prev && start < prev->vm_end)
 536                         start = prev->vm_end;
 537                 error = unmapped_error;
 538                 if (start >= end)
 539                         goto out;
 540                 if (prev)
 541                         vma = prev->vm_next;
 542                 else    /* madvise_remove dropped mmap_sem */
 543                         vma = find_vma(current->mm, start);
 544         }
 545 out:
 546         blk_finish_plug(&plug);
 547         if (write)
 548                 up_write(&current->mm->mmap_sem);
 549         else
 550                 up_read(&current->mm->mmap_sem);
 551
 552         return error;
 553 }