mm/truncate.c

   1 /*
   2  * mm/truncate.c - code for taking down pages from address_spaces
   3  *
   4  * Copyright (C) 2002, Linus Torvalds
   5  *
   6  * 10Sep2002    akpm@zip.com.au
   7  *              Initial version.
   8  */
   9
  10 #include <linux/kernel.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/mm.h>
  13 #include <linux/swap.h>
  14 #include <linux/module.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/highmem.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/task_io_accounting_ops.h>
  19 #include <linux/buffer_head.h>  /* grr. try_to_release_page,
  20                                    do_invalidatepage */
  21
  22
  23 /**
  24  * do_invalidatepage - invalidate part or all of a page
  25  * @page: the page which is affected
  26  * @offset: the index of the truncation point
  27  *
  28  * do_invalidatepage() is called when all or part of the page has become
  29  * invalidated by a truncate operation.
  30  *
  31  * do_invalidatepage() does not have to release all buffers, but it must
  32  * ensure that no dirty buffer is left outside @offset and that no I/O
  33  * is underway against any of the blocks which are outside the truncation
  34  * point.  Because the caller is about to free (and possibly reuse) those
  35  * blocks on-disk.
  36  */
  37 void do_invalidatepage(struct page *page, unsigned long offset)
  38 {
  39         void (*invalidatepage)(struct page *, unsigned long);
  40         invalidatepage = page->mapping->a_ops->invalidatepage;
  41 #ifdef CONFIG_BLOCK
  42         if (!invalidatepage)
  43                 invalidatepage = block_invalidatepage;
  44 #endif
  45         if (invalidatepage)
  46                 (*invalidatepage)(page, offset);
  47 }
  48
  49 static inline void truncate_partial_page(struct page *page, unsigned partial)
  50 {
  51         zero_user_segment(page, partial, PAGE_CACHE_SIZE);
  52         if (PagePrivate(page))
  53                 do_invalidatepage(page, partial);
  54 }
  55
  56 /*
  57  * This cancels just the dirty bit on the kernel page itself, it
  58  * does NOT actually remove dirty bits on any mmap's that may be
  59  * around. It also leaves the page tagged dirty, so any sync
  60  * activity will still find it on the dirty lists, and in particular,
  61  * clear_page_dirty_for_io() will still look at the dirty bits in
  62  * the VM.
  63  *
  64  * Doing this should *normally* only ever be done when a page
  65  * is truncated, and is not actually mapped anywhere at all. However,
  66  * fs/buffer.c does this when it notices that somebody has cleaned
  67  * out all the buffers on a page without actually doing it through
  68  * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
  69  */
  70 void cancel_dirty_page(struct page *page, unsigned int account_size)
  71 {
  72         if (TestClearPageDirty(page)) {
  73                 struct address_space *mapping = page->mapping;
  74                 if (mapping && mapping_cap_account_dirty(mapping)) {
  75                         dec_zone_page_state(page, NR_FILE_DIRTY);
  76                         dec_bdi_stat(mapping->backing_dev_info,
  77                                         BDI_RECLAIMABLE);
  78                         if (account_size)
  79                                 task_io_account_cancelled_write(account_size);
  80                 }
  81         }
  82 }
  83 EXPORT_SYMBOL(cancel_dirty_page);
  84
  85 /*
  86  * If truncate cannot remove the fs-private metadata from the page, the page
  87  * becomes orphaned.  It will be left on the LRU and may even be mapped into
  88  * user pagetables if we're racing with filemap_fault().
  89  *
  90  * We need to bale out if page->mapping is no longer equal to the original
  91  * mapping.  This happens a) when the VM reclaimed the page while we waited on
  92  * its lock, b) when a concurrent invalidate_mapping_pages got there first and
  93  * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
  94  */
  95 static void
  96 truncate_complete_page(struct address_space *mapping, struct page *page)
  97 {
  98         if (page->mapping != mapping)
  99                 return;
 100
 101         if (PagePrivate(page))
 102                 do_invalidatepage(page, 0);
 103
 104         cancel_dirty_page(page, PAGE_CACHE_SIZE);
 105
 106         remove_from_page_cache(page);
 107         ClearPageUptodate(page);
 108         ClearPageMappedToDisk(page);
 109         page_cache_release(page);       /* pagecache ref */
 110 }
 111
 112 /*
 113  * This is for invalidate_mapping_pages().  That function can be called at
 114  * any time, and is not supposed to throw away dirty pages.  But pages can
 115  * be marked dirty at any time too, so use remove_mapping which safely
 116  * discards clean, unused pages.
