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 & end byte offsets
 138  * @mapping: mapping to truncate
 139  * @lstart: offset from which to truncate
 140  * @lend: offset to which to truncate
 141  *
 142  * Truncate the page cache, removing the pages that are between
 143  * specified offsets (and zeroing out partial page
 144  * (if lstart is not page aligned)).
 145  *
 146  * Truncate takes two passes - the first pass is nonblocking.  It will not
 147  * block on page locks and it will not block on writeback.  The second pass
 148  * will wait.  This is to prevent as much IO as possible in the affected region.
 149  * The first pass will remove most pages, so the search cost of the second pass
 150  * is low.
 151  *
 152  * When looking at page->index outside the page lock we need to be careful to
 153  * copy it into a local to avoid races (it could change at any time).
 154  *
 155  * We pass down the cache-hot hint to the page freeing code.  Even if the
 156  * mapping is large, it is probably the case that the final pages are the most
 157  * recently touched, and freeing happens in ascending file offset order.
 158  */
 159 void truncate_inode_pages_range(struct address_space *mapping,
 160                                 loff_t lstart, loff_t lend)
 161 {
 162         const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 163         pgoff_t end;
 164         const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
 165         struct pagevec pvec;
 166         pgoff_t next;
 167         int i;
 168
 169         if (mapping->nrpages == 0)
 170                 return;
 171
 172         BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
 173         end = (lend >> PAGE_CACHE_SHIFT);
 174
 175         pagevec_init(&pvec, 0);
 176         next = start;
 177         while (next <= end &&
 178                pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 179                 for (i = 0; i < pagevec_count(&pvec); i++) {
 180                         struct page *page = pvec.pages[i];
 181                         pgoff_t page_index = page->index;
 182
 183                         if (page_index > end) {
 184                                 next = page_index;
 185                                 break;
 186                         }
 187
 188                         if (page_index > next)
 189                                 next = page_index;
 190                         next++;
 191                         if (TestSetPageLocked(page))
 192                                 continue;
 193                         if (PageWriteback(page)) {
 194                                 unlock_page(page);
 195                                 continue;
 196                         }
 197                         if (page_mapped(page)) {
 198                                 unmap_mapping_range(mapping,
 199                                   (loff_t)page_index<<PAGE_CACHE_SHIFT,
 200                                   PAGE_CACHE_SIZE, 0);
 201                         }
 202                         truncate_complete_page(mapping, page);
 203                         unlock_page(page);
 204                 }
 205                 pagevec_release(&pvec);
 206                 cond_resched();
 207         }
 208
 209         if (partial) {
 210                 struct page *page = find_lock_page(mapping, start - 1);
 211                 if (page) {
 212                         wait_on_page_writeback(page);
 213                         truncate_partial_page(page, partial);
 214                         unlock_page(page);
 215                         page_cache_release(page);
 216                 }
 217         }
 218
 219         next = start;
 220         for ( ; ; ) {
 221                 cond_resched();
 222                 if (!pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 223                         if (next == start)
 224                                 break;
 225                         next = start;
 226                         continue;
 227                 }
 228                 if (pvec.pages[0]->index > end) {
 229                         pagevec_release(&pvec);
 230                         break;
 231                 }
 232                 for (i = 0; i < pagevec_count(&pvec); i++) {
 233                         struct page *page = pvec.pages[i];
 234
 235                         if (page->index > end)
 236                                 break;
 237                         lock_page(page);
 238                         wait_on_page_writeback(page);
 239                         if (page_mapped(page)) {
 240                                 unmap_mapping_range(mapping,
 241                                   (loff_t)page->index<<PAGE_CACHE_SHIFT,
 242                                   PAGE_CACHE_SIZE, 0);
 243                         }
 244                         if (page->index > next)
 245                                 next = page->index;
 246                         next++;
 247                         truncate_complete_page(mapping, page);
 248                         unlock_page(page);
 249                 }
 250                 pagevec_release(&pvec);
 251         }
 252 }
 253 EXPORT_SYMBOL(truncate_inode_pages_range);
 254
 255 /**
 256  * truncate_inode_pages - truncate *all* the pages from an offset
 257  * @mapping: mapping to truncate
 258  * @lstart: offset from which to truncate
 259  *
 260  * Called under (and serialised by) inode->i_mutex.
 261  */
 262 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
 263 {
 264         truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
 265 }
 266 EXPORT_SYMBOL(truncate_inode_pages);
 267
 268 unsigned long __invalidate_mapping_pages(struct address_space *mapping,
 269                                 pgoff_t start, pgoff_t end, bool be_atomic)
 270 {
 271         struct pagevec pvec;
 272         pgoff_t next = start;
 273         unsigned long ret = 0;
 274         int i;
 275
 276         pagevec_init(&pvec, 0);
 277         while (next <= end &&
 278                         pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
 279                 for (i = 0; i < pagevec_count(&pvec); i++) {
 280                         struct page *page = pvec.pages[i];
 281                         pgoff_t index;
 282                         int lock_failed;
 283
 284                         lock_failed = TestSetPageLocked(page);
 285
 286                         /*
 287                          * We really shouldn't be looking at the ->index of an
 288                          * unlocked page.  But we're not allowed to lock these
 289                          * pages.  So we rely upon nobody altering the ->index
 290                          * of this (pinned-by-us) page.
