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 *
  29 __nilfs_get_page_block(struct page *page, unsigned long block, pgoff_t index,
  30                        int blkbits, unsigned long b_state)
  31
  32 {
  33         unsigned long first_block;
  34         struct buffer_head *bh;
  35
  36         if (!page_has_buffers(page))
  37                 create_empty_buffers(page, 1 << blkbits, b_state);
  38
  39         first_block = (unsigned long)index << (PAGE_SHIFT - blkbits);
  40         bh = nilfs_page_get_nth_block(page, 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 page *page;
  55         struct buffer_head *bh;
  56
  57         page = grab_cache_page(mapping, index);
  58         if (unlikely(!page))
  59                 return NULL;
  60
  61         bh = __nilfs_get_page_block(page, blkoff, index, blkbits, b_state);
  62         if (unlikely(!bh)) {
  63                 unlock_page(page);
  64                 put_page(page);
  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_page -- copy the page with buffers
 188  * @dst: destination page
 189  * @src: source page
 190  * @copy_dirty: flag whether to copy dirty states on the page's buffer heads.
 191  *
 192  * This function is for both data pages and btnode pages.  The dirty flag
 193  * should be treated by caller.  The page must not be under i/o.
 194  * Both src and dst page must be locked
 195  */
 196 static void nilfs_copy_page(struct page *dst, struct page *src, int copy_dirty)
 197 {
 198         struct buffer_head *dbh, *dbufs, *sbh;
 199         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
 200
 201         BUG_ON(PageWriteback(dst));
 202
 203         sbh = page_buffers(src);
 204         if (!page_has_buffers(dst))
 205                 create_empty_buffers(dst, sbh->b_size, 0);
 206
 207         if (copy_dirty)
 208                 mask |= BIT(BH_Dirty);
 209
 210         dbh = dbufs = page_buffers(dst);
 211         do {
 212                 lock_buffer(sbh);
 213                 lock_buffer(dbh);
 214                 dbh->b_state = sbh->b_state & mask;
 215                 dbh->b_blocknr = sbh->b_blocknr;
 216                 dbh->b_bdev = sbh->b_bdev;
 217                 sbh = sbh->b_this_page;
 218                 dbh = dbh->b_this_page;
 219         } while (dbh != dbufs);
 220
 221         copy_highpage(dst, src);
 222
 223         if (PageUptodate(src) && !PageUptodate(dst))
 224                 SetPageUptodate(dst);
 225         else if (!PageUptodate(src) && PageUptodate(dst))
 226                 ClearPageUptodate(dst);
 227         if (PageMappedToDisk(src) && !PageMappedToDisk(dst))
 228                 SetPageMappedToDisk(dst);
 229         else if (!PageMappedToDisk(src) && PageMappedToDisk(dst))
 230                 ClearPageMappedToDisk(dst);
 231
 232         do {
 233                 unlock_buffer(sbh);
 234                 unlock_buffer(dbh);
 235                 sbh = sbh->b_this_page;
 236                 dbh = dbh->b_this_page;
 237         } while (dbh != dbufs);
 238 }
 239
 240 int nilfs_copy_dirty_pages(struct address_space *dmap,
 241                            struct address_space *smap)
 242 {
 243         struct folio_batch fbatch;
 244         unsigned int i;
 245         pgoff_t index = 0;
 246         int err = 0;
 247
 248         folio_batch_init(&fbatch);
 249 repeat:
 250         if (!filemap_get_folios_tag(smap, &index, (pgoff_t)-1,
 251                                 PAGECACHE_TAG_DIRTY, &fbatch))
 252                 return 0;
 253
 254         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 255                 struct folio *folio = fbatch.folios[i], *dfolio;
 256
 257                 folio_lock(folio);
 258                 if (unlikely(!folio_test_dirty(folio)))
 259                         NILFS_PAGE_BUG(&folio->page, "inconsistent dirty state");
 260
 261                 dfolio = filemap_grab_folio(dmap, folio->index);
 262                 if (unlikely(IS_ERR(dfolio))) {
 263                         /* No empty page is added to the page cache */
 264                         folio_unlock(folio);
 265                         err = PTR_ERR(dfolio);
 266                         break;
 267                 }
 268                 if (unlikely(!folio_buffers(folio)))
 269                         NILFS_PAGE_BUG(&folio->page,
 270                                        "found empty page in dat page cache");
 271
 272                 nilfs_copy_page(&dfolio->page, &folio->page, 1);
 273                 filemap_dirty_folio(folio_mapping(dfolio), dfolio);
 274
 275                 folio_unlock(dfolio);
 276                 folio_put(dfolio);
 277                 folio_unlock(folio);
 278         }
 279         folio_batch_release(&fbatch);
 280         cond_resched();
 281
 282         if (likely(!err))
 283                 goto repeat;
 284         return err;
 285 }
 286
 287 /**
 288  * nilfs_copy_back_pages -- copy back pages to original cache from shadow cache
 289  * @dmap: destination page cache
 290  * @smap: source page cache
 291  *
 292  * No pages must be added to the cache during this process.