 117  *
 118  * Returns non-zero if the page was successfully invalidated.
 119  */
 120 static int
 121 invalidate_complete_page(struct address_space *mapping, struct page *page)
 122 {
 123         int ret;
 124
 125         if (page->mapping != mapping)
 126                 return 0;
 127
 128         if (PagePrivate(page) && !try_to_release_page(page, 0))
 129                 return 0;
 130
 131         ret = remove_mapping(mapping, page);
 132
 133         return ret;
 134 }
 135
 136 /**
 137  * truncate_inode_pages - truncate range of pages specified by start and
 138  * end byte offsets
 139  * @mapping: mapping to truncate
 140  * @lstart: offset from which to truncate
 141  * @lend: offset to which to truncate
 142  *
 143  * Truncate the page cache, removing the pages that are between
 144  * specified offsets (and zeroing out partial page
 145  * (if lstart is not page aligned)).
 146  *
 147  * Truncate takes two passes - the first pass is nonblocking.  It will not
 148  * block on page locks and it will not block on writeback.  The second pass
 149  * will wait.  This is to prevent as much IO as possible in the affected region.
 150  * The first pass will remove most pages, so the search cost of the second pass
 151  * is low.
 152  *
 153  * When looking at page->index outside the page lock we need to be careful to
 154  * copy it into a local to avoid races (it could change at any time).
 155  *
 156  * We pass down the cache-hot hint to the page freeing code.  Even if the
 157  * mapping is large, it is probably the case that the final pages are the most
 158  * recently touched, and freeing happens in ascending file offset order.
 159  */
 160 void truncate_inode_pages_range(struct address_space *mapping,
 161                                 loff_t lstart, loff_t lend)
 162 {
 163         const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 164         pgoff_t end;
 165         const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
 166         struct pagevec pvec;
 167         pgoff_t next;
 168         int i;
 169
 170         if (mapping->nrpages == 0)
 171                 return;
 172
 173         BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
 174         end = (lend >> PAGE_CACHE_SHIFT);
 175
 176         pagevec_init(&pvec, 0);
 177         next = start;
 178         while (next <= end &&
 179                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 180                 for (i = 0; i < pagevec_count(&pvec); i++) {
 181                         struct page *page = pvec.pages[i];
 182                         pgoff_t page_index = page->index;
 183
 184                         if (page_index > end) {
 185                                 next = page_index;
 186                                 break;
 187                         }
 188
 189                         if (page_index > next)
 190                                 next = page_index;
 191                         next++;
 192                         if (TestSetPageLocked(page))
 193                                 continue;
 194                         if (PageWriteback(page)) {
 195                                 unlock_page(page);
 196                                 continue;
 197                         }
 198                         if (page_mapped(page)) {
 199                                 unmap_mapping_range(mapping,
 200                                   (loff_t)page_index<<PAGE_CACHE_SHIFT,
 201                                   PAGE_CACHE_SIZE, 0);
 202                         }
 203                         truncate_complete_page(mapping, page);
 204                         unlock_page(page);
 205                 }
 206                 pagevec_release(&pvec);
 207                 cond_resched();
 208         }
 209
 210         if (partial) {
 211                 struct page *page = find_lock_page(mapping, start - 1);
 212                 if (page) {
 213                         wait_on_page_writeback(page);
 214                         truncate_partial_page(page, partial);
 215                         unlock_page(page);
 216                         page_cache_release(page);
 217                 }
 218         }
 219
 220         next = start;
 221         for ( ; ; ) {
 222                 cond_resched();
 223                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 224                         if (next == start)
 225                                 break;
 226                         next = start;
 227                         continue;
 228                 }
 229                 if (pvec.pages[0]->index > end) {
 230                         pagevec_release(&pvec);
 231                         break;
 232                 }
 233                 for (i = 0; i < pagevec_count(&pvec); i++) {
 234                         struct page *page = pvec.pages[i];
 235
 236                         if (page->index > end)
 237                                 break;
 238                         lock_page(page);
 239                         wait_on_page_writeback(page);
 240                         if (page_mapped(page)) {
 241                                 unmap_mapping_range(mapping,
 242                                   (loff_t)page->index<<PAGE_CACHE_SHIFT,
 243                                   PAGE_CACHE_SIZE, 0);
 244                         }
 245                         if (page->index > next)
 246                                 next = page->index;
 247                         next++;
 248                         truncate_complete_page(mapping, page);
 249                         unlock_page(page);
 250                 }
 251                 pagevec_release(&pvec);
 252         }
 253 }
 254 EXPORT_SYMBOL(truncate_inode_pages_range);
 255
 256 /**
 257  * truncate_inode_pages - truncate *all* the pages from an offset
 258  * @mapping: mapping to truncate
 259  * @lstart: offset from which to truncate
 260  *
 261  * Called under (and serialised by) inode->i_mutex.
