fs/nilfs2/page.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Buffer/page management specific to NILFS
   4  *
   5  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   6  *
   7  * Written by Ryusuke Konishi and Seiji Kihara.
   8  */
   9
  10 #include <linux/pagemap.h>
  11 #include <linux/writeback.h>
  12 #include <linux/swap.h>
  13 #include <linux/bitops.h>
  14 #include <linux/page-flags.h>
  15 #include <linux/list.h>
  16 #include <linux/highmem.h>
  17 #include <linux/pagevec.h>
  18 #include <linux/gfp.h>
  19 #include "nilfs.h"
  20 #include "page.h"
  21 #include "mdt.h"
  22
  23
  24 #define NILFS_BUFFER_INHERENT_BITS                                      \
  25         (BIT(BH_Uptodate) | BIT(BH_Mapped) | BIT(BH_NILFS_Node) |       \
  26          BIT(BH_NILFS_Volatile) | BIT(BH_NILFS_Checked))
  27
  28 static struct buffer_head *__nilfs_get_folio_block(struct folio *folio,
  29                 unsigned long block, pgoff_t index, int blkbits,
  30                 unsigned long b_state)
  31
  32 {
  33         unsigned long first_block;
  34         struct buffer_head *bh = folio_buffers(folio);
  35
  36         if (!bh)
  37                 bh = create_empty_buffers(folio, 1 << blkbits, b_state);
  38
  39         first_block = (unsigned long)index << (PAGE_SHIFT - blkbits);
  40         bh = get_nth_bh(bh, block - first_block);
  41
  42         touch_buffer(bh);
  43         wait_on_buffer(bh);
  44         return bh;
  45 }
  46
  47 struct buffer_head *nilfs_grab_buffer(struct inode *inode,
  48                                       struct address_space *mapping,
  49                                       unsigned long blkoff,
  50                                       unsigned long b_state)
  51 {
  52         int blkbits = inode->i_blkbits;
  53         pgoff_t index = blkoff >> (PAGE_SHIFT - blkbits);
  54         struct folio *folio;
  55         struct buffer_head *bh;
  56
  57         folio = filemap_grab_folio(mapping, index);
  58         if (IS_ERR(folio))
  59                 return NULL;
  60
  61         bh = __nilfs_get_folio_block(folio, blkoff, index, blkbits, b_state);
  62         if (unlikely(!bh)) {
  63                 folio_unlock(folio);
  64                 folio_put(folio);
  65                 return NULL;
  66         }
  67         return bh;
  68 }
  69
  70 /**
  71  * nilfs_forget_buffer - discard dirty state
  72  * @bh: buffer head of the buffer to be discarded
  73  */
  74 void nilfs_forget_buffer(struct buffer_head *bh)
  75 {
  76         struct page *page = bh->b_page;
  77         const unsigned long clear_bits =
  78                 (BIT(BH_Uptodate) | BIT(BH_Dirty) | BIT(BH_Mapped) |
  79                  BIT(BH_Async_Write) | BIT(BH_NILFS_Volatile) |
  80                  BIT(BH_NILFS_Checked) | BIT(BH_NILFS_Redirected));
  81
  82         lock_buffer(bh);
  83         set_mask_bits(&bh->b_state, clear_bits, 0);
  84         if (nilfs_page_buffers_clean(page))
  85                 __nilfs_clear_page_dirty(page);
  86
  87         bh->b_blocknr = -1;
  88         ClearPageUptodate(page);
  89         ClearPageMappedToDisk(page);
  90         unlock_buffer(bh);
  91         brelse(bh);
  92 }
  93
  94 /**
  95  * nilfs_copy_buffer -- copy buffer data and flags
  96  * @dbh: destination buffer
  97  * @sbh: source buffer
  98  */
  99 void nilfs_copy_buffer(struct buffer_head *dbh, struct buffer_head *sbh)
 100 {
 101         void *kaddr0, *kaddr1;
 102         unsigned long bits;
 103         struct page *spage = sbh->b_page, *dpage = dbh->b_page;
 104         struct buffer_head *bh;
 105
 106         kaddr0 = kmap_atomic(spage);
 107         kaddr1 = kmap_atomic(dpage);
 108         memcpy(kaddr1 + bh_offset(dbh), kaddr0 + bh_offset(sbh), sbh->b_size);
 109         kunmap_atomic(kaddr1);
 110         kunmap_atomic(kaddr0);
 111
 112         dbh->b_state = sbh->b_state & NILFS_BUFFER_INHERENT_BITS;
 113         dbh->b_blocknr = sbh->b_blocknr;
 114         dbh->b_bdev = sbh->b_bdev;
 115
 116         bh = dbh;
 117         bits = sbh->b_state & (BIT(BH_Uptodate) | BIT(BH_Mapped));
 118         while ((bh = bh->b_this_page) != dbh) {
 119                 lock_buffer(bh);
 120                 bits &= bh->b_state;
 121                 unlock_buffer(bh);
 122         }
 123         if (bits & BIT(BH_Uptodate))
 124                 SetPageUptodate(dpage);
 125         else
 126                 ClearPageUptodate(dpage);
 127         if (bits & BIT(BH_Mapped))
 128                 SetPageMappedToDisk(dpage);
 129         else
 130                 ClearPageMappedToDisk(dpage);
 131 }
 132
 133 /**
 134  * nilfs_page_buffers_clean - check if a page has dirty buffers or not.