 293  * This must be ensured by the caller.
 294  */
 295 void nilfs_copy_back_pages(struct address_space *dmap,
 296                            struct address_space *smap)
 297 {
 298         struct folio_batch fbatch;
 299         unsigned int i, n;
 300         pgoff_t start = 0;
 301
 302         folio_batch_init(&fbatch);
 303 repeat:
 304         n = filemap_get_folios(smap, &start, ~0UL, &fbatch);
 305         if (!n)
 306                 return;
 307
 308         for (i = 0; i < folio_batch_count(&fbatch); i++) {
 309                 struct folio *folio = fbatch.folios[i], *dfolio;
 310                 pgoff_t index = folio->index;
 311
 312                 folio_lock(folio);
 313                 dfolio = filemap_lock_folio(dmap, index);
 314                 if (!IS_ERR(dfolio)) {
 315                         /* overwrite existing folio in the destination cache */
 316                         WARN_ON(folio_test_dirty(dfolio));
 317                         nilfs_copy_page(&dfolio->page, &folio->page, 0);
 318                         folio_unlock(dfolio);
 319                         folio_put(dfolio);
 320                         /* Do we not need to remove folio from smap here? */
 321                 } else {
 322                         struct folio *f;
 323
 324                         /* move the folio to the destination cache */
 325                         xa_lock_irq(&smap->i_pages);
 326                         f = __xa_erase(&smap->i_pages, index);
 327                         WARN_ON(folio != f);
 328                         smap->nrpages--;
 329                         xa_unlock_irq(&smap->i_pages);
 330
 331                         xa_lock_irq(&dmap->i_pages);
 332                         f = __xa_store(&dmap->i_pages, index, folio, GFP_NOFS);
 333                         if (unlikely(f)) {
 334                                 /* Probably -ENOMEM */
 335                                 folio->mapping = NULL;
 336                                 folio_put(folio);
 337                         } else {
 338                                 folio->mapping = dmap;
 339                                 dmap->nrpages++;
 340                                 if (folio_test_dirty(folio))
 341                                         __xa_set_mark(&dmap->i_pages, index,
 342                                                         PAGECACHE_TAG_DIRTY);
 343                         }
 344                         xa_unlock_irq(&dmap->i_pages);
 345                 }
 346                 folio_unlock(folio);
 347         }
 348         folio_batch_release(&fbatch);
 349         cond_resched();
 350
 351         goto repeat;
 352 }
 353
 354 /**
 355  * nilfs_clear_dirty_pages - discard dirty pages in address space
 356  * @mapping: address space with dirty pages for discarding
 357  * @silent: suppress [true] or print [false] warning messages
 358  */
 359 void nilfs_clear_dirty_pages(struct address_space *mapping, bool silent)
 360 {
 361         struct folio_batch fbatch;
 362         unsigned int i;
 363         pgoff_t index = 0;
 364
 365         folio_batch_init(&fbatch);
 366
 367         while (filemap_get_folios_tag(mapping, &index, (pgoff_t)-1,
 368                                 PAGECACHE_TAG_DIRTY, &fbatch)) {
 369                 for (i = 0; i < folio_batch_count(&fbatch); i++) {
 370                         struct folio *folio = fbatch.folios[i];
 371
 372                         folio_lock(folio);
 373
 374                         /*
 375                          * This folio may have been removed from the address
 376                          * space by truncation or invalidation when the lock
 377                          * was acquired.  Skip processing in that case.