 262  */
 263 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
 264 {
 265         truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
 266 }
 267 EXPORT_SYMBOL(truncate_inode_pages);
 268
 269 unsigned long __invalidate_mapping_pages(struct address_space *mapping,
 270                                 pgoff_t start, pgoff_t end, bool be_atomic)
 271 {
 272         struct pagevec pvec;
 273         pgoff_t next = start;
 274         unsigned long ret = 0;
 275         int i;
 276
 277         pagevec_init(&pvec, 0);
 278         while (next <= end &&
 279                         pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 280                 for (i = 0; i < pagevec_count(&pvec); i++) {
 281                         struct page *page = pvec.pages[i];
 282                         pgoff_t index;
 283                         int lock_failed;
 284
 285                         lock_failed = TestSetPageLocked(page);
 286
 287                         /*
 288                          * We really shouldn't be looking at the ->index of an
 289                          * unlocked page.  But we're not allowed to lock these
 290                          * pages.  So we rely upon nobody altering the ->index
 291                          * of this (pinned-by-us) page.
 292                          */
 293                         index = page->index;
 294                         if (index > next)
 295                                 next = index;
 296                         next++;
 297                         if (lock_failed)
 298                                 continue;
 299
 300                         if (PageDirty(page) || PageWriteback(page))
 301                                 goto unlock;
 302                         if (page_mapped(page))
 303                                 goto unlock;
 304                         ret += invalidate_complete_page(mapping, page);
 305 unlock:
 306                         unlock_page(page);
 307                         if (next > end)
 308                                 break;
 309                 }
 310                 pagevec_release(&pvec);
 311                 if (likely(!be_atomic))
 312                         cond_resched();
 313         }
 314         return ret;
 315 }
 316
 317 /**
 318  * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
 319  * @mapping: the address_space which holds the pages to invalidate
 320  * @start: the offset 'from' which to invalidate
 321  * @end: the offset 'to' which to invalidate (inclusive)
 322  *
 323  * This function only removes the unlocked pages, if you want to
 324  * remove all the pages of one inode, you must call truncate_inode_pages.
 325  *
 326  * invalidate_mapping_pages() will not block on IO activity. It will not
 327  * invalidate pages which are dirty, locked, under writeback or mapped into
 328  * pagetables.
 329  */
 330 unsigned long invalidate_mapping_pages(struct address_space *mapping,
 331                                 pgoff_t start, pgoff_t end)
 332 {
 333         return __invalidate_mapping_pages(mapping, start, end, false);
 334 }
 335 EXPORT_SYMBOL(invalidate_mapping_pages);
 336
 337 /*
 338  * This is like invalidate_complete_page(), except it ignores the page's
 339  * refcount.  We do this because invalidate_inode_pages2() needs stronger
 340  * invalidation guarantees, and cannot afford to leave pages behind because
 341  * shrink_page_list() has a temp ref on them, or because they're transiently
 342  * sitting in the lru_cache_add() pagevecs.