 135  * @page: page to be checked
 136  *
 137  * nilfs_page_buffers_clean() returns zero if the page has dirty buffers.
 138  * Otherwise, it returns non-zero value.
 139  */
 140 int nilfs_page_buffers_clean(struct page *page)
 141 {
 142         struct buffer_head *bh, *head;
 143
 144         bh = head = page_buffers(page);
 145         do {
 146                 if (buffer_dirty(bh))
 147                         return 0;
 148                 bh = bh->b_this_page;
 149         } while (bh != head);
 150         return 1;
 151 }
 152
 153 void nilfs_page_bug(struct page *page)
 154 {
 155         struct address_space *m;
 156         unsigned long ino;
 157
 158         if (unlikely(!page)) {
 159                 printk(KERN_CRIT "NILFS_PAGE_BUG(NULL)\n");
 160                 return;
 161         }
 162
 163         m = page->mapping;
 164         ino = m ? m->host->i_ino : 0;
 165
 166         printk(KERN_CRIT "NILFS_PAGE_BUG(%p): cnt=%d index#=%llu flags=0x%lx "
 167                "mapping=%p ino=%lu\n",
 168                page, page_ref_count(page),
 169                (unsigned long long)page->index, page->flags, m, ino);
 170
 171         if (page_has_buffers(page)) {
 172                 struct buffer_head *bh, *head;
 173                 int i = 0;
 174
 175                 bh = head = page_buffers(page);
 176                 do {
 177                         printk(KERN_CRIT
 178                                " BH[%d] %p: cnt=%d block#=%llu state=0x%lx\n",
 179                                i++, bh, atomic_read(&bh->b_count),
 180                                (unsigned long long)bh->b_blocknr, bh->b_state);
 181                         bh = bh->b_this_page;
 182                 } while (bh != head);
 183         }
 184 }
 185
 186 /**
 187  * nilfs_copy_folio -- copy the folio with buffers
 188  * @dst: destination folio
 189  * @src: source folio
 190  * @copy_dirty: flag whether to copy dirty states on the folio's buffer heads.
 191  *
 192  * This function is for both data folios and btnode folios.  The dirty flag
 193  * should be treated by caller.  The folio must not be under i/o.