 291                          */
 292                         index = page->index;
 293                         if (index > next)
 294                                 next = index;
 295                         next++;
 296                         if (lock_failed)
 297                                 continue;
 298
 299                         if (PageDirty(page) || PageWriteback(page))
 300                                 goto unlock;
 301                         if (page_mapped(page))
 302                                 goto unlock;
 303                         ret += invalidate_complete_page(mapping, page);
 304 unlock:
 305                         unlock_page(page);
 306                         if (next > end)
 307                                 break;
 308                 }
 309                 pagevec_release(&pvec);
 310                 if (likely(!be_atomic))
 311                         cond_resched();
 312         }
 313         return ret;
 314 }
 315
 316 /**
 317  * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
 318  * @mapping: the address_space which holds the pages to invalidate
 319  * @start: the offset 'from' which to invalidate
 320  * @end: the offset 'to' which to invalidate (inclusive)
 321  *
 322  * This function only removes the unlocked pages, if you want to
 323  * remove all the pages of one inode, you must call truncate_inode_pages.
 324  *
 325  * invalidate_mapping_pages() will not block on IO activity. It will not
 326  * invalidate pages which are dirty, locked, under writeback or mapped into
 327  * pagetables.
 328  */
 329 unsigned long invalidate_mapping_pages(struct address_space *mapping,
 330                                 pgoff_t start, pgoff_t end)
 331 {
 332         return __invalidate_mapping_pages(mapping, start, end, false);
 333 }
 334 EXPORT_SYMBOL(invalidate_mapping_pages);
 335
 336 /*
 337  * This is like invalidate_complete_page(), except it ignores the page's
 338  * refcount.  We do this because invalidate_inode_pages2() needs stronger
 339  * invalidation guarantees, and cannot afford to leave pages behind because
 340  * shrink_page_list() has a temp ref on them, or because they're transiently
 341  * sitting in the lru_cache_add() pagevecs.
 342  */
 343 static int
 344 invalidate_complete_page2(struct address_space *mapping, struct page *page)
 345 {
 346         if (page->mapping != mapping)
 347                 return 0;
 348
 349         if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
 350                 return 0;
 351
 352         write_lock_irq(&mapping->tree_lock);
 353         if (PageDirty(page))
 354                 goto failed;
 355
 356         BUG_ON(PagePrivate(page));
 357         __remove_from_page_cache(page);
 358         write_unlock_irq(&mapping->tree_lock);
 359         ClearPageUptodate(page);
 360         page_cache_release(page);       /* pagecache ref */
 361         return 1;
 362 failed:
 363         write_unlock_irq(&mapping->tree_lock);
 364         return 0;
 365 }
 366
 367 static int do_launder_page(struct address_space *mapping, struct page *page)
 368 {
 369         if (!PageDirty(page))
 370                 return 0;
 371         if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
 372                 return 0;
 373         return mapping->a_ops->launder_page(page);
 374 }
 375
 376 /**
 377  * invalidate_inode_pages2_range - remove range of pages from an address_space
 378  * @mapping: the address_space
 379  * @start: the page offset 'from' which to invalidate
 380  * @end: the page offset 'to' which to invalidate (inclusive)
 381  *
 382  * Any pages which are found to be mapped into pagetables are unmapped prior to
 383  * invalidation.
 384  *
 385  * Returns -EIO if any pages could not be invalidated.
 386  */
 387 int invalidate_inode_pages2_range(struct address_space *mapping,
 388                                   pgoff_t start, pgoff_t end)
 389 {
 390         struct pagevec pvec;
 391         pgoff_t next;
 392         int i;
 393         int ret = 0;
 394         int did_range_unmap = 0;
 395         int wrapped = 0;
 396
 397         pagevec_init(&pvec, 0);
 398         next = start;
 399         while (next <= end && !wrapped &&
 400                 pagevec_lookup(&pvec, mapping, next,
 401                         min(end - next, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
 402                 for (i = 0; i < pagevec_count(&pvec); i++) {
 403                         struct page *page = pvec.pages[i];
 404                         pgoff_t page_index;
 405
 406                         lock_page(page);
 407                         if (page->mapping != mapping) {
 408                                 unlock_page(page);
 409                                 continue;
 410                         }
 411                         page_index = page->index;
 412                         next = page_index + 1;
 413                         if (next == 0)
 414                                 wrapped = 1;
 415                         if (page_index > end) {
 416                                 unlock_page(page);
 417                                 break;
 418                         }
 419                         wait_on_page_writeback(page);
 420                         if (page_mapped(page)) {
 421                                 if (!did_range_unmap) {
 422                                         /*
 423                                          * Zap the rest of the file in one hit.
 424                                          */
 425                                         unmap_mapping_range(mapping,
 426                                            (loff_t)page_index<<PAGE_CACHE_SHIFT,
 427                                            (loff_t)(end - page_index + 1)
 428                                                         << PAGE_CACHE_SHIFT,
 429                                             0);
 430                                         did_range_unmap = 1;
 431                                 } else {
 432                                         /*
 433                                          * Just zap this page
 434                                          */
 435                                         unmap_mapping_range(mapping,
 436                                           (loff_t)page_index<<PAGE_CACHE_SHIFT,
 437                                           PAGE_CACHE_SIZE, 0);
 438                                 }
 439                         }
 440                         BUG_ON(page_mapped(page));
 441                         ret = do_launder_page(mapping, page);
 442                         if (ret == 0 && !invalidate_complete_page2(mapping, page))
 443                                 ret = -EIO;
 444                         unlock_page(page);
 445                 }
 446                 pagevec_release(&pvec);
 447                 cond_resched();
 448         }
 449         return ret;
 450 }
 451 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
 452
 453 /**
 454  * invalidate_inode_pages2 - remove all pages from an address_space
 455  * @mapping: the address_space
 456  *
 457  * Any pages which are found to be mapped into pagetables are unmapped prior to
 458  * invalidation.
 459  *
 460  * Returns -EIO if any pages could not be invalidated.
 461  */
 462 int invalidate_inode_pages2(struct address_space *mapping)
 463 {
 464         return invalidate_inode_pages2_range(mapping, 0, -1);
 465 }
 466 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);