 378                          */
 379                         if (likely(folio->mapping == mapping))
 380                                 nilfs_clear_dirty_page(&folio->page, silent);
 381
 382                         folio_unlock(folio);
 383                 }
 384                 folio_batch_release(&fbatch);
 385                 cond_resched();
 386         }
 387 }
 388
 389 /**
 390  * nilfs_clear_dirty_page - discard dirty page
 391  * @page: dirty page that will be discarded
 392  * @silent: suppress [true] or print [false] warning messages
 393  */
 394 void nilfs_clear_dirty_page(struct page *page, bool silent)
 395 {
 396         struct inode *inode = page->mapping->host;
 397         struct super_block *sb = inode->i_sb;
 398
 399         BUG_ON(!PageLocked(page));
 400
 401         if (!silent)
 402                 nilfs_warn(sb, "discard dirty page: offset=%lld, ino=%lu",
 403                            page_offset(page), inode->i_ino);
 404
 405         ClearPageUptodate(page);
 406         ClearPageMappedToDisk(page);
 407
 408         if (page_has_buffers(page)) {
 409                 struct buffer_head *bh, *head;
 410                 const unsigned long clear_bits =
 411                         (BIT(BH_Uptodate) | BIT(BH_Dirty) | BIT(BH_Mapped) |
 412                          BIT(BH_Async_Write) | BIT(BH_NILFS_Volatile) |
 413                          BIT(BH_NILFS_Checked) | BIT(BH_NILFS_Redirected));
 414
 415                 bh = head = page_buffers(page);
 416                 do {
 417                         lock_buffer(bh);
 418                         if (!silent)
 419                                 nilfs_warn(sb,
 420                                            "discard dirty block: blocknr=%llu, size=%zu",
 421                                            (u64)bh->b_blocknr, bh->b_size);
 422
 423                         set_mask_bits(&bh->b_state, clear_bits, 0);
 424                         unlock_buffer(bh);
 425                 } while (bh = bh->b_this_page, bh != head);
 426         }
 427
 428         __nilfs_clear_page_dirty(page);
 429 }
 430
 431 unsigned int nilfs_page_count_clean_buffers(struct page *page,
 432                                             unsigned int from, unsigned int to)
 433 {
 434         unsigned int block_start, block_end;
 435         struct buffer_head *bh, *head;
 436         unsigned int nc = 0;
 437
 438         for (bh = head = page_buffers(page), block_start = 0;
 439              bh != head || !block_start;
 440              block_start = block_end, bh = bh->b_this_page) {
 441                 block_end = block_start + bh->b_size;
 442                 if (block_end > from && block_start < to && !buffer_dirty(bh))
 443                         nc++;
 444         }
 445         return nc;
 446 }
 447
 448 /*
 449  * NILFS2 needs clear_page_dirty() in the following two cases:
 450  *
 451  * 1) For B-tree node pages and data pages of DAT file, NILFS2 clears dirty
 452  *    flag of pages when it copies back pages from shadow cache to the
 453  *    original cache.
 454  *
 455  * 2) Some B-tree operations like insertion or deletion may dispose buffers
 456  *    in dirty state, and this needs to cancel the dirty state of their pages.
 457  */
 458 int __nilfs_clear_page_dirty(struct page *page)
 459 {
 460         struct address_space *mapping = page->mapping;
 461
 462         if (mapping) {
 463                 xa_lock_irq(&mapping->i_pages);
 464                 if (test_bit(PG_dirty, &page->flags)) {
 465                         __xa_clear_mark(&mapping->i_pages, page_index(page),
 466                                              PAGECACHE_TAG_DIRTY);
 467                         xa_unlock_irq(&mapping->i_pages);
 468                         return clear_page_dirty_for_io(page);
 469                 }
 470                 xa_unlock_irq(&mapping->i_pages);
 471                 return 0;
 472         }
 473         return TestClearPageDirty(page);
 474 }
 475
 476 /**
 477  * nilfs_find_uncommitted_extent - find extent of uncommitted data
 478  * @inode: inode
 479  * @start_blk: start block offset (in)
 480  * @blkoff: start offset of the found extent (out)
 481  *
 482  * This function searches an extent of buffers marked "delayed" which
 483  * starts from a block offset equal to or larger than @start_blk.  If
 484  * such an extent was found, this will store the start offset in
 485  * @blkoff and return its length in blocks.  Otherwise, zero is
 486  * returned.
 487  */
 488 unsigned long nilfs_find_uncommitted_extent(struct inode *inode,
 489                                             sector_t start_blk,
 490                                             sector_t *blkoff)
 491 {
 492         unsigned int i, nr_folios;
 493         pgoff_t index;
 494         unsigned long length = 0;
 495         struct folio_batch fbatch;
 496         struct folio *folio;
 497
 498         if (inode->i_mapping->nrpages == 0)
 499                 return 0;
 500
 501         index = start_blk >> (PAGE_SHIFT - inode->i_blkbits);
 502
 503         folio_batch_init(&fbatch);
 504
 505 repeat:
 506         nr_folios = filemap_get_folios_contig(inode->i_mapping, &index, ULONG_MAX,
 507                         &fbatch);
 508         if (nr_folios == 0)
 509                 return length;
 510
 511         i = 0;
 512         do {
 513                 folio = fbatch.folios[i];
 514
 515                 folio_lock(folio);
 516                 if (folio_buffers(folio)) {
 517                         struct buffer_head *bh, *head;
 518                         sector_t b;
 519
 520                         b = folio->index << (PAGE_SHIFT - inode->i_blkbits);
 521                         bh = head = folio_buffers(folio);
 522                         do {
 523                                 if (b < start_blk)
 524                                         continue;
 525                                 if (buffer_delay(bh)) {
 526                                         if (length == 0)
 527                                                 *blkoff = b;
 528                                         length++;
 529                                 } else if (length > 0) {
 530                                         goto out_locked;
 531                                 }
 532                         } while (++b, bh = bh->b_this_page, bh != head);
 533                 } else {
 534                         if (length > 0)
 535                                 goto out_locked;
 536                 }
 537                 folio_unlock(folio);
 538
 539         } while (++i < nr_folios);
 540
 541         folio_batch_release(&fbatch);
 542         cond_resched();
 543         goto repeat;
 544
 545 out_locked:
 546         folio_unlock(folio);
 547         folio_batch_release(&fbatch);
 548         return length;
 549 }