 194  * Both src and dst folio must be locked
 195  */
 196 static void nilfs_copy_folio(struct folio *dst, struct folio *src,
 197                 bool copy_dirty)
 198 {
 199         struct buffer_head *dbh, *dbufs, *sbh;
 200         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
 201
 202         BUG_ON(folio_test_writeback(dst));
 203
 204         sbh = folio_buffers(src);
 205         dbh = folio_buffers(dst);
 206         if (!dbh)
 207                 dbh = create_empty_buffers(dst, sbh->b_size, 0);
 208
 209         if (copy_dirty)
 210                 mask |= BIT(BH_Dirty);
 211
 212         dbufs = dbh;
 213         do {
 214                 lock_buffer(sbh);
 215                 lock_buffer(dbh);
 216                 dbh->b_state = sbh->b_state & mask;
 217                 dbh->b_blocknr = sbh->b_blocknr;
 218                 dbh->b_bdev = sbh->b_bdev;
 219                 sbh = sbh->b_this_page;
 220                 dbh = dbh->b_this_page;
 221         } while (dbh != dbufs);
 222
 223         folio_copy(dst, src);
 224
 225         if (folio_test_uptodate(src) && !folio_test_uptodate(dst))
 226                 folio_mark_uptodate(dst);
 227         else if (!folio_test_uptodate(src) && folio_test_uptodate(dst))
 228                 folio_clear_uptodate(dst);
 229         if (folio_test_mappedtodisk(src) && !folio_test_mappedtodisk(dst))
 230                 folio_set_mappedtodisk(dst);
 231         else if (!folio_test_mappedtodisk(src) && folio_test_mappedtodisk(dst))
 232                 folio_clear_mappedtodisk(dst);
 233
 234         do {
 235                 unlock_buffer(sbh);
 236                 unlock_buffer(dbh);
 237                 sbh = sbh->b_this_page;
 238                 dbh = dbh->b_this_page;
 239         } while (dbh != dbufs);
 240 }
 241
 242 int nilfs_copy_dirty_pages(struct address_space *dmap,
 243                            struct address_space *smap)
 244 {
 245         struct folio_batch fbatch;
 246         unsigned int i;
 247         pgoff_t index = 0;
 248         int err = 0;
 249
 250         folio_batch_init(&fbatch);
 251 repeat:
 252         if (!filemap_get_folios_tag(smap, &index, (pgoff_t)-1,
 253                                 PAGECACHE_TAG_DIRTY, &fbatch))
 254                 return 0;
 255
 256         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 257                 struct folio *folio = fbatch.folios[i], *dfolio;
 258
 259                 folio_lock(folio);
 260                 if (unlikely(!folio_test_dirty(folio)))
 261                         NILFS_PAGE_BUG(&folio->page, "inconsistent dirty state");
 262
 263                 dfolio = filemap_grab_folio(dmap, folio->index);
 264                 if (unlikely(IS_ERR(dfolio))) {
 265                         /* No empty page is added to the page cache */
 266                         folio_unlock(folio);
 267                         err = PTR_ERR(dfolio);
 268                         break;
 269                 }
 270                 if (unlikely(!folio_buffers(folio)))
 271                         NILFS_PAGE_BUG(&folio->page,
 272                                        "found empty page in dat page cache");
 273
 274                 nilfs_copy_folio(dfolio, folio, true);
 275                 filemap_dirty_folio(folio_mapping(dfolio), dfolio);
 276
 277                 folio_unlock(dfolio);
 278                 folio_put(dfolio);
 279                 folio_unlock(folio);
 280         }
 281         folio_batch_release(&fbatch);
 282         cond_resched();
 283
 284         if (likely(!err))
 285                 goto repeat;
 286         return err;
 287 }
 288
 289 /**
 290  * nilfs_copy_back_pages -- copy back pages to original cache from shadow cache
 291  * @dmap: destination page cache
 292  * @smap: source page cache
 293  *
 294  * No pages must be added to the cache during this process.
 295  * This must be ensured by the caller.