 343  */
 344 static int
 345 invalidate_complete_page2(struct address_space *mapping, struct page *page)
 346 {
 347         if (page->mapping != mapping)
 348                 return 0;
 349
 350         if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
 351                 return 0;
 352
 353         write_lock_irq(&mapping->tree_lock);
 354         if (PageDirty(page))
 355                 goto failed;
 356
 357         BUG_ON(PagePrivate(page));
 358         __remove_from_page_cache(page);
 359         write_unlock_irq(&mapping->tree_lock);
 360         ClearPageUptodate(page);
 361         page_cache_release(page);       /* pagecache ref */
 362         return 1;
 363 failed:
 364         write_unlock_irq(&mapping->tree_lock);
 365         return 0;
 366 }
 367
 368 static int do_launder_page(struct address_space *mapping, struct page *page)
 369 {
 370         if (!PageDirty(page))
 371                 return 0;
 372         if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
 373                 return 0;
 374         return mapping->a_ops->launder_page(page);
 375 }
 376
 377 /**
 378  * invalidate_inode_pages2_range - remove range of pages from an address_space
 379  * @mapping: the address_space
 380  * @start: the page offset 'from' which to invalidate
 381  * @end: the page offset 'to' which to invalidate (inclusive)
 382  *
 383  * Any pages which are found to be mapped into pagetables are unmapped prior to
 384  * invalidation.
 385  *
 386  * Returns -EIO if any pages could not be invalidated.
 387  */
 388 int invalidate_inode_pages2_range(struct address_space *mapping,
 389                                   pgoff_t start, pgoff_t end)
 390 {
 391         struct pagevec pvec;
 392         pgoff_t next;
 393         int i;
 394         int ret = 0;
 395         int did_range_unmap = 0;
 396         int wrapped = 0;
 397
 398         pagevec_init(&pvec, 0);
 399         next = start;
 400         while (next <= end && !wrapped &&
 401                 pagevec_lookup(&pvec, mapping, next,
 402                         min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
 403                 for (i = 0; i < pagevec_count(&pvec); i++) {
 404                         struct page *page = pvec.pages[i];
 405                         pgoff_t page_index;
 406
 407                         lock_page(page);
 408                         if (page->mapping != mapping) {
 409                                 unlock_page(page);
 410                                 continue;
 411                         }
 412                         page_index = page->index;
 413                         next = page_index + 1;
 414                         if (next == 0)
 415                                 wrapped = 1;
 416                         if (page_index > end) {
 417                                 unlock_page(page);
 418                                 break;
 419                         }
 420                         wait_on_page_writeback(page);
 421                         if (page_mapped(page)) {
 422                                 if (!did_range_unmap) {
 423                                         /*
 424                                          * Zap the rest of the file in one hit.
 425                                          */
 426                                         unmap_mapping_range(mapping,
 427                                            (loff_t)page_index<<PAGE_CACHE_SHIFT,
 428                                            (loff_t)(end - page_index + 1)
 429                                                         << PAGE_CACHE_SHIFT,
 430                                             0);
 431                                         did_range_unmap = 1;
 432                                 } else {
 433                                         /*
 434                                          * Just zap this page
 435                                          */
 436                                         unmap_mapping_range(mapping,
 437                                           (loff_t)page_index<<PAGE_CACHE_SHIFT,
 438                                           PAGE_CACHE_SIZE, 0);
 439                                 }
 440                         }
 441                         BUG_ON(page_mapped(page));
 442                         ret = do_launder_page(mapping, page);
 443                         if (ret == 0 && !invalidate_complete_page2(mapping, page))
 444                                 ret = -EIO;
 445                         unlock_page(page);
 446                 }
 447                 pagevec_release(&pvec);
 448                 cond_resched();
 449         }
 450         return ret;
 451 }
 452 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
 453
 454 /**
 455  * invalidate_inode_pages2 - remove all pages from an address_space
 456  * @mapping: the address_space
 457  *
 458  * Any pages which are found to be mapped into pagetables are unmapped prior to
 459  * invalidation.
 460  *
 461  * Returns -EIO if any pages could not be invalidated.
 462  */
 463 int invalidate_inode_pages2(struct address_space *mapping)
 464 {
 465         return invalidate_inode_pages2_range(mapping, 0, -1);
 466 }
 467 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);