 296  */
 297 void nilfs_copy_back_pages(struct address_space *dmap,
 298                            struct address_space *smap)
 299 {
 300         struct folio_batch fbatch;
 301         unsigned int i, n;
 302         pgoff_t start = 0;
 303
 304         folio_batch_init(&fbatch);
 305 repeat:
 306         n = filemap_get_folios(smap, &start, ~0UL, &fbatch);
 307         if (!n)
 308                 return;
 309
 310         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 311                 struct folio *folio = fbatch.folios[i], *dfolio;
 312                 pgoff_t index = folio->index;
 313
 314                 folio_lock(folio);
 315                 dfolio = filemap_lock_folio(dmap, index);
 316                 if (!IS_ERR(dfolio)) {
 317                         /* overwrite existing folio in the destination cache */
 318                         WARN_ON(folio_test_dirty(dfolio));
 319                         nilfs_copy_folio(dfolio, folio, false);
 320                         folio_unlock(dfolio);
 321                         folio_put(dfolio);
 322                         /* Do we not need to remove folio from smap here? */
 323                 } else {
 324                         struct folio *f;
 325
 326                         /* move the folio to the destination cache */
 327                         xa_lock_irq(&smap->i_pages);
 328                         f = __xa_erase(&smap->i_pages, index);
 329                         WARN_ON(folio != f);
 330                         smap->nrpages--;
 331                         xa_unlock_irq(&smap->i_pages);
 332
 333                         xa_lock_irq(&dmap->i_pages);
 334                         f = __xa_store(&dmap->i_pages, index, folio, GFP_NOFS);
 335                         if (unlikely(f)) {
 336                                 /* Probably -ENOMEM */
 337                                 folio->mapping = NULL;
 338                                 folio_put(folio);
 339                         } else {
 340                                 folio->mapping = dmap;
 341                                 dmap->nrpages++;
 342                                 if (folio_test_dirty(folio))
 343                                         __xa_set_mark(&dmap->i_pages, index,
 344                                                         PAGECACHE_TAG_DIRTY);
 345                         }
 346                         xa_unlock_irq(&dmap->i_pages);
 347                 }
 348                 folio_unlock(folio);
 349         }
 350         folio_batch_release(&fbatch);
 351         cond_resched();
 352
 353         goto repeat;
 354 }
 355
 356 /**
 357  * nilfs_clear_dirty_pages - discard dirty pages in address space
 358  * @mapping: address space with dirty pages for discarding
 359  * @silent: suppress [true] or print [false] warning messages
 360  */
 361 void nilfs_clear_dirty_pages(struct address_space *mapping, bool silent)
 362 {
 363         struct folio_batch fbatch;
 364         unsigned int i;
 365         pgoff_t index = 0;
 366
 367         folio_batch_init(&fbatch);
 368
 369         while (filemap_get_folios_tag(mapping, &index, (pgoff_t)-1,
 370                                 PAGECACHE_TAG_DIRTY, &fbatch)) {
 371                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 372                         struct folio *folio = fbatch.folios[i];
 373
 374                         folio_lock(folio);
 375
 376                         /*
 377                          * This folio may have been removed from the address
 378                          * space by truncation or invalidation when the lock
 379                          * was acquired.  Skip processing in that case.
 380                          */
 381                         if (likely(folio->mapping == mapping))
 382                                 nilfs_clear_dirty_page(&folio->page, silent);
 383
 384                         folio_unlock(folio);
 385                 }
 386                 folio_batch_release(&fbatch);
 387                 cond_resched();
 388         }
 389 }
 390
 391 /**
 392  * nilfs_clear_dirty_page - discard dirty page
 393  * @page: dirty page that will be discarded
 394  * @silent: suppress [true] or print [false] warning messages
 395  */
 396 void nilfs_clear_dirty_page(struct page *page, bool silent)
 397 {
 398         struct inode *inode = page->mapping->host;
 399         struct super_block *sb = inode->i_sb;
 400
 401         BUG_ON(!PageLocked(page));
 402
 403         if (!silent)
 404                 nilfs_warn(sb, "discard dirty page: offset=%lld, ino=%lu",
 405                            page_offset(page), inode->i_ino);
 406
 407         ClearPageUptodate(page);
 408         ClearPageMappedToDisk(page);
 409
 410         if (page_has_buffers(page)) {
 411                 struct buffer_head *bh, *head;
 412                 const unsigned long clear_bits =
 413                         (BIT(BH_Uptodate) | BIT(BH_Dirty) | BIT(BH_Mapped) |
 414                          BIT(BH_Async_Write) | BIT(BH_NILFS_Volatile) |
 415                          BIT(BH_NILFS_Checked) | BIT(BH_NILFS_Redirected));
 416
 417                 bh = head = page_buffers(page);
 418                 do {
 419                         lock_buffer(bh);
 420                         if (!silent)
 421                                 nilfs_warn(sb,
 422                                            "discard dirty block: blocknr=%llu, size=%zu",
 423                                            (u64)bh->b_blocknr, bh->b_size);
 424
 425                         set_mask_bits(&bh->b_state, clear_bits, 0);
 426                         unlock_buffer(bh);
 427                 } while (bh = bh->b_this_page, bh != head);
 428         }
 429
 430         __nilfs_clear_page_dirty(page);
 431 }
 432
 433 unsigned int nilfs_page_count_clean_buffers(struct page *page,
 434                                             unsigned int from, unsigned int to)
 435 {
 436         unsigned int block_start, block_end;
 437         struct buffer_head *bh, *head;
 438         unsigned int nc = 0;
 439
 440         for (bh = head = page_buffers(page), block_start = 0;
 441              bh != head || !block_start;
 442              block_start = block_end, bh = bh->b_this_page) {
 443                 block_end = block_start + bh->b_size;
 444                 if (block_end > from && block_start < to && !buffer_dirty(bh))
 445                         nc++;
 446         }
 447         return nc;
 448 }
 449
 450 /*
 451  * NILFS2 needs clear_page_dirty() in the following two cases:
 452  *
 453  * 1) For B-tree node pages and data pages of DAT file, NILFS2 clears dirty
 454  *    flag of pages when it copies back pages from shadow cache to the
 455  *    original cache.
 456  *
 457  * 2) Some B-tree operations like insertion or deletion may dispose buffers
 458  *    in dirty state, and this needs to cancel the dirty state of their pages.
 459  */
 460 int __nilfs_clear_page_dirty(struct page *page)
 461 {
 462         struct address_space *mapping = page->mapping;
 463
 464         if (mapping) {
 465                 xa_lock_irq(&mapping->i_pages);
 466                 if (test_bit(PG_dirty, &page->flags)) {
 467                         __xa_clear_mark(&mapping->i_pages, page_index(page),
 468                                              PAGECACHE_TAG_DIRTY);
 469                         xa_unlock_irq(&mapping->i_pages);
 470                         return clear_page_dirty_for_io(page);
 471                 }
 472                 xa_unlock_irq(&mapping->i_pages);
 473                 return 0;
 474         }
 475         return TestClearPageDirty(page);
 476 }
 477
 478 /**
 479  * nilfs_find_uncommitted_extent - find extent of uncommitted data
 480  * @inode: inode
 481  * @start_blk: start block offset (in)
 482  * @blkoff: start offset of the found extent (out)
 483  *
 484  * This function searches an extent of buffers marked "delayed" which
 485  * starts from a block offset equal to or larger than @start_blk.  If
 486  * such an extent was found, this will store the start offset in
 487  * @blkoff and return its length in blocks.  Otherwise, zero is
 488  * returned.
 489  */
 490 unsigned long nilfs_find_uncommitted_extent(struct inode *inode,
 491                                             sector_t start_blk,
 492                                             sector_t *blkoff)
 493 {
 494         unsigned int i, nr_folios;
 495         pgoff_t index;
 496         unsigned long length = 0;
 497         struct folio_batch fbatch;
 498         struct folio *folio;
 499
 500         if (inode->i_mapping->nrpages == 0)
 501                 return 0;
 502
 503         index = start_blk >> (PAGE_SHIFT - inode->i_blkbits);
 504
 505         folio_batch_init(&fbatch);
 506
 507 repeat:
 508         nr_folios = filemap_get_folios_contig(inode->i_mapping, &index, ULONG_MAX,
 509                         &fbatch);
 510         if (nr_folios == 0)
 511                 return length;
 512
 513         i = 0;
 514         do {
 515                 folio = fbatch.folios[i];
 516
 517                 folio_lock(folio);
 518                 if (folio_buffers(folio)) {
 519                         struct buffer_head *bh, *head;
 520                         sector_t b;
 521
 522                         b = folio->index << (PAGE_SHIFT - inode->i_blkbits);
 523                         bh = head = folio_buffers(folio);
 524                         do {
 525                                 if (b < start_blk)
 526                                         continue;
 527                                 if (buffer_delay(bh)) {
 528                                         if (length == 0)
 529                                                 *blkoff = b;
 530                                         length++;
 531                                 } else if (length > 0) {
 532                                         goto out_locked;
 533                                 }
 534                         } while (++b, bh = bh->b_this_page, bh != head);
 535                 } else {
 536                         if (length > 0)
 537                                 goto out_locked;
 538                 }
 539                 folio_unlock(folio);
 540
 541         } while (++i < nr_folios);
 542
 543         folio_batch_release(&fbatch);
 544         cond_resched();
 545         goto repeat;
 546
 547 out_locked:
 548         folio_unlock(folio);
 549         folio_batch_release(&fbatch);
 550         return length;
 